Merge pull request #1049 from router-for-me/codex

feat(codex): add config toggle for codex instructions injection

.dockerignore

@@ -13,8 +13,6 @@ Dockerfile
 docs/*
 README.md
 README_CN.md
-MANAGEMENT_API.md
-MANAGEMENT_API_CN.md
 LICENSE
 
 # Runtime data folders (should be mounted as volumes)
@@ -25,10 +23,14 @@ config.yaml
 
 # Development/editor
 bin/*
-.claude/*
 .vscode/*
+.claude/*
+.codex/*
 .gemini/*
 .serena/*
 .agent/*
+.agents/*
+.opencode/*
 .bmad/*
 _bmad/*
+_bmad-output/*

.github/ISSUE_TEMPLATE/bug_report.md

@@ -7,6 +7,13 @@ assignees: ''
 
 ---
 
+**Is it a request payload issue?**
+[  ] Yes, this is a request payload issue. I am using a client/cURL to send a request payload, but I received an unexpected error.
+[  ] No, it's another issue.
+
+**If it's a request payload issue, you MUST know**
+Our team doesn't have any GODs or ORACLEs or MIND READERs. Please make sure to attach the request log or curl payload.
+
 **Describe the bug**
 A clear and concise description of what the bug is.
 

.gitignore

@@ -11,11 +11,15 @@ bin/*
 logs/*
 conv/*
 temp/*
+refs/*
+
+# Storage backends
 pgstore/*
 gitstore/*
 objectstore/*
+
+# Static assets
 static/*
-refs/*
 
 # Authentication data
 auths/*
@@ -29,12 +33,17 @@ GEMINI.md
 
 # Tooling metadata
 .vscode/*
+.codex/*
 .claude/*
 .gemini/*
 .serena/*
 .agent/*
+.agents/*
+.agents/*
+.opencode/*
 .bmad/*
 _bmad/*
+_bmad-output/*
 
 # macOS
 .DS_Store

README.md

@@ -10,14 +10,29 @@ So you can use local or multi-account CLI access with OpenAI(include Responses)/
 
 ## Sponsor
 
-[![z.ai](https://assets.router-for.me/english.png)](https://z.ai/subscribe?ic=8JVLJQFSKB)
+[![z.ai](https://assets.router-for.me/english-4.7.png)](https://z.ai/subscribe?ic=8JVLJQFSKB)
 
 This project is sponsored by Z.ai, supporting us with their GLM CODING PLAN.
 
-GLM CODING PLAN is a subscription service designed for AI coding, starting at just $3/month. It provides access to their flagship GLM-4.6 model across 10+ popular AI coding tools (Claude Code, Cline, Roo Code, etc.), offering developers top-tier, fast, and stable coding experiences.
+GLM CODING PLAN is a subscription service designed for AI coding, starting at just $3/month. It provides access to their flagship GLM-4.7 model across 10+ popular AI coding tools (Claude Code, Cline, Roo Code, etc.), offering developers top-tier, fast, and stable coding experiences.
 
 Get 10% OFF GLM CODING PLAN：https://z.ai/subscribe?ic=8JVLJQFSKB
 
+---
+
+<table>
+<tbody>
+<tr>
+<td width="180"><a href="https://www.packyapi.com/register?aff=cliproxyapi"><img src="./assets/packycode.png" alt="PackyCode" width="150"></a></td>
+<td>Thanks to PackyCode for sponsoring this project! PackyCode is a reliable and efficient API relay service provider, offering relay services for Claude Code, Codex, Gemini, and more. PackyCode provides special discounts for our software users: register using <a href="https://www.packyapi.com/register?aff=cliproxyapi">this link</a> and enter the "cliproxyapi" promo code during recharge to get 10% off.</td>
+</tr>
+<tr>
+<td width="180"><a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa"><img src="./assets/cubence.png" alt="Cubence" width="150"></a></td>
+<td>Thanks to Cubence for sponsoring this project! Cubence is a reliable and efficient API relay service provider, offering relay services for Claude Code, Codex, Gemini, and more. Cubence provides special discounts for our software users: register using <a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa">this link</a> and enter the "CLIPROXYAPI" promo code during recharge to get 10% off.</td>
+</tr>
+</tbody>
+</table>
+
 ## Overview
 
 - OpenAI/Gemini/Claude compatible API endpoints for CLI models
@@ -99,9 +114,36 @@ CLI wrapper for instant switching between multiple Claude accounts and alternati
 
 Native macOS GUI for managing CLIProxyAPI: configure providers, model mappings, and endpoints via OAuth - no API keys needed.
 
+### [Quotio](https://github.com/nguyenphutrong/quotio)
+
+Native macOS menu bar app that unifies Claude, Gemini, OpenAI, Qwen, and Antigravity subscriptions with real-time quota tracking and smart auto-failover for AI coding tools like Claude Code, OpenCode, and Droid - no API keys needed.
+
+### [CodMate](https://github.com/loocor/CodMate)
+
+Native macOS SwiftUI app for managing CLI AI sessions (Codex, Claude Code, Gemini CLI) with unified provider management, Git review, project organization, global search, and terminal integration. Integrates CLIProxyAPI to provide OAuth authentication for Codex, Claude, Gemini, Antigravity, and Qwen Code, with built-in and third-party provider rerouting through a single proxy endpoint - no API keys needed for OAuth providers.
+
+### [ProxyPilot](https://github.com/Finesssee/ProxyPilot)
+
+Windows-native CLIProxyAPI fork with TUI, system tray, and multi-provider OAuth for AI coding tools - no API keys needed.
+
+### [Claude Proxy VSCode](https://github.com/uzhao/claude-proxy-vscode)
+
+VSCode extension for quick switching between Claude Code models, featuring integrated CLIProxyAPI as its backend with automatic background lifecycle management.
+
 > [!NOTE]  
 > If you developed a project based on CLIProxyAPI, please open a PR to add it to this list.
 
+## More choices
+
+Those projects are ports of CLIProxyAPI or inspired by it:
+
+### [9Router](https://github.com/decolua/9router)
+
+A Next.js implementation inspired by CLIProxyAPI, easy to install and use, built from scratch with format translation (OpenAI/Claude/Gemini/Ollama), combo system with auto-fallback, multi-account management with exponential backoff, a Next.js web dashboard, and support for CLI tools (Cursor, Claude Code, Cline, RooCode) - no API keys needed.
+
+> [!NOTE]  
+> If you have developed a port of CLIProxyAPI or a project inspired by it, please open a PR to add it to this list.
+
 ## License
 
 This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.

README_CN.md

@@ -10,14 +10,30 @@
 
 ## 赞助商
 
-[![bigmodel.cn](https://assets.router-for.me/chinese.png)](https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII)
+[![bigmodel.cn](https://assets.router-for.me/chinese-4.7.png)](https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII)
 
 本项目由 Z智谱 提供赞助, 他们通过 GLM CODING PLAN 对本项目提供技术支持。
 
-GLM CODING PLAN 是专为AI编码打造的订阅套餐，每月最低仅需20元，即可在十余款主流AI编码工具如 Claude Code、Cline、Roo Code 中畅享智谱旗舰模型GLM-4.6，为开发者提供顶尖的编码体验。
+GLM CODING PLAN 是专为AI编码打造的订阅套餐，每月最低仅需20元，即可在十余款主流AI编码工具如 Claude Code、Cline、Roo Code 中畅享智谱旗舰模型GLM-4.7，为开发者提供顶尖的编码体验。
 
 智谱AI为本软件提供了特别优惠，使用以下链接购买可以享受九折优惠：https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII
 
+---
+
+<table>
+<tbody>
+<tr>
+<td width="180"><a href="https://www.packyapi.com/register?aff=cliproxyapi"><img src="./assets/packycode.png" alt="PackyCode" width="150"></a></td>
+<td>感谢 PackyCode 对本项目的赞助！PackyCode 是一家可靠高效的 API 中转服务商，提供 Claude Code、Codex、Gemini 等多种服务的中转。PackyCode 为本软件用户提供了特别优惠：使用<a href="https://www.packyapi.com/register?aff=cliproxyapi">此链接</a>注册，并在充值时输入 "cliproxyapi" 优惠码即可享受九折优惠。</td>
+</tr>
+<tr>
+<td width="180"><a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa"><img src="./assets/cubence.png" alt="Cubence" width="150"></a></td>
+<td>感谢 Cubence 对本项目的赞助！Cubence 是一家可靠高效的 API 中转服务商，提供 Claude Code、Codex、Gemini 等多种服务的中转。Cubence 为本软件用户提供了特别优惠：使用<a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa">此链接</a>注册，并在充值时输入 "CLIPROXYAPI" 优惠码即可享受九折优惠。</td>
+</tr>
+</tbody>
+</table>
+
+
 ## 功能特性
 
 - 为 CLI 模型提供 OpenAI/Gemini/Claude/Codex 兼容的 API 端点
@@ -97,9 +113,36 @@ CLI 封装器，用于通过 CLIProxyAPI OAuth 即时切换多个 Claude 账户
 
 基于 macOS 平台的原生 CLIProxyAPI GUI：配置供应商、模型映射以及OAuth端点，无需 API 密钥。
 
+### [Quotio](https://github.com/nguyenphutrong/quotio)
+
+原生 macOS 菜单栏应用，统一管理 Claude、Gemini、OpenAI、Qwen 和 Antigravity 订阅，提供实时配额追踪和智能自动故障转移，支持 Claude Code、OpenCode 和 Droid 等 AI 编程工具，无需 API 密钥。
+
+### [CodMate](https://github.com/loocor/CodMate)
+
+原生 macOS SwiftUI 应用，用于管理 CLI AI 会话（Claude Code、Codex、Gemini CLI），提供统一的提供商管理、Git 审查、项目组织、全局搜索和终端集成。集成 CLIProxyAPI 为 Codex、Claude、Gemini、Antigravity 和 Qwen Code 提供统一的 OAuth 认证，支持内置和第三方提供商通过单一代理端点重路由 - OAuth 提供商无需 API 密钥。
+
+### [ProxyPilot](https://github.com/Finesssee/ProxyPilot)
+
+原生 Windows CLIProxyAPI 分支，集成 TUI、系统托盘及多服务商 OAuth 认证，专为 AI 编程工具打造，无需 API 密钥。
+
+### [Claude Proxy VSCode](https://github.com/uzhao/claude-proxy-vscode)
+
+一款 VSCode 扩展，提供了在 VSCode 中快速切换 Claude Code 模型的功能，内置 CLIProxyAPI 作为其后端，支持后台自动启动和关闭。
+
 > [!NOTE]  
 > 如果你开发了基于 CLIProxyAPI 的项目，请提交一个 PR（拉取请求）将其添加到此列表中。
 
+## 更多选择
+
+以下项目是 CLIProxyAPI 的移植版或受其启发：
+
+### [9Router](https://github.com/decolua/9router)
+
+基于 Next.js 的实现，灵感来自 CLIProxyAPI，易于安装使用；自研格式转换（OpenAI/Claude/Gemini/Ollama）、组合系统与自动回退、多账户管理（指数退避）、Next.js Web 控制台，并支持 Cursor、Claude Code、Cline、RooCode 等 CLI 工具，无需 API 密钥。
+
+> [!NOTE]  
+> 如果你开发了 CLIProxyAPI 的移植或衍生项目，请提交 PR 将其添加到此列表中。
+
 ## 许可证
 
 此项目根据 MIT 许可证授权 - 有关详细信息，请参阅 [LICENSE](LICENSE) 文件。

cmd/server/main.go

@@ -61,6 +61,7 @@ func main() {
 	var iflowLogin bool
 	var iflowCookie bool
 	var noBrowser bool
+	var oauthCallbackPort int
 	var antigravityLogin bool
 	var projectID string
 	var vertexImport string
@@ -75,6 +76,7 @@ func main() {
 	flag.BoolVar(&iflowLogin, "iflow-login", false, "Login to iFlow using OAuth")
 	flag.BoolVar(&iflowCookie, "iflow-cookie", false, "Login to iFlow using Cookie")
 	flag.BoolVar(&noBrowser, "no-browser", false, "Don't open browser automatically for OAuth")
+	flag.IntVar(&oauthCallbackPort, "oauth-callback-port", 0, "Override OAuth callback port (defaults to provider-specific port)")
 	flag.BoolVar(&antigravityLogin, "antigravity-login", false, "Login to Antigravity using OAuth")
 	flag.StringVar(&projectID, "project_id", "", "Project ID (Gemini only, not required)")
 	flag.StringVar(&configPath, "config", DefaultConfigPath, "Configure File Path")
@@ -405,7 +407,7 @@ func main() {
 	usage.SetStatisticsEnabled(cfg.UsageStatisticsEnabled)
 	coreauth.SetQuotaCooldownDisabled(cfg.DisableCooling)
 
-	if err = logging.ConfigureLogOutput(cfg.LoggingToFile, cfg.LogsMaxTotalSizeMB); err != nil {
+	if err = logging.ConfigureLogOutput(cfg); err != nil {
 		log.Errorf("failed to configure log output: %v", err)
 		return
 	}
@@ -425,7 +427,8 @@ func main() {
 
 	// Create login options to be used in authentication flows.
 	options := &cmd.LoginOptions{
-		NoBrowser: noBrowser,
+		NoBrowser:    noBrowser,
+		CallbackPort: oauthCallbackPort,
 	}
 
 	// Register the shared token store once so all components use the same persistence backend.

config.example.yaml

@@ -35,10 +35,14 @@ auth-dir: "~/.cli-proxy-api"
 api-keys:
   - "your-api-key-1"
   - "your-api-key-2"
+  - "your-api-key-3"
 
 # Enable debug logging
 debug: false
 
+# When true, disable high-overhead HTTP middleware features to reduce per-request memory usage under high concurrency.
+commercial-mode: false
+
 # When true, write application logs to rotating files instead of stdout
 logging-to-file: false
 
@@ -66,9 +70,25 @@ quota-exceeded:
   switch-project: true # Whether to automatically switch to another project when a quota is exceeded
   switch-preview-model: true # Whether to automatically switch to a preview model when a quota is exceeded
 
+# Routing strategy for selecting credentials when multiple match.
+routing:
+  strategy: "round-robin" # round-robin (default), fill-first
+
 # When true, enable authentication for the WebSocket API (/v1/ws).
 ws-auth: false
 
+# When > 0, emit blank lines every N seconds for non-streaming responses to prevent idle timeouts.
+nonstream-keepalive-interval: 0
+
+# Streaming behavior (SSE keep-alives + safe bootstrap retries).
+# streaming:
+#   keepalive-seconds: 15   # Default: 0 (disabled). <= 0 disables keep-alives.
+#   bootstrap-retries: 1    # Default: 0 (disabled). Retries before first byte is sent.
+
+# When true, enable custom Codex instructions injection for Codex API requests.
+# When false (default), CodexInstructionsForModel returns immediately without modification.
+codex-instructions-enabled: false
+
 # Gemini API keys
 # gemini-api-key:
 #   - api-key: "AIzaSy...01"
@@ -77,6 +97,9 @@ ws-auth: false
 #     headers:
 #       X-Custom-Header: "custom-value"
 #     proxy-url: "socks5://proxy.example.com:1080"
+#     models:
+#       - name: "gemini-2.5-flash" # upstream model name
+#         alias: "gemini-flash"    # client alias mapped to the upstream model
 #     excluded-models:
 #       - "gemini-2.5-pro"     # exclude specific models from this provider (exact match)
 #       - "gemini-2.5-*"       # wildcard matching prefix (e.g. gemini-2.5-flash, gemini-2.5-pro)
@@ -92,6 +115,9 @@ ws-auth: false
 #     headers:
 #       X-Custom-Header: "custom-value"
 #     proxy-url: "socks5://proxy.example.com:1080" # optional: per-key proxy override
+#     models:
+#       - name: "gpt-5-codex"   # upstream model name
+#         alias: "codex-latest" # client alias mapped to the upstream model
 #     excluded-models:
 #       - "gpt-5.1"         # exclude specific models (exact match)
 #       - "gpt-5-*"         # wildcard matching prefix (e.g. gpt-5-medium, gpt-5-codex)
@@ -109,7 +135,7 @@ ws-auth: false
 #     proxy-url: "socks5://proxy.example.com:1080" # optional: per-key proxy override
 #     models:
 #       - name: "claude-3-5-sonnet-20241022" # upstream model name
-#         alias: "claude-sonnet-latest" # client alias mapped to the upstream model
+#         alias: "claude-sonnet-latest"      # client alias mapped to the upstream model
 #     excluded-models:
 #       - "claude-opus-4-5-20251101" # exclude specific models (exact match)
 #       - "claude-3-*"               # wildcard matching prefix (e.g. claude-3-7-sonnet-20250219)
@@ -140,9 +166,9 @@ ws-auth: false
 #     headers:
 #       X-Custom-Header: "custom-value"
 #     models:                                     # optional: map aliases to upstream model names
-#       - name: "gemini-2.0-flash"                # upstream model name
+#       - name: "gemini-2.5-flash"                # upstream model name
 #         alias: "vertex-flash"                   # client-visible alias
-#       - name: "gemini-1.5-pro"
+#       - name: "gemini-2.5-pro"
 #         alias: "vertex-pro"
 
 # Amp Integration
@@ -151,6 +177,18 @@ ws-auth: false
 #   upstream-url: "https://ampcode.com"
 #   # Optional: Override API key for Amp upstream (otherwise uses env or file)
 #   upstream-api-key: ""
+#   # Per-client upstream API key mapping
+#   # Maps client API keys (from top-level api-keys) to different Amp upstream API keys.
+#   # Useful when different clients need to use different Amp accounts/quotas.
+#   # If a client key isn't mapped, falls back to upstream-api-key (default behavior).
+#   upstream-api-keys:
+#     - upstream-api-key: "amp_key_for_team_a"    # Upstream key to use for these clients
+#       api-keys:                                 # Client keys that use this upstream key
+#         - "your-api-key-1"
+#         - "your-api-key-2"
+#     - upstream-api-key: "amp_key_for_team_b"
+#       api-keys:
+#         - "your-api-key-3"
 #   # Restrict Amp management routes (/api/auth, /api/user, etc.) to localhost only (default: false)
 #   restrict-management-to-localhost: false
 #   # Force model mappings to run before checking local API keys (default: false)
@@ -160,12 +198,56 @@ ws-auth: false
 #   # Useful when Amp CLI requests models you don't have access to (e.g., Claude Opus 4.5)
 #   # but you have a similar model available (e.g., Claude Sonnet 4).
 #   model-mappings:
-#     - from: "claude-opus-4.5"       # Model requested by Amp CLI
-#       to: "claude-sonnet-4"         # Route to this available model instead
-#     - from: "gpt-5"
-#       to: "gemini-2.5-pro"
-#     - from: "claude-3-opus-20240229"
-#       to: "claude-3-5-sonnet-20241022"
+#     - from: "claude-opus-4-5-20251101"          # Model requested by Amp CLI
+#       to: "gemini-claude-opus-4-5-thinking"     # Route to this available model instead
+#     - from: "claude-sonnet-4-5-20250929"
+#       to: "gemini-claude-sonnet-4-5-thinking"
+#     - from: "claude-haiku-4-5-20251001"
+#       to: "gemini-2.5-flash"
+
+# Global OAuth model name aliases (per channel)
+# These aliases rename model IDs for both model listing and request routing.
+# Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow.
+# NOTE: Aliases do not apply to gemini-api-key, codex-api-key, claude-api-key, openai-compatibility, vertex-api-key, or ampcode.
+# You can repeat the same name with different aliases to expose multiple client model names.
+oauth-model-alias:
+  antigravity:
+    - name: "rev19-uic3-1p"
+      alias: "gemini-2.5-computer-use-preview-10-2025"
+    - name: "gemini-3-pro-image"
+      alias: "gemini-3-pro-image-preview"
+    - name: "gemini-3-pro-high"
+      alias: "gemini-3-pro-preview"
+    - name: "gemini-3-flash"
+      alias: "gemini-3-flash-preview"
+    - name: "claude-sonnet-4-5"
+      alias: "gemini-claude-sonnet-4-5"
+    - name: "claude-sonnet-4-5-thinking"
+      alias: "gemini-claude-sonnet-4-5-thinking"
+    - name: "claude-opus-4-5-thinking"
+      alias: "gemini-claude-opus-4-5-thinking"
+#   gemini-cli:
+#     - name: "gemini-2.5-pro"          # original model name under this channel
+#       alias: "g2.5p"                  # client-visible alias
+#       fork: true                      # when true, keep original and also add the alias as an extra model (default: false)
+#   vertex:
+#     - name: "gemini-2.5-pro"
+#       alias: "g2.5p"
+#   aistudio:
+#     - name: "gemini-2.5-pro"
+#       alias: "g2.5p"
+#   claude:
+#     - name: "claude-sonnet-4-5-20250929"
+#       alias: "cs4.5"
+#   codex:
+#     - name: "gpt-5"
+#       alias: "g5"
+#   qwen:
+#     - name: "qwen3-coder-plus"
+#       alias: "qwen-plus"
+#   iflow:
+#     - name: "glm-4.7"
+#       alias: "glm-god"
 
 # OAuth provider excluded models
 # oauth-excluded-models:
@@ -197,9 +279,21 @@ ws-auth: false
 #           protocol: "gemini" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex
 #       params: # JSON path (gjson/sjson syntax) -> value
 #         "generationConfig.thinkingConfig.thinkingBudget": 32768
+#   default-raw: # Default raw rules set parameters using raw JSON when missing (must be valid JSON).
+#     - models:
+#         - name: "gemini-2.5-pro" # Supports wildcards (e.g., "gemini-*")
+#           protocol: "gemini" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex
+#       params: # JSON path (gjson/sjson syntax) -> raw JSON value (strings are used as-is, must be valid JSON)
+#         "generationConfig.responseJsonSchema": "{\"type\":\"object\",\"properties\":{\"answer\":{\"type\":\"string\"}}}"
 #   override: # Override rules always set parameters, overwriting any existing values.
 #     - models:
 #         - name: "gpt-*" # Supports wildcards (e.g., "gpt-*")
 #           protocol: "codex" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex
 #       params: # JSON path (gjson/sjson syntax) -> value
 #         "reasoning.effort": "high"
+#   override-raw: # Override raw rules always set parameters using raw JSON (must be valid JSON).
+#     - models:
+#         - name: "gpt-*" # Supports wildcards (e.g., "gpt-*")
+#           protocol: "codex" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex
+#       params: # JSON path (gjson/sjson syntax) -> raw JSON value (strings are used as-is, must be valid JSON)
+#         "response_format": "{\"type\":\"json_schema\",\"json_schema\":{\"name\":\"answer\",\"schema\":{\"type\":\"object\"}}}"

docker-build.sh

@@ -5,9 +5,115 @@
 # This script automates the process of building and running the Docker container
 # with version information dynamically injected at build time.
 
-# Exit immediately if a command exits with a non-zero status.
+# Hidden feature: Preserve usage statistics across rebuilds
+# Usage: ./docker-build.sh --with-usage
+# First run prompts for management API key, saved to temp/stats/.api_secret
+
 set -euo pipefail
 
+STATS_DIR="temp/stats"
+STATS_FILE="${STATS_DIR}/.usage_backup.json"
+SECRET_FILE="${STATS_DIR}/.api_secret"
+WITH_USAGE=false
+
+get_port() {
+  if [[ -f "config.yaml" ]]; then
+    grep -E "^port:" config.yaml | sed -E 's/^port: *["'"'"']?([0-9]+)["'"'"']?.*$/\1/'
+  else
+    echo "8317"
+  fi
+}
+
+export_stats_api_secret() {
+  if [[ -f "${SECRET_FILE}" ]]; then
+    API_SECRET=$(cat "${SECRET_FILE}")
+  else
+    if [[ ! -d "${STATS_DIR}" ]]; then
+      mkdir -p "${STATS_DIR}"
+    fi
+    echo "First time using --with-usage. Management API key required."
+    read -r -p "Enter management key: " -s API_SECRET
+    echo
+    echo "${API_SECRET}" > "${SECRET_FILE}"
+    chmod 600 "${SECRET_FILE}"
+  fi
+}
+
+check_container_running() {
+  local port
+  port=$(get_port)
+
+  if ! curl -s -o /dev/null -w "%{http_code}" "http://localhost:${port}/" | grep -q "200"; then
+    echo "Error: cli-proxy-api service is not responding at localhost:${port}"
+    echo "Please start the container first or use without --with-usage flag."
+    exit 1
+  fi
+}
+
+export_stats() {
+  local port
+  port=$(get_port)
+
+  if [[ ! -d "${STATS_DIR}" ]]; then
+    mkdir -p "${STATS_DIR}"
+  fi
+  check_container_running
+  echo "Exporting usage statistics..."
+  EXPORT_RESPONSE=$(curl -s -w "\n%{http_code}" -H "X-Management-Key: ${API_SECRET}" \
+    "http://localhost:${port}/v0/management/usage/export")
+  HTTP_CODE=$(echo "${EXPORT_RESPONSE}" | tail -n1)
+  RESPONSE_BODY=$(echo "${EXPORT_RESPONSE}" | sed '$d')
+
+  if [[ "${HTTP_CODE}" != "200" ]]; then
+    echo "Export failed (HTTP ${HTTP_CODE}): ${RESPONSE_BODY}"
+    exit 1
+  fi
+
+  echo "${RESPONSE_BODY}" > "${STATS_FILE}"
+  echo "Statistics exported to ${STATS_FILE}"
+}
+
+import_stats() {
+  local port
+  port=$(get_port)
+
+  echo "Importing usage statistics..."
+  IMPORT_RESPONSE=$(curl -s -w "\n%{http_code}" -X POST \
+    -H "X-Management-Key: ${API_SECRET}" \
+    -H "Content-Type: application/json" \
+    -d @"${STATS_FILE}" \
+    "http://localhost:${port}/v0/management/usage/import")
+  IMPORT_CODE=$(echo "${IMPORT_RESPONSE}" | tail -n1)
+  IMPORT_BODY=$(echo "${IMPORT_RESPONSE}" | sed '$d')
+
+  if [[ "${IMPORT_CODE}" == "200" ]]; then
+    echo "Statistics imported successfully"
+  else
+    echo "Import failed (HTTP ${IMPORT_CODE}): ${IMPORT_BODY}"
+  fi
+
+  rm -f "${STATS_FILE}"
+}
+
+wait_for_service() {
+  local port
+  port=$(get_port)
+
+  echo "Waiting for service to be ready..."
+  for i in {1..30}; do
+    if curl -s -o /dev/null -w "%{http_code}" "http://localhost:${port}/" | grep -q "200"; then
+      break
+    fi
+    sleep 1
+  done
+  sleep 2
+}
+
+if [[ "${1:-}" == "--with-usage" ]]; then
+  WITH_USAGE=true
+  export_stats_api_secret
+fi
+
 # --- Step 1: Choose Environment ---
 echo "Please select an option:"
 echo "1) Run using Pre-built Image (Recommended)"
@@ -18,7 +124,14 @@ read -r -p "Enter choice [1-2]: " choice
 case "$choice" in
   1)
     echo "--- Running with Pre-built Image ---"
+    if [[ "${WITH_USAGE}" == "true" ]]; then
+      export_stats
+    fi
     docker compose up -d --remove-orphans --no-build
+    if [[ "${WITH_USAGE}" == "true" ]]; then
+      wait_for_service
+      import_stats
+    fi
     echo "Services are starting from remote image."
     echo "Run 'docker compose logs -f' to see the logs."
     ;;
@@ -38,16 +151,25 @@ case "$choice" in
 
     # Build and start the services with a local-only image tag
     export CLI_PROXY_IMAGE="cli-proxy-api:local"
-    
+
     echo "Building the Docker image..."
     docker compose build \
       --build-arg VERSION="${VERSION}" \
       --build-arg COMMIT="${COMMIT}" \
       --build-arg BUILD_DATE="${BUILD_DATE}"
 
+    if [[ "${WITH_USAGE}" == "true" ]]; then
+      export_stats
+    fi
+
     echo "Starting the services..."
     docker compose up -d --remove-orphans --pull never
 
+    if [[ "${WITH_USAGE}" == "true" ]]; then
+      wait_for_service
+      import_stats
+    fi
+
     echo "Build complete. Services are starting."
     echo "Run 'docker compose logs -f' to see the logs."
     ;;
@@ -55,4 +177,4 @@ case "$choice" in
     echo "Invalid choice. Please enter 1 or 2."
     exit 1
     ;;
-esac
+esac

docker-compose.yml

@@ -22,7 +22,7 @@ services:
       - "51121:51121"
       - "11451:11451"
     volumes:
-      - ./config.yaml:/CLIProxyAPI/config.yaml
-      - ./auths:/root/.cli-proxy-api
-      - ./logs:/CLIProxyAPI/logs
+      - ${CLI_PROXY_CONFIG_PATH:-./config.yaml}:/CLIProxyAPI/config.yaml
+      - ${CLI_PROXY_AUTH_PATH:-./auths}:/root/.cli-proxy-api
+      - ${CLI_PROXY_LOG_PATH:-./logs}:/CLIProxyAPI/logs
     restart: unless-stopped

examples/custom-provider/main.go

@@ -14,6 +14,7 @@ import (
 	"bytes"
 	"context"
 	"errors"
+	"fmt"
 	"io"
 	"net/http"
 	"net/url"
@@ -122,7 +123,9 @@ func (MyExecutor) Execute(ctx context.Context, a *coreauth.Auth, req clipexec.Re
 	httpReq.Header.Set("Content-Type", "application/json")
 
 	// Inject credentials via PrepareRequest hook.
-	_ = (MyExecutor{}).PrepareRequest(httpReq, a)
+	if errPrep := (MyExecutor{}).PrepareRequest(httpReq, a); errPrep != nil {
+		return clipexec.Response{}, errPrep
+	}
 
 	resp, errDo := client.Do(httpReq)
 	if errDo != nil {
@@ -130,13 +133,28 @@ func (MyExecutor) Execute(ctx context.Context, a *coreauth.Auth, req clipexec.Re
 	}
 	defer func() {
 		if errClose := resp.Body.Close(); errClose != nil {
-			// Best-effort close; log if needed in real projects.
+			fmt.Fprintf(os.Stderr, "close response body error: %v\n", errClose)
 		}
 	}()
 	body, _ := io.ReadAll(resp.Body)
 	return clipexec.Response{Payload: body}, nil
 }
 
+func (MyExecutor) HttpRequest(ctx context.Context, a *coreauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("myprov executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if errPrep := (MyExecutor{}).PrepareRequest(httpReq, a); errPrep != nil {
+		return nil, errPrep
+	}
+	client := buildHTTPClient(a)
+	return client.Do(httpReq)
+}
+
 func (MyExecutor) CountTokens(context.Context, *coreauth.Auth, clipexec.Request, clipexec.Options) (clipexec.Response, error) {
 	return clipexec.Response{}, errors.New("count tokens not implemented")
 }
@@ -199,8 +217,8 @@ func main() {
 	ctx, cancel := context.WithCancel(context.Background())
 	defer cancel()
 
-	if err := svc.Run(ctx); err != nil && !errors.Is(err, context.Canceled) {
-		panic(err)
+	if errRun := svc.Run(ctx); errRun != nil && !errors.Is(errRun, context.Canceled) {
+		panic(errRun)
 	}
 	_ = os.Stderr // keep os import used (demo only)
 	_ = time.Second

examples/http-request/main.go

@@ -0,0 +1,140 @@
+// Package main demonstrates how to use coreauth.Manager.HttpRequest/NewHttpRequest
+// to execute arbitrary HTTP requests with provider credentials injected.
+//
+// This example registers a minimal custom executor that injects an Authorization
+// header from auth.Attributes["api_key"], then performs two requests against
+// httpbin.org to show the injected headers.
+package main
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	clipexec "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+	log "github.com/sirupsen/logrus"
+)
+
+const providerKey = "echo"
+
+// EchoExecutor is a minimal provider implementation for demonstration purposes.
+type EchoExecutor struct{}
+
+func (EchoExecutor) Identifier() string { return providerKey }
+
+func (EchoExecutor) PrepareRequest(req *http.Request, auth *coreauth.Auth) error {
+	if req == nil || auth == nil {
+		return nil
+	}
+	if auth.Attributes != nil {
+		if apiKey := strings.TrimSpace(auth.Attributes["api_key"]); apiKey != "" {
+			req.Header.Set("Authorization", "Bearer "+apiKey)
+		}
+	}
+	return nil
+}
+
+func (EchoExecutor) HttpRequest(ctx context.Context, auth *coreauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("echo executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if errPrep := (EchoExecutor{}).PrepareRequest(httpReq, auth); errPrep != nil {
+		return nil, errPrep
+	}
+	return http.DefaultClient.Do(httpReq)
+}
+
+func (EchoExecutor) Execute(context.Context, *coreauth.Auth, clipexec.Request, clipexec.Options) (clipexec.Response, error) {
+	return clipexec.Response{}, errors.New("echo executor: Execute not implemented")
+}
+
+func (EchoExecutor) ExecuteStream(context.Context, *coreauth.Auth, clipexec.Request, clipexec.Options) (<-chan clipexec.StreamChunk, error) {
+	return nil, errors.New("echo executor: ExecuteStream not implemented")
+}
+
+func (EchoExecutor) Refresh(context.Context, *coreauth.Auth) (*coreauth.Auth, error) {
+	return nil, errors.New("echo executor: Refresh not implemented")
+}
+
+func (EchoExecutor) CountTokens(context.Context, *coreauth.Auth, clipexec.Request, clipexec.Options) (clipexec.Response, error) {
+	return clipexec.Response{}, errors.New("echo executor: CountTokens not implemented")
+}
+
+func main() {
+	log.SetLevel(log.InfoLevel)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	core := coreauth.NewManager(nil, nil, nil)
+	core.RegisterExecutor(EchoExecutor{})
+
+	auth := &coreauth.Auth{
+		ID:       "demo-echo",
+		Provider: providerKey,
+		Attributes: map[string]string{
+			"api_key": "demo-api-key",
+		},
+	}
+
+	// Example 1: Build a prepared request and execute it using your own http.Client.
+	reqPrepared, errReqPrepared := core.NewHttpRequest(
+		ctx,
+		auth,
+		http.MethodGet,
+		"https://httpbin.org/anything",
+		nil,
+		http.Header{"X-Example": []string{"prepared"}},
+	)
+	if errReqPrepared != nil {
+		panic(errReqPrepared)
+	}
+	respPrepared, errDoPrepared := http.DefaultClient.Do(reqPrepared)
+	if errDoPrepared != nil {
+		panic(errDoPrepared)
+	}
+	defer func() {
+		if errClose := respPrepared.Body.Close(); errClose != nil {
+			log.Errorf("close response body error: %v", errClose)
+		}
+	}()
+	bodyPrepared, errReadPrepared := io.ReadAll(respPrepared.Body)
+	if errReadPrepared != nil {
+		panic(errReadPrepared)
+	}
+	fmt.Printf("Prepared request status: %d\n%s\n\n", respPrepared.StatusCode, bodyPrepared)
+
+	// Example 2: Execute a raw request via core.HttpRequest (auto inject + do).
+	rawBody := []byte(`{"hello":"world"}`)
+	rawReq, errRawReq := http.NewRequestWithContext(ctx, http.MethodPost, "https://httpbin.org/anything", bytes.NewReader(rawBody))
+	if errRawReq != nil {
+		panic(errRawReq)
+	}
+	rawReq.Header.Set("Content-Type", "application/json")
+	rawReq.Header.Set("X-Example", "executed")
+
+	respExec, errDoExec := core.HttpRequest(ctx, auth, rawReq)
+	if errDoExec != nil {
+		panic(errDoExec)
+	}
+	defer func() {
+		if errClose := respExec.Body.Close(); errClose != nil {
+			log.Errorf("close response body error: %v", errClose)
+		}
+	}()
+	bodyExec, errReadExec := io.ReadAll(respExec.Body)
+	if errReadExec != nil {
+		panic(errReadExec)
+	}
+	fmt.Printf("Manager HttpRequest status: %d\n%s\n", respExec.StatusCode, bodyExec)
+}

internal/api/handlers/management/api_tools.go

@@ -0,0 +1,704 @@
+package management
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/url"
+	"strings"
+	"time"
+
+	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/runtime/geminicli"
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/net/proxy"
+	"golang.org/x/oauth2"
+	"golang.org/x/oauth2/google"
+)
+
+const defaultAPICallTimeout = 60 * time.Second
+
+const (
+	geminiOAuthClientID     = "681255809395-oo8ft2oprdrnp9e3aqf6av3hmdib135j.apps.googleusercontent.com"
+	geminiOAuthClientSecret = "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"
+)
+
+var geminiOAuthScopes = []string{
+	"https://www.googleapis.com/auth/cloud-platform",
+	"https://www.googleapis.com/auth/userinfo.email",
+	"https://www.googleapis.com/auth/userinfo.profile",
+}
+
+const (
+	antigravityOAuthClientID     = "1071006060591-tmhssin2h21lcre235vtolojh4g403ep.apps.googleusercontent.com"
+	antigravityOAuthClientSecret = "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"
+)
+
+var antigravityOAuthTokenURL = "https://oauth2.googleapis.com/token"
+
+type apiCallRequest struct {
+	AuthIndexSnake  *string           `json:"auth_index"`
+	AuthIndexCamel  *string           `json:"authIndex"`
+	AuthIndexPascal *string           `json:"AuthIndex"`
+	Method          string            `json:"method"`
+	URL             string            `json:"url"`
+	Header          map[string]string `json:"header"`
+	Data            string            `json:"data"`
+}
+
+type apiCallResponse struct {
+	StatusCode int                 `json:"status_code"`
+	Header     map[string][]string `json:"header"`
+	Body       string              `json:"body"`
+}
+
+// APICall makes a generic HTTP request on behalf of the management API caller.
+// It is protected by the management middleware.
+//
+// Endpoint:
+//
+//	POST /v0/management/api-call
+//
+// Authentication:
+//
+//	Same as other management APIs (requires a management key and remote-management rules).
+//	You can provide the key via:
+//	- Authorization: Bearer <key>
+//	- X-Management-Key: <key>
+//
+// Request JSON:
+//   - auth_index / authIndex / AuthIndex (optional):
+//     The credential "auth_index" from GET /v0/management/auth-files (or other endpoints returning it).
+//     If omitted or not found, credential-specific proxy/token substitution is skipped.
+//   - method (required): HTTP method, e.g. GET, POST, PUT, PATCH, DELETE.
+//   - url (required): Absolute URL including scheme and host, e.g. "https://api.example.com/v1/ping".
+//   - header (optional): Request headers map.
+//     Supports magic variable "$TOKEN$" which is replaced using the selected credential:
+//     1) metadata.access_token
+//     2) attributes.api_key
+//     3) metadata.token / metadata.id_token / metadata.cookie
+//     Example: {"Authorization":"Bearer $TOKEN$"}.
+//     Note: if you need to override the HTTP Host header, set header["Host"].
+//   - data (optional): Raw request body as string (useful for POST/PUT/PATCH).
+//
+// Proxy selection (highest priority first):
+//  1. Selected credential proxy_url
+//  2. Global config proxy-url
+//  3. Direct connect (environment proxies are not used)
+//
+// Response JSON (returned with HTTP 200 when the APICall itself succeeds):
+//   - status_code: Upstream HTTP status code.
+//   - header: Upstream response headers.
+//   - body: Upstream response body as string.
+//
+// Example:
+//
+//	curl -sS -X POST "http://127.0.0.1:8317/v0/management/api-call" \
+//	  -H "Authorization: Bearer <MANAGEMENT_KEY>" \
+//	  -H "Content-Type: application/json" \
+//	  -d '{"auth_index":"<AUTH_INDEX>","method":"GET","url":"https://api.example.com/v1/ping","header":{"Authorization":"Bearer $TOKEN$"}}'
+//
+//	curl -sS -X POST "http://127.0.0.1:8317/v0/management/api-call" \
+//	  -H "Authorization: Bearer 831227" \
+//	  -H "Content-Type: application/json" \
+//	  -d '{"auth_index":"<AUTH_INDEX>","method":"POST","url":"https://api.example.com/v1/fetchAvailableModels","header":{"Authorization":"Bearer $TOKEN$","Content-Type":"application/json","User-Agent":"cliproxyapi"},"data":"{}"}'
+func (h *Handler) APICall(c *gin.Context) {
+	var body apiCallRequest
+	if errBindJSON := c.ShouldBindJSON(&body); errBindJSON != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid body"})
+		return
+	}
+
+	method := strings.ToUpper(strings.TrimSpace(body.Method))
+	if method == "" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "missing method"})
+		return
+	}
+
+	urlStr := strings.TrimSpace(body.URL)
+	if urlStr == "" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "missing url"})
+		return
+	}
+	parsedURL, errParseURL := url.Parse(urlStr)
+	if errParseURL != nil || parsedURL.Scheme == "" || parsedURL.Host == "" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid url"})
+		return
+	}
+
+	authIndex := firstNonEmptyString(body.AuthIndexSnake, body.AuthIndexCamel, body.AuthIndexPascal)
+	auth := h.authByIndex(authIndex)
+
+	reqHeaders := body.Header
+	if reqHeaders == nil {
+		reqHeaders = map[string]string{}
+	}
+
+	var hostOverride string
+	var token string
+	var tokenResolved bool
+	var tokenErr error
+	for key, value := range reqHeaders {
+		if !strings.Contains(value, "$TOKEN$") {
+			continue
+		}
+		if !tokenResolved {
+			token, tokenErr = h.resolveTokenForAuth(c.Request.Context(), auth)
+			tokenResolved = true
+		}
+		if auth != nil && token == "" {
+			if tokenErr != nil {
+				c.JSON(http.StatusBadRequest, gin.H{"error": "auth token refresh failed"})
+				return
+			}
+			c.JSON(http.StatusBadRequest, gin.H{"error": "auth token not found"})
+			return
+		}
+		if token == "" {
+			continue
+		}
+		reqHeaders[key] = strings.ReplaceAll(value, "$TOKEN$", token)
+	}
+
+	var requestBody io.Reader
+	if body.Data != "" {
+		requestBody = strings.NewReader(body.Data)
+	}
+
+	req, errNewRequest := http.NewRequestWithContext(c.Request.Context(), method, urlStr, requestBody)
+	if errNewRequest != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "failed to build request"})
+		return
+	}
+
+	for key, value := range reqHeaders {
+		if strings.EqualFold(key, "host") {
+			hostOverride = strings.TrimSpace(value)
+			continue
+		}
+		req.Header.Set(key, value)
+	}
+	if hostOverride != "" {
+		req.Host = hostOverride
+	}
+
+	httpClient := &http.Client{
+		Timeout: defaultAPICallTimeout,
+	}
+	httpClient.Transport = h.apiCallTransport(auth)
+
+	resp, errDo := httpClient.Do(req)
+	if errDo != nil {
+		log.WithError(errDo).Debug("management APICall request failed")
+		c.JSON(http.StatusBadGateway, gin.H{"error": "request failed"})
+		return
+	}
+	defer func() {
+		if errClose := resp.Body.Close(); errClose != nil {
+			log.Errorf("response body close error: %v", errClose)
+		}
+	}()
+
+	respBody, errReadAll := io.ReadAll(resp.Body)
+	if errReadAll != nil {
+		c.JSON(http.StatusBadGateway, gin.H{"error": "failed to read response"})
+		return
+	}
+
+	c.JSON(http.StatusOK, apiCallResponse{
+		StatusCode: resp.StatusCode,
+		Header:     resp.Header,
+		Body:       string(respBody),
+	})
+}
+
+func firstNonEmptyString(values ...*string) string {
+	for _, v := range values {
+		if v == nil {
+			continue
+		}
+		if out := strings.TrimSpace(*v); out != "" {
+			return out
+		}
+	}
+	return ""
+}
+
+func tokenValueForAuth(auth *coreauth.Auth) string {
+	if auth == nil {
+		return ""
+	}
+	if v := tokenValueFromMetadata(auth.Metadata); v != "" {
+		return v
+	}
+	if auth.Attributes != nil {
+		if v := strings.TrimSpace(auth.Attributes["api_key"]); v != "" {
+			return v
+		}
+	}
+	if shared := geminicli.ResolveSharedCredential(auth.Runtime); shared != nil {
+		if v := tokenValueFromMetadata(shared.MetadataSnapshot()); v != "" {
+			return v
+		}
+	}
+	return ""
+}
+
+func (h *Handler) resolveTokenForAuth(ctx context.Context, auth *coreauth.Auth) (string, error) {
+	if auth == nil {
+		return "", nil
+	}
+
+	provider := strings.ToLower(strings.TrimSpace(auth.Provider))
+	if provider == "gemini-cli" {
+		token, errToken := h.refreshGeminiOAuthAccessToken(ctx, auth)
+		return token, errToken
+	}
+	if provider == "antigravity" {
+		token, errToken := h.refreshAntigravityOAuthAccessToken(ctx, auth)
+		return token, errToken
+	}
+
+	return tokenValueForAuth(auth), nil
+}
+
+func (h *Handler) refreshGeminiOAuthAccessToken(ctx context.Context, auth *coreauth.Auth) (string, error) {
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	if auth == nil {
+		return "", nil
+	}
+
+	metadata, updater := geminiOAuthMetadata(auth)
+	if len(metadata) == 0 {
+		return "", fmt.Errorf("gemini oauth metadata missing")
+	}
+
+	base := make(map[string]any)
+	if tokenRaw, ok := metadata["token"].(map[string]any); ok && tokenRaw != nil {
+		base = cloneMap(tokenRaw)
+	}
+
+	var token oauth2.Token
+	if len(base) > 0 {
+		if raw, errMarshal := json.Marshal(base); errMarshal == nil {
+			_ = json.Unmarshal(raw, &token)
+		}
+	}
+
+	if token.AccessToken == "" {
+		token.AccessToken = stringValue(metadata, "access_token")
+	}
+	if token.RefreshToken == "" {
+		token.RefreshToken = stringValue(metadata, "refresh_token")
+	}
+	if token.TokenType == "" {
+		token.TokenType = stringValue(metadata, "token_type")
+	}
+	if token.Expiry.IsZero() {
+		if expiry := stringValue(metadata, "expiry"); expiry != "" {
+			if ts, errParseTime := time.Parse(time.RFC3339, expiry); errParseTime == nil {
+				token.Expiry = ts
+			}
+		}
+	}
+
+	conf := &oauth2.Config{
+		ClientID:     geminiOAuthClientID,
+		ClientSecret: geminiOAuthClientSecret,
+		Scopes:       geminiOAuthScopes,
+		Endpoint:     google.Endpoint,
+	}
+
+	ctxToken := ctx
+	httpClient := &http.Client{
+		Timeout:   defaultAPICallTimeout,
+		Transport: h.apiCallTransport(auth),
+	}
+	ctxToken = context.WithValue(ctxToken, oauth2.HTTPClient, httpClient)
+
+	src := conf.TokenSource(ctxToken, &token)
+	currentToken, errToken := src.Token()
+	if errToken != nil {
+		return "", errToken
+	}
+
+	merged := buildOAuthTokenMap(base, currentToken)
+	fields := buildOAuthTokenFields(currentToken, merged)
+	if updater != nil {
+		updater(fields)
+	}
+	return strings.TrimSpace(currentToken.AccessToken), nil
+}
+
+func (h *Handler) refreshAntigravityOAuthAccessToken(ctx context.Context, auth *coreauth.Auth) (string, error) {
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	if auth == nil {
+		return "", nil
+	}
+
+	metadata := auth.Metadata
+	if len(metadata) == 0 {
+		return "", fmt.Errorf("antigravity oauth metadata missing")
+	}
+
+	current := strings.TrimSpace(tokenValueFromMetadata(metadata))
+	if current != "" && !antigravityTokenNeedsRefresh(metadata) {
+		return current, nil
+	}
+
+	refreshToken := stringValue(metadata, "refresh_token")
+	if refreshToken == "" {
+		return "", fmt.Errorf("antigravity refresh token missing")
+	}
+
+	tokenURL := strings.TrimSpace(antigravityOAuthTokenURL)
+	if tokenURL == "" {
+		tokenURL = "https://oauth2.googleapis.com/token"
+	}
+	form := url.Values{}
+	form.Set("client_id", antigravityOAuthClientID)
+	form.Set("client_secret", antigravityOAuthClientSecret)
+	form.Set("grant_type", "refresh_token")
+	form.Set("refresh_token", refreshToken)
+
+	req, errReq := http.NewRequestWithContext(ctx, http.MethodPost, tokenURL, strings.NewReader(form.Encode()))
+	if errReq != nil {
+		return "", errReq
+	}
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+
+	httpClient := &http.Client{
+		Timeout:   defaultAPICallTimeout,
+		Transport: h.apiCallTransport(auth),
+	}
+	resp, errDo := httpClient.Do(req)
+	if errDo != nil {
+		return "", errDo
+	}
+	defer func() {
+		if errClose := resp.Body.Close(); errClose != nil {
+			log.Errorf("response body close error: %v", errClose)
+		}
+	}()
+
+	bodyBytes, errRead := io.ReadAll(resp.Body)
+	if errRead != nil {
+		return "", errRead
+	}
+	if resp.StatusCode < http.StatusOK || resp.StatusCode >= http.StatusMultipleChoices {
+		return "", fmt.Errorf("antigravity oauth token refresh failed: status %d: %s", resp.StatusCode, strings.TrimSpace(string(bodyBytes)))
+	}
+
+	var tokenResp struct {
+		AccessToken  string `json:"access_token"`
+		RefreshToken string `json:"refresh_token"`
+		ExpiresIn    int64  `json:"expires_in"`
+		TokenType    string `json:"token_type"`
+	}
+	if errUnmarshal := json.Unmarshal(bodyBytes, &tokenResp); errUnmarshal != nil {
+		return "", errUnmarshal
+	}
+
+	if strings.TrimSpace(tokenResp.AccessToken) == "" {
+		return "", fmt.Errorf("antigravity oauth token refresh returned empty access_token")
+	}
+
+	if auth.Metadata == nil {
+		auth.Metadata = make(map[string]any)
+	}
+	now := time.Now()
+	auth.Metadata["access_token"] = strings.TrimSpace(tokenResp.AccessToken)
+	if strings.TrimSpace(tokenResp.RefreshToken) != "" {
+		auth.Metadata["refresh_token"] = strings.TrimSpace(tokenResp.RefreshToken)
+	}
+	if tokenResp.ExpiresIn > 0 {
+		auth.Metadata["expires_in"] = tokenResp.ExpiresIn
+		auth.Metadata["timestamp"] = now.UnixMilli()
+		auth.Metadata["expired"] = now.Add(time.Duration(tokenResp.ExpiresIn) * time.Second).Format(time.RFC3339)
+	}
+	auth.Metadata["type"] = "antigravity"
+
+	if h != nil && h.authManager != nil {
+		auth.LastRefreshedAt = now
+		auth.UpdatedAt = now
+		_, _ = h.authManager.Update(ctx, auth)
+	}
+
+	return strings.TrimSpace(tokenResp.AccessToken), nil
+}
+
+func antigravityTokenNeedsRefresh(metadata map[string]any) bool {
+	// Refresh a bit early to avoid requests racing token expiry.
+	const skew = 30 * time.Second
+
+	if metadata == nil {
+		return true
+	}
+	if expStr, ok := metadata["expired"].(string); ok {
+		if ts, errParse := time.Parse(time.RFC3339, strings.TrimSpace(expStr)); errParse == nil {
+			return !ts.After(time.Now().Add(skew))
+		}
+	}
+	expiresIn := int64Value(metadata["expires_in"])
+	timestampMs := int64Value(metadata["timestamp"])
+	if expiresIn > 0 && timestampMs > 0 {
+		exp := time.UnixMilli(timestampMs).Add(time.Duration(expiresIn) * time.Second)
+		return !exp.After(time.Now().Add(skew))
+	}
+	return true
+}
+
+func int64Value(raw any) int64 {
+	switch typed := raw.(type) {
+	case int:
+		return int64(typed)
+	case int32:
+		return int64(typed)
+	case int64:
+		return typed
+	case uint:
+		return int64(typed)
+	case uint32:
+		return int64(typed)
+	case uint64:
+		if typed > uint64(^uint64(0)>>1) {
+			return 0
+		}
+		return int64(typed)
+	case float32:
+		return int64(typed)
+	case float64:
+		return int64(typed)
+	case json.Number:
+		if i, errParse := typed.Int64(); errParse == nil {
+			return i
+		}
+	case string:
+		if s := strings.TrimSpace(typed); s != "" {
+			if i, errParse := json.Number(s).Int64(); errParse == nil {
+				return i
+			}
+		}
+	}
+	return 0
+}
+
+func geminiOAuthMetadata(auth *coreauth.Auth) (map[string]any, func(map[string]any)) {
+	if auth == nil {
+		return nil, nil
+	}
+	if shared := geminicli.ResolveSharedCredential(auth.Runtime); shared != nil {
+		snapshot := shared.MetadataSnapshot()
+		return snapshot, func(fields map[string]any) { shared.MergeMetadata(fields) }
+	}
+	return auth.Metadata, func(fields map[string]any) {
+		if auth.Metadata == nil {
+			auth.Metadata = make(map[string]any)
+		}
+		for k, v := range fields {
+			auth.Metadata[k] = v
+		}
+	}
+}
+
+func stringValue(metadata map[string]any, key string) string {
+	if len(metadata) == 0 || key == "" {
+		return ""
+	}
+	if v, ok := metadata[key].(string); ok {
+		return strings.TrimSpace(v)
+	}
+	return ""
+}
+
+func cloneMap(in map[string]any) map[string]any {
+	if len(in) == 0 {
+		return nil
+	}
+	out := make(map[string]any, len(in))
+	for k, v := range in {
+		out[k] = v
+	}
+	return out
+}
+
+func buildOAuthTokenMap(base map[string]any, tok *oauth2.Token) map[string]any {
+	merged := cloneMap(base)
+	if merged == nil {
+		merged = make(map[string]any)
+	}
+	if tok == nil {
+		return merged
+	}
+	if raw, errMarshal := json.Marshal(tok); errMarshal == nil {
+		var tokenMap map[string]any
+		if errUnmarshal := json.Unmarshal(raw, &tokenMap); errUnmarshal == nil {
+			for k, v := range tokenMap {
+				merged[k] = v
+			}
+		}
+	}
+	return merged
+}
+
+func buildOAuthTokenFields(tok *oauth2.Token, merged map[string]any) map[string]any {
+	fields := make(map[string]any, 5)
+	if tok != nil && tok.AccessToken != "" {
+		fields["access_token"] = tok.AccessToken
+	}
+	if tok != nil && tok.TokenType != "" {
+		fields["token_type"] = tok.TokenType
+	}
+	if tok != nil && tok.RefreshToken != "" {
+		fields["refresh_token"] = tok.RefreshToken
+	}
+	if tok != nil && !tok.Expiry.IsZero() {
+		fields["expiry"] = tok.Expiry.Format(time.RFC3339)
+	}
+	if len(merged) > 0 {
+		fields["token"] = cloneMap(merged)
+	}
+	return fields
+}
+
+func tokenValueFromMetadata(metadata map[string]any) string {
+	if len(metadata) == 0 {
+		return ""
+	}
+	if v, ok := metadata["accessToken"].(string); ok && strings.TrimSpace(v) != "" {
+		return strings.TrimSpace(v)
+	}
+	if v, ok := metadata["access_token"].(string); ok && strings.TrimSpace(v) != "" {
+		return strings.TrimSpace(v)
+	}
+	if tokenRaw, ok := metadata["token"]; ok && tokenRaw != nil {
+		switch typed := tokenRaw.(type) {
+		case string:
+			if v := strings.TrimSpace(typed); v != "" {
+				return v
+			}
+		case map[string]any:
+			if v, ok := typed["access_token"].(string); ok && strings.TrimSpace(v) != "" {
+				return strings.TrimSpace(v)
+			}
+			if v, ok := typed["accessToken"].(string); ok && strings.TrimSpace(v) != "" {
+				return strings.TrimSpace(v)
+			}
+		case map[string]string:
+			if v := strings.TrimSpace(typed["access_token"]); v != "" {
+				return v
+			}
+			if v := strings.TrimSpace(typed["accessToken"]); v != "" {
+				return v
+			}
+		}
+	}
+	if v, ok := metadata["token"].(string); ok && strings.TrimSpace(v) != "" {
+		return strings.TrimSpace(v)
+	}
+	if v, ok := metadata["id_token"].(string); ok && strings.TrimSpace(v) != "" {
+		return strings.TrimSpace(v)
+	}
+	if v, ok := metadata["cookie"].(string); ok && strings.TrimSpace(v) != "" {
+		return strings.TrimSpace(v)
+	}
+	return ""
+}
+
+func (h *Handler) authByIndex(authIndex string) *coreauth.Auth {
+	authIndex = strings.TrimSpace(authIndex)
+	if authIndex == "" || h == nil || h.authManager == nil {
+		return nil
+	}
+	auths := h.authManager.List()
+	for _, auth := range auths {
+		if auth == nil {
+			continue
+		}
+		auth.EnsureIndex()
+		if auth.Index == authIndex {
+			return auth
+		}
+	}
+	return nil
+}
+
+func (h *Handler) apiCallTransport(auth *coreauth.Auth) http.RoundTripper {
+	var proxyCandidates []string
+	if auth != nil {
+		if proxyStr := strings.TrimSpace(auth.ProxyURL); proxyStr != "" {
+			proxyCandidates = append(proxyCandidates, proxyStr)
+		}
+	}
+	if h != nil && h.cfg != nil {
+		if proxyStr := strings.TrimSpace(h.cfg.ProxyURL); proxyStr != "" {
+			proxyCandidates = append(proxyCandidates, proxyStr)
+		}
+	}
+
+	for _, proxyStr := range proxyCandidates {
+		if transport := buildProxyTransport(proxyStr); transport != nil {
+			return transport
+		}
+	}
+
+	transport, ok := http.DefaultTransport.(*http.Transport)
+	if !ok || transport == nil {
+		return &http.Transport{Proxy: nil}
+	}
+	clone := transport.Clone()
+	clone.Proxy = nil
+	return clone
+}
+
+func buildProxyTransport(proxyStr string) *http.Transport {
+	proxyStr = strings.TrimSpace(proxyStr)
+	if proxyStr == "" {
+		return nil
+	}
+
+	proxyURL, errParse := url.Parse(proxyStr)
+	if errParse != nil {
+		log.WithError(errParse).Debug("parse proxy URL failed")
+		return nil
+	}
+	if proxyURL.Scheme == "" || proxyURL.Host == "" {
+		log.Debug("proxy URL missing scheme/host")
+		return nil
+	}
+
+	if proxyURL.Scheme == "socks5" {
+		var proxyAuth *proxy.Auth
+		if proxyURL.User != nil {
+			username := proxyURL.User.Username()
+			password, _ := proxyURL.User.Password()
+			proxyAuth = &proxy.Auth{User: username, Password: password}
+		}
+		dialer, errSOCKS5 := proxy.SOCKS5("tcp", proxyURL.Host, proxyAuth, proxy.Direct)
+		if errSOCKS5 != nil {
+			log.WithError(errSOCKS5).Debug("create SOCKS5 dialer failed")
+			return nil
+		}
+		return &http.Transport{
+			Proxy: nil,
+			DialContext: func(ctx context.Context, network, addr string) (net.Conn, error) {
+				return dialer.Dial(network, addr)
+			},
+		}
+	}
+
+	if proxyURL.Scheme == "http" || proxyURL.Scheme == "https" {
+		return &http.Transport{Proxy: http.ProxyURL(proxyURL)}
+	}
+
+	log.Debugf("unsupported proxy scheme: %s", proxyURL.Scheme)
+	return nil
+}

internal/api/handlers/management/api_tools_test.go

@@ -0,0 +1,173 @@
+package management
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+)
+
+type memoryAuthStore struct {
+	mu    sync.Mutex
+	items map[string]*coreauth.Auth
+}
+
+func (s *memoryAuthStore) List(ctx context.Context) ([]*coreauth.Auth, error) {
+	_ = ctx
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	out := make([]*coreauth.Auth, 0, len(s.items))
+	for _, a := range s.items {
+		out = append(out, a.Clone())
+	}
+	return out, nil
+}
+
+func (s *memoryAuthStore) Save(ctx context.Context, auth *coreauth.Auth) (string, error) {
+	_ = ctx
+	if auth == nil {
+		return "", nil
+	}
+	s.mu.Lock()
+	if s.items == nil {
+		s.items = make(map[string]*coreauth.Auth)
+	}
+	s.items[auth.ID] = auth.Clone()
+	s.mu.Unlock()
+	return auth.ID, nil
+}
+
+func (s *memoryAuthStore) Delete(ctx context.Context, id string) error {
+	_ = ctx
+	s.mu.Lock()
+	delete(s.items, id)
+	s.mu.Unlock()
+	return nil
+}
+
+func TestResolveTokenForAuth_Antigravity_RefreshesExpiredToken(t *testing.T) {
+	var callCount int
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		callCount++
+		if r.Method != http.MethodPost {
+			t.Fatalf("expected POST, got %s", r.Method)
+		}
+		if ct := r.Header.Get("Content-Type"); !strings.HasPrefix(ct, "application/x-www-form-urlencoded") {
+			t.Fatalf("unexpected content-type: %s", ct)
+		}
+		bodyBytes, _ := io.ReadAll(r.Body)
+		_ = r.Body.Close()
+		values, err := url.ParseQuery(string(bodyBytes))
+		if err != nil {
+			t.Fatalf("parse form: %v", err)
+		}
+		if values.Get("grant_type") != "refresh_token" {
+			t.Fatalf("unexpected grant_type: %s", values.Get("grant_type"))
+		}
+		if values.Get("refresh_token") != "rt" {
+			t.Fatalf("unexpected refresh_token: %s", values.Get("refresh_token"))
+		}
+		if values.Get("client_id") != antigravityOAuthClientID {
+			t.Fatalf("unexpected client_id: %s", values.Get("client_id"))
+		}
+		if values.Get("client_secret") != antigravityOAuthClientSecret {
+			t.Fatalf("unexpected client_secret")
+		}
+
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"access_token":  "new-token",
+			"refresh_token": "rt2",
+			"expires_in":    int64(3600),
+			"token_type":    "Bearer",
+		})
+	}))
+	t.Cleanup(srv.Close)
+
+	originalURL := antigravityOAuthTokenURL
+	antigravityOAuthTokenURL = srv.URL
+	t.Cleanup(func() { antigravityOAuthTokenURL = originalURL })
+
+	store := &memoryAuthStore{}
+	manager := coreauth.NewManager(store, nil, nil)
+
+	auth := &coreauth.Auth{
+		ID:       "antigravity-test.json",
+		FileName: "antigravity-test.json",
+		Provider: "antigravity",
+		Metadata: map[string]any{
+			"type":          "antigravity",
+			"access_token":  "old-token",
+			"refresh_token": "rt",
+			"expires_in":    int64(3600),
+			"timestamp":     time.Now().Add(-2 * time.Hour).UnixMilli(),
+			"expired":       time.Now().Add(-1 * time.Hour).Format(time.RFC3339),
+		},
+	}
+	if _, err := manager.Register(context.Background(), auth); err != nil {
+		t.Fatalf("register auth: %v", err)
+	}
+
+	h := &Handler{authManager: manager}
+	token, err := h.resolveTokenForAuth(context.Background(), auth)
+	if err != nil {
+		t.Fatalf("resolveTokenForAuth: %v", err)
+	}
+	if token != "new-token" {
+		t.Fatalf("expected refreshed token, got %q", token)
+	}
+	if callCount != 1 {
+		t.Fatalf("expected 1 refresh call, got %d", callCount)
+	}
+
+	updated, ok := manager.GetByID(auth.ID)
+	if !ok || updated == nil {
+		t.Fatalf("expected auth in manager after update")
+	}
+	if got := tokenValueFromMetadata(updated.Metadata); got != "new-token" {
+		t.Fatalf("expected manager metadata updated, got %q", got)
+	}
+}
+
+func TestResolveTokenForAuth_Antigravity_SkipsRefreshWhenTokenValid(t *testing.T) {
+	var callCount int
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		callCount++
+		w.WriteHeader(http.StatusInternalServerError)
+	}))
+	t.Cleanup(srv.Close)
+
+	originalURL := antigravityOAuthTokenURL
+	antigravityOAuthTokenURL = srv.URL
+	t.Cleanup(func() { antigravityOAuthTokenURL = originalURL })
+
+	auth := &coreauth.Auth{
+		ID:       "antigravity-valid.json",
+		FileName: "antigravity-valid.json",
+		Provider: "antigravity",
+		Metadata: map[string]any{
+			"type":         "antigravity",
+			"access_token": "ok-token",
+			"expired":      time.Now().Add(30 * time.Minute).Format(time.RFC3339),
+		},
+	}
+	h := &Handler{}
+	token, err := h.resolveTokenForAuth(context.Background(), auth)
+	if err != nil {
+		t.Fatalf("resolveTokenForAuth: %v", err)
+	}
+	if token != "ok-token" {
+		t.Fatalf("expected existing token, got %q", token)
+	}
+	if callCount != 0 {
+		t.Fatalf("expected no refresh calls, got %d", callCount)
+	}
+}

internal/api/handlers/management/auth_files.go

@@ -197,6 +197,19 @@ func stopCallbackForwarder(port int) {
 	stopForwarderInstance(port, forwarder)
 }
 
+func stopCallbackForwarderInstance(port int, forwarder *callbackForwarder) {
+	if forwarder == nil {
+		return
+	}
+	callbackForwardersMu.Lock()
+	if current := callbackForwarders[port]; current == forwarder {
+		delete(callbackForwarders, port)
+	}
+	callbackForwardersMu.Unlock()
+
+	stopForwarderInstance(port, forwarder)
+}
+
 func stopForwarderInstance(port int, forwarder *callbackForwarder) {
 	if forwarder == nil || forwarder.server == nil {
 		return
@@ -414,9 +427,52 @@ func (h *Handler) buildAuthFileEntry(auth *coreauth.Auth) gin.H {
 			log.WithError(err).Warnf("failed to stat auth file %s", path)
 		}
 	}
+	if claims := extractCodexIDTokenClaims(auth); claims != nil {
+		entry["id_token"] = claims
+	}
 	return entry
 }
 
+func extractCodexIDTokenClaims(auth *coreauth.Auth) gin.H {
+	if auth == nil || auth.Metadata == nil {
+		return nil
+	}
+	if !strings.EqualFold(strings.TrimSpace(auth.Provider), "codex") {
+		return nil
+	}
+	idTokenRaw, ok := auth.Metadata["id_token"].(string)
+	if !ok {
+		return nil
+	}
+	idToken := strings.TrimSpace(idTokenRaw)
+	if idToken == "" {
+		return nil
+	}
+	claims, err := codex.ParseJWTToken(idToken)
+	if err != nil || claims == nil {
+		return nil
+	}
+
+	result := gin.H{}
+	if v := strings.TrimSpace(claims.CodexAuthInfo.ChatgptAccountID); v != "" {
+		result["chatgpt_account_id"] = v
+	}
+	if v := strings.TrimSpace(claims.CodexAuthInfo.ChatgptPlanType); v != "" {
+		result["plan_type"] = v
+	}
+	if v := claims.CodexAuthInfo.ChatgptSubscriptionActiveStart; v != nil {
+		result["chatgpt_subscription_active_start"] = v
+	}
+	if v := claims.CodexAuthInfo.ChatgptSubscriptionActiveUntil; v != nil {
+		result["chatgpt_subscription_active_until"] = v
+	}
+
+	if len(result) == 0 {
+		return nil
+	}
+	return result
+}
+
 func authEmail(auth *coreauth.Auth) string {
 	if auth == nil {
 		return ""
@@ -785,6 +841,7 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 	RegisterOAuthSession(state, "anthropic")
 
 	isWebUI := isWebUIRequest(c)
+	var forwarder *callbackForwarder
 	if isWebUI {
 		targetURL, errTarget := h.managementCallbackURL("/anthropic/callback")
 		if errTarget != nil {
@@ -792,7 +849,8 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 			c.JSON(http.StatusInternalServerError, gin.H{"error": "callback server unavailable"})
 			return
 		}
-		if _, errStart := startCallbackForwarder(anthropicCallbackPort, "anthropic", targetURL); errStart != nil {
+		var errStart error
+		if forwarder, errStart = startCallbackForwarder(anthropicCallbackPort, "anthropic", targetURL); errStart != nil {
 			log.WithError(errStart).Error("failed to start anthropic callback forwarder")
 			c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to start callback server"})
 			return
@@ -801,7 +859,7 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 
 	go func() {
 		if isWebUI {
-			defer stopCallbackForwarder(anthropicCallbackPort)
+			defer stopCallbackForwarderInstance(anthropicCallbackPort, forwarder)
 		}
 
 		// Helper: wait for callback file
@@ -809,6 +867,9 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 		waitForFile := func(path string, timeout time.Duration) (map[string]string, error) {
 			deadline := time.Now().Add(timeout)
 			for {
+				if !IsOAuthSessionPending(state, "anthropic") {
+					return nil, errOAuthSessionNotPending
+				}
 				if time.Now().After(deadline) {
 					SetOAuthSessionError(state, "Timeout waiting for OAuth callback")
 					return nil, fmt.Errorf("timeout waiting for OAuth callback")
@@ -828,6 +889,9 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 		// Wait up to 5 minutes
 		resultMap, errWait := waitForFile(waitFile, 5*time.Minute)
 		if errWait != nil {
+			if errors.Is(errWait, errOAuthSessionNotPending) {
+				return
+			}
 			authErr := claude.NewAuthenticationError(claude.ErrCallbackTimeout, errWait)
 			log.Error(claude.GetUserFriendlyMessage(authErr))
 			return
@@ -933,6 +997,7 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 		}
 		fmt.Println("You can now use Claude services through this CLI")
 		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("anthropic")
 	}()
 
 	c.JSON(200, gin.H{"status": "ok", "url": authURL, "state": state})
@@ -968,6 +1033,7 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 	RegisterOAuthSession(state, "gemini")
 
 	isWebUI := isWebUIRequest(c)
+	var forwarder *callbackForwarder
 	if isWebUI {
 		targetURL, errTarget := h.managementCallbackURL("/google/callback")
 		if errTarget != nil {
@@ -975,7 +1041,8 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			c.JSON(http.StatusInternalServerError, gin.H{"error": "callback server unavailable"})
 			return
 		}
-		if _, errStart := startCallbackForwarder(geminiCallbackPort, "gemini", targetURL); errStart != nil {
+		var errStart error
+		if forwarder, errStart = startCallbackForwarder(geminiCallbackPort, "gemini", targetURL); errStart != nil {
 			log.WithError(errStart).Error("failed to start gemini callback forwarder")
 			c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to start callback server"})
 			return
@@ -984,7 +1051,7 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 
 	go func() {
 		if isWebUI {
-			defer stopCallbackForwarder(geminiCallbackPort)
+			defer stopCallbackForwarderInstance(geminiCallbackPort, forwarder)
 		}
 
 		// Wait for callback file written by server route
@@ -993,6 +1060,9 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 		deadline := time.Now().Add(5 * time.Minute)
 		var authCode string
 		for {
+			if !IsOAuthSessionPending(state, "gemini") {
+				return
+			}
 			if time.Now().After(deadline) {
 				log.Error("oauth flow timed out")
 				SetOAuthSessionError(state, "OAuth flow timed out")
@@ -1093,7 +1163,9 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 
 		// Initialize authenticated HTTP client via GeminiAuth to honor proxy settings
 		gemAuth := geminiAuth.NewGeminiAuth()
-		gemClient, errGetClient := gemAuth.GetAuthenticatedClient(ctx, &ts, h.cfg, true)
+		gemClient, errGetClient := gemAuth.GetAuthenticatedClient(ctx, &ts, h.cfg, &geminiAuth.WebLoginOptions{
+			NoBrowser: true,
+		})
 		if errGetClient != nil {
 			log.Errorf("failed to get authenticated client: %v", errGetClient)
 			SetOAuthSessionError(state, "Failed to get authenticated client")
@@ -1166,6 +1238,7 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 		}
 
 		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("gemini")
 		fmt.Printf("You can now use Gemini CLI services through this CLI; token saved to %s\n", savedPath)
 	}()
 
@@ -1207,6 +1280,7 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 	RegisterOAuthSession(state, "codex")
 
 	isWebUI := isWebUIRequest(c)
+	var forwarder *callbackForwarder
 	if isWebUI {
 		targetURL, errTarget := h.managementCallbackURL("/codex/callback")
 		if errTarget != nil {
@@ -1214,7 +1288,8 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 			c.JSON(http.StatusInternalServerError, gin.H{"error": "callback server unavailable"})
 			return
 		}
-		if _, errStart := startCallbackForwarder(codexCallbackPort, "codex", targetURL); errStart != nil {
+		var errStart error
+		if forwarder, errStart = startCallbackForwarder(codexCallbackPort, "codex", targetURL); errStart != nil {
 			log.WithError(errStart).Error("failed to start codex callback forwarder")
 			c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to start callback server"})
 			return
@@ -1223,7 +1298,7 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 
 	go func() {
 		if isWebUI {
-			defer stopCallbackForwarder(codexCallbackPort)
+			defer stopCallbackForwarderInstance(codexCallbackPort, forwarder)
 		}
 
 		// Wait for callback file
@@ -1231,6 +1306,9 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 		deadline := time.Now().Add(5 * time.Minute)
 		var code string
 		for {
+			if !IsOAuthSessionPending(state, "codex") {
+				return
+			}
 			if time.Now().After(deadline) {
 				authErr := codex.NewAuthenticationError(codex.ErrCallbackTimeout, fmt.Errorf("timeout waiting for OAuth callback"))
 				log.Error(codex.GetUserFriendlyMessage(authErr))
@@ -1346,6 +1424,7 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 		}
 		fmt.Println("You can now use Codex services through this CLI")
 		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("codex")
 	}()
 
 	c.JSON(200, gin.H{"status": "ok", "url": authURL, "state": state})
@@ -1391,6 +1470,7 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 	RegisterOAuthSession(state, "antigravity")
 
 	isWebUI := isWebUIRequest(c)
+	var forwarder *callbackForwarder
 	if isWebUI {
 		targetURL, errTarget := h.managementCallbackURL("/antigravity/callback")
 		if errTarget != nil {
@@ -1398,7 +1478,8 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 			c.JSON(http.StatusInternalServerError, gin.H{"error": "callback server unavailable"})
 			return
 		}
-		if _, errStart := startCallbackForwarder(antigravityCallbackPort, "antigravity", targetURL); errStart != nil {
+		var errStart error
+		if forwarder, errStart = startCallbackForwarder(antigravityCallbackPort, "antigravity", targetURL); errStart != nil {
 			log.WithError(errStart).Error("failed to start antigravity callback forwarder")
 			c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to start callback server"})
 			return
@@ -1407,13 +1488,16 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 
 	go func() {
 		if isWebUI {
-			defer stopCallbackForwarder(antigravityCallbackPort)
+			defer stopCallbackForwarderInstance(antigravityCallbackPort, forwarder)
 		}
 
 		waitFile := filepath.Join(h.cfg.AuthDir, fmt.Sprintf(".oauth-antigravity-%s.oauth", state))
 		deadline := time.Now().Add(5 * time.Minute)
 		var authCode string
 		for {
+			if !IsOAuthSessionPending(state, "antigravity") {
+				return
+			}
 			if time.Now().After(deadline) {
 				log.Error("oauth flow timed out")
 				SetOAuthSessionError(state, "OAuth flow timed out")
@@ -1576,6 +1660,7 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 		}
 
 		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("antigravity")
 		fmt.Printf("Authentication successful! Token saved to %s\n", savedPath)
 		if projectID != "" {
 			fmt.Printf("Using GCP project: %s\n", projectID)
@@ -1653,6 +1738,7 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 	RegisterOAuthSession(state, "iflow")
 
 	isWebUI := isWebUIRequest(c)
+	var forwarder *callbackForwarder
 	if isWebUI {
 		targetURL, errTarget := h.managementCallbackURL("/iflow/callback")
 		if errTarget != nil {
@@ -1660,7 +1746,8 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 			c.JSON(http.StatusInternalServerError, gin.H{"status": "error", "error": "callback server unavailable"})
 			return
 		}
-		if _, errStart := startCallbackForwarder(iflowauth.CallbackPort, "iflow", targetURL); errStart != nil {
+		var errStart error
+		if forwarder, errStart = startCallbackForwarder(iflowauth.CallbackPort, "iflow", targetURL); errStart != nil {
 			log.WithError(errStart).Error("failed to start iflow callback forwarder")
 			c.JSON(http.StatusInternalServerError, gin.H{"status": "error", "error": "failed to start callback server"})
 			return
@@ -1669,7 +1756,7 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 
 	go func() {
 		if isWebUI {
-			defer stopCallbackForwarder(iflowauth.CallbackPort)
+			defer stopCallbackForwarderInstance(iflowauth.CallbackPort, forwarder)
 		}
 		fmt.Println("Waiting for authentication...")
 
@@ -1677,6 +1764,9 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 		deadline := time.Now().Add(5 * time.Minute)
 		var resultMap map[string]string
 		for {
+			if !IsOAuthSessionPending(state, "iflow") {
+				return
+			}
 			if time.Now().After(deadline) {
 				SetOAuthSessionError(state, "Authentication failed")
 				fmt.Println("Authentication failed: timeout waiting for callback")
@@ -1743,6 +1833,7 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 		}
 		fmt.Println("You can now use iFlow services through this CLI")
 		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("iflow")
 	}()
 
 	c.JSON(http.StatusOK, gin.H{"status": "ok", "url": authURL, "state": state})

internal/api/handlers/management/config_basic.go

@@ -202,6 +202,26 @@ func (h *Handler) PutLoggingToFile(c *gin.Context) {
 	h.updateBoolField(c, func(v bool) { h.cfg.LoggingToFile = v })
 }
 
+// LogsMaxTotalSizeMB
+func (h *Handler) GetLogsMaxTotalSizeMB(c *gin.Context) {
+	c.JSON(200, gin.H{"logs-max-total-size-mb": h.cfg.LogsMaxTotalSizeMB})
+}
+func (h *Handler) PutLogsMaxTotalSizeMB(c *gin.Context) {
+	var body struct {
+		Value *int `json:"value"`
+	}
+	if errBindJSON := c.ShouldBindJSON(&body); errBindJSON != nil || body.Value == nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid body"})
+		return
+	}
+	value := *body.Value
+	if value < 0 {
+		value = 0
+	}
+	h.cfg.LogsMaxTotalSizeMB = value
+	h.persist(c)
+}
+
 // Request log
 func (h *Handler) GetRequestLog(c *gin.Context) { c.JSON(200, gin.H{"request-log": h.cfg.RequestLog}) }
 func (h *Handler) PutRequestLog(c *gin.Context) {
@@ -232,6 +252,52 @@ func (h *Handler) PutMaxRetryInterval(c *gin.Context) {
 	h.updateIntField(c, func(v int) { h.cfg.MaxRetryInterval = v })
 }
 
+// ForceModelPrefix
+func (h *Handler) GetForceModelPrefix(c *gin.Context) {
+	c.JSON(200, gin.H{"force-model-prefix": h.cfg.ForceModelPrefix})
+}
+func (h *Handler) PutForceModelPrefix(c *gin.Context) {
+	h.updateBoolField(c, func(v bool) { h.cfg.ForceModelPrefix = v })
+}
+
+func normalizeRoutingStrategy(strategy string) (string, bool) {
+	normalized := strings.ToLower(strings.TrimSpace(strategy))
+	switch normalized {
+	case "", "round-robin", "roundrobin", "rr":
+		return "round-robin", true
+	case "fill-first", "fillfirst", "ff":
+		return "fill-first", true
+	default:
+		return "", false
+	}
+}
+
+// RoutingStrategy
+func (h *Handler) GetRoutingStrategy(c *gin.Context) {
+	strategy, ok := normalizeRoutingStrategy(h.cfg.Routing.Strategy)
+	if !ok {
+		c.JSON(200, gin.H{"strategy": strings.TrimSpace(h.cfg.Routing.Strategy)})
+		return
+	}
+	c.JSON(200, gin.H{"strategy": strategy})
+}
+func (h *Handler) PutRoutingStrategy(c *gin.Context) {
+	var body struct {
+		Value *string `json:"value"`
+	}
+	if errBindJSON := c.ShouldBindJSON(&body); errBindJSON != nil || body.Value == nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid body"})
+		return
+	}
+	normalized, ok := normalizeRoutingStrategy(*body.Value)
+	if !ok {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid strategy"})
+		return
+	}
+	h.cfg.Routing.Strategy = normalized
+	h.persist(c)
+}
+
 // Proxy URL
 func (h *Handler) GetProxyURL(c *gin.Context) { c.JSON(200, gin.H{"proxy-url": h.cfg.ProxyURL}) }
 func (h *Handler) PutProxyURL(c *gin.Context) {

internal/api/handlers/management/config_lists.go

@@ -145,71 +145,74 @@ func (h *Handler) PutGeminiKeys(c *gin.Context) {
 	h.persist(c)
 }
 func (h *Handler) PatchGeminiKey(c *gin.Context) {
+	type geminiKeyPatch struct {
+		APIKey         *string            `json:"api-key"`
+		Prefix         *string            `json:"prefix"`
+		BaseURL        *string            `json:"base-url"`
+		ProxyURL       *string            `json:"proxy-url"`
+		Headers        *map[string]string `json:"headers"`
+		ExcludedModels *[]string          `json:"excluded-models"`
+	}
 	var body struct {
-		Index *int              `json:"index"`
-		Match *string           `json:"match"`
-		Value *config.GeminiKey `json:"value"`
+		Index *int            `json:"index"`
+		Match *string         `json:"match"`
+		Value *geminiKeyPatch `json:"value"`
 	}
 	if err := c.ShouldBindJSON(&body); err != nil || body.Value == nil {
 		c.JSON(400, gin.H{"error": "invalid body"})
 		return
 	}
-	value := *body.Value
-	value.APIKey = strings.TrimSpace(value.APIKey)
-	value.BaseURL = strings.TrimSpace(value.BaseURL)
-	value.ProxyURL = strings.TrimSpace(value.ProxyURL)
-	value.ExcludedModels = config.NormalizeExcludedModels(value.ExcludedModels)
-	if value.APIKey == "" {
-		// Treat empty API key as delete.
-		if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.GeminiKey) {
-			h.cfg.GeminiKey = append(h.cfg.GeminiKey[:*body.Index], h.cfg.GeminiKey[*body.Index+1:]...)
-			h.cfg.SanitizeGeminiKeys()
-			h.persist(c)
-			return
-		}
-		if body.Match != nil {
-			match := strings.TrimSpace(*body.Match)
-			if match != "" {
-				out := make([]config.GeminiKey, 0, len(h.cfg.GeminiKey))
-				removed := false
-				for i := range h.cfg.GeminiKey {
-					if !removed && h.cfg.GeminiKey[i].APIKey == match {
-						removed = true
-						continue
-					}
-					out = append(out, h.cfg.GeminiKey[i])
-				}
-				if removed {
-					h.cfg.GeminiKey = out
-					h.cfg.SanitizeGeminiKeys()
-					h.persist(c)
-					return
+	targetIndex := -1
+	if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.GeminiKey) {
+		targetIndex = *body.Index
+	}
+	if targetIndex == -1 && body.Match != nil {
+		match := strings.TrimSpace(*body.Match)
+		if match != "" {
+			for i := range h.cfg.GeminiKey {
+				if h.cfg.GeminiKey[i].APIKey == match {
+					targetIndex = i
+					break
 				}
 			}
 		}
+	}
+	if targetIndex == -1 {
 		c.JSON(404, gin.H{"error": "item not found"})
 		return
 	}
 
-	if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.GeminiKey) {
-		h.cfg.GeminiKey[*body.Index] = value
-		h.cfg.SanitizeGeminiKeys()
-		h.persist(c)
-		return
-	}
-	if body.Match != nil {
-		match := strings.TrimSpace(*body.Match)
-		for i := range h.cfg.GeminiKey {
-			if h.cfg.GeminiKey[i].APIKey == match {
-				h.cfg.GeminiKey[i] = value
-				h.cfg.SanitizeGeminiKeys()
-				h.persist(c)
-				return
-			}
+	entry := h.cfg.GeminiKey[targetIndex]
+	if body.Value.APIKey != nil {
+		trimmed := strings.TrimSpace(*body.Value.APIKey)
+		if trimmed == "" {
+			h.cfg.GeminiKey = append(h.cfg.GeminiKey[:targetIndex], h.cfg.GeminiKey[targetIndex+1:]...)
+			h.cfg.SanitizeGeminiKeys()
+			h.persist(c)
+			return
 		}
+		entry.APIKey = trimmed
 	}
-	c.JSON(404, gin.H{"error": "item not found"})
+	if body.Value.Prefix != nil {
+		entry.Prefix = strings.TrimSpace(*body.Value.Prefix)
+	}
+	if body.Value.BaseURL != nil {
+		entry.BaseURL = strings.TrimSpace(*body.Value.BaseURL)
+	}
+	if body.Value.ProxyURL != nil {
+		entry.ProxyURL = strings.TrimSpace(*body.Value.ProxyURL)
+	}
+	if body.Value.Headers != nil {
+		entry.Headers = config.NormalizeHeaders(*body.Value.Headers)
+	}
+	if body.Value.ExcludedModels != nil {
+		entry.ExcludedModels = config.NormalizeExcludedModels(*body.Value.ExcludedModels)
+	}
+	h.cfg.GeminiKey[targetIndex] = entry
+	h.cfg.SanitizeGeminiKeys()
+	h.persist(c)
 }
+
 func (h *Handler) DeleteGeminiKey(c *gin.Context) {
 	if val := strings.TrimSpace(c.Query("api-key")); val != "" {
 		out := make([]config.GeminiKey, 0, len(h.cfg.GeminiKey))
@@ -268,35 +271,70 @@ func (h *Handler) PutClaudeKeys(c *gin.Context) {
 	h.persist(c)
 }
 func (h *Handler) PatchClaudeKey(c *gin.Context) {
+	type claudeKeyPatch struct {
+		APIKey         *string               `json:"api-key"`
+		Prefix         *string               `json:"prefix"`
+		BaseURL        *string               `json:"base-url"`
+		ProxyURL       *string               `json:"proxy-url"`
+		Models         *[]config.ClaudeModel `json:"models"`
+		Headers        *map[string]string    `json:"headers"`
+		ExcludedModels *[]string             `json:"excluded-models"`
+	}
 	var body struct {
-		Index *int              `json:"index"`
-		Match *string           `json:"match"`
-		Value *config.ClaudeKey `json:"value"`
+		Index *int            `json:"index"`
+		Match *string         `json:"match"`
+		Value *claudeKeyPatch `json:"value"`
 	}
 	if err := c.ShouldBindJSON(&body); err != nil || body.Value == nil {
 		c.JSON(400, gin.H{"error": "invalid body"})
 		return
 	}
-	value := *body.Value
-	normalizeClaudeKey(&value)
+	targetIndex := -1
 	if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.ClaudeKey) {
-		h.cfg.ClaudeKey[*body.Index] = value
-		h.cfg.SanitizeClaudeKeys()
-		h.persist(c)
-		return
+		targetIndex = *body.Index
 	}
-	if body.Match != nil {
+	if targetIndex == -1 && body.Match != nil {
+		match := strings.TrimSpace(*body.Match)
 		for i := range h.cfg.ClaudeKey {
-			if h.cfg.ClaudeKey[i].APIKey == *body.Match {
-				h.cfg.ClaudeKey[i] = value
-				h.cfg.SanitizeClaudeKeys()
-				h.persist(c)
-				return
+			if h.cfg.ClaudeKey[i].APIKey == match {
+				targetIndex = i
+				break
 			}
 		}
 	}
-	c.JSON(404, gin.H{"error": "item not found"})
+	if targetIndex == -1 {
+		c.JSON(404, gin.H{"error": "item not found"})
+		return
+	}
+
+	entry := h.cfg.ClaudeKey[targetIndex]
+	if body.Value.APIKey != nil {
+		entry.APIKey = strings.TrimSpace(*body.Value.APIKey)
+	}
+	if body.Value.Prefix != nil {
+		entry.Prefix = strings.TrimSpace(*body.Value.Prefix)
+	}
+	if body.Value.BaseURL != nil {
+		entry.BaseURL = strings.TrimSpace(*body.Value.BaseURL)
+	}
+	if body.Value.ProxyURL != nil {
+		entry.ProxyURL = strings.TrimSpace(*body.Value.ProxyURL)
+	}
+	if body.Value.Models != nil {
+		entry.Models = append([]config.ClaudeModel(nil), (*body.Value.Models)...)
+	}
+	if body.Value.Headers != nil {
+		entry.Headers = config.NormalizeHeaders(*body.Value.Headers)
+	}
+	if body.Value.ExcludedModels != nil {
+		entry.ExcludedModels = config.NormalizeExcludedModels(*body.Value.ExcludedModels)
+	}
+	normalizeClaudeKey(&entry)
+	h.cfg.ClaudeKey[targetIndex] = entry
+	h.cfg.SanitizeClaudeKeys()
+	h.persist(c)
 }
+
 func (h *Handler) DeleteClaudeKey(c *gin.Context) {
 	if val := c.Query("api-key"); val != "" {
 		out := make([]config.ClaudeKey, 0, len(h.cfg.ClaudeKey))
@@ -356,62 +394,73 @@ func (h *Handler) PutOpenAICompat(c *gin.Context) {
 	h.persist(c)
 }
 func (h *Handler) PatchOpenAICompat(c *gin.Context) {
+	type openAICompatPatch struct {
+		Name          *string                             `json:"name"`
+		Prefix        *string                             `json:"prefix"`
+		BaseURL       *string                             `json:"base-url"`
+		APIKeyEntries *[]config.OpenAICompatibilityAPIKey `json:"api-key-entries"`
+		Models        *[]config.OpenAICompatibilityModel  `json:"models"`
+		Headers       *map[string]string                  `json:"headers"`
+	}
 	var body struct {
-		Name  *string                     `json:"name"`
-		Index *int                        `json:"index"`
-		Value *config.OpenAICompatibility `json:"value"`
+		Name  *string            `json:"name"`
+		Index *int               `json:"index"`
+		Value *openAICompatPatch `json:"value"`
 	}
 	if err := c.ShouldBindJSON(&body); err != nil || body.Value == nil {
 		c.JSON(400, gin.H{"error": "invalid body"})
 		return
 	}
-	normalizeOpenAICompatibilityEntry(body.Value)
-	// If base-url becomes empty, delete the provider instead of updating
-	if strings.TrimSpace(body.Value.BaseURL) == "" {
-		if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.OpenAICompatibility) {
-			h.cfg.OpenAICompatibility = append(h.cfg.OpenAICompatibility[:*body.Index], h.cfg.OpenAICompatibility[*body.Index+1:]...)
+	targetIndex := -1
+	if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.OpenAICompatibility) {
+		targetIndex = *body.Index
+	}
+	if targetIndex == -1 && body.Name != nil {
+		match := strings.TrimSpace(*body.Name)
+		for i := range h.cfg.OpenAICompatibility {
+			if h.cfg.OpenAICompatibility[i].Name == match {
+				targetIndex = i
+				break
+			}
+		}
+	}
+	if targetIndex == -1 {
+		c.JSON(404, gin.H{"error": "item not found"})
+		return
+	}
+
+	entry := h.cfg.OpenAICompatibility[targetIndex]
+	if body.Value.Name != nil {
+		entry.Name = strings.TrimSpace(*body.Value.Name)
+	}
+	if body.Value.Prefix != nil {
+		entry.Prefix = strings.TrimSpace(*body.Value.Prefix)
+	}
+	if body.Value.BaseURL != nil {
+		trimmed := strings.TrimSpace(*body.Value.BaseURL)
+		if trimmed == "" {
+			h.cfg.OpenAICompatibility = append(h.cfg.OpenAICompatibility[:targetIndex], h.cfg.OpenAICompatibility[targetIndex+1:]...)
 			h.cfg.SanitizeOpenAICompatibility()
 			h.persist(c)
 			return
 		}
-		if body.Name != nil {
-			out := make([]config.OpenAICompatibility, 0, len(h.cfg.OpenAICompatibility))
-			removed := false
-			for i := range h.cfg.OpenAICompatibility {
-				if !removed && h.cfg.OpenAICompatibility[i].Name == *body.Name {
-					removed = true
-					continue
-				}
-				out = append(out, h.cfg.OpenAICompatibility[i])
-			}
-			if removed {
-				h.cfg.OpenAICompatibility = out
-				h.cfg.SanitizeOpenAICompatibility()
-				h.persist(c)
-				return
-			}
-		}
-		c.JSON(404, gin.H{"error": "item not found"})
-		return
+		entry.BaseURL = trimmed
 	}
-	if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.OpenAICompatibility) {
-		h.cfg.OpenAICompatibility[*body.Index] = *body.Value
-		h.cfg.SanitizeOpenAICompatibility()
-		h.persist(c)
-		return
+	if body.Value.APIKeyEntries != nil {
+		entry.APIKeyEntries = append([]config.OpenAICompatibilityAPIKey(nil), (*body.Value.APIKeyEntries)...)
 	}
-	if body.Name != nil {
-		for i := range h.cfg.OpenAICompatibility {
-			if h.cfg.OpenAICompatibility[i].Name == *body.Name {
-				h.cfg.OpenAICompatibility[i] = *body.Value
-				h.cfg.SanitizeOpenAICompatibility()
-				h.persist(c)
-				return
-			}
-		}
+	if body.Value.Models != nil {
+		entry.Models = append([]config.OpenAICompatibilityModel(nil), (*body.Value.Models)...)
 	}
-	c.JSON(404, gin.H{"error": "item not found"})
+	if body.Value.Headers != nil {
+		entry.Headers = config.NormalizeHeaders(*body.Value.Headers)
+	}
+	normalizeOpenAICompatibilityEntry(&entry)
+	h.cfg.OpenAICompatibility[targetIndex] = entry
+	h.cfg.SanitizeOpenAICompatibility()
+	h.persist(c)
 }
+
 func (h *Handler) DeleteOpenAICompat(c *gin.Context) {
 	if name := c.Query("name"); name != "" {
 		out := make([]config.OpenAICompatibility, 0, len(h.cfg.OpenAICompatibility))
@@ -438,6 +487,137 @@ func (h *Handler) DeleteOpenAICompat(c *gin.Context) {
 	c.JSON(400, gin.H{"error": "missing name or index"})
 }
 
+// vertex-api-key: []VertexCompatKey
+func (h *Handler) GetVertexCompatKeys(c *gin.Context) {
+	c.JSON(200, gin.H{"vertex-api-key": h.cfg.VertexCompatAPIKey})
+}
+func (h *Handler) PutVertexCompatKeys(c *gin.Context) {
+	data, err := c.GetRawData()
+	if err != nil {
+		c.JSON(400, gin.H{"error": "failed to read body"})
+		return
+	}
+	var arr []config.VertexCompatKey
+	if err = json.Unmarshal(data, &arr); err != nil {
+		var obj struct {
+			Items []config.VertexCompatKey `json:"items"`
+		}
+		if err2 := json.Unmarshal(data, &obj); err2 != nil || len(obj.Items) == 0 {
+			c.JSON(400, gin.H{"error": "invalid body"})
+			return
+		}
+		arr = obj.Items
+	}
+	for i := range arr {
+		normalizeVertexCompatKey(&arr[i])
+	}
+	h.cfg.VertexCompatAPIKey = arr
+	h.cfg.SanitizeVertexCompatKeys()
+	h.persist(c)
+}
+func (h *Handler) PatchVertexCompatKey(c *gin.Context) {
+	type vertexCompatPatch struct {
+		APIKey   *string                     `json:"api-key"`
+		Prefix   *string                     `json:"prefix"`
+		BaseURL  *string                     `json:"base-url"`
+		ProxyURL *string                     `json:"proxy-url"`
+		Headers  *map[string]string          `json:"headers"`
+		Models   *[]config.VertexCompatModel `json:"models"`
+	}
+	var body struct {
+		Index *int               `json:"index"`
+		Match *string            `json:"match"`
+		Value *vertexCompatPatch `json:"value"`
+	}
+	if errBindJSON := c.ShouldBindJSON(&body); errBindJSON != nil || body.Value == nil {
+		c.JSON(400, gin.H{"error": "invalid body"})
+		return
+	}
+	targetIndex := -1
+	if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.VertexCompatAPIKey) {
+		targetIndex = *body.Index
+	}
+	if targetIndex == -1 && body.Match != nil {
+		match := strings.TrimSpace(*body.Match)
+		if match != "" {
+			for i := range h.cfg.VertexCompatAPIKey {
+				if h.cfg.VertexCompatAPIKey[i].APIKey == match {
+					targetIndex = i
+					break
+				}
+			}
+		}
+	}
+	if targetIndex == -1 {
+		c.JSON(404, gin.H{"error": "item not found"})
+		return
+	}
+
+	entry := h.cfg.VertexCompatAPIKey[targetIndex]
+	if body.Value.APIKey != nil {
+		trimmed := strings.TrimSpace(*body.Value.APIKey)
+		if trimmed == "" {
+			h.cfg.VertexCompatAPIKey = append(h.cfg.VertexCompatAPIKey[:targetIndex], h.cfg.VertexCompatAPIKey[targetIndex+1:]...)
+			h.cfg.SanitizeVertexCompatKeys()
+			h.persist(c)
+			return
+		}
+		entry.APIKey = trimmed
+	}
+	if body.Value.Prefix != nil {
+		entry.Prefix = strings.TrimSpace(*body.Value.Prefix)
+	}
+	if body.Value.BaseURL != nil {
+		trimmed := strings.TrimSpace(*body.Value.BaseURL)
+		if trimmed == "" {
+			h.cfg.VertexCompatAPIKey = append(h.cfg.VertexCompatAPIKey[:targetIndex], h.cfg.VertexCompatAPIKey[targetIndex+1:]...)
+			h.cfg.SanitizeVertexCompatKeys()
+			h.persist(c)
+			return
+		}
+		entry.BaseURL = trimmed
+	}
+	if body.Value.ProxyURL != nil {
+		entry.ProxyURL = strings.TrimSpace(*body.Value.ProxyURL)
+	}
+	if body.Value.Headers != nil {
+		entry.Headers = config.NormalizeHeaders(*body.Value.Headers)
+	}
+	if body.Value.Models != nil {
+		entry.Models = append([]config.VertexCompatModel(nil), (*body.Value.Models)...)
+	}
+	normalizeVertexCompatKey(&entry)
+	h.cfg.VertexCompatAPIKey[targetIndex] = entry
+	h.cfg.SanitizeVertexCompatKeys()
+	h.persist(c)
+}
+
+func (h *Handler) DeleteVertexCompatKey(c *gin.Context) {
+	if val := strings.TrimSpace(c.Query("api-key")); val != "" {
+		out := make([]config.VertexCompatKey, 0, len(h.cfg.VertexCompatAPIKey))
+		for _, v := range h.cfg.VertexCompatAPIKey {
+			if v.APIKey != val {
+				out = append(out, v)
+			}
+		}
+		h.cfg.VertexCompatAPIKey = out
+		h.cfg.SanitizeVertexCompatKeys()
+		h.persist(c)
+		return
+	}
+	if idxStr := c.Query("index"); idxStr != "" {
+		var idx int
+		_, errScan := fmt.Sscanf(idxStr, "%d", &idx)
+		if errScan == nil && idx >= 0 && idx < len(h.cfg.VertexCompatAPIKey) {
+			h.cfg.VertexCompatAPIKey = append(h.cfg.VertexCompatAPIKey[:idx], h.cfg.VertexCompatAPIKey[idx+1:]...)
+			h.cfg.SanitizeVertexCompatKeys()
+			h.persist(c)
+			return
+		}
+	}
+	c.JSON(400, gin.H{"error": "missing api-key or index"})
+}
+
 // oauth-excluded-models: map[string][]string
 func (h *Handler) GetOAuthExcludedModels(c *gin.Context) {
 	c.JSON(200, gin.H{"oauth-excluded-models": config.NormalizeOAuthExcludedModels(h.cfg.OAuthExcludedModels)})
@@ -523,6 +703,103 @@ func (h *Handler) DeleteOAuthExcludedModels(c *gin.Context) {
 	h.persist(c)
 }
 
+// oauth-model-alias: map[string][]OAuthModelAlias
+func (h *Handler) GetOAuthModelAlias(c *gin.Context) {
+	c.JSON(200, gin.H{"oauth-model-alias": sanitizedOAuthModelAlias(h.cfg.OAuthModelAlias)})
+}
+
+func (h *Handler) PutOAuthModelAlias(c *gin.Context) {
+	data, err := c.GetRawData()
+	if err != nil {
+		c.JSON(400, gin.H{"error": "failed to read body"})
+		return
+	}
+	var entries map[string][]config.OAuthModelAlias
+	if err = json.Unmarshal(data, &entries); err != nil {
+		var wrapper struct {
+			Items map[string][]config.OAuthModelAlias `json:"items"`
+		}
+		if err2 := json.Unmarshal(data, &wrapper); err2 != nil {
+			c.JSON(400, gin.H{"error": "invalid body"})
+			return
+		}
+		entries = wrapper.Items
+	}
+	h.cfg.OAuthModelAlias = sanitizedOAuthModelAlias(entries)
+	h.persist(c)
+}
+
+func (h *Handler) PatchOAuthModelAlias(c *gin.Context) {
+	var body struct {
+		Provider *string                  `json:"provider"`
+		Channel  *string                  `json:"channel"`
+		Aliases  []config.OAuthModelAlias `json:"aliases"`
+	}
+	if errBindJSON := c.ShouldBindJSON(&body); errBindJSON != nil {
+		c.JSON(400, gin.H{"error": "invalid body"})
+		return
+	}
+	channelRaw := ""
+	if body.Channel != nil {
+		channelRaw = *body.Channel
+	} else if body.Provider != nil {
+		channelRaw = *body.Provider
+	}
+	channel := strings.ToLower(strings.TrimSpace(channelRaw))
+	if channel == "" {
+		c.JSON(400, gin.H{"error": "invalid channel"})
+		return
+	}
+
+	normalizedMap := sanitizedOAuthModelAlias(map[string][]config.OAuthModelAlias{channel: body.Aliases})
+	normalized := normalizedMap[channel]
+	if len(normalized) == 0 {
+		if h.cfg.OAuthModelAlias == nil {
+			c.JSON(404, gin.H{"error": "channel not found"})
+			return
+		}
+		if _, ok := h.cfg.OAuthModelAlias[channel]; !ok {
+			c.JSON(404, gin.H{"error": "channel not found"})
+			return
+		}
+		delete(h.cfg.OAuthModelAlias, channel)
+		if len(h.cfg.OAuthModelAlias) == 0 {
+			h.cfg.OAuthModelAlias = nil
+		}
+		h.persist(c)
+		return
+	}
+	if h.cfg.OAuthModelAlias == nil {
+		h.cfg.OAuthModelAlias = make(map[string][]config.OAuthModelAlias)
+	}
+	h.cfg.OAuthModelAlias[channel] = normalized
+	h.persist(c)
+}
+
+func (h *Handler) DeleteOAuthModelAlias(c *gin.Context) {
+	channel := strings.ToLower(strings.TrimSpace(c.Query("channel")))
+	if channel == "" {
+		channel = strings.ToLower(strings.TrimSpace(c.Query("provider")))
+	}
+	if channel == "" {
+		c.JSON(400, gin.H{"error": "missing channel"})
+		return
+	}
+	if h.cfg.OAuthModelAlias == nil {
+		c.JSON(404, gin.H{"error": "channel not found"})
+		return
+	}
+	if _, ok := h.cfg.OAuthModelAlias[channel]; !ok {
+		c.JSON(404, gin.H{"error": "channel not found"})
+		return
+	}
+	delete(h.cfg.OAuthModelAlias, channel)
+	if len(h.cfg.OAuthModelAlias) == 0 {
+		h.cfg.OAuthModelAlias = nil
+	}
+	h.persist(c)
+}
+
 // codex-api-key: []CodexKey
 func (h *Handler) GetCodexKeys(c *gin.Context) {
 	c.JSON(200, gin.H{"codex-api-key": h.cfg.CodexKey})
@@ -548,11 +825,7 @@ func (h *Handler) PutCodexKeys(c *gin.Context) {
 	filtered := make([]config.CodexKey, 0, len(arr))
 	for i := range arr {
 		entry := arr[i]
-		entry.APIKey = strings.TrimSpace(entry.APIKey)
-		entry.BaseURL = strings.TrimSpace(entry.BaseURL)
-		entry.ProxyURL = strings.TrimSpace(entry.ProxyURL)
-		entry.Headers = config.NormalizeHeaders(entry.Headers)
-		entry.ExcludedModels = config.NormalizeExcludedModels(entry.ExcludedModels)
+		normalizeCodexKey(&entry)
 		if entry.BaseURL == "" {
 			continue
 		}
@@ -563,66 +836,77 @@ func (h *Handler) PutCodexKeys(c *gin.Context) {
 	h.persist(c)
 }
 func (h *Handler) PatchCodexKey(c *gin.Context) {
+	type codexKeyPatch struct {
+		APIKey         *string              `json:"api-key"`
+		Prefix         *string              `json:"prefix"`
+		BaseURL        *string              `json:"base-url"`
+		ProxyURL       *string              `json:"proxy-url"`
+		Models         *[]config.CodexModel `json:"models"`
+		Headers        *map[string]string   `json:"headers"`
+		ExcludedModels *[]string            `json:"excluded-models"`
+	}
 	var body struct {
-		Index *int             `json:"index"`
-		Match *string          `json:"match"`
-		Value *config.CodexKey `json:"value"`
+		Index *int           `json:"index"`
+		Match *string        `json:"match"`
+		Value *codexKeyPatch `json:"value"`
 	}
 	if err := c.ShouldBindJSON(&body); err != nil || body.Value == nil {
 		c.JSON(400, gin.H{"error": "invalid body"})
 		return
 	}
-	value := *body.Value
-	value.APIKey = strings.TrimSpace(value.APIKey)
-	value.BaseURL = strings.TrimSpace(value.BaseURL)
-	value.ProxyURL = strings.TrimSpace(value.ProxyURL)
-	value.Headers = config.NormalizeHeaders(value.Headers)
-	value.ExcludedModels = config.NormalizeExcludedModels(value.ExcludedModels)
-	// If base-url becomes empty, delete instead of update
-	if value.BaseURL == "" {
-		if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.CodexKey) {
-			h.cfg.CodexKey = append(h.cfg.CodexKey[:*body.Index], h.cfg.CodexKey[*body.Index+1:]...)
-			h.cfg.SanitizeCodexKeys()
-			h.persist(c)
-			return
-		}
-		if body.Match != nil {
-			out := make([]config.CodexKey, 0, len(h.cfg.CodexKey))
-			removed := false
-			for i := range h.cfg.CodexKey {
-				if !removed && h.cfg.CodexKey[i].APIKey == *body.Match {
-					removed = true
-					continue
-				}
-				out = append(out, h.cfg.CodexKey[i])
-			}
-			if removed {
-				h.cfg.CodexKey = out
-				h.cfg.SanitizeCodexKeys()
-				h.persist(c)
-				return
-			}
-		}
-	} else {
-		if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.CodexKey) {
-			h.cfg.CodexKey[*body.Index] = value
-			h.cfg.SanitizeCodexKeys()
-			h.persist(c)
-			return
-		}
-		if body.Match != nil {
-			for i := range h.cfg.CodexKey {
-				if h.cfg.CodexKey[i].APIKey == *body.Match {
-					h.cfg.CodexKey[i] = value
-					h.cfg.SanitizeCodexKeys()
-					h.persist(c)
-					return
-				}
+	targetIndex := -1
+	if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.CodexKey) {
+		targetIndex = *body.Index
+	}
+	if targetIndex == -1 && body.Match != nil {
+		match := strings.TrimSpace(*body.Match)
+		for i := range h.cfg.CodexKey {
+			if h.cfg.CodexKey[i].APIKey == match {
+				targetIndex = i
+				break
 			}
 		}
 	}
-	c.JSON(404, gin.H{"error": "item not found"})
+	if targetIndex == -1 {
+		c.JSON(404, gin.H{"error": "item not found"})
+		return
+	}
+
+	entry := h.cfg.CodexKey[targetIndex]
+	if body.Value.APIKey != nil {
+		entry.APIKey = strings.TrimSpace(*body.Value.APIKey)
+	}
+	if body.Value.Prefix != nil {
+		entry.Prefix = strings.TrimSpace(*body.Value.Prefix)
+	}
+	if body.Value.BaseURL != nil {
+		trimmed := strings.TrimSpace(*body.Value.BaseURL)
+		if trimmed == "" {
+			h.cfg.CodexKey = append(h.cfg.CodexKey[:targetIndex], h.cfg.CodexKey[targetIndex+1:]...)
+			h.cfg.SanitizeCodexKeys()
+			h.persist(c)
+			return
+		}
+		entry.BaseURL = trimmed
+	}
+	if body.Value.ProxyURL != nil {
+		entry.ProxyURL = strings.TrimSpace(*body.Value.ProxyURL)
+	}
+	if body.Value.Models != nil {
+		entry.Models = append([]config.CodexModel(nil), (*body.Value.Models)...)
+	}
+	if body.Value.Headers != nil {
+		entry.Headers = config.NormalizeHeaders(*body.Value.Headers)
+	}
+	if body.Value.ExcludedModels != nil {
+		entry.ExcludedModels = config.NormalizeExcludedModels(*body.Value.ExcludedModels)
+	}
+	normalizeCodexKey(&entry)
+	h.cfg.CodexKey[targetIndex] = entry
+	h.cfg.SanitizeCodexKeys()
+	h.persist(c)
 }
+
 func (h *Handler) DeleteCodexKey(c *gin.Context) {
 	if val := c.Query("api-key"); val != "" {
 		out := make([]config.CodexKey, 0, len(h.cfg.CodexKey))
@@ -707,6 +991,79 @@ func normalizeClaudeKey(entry *config.ClaudeKey) {
 	entry.Models = normalized
 }
 
+func normalizeCodexKey(entry *config.CodexKey) {
+	if entry == nil {
+		return
+	}
+	entry.APIKey = strings.TrimSpace(entry.APIKey)
+	entry.Prefix = strings.TrimSpace(entry.Prefix)
+	entry.BaseURL = strings.TrimSpace(entry.BaseURL)
+	entry.ProxyURL = strings.TrimSpace(entry.ProxyURL)
+	entry.Headers = config.NormalizeHeaders(entry.Headers)
+	entry.ExcludedModels = config.NormalizeExcludedModels(entry.ExcludedModels)
+	if len(entry.Models) == 0 {
+		return
+	}
+	normalized := make([]config.CodexModel, 0, len(entry.Models))
+	for i := range entry.Models {
+		model := entry.Models[i]
+		model.Name = strings.TrimSpace(model.Name)
+		model.Alias = strings.TrimSpace(model.Alias)
+		if model.Name == "" && model.Alias == "" {
+			continue
+		}
+		normalized = append(normalized, model)
+	}
+	entry.Models = normalized
+}
+
+func normalizeVertexCompatKey(entry *config.VertexCompatKey) {
+	if entry == nil {
+		return
+	}
+	entry.APIKey = strings.TrimSpace(entry.APIKey)
+	entry.Prefix = strings.TrimSpace(entry.Prefix)
+	entry.BaseURL = strings.TrimSpace(entry.BaseURL)
+	entry.ProxyURL = strings.TrimSpace(entry.ProxyURL)
+	entry.Headers = config.NormalizeHeaders(entry.Headers)
+	if len(entry.Models) == 0 {
+		return
+	}
+	normalized := make([]config.VertexCompatModel, 0, len(entry.Models))
+	for i := range entry.Models {
+		model := entry.Models[i]
+		model.Name = strings.TrimSpace(model.Name)
+		model.Alias = strings.TrimSpace(model.Alias)
+		if model.Name == "" || model.Alias == "" {
+			continue
+		}
+		normalized = append(normalized, model)
+	}
+	entry.Models = normalized
+}
+
+func sanitizedOAuthModelAlias(entries map[string][]config.OAuthModelAlias) map[string][]config.OAuthModelAlias {
+	if len(entries) == 0 {
+		return nil
+	}
+	copied := make(map[string][]config.OAuthModelAlias, len(entries))
+	for channel, aliases := range entries {
+		if len(aliases) == 0 {
+			continue
+		}
+		copied[channel] = append([]config.OAuthModelAlias(nil), aliases...)
+	}
+	if len(copied) == 0 {
+		return nil
+	}
+	cfg := config.Config{OAuthModelAlias: copied}
+	cfg.SanitizeOAuthModelAlias()
+	if len(cfg.OAuthModelAlias) == 0 {
+		return nil
+	}
+	return cfg.OAuthModelAlias
+}
+
 // GetAmpCode returns the complete ampcode configuration.
 func (h *Handler) GetAmpCode(c *gin.Context) {
 	if h == nil || h.cfg == nil {
@@ -858,3 +1215,151 @@ func (h *Handler) GetAmpForceModelMappings(c *gin.Context) {
 func (h *Handler) PutAmpForceModelMappings(c *gin.Context) {
 	h.updateBoolField(c, func(v bool) { h.cfg.AmpCode.ForceModelMappings = v })
 }
+
+// GetAmpUpstreamAPIKeys returns the ampcode upstream API keys mapping.
+func (h *Handler) GetAmpUpstreamAPIKeys(c *gin.Context) {
+	if h == nil || h.cfg == nil {
+		c.JSON(200, gin.H{"upstream-api-keys": []config.AmpUpstreamAPIKeyEntry{}})
+		return
+	}
+	c.JSON(200, gin.H{"upstream-api-keys": h.cfg.AmpCode.UpstreamAPIKeys})
+}
+
+// PutAmpUpstreamAPIKeys replaces all ampcode upstream API keys mappings.
+func (h *Handler) PutAmpUpstreamAPIKeys(c *gin.Context) {
+	var body struct {
+		Value []config.AmpUpstreamAPIKeyEntry `json:"value"`
+	}
+	if err := c.ShouldBindJSON(&body); err != nil {
+		c.JSON(400, gin.H{"error": "invalid body"})
+		return
+	}
+	// Normalize entries: trim whitespace, filter empty
+	normalized := normalizeAmpUpstreamAPIKeyEntries(body.Value)
+	h.cfg.AmpCode.UpstreamAPIKeys = normalized
+	h.persist(c)
+}
+
+// PatchAmpUpstreamAPIKeys adds or updates upstream API keys entries.
+// Matching is done by upstream-api-key value.
+func (h *Handler) PatchAmpUpstreamAPIKeys(c *gin.Context) {
+	var body struct {
+		Value []config.AmpUpstreamAPIKeyEntry `json:"value"`
+	}
+	if err := c.ShouldBindJSON(&body); err != nil {
+		c.JSON(400, gin.H{"error": "invalid body"})
+		return
+	}
+
+	existing := make(map[string]int)
+	for i, entry := range h.cfg.AmpCode.UpstreamAPIKeys {
+		existing[strings.TrimSpace(entry.UpstreamAPIKey)] = i
+	}
+
+	for _, newEntry := range body.Value {
+		upstreamKey := strings.TrimSpace(newEntry.UpstreamAPIKey)
+		if upstreamKey == "" {
+			continue
+		}
+		normalizedEntry := config.AmpUpstreamAPIKeyEntry{
+			UpstreamAPIKey: upstreamKey,
+			APIKeys:        normalizeAPIKeysList(newEntry.APIKeys),
+		}
+		if idx, ok := existing[upstreamKey]; ok {
+			h.cfg.AmpCode.UpstreamAPIKeys[idx] = normalizedEntry
+		} else {
+			h.cfg.AmpCode.UpstreamAPIKeys = append(h.cfg.AmpCode.UpstreamAPIKeys, normalizedEntry)
+			existing[upstreamKey] = len(h.cfg.AmpCode.UpstreamAPIKeys) - 1
+		}
+	}
+	h.persist(c)
+}
+
+// DeleteAmpUpstreamAPIKeys removes specified upstream API keys entries.
+// Body must be JSON: {"value": ["<upstream-api-key>", ...]}.
+// If "value" is an empty array, clears all entries.
+// If JSON is invalid or "value" is missing/null, returns 400 and does not persist any change.
+func (h *Handler) DeleteAmpUpstreamAPIKeys(c *gin.Context) {
+	var body struct {
+		Value []string `json:"value"`
+	}
+	if err := c.ShouldBindJSON(&body); err != nil {
+		c.JSON(400, gin.H{"error": "invalid body"})
+		return
+	}
+
+	if body.Value == nil {
+		c.JSON(400, gin.H{"error": "missing value"})
+		return
+	}
+
+	// Empty array means clear all
+	if len(body.Value) == 0 {
+		h.cfg.AmpCode.UpstreamAPIKeys = nil
+		h.persist(c)
+		return
+	}
+
+	toRemove := make(map[string]bool)
+	for _, key := range body.Value {
+		trimmed := strings.TrimSpace(key)
+		if trimmed == "" {
+			continue
+		}
+		toRemove[trimmed] = true
+	}
+	if len(toRemove) == 0 {
+		c.JSON(400, gin.H{"error": "empty value"})
+		return
+	}
+
+	newEntries := make([]config.AmpUpstreamAPIKeyEntry, 0, len(h.cfg.AmpCode.UpstreamAPIKeys))
+	for _, entry := range h.cfg.AmpCode.UpstreamAPIKeys {
+		if !toRemove[strings.TrimSpace(entry.UpstreamAPIKey)] {
+			newEntries = append(newEntries, entry)
+		}
+	}
+	h.cfg.AmpCode.UpstreamAPIKeys = newEntries
+	h.persist(c)
+}
+
+// normalizeAmpUpstreamAPIKeyEntries normalizes a list of upstream API key entries.
+func normalizeAmpUpstreamAPIKeyEntries(entries []config.AmpUpstreamAPIKeyEntry) []config.AmpUpstreamAPIKeyEntry {
+	if len(entries) == 0 {
+		return nil
+	}
+	out := make([]config.AmpUpstreamAPIKeyEntry, 0, len(entries))
+	for _, entry := range entries {
+		upstreamKey := strings.TrimSpace(entry.UpstreamAPIKey)
+		if upstreamKey == "" {
+			continue
+		}
+		apiKeys := normalizeAPIKeysList(entry.APIKeys)
+		out = append(out, config.AmpUpstreamAPIKeyEntry{
+			UpstreamAPIKey: upstreamKey,
+			APIKeys:        apiKeys,
+		})
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
+// normalizeAPIKeysList trims and filters empty strings from a list of API keys.
+func normalizeAPIKeysList(keys []string) []string {
+	if len(keys) == 0 {
+		return nil
+	}
+	out := make([]string, 0, len(keys))
+	for _, k := range keys {
+		trimmed := strings.TrimSpace(k)
+		if trimmed != "" {
+			out = append(out, trimmed)
+		}
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}

internal/api/handlers/management/handler.go

@@ -24,8 +24,15 @@ import (
 type attemptInfo struct {
 	count        int
 	blockedUntil time.Time
+	lastActivity time.Time // track last activity for cleanup
 }
 
+// attemptCleanupInterval controls how often stale IP entries are purged
+const attemptCleanupInterval = 1 * time.Hour
+
+// attemptMaxIdleTime controls how long an IP can be idle before cleanup
+const attemptMaxIdleTime = 2 * time.Hour
+
 // Handler aggregates config reference, persistence path and helpers.
 type Handler struct {
 	cfg                 *config.Config
@@ -47,7 +54,7 @@ func NewHandler(cfg *config.Config, configFilePath string, manager *coreauth.Man
 	envSecret, _ := os.LookupEnv("MANAGEMENT_PASSWORD")
 	envSecret = strings.TrimSpace(envSecret)
 
-	return &Handler{
+	h := &Handler{
 		cfg:                 cfg,
 		configFilePath:      configFilePath,
 		failedAttempts:      make(map[string]*attemptInfo),
@@ -57,6 +64,43 @@ func NewHandler(cfg *config.Config, configFilePath string, manager *coreauth.Man
 		allowRemoteOverride: envSecret != "",
 		envSecret:           envSecret,
 	}
+	h.startAttemptCleanup()
+	return h
+}
+
+// startAttemptCleanup launches a background goroutine that periodically
+// removes stale IP entries from failedAttempts to prevent memory leaks.
+func (h *Handler) startAttemptCleanup() {
+	go func() {
+		ticker := time.NewTicker(attemptCleanupInterval)
+		defer ticker.Stop()
+		for range ticker.C {
+			h.purgeStaleAttempts()
+		}
+	}()
+}
+
+// purgeStaleAttempts removes IP entries that have been idle beyond attemptMaxIdleTime
+// and whose ban (if any) has expired.
+func (h *Handler) purgeStaleAttempts() {
+	now := time.Now()
+	h.attemptsMu.Lock()
+	defer h.attemptsMu.Unlock()
+	for ip, ai := range h.failedAttempts {
+		// Skip if still banned
+		if !ai.blockedUntil.IsZero() && now.Before(ai.blockedUntil) {
+			continue
+		}
+		// Remove if idle too long
+		if now.Sub(ai.lastActivity) > attemptMaxIdleTime {
+			delete(h.failedAttempts, ip)
+		}
+	}
+}
+
+// NewHandler creates a new management handler instance.
+func NewHandlerWithoutConfigFilePath(cfg *config.Config, manager *coreauth.Manager) *Handler {
+	return NewHandler(cfg, "", manager)
 }
 
 // SetConfig updates the in-memory config reference when the server hot-reloads.
@@ -144,6 +188,7 @@ func (h *Handler) Middleware() gin.HandlerFunc {
 					h.failedAttempts[clientIP] = aip
 				}
 				aip.count++
+				aip.lastActivity = time.Now()
 				if aip.count >= maxFailures {
 					aip.blockedUntil = time.Now().Add(banDuration)
 					aip.count = 0

internal/api/handlers/management/logs.go

@@ -209,6 +209,94 @@ func (h *Handler) GetRequestErrorLogs(c *gin.Context) {
 	c.JSON(http.StatusOK, gin.H{"files": files})
 }
 
+// GetRequestLogByID finds and downloads a request log file by its request ID.
+// The ID is matched against the suffix of log file names (format: *-{requestID}.log).
+func (h *Handler) GetRequestLogByID(c *gin.Context) {
+	if h == nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "handler unavailable"})
+		return
+	}
+	if h.cfg == nil {
+		c.JSON(http.StatusServiceUnavailable, gin.H{"error": "configuration unavailable"})
+		return
+	}
+
+	dir := h.logDirectory()
+	if strings.TrimSpace(dir) == "" {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "log directory not configured"})
+		return
+	}
+
+	requestID := strings.TrimSpace(c.Param("id"))
+	if requestID == "" {
+		requestID = strings.TrimSpace(c.Query("id"))
+	}
+	if requestID == "" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "missing request ID"})
+		return
+	}
+	if strings.ContainsAny(requestID, "/\\") {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid request ID"})
+		return
+	}
+
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		if os.IsNotExist(err) {
+			c.JSON(http.StatusNotFound, gin.H{"error": "log directory not found"})
+			return
+		}
+		c.JSON(http.StatusInternalServerError, gin.H{"error": fmt.Sprintf("failed to list log directory: %v", err)})
+		return
+	}
+
+	suffix := "-" + requestID + ".log"
+	var matchedFile string
+	for _, entry := range entries {
+		if entry.IsDir() {
+			continue
+		}
+		name := entry.Name()
+		if strings.HasSuffix(name, suffix) {
+			matchedFile = name
+			break
+		}
+	}
+
+	if matchedFile == "" {
+		c.JSON(http.StatusNotFound, gin.H{"error": "log file not found for the given request ID"})
+		return
+	}
+
+	dirAbs, errAbs := filepath.Abs(dir)
+	if errAbs != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": fmt.Sprintf("failed to resolve log directory: %v", errAbs)})
+		return
+	}
+	fullPath := filepath.Clean(filepath.Join(dirAbs, matchedFile))
+	prefix := dirAbs + string(os.PathSeparator)
+	if !strings.HasPrefix(fullPath, prefix) {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid log file path"})
+		return
+	}
+
+	info, errStat := os.Stat(fullPath)
+	if errStat != nil {
+		if os.IsNotExist(errStat) {
+			c.JSON(http.StatusNotFound, gin.H{"error": "log file not found"})
+			return
+		}
+		c.JSON(http.StatusInternalServerError, gin.H{"error": fmt.Sprintf("failed to read log file: %v", errStat)})
+		return
+	}
+	if info.IsDir() {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid log file"})
+		return
+	}
+
+	c.FileAttachment(fullPath, matchedFile)
+}
+
 // DownloadRequestErrorLog downloads a specific error request log file by name.
 func (h *Handler) DownloadRequestErrorLog(c *gin.Context) {
 	if h == nil {

internal/api/handlers/management/oauth_sessions.go

@@ -111,6 +111,27 @@ func (s *oauthSessionStore) Complete(state string) {
 	delete(s.sessions, state)
 }
 
+func (s *oauthSessionStore) CompleteProvider(provider string) int {
+	provider = strings.ToLower(strings.TrimSpace(provider))
+	if provider == "" {
+		return 0
+	}
+	now := time.Now()
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	s.purgeExpiredLocked(now)
+	removed := 0
+	for state, session := range s.sessions {
+		if strings.EqualFold(session.Provider, provider) {
+			delete(s.sessions, state)
+			removed++
+		}
+	}
+	return removed
+}
+
 func (s *oauthSessionStore) Get(state string) (oauthSession, bool) {
 	state = strings.TrimSpace(state)
 	now := time.Now()
@@ -153,6 +174,10 @@ func SetOAuthSessionError(state, message string) { oauthSessions.SetError(state,
 
 func CompleteOAuthSession(state string) { oauthSessions.Complete(state) }
 
+func CompleteOAuthSessionsByProvider(provider string) int {
+	return oauthSessions.CompleteProvider(provider)
+}
+
 func GetOAuthSession(state string) (provider string, status string, ok bool) {
 	session, ok := oauthSessions.Get(state)
 	if !ok {

internal/api/handlers/management/usage.go

@@ -1,12 +1,25 @@
 package management
 
 import (
+	"encoding/json"
 	"net/http"
+	"time"
 
 	"github.com/gin-gonic/gin"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/usage"
 )
 
+type usageExportPayload struct {
+	Version    int                      `json:"version"`
+	ExportedAt time.Time                `json:"exported_at"`
+	Usage      usage.StatisticsSnapshot `json:"usage"`
+}
+
+type usageImportPayload struct {
+	Version int                      `json:"version"`
+	Usage   usage.StatisticsSnapshot `json:"usage"`
+}
+
 // GetUsageStatistics returns the in-memory request statistics snapshot.
 func (h *Handler) GetUsageStatistics(c *gin.Context) {
 	var snapshot usage.StatisticsSnapshot
@@ -18,3 +31,49 @@ func (h *Handler) GetUsageStatistics(c *gin.Context) {
 		"failed_requests": snapshot.FailureCount,
 	})
 }
+
+// ExportUsageStatistics returns a complete usage snapshot for backup/migration.
+func (h *Handler) ExportUsageStatistics(c *gin.Context) {
+	var snapshot usage.StatisticsSnapshot
+	if h != nil && h.usageStats != nil {
+		snapshot = h.usageStats.Snapshot()
+	}
+	c.JSON(http.StatusOK, usageExportPayload{
+		Version:    1,
+		ExportedAt: time.Now().UTC(),
+		Usage:      snapshot,
+	})
+}
+
+// ImportUsageStatistics merges a previously exported usage snapshot into memory.
+func (h *Handler) ImportUsageStatistics(c *gin.Context) {
+	if h == nil || h.usageStats == nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "usage statistics unavailable"})
+		return
+	}
+
+	data, err := c.GetRawData()
+	if err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "failed to read request body"})
+		return
+	}
+
+	var payload usageImportPayload
+	if err := json.Unmarshal(data, &payload); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid json"})
+		return
+	}
+	if payload.Version != 0 && payload.Version != 1 {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "unsupported version"})
+		return
+	}
+
+	result := h.usageStats.MergeSnapshot(payload.Usage)
+	snapshot := h.usageStats.Snapshot()
+	c.JSON(http.StatusOK, gin.H{
+		"added":           result.Added,
+		"skipped":         result.Skipped,
+		"total_requests":  snapshot.TotalRequests,
+		"failed_requests": snapshot.FailureCount,
+	})
+}

internal/api/middleware/request_logging.go

@@ -98,10 +98,11 @@ func captureRequestInfo(c *gin.Context) (*RequestInfo, error) {
 	}
 
 	return &RequestInfo{
-		URL:     url,
-		Method:  method,
-		Headers: headers,
-		Body:    body,
+		URL:       url,
+		Method:    method,
+		Headers:   headers,
+		Body:      body,
+		RequestID: logging.GetGinRequestID(c),
 	}, nil
 }
 

internal/api/middleware/response_writer.go

@@ -15,10 +15,11 @@ import (
 
 // RequestInfo holds essential details of an incoming HTTP request for logging purposes.
 type RequestInfo struct {
-	URL     string              // URL is the request URL.
-	Method  string              // Method is the HTTP method (e.g., GET, POST).
-	Headers map[string][]string // Headers contains the request headers.
-	Body    []byte              // Body is the raw request body.
+	URL       string              // URL is the request URL.
+	Method    string              // Method is the HTTP method (e.g., GET, POST).
+	Headers   map[string][]string // Headers contains the request headers.
+	Body      []byte              // Body is the raw request body.
+	RequestID string              // RequestID is the unique identifier for the request.
 }
 
 // ResponseWriterWrapper wraps the standard gin.ResponseWriter to intercept and log response data.
@@ -149,6 +150,7 @@ func (w *ResponseWriterWrapper) WriteHeader(statusCode int) {
 			w.requestInfo.Method,
 			w.requestInfo.Headers,
 			w.requestInfo.Body,
+			w.requestInfo.RequestID,
 		)
 		if err == nil {
 			w.streamWriter = streamWriter
@@ -346,7 +348,7 @@ func (w *ResponseWriterWrapper) logRequest(statusCode int, headers map[string][]
 	}
 
 	if loggerWithOptions, ok := w.logger.(interface {
-		LogRequestWithOptions(string, string, map[string][]string, []byte, int, map[string][]string, []byte, []byte, []byte, []*interfaces.ErrorMessage, bool) error
+		LogRequestWithOptions(string, string, map[string][]string, []byte, int, map[string][]string, []byte, []byte, []byte, []*interfaces.ErrorMessage, bool, string) error
 	}); ok {
 		return loggerWithOptions.LogRequestWithOptions(
 			w.requestInfo.URL,
@@ -360,6 +362,7 @@ func (w *ResponseWriterWrapper) logRequest(statusCode int, headers map[string][]
 			apiResponseBody,
 			apiResponseErrors,
 			forceLog,
+			w.requestInfo.RequestID,
 		)
 	}
 
@@ -374,5 +377,6 @@ func (w *ResponseWriterWrapper) logRequest(statusCode int, headers map[string][]
 		apiRequestBody,
 		apiResponseBody,
 		apiResponseErrors,
+		w.requestInfo.RequestID,
 	)
 }

internal/api/modules/amp/amp.go

@@ -227,11 +227,20 @@ func (m *AmpModule) OnConfigUpdated(cfg *config.Config) error {
 			}
 		}
 
-		// Check API key change
+		// Check API key change (both default and per-client mappings)
 		apiKeyChanged := m.hasAPIKeyChanged(oldSettings, &newSettings)
-		if apiKeyChanged {
+		upstreamAPIKeysChanged := m.hasUpstreamAPIKeysChanged(oldSettings, &newSettings)
+		if apiKeyChanged || upstreamAPIKeysChanged {
 			if m.secretSource != nil {
-				if ms, ok := m.secretSource.(*MultiSourceSecret); ok {
+				if ms, ok := m.secretSource.(*MappedSecretSource); ok {
+					if apiKeyChanged {
+						ms.UpdateDefaultExplicitKey(newSettings.UpstreamAPIKey)
+						ms.InvalidateCache()
+					}
+					if upstreamAPIKeysChanged {
+						ms.UpdateMappings(newSettings.UpstreamAPIKeys)
+					}
+				} else if ms, ok := m.secretSource.(*MultiSourceSecret); ok {
 					ms.UpdateExplicitKey(newSettings.UpstreamAPIKey)
 					ms.InvalidateCache()
 				}
@@ -251,10 +260,22 @@ func (m *AmpModule) OnConfigUpdated(cfg *config.Config) error {
 
 func (m *AmpModule) enableUpstreamProxy(upstreamURL string, settings *config.AmpCode) error {
 	if m.secretSource == nil {
-		m.secretSource = NewMultiSourceSecret(settings.UpstreamAPIKey, 0 /* default 5min */)
+		// Create MultiSourceSecret as the default source, then wrap with MappedSecretSource
+		defaultSource := NewMultiSourceSecret(settings.UpstreamAPIKey, 0 /* default 5min */)
+		mappedSource := NewMappedSecretSource(defaultSource)
+		mappedSource.UpdateMappings(settings.UpstreamAPIKeys)
+		m.secretSource = mappedSource
+	} else if ms, ok := m.secretSource.(*MappedSecretSource); ok {
+		ms.UpdateDefaultExplicitKey(settings.UpstreamAPIKey)
+		ms.InvalidateCache()
+		ms.UpdateMappings(settings.UpstreamAPIKeys)
 	} else if ms, ok := m.secretSource.(*MultiSourceSecret); ok {
+		// Legacy path: wrap existing MultiSourceSecret with MappedSecretSource
 		ms.UpdateExplicitKey(settings.UpstreamAPIKey)
 		ms.InvalidateCache()
+		mappedSource := NewMappedSecretSource(ms)
+		mappedSource.UpdateMappings(settings.UpstreamAPIKeys)
+		m.secretSource = mappedSource
 	}
 
 	proxy, err := createReverseProxy(upstreamURL, m.secretSource)
@@ -279,16 +300,23 @@ func (m *AmpModule) hasModelMappingsChanged(old *config.AmpCode, new *config.Amp
 		return true
 	}
 
-	// Build map for efficient comparison
-	oldMap := make(map[string]string, len(old.ModelMappings))
+	// Build map for efficient and robust comparison
+	type mappingInfo struct {
+		to    string
+		regex bool
+	}
+	oldMap := make(map[string]mappingInfo, len(old.ModelMappings))
 	for _, mapping := range old.ModelMappings {
-		oldMap[strings.TrimSpace(mapping.From)] = strings.TrimSpace(mapping.To)
+		oldMap[strings.TrimSpace(mapping.From)] = mappingInfo{
+			to:    strings.TrimSpace(mapping.To),
+			regex: mapping.Regex,
+		}
 	}
 
 	for _, mapping := range new.ModelMappings {
 		from := strings.TrimSpace(mapping.From)
 		to := strings.TrimSpace(mapping.To)
-		if oldTo, exists := oldMap[from]; !exists || oldTo != to {
+		if oldVal, exists := oldMap[from]; !exists || oldVal.to != to || oldVal.regex != mapping.Regex {
 			return true
 		}
 	}
@@ -306,6 +334,66 @@ func (m *AmpModule) hasAPIKeyChanged(old *config.AmpCode, new *config.AmpCode) b
 	return oldKey != newKey
 }
 
+// hasUpstreamAPIKeysChanged compares old and new per-client upstream API key mappings.
+func (m *AmpModule) hasUpstreamAPIKeysChanged(old *config.AmpCode, new *config.AmpCode) bool {
+	if old == nil {
+		return len(new.UpstreamAPIKeys) > 0
+	}
+
+	if len(old.UpstreamAPIKeys) != len(new.UpstreamAPIKeys) {
+		return true
+	}
+
+	// Build map for comparison: upstreamKey -> set of clientKeys
+	type entryInfo struct {
+		upstreamKey string
+		clientKeys  map[string]struct{}
+	}
+	oldEntries := make([]entryInfo, len(old.UpstreamAPIKeys))
+	for i, entry := range old.UpstreamAPIKeys {
+		clientKeys := make(map[string]struct{}, len(entry.APIKeys))
+		for _, k := range entry.APIKeys {
+			trimmed := strings.TrimSpace(k)
+			if trimmed == "" {
+				continue
+			}
+			clientKeys[trimmed] = struct{}{}
+		}
+		oldEntries[i] = entryInfo{
+			upstreamKey: strings.TrimSpace(entry.UpstreamAPIKey),
+			clientKeys:  clientKeys,
+		}
+	}
+
+	for i, newEntry := range new.UpstreamAPIKeys {
+		if i >= len(oldEntries) {
+			return true
+		}
+		oldE := oldEntries[i]
+		if strings.TrimSpace(newEntry.UpstreamAPIKey) != oldE.upstreamKey {
+			return true
+		}
+		newKeys := make(map[string]struct{}, len(newEntry.APIKeys))
+		for _, k := range newEntry.APIKeys {
+			trimmed := strings.TrimSpace(k)
+			if trimmed == "" {
+				continue
+			}
+			newKeys[trimmed] = struct{}{}
+		}
+		if len(newKeys) != len(oldE.clientKeys) {
+			return true
+		}
+		for k := range newKeys {
+			if _, ok := oldE.clientKeys[k]; !ok {
+				return true
+			}
+		}
+	}
+
+	return false
+}
+
 // GetModelMapper returns the model mapper instance (for testing/debugging).
 func (m *AmpModule) GetModelMapper() *DefaultModelMapper {
 	return m.modelMapper

internal/api/modules/amp/amp_test.go

@@ -312,3 +312,41 @@ func TestAmpModule_ProviderAliasesAlwaysRegistered(t *testing.T) {
 		})
 	}
 }
+
+func TestAmpModule_hasUpstreamAPIKeysChanged_DetectsRemovedKeyWithDuplicateInput(t *testing.T) {
+	m := &AmpModule{}
+
+	oldCfg := &config.AmpCode{
+		UpstreamAPIKeys: []config.AmpUpstreamAPIKeyEntry{
+			{UpstreamAPIKey: "u1", APIKeys: []string{"k1", "k2"}},
+		},
+	}
+	newCfg := &config.AmpCode{
+		UpstreamAPIKeys: []config.AmpUpstreamAPIKeyEntry{
+			{UpstreamAPIKey: "u1", APIKeys: []string{"k1", "k1"}},
+		},
+	}
+
+	if !m.hasUpstreamAPIKeysChanged(oldCfg, newCfg) {
+		t.Fatal("expected change to be detected when k2 is removed but new list contains duplicates")
+	}
+}
+
+func TestAmpModule_hasUpstreamAPIKeysChanged_IgnoresEmptyAndWhitespaceKeys(t *testing.T) {
+	m := &AmpModule{}
+
+	oldCfg := &config.AmpCode{
+		UpstreamAPIKeys: []config.AmpUpstreamAPIKeyEntry{
+			{UpstreamAPIKey: "u1", APIKeys: []string{"k1", "k2"}},
+		},
+	}
+	newCfg := &config.AmpCode{
+		UpstreamAPIKeys: []config.AmpUpstreamAPIKeyEntry{
+			{UpstreamAPIKey: "u1", APIKeys: []string{"  k1  ", "", "k2", "   "}},
+		},
+	}
+
+	if m.hasUpstreamAPIKeysChanged(oldCfg, newCfg) {
+		t.Fatal("expected no change when only whitespace/empty entries differ")
+	}
+}

internal/api/modules/amp/fallback_handlers.go

@@ -8,6 +8,7 @@ import (
 	"time"
 
 	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
@@ -134,10 +135,11 @@ func (fh *FallbackHandler) WrapHandler(handler gin.HandlerFunc) gin.HandlerFunc
 		}
 
 		// Normalize model (handles dynamic thinking suffixes)
-		normalizedModel, thinkingMetadata := util.NormalizeThinkingModel(modelName)
+		suffixResult := thinking.ParseSuffix(modelName)
+		normalizedModel := suffixResult.ModelName
 		thinkingSuffix := ""
-		if thinkingMetadata != nil && strings.HasPrefix(modelName, normalizedModel) {
-			thinkingSuffix = modelName[len(normalizedModel):]
+		if suffixResult.HasSuffix {
+			thinkingSuffix = "(" + suffixResult.RawSuffix + ")"
 		}
 
 		resolveMappedModel := func() (string, []string) {
@@ -157,13 +159,13 @@ func (fh *FallbackHandler) WrapHandler(handler gin.HandlerFunc) gin.HandlerFunc
 			// Preserve dynamic thinking suffix (e.g. "(xhigh)") when mapping applies, unless the target
 			// already specifies its own thinking suffix.
 			if thinkingSuffix != "" {
-				_, mappedThinkingMetadata := util.NormalizeThinkingModel(mappedModel)
-				if mappedThinkingMetadata == nil {
+				mappedSuffixResult := thinking.ParseSuffix(mappedModel)
+				if !mappedSuffixResult.HasSuffix {
 					mappedModel += thinkingSuffix
 				}
 			}
 
-			mappedBaseModel, _ := util.NormalizeThinkingModel(mappedModel)
+			mappedBaseModel := thinking.ParseSuffix(mappedModel).ModelName
 			mappedProviders := util.GetProviderName(mappedBaseModel)
 			if len(mappedProviders) == 0 {
 				return "", nil

internal/api/modules/amp/model_mapping.go

@@ -3,10 +3,12 @@
 package amp
 
 import (
+	"regexp"
 	"strings"
 	"sync"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	log "github.com/sirupsen/logrus"
 )
@@ -26,13 +28,15 @@ type ModelMapper interface {
 // DefaultModelMapper implements ModelMapper with thread-safe mapping storage.
 type DefaultModelMapper struct {
 	mu       sync.RWMutex
-	mappings map[string]string // from -> to (normalized lowercase keys)
+	mappings map[string]string // exact: from -> to (normalized lowercase keys)
+	regexps  []regexMapping    // regex rules evaluated in order
 }
 
 // NewModelMapper creates a new model mapper with the given initial mappings.
 func NewModelMapper(mappings []config.AmpModelMapping) *DefaultModelMapper {
 	m := &DefaultModelMapper{
 		mappings: make(map[string]string),
+		regexps:  nil,
 	}
 	m.UpdateMappings(mappings)
 	return m
@@ -41,6 +45,11 @@ func NewModelMapper(mappings []config.AmpModelMapping) *DefaultModelMapper {
 // MapModel checks if a mapping exists for the requested model and if the
 // target model has available local providers. Returns the mapped model name
 // or empty string if no valid mapping exists.
+//
+// If the requested model contains a thinking suffix (e.g., "g25p(8192)"),
+// the suffix is preserved in the returned model name (e.g., "gemini-2.5-pro(8192)").
+// However, if the mapping target already contains a suffix, the config suffix
+// takes priority over the user's suffix.
 func (m *DefaultModelMapper) MapModel(requestedModel string) string {
 	if requestedModel == "" {
 		return ""
@@ -49,23 +58,52 @@ func (m *DefaultModelMapper) MapModel(requestedModel string) string {
 	m.mu.RLock()
 	defer m.mu.RUnlock()
 
-	// Normalize the requested model for lookup
-	normalizedRequest := strings.ToLower(strings.TrimSpace(requestedModel))
+	// Extract thinking suffix from requested model using ParseSuffix
+	requestResult := thinking.ParseSuffix(requestedModel)
+	baseModel := requestResult.ModelName
 
-	// Check for direct mapping
-	targetModel, exists := m.mappings[normalizedRequest]
+	// Normalize the base model for lookup (case-insensitive)
+	normalizedBase := strings.ToLower(strings.TrimSpace(baseModel))
+
+	// Check for direct mapping using base model name
+	targetModel, exists := m.mappings[normalizedBase]
 	if !exists {
-		return ""
+		// Try regex mappings in order using base model only
+		// (suffix is handled separately via ParseSuffix)
+		for _, rm := range m.regexps {
+			if rm.re.MatchString(baseModel) {
+				targetModel = rm.to
+				exists = true
+				break
+			}
+		}
+		if !exists {
+			return ""
+		}
 	}
 
-	// Verify target model has available providers
-	normalizedTarget, _ := util.NormalizeThinkingModel(targetModel)
-	providers := util.GetProviderName(normalizedTarget)
+	// Check if target model already has a thinking suffix (config priority)
+	targetResult := thinking.ParseSuffix(targetModel)
+
+	// Verify target model has available providers (use base model for lookup)
+	providers := util.GetProviderName(targetResult.ModelName)
 	if len(providers) == 0 {
 		log.Debugf("amp model mapping: target model %s has no available providers, skipping mapping", targetModel)
 		return ""
 	}
 
+	// Suffix handling: config suffix takes priority, otherwise preserve user suffix
+	if targetResult.HasSuffix {
+		// Config's "to" already contains a suffix - use it as-is (config priority)
+		return targetModel
+	}
+
+	// Preserve user's thinking suffix on the mapped model
+	// (skip empty suffixes to avoid returning "model()")
+	if requestResult.HasSuffix && requestResult.RawSuffix != "" {
+		return targetModel + "(" + requestResult.RawSuffix + ")"
+	}
+
 	// Note: Detailed routing log is handled by logAmpRouting in fallback_handlers.go
 	return targetModel
 }
@@ -78,6 +116,7 @@ func (m *DefaultModelMapper) UpdateMappings(mappings []config.AmpModelMapping) {
 
 	// Clear and rebuild mappings
 	m.mappings = make(map[string]string, len(mappings))
+	m.regexps = make([]regexMapping, 0, len(mappings))
 
 	for _, mapping := range mappings {
 		from := strings.TrimSpace(mapping.From)
@@ -88,16 +127,30 @@ func (m *DefaultModelMapper) UpdateMappings(mappings []config.AmpModelMapping) {
 			continue
 		}
 
-		// Store with normalized lowercase key for case-insensitive lookup
-		normalizedFrom := strings.ToLower(from)
-		m.mappings[normalizedFrom] = to
-
-		log.Debugf("amp model mapping registered: %s -> %s", from, to)
+		if mapping.Regex {
+			// Compile case-insensitive regex; wrap with (?i) to match behavior of exact lookups
+			pattern := "(?i)" + from
+			re, err := regexp.Compile(pattern)
+			if err != nil {
+				log.Warnf("amp model mapping: invalid regex %q: %v", from, err)
+				continue
+			}
+			m.regexps = append(m.regexps, regexMapping{re: re, to: to})
+			log.Debugf("amp model regex mapping registered: /%s/ -> %s", from, to)
+		} else {
+			// Store with normalized lowercase key for case-insensitive lookup
+			normalizedFrom := strings.ToLower(from)
+			m.mappings[normalizedFrom] = to
+			log.Debugf("amp model mapping registered: %s -> %s", from, to)
+		}
 	}
 
 	if len(m.mappings) > 0 {
 		log.Infof("amp model mapping: loaded %d mapping(s)", len(m.mappings))
 	}
+	if n := len(m.regexps); n > 0 {
+		log.Infof("amp model mapping: loaded %d regex mapping(s)", n)
+	}
 }
 
 // GetMappings returns a copy of current mappings (for debugging/status).
@@ -111,3 +164,8 @@ func (m *DefaultModelMapper) GetMappings() map[string]string {
 	}
 	return result
 }
+
+type regexMapping struct {
+	re *regexp.Regexp
+	to string
+}

internal/api/modules/amp/model_mapping_test.go

@@ -203,3 +203,173 @@ func TestModelMapper_GetMappings_ReturnsCopy(t *testing.T) {
 		t.Error("Original map was modified")
 	}
 }
+
+func TestModelMapper_Regex_MatchBaseWithoutParens(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient("test-client-regex-1", "gemini", []*registry.ModelInfo{
+		{ID: "gemini-2.5-pro", OwnedBy: "google", Type: "gemini"},
+	})
+	defer reg.UnregisterClient("test-client-regex-1")
+
+	mappings := []config.AmpModelMapping{
+		{From: "^gpt-5$", To: "gemini-2.5-pro", Regex: true},
+	}
+
+	mapper := NewModelMapper(mappings)
+
+	// Incoming model has reasoning suffix, regex matches base, suffix is preserved
+	result := mapper.MapModel("gpt-5(high)")
+	if result != "gemini-2.5-pro(high)" {
+		t.Errorf("Expected gemini-2.5-pro(high), got %s", result)
+	}
+}
+
+func TestModelMapper_Regex_ExactPrecedence(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient("test-client-regex-2", "claude", []*registry.ModelInfo{
+		{ID: "claude-sonnet-4", OwnedBy: "anthropic", Type: "claude"},
+	})
+	reg.RegisterClient("test-client-regex-3", "gemini", []*registry.ModelInfo{
+		{ID: "gemini-2.5-pro", OwnedBy: "google", Type: "gemini"},
+	})
+	defer reg.UnregisterClient("test-client-regex-2")
+	defer reg.UnregisterClient("test-client-regex-3")
+
+	mappings := []config.AmpModelMapping{
+		{From: "gpt-5", To: "claude-sonnet-4"},                 // exact
+		{From: "^gpt-5.*$", To: "gemini-2.5-pro", Regex: true}, // regex
+	}
+
+	mapper := NewModelMapper(mappings)
+
+	// Exact match should win over regex
+	result := mapper.MapModel("gpt-5")
+	if result != "claude-sonnet-4" {
+		t.Errorf("Expected claude-sonnet-4, got %s", result)
+	}
+}
+
+func TestModelMapper_Regex_InvalidPattern_Skipped(t *testing.T) {
+	// Invalid regex should be skipped and not cause panic
+	mappings := []config.AmpModelMapping{
+		{From: "(", To: "target", Regex: true},
+	}
+
+	mapper := NewModelMapper(mappings)
+
+	result := mapper.MapModel("anything")
+	if result != "" {
+		t.Errorf("Expected empty result due to invalid regex, got %s", result)
+	}
+}
+
+func TestModelMapper_Regex_CaseInsensitive(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient("test-client-regex-4", "claude", []*registry.ModelInfo{
+		{ID: "claude-sonnet-4", OwnedBy: "anthropic", Type: "claude"},
+	})
+	defer reg.UnregisterClient("test-client-regex-4")
+
+	mappings := []config.AmpModelMapping{
+		{From: "^CLAUDE-OPUS-.*$", To: "claude-sonnet-4", Regex: true},
+	}
+
+	mapper := NewModelMapper(mappings)
+
+	result := mapper.MapModel("claude-opus-4.5")
+	if result != "claude-sonnet-4" {
+		t.Errorf("Expected claude-sonnet-4, got %s", result)
+	}
+}
+
+func TestModelMapper_SuffixPreservation(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+
+	// Register test models
+	reg.RegisterClient("test-client-suffix", "gemini", []*registry.ModelInfo{
+		{ID: "gemini-2.5-pro", OwnedBy: "google", Type: "gemini"},
+	})
+	reg.RegisterClient("test-client-suffix-2", "claude", []*registry.ModelInfo{
+		{ID: "claude-sonnet-4", OwnedBy: "anthropic", Type: "claude"},
+	})
+	defer reg.UnregisterClient("test-client-suffix")
+	defer reg.UnregisterClient("test-client-suffix-2")
+
+	tests := []struct {
+		name     string
+		mappings []config.AmpModelMapping
+		input    string
+		want     string
+	}{
+		{
+			name:     "numeric suffix preserved",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p(8192)",
+			want:     "gemini-2.5-pro(8192)",
+		},
+		{
+			name:     "level suffix preserved",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p(high)",
+			want:     "gemini-2.5-pro(high)",
+		},
+		{
+			name:     "no suffix unchanged",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p",
+			want:     "gemini-2.5-pro",
+		},
+		{
+			name:     "config suffix takes priority",
+			mappings: []config.AmpModelMapping{{From: "alias", To: "gemini-2.5-pro(medium)"}},
+			input:    "alias(high)",
+			want:     "gemini-2.5-pro(medium)",
+		},
+		{
+			name:     "regex with suffix preserved",
+			mappings: []config.AmpModelMapping{{From: "^g25.*", To: "gemini-2.5-pro", Regex: true}},
+			input:    "g25p(8192)",
+			want:     "gemini-2.5-pro(8192)",
+		},
+		{
+			name:     "auto suffix preserved",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p(auto)",
+			want:     "gemini-2.5-pro(auto)",
+		},
+		{
+			name:     "none suffix preserved",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p(none)",
+			want:     "gemini-2.5-pro(none)",
+		},
+		{
+			name:     "case insensitive base lookup with suffix",
+			mappings: []config.AmpModelMapping{{From: "G25P", To: "gemini-2.5-pro"}},
+			input:    "g25p(high)",
+			want:     "gemini-2.5-pro(high)",
+		},
+		{
+			name:     "empty suffix filtered out",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p()",
+			want:     "gemini-2.5-pro",
+		},
+		{
+			name:     "incomplete suffix treated as no suffix",
+			mappings: []config.AmpModelMapping{{From: "g25p(high", To: "gemini-2.5-pro"}},
+			input:    "g25p(high",
+			want:     "gemini-2.5-pro",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			mapper := NewModelMapper(tt.mappings)
+			got := mapper.MapModel(tt.input)
+			if got != tt.want {
+				t.Errorf("MapModel(%q) = %q, want %q", tt.input, got, tt.want)
+			}
+		})
+	}
+}

internal/api/modules/amp/proxy.go

@@ -15,6 +15,33 @@ import (
 	log "github.com/sirupsen/logrus"
 )
 
+func removeQueryValuesMatching(req *http.Request, key string, match string) {
+	if req == nil || req.URL == nil || match == "" {
+		return
+	}
+
+	q := req.URL.Query()
+	values, ok := q[key]
+	if !ok || len(values) == 0 {
+		return
+	}
+
+	kept := make([]string, 0, len(values))
+	for _, v := range values {
+		if v == match {
+			continue
+		}
+		kept = append(kept, v)
+	}
+
+	if len(kept) == 0 {
+		q.Del(key)
+	} else {
+		q[key] = kept
+	}
+	req.URL.RawQuery = q.Encode()
+}
+
 // readCloser wraps a reader and forwards Close to a separate closer.
 // Used to restore peeked bytes while preserving upstream body Close behavior.
 type readCloser struct {
@@ -45,6 +72,14 @@ func createReverseProxy(upstreamURL string, secretSource SecretSource) (*httputi
 		// We will set our own Authorization using the configured upstream-api-key
 		req.Header.Del("Authorization")
 		req.Header.Del("X-Api-Key")
+		req.Header.Del("X-Goog-Api-Key")
+
+		// Remove query-based credentials if they match the authenticated client API key.
+		// This prevents leaking client auth material to the Amp upstream while avoiding
+		// breaking unrelated upstream query parameters.
+		clientKey := getClientAPIKeyFromContext(req.Context())
+		removeQueryValuesMatching(req, "key", clientKey)
+		removeQueryValuesMatching(req, "auth_token", clientKey)
 
 		// Preserve correlation headers for debugging
 		if req.Header.Get("X-Request-ID") == "" {

internal/api/modules/amp/proxy_test.go

@@ -3,11 +3,15 @@ package amp
 import (
 	"bytes"
 	"compress/gzip"
+	"context"
 	"fmt"
 	"io"
 	"net/http"
 	"net/http/httptest"
+	"strings"
 	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 )
 
 // Helper: compress data with gzip
@@ -306,6 +310,159 @@ func TestReverseProxy_EmptySecret(t *testing.T) {
 	}
 }
 
+func TestReverseProxy_StripsClientCredentialsFromHeadersAndQuery(t *testing.T) {
+	type captured struct {
+		headers http.Header
+		query   string
+	}
+	got := make(chan captured, 1)
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		got <- captured{headers: r.Header.Clone(), query: r.URL.RawQuery}
+		w.WriteHeader(200)
+		w.Write([]byte(`ok`))
+	}))
+	defer upstream.Close()
+
+	proxy, err := createReverseProxy(upstream.URL, NewStaticSecretSource("upstream"))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		// Simulate clientAPIKeyMiddleware injection (per-request)
+		ctx := context.WithValue(r.Context(), clientAPIKeyContextKey{}, "client-key")
+		proxy.ServeHTTP(w, r.WithContext(ctx))
+	}))
+	defer srv.Close()
+
+	req, err := http.NewRequest(http.MethodGet, srv.URL+"/test?key=client-key&key=keep&auth_token=client-key&foo=bar", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	req.Header.Set("Authorization", "Bearer client-key")
+	req.Header.Set("X-Api-Key", "client-key")
+	req.Header.Set("X-Goog-Api-Key", "client-key")
+
+	res, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	res.Body.Close()
+
+	c := <-got
+
+	// These are client-provided credentials and must not reach the upstream.
+	if v := c.headers.Get("X-Goog-Api-Key"); v != "" {
+		t.Fatalf("X-Goog-Api-Key should be stripped, got: %q", v)
+	}
+
+	// We inject upstream Authorization/X-Api-Key, so the client auth must not survive.
+	if v := c.headers.Get("Authorization"); v != "Bearer upstream" {
+		t.Fatalf("Authorization should be upstream-injected, got: %q", v)
+	}
+	if v := c.headers.Get("X-Api-Key"); v != "upstream" {
+		t.Fatalf("X-Api-Key should be upstream-injected, got: %q", v)
+	}
+
+	// Query-based credentials should be stripped only when they match the authenticated client key.
+	// Should keep unrelated values and parameters.
+	if strings.Contains(c.query, "auth_token=client-key") || strings.Contains(c.query, "key=client-key") {
+		t.Fatalf("query credentials should be stripped, got raw query: %q", c.query)
+	}
+	if !strings.Contains(c.query, "key=keep") || !strings.Contains(c.query, "foo=bar") {
+		t.Fatalf("expected query to keep non-credential params, got raw query: %q", c.query)
+	}
+}
+
+func TestReverseProxy_InjectsMappedSecret_FromRequestContext(t *testing.T) {
+	gotHeaders := make(chan http.Header, 1)
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotHeaders <- r.Header.Clone()
+		w.WriteHeader(200)
+		w.Write([]byte(`ok`))
+	}))
+	defer upstream.Close()
+
+	defaultSource := NewStaticSecretSource("default")
+	mapped := NewMappedSecretSource(defaultSource)
+	mapped.UpdateMappings([]config.AmpUpstreamAPIKeyEntry{
+		{
+			UpstreamAPIKey: "u1",
+			APIKeys:        []string{"k1"},
+		},
+	})
+
+	proxy, err := createReverseProxy(upstream.URL, mapped)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		// Simulate clientAPIKeyMiddleware injection (per-request)
+		ctx := context.WithValue(r.Context(), clientAPIKeyContextKey{}, "k1")
+		proxy.ServeHTTP(w, r.WithContext(ctx))
+	}))
+	defer srv.Close()
+
+	res, err := http.Get(srv.URL + "/test")
+	if err != nil {
+		t.Fatal(err)
+	}
+	res.Body.Close()
+
+	hdr := <-gotHeaders
+	if hdr.Get("X-Api-Key") != "u1" {
+		t.Fatalf("X-Api-Key missing or wrong, got: %q", hdr.Get("X-Api-Key"))
+	}
+	if hdr.Get("Authorization") != "Bearer u1" {
+		t.Fatalf("Authorization missing or wrong, got: %q", hdr.Get("Authorization"))
+	}
+}
+
+func TestReverseProxy_MappedSecret_FallsBackToDefault(t *testing.T) {
+	gotHeaders := make(chan http.Header, 1)
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotHeaders <- r.Header.Clone()
+		w.WriteHeader(200)
+		w.Write([]byte(`ok`))
+	}))
+	defer upstream.Close()
+
+	defaultSource := NewStaticSecretSource("default")
+	mapped := NewMappedSecretSource(defaultSource)
+	mapped.UpdateMappings([]config.AmpUpstreamAPIKeyEntry{
+		{
+			UpstreamAPIKey: "u1",
+			APIKeys:        []string{"k1"},
+		},
+	})
+
+	proxy, err := createReverseProxy(upstream.URL, mapped)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		ctx := context.WithValue(r.Context(), clientAPIKeyContextKey{}, "k2")
+		proxy.ServeHTTP(w, r.WithContext(ctx))
+	}))
+	defer srv.Close()
+
+	res, err := http.Get(srv.URL + "/test")
+	if err != nil {
+		t.Fatal(err)
+	}
+	res.Body.Close()
+
+	hdr := <-gotHeaders
+	if hdr.Get("X-Api-Key") != "default" {
+		t.Fatalf("X-Api-Key fallback missing or wrong, got: %q", hdr.Get("X-Api-Key"))
+	}
+	if hdr.Get("Authorization") != "Bearer default" {
+		t.Fatalf("Authorization fallback missing or wrong, got: %q", hdr.Get("Authorization"))
+	}
+}
+
 func TestReverseProxy_ErrorHandler(t *testing.T) {
 	// Point proxy to a non-routable address to trigger error
 	proxy, err := createReverseProxy("http://127.0.0.1:1", NewStaticSecretSource(""))

internal/api/modules/amp/response_rewriter.go

@@ -69,7 +69,30 @@ func (rw *ResponseRewriter) Flush() {
 var modelFieldPaths = []string{"model", "modelVersion", "response.modelVersion", "message.model"}
 
 // rewriteModelInResponse replaces all occurrences of the mapped model with the original model in JSON
+// It also suppresses "thinking" blocks if "tool_use" is present to ensure Amp client compatibility
 func (rw *ResponseRewriter) rewriteModelInResponse(data []byte) []byte {
+	// 1. Amp Compatibility: Suppress thinking blocks if tool use is detected
+	// The Amp client struggles when both thinking and tool_use blocks are present
+	if gjson.GetBytes(data, `content.#(type=="tool_use")`).Exists() {
+		filtered := gjson.GetBytes(data, `content.#(type!="thinking")#`)
+		if filtered.Exists() {
+			originalCount := gjson.GetBytes(data, "content.#").Int()
+			filteredCount := filtered.Get("#").Int()
+
+			if originalCount > filteredCount {
+				var err error
+				data, err = sjson.SetBytes(data, "content", filtered.Value())
+				if err != nil {
+					log.Warnf("Amp ResponseRewriter: failed to suppress thinking blocks: %v", err)
+				} else {
+					log.Debugf("Amp ResponseRewriter: Suppressed %d thinking blocks due to tool usage", originalCount-filteredCount)
+					// Log the result for verification
+					log.Debugf("Amp ResponseRewriter: Resulting content: %s", gjson.GetBytes(data, "content").String())
+				}
+			}
+		}
+	}
+
 	if rw.originalModel == "" {
 		return data
 	}

internal/api/modules/amp/routes.go

@@ -1,6 +1,7 @@
 package amp
 
 import (
+	"context"
 	"errors"
 	"net"
 	"net/http"
@@ -16,6 +17,37 @@ import (
 	log "github.com/sirupsen/logrus"
 )
 
+// clientAPIKeyContextKey is the context key used to pass the client API key
+// from gin.Context to the request context for SecretSource lookup.
+type clientAPIKeyContextKey struct{}
+
+// clientAPIKeyMiddleware injects the authenticated client API key from gin.Context["apiKey"]
+// into the request context so that SecretSource can look it up for per-client upstream routing.
+func clientAPIKeyMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		// Extract the client API key from gin context (set by AuthMiddleware)
+		if apiKey, exists := c.Get("apiKey"); exists {
+			if keyStr, ok := apiKey.(string); ok && keyStr != "" {
+				// Inject into request context for SecretSource.Get(ctx) to read
+				ctx := context.WithValue(c.Request.Context(), clientAPIKeyContextKey{}, keyStr)
+				c.Request = c.Request.WithContext(ctx)
+			}
+		}
+		c.Next()
+	}
+}
+
+// getClientAPIKeyFromContext retrieves the client API key from request context.
+// Returns empty string if not present.
+func getClientAPIKeyFromContext(ctx context.Context) string {
+	if val := ctx.Value(clientAPIKeyContextKey{}); val != nil {
+		if keyStr, ok := val.(string); ok {
+			return keyStr
+		}
+	}
+	return ""
+}
+
 // localhostOnlyMiddleware returns a middleware that dynamically checks the module's
 // localhost restriction setting. This allows hot-reload of the restriction without restarting.
 func (m *AmpModule) localhostOnlyMiddleware() gin.HandlerFunc {
@@ -129,6 +161,9 @@ func (m *AmpModule) registerManagementRoutes(engine *gin.Engine, baseHandler *ha
 		authWithBypass = wrapManagementAuth(auth, "/threads", "/auth", "/docs", "/settings")
 	}
 
+	// Inject client API key into request context for per-client upstream routing
+	ampAPI.Use(clientAPIKeyMiddleware())
+
 	// Dynamic proxy handler that uses m.getProxy() for hot-reload support
 	proxyHandler := func(c *gin.Context) {
 		// Swallow ErrAbortHandler panics from ReverseProxy copyResponse to avoid noisy stack traces
@@ -175,6 +210,8 @@ func (m *AmpModule) registerManagementRoutes(engine *gin.Engine, baseHandler *ha
 	if authWithBypass != nil {
 		rootMiddleware = append(rootMiddleware, authWithBypass)
 	}
+	// Add clientAPIKeyMiddleware after auth for per-client upstream routing
+	rootMiddleware = append(rootMiddleware, clientAPIKeyMiddleware())
 	engine.GET("/threads", append(rootMiddleware, proxyHandler)...)
 	engine.GET("/threads/*path", append(rootMiddleware, proxyHandler)...)
 	engine.GET("/docs", append(rootMiddleware, proxyHandler)...)
@@ -244,6 +281,8 @@ func (m *AmpModule) registerProviderAliases(engine *gin.Engine, baseHandler *han
 	if auth != nil {
 		ampProviders.Use(auth)
 	}
+	// Inject client API key into request context for per-client upstream routing
+	ampProviders.Use(clientAPIKeyMiddleware())
 
 	provider := ampProviders.Group("/:provider")
 

internal/api/modules/amp/secret.go

@@ -9,6 +9,9 @@ import (
 	"strings"
 	"sync"
 	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	log "github.com/sirupsen/logrus"
 )
 
 // SecretSource provides Amp API keys with configurable precedence and caching
@@ -164,3 +167,82 @@ func NewStaticSecretSource(key string) *StaticSecretSource {
 func (s *StaticSecretSource) Get(ctx context.Context) (string, error) {
 	return s.key, nil
 }
+
+// MappedSecretSource wraps a default SecretSource and adds per-client API key mapping.
+// When a request context contains a client API key that matches a configured mapping,
+// the corresponding upstream key is returned. Otherwise, falls back to the default source.
+type MappedSecretSource struct {
+	defaultSource SecretSource
+	mu            sync.RWMutex
+	lookup        map[string]string // clientKey -> upstreamKey
+}
+
+// NewMappedSecretSource creates a MappedSecretSource wrapping the given default source.
+func NewMappedSecretSource(defaultSource SecretSource) *MappedSecretSource {
+	return &MappedSecretSource{
+		defaultSource: defaultSource,
+		lookup:        make(map[string]string),
+	}
+}
+
+// Get retrieves the Amp API key, checking per-client mappings first.
+// If the request context contains a client API key that matches a configured mapping,
+// returns the corresponding upstream key. Otherwise, falls back to the default source.
+func (s *MappedSecretSource) Get(ctx context.Context) (string, error) {
+	// Try to get client API key from request context
+	clientKey := getClientAPIKeyFromContext(ctx)
+	if clientKey != "" {
+		s.mu.RLock()
+		if upstreamKey, ok := s.lookup[clientKey]; ok && upstreamKey != "" {
+			s.mu.RUnlock()
+			return upstreamKey, nil
+		}
+		s.mu.RUnlock()
+	}
+
+	// Fall back to default source
+	return s.defaultSource.Get(ctx)
+}
+
+// UpdateMappings rebuilds the client-to-upstream key mapping from configuration entries.
+// If the same client key appears in multiple entries, logs a warning and uses the first one.
+func (s *MappedSecretSource) UpdateMappings(entries []config.AmpUpstreamAPIKeyEntry) {
+	newLookup := make(map[string]string)
+
+	for _, entry := range entries {
+		upstreamKey := strings.TrimSpace(entry.UpstreamAPIKey)
+		if upstreamKey == "" {
+			continue
+		}
+		for _, clientKey := range entry.APIKeys {
+			trimmedKey := strings.TrimSpace(clientKey)
+			if trimmedKey == "" {
+				continue
+			}
+			if _, exists := newLookup[trimmedKey]; exists {
+				// Log warning for duplicate client key, first one wins
+				log.Warnf("amp upstream-api-keys: client API key appears in multiple entries; using first mapping.")
+				continue
+			}
+			newLookup[trimmedKey] = upstreamKey
+		}
+	}
+
+	s.mu.Lock()
+	s.lookup = newLookup
+	s.mu.Unlock()
+}
+
+// UpdateDefaultExplicitKey updates the explicit key on the underlying MultiSourceSecret (if applicable).
+func (s *MappedSecretSource) UpdateDefaultExplicitKey(key string) {
+	if ms, ok := s.defaultSource.(*MultiSourceSecret); ok {
+		ms.UpdateExplicitKey(key)
+	}
+}
+
+// InvalidateCache invalidates cache on the underlying MultiSourceSecret (if applicable).
+func (s *MappedSecretSource) InvalidateCache() {
+	if ms, ok := s.defaultSource.(*MultiSourceSecret); ok {
+		ms.InvalidateCache()
+	}
+}

internal/api/modules/amp/secret_test.go

@@ -8,6 +8,10 @@ import (
 	"sync"
 	"testing"
 	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	log "github.com/sirupsen/logrus"
+	"github.com/sirupsen/logrus/hooks/test"
 )
 
 func TestMultiSourceSecret_PrecedenceOrder(t *testing.T) {
@@ -278,3 +282,85 @@ func TestMultiSourceSecret_CacheEmptyResult(t *testing.T) {
 		t.Fatalf("after cache expiry, expected new-value, got %q", got3)
 	}
 }
+
+func TestMappedSecretSource_UsesMappingFromContext(t *testing.T) {
+	defaultSource := NewStaticSecretSource("default")
+	s := NewMappedSecretSource(defaultSource)
+	s.UpdateMappings([]config.AmpUpstreamAPIKeyEntry{
+		{
+			UpstreamAPIKey: "u1",
+			APIKeys:        []string{"k1"},
+		},
+	})
+
+	ctx := context.WithValue(context.Background(), clientAPIKeyContextKey{}, "k1")
+	got, err := s.Get(ctx)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if got != "u1" {
+		t.Fatalf("want u1, got %q", got)
+	}
+
+	ctx = context.WithValue(context.Background(), clientAPIKeyContextKey{}, "k2")
+	got, err = s.Get(ctx)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if got != "default" {
+		t.Fatalf("want default fallback, got %q", got)
+	}
+}
+
+func TestMappedSecretSource_DuplicateClientKey_FirstWins(t *testing.T) {
+	defaultSource := NewStaticSecretSource("default")
+	s := NewMappedSecretSource(defaultSource)
+	s.UpdateMappings([]config.AmpUpstreamAPIKeyEntry{
+		{
+			UpstreamAPIKey: "u1",
+			APIKeys:        []string{"k1"},
+		},
+		{
+			UpstreamAPIKey: "u2",
+			APIKeys:        []string{"k1"},
+		},
+	})
+
+	ctx := context.WithValue(context.Background(), clientAPIKeyContextKey{}, "k1")
+	got, err := s.Get(ctx)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if got != "u1" {
+		t.Fatalf("want u1 (first wins), got %q", got)
+	}
+}
+
+func TestMappedSecretSource_DuplicateClientKey_LogsWarning(t *testing.T) {
+	hook := test.NewLocal(log.StandardLogger())
+	defer hook.Reset()
+
+	defaultSource := NewStaticSecretSource("default")
+	s := NewMappedSecretSource(defaultSource)
+	s.UpdateMappings([]config.AmpUpstreamAPIKeyEntry{
+		{
+			UpstreamAPIKey: "u1",
+			APIKeys:        []string{"k1"},
+		},
+		{
+			UpstreamAPIKey: "u2",
+			APIKeys:        []string{"k1"},
+		},
+	})
+
+	foundWarning := false
+	for _, entry := range hook.AllEntries() {
+		if entry.Level == log.WarnLevel && entry.Message == "amp upstream-api-keys: client API key appears in multiple entries; using first mapping." {
+			foundWarning = true
+			break
+		}
+	}
+	if !foundWarning {
+		t.Fatal("expected warning log for duplicate client key, but none was found")
+	}
+}

internal/api/server.go

@@ -26,6 +26,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/logging"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/managementasset"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/usage"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	sdkaccess "github.com/router-for-me/CLIProxyAPI/v6/sdk/access"
@@ -33,6 +34,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/sdk/api/handlers/claude"
 	"github.com/router-for-me/CLIProxyAPI/v6/sdk/api/handlers/gemini"
 	"github.com/router-for-me/CLIProxyAPI/v6/sdk/api/handlers/openai"
+	sdkAuth "github.com/router-for-me/CLIProxyAPI/v6/sdk/auth"
 	"github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	log "github.com/sirupsen/logrus"
 	"gopkg.in/yaml.v3"
@@ -209,13 +211,15 @@ func NewServer(cfg *config.Config, authManager *auth.Manager, accessManager *sdk
 	// Resolve logs directory relative to the configuration file directory.
 	var requestLogger logging.RequestLogger
 	var toggle func(bool)
-	if optionState.requestLoggerFactory != nil {
-		requestLogger = optionState.requestLoggerFactory(cfg, configFilePath)
-	}
-	if requestLogger != nil {
-		engine.Use(middleware.RequestLoggingMiddleware(requestLogger))
-		if setter, ok := requestLogger.(interface{ SetEnabled(bool) }); ok {
-			toggle = setter.SetEnabled
+	if !cfg.CommercialMode {
+		if optionState.requestLoggerFactory != nil {
+			requestLogger = optionState.requestLoggerFactory(cfg, configFilePath)
+		}
+		if requestLogger != nil {
+			engine.Use(middleware.RequestLoggingMiddleware(requestLogger))
+			if setter, ok := requestLogger.(interface{ SetEnabled(bool) }); ok {
+				toggle = setter.SetEnabled
+			}
 		}
 	}
 
@@ -251,6 +255,7 @@ func NewServer(cfg *config.Config, authManager *auth.Manager, accessManager *sdk
 	}
 	managementasset.SetCurrentConfig(cfg)
 	auth.SetQuotaCooldownDisabled(cfg.DisableCooling)
+	misc.SetCodexInstructionsEnabled(cfg.CodexInstructionsEnabled)
 	// Initialize management handler
 	s.mgmt = managementHandlers.NewHandler(cfg, configFilePath, authManager)
 	if optionState.localPassword != "" {
@@ -474,6 +479,8 @@ func (s *Server) registerManagementRoutes() {
 	mgmt.Use(s.managementAvailabilityMiddleware(), s.mgmt.Middleware())
 	{
 		mgmt.GET("/usage", s.mgmt.GetUsageStatistics)
+		mgmt.GET("/usage/export", s.mgmt.ExportUsageStatistics)
+		mgmt.POST("/usage/import", s.mgmt.ImportUsageStatistics)
 		mgmt.GET("/config", s.mgmt.GetConfig)
 		mgmt.GET("/config.yaml", s.mgmt.GetConfigYAML)
 		mgmt.PUT("/config.yaml", s.mgmt.PutConfigYAML)
@@ -487,6 +494,10 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.PUT("/logging-to-file", s.mgmt.PutLoggingToFile)
 		mgmt.PATCH("/logging-to-file", s.mgmt.PutLoggingToFile)
 
+		mgmt.GET("/logs-max-total-size-mb", s.mgmt.GetLogsMaxTotalSizeMB)
+		mgmt.PUT("/logs-max-total-size-mb", s.mgmt.PutLogsMaxTotalSizeMB)
+		mgmt.PATCH("/logs-max-total-size-mb", s.mgmt.PutLogsMaxTotalSizeMB)
+
 		mgmt.GET("/usage-statistics-enabled", s.mgmt.GetUsageStatisticsEnabled)
 		mgmt.PUT("/usage-statistics-enabled", s.mgmt.PutUsageStatisticsEnabled)
 		mgmt.PATCH("/usage-statistics-enabled", s.mgmt.PutUsageStatisticsEnabled)
@@ -496,6 +507,8 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.PATCH("/proxy-url", s.mgmt.PutProxyURL)
 		mgmt.DELETE("/proxy-url", s.mgmt.DeleteProxyURL)
 
+		mgmt.POST("/api-call", s.mgmt.APICall)
+
 		mgmt.GET("/quota-exceeded/switch-project", s.mgmt.GetSwitchProject)
 		mgmt.PUT("/quota-exceeded/switch-project", s.mgmt.PutSwitchProject)
 		mgmt.PATCH("/quota-exceeded/switch-project", s.mgmt.PutSwitchProject)
@@ -518,6 +531,7 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.DELETE("/logs", s.mgmt.DeleteLogs)
 		mgmt.GET("/request-error-logs", s.mgmt.GetRequestErrorLogs)
 		mgmt.GET("/request-error-logs/:name", s.mgmt.DownloadRequestErrorLog)
+		mgmt.GET("/request-log-by-id/:id", s.mgmt.GetRequestLogByID)
 		mgmt.GET("/request-log", s.mgmt.GetRequestLog)
 		mgmt.PUT("/request-log", s.mgmt.PutRequestLog)
 		mgmt.PATCH("/request-log", s.mgmt.PutRequestLog)
@@ -544,6 +558,10 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.GET("/ampcode/force-model-mappings", s.mgmt.GetAmpForceModelMappings)
 		mgmt.PUT("/ampcode/force-model-mappings", s.mgmt.PutAmpForceModelMappings)
 		mgmt.PATCH("/ampcode/force-model-mappings", s.mgmt.PutAmpForceModelMappings)
+		mgmt.GET("/ampcode/upstream-api-keys", s.mgmt.GetAmpUpstreamAPIKeys)
+		mgmt.PUT("/ampcode/upstream-api-keys", s.mgmt.PutAmpUpstreamAPIKeys)
+		mgmt.PATCH("/ampcode/upstream-api-keys", s.mgmt.PatchAmpUpstreamAPIKeys)
+		mgmt.DELETE("/ampcode/upstream-api-keys", s.mgmt.DeleteAmpUpstreamAPIKeys)
 
 		mgmt.GET("/request-retry", s.mgmt.GetRequestRetry)
 		mgmt.PUT("/request-retry", s.mgmt.PutRequestRetry)
@@ -552,6 +570,14 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.PUT("/max-retry-interval", s.mgmt.PutMaxRetryInterval)
 		mgmt.PATCH("/max-retry-interval", s.mgmt.PutMaxRetryInterval)
 
+		mgmt.GET("/force-model-prefix", s.mgmt.GetForceModelPrefix)
+		mgmt.PUT("/force-model-prefix", s.mgmt.PutForceModelPrefix)
+		mgmt.PATCH("/force-model-prefix", s.mgmt.PutForceModelPrefix)
+
+		mgmt.GET("/routing/strategy", s.mgmt.GetRoutingStrategy)
+		mgmt.PUT("/routing/strategy", s.mgmt.PutRoutingStrategy)
+		mgmt.PATCH("/routing/strategy", s.mgmt.PutRoutingStrategy)
+
 		mgmt.GET("/claude-api-key", s.mgmt.GetClaudeKeys)
 		mgmt.PUT("/claude-api-key", s.mgmt.PutClaudeKeys)
 		mgmt.PATCH("/claude-api-key", s.mgmt.PatchClaudeKey)
@@ -567,11 +593,21 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.PATCH("/openai-compatibility", s.mgmt.PatchOpenAICompat)
 		mgmt.DELETE("/openai-compatibility", s.mgmt.DeleteOpenAICompat)
 
+		mgmt.GET("/vertex-api-key", s.mgmt.GetVertexCompatKeys)
+		mgmt.PUT("/vertex-api-key", s.mgmt.PutVertexCompatKeys)
+		mgmt.PATCH("/vertex-api-key", s.mgmt.PatchVertexCompatKey)
+		mgmt.DELETE("/vertex-api-key", s.mgmt.DeleteVertexCompatKey)
+
 		mgmt.GET("/oauth-excluded-models", s.mgmt.GetOAuthExcludedModels)
 		mgmt.PUT("/oauth-excluded-models", s.mgmt.PutOAuthExcludedModels)
 		mgmt.PATCH("/oauth-excluded-models", s.mgmt.PatchOAuthExcludedModels)
 		mgmt.DELETE("/oauth-excluded-models", s.mgmt.DeleteOAuthExcludedModels)
 
+		mgmt.GET("/oauth-model-alias", s.mgmt.GetOAuthModelAlias)
+		mgmt.PUT("/oauth-model-alias", s.mgmt.PutOAuthModelAlias)
+		mgmt.PATCH("/oauth-model-alias", s.mgmt.PatchOAuthModelAlias)
+		mgmt.DELETE("/oauth-model-alias", s.mgmt.DeleteOAuthModelAlias)
+
 		mgmt.GET("/auth-files", s.mgmt.ListAuthFiles)
 		mgmt.GET("/auth-files/models", s.mgmt.GetAuthFileModels)
 		mgmt.GET("/auth-files/download", s.mgmt.DownloadAuthFile)
@@ -845,7 +881,7 @@ func (s *Server) UpdateClients(cfg *config.Config) {
 	}
 
 	if oldCfg == nil || oldCfg.LoggingToFile != cfg.LoggingToFile || oldCfg.LogsMaxTotalSizeMB != cfg.LogsMaxTotalSizeMB {
-		if err := logging.ConfigureLogOutput(cfg.LoggingToFile, cfg.LogsMaxTotalSizeMB); err != nil {
+		if err := logging.ConfigureLogOutput(cfg); err != nil {
 			log.Errorf("failed to reconfigure log output: %v", err)
 		} else {
 			if oldCfg == nil {
@@ -878,6 +914,16 @@ func (s *Server) UpdateClients(cfg *config.Config) {
 			log.Debugf("disable_cooling toggled to %t", cfg.DisableCooling)
 		}
 	}
+
+	if oldCfg == nil || oldCfg.CodexInstructionsEnabled != cfg.CodexInstructionsEnabled {
+		misc.SetCodexInstructionsEnabled(cfg.CodexInstructionsEnabled)
+		if oldCfg != nil {
+			log.Debugf("codex_instructions_enabled updated from %t to %t", oldCfg.CodexInstructionsEnabled, cfg.CodexInstructionsEnabled)
+		} else {
+			log.Debugf("codex_instructions_enabled toggled to %t", cfg.CodexInstructionsEnabled)
+		}
+	}
+
 	if s.handlers != nil && s.handlers.AuthManager != nil {
 		s.handlers.AuthManager.SetRetryConfig(cfg.RequestRetry, time.Duration(cfg.MaxRetryInterval)*time.Second)
 	}
@@ -955,8 +1001,12 @@ func (s *Server) UpdateClients(cfg *config.Config) {
 		log.Warnf("amp module is nil, skipping config update")
 	}
 
-	// Count client sources from configuration and auth directory
-	authFiles := util.CountAuthFiles(cfg.AuthDir)
+	// Count client sources from configuration and auth store.
+	tokenStore := sdkAuth.GetTokenStore()
+	if dirSetter, ok := tokenStore.(interface{ SetBaseDir(string) }); ok {
+		dirSetter.SetBaseDir(cfg.AuthDir)
+	}
+	authEntries := util.CountAuthFiles(context.Background(), tokenStore)
 	geminiAPIKeyCount := len(cfg.GeminiKey)
 	claudeAPIKeyCount := len(cfg.ClaudeKey)
 	codexAPIKeyCount := len(cfg.CodexKey)
@@ -967,10 +1017,10 @@ func (s *Server) UpdateClients(cfg *config.Config) {
 		openAICompatCount += len(entry.APIKeyEntries)
 	}
 
-	total := authFiles + geminiAPIKeyCount + claudeAPIKeyCount + codexAPIKeyCount + vertexAICompatCount + openAICompatCount
-	fmt.Printf("server clients and configuration updated: %d clients (%d auth files + %d Gemini API keys + %d Claude API keys + %d Codex keys + %d Vertex-compat + %d OpenAI-compat)\n",
+	total := authEntries + geminiAPIKeyCount + claudeAPIKeyCount + codexAPIKeyCount + vertexAICompatCount + openAICompatCount
+	fmt.Printf("server clients and configuration updated: %d clients (%d auth entries + %d Gemini API keys + %d Claude API keys + %d Codex keys + %d Vertex-compat + %d OpenAI-compat)\n",
 		total,
-		authFiles,
+		authEntries,
 		geminiAPIKeyCount,
 		claudeAPIKeyCount,
 		codexAPIKeyCount,

internal/auth/gemini/gemini_auth.go

@@ -18,6 +18,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/codex"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/browser"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
@@ -28,8 +29,9 @@ import (
 )
 
 const (
-	geminiOauthClientID     = "681255809395-oo8ft2oprdrnp9e3aqf6av3hmdib135j.apps.googleusercontent.com"
-	geminiOauthClientSecret = "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"
+	geminiOauthClientID       = "681255809395-oo8ft2oprdrnp9e3aqf6av3hmdib135j.apps.googleusercontent.com"
+	geminiOauthClientSecret   = "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"
+	geminiDefaultCallbackPort = 8085
 )
 
 var (
@@ -46,6 +48,13 @@ var (
 type GeminiAuth struct {
 }
 
+// WebLoginOptions customizes the interactive OAuth flow.
+type WebLoginOptions struct {
+	NoBrowser    bool
+	CallbackPort int
+	Prompt       func(string) (string, error)
+}
+
 // NewGeminiAuth creates a new instance of GeminiAuth.
 func NewGeminiAuth() *GeminiAuth {
 	return &GeminiAuth{}
@@ -59,12 +68,18 @@ func NewGeminiAuth() *GeminiAuth {
 //   - ctx: The context for the HTTP client
 //   - ts: The Gemini token storage containing authentication tokens
 //   - cfg: The configuration containing proxy settings
-//   - noBrowser: Optional parameter to disable browser opening
+//   - opts: Optional parameters to customize browser and prompt behavior
 //
 // Returns:
 //   - *http.Client: An HTTP client configured with authentication
 //   - error: An error if the client configuration fails, nil otherwise
-func (g *GeminiAuth) GetAuthenticatedClient(ctx context.Context, ts *GeminiTokenStorage, cfg *config.Config, noBrowser ...bool) (*http.Client, error) {
+func (g *GeminiAuth) GetAuthenticatedClient(ctx context.Context, ts *GeminiTokenStorage, cfg *config.Config, opts *WebLoginOptions) (*http.Client, error) {
+	callbackPort := geminiDefaultCallbackPort
+	if opts != nil && opts.CallbackPort > 0 {
+		callbackPort = opts.CallbackPort
+	}
+	callbackURL := fmt.Sprintf("http://localhost:%d/oauth2callback", callbackPort)
+
 	// Configure proxy settings for the HTTP client if a proxy URL is provided.
 	proxyURL, err := url.Parse(cfg.ProxyURL)
 	if err == nil {
@@ -99,7 +114,7 @@ func (g *GeminiAuth) GetAuthenticatedClient(ctx context.Context, ts *GeminiToken
 	conf := &oauth2.Config{
 		ClientID:     geminiOauthClientID,
 		ClientSecret: geminiOauthClientSecret,
-		RedirectURL:  "http://localhost:8085/oauth2callback", // This will be used by the local server.
+		RedirectURL:  callbackURL, // This will be used by the local server.
 		Scopes:       geminiOauthScopes,
 		Endpoint:     google.Endpoint,
 	}
@@ -109,7 +124,7 @@ func (g *GeminiAuth) GetAuthenticatedClient(ctx context.Context, ts *GeminiToken
 	// If no token is found in storage, initiate the web-based OAuth flow.
 	if ts.Token == nil {
 		fmt.Printf("Could not load token from file, starting OAuth flow.\n")
-		token, err = g.getTokenFromWeb(ctx, conf, noBrowser...)
+		token, err = g.getTokenFromWeb(ctx, conf, opts)
 		if err != nil {
 			return nil, fmt.Errorf("failed to get token from web: %w", err)
 		}
@@ -205,35 +220,50 @@ func (g *GeminiAuth) createTokenStorage(ctx context.Context, config *oauth2.Conf
 // Parameters:
 //   - ctx: The context for the HTTP client
 //   - config: The OAuth2 configuration
-//   - noBrowser: Optional parameter to disable browser opening
+//   - opts: Optional parameters to customize browser and prompt behavior
 //
 // Returns:
 //   - *oauth2.Token: The OAuth2 token obtained from the authorization flow
 //   - error: An error if the token acquisition fails, nil otherwise
-func (g *GeminiAuth) getTokenFromWeb(ctx context.Context, config *oauth2.Config, noBrowser ...bool) (*oauth2.Token, error) {
+func (g *GeminiAuth) getTokenFromWeb(ctx context.Context, config *oauth2.Config, opts *WebLoginOptions) (*oauth2.Token, error) {
+	callbackPort := geminiDefaultCallbackPort
+	if opts != nil && opts.CallbackPort > 0 {
+		callbackPort = opts.CallbackPort
+	}
+	callbackURL := fmt.Sprintf("http://localhost:%d/oauth2callback", callbackPort)
+
 	// Use a channel to pass the authorization code from the HTTP handler to the main function.
-	codeChan := make(chan string)
-	errChan := make(chan error)
+	codeChan := make(chan string, 1)
+	errChan := make(chan error, 1)
 
 	// Create a new HTTP server with its own multiplexer.
 	mux := http.NewServeMux()
-	server := &http.Server{Addr: ":8085", Handler: mux}
-	config.RedirectURL = "http://localhost:8085/oauth2callback"
+	server := &http.Server{Addr: fmt.Sprintf(":%d", callbackPort), Handler: mux}
+	config.RedirectURL = callbackURL
 
 	mux.HandleFunc("/oauth2callback", func(w http.ResponseWriter, r *http.Request) {
 		if err := r.URL.Query().Get("error"); err != "" {
 			_, _ = fmt.Fprintf(w, "Authentication failed: %s", err)
-			errChan <- fmt.Errorf("authentication failed via callback: %s", err)
+			select {
+			case errChan <- fmt.Errorf("authentication failed via callback: %s", err):
+			default:
+			}
 			return
 		}
 		code := r.URL.Query().Get("code")
 		if code == "" {
 			_, _ = fmt.Fprint(w, "Authentication failed: code not found.")
-			errChan <- fmt.Errorf("code not found in callback")
+			select {
+			case errChan <- fmt.Errorf("code not found in callback"):
+			default:
+			}
 			return
 		}
 		_, _ = fmt.Fprint(w, "<html><body><h1>Authentication successful!</h1><p>You can close this window.</p></body></html>")
-		codeChan <- code
+		select {
+		case codeChan <- code:
+		default:
+		}
 	})
 
 	// Start the server in a goroutine.
@@ -250,19 +280,24 @@ func (g *GeminiAuth) getTokenFromWeb(ctx context.Context, config *oauth2.Config,
 	// Open the authorization URL in the user's browser.
 	authURL := config.AuthCodeURL("state-token", oauth2.AccessTypeOffline, oauth2.SetAuthURLParam("prompt", "consent"))
 
-	if len(noBrowser) == 1 && !noBrowser[0] {
+	noBrowser := false
+	if opts != nil {
+		noBrowser = opts.NoBrowser
+	}
+
+	if !noBrowser {
 		fmt.Println("Opening browser for authentication...")
 
 		// Check if browser is available
 		if !browser.IsAvailable() {
 			log.Warn("No browser available on this system")
-			util.PrintSSHTunnelInstructions(8085)
+			util.PrintSSHTunnelInstructions(callbackPort)
 			fmt.Printf("Please manually open this URL in your browser:\n\n%s\n", authURL)
 		} else {
 			if err := browser.OpenURL(authURL); err != nil {
 				authErr := codex.NewAuthenticationError(codex.ErrBrowserOpenFailed, err)
 				log.Warn(codex.GetUserFriendlyMessage(authErr))
-				util.PrintSSHTunnelInstructions(8085)
+				util.PrintSSHTunnelInstructions(callbackPort)
 				fmt.Printf("Please manually open this URL in your browser:\n\n%s\n", authURL)
 
 				// Log platform info for debugging
@@ -273,7 +308,7 @@ func (g *GeminiAuth) getTokenFromWeb(ctx context.Context, config *oauth2.Config,
 			}
 		}
 	} else {
-		util.PrintSSHTunnelInstructions(8085)
+		util.PrintSSHTunnelInstructions(callbackPort)
 		fmt.Printf("Please open this URL in your browser:\n\n%s\n", authURL)
 	}
 
@@ -281,13 +316,60 @@ func (g *GeminiAuth) getTokenFromWeb(ctx context.Context, config *oauth2.Config,
 
 	// Wait for the authorization code or an error.
 	var authCode string
-	select {
-	case code := <-codeChan:
-		authCode = code
-	case err := <-errChan:
-		return nil, err
-	case <-time.After(5 * time.Minute): // Timeout
-		return nil, fmt.Errorf("oauth flow timed out")
+	timeoutTimer := time.NewTimer(5 * time.Minute)
+	defer timeoutTimer.Stop()
+
+	var manualPromptTimer *time.Timer
+	var manualPromptC <-chan time.Time
+	if opts != nil && opts.Prompt != nil {
+		manualPromptTimer = time.NewTimer(15 * time.Second)
+		manualPromptC = manualPromptTimer.C
+		defer manualPromptTimer.Stop()
+	}
+
+waitForCallback:
+	for {
+		select {
+		case code := <-codeChan:
+			authCode = code
+			break waitForCallback
+		case err := <-errChan:
+			return nil, err
+		case <-manualPromptC:
+			manualPromptC = nil
+			if manualPromptTimer != nil {
+				manualPromptTimer.Stop()
+			}
+			select {
+			case code := <-codeChan:
+				authCode = code
+				break waitForCallback
+			case err := <-errChan:
+				return nil, err
+			default:
+			}
+			input, err := opts.Prompt("Paste the Gemini callback URL (or press Enter to keep waiting): ")
+			if err != nil {
+				return nil, err
+			}
+			parsed, err := misc.ParseOAuthCallback(input)
+			if err != nil {
+				return nil, err
+			}
+			if parsed == nil {
+				continue
+			}
+			if parsed.Error != "" {
+				return nil, fmt.Errorf("authentication failed via callback: %s", parsed.Error)
+			}
+			if parsed.Code == "" {
+				return nil, fmt.Errorf("code not found in callback")
+			}
+			authCode = parsed.Code
+			break waitForCallback
+		case <-timeoutTimer.C:
+			return nil, fmt.Errorf("oauth flow timed out")
+		}
 	}
 
 	// Shutdown the server.

internal/cache/signature_cache.go

@@ -0,0 +1,169 @@
+package cache
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"sync"
+	"time"
+)
+
+// SignatureEntry holds a cached thinking signature with timestamp
+type SignatureEntry struct {
+	Signature string
+	Timestamp time.Time
+}
+
+const (
+	// SignatureCacheTTL is how long signatures are valid
+	SignatureCacheTTL = 3 * time.Hour
+
+	// SignatureTextHashLen is the length of the hash key (16 hex chars = 64-bit key space)
+	SignatureTextHashLen = 16
+
+	// MinValidSignatureLen is the minimum length for a signature to be considered valid
+	MinValidSignatureLen = 50
+
+	// SessionCleanupInterval controls how often stale sessions are purged
+	SessionCleanupInterval = 10 * time.Minute
+)
+
+// signatureCache stores signatures by sessionId -> textHash -> SignatureEntry
+var signatureCache sync.Map
+
+// sessionCleanupOnce ensures the background cleanup goroutine starts only once
+var sessionCleanupOnce sync.Once
+
+// sessionCache is the inner map type
+type sessionCache struct {
+	mu      sync.RWMutex
+	entries map[string]SignatureEntry
+}
+
+// hashText creates a stable, Unicode-safe key from text content
+func hashText(text string) string {
+	h := sha256.Sum256([]byte(text))
+	return hex.EncodeToString(h[:])[:SignatureTextHashLen]
+}
+
+// getOrCreateSession gets or creates a session cache
+func getOrCreateSession(sessionID string) *sessionCache {
+	// Start background cleanup on first access
+	sessionCleanupOnce.Do(startSessionCleanup)
+
+	if val, ok := signatureCache.Load(sessionID); ok {
+		return val.(*sessionCache)
+	}
+	sc := &sessionCache{entries: make(map[string]SignatureEntry)}
+	actual, _ := signatureCache.LoadOrStore(sessionID, sc)
+	return actual.(*sessionCache)
+}
+
+// startSessionCleanup launches a background goroutine that periodically
+// removes sessions where all entries have expired.
+func startSessionCleanup() {
+	go func() {
+		ticker := time.NewTicker(SessionCleanupInterval)
+		defer ticker.Stop()
+		for range ticker.C {
+			purgeExpiredSessions()
+		}
+	}()
+}
+
+// purgeExpiredSessions removes sessions with no valid (non-expired) entries.
+func purgeExpiredSessions() {
+	now := time.Now()
+	signatureCache.Range(func(key, value any) bool {
+		sc := value.(*sessionCache)
+		sc.mu.Lock()
+		// Remove expired entries
+		for k, entry := range sc.entries {
+			if now.Sub(entry.Timestamp) > SignatureCacheTTL {
+				delete(sc.entries, k)
+			}
+		}
+		isEmpty := len(sc.entries) == 0
+		sc.mu.Unlock()
+		// Remove session if empty
+		if isEmpty {
+			signatureCache.Delete(key)
+		}
+		return true
+	})
+}
+
+// CacheSignature stores a thinking signature for a given session and text.
+// Used for Claude models that require signed thinking blocks in multi-turn conversations.
+func CacheSignature(sessionID, text, signature string) {
+	if sessionID == "" || text == "" || signature == "" {
+		return
+	}
+	if len(signature) < MinValidSignatureLen {
+		return
+	}
+
+	sc := getOrCreateSession(sessionID)
+	textHash := hashText(text)
+
+	sc.mu.Lock()
+	defer sc.mu.Unlock()
+
+	sc.entries[textHash] = SignatureEntry{
+		Signature: signature,
+		Timestamp: time.Now(),
+	}
+}
+
+// GetCachedSignature retrieves a cached signature for a given session and text.
+// Returns empty string if not found or expired.
+func GetCachedSignature(sessionID, text string) string {
+	if sessionID == "" || text == "" {
+		return ""
+	}
+
+	val, ok := signatureCache.Load(sessionID)
+	if !ok {
+		return ""
+	}
+	sc := val.(*sessionCache)
+
+	textHash := hashText(text)
+
+	now := time.Now()
+
+	sc.mu.Lock()
+	entry, exists := sc.entries[textHash]
+	if !exists {
+		sc.mu.Unlock()
+		return ""
+	}
+	if now.Sub(entry.Timestamp) > SignatureCacheTTL {
+		delete(sc.entries, textHash)
+		sc.mu.Unlock()
+		return ""
+	}
+
+	// Refresh TTL on access (sliding expiration).
+	entry.Timestamp = now
+	sc.entries[textHash] = entry
+	sc.mu.Unlock()
+
+	return entry.Signature
+}
+
+// ClearSignatureCache clears signature cache for a specific session or all sessions.
+func ClearSignatureCache(sessionID string) {
+	if sessionID != "" {
+		signatureCache.Delete(sessionID)
+	} else {
+		signatureCache.Range(func(key, _ any) bool {
+			signatureCache.Delete(key)
+			return true
+		})
+	}
+}
+
+// HasValidSignature checks if a signature is valid (non-empty and long enough)
+func HasValidSignature(signature string) bool {
+	return signature != "" && len(signature) >= MinValidSignatureLen
+}

internal/cache/signature_cache_test.go

@@ -0,0 +1,216 @@
+package cache
+
+import (
+	"testing"
+	"time"
+)
+
+func TestCacheSignature_BasicStorageAndRetrieval(t *testing.T) {
+	ClearSignatureCache("")
+
+	sessionID := "test-session-1"
+	text := "This is some thinking text content"
+	signature := "abc123validSignature1234567890123456789012345678901234567890"
+
+	// Store signature
+	CacheSignature(sessionID, text, signature)
+
+	// Retrieve signature
+	retrieved := GetCachedSignature(sessionID, text)
+	if retrieved != signature {
+		t.Errorf("Expected signature '%s', got '%s'", signature, retrieved)
+	}
+}
+
+func TestCacheSignature_DifferentSessions(t *testing.T) {
+	ClearSignatureCache("")
+
+	text := "Same text in different sessions"
+	sig1 := "signature1_1234567890123456789012345678901234567890123456"
+	sig2 := "signature2_1234567890123456789012345678901234567890123456"
+
+	CacheSignature("session-a", text, sig1)
+	CacheSignature("session-b", text, sig2)
+
+	if GetCachedSignature("session-a", text) != sig1 {
+		t.Error("Session-a signature mismatch")
+	}
+	if GetCachedSignature("session-b", text) != sig2 {
+		t.Error("Session-b signature mismatch")
+	}
+}
+
+func TestCacheSignature_NotFound(t *testing.T) {
+	ClearSignatureCache("")
+
+	// Non-existent session
+	if got := GetCachedSignature("nonexistent", "some text"); got != "" {
+		t.Errorf("Expected empty string for nonexistent session, got '%s'", got)
+	}
+
+	// Existing session but different text
+	CacheSignature("session-x", "text-a", "sigA12345678901234567890123456789012345678901234567890")
+	if got := GetCachedSignature("session-x", "text-b"); got != "" {
+		t.Errorf("Expected empty string for different text, got '%s'", got)
+	}
+}
+
+func TestCacheSignature_EmptyInputs(t *testing.T) {
+	ClearSignatureCache("")
+
+	// All empty/invalid inputs should be no-ops
+	CacheSignature("", "text", "sig12345678901234567890123456789012345678901234567890")
+	CacheSignature("session", "", "sig12345678901234567890123456789012345678901234567890")
+	CacheSignature("session", "text", "")
+	CacheSignature("session", "text", "short") // Too short
+
+	if got := GetCachedSignature("session", "text"); got != "" {
+		t.Errorf("Expected empty after invalid cache attempts, got '%s'", got)
+	}
+}
+
+func TestCacheSignature_ShortSignatureRejected(t *testing.T) {
+	ClearSignatureCache("")
+
+	sessionID := "test-short-sig"
+	text := "Some text"
+	shortSig := "abc123" // Less than 50 chars
+
+	CacheSignature(sessionID, text, shortSig)
+
+	if got := GetCachedSignature(sessionID, text); got != "" {
+		t.Errorf("Short signature should be rejected, got '%s'", got)
+	}
+}
+
+func TestClearSignatureCache_SpecificSession(t *testing.T) {
+	ClearSignatureCache("")
+
+	sig := "validSig1234567890123456789012345678901234567890123456"
+	CacheSignature("session-1", "text", sig)
+	CacheSignature("session-2", "text", sig)
+
+	ClearSignatureCache("session-1")
+
+	if got := GetCachedSignature("session-1", "text"); got != "" {
+		t.Error("session-1 should be cleared")
+	}
+	if got := GetCachedSignature("session-2", "text"); got != sig {
+		t.Error("session-2 should still exist")
+	}
+}
+
+func TestClearSignatureCache_AllSessions(t *testing.T) {
+	ClearSignatureCache("")
+
+	sig := "validSig1234567890123456789012345678901234567890123456"
+	CacheSignature("session-1", "text", sig)
+	CacheSignature("session-2", "text", sig)
+
+	ClearSignatureCache("")
+
+	if got := GetCachedSignature("session-1", "text"); got != "" {
+		t.Error("session-1 should be cleared")
+	}
+	if got := GetCachedSignature("session-2", "text"); got != "" {
+		t.Error("session-2 should be cleared")
+	}
+}
+
+func TestHasValidSignature(t *testing.T) {
+	tests := []struct {
+		name      string
+		signature string
+		expected  bool
+	}{
+		{"valid long signature", "abc123validSignature1234567890123456789012345678901234567890", true},
+		{"exactly 50 chars", "12345678901234567890123456789012345678901234567890", true},
+		{"49 chars - invalid", "1234567890123456789012345678901234567890123456789", false},
+		{"empty string", "", false},
+		{"short signature", "abc", false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := HasValidSignature(tt.signature)
+			if result != tt.expected {
+				t.Errorf("HasValidSignature(%q) = %v, expected %v", tt.signature, result, tt.expected)
+			}
+		})
+	}
+}
+
+func TestCacheSignature_TextHashCollisionResistance(t *testing.T) {
+	ClearSignatureCache("")
+
+	sessionID := "hash-test-session"
+
+	// Different texts should produce different hashes
+	text1 := "First thinking text"
+	text2 := "Second thinking text"
+	sig1 := "signature1_1234567890123456789012345678901234567890123456"
+	sig2 := "signature2_1234567890123456789012345678901234567890123456"
+
+	CacheSignature(sessionID, text1, sig1)
+	CacheSignature(sessionID, text2, sig2)
+
+	if GetCachedSignature(sessionID, text1) != sig1 {
+		t.Error("text1 signature mismatch")
+	}
+	if GetCachedSignature(sessionID, text2) != sig2 {
+		t.Error("text2 signature mismatch")
+	}
+}
+
+func TestCacheSignature_UnicodeText(t *testing.T) {
+	ClearSignatureCache("")
+
+	sessionID := "unicode-session"
+	text := "한글 텍스트와 이모지 🎉 그리고 特殊文字"
+	sig := "unicodeSig123456789012345678901234567890123456789012345"
+
+	CacheSignature(sessionID, text, sig)
+
+	if got := GetCachedSignature(sessionID, text); got != sig {
+		t.Errorf("Unicode text signature retrieval failed, got '%s'", got)
+	}
+}
+
+func TestCacheSignature_Overwrite(t *testing.T) {
+	ClearSignatureCache("")
+
+	sessionID := "overwrite-session"
+	text := "Same text"
+	sig1 := "firstSignature12345678901234567890123456789012345678901"
+	sig2 := "secondSignature1234567890123456789012345678901234567890"
+
+	CacheSignature(sessionID, text, sig1)
+	CacheSignature(sessionID, text, sig2) // Overwrite
+
+	if got := GetCachedSignature(sessionID, text); got != sig2 {
+		t.Errorf("Expected overwritten signature '%s', got '%s'", sig2, got)
+	}
+}
+
+// Note: TTL expiration test is tricky to test without mocking time
+// We test the logic path exists but actual expiration would require time manipulation
+func TestCacheSignature_ExpirationLogic(t *testing.T) {
+	ClearSignatureCache("")
+
+	// This test verifies the expiration check exists
+	// In a real scenario, we'd mock time.Now()
+	sessionID := "expiration-test"
+	text := "text"
+	sig := "validSig1234567890123456789012345678901234567890123456"
+
+	CacheSignature(sessionID, text, sig)
+
+	// Fresh entry should be retrievable
+	if got := GetCachedSignature(sessionID, text); got != sig {
+		t.Errorf("Fresh entry should be retrievable, got '%s'", got)
+	}
+
+	// We can't easily test actual expiration without time mocking
+	// but the logic is verified by the implementation
+	_ = time.Now() // Acknowledge we're not testing time passage
+}

internal/cmd/anthropic_login.go

@@ -24,12 +24,18 @@ func DoClaudeLogin(cfg *config.Config, options *LoginOptions) {
 		options = &LoginOptions{}
 	}
 
+	promptFn := options.Prompt
+	if promptFn == nil {
+		promptFn = defaultProjectPrompt()
+	}
+
 	manager := newAuthManager()
 
 	authOpts := &sdkAuth.LoginOptions{
-		NoBrowser: options.NoBrowser,
-		Metadata:  map[string]string{},
-		Prompt:    options.Prompt,
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       promptFn,
 	}
 
 	_, savedPath, err := manager.Login(context.Background(), "claude", cfg, authOpts)

internal/cmd/antigravity_login.go

@@ -15,11 +15,17 @@ func DoAntigravityLogin(cfg *config.Config, options *LoginOptions) {
 		options = &LoginOptions{}
 	}
 
+	promptFn := options.Prompt
+	if promptFn == nil {
+		promptFn = defaultProjectPrompt()
+	}
+
 	manager := newAuthManager()
 	authOpts := &sdkAuth.LoginOptions{
-		NoBrowser: options.NoBrowser,
-		Metadata:  map[string]string{},
-		Prompt:    options.Prompt,
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       promptFn,
 	}
 
 	record, savedPath, err := manager.Login(context.Background(), "antigravity", cfg, authOpts)

internal/cmd/iflow_login.go

@@ -20,19 +20,14 @@ func DoIFlowLogin(cfg *config.Config, options *LoginOptions) {
 
 	promptFn := options.Prompt
 	if promptFn == nil {
-		promptFn = func(prompt string) (string, error) {
-			fmt.Println()
-			fmt.Println(prompt)
-			var value string
-			_, err := fmt.Scanln(&value)
-			return value, err
-		}
+		promptFn = defaultProjectPrompt()
 	}
 
 	authOpts := &sdkAuth.LoginOptions{
-		NoBrowser: options.NoBrowser,
-		Metadata:  map[string]string{},
-		Prompt:    promptFn,
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       promptFn,
 	}
 
 	_, savedPath, err := manager.Login(context.Background(), "iflow", cfg, authOpts)

internal/cmd/login.go

@@ -55,11 +55,23 @@ func DoLogin(cfg *config.Config, projectID string, options *LoginOptions) {
 
 	ctx := context.Background()
 
+	promptFn := options.Prompt
+	if promptFn == nil {
+		promptFn = defaultProjectPrompt()
+	}
+
+	trimmedProjectID := strings.TrimSpace(projectID)
+	callbackPrompt := promptFn
+	if trimmedProjectID == "" {
+		callbackPrompt = nil
+	}
+
 	loginOpts := &sdkAuth.LoginOptions{
-		NoBrowser: options.NoBrowser,
-		ProjectID: strings.TrimSpace(projectID),
-		Metadata:  map[string]string{},
-		Prompt:    options.Prompt,
+		NoBrowser:    options.NoBrowser,
+		ProjectID:    trimmedProjectID,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       callbackPrompt,
 	}
 
 	authenticator := sdkAuth.NewGeminiAuthenticator()
@@ -76,7 +88,11 @@ func DoLogin(cfg *config.Config, projectID string, options *LoginOptions) {
 	}
 
 	geminiAuth := gemini.NewGeminiAuth()
-	httpClient, errClient := geminiAuth.GetAuthenticatedClient(ctx, storage, cfg, options.NoBrowser)
+	httpClient, errClient := geminiAuth.GetAuthenticatedClient(ctx, storage, cfg, &gemini.WebLoginOptions{
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Prompt:       callbackPrompt,
+	})
 	if errClient != nil {
 		log.Errorf("Gemini authentication failed: %v", errClient)
 		return
@@ -90,12 +106,7 @@ func DoLogin(cfg *config.Config, projectID string, options *LoginOptions) {
 		return
 	}
 
-	promptFn := options.Prompt
-	if promptFn == nil {
-		promptFn = defaultProjectPrompt()
-	}
-
-	selectedProjectID := promptForProjectSelection(projects, strings.TrimSpace(projectID), promptFn)
+	selectedProjectID := promptForProjectSelection(projects, trimmedProjectID, promptFn)
 	projectSelections, errSelection := resolveProjectSelections(selectedProjectID, projects)
 	if errSelection != nil {
 		log.Errorf("Invalid project selection: %v", errSelection)

internal/cmd/openai_login.go

@@ -19,6 +19,9 @@ type LoginOptions struct {
 	// NoBrowser indicates whether to skip opening the browser automatically.
 	NoBrowser bool
 
+	// CallbackPort overrides the local OAuth callback port when set (>0).
+	CallbackPort int
+
 	// Prompt allows the caller to provide interactive input when needed.
 	Prompt func(prompt string) (string, error)
 }
@@ -35,12 +38,18 @@ func DoCodexLogin(cfg *config.Config, options *LoginOptions) {
 		options = &LoginOptions{}
 	}
 
+	promptFn := options.Prompt
+	if promptFn == nil {
+		promptFn = defaultProjectPrompt()
+	}
+
 	manager := newAuthManager()
 
 	authOpts := &sdkAuth.LoginOptions{
-		NoBrowser: options.NoBrowser,
-		Metadata:  map[string]string{},
-		Prompt:    options.Prompt,
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       promptFn,
 	}
 
 	_, savedPath, err := manager.Login(context.Background(), "codex", cfg, authOpts)

internal/cmd/qwen_login.go

@@ -36,9 +36,10 @@ func DoQwenLogin(cfg *config.Config, options *LoginOptions) {
 	}
 
 	authOpts := &sdkAuth.LoginOptions{
-		NoBrowser: options.NoBrowser,
-		Metadata:  map[string]string{},
-		Prompt:    promptFn,
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       promptFn,
 	}
 
 	_, savedPath, err := manager.Login(context.Background(), "qwen", cfg, authOpts)

internal/config/config.go

@@ -6,12 +6,14 @@ package config
 
 import (
 	"bytes"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"os"
 	"strings"
 	"syscall"
 
+	log "github.com/sirupsen/logrus"
 	"golang.org/x/crypto/bcrypt"
 	"gopkg.in/yaml.v3"
 )
@@ -39,6 +41,9 @@ type Config struct {
 	// Debug enables or disables debug-level logging and other debug features.
 	Debug bool `yaml:"debug" json:"debug"`
 
+	// CommercialMode disables high-overhead HTTP middleware features to minimize per-request memory usage.
+	CommercialMode bool `yaml:"commercial-mode" json:"commercial-mode"`
+
 	// LoggingToFile controls whether application logs are written to rotating files or stdout.
 	LoggingToFile bool `yaml:"logging-to-file" json:"logging-to-file"`
 
@@ -60,9 +65,17 @@ type Config struct {
 	// QuotaExceeded defines the behavior when a quota is exceeded.
 	QuotaExceeded QuotaExceeded `yaml:"quota-exceeded" json:"quota-exceeded"`
 
+	// Routing controls credential selection behavior.
+	Routing RoutingConfig `yaml:"routing" json:"routing"`
+
 	// WebsocketAuth enables or disables authentication for the WebSocket API.
 	WebsocketAuth bool `yaml:"ws-auth" json:"ws-auth"`
 
+	// CodexInstructionsEnabled controls whether custom Codex instructions are injected.
+	// When false (default), CodexInstructionsForModel returns immediately without modification.
+	// When true, the original instruction injection logic is used.
+	CodexInstructionsEnabled bool `yaml:"codex-instructions-enabled" json:"codex-instructions-enabled"`
+
 	// GeminiKey defines Gemini API key configurations with optional routing overrides.
 	GeminiKey []GeminiKey `yaml:"gemini-api-key" json:"gemini-api-key"`
 
@@ -85,6 +98,14 @@ type Config struct {
 	// OAuthExcludedModels defines per-provider global model exclusions applied to OAuth/file-backed auth entries.
 	OAuthExcludedModels map[string][]string `yaml:"oauth-excluded-models,omitempty" json:"oauth-excluded-models,omitempty"`
 
+	// OAuthModelAlias defines global model name aliases for OAuth/file-backed auth channels.
+	// These aliases affect both model listing and model routing for supported channels:
+	// gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow.
+	//
+	// NOTE: This does not apply to existing per-credential model alias features under:
+	// gemini-api-key, codex-api-key, claude-api-key, openai-compatibility, vertex-api-key, and ampcode.
+	OAuthModelAlias map[string][]OAuthModelAlias `yaml:"oauth-model-alias,omitempty" json:"oauth-model-alias,omitempty"`
+
 	// Payload defines default and override rules for provider payload parameters.
 	Payload PayloadConfig `yaml:"payload" json:"payload"`
 
@@ -124,6 +145,23 @@ type QuotaExceeded struct {
 	SwitchPreviewModel bool `yaml:"switch-preview-model" json:"switch-preview-model"`
 }
 
+// RoutingConfig configures how credentials are selected for requests.
+type RoutingConfig struct {
+	// Strategy selects the credential selection strategy.
+	// Supported values: "round-robin" (default), "fill-first".
+	Strategy string `yaml:"strategy,omitempty" json:"strategy,omitempty"`
+}
+
+// OAuthModelAlias defines a model ID alias for a specific channel.
+// It maps the upstream model name (Name) to the client-visible alias (Alias).
+// When Fork is true, the alias is added as an additional model in listings while
+// keeping the original model ID available.
+type OAuthModelAlias struct {
+	Name  string `yaml:"name" json:"name"`
+	Alias string `yaml:"alias" json:"alias"`
+	Fork  bool   `yaml:"fork,omitempty" json:"fork,omitempty"`
+}
+
 // AmpModelMapping defines a model name mapping for Amp CLI requests.
 // When Amp requests a model that isn't available locally, this mapping
 // allows routing to an alternative model that IS available.
@@ -134,6 +172,11 @@ type AmpModelMapping struct {
 	// To is the target model name to route to (e.g., "claude-sonnet-4").
 	// The target model must have available providers in the registry.
 	To string `yaml:"to" json:"to"`
+
+	// Regex indicates whether the 'from' field should be interpreted as a regular
+	// expression for matching model names. When true, this mapping is evaluated
+	// after exact matches and in the order provided. Defaults to false (exact match).
+	Regex bool `yaml:"regex,omitempty" json:"regex,omitempty"`
 }
 
 // AmpCode groups Amp CLI integration settings including upstream routing,
@@ -145,6 +188,11 @@ type AmpCode struct {
 	// UpstreamAPIKey optionally overrides the Authorization header when proxying Amp upstream calls.
 	UpstreamAPIKey string `yaml:"upstream-api-key" json:"upstream-api-key"`
 
+	// UpstreamAPIKeys maps client API keys (from top-level api-keys) to upstream API keys.
+	// When a client authenticates with a key that matches an entry, that upstream key is used.
+	// If no match is found, falls back to UpstreamAPIKey (default behavior).
+	UpstreamAPIKeys []AmpUpstreamAPIKeyEntry `yaml:"upstream-api-keys,omitempty" json:"upstream-api-keys,omitempty"`
+
 	// RestrictManagementToLocalhost restricts Amp management routes (/api/user, /api/threads, etc.)
 	// to only accept connections from localhost (127.0.0.1, ::1). When true, prevents drive-by
 	// browser attacks and remote access to management endpoints. Default: false (API key auth is sufficient).
@@ -160,12 +208,27 @@ type AmpCode struct {
 	ForceModelMappings bool `yaml:"force-model-mappings" json:"force-model-mappings"`
 }
 
+// AmpUpstreamAPIKeyEntry maps a set of client API keys to a specific upstream API key.
+// When a request is authenticated with one of the APIKeys, the corresponding UpstreamAPIKey
+// is used for the upstream Amp request.
+type AmpUpstreamAPIKeyEntry struct {
+	// UpstreamAPIKey is the API key to use when proxying to the Amp upstream.
+	UpstreamAPIKey string `yaml:"upstream-api-key" json:"upstream-api-key"`
+
+	// APIKeys are the client API keys (from top-level api-keys) that map to this upstream key.
+	APIKeys []string `yaml:"api-keys" json:"api-keys"`
+}
+
 // PayloadConfig defines default and override parameter rules applied to provider payloads.
 type PayloadConfig struct {
 	// Default defines rules that only set parameters when they are missing in the payload.
 	Default []PayloadRule `yaml:"default" json:"default"`
+	// DefaultRaw defines rules that set raw JSON values only when they are missing.
+	DefaultRaw []PayloadRule `yaml:"default-raw" json:"default-raw"`
 	// Override defines rules that always set parameters, overwriting any existing values.
 	Override []PayloadRule `yaml:"override" json:"override"`
+	// OverrideRaw defines rules that always set raw JSON values, overwriting any existing values.
+	OverrideRaw []PayloadRule `yaml:"override-raw" json:"override-raw"`
 }
 
 // PayloadRule describes a single rule targeting a list of models with parameter updates.
@@ -173,6 +236,7 @@ type PayloadRule struct {
 	// Models lists model entries with name pattern and protocol constraint.
 	Models []PayloadModelRule `yaml:"models" json:"models"`
 	// Params maps JSON paths (gjson/sjson syntax) to values written into the payload.
+	// For *-raw rules, values are treated as raw JSON fragments (strings are used as-is).
 	Params map[string]any `yaml:"params" json:"params"`
 }
 
@@ -190,6 +254,10 @@ type ClaudeKey struct {
 	// APIKey is the authentication key for accessing Claude API services.
 	APIKey string `yaml:"api-key" json:"api-key"`
 
+	// Priority controls selection preference when multiple credentials match.
+	// Higher values are preferred; defaults to 0.
+	Priority int `yaml:"priority,omitempty" json:"priority,omitempty"`
+
 	// Prefix optionally namespaces models for this credential (e.g., "teamA/claude-sonnet-4").
 	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
 
@@ -210,6 +278,9 @@ type ClaudeKey struct {
 	ExcludedModels []string `yaml:"excluded-models,omitempty" json:"excluded-models,omitempty"`
 }
 
+func (k ClaudeKey) GetAPIKey() string  { return k.APIKey }
+func (k ClaudeKey) GetBaseURL() string { return k.BaseURL }
+
 // ClaudeModel describes a mapping between an alias and the actual upstream model name.
 type ClaudeModel struct {
 	// Name is the upstream model identifier used when issuing requests.
@@ -219,12 +290,19 @@ type ClaudeModel struct {
 	Alias string `yaml:"alias" json:"alias"`
 }
 
+func (m ClaudeModel) GetName() string  { return m.Name }
+func (m ClaudeModel) GetAlias() string { return m.Alias }
+
 // CodexKey represents the configuration for a Codex API key,
 // including the API key itself and an optional base URL for the API endpoint.
 type CodexKey struct {
 	// APIKey is the authentication key for accessing Codex API services.
 	APIKey string `yaml:"api-key" json:"api-key"`
 
+	// Priority controls selection preference when multiple credentials match.
+	// Higher values are preferred; defaults to 0.
+	Priority int `yaml:"priority,omitempty" json:"priority,omitempty"`
+
 	// Prefix optionally namespaces models for this credential (e.g., "teamA/gpt-5-codex").
 	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
 
@@ -235,6 +313,9 @@ type CodexKey struct {
 	// ProxyURL overrides the global proxy setting for this API key if provided.
 	ProxyURL string `yaml:"proxy-url" json:"proxy-url"`
 
+	// Models defines upstream model names and aliases for request routing.
+	Models []CodexModel `yaml:"models" json:"models"`
+
 	// Headers optionally adds extra HTTP headers for requests sent with this key.
 	Headers map[string]string `yaml:"headers,omitempty" json:"headers,omitempty"`
 
@@ -242,12 +323,31 @@ type CodexKey struct {
 	ExcludedModels []string `yaml:"excluded-models,omitempty" json:"excluded-models,omitempty"`
 }
 
+func (k CodexKey) GetAPIKey() string  { return k.APIKey }
+func (k CodexKey) GetBaseURL() string { return k.BaseURL }
+
+// CodexModel describes a mapping between an alias and the actual upstream model name.
+type CodexModel struct {
+	// Name is the upstream model identifier used when issuing requests.
+	Name string `yaml:"name" json:"name"`
+
+	// Alias is the client-facing model name that maps to Name.
+	Alias string `yaml:"alias" json:"alias"`
+}
+
+func (m CodexModel) GetName() string  { return m.Name }
+func (m CodexModel) GetAlias() string { return m.Alias }
+
 // GeminiKey represents the configuration for a Gemini API key,
 // including optional overrides for upstream base URL, proxy routing, and headers.
 type GeminiKey struct {
 	// APIKey is the authentication key for accessing Gemini API services.
 	APIKey string `yaml:"api-key" json:"api-key"`
 
+	// Priority controls selection preference when multiple credentials match.
+	// Higher values are preferred; defaults to 0.
+	Priority int `yaml:"priority,omitempty" json:"priority,omitempty"`
+
 	// Prefix optionally namespaces models for this credential (e.g., "teamA/gemini-3-pro-preview").
 	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
 
@@ -257,6 +357,9 @@ type GeminiKey struct {
 	// ProxyURL optionally overrides the global proxy for this API key.
 	ProxyURL string `yaml:"proxy-url,omitempty" json:"proxy-url,omitempty"`
 
+	// Models defines upstream model names and aliases for request routing.
+	Models []GeminiModel `yaml:"models,omitempty" json:"models,omitempty"`
+
 	// Headers optionally adds extra HTTP headers for requests sent with this key.
 	Headers map[string]string `yaml:"headers,omitempty" json:"headers,omitempty"`
 
@@ -264,12 +367,31 @@ type GeminiKey struct {
 	ExcludedModels []string `yaml:"excluded-models,omitempty" json:"excluded-models,omitempty"`
 }
 
+func (k GeminiKey) GetAPIKey() string  { return k.APIKey }
+func (k GeminiKey) GetBaseURL() string { return k.BaseURL }
+
+// GeminiModel describes a mapping between an alias and the actual upstream model name.
+type GeminiModel struct {
+	// Name is the upstream model identifier used when issuing requests.
+	Name string `yaml:"name" json:"name"`
+
+	// Alias is the client-facing model name that maps to Name.
+	Alias string `yaml:"alias" json:"alias"`
+}
+
+func (m GeminiModel) GetName() string  { return m.Name }
+func (m GeminiModel) GetAlias() string { return m.Alias }
+
 // OpenAICompatibility represents the configuration for OpenAI API compatibility
 // with external providers, allowing model aliases to be routed through OpenAI API format.
 type OpenAICompatibility struct {
 	// Name is the identifier for this OpenAI compatibility configuration.
 	Name string `yaml:"name" json:"name"`
 
+	// Priority controls selection preference when multiple providers or credentials match.
+	// Higher values are preferred; defaults to 0.
+	Priority int `yaml:"priority,omitempty" json:"priority,omitempty"`
+
 	// Prefix optionally namespaces model aliases for this provider (e.g., "teamA/kimi-k2").
 	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
 
@@ -305,6 +427,9 @@ type OpenAICompatibilityModel struct {
 	Alias string `yaml:"alias" json:"alias"`
 }
 
+func (m OpenAICompatibilityModel) GetName() string  { return m.Name }
+func (m OpenAICompatibilityModel) GetAlias() string { return m.Alias }
+
 // LoadConfig reads a YAML configuration file from the given path,
 // unmarshals it into a Config struct, applies environment variable overrides,
 // and returns it.
@@ -323,6 +448,15 @@ func LoadConfig(configFile string) (*Config, error) {
 // If optional is true and the file is missing, it returns an empty Config.
 // If optional is true and the file is empty or invalid, it returns an empty Config.
 func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
+	// Perform oauth-model-alias migration before loading config.
+	// This migrates oauth-model-mappings to oauth-model-alias if needed.
+	if migrated, err := MigrateOAuthModelAlias(configFile); err != nil {
+		// Log warning but don't fail - config loading should still work
+		fmt.Printf("Warning: oauth-model-alias migration failed: %v\n", err)
+	} else if migrated {
+		fmt.Println("Migrated oauth-model-mappings to oauth-model-alias")
+	}
+
 	// Read the entire configuration file into memory.
 	data, err := os.ReadFile(configFile)
 	if err != nil {
@@ -415,6 +549,12 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 	// Normalize OAuth provider model exclusion map.
 	cfg.OAuthExcludedModels = NormalizeOAuthExcludedModels(cfg.OAuthExcludedModels)
 
+	// Normalize global OAuth model name aliases.
+	cfg.SanitizeOAuthModelAlias()
+
+	// Validate raw payload rules and drop invalid entries.
+	cfg.SanitizePayloadRules()
+
 	if cfg.legacyMigrationPending {
 		fmt.Println("Detected legacy configuration keys, attempting to persist the normalized config...")
 		if !optional && configFile != "" {
@@ -431,6 +571,99 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 	return &cfg, nil
 }
 
+// SanitizePayloadRules validates raw JSON payload rule params and drops invalid rules.
+func (cfg *Config) SanitizePayloadRules() {
+	if cfg == nil {
+		return
+	}
+	cfg.Payload.DefaultRaw = sanitizePayloadRawRules(cfg.Payload.DefaultRaw, "default-raw")
+	cfg.Payload.OverrideRaw = sanitizePayloadRawRules(cfg.Payload.OverrideRaw, "override-raw")
+}
+
+func sanitizePayloadRawRules(rules []PayloadRule, section string) []PayloadRule {
+	if len(rules) == 0 {
+		return rules
+	}
+	out := make([]PayloadRule, 0, len(rules))
+	for i := range rules {
+		rule := rules[i]
+		if len(rule.Params) == 0 {
+			continue
+		}
+		invalid := false
+		for path, value := range rule.Params {
+			raw, ok := payloadRawString(value)
+			if !ok {
+				continue
+			}
+			trimmed := bytes.TrimSpace(raw)
+			if len(trimmed) == 0 || !json.Valid(trimmed) {
+				log.WithFields(log.Fields{
+					"section":    section,
+					"rule_index": i + 1,
+					"param":      path,
+				}).Warn("payload rule dropped: invalid raw JSON")
+				invalid = true
+				break
+			}
+		}
+		if invalid {
+			continue
+		}
+		out = append(out, rule)
+	}
+	return out
+}
+
+func payloadRawString(value any) ([]byte, bool) {
+	switch typed := value.(type) {
+	case string:
+		return []byte(typed), true
+	case []byte:
+		return typed, true
+	default:
+		return nil, false
+	}
+}
+
+// SanitizeOAuthModelAlias normalizes and deduplicates global OAuth model name aliases.
+// It trims whitespace, normalizes channel keys to lower-case, drops empty entries,
+// allows multiple aliases per upstream name, and ensures aliases are unique within each channel.
+func (cfg *Config) SanitizeOAuthModelAlias() {
+	if cfg == nil || len(cfg.OAuthModelAlias) == 0 {
+		return
+	}
+	out := make(map[string][]OAuthModelAlias, len(cfg.OAuthModelAlias))
+	for rawChannel, aliases := range cfg.OAuthModelAlias {
+		channel := strings.ToLower(strings.TrimSpace(rawChannel))
+		if channel == "" || len(aliases) == 0 {
+			continue
+		}
+		seenAlias := make(map[string]struct{}, len(aliases))
+		clean := make([]OAuthModelAlias, 0, len(aliases))
+		for _, entry := range aliases {
+			name := strings.TrimSpace(entry.Name)
+			alias := strings.TrimSpace(entry.Alias)
+			if name == "" || alias == "" {
+				continue
+			}
+			if strings.EqualFold(name, alias) {
+				continue
+			}
+			aliasKey := strings.ToLower(alias)
+			if _, ok := seenAlias[aliasKey]; ok {
+				continue
+			}
+			seenAlias[aliasKey] = struct{}{}
+			clean = append(clean, OAuthModelAlias{Name: name, Alias: alias, Fork: entry.Fork})
+		}
+		if len(clean) > 0 {
+			out[channel] = clean
+		}
+	}
+	cfg.OAuthModelAlias = out
+}
+
 // SanitizeOpenAICompatibility removes OpenAI-compatibility provider entries that are
 // not actionable, specifically those missing a BaseURL. It trims whitespace before
 // evaluation and preserves the relative order of remaining entries.
@@ -799,8 +1032,8 @@ func getOrCreateMapValue(mapNode *yaml.Node, key string) *yaml.Node {
 }
 
 // mergeMappingPreserve merges keys from src into dst mapping node while preserving
-// key order and comments of existing keys in dst. Unknown keys from src are appended
-// to dst at the end, copying their node structure from src.
+// key order and comments of existing keys in dst. New keys are only added if their
+// value is non-zero to avoid polluting the config with defaults.
 func mergeMappingPreserve(dst, src *yaml.Node) {
 	if dst == nil || src == nil {
 		return
@@ -811,20 +1044,19 @@ func mergeMappingPreserve(dst, src *yaml.Node) {
 		copyNodeShallow(dst, src)
 		return
 	}
-	// Build a lookup of existing keys in dst
 	for i := 0; i+1 < len(src.Content); i += 2 {
 		sk := src.Content[i]
 		sv := src.Content[i+1]
 		idx := findMapKeyIndex(dst, sk.Value)
 		if idx >= 0 {
-			// Merge into existing value node
+			// Merge into existing value node (always update, even to zero values)
 			dv := dst.Content[idx+1]
 			mergeNodePreserve(dv, sv)
 		} else {
-			if shouldSkipEmptyCollectionOnPersist(sk.Value, sv) {
+			// New key: only add if value is non-zero to avoid polluting config with defaults
+			if isZeroValueNode(sv) {
 				continue
 			}
-			// Append new key/value pair by deep-copying from src
 			dst.Content = append(dst.Content, deepCopyNode(sk), deepCopyNode(sv))
 		}
 	}
@@ -907,32 +1139,49 @@ func findMapKeyIndex(mapNode *yaml.Node, key string) int {
 	return -1
 }
 
-func shouldSkipEmptyCollectionOnPersist(key string, node *yaml.Node) bool {
-	switch key {
-	case "generative-language-api-key",
-		"gemini-api-key",
-		"vertex-api-key",
-		"claude-api-key",
-		"codex-api-key",
-		"openai-compatibility":
-		return isEmptyCollectionNode(node)
-	default:
-		return false
-	}
-}
-
-func isEmptyCollectionNode(node *yaml.Node) bool {
+// isZeroValueNode returns true if the YAML node represents a zero/default value
+// that should not be written as a new key to preserve config cleanliness.
+// For mappings and sequences, recursively checks if all children are zero values.
+func isZeroValueNode(node *yaml.Node) bool {
 	if node == nil {
 		return true
 	}
 	switch node.Kind {
-	case yaml.SequenceNode:
-		return len(node.Content) == 0
 	case yaml.ScalarNode:
-		return node.Tag == "!!null"
-	default:
-		return false
+		switch node.Tag {
+		case "!!bool":
+			return node.Value == "false"
+		case "!!int", "!!float":
+			return node.Value == "0" || node.Value == "0.0"
+		case "!!str":
+			return node.Value == ""
+		case "!!null":
+			return true
+		}
+	case yaml.SequenceNode:
+		if len(node.Content) == 0 {
+			return true
+		}
+		// Check if all elements are zero values
+		for _, child := range node.Content {
+			if !isZeroValueNode(child) {
+				return false
+			}
+		}
+		return true
+	case yaml.MappingNode:
+		if len(node.Content) == 0 {
+			return true
+		}
+		// Check if all values are zero values (values are at odd indices)
+		for i := 1; i < len(node.Content); i += 2 {
+			if !isZeroValueNode(node.Content[i]) {
+				return false
+			}
+		}
+		return true
 	}
+	return false
 }
 
 // deepCopyNode creates a deep copy of a yaml.Node graph.

internal/config/oauth_model_alias_migration.go

@@ -0,0 +1,258 @@
+package config
+
+import (
+	"os"
+	"strings"
+
+	"gopkg.in/yaml.v3"
+)
+
+// antigravityModelConversionTable maps old built-in aliases to actual model names
+// for the antigravity channel during migration.
+var antigravityModelConversionTable = map[string]string{
+	"gemini-2.5-computer-use-preview-10-2025": "rev19-uic3-1p",
+	"gemini-3-pro-image-preview":              "gemini-3-pro-image",
+	"gemini-3-pro-preview":                    "gemini-3-pro-high",
+	"gemini-3-flash-preview":                  "gemini-3-flash",
+	"gemini-claude-sonnet-4-5":                "claude-sonnet-4-5",
+	"gemini-claude-sonnet-4-5-thinking":       "claude-sonnet-4-5-thinking",
+	"gemini-claude-opus-4-5-thinking":         "claude-opus-4-5-thinking",
+}
+
+// defaultAntigravityAliases returns the default oauth-model-alias configuration
+// for the antigravity channel when neither field exists.
+func defaultAntigravityAliases() []OAuthModelAlias {
+	return []OAuthModelAlias{
+		{Name: "rev19-uic3-1p", Alias: "gemini-2.5-computer-use-preview-10-2025"},
+		{Name: "gemini-3-pro-image", Alias: "gemini-3-pro-image-preview"},
+		{Name: "gemini-3-pro-high", Alias: "gemini-3-pro-preview"},
+		{Name: "gemini-3-flash", Alias: "gemini-3-flash-preview"},
+		{Name: "claude-sonnet-4-5", Alias: "gemini-claude-sonnet-4-5"},
+		{Name: "claude-sonnet-4-5-thinking", Alias: "gemini-claude-sonnet-4-5-thinking"},
+		{Name: "claude-opus-4-5-thinking", Alias: "gemini-claude-opus-4-5-thinking"},
+	}
+}
+
+// MigrateOAuthModelAlias checks for and performs migration from oauth-model-mappings
+// to oauth-model-alias at startup. Returns true if migration was performed.
+//
+// Migration flow:
+// 1. Check if oauth-model-alias exists -> skip migration
+// 2. Check if oauth-model-mappings exists -> convert and migrate
+//   - For antigravity channel, convert old built-in aliases to actual model names
+//
+// 3. Neither exists -> add default antigravity config
+func MigrateOAuthModelAlias(configFile string) (bool, error) {
+	data, err := os.ReadFile(configFile)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return false, nil
+		}
+		return false, err
+	}
+	if len(data) == 0 {
+		return false, nil
+	}
+
+	// Parse YAML into node tree to preserve structure
+	var root yaml.Node
+	if err := yaml.Unmarshal(data, &root); err != nil {
+		return false, nil
+	}
+	if root.Kind != yaml.DocumentNode || len(root.Content) == 0 {
+		return false, nil
+	}
+	rootMap := root.Content[0]
+	if rootMap == nil || rootMap.Kind != yaml.MappingNode {
+		return false, nil
+	}
+
+	// Check if oauth-model-alias already exists
+	if findMapKeyIndex(rootMap, "oauth-model-alias") >= 0 {
+		return false, nil
+	}
+
+	// Check if oauth-model-mappings exists
+	oldIdx := findMapKeyIndex(rootMap, "oauth-model-mappings")
+	if oldIdx >= 0 {
+		// Migrate from old field
+		return migrateFromOldField(configFile, &root, rootMap, oldIdx)
+	}
+
+	// Neither field exists - add default antigravity config
+	return addDefaultAntigravityConfig(configFile, &root, rootMap)
+}
+
+// migrateFromOldField converts oauth-model-mappings to oauth-model-alias
+func migrateFromOldField(configFile string, root *yaml.Node, rootMap *yaml.Node, oldIdx int) (bool, error) {
+	if oldIdx+1 >= len(rootMap.Content) {
+		return false, nil
+	}
+	oldValue := rootMap.Content[oldIdx+1]
+	if oldValue == nil || oldValue.Kind != yaml.MappingNode {
+		return false, nil
+	}
+
+	// Parse the old aliases
+	oldAliases := parseOldAliasNode(oldValue)
+	if len(oldAliases) == 0 {
+		// Remove the old field and write
+		removeMapKeyByIndex(rootMap, oldIdx)
+		return writeYAMLNode(configFile, root)
+	}
+
+	// Convert model names for antigravity channel
+	newAliases := make(map[string][]OAuthModelAlias, len(oldAliases))
+	for channel, entries := range oldAliases {
+		converted := make([]OAuthModelAlias, 0, len(entries))
+		for _, entry := range entries {
+			newEntry := OAuthModelAlias{
+				Name:  entry.Name,
+				Alias: entry.Alias,
+				Fork:  entry.Fork,
+			}
+			// Convert model names for antigravity channel
+			if strings.EqualFold(channel, "antigravity") {
+				if actual, ok := antigravityModelConversionTable[entry.Name]; ok {
+					newEntry.Name = actual
+				}
+			}
+			converted = append(converted, newEntry)
+		}
+		newAliases[channel] = converted
+	}
+
+	// Build new node
+	newNode := buildOAuthModelAliasNode(newAliases)
+
+	// Replace old key with new key and value
+	rootMap.Content[oldIdx].Value = "oauth-model-alias"
+	rootMap.Content[oldIdx+1] = newNode
+
+	return writeYAMLNode(configFile, root)
+}
+
+// addDefaultAntigravityConfig adds the default antigravity configuration
+func addDefaultAntigravityConfig(configFile string, root *yaml.Node, rootMap *yaml.Node) (bool, error) {
+	defaults := map[string][]OAuthModelAlias{
+		"antigravity": defaultAntigravityAliases(),
+	}
+	newNode := buildOAuthModelAliasNode(defaults)
+
+	// Add new key-value pair
+	keyNode := &yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "oauth-model-alias"}
+	rootMap.Content = append(rootMap.Content, keyNode, newNode)
+
+	return writeYAMLNode(configFile, root)
+}
+
+// parseOldAliasNode parses the old oauth-model-mappings node structure
+func parseOldAliasNode(node *yaml.Node) map[string][]OAuthModelAlias {
+	if node == nil || node.Kind != yaml.MappingNode {
+		return nil
+	}
+	result := make(map[string][]OAuthModelAlias)
+	for i := 0; i+1 < len(node.Content); i += 2 {
+		channelNode := node.Content[i]
+		entriesNode := node.Content[i+1]
+		if channelNode == nil || entriesNode == nil {
+			continue
+		}
+		channel := strings.ToLower(strings.TrimSpace(channelNode.Value))
+		if channel == "" || entriesNode.Kind != yaml.SequenceNode {
+			continue
+		}
+		entries := make([]OAuthModelAlias, 0, len(entriesNode.Content))
+		for _, entryNode := range entriesNode.Content {
+			if entryNode == nil || entryNode.Kind != yaml.MappingNode {
+				continue
+			}
+			entry := parseAliasEntry(entryNode)
+			if entry.Name != "" && entry.Alias != "" {
+				entries = append(entries, entry)
+			}
+		}
+		if len(entries) > 0 {
+			result[channel] = entries
+		}
+	}
+	return result
+}
+
+// parseAliasEntry parses a single alias entry node
+func parseAliasEntry(node *yaml.Node) OAuthModelAlias {
+	var entry OAuthModelAlias
+	for i := 0; i+1 < len(node.Content); i += 2 {
+		keyNode := node.Content[i]
+		valNode := node.Content[i+1]
+		if keyNode == nil || valNode == nil {
+			continue
+		}
+		switch strings.ToLower(strings.TrimSpace(keyNode.Value)) {
+		case "name":
+			entry.Name = strings.TrimSpace(valNode.Value)
+		case "alias":
+			entry.Alias = strings.TrimSpace(valNode.Value)
+		case "fork":
+			entry.Fork = strings.ToLower(strings.TrimSpace(valNode.Value)) == "true"
+		}
+	}
+	return entry
+}
+
+// buildOAuthModelAliasNode creates a YAML node for oauth-model-alias
+func buildOAuthModelAliasNode(aliases map[string][]OAuthModelAlias) *yaml.Node {
+	node := &yaml.Node{Kind: yaml.MappingNode, Tag: "!!map"}
+	for channel, entries := range aliases {
+		channelNode := &yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: channel}
+		entriesNode := &yaml.Node{Kind: yaml.SequenceNode, Tag: "!!seq"}
+		for _, entry := range entries {
+			entryNode := &yaml.Node{Kind: yaml.MappingNode, Tag: "!!map"}
+			entryNode.Content = append(entryNode.Content,
+				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "name"},
+				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: entry.Name},
+				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "alias"},
+				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: entry.Alias},
+			)
+			if entry.Fork {
+				entryNode.Content = append(entryNode.Content,
+					&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "fork"},
+					&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!bool", Value: "true"},
+				)
+			}
+			entriesNode.Content = append(entriesNode.Content, entryNode)
+		}
+		node.Content = append(node.Content, channelNode, entriesNode)
+	}
+	return node
+}
+
+// removeMapKeyByIndex removes a key-value pair from a mapping node by index
+func removeMapKeyByIndex(mapNode *yaml.Node, keyIdx int) {
+	if mapNode == nil || mapNode.Kind != yaml.MappingNode {
+		return
+	}
+	if keyIdx < 0 || keyIdx+1 >= len(mapNode.Content) {
+		return
+	}
+	mapNode.Content = append(mapNode.Content[:keyIdx], mapNode.Content[keyIdx+2:]...)
+}
+
+// writeYAMLNode writes the YAML node tree back to file
+func writeYAMLNode(configFile string, root *yaml.Node) (bool, error) {
+	f, err := os.Create(configFile)
+	if err != nil {
+		return false, err
+	}
+	defer f.Close()
+
+	enc := yaml.NewEncoder(f)
+	enc.SetIndent(2)
+	if err := enc.Encode(root); err != nil {
+		return false, err
+	}
+	if err := enc.Close(); err != nil {
+		return false, err
+	}
+	return true, nil
+}

internal/config/oauth_model_alias_migration_test.go

@@ -0,0 +1,225 @@
+package config
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gopkg.in/yaml.v3"
+)
+
+func TestMigrateOAuthModelAlias_SkipsIfNewFieldExists(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	content := `oauth-model-alias:
+  gemini-cli:
+    - name: "gemini-2.5-pro"
+      alias: "g2.5p"
+`
+	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if migrated {
+		t.Fatal("expected no migration when oauth-model-alias already exists")
+	}
+
+	// Verify file unchanged
+	data, _ := os.ReadFile(configFile)
+	if !strings.Contains(string(data), "oauth-model-alias:") {
+		t.Fatal("file should still contain oauth-model-alias")
+	}
+}
+
+func TestMigrateOAuthModelAlias_MigratesOldField(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	content := `oauth-model-mappings:
+  gemini-cli:
+    - name: "gemini-2.5-pro"
+      alias: "g2.5p"
+      fork: true
+`
+	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !migrated {
+		t.Fatal("expected migration to occur")
+	}
+
+	// Verify new field exists and old field removed
+	data, _ := os.ReadFile(configFile)
+	if strings.Contains(string(data), "oauth-model-mappings:") {
+		t.Fatal("old field should be removed")
+	}
+	if !strings.Contains(string(data), "oauth-model-alias:") {
+		t.Fatal("new field should exist")
+	}
+
+	// Parse and verify structure
+	var root yaml.Node
+	if err := yaml.Unmarshal(data, &root); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestMigrateOAuthModelAlias_ConvertsAntigravityModels(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	// Use old model names that should be converted
+	content := `oauth-model-mappings:
+  antigravity:
+    - name: "gemini-2.5-computer-use-preview-10-2025"
+      alias: "computer-use"
+    - name: "gemini-3-pro-preview"
+      alias: "g3p"
+`
+	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !migrated {
+		t.Fatal("expected migration to occur")
+	}
+
+	// Verify model names were converted
+	data, _ := os.ReadFile(configFile)
+	content = string(data)
+	if !strings.Contains(content, "rev19-uic3-1p") {
+		t.Fatal("expected gemini-2.5-computer-use-preview-10-2025 to be converted to rev19-uic3-1p")
+	}
+	if !strings.Contains(content, "gemini-3-pro-high") {
+		t.Fatal("expected gemini-3-pro-preview to be converted to gemini-3-pro-high")
+	}
+}
+
+func TestMigrateOAuthModelAlias_AddsDefaultIfNeitherExists(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	content := `debug: true
+port: 8080
+`
+	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !migrated {
+		t.Fatal("expected migration to add default config")
+	}
+
+	// Verify default antigravity config was added
+	data, _ := os.ReadFile(configFile)
+	content = string(data)
+	if !strings.Contains(content, "oauth-model-alias:") {
+		t.Fatal("expected oauth-model-alias to be added")
+	}
+	if !strings.Contains(content, "antigravity:") {
+		t.Fatal("expected antigravity channel to be added")
+	}
+	if !strings.Contains(content, "rev19-uic3-1p") {
+		t.Fatal("expected default antigravity aliases to include rev19-uic3-1p")
+	}
+}
+
+func TestMigrateOAuthModelAlias_PreservesOtherConfig(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	content := `debug: true
+port: 8080
+oauth-model-mappings:
+  gemini-cli:
+    - name: "test"
+      alias: "t"
+api-keys:
+  - "key1"
+  - "key2"
+`
+	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !migrated {
+		t.Fatal("expected migration to occur")
+	}
+
+	// Verify other config preserved
+	data, _ := os.ReadFile(configFile)
+	content = string(data)
+	if !strings.Contains(content, "debug: true") {
+		t.Fatal("expected debug field to be preserved")
+	}
+	if !strings.Contains(content, "port: 8080") {
+		t.Fatal("expected port field to be preserved")
+	}
+	if !strings.Contains(content, "api-keys:") {
+		t.Fatal("expected api-keys field to be preserved")
+	}
+}
+
+func TestMigrateOAuthModelAlias_NonexistentFile(t *testing.T) {
+	t.Parallel()
+
+	migrated, err := MigrateOAuthModelAlias("/nonexistent/path/config.yaml")
+	if err != nil {
+		t.Fatalf("unexpected error for nonexistent file: %v", err)
+	}
+	if migrated {
+		t.Fatal("expected no migration for nonexistent file")
+	}
+}
+
+func TestMigrateOAuthModelAlias_EmptyFile(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	if err := os.WriteFile(configFile, []byte(""), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if migrated {
+		t.Fatal("expected no migration for empty file")
+	}
+}

internal/config/oauth_model_alias_test.go

@@ -0,0 +1,56 @@
+package config
+
+import "testing"
+
+func TestSanitizeOAuthModelAlias_PreservesForkFlag(t *testing.T) {
+	cfg := &Config{
+		OAuthModelAlias: map[string][]OAuthModelAlias{
+			" CoDeX ": {
+				{Name: " gpt-5 ", Alias: " g5 ", Fork: true},
+				{Name: "gpt-6", Alias: "g6"},
+			},
+		},
+	}
+
+	cfg.SanitizeOAuthModelAlias()
+
+	aliases := cfg.OAuthModelAlias["codex"]
+	if len(aliases) != 2 {
+		t.Fatalf("expected 2 sanitized aliases, got %d", len(aliases))
+	}
+	if aliases[0].Name != "gpt-5" || aliases[0].Alias != "g5" || !aliases[0].Fork {
+		t.Fatalf("expected first alias to be gpt-5->g5 fork=true, got name=%q alias=%q fork=%v", aliases[0].Name, aliases[0].Alias, aliases[0].Fork)
+	}
+	if aliases[1].Name != "gpt-6" || aliases[1].Alias != "g6" || aliases[1].Fork {
+		t.Fatalf("expected second alias to be gpt-6->g6 fork=false, got name=%q alias=%q fork=%v", aliases[1].Name, aliases[1].Alias, aliases[1].Fork)
+	}
+}
+
+func TestSanitizeOAuthModelAlias_AllowsMultipleAliasesForSameName(t *testing.T) {
+	cfg := &Config{
+		OAuthModelAlias: map[string][]OAuthModelAlias{
+			"antigravity": {
+				{Name: "gemini-claude-opus-4-5-thinking", Alias: "claude-opus-4-5-20251101", Fork: true},
+				{Name: "gemini-claude-opus-4-5-thinking", Alias: "claude-opus-4-5-20251101-thinking", Fork: true},
+				{Name: "gemini-claude-opus-4-5-thinking", Alias: "claude-opus-4-5", Fork: true},
+			},
+		},
+	}
+
+	cfg.SanitizeOAuthModelAlias()
+
+	aliases := cfg.OAuthModelAlias["antigravity"]
+	expected := []OAuthModelAlias{
+		{Name: "gemini-claude-opus-4-5-thinking", Alias: "claude-opus-4-5-20251101", Fork: true},
+		{Name: "gemini-claude-opus-4-5-thinking", Alias: "claude-opus-4-5-20251101-thinking", Fork: true},
+		{Name: "gemini-claude-opus-4-5-thinking", Alias: "claude-opus-4-5", Fork: true},
+	}
+	if len(aliases) != len(expected) {
+		t.Fatalf("expected %d sanitized aliases, got %d", len(expected), len(aliases))
+	}
+	for i, exp := range expected {
+		if aliases[i].Name != exp.Name || aliases[i].Alias != exp.Alias || aliases[i].Fork != exp.Fork {
+			t.Fatalf("expected alias %d to be name=%q alias=%q fork=%v, got name=%q alias=%q fork=%v", i, exp.Name, exp.Alias, exp.Fork, aliases[i].Name, aliases[i].Alias, aliases[i].Fork)
+		}
+	}
+}

internal/config/sdk_config.go

@@ -22,6 +22,25 @@ type SDKConfig struct {
 
 	// Access holds request authentication provider configuration.
 	Access AccessConfig `yaml:"auth,omitempty" json:"auth,omitempty"`
+
+	// Streaming configures server-side streaming behavior (keep-alives and safe bootstrap retries).
+	Streaming StreamingConfig `yaml:"streaming" json:"streaming"`
+
+	// NonStreamKeepAliveInterval controls how often blank lines are emitted for non-streaming responses.
+	// <= 0 disables keep-alives. Value is in seconds.
+	NonStreamKeepAliveInterval int `yaml:"nonstream-keepalive-interval,omitempty" json:"nonstream-keepalive-interval,omitempty"`
+}
+
+// StreamingConfig holds server streaming behavior configuration.
+type StreamingConfig struct {
+	// KeepAliveSeconds controls how often the server emits SSE heartbeats (": keep-alive\n\n").
+	// <= 0 disables keep-alives. Default is 0.
+	KeepAliveSeconds int `yaml:"keepalive-seconds,omitempty" json:"keepalive-seconds,omitempty"`
+
+	// BootstrapRetries controls how many times the server may retry a streaming request before any bytes are sent,
+	// to allow auth rotation / transient recovery.
+	// <= 0 disables bootstrap retries. Default is 0.
+	BootstrapRetries int `yaml:"bootstrap-retries,omitempty" json:"bootstrap-retries,omitempty"`
 }
 
 // AccessConfig groups request authentication providers.

internal/config/vertex_compat.go

@@ -13,6 +13,10 @@ type VertexCompatKey struct {
 	// Maps to the x-goog-api-key header.
 	APIKey string `yaml:"api-key" json:"api-key"`
 
+	// Priority controls selection preference when multiple credentials match.
+	// Higher values are preferred; defaults to 0.
+	Priority int `yaml:"priority,omitempty" json:"priority,omitempty"`
+
 	// Prefix optionally namespaces model aliases for this credential (e.g., "teamA/vertex-pro").
 	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
 
@@ -32,6 +36,9 @@ type VertexCompatKey struct {
 	Models []VertexCompatModel `yaml:"models,omitempty" json:"models,omitempty"`
 }
 
+func (k VertexCompatKey) GetAPIKey() string  { return k.APIKey }
+func (k VertexCompatKey) GetBaseURL() string { return k.BaseURL }
+
 // VertexCompatModel represents a model configuration for Vertex compatibility,
 // including the actual model name and its alias for API routing.
 type VertexCompatModel struct {
@@ -42,6 +49,9 @@ type VertexCompatModel struct {
 	Alias string `yaml:"alias" json:"alias"`
 }
 
+func (m VertexCompatModel) GetName() string  { return m.Name }
+func (m VertexCompatModel) GetAlias() string { return m.Alias }
+
 // SanitizeVertexCompatKeys deduplicates and normalizes Vertex-compatible API key credentials.
 func (cfg *Config) SanitizeVertexCompatKeys() {
 	if cfg == nil {

internal/logging/gin_logger.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"net/http"
 	"runtime/debug"
+	"strings"
 	"time"
 
 	"github.com/gin-gonic/gin"
@@ -14,11 +15,24 @@ import (
 	log "github.com/sirupsen/logrus"
 )
 
+// aiAPIPrefixes defines path prefixes for AI API requests that should have request ID tracking.
+var aiAPIPrefixes = []string{
+	"/v1/chat/completions",
+	"/v1/completions",
+	"/v1/messages",
+	"/v1/responses",
+	"/v1beta/models/",
+	"/api/provider/",
+}
+
 const skipGinLogKey = "__gin_skip_request_logging__"
 
 // GinLogrusLogger returns a Gin middleware handler that logs HTTP requests and responses
 // using logrus. It captures request details including method, path, status code, latency,
-// client IP, and any error messages, formatting them in a Gin-style log format.
+// client IP, and any error messages. Request ID is only added for AI API requests.
+//
+// Output format (AI API): [2025-12-23 20:14:10] [info ] | a1b2c3d4 | 200 |       23.559s | ...
+// Output format (others): [2025-12-23 20:14:10] [info ] | -------- | 200 |       23.559s | ...
 //
 // Returns:
 //   - gin.HandlerFunc: A middleware handler for request logging
@@ -28,6 +42,15 @@ func GinLogrusLogger() gin.HandlerFunc {
 		path := c.Request.URL.Path
 		raw := util.MaskSensitiveQuery(c.Request.URL.RawQuery)
 
+		// Only generate request ID for AI API paths
+		var requestID string
+		if isAIAPIPath(path) {
+			requestID = GenerateRequestID()
+			SetGinRequestID(c, requestID)
+			ctx := WithRequestID(c.Request.Context(), requestID)
+			c.Request = c.Request.WithContext(ctx)
+		}
+
 		c.Next()
 
 		if shouldSkipGinRequestLogging(c) {
@@ -49,23 +72,38 @@ func GinLogrusLogger() gin.HandlerFunc {
 		clientIP := c.ClientIP()
 		method := c.Request.Method
 		errorMessage := c.Errors.ByType(gin.ErrorTypePrivate).String()
-		timestamp := time.Now().Format("2006/01/02 - 15:04:05")
-		logLine := fmt.Sprintf("[GIN] %s | %3d | %13v | %15s | %-7s \"%s\"", timestamp, statusCode, latency, clientIP, method, path)
+
+		if requestID == "" {
+			requestID = "--------"
+		}
+		logLine := fmt.Sprintf("%3d | %13v | %15s | %-7s \"%s\"", statusCode, latency, clientIP, method, path)
 		if errorMessage != "" {
 			logLine = logLine + " | " + errorMessage
 		}
 
+		entry := log.WithField("request_id", requestID)
+
 		switch {
 		case statusCode >= http.StatusInternalServerError:
-			log.Error(logLine)
+			entry.Error(logLine)
 		case statusCode >= http.StatusBadRequest:
-			log.Warn(logLine)
+			entry.Warn(logLine)
 		default:
-			log.Info(logLine)
+			entry.Info(logLine)
 		}
 	}
 }
 
+// isAIAPIPath checks if the given path is an AI API endpoint that should have request ID tracking.
+func isAIAPIPath(path string) bool {
+	for _, prefix := range aiAPIPrefixes {
+		if strings.HasPrefix(path, prefix) {
+			return true
+		}
+	}
+	return false
+}
+
 // GinLogrusRecovery returns a Gin middleware handler that recovers from panics and logs
 // them using logrus. When a panic occurs, it captures the panic value, stack trace,
 // and request path, then returns a 500 Internal Server Error response to the client.

internal/logging/global_logger.go

@@ -10,6 +10,7 @@ import (
 	"sync"
 
 	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	log "github.com/sirupsen/logrus"
 	"gopkg.in/natefinch/lumberjack.v2"
@@ -24,9 +25,13 @@ var (
 )
 
 // LogFormatter defines a custom log format for logrus.
-// This formatter adds timestamp, level, and source location to each log entry.
+// This formatter adds timestamp, level, request ID, and source location to each log entry.
+// Format: [2025-12-23 20:14:04] [debug] [manager.go:524] | a1b2c3d4 | Use API key sk-9...0RHO for model gpt-5.2
 type LogFormatter struct{}
 
+// logFieldOrder defines the display order for common log fields.
+var logFieldOrder = []string{"provider", "model", "mode", "budget", "level", "original_value", "min", "max", "clamped_to", "error"}
+
 // Format renders a single log entry with custom formatting.
 func (m *LogFormatter) Format(entry *log.Entry) ([]byte, error) {
 	var buffer *bytes.Buffer
@@ -39,11 +44,36 @@ func (m *LogFormatter) Format(entry *log.Entry) ([]byte, error) {
 	timestamp := entry.Time.Format("2006-01-02 15:04:05")
 	message := strings.TrimRight(entry.Message, "\r\n")
 
+	reqID := "--------"
+	if id, ok := entry.Data["request_id"].(string); ok && id != "" {
+		reqID = id
+	}
+
+	level := entry.Level.String()
+	if level == "warning" {
+		level = "warn"
+	}
+	levelStr := fmt.Sprintf("%-5s", level)
+
+	// Build fields string (only print fields in logFieldOrder)
+	var fieldsStr string
+	if len(entry.Data) > 0 {
+		var fields []string
+		for _, k := range logFieldOrder {
+			if v, ok := entry.Data[k]; ok {
+				fields = append(fields, fmt.Sprintf("%s=%v", k, v))
+			}
+		}
+		if len(fields) > 0 {
+			fieldsStr = " " + strings.Join(fields, " ")
+		}
+	}
+
 	var formatted string
 	if entry.Caller != nil {
-		formatted = fmt.Sprintf("[%s] [%s] [%s:%d] %s\n", timestamp, entry.Level, filepath.Base(entry.Caller.File), entry.Caller.Line, message)
+		formatted = fmt.Sprintf("[%s] [%s] [%s] [%s:%d] %s%s\n", timestamp, reqID, levelStr, filepath.Base(entry.Caller.File), entry.Caller.Line, message, fieldsStr)
 	} else {
-		formatted = fmt.Sprintf("[%s] [%s] %s\n", timestamp, entry.Level, message)
+		formatted = fmt.Sprintf("[%s] [%s] [%s] %s%s\n", timestamp, reqID, levelStr, message, fieldsStr)
 	}
 	buffer.WriteString(formatted)
 
@@ -71,10 +101,30 @@ func SetupBaseLogger() {
 	})
 }
 
+// isDirWritable checks if the specified directory exists and is writable by attempting to create and remove a test file.
+func isDirWritable(dir string) bool {
+	info, err := os.Stat(dir)
+	if err != nil || !info.IsDir() {
+		return false
+	}
+
+	testFile := filepath.Join(dir, ".perm_test")
+	f, err := os.Create(testFile)
+	if err != nil {
+		return false
+	}
+
+	defer func() {
+		_ = f.Close()
+		_ = os.Remove(testFile)
+	}()
+	return true
+}
+
 // ConfigureLogOutput switches the global log destination between rotating files and stdout.
 // When logsMaxTotalSizeMB > 0, a background cleaner removes the oldest log files in the logs directory
 // until the total size is within the limit.
-func ConfigureLogOutput(loggingToFile bool, logsMaxTotalSizeMB int) error {
+func ConfigureLogOutput(cfg *config.Config) error {
 	SetupBaseLogger()
 
 	writerMu.Lock()
@@ -83,10 +133,12 @@ func ConfigureLogOutput(loggingToFile bool, logsMaxTotalSizeMB int) error {
 	logDir := "logs"
 	if base := util.WritablePath(); base != "" {
 		logDir = filepath.Join(base, "logs")
+	} else if !isDirWritable(logDir) {
+		logDir = filepath.Join(cfg.AuthDir, "logs")
 	}
 
 	protectedPath := ""
-	if loggingToFile {
+	if cfg.LoggingToFile {
 		if err := os.MkdirAll(logDir, 0o755); err != nil {
 			return fmt.Errorf("logging: failed to create log directory: %w", err)
 		}
@@ -110,7 +162,7 @@ func ConfigureLogOutput(loggingToFile bool, logsMaxTotalSizeMB int) error {
 		log.SetOutput(os.Stdout)
 	}
 
-	configureLogDirCleanerLocked(logDir, logsMaxTotalSizeMB, protectedPath)
+	configureLogDirCleanerLocked(logDir, cfg.LogsMaxTotalSizeMB, protectedPath)
 	return nil
 }
 

internal/logging/request_logger.go

@@ -43,10 +43,11 @@ type RequestLogger interface {
 	//   - response: The raw response data
 	//   - apiRequest: The API request data
 	//   - apiResponse: The API response data
+	//   - requestID: Optional request ID for log file naming
 	//
 	// Returns:
 	//   - error: An error if logging fails, nil otherwise
-	LogRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage) error
+	LogRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, requestID string) error
 
 	// LogStreamingRequest initiates logging for a streaming request and returns a writer for chunks.
 	//
@@ -55,11 +56,12 @@ type RequestLogger interface {
 	//   - method: The HTTP method
 	//   - headers: The request headers
 	//   - body: The request body
+	//   - requestID: Optional request ID for log file naming
 	//
 	// Returns:
 	//   - StreamingLogWriter: A writer for streaming response chunks
 	//   - error: An error if logging initialization fails, nil otherwise
-	LogStreamingRequest(url, method string, headers map[string][]string, body []byte) (StreamingLogWriter, error)
+	LogStreamingRequest(url, method string, headers map[string][]string, body []byte, requestID string) (StreamingLogWriter, error)
 
 	// IsEnabled returns whether request logging is currently enabled.
 	//
@@ -177,20 +179,21 @@ func (l *FileRequestLogger) SetEnabled(enabled bool) {
 //   - response: The raw response data
 //   - apiRequest: The API request data
 //   - apiResponse: The API response data
+//   - requestID: Optional request ID for log file naming
 //
 // Returns:
 //   - error: An error if logging fails, nil otherwise
-func (l *FileRequestLogger) LogRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage) error {
-	return l.logRequest(url, method, requestHeaders, body, statusCode, responseHeaders, response, apiRequest, apiResponse, apiResponseErrors, false)
+func (l *FileRequestLogger) LogRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, requestID string) error {
+	return l.logRequest(url, method, requestHeaders, body, statusCode, responseHeaders, response, apiRequest, apiResponse, apiResponseErrors, false, requestID)
 }
 
 // LogRequestWithOptions logs a request with optional forced logging behavior.
 // The force flag allows writing error logs even when regular request logging is disabled.
-func (l *FileRequestLogger) LogRequestWithOptions(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, force bool) error {
-	return l.logRequest(url, method, requestHeaders, body, statusCode, responseHeaders, response, apiRequest, apiResponse, apiResponseErrors, force)
+func (l *FileRequestLogger) LogRequestWithOptions(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, force bool, requestID string) error {
+	return l.logRequest(url, method, requestHeaders, body, statusCode, responseHeaders, response, apiRequest, apiResponse, apiResponseErrors, force, requestID)
 }
 
-func (l *FileRequestLogger) logRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, force bool) error {
+func (l *FileRequestLogger) logRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, force bool, requestID string) error {
 	if !l.enabled && !force {
 		return nil
 	}
@@ -200,10 +203,10 @@ func (l *FileRequestLogger) logRequest(url, method string, requestHeaders map[st
 		return fmt.Errorf("failed to create logs directory: %w", errEnsure)
 	}
 
-	// Generate filename
-	filename := l.generateFilename(url)
+	// Generate filename with request ID
+	filename := l.generateFilename(url, requestID)
 	if force && !l.enabled {
-		filename = l.generateErrorFilename(url)
+		filename = l.generateErrorFilename(url, requestID)
 	}
 	filePath := filepath.Join(l.logsDir, filename)
 
@@ -271,11 +274,12 @@ func (l *FileRequestLogger) logRequest(url, method string, requestHeaders map[st
 //   - method: The HTTP method
 //   - headers: The request headers
 //   - body: The request body
+//   - requestID: Optional request ID for log file naming
 //
 // Returns:
 //   - StreamingLogWriter: A writer for streaming response chunks
 //   - error: An error if logging initialization fails, nil otherwise
-func (l *FileRequestLogger) LogStreamingRequest(url, method string, headers map[string][]string, body []byte) (StreamingLogWriter, error) {
+func (l *FileRequestLogger) LogStreamingRequest(url, method string, headers map[string][]string, body []byte, requestID string) (StreamingLogWriter, error) {
 	if !l.enabled {
 		return &NoOpStreamingLogWriter{}, nil
 	}
@@ -285,8 +289,8 @@ func (l *FileRequestLogger) LogStreamingRequest(url, method string, headers map[
 		return nil, fmt.Errorf("failed to create logs directory: %w", err)
 	}
 
-	// Generate filename
-	filename := l.generateFilename(url)
+	// Generate filename with request ID
+	filename := l.generateFilename(url, requestID)
 	filePath := filepath.Join(l.logsDir, filename)
 
 	requestHeaders := make(map[string][]string, len(headers))
@@ -330,8 +334,8 @@ func (l *FileRequestLogger) LogStreamingRequest(url, method string, headers map[
 }
 
 // generateErrorFilename creates a filename with an error prefix to differentiate forced error logs.
-func (l *FileRequestLogger) generateErrorFilename(url string) string {
-	return fmt.Sprintf("error-%s", l.generateFilename(url))
+func (l *FileRequestLogger) generateErrorFilename(url string, requestID ...string) string {
+	return fmt.Sprintf("error-%s", l.generateFilename(url, requestID...))
 }
 
 // ensureLogsDir creates the logs directory if it doesn't exist.
@@ -346,13 +350,15 @@ func (l *FileRequestLogger) ensureLogsDir() error {
 }
 
 // generateFilename creates a sanitized filename from the URL path and current timestamp.
+// Format: v1-responses-2025-12-23T195811-a1b2c3d4.log
 //
 // Parameters:
 //   - url: The request URL
+//   - requestID: Optional request ID to include in filename
 //
 // Returns:
 //   - string: A sanitized filename for the log file
-func (l *FileRequestLogger) generateFilename(url string) string {
+func (l *FileRequestLogger) generateFilename(url string, requestID ...string) string {
 	// Extract path from URL
 	path := url
 	if strings.Contains(url, "?") {
@@ -368,12 +374,18 @@ func (l *FileRequestLogger) generateFilename(url string) string {
 	sanitized := l.sanitizeForFilename(path)
 
 	// Add timestamp
-	timestamp := time.Now().Format("2006-01-02T150405-.000000000")
-	timestamp = strings.Replace(timestamp, ".", "", -1)
+	timestamp := time.Now().Format("2006-01-02T150405")
 
-	id := requestLogID.Add(1)
+	// Use request ID if provided, otherwise use sequential ID
+	var idPart string
+	if len(requestID) > 0 && requestID[0] != "" {
+		idPart = requestID[0]
+	} else {
+		id := requestLogID.Add(1)
+		idPart = fmt.Sprintf("%d", id)
+	}
 
-	return fmt.Sprintf("%s-%s-%d.log", sanitized, timestamp, id)
+	return fmt.Sprintf("%s-%s-%s.log", sanitized, timestamp, idPart)
 }
 
 // sanitizeForFilename replaces characters that are not safe for filenames.

internal/logging/requestid.go

@@ -0,0 +1,61 @@
+package logging
+
+import (
+	"context"
+	"crypto/rand"
+	"encoding/hex"
+
+	"github.com/gin-gonic/gin"
+)
+
+// requestIDKey is the context key for storing/retrieving request IDs.
+type requestIDKey struct{}
+
+// ginRequestIDKey is the Gin context key for request IDs.
+const ginRequestIDKey = "__request_id__"
+
+// GenerateRequestID creates a new 8-character hex request ID.
+func GenerateRequestID() string {
+	b := make([]byte, 4)
+	if _, err := rand.Read(b); err != nil {
+		return "00000000"
+	}
+	return hex.EncodeToString(b)
+}
+
+// WithRequestID returns a new context with the request ID attached.
+func WithRequestID(ctx context.Context, requestID string) context.Context {
+	return context.WithValue(ctx, requestIDKey{}, requestID)
+}
+
+// GetRequestID retrieves the request ID from the context.
+// Returns empty string if not found.
+func GetRequestID(ctx context.Context) string {
+	if ctx == nil {
+		return ""
+	}
+	if id, ok := ctx.Value(requestIDKey{}).(string); ok {
+		return id
+	}
+	return ""
+}
+
+// SetGinRequestID stores the request ID in the Gin context.
+func SetGinRequestID(c *gin.Context, requestID string) {
+	if c != nil {
+		c.Set(ginRequestIDKey, requestID)
+	}
+}
+
+// GetGinRequestID retrieves the request ID from the Gin context.
+func GetGinRequestID(c *gin.Context) string {
+	if c == nil {
+		return ""
+	}
+	if id, exists := c.Get(ginRequestIDKey); exists {
+		if s, ok := id.(string); ok {
+			return s
+		}
+	}
+	return ""
+}

internal/managementasset/updater.go

@@ -24,10 +24,11 @@ import (
 )
 
 const (
-	defaultManagementReleaseURL = "https://api.github.com/repos/router-for-me/Cli-Proxy-API-Management-Center/releases/latest"
-	managementAssetName         = "management.html"
-	httpUserAgent               = "CLIProxyAPI-management-updater"
-	updateCheckInterval         = 3 * time.Hour
+	defaultManagementReleaseURL  = "https://api.github.com/repos/router-for-me/Cli-Proxy-API-Management-Center/releases/latest"
+	defaultManagementFallbackURL = "https://cpamc.router-for.me/"
+	managementAssetName          = "management.html"
+	httpUserAgent                = "CLIProxyAPI-management-updater"
+	updateCheckInterval          = 3 * time.Hour
 )
 
 // ManagementFileName exposes the control panel asset filename.
@@ -198,6 +199,16 @@ func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL
 		return
 	}
 
+	localPath := filepath.Join(staticDir, managementAssetName)
+	localFileMissing := false
+	if _, errStat := os.Stat(localPath); errStat != nil {
+		if errors.Is(errStat, os.ErrNotExist) {
+			localFileMissing = true
+		} else {
+			log.WithError(errStat).Debug("failed to stat local management asset")
+		}
+	}
+
 	// Rate limiting: check only once every 3 hours
 	lastUpdateCheckMu.Lock()
 	now := time.Now()
@@ -210,15 +221,14 @@ func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL
 	lastUpdateCheckTime = now
 	lastUpdateCheckMu.Unlock()
 
-	if err := os.MkdirAll(staticDir, 0o755); err != nil {
-		log.WithError(err).Warn("failed to prepare static directory for management asset")
+	if errMkdirAll := os.MkdirAll(staticDir, 0o755); errMkdirAll != nil {
+		log.WithError(errMkdirAll).Warn("failed to prepare static directory for management asset")
 		return
 	}
 
 	releaseURL := resolveReleaseURL(panelRepository)
 	client := newHTTPClient(proxyURL)
 
-	localPath := filepath.Join(staticDir, managementAssetName)
 	localHash, err := fileSHA256(localPath)
 	if err != nil {
 		if !errors.Is(err, os.ErrNotExist) {
@@ -229,6 +239,13 @@ func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL
 
 	asset, remoteHash, err := fetchLatestAsset(ctx, client, releaseURL)
 	if err != nil {
+		if localFileMissing {
+			log.WithError(err).Warn("failed to fetch latest management release information, trying fallback page")
+			if ensureFallbackManagementHTML(ctx, client, localPath) {
+				return
+			}
+			return
+		}
 		log.WithError(err).Warn("failed to fetch latest management release information")
 		return
 	}
@@ -240,6 +257,13 @@ func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL
 
 	data, downloadedHash, err := downloadAsset(ctx, client, asset.BrowserDownloadURL)
 	if err != nil {
+		if localFileMissing {
+			log.WithError(err).Warn("failed to download management asset, trying fallback page")
+			if ensureFallbackManagementHTML(ctx, client, localPath) {
+				return
+			}
+			return
+		}
 		log.WithError(err).Warn("failed to download management asset")
 		return
 	}
@@ -256,6 +280,22 @@ func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL
 	log.Infof("management asset updated successfully (hash=%s)", downloadedHash)
 }
 
+func ensureFallbackManagementHTML(ctx context.Context, client *http.Client, localPath string) bool {
+	data, downloadedHash, err := downloadAsset(ctx, client, defaultManagementFallbackURL)
+	if err != nil {
+		log.WithError(err).Warn("failed to download fallback management control panel page")
+		return false
+	}
+
+	if err = atomicWriteFile(localPath, data); err != nil {
+		log.WithError(err).Warn("failed to persist fallback management control panel page")
+		return false
+	}
+
+	log.Infof("management asset updated from fallback page successfully (hash=%s)", downloadedHash)
+	return true
+}
+
 func resolveReleaseURL(repo string) string {
 	repo = strings.TrimSpace(repo)
 	if repo == "" {

internal/misc/codex_instructions.go

@@ -7,12 +7,93 @@ import (
 	"embed"
 	_ "embed"
 	"strings"
+	"sync/atomic"
+
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
 )
 
+// codexInstructionsEnabled controls whether CodexInstructionsForModel returns custom instructions.
+// When false (default), CodexInstructionsForModel returns (true, "") immediately.
+// Set via SetCodexInstructionsEnabled from config.
+var codexInstructionsEnabled atomic.Bool
+
+// SetCodexInstructionsEnabled sets whether codex instructions processing is enabled.
+func SetCodexInstructionsEnabled(enabled bool) {
+	codexInstructionsEnabled.Store(enabled)
+}
+
+// GetCodexInstructionsEnabled returns whether codex instructions processing is enabled.
+func GetCodexInstructionsEnabled() bool {
+	return codexInstructionsEnabled.Load()
+}
+
 //go:embed codex_instructions
 var codexInstructionsDir embed.FS
 
-func CodexInstructionsForModel(modelName, systemInstructions string) (bool, string) {
+//go:embed opencode_codex_instructions.txt
+var opencodeCodexInstructions string
+
+const (
+	codexUserAgentKey  = "__cpa_user_agent"
+	userAgentOpenAISDK = "ai-sdk/openai/"
+)
+
+func InjectCodexUserAgent(raw []byte, userAgent string) []byte {
+	if len(raw) == 0 {
+		return raw
+	}
+	trimmed := strings.TrimSpace(userAgent)
+	if trimmed == "" {
+		return raw
+	}
+	updated, err := sjson.SetBytes(raw, codexUserAgentKey, trimmed)
+	if err != nil {
+		return raw
+	}
+	return updated
+}
+
+func ExtractCodexUserAgent(raw []byte) string {
+	if len(raw) == 0 {
+		return ""
+	}
+	return strings.TrimSpace(gjson.GetBytes(raw, codexUserAgentKey).String())
+}
+
+func StripCodexUserAgent(raw []byte) []byte {
+	if len(raw) == 0 {
+		return raw
+	}
+	if !gjson.GetBytes(raw, codexUserAgentKey).Exists() {
+		return raw
+	}
+	updated, err := sjson.DeleteBytes(raw, codexUserAgentKey)
+	if err != nil {
+		return raw
+	}
+	return updated
+}
+
+func codexInstructionsForOpenCode(systemInstructions string) (bool, string) {
+	if opencodeCodexInstructions == "" {
+		return false, ""
+	}
+	if strings.HasPrefix(systemInstructions, opencodeCodexInstructions) {
+		return true, ""
+	}
+	return false, opencodeCodexInstructions
+}
+
+func useOpenCodeInstructions(userAgent string) bool {
+	return strings.Contains(strings.ToLower(userAgent), userAgentOpenAISDK)
+}
+
+func IsOpenCodeUserAgent(userAgent string) bool {
+	return useOpenCodeInstructions(userAgent)
+}
+
+func codexInstructionsForCodex(modelName, systemInstructions string) (bool, string) {
 	entries, _ := codexInstructionsDir.ReadDir("codex_instructions")
 
 	lastPrompt := ""
@@ -57,3 +138,13 @@ func CodexInstructionsForModel(modelName, systemInstructions string) (bool, stri
 		return false, lastPrompt
 	}
 }
+
+func CodexInstructionsForModel(modelName, systemInstructions, userAgent string) (bool, string) {
+	if !GetCodexInstructionsEnabled() {
+		return true, ""
+	}
+	if IsOpenCodeUserAgent(userAgent) {
+		return codexInstructionsForOpenCode(systemInstructions)
+	}
+	return codexInstructionsForCodex(modelName, systemInstructions)
+}

internal/misc/oauth.go

@@ -4,6 +4,8 @@ import (
 	"crypto/rand"
 	"encoding/hex"
 	"fmt"
+	"net/url"
+	"strings"
 )
 
 // GenerateRandomState generates a cryptographically secure random state parameter
@@ -19,3 +21,83 @@ func GenerateRandomState() (string, error) {
 	}
 	return hex.EncodeToString(bytes), nil
 }
+
+// OAuthCallback captures the parsed OAuth callback parameters.
+type OAuthCallback struct {
+	Code             string
+	State            string
+	Error            string
+	ErrorDescription string
+}
+
+// ParseOAuthCallback extracts OAuth parameters from a callback URL.
+// It returns nil when the input is empty.
+func ParseOAuthCallback(input string) (*OAuthCallback, error) {
+	trimmed := strings.TrimSpace(input)
+	if trimmed == "" {
+		return nil, nil
+	}
+
+	candidate := trimmed
+	if !strings.Contains(candidate, "://") {
+		if strings.HasPrefix(candidate, "?") {
+			candidate = "http://localhost" + candidate
+		} else if strings.ContainsAny(candidate, "/?#") || strings.Contains(candidate, ":") {
+			candidate = "http://" + candidate
+		} else if strings.Contains(candidate, "=") {
+			candidate = "http://localhost/?" + candidate
+		} else {
+			return nil, fmt.Errorf("invalid callback URL")
+		}
+	}
+
+	parsedURL, err := url.Parse(candidate)
+	if err != nil {
+		return nil, err
+	}
+
+	query := parsedURL.Query()
+	code := strings.TrimSpace(query.Get("code"))
+	state := strings.TrimSpace(query.Get("state"))
+	errCode := strings.TrimSpace(query.Get("error"))
+	errDesc := strings.TrimSpace(query.Get("error_description"))
+
+	if parsedURL.Fragment != "" {
+		if fragQuery, errFrag := url.ParseQuery(parsedURL.Fragment); errFrag == nil {
+			if code == "" {
+				code = strings.TrimSpace(fragQuery.Get("code"))
+			}
+			if state == "" {
+				state = strings.TrimSpace(fragQuery.Get("state"))
+			}
+			if errCode == "" {
+				errCode = strings.TrimSpace(fragQuery.Get("error"))
+			}
+			if errDesc == "" {
+				errDesc = strings.TrimSpace(fragQuery.Get("error_description"))
+			}
+		}
+	}
+
+	if code != "" && state == "" && strings.Contains(code, "#") {
+		parts := strings.SplitN(code, "#", 2)
+		code = parts[0]
+		state = parts[1]
+	}
+
+	if errCode == "" && errDesc != "" {
+		errCode = errDesc
+		errDesc = ""
+	}
+
+	if code == "" && errCode == "" {
+		return nil, fmt.Errorf("callback URL missing code")
+	}
+
+	return &OAuthCallback{
+		Code:             code,
+		State:            state,
+		Error:            errCode,
+		ErrorDescription: errDesc,
+	}, nil
+}

internal/misc/opencode_codex_instructions.txt

@@ -0,0 +1,318 @@
+You are a coding agent running in the opencode, a terminal-based coding assistant. opencode is an open source project. You are expected to be precise, safe, and helpful.
+
+Your capabilities:
+
+- Receive user prompts and other context provided by the harness, such as files in the workspace.
+- Communicate with the user by streaming thinking & responses, and by making & updating plans.
+- Emit function calls to run terminal commands and apply edits. Depending on how this specific run is configured, you can request that these function calls be escalated to the user for approval before running. More on this in the "Sandbox and approvals" section.
+
+Within this context, Codex refers to the open-source agentic coding interface (not the old Codex language model built by OpenAI).
+
+# How you work
+
+## Personality
+
+Your default personality and tone is concise, direct, and friendly. You communicate efficiently, always keeping the user clearly informed about ongoing actions without unnecessary detail. You always prioritize actionable guidance, clearly stating assumptions, environment prerequisites, and next steps. Unless explicitly asked, you avoid excessively verbose explanations about your work.
+
+# AGENTS.md spec
+- Repos often contain AGENTS.md files. These files can appear anywhere within the repository.
+- These files are a way for humans to give you (the agent) instructions or tips for working within the container.
+- Some examples might be: coding conventions, info about how code is organized, or instructions for how to run or test code.
+- Instructions in AGENTS.md files:
+    - The scope of an AGENTS.md file is the entire directory tree rooted at the folder that contains it.
+    - For every file you touch in the final patch, you must obey instructions in any AGENTS.md file whose scope includes that file.
+    - Instructions about code style, structure, naming, etc. apply only to code within the AGENTS.md file's scope, unless the file states otherwise.
+    - More-deeply-nested AGENTS.md files take precedence in the case of conflicting instructions.
+    - Direct system/developer/user instructions (as part of a prompt) take precedence over AGENTS.md instructions.
+- The contents of the AGENTS.md file at the root of the repo and any directories from the CWD up to the root are included with the developer message and don't need to be re-read. When working in a subdirectory of CWD, or a directory outside the CWD, check for any AGENTS.md files that may be applicable.
+
+## Responsiveness
+
+### Preamble messages
+
+Before making tool calls, send a brief preamble to the user explaining what you’re about to do. When sending preamble messages, follow these principles and examples:
+
+- **Logically group related actions**: if you’re about to run several related commands, describe them together in one preamble rather than sending a separate note for each.
+- **Keep it concise**: be no more than 1-2 sentences, focused on immediate, tangible next steps. (8–12 words for quick updates).
+- **Build on prior context**: if this is not your first tool call, use the preamble message to connect the dots with what’s been done so far and create a sense of momentum and clarity for the user to understand your next actions.
+- **Keep your tone light, friendly and curious**: add small touches of personality in preambles feel collaborative and engaging.
+- **Exception**: Avoid adding a preamble for every trivial read (e.g., `cat` a single file) unless it’s part of a larger grouped action.
+
+**Examples:**
+
+- “I’ve explored the repo; now checking the API route definitions.”
+- “Next, I’ll patch the config and update the related tests.”
+- “I’m about to scaffold the CLI commands and helper functions.”
+- “Ok cool, so I’ve wrapped my head around the repo. Now digging into the API routes.”
+- “Config’s looking tidy. Next up is editing helpers to keep things in sync.”
+- “Finished poking at the DB gateway. I will now chase down error handling.”
+- “Alright, build pipeline order is interesting. Checking how it reports failures.”
+- “Spotted a clever caching util; now hunting where it gets used.”
+
+## Planning
+
+You have access to an `todowrite` tool which tracks steps and progress and renders them to the user. Using the tool helps demonstrate that you've understood the task and convey how you're approaching it. Plans can help to make complex, ambiguous, or multi-phase work clearer and more collaborative for the user. A good plan should break the task into meaningful, logically ordered steps that are easy to verify as you go.
+
+Note that plans are not for padding out simple work with filler steps or stating the obvious. The content of your plan should not involve doing anything that you aren't capable of doing (i.e. don't try to test things that you can't test). Do not use plans for simple or single-step queries that you can just do or answer immediately.
+
+Do not repeat the full contents of the plan after an `todowrite` call — the harness already displays it. Instead, summarize the change made and highlight any important context or next step.
+
+Before running a command, consider whether or not you have completed the
+previous step, and make sure to mark it as completed before moving on to the
+next step. It may be the case that you complete all steps in your plan after a
+single pass of implementation. If this is the case, you can simply mark all the
+planned steps as completed. Sometimes, you may need to change plans in the
+middle of a task: call `todowrite` with the updated plan and make sure to provide an `explanation` of the rationale when doing so.
+
+Use a plan when:
+
+- The task is non-trivial and will require multiple actions over a long time horizon.
+- There are logical phases or dependencies where sequencing matters.
+- The work has ambiguity that benefits from outlining high-level goals.
+- You want intermediate checkpoints for feedback and validation.
+- When the user asked you to do more than one thing in a single prompt
+- The user has asked you to use the plan tool (aka "TODOs")
+- You generate additional steps while working, and plan to do them before yielding to the user
+
+### Examples
+
+**High-quality plans**
+
+Example 1:
+
+1. Add CLI entry with file args
+2. Parse Markdown via CommonMark library
+3. Apply semantic HTML template
+4. Handle code blocks, images, links
+5. Add error handling for invalid files
+
+Example 2:
+
+1. Define CSS variables for colors
+2. Add toggle with localStorage state
+3. Refactor components to use variables
+4. Verify all views for readability
+5. Add smooth theme-change transition
+
+Example 3:
+
+1. Set up Node.js + WebSocket server
+2. Add join/leave broadcast events
+3. Implement messaging with timestamps
+4. Add usernames + mention highlighting
+5. Persist messages in lightweight DB
+6. Add typing indicators + unread count
+
+**Low-quality plans**
+
+Example 1:
+
+1. Create CLI tool
+2. Add Markdown parser
+3. Convert to HTML
+
+Example 2:
+
+1. Add dark mode toggle
+2. Save preference
+3. Make styles look good
+
+Example 3:
+
+1. Create single-file HTML game
+2. Run quick sanity check
+3. Summarize usage instructions
+
+If you need to write a plan, only write high quality plans, not low quality ones.
+
+## Task execution
+
+You are a coding agent. Please keep going until the query is completely resolved, before ending your turn and yielding back to the user. Only terminate your turn when you are sure that the problem is solved. Autonomously resolve the query to the best of your ability, using the tools available to you, before coming back to the user. Do NOT guess or make up an answer.
+
+You MUST adhere to the following criteria when solving queries:
+
+- Working on the repo(s) in the current environment is allowed, even if they are proprietary.
+- Analyzing code for vulnerabilities is allowed.
+- Showing user code and tool call details is allowed.
+- Use the `edit` tool to edit files
+
+If completing the user's task requires writing or modifying files, your code and final answer should follow these coding guidelines, though user instructions (i.e. AGENTS.md) may override these guidelines:
+
+- Fix the problem at the root cause rather than applying surface-level patches, when possible.
+- Avoid unneeded complexity in your solution.
+- Do not attempt to fix unrelated bugs or broken tests. It is not your responsibility to fix them. (You may mention them to the user in your final message though.)
+- Update documentation as necessary.
+- Keep changes consistent with the style of the existing codebase. Changes should be minimal and focused on the task.
+- Use `git log` and `git blame` to search the history of the codebase if additional context is required.
+- NEVER add copyright or license headers unless specifically requested.
+- Do not waste tokens by re-reading files after calling `edit` on them. The tool call will fail if it didn't work. The same goes for making folders, deleting folders, etc.
+- Do not `git commit` your changes or create new git branches unless explicitly requested.
+- Do not add inline comments within code unless explicitly requested.
+- Do not use one-letter variable names unless explicitly requested.
+- NEVER output inline citations like "【F:README.md†L5-L14】" in your outputs. The CLI is not able to render these so they will just be broken in the UI. Instead, if you output valid filepaths, users will be able to click on them to open the files in their editor.
+
+## Sandbox and approvals
+
+The Codex CLI harness supports several different sandboxing, and approval configurations that the user can choose from.
+
+Filesystem sandboxing prevents you from editing files without user approval. The options are:
+
+- **read-only**: You can only read files.
+- **workspace-write**: You can read files. You can write to files in your workspace folder, but not outside it.
+- **danger-full-access**: No filesystem sandboxing.
+
+Network sandboxing prevents you from accessing network without approval. Options are
+
+- **restricted**
+- **enabled**
+
+Approvals are your mechanism to get user consent to perform more privileged actions. Although they introduce friction to the user because your work is paused until the user responds, you should leverage them to accomplish your important work. Do not let these settings or the sandbox deter you from attempting to accomplish the user's task. Approval options are
+
+- **untrusted**: The harness will escalate most commands for user approval, apart from a limited allowlist of safe "read" commands.
+- **on-failure**: The harness will allow all commands to run in the sandbox (if enabled), and failures will be escalated to the user for approval to run again without the sandbox.
+- **on-request**: Commands will be run in the sandbox by default, and you can specify in your tool call if you want to escalate a command to run without sandboxing. (Note that this mode is not always available. If it is, you'll see parameters for it in the `shell` command description.)
+- **never**: This is a non-interactive mode where you may NEVER ask the user for approval to run commands. Instead, you must always persist and work around constraints to solve the task for the user. You MUST do your utmost best to finish the task and validate your work before yielding. If this mode is pared with `danger-full-access`, take advantage of it to deliver the best outcome for the user. Further, in this mode, your default testing philosophy is overridden: Even if you don't see local patterns for testing, you may add tests and scripts to validate your work. Just remove them before yielding.
+
+When you are running with approvals `on-request`, and sandboxing enabled, here are scenarios where you'll need to request approval:
+
+- You need to run a command that writes to a directory that requires it (e.g. running tests that write to /tmp)
+- You need to run a GUI app (e.g., open/xdg-open/osascript) to open browsers or files.
+- You are running sandboxed and need to run a command that requires network access (e.g. installing packages)
+- If you run a command that is important to solving the user's query, but it fails because of sandboxing, rerun the command with approval.
+- You are about to take a potentially destructive action such as an `rm` or `git reset` that the user did not explicitly ask for
+- (For all of these, you should weigh alternative paths that do not require approval.)
+
+Note that when sandboxing is set to read-only, you'll need to request approval for any command that isn't a read.
+
+You will be told what filesystem sandboxing, network sandboxing, and approval mode are active in a developer or user message. If you are not told about this, assume that you are running with workspace-write, network sandboxing ON, and approval on-failure.
+
+## Validating your work
+
+If the codebase has tests or the ability to build or run, consider using them to verify that your work is complete.
+
+When testing, your philosophy should be to start as specific as possible to the code you changed so that you can catch issues efficiently, then make your way to broader tests as you build confidence. If there's no test for the code you changed, and if the adjacent patterns in the codebases show that there's a logical place for you to add a test, you may do so. However, do not add tests to codebases with no tests.
+
+Similarly, once you're confident in correctness, you can suggest or use formatting commands to ensure that your code is well formatted. If there are issues you can iterate up to 3 times to get formatting right, but if you still can't manage it's better to save the user time and present them a correct solution where you call out the formatting in your final message. If the codebase does not have a formatter configured, do not add one.
+
+For all of testing, running, building, and formatting, do not attempt to fix unrelated bugs. It is not your responsibility to fix them. (You may mention them to the user in your final message though.)
+
+Be mindful of whether to run validation commands proactively. In the absence of behavioral guidance:
+
+- When running in non-interactive approval modes like **never** or **on-failure**, proactively run tests, lint and do whatever you need to ensure you've completed the task.
+- When working in interactive approval modes like **untrusted**, or **on-request**, hold off on running tests or lint commands until the user is ready for you to finalize your output, because these commands take time to run and slow down iteration. Instead suggest what you want to do next, and let the user confirm first.
+- When working on test-related tasks, such as adding tests, fixing tests, or reproducing a bug to verify behavior, you may proactively run tests regardless of approval mode. Use your judgement to decide whether this is a test-related task.
+
+## Ambition vs. precision
+
+For tasks that have no prior context (i.e. the user is starting something brand new), you should feel free to be ambitious and demonstrate creativity with your implementation.
+
+If you're operating in an existing codebase, you should make sure you do exactly what the user asks with surgical precision. Treat the surrounding codebase with respect, and don't overstep (i.e. changing filenames or variables unnecessarily). You should balance being sufficiently ambitious and proactive when completing tasks of this nature.
+
+You should use judicious initiative to decide on the right level of detail and complexity to deliver based on the user's needs. This means showing good judgment that you're capable of doing the right extras without gold-plating. This might be demonstrated by high-value, creative touches when scope of the task is vague; while being surgical and targeted when scope is tightly specified.
+
+## Sharing progress updates
+
+For especially longer tasks that you work on (i.e. requiring many tool calls, or a plan with multiple steps), you should provide progress updates back to the user at reasonable intervals. These updates should be structured as a concise sentence or two (no more than 8-10 words long) recapping progress so far in plain language: this update demonstrates your understanding of what needs to be done, progress so far (i.e. files explores, subtasks complete), and where you're going next.
+
+Before doing large chunks of work that may incur latency as experienced by the user (i.e. writing a new file), you should send a concise message to the user with an update indicating what you're about to do to ensure they know what you're spending time on. Don't start editing or writing large files before informing the user what you are doing and why.
+
+The messages you send before tool calls should describe what is immediately about to be done next in very concise language. If there was previous work done, this preamble message should also include a note about the work done so far to bring the user along.
+
+## Presenting your work and final message
+
+Your final message should read naturally, like an update from a concise teammate. For casual conversation, brainstorming tasks, or quick questions from the user, respond in a friendly, conversational tone. You should ask questions, suggest ideas, and adapt to the user’s style. If you've finished a large amount of work, when describing what you've done to the user, you should follow the final answer formatting guidelines to communicate substantive changes. You don't need to add structured formatting for one-word answers, greetings, or purely conversational exchanges.
+
+You can skip heavy formatting for single, simple actions or confirmations. In these cases, respond in plain sentences with any relevant next step or quick option. Reserve multisection structured responses for results that need grouping or explanation.
+
+The user is working on the same computer as you, and has access to your work. As such there's no need to show the full contents of large files you have already written unless the user explicitly asks for them. Similarly, if you've created or modified files using `edit`, there's no need to tell users to "save the file" or "copy the code into a file"—just reference the file path.
+
+If there's something that you think you could help with as a logical next step, concisely ask the user if they want you to do so. Good examples of this are running tests, committing changes, or building out the next logical component. If there’s something that you couldn't do (even with approval) but that the user might want to do (such as verifying changes by running the app), include those instructions succinctly.
+
+Brevity is very important as a default. You should be very concise (i.e. no more than 10 lines), but can relax this requirement for tasks where additional detail and comprehensiveness is important for the user's understanding.
+
+### Final answer structure and style guidelines
+
+You are producing plain text that will later be styled by the CLI. Follow these rules exactly. Formatting should make results easy to scan, but not feel mechanical. Use judgment to decide how much structure adds value.
+
+**Section Headers**
+
+- Use only when they improve clarity — they are not mandatory for every answer.
+- Choose descriptive names that fit the content
+- Keep headers short (1–3 words) and in `**Title Case**`. Always start headers with `**` and end with `**`
+- Leave no blank line before the first bullet under a header.
+- Section headers should only be used where they genuinely improve scannability; avoid fragmenting the answer.
+
+**Bullets**
+
+- Use `-` followed by a space for every bullet.
+- Merge related points when possible; avoid a bullet for every trivial detail.
+- Keep bullets to one line unless breaking for clarity is unavoidable.
+- Group into short lists (4–6 bullets) ordered by importance.
+- Use consistent keyword phrasing and formatting across sections.
+
+**Monospace**
+
+- Wrap all commands, file paths, env vars, and code identifiers in backticks (`` `...` ``).
+- Apply to inline examples and to bullet keywords if the keyword itself is a literal file/command.
+- Never mix monospace and bold markers; choose one based on whether it’s a keyword (`**`) or inline code/path (`` ` ``).
+
+**File References**
+When referencing files in your response, make sure to include the relevant start line and always follow the below rules:
+  * Use inline code to make file paths clickable.
+  * Each reference should have a standalone path. Even if it's the same file.
+  * Accepted: absolute, workspace‑relative, a/ or b/ diff prefixes, or bare filename/suffix.
+  * Line/column (1‑based, optional): :line[:column] or #Lline[Ccolumn] (column defaults to 1).
+  * Do not use URIs like file://, vscode://, or https://.
+  * Do not provide range of lines
+  * Examples: src/app.ts, src/app.ts:42, b/server/index.js#L10, C:\repo\project\main.rs:12:5
+
+**Structure**
+
+- Place related bullets together; don’t mix unrelated concepts in the same section.
+- Order sections from general → specific → supporting info.
+- For subsections (e.g., “Binaries” under “Rust Workspace”), introduce with a bolded keyword bullet, then list items under it.
+- Match structure to complexity:
+  - Multi-part or detailed results → use clear headers and grouped bullets.
+  - Simple results → minimal headers, possibly just a short list or paragraph.
+
+**Tone**
+
+- Keep the voice collaborative and natural, like a coding partner handing off work.
+- Be concise and factual — no filler or conversational commentary and avoid unnecessary repetition
+- Use present tense and active voice (e.g., “Runs tests” not “This will run tests”).
+- Keep descriptions self-contained; don’t refer to “above” or “below”.
+- Use parallel structure in lists for consistency.
+
+**Don’t**
+
+- Don’t use literal words “bold” or “monospace” in the content.
+- Don’t nest bullets or create deep hierarchies.
+- Don’t output ANSI escape codes directly — the CLI renderer applies them.
+- Don’t cram unrelated keywords into a single bullet; split for clarity.
+- Don’t let keyword lists run long — wrap or reformat for scannability.
+
+Generally, ensure your final answers adapt their shape and depth to the request. For example, answers to code explanations should have a precise, structured explanation with code references that answer the question directly. For tasks with a simple implementation, lead with the outcome and supplement only with what’s needed for clarity. Larger changes can be presented as a logical walkthrough of your approach, grouping related steps, explaining rationale where it adds value, and highlighting next actions to accelerate the user. Your answers should provide the right level of detail while being easily scannable.
+
+For casual greetings, acknowledgements, or other one-off conversational messages that are not delivering substantive information or structured results, respond naturally without section headers or bullet formatting.
+
+# Tool Guidelines
+
+## Shell commands
+
+When using the shell, you must adhere to the following guidelines:
+
+- When searching for text or files, prefer using `rg` or `rg --files` respectively because `rg` is much faster than alternatives like `grep`. (If the `rg` command is not found, then use alternatives.)
+- Read files in chunks with a max chunk size of 250 lines. Do not use python scripts to attempt to output larger chunks of a file. Command line output will be truncated after 10 kilobytes or 256 lines of output, regardless of the command used.
+
+## `todowrite`
+
+A tool named `todowrite` is available to you. You can use it to keep an up‑to‑date, step‑by‑step plan for the task.
+
+To create a new plan, call `todowrite` with a short list of 1‑sentence steps (no more than 5-7 words each) with a `status` for each step (`pending`, `in_progress`, or `completed`).
+
+When steps have been completed, use `todowrite` to mark each finished step as
+`completed` and the next step you are working on as `in_progress`. There should
+always be exactly one `in_progress` step until everything is done. You can mark
+multiple items as complete in a single `todowrite` call.
+
+If all steps are complete, ensure you call `todowrite` to mark all steps as `completed`.

internal/registry/model_definitions.go

@@ -27,7 +27,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4.5 Sonnet",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-opus-4-5-20251101",
@@ -39,7 +39,7 @@ func GetClaudeModels() []*ModelInfo {
 			Description:         "Premium model combining maximum intelligence with practical performance",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-opus-4-1-20250805",
@@ -50,7 +50,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4.1 Opus",
 			ContextLength:       200000,
 			MaxCompletionTokens: 32000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-opus-4-20250514",
@@ -61,7 +61,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4 Opus",
 			ContextLength:       200000,
 			MaxCompletionTokens: 32000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-sonnet-4-20250514",
@@ -72,7 +72,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4 Sonnet",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-3-7-sonnet-20250219",
@@ -83,7 +83,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 3.7 Sonnet",
 			ContextLength:       128000,
 			MaxCompletionTokens: 8192,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-3-5-haiku-20241022",
@@ -432,7 +432,7 @@ func GetAIStudioModels() []*ModelInfo {
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
 		},
 		{
 			ID:                         "gemini-3-flash-preview",
@@ -447,7 +447,7 @@ func GetAIStudioModels() []*ModelInfo {
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
 		},
 		{
 			ID:                         "gemini-pro-latest",
@@ -727,6 +727,7 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "qwen3-max-preview", DisplayName: "Qwen3-Max-Preview", Description: "Qwen3 Max preview build", Created: 1757030400},
 		{ID: "kimi-k2-0905", DisplayName: "Kimi-K2-Instruct-0905", Description: "Moonshot Kimi K2 instruct 0905", Created: 1757030400},
 		{ID: "glm-4.6", DisplayName: "GLM-4.6", Description: "Zhipu GLM 4.6 general model", Created: 1759190400, Thinking: iFlowThinkingSupport},
+		{ID: "glm-4.7", DisplayName: "GLM-4.7", Description: "Zhipu GLM 4.7 general model", Created: 1766448000, Thinking: iFlowThinkingSupport},
 		{ID: "kimi-k2", DisplayName: "Kimi-K2", Description: "Moonshot Kimi K2 general model", Created: 1752192000},
 		{ID: "kimi-k2-thinking", DisplayName: "Kimi-K2-Thinking", Description: "Moonshot Kimi K2 thinking model", Created: 1762387200},
 		{ID: "deepseek-v3.2-chat", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2 Chat", Created: 1764576000},
@@ -739,7 +740,9 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "qwen3-235b-a22b-thinking-2507", DisplayName: "Qwen3-235B-A22B-Thinking", Description: "Qwen3 235B A22B Thinking (2507)", Created: 1753401600},
 		{ID: "qwen3-235b-a22b-instruct", DisplayName: "Qwen3-235B-A22B-Instruct", Description: "Qwen3 235B A22B Instruct", Created: 1753401600},
 		{ID: "qwen3-235b", DisplayName: "Qwen3-235B-A22B", Description: "Qwen3 235B A22B", Created: 1753401600},
-		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000},
+		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000, Thinking: iFlowThinkingSupport},
+		{ID: "minimax-m2.1", DisplayName: "MiniMax-M2.1", Description: "MiniMax M2.1", Created: 1766448000, Thinking: iFlowThinkingSupport},
+		{ID: "iflow-rome-30ba3b", DisplayName: "iFlow-ROME", Description: "iFlow Rome 30BA3B model", Created: 1736899200},
 	}
 	models := make([]*ModelInfo, 0, len(entries))
 	for _, entry := range entries {
@@ -766,16 +769,54 @@ type AntigravityModelConfig struct {
 }
 
 // GetAntigravityModelConfig returns static configuration for antigravity models.
-// Keys use the ALIASED model names (after modelName2Alias conversion) for direct lookup.
+// Keys use upstream model names returned by the Antigravity models endpoint.
 func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 	return map[string]*AntigravityModelConfig{
-		"gemini-2.5-flash":                        {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}, Name: "models/gemini-2.5-flash"},
-		"gemini-2.5-flash-lite":                   {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}, Name: "models/gemini-2.5-flash-lite"},
-		"gemini-2.5-computer-use-preview-10-2025": {Name: "models/gemini-2.5-computer-use-preview-10-2025"},
-		"gemini-3-pro-preview":                    {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}, Name: "models/gemini-3-pro-preview"},
-		"gemini-3-pro-image-preview":              {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}, Name: "models/gemini-3-pro-image-preview"},
-		"gemini-3-flash-preview":                  {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}, Name: "models/gemini-3-flash-preview"},
-		"gemini-claude-sonnet-4-5-thinking":       {Thinking: &ThinkingSupport{Min: 1024, Max: 200000, ZeroAllowed: false, DynamicAllowed: true}, MaxCompletionTokens: 64000},
-		"gemini-claude-opus-4-5-thinking":         {Thinking: &ThinkingSupport{Min: 1024, Max: 200000, ZeroAllowed: false, DynamicAllowed: true}, MaxCompletionTokens: 64000},
+		"gemini-2.5-flash":           {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}, Name: "models/gemini-2.5-flash"},
+		"gemini-2.5-flash-lite":      {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}, Name: "models/gemini-2.5-flash-lite"},
+		"rev19-uic3-1p":              {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true}, Name: "models/rev19-uic3-1p"},
+		"gemini-3-pro-high":          {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}, Name: "models/gemini-3-pro-high"},
+		"gemini-3-pro-image":         {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}, Name: "models/gemini-3-pro-image"},
+		"gemini-3-flash":             {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}, Name: "models/gemini-3-flash"},
+		"claude-sonnet-4-5-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
+		"claude-opus-4-5-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 	}
 }
+
+// LookupStaticModelInfo searches all static model definitions for a model by ID.
+// Returns nil if no matching model is found.
+func LookupStaticModelInfo(modelID string) *ModelInfo {
+	if modelID == "" {
+		return nil
+	}
+
+	allModels := [][]*ModelInfo{
+		GetClaudeModels(),
+		GetGeminiModels(),
+		GetGeminiVertexModels(),
+		GetGeminiCLIModels(),
+		GetAIStudioModels(),
+		GetOpenAIModels(),
+		GetQwenModels(),
+		GetIFlowModels(),
+	}
+	for _, models := range allModels {
+		for _, m := range models {
+			if m != nil && m.ID == modelID {
+				return m
+			}
+		}
+	}
+
+	// Check Antigravity static config
+	if cfg := GetAntigravityModelConfig()[modelID]; cfg != nil && cfg.Thinking != nil {
+		return &ModelInfo{
+			ID:                  modelID,
+			Name:                cfg.Name,
+			Thinking:            cfg.Thinking,
+			MaxCompletionTokens: cfg.MaxCompletionTokens,
+		}
+	}
+
+	return nil
+}

internal/registry/model_registry.go

@@ -4,6 +4,7 @@
 package registry
 
 import (
+	"context"
 	"fmt"
 	"sort"
 	"strings"
@@ -50,6 +51,11 @@ type ModelInfo struct {
 	// Thinking holds provider-specific reasoning/thinking budget capabilities.
 	// This is optional and currently used for Gemini thinking budget normalization.
 	Thinking *ThinkingSupport `json:"thinking,omitempty"`
+
+	// UserDefined indicates this model was defined through config file's models[]
+	// array (e.g., openai-compatibility.*.models[], *-api-key.models[]).
+	// UserDefined models have thinking configuration passed through without validation.
+	UserDefined bool `json:"-"`
 }
 
 // ThinkingSupport describes a model family's supported internal reasoning budget range.
@@ -84,6 +90,13 @@ type ModelRegistration struct {
 	SuspendedClients map[string]string
 }
 
+// ModelRegistryHook provides optional callbacks for external integrations to track model list changes.
+// Hook implementations must be non-blocking and resilient; calls are executed asynchronously and panics are recovered.
+type ModelRegistryHook interface {
+	OnModelsRegistered(ctx context.Context, provider, clientID string, models []*ModelInfo)
+	OnModelsUnregistered(ctx context.Context, provider, clientID string)
+}
+
 // ModelRegistry manages the global registry of available models
 type ModelRegistry struct {
 	// models maps model ID to registration information
@@ -97,6 +110,8 @@ type ModelRegistry struct {
 	clientProviders map[string]string
 	// mutex ensures thread-safe access to the registry
 	mutex *sync.RWMutex
+	// hook is an optional callback sink for model registration changes
+	hook ModelRegistryHook
 }
 
 // Global model registry instance
@@ -117,6 +132,68 @@ func GetGlobalRegistry() *ModelRegistry {
 	return globalRegistry
 }
 
+// LookupModelInfo searches the dynamic registry first, then falls back to static model definitions.
+//
+// This helper exists because some code paths only have a model ID and still need Thinking and
+// max completion token metadata even when the dynamic registry hasn't been populated.
+func LookupModelInfo(modelID string) *ModelInfo {
+	modelID = strings.TrimSpace(modelID)
+	if modelID == "" {
+		return nil
+	}
+	if info := GetGlobalRegistry().GetModelInfo(modelID); info != nil {
+		return info
+	}
+	return LookupStaticModelInfo(modelID)
+}
+
+// SetHook sets an optional hook for observing model registration changes.
+func (r *ModelRegistry) SetHook(hook ModelRegistryHook) {
+	if r == nil {
+		return
+	}
+	r.mutex.Lock()
+	defer r.mutex.Unlock()
+	r.hook = hook
+}
+
+const defaultModelRegistryHookTimeout = 5 * time.Second
+
+func (r *ModelRegistry) triggerModelsRegistered(provider, clientID string, models []*ModelInfo) {
+	hook := r.hook
+	if hook == nil {
+		return
+	}
+	modelsCopy := cloneModelInfosUnique(models)
+	go func() {
+		defer func() {
+			if recovered := recover(); recovered != nil {
+				log.Errorf("model registry hook OnModelsRegistered panic: %v", recovered)
+			}
+		}()
+		ctx, cancel := context.WithTimeout(context.Background(), defaultModelRegistryHookTimeout)
+		defer cancel()
+		hook.OnModelsRegistered(ctx, provider, clientID, modelsCopy)
+	}()
+}
+
+func (r *ModelRegistry) triggerModelsUnregistered(provider, clientID string) {
+	hook := r.hook
+	if hook == nil {
+		return
+	}
+	go func() {
+		defer func() {
+			if recovered := recover(); recovered != nil {
+				log.Errorf("model registry hook OnModelsUnregistered panic: %v", recovered)
+			}
+		}()
+		ctx, cancel := context.WithTimeout(context.Background(), defaultModelRegistryHookTimeout)
+		defer cancel()
+		hook.OnModelsUnregistered(ctx, provider, clientID)
+	}()
+}
+
 // RegisterClient registers a client and its supported models
 // Parameters:
 //   - clientID: Unique identifier for the client
@@ -177,6 +254,7 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 		} else {
 			delete(r.clientProviders, clientID)
 		}
+		r.triggerModelsRegistered(provider, clientID, models)
 		log.Debugf("Registered client %s from provider %s with %d models", clientID, clientProvider, len(rawModelIDs))
 		misc.LogCredentialSeparator()
 		return
@@ -310,6 +388,7 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 		delete(r.clientProviders, clientID)
 	}
 
+	r.triggerModelsRegistered(provider, clientID, models)
 	if len(added) == 0 && len(removed) == 0 && !providerChanged {
 		// Only metadata (e.g., display name) changed; skip separator when no log output.
 		return
@@ -400,6 +479,25 @@ func cloneModelInfo(model *ModelInfo) *ModelInfo {
 	return &copyModel
 }
 
+func cloneModelInfosUnique(models []*ModelInfo) []*ModelInfo {
+	if len(models) == 0 {
+		return nil
+	}
+	cloned := make([]*ModelInfo, 0, len(models))
+	seen := make(map[string]struct{}, len(models))
+	for _, model := range models {
+		if model == nil || model.ID == "" {
+			continue
+		}
+		if _, exists := seen[model.ID]; exists {
+			continue
+		}
+		seen[model.ID] = struct{}{}
+		cloned = append(cloned, cloneModelInfo(model))
+	}
+	return cloned
+}
+
 // UnregisterClient removes a client and decrements counts for its models
 // Parameters:
 //   - clientID: Unique identifier for the client to remove
@@ -460,6 +558,7 @@ func (r *ModelRegistry) unregisterClientInternal(clientID string) {
 	log.Debugf("Unregistered client %s", clientID)
 	// Separator line after completing client unregistration (after the summary line)
 	misc.LogCredentialSeparator()
+	r.triggerModelsUnregistered(provider, clientID)
 }
 
 // SetModelQuotaExceeded marks a model as quota exceeded for a specific client
@@ -625,6 +724,131 @@ func (r *ModelRegistry) GetAvailableModels(handlerType string) []map[string]any
 	return models
 }
 
+// GetAvailableModelsByProvider returns models available for the given provider identifier.
+// Parameters:
+//   - provider: Provider identifier (e.g., "codex", "gemini", "antigravity")
+//
+// Returns:
+//   - []*ModelInfo: List of available models for the provider
+func (r *ModelRegistry) GetAvailableModelsByProvider(provider string) []*ModelInfo {
+	provider = strings.ToLower(strings.TrimSpace(provider))
+	if provider == "" {
+		return nil
+	}
+
+	r.mutex.RLock()
+	defer r.mutex.RUnlock()
+
+	type providerModel struct {
+		count int
+		info  *ModelInfo
+	}
+
+	providerModels := make(map[string]*providerModel)
+
+	for clientID, clientProvider := range r.clientProviders {
+		if clientProvider != provider {
+			continue
+		}
+		modelIDs := r.clientModels[clientID]
+		if len(modelIDs) == 0 {
+			continue
+		}
+		clientInfos := r.clientModelInfos[clientID]
+		for _, modelID := range modelIDs {
+			modelID = strings.TrimSpace(modelID)
+			if modelID == "" {
+				continue
+			}
+			entry := providerModels[modelID]
+			if entry == nil {
+				entry = &providerModel{}
+				providerModels[modelID] = entry
+			}
+			entry.count++
+			if entry.info == nil {
+				if clientInfos != nil {
+					if info := clientInfos[modelID]; info != nil {
+						entry.info = info
+					}
+				}
+				if entry.info == nil {
+					if reg, ok := r.models[modelID]; ok && reg != nil && reg.Info != nil {
+						entry.info = reg.Info
+					}
+				}
+			}
+		}
+	}
+
+	if len(providerModels) == 0 {
+		return nil
+	}
+
+	quotaExpiredDuration := 5 * time.Minute
+	now := time.Now()
+	result := make([]*ModelInfo, 0, len(providerModels))
+
+	for modelID, entry := range providerModels {
+		if entry == nil || entry.count <= 0 {
+			continue
+		}
+		registration, ok := r.models[modelID]
+
+		expiredClients := 0
+		cooldownSuspended := 0
+		otherSuspended := 0
+		if ok && registration != nil {
+			if registration.QuotaExceededClients != nil {
+				for clientID, quotaTime := range registration.QuotaExceededClients {
+					if clientID == "" {
+						continue
+					}
+					if p, okProvider := r.clientProviders[clientID]; !okProvider || p != provider {
+						continue
+					}
+					if quotaTime != nil && now.Sub(*quotaTime) < quotaExpiredDuration {
+						expiredClients++
+					}
+				}
+			}
+			if registration.SuspendedClients != nil {
+				for clientID, reason := range registration.SuspendedClients {
+					if clientID == "" {
+						continue
+					}
+					if p, okProvider := r.clientProviders[clientID]; !okProvider || p != provider {
+						continue
+					}
+					if strings.EqualFold(reason, "quota") {
+						cooldownSuspended++
+						continue
+					}
+					otherSuspended++
+				}
+			}
+		}
+
+		availableClients := entry.count
+		effectiveClients := availableClients - expiredClients - otherSuspended
+		if effectiveClients < 0 {
+			effectiveClients = 0
+		}
+
+		if effectiveClients > 0 || (availableClients > 0 && (expiredClients > 0 || cooldownSuspended > 0) && otherSuspended == 0) {
+			if entry.info != nil {
+				result = append(result, entry.info)
+				continue
+			}
+			if ok && registration != nil && registration.Info != nil {
+				result = append(result, registration.Info)
+			}
+		}
+	}
+
+	return result
+}
+
 // GetModelCount returns the number of available clients for a specific model
 // Parameters:
 //   - modelID: The model ID to check

internal/registry/model_registry_hook_test.go

@@ -0,0 +1,204 @@
+package registry
+
+import (
+	"context"
+	"sync"
+	"testing"
+	"time"
+)
+
+func newTestModelRegistry() *ModelRegistry {
+	return &ModelRegistry{
+		models:           make(map[string]*ModelRegistration),
+		clientModels:     make(map[string][]string),
+		clientModelInfos: make(map[string]map[string]*ModelInfo),
+		clientProviders:  make(map[string]string),
+		mutex:            &sync.RWMutex{},
+	}
+}
+
+type registeredCall struct {
+	provider string
+	clientID string
+	models   []*ModelInfo
+}
+
+type unregisteredCall struct {
+	provider string
+	clientID string
+}
+
+type capturingHook struct {
+	registeredCh   chan registeredCall
+	unregisteredCh chan unregisteredCall
+}
+
+func (h *capturingHook) OnModelsRegistered(ctx context.Context, provider, clientID string, models []*ModelInfo) {
+	h.registeredCh <- registeredCall{provider: provider, clientID: clientID, models: models}
+}
+
+func (h *capturingHook) OnModelsUnregistered(ctx context.Context, provider, clientID string) {
+	h.unregisteredCh <- unregisteredCall{provider: provider, clientID: clientID}
+}
+
+func TestModelRegistryHook_OnModelsRegisteredCalled(t *testing.T) {
+	r := newTestModelRegistry()
+	hook := &capturingHook{
+		registeredCh:   make(chan registeredCall, 1),
+		unregisteredCh: make(chan unregisteredCall, 1),
+	}
+	r.SetHook(hook)
+
+	inputModels := []*ModelInfo{
+		{ID: "m1", DisplayName: "Model One"},
+		{ID: "m2", DisplayName: "Model Two"},
+	}
+	r.RegisterClient("client-1", "OpenAI", inputModels)
+
+	select {
+	case call := <-hook.registeredCh:
+		if call.provider != "openai" {
+			t.Fatalf("provider mismatch: got %q, want %q", call.provider, "openai")
+		}
+		if call.clientID != "client-1" {
+			t.Fatalf("clientID mismatch: got %q, want %q", call.clientID, "client-1")
+		}
+		if len(call.models) != 2 {
+			t.Fatalf("models length mismatch: got %d, want %d", len(call.models), 2)
+		}
+		if call.models[0] == nil || call.models[0].ID != "m1" {
+			t.Fatalf("models[0] mismatch: got %#v, want ID=%q", call.models[0], "m1")
+		}
+		if call.models[1] == nil || call.models[1].ID != "m2" {
+			t.Fatalf("models[1] mismatch: got %#v, want ID=%q", call.models[1], "m2")
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("timeout waiting for OnModelsRegistered hook call")
+	}
+}
+
+func TestModelRegistryHook_OnModelsUnregisteredCalled(t *testing.T) {
+	r := newTestModelRegistry()
+	hook := &capturingHook{
+		registeredCh:   make(chan registeredCall, 1),
+		unregisteredCh: make(chan unregisteredCall, 1),
+	}
+	r.SetHook(hook)
+
+	r.RegisterClient("client-1", "OpenAI", []*ModelInfo{{ID: "m1"}})
+	select {
+	case <-hook.registeredCh:
+	case <-time.After(2 * time.Second):
+		t.Fatal("timeout waiting for OnModelsRegistered hook call")
+	}
+
+	r.UnregisterClient("client-1")
+
+	select {
+	case call := <-hook.unregisteredCh:
+		if call.provider != "openai" {
+			t.Fatalf("provider mismatch: got %q, want %q", call.provider, "openai")
+		}
+		if call.clientID != "client-1" {
+			t.Fatalf("clientID mismatch: got %q, want %q", call.clientID, "client-1")
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("timeout waiting for OnModelsUnregistered hook call")
+	}
+}
+
+type blockingHook struct {
+	started chan struct{}
+	unblock chan struct{}
+}
+
+func (h *blockingHook) OnModelsRegistered(ctx context.Context, provider, clientID string, models []*ModelInfo) {
+	select {
+	case <-h.started:
+	default:
+		close(h.started)
+	}
+	<-h.unblock
+}
+
+func (h *blockingHook) OnModelsUnregistered(ctx context.Context, provider, clientID string) {}
+
+func TestModelRegistryHook_DoesNotBlockRegisterClient(t *testing.T) {
+	r := newTestModelRegistry()
+	hook := &blockingHook{
+		started: make(chan struct{}),
+		unblock: make(chan struct{}),
+	}
+	r.SetHook(hook)
+	defer close(hook.unblock)
+
+	done := make(chan struct{})
+	go func() {
+		r.RegisterClient("client-1", "OpenAI", []*ModelInfo{{ID: "m1"}})
+		close(done)
+	}()
+
+	select {
+	case <-hook.started:
+	case <-time.After(2 * time.Second):
+		t.Fatal("timeout waiting for hook to start")
+	}
+
+	select {
+	case <-done:
+	case <-time.After(200 * time.Millisecond):
+		t.Fatal("RegisterClient appears to be blocked by hook")
+	}
+
+	if !r.ClientSupportsModel("client-1", "m1") {
+		t.Fatal("model registration failed; expected client to support model")
+	}
+}
+
+type panicHook struct {
+	registeredCalled   chan struct{}
+	unregisteredCalled chan struct{}
+}
+
+func (h *panicHook) OnModelsRegistered(ctx context.Context, provider, clientID string, models []*ModelInfo) {
+	if h.registeredCalled != nil {
+		h.registeredCalled <- struct{}{}
+	}
+	panic("boom")
+}
+
+func (h *panicHook) OnModelsUnregistered(ctx context.Context, provider, clientID string) {
+	if h.unregisteredCalled != nil {
+		h.unregisteredCalled <- struct{}{}
+	}
+	panic("boom")
+}
+
+func TestModelRegistryHook_PanicDoesNotAffectRegistry(t *testing.T) {
+	r := newTestModelRegistry()
+	hook := &panicHook{
+		registeredCalled:   make(chan struct{}, 1),
+		unregisteredCalled: make(chan struct{}, 1),
+	}
+	r.SetHook(hook)
+
+	r.RegisterClient("client-1", "OpenAI", []*ModelInfo{{ID: "m1"}})
+
+	select {
+	case <-hook.registeredCalled:
+	case <-time.After(2 * time.Second):
+		t.Fatal("timeout waiting for OnModelsRegistered hook call")
+	}
+
+	if !r.ClientSupportsModel("client-1", "m1") {
+		t.Fatal("model registration failed; expected client to support model")
+	}
+
+	r.UnregisterClient("client-1")
+
+	select {
+	case <-hook.unregisteredCalled:
+	case <-time.After(2 * time.Second):
+		t.Fatal("timeout waiting for OnModelsUnregistered hook call")
+	}
+}

internal/runtime/executor/aistudio_executor.go

@@ -8,12 +8,13 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"io"
 	"net/http"
 	"net/url"
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/wsrelay"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -50,16 +51,76 @@ func (e *AIStudioExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth)
 	return nil
 }
 
+// HttpRequest forwards an arbitrary HTTP request through the websocket relay.
+func (e *AIStudioExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("aistudio executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	if e.relay == nil {
+		return nil, fmt.Errorf("aistudio executor: ws relay is nil")
+	}
+	if auth == nil || auth.ID == "" {
+		return nil, fmt.Errorf("aistudio executor: missing auth")
+	}
+	httpReq := req.WithContext(ctx)
+	if httpReq.URL == nil || strings.TrimSpace(httpReq.URL.String()) == "" {
+		return nil, fmt.Errorf("aistudio executor: request URL is empty")
+	}
+
+	var body []byte
+	if httpReq.Body != nil {
+		b, errRead := io.ReadAll(httpReq.Body)
+		if errRead != nil {
+			return nil, errRead
+		}
+		body = b
+		httpReq.Body = io.NopCloser(bytes.NewReader(b))
+	}
+
+	wsReq := &wsrelay.HTTPRequest{
+		Method:  httpReq.Method,
+		URL:     httpReq.URL.String(),
+		Headers: httpReq.Header.Clone(),
+		Body:    body,
+	}
+	wsResp, errRelay := e.relay.NonStream(ctx, auth.ID, wsReq)
+	if errRelay != nil {
+		return nil, errRelay
+	}
+	if wsResp == nil {
+		return nil, fmt.Errorf("aistudio executor: ws response is nil")
+	}
+
+	statusText := http.StatusText(wsResp.Status)
+	if statusText == "" {
+		statusText = "Unknown"
+	}
+	resp := &http.Response{
+		StatusCode:    wsResp.Status,
+		Status:        fmt.Sprintf("%d %s", wsResp.Status, statusText),
+		Header:        wsResp.Headers.Clone(),
+		Body:          io.NopCloser(bytes.NewReader(wsResp.Body)),
+		ContentLength: int64(len(wsResp.Body)),
+		Request:       httpReq,
+	}
+	return resp, nil
+}
+
 // Execute performs a non-streaming request to the AI Studio API.
 func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	translatedReq, body, err := e.translateRequest(req, opts, false)
 	if err != nil {
 		return resp, err
 	}
-	endpoint := e.buildEndpoint(req.Model, body.action, opts.Alt)
+
+	endpoint := e.buildEndpoint(baseModel, body.action, opts.Alt)
 	wsReq := &wsrelay.HTTPRequest{
 		Method:  http.MethodPost,
 		URL:     endpoint,
@@ -106,14 +167,16 @@ func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth,
 
 // ExecuteStream performs a streaming request to the AI Studio API.
 func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	translatedReq, body, err := e.translateRequest(req, opts, true)
 	if err != nil {
 		return nil, err
 	}
-	endpoint := e.buildEndpoint(req.Model, body.action, opts.Alt)
+
+	endpoint := e.buildEndpoint(baseModel, body.action, opts.Alt)
 	wsReq := &wsrelay.HTTPRequest{
 		Method:  http.MethodPost,
 		URL:     endpoint,
@@ -254,6 +317,7 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 
 // CountTokens counts tokens for the given request using the AI Studio API.
 func (e *AIStudioExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 	_, body, err := e.translateRequest(req, opts, false)
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
@@ -263,7 +327,7 @@ func (e *AIStudioExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.A
 	body.payload, _ = sjson.DeleteBytes(body.payload, "tools")
 	body.payload, _ = sjson.DeleteBytes(body.payload, "safetySettings")
 
-	endpoint := e.buildEndpoint(req.Model, "countTokens", "")
+	endpoint := e.buildEndpoint(baseModel, "countTokens", "")
 	wsReq := &wsrelay.HTTPRequest{
 		Method:  http.MethodPost,
 		URL:     endpoint,
@@ -319,17 +383,22 @@ type translatedPayload struct {
 }
 
 func (e *AIStudioExecutor) translateRequest(req cliproxyexecutor.Request, opts cliproxyexecutor.Options, stream bool) ([]byte, translatedPayload, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	payload := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
-	payload = ApplyThinkingMetadata(payload, req.Metadata, req.Model)
-	payload = util.ApplyGemini3ThinkingLevelFromMetadata(req.Model, req.Metadata, payload)
-	payload = util.ApplyDefaultThinkingIfNeeded(req.Model, payload)
-	payload = util.ConvertThinkingLevelToBudget(payload, req.Model, true)
-	payload = util.NormalizeGeminiThinkingBudget(req.Model, payload, true)
-	payload = util.StripThinkingConfigIfUnsupported(req.Model, payload)
-	payload = fixGeminiImageAspectRatio(req.Model, payload)
-	payload = applyPayloadConfig(e.cfg, req.Model, payload)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
+	}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, stream)
+	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
+	payload, err := thinking.ApplyThinking(payload, req.Model, "gemini")
+	if err != nil {
+		return nil, translatedPayload{}, err
+	}
+	payload = fixGeminiImageAspectRatio(baseModel, payload)
+	payload = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", payload, originalTranslated)
 	payload, _ = sjson.DeleteBytes(payload, "generationConfig.maxOutputTokens")
 	payload, _ = sjson.DeleteBytes(payload, "generationConfig.responseMimeType")
 	payload, _ = sjson.DeleteBytes(payload, "generationConfig.responseJsonSchema")

internal/runtime/executor/antigravity_executor.go

@@ -7,7 +7,10 @@ import (
 	"bufio"
 	"bytes"
 	"context"
+	"crypto/sha256"
+	"encoding/binary"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"math/rand"
@@ -15,12 +18,15 @@ import (
 	"net/url"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/google/uuid"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	sdkAuth "github.com/router-for-me/CLIProxyAPI/v6/sdk/auth"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
@@ -30,21 +36,25 @@ import (
 )
 
 const (
-	antigravityBaseURLDaily = "https://daily-cloudcode-pa.sandbox.googleapis.com"
-	// antigravityBaseURLAutopush     = "https://autopush-cloudcode-pa.sandbox.googleapis.com"
-	antigravityBaseURLProd     = "https://cloudcode-pa.googleapis.com"
-	antigravityCountTokensPath = "/v1internal:countTokens"
-	antigravityStreamPath      = "/v1internal:streamGenerateContent"
-	antigravityGeneratePath    = "/v1internal:generateContent"
-	antigravityModelsPath      = "/v1internal:fetchAvailableModels"
-	antigravityClientID        = "1071006060591-tmhssin2h21lcre235vtolojh4g403ep.apps.googleusercontent.com"
-	antigravityClientSecret    = "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"
-	defaultAntigravityAgent    = "antigravity/1.11.5 windows/amd64"
-	antigravityAuthType        = "antigravity"
-	refreshSkew                = 3000 * time.Second
+	antigravityBaseURLDaily        = "https://daily-cloudcode-pa.googleapis.com"
+	antigravitySandboxBaseURLDaily = "https://daily-cloudcode-pa.sandbox.googleapis.com"
+	antigravityBaseURLProd         = "https://cloudcode-pa.googleapis.com"
+	antigravityCountTokensPath     = "/v1internal:countTokens"
+	antigravityStreamPath          = "/v1internal:streamGenerateContent"
+	antigravityGeneratePath        = "/v1internal:generateContent"
+	antigravityModelsPath          = "/v1internal:fetchAvailableModels"
+	antigravityClientID            = "1071006060591-tmhssin2h21lcre235vtolojh4g403ep.apps.googleusercontent.com"
+	antigravityClientSecret        = "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"
+	defaultAntigravityAgent        = "antigravity/1.104.0 darwin/arm64"
+	antigravityAuthType            = "antigravity"
+	refreshSkew                    = 3000 * time.Second
+	systemInstruction              = "You are Antigravity, a powerful agentic AI coding assistant designed by the Google Deepmind team working on Advanced Agentic Coding.You are pair programming with a USER to solve their coding task. The task may require creating a new codebase, modifying or debugging an existing codebase, or simply answering a question.**Absolute paths only****Proactiveness**"
 )
 
-var randSource = rand.New(rand.NewSource(time.Now().UnixNano()))
+var (
+	randSource      = rand.New(rand.NewSource(time.Now().UnixNano()))
+	randSourceMutex sync.Mutex
+)
 
 // AntigravityExecutor proxies requests to the antigravity upstream.
 type AntigravityExecutor struct {
@@ -65,12 +75,44 @@ func NewAntigravityExecutor(cfg *config.Config) *AntigravityExecutor {
 // Identifier returns the executor identifier.
 func (e *AntigravityExecutor) Identifier() string { return antigravityAuthType }
 
-// PrepareRequest prepares the HTTP request for execution (no-op for Antigravity).
-func (e *AntigravityExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Antigravity credentials into the outgoing HTTP request.
+func (e *AntigravityExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	token, _, errToken := e.ensureAccessToken(req.Context(), auth)
+	if errToken != nil {
+		return errToken
+	}
+	if strings.TrimSpace(token) == "" {
+		return statusErr{code: http.StatusUnauthorized, msg: "missing access token"}
+	}
+	req.Header.Set("Authorization", "Bearer "+token)
+	return nil
+}
+
+// HttpRequest injects Antigravity credentials into the request and executes it.
+func (e *AntigravityExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("antigravity executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 // Execute performs a non-streaming request to the Antigravity API.
 func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	if strings.Contains(req.Model, "claude") {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+	isClaude := strings.Contains(strings.ToLower(baseModel), "claude")
+
+	if isClaude || strings.Contains(baseModel, "gemini-3-pro") {
 		return e.executeClaudeNonStream(ctx, auth, req, opts)
 	}
 
@@ -82,18 +124,25 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 		auth = updatedAuth
 	}
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("antigravity")
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
 
-	translated = applyThinkingMetadataCLI(translated, req.Metadata, req.Model)
-	translated = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, translated)
-	translated = util.ApplyDefaultThinkingIfNeededCLI(req.Model, translated)
-	translated = normalizeAntigravityThinking(req.Model, translated)
-	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, "antigravity", "request", translated)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
+	}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	if err != nil {
+		return resp, err
+	}
+
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated)
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
@@ -103,7 +152,7 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	var lastErr error
 
 	for idx, baseURL := range baseURLs {
-		httpReq, errReq := e.buildRequest(ctx, auth, token, req.Model, translated, false, opts.Alt, baseURL)
+		httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, translated, false, opts.Alt, baseURL)
 		if errReq != nil {
 			err = errReq
 			return resp, err
@@ -112,6 +161,9 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
 			recordAPIResponseError(ctx, e.cfg, errDo)
+			if errors.Is(errDo, context.Canceled) || errors.Is(errDo, context.DeadlineExceeded) {
+				return resp, errDo
+			}
 			lastStatus = 0
 			lastBody = nil
 			lastErr = errDo
@@ -144,7 +196,13 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 				log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 				continue
 			}
-			err = statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			if httpResp.StatusCode == http.StatusTooManyRequests {
+				if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+					sErr.retryAfter = retryAfter
+				}
+			}
+			err = sErr
 			return resp, err
 		}
 
@@ -158,7 +216,13 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 
 	switch {
 	case lastStatus != 0:
-		err = statusErr{code: lastStatus, msg: string(lastBody)}
+		sErr := statusErr{code: lastStatus, msg: string(lastBody)}
+		if lastStatus == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(lastBody); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		err = sErr
 	case lastErr != nil:
 		err = lastErr
 	default:
@@ -169,6 +233,8 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 
 // executeClaudeNonStream performs a claude non-streaming request to the Antigravity API.
 func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
 	if errToken != nil {
 		return resp, errToken
@@ -177,18 +243,25 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 		auth = updatedAuth
 	}
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("antigravity")
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 
-	translated = applyThinkingMetadataCLI(translated, req.Metadata, req.Model)
-	translated = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, translated)
-	translated = util.ApplyDefaultThinkingIfNeededCLI(req.Model, translated)
-	translated = normalizeAntigravityThinking(req.Model, translated)
-	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, "antigravity", "request", translated)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
+	}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+
+	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	if err != nil {
+		return resp, err
+	}
+
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated)
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
@@ -198,7 +271,7 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	var lastErr error
 
 	for idx, baseURL := range baseURLs {
-		httpReq, errReq := e.buildRequest(ctx, auth, token, req.Model, translated, true, opts.Alt, baseURL)
+		httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, translated, true, opts.Alt, baseURL)
 		if errReq != nil {
 			err = errReq
 			return resp, err
@@ -207,6 +280,9 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
 			recordAPIResponseError(ctx, e.cfg, errDo)
+			if errors.Is(errDo, context.Canceled) || errors.Is(errDo, context.DeadlineExceeded) {
+				return resp, errDo
+			}
 			lastStatus = 0
 			lastBody = nil
 			lastErr = errDo
@@ -225,6 +301,14 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 			}
 			if errRead != nil {
 				recordAPIResponseError(ctx, e.cfg, errRead)
+				if errors.Is(errRead, context.Canceled) || errors.Is(errRead, context.DeadlineExceeded) {
+					err = errRead
+					return resp, err
+				}
+				if errCtx := ctx.Err(); errCtx != nil {
+					err = errCtx
+					return resp, err
+				}
 				lastStatus = 0
 				lastBody = nil
 				lastErr = errRead
@@ -243,7 +327,13 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 				log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 				continue
 			}
-			err = statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			if httpResp.StatusCode == http.StatusTooManyRequests {
+				if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+					sErr.retryAfter = retryAfter
+				}
+			}
+			err = sErr
 			return resp, err
 		}
 
@@ -308,7 +398,13 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 
 	switch {
 	case lastStatus != 0:
-		err = statusErr{code: lastStatus, msg: string(lastBody)}
+		sErr := statusErr{code: lastStatus, msg: string(lastBody)}
+		if lastStatus == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(lastBody); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		err = sErr
 	case lastErr != nil:
 		err = lastErr
 	default:
@@ -501,6 +597,8 @@ func (e *AntigravityExecutor) convertStreamToNonStream(stream []byte) []byte {
 
 // ExecuteStream performs a streaming request to the Antigravity API.
 func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	ctx = context.WithValue(ctx, "alt", "")
 
 	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
@@ -511,18 +609,25 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 		auth = updatedAuth
 	}
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("antigravity")
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 
-	translated = applyThinkingMetadataCLI(translated, req.Metadata, req.Model)
-	translated = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, translated)
-	translated = util.ApplyDefaultThinkingIfNeededCLI(req.Model, translated)
-	translated = normalizeAntigravityThinking(req.Model, translated)
-	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, "antigravity", "request", translated)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
+	}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+
+	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	if err != nil {
+		return nil, err
+	}
+
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated)
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
@@ -532,7 +637,7 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	var lastErr error
 
 	for idx, baseURL := range baseURLs {
-		httpReq, errReq := e.buildRequest(ctx, auth, token, req.Model, translated, true, opts.Alt, baseURL)
+		httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, translated, true, opts.Alt, baseURL)
 		if errReq != nil {
 			err = errReq
 			return nil, err
@@ -541,6 +646,9 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
 			recordAPIResponseError(ctx, e.cfg, errDo)
+			if errors.Is(errDo, context.Canceled) || errors.Is(errDo, context.DeadlineExceeded) {
+				return nil, errDo
+			}
 			lastStatus = 0
 			lastBody = nil
 			lastErr = errDo
@@ -559,6 +667,14 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 			}
 			if errRead != nil {
 				recordAPIResponseError(ctx, e.cfg, errRead)
+				if errors.Is(errRead, context.Canceled) || errors.Is(errRead, context.DeadlineExceeded) {
+					err = errRead
+					return nil, err
+				}
+				if errCtx := ctx.Err(); errCtx != nil {
+					err = errCtx
+					return nil, err
+				}
 				lastStatus = 0
 				lastBody = nil
 				lastErr = errRead
@@ -577,7 +693,13 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 				log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 				continue
 			}
-			err = statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			if httpResp.StatusCode == http.StatusTooManyRequests {
+				if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+					sErr.retryAfter = retryAfter
+				}
+			}
+			err = sErr
 			return nil, err
 		}
 
@@ -632,7 +754,13 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 
 	switch {
 	case lastStatus != 0:
-		err = statusErr{code: lastStatus, msg: string(lastBody)}
+		sErr := statusErr{code: lastStatus, msg: string(lastBody)}
+		if lastStatus == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(lastBody); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		err = sErr
 	case lastErr != nil:
 		err = lastErr
 	default:
@@ -655,6 +783,8 @@ func (e *AntigravityExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Au
 
 // CountTokens counts tokens for the given request using the Antigravity API.
 func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
 	if errToken != nil {
 		return cliproxyexecutor.Response{}, errToken
@@ -670,6 +800,18 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 	to := sdktranslator.FromString("antigravity")
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
 
+	// Prepare payload once (doesn't depend on baseURL)
+	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	payload, err := thinking.ApplyThinking(payload, req.Model, "antigravity")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
+
+	payload = deleteJSONField(payload, "project")
+	payload = deleteJSONField(payload, "model")
+	payload = deleteJSONField(payload, "request.safetySettings")
+
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
 
@@ -685,14 +827,6 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 	var lastErr error
 
 	for idx, baseURL := range baseURLs {
-		payload := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-		payload = applyThinkingMetadataCLI(payload, req.Metadata, req.Model)
-		payload = util.ApplyDefaultThinkingIfNeededCLI(req.Model, payload)
-		payload = normalizeAntigravityThinking(req.Model, payload)
-		payload = deleteJSONField(payload, "project")
-		payload = deleteJSONField(payload, "model")
-		payload = deleteJSONField(payload, "request.safetySettings")
-
 		base := strings.TrimSuffix(baseURL, "/")
 		if base == "" {
 			base = buildBaseURL(auth)
@@ -733,6 +867,9 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
 			recordAPIResponseError(ctx, e.cfg, errDo)
+			if errors.Is(errDo, context.Canceled) || errors.Is(errDo, context.DeadlineExceeded) {
+				return cliproxyexecutor.Response{}, errDo
+			}
 			lastStatus = 0
 			lastBody = nil
 			lastErr = errDo
@@ -767,12 +904,24 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 			log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 			continue
 		}
-		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+		sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+		if httpResp.StatusCode == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		return cliproxyexecutor.Response{}, sErr
 	}
 
 	switch {
 	case lastStatus != 0:
-		return cliproxyexecutor.Response{}, statusErr{code: lastStatus, msg: string(lastBody)}
+		sErr := statusErr{code: lastStatus, msg: string(lastBody)}
+		if lastStatus == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(lastBody); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		return cliproxyexecutor.Response{}, sErr
 	case lastErr != nil:
 		return cliproxyexecutor.Response{}, lastErr
 	default:
@@ -809,6 +958,9 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
+			if errors.Is(errDo, context.Canceled) || errors.Is(errDo, context.DeadlineExceeded) {
+				return nil
+			}
 			if idx+1 < len(baseURLs) {
 				log.Debugf("antigravity executor: models request error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 				continue
@@ -844,35 +996,40 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 		modelConfig := registry.GetAntigravityModelConfig()
 		models := make([]*registry.ModelInfo, 0, len(result.Map()))
 		for originalName := range result.Map() {
-			aliasName := modelName2Alias(originalName)
-			if aliasName != "" {
-				cfg := modelConfig[aliasName]
-				modelName := aliasName
-				if cfg != nil && cfg.Name != "" {
-					modelName = cfg.Name
-				}
-				modelInfo := &registry.ModelInfo{
-					ID:          aliasName,
-					Name:        modelName,
-					Description: aliasName,
-					DisplayName: aliasName,
-					Version:     aliasName,
-					Object:      "model",
-					Created:     now,
-					OwnedBy:     antigravityAuthType,
-					Type:        antigravityAuthType,
-				}
-				// Look up Thinking support from static config using alias name
-				if cfg != nil {
-					if cfg.Thinking != nil {
-						modelInfo.Thinking = cfg.Thinking
-					}
-					if cfg.MaxCompletionTokens > 0 {
-						modelInfo.MaxCompletionTokens = cfg.MaxCompletionTokens
-					}
-				}
-				models = append(models, modelInfo)
+			modelID := strings.TrimSpace(originalName)
+			if modelID == "" {
+				continue
 			}
+			switch modelID {
+			case "chat_20706", "chat_23310", "gemini-2.5-flash-thinking", "gemini-3-pro-low", "gemini-2.5-pro":
+				continue
+			}
+			cfg := modelConfig[modelID]
+			modelName := modelID
+			if cfg != nil && cfg.Name != "" {
+				modelName = cfg.Name
+			}
+			modelInfo := &registry.ModelInfo{
+				ID:          modelID,
+				Name:        modelName,
+				Description: modelID,
+				DisplayName: modelID,
+				Version:     modelID,
+				Object:      "model",
+				Created:     now,
+				OwnedBy:     antigravityAuthType,
+				Type:        antigravityAuthType,
+			}
+			// Look up Thinking support from static config using upstream model name.
+			if cfg != nil {
+				if cfg.Thinking != nil {
+					modelInfo.Thinking = cfg.Thinking
+				}
+				if cfg.MaxCompletionTokens > 0 {
+					modelInfo.MaxCompletionTokens = cfg.MaxCompletionTokens
+				}
+			}
+			models = append(models, modelInfo)
 		}
 		return models
 	}
@@ -888,7 +1045,13 @@ func (e *AntigravityExecutor) ensureAccessToken(ctx context.Context, auth *clipr
 	if accessToken != "" && expiry.After(time.Now().Add(refreshSkew)) {
 		return accessToken, nil, nil
 	}
-	updated, errRefresh := e.refreshToken(ctx, auth.Clone())
+	refreshCtx := context.Background()
+	if ctx != nil {
+		if rt, ok := ctx.Value("cliproxy.roundtripper").(http.RoundTripper); ok && rt != nil {
+			refreshCtx = context.WithValue(refreshCtx, "cliproxy.roundtripper", rt)
+		}
+	}
+	updated, errRefresh := e.refreshToken(refreshCtx, auth.Clone())
 	if errRefresh != nil {
 		return "", nil, errRefresh
 	}
@@ -935,7 +1098,13 @@ func (e *AntigravityExecutor) refreshToken(ctx context.Context, auth *cliproxyau
 	}
 
 	if httpResp.StatusCode < http.StatusOK || httpResp.StatusCode >= http.StatusMultipleChoices {
-		return auth, statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+		sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+		if httpResp.StatusCode == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		return auth, sErr
 	}
 
 	var tokenResp struct {
@@ -956,12 +1125,49 @@ func (e *AntigravityExecutor) refreshToken(ctx context.Context, auth *cliproxyau
 		auth.Metadata["refresh_token"] = tokenResp.RefreshToken
 	}
 	auth.Metadata["expires_in"] = tokenResp.ExpiresIn
-	auth.Metadata["timestamp"] = time.Now().UnixMilli()
-	auth.Metadata["expired"] = time.Now().Add(time.Duration(tokenResp.ExpiresIn) * time.Second).Format(time.RFC3339)
+	now := time.Now()
+	auth.Metadata["timestamp"] = now.UnixMilli()
+	auth.Metadata["expired"] = now.Add(time.Duration(tokenResp.ExpiresIn) * time.Second).Format(time.RFC3339)
 	auth.Metadata["type"] = antigravityAuthType
+	if errProject := e.ensureAntigravityProjectID(ctx, auth, tokenResp.AccessToken); errProject != nil {
+		log.Warnf("antigravity executor: ensure project id failed: %v", errProject)
+	}
 	return auth, nil
 }
 
+func (e *AntigravityExecutor) ensureAntigravityProjectID(ctx context.Context, auth *cliproxyauth.Auth, accessToken string) error {
+	if auth == nil {
+		return nil
+	}
+
+	if auth.Metadata["project_id"] != nil {
+		return nil
+	}
+
+	token := strings.TrimSpace(accessToken)
+	if token == "" {
+		token = metaStringValue(auth.Metadata, "access_token")
+	}
+	if token == "" {
+		return nil
+	}
+
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	projectID, errFetch := sdkAuth.FetchAntigravityProjectID(ctx, token, httpClient)
+	if errFetch != nil {
+		return errFetch
+	}
+	if strings.TrimSpace(projectID) == "" {
+		return nil
+	}
+	if auth.Metadata == nil {
+		auth.Metadata = make(map[string]any)
+	}
+	auth.Metadata["project_id"] = strings.TrimSpace(projectID)
+
+	return nil
+}
+
 func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyauth.Auth, token, modelName string, payload []byte, stream bool, alt, baseURL string) (*http.Request, error) {
 	if token == "" {
 		return nil, statusErr{code: http.StatusUnauthorized, msg: "missing access token"}
@@ -998,7 +1204,7 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 		}
 	}
 	payload = geminiToAntigravity(modelName, payload, projectID)
-	payload, _ = sjson.SetBytes(payload, "model", alias2ModelName(modelName))
+	payload, _ = sjson.SetBytes(payload, "model", modelName)
 
 	if strings.Contains(modelName, "claude") {
 		strJSON := string(payload)
@@ -1010,11 +1216,24 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 
 		// Use the centralized schema cleaner to handle unsupported keywords,
 		// const->enum conversion, and flattening of types/anyOf.
-		strJSON = util.CleanJSONSchemaForGemini(strJSON)
+		strJSON = util.CleanJSONSchemaForAntigravity(strJSON)
 
 		payload = []byte(strJSON)
 	}
 
+	if strings.Contains(modelName, "claude") || strings.Contains(modelName, "gemini-3-pro-preview") {
+		systemInstructionPartsResult := gjson.GetBytes(payload, "request.systemInstruction.parts")
+		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.role", "user")
+		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.parts.0.text", systemInstruction)
+		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.parts.1.text", fmt.Sprintf("Please ignore following [ignore]%s[/ignore]", systemInstruction))
+
+		if systemInstructionPartsResult.Exists() && systemInstructionPartsResult.IsArray() {
+			for _, partResult := range systemInstructionPartsResult.Array() {
+				payload, _ = sjson.SetRawBytes(payload, "request.systemInstruction.parts.-1", []byte(partResult.Raw))
+			}
+		}
+	}
+
 	httpReq, errReq := http.NewRequestWithContext(ctx, http.MethodPost, requestURL.String(), bytes.NewReader(payload))
 	if errReq != nil {
 		return nil, errReq
@@ -1149,8 +1368,8 @@ func antigravityBaseURLFallbackOrder(auth *cliproxyauth.Auth) []string {
 		return []string{base}
 	}
 	return []string{
+		antigravitySandboxBaseURLDaily,
 		antigravityBaseURLDaily,
-		// antigravityBaseURLAutopush,
 		antigravityBaseURLProd,
 	}
 }
@@ -1178,6 +1397,7 @@ func resolveCustomAntigravityBaseURL(auth *cliproxyauth.Auth) string {
 func geminiToAntigravity(modelName string, payload []byte, projectID string) []byte {
 	template, _ := sjson.Set(string(payload), "model", modelName)
 	template, _ = sjson.Set(template, "userAgent", "antigravity")
+	template, _ = sjson.Set(template, "requestType", "agent")
 
 	// Use real project ID from auth if available, otherwise generate random (legacy fallback)
 	if projectID != "" {
@@ -1186,7 +1406,7 @@ func geminiToAntigravity(modelName string, payload []byte, projectID string) []b
 		template, _ = sjson.Set(template, "project", generateProjectID())
 	}
 	template, _ = sjson.Set(template, "requestId", generateRequestID())
-	template, _ = sjson.Set(template, "request.sessionId", generateSessionID())
+	template, _ = sjson.Set(template, "request.sessionId", generateStableSessionID(payload))
 
 	template, _ = sjson.Delete(template, "request.safetySettings")
 	template, _ = sjson.Set(template, "request.toolConfig.functionCallingConfig.mode", "VALIDATED")
@@ -1222,121 +1442,36 @@ func generateRequestID() string {
 }
 
 func generateSessionID() string {
+	randSourceMutex.Lock()
 	n := randSource.Int63n(9_000_000_000_000_000_000)
+	randSourceMutex.Unlock()
 	return "-" + strconv.FormatInt(n, 10)
 }
 
+func generateStableSessionID(payload []byte) string {
+	contents := gjson.GetBytes(payload, "request.contents")
+	if contents.IsArray() {
+		for _, content := range contents.Array() {
+			if content.Get("role").String() == "user" {
+				text := content.Get("parts.0.text").String()
+				if text != "" {
+					h := sha256.Sum256([]byte(text))
+					n := int64(binary.BigEndian.Uint64(h[:8])) & 0x7FFFFFFFFFFFFFFF
+					return "-" + strconv.FormatInt(n, 10)
+				}
+			}
+		}
+	}
+	return generateSessionID()
+}
+
 func generateProjectID() string {
 	adjectives := []string{"useful", "bright", "swift", "calm", "bold"}
 	nouns := []string{"fuze", "wave", "spark", "flow", "core"}
+	randSourceMutex.Lock()
 	adj := adjectives[randSource.Intn(len(adjectives))]
 	noun := nouns[randSource.Intn(len(nouns))]
+	randSourceMutex.Unlock()
 	randomPart := strings.ToLower(uuid.NewString())[:5]
 	return adj + "-" + noun + "-" + randomPart
 }
-
-func modelName2Alias(modelName string) string {
-	switch modelName {
-	case "rev19-uic3-1p":
-		return "gemini-2.5-computer-use-preview-10-2025"
-	case "gemini-3-pro-image":
-		return "gemini-3-pro-image-preview"
-	case "gemini-3-pro-high":
-		return "gemini-3-pro-preview"
-	case "gemini-3-flash":
-		return "gemini-3-flash-preview"
-	case "claude-sonnet-4-5":
-		return "gemini-claude-sonnet-4-5"
-	case "claude-sonnet-4-5-thinking":
-		return "gemini-claude-sonnet-4-5-thinking"
-	case "claude-opus-4-5-thinking":
-		return "gemini-claude-opus-4-5-thinking"
-	case "chat_20706", "chat_23310", "gemini-2.5-flash-thinking", "gemini-3-pro-low", "gemini-2.5-pro":
-		return ""
-	default:
-		return modelName
-	}
-}
-
-func alias2ModelName(modelName string) string {
-	switch modelName {
-	case "gemini-2.5-computer-use-preview-10-2025":
-		return "rev19-uic3-1p"
-	case "gemini-3-pro-image-preview":
-		return "gemini-3-pro-image"
-	case "gemini-3-pro-preview":
-		return "gemini-3-pro-high"
-	case "gemini-3-flash-preview":
-		return "gemini-3-flash"
-	case "gemini-claude-sonnet-4-5":
-		return "claude-sonnet-4-5"
-	case "gemini-claude-sonnet-4-5-thinking":
-		return "claude-sonnet-4-5-thinking"
-	case "gemini-claude-opus-4-5-thinking":
-		return "claude-opus-4-5-thinking"
-	default:
-		return modelName
-	}
-}
-
-// normalizeAntigravityThinking clamps or removes thinking config based on model support.
-// For Claude models, it additionally ensures thinking budget < max_tokens.
-func normalizeAntigravityThinking(model string, payload []byte) []byte {
-	payload = util.StripThinkingConfigIfUnsupported(model, payload)
-	if !util.ModelSupportsThinking(model) {
-		return payload
-	}
-	budget := gjson.GetBytes(payload, "request.generationConfig.thinkingConfig.thinkingBudget")
-	if !budget.Exists() {
-		return payload
-	}
-	raw := int(budget.Int())
-	normalized := util.NormalizeThinkingBudget(model, raw)
-
-	isClaude := strings.Contains(strings.ToLower(model), "claude")
-	if isClaude {
-		effectiveMax, setDefaultMax := antigravityEffectiveMaxTokens(model, payload)
-		if effectiveMax > 0 && normalized >= effectiveMax {
-			normalized = effectiveMax - 1
-		}
-		minBudget := antigravityMinThinkingBudget(model)
-		if minBudget > 0 && normalized >= 0 && normalized < minBudget {
-			// Budget is below minimum, remove thinking config entirely
-			payload, _ = sjson.DeleteBytes(payload, "request.generationConfig.thinkingConfig")
-			return payload
-		}
-		if setDefaultMax {
-			if res, errSet := sjson.SetBytes(payload, "request.generationConfig.maxOutputTokens", effectiveMax); errSet == nil {
-				payload = res
-			}
-		}
-	}
-
-	updated, err := sjson.SetBytes(payload, "request.generationConfig.thinkingConfig.thinkingBudget", normalized)
-	if err != nil {
-		return payload
-	}
-	return updated
-}
-
-// antigravityEffectiveMaxTokens returns the max tokens to cap thinking:
-// prefer request-provided maxOutputTokens; otherwise fall back to model default.
-// The boolean indicates whether the value came from the model default (and thus should be written back).
-func antigravityEffectiveMaxTokens(model string, payload []byte) (max int, fromModel bool) {
-	if maxTok := gjson.GetBytes(payload, "request.generationConfig.maxOutputTokens"); maxTok.Exists() && maxTok.Int() > 0 {
-		return int(maxTok.Int()), false
-	}
-	if modelInfo := registry.GetGlobalRegistry().GetModelInfo(model); modelInfo != nil && modelInfo.MaxCompletionTokens > 0 {
-		return modelInfo.MaxCompletionTokens, true
-	}
-	return 0, false
-}
-
-// antigravityMinThinkingBudget returns the minimum thinking budget for a model.
-// Falls back to -1 if no model info is found.
-func antigravityMinThinkingBudget(model string) int {
-	if modelInfo := registry.GetGlobalRegistry().GetModelInfo(model); modelInfo != nil && modelInfo.Thinking != nil {
-		return modelInfo.Thinking.Min
-	}
-	return -1
-}

internal/runtime/executor/cache_helpers.go

@@ -1,10 +1,68 @@
 package executor
 
-import "time"
+import (
+	"sync"
+	"time"
+)
 
 type codexCache struct {
 	ID     string
 	Expire time.Time
 }
 
-var codexCacheMap = map[string]codexCache{}
+// codexCacheMap stores prompt cache IDs keyed by model+user_id.
+// Protected by codexCacheMu. Entries expire after 1 hour.
+var (
+	codexCacheMap = make(map[string]codexCache)
+	codexCacheMu  sync.RWMutex
+)
+
+// codexCacheCleanupInterval controls how often expired entries are purged.
+const codexCacheCleanupInterval = 15 * time.Minute
+
+// codexCacheCleanupOnce ensures the background cleanup goroutine starts only once.
+var codexCacheCleanupOnce sync.Once
+
+// startCodexCacheCleanup launches a background goroutine that periodically
+// removes expired entries from codexCacheMap to prevent memory leaks.
+func startCodexCacheCleanup() {
+	go func() {
+		ticker := time.NewTicker(codexCacheCleanupInterval)
+		defer ticker.Stop()
+		for range ticker.C {
+			purgeExpiredCodexCache()
+		}
+	}()
+}
+
+// purgeExpiredCodexCache removes entries that have expired.
+func purgeExpiredCodexCache() {
+	now := time.Now()
+	codexCacheMu.Lock()
+	defer codexCacheMu.Unlock()
+	for key, cache := range codexCacheMap {
+		if cache.Expire.Before(now) {
+			delete(codexCacheMap, key)
+		}
+	}
+}
+
+// getCodexCache retrieves a cached entry, returning ok=false if not found or expired.
+func getCodexCache(key string) (codexCache, bool) {
+	codexCacheCleanupOnce.Do(startCodexCacheCleanup)
+	codexCacheMu.RLock()
+	cache, ok := codexCacheMap[key]
+	codexCacheMu.RUnlock()
+	if !ok || cache.Expire.Before(time.Now()) {
+		return codexCache{}, false
+	}
+	return cache, true
+}
+
+// setCodexCache stores a cache entry.
+func setCodexCache(key string, cache codexCache) {
+	codexCacheCleanupOnce.Do(startCodexCacheCleanup)
+	codexCacheMu.Lock()
+	codexCacheMap[key] = cache
+	codexCacheMu.Unlock()
+}

internal/runtime/executor/claude_executor.go

@@ -18,6 +18,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -35,54 +36,103 @@ type ClaudeExecutor struct {
 	cfg *config.Config
 }
 
+const claudeToolPrefix = "proxy_"
+
 func NewClaudeExecutor(cfg *config.Config) *ClaudeExecutor { return &ClaudeExecutor{cfg: cfg} }
 
 func (e *ClaudeExecutor) Identifier() string { return "claude" }
 
-func (e *ClaudeExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Claude credentials into the outgoing HTTP request.
+func (e *ClaudeExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	apiKey, _ := claudeCreds(auth)
+	if strings.TrimSpace(apiKey) == "" {
+		return nil
+	}
+	useAPIKey := auth != nil && auth.Attributes != nil && strings.TrimSpace(auth.Attributes["api_key"]) != ""
+	isAnthropicBase := req.URL != nil && strings.EqualFold(req.URL.Scheme, "https") && strings.EqualFold(req.URL.Host, "api.anthropic.com")
+	if isAnthropicBase && useAPIKey {
+		req.Header.Del("Authorization")
+		req.Header.Set("x-api-key", apiKey)
+	} else {
+		req.Header.Del("x-api-key")
+		req.Header.Set("Authorization", "Bearer "+apiKey)
+	}
+	var attrs map[string]string
+	if auth != nil {
+		attrs = auth.Attributes
+	}
+	util.ApplyCustomHeadersFromAttrs(req, attrs)
+	return nil
+}
+
+// HttpRequest injects Claude credentials into the request and executes it.
+func (e *ClaudeExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("claude executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	apiKey, baseURL := claudeCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	apiKey, baseURL := claudeCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://api.anthropic.com"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("claude")
 	// Use streaming translation to preserve function calling, except for claude.
 	stream := from != to
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-	if upstreamModel == "" {
-		upstreamModel = req.Model
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
-		upstreamModel = modelOverride
-	} else if !strings.EqualFold(upstreamModel, req.Model) {
-		if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
-			upstreamModel = modelOverride
-		}
-	}
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
-	// Inject thinking config based on model metadata for thinking variants
-	body = e.injectThinkingConfig(req.Model, req.Metadata, body)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, stream)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	if !strings.HasPrefix(upstreamModel, "claude-3-5-haiku") {
+	body, err = thinking.ApplyThinking(body, req.Model, "claude")
+	if err != nil {
+		return resp, err
+	}
+
+	if !strings.HasPrefix(baseModel, "claude-3-5-haiku") {
 		body = checkSystemInstructions(body)
 	}
-	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+
+	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
+	body = disableThinkingIfToolChoiceForced(body)
 
 	// Ensure max_tokens > thinking.budget_tokens when thinking is enabled
-	body = ensureMaxTokensForThinking(req.Model, body)
+	body = ensureMaxTokensForThinking(baseModel, body)
 
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
+	bodyForTranslation := body
+	bodyForUpstream := body
+	if isClaudeOAuthToken(apiKey) {
+		bodyForUpstream = applyClaudeToolPrefix(body, claudeToolPrefix)
+	}
 
 	url := fmt.Sprintf("%s/v1/messages?beta=true", baseURL)
-	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(bodyForUpstream))
 	if err != nil {
 		return resp, err
 	}
@@ -97,7 +147,7 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 		URL:       url,
 		Method:    http.MethodPost,
 		Headers:   httpReq.Header.Clone(),
-		Body:      body,
+		Body:      bodyForUpstream,
 		Provider:  e.Identifier(),
 		AuthID:    authID,
 		AuthLabel: authLabel,
@@ -151,49 +201,69 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	} else {
 		reporter.publish(ctx, parseClaudeUsage(data))
 	}
+	if isClaudeOAuthToken(apiKey) {
+		data = stripClaudeToolPrefixFromResponse(data, claudeToolPrefix)
+	}
 	var param any
-	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
+	out := sdktranslator.TranslateNonStream(
+		ctx,
+		to,
+		from,
+		req.Model,
+		bytes.Clone(opts.OriginalRequest),
+		bodyForTranslation,
+		data,
+		&param,
+	)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
 }
 
 func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	apiKey, baseURL := claudeCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	apiKey, baseURL := claudeCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://api.anthropic.com"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("claude")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-	if upstreamModel == "" {
-		upstreamModel = req.Model
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
-		upstreamModel = modelOverride
-	} else if !strings.EqualFold(upstreamModel, req.Model) {
-		if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
-			upstreamModel = modelOverride
-		}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
+
+	body, err = thinking.ApplyThinking(body, req.Model, "claude")
+	if err != nil {
+		return nil, err
 	}
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
-	// Inject thinking config based on model metadata for thinking variants
-	body = e.injectThinkingConfig(req.Model, req.Metadata, body)
+
 	body = checkSystemInstructions(body)
-	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+
+	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
+	body = disableThinkingIfToolChoiceForced(body)
 
 	// Ensure max_tokens > thinking.budget_tokens when thinking is enabled
-	body = ensureMaxTokensForThinking(req.Model, body)
+	body = ensureMaxTokensForThinking(baseModel, body)
 
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
+	bodyForTranslation := body
+	bodyForUpstream := body
+	if isClaudeOAuthToken(apiKey) {
+		bodyForUpstream = applyClaudeToolPrefix(body, claudeToolPrefix)
+	}
 
 	url := fmt.Sprintf("%s/v1/messages?beta=true", baseURL)
-	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(bodyForUpstream))
 	if err != nil {
 		return nil, err
 	}
@@ -208,7 +278,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 		URL:       url,
 		Method:    http.MethodPost,
 		Headers:   httpReq.Header.Clone(),
-		Body:      body,
+		Body:      bodyForUpstream,
 		Provider:  e.Identifier(),
 		AuthID:    authID,
 		AuthLabel: authLabel,
@@ -261,6 +331,9 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 				if detail, ok := parseClaudeStreamUsage(line); ok {
 					reporter.publish(ctx, detail)
 				}
+				if isClaudeOAuthToken(apiKey) {
+					line = stripClaudeToolPrefixFromStreamLine(line, claudeToolPrefix)
+				}
 				// Forward the line as-is to preserve SSE format
 				cloned := make([]byte, len(line)+1)
 				copy(cloned, line)
@@ -285,7 +358,19 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 			if detail, ok := parseClaudeStreamUsage(line); ok {
 				reporter.publish(ctx, detail)
 			}
-			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(line), &param)
+			if isClaudeOAuthToken(apiKey) {
+				line = stripClaudeToolPrefixFromStreamLine(line, claudeToolPrefix)
+			}
+			chunks := sdktranslator.TranslateStream(
+				ctx,
+				to,
+				from,
+				req.Model,
+				bytes.Clone(opts.OriginalRequest),
+				bodyForTranslation,
+				bytes.Clone(line),
+				&param,
+			)
 			for i := range chunks {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
 			}
@@ -300,8 +385,9 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 }
 
 func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
-	apiKey, baseURL := claudeCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	apiKey, baseURL := claudeCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://api.anthropic.com"
 	}
@@ -310,27 +396,19 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	to := sdktranslator.FromString("claude")
 	// Use streaming translation to preserve function calling, except for claude.
 	stream := from != to
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-	if upstreamModel == "" {
-		upstreamModel = req.Model
-	}
-	if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
-		upstreamModel = modelOverride
-	} else if !strings.EqualFold(upstreamModel, req.Model) {
-		if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
-			upstreamModel = modelOverride
-		}
-	}
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	if !strings.HasPrefix(upstreamModel, "claude-3-5-haiku") {
+	if !strings.HasPrefix(baseModel, "claude-3-5-haiku") {
 		body = checkSystemInstructions(body)
 	}
 
 	// Extract betas from body and convert to header (for count_tokens too)
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
+	if isClaudeOAuthToken(apiKey) {
+		body = applyClaudeToolPrefix(body, claudeToolPrefix)
+	}
 
 	url := fmt.Sprintf("%s/v1/messages/count_tokens?beta=true", baseURL)
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
@@ -450,15 +528,17 @@ func extractAndRemoveBetas(body []byte) ([]string, []byte) {
 	return betas, body
 }
 
-// injectThinkingConfig adds thinking configuration based on metadata using the unified flow.
-// It uses util.ResolveClaudeThinkingConfig which internally calls ResolveThinkingConfigFromMetadata
-// and NormalizeThinkingBudget, ensuring consistency with other executors like Gemini.
-func (e *ClaudeExecutor) injectThinkingConfig(modelName string, metadata map[string]any, body []byte) []byte {
-	budget, ok := util.ResolveClaudeThinkingConfig(modelName, metadata)
-	if !ok {
-		return body
+// disableThinkingIfToolChoiceForced checks if tool_choice forces tool use and disables thinking.
+// Anthropic API does not allow thinking when tool_choice is set to "any" or a specific tool.
+// See: https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking#important-considerations
+func disableThinkingIfToolChoiceForced(body []byte) []byte {
+	toolChoiceType := gjson.GetBytes(body, "tool_choice.type").String()
+	// "auto" is allowed with thinking, but "any" or "tool" (specific tool) are not
+	if toolChoiceType == "any" || toolChoiceType == "tool" {
+		// Remove thinking configuration entirely to avoid API error
+		body, _ = sjson.DeleteBytes(body, "thinking")
 	}
-	return util.ApplyClaudeThinkingConfig(body, budget)
+	return body
 }
 
 // ensureMaxTokensForThinking ensures max_tokens > thinking.budget_tokens when thinking is enabled.
@@ -480,7 +560,7 @@ func ensureMaxTokensForThinking(modelName string, body []byte) []byte {
 
 	// Look up the model's max completion tokens from the registry
 	maxCompletionTokens := 0
-	if modelInfo := registry.GetGlobalRegistry().GetModelInfo(modelName); modelInfo != nil {
+	if modelInfo := registry.LookupModelInfo(modelName); modelInfo != nil {
 		maxCompletionTokens = modelInfo.MaxCompletionTokens
 	}
 
@@ -497,51 +577,6 @@ func ensureMaxTokensForThinking(modelName string, body []byte) []byte {
 	return body
 }
 
-func (e *ClaudeExecutor) resolveUpstreamModel(alias string, auth *cliproxyauth.Auth) string {
-	trimmed := strings.TrimSpace(alias)
-	if trimmed == "" {
-		return ""
-	}
-
-	entry := e.resolveClaudeConfig(auth)
-	if entry == nil {
-		return ""
-	}
-
-	normalizedModel, metadata := util.NormalizeThinkingModel(trimmed)
-
-	// Candidate names to match against configured aliases/names.
-	candidates := []string{strings.TrimSpace(normalizedModel)}
-	if !strings.EqualFold(normalizedModel, trimmed) {
-		candidates = append(candidates, trimmed)
-	}
-	if original := util.ResolveOriginalModel(normalizedModel, metadata); original != "" && !strings.EqualFold(original, normalizedModel) {
-		candidates = append(candidates, original)
-	}
-
-	for i := range entry.Models {
-		model := entry.Models[i]
-		name := strings.TrimSpace(model.Name)
-		modelAlias := strings.TrimSpace(model.Alias)
-
-		for _, candidate := range candidates {
-			if candidate == "" {
-				continue
-			}
-			if modelAlias != "" && strings.EqualFold(modelAlias, candidate) {
-				if name != "" {
-					return name
-				}
-				return candidate
-			}
-			if name != "" && strings.EqualFold(name, candidate) {
-				return name
-			}
-		}
-	}
-	return ""
-}
-
 func (e *ClaudeExecutor) resolveClaudeConfig(auth *cliproxyauth.Auth) *config.ClaudeKey {
 	if auth == nil || e.cfg == nil {
 		return nil
@@ -662,7 +697,14 @@ func decodeResponseBody(body io.ReadCloser, contentEncoding string) (io.ReadClos
 }
 
 func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string, stream bool, extraBetas []string) {
-	r.Header.Set("Authorization", "Bearer "+apiKey)
+	useAPIKey := auth != nil && auth.Attributes != nil && strings.TrimSpace(auth.Attributes["api_key"]) != ""
+	isAnthropicBase := r.URL != nil && strings.EqualFold(r.URL.Scheme, "https") && strings.EqualFold(r.URL.Host, "api.anthropic.com")
+	if isAnthropicBase && useAPIKey {
+		r.Header.Del("Authorization")
+		r.Header.Set("x-api-key", apiKey)
+	} else {
+		r.Header.Set("Authorization", "Bearer "+apiKey)
+	}
 	r.Header.Set("Content-Type", "application/json")
 
 	var ginHeaders http.Header
@@ -755,3 +797,107 @@ func checkSystemInstructions(payload []byte) []byte {
 	}
 	return payload
 }
+
+func isClaudeOAuthToken(apiKey string) bool {
+	return strings.Contains(apiKey, "sk-ant-oat")
+}
+
+func applyClaudeToolPrefix(body []byte, prefix string) []byte {
+	if prefix == "" {
+		return body
+	}
+
+	if tools := gjson.GetBytes(body, "tools"); tools.Exists() && tools.IsArray() {
+		tools.ForEach(func(index, tool gjson.Result) bool {
+			name := tool.Get("name").String()
+			if name == "" || strings.HasPrefix(name, prefix) {
+				return true
+			}
+			path := fmt.Sprintf("tools.%d.name", index.Int())
+			body, _ = sjson.SetBytes(body, path, prefix+name)
+			return true
+		})
+	}
+
+	if gjson.GetBytes(body, "tool_choice.type").String() == "tool" {
+		name := gjson.GetBytes(body, "tool_choice.name").String()
+		if name != "" && !strings.HasPrefix(name, prefix) {
+			body, _ = sjson.SetBytes(body, "tool_choice.name", prefix+name)
+		}
+	}
+
+	if messages := gjson.GetBytes(body, "messages"); messages.Exists() && messages.IsArray() {
+		messages.ForEach(func(msgIndex, msg gjson.Result) bool {
+			content := msg.Get("content")
+			if !content.Exists() || !content.IsArray() {
+				return true
+			}
+			content.ForEach(func(contentIndex, part gjson.Result) bool {
+				if part.Get("type").String() != "tool_use" {
+					return true
+				}
+				name := part.Get("name").String()
+				if name == "" || strings.HasPrefix(name, prefix) {
+					return true
+				}
+				path := fmt.Sprintf("messages.%d.content.%d.name", msgIndex.Int(), contentIndex.Int())
+				body, _ = sjson.SetBytes(body, path, prefix+name)
+				return true
+			})
+			return true
+		})
+	}
+
+	return body
+}
+
+func stripClaudeToolPrefixFromResponse(body []byte, prefix string) []byte {
+	if prefix == "" {
+		return body
+	}
+	content := gjson.GetBytes(body, "content")
+	if !content.Exists() || !content.IsArray() {
+		return body
+	}
+	content.ForEach(func(index, part gjson.Result) bool {
+		if part.Get("type").String() != "tool_use" {
+			return true
+		}
+		name := part.Get("name").String()
+		if !strings.HasPrefix(name, prefix) {
+			return true
+		}
+		path := fmt.Sprintf("content.%d.name", index.Int())
+		body, _ = sjson.SetBytes(body, path, strings.TrimPrefix(name, prefix))
+		return true
+	})
+	return body
+}
+
+func stripClaudeToolPrefixFromStreamLine(line []byte, prefix string) []byte {
+	if prefix == "" {
+		return line
+	}
+	payload := jsonPayload(line)
+	if len(payload) == 0 || !gjson.ValidBytes(payload) {
+		return line
+	}
+	contentBlock := gjson.GetBytes(payload, "content_block")
+	if !contentBlock.Exists() || contentBlock.Get("type").String() != "tool_use" {
+		return line
+	}
+	name := contentBlock.Get("name").String()
+	if !strings.HasPrefix(name, prefix) {
+		return line
+	}
+	updated, err := sjson.SetBytes(payload, "content_block.name", strings.TrimPrefix(name, prefix))
+	if err != nil {
+		return line
+	}
+
+	trimmed := bytes.TrimSpace(line)
+	if bytes.HasPrefix(trimmed, []byte("data:")) {
+		return append([]byte("data: "), updated...)
+	}
+	return updated
+}

internal/runtime/executor/claude_executor_test.go

@@ -0,0 +1,51 @@
+package executor
+
+import (
+	"bytes"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestApplyClaudeToolPrefix(t *testing.T) {
+	input := []byte(`{"tools":[{"name":"alpha"},{"name":"proxy_bravo"}],"tool_choice":{"type":"tool","name":"charlie"},"messages":[{"role":"assistant","content":[{"type":"tool_use","name":"delta","id":"t1","input":{}}]}]}`)
+	out := applyClaudeToolPrefix(input, "proxy_")
+
+	if got := gjson.GetBytes(out, "tools.0.name").String(); got != "proxy_alpha" {
+		t.Fatalf("tools.0.name = %q, want %q", got, "proxy_alpha")
+	}
+	if got := gjson.GetBytes(out, "tools.1.name").String(); got != "proxy_bravo" {
+		t.Fatalf("tools.1.name = %q, want %q", got, "proxy_bravo")
+	}
+	if got := gjson.GetBytes(out, "tool_choice.name").String(); got != "proxy_charlie" {
+		t.Fatalf("tool_choice.name = %q, want %q", got, "proxy_charlie")
+	}
+	if got := gjson.GetBytes(out, "messages.0.content.0.name").String(); got != "proxy_delta" {
+		t.Fatalf("messages.0.content.0.name = %q, want %q", got, "proxy_delta")
+	}
+}
+
+func TestStripClaudeToolPrefixFromResponse(t *testing.T) {
+	input := []byte(`{"content":[{"type":"tool_use","name":"proxy_alpha","id":"t1","input":{}},{"type":"tool_use","name":"bravo","id":"t2","input":{}}]}`)
+	out := stripClaudeToolPrefixFromResponse(input, "proxy_")
+
+	if got := gjson.GetBytes(out, "content.0.name").String(); got != "alpha" {
+		t.Fatalf("content.0.name = %q, want %q", got, "alpha")
+	}
+	if got := gjson.GetBytes(out, "content.1.name").String(); got != "bravo" {
+		t.Fatalf("content.1.name = %q, want %q", got, "bravo")
+	}
+}
+
+func TestStripClaudeToolPrefixFromStreamLine(t *testing.T) {
+	line := []byte(`data: {"type":"content_block_start","content_block":{"type":"tool_use","name":"proxy_alpha","id":"t1"},"index":0}`)
+	out := stripClaudeToolPrefixFromStreamLine(line, "proxy_")
+
+	payload := bytes.TrimSpace(out)
+	if bytes.HasPrefix(payload, []byte("data:")) {
+		payload = bytes.TrimSpace(payload[len("data:"):])
+	}
+	if got := gjson.GetBytes(payload, "content_block.name").String(); got != "alpha" {
+		t.Fatalf("content_block.name = %q, want %q", got, "alpha")
+	}
+}

internal/runtime/executor/codex_executor.go

@@ -13,6 +13,7 @@ import (
 	codexauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/codex"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -38,31 +39,73 @@ func NewCodexExecutor(cfg *config.Config) *CodexExecutor { return &CodexExecutor
 
 func (e *CodexExecutor) Identifier() string { return "codex" }
 
-func (e *CodexExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Codex credentials into the outgoing HTTP request.
+func (e *CodexExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	apiKey, _ := codexCreds(auth)
+	if strings.TrimSpace(apiKey) != "" {
+		req.Header.Set("Authorization", "Bearer "+apiKey)
+	}
+	var attrs map[string]string
+	if auth != nil {
+		attrs = auth.Attributes
+	}
+	util.ApplyCustomHeadersFromAttrs(req, attrs)
+	return nil
+}
+
+// HttpRequest injects Codex credentials into the request and executes it.
+func (e *CodexExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("codex executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	apiKey, baseURL := codexCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	apiKey, baseURL := codexCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://chatgpt.com/backend-api/codex"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
 
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
-	body = NormalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
-		return resp, errValidate
+	userAgent := codexUserAgent(ctx)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	body = applyPayloadConfig(e.cfg, req.Model, body)
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	originalPayload = misc.InjectCodexUserAgent(originalPayload, userAgent)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := misc.InjectCodexUserAgent(bytes.Clone(req.Payload), userAgent)
+	body = sdktranslator.TranslateRequest(from, to, baseModel, body, false)
+	body = misc.StripCodexUserAgent(body)
+
+	body, err = thinking.ApplyThinking(body, req.Model, "codex")
+	if err != nil {
+		return resp, err
+	}
+
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 	body, _ = sjson.SetBytes(body, "stream", true)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
+	body, _ = sjson.DeleteBytes(body, "prompt_cache_retention")
 
 	url := strings.TrimSuffix(baseURL, "/") + "/responses"
 	httpReq, err := e.cacheHelper(ctx, from, url, req, body)
@@ -129,7 +172,7 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 		}
 
 		var param any
-		out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, line, &param)
+		out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(originalPayload), body, line, &param)
 		resp = cliproxyexecutor.Response{Payload: []byte(out)}
 		return resp, nil
 	}
@@ -138,28 +181,38 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 }
 
 func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	apiKey, baseURL := codexCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	apiKey, baseURL := codexCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://chatgpt.com/backend-api/codex"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
 
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
-	body = NormalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
-		return nil, errValidate
+	userAgent := codexUserAgent(ctx)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	body = applyPayloadConfig(e.cfg, req.Model, body)
+	originalPayload = misc.InjectCodexUserAgent(originalPayload, userAgent)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := misc.InjectCodexUserAgent(bytes.Clone(req.Payload), userAgent)
+	body = sdktranslator.TranslateRequest(from, to, baseModel, body, true)
+	body = misc.StripCodexUserAgent(body)
+
+	body, err = thinking.ApplyThinking(body, req.Model, "codex")
+	if err != nil {
+		return nil, err
+	}
+
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	body, _ = sjson.DeleteBytes(body, "prompt_cache_retention")
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	url := strings.TrimSuffix(baseURL, "/") + "/responses"
 	httpReq, err := e.cacheHelper(ctx, from, url, req, body)
@@ -231,7 +284,7 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 				}
 			}
 
-			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(line), &param)
+			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(originalPayload), body, bytes.Clone(line), &param)
 			for i := range chunks {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
 			}
@@ -246,20 +299,26 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 }
 
 func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	userAgent := codexUserAgent(ctx)
+	body := misc.InjectCodexUserAgent(bytes.Clone(req.Payload), userAgent)
+	body = sdktranslator.TranslateRequest(from, to, baseModel, body, false)
+	body = misc.StripCodexUserAgent(body)
 
-	modelForCounting := req.Model
+	body, err := thinking.ApplyThinking(body, req.Model, "codex")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
 
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
+	body, _ = sjson.DeleteBytes(body, "prompt_cache_retention")
 	body, _ = sjson.SetBytes(body, "stream", false)
 
-	enc, err := tokenizerForCodexModel(modelForCounting)
+	enc, err := tokenizerForCodexModel(baseModel)
 	if err != nil {
 		return cliproxyexecutor.Response{}, fmt.Errorf("codex executor: tokenizer init failed: %w", err)
 	}
@@ -440,14 +499,14 @@ func (e *CodexExecutor) cacheHelper(ctx context.Context, from sdktranslator.Form
 	if from == "claude" {
 		userIDResult := gjson.GetBytes(req.Payload, "metadata.user_id")
 		if userIDResult.Exists() {
-			var hasKey bool
 			key := fmt.Sprintf("%s-%s", req.Model, userIDResult.String())
-			if cache, hasKey = codexCacheMap[key]; !hasKey || cache.Expire.Before(time.Now()) {
+			var ok bool
+			if cache, ok = getCodexCache(key); !ok {
 				cache = codexCache{
 					ID:     uuid.New().String(),
 					Expire: time.Now().Add(1 * time.Hour),
 				}
-				codexCacheMap[key] = cache
+				setCodexCache(key, cache)
 			}
 		}
 	} else if from == "openai-response" {
@@ -505,6 +564,16 @@ func applyCodexHeaders(r *http.Request, auth *cliproxyauth.Auth, token string) {
 	util.ApplyCustomHeadersFromAttrs(r, attrs)
 }
 
+func codexUserAgent(ctx context.Context) string {
+	if ctx == nil {
+		return ""
+	}
+	if ginCtx, ok := ctx.Value("gin").(*gin.Context); ok && ginCtx != nil && ginCtx.Request != nil {
+		return strings.TrimSpace(ginCtx.Request.UserAgent())
+	}
+	return ""
+}
+
 func codexCreds(a *cliproxyauth.Auth) (apiKey, baseURL string) {
 	if a == nil {
 		return "", ""
@@ -520,3 +589,42 @@ func codexCreds(a *cliproxyauth.Auth) (apiKey, baseURL string) {
 	}
 	return
 }
+
+func (e *CodexExecutor) resolveCodexConfig(auth *cliproxyauth.Auth) *config.CodexKey {
+	if auth == nil || e.cfg == nil {
+		return nil
+	}
+	var attrKey, attrBase string
+	if auth.Attributes != nil {
+		attrKey = strings.TrimSpace(auth.Attributes["api_key"])
+		attrBase = strings.TrimSpace(auth.Attributes["base_url"])
+	}
+	for i := range e.cfg.CodexKey {
+		entry := &e.cfg.CodexKey[i]
+		cfgKey := strings.TrimSpace(entry.APIKey)
+		cfgBase := strings.TrimSpace(entry.BaseURL)
+		if attrKey != "" && attrBase != "" {
+			if strings.EqualFold(cfgKey, attrKey) && strings.EqualFold(cfgBase, attrBase) {
+				return entry
+			}
+			continue
+		}
+		if attrKey != "" && strings.EqualFold(cfgKey, attrKey) {
+			if cfgBase == "" || strings.EqualFold(cfgBase, attrBase) {
+				return entry
+			}
+		}
+		if attrKey == "" && attrBase != "" && strings.EqualFold(cfgBase, attrBase) {
+			return entry
+		}
+	}
+	if attrKey != "" {
+		for i := range e.cfg.CodexKey {
+			entry := &e.cfg.CodexKey[i]
+			if strings.EqualFold(strings.TrimSpace(entry.APIKey), attrKey) {
+				return entry
+			}
+		}
+	}
+	return nil
+}

internal/runtime/executor/gemini_cli_executor.go

@@ -20,6 +20,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/runtime/geminicli"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -63,28 +64,72 @@ func NewGeminiCLIExecutor(cfg *config.Config) *GeminiCLIExecutor {
 // Identifier returns the executor identifier.
 func (e *GeminiCLIExecutor) Identifier() string { return "gemini-cli" }
 
-// PrepareRequest prepares the HTTP request for execution (no-op for Gemini CLI).
-func (e *GeminiCLIExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Gemini CLI credentials into the outgoing HTTP request.
+func (e *GeminiCLIExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	tokenSource, _, errSource := prepareGeminiCLITokenSource(req.Context(), e.cfg, auth)
+	if errSource != nil {
+		return errSource
+	}
+	tok, errTok := tokenSource.Token()
+	if errTok != nil {
+		return errTok
+	}
+	if strings.TrimSpace(tok.AccessToken) == "" {
+		return statusErr{code: http.StatusUnauthorized, msg: "missing access token"}
+	}
+	req.Header.Set("Authorization", "Bearer "+tok.AccessToken)
+	applyGeminiCLIHeaders(req)
+	return nil
+}
+
+// HttpRequest injects Gemini CLI credentials into the request and executes it.
+func (e *GeminiCLIExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("gemini-cli executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 // Execute performs a non-streaming request to the Gemini CLI API.
 func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	tokenSource, baseTokenData, err := prepareGeminiCLITokenSource(ctx, e.cfg, auth)
 	if err != nil {
 		return resp, err
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini-cli")
-	basePayload := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	basePayload = applyThinkingMetadataCLI(basePayload, req.Metadata, req.Model)
-	basePayload = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, basePayload)
-	basePayload = util.ApplyDefaultThinkingIfNeededCLI(req.Model, basePayload)
-	basePayload = util.NormalizeGeminiCLIThinkingBudget(req.Model, basePayload)
-	basePayload = util.StripThinkingConfigIfUnsupported(req.Model, basePayload)
-	basePayload = fixGeminiCLIImageAspectRatio(req.Model, basePayload)
-	basePayload = applyPayloadConfigWithRoot(e.cfg, req.Model, "gemini", "request", basePayload)
+
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
+	}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, "gemini-cli")
+	if err != nil {
+		return resp, err
+	}
+
+	basePayload = fixGeminiCLIImageAspectRatio(baseModel, basePayload)
+	basePayload = applyPayloadConfigWithRoot(e.cfg, baseModel, "gemini", "request", basePayload, originalTranslated)
 
 	action := "generateContent"
 	if req.Metadata != nil {
@@ -94,9 +139,9 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 	}
 
 	projectID := resolveGeminiProjectID(auth)
-	models := cliPreviewFallbackOrder(req.Model)
-	if len(models) == 0 || models[0] != req.Model {
-		models = append([]string{req.Model}, models...)
+	models := cliPreviewFallbackOrder(baseModel)
+	if len(models) == 0 || models[0] != baseModel {
+		models = append([]string{baseModel}, models...)
 	}
 
 	httpClient := newHTTPClient(ctx, e.cfg, auth, 0)
@@ -207,29 +252,39 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 
 // ExecuteStream performs a streaming request to the Gemini CLI API.
 func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	tokenSource, baseTokenData, err := prepareGeminiCLITokenSource(ctx, e.cfg, auth)
 	if err != nil {
 		return nil, err
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini-cli")
-	basePayload := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	basePayload = applyThinkingMetadataCLI(basePayload, req.Metadata, req.Model)
-	basePayload = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, basePayload)
-	basePayload = util.ApplyDefaultThinkingIfNeededCLI(req.Model, basePayload)
-	basePayload = util.NormalizeGeminiCLIThinkingBudget(req.Model, basePayload)
-	basePayload = util.StripThinkingConfigIfUnsupported(req.Model, basePayload)
-	basePayload = fixGeminiCLIImageAspectRatio(req.Model, basePayload)
-	basePayload = applyPayloadConfigWithRoot(e.cfg, req.Model, "gemini", "request", basePayload)
+
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
+	}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+
+	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, "gemini-cli")
+	if err != nil {
+		return nil, err
+	}
+
+	basePayload = fixGeminiCLIImageAspectRatio(baseModel, basePayload)
+	basePayload = applyPayloadConfigWithRoot(e.cfg, baseModel, "gemini", "request", basePayload, originalTranslated)
 
 	projectID := resolveGeminiProjectID(auth)
 
-	models := cliPreviewFallbackOrder(req.Model)
-	if len(models) == 0 || models[0] != req.Model {
-		models = append([]string{req.Model}, models...)
+	models := cliPreviewFallbackOrder(baseModel)
+	if len(models) == 0 || models[0] != baseModel {
+		models = append([]string{baseModel}, models...)
 	}
 
 	httpClient := newHTTPClient(ctx, e.cfg, auth, 0)
@@ -318,7 +373,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 
 		out := make(chan cliproxyexecutor.StreamChunk)
 		stream = out
-		go func(resp *http.Response, reqBody []byte, attempt string) {
+		go func(resp *http.Response, reqBody []byte, attemptModel string) {
 			defer close(out)
 			defer func() {
 				if errClose := resp.Body.Close(); errClose != nil {
@@ -336,14 +391,14 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 						reporter.publish(ctx, detail)
 					}
 					if bytes.HasPrefix(line, dataTag) {
-						segments := sdktranslator.TranslateStream(respCtx, to, from, attempt, bytes.Clone(opts.OriginalRequest), reqBody, bytes.Clone(line), &param)
+						segments := sdktranslator.TranslateStream(respCtx, to, from, attemptModel, bytes.Clone(opts.OriginalRequest), reqBody, bytes.Clone(line), &param)
 						for i := range segments {
 							out <- cliproxyexecutor.StreamChunk{Payload: []byte(segments[i])}
 						}
 					}
 				}
 
-				segments := sdktranslator.TranslateStream(respCtx, to, from, attempt, bytes.Clone(opts.OriginalRequest), reqBody, bytes.Clone([]byte("[DONE]")), &param)
+				segments := sdktranslator.TranslateStream(respCtx, to, from, attemptModel, bytes.Clone(opts.OriginalRequest), reqBody, bytes.Clone([]byte("[DONE]")), &param)
 				for i := range segments {
 					out <- cliproxyexecutor.StreamChunk{Payload: []byte(segments[i])}
 				}
@@ -365,12 +420,12 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 			appendAPIResponseChunk(ctx, e.cfg, data)
 			reporter.publish(ctx, parseGeminiCLIUsage(data))
 			var param any
-			segments := sdktranslator.TranslateStream(respCtx, to, from, attempt, bytes.Clone(opts.OriginalRequest), reqBody, data, &param)
+			segments := sdktranslator.TranslateStream(respCtx, to, from, attemptModel, bytes.Clone(opts.OriginalRequest), reqBody, data, &param)
 			for i := range segments {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(segments[i])}
 			}
 
-			segments = sdktranslator.TranslateStream(respCtx, to, from, attempt, bytes.Clone(opts.OriginalRequest), reqBody, bytes.Clone([]byte("[DONE]")), &param)
+			segments = sdktranslator.TranslateStream(respCtx, to, from, attemptModel, bytes.Clone(opts.OriginalRequest), reqBody, bytes.Clone([]byte("[DONE]")), &param)
 			for i := range segments {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(segments[i])}
 			}
@@ -391,6 +446,8 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 
 // CountTokens counts tokens for the given request using the Gemini CLI API.
 func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	tokenSource, baseTokenData, err := prepareGeminiCLITokenSource(ctx, e.cfg, auth)
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
@@ -399,9 +456,9 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini-cli")
 
-	models := cliPreviewFallbackOrder(req.Model)
-	if len(models) == 0 || models[0] != req.Model {
-		models = append([]string{req.Model}, models...)
+	models := cliPreviewFallbackOrder(baseModel)
+	if len(models) == 0 || models[0] != baseModel {
+		models = append([]string{baseModel}, models...)
 	}
 
 	httpClient := newHTTPClient(ctx, e.cfg, auth, 0)
@@ -417,15 +474,20 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 	var lastStatus int
 	var lastBody []byte
 
-	for _, attemptModel := range models {
-		payload := sdktranslator.TranslateRequest(from, to, attemptModel, bytes.Clone(req.Payload), false)
-		payload = applyThinkingMetadataCLI(payload, req.Metadata, req.Model)
-		payload = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, payload)
+	// The loop variable attemptModel is only used as the concrete model id sent to the upstream
+	// Gemini CLI endpoint when iterating fallback variants.
+	for range models {
+		payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+		payload, err = thinking.ApplyThinking(payload, req.Model, "gemini-cli")
+		if err != nil {
+			return cliproxyexecutor.Response{}, err
+		}
+
 		payload = deleteJSONField(payload, "project")
 		payload = deleteJSONField(payload, "model")
 		payload = deleteJSONField(payload, "request.safetySettings")
-		payload = util.StripThinkingConfigIfUnsupported(req.Model, payload)
-		payload = fixGeminiCLIImageAspectRatio(attemptModel, payload)
+		payload = fixGeminiCLIImageAspectRatio(baseModel, payload)
 
 		tok, errTok := tokenSource.Token()
 		if errTok != nil {

internal/runtime/executor/gemini_executor.go

@@ -13,6 +13,7 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -55,8 +56,38 @@ func NewGeminiExecutor(cfg *config.Config) *GeminiExecutor {
 // Identifier returns the executor identifier.
 func (e *GeminiExecutor) Identifier() string { return "gemini" }
 
-// PrepareRequest prepares the HTTP request for execution (no-op for Gemini).
-func (e *GeminiExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Gemini credentials into the outgoing HTTP request.
+func (e *GeminiExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	apiKey, bearer := geminiCreds(auth)
+	if apiKey != "" {
+		req.Header.Set("x-goog-api-key", apiKey)
+		req.Header.Del("Authorization")
+	} else if bearer != "" {
+		req.Header.Set("Authorization", "Bearer "+bearer)
+		req.Header.Del("x-goog-api-key")
+	}
+	applyGeminiHeaders(req, auth)
+	return nil
+}
+
+// HttpRequest injects Gemini credentials into the request and executes it.
+func (e *GeminiExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("gemini executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 // Execute performs a non-streaming request to the Gemini API.
 // It translates the request to Gemini format, sends it to the API, and translates
@@ -72,24 +103,31 @@ func (e *GeminiExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) e
 //   - cliproxyexecutor.Response: The response from the API
 //   - error: An error if the request fails
 func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	apiKey, bearer := geminiCreds(auth)
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-
 	// Official Gemini API via API key or OAuth bearer
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = ApplyThinkingMetadata(body, req.Metadata, req.Model)
-	body = util.ApplyDefaultThinkingIfNeeded(req.Model, body)
-	body = util.NormalizeGeminiThinkingBudget(req.Model, body)
-	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
-	body = fixGeminiImageAspectRatio(req.Model, body)
-	body = applyPayloadConfig(e.cfg, req.Model, body)
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
+	}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return resp, err
+	}
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	action := "generateContent"
 	if req.Metadata != nil {
@@ -98,7 +136,7 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 		}
 	}
 	baseURL := resolveGeminiBaseURL(auth)
-	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, upstreamModel, action)
+	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, baseModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -168,26 +206,33 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 
 // ExecuteStream performs a streaming request to the Gemini API.
 func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	apiKey, bearer := geminiCreds(auth)
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	body = ApplyThinkingMetadata(body, req.Metadata, req.Model)
-	body = util.ApplyDefaultThinkingIfNeeded(req.Model, body)
-	body = util.NormalizeGeminiThinkingBudget(req.Model, body)
-	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
-	body = fixGeminiImageAspectRatio(req.Model, body)
-	body = applyPayloadConfig(e.cfg, req.Model, body)
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
+	}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return nil, err
+	}
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	baseURL := resolveGeminiBaseURL(auth)
-	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, upstreamModel, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, baseModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
@@ -285,21 +330,28 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 
 // CountTokens counts tokens for the given request using the Gemini API.
 func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	apiKey, bearer := geminiCreds(auth)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	translatedReq = ApplyThinkingMetadata(translatedReq, req.Metadata, req.Model)
-	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
-	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
+	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
+
+	translatedReq = fixGeminiImageAspectRatio(baseModel, translatedReq)
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "safetySettings")
+	translatedReq, _ = sjson.SetBytes(translatedReq, "model", baseModel)
 
 	baseURL := resolveGeminiBaseURL(auth)
-	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, req.Model, "countTokens")
+	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, baseModel, "countTokens")
 
 	requestBody := bytes.NewReader(translatedReq)
 
@@ -398,6 +450,45 @@ func resolveGeminiBaseURL(auth *cliproxyauth.Auth) string {
 	return base
 }
 
+func (e *GeminiExecutor) resolveGeminiConfig(auth *cliproxyauth.Auth) *config.GeminiKey {
+	if auth == nil || e.cfg == nil {
+		return nil
+	}
+	var attrKey, attrBase string
+	if auth.Attributes != nil {
+		attrKey = strings.TrimSpace(auth.Attributes["api_key"])
+		attrBase = strings.TrimSpace(auth.Attributes["base_url"])
+	}
+	for i := range e.cfg.GeminiKey {
+		entry := &e.cfg.GeminiKey[i]
+		cfgKey := strings.TrimSpace(entry.APIKey)
+		cfgBase := strings.TrimSpace(entry.BaseURL)
+		if attrKey != "" && attrBase != "" {
+			if strings.EqualFold(cfgKey, attrKey) && strings.EqualFold(cfgBase, attrBase) {
+				return entry
+			}
+			continue
+		}
+		if attrKey != "" && strings.EqualFold(cfgKey, attrKey) {
+			if cfgBase == "" || strings.EqualFold(cfgBase, attrBase) {
+				return entry
+			}
+		}
+		if attrKey == "" && attrBase != "" && strings.EqualFold(cfgBase, attrBase) {
+			return entry
+		}
+	}
+	if attrKey != "" {
+		for i := range e.cfg.GeminiKey {
+			entry := &e.cfg.GeminiKey[i]
+			if strings.EqualFold(strings.TrimSpace(entry.APIKey), attrKey) {
+				return entry
+			}
+		}
+	}
+	return nil
+}
+
 func applyGeminiHeaders(req *http.Request, auth *cliproxyauth.Auth) {
 	var attrs map[string]string
 	if auth != nil {

internal/runtime/executor/gemini_vertex_executor.go

@@ -15,7 +15,7 @@ import (
 
 	vertexauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/vertex"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
@@ -50,11 +50,49 @@ func NewGeminiVertexExecutor(cfg *config.Config) *GeminiVertexExecutor {
 // Identifier returns the executor identifier.
 func (e *GeminiVertexExecutor) Identifier() string { return "vertex" }
 
-// PrepareRequest prepares the HTTP request for execution (no-op for Vertex).
-func (e *GeminiVertexExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error {
+// PrepareRequest injects Vertex credentials into the outgoing HTTP request.
+func (e *GeminiVertexExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	apiKey, _ := vertexAPICreds(auth)
+	if strings.TrimSpace(apiKey) != "" {
+		req.Header.Set("x-goog-api-key", apiKey)
+		req.Header.Del("Authorization")
+		return nil
+	}
+	_, _, saJSON, errCreds := vertexCreds(auth)
+	if errCreds != nil {
+		return errCreds
+	}
+	token, errToken := vertexAccessToken(req.Context(), e.cfg, auth, saJSON)
+	if errToken != nil {
+		return errToken
+	}
+	if strings.TrimSpace(token) == "" {
+		return statusErr{code: http.StatusUnauthorized, msg: "missing access token"}
+	}
+	req.Header.Set("Authorization", "Bearer "+token)
+	req.Header.Del("x-goog-api-key")
 	return nil
 }
 
+// HttpRequest injects Vertex credentials into the request and executes it.
+func (e *GeminiVertexExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("vertex executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
+
 // Execute performs a non-streaming request to the Vertex AI API.
 func (e *GeminiVertexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
 	// Try API key authentication first
@@ -117,27 +155,29 @@ func (e *GeminiVertexExecutor) Refresh(_ context.Context, auth *cliproxyauth.Aut
 // executeWithServiceAccount handles authentication using service account credentials.
 // This method contains the original service account authentication logic.
 func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (resp cliproxyexecutor.Response, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		body = util.ApplyGeminiThinkingConfig(body, budgetOverride, includeOverride)
+
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	body = util.ApplyDefaultThinkingIfNeeded(req.Model, body)
-	body = util.NormalizeGeminiThinkingBudget(req.Model, body)
-	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
-	body = fixGeminiImageAspectRatio(req.Model, body)
-	body = applyPayloadConfig(e.cfg, req.Model, body)
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return resp, err
+	}
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	action := "generateContent"
 	if req.Metadata != nil {
@@ -146,7 +186,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 		}
 	}
 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, action)
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, baseModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -217,27 +257,29 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 
 // executeWithAPIKey handles authentication using API key credentials.
 func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (resp cliproxyexecutor.Response, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		body = util.ApplyGeminiThinkingConfig(body, budgetOverride, includeOverride)
+
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	body = util.ApplyDefaultThinkingIfNeeded(req.Model, body)
-	body = util.NormalizeGeminiThinkingBudget(req.Model, body)
-	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
-	body = fixGeminiImageAspectRatio(req.Model, body)
-	body = applyPayloadConfig(e.cfg, req.Model, body)
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return resp, err
+	}
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	action := "generateContent"
 	if req.Metadata != nil {
@@ -250,7 +292,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	if baseURL == "" {
 		baseURL = "https://generativelanguage.googleapis.com"
 	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, upstreamModel, action)
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -318,30 +360,32 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 
 // executeStreamWithServiceAccount handles streaming authentication using service account credentials.
 func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		body = util.ApplyGeminiThinkingConfig(body, budgetOverride, includeOverride)
+
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	body = util.ApplyDefaultThinkingIfNeeded(req.Model, body)
-	body = util.NormalizeGeminiThinkingBudget(req.Model, body)
-	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
-	body = fixGeminiImageAspectRatio(req.Model, body)
-	body = applyPayloadConfig(e.cfg, req.Model, body)
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return nil, err
+	}
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, baseModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
@@ -435,33 +479,35 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 
 // executeStreamWithAPIKey handles streaming authentication using API key credentials.
 func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		body = util.ApplyGeminiThinkingConfig(body, budgetOverride, includeOverride)
+
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	body = util.ApplyDefaultThinkingIfNeeded(req.Model, body)
-	body = util.NormalizeGeminiThinkingBudget(req.Model, body)
-	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
-	body = fixGeminiImageAspectRatio(req.Model, body)
-	body = applyPayloadConfig(e.cfg, req.Model, body)
-	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+
+	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	if err != nil {
+		return nil, err
+	}
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	// For API key auth, use simpler URL format without project/location
 	if baseURL == "" {
 		baseURL = "https://generativelanguage.googleapis.com"
 	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, upstreamModel, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
@@ -552,28 +598,27 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 
 // countTokensWithServiceAccount counts tokens using service account credentials.
 func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (cliproxyexecutor.Response, error) {
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		translatedReq = util.ApplyGeminiThinkingConfig(translatedReq, budgetOverride, includeOverride)
+
+	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
 	}
-	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
-	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
-	translatedReq, _ = sjson.SetBytes(translatedReq, "model", upstreamModel)
+
+	translatedReq = fixGeminiImageAspectRatio(baseModel, translatedReq)
+	translatedReq, _ = sjson.SetBytes(translatedReq, "model", baseModel)
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "safetySettings")
 
 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, "countTokens")
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, baseModel, "countTokens")
 
 	httpReq, errNewReq := http.NewRequestWithContext(respCtx, http.MethodPost, url, bytes.NewReader(translatedReq))
 	if errNewReq != nil {
@@ -630,10 +675,6 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 		return cliproxyexecutor.Response{}, errRead
 	}
 	appendAPIResponseChunk(ctx, e.cfg, data)
-	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
-		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(data)}
-	}
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
 	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
@@ -641,21 +682,20 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 
 // countTokensWithAPIKey handles token counting using API key credentials.
 func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (cliproxyexecutor.Response, error) {
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		translatedReq = util.ApplyGeminiThinkingConfig(translatedReq, budgetOverride, includeOverride)
+
+	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
 	}
-	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
-	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
-	translatedReq, _ = sjson.SetBytes(translatedReq, "model", upstreamModel)
+
+	translatedReq = fixGeminiImageAspectRatio(baseModel, translatedReq)
+	translatedReq, _ = sjson.SetBytes(translatedReq, "model", baseModel)
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
@@ -665,7 +705,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 	if baseURL == "" {
 		baseURL = "https://generativelanguage.googleapis.com"
 	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, req.Model, "countTokens")
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, "countTokens")
 
 	httpReq, errNewReq := http.NewRequestWithContext(respCtx, http.MethodPost, url, bytes.NewReader(translatedReq))
 	if errNewReq != nil {
@@ -719,10 +759,6 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 		return cliproxyexecutor.Response{}, errRead
 	}
 	appendAPIResponseChunk(ctx, e.cfg, data)
-	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
-		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(data)}
-	}
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
 	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
@@ -808,3 +844,43 @@ func vertexAccessToken(ctx context.Context, cfg *config.Config, auth *cliproxyau
 	}
 	return tok.AccessToken, nil
 }
+
+// resolveVertexConfig finds the matching vertex-api-key configuration entry for the given auth.
+func (e *GeminiVertexExecutor) resolveVertexConfig(auth *cliproxyauth.Auth) *config.VertexCompatKey {
+	if auth == nil || e.cfg == nil {
+		return nil
+	}
+	var attrKey, attrBase string
+	if auth.Attributes != nil {
+		attrKey = strings.TrimSpace(auth.Attributes["api_key"])
+		attrBase = strings.TrimSpace(auth.Attributes["base_url"])
+	}
+	for i := range e.cfg.VertexCompatAPIKey {
+		entry := &e.cfg.VertexCompatAPIKey[i]
+		cfgKey := strings.TrimSpace(entry.APIKey)
+		cfgBase := strings.TrimSpace(entry.BaseURL)
+		if attrKey != "" && attrBase != "" {
+			if strings.EqualFold(cfgKey, attrKey) && strings.EqualFold(cfgBase, attrBase) {
+				return entry
+			}
+			continue
+		}
+		if attrKey != "" && strings.EqualFold(cfgKey, attrKey) {
+			if cfgBase == "" || strings.EqualFold(cfgBase, attrBase) {
+				return entry
+			}
+		}
+		if attrKey == "" && attrBase != "" && strings.EqualFold(cfgBase, attrBase) {
+			return entry
+		}
+	}
+	if attrKey != "" {
+		for i := range e.cfg.VertexCompatAPIKey {
+			entry := &e.cfg.VertexCompatAPIKey[i]
+			if strings.EqualFold(strings.TrimSpace(entry.APIKey), attrKey) {
+				return entry
+			}
+		}
+	}
+	return nil
+}

internal/runtime/executor/iflow_executor.go

@@ -12,6 +12,7 @@ import (
 
 	iflowauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/iflow"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -37,11 +38,38 @@ func NewIFlowExecutor(cfg *config.Config) *IFlowExecutor { return &IFlowExecutor
 // Identifier returns the provider key.
 func (e *IFlowExecutor) Identifier() string { return "iflow" }
 
-// PrepareRequest implements ProviderExecutor but requires no preprocessing.
-func (e *IFlowExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects iFlow credentials into the outgoing HTTP request.
+func (e *IFlowExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	apiKey, _ := iflowCreds(auth)
+	if strings.TrimSpace(apiKey) != "" {
+		req.Header.Set("Authorization", "Bearer "+apiKey)
+	}
+	return nil
+}
+
+// HttpRequest injects iFlow credentials into the request and executes it.
+func (e *IFlowExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("iflow executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 // Execute performs a non-streaming chat completion request.
 func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	apiKey, baseURL := iflowCreds(auth)
 	if strings.TrimSpace(apiKey) == "" {
 		err = fmt.Errorf("iflow executor: missing api key")
@@ -51,23 +79,26 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 		baseURL = iflowauth.DefaultAPIBaseURL
 	}
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-	if upstreamModel != "" {
-		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	body = NormalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
-		return resp, errValidate
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
+
+	body, err = thinking.ApplyThinking(body, req.Model, "iflow")
+	if err != nil {
+		return resp, err
 	}
-	body = applyIFlowThinkingConfig(body)
-	body = applyPayloadConfig(e.cfg, req.Model, body)
+
+	body = preserveReasoningContentInMessages(body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	endpoint := strings.TrimSuffix(baseURL, "/") + iflowDefaultEndpoint
 
@@ -126,6 +157,8 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	reporter.ensurePublished(ctx)
 
 	var param any
+	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
+	// the original model name in the response for client compatibility.
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
@@ -133,6 +166,8 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 
 // ExecuteStream performs a streaming chat completion request.
 func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	apiKey, baseURL := iflowCreds(auth)
 	if strings.TrimSpace(apiKey) == "" {
 		err = fmt.Errorf("iflow executor: missing api key")
@@ -142,29 +177,31 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 		baseURL = iflowauth.DefaultAPIBaseURL
 	}
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
+	}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-	if upstreamModel != "" {
-		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	body, err = thinking.ApplyThinking(body, req.Model, "iflow")
+	if err != nil {
+		return nil, err
 	}
-	body = NormalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
-		return nil, errValidate
-	}
-	body = applyIFlowThinkingConfig(body)
+
+	body = preserveReasoningContentInMessages(body)
 	// Ensure tools array exists to avoid provider quirks similar to Qwen's behaviour.
 	toolsResult := gjson.GetBytes(body, "tools")
 	if toolsResult.Exists() && toolsResult.IsArray() && len(toolsResult.Array()) == 0 {
 		body = ensureToolsArray(body)
 	}
-	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	endpoint := strings.TrimSuffix(baseURL, "/") + iflowDefaultEndpoint
 
@@ -247,11 +284,13 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 }
 
 func (e *IFlowExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	enc, err := tokenizerForModel(req.Model)
+	enc, err := tokenizerForModel(baseModel)
 	if err != nil {
 		return cliproxyexecutor.Response{}, fmt.Errorf("iflow executor: tokenizer init failed: %w", err)
 	}
@@ -445,20 +484,47 @@ func ensureToolsArray(body []byte) []byte {
 	return updated
 }
 
-// applyIFlowThinkingConfig converts normalized reasoning_effort to iFlow chat_template_kwargs.enable_thinking.
-// This should be called after NormalizeThinkingConfig has processed the payload.
-// iFlow only supports boolean enable_thinking, so any non-"none" effort enables thinking.
-func applyIFlowThinkingConfig(body []byte) []byte {
-	effort := gjson.GetBytes(body, "reasoning_effort")
-	if !effort.Exists() {
+// preserveReasoningContentInMessages checks if reasoning_content from assistant messages
+// is preserved in conversation history for iFlow models that support thinking.
+// This is helpful for multi-turn conversations where the model may benefit from seeing
+// its previous reasoning to maintain coherent thought chains.
+//
+// For GLM-4.6/4.7 and MiniMax M2/M2.1, it is recommended to include the full assistant
+// response (including reasoning_content) in message history for better context continuity.
+func preserveReasoningContentInMessages(body []byte) []byte {
+	model := strings.ToLower(gjson.GetBytes(body, "model").String())
+
+	// Only apply to models that support thinking with history preservation
+	needsPreservation := strings.HasPrefix(model, "glm-4") || strings.HasPrefix(model, "minimax-m2")
+
+	if !needsPreservation {
 		return body
 	}
 
-	val := strings.ToLower(strings.TrimSpace(effort.String()))
-	enableThinking := val != "none" && val != ""
+	messages := gjson.GetBytes(body, "messages")
+	if !messages.Exists() || !messages.IsArray() {
+		return body
+	}
 
-	body, _ = sjson.DeleteBytes(body, "reasoning_effort")
-	body, _ = sjson.SetBytes(body, "chat_template_kwargs.enable_thinking", enableThinking)
+	// Check if any assistant message already has reasoning_content preserved
+	hasReasoningContent := false
+	messages.ForEach(func(_, msg gjson.Result) bool {
+		role := msg.Get("role").String()
+		if role == "assistant" {
+			rc := msg.Get("reasoning_content")
+			if rc.Exists() && rc.String() != "" {
+				hasReasoningContent = true
+				return false // stop iteration
+			}
+		}
+		return true
+	})
+
+	// If reasoning content is already present, the messages are properly formatted
+	// No need to modify - the client has correctly preserved reasoning in history
+	if hasReasoningContent {
+		log.Debugf("iflow executor: reasoning_content found in message history for %s", model)
+	}
 
 	return body
 }

internal/runtime/executor/iflow_executor_test.go

@@ -0,0 +1,67 @@
+package executor
+
+import (
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+)
+
+func TestIFlowExecutorParseSuffix(t *testing.T) {
+	tests := []struct {
+		name      string
+		model     string
+		wantBase  string
+		wantLevel string
+	}{
+		{"no suffix", "glm-4", "glm-4", ""},
+		{"glm with suffix", "glm-4.1-flash(high)", "glm-4.1-flash", "high"},
+		{"minimax no suffix", "minimax-m2", "minimax-m2", ""},
+		{"minimax with suffix", "minimax-m2.1(medium)", "minimax-m2.1", "medium"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := thinking.ParseSuffix(tt.model)
+			if result.ModelName != tt.wantBase {
+				t.Errorf("ParseSuffix(%q).ModelName = %q, want %q", tt.model, result.ModelName, tt.wantBase)
+			}
+		})
+	}
+}
+
+func TestPreserveReasoningContentInMessages(t *testing.T) {
+	tests := []struct {
+		name  string
+		input []byte
+		want  []byte // nil means output should equal input
+	}{
+		{
+			"non-glm model passthrough",
+			[]byte(`{"model":"gpt-4","messages":[]}`),
+			nil,
+		},
+		{
+			"glm model with empty messages",
+			[]byte(`{"model":"glm-4","messages":[]}`),
+			nil,
+		},
+		{
+			"glm model preserves existing reasoning_content",
+			[]byte(`{"model":"glm-4","messages":[{"role":"assistant","content":"hi","reasoning_content":"thinking..."}]}`),
+			nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := preserveReasoningContentInMessages(tt.input)
+			want := tt.want
+			if want == nil {
+				want = tt.input
+			}
+			if string(got) != string(want) {
+				t.Errorf("preserveReasoningContentInMessages() = %s, want %s", got, want)
+			}
+		})
+	}
+}

internal/runtime/executor/logging_helpers.go

@@ -304,11 +304,7 @@ func formatAuthInfo(info upstreamRequestLog) string {
 			parts = append(parts, "type=api_key")
 		}
 	case "oauth":
-		if authValue != "" {
-			parts = append(parts, fmt.Sprintf("type=oauth account=%s", authValue))
-		} else {
-			parts = append(parts, "type=oauth")
-		}
+		parts = append(parts, "type=oauth")
 	default:
 		if authType != "" {
 			if authValue != "" {

internal/runtime/executor/openai_compat_executor.go

@@ -11,6 +11,7 @@ import (
 	"time"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -35,13 +36,43 @@ func NewOpenAICompatExecutor(provider string, cfg *config.Config) *OpenAICompatE
 // Identifier implements cliproxyauth.ProviderExecutor.
 func (e *OpenAICompatExecutor) Identifier() string { return e.provider }
 
-// PrepareRequest is a no-op for now (credentials are added via headers at execution time).
-func (e *OpenAICompatExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error {
+// PrepareRequest injects OpenAI-compatible credentials into the outgoing HTTP request.
+func (e *OpenAICompatExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	_, apiKey := e.resolveCredentials(auth)
+	if strings.TrimSpace(apiKey) != "" {
+		req.Header.Set("Authorization", "Bearer "+apiKey)
+	}
+	var attrs map[string]string
+	if auth != nil {
+		attrs = auth.Attributes
+	}
+	util.ApplyCustomHeadersFromAttrs(req, attrs)
 	return nil
 }
 
+// HttpRequest injects OpenAI-compatible credentials into the request and executes it.
+func (e *OpenAICompatExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("openai compat executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
+
 func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	baseURL, apiKey := e.resolveCredentials(auth)
@@ -53,21 +84,17 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	// Translate inbound request to OpenAI format
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), opts.Stream)
-	modelOverride := e.resolveUpstreamModel(req.Model, auth)
-	if modelOverride != "" {
-		translated = e.overrideModel(translated, modelOverride)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
-	allowCompat := e.allowCompatReasoningEffort(req.Model, auth)
-	translated = ApplyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-	if upstreamModel != "" && modelOverride == "" {
-		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
-	}
-	translated = NormalizeThinkingConfig(translated, upstreamModel, allowCompat)
-	if errValidate := ValidateThinkingConfig(translated, upstreamModel); errValidate != nil {
-		return resp, errValidate
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, opts.Stream)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), opts.Stream)
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated)
+
+	translated, err = thinking.ApplyThinking(translated, req.Model, "openai")
+	if err != nil {
+		return resp, err
 	}
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
@@ -139,7 +166,9 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 }
 
 func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	baseURL, apiKey := e.resolveCredentials(auth)
@@ -147,23 +176,20 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 		err = statusErr{code: http.StatusUnauthorized, msg: "missing provider baseURL"}
 		return nil, err
 	}
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	modelOverride := e.resolveUpstreamModel(req.Model, auth)
-	if modelOverride != "" {
-		translated = e.overrideModel(translated, modelOverride)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
-	allowCompat := e.allowCompatReasoningEffort(req.Model, auth)
-	translated = ApplyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-	if upstreamModel != "" && modelOverride == "" {
-		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
-	}
-	translated = NormalizeThinkingConfig(translated, upstreamModel, allowCompat)
-	if errValidate := ValidateThinkingConfig(translated, upstreamModel); errValidate != nil {
-		return nil, errValidate
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated)
+
+	translated, err = thinking.ApplyThinking(translated, req.Model, "openai")
+	if err != nil {
+		return nil, err
 	}
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
@@ -239,6 +265,11 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 			if len(line) == 0 {
 				continue
 			}
+
+			if !bytes.HasPrefix(line, []byte("data:")) {
+				continue
+			}
+
 			// OpenAI-compatible streams are SSE: lines typically prefixed with "data: ".
 			// Pass through translator; it yields one or more chunks for the target schema.
 			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, bytes.Clone(line), &param)
@@ -258,14 +289,17 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 }
 
 func (e *OpenAICompatExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	modelForCounting := req.Model
-	if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
-		translated = e.overrideModel(translated, modelOverride)
-		modelForCounting = modelOverride
+	modelForCounting := baseModel
+
+	translated, err := thinking.ApplyThinking(translated, req.Model, "openai")
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
 	}
 
 	enc, err := tokenizerForModel(modelForCounting)
@@ -301,53 +335,6 @@ func (e *OpenAICompatExecutor) resolveCredentials(auth *cliproxyauth.Auth) (base
 	return
 }
 
-func (e *OpenAICompatExecutor) resolveUpstreamModel(alias string, auth *cliproxyauth.Auth) string {
-	if alias == "" || auth == nil || e.cfg == nil {
-		return ""
-	}
-	compat := e.resolveCompatConfig(auth)
-	if compat == nil {
-		return ""
-	}
-	for i := range compat.Models {
-		model := compat.Models[i]
-		if model.Alias != "" {
-			if strings.EqualFold(model.Alias, alias) {
-				if model.Name != "" {
-					return model.Name
-				}
-				return alias
-			}
-			continue
-		}
-		if strings.EqualFold(model.Name, alias) {
-			return model.Name
-		}
-	}
-	return ""
-}
-
-func (e *OpenAICompatExecutor) allowCompatReasoningEffort(model string, auth *cliproxyauth.Auth) bool {
-	trimmed := strings.TrimSpace(model)
-	if trimmed == "" || e == nil || e.cfg == nil {
-		return false
-	}
-	compat := e.resolveCompatConfig(auth)
-	if compat == nil || len(compat.Models) == 0 {
-		return false
-	}
-	for i := range compat.Models {
-		entry := compat.Models[i]
-		if strings.EqualFold(strings.TrimSpace(entry.Alias), trimmed) {
-			return true
-		}
-		if strings.EqualFold(strings.TrimSpace(entry.Name), trimmed) {
-			return true
-		}
-	}
-	return false
-}
-
 func (e *OpenAICompatExecutor) resolveCompatConfig(auth *cliproxyauth.Auth) *config.OpenAICompatibility {
 	if auth == nil || e.cfg == nil {
 		return nil

internal/runtime/executor/payload_helpers.go

@@ -1,103 +1,24 @@
 package executor
 
 import (
-	"fmt"
-	"net/http"
+	"encoding/json"
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
 
-// ApplyThinkingMetadata applies thinking config from model suffix metadata (e.g., (high), (8192))
-// for standard Gemini format payloads. It normalizes the budget when the model supports thinking.
-func ApplyThinkingMetadata(payload []byte, metadata map[string]any, model string) []byte {
-	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
-	if !ok || (budgetOverride == nil && includeOverride == nil) {
-		return payload
-	}
-	if !util.ModelSupportsThinking(model) {
-		return payload
-	}
-	if budgetOverride != nil {
-		norm := util.NormalizeThinkingBudget(model, *budgetOverride)
-		budgetOverride = &norm
-	}
-	return util.ApplyGeminiThinkingConfig(payload, budgetOverride, includeOverride)
-}
-
-// applyThinkingMetadataCLI applies thinking config from model suffix metadata (e.g., (high), (8192))
-// for Gemini CLI format payloads (nested under "request"). It normalizes the budget when the model supports thinking.
-func applyThinkingMetadataCLI(payload []byte, metadata map[string]any, model string) []byte {
-	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
-	if !ok || (budgetOverride == nil && includeOverride == nil) {
-		return payload
-	}
-	if !util.ModelSupportsThinking(model) {
-		return payload
-	}
-	if budgetOverride != nil {
-		norm := util.NormalizeThinkingBudget(model, *budgetOverride)
-		budgetOverride = &norm
-	}
-	return util.ApplyGeminiCLIThinkingConfig(payload, budgetOverride, includeOverride)
-}
-
-// ApplyReasoningEffortMetadata applies reasoning effort overrides from metadata to the given JSON path.
-// Metadata values take precedence over any existing field when the model supports thinking, intentionally
-// overwriting caller-provided values to honor suffix/default metadata priority.
-func ApplyReasoningEffortMetadata(payload []byte, metadata map[string]any, model, field string, allowCompat bool) []byte {
-	if len(metadata) == 0 {
-		return payload
-	}
-	if field == "" {
-		return payload
-	}
-	baseModel := util.ResolveOriginalModel(model, metadata)
-	if baseModel == "" {
-		baseModel = model
-	}
-	if !util.ModelSupportsThinking(baseModel) && !allowCompat {
-		return payload
-	}
-	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
-		if util.ModelUsesThinkingLevels(baseModel) || allowCompat {
-			if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
-				return updated
-			}
-		}
-	}
-	// Fallback: numeric thinking_budget suffix for level-based (OpenAI-style) models.
-	if util.ModelUsesThinkingLevels(baseModel) || allowCompat {
-		if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
-			if effort, ok := util.ThinkingBudgetToEffort(baseModel, *budget); ok && effort != "" {
-				if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
-					return updated
-				}
-			}
-		}
-	}
-	return payload
-}
-
-// applyPayloadConfig applies payload default and override rules from configuration
-// to the given JSON payload for the specified model.
-// Defaults only fill missing fields, while overrides always overwrite existing values.
-func applyPayloadConfig(cfg *config.Config, model string, payload []byte) []byte {
-	return applyPayloadConfigWithRoot(cfg, model, "", "", payload)
-}
-
 // applyPayloadConfigWithRoot behaves like applyPayloadConfig but treats all parameter
 // paths as relative to the provided root path (for example, "request" for Gemini CLI)
-// and restricts matches to the given protocol when supplied.
-func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string, payload []byte) []byte {
+// and restricts matches to the given protocol when supplied. Defaults are checked
+// against the original payload when provided.
+func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string, payload, original []byte) []byte {
 	if cfg == nil || len(payload) == 0 {
 		return payload
 	}
 	rules := cfg.Payload
-	if len(rules.Default) == 0 && len(rules.Override) == 0 {
+	if len(rules.Default) == 0 && len(rules.DefaultRaw) == 0 && len(rules.Override) == 0 && len(rules.OverrideRaw) == 0 {
 		return payload
 	}
 	model = strings.TrimSpace(model)
@@ -105,6 +26,11 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 		return payload
 	}
 	out := payload
+	source := original
+	if len(source) == 0 {
+		source = payload
+	}
+	appliedDefaults := make(map[string]struct{})
 	// Apply default rules: first write wins per field across all matching rules.
 	for i := range rules.Default {
 		rule := &rules.Default[i]
@@ -116,7 +42,10 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 			if fullPath == "" {
 				continue
 			}
-			if gjson.GetBytes(out, fullPath).Exists() {
+			if gjson.GetBytes(source, fullPath).Exists() {
+				continue
+			}
+			if _, ok := appliedDefaults[fullPath]; ok {
 				continue
 			}
 			updated, errSet := sjson.SetBytes(out, fullPath, value)
@@ -124,6 +53,36 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 				continue
 			}
 			out = updated
+			appliedDefaults[fullPath] = struct{}{}
+		}
+	}
+	// Apply default raw rules: first write wins per field across all matching rules.
+	for i := range rules.DefaultRaw {
+		rule := &rules.DefaultRaw[i]
+		if !payloadRuleMatchesModel(rule, model, protocol) {
+			continue
+		}
+		for path, value := range rule.Params {
+			fullPath := buildPayloadPath(root, path)
+			if fullPath == "" {
+				continue
+			}
+			if gjson.GetBytes(source, fullPath).Exists() {
+				continue
+			}
+			if _, ok := appliedDefaults[fullPath]; ok {
+				continue
+			}
+			rawValue, ok := payloadRawValue(value)
+			if !ok {
+				continue
+			}
+			updated, errSet := sjson.SetRawBytes(out, fullPath, rawValue)
+			if errSet != nil {
+				continue
+			}
+			out = updated
+			appliedDefaults[fullPath] = struct{}{}
 		}
 	}
 	// Apply override rules: last write wins per field across all matching rules.
@@ -144,6 +103,28 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 			out = updated
 		}
 	}
+	// Apply override raw rules: last write wins per field across all matching rules.
+	for i := range rules.OverrideRaw {
+		rule := &rules.OverrideRaw[i]
+		if !payloadRuleMatchesModel(rule, model, protocol) {
+			continue
+		}
+		for path, value := range rule.Params {
+			fullPath := buildPayloadPath(root, path)
+			if fullPath == "" {
+				continue
+			}
+			rawValue, ok := payloadRawValue(value)
+			if !ok {
+				continue
+			}
+			updated, errSet := sjson.SetRawBytes(out, fullPath, rawValue)
+			if errSet != nil {
+				continue
+			}
+			out = updated
+		}
+	}
 	return out
 }
 
@@ -187,6 +168,24 @@ func buildPayloadPath(root, path string) string {
 	return r + "." + p
 }
 
+func payloadRawValue(value any) ([]byte, bool) {
+	if value == nil {
+		return nil, false
+	}
+	switch typed := value.(type) {
+	case string:
+		return []byte(typed), true
+	case []byte:
+		return typed, true
+	default:
+		raw, errMarshal := json.Marshal(typed)
+		if errMarshal != nil {
+			return nil, false
+		}
+		return raw, true
+	}
+}
+
 // matchModelPattern performs simple wildcard matching where '*' matches zero or more characters.
 // Examples:
 //
@@ -231,102 +230,3 @@ func matchModelPattern(pattern, model string) bool {
 	}
 	return pi == len(pattern)
 }
-
-// NormalizeThinkingConfig normalizes thinking-related fields in the payload
-// based on model capabilities. For models without thinking support, it strips
-// reasoning fields. For models with level-based thinking, it validates and
-// normalizes the reasoning effort level. For models with numeric budget thinking,
-// it strips the effort string fields.
-func NormalizeThinkingConfig(payload []byte, model string, allowCompat bool) []byte {
-	if len(payload) == 0 || model == "" {
-		return payload
-	}
-
-	if !util.ModelSupportsThinking(model) {
-		if allowCompat {
-			return payload
-		}
-		return StripThinkingFields(payload, false)
-	}
-
-	if util.ModelUsesThinkingLevels(model) {
-		return NormalizeReasoningEffortLevel(payload, model)
-	}
-
-	// Model supports thinking but uses numeric budgets, not levels.
-	// Strip effort string fields since they are not applicable.
-	return StripThinkingFields(payload, true)
-}
-
-// StripThinkingFields removes thinking-related fields from the payload for
-// models that do not support thinking. If effortOnly is true, only removes
-// effort string fields (for models using numeric budgets).
-func StripThinkingFields(payload []byte, effortOnly bool) []byte {
-	fieldsToRemove := []string{
-		"reasoning_effort",
-		"reasoning.effort",
-	}
-	if !effortOnly {
-		fieldsToRemove = append([]string{"reasoning", "thinking"}, fieldsToRemove...)
-	}
-	out := payload
-	for _, field := range fieldsToRemove {
-		if gjson.GetBytes(out, field).Exists() {
-			out, _ = sjson.DeleteBytes(out, field)
-		}
-	}
-	return out
-}
-
-// NormalizeReasoningEffortLevel validates and normalizes the reasoning_effort
-// or reasoning.effort field for level-based thinking models.
-func NormalizeReasoningEffortLevel(payload []byte, model string) []byte {
-	out := payload
-
-	if effort := gjson.GetBytes(out, "reasoning_effort"); effort.Exists() {
-		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
-			out, _ = sjson.SetBytes(out, "reasoning_effort", normalized)
-		}
-	}
-
-	if effort := gjson.GetBytes(out, "reasoning.effort"); effort.Exists() {
-		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
-			out, _ = sjson.SetBytes(out, "reasoning.effort", normalized)
-		}
-	}
-
-	return out
-}
-
-// ValidateThinkingConfig checks for unsupported reasoning levels on level-based models.
-// Returns a statusErr with 400 when an unsupported level is supplied to avoid silently
-// downgrading requests.
-func ValidateThinkingConfig(payload []byte, model string) error {
-	if len(payload) == 0 || model == "" {
-		return nil
-	}
-	if !util.ModelSupportsThinking(model) || !util.ModelUsesThinkingLevels(model) {
-		return nil
-	}
-
-	levels := util.GetModelThinkingLevels(model)
-	checkField := func(path string) error {
-		if effort := gjson.GetBytes(payload, path); effort.Exists() {
-			if _, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); !ok {
-				return statusErr{
-					code: http.StatusBadRequest,
-					msg:  fmt.Sprintf("unsupported reasoning effort level %q for model %s (supported: %s)", effort.String(), model, strings.Join(levels, ", ")),
-				}
-			}
-		}
-		return nil
-	}
-
-	if err := checkField("reasoning_effort"); err != nil {
-		return err
-	}
-	if err := checkField("reasoning.effort"); err != nil {
-		return err
-	}
-	return nil
-}

internal/runtime/executor/qwen_executor.go

@@ -12,7 +12,7 @@ import (
 
 	qwenauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/qwen"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
@@ -37,30 +37,61 @@ func NewQwenExecutor(cfg *config.Config) *QwenExecutor { return &QwenExecutor{cf
 
 func (e *QwenExecutor) Identifier() string { return "qwen" }
 
-func (e *QwenExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Qwen credentials into the outgoing HTTP request.
+func (e *QwenExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	token, _ := qwenCreds(auth)
+	if strings.TrimSpace(token) != "" {
+		req.Header.Set("Authorization", "Bearer "+token)
+	}
+	return nil
+}
+
+// HttpRequest injects Qwen credentials into the request and executes it.
+func (e *QwenExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("qwen executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	token, baseURL := qwenCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	token, baseURL := qwenCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://portal.qwen.ai/v1"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-	if upstreamModel != "" {
-		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	body = NormalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
-		return resp, errValidate
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
+
+	body, err = thinking.ApplyThinking(body, req.Model, "openai")
+	if err != nil {
+		return resp, err
 	}
-	body = applyPayloadConfig(e.cfg, req.Model, body)
+
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
@@ -113,33 +144,39 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	reporter.publish(ctx, parseOpenAIUsage(data))
 	var param any
+	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
+	// the original model name in the response for client compatibility.
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
 }
 
 func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	token, baseURL := qwenCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	token, baseURL := qwenCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://portal.qwen.ai/v1"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
+	}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
-	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-	if upstreamModel != "" {
-		body, _ = sjson.SetBytes(body, "model", upstreamModel)
-	}
-	body = NormalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
-		return nil, errValidate
+	body, err = thinking.ApplyThinking(body, req.Model, "openai")
+	if err != nil {
+		return nil, err
 	}
+
 	toolsResult := gjson.GetBytes(body, "tools")
 	// I'm addressing the Qwen3 "poisoning" issue, which is caused by the model needing a tool to be defined. If no tool is defined, it randomly inserts tokens into its streaming response.
 	// This will have no real consequences. It's just to scare Qwen3.
@@ -147,7 +184,7 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		body, _ = sjson.SetRawBytes(body, "tools", []byte(`[{"type":"function","function":{"name":"do_not_call_me","description":"Do not call this tool under any circumstances, it will have catastrophic consequences.","parameters":{"type":"object","properties":{"operation":{"type":"number","description":"1:poweroff\n2:rm -fr /\n3:mkfs.ext4 /dev/sda1"}},"required":["operation"]}}}]`))
 	}
 	body, _ = sjson.SetBytes(body, "stream_options.include_usage", true)
-	body = applyPayloadConfig(e.cfg, req.Model, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
@@ -227,13 +264,15 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 }
 
 func (e *QwenExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
 	modelName := gjson.GetBytes(body, "model").String()
 	if strings.TrimSpace(modelName) == "" {
-		modelName = req.Model
+		modelName = baseModel
 	}
 
 	enc, err := tokenizerForModel(modelName)

internal/runtime/executor/qwen_executor_test.go

@@ -0,0 +1,30 @@
+package executor
+
+import (
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+)
+
+func TestQwenExecutorParseSuffix(t *testing.T) {
+	tests := []struct {
+		name      string
+		model     string
+		wantBase  string
+		wantLevel string
+	}{
+		{"no suffix", "qwen-max", "qwen-max", ""},
+		{"with level suffix", "qwen-max(high)", "qwen-max", "high"},
+		{"with budget suffix", "qwen-max(16384)", "qwen-max", "16384"},
+		{"complex model name", "qwen-plus-latest(medium)", "qwen-plus-latest", "medium"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := thinking.ParseSuffix(tt.model)
+			if result.ModelName != tt.wantBase {
+				t.Errorf("ParseSuffix(%q).ModelName = %q, want %q", tt.model, result.ModelName, tt.wantBase)
+			}
+		})
+	}
+}

internal/runtime/executor/thinking_providers.go

@@ -0,0 +1,11 @@
+package executor
+
+import (
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/antigravity"
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/claude"
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/codex"
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/gemini"
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/geminicli"
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/iflow"
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/openai"
+)

internal/runtime/executor/usage_helpers.go

@@ -19,7 +19,7 @@ type usageReporter struct {
 	provider    string
 	model       string
 	authID      string
-	authIndex   uint64
+	authIndex   string
 	apiKey      string
 	source      string
 	requestedAt time.Time
@@ -275,6 +275,20 @@ func parseClaudeStreamUsage(line []byte) (usage.Detail, bool) {
 	return detail, true
 }
 
+func parseGeminiFamilyUsageDetail(node gjson.Result) usage.Detail {
+	detail := usage.Detail{
+		InputTokens:     node.Get("promptTokenCount").Int(),
+		OutputTokens:    node.Get("candidatesTokenCount").Int(),
+		ReasoningTokens: node.Get("thoughtsTokenCount").Int(),
+		TotalTokens:     node.Get("totalTokenCount").Int(),
+		CachedTokens:    node.Get("cachedContentTokenCount").Int(),
+	}
+	if detail.TotalTokens == 0 {
+		detail.TotalTokens = detail.InputTokens + detail.OutputTokens + detail.ReasoningTokens
+	}
+	return detail
+}
+
 func parseGeminiCLIUsage(data []byte) usage.Detail {
 	usageNode := gjson.ParseBytes(data)
 	node := usageNode.Get("response.usageMetadata")
@@ -284,16 +298,7 @@ func parseGeminiCLIUsage(data []byte) usage.Detail {
 	if !node.Exists() {
 		return usage.Detail{}
 	}
-	detail := usage.Detail{
-		InputTokens:     node.Get("promptTokenCount").Int(),
-		OutputTokens:    node.Get("candidatesTokenCount").Int(),
-		ReasoningTokens: node.Get("thoughtsTokenCount").Int(),
-		TotalTokens:     node.Get("totalTokenCount").Int(),
-	}
-	if detail.TotalTokens == 0 {
-		detail.TotalTokens = detail.InputTokens + detail.OutputTokens + detail.ReasoningTokens
-	}
-	return detail
+	return parseGeminiFamilyUsageDetail(node)
 }
 
 func parseGeminiUsage(data []byte) usage.Detail {
@@ -305,16 +310,7 @@ func parseGeminiUsage(data []byte) usage.Detail {
 	if !node.Exists() {
 		return usage.Detail{}
 	}
-	detail := usage.Detail{
-		InputTokens:     node.Get("promptTokenCount").Int(),
-		OutputTokens:    node.Get("candidatesTokenCount").Int(),
-		ReasoningTokens: node.Get("thoughtsTokenCount").Int(),
-		TotalTokens:     node.Get("totalTokenCount").Int(),
-	}
-	if detail.TotalTokens == 0 {
-		detail.TotalTokens = detail.InputTokens + detail.OutputTokens + detail.ReasoningTokens
-	}
-	return detail
+	return parseGeminiFamilyUsageDetail(node)
 }
 
 func parseGeminiStreamUsage(line []byte) (usage.Detail, bool) {
@@ -329,16 +325,7 @@ func parseGeminiStreamUsage(line []byte) (usage.Detail, bool) {
 	if !node.Exists() {
 		return usage.Detail{}, false
 	}
-	detail := usage.Detail{
-		InputTokens:     node.Get("promptTokenCount").Int(),
-		OutputTokens:    node.Get("candidatesTokenCount").Int(),
-		ReasoningTokens: node.Get("thoughtsTokenCount").Int(),
-		TotalTokens:     node.Get("totalTokenCount").Int(),
-	}
-	if detail.TotalTokens == 0 {
-		detail.TotalTokens = detail.InputTokens + detail.OutputTokens + detail.ReasoningTokens
-	}
-	return detail, true
+	return parseGeminiFamilyUsageDetail(node), true
 }
 
 func parseGeminiCLIStreamUsage(line []byte) (usage.Detail, bool) {
@@ -353,16 +340,7 @@ func parseGeminiCLIStreamUsage(line []byte) (usage.Detail, bool) {
 	if !node.Exists() {
 		return usage.Detail{}, false
 	}
-	detail := usage.Detail{
-		InputTokens:     node.Get("promptTokenCount").Int(),
-		OutputTokens:    node.Get("candidatesTokenCount").Int(),
-		ReasoningTokens: node.Get("thoughtsTokenCount").Int(),
-		TotalTokens:     node.Get("totalTokenCount").Int(),
-	}
-	if detail.TotalTokens == 0 {
-		detail.TotalTokens = detail.InputTokens + detail.OutputTokens + detail.ReasoningTokens
-	}
-	return detail, true
+	return parseGeminiFamilyUsageDetail(node), true
 }
 
 func parseAntigravityUsage(data []byte) usage.Detail {
@@ -377,16 +355,7 @@ func parseAntigravityUsage(data []byte) usage.Detail {
 	if !node.Exists() {
 		return usage.Detail{}
 	}
-	detail := usage.Detail{
-		InputTokens:     node.Get("promptTokenCount").Int(),
-		OutputTokens:    node.Get("candidatesTokenCount").Int(),
-		ReasoningTokens: node.Get("thoughtsTokenCount").Int(),
-		TotalTokens:     node.Get("totalTokenCount").Int(),
-	}
-	if detail.TotalTokens == 0 {
-		detail.TotalTokens = detail.InputTokens + detail.OutputTokens + detail.ReasoningTokens
-	}
-	return detail
+	return parseGeminiFamilyUsageDetail(node)
 }
 
 func parseAntigravityStreamUsage(line []byte) (usage.Detail, bool) {
@@ -404,16 +373,7 @@ func parseAntigravityStreamUsage(line []byte) (usage.Detail, bool) {
 	if !node.Exists() {
 		return usage.Detail{}, false
 	}
-	detail := usage.Detail{
-		InputTokens:     node.Get("promptTokenCount").Int(),
-		OutputTokens:    node.Get("candidatesTokenCount").Int(),
-		ReasoningTokens: node.Get("thoughtsTokenCount").Int(),
-		TotalTokens:     node.Get("totalTokenCount").Int(),
-	}
-	if detail.TotalTokens == 0 {
-		detail.TotalTokens = detail.InputTokens + detail.OutputTokens + detail.ReasoningTokens
-	}
-	return detail, true
+	return parseGeminiFamilyUsageDetail(node), true
 }
 
 var stopChunkWithoutUsage sync.Map
@@ -522,12 +482,16 @@ func StripUsageMetadataFromJSON(rawJSON []byte) ([]byte, bool) {
 	cleaned := jsonBytes
 	var changed bool
 
-	if gjson.GetBytes(cleaned, "usageMetadata").Exists() {
+	if usageMetadata = gjson.GetBytes(cleaned, "usageMetadata"); usageMetadata.Exists() {
+		// Rename usageMetadata to cpaUsageMetadata in the message_start event of Claude
+		cleaned, _ = sjson.SetRawBytes(cleaned, "cpaUsageMetadata", []byte(usageMetadata.Raw))
 		cleaned, _ = sjson.DeleteBytes(cleaned, "usageMetadata")
 		changed = true
 	}
 
-	if gjson.GetBytes(cleaned, "response.usageMetadata").Exists() {
+	if usageMetadata = gjson.GetBytes(cleaned, "response.usageMetadata"); usageMetadata.Exists() {
+		// Rename usageMetadata to cpaUsageMetadata in the message_start event of Claude
+		cleaned, _ = sjson.SetRawBytes(cleaned, "response.cpaUsageMetadata", []byte(usageMetadata.Raw))
 		cleaned, _ = sjson.DeleteBytes(cleaned, "response.usageMetadata")
 		changed = true
 	}

internal/thinking/apply.go

@@ -0,0 +1,451 @@
+// Package thinking provides unified thinking configuration processing.
+package thinking
+
+import (
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	log "github.com/sirupsen/logrus"
+	"github.com/tidwall/gjson"
+)
+
+// providerAppliers maps provider names to their ProviderApplier implementations.
+var providerAppliers = map[string]ProviderApplier{
+	"gemini":      nil,
+	"gemini-cli":  nil,
+	"claude":      nil,
+	"openai":      nil,
+	"codex":       nil,
+	"iflow":       nil,
+	"antigravity": nil,
+}
+
+// GetProviderApplier returns the ProviderApplier for the given provider name.
+// Returns nil if the provider is not registered.
+func GetProviderApplier(provider string) ProviderApplier {
+	return providerAppliers[provider]
+}
+
+// RegisterProvider registers a provider applier by name.
+func RegisterProvider(name string, applier ProviderApplier) {
+	providerAppliers[name] = applier
+}
+
+// IsUserDefinedModel reports whether the model is a user-defined model that should
+// have thinking configuration passed through without validation.
+//
+// User-defined models are configured via config file's models[] array
+// (e.g., openai-compatibility.*.models[], *-api-key.models[]). These models
+// are marked with UserDefined=true at registration time.
+//
+// User-defined models should have their thinking configuration applied directly,
+// letting the upstream service validate the configuration.
+func IsUserDefinedModel(modelInfo *registry.ModelInfo) bool {
+	if modelInfo == nil {
+		return true
+	}
+	return modelInfo.UserDefined
+}
+
+// ApplyThinking applies thinking configuration to a request body.
+//
+// This is the unified entry point for all providers. It follows the processing
+// order defined in FR25: route check → model capability query → config extraction
+// → validation → application.
+//
+// Suffix Priority: When the model name includes a thinking suffix (e.g., "gemini-2.5-pro(8192)"),
+// the suffix configuration takes priority over any thinking parameters in the request body.
+// This enables users to override thinking settings via the model name without modifying their
+// request payload.
+//
+// Parameters:
+//   - body: Original request body JSON
+//   - model: Model name, optionally with thinking suffix (e.g., "claude-sonnet-4-5(16384)")
+//   - provider: Provider name (gemini, gemini-cli, antigravity, claude, openai, codex, iflow)
+//
+// Returns:
+//   - Modified request body JSON with thinking configuration applied
+//   - Error if validation fails (ThinkingError). On error, the original body
+//     is returned (not nil) to enable defensive programming patterns.
+//
+// Passthrough behavior (returns original body without error):
+//   - Unknown provider (not in providerAppliers map)
+//   - modelInfo.Thinking is nil (model doesn't support thinking)
+//
+// Note: Unknown models (modelInfo is nil) are treated as user-defined models: we skip
+// validation and still apply the thinking config so the upstream can validate it.
+//
+// Example:
+//
+//	// With suffix - suffix config takes priority
+//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro(8192)", "gemini")
+//
+//	// Without suffix - uses body config
+//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro", "gemini")
+func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
+	// 1. Route check: Get provider applier
+	applier := GetProviderApplier(provider)
+	if applier == nil {
+		log.WithFields(log.Fields{
+			"provider": provider,
+			"model":    model,
+		}).Debug("thinking: unknown provider, passthrough |")
+		return body, nil
+	}
+
+	// 2. Parse suffix and get modelInfo
+	suffixResult := ParseSuffix(model)
+	baseModel := suffixResult.ModelName
+	modelInfo := registry.LookupModelInfo(baseModel)
+
+	// 3. Model capability check
+	// Unknown models are treated as user-defined so thinking config can still be applied.
+	// The upstream service is responsible for validating the configuration.
+	if IsUserDefinedModel(modelInfo) {
+		return applyUserDefinedModel(body, modelInfo, provider, suffixResult)
+	}
+	if modelInfo.Thinking == nil {
+		config := extractThinkingConfig(body, provider)
+		if hasThinkingConfig(config) {
+			log.WithFields(log.Fields{
+				"model":    baseModel,
+				"provider": provider,
+			}).Debug("thinking: model does not support thinking, stripping config |")
+			return StripThinkingConfig(body, provider), nil
+		}
+		log.WithFields(log.Fields{
+			"provider": provider,
+			"model":    baseModel,
+		}).Debug("thinking: model does not support thinking, passthrough |")
+		return body, nil
+	}
+
+	// 4. Get config: suffix priority over body
+	var config ThinkingConfig
+	if suffixResult.HasSuffix {
+		config = parseSuffixToConfig(suffixResult.RawSuffix, provider, model)
+		log.WithFields(log.Fields{
+			"provider": provider,
+			"model":    model,
+			"mode":     config.Mode,
+			"budget":   config.Budget,
+			"level":    config.Level,
+		}).Debug("thinking: config from model suffix |")
+	} else {
+		config = extractThinkingConfig(body, provider)
+		if hasThinkingConfig(config) {
+			log.WithFields(log.Fields{
+				"provider": provider,
+				"model":    modelInfo.ID,
+				"mode":     config.Mode,
+				"budget":   config.Budget,
+				"level":    config.Level,
+			}).Debug("thinking: original config from request |")
+		}
+	}
+
+	if !hasThinkingConfig(config) {
+		log.WithFields(log.Fields{
+			"provider": provider,
+			"model":    modelInfo.ID,
+		}).Debug("thinking: no config found, passthrough |")
+		return body, nil
+	}
+
+	// 5. Validate and normalize configuration
+	validated, err := ValidateConfig(config, modelInfo, provider)
+	if err != nil {
+		log.WithFields(log.Fields{
+			"provider": provider,
+			"model":    modelInfo.ID,
+			"error":    err.Error(),
+		}).Warn("thinking: validation failed |")
+		// Return original body on validation failure (defensive programming).
+		// This ensures callers who ignore the error won't receive nil body.
+		// The upstream service will decide how to handle the unmodified request.
+		return body, err
+	}
+
+	// Defensive check: ValidateConfig should never return (nil, nil)
+	if validated == nil {
+		log.WithFields(log.Fields{
+			"provider": provider,
+			"model":    modelInfo.ID,
+		}).Warn("thinking: ValidateConfig returned nil config without error, passthrough |")
+		return body, nil
+	}
+
+	log.WithFields(log.Fields{
+		"provider": provider,
+		"model":    modelInfo.ID,
+		"mode":     validated.Mode,
+		"budget":   validated.Budget,
+		"level":    validated.Level,
+	}).Debug("thinking: processed config to apply |")
+
+	// 6. Apply configuration using provider-specific applier
+	return applier.Apply(body, *validated, modelInfo)
+}
+
+// parseSuffixToConfig converts a raw suffix string to ThinkingConfig.
+//
+// Parsing priority:
+//  1. Special values: "none" → ModeNone, "auto"/"-1" → ModeAuto
+//  2. Level names: "minimal", "low", "medium", "high", "xhigh" → ModeLevel
+//  3. Numeric values: positive integers → ModeBudget, 0 → ModeNone
+//
+// If none of the above match, returns empty ThinkingConfig (treated as no config).
+func parseSuffixToConfig(rawSuffix, provider, model string) ThinkingConfig {
+	// 1. Try special values first (none, auto, -1)
+	if mode, ok := ParseSpecialSuffix(rawSuffix); ok {
+		switch mode {
+		case ModeNone:
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		case ModeAuto:
+			return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+		}
+	}
+
+	// 2. Try level parsing (minimal, low, medium, high, xhigh)
+	if level, ok := ParseLevelSuffix(rawSuffix); ok {
+		return ThinkingConfig{Mode: ModeLevel, Level: level}
+	}
+
+	// 3. Try numeric parsing
+	if budget, ok := ParseNumericSuffix(rawSuffix); ok {
+		if budget == 0 {
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		}
+		return ThinkingConfig{Mode: ModeBudget, Budget: budget}
+	}
+
+	// Unknown suffix format - return empty config
+	log.WithFields(log.Fields{
+		"provider":   provider,
+		"model":      model,
+		"raw_suffix": rawSuffix,
+	}).Debug("thinking: unknown suffix format, treating as no config |")
+	return ThinkingConfig{}
+}
+
+// applyUserDefinedModel applies thinking configuration for user-defined models
+// without ThinkingSupport validation.
+func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, provider string, suffixResult SuffixResult) ([]byte, error) {
+	// Get model ID for logging
+	modelID := ""
+	if modelInfo != nil {
+		modelID = modelInfo.ID
+	} else {
+		modelID = suffixResult.ModelName
+	}
+
+	// Get config: suffix priority over body
+	var config ThinkingConfig
+	if suffixResult.HasSuffix {
+		config = parseSuffixToConfig(suffixResult.RawSuffix, provider, modelID)
+	} else {
+		config = extractThinkingConfig(body, provider)
+	}
+
+	if !hasThinkingConfig(config) {
+		log.WithFields(log.Fields{
+			"model":    modelID,
+			"provider": provider,
+		}).Debug("thinking: user-defined model, passthrough (no config) |")
+		return body, nil
+	}
+
+	applier := GetProviderApplier(provider)
+	if applier == nil {
+		log.WithFields(log.Fields{
+			"model":    modelID,
+			"provider": provider,
+		}).Debug("thinking: user-defined model, passthrough (unknown provider) |")
+		return body, nil
+	}
+
+	log.WithFields(log.Fields{
+		"provider": provider,
+		"model":    modelID,
+		"mode":     config.Mode,
+		"budget":   config.Budget,
+		"level":    config.Level,
+	}).Debug("thinking: applying config for user-defined model (skip validation)")
+
+	return applier.Apply(body, config, modelInfo)
+}
+
+// extractThinkingConfig extracts provider-specific thinking config from request body.
+func extractThinkingConfig(body []byte, provider string) ThinkingConfig {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		return ThinkingConfig{}
+	}
+
+	switch provider {
+	case "claude":
+		return extractClaudeConfig(body)
+	case "gemini", "gemini-cli", "antigravity":
+		return extractGeminiConfig(body, provider)
+	case "openai":
+		return extractOpenAIConfig(body)
+	case "codex":
+		return extractCodexConfig(body)
+	case "iflow":
+		return extractIFlowConfig(body)
+	default:
+		return ThinkingConfig{}
+	}
+}
+
+func hasThinkingConfig(config ThinkingConfig) bool {
+	return config.Mode != ModeBudget || config.Budget != 0 || config.Level != ""
+}
+
+// extractClaudeConfig extracts thinking configuration from Claude format request body.
+//
+// Claude API format:
+//   - thinking.type: "enabled" or "disabled"
+//   - thinking.budget_tokens: integer (-1=auto, 0=disabled, >0=budget)
+//
+// Priority: thinking.type="disabled" takes precedence over budget_tokens.
+// When type="enabled" without budget_tokens, returns ModeAuto to indicate
+// the user wants thinking enabled but didn't specify a budget.
+func extractClaudeConfig(body []byte) ThinkingConfig {
+	thinkingType := gjson.GetBytes(body, "thinking.type").String()
+	if thinkingType == "disabled" {
+		return ThinkingConfig{Mode: ModeNone, Budget: 0}
+	}
+
+	// Check budget_tokens
+	if budget := gjson.GetBytes(body, "thinking.budget_tokens"); budget.Exists() {
+		value := int(budget.Int())
+		switch value {
+		case 0:
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		case -1:
+			return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+		default:
+			return ThinkingConfig{Mode: ModeBudget, Budget: value}
+		}
+	}
+
+	// If type="enabled" but no budget_tokens, treat as auto (user wants thinking but no budget specified)
+	if thinkingType == "enabled" {
+		return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+	}
+
+	return ThinkingConfig{}
+}
+
+// extractGeminiConfig extracts thinking configuration from Gemini format request body.
+//
+// Gemini API format:
+//   - generationConfig.thinkingConfig.thinkingLevel: "none", "auto", or level name (Gemini 3)
+//   - generationConfig.thinkingConfig.thinkingBudget: integer (Gemini 2.5)
+//
+// For gemini-cli and antigravity providers, the path is prefixed with "request.".
+//
+// Priority: thinkingLevel is checked first (Gemini 3 format), then thinkingBudget (Gemini 2.5 format).
+// This allows newer Gemini 3 level-based configs to take precedence.
+func extractGeminiConfig(body []byte, provider string) ThinkingConfig {
+	prefix := "generationConfig.thinkingConfig"
+	if provider == "gemini-cli" || provider == "antigravity" {
+		prefix = "request.generationConfig.thinkingConfig"
+	}
+
+	// Check thinkingLevel first (Gemini 3 format takes precedence)
+	if level := gjson.GetBytes(body, prefix+".thinkingLevel"); level.Exists() {
+		value := level.String()
+		switch value {
+		case "none":
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		case "auto":
+			return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+		default:
+			return ThinkingConfig{Mode: ModeLevel, Level: ThinkingLevel(value)}
+		}
+	}
+
+	// Check thinkingBudget (Gemini 2.5 format)
+	if budget := gjson.GetBytes(body, prefix+".thinkingBudget"); budget.Exists() {
+		value := int(budget.Int())
+		switch value {
+		case 0:
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		case -1:
+			return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+		default:
+			return ThinkingConfig{Mode: ModeBudget, Budget: value}
+		}
+	}
+
+	return ThinkingConfig{}
+}
+
+// extractOpenAIConfig extracts thinking configuration from OpenAI format request body.
+//
+// OpenAI API format:
+//   - reasoning_effort: "none", "low", "medium", "high" (discrete levels)
+//
+// OpenAI uses level-based thinking configuration only, no numeric budget support.
+// The "none" value is treated specially to return ModeNone.
+func extractOpenAIConfig(body []byte) ThinkingConfig {
+	// Check reasoning_effort (OpenAI Chat Completions format)
+	if effort := gjson.GetBytes(body, "reasoning_effort"); effort.Exists() {
+		value := effort.String()
+		if value == "none" {
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		}
+		return ThinkingConfig{Mode: ModeLevel, Level: ThinkingLevel(value)}
+	}
+
+	return ThinkingConfig{}
+}
+
+// extractCodexConfig extracts thinking configuration from Codex format request body.
+//
+// Codex API format (OpenAI Responses API):
+//   - reasoning.effort: "none", "low", "medium", "high"
+//
+// This is similar to OpenAI but uses nested field "reasoning.effort" instead of "reasoning_effort".
+func extractCodexConfig(body []byte) ThinkingConfig {
+	// Check reasoning.effort (Codex / OpenAI Responses API format)
+	if effort := gjson.GetBytes(body, "reasoning.effort"); effort.Exists() {
+		value := effort.String()
+		if value == "none" {
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		}
+		return ThinkingConfig{Mode: ModeLevel, Level: ThinkingLevel(value)}
+	}
+
+	return ThinkingConfig{}
+}
+
+// extractIFlowConfig extracts thinking configuration from iFlow format request body.
+//
+// iFlow API format (supports multiple model families):
+//   - GLM format: chat_template_kwargs.enable_thinking (boolean)
+//   - MiniMax format: reasoning_split (boolean)
+//
+// Returns ModeBudget with Budget=1 as a sentinel value indicating "enabled".
+// The actual budget/configuration is determined by the iFlow applier based on model capabilities.
+// Budget=1 is used because iFlow models don't use numeric budgets; they only support on/off.
+func extractIFlowConfig(body []byte) ThinkingConfig {
+	// GLM format: chat_template_kwargs.enable_thinking
+	if enabled := gjson.GetBytes(body, "chat_template_kwargs.enable_thinking"); enabled.Exists() {
+		if enabled.Bool() {
+			// Budget=1 is a sentinel meaning "enabled" (iFlow doesn't use numeric budgets)
+			return ThinkingConfig{Mode: ModeBudget, Budget: 1}
+		}
+		return ThinkingConfig{Mode: ModeNone, Budget: 0}
+	}
+
+	// MiniMax format: reasoning_split
+	if split := gjson.GetBytes(body, "reasoning_split"); split.Exists() {
+		if split.Bool() {
+			// Budget=1 is a sentinel meaning "enabled" (iFlow doesn't use numeric budgets)
+			return ThinkingConfig{Mode: ModeBudget, Budget: 1}
+		}
+		return ThinkingConfig{Mode: ModeNone, Budget: 0}
+	}
+
+	return ThinkingConfig{}
+}

internal/thinking/convert.go

@@ -0,0 +1,142 @@
+package thinking
+
+import (
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+)
+
+// levelToBudgetMap defines the standard Level → Budget mapping.
+// All keys are lowercase; lookups should use strings.ToLower.
+var levelToBudgetMap = map[string]int{
+	"none":    0,
+	"auto":    -1,
+	"minimal": 512,
+	"low":     1024,
+	"medium":  8192,
+	"high":    24576,
+	"xhigh":   32768,
+}
+
+// ConvertLevelToBudget converts a thinking level to a budget value.
+//
+// This is a semantic conversion that maps discrete levels to numeric budgets.
+// Level matching is case-insensitive.
+//
+// Level → Budget mapping:
+//   - none    → 0
+//   - auto    → -1
+//   - minimal → 512
+//   - low     → 1024
+//   - medium  → 8192
+//   - high    → 24576
+//   - xhigh   → 32768
+//
+// Returns:
+//   - budget: The converted budget value
+//   - ok: true if level is valid, false otherwise
+func ConvertLevelToBudget(level string) (int, bool) {
+	budget, ok := levelToBudgetMap[strings.ToLower(level)]
+	return budget, ok
+}
+
+// BudgetThreshold constants define the upper bounds for each thinking level.
+// These are used by ConvertBudgetToLevel for range-based mapping.
+const (
+	// ThresholdMinimal is the upper bound for "minimal" level (1-512)
+	ThresholdMinimal = 512
+	// ThresholdLow is the upper bound for "low" level (513-1024)
+	ThresholdLow = 1024
+	// ThresholdMedium is the upper bound for "medium" level (1025-8192)
+	ThresholdMedium = 8192
+	// ThresholdHigh is the upper bound for "high" level (8193-24576)
+	ThresholdHigh = 24576
+)
+
+// ConvertBudgetToLevel converts a budget value to the nearest thinking level.
+//
+// This is a semantic conversion that maps numeric budgets to discrete levels.
+// Uses threshold-based mapping for range conversion.
+//
+// Budget → Level thresholds:
+//   - -1        → auto
+//   - 0         → none
+//   - 1-512     → minimal
+//   - 513-1024  → low
+//   - 1025-8192 → medium
+//   - 8193-24576 → high
+//   - 24577+    → xhigh
+//
+// Returns:
+//   - level: The converted thinking level string
+//   - ok: true if budget is valid, false for invalid negatives (< -1)
+func ConvertBudgetToLevel(budget int) (string, bool) {
+	switch {
+	case budget < -1:
+		// Invalid negative values
+		return "", false
+	case budget == -1:
+		return string(LevelAuto), true
+	case budget == 0:
+		return string(LevelNone), true
+	case budget <= ThresholdMinimal:
+		return string(LevelMinimal), true
+	case budget <= ThresholdLow:
+		return string(LevelLow), true
+	case budget <= ThresholdMedium:
+		return string(LevelMedium), true
+	case budget <= ThresholdHigh:
+		return string(LevelHigh), true
+	default:
+		return string(LevelXHigh), true
+	}
+}
+
+// ModelCapability describes the thinking format support of a model.
+type ModelCapability int
+
+const (
+	// CapabilityUnknown indicates modelInfo is nil (passthrough behavior, internal use).
+	CapabilityUnknown ModelCapability = iota - 1
+	// CapabilityNone indicates model doesn't support thinking (Thinking is nil).
+	CapabilityNone
+	// CapabilityBudgetOnly indicates the model supports numeric budgets only.
+	CapabilityBudgetOnly
+	// CapabilityLevelOnly indicates the model supports discrete levels only.
+	CapabilityLevelOnly
+	// CapabilityHybrid indicates the model supports both budgets and levels.
+	CapabilityHybrid
+)
+
+// detectModelCapability determines the thinking format capability of a model.
+//
+// This is an internal function used by validation and conversion helpers.
+// It analyzes the model's ThinkingSupport configuration to classify the model:
+//   - CapabilityNone: modelInfo.Thinking is nil (model doesn't support thinking)
+//   - CapabilityBudgetOnly: Has Min/Max but no Levels (Claude, Gemini 2.5)
+//   - CapabilityLevelOnly: Has Levels but no Min/Max (OpenAI, iFlow)
+//   - CapabilityHybrid: Has both Min/Max and Levels (Gemini 3)
+//
+// Note: Returns a special sentinel value when modelInfo itself is nil (unknown model).
+func detectModelCapability(modelInfo *registry.ModelInfo) ModelCapability {
+	if modelInfo == nil {
+		return CapabilityUnknown // sentinel for "passthrough" behavior
+	}
+	if modelInfo.Thinking == nil {
+		return CapabilityNone
+	}
+	support := modelInfo.Thinking
+	hasBudget := support.Min > 0 || support.Max > 0
+	hasLevels := len(support.Levels) > 0
+
+	switch {
+	case hasBudget && hasLevels:
+		return CapabilityHybrid
+	case hasBudget:
+		return CapabilityBudgetOnly
+	case hasLevels:
+		return CapabilityLevelOnly
+	default:
+		return CapabilityNone
+	}
+}

internal/thinking/errors.go

@@ -0,0 +1,78 @@
+// Package thinking provides unified thinking configuration processing logic.
+package thinking
+
+import "net/http"
+
+// ErrorCode represents the type of thinking configuration error.
+type ErrorCode string
+
+// Error codes for thinking configuration processing.
+const (
+	// ErrInvalidSuffix indicates the suffix format cannot be parsed.
+	// Example: "model(abc" (missing closing parenthesis)
+	ErrInvalidSuffix ErrorCode = "INVALID_SUFFIX"
+
+	// ErrUnknownLevel indicates the level value is not in the valid list.
+	// Example: "model(ultra)" where "ultra" is not a valid level
+	ErrUnknownLevel ErrorCode = "UNKNOWN_LEVEL"
+
+	// ErrThinkingNotSupported indicates the model does not support thinking.
+	// Example: claude-haiku-4-5 does not have thinking capability
+	ErrThinkingNotSupported ErrorCode = "THINKING_NOT_SUPPORTED"
+
+	// ErrLevelNotSupported indicates the model does not support level mode.
+	// Example: using level with a budget-only model
+	ErrLevelNotSupported ErrorCode = "LEVEL_NOT_SUPPORTED"
+
+	// ErrProviderMismatch indicates the provider does not match the model.
+	// Example: applying Claude format to a Gemini model
+	ErrProviderMismatch ErrorCode = "PROVIDER_MISMATCH"
+)
+
+// ThinkingError represents an error that occurred during thinking configuration processing.
+//
+// This error type provides structured information about the error, including:
+//   - Code: A machine-readable error code for programmatic handling
+//   - Message: A human-readable description of the error
+//   - Model: The model name related to the error (optional)
+//   - Details: Additional context information (optional)
+type ThinkingError struct {
+	// Code is the machine-readable error code
+	Code ErrorCode
+	// Message is the human-readable error description.
+	// Should be lowercase, no trailing period, with context if applicable.
+	Message string
+	// Model is the model name related to this error (optional)
+	Model string
+	// Details contains additional context information (optional)
+	Details map[string]interface{}
+}
+
+// Error implements the error interface.
+// Returns the message directly without code prefix.
+// Use Code field for programmatic error handling.
+func (e *ThinkingError) Error() string {
+	return e.Message
+}
+
+// NewThinkingError creates a new ThinkingError with the given code and message.
+func NewThinkingError(code ErrorCode, message string) *ThinkingError {
+	return &ThinkingError{
+		Code:    code,
+		Message: message,
+	}
+}
+
+// NewThinkingErrorWithModel creates a new ThinkingError with model context.
+func NewThinkingErrorWithModel(code ErrorCode, message, model string) *ThinkingError {
+	return &ThinkingError{
+		Code:    code,
+		Message: message,
+		Model:   model,
+	}
+}
+
+// StatusCode implements a portable status code interface for HTTP handlers.
+func (e *ThinkingError) StatusCode() int {
+	return http.StatusBadRequest
+}

internal/thinking/provider/antigravity/apply.go

@@ -0,0 +1,201 @@
+// Package antigravity implements thinking configuration for Antigravity API format.
+//
+// Antigravity uses request.generationConfig.thinkingConfig.* path (same as gemini-cli)
+// but requires additional normalization for Claude models:
+//   - Ensure thinking budget < max_tokens
+//   - Remove thinkingConfig if budget < minimum allowed
+package antigravity
+
+import (
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier applies thinking configuration for Antigravity API format.
+type Applier struct{}
+
+var _ thinking.ProviderApplier = (*Applier)(nil)
+
+// NewApplier creates a new Antigravity thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("antigravity", NewApplier())
+}
+
+// Apply applies thinking configuration to Antigravity request body.
+//
+// For Claude models, additional constraints are applied:
+//   - Ensure thinking budget < max_tokens
+//   - Remove thinkingConfig if budget < minimum allowed
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return a.applyCompatible(body, config, modelInfo)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	isClaude := strings.Contains(strings.ToLower(modelInfo.ID), "claude")
+
+	// ModeAuto: Always use Budget format with thinkingBudget=-1
+	if config.Mode == thinking.ModeAuto {
+		return a.applyBudgetFormat(body, config, modelInfo, isClaude)
+	}
+	if config.Mode == thinking.ModeBudget {
+		return a.applyBudgetFormat(body, config, modelInfo, isClaude)
+	}
+
+	// For non-auto modes, choose format based on model capabilities
+	support := modelInfo.Thinking
+	if len(support.Levels) > 0 {
+		return a.applyLevelFormat(body, config)
+	}
+	return a.applyBudgetFormat(body, config, modelInfo, isClaude)
+}
+
+func (a *Applier) applyCompatible(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	isClaude := false
+	if modelInfo != nil {
+		isClaude = strings.Contains(strings.ToLower(modelInfo.ID), "claude")
+	}
+
+	if config.Mode == thinking.ModeAuto {
+		return a.applyBudgetFormat(body, config, modelInfo, isClaude)
+	}
+
+	if config.Mode == thinking.ModeLevel || (config.Mode == thinking.ModeNone && config.Level != "") {
+		return a.applyLevelFormat(body, config)
+	}
+
+	return a.applyBudgetFormat(body, config, modelInfo, isClaude)
+}
+
+func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingBudget")
+	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
+
+	if config.Mode == thinking.ModeNone {
+		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", false)
+		if config.Level != "" {
+			result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", string(config.Level))
+		}
+		return result, nil
+	}
+
+	// Only handle ModeLevel - budget conversion should be done by upper layer
+	if config.Mode != thinking.ModeLevel {
+		return body, nil
+	}
+
+	level := string(config.Level)
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", level)
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", true)
+	return result, nil
+}
+
+func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo, isClaude bool) ([]byte, error) {
+	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingLevel")
+	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
+
+	budget := config.Budget
+	includeThoughts := false
+	switch config.Mode {
+	case thinking.ModeNone:
+		includeThoughts = false
+	case thinking.ModeAuto:
+		includeThoughts = true
+	default:
+		includeThoughts = budget > 0
+	}
+
+	// Apply Claude-specific constraints
+	if isClaude && modelInfo != nil {
+		budget, result = a.normalizeClaudeBudget(budget, result, modelInfo)
+		// Check if budget was removed entirely
+		if budget == -2 {
+			return result, nil
+		}
+	}
+
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
+	return result, nil
+}
+
+// normalizeClaudeBudget applies Claude-specific constraints to thinking budget.
+//
+// It handles:
+//   - Ensuring thinking budget < max_tokens
+//   - Removing thinkingConfig if budget < minimum allowed
+//
+// Returns the normalized budget and updated payload.
+// Returns budget=-2 as a sentinel indicating thinkingConfig was removed entirely.
+func (a *Applier) normalizeClaudeBudget(budget int, payload []byte, modelInfo *registry.ModelInfo) (int, []byte) {
+	if modelInfo == nil {
+		return budget, payload
+	}
+
+	// Get effective max tokens
+	effectiveMax, setDefaultMax := a.effectiveMaxTokens(payload, modelInfo)
+	if effectiveMax > 0 && budget >= effectiveMax {
+		budget = effectiveMax - 1
+	}
+
+	// Check minimum budget
+	minBudget := 0
+	if modelInfo.Thinking != nil {
+		minBudget = modelInfo.Thinking.Min
+	}
+	if minBudget > 0 && budget >= 0 && budget < minBudget {
+		// Budget is below minimum, remove thinking config entirely
+		payload, _ = sjson.DeleteBytes(payload, "request.generationConfig.thinkingConfig")
+		return -2, payload
+	}
+
+	// Set default max tokens if needed
+	if setDefaultMax && effectiveMax > 0 {
+		payload, _ = sjson.SetBytes(payload, "request.generationConfig.maxOutputTokens", effectiveMax)
+	}
+
+	return budget, payload
+}
+
+// effectiveMaxTokens returns the max tokens to cap thinking:
+// prefer request-provided maxOutputTokens; otherwise fall back to model default.
+// The boolean indicates whether the value came from the model default (and thus should be written back).
+func (a *Applier) effectiveMaxTokens(payload []byte, modelInfo *registry.ModelInfo) (max int, fromModel bool) {
+	if maxTok := gjson.GetBytes(payload, "request.generationConfig.maxOutputTokens"); maxTok.Exists() && maxTok.Int() > 0 {
+		return int(maxTok.Int()), false
+	}
+	if modelInfo != nil && modelInfo.MaxCompletionTokens > 0 {
+		return modelInfo.MaxCompletionTokens, true
+	}
+	return 0, false
+}

internal/thinking/provider/claude/apply.go

@@ -0,0 +1,109 @@
+// Package claude implements thinking configuration scaffolding for Claude models.
+//
+// Claude models use the thinking.budget_tokens format with values in the range
+// 1024-128000. Some Claude models support ZeroAllowed (sonnet-4-5, opus-4-5),
+// while older models do not.
+// See: _bmad-output/planning-artifacts/architecture.md#Epic-6
+package claude
+
+import (
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier implements thinking.ProviderApplier for Claude models.
+// This applier is stateless and holds no configuration.
+type Applier struct{}
+
+// NewApplier creates a new Claude thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("claude", NewApplier())
+}
+
+// Apply applies thinking configuration to Claude request body.
+//
+// IMPORTANT: This method expects config to be pre-validated by thinking.ValidateConfig.
+// ValidateConfig handles:
+//   - Mode conversion (Level→Budget, Auto→Budget)
+//   - Budget clamping to model range
+//   - ZeroAllowed constraint enforcement
+//
+// Apply only processes ModeBudget and ModeNone; other modes are passed through unchanged.
+//
+// Expected output format when enabled:
+//
+//	{
+//	  "thinking": {
+//	    "type": "enabled",
+//	    "budget_tokens": 16384
+//	  }
+//	}
+//
+// Expected output format when disabled:
+//
+//	{
+//	  "thinking": {
+//	    "type": "disabled"
+//	  }
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return applyCompatibleClaude(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	// Only process ModeBudget and ModeNone; other modes pass through
+	// (caller should use ValidateConfig first to normalize modes)
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	// Budget is expected to be pre-validated by ValidateConfig (clamped, ZeroAllowed enforced)
+	// Decide enabled/disabled based on budget value
+	if config.Budget == 0 {
+		result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		return result, nil
+	}
+
+	result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+	result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+	return result, nil
+}
+
+func applyCompatibleClaude(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	switch config.Mode {
+	case thinking.ModeNone:
+		result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		return result, nil
+	case thinking.ModeAuto:
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		return result, nil
+	default:
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+		return result, nil
+	}
+}

internal/thinking/provider/codex/apply.go

@@ -0,0 +1,131 @@
+// Package codex implements thinking configuration for Codex (OpenAI Responses API) models.
+//
+// Codex models use the reasoning.effort format with discrete levels
+// (low/medium/high). This is similar to OpenAI but uses nested field
+// "reasoning.effort" instead of "reasoning_effort".
+// See: _bmad-output/planning-artifacts/architecture.md#Epic-8
+package codex
+
+import (
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier implements thinking.ProviderApplier for Codex models.
+//
+// Codex-specific behavior:
+//   - Output format: reasoning.effort (string: low/medium/high/xhigh)
+//   - Level-only mode: no numeric budget support
+//   - Some models support ZeroAllowed (gpt-5.1, gpt-5.2)
+type Applier struct{}
+
+var _ thinking.ProviderApplier = (*Applier)(nil)
+
+// NewApplier creates a new Codex thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("codex", NewApplier())
+}
+
+// Apply applies thinking configuration to Codex request body.
+//
+// Expected output format:
+//
+//	{
+//	  "reasoning": {
+//	    "effort": "high"
+//	  }
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return applyCompatibleCodex(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	// Only handle ModeLevel and ModeNone; other modes pass through unchanged.
+	if config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	if config.Mode == thinking.ModeLevel {
+		result, _ := sjson.SetBytes(body, "reasoning.effort", string(config.Level))
+		return result, nil
+	}
+
+	effort := ""
+	support := modelInfo.Thinking
+	if config.Budget == 0 {
+		if support.ZeroAllowed || hasLevel(support.Levels, string(thinking.LevelNone)) {
+			effort = string(thinking.LevelNone)
+		}
+	}
+	if effort == "" && config.Level != "" {
+		effort = string(config.Level)
+	}
+	if effort == "" && len(support.Levels) > 0 {
+		effort = support.Levels[0]
+	}
+	if effort == "" {
+		return body, nil
+	}
+
+	result, _ := sjson.SetBytes(body, "reasoning.effort", effort)
+	return result, nil
+}
+
+func applyCompatibleCodex(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	var effort string
+	switch config.Mode {
+	case thinking.ModeLevel:
+		if config.Level == "" {
+			return body, nil
+		}
+		effort = string(config.Level)
+	case thinking.ModeNone:
+		effort = string(thinking.LevelNone)
+		if config.Level != "" {
+			effort = string(config.Level)
+		}
+	case thinking.ModeAuto:
+		// Auto mode for user-defined models: pass through as "auto"
+		effort = string(thinking.LevelAuto)
+	case thinking.ModeBudget:
+		// Budget mode: convert budget to level using threshold mapping
+		level, ok := thinking.ConvertBudgetToLevel(config.Budget)
+		if !ok {
+			return body, nil
+		}
+		effort = level
+	default:
+		return body, nil
+	}
+
+	result, _ := sjson.SetBytes(body, "reasoning.effort", effort)
+	return result, nil
+}
+
+func hasLevel(levels []string, target string) bool {
+	for _, level := range levels {
+		if strings.EqualFold(strings.TrimSpace(level), target) {
+			return true
+		}
+	}
+	return false
+}

internal/thinking/provider/gemini/apply.go

@@ -0,0 +1,169 @@
+// Package gemini implements thinking configuration for Gemini models.
+//
+// Gemini models have two formats:
+//   - Gemini 2.5: Uses thinkingBudget (numeric)
+//   - Gemini 3.x: Uses thinkingLevel (string: minimal/low/medium/high)
+//     or thinkingBudget=-1 for auto/dynamic mode
+//
+// Output format is determined by ThinkingConfig.Mode and ThinkingSupport.Levels:
+//   - ModeAuto: Always uses thinkingBudget=-1 (both Gemini 2.5 and 3.x)
+//   - len(Levels) > 0: Uses thinkingLevel (Gemini 3.x discrete levels)
+//   - len(Levels) == 0: Uses thinkingBudget (Gemini 2.5)
+package gemini
+
+import (
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier applies thinking configuration for Gemini models.
+//
+// Gemini-specific behavior:
+//   - Gemini 2.5: thinkingBudget format, flash series supports ZeroAllowed
+//   - Gemini 3.x: thinkingLevel format, cannot be disabled
+//   - Use ThinkingSupport.Levels to decide output format
+type Applier struct{}
+
+// NewApplier creates a new Gemini thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("gemini", NewApplier())
+}
+
+// Apply applies thinking configuration to Gemini request body.
+//
+// Expected output format (Gemini 2.5):
+//
+//	{
+//	  "generationConfig": {
+//	    "thinkingConfig": {
+//	      "thinkingBudget": 8192,
+//	      "includeThoughts": true
+//	    }
+//	  }
+//	}
+//
+// Expected output format (Gemini 3.x):
+//
+//	{
+//	  "generationConfig": {
+//	    "thinkingConfig": {
+//	      "thinkingLevel": "high",
+//	      "includeThoughts": true
+//	    }
+//	  }
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return a.applyCompatible(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	// Choose format based on config.Mode and model capabilities:
+	// - ModeLevel: use Level format (validation will reject unsupported levels)
+	// - ModeNone: use Level format if model has Levels, else Budget format
+	// - ModeBudget/ModeAuto: use Budget format
+	switch config.Mode {
+	case thinking.ModeLevel:
+		return a.applyLevelFormat(body, config)
+	case thinking.ModeNone:
+		// ModeNone: route based on model capability (has Levels or not)
+		if len(modelInfo.Thinking.Levels) > 0 {
+			return a.applyLevelFormat(body, config)
+		}
+		return a.applyBudgetFormat(body, config)
+	default:
+		return a.applyBudgetFormat(body, config)
+	}
+}
+
+func (a *Applier) applyCompatible(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	if config.Mode == thinking.ModeAuto {
+		return a.applyBudgetFormat(body, config)
+	}
+
+	if config.Mode == thinking.ModeLevel || (config.Mode == thinking.ModeNone && config.Level != "") {
+		return a.applyLevelFormat(body, config)
+	}
+
+	return a.applyBudgetFormat(body, config)
+}
+
+func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	// ModeNone semantics:
+	//   - ModeNone + Budget=0: completely disable thinking (not possible for Level-only models)
+	//   - ModeNone + Budget>0: forced to think but hide output (includeThoughts=false)
+	// ValidateConfig sets config.Level to the lowest level when ModeNone + Budget > 0.
+
+	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	result, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig.thinkingBudget")
+	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.include_thoughts")
+
+	if config.Mode == thinking.ModeNone {
+		result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", false)
+		if config.Level != "" {
+			result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingLevel", string(config.Level))
+		}
+		return result, nil
+	}
+
+	// Only handle ModeLevel - budget conversion should be done by upper layer
+	if config.Mode != thinking.ModeLevel {
+		return body, nil
+	}
+
+	level := string(config.Level)
+	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingLevel", level)
+	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", true)
+	return result, nil
+}
+
+func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	result, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig.thinkingLevel")
+	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.include_thoughts")
+
+	budget := config.Budget
+	// ModeNone semantics:
+	//   - ModeNone + Budget=0: completely disable thinking
+	//   - ModeNone + Budget>0: forced to think but hide output (includeThoughts=false)
+	// When ZeroAllowed=false, ValidateConfig clamps Budget to Min while preserving ModeNone.
+	includeThoughts := false
+	switch config.Mode {
+	case thinking.ModeNone:
+		includeThoughts = false
+	case thinking.ModeAuto:
+		includeThoughts = true
+	default:
+		includeThoughts = budget > 0
+	}
+
+	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingBudget", budget)
+	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", includeThoughts)
+	return result, nil
+}

internal/thinking/provider/geminicli/apply.go

@@ -0,0 +1,126 @@
+// Package geminicli implements thinking configuration for Gemini CLI API format.
+//
+// Gemini CLI uses request.generationConfig.thinkingConfig.* path instead of
+// generationConfig.thinkingConfig.* used by standard Gemini API.
+package geminicli
+
+import (
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier applies thinking configuration for Gemini CLI API format.
+type Applier struct{}
+
+var _ thinking.ProviderApplier = (*Applier)(nil)
+
+// NewApplier creates a new Gemini CLI thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("gemini-cli", NewApplier())
+}
+
+// Apply applies thinking configuration to Gemini CLI request body.
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return a.applyCompatible(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	// ModeAuto: Always use Budget format with thinkingBudget=-1
+	if config.Mode == thinking.ModeAuto {
+		return a.applyBudgetFormat(body, config)
+	}
+	if config.Mode == thinking.ModeBudget {
+		return a.applyBudgetFormat(body, config)
+	}
+
+	// For non-auto modes, choose format based on model capabilities
+	support := modelInfo.Thinking
+	if len(support.Levels) > 0 {
+		return a.applyLevelFormat(body, config)
+	}
+	return a.applyBudgetFormat(body, config)
+}
+
+func (a *Applier) applyCompatible(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	if config.Mode == thinking.ModeAuto {
+		return a.applyBudgetFormat(body, config)
+	}
+
+	if config.Mode == thinking.ModeLevel || (config.Mode == thinking.ModeNone && config.Level != "") {
+		return a.applyLevelFormat(body, config)
+	}
+
+	return a.applyBudgetFormat(body, config)
+}
+
+func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingBudget")
+	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
+
+	if config.Mode == thinking.ModeNone {
+		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", false)
+		if config.Level != "" {
+			result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", string(config.Level))
+		}
+		return result, nil
+	}
+
+	// Only handle ModeLevel - budget conversion should be done by upper layer
+	if config.Mode != thinking.ModeLevel {
+		return body, nil
+	}
+
+	level := string(config.Level)
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", level)
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", true)
+	return result, nil
+}
+
+func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingLevel")
+	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
+
+	budget := config.Budget
+	includeThoughts := false
+	switch config.Mode {
+	case thinking.ModeNone:
+		includeThoughts = false
+	case thinking.ModeAuto:
+		includeThoughts = true
+	default:
+		includeThoughts = budget > 0
+	}
+
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
+	return result, nil
+}

internal/thinking/provider/iflow/apply.go

@@ -0,0 +1,156 @@
+// Package iflow implements thinking configuration for iFlow models (GLM, MiniMax).
+//
+// iFlow models use boolean toggle semantics:
+//   - GLM models: chat_template_kwargs.enable_thinking (boolean)
+//   - MiniMax models: reasoning_split (boolean)
+//
+// Level values are converted to boolean: none=false, all others=true
+// See: _bmad-output/planning-artifacts/architecture.md#Epic-9
+package iflow
+
+import (
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier implements thinking.ProviderApplier for iFlow models.
+//
+// iFlow-specific behavior:
+//   - GLM models: enable_thinking boolean + clear_thinking=false
+//   - MiniMax models: reasoning_split boolean
+//   - Level to boolean: none=false, others=true
+//   - No quantized support (only on/off)
+type Applier struct{}
+
+var _ thinking.ProviderApplier = (*Applier)(nil)
+
+// NewApplier creates a new iFlow thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("iflow", NewApplier())
+}
+
+// Apply applies thinking configuration to iFlow request body.
+//
+// Expected output format (GLM):
+//
+//	{
+//	  "chat_template_kwargs": {
+//	    "enable_thinking": true,
+//	    "clear_thinking": false
+//	  }
+//	}
+//
+// Expected output format (MiniMax):
+//
+//	{
+//	  "reasoning_split": true
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return body, nil
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	if isGLMModel(modelInfo.ID) {
+		return applyGLM(body, config), nil
+	}
+
+	if isMiniMaxModel(modelInfo.ID) {
+		return applyMiniMax(body, config), nil
+	}
+
+	return body, nil
+}
+
+// configToBoolean converts ThinkingConfig to boolean for iFlow models.
+//
+// Conversion rules:
+//   - ModeNone: false
+//   - ModeAuto: true
+//   - ModeBudget + Budget=0: false
+//   - ModeBudget + Budget>0: true
+//   - ModeLevel + Level="none": false
+//   - ModeLevel + any other level: true
+//   - Default (unknown mode): true
+func configToBoolean(config thinking.ThinkingConfig) bool {
+	switch config.Mode {
+	case thinking.ModeNone:
+		return false
+	case thinking.ModeAuto:
+		return true
+	case thinking.ModeBudget:
+		return config.Budget > 0
+	case thinking.ModeLevel:
+		return config.Level != thinking.LevelNone
+	default:
+		return true
+	}
+}
+
+// applyGLM applies thinking configuration for GLM models.
+//
+// Output format when enabled:
+//
+//	{"chat_template_kwargs": {"enable_thinking": true, "clear_thinking": false}}
+//
+// Output format when disabled:
+//
+//	{"chat_template_kwargs": {"enable_thinking": false}}
+//
+// Note: clear_thinking is only set when thinking is enabled, to preserve
+// thinking output in the response.
+func applyGLM(body []byte, config thinking.ThinkingConfig) []byte {
+	enableThinking := configToBoolean(config)
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	result, _ := sjson.SetBytes(body, "chat_template_kwargs.enable_thinking", enableThinking)
+
+	// clear_thinking only needed when thinking is enabled
+	if enableThinking {
+		result, _ = sjson.SetBytes(result, "chat_template_kwargs.clear_thinking", false)
+	}
+
+	return result
+}
+
+// applyMiniMax applies thinking configuration for MiniMax models.
+//
+// Output format:
+//
+//	{"reasoning_split": true/false}
+func applyMiniMax(body []byte, config thinking.ThinkingConfig) []byte {
+	reasoningSplit := configToBoolean(config)
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	result, _ := sjson.SetBytes(body, "reasoning_split", reasoningSplit)
+
+	return result
+}
+
+// isGLMModel determines if the model is a GLM series model.
+// GLM models use chat_template_kwargs.enable_thinking format.
+func isGLMModel(modelID string) bool {
+	return strings.HasPrefix(strings.ToLower(modelID), "glm")
+}
+
+// isMiniMaxModel determines if the model is a MiniMax series model.
+// MiniMax models use reasoning_split format.
+func isMiniMaxModel(modelID string) bool {
+	return strings.HasPrefix(strings.ToLower(modelID), "minimax")
+}

internal/thinking/provider/openai/apply.go

@@ -0,0 +1,128 @@
+// Package openai implements thinking configuration for OpenAI/Codex models.
+//
+// OpenAI models use the reasoning_effort format with discrete levels
+// (low/medium/high). Some models support xhigh and none levels.
+// See: _bmad-output/planning-artifacts/architecture.md#Epic-8
+package openai
+
+import (
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier implements thinking.ProviderApplier for OpenAI models.
+//
+// OpenAI-specific behavior:
+//   - Output format: reasoning_effort (string: low/medium/high/xhigh)
+//   - Level-only mode: no numeric budget support
+//   - Some models support ZeroAllowed (gpt-5.1, gpt-5.2)
+type Applier struct{}
+
+var _ thinking.ProviderApplier = (*Applier)(nil)
+
+// NewApplier creates a new OpenAI thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("openai", NewApplier())
+}
+
+// Apply applies thinking configuration to OpenAI request body.
+//
+// Expected output format:
+//
+//	{
+//	  "reasoning_effort": "high"
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return applyCompatibleOpenAI(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	// Only handle ModeLevel and ModeNone; other modes pass through unchanged.
+	if config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	if config.Mode == thinking.ModeLevel {
+		result, _ := sjson.SetBytes(body, "reasoning_effort", string(config.Level))
+		return result, nil
+	}
+
+	effort := ""
+	support := modelInfo.Thinking
+	if config.Budget == 0 {
+		if support.ZeroAllowed || hasLevel(support.Levels, string(thinking.LevelNone)) {
+			effort = string(thinking.LevelNone)
+		}
+	}
+	if effort == "" && config.Level != "" {
+		effort = string(config.Level)
+	}
+	if effort == "" && len(support.Levels) > 0 {
+		effort = support.Levels[0]
+	}
+	if effort == "" {
+		return body, nil
+	}
+
+	result, _ := sjson.SetBytes(body, "reasoning_effort", effort)
+	return result, nil
+}
+
+func applyCompatibleOpenAI(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	var effort string
+	switch config.Mode {
+	case thinking.ModeLevel:
+		if config.Level == "" {
+			return body, nil
+		}
+		effort = string(config.Level)
+	case thinking.ModeNone:
+		effort = string(thinking.LevelNone)
+		if config.Level != "" {
+			effort = string(config.Level)
+		}
+	case thinking.ModeAuto:
+		// Auto mode for user-defined models: pass through as "auto"
+		effort = string(thinking.LevelAuto)
+	case thinking.ModeBudget:
+		// Budget mode: convert budget to level using threshold mapping
+		level, ok := thinking.ConvertBudgetToLevel(config.Budget)
+		if !ok {
+			return body, nil
+		}
+		effort = level
+	default:
+		return body, nil
+	}
+
+	result, _ := sjson.SetBytes(body, "reasoning_effort", effort)
+	return result, nil
+}
+
+func hasLevel(levels []string, target string) bool {
+	for _, level := range levels {
+		if strings.EqualFold(strings.TrimSpace(level), target) {
+			return true
+		}
+	}
+	return false
+}

internal/thinking/strip.go

@@ -0,0 +1,54 @@
+// Package thinking provides unified thinking configuration processing.
+package thinking
+
+import (
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// StripThinkingConfig removes thinking configuration fields from request body.
+//
+// This function is used when a model doesn't support thinking but the request
+// contains thinking configuration. The configuration is silently removed to
+// prevent upstream API errors.
+//
+// Parameters:
+//   - body: Original request body JSON
+//   - provider: Provider name (determines which fields to strip)
+//
+// Returns:
+//   - Modified request body JSON with thinking configuration removed
+//   - Original body is returned unchanged if:
+//   - body is empty or invalid JSON
+//   - provider is unknown
+//   - no thinking configuration found
+func StripThinkingConfig(body []byte, provider string) []byte {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		return body
+	}
+
+	switch provider {
+	case "claude":
+		result, _ := sjson.DeleteBytes(body, "thinking")
+		return result
+	case "gemini":
+		result, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig")
+		return result
+	case "gemini-cli", "antigravity":
+		result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig")
+		return result
+	case "openai":
+		result, _ := sjson.DeleteBytes(body, "reasoning_effort")
+		return result
+	case "codex":
+		result, _ := sjson.DeleteBytes(body, "reasoning.effort")
+		return result
+	case "iflow":
+		result, _ := sjson.DeleteBytes(body, "chat_template_kwargs.enable_thinking")
+		result, _ = sjson.DeleteBytes(result, "chat_template_kwargs.clear_thinking")
+		result, _ = sjson.DeleteBytes(result, "reasoning_split")
+		return result
+	default:
+		return body
+	}
+}

internal/thinking/suffix.go

@@ -0,0 +1,146 @@
+// Package thinking provides unified thinking configuration processing.
+//
+// This file implements suffix parsing functionality for extracting
+// thinking configuration from model names in the format model(value).
+package thinking
+
+import (
+	"strconv"
+	"strings"
+)
+
+// ParseSuffix extracts thinking suffix from a model name.
+//
+// The suffix format is: model-name(value)
+// Examples:
+//   - "claude-sonnet-4-5(16384)" -> ModelName="claude-sonnet-4-5", RawSuffix="16384"
+//   - "gpt-5.2(high)" -> ModelName="gpt-5.2", RawSuffix="high"
+//   - "gemini-2.5-pro" -> ModelName="gemini-2.5-pro", HasSuffix=false
+//
+// This function only extracts the suffix; it does not validate or interpret
+// the suffix content. Use ParseNumericSuffix, ParseLevelSuffix, etc. for
+// content interpretation.
+func ParseSuffix(model string) SuffixResult {
+	// Find the last opening parenthesis
+	lastOpen := strings.LastIndex(model, "(")
+	if lastOpen == -1 {
+		return SuffixResult{ModelName: model, HasSuffix: false}
+	}
+
+	// Check if the string ends with a closing parenthesis
+	if !strings.HasSuffix(model, ")") {
+		return SuffixResult{ModelName: model, HasSuffix: false}
+	}
+
+	// Extract components
+	modelName := model[:lastOpen]
+	rawSuffix := model[lastOpen+1 : len(model)-1]
+
+	return SuffixResult{
+		ModelName: modelName,
+		HasSuffix: true,
+		RawSuffix: rawSuffix,
+	}
+}
+
+// ParseNumericSuffix attempts to parse a raw suffix as a numeric budget value.
+//
+// This function parses the raw suffix content (from ParseSuffix.RawSuffix) as an integer.
+// Only non-negative integers are considered valid numeric suffixes.
+//
+// Platform note: The budget value uses Go's int type, which is 32-bit on 32-bit
+// systems and 64-bit on 64-bit systems. Values exceeding the platform's int range
+// will return ok=false.
+//
+// Leading zeros are accepted: "08192" parses as 8192.
+//
+// Examples:
+//   - "8192" -> budget=8192, ok=true
+//   - "0" -> budget=0, ok=true (represents ModeNone)
+//   - "08192" -> budget=8192, ok=true (leading zeros accepted)
+//   - "-1" -> budget=0, ok=false (negative numbers are not valid numeric suffixes)
+//   - "high" -> budget=0, ok=false (not a number)
+//   - "9223372036854775808" -> budget=0, ok=false (overflow on 64-bit systems)
+//
+// For special handling of -1 as auto mode, use ParseSpecialSuffix instead.
+func ParseNumericSuffix(rawSuffix string) (budget int, ok bool) {
+	if rawSuffix == "" {
+		return 0, false
+	}
+
+	value, err := strconv.Atoi(rawSuffix)
+	if err != nil {
+		return 0, false
+	}
+
+	// Negative numbers are not valid numeric suffixes
+	// -1 should be handled by special value parsing as "auto"
+	if value < 0 {
+		return 0, false
+	}
+
+	return value, true
+}
+
+// ParseSpecialSuffix attempts to parse a raw suffix as a special thinking mode value.
+//
+// This function handles special strings that represent a change in thinking mode:
+//   - "none" -> ModeNone (disables thinking)
+//   - "auto" -> ModeAuto (automatic/dynamic thinking)
+//   - "-1"   -> ModeAuto (numeric representation of auto mode)
+//
+// String values are case-insensitive.
+func ParseSpecialSuffix(rawSuffix string) (mode ThinkingMode, ok bool) {
+	if rawSuffix == "" {
+		return ModeBudget, false
+	}
+
+	// Case-insensitive matching
+	switch strings.ToLower(rawSuffix) {
+	case "none":
+		return ModeNone, true
+	case "auto", "-1":
+		return ModeAuto, true
+	default:
+		return ModeBudget, false
+	}
+}
+
+// ParseLevelSuffix attempts to parse a raw suffix as a discrete thinking level.
+//
+// This function parses the raw suffix content (from ParseSuffix.RawSuffix) as a level.
+// Only discrete effort levels are valid: minimal, low, medium, high, xhigh.
+// Level matching is case-insensitive.
+//
+// Special values (none, auto) are NOT handled by this function; use ParseSpecialSuffix
+// instead. This separation allows callers to prioritize special value handling.
+//
+// Examples:
+//   - "high" -> level=LevelHigh, ok=true
+//   - "HIGH" -> level=LevelHigh, ok=true (case insensitive)
+//   - "medium" -> level=LevelMedium, ok=true
+//   - "none" -> level="", ok=false (special value, use ParseSpecialSuffix)
+//   - "auto" -> level="", ok=false (special value, use ParseSpecialSuffix)
+//   - "8192" -> level="", ok=false (numeric, use ParseNumericSuffix)
+//   - "ultra" -> level="", ok=false (unknown level)
+func ParseLevelSuffix(rawSuffix string) (level ThinkingLevel, ok bool) {
+	if rawSuffix == "" {
+		return "", false
+	}
+
+	// Case-insensitive matching
+	switch strings.ToLower(rawSuffix) {
+	case "minimal":
+		return LevelMinimal, true
+	case "low":
+		return LevelLow, true
+	case "medium":
+		return LevelMedium, true
+	case "high":
+		return LevelHigh, true
+	case "xhigh":
+		return LevelXHigh, true
+	default:
+		return "", false
+	}
+}

internal/thinking/text.go

@@ -0,0 +1,41 @@
+package thinking
+
+import (
+	"github.com/tidwall/gjson"
+)
+
+// GetThinkingText extracts the thinking text from a content part.
+// Handles various formats:
+// - Simple string: { "thinking": "text" } or { "text": "text" }
+// - Wrapped object: { "thinking": { "text": "text", "cache_control": {...} } }
+// - Gemini-style: { "thought": true, "text": "text" }
+// Returns the extracted text string.
+func GetThinkingText(part gjson.Result) string {
+	// Try direct text field first (Gemini-style)
+	if text := part.Get("text"); text.Exists() && text.Type == gjson.String {
+		return text.String()
+	}
+
+	// Try thinking field
+	thinkingField := part.Get("thinking")
+	if !thinkingField.Exists() {
+		return ""
+	}
+
+	// thinking is a string
+	if thinkingField.Type == gjson.String {
+		return thinkingField.String()
+	}
+
+	// thinking is an object with inner text/thinking
+	if thinkingField.IsObject() {
+		if inner := thinkingField.Get("text"); inner.Exists() && inner.Type == gjson.String {
+			return inner.String()
+		}
+		if inner := thinkingField.Get("thinking"); inner.Exists() && inner.Type == gjson.String {
+			return inner.String()
+		}
+	}
+
+	return ""
+}

internal/thinking/types.go

@@ -0,0 +1,116 @@
+// Package thinking provides unified thinking configuration processing.
+//
+// This package offers a unified interface for parsing, validating, and applying
+// thinking configurations across various AI providers (Claude, Gemini, OpenAI, iFlow).
+package thinking
+
+import "github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+
+// ThinkingMode represents the type of thinking configuration mode.
+type ThinkingMode int
+
+const (
+	// ModeBudget indicates using a numeric budget (corresponds to suffix "(1000)" etc.)
+	ModeBudget ThinkingMode = iota
+	// ModeLevel indicates using a discrete level (corresponds to suffix "(high)" etc.)
+	ModeLevel
+	// ModeNone indicates thinking is disabled (corresponds to suffix "(none)" or budget=0)
+	ModeNone
+	// ModeAuto indicates automatic/dynamic thinking (corresponds to suffix "(auto)" or budget=-1)
+	ModeAuto
+)
+
+// String returns the string representation of ThinkingMode.
+func (m ThinkingMode) String() string {
+	switch m {
+	case ModeBudget:
+		return "budget"
+	case ModeLevel:
+		return "level"
+	case ModeNone:
+		return "none"
+	case ModeAuto:
+		return "auto"
+	default:
+		return "unknown"
+	}
+}
+
+// ThinkingLevel represents a discrete thinking level.
+type ThinkingLevel string
+
+const (
+	// LevelNone disables thinking
+	LevelNone ThinkingLevel = "none"
+	// LevelAuto enables automatic/dynamic thinking
+	LevelAuto ThinkingLevel = "auto"
+	// LevelMinimal sets minimal thinking effort
+	LevelMinimal ThinkingLevel = "minimal"
+	// LevelLow sets low thinking effort
+	LevelLow ThinkingLevel = "low"
+	// LevelMedium sets medium thinking effort
+	LevelMedium ThinkingLevel = "medium"
+	// LevelHigh sets high thinking effort
+	LevelHigh ThinkingLevel = "high"
+	// LevelXHigh sets extra-high thinking effort
+	LevelXHigh ThinkingLevel = "xhigh"
+)
+
+// ThinkingConfig represents a unified thinking configuration.
+//
+// This struct is used to pass thinking configuration information between components.
+// Depending on Mode, either Budget or Level field is effective:
+//   - ModeNone: Budget=0, Level is ignored
+//   - ModeAuto: Budget=-1, Level is ignored
+//   - ModeBudget: Budget is a positive integer, Level is ignored
+//   - ModeLevel: Budget is ignored, Level is a valid level
+type ThinkingConfig struct {
+	// Mode specifies the configuration mode
+	Mode ThinkingMode
+	// Budget is the thinking budget (token count), only effective when Mode is ModeBudget.
+	// Special values: 0 means disabled, -1 means automatic
+	Budget int
+	// Level is the thinking level, only effective when Mode is ModeLevel
+	Level ThinkingLevel
+}
+
+// SuffixResult represents the result of parsing a model name for thinking suffix.
+//
+// A thinking suffix is specified in the format model-name(value), where value
+// can be a numeric budget (e.g., "16384") or a level name (e.g., "high").
+type SuffixResult struct {
+	// ModelName is the model name with the suffix removed.
+	// If no suffix was found, this equals the original input.
+	ModelName string
+
+	// HasSuffix indicates whether a valid suffix was found.
+	HasSuffix bool
+
+	// RawSuffix is the content inside the parentheses, without the parentheses.
+	// Empty string if HasSuffix is false.
+	RawSuffix string
+}
+
+// ProviderApplier defines the interface for provider-specific thinking configuration application.
+//
+// Types implementing this interface are responsible for converting a unified ThinkingConfig
+// into provider-specific format and applying it to the request body.
+//
+// Implementation requirements:
+//   - Apply method must be idempotent
+//   - Must not modify the input config or modelInfo
+//   - Returns a modified copy of the request body
+//   - Returns appropriate ThinkingError for unsupported configurations
+type ProviderApplier interface {
+	// Apply applies the thinking configuration to the request body.
+	//
+	// Parameters:
+	//   - body: Original request body JSON
+	//   - config: Unified thinking configuration
+	//   - modelInfo: Model registry information containing ThinkingSupport properties
+	//
+	// Returns:
+	//   - Modified request body JSON
+	//   - ThinkingError if the configuration is invalid or unsupported
+	Apply(body []byte, config ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error)
+}

internal/thinking/validate.go

@@ -0,0 +1,259 @@
+// Package thinking provides unified thinking configuration processing logic.
+package thinking
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	log "github.com/sirupsen/logrus"
+)
+
+// ClampBudget clamps a budget value to the model's supported range.
+//
+// Logging:
+//   - Warn when value=0 but ZeroAllowed=false
+//   - Debug when value is clamped to min/max
+//
+// Fields: provider, model, original_value, clamped_to, min, max
+func ClampBudget(value int, modelInfo *registry.ModelInfo, provider string) int {
+	model := "unknown"
+	support := (*registry.ThinkingSupport)(nil)
+	if modelInfo != nil {
+		if modelInfo.ID != "" {
+			model = modelInfo.ID
+		}
+		support = modelInfo.Thinking
+	}
+	if support == nil {
+		return value
+	}
+
+	// Auto value (-1) passes through without clamping.
+	if value == -1 {
+		return value
+	}
+
+	min := support.Min
+	max := support.Max
+	if value == 0 && !support.ZeroAllowed {
+		log.WithFields(log.Fields{
+			"provider":       provider,
+			"model":          model,
+			"original_value": value,
+			"clamped_to":     min,
+			"min":            min,
+			"max":            max,
+		}).Warn("thinking: budget zero not allowed |")
+		return min
+	}
+
+	// Some models are level-only and do not define numeric budget ranges.
+	if min == 0 && max == 0 {
+		return value
+	}
+
+	if value < min {
+		if value == 0 && support.ZeroAllowed {
+			return 0
+		}
+		logClamp(provider, model, value, min, min, max)
+		return min
+	}
+	if value > max {
+		logClamp(provider, model, value, max, min, max)
+		return max
+	}
+	return value
+}
+
+// ValidateConfig validates a thinking configuration against model capabilities.
+//
+// This function performs comprehensive validation:
+//   - Checks if the model supports thinking
+//   - Auto-converts between Budget and Level formats based on model capability
+//   - Validates that requested level is in the model's supported levels list
+//   - Clamps budget values to model's allowed range
+//
+// Parameters:
+//   - config: The thinking configuration to validate
+//   - support: Model's ThinkingSupport properties (nil means no thinking support)
+//
+// Returns:
+//   - Normalized ThinkingConfig with clamped values
+//   - ThinkingError if validation fails (ErrThinkingNotSupported, ErrLevelNotSupported, etc.)
+//
+// Auto-conversion behavior:
+//   - Budget-only model + Level config → Level converted to Budget
+//   - Level-only model + Budget config → Budget converted to Level
+//   - Hybrid model → preserve original format
+func ValidateConfig(config ThinkingConfig, modelInfo *registry.ModelInfo, provider string) (*ThinkingConfig, error) {
+	normalized := config
+
+	model := "unknown"
+	support := (*registry.ThinkingSupport)(nil)
+	if modelInfo != nil {
+		if modelInfo.ID != "" {
+			model = modelInfo.ID
+		}
+		support = modelInfo.Thinking
+	}
+
+	if support == nil {
+		if config.Mode != ModeNone {
+			return nil, NewThinkingErrorWithModel(ErrThinkingNotSupported, "thinking not supported for this model", model)
+		}
+		return &normalized, nil
+	}
+
+	capability := detectModelCapability(modelInfo)
+	switch capability {
+	case CapabilityBudgetOnly:
+		if normalized.Mode == ModeLevel {
+			if normalized.Level == LevelAuto {
+				break
+			}
+			budget, ok := ConvertLevelToBudget(string(normalized.Level))
+			if !ok {
+				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("unknown level: %s", normalized.Level))
+			}
+			normalized.Mode = ModeBudget
+			normalized.Budget = budget
+			normalized.Level = ""
+		}
+	case CapabilityLevelOnly:
+		if normalized.Mode == ModeBudget {
+			level, ok := ConvertBudgetToLevel(normalized.Budget)
+			if !ok {
+				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("budget %d cannot be converted to a valid level", normalized.Budget))
+			}
+			normalized.Mode = ModeLevel
+			normalized.Level = ThinkingLevel(level)
+			normalized.Budget = 0
+		}
+	case CapabilityHybrid:
+	}
+
+	if normalized.Mode == ModeLevel && normalized.Level == LevelNone {
+		normalized.Mode = ModeNone
+		normalized.Budget = 0
+		normalized.Level = ""
+	}
+	if normalized.Mode == ModeLevel && normalized.Level == LevelAuto {
+		normalized.Mode = ModeAuto
+		normalized.Budget = -1
+		normalized.Level = ""
+	}
+	if normalized.Mode == ModeBudget && normalized.Budget == 0 {
+		normalized.Mode = ModeNone
+		normalized.Level = ""
+	}
+
+	if len(support.Levels) > 0 && normalized.Mode == ModeLevel {
+		if !isLevelSupported(string(normalized.Level), support.Levels) {
+			validLevels := normalizeLevels(support.Levels)
+			message := fmt.Sprintf("level %q not supported, valid levels: %s", strings.ToLower(string(normalized.Level)), strings.Join(validLevels, ", "))
+			return nil, NewThinkingError(ErrLevelNotSupported, message)
+		}
+	}
+
+	// Convert ModeAuto to mid-range if dynamic not allowed
+	if normalized.Mode == ModeAuto && !support.DynamicAllowed {
+		normalized = convertAutoToMidRange(normalized, support, provider, model)
+	}
+
+	if normalized.Mode == ModeNone && provider == "claude" {
+		// Claude supports explicit disable via thinking.type="disabled".
+		// Keep Budget=0 so applier can omit budget_tokens.
+		normalized.Budget = 0
+		normalized.Level = ""
+	} else {
+		switch normalized.Mode {
+		case ModeBudget, ModeAuto, ModeNone:
+			normalized.Budget = ClampBudget(normalized.Budget, modelInfo, provider)
+		}
+
+		// ModeNone with clamped Budget > 0: set Level to lowest for Level-only/Hybrid models
+		// This ensures Apply layer doesn't need to access support.Levels
+		if normalized.Mode == ModeNone && normalized.Budget > 0 && len(support.Levels) > 0 {
+			normalized.Level = ThinkingLevel(support.Levels[0])
+		}
+	}
+
+	return &normalized, nil
+}
+
+func isLevelSupported(level string, supported []string) bool {
+	for _, candidate := range supported {
+		if strings.EqualFold(level, strings.TrimSpace(candidate)) {
+			return true
+		}
+	}
+	return false
+}
+
+func normalizeLevels(levels []string) []string {
+	normalized := make([]string, 0, len(levels))
+	for _, level := range levels {
+		normalized = append(normalized, strings.ToLower(strings.TrimSpace(level)))
+	}
+	return normalized
+}
+
+// convertAutoToMidRange converts ModeAuto to a mid-range value when dynamic is not allowed.
+//
+// This function handles the case where a model does not support dynamic/auto thinking.
+// The auto mode is silently converted to a fixed value based on model capability:
+//   - Level-only models: convert to ModeLevel with LevelMedium
+//   - Budget models: convert to ModeBudget with mid = (Min + Max) / 2
+//
+// Logging:
+//   - Debug level when conversion occurs
+//   - Fields: original_mode, clamped_to, reason
+func convertAutoToMidRange(config ThinkingConfig, support *registry.ThinkingSupport, provider, model string) ThinkingConfig {
+	// For level-only models (has Levels but no Min/Max range), use ModeLevel with medium
+	if len(support.Levels) > 0 && support.Min == 0 && support.Max == 0 {
+		config.Mode = ModeLevel
+		config.Level = LevelMedium
+		config.Budget = 0
+		log.WithFields(log.Fields{
+			"provider":      provider,
+			"model":         model,
+			"original_mode": "auto",
+			"clamped_to":    string(LevelMedium),
+		}).Debug("thinking: mode converted, dynamic not allowed, using medium level |")
+		return config
+	}
+
+	// For budget models, use mid-range budget
+	mid := (support.Min + support.Max) / 2
+	if mid <= 0 && support.ZeroAllowed {
+		config.Mode = ModeNone
+		config.Budget = 0
+	} else if mid <= 0 {
+		config.Mode = ModeBudget
+		config.Budget = support.Min
+	} else {
+		config.Mode = ModeBudget
+		config.Budget = mid
+	}
+	log.WithFields(log.Fields{
+		"provider":      provider,
+		"model":         model,
+		"original_mode": "auto",
+		"clamped_to":    config.Budget,
+	}).Debug("thinking: mode converted, dynamic not allowed |")
+	return config
+}
+
+// logClamp logs a debug message when budget clamping occurs.
+func logClamp(provider, model string, original, clampedTo, min, max int) {
+	log.WithFields(log.Fields{
+		"provider":       provider,
+		"model":          model,
+		"original_value": original,
+		"min":            min,
+		"max":            max,
+		"clamped_to":     clampedTo,
+	}).Debug("thinking: budget clamped |")
+}

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -7,15 +7,41 @@ package claude
 
 import (
 	"bytes"
+	"crypto/sha256"
+	"encoding/hex"
 	"strings"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/cache"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
 
-const geminiCLIClaudeThoughtSignature = "skip_thought_signature_validator"
+// deriveSessionID generates a stable session ID from the request.
+// Uses the hash of the first user message to identify the conversation.
+func deriveSessionID(rawJSON []byte) string {
+	messages := gjson.GetBytes(rawJSON, "messages")
+	if !messages.IsArray() {
+		return ""
+	}
+	for _, msg := range messages.Array() {
+		if msg.Get("role").String() == "user" {
+			content := msg.Get("content").String()
+			if content == "" {
+				// Try to get text from content array
+				content = msg.Get("content.0.text").String()
+			}
+			if content != "" {
+				h := sha256.Sum256([]byte(content))
+				return hex.EncodeToString(h[:16])
+			}
+		}
+	}
+	return ""
+}
 
 // ConvertClaudeRequestToAntigravity parses and transforms a Claude Code API request into Gemini CLI API format.
 // It extracts the model name, system instruction, message contents, and tool declarations
@@ -37,7 +63,9 @@ const geminiCLIClaudeThoughtSignature = "skip_thought_signature_validator"
 //   - []byte: The transformed request data in Gemini CLI API format
 func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _ bool) []byte {
 	rawJSON := bytes.Clone(inputRawJSON)
-	rawJSON = bytes.Replace(rawJSON, []byte(`"url":{"type":"string","format":"uri",`), []byte(`"url":{"type":"string",`), -1)
+
+	// Derive session ID for signature caching
+	sessionID := deriveSessionID(rawJSON)
 
 	// system instruction
 	systemInstructionJSON := ""
@@ -59,21 +87,28 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 				hasSystemInstruction = true
 			}
 		}
+	} else if systemResult.Type == gjson.String {
+		systemInstructionJSON = `{"role":"user","parts":[{"text":""}]}`
+		systemInstructionJSON, _ = sjson.Set(systemInstructionJSON, "parts.0.text", systemResult.String())
+		hasSystemInstruction = true
 	}
 
 	// contents
 	contentsJSON := "[]"
 	hasContents := false
+
 	messagesResult := gjson.GetBytes(rawJSON, "messages")
 	if messagesResult.IsArray() {
 		messageResults := messagesResult.Array()
-		for i := 0; i < len(messageResults); i++ {
+		numMessages := len(messageResults)
+		for i := 0; i < numMessages; i++ {
 			messageResult := messageResults[i]
 			roleResult := messageResult.Get("role")
 			if roleResult.Type != gjson.String {
 				continue
 			}
-			role := roleResult.String()
+			originalRole := roleResult.String()
+			role := originalRole
 			if role == "assistant" {
 				role = "model"
 			}
@@ -82,20 +117,57 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 			contentsResult := messageResult.Get("content")
 			if contentsResult.IsArray() {
 				contentResults := contentsResult.Array()
-				for j := 0; j < len(contentResults); j++ {
+				numContents := len(contentResults)
+				var currentMessageThinkingSignature string
+				for j := 0; j < numContents; j++ {
 					contentResult := contentResults[j]
 					contentTypeResult := contentResult.Get("type")
 					if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "thinking" {
-						prompt := contentResult.Get("thinking").String()
+						// Use GetThinkingText to handle wrapped thinking objects
+						thinkingText := thinking.GetThinkingText(contentResult)
 						signatureResult := contentResult.Get("signature")
-						signature := geminiCLIClaudeThoughtSignature
-						if signatureResult.Exists() {
-							signature = signatureResult.String()
+						clientSignature := ""
+						if signatureResult.Exists() && signatureResult.String() != "" {
+							clientSignature = signatureResult.String()
 						}
+
+						// Always try cached signature first (more reliable than client-provided)
+						// Client may send stale or invalid signatures from different sessions
+						signature := ""
+						if sessionID != "" && thinkingText != "" {
+							if cachedSig := cache.GetCachedSignature(sessionID, thinkingText); cachedSig != "" {
+								signature = cachedSig
+								// log.Debugf("Using cached signature for thinking block")
+							}
+						}
+
+						// Fallback to client signature only if cache miss and client signature is valid
+						if signature == "" && cache.HasValidSignature(clientSignature) {
+							signature = clientSignature
+							// log.Debugf("Using client-provided signature for thinking block")
+						}
+
+						// Store for subsequent tool_use in the same message
+						if cache.HasValidSignature(signature) {
+							currentMessageThinkingSignature = signature
+						}
+
+						// Skip trailing unsigned thinking blocks on last assistant message
+						isUnsigned := !cache.HasValidSignature(signature)
+
+						// If unsigned, skip entirely (don't convert to text)
+						// Claude requires assistant messages to start with thinking blocks when thinking is enabled
+						// Converting to text would break this requirement
+						if isUnsigned {
+							// log.Debugf("Dropping unsigned thinking block (no valid signature)")
+							continue
+						}
+
+						// Valid signature, send as thought block
 						partJSON := `{}`
 						partJSON, _ = sjson.Set(partJSON, "thought", true)
-						if prompt != "" {
-							partJSON, _ = sjson.Set(partJSON, "text", prompt)
+						if thinkingText != "" {
+							partJSON, _ = sjson.Set(partJSON, "text", thinkingText)
 						}
 						if signature != "" {
 							partJSON, _ = sjson.Set(partJSON, "thoughtSignature", signature)
@@ -109,24 +181,46 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 						}
 						clientContentJSON, _ = sjson.SetRaw(clientContentJSON, "parts.-1", partJSON)
 					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "tool_use" {
+						// NOTE: Do NOT inject dummy thinking blocks here.
+						// Antigravity API validates signatures, so dummy values are rejected.
+
 						functionName := contentResult.Get("name").String()
-						functionArgs := contentResult.Get("input").String()
+						argsResult := contentResult.Get("input")
 						functionID := contentResult.Get("id").String()
-						if gjson.Valid(functionArgs) {
-							argsResult := gjson.Parse(functionArgs)
-							if argsResult.IsObject() {
-								partJSON := `{}`
-								if !strings.Contains(modelName, "claude") {
-									partJSON, _ = sjson.Set(partJSON, "thoughtSignature", geminiCLIClaudeThoughtSignature)
-								}
-								if functionID != "" {
-									partJSON, _ = sjson.Set(partJSON, "functionCall.id", functionID)
-								}
-								partJSON, _ = sjson.Set(partJSON, "functionCall.name", functionName)
-								partJSON, _ = sjson.SetRaw(partJSON, "functionCall.args", argsResult.Raw)
-								clientContentJSON, _ = sjson.SetRaw(clientContentJSON, "parts.-1", partJSON)
+
+						// Handle both object and string input formats
+						var argsRaw string
+						if argsResult.IsObject() {
+							argsRaw = argsResult.Raw
+						} else if argsResult.Type == gjson.String {
+							// Input is a JSON string, parse and validate it
+							parsed := gjson.Parse(argsResult.String())
+							if parsed.IsObject() {
+								argsRaw = parsed.Raw
 							}
 						}
+
+						if argsRaw != "" {
+							partJSON := `{}`
+
+							// Use skip_thought_signature_validator for tool calls without valid thinking signature
+							// This is the approach used in opencode-google-antigravity-auth for Gemini
+							// and also works for Claude through Antigravity API
+							const skipSentinel = "skip_thought_signature_validator"
+							if cache.HasValidSignature(currentMessageThinkingSignature) {
+								partJSON, _ = sjson.Set(partJSON, "thoughtSignature", currentMessageThinkingSignature)
+							} else {
+								// No valid signature - use skip sentinel to bypass validation
+								partJSON, _ = sjson.Set(partJSON, "thoughtSignature", skipSentinel)
+							}
+
+							if functionID != "" {
+								partJSON, _ = sjson.Set(partJSON, "functionCall.id", functionID)
+							}
+							partJSON, _ = sjson.Set(partJSON, "functionCall.name", functionName)
+							partJSON, _ = sjson.SetRaw(partJSON, "functionCall.args", argsRaw)
+							clientContentJSON, _ = sjson.SetRaw(clientContentJSON, "parts.-1", partJSON)
+						}
 					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "tool_result" {
 						toolCallID := contentResult.Get("tool_use_id").String()
 						if toolCallID != "" {
@@ -180,6 +274,37 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 						}
 					}
 				}
+
+				// Reorder parts for 'model' role to ensure thinking block is first
+				if role == "model" {
+					partsResult := gjson.Get(clientContentJSON, "parts")
+					if partsResult.IsArray() {
+						parts := partsResult.Array()
+						var thinkingParts []gjson.Result
+						var otherParts []gjson.Result
+						for _, part := range parts {
+							if part.Get("thought").Bool() {
+								thinkingParts = append(thinkingParts, part)
+							} else {
+								otherParts = append(otherParts, part)
+							}
+						}
+						if len(thinkingParts) > 0 {
+							firstPartIsThinking := parts[0].Get("thought").Bool()
+							if !firstPartIsThinking || len(thinkingParts) > 1 {
+								var newParts []interface{}
+								for _, p := range thinkingParts {
+									newParts = append(newParts, p.Value())
+								}
+								for _, p := range otherParts {
+									newParts = append(newParts, p.Value())
+								}
+								clientContentJSON, _ = sjson.Set(clientContentJSON, "parts", newParts)
+							}
+						}
+					}
+				}
+
 				contentsJSON, _ = sjson.SetRaw(contentsJSON, "-1", clientContentJSON)
 				hasContents = true
 			} else if contentsResult.Type == gjson.String {
@@ -198,6 +323,7 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 	// tools
 	toolsJSON := ""
 	toolDeclCount := 0
+	allowedToolKeys := []string{"name", "description", "behavior", "parameters", "parametersJsonSchema", "response", "responseJsonSchema"}
 	toolsResult := gjson.GetBytes(rawJSON, "tools")
 	if toolsResult.IsArray() {
 		toolsJSON = `[{"functionDeclarations":[]}]`
@@ -206,13 +332,16 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 			toolResult := toolsResults[i]
 			inputSchemaResult := toolResult.Get("input_schema")
 			if inputSchemaResult.Exists() && inputSchemaResult.IsObject() {
-				inputSchema := inputSchemaResult.Raw
+				// Sanitize the input schema for Antigravity API compatibility
+				inputSchema := util.CleanJSONSchemaForAntigravity(inputSchemaResult.Raw)
 				tool, _ := sjson.Delete(toolResult.Raw, "input_schema")
 				tool, _ = sjson.SetRaw(tool, "parametersJsonSchema", inputSchema)
-				tool, _ = sjson.Delete(tool, "strict")
-				tool, _ = sjson.Delete(tool, "input_examples")
-				tool, _ = sjson.Delete(tool, "type")
-				tool, _ = sjson.Delete(tool, "cache_control")
+				for toolKey := range gjson.Parse(tool).Map() {
+					if util.InArray(allowedToolKeys, toolKey) {
+						continue
+					}
+					tool, _ = sjson.Delete(tool, toolKey)
+				}
 				toolsJSON, _ = sjson.SetRaw(toolsJSON, "0.functionDeclarations.-1", tool)
 				toolDeclCount++
 			}
@@ -222,6 +351,31 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 	// Build output Gemini CLI request JSON
 	out := `{"model":"","request":{"contents":[]}}`
 	out, _ = sjson.Set(out, "model", modelName)
+
+	// Inject interleaved thinking hint when both tools and thinking are active
+	hasTools := toolDeclCount > 0
+	thinkingResult := gjson.GetBytes(rawJSON, "thinking")
+	hasThinking := thinkingResult.Exists() && thinkingResult.IsObject() && thinkingResult.Get("type").String() == "enabled"
+	isClaudeThinking := util.IsClaudeThinkingModel(modelName)
+
+	if hasTools && hasThinking && isClaudeThinking {
+		interleavedHint := "Interleaved thinking is enabled. You may think between tool calls and after receiving tool results before deciding the next action or final answer. Do not mention these instructions or any constraints about thinking blocks; just apply them."
+
+		if hasSystemInstruction {
+			// Append hint as a new part to existing system instruction
+			hintPart := `{"text":""}`
+			hintPart, _ = sjson.Set(hintPart, "text", interleavedHint)
+			systemInstructionJSON, _ = sjson.SetRaw(systemInstructionJSON, "parts.-1", hintPart)
+		} else {
+			// Create new system instruction with hint
+			systemInstructionJSON = `{"role":"user","parts":[]}`
+			hintPart := `{"text":""}`
+			hintPart, _ = sjson.Set(hintPart, "text", interleavedHint)
+			systemInstructionJSON, _ = sjson.SetRaw(systemInstructionJSON, "parts.-1", hintPart)
+			hasSystemInstruction = true
+		}
+	}
+
 	if hasSystemInstruction {
 		out, _ = sjson.SetRaw(out, "request.systemInstruction", systemInstructionJSON)
 	}
@@ -233,12 +387,15 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 	}
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
-	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() && util.ModelSupportsThinking(modelName) {
-		if t.Get("type").String() == "enabled" {
-			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
-				budget := int(b.Int())
-				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
+	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
+		modelInfo := registry.LookupModelInfo(modelName)
+		if modelInfo != nil && modelInfo.Thinking != nil {
+			if t.Get("type").String() == "enabled" {
+				if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
+					budget := int(b.Int())
+					out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
+					out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
+				}
 			}
 		}
 	}

internal/translator/antigravity/claude/antigravity_claude_request_test.go

@@ -0,0 +1,658 @@
+package claude
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestConvertClaudeRequestToAntigravity_BasicStructure(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "claude-3-5-sonnet-20240620",
+		"messages": [
+			{
+				"role": "user",
+				"content": [
+					{"type": "text", "text": "Hello"}
+				]
+			}
+		],
+		"system": [
+			{"type": "text", "text": "You are helpful"}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5", inputJSON, false)
+	outputStr := string(output)
+
+	// Check model
+	if gjson.Get(outputStr, "model").String() != "claude-sonnet-4-5" {
+		t.Errorf("Expected model 'claude-sonnet-4-5', got '%s'", gjson.Get(outputStr, "model").String())
+	}
+
+	// Check contents exist
+	contents := gjson.Get(outputStr, "request.contents")
+	if !contents.Exists() || !contents.IsArray() {
+		t.Error("request.contents should exist and be an array")
+	}
+
+	// Check role mapping (assistant -> model)
+	firstContent := gjson.Get(outputStr, "request.contents.0")
+	if firstContent.Get("role").String() != "user" {
+		t.Errorf("Expected role 'user', got '%s'", firstContent.Get("role").String())
+	}
+
+	// Check systemInstruction
+	sysInstruction := gjson.Get(outputStr, "request.systemInstruction")
+	if !sysInstruction.Exists() {
+		t.Error("systemInstruction should exist")
+	}
+	if sysInstruction.Get("parts.0.text").String() != "You are helpful" {
+		t.Error("systemInstruction text mismatch")
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_RoleMapping(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "claude-3-5-sonnet-20240620",
+		"messages": [
+			{"role": "user", "content": [{"type": "text", "text": "Hi"}]},
+			{"role": "assistant", "content": [{"type": "text", "text": "Hello"}]}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5", inputJSON, false)
+	outputStr := string(output)
+
+	// assistant should be mapped to model
+	secondContent := gjson.Get(outputStr, "request.contents.1")
+	if secondContent.Get("role").String() != "model" {
+		t.Errorf("Expected role 'model' (mapped from 'assistant'), got '%s'", secondContent.Get("role").String())
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ThinkingBlocks(t *testing.T) {
+	// Valid signature must be at least 50 characters
+	validSignature := "abc123validSignature1234567890123456789012345678901234567890"
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5-thinking",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{"type": "thinking", "thinking": "Let me think...", "signature": "` + validSignature + `"},
+					{"type": "text", "text": "Answer"}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	// Check thinking block conversion
+	firstPart := gjson.Get(outputStr, "request.contents.0.parts.0")
+	if !firstPart.Get("thought").Bool() {
+		t.Error("thinking block should have thought: true")
+	}
+	if firstPart.Get("text").String() != "Let me think..." {
+		t.Error("thinking text mismatch")
+	}
+	if firstPart.Get("thoughtSignature").String() != validSignature {
+		t.Errorf("Expected thoughtSignature '%s', got '%s'", validSignature, firstPart.Get("thoughtSignature").String())
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ThinkingBlockWithoutSignature(t *testing.T) {
+	// Unsigned thinking blocks should be removed entirely (not converted to text)
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5-thinking",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{"type": "thinking", "thinking": "Let me think..."},
+					{"type": "text", "text": "Answer"}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	// Without signature, thinking block should be removed (not converted to text)
+	parts := gjson.Get(outputStr, "request.contents.0.parts").Array()
+	if len(parts) != 1 {
+		t.Fatalf("Expected 1 part (thinking removed), got %d", len(parts))
+	}
+
+	// Only text part should remain
+	if parts[0].Get("thought").Bool() {
+		t.Error("Thinking block should be removed, not preserved")
+	}
+	if parts[0].Get("text").String() != "Answer" {
+		t.Errorf("Expected text 'Answer', got '%s'", parts[0].Get("text").String())
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ToolDeclarations(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "claude-3-5-sonnet-20240620",
+		"messages": [],
+		"tools": [
+			{
+				"name": "test_tool",
+				"description": "A test tool",
+				"input_schema": {
+					"type": "object",
+					"properties": {
+						"name": {"type": "string"}
+					},
+					"required": ["name"]
+				}
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("gemini-1.5-pro", inputJSON, false)
+	outputStr := string(output)
+
+	// Check tools structure
+	tools := gjson.Get(outputStr, "request.tools")
+	if !tools.Exists() {
+		t.Error("Tools should exist in output")
+	}
+
+	funcDecl := gjson.Get(outputStr, "request.tools.0.functionDeclarations.0")
+	if funcDecl.Get("name").String() != "test_tool" {
+		t.Errorf("Expected tool name 'test_tool', got '%s'", funcDecl.Get("name").String())
+	}
+
+	// Check input_schema renamed to parametersJsonSchema
+	if funcDecl.Get("parametersJsonSchema").Exists() {
+		t.Log("parametersJsonSchema exists (expected)")
+	}
+	if funcDecl.Get("input_schema").Exists() {
+		t.Error("input_schema should be removed")
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ToolUse(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "claude-3-5-sonnet-20240620",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{
+						"type": "tool_use",
+						"id": "call_123",
+						"name": "get_weather",
+						"input": "{\"location\": \"Paris\"}"
+					}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5", inputJSON, false)
+	outputStr := string(output)
+
+	// Now we expect only 1 part (tool_use), no dummy thinking block injected
+	parts := gjson.Get(outputStr, "request.contents.0.parts").Array()
+	if len(parts) != 1 {
+		t.Fatalf("Expected 1 part (tool only, no dummy injection), got %d", len(parts))
+	}
+
+	// Check function call conversion at parts[0]
+	funcCall := parts[0].Get("functionCall")
+	if !funcCall.Exists() {
+		t.Error("functionCall should exist at parts[0]")
+	}
+	if funcCall.Get("name").String() != "get_weather" {
+		t.Errorf("Expected function name 'get_weather', got '%s'", funcCall.Get("name").String())
+	}
+	if funcCall.Get("id").String() != "call_123" {
+		t.Errorf("Expected function id 'call_123', got '%s'", funcCall.Get("id").String())
+	}
+	// Verify skip_thought_signature_validator is added (bypass for tools without valid thinking)
+	expectedSig := "skip_thought_signature_validator"
+	actualSig := parts[0].Get("thoughtSignature").String()
+	if actualSig != expectedSig {
+		t.Errorf("Expected thoughtSignature '%s', got '%s'", expectedSig, actualSig)
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ToolUse_WithSignature(t *testing.T) {
+	validSignature := "abc123validSignature1234567890123456789012345678901234567890"
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5-thinking",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{"type": "thinking", "thinking": "Let me think...", "signature": "` + validSignature + `"},
+					{
+						"type": "tool_use",
+						"id": "call_123",
+						"name": "get_weather",
+						"input": "{\"location\": \"Paris\"}"
+					}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	// Check function call has the signature from the preceding thinking block
+	part := gjson.Get(outputStr, "request.contents.0.parts.1")
+	if part.Get("functionCall.name").String() != "get_weather" {
+		t.Errorf("Expected functionCall, got %s", part.Raw)
+	}
+	if part.Get("thoughtSignature").String() != validSignature {
+		t.Errorf("Expected thoughtSignature '%s' on tool_use, got '%s'", validSignature, part.Get("thoughtSignature").String())
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ReorderThinking(t *testing.T) {
+	// Case: text block followed by thinking block -> should be reordered to thinking first
+	validSignature := "abc123validSignature1234567890123456789012345678901234567890"
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5-thinking",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{"type": "text", "text": "Here is the plan."},
+					{"type": "thinking", "thinking": "Planning...", "signature": "` + validSignature + `"}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	// Verify order: Thinking block MUST be first
+	parts := gjson.Get(outputStr, "request.contents.0.parts").Array()
+	if len(parts) != 2 {
+		t.Fatalf("Expected 2 parts, got %d", len(parts))
+	}
+
+	if !parts[0].Get("thought").Bool() {
+		t.Error("First part should be thinking block after reordering")
+	}
+	if parts[1].Get("text").String() != "Here is the plan." {
+		t.Error("Second part should be text block")
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ToolResult(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "claude-3-5-sonnet-20240620",
+		"messages": [
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "tool_result",
+						"tool_use_id": "get_weather-call-123",
+						"content": "22C sunny"
+					}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5", inputJSON, false)
+	outputStr := string(output)
+
+	// Check function response conversion
+	funcResp := gjson.Get(outputStr, "request.contents.0.parts.0.functionResponse")
+	if !funcResp.Exists() {
+		t.Error("functionResponse should exist")
+	}
+	if funcResp.Get("id").String() != "get_weather-call-123" {
+		t.Errorf("Expected function id, got '%s'", funcResp.Get("id").String())
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ThinkingConfig(t *testing.T) {
+	// Note: This test requires the model to be registered in the registry
+	// with Thinking metadata. If the registry is not populated in test environment,
+	// thinkingConfig won't be added. We'll test the basic structure only.
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5-thinking",
+		"messages": [],
+		"thinking": {
+			"type": "enabled",
+			"budget_tokens": 8000
+		}
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	// Check thinking config conversion (only if model supports thinking in registry)
+	thinkingConfig := gjson.Get(outputStr, "request.generationConfig.thinkingConfig")
+	if thinkingConfig.Exists() {
+		if thinkingConfig.Get("thinkingBudget").Int() != 8000 {
+			t.Errorf("Expected thinkingBudget 8000, got %d", thinkingConfig.Get("thinkingBudget").Int())
+		}
+		if !thinkingConfig.Get("include_thoughts").Bool() {
+			t.Error("include_thoughts should be true")
+		}
+	} else {
+		t.Log("thinkingConfig not present - model may not be registered in test registry")
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ImageContent(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "claude-3-5-sonnet-20240620",
+		"messages": [
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "image",
+						"source": {
+							"type": "base64",
+							"media_type": "image/png",
+							"data": "iVBORw0KGgoAAAANSUhEUg=="
+						}
+					}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5", inputJSON, false)
+	outputStr := string(output)
+
+	// Check inline data conversion
+	inlineData := gjson.Get(outputStr, "request.contents.0.parts.0.inlineData")
+	if !inlineData.Exists() {
+		t.Error("inlineData should exist")
+	}
+	if inlineData.Get("mime_type").String() != "image/png" {
+		t.Error("mime_type mismatch")
+	}
+	if !strings.Contains(inlineData.Get("data").String(), "iVBORw0KGgo") {
+		t.Error("data mismatch")
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_GenerationConfig(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "claude-3-5-sonnet-20240620",
+		"messages": [],
+		"temperature": 0.7,
+		"top_p": 0.9,
+		"top_k": 40,
+		"max_tokens": 2000
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5", inputJSON, false)
+	outputStr := string(output)
+
+	genConfig := gjson.Get(outputStr, "request.generationConfig")
+	if genConfig.Get("temperature").Float() != 0.7 {
+		t.Errorf("Expected temperature 0.7, got %f", genConfig.Get("temperature").Float())
+	}
+	if genConfig.Get("topP").Float() != 0.9 {
+		t.Errorf("Expected topP 0.9, got %f", genConfig.Get("topP").Float())
+	}
+	if genConfig.Get("topK").Float() != 40 {
+		t.Errorf("Expected topK 40, got %f", genConfig.Get("topK").Float())
+	}
+	if genConfig.Get("maxOutputTokens").Float() != 2000 {
+		t.Errorf("Expected maxOutputTokens 2000, got %f", genConfig.Get("maxOutputTokens").Float())
+	}
+}
+
+// ============================================================================
+// Trailing Unsigned Thinking Block Removal
+// ============================================================================
+
+func TestConvertClaudeRequestToAntigravity_TrailingUnsignedThinking_Removed(t *testing.T) {
+	// Last assistant message ends with unsigned thinking block - should be removed
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5-thinking",
+		"messages": [
+			{
+				"role": "user",
+				"content": [{"type": "text", "text": "Hello"}]
+			},
+			{
+				"role": "assistant",
+				"content": [
+					{"type": "text", "text": "Here is my answer"},
+					{"type": "thinking", "thinking": "I should think more..."}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	// The last part of the last assistant message should NOT be a thinking block
+	lastMessageParts := gjson.Get(outputStr, "request.contents.1.parts")
+	if !lastMessageParts.IsArray() {
+		t.Fatal("Last message should have parts array")
+	}
+	parts := lastMessageParts.Array()
+	if len(parts) == 0 {
+		t.Fatal("Last message should have at least one part")
+	}
+
+	// The unsigned thinking should be removed, leaving only the text
+	lastPart := parts[len(parts)-1]
+	if lastPart.Get("thought").Bool() {
+		t.Error("Trailing unsigned thinking block should be removed")
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_TrailingSignedThinking_Kept(t *testing.T) {
+	// Last assistant message ends with signed thinking block - should be kept
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5-thinking",
+		"messages": [
+			{
+				"role": "user",
+				"content": [{"type": "text", "text": "Hello"}]
+			},
+			{
+				"role": "assistant",
+				"content": [
+					{"type": "text", "text": "Here is my answer"},
+					{"type": "thinking", "thinking": "Valid thinking...", "signature": "abc123validSignature1234567890123456789012345678901234567890"}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	// The signed thinking block should be preserved
+	lastMessageParts := gjson.Get(outputStr, "request.contents.1.parts")
+	parts := lastMessageParts.Array()
+	if len(parts) < 2 {
+		t.Error("Signed thinking block should be preserved")
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_MiddleUnsignedThinking_Removed(t *testing.T) {
+	// Middle message has unsigned thinking - should be removed entirely
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5-thinking",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{"type": "thinking", "thinking": "Middle thinking..."},
+					{"type": "text", "text": "Answer"}
+				]
+			},
+			{
+				"role": "user",
+				"content": [{"type": "text", "text": "Follow up"}]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	// Unsigned thinking should be removed entirely
+	parts := gjson.Get(outputStr, "request.contents.0.parts").Array()
+	if len(parts) != 1 {
+		t.Fatalf("Expected 1 part (thinking removed), got %d", len(parts))
+	}
+
+	// Only text part should remain
+	if parts[0].Get("thought").Bool() {
+		t.Error("Thinking block should be removed, not preserved")
+	}
+	if parts[0].Get("text").String() != "Answer" {
+		t.Errorf("Expected text 'Answer', got '%s'", parts[0].Get("text").String())
+	}
+}
+
+// ============================================================================
+// Tool + Thinking System Hint Injection
+// ============================================================================
+
+func TestConvertClaudeRequestToAntigravity_ToolAndThinking_HintInjected(t *testing.T) {
+	// When both tools and thinking are enabled, hint should be injected into system instruction
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5-thinking",
+		"messages": [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}],
+		"system": [{"type": "text", "text": "You are helpful."}],
+		"tools": [
+			{
+				"name": "get_weather",
+				"description": "Get weather",
+				"input_schema": {"type": "object", "properties": {"location": {"type": "string"}}}
+			}
+		],
+		"thinking": {"type": "enabled", "budget_tokens": 8000}
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	// System instruction should contain the interleaved thinking hint
+	sysInstruction := gjson.Get(outputStr, "request.systemInstruction")
+	if !sysInstruction.Exists() {
+		t.Fatal("systemInstruction should exist")
+	}
+
+	// Check if hint is appended
+	sysText := sysInstruction.Get("parts").Array()
+	found := false
+	for _, part := range sysText {
+		if strings.Contains(part.Get("text").String(), "Interleaved thinking is enabled") {
+			found = true
+			break
+		}
+	}
+	if !found {
+		t.Errorf("Interleaved thinking hint should be injected when tools and thinking are both active, got: %v", sysInstruction.Raw)
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ToolsOnly_NoHint(t *testing.T) {
+	// When only tools are present (no thinking), hint should NOT be injected
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5",
+		"messages": [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}],
+		"system": [{"type": "text", "text": "You are helpful."}],
+		"tools": [
+			{
+				"name": "get_weather",
+				"description": "Get weather",
+				"input_schema": {"type": "object", "properties": {"location": {"type": "string"}}}
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5", inputJSON, false)
+	outputStr := string(output)
+
+	// System instruction should NOT contain the hint
+	sysInstruction := gjson.Get(outputStr, "request.systemInstruction")
+	if sysInstruction.Exists() {
+		for _, part := range sysInstruction.Get("parts").Array() {
+			if strings.Contains(part.Get("text").String(), "Interleaved thinking is enabled") {
+				t.Error("Hint should NOT be injected when only tools are present (no thinking)")
+			}
+		}
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ThinkingOnly_NoHint(t *testing.T) {
+	// When only thinking is enabled (no tools), hint should NOT be injected
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5-thinking",
+		"messages": [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}],
+		"system": [{"type": "text", "text": "You are helpful."}],
+		"thinking": {"type": "enabled", "budget_tokens": 8000}
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	// System instruction should NOT contain the hint (no tools)
+	sysInstruction := gjson.Get(outputStr, "request.systemInstruction")
+	if sysInstruction.Exists() {
+		for _, part := range sysInstruction.Get("parts").Array() {
+			if strings.Contains(part.Get("text").String(), "Interleaved thinking is enabled") {
+				t.Error("Hint should NOT be injected when only thinking is present (no tools)")
+			}
+		}
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ToolAndThinking_NoExistingSystem(t *testing.T) {
+	// When tools + thinking but no system instruction, should create one with hint
+	inputJSON := []byte(`{
+		"model": "claude-sonnet-4-5-thinking",
+		"messages": [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}],
+		"tools": [
+			{
+				"name": "get_weather",
+				"description": "Get weather",
+				"input_schema": {"type": "object", "properties": {"location": {"type": "string"}}}
+			}
+		],
+		"thinking": {"type": "enabled", "budget_tokens": 8000}
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	// System instruction should be created with hint
+	sysInstruction := gjson.Get(outputStr, "request.systemInstruction")
+	if !sysInstruction.Exists() {
+		t.Fatal("systemInstruction should be created when tools + thinking are active")
+	}
+
+	sysText := sysInstruction.Get("parts").Array()
+	found := false
+	for _, part := range sysText {
+		if strings.Contains(part.Get("text").String(), "Interleaved thinking is enabled") {
+			found = true
+			break
+		}
+	}
+	if !found {
+		t.Errorf("Interleaved thinking hint should be in created systemInstruction, got: %v", sysInstruction.Raw)
+	}
+}

internal/translator/antigravity/claude/antigravity_claude_response.go

@@ -14,6 +14,9 @@ import (
 	"sync/atomic"
 	"time"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/cache"
+	log "github.com/sirupsen/logrus"
+
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -32,9 +35,14 @@ type Params struct {
 	CandidatesTokenCount int64  // Cached candidate token count from usage metadata
 	ThoughtsTokenCount   int64  // Cached thinking token count from usage metadata
 	TotalTokenCount      int64  // Cached total token count from usage metadata
+	CachedTokenCount     int64  // Cached content token count (indicates prompt caching)
 	HasSentFinalEvents   bool   // Indicates if final content/message events have been sent
 	HasToolUse           bool   // Indicates if tool use was observed in the stream
 	HasContent           bool   // Tracks whether any content (text, thinking, or tool use) has been output
+
+	// Signature caching support
+	SessionID           string          // Session ID derived from request for signature caching
+	CurrentThinkingText strings.Builder // Accumulates thinking text for signature caching
 }
 
 // toolUseIDCounter provides a process-wide unique counter for tool use identifiers.
@@ -62,6 +70,7 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 			HasFirstResponse: false,
 			ResponseType:     0,
 			ResponseIndex:    0,
+			SessionID:        deriveSessionID(originalRequestRawJSON),
 		}
 	}
 
@@ -90,6 +99,14 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 		// This follows the Claude Code API specification for streaming message initialization
 		messageStartTemplate := `{"type": "message_start", "message": {"id": "msg_1nZdL29xx5MUA1yADyHTEsnR8uuvGzszyY", "type": "message", "role": "assistant", "content": [], "model": "claude-3-5-sonnet-20241022", "stop_reason": null, "stop_sequence": null, "usage": {"input_tokens": 0, "output_tokens": 0}}}`
 
+		// Use cpaUsageMetadata within the message_start event for Claude.
+		if promptTokenCount := gjson.GetBytes(rawJSON, "response.cpaUsageMetadata.promptTokenCount"); promptTokenCount.Exists() {
+			messageStartTemplate, _ = sjson.Set(messageStartTemplate, "message.usage.input_tokens", promptTokenCount.Int())
+		}
+		if candidatesTokenCount := gjson.GetBytes(rawJSON, "response.cpaUsageMetadata.candidatesTokenCount"); candidatesTokenCount.Exists() {
+			messageStartTemplate, _ = sjson.Set(messageStartTemplate, "message.usage.output_tokens", candidatesTokenCount.Int())
+		}
+
 		// Override default values with actual response metadata if available from the Gemini CLI response
 		if modelVersionResult := gjson.GetBytes(rawJSON, "response.modelVersion"); modelVersionResult.Exists() {
 			messageStartTemplate, _ = sjson.Set(messageStartTemplate, "message.model", modelVersionResult.String())
@@ -119,11 +136,20 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 				// Process thinking content (internal reasoning)
 				if partResult.Get("thought").Bool() {
 					if thoughtSignature := partResult.Get("thoughtSignature"); thoughtSignature.Exists() && thoughtSignature.String() != "" {
+						// log.Debug("Branch: signature_delta")
+
+						if params.SessionID != "" && params.CurrentThinkingText.Len() > 0 {
+							cache.CacheSignature(params.SessionID, params.CurrentThinkingText.String(), thoughtSignature.String())
+							// log.Debugf("Cached signature for thinking block (sessionID=%s, textLen=%d)", params.SessionID, params.CurrentThinkingText.Len())
+							params.CurrentThinkingText.Reset()
+						}
+
 						output = output + "event: content_block_delta\n"
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"signature_delta","signature":""}}`, params.ResponseIndex), "delta.signature", thoughtSignature.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
 						params.HasContent = true
 					} else if params.ResponseType == 2 { // Continue existing thinking block if already in thinking state
+						params.CurrentThinkingText.WriteString(partTextResult.String())
 						output = output + "event: content_block_delta\n"
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"thinking_delta","thinking":""}}`, params.ResponseIndex), "delta.thinking", partTextResult.String())
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
@@ -152,6 +178,9 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
 						params.ResponseType = 2 // Set state to thinking
 						params.HasContent = true
+						// Start accumulating thinking text for signature caching
+						params.CurrentThinkingText.Reset()
+						params.CurrentThinkingText.WriteString(partTextResult.String())
 					}
 				} else {
 					finishReasonResult := gjson.GetBytes(rawJSON, "response.candidates.0.finishReason")
@@ -250,7 +279,8 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 
 	if usageResult := gjson.GetBytes(rawJSON, "response.usageMetadata"); usageResult.Exists() {
 		params.HasUsageMetadata = true
-		params.PromptTokenCount = usageResult.Get("promptTokenCount").Int()
+		params.CachedTokenCount = usageResult.Get("cachedContentTokenCount").Int()
+		params.PromptTokenCount = usageResult.Get("promptTokenCount").Int() - params.CachedTokenCount
 		params.CandidatesTokenCount = usageResult.Get("candidatesTokenCount").Int()
 		params.ThoughtsTokenCount = usageResult.Get("thoughtsTokenCount").Int()
 		params.TotalTokenCount = usageResult.Get("totalTokenCount").Int()
@@ -302,6 +332,14 @@ func appendFinalEvents(params *Params, output *string, force bool) {
 	*output = *output + "event: message_delta\n"
 	*output = *output + "data: "
 	delta := fmt.Sprintf(`{"type":"message_delta","delta":{"stop_reason":"%s","stop_sequence":null},"usage":{"input_tokens":%d,"output_tokens":%d}}`, stopReason, params.PromptTokenCount, usageOutputTokens)
+	// Add cache_read_input_tokens if cached tokens are present (indicates prompt caching is working)
+	if params.CachedTokenCount > 0 {
+		var err error
+		delta, err = sjson.Set(delta, "usage.cache_read_input_tokens", params.CachedTokenCount)
+		if err != nil {
+			log.Warnf("antigravity claude response: failed to set cache_read_input_tokens: %v", err)
+		}
+	}
 	*output = *output + delta + "\n\n\n"
 
 	params.HasSentFinalEvents = true
@@ -341,6 +379,7 @@ func ConvertAntigravityResponseToClaudeNonStream(_ context.Context, _ string, or
 	candidateTokens := root.Get("response.usageMetadata.candidatesTokenCount").Int()
 	thoughtTokens := root.Get("response.usageMetadata.thoughtsTokenCount").Int()
 	totalTokens := root.Get("response.usageMetadata.totalTokenCount").Int()
+	cachedTokens := root.Get("response.usageMetadata.cachedContentTokenCount").Int()
 	outputTokens := candidateTokens + thoughtTokens
 	if outputTokens == 0 && totalTokens > 0 {
 		outputTokens = totalTokens - promptTokens
@@ -354,6 +393,14 @@ func ConvertAntigravityResponseToClaudeNonStream(_ context.Context, _ string, or
 	responseJSON, _ = sjson.Set(responseJSON, "model", root.Get("response.modelVersion").String())
 	responseJSON, _ = sjson.Set(responseJSON, "usage.input_tokens", promptTokens)
 	responseJSON, _ = sjson.Set(responseJSON, "usage.output_tokens", outputTokens)
+	// Add cache_read_input_tokens if cached tokens are present (indicates prompt caching is working)
+	if cachedTokens > 0 {
+		var err error
+		responseJSON, err = sjson.Set(responseJSON, "usage.cache_read_input_tokens", cachedTokens)
+		if err != nil {
+			log.Warnf("antigravity claude response: failed to set cache_read_input_tokens: %v", err)
+		}
+	}
 
 	contentArrayInitialized := false
 	ensureContentArray := func() {
@@ -432,7 +479,7 @@ func ConvertAntigravityResponseToClaudeNonStream(_ context.Context, _ string, or
 				toolBlock, _ = sjson.Set(toolBlock, "id", fmt.Sprintf("tool_%d", toolIDCounter))
 				toolBlock, _ = sjson.Set(toolBlock, "name", name)
 
-				if args := functionCall.Get("args"); args.Exists() && args.Raw != "" && gjson.Valid(args.Raw) {
+				if args := functionCall.Get("args"); args.Exists() && args.Raw != "" && gjson.Valid(args.Raw) && args.IsObject() {
 					toolBlock, _ = sjson.SetRaw(toolBlock, "input", args.Raw)
 				}
 

internal/translator/antigravity/claude/antigravity_claude_response_test.go

@@ -0,0 +1,316 @@
+package claude
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/cache"
+)
+
+// ============================================================================
+// Signature Caching Tests
+// ============================================================================
+
+func TestConvertAntigravityResponseToClaude_SessionIDDerived(t *testing.T) {
+	cache.ClearSignatureCache("")
+
+	// Request with user message - should derive session ID
+	requestJSON := []byte(`{
+		"messages": [
+			{"role": "user", "content": [{"type": "text", "text": "Hello world"}]}
+		]
+	}`)
+
+	// First response chunk with thinking
+	responseJSON := []byte(`{
+		"response": {
+			"candidates": [{
+				"content": {
+					"parts": [{"text": "Let me think...", "thought": true}]
+				}
+			}]
+		}
+	}`)
+
+	var param any
+	ctx := context.Background()
+	ConvertAntigravityResponseToClaude(ctx, "claude-sonnet-4-5-thinking", requestJSON, requestJSON, responseJSON, &param)
+
+	// Verify session ID was set
+	params := param.(*Params)
+	if params.SessionID == "" {
+		t.Error("SessionID should be derived from request")
+	}
+}
+
+func TestConvertAntigravityResponseToClaude_ThinkingTextAccumulated(t *testing.T) {
+	cache.ClearSignatureCache("")
+
+	requestJSON := []byte(`{
+		"messages": [{"role": "user", "content": [{"type": "text", "text": "Test"}]}]
+	}`)
+
+	// First thinking chunk
+	chunk1 := []byte(`{
+		"response": {
+			"candidates": [{
+				"content": {
+					"parts": [{"text": "First part of thinking...", "thought": true}]
+				}
+			}]
+		}
+	}`)
+
+	// Second thinking chunk (continuation)
+	chunk2 := []byte(`{
+		"response": {
+			"candidates": [{
+				"content": {
+					"parts": [{"text": " Second part of thinking...", "thought": true}]
+				}
+			}]
+		}
+	}`)
+
+	var param any
+	ctx := context.Background()
+
+	// Process first chunk - starts new thinking block
+	ConvertAntigravityResponseToClaude(ctx, "claude-sonnet-4-5-thinking", requestJSON, requestJSON, chunk1, &param)
+	params := param.(*Params)
+
+	if params.CurrentThinkingText.Len() == 0 {
+		t.Error("Thinking text should be accumulated after first chunk")
+	}
+
+	// Process second chunk - continues thinking block
+	ConvertAntigravityResponseToClaude(ctx, "claude-sonnet-4-5-thinking", requestJSON, requestJSON, chunk2, &param)
+
+	text := params.CurrentThinkingText.String()
+	if !strings.Contains(text, "First part") || !strings.Contains(text, "Second part") {
+		t.Errorf("Thinking text should accumulate both parts, got: %s", text)
+	}
+}
+
+func TestConvertAntigravityResponseToClaude_SignatureCached(t *testing.T) {
+	cache.ClearSignatureCache("")
+
+	requestJSON := []byte(`{
+		"messages": [{"role": "user", "content": [{"type": "text", "text": "Cache test"}]}]
+	}`)
+
+	// Thinking chunk
+	thinkingChunk := []byte(`{
+		"response": {
+			"candidates": [{
+				"content": {
+					"parts": [{"text": "My thinking process here", "thought": true}]
+				}
+			}]
+		}
+	}`)
+
+	// Signature chunk
+	validSignature := "abc123validSignature1234567890123456789012345678901234567890"
+	signatureChunk := []byte(`{
+		"response": {
+			"candidates": [{
+				"content": {
+					"parts": [{"text": "", "thought": true, "thoughtSignature": "` + validSignature + `"}]
+				}
+			}]
+		}
+	}`)
+
+	var param any
+	ctx := context.Background()
+
+	// Process thinking chunk
+	ConvertAntigravityResponseToClaude(ctx, "claude-sonnet-4-5-thinking", requestJSON, requestJSON, thinkingChunk, &param)
+	params := param.(*Params)
+	sessionID := params.SessionID
+	thinkingText := params.CurrentThinkingText.String()
+
+	if sessionID == "" {
+		t.Fatal("SessionID should be set")
+	}
+	if thinkingText == "" {
+		t.Fatal("Thinking text should be accumulated")
+	}
+
+	// Process signature chunk - should cache the signature
+	ConvertAntigravityResponseToClaude(ctx, "claude-sonnet-4-5-thinking", requestJSON, requestJSON, signatureChunk, &param)
+
+	// Verify signature was cached
+	cachedSig := cache.GetCachedSignature(sessionID, thinkingText)
+	if cachedSig != validSignature {
+		t.Errorf("Expected cached signature '%s', got '%s'", validSignature, cachedSig)
+	}
+
+	// Verify thinking text was reset after caching
+	if params.CurrentThinkingText.Len() != 0 {
+		t.Error("Thinking text should be reset after signature is cached")
+	}
+}
+
+func TestConvertAntigravityResponseToClaude_MultipleThinkingBlocks(t *testing.T) {
+	cache.ClearSignatureCache("")
+
+	requestJSON := []byte(`{
+		"messages": [{"role": "user", "content": [{"type": "text", "text": "Multi block test"}]}]
+	}`)
+
+	validSig1 := "signature1_12345678901234567890123456789012345678901234567"
+	validSig2 := "signature2_12345678901234567890123456789012345678901234567"
+
+	// First thinking block with signature
+	block1Thinking := []byte(`{
+		"response": {
+			"candidates": [{
+				"content": {
+					"parts": [{"text": "First thinking block", "thought": true}]
+				}
+			}]
+		}
+	}`)
+	block1Sig := []byte(`{
+		"response": {
+			"candidates": [{
+				"content": {
+					"parts": [{"text": "", "thought": true, "thoughtSignature": "` + validSig1 + `"}]
+				}
+			}]
+		}
+	}`)
+
+	// Text content (breaks thinking)
+	textBlock := []byte(`{
+		"response": {
+			"candidates": [{
+				"content": {
+					"parts": [{"text": "Regular text output"}]
+				}
+			}]
+		}
+	}`)
+
+	// Second thinking block with signature
+	block2Thinking := []byte(`{
+		"response": {
+			"candidates": [{
+				"content": {
+					"parts": [{"text": "Second thinking block", "thought": true}]
+				}
+			}]
+		}
+	}`)
+	block2Sig := []byte(`{
+		"response": {
+			"candidates": [{
+				"content": {
+					"parts": [{"text": "", "thought": true, "thoughtSignature": "` + validSig2 + `"}]
+				}
+			}]
+		}
+	}`)
+
+	var param any
+	ctx := context.Background()
+
+	// Process first thinking block
+	ConvertAntigravityResponseToClaude(ctx, "claude-sonnet-4-5-thinking", requestJSON, requestJSON, block1Thinking, &param)
+	params := param.(*Params)
+	sessionID := params.SessionID
+	firstThinkingText := params.CurrentThinkingText.String()
+
+	ConvertAntigravityResponseToClaude(ctx, "claude-sonnet-4-5-thinking", requestJSON, requestJSON, block1Sig, &param)
+
+	// Verify first signature cached
+	if cache.GetCachedSignature(sessionID, firstThinkingText) != validSig1 {
+		t.Error("First thinking block signature should be cached")
+	}
+
+	// Process text (transitions out of thinking)
+	ConvertAntigravityResponseToClaude(ctx, "claude-sonnet-4-5-thinking", requestJSON, requestJSON, textBlock, &param)
+
+	// Process second thinking block
+	ConvertAntigravityResponseToClaude(ctx, "claude-sonnet-4-5-thinking", requestJSON, requestJSON, block2Thinking, &param)
+	secondThinkingText := params.CurrentThinkingText.String()
+
+	ConvertAntigravityResponseToClaude(ctx, "claude-sonnet-4-5-thinking", requestJSON, requestJSON, block2Sig, &param)
+
+	// Verify second signature cached
+	if cache.GetCachedSignature(sessionID, secondThinkingText) != validSig2 {
+		t.Error("Second thinking block signature should be cached")
+	}
+}
+
+func TestDeriveSessionIDFromRequest(t *testing.T) {
+	tests := []struct {
+		name      string
+		input     []byte
+		wantEmpty bool
+	}{
+		{
+			name:      "valid user message",
+			input:     []byte(`{"messages": [{"role": "user", "content": "Hello"}]}`),
+			wantEmpty: false,
+		},
+		{
+			name:      "user message with content array",
+			input:     []byte(`{"messages": [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}]}`),
+			wantEmpty: false,
+		},
+		{
+			name:      "no user message",
+			input:     []byte(`{"messages": [{"role": "assistant", "content": "Hi"}]}`),
+			wantEmpty: true,
+		},
+		{
+			name:      "empty messages",
+			input:     []byte(`{"messages": []}`),
+			wantEmpty: true,
+		},
+		{
+			name:      "no messages field",
+			input:     []byte(`{}`),
+			wantEmpty: true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := deriveSessionID(tt.input)
+			if tt.wantEmpty && result != "" {
+				t.Errorf("Expected empty session ID, got '%s'", result)
+			}
+			if !tt.wantEmpty && result == "" {
+				t.Error("Expected non-empty session ID")
+			}
+		})
+	}
+}
+
+func TestDeriveSessionIDFromRequest_Deterministic(t *testing.T) {
+	input := []byte(`{"messages": [{"role": "user", "content": "Same message"}]}`)
+
+	id1 := deriveSessionID(input)
+	id2 := deriveSessionID(input)
+
+	if id1 != id2 {
+		t.Errorf("Session ID should be deterministic: '%s' != '%s'", id1, id2)
+	}
+}
+
+func TestDeriveSessionIDFromRequest_DifferentMessages(t *testing.T) {
+	input1 := []byte(`{"messages": [{"role": "user", "content": "Message A"}]}`)
+	input2 := []byte(`{"messages": [{"role": "user", "content": "Message B"}]}`)
+
+	id1 := deriveSessionID(input1)
+	id2 := deriveSessionID(input2)
+
+	if id1 == id2 {
+		t.Error("Different messages should produce different session IDs")
+	}
+}

internal/translator/antigravity/gemini/antigravity_gemini_request.go

@@ -7,7 +7,6 @@ package gemini
 
 import (
 	"bytes"
-	"encoding/json"
 	"fmt"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
@@ -98,16 +97,34 @@ func ConvertGeminiRequestToAntigravity(_ string, inputRawJSON []byte, _ bool) []
 		}
 	}
 
-	gjson.GetBytes(rawJSON, "request.contents").ForEach(func(key, content gjson.Result) bool {
+	// Gemini-specific handling: add skip_thought_signature_validator to functionCall parts
+	// and remove thinking blocks entirely (Gemini doesn't need to preserve them)
+	const skipSentinel = "skip_thought_signature_validator"
+
+	gjson.GetBytes(rawJSON, "request.contents").ForEach(func(contentIdx, content gjson.Result) bool {
 		if content.Get("role").String() == "model" {
-			content.Get("parts").ForEach(func(partKey, part gjson.Result) bool {
+			// First pass: collect indices of thinking parts to remove
+			var thinkingIndicesToRemove []int64
+			content.Get("parts").ForEach(func(partIdx, part gjson.Result) bool {
+				// Mark thinking blocks for removal
+				if part.Get("thought").Bool() {
+					thinkingIndicesToRemove = append(thinkingIndicesToRemove, partIdx.Int())
+				}
+				// Add skip sentinel to functionCall parts
 				if part.Get("functionCall").Exists() {
-					rawJSON, _ = sjson.SetBytes(rawJSON, fmt.Sprintf("request.contents.%d.parts.%d.thoughtSignature", key.Int(), partKey.Int()), "skip_thought_signature_validator")
-				} else if part.Get("thoughtSignature").Exists() {
-					rawJSON, _ = sjson.SetBytes(rawJSON, fmt.Sprintf("request.contents.%d.parts.%d.thoughtSignature", key.Int(), partKey.Int()), "skip_thought_signature_validator")
+					existingSig := part.Get("thoughtSignature").String()
+					if existingSig == "" || len(existingSig) < 50 {
+						rawJSON, _ = sjson.SetBytes(rawJSON, fmt.Sprintf("request.contents.%d.parts.%d.thoughtSignature", contentIdx.Int(), partIdx.Int()), skipSentinel)
+					}
 				}
 				return true
 			})
+
+			// Remove thinking blocks in reverse order to preserve indices
+			for i := len(thinkingIndicesToRemove) - 1; i >= 0; i-- {
+				idx := thinkingIndicesToRemove[i]
+				rawJSON, _ = sjson.DeleteBytes(rawJSON, fmt.Sprintf("request.contents.%d.parts.%d", contentIdx.Int(), idx))
+			}
 		}
 		return true
 	})
@@ -117,41 +134,31 @@ func ConvertGeminiRequestToAntigravity(_ string, inputRawJSON []byte, _ bool) []
 
 // FunctionCallGroup represents a group of function calls and their responses
 type FunctionCallGroup struct {
-	ModelContent    map[string]interface{}
-	FunctionCalls   []gjson.Result
 	ResponsesNeeded int
 }
 
-// parseFunctionResponse attempts to unmarshal a function response part.
-// Falls back to gjson extraction if standard json.Unmarshal fails.
-func parseFunctionResponse(response gjson.Result) map[string]interface{} {
-	var responseMap map[string]interface{}
-	err := json.Unmarshal([]byte(response.Raw), &responseMap)
-	if err == nil {
-		return responseMap
+// parseFunctionResponseRaw attempts to normalize a function response part into a JSON object string.
+// Falls back to a minimal "functionResponse" object when parsing fails.
+func parseFunctionResponseRaw(response gjson.Result) string {
+	if response.IsObject() && gjson.Valid(response.Raw) {
+		return response.Raw
 	}
 
-	log.Debugf("unmarshal function response failed, using fallback: %v", err)
+	log.Debugf("parse function response failed, using fallback")
 	funcResp := response.Get("functionResponse")
 	if funcResp.Exists() {
-		fr := map[string]interface{}{
-			"name": funcResp.Get("name").String(),
-			"response": map[string]interface{}{
-				"result": funcResp.Get("response").String(),
-			},
-		}
+		fr := `{"functionResponse":{"name":"","response":{"result":""}}}`
+		fr, _ = sjson.Set(fr, "functionResponse.name", funcResp.Get("name").String())
+		fr, _ = sjson.Set(fr, "functionResponse.response.result", funcResp.Get("response").String())
 		if id := funcResp.Get("id").String(); id != "" {
-			fr["id"] = id
+			fr, _ = sjson.Set(fr, "functionResponse.id", id)
 		}
-		return map[string]interface{}{"functionResponse": fr}
+		return fr
 	}
 
-	return map[string]interface{}{
-		"functionResponse": map[string]interface{}{
-			"name":     "unknown",
-			"response": map[string]interface{}{"result": response.String()},
-		},
-	}
+	fr := `{"functionResponse":{"name":"unknown","response":{"result":""}}}`
+	fr, _ = sjson.Set(fr, "functionResponse.response.result", response.String())
+	return fr
 }
 
 // fixCLIToolResponse performs sophisticated tool response format conversion and grouping.
@@ -178,7 +185,7 @@ func fixCLIToolResponse(input string) (string, error) {
 	}
 
 	// Initialize data structures for processing and grouping
-	var newContents []interface{}          // Final processed contents array
+	contentsWrapper := `{"contents":[]}`
 	var pendingGroups []*FunctionCallGroup // Groups awaiting completion with responses
 	var collectedResponses []gjson.Result  // Standalone responses to be matched
 
@@ -210,17 +217,16 @@ func fixCLIToolResponse(input string) (string, error) {
 					collectedResponses = collectedResponses[group.ResponsesNeeded:]
 
 					// Create merged function response content
-					var responseParts []interface{}
+					functionResponseContent := `{"parts":[],"role":"function"}`
 					for _, response := range groupResponses {
-						responseParts = append(responseParts, parseFunctionResponse(response))
+						partRaw := parseFunctionResponseRaw(response)
+						if partRaw != "" {
+							functionResponseContent, _ = sjson.SetRaw(functionResponseContent, "parts.-1", partRaw)
+						}
 					}
 
-					if len(responseParts) > 0 {
-						functionResponseContent := map[string]interface{}{
-							"parts": responseParts,
-							"role":  "function",
-						}
-						newContents = append(newContents, functionResponseContent)
+					if gjson.Get(functionResponseContent, "parts.#").Int() > 0 {
+						contentsWrapper, _ = sjson.SetRaw(contentsWrapper, "contents.-1", functionResponseContent)
 					}
 
 					// Remove this group as it's been satisfied
@@ -234,50 +240,42 @@ func fixCLIToolResponse(input string) (string, error) {
 
 		// If this is a model with function calls, create a new group
 		if role == "model" {
-			var functionCallsInThisModel []gjson.Result
+			functionCallsCount := 0
 			parts.ForEach(func(_, part gjson.Result) bool {
 				if part.Get("functionCall").Exists() {
-					functionCallsInThisModel = append(functionCallsInThisModel, part)
+					functionCallsCount++
 				}
 				return true
 			})
 
-			if len(functionCallsInThisModel) > 0 {
+			if functionCallsCount > 0 {
 				// Add the model content
-				var contentMap map[string]interface{}
-				errUnmarshal := json.Unmarshal([]byte(value.Raw), &contentMap)
-				if errUnmarshal != nil {
-					log.Warnf("failed to unmarshal model content: %v\n", errUnmarshal)
+				if !value.IsObject() {
+					log.Warnf("failed to parse model content")
 					return true
 				}
-				newContents = append(newContents, contentMap)
+				contentsWrapper, _ = sjson.SetRaw(contentsWrapper, "contents.-1", value.Raw)
 
 				// Create a new group for tracking responses
 				group := &FunctionCallGroup{
-					ModelContent:    contentMap,
-					FunctionCalls:   functionCallsInThisModel,
-					ResponsesNeeded: len(functionCallsInThisModel),
+					ResponsesNeeded: functionCallsCount,
 				}
 				pendingGroups = append(pendingGroups, group)
 			} else {
 				// Regular model content without function calls
-				var contentMap map[string]interface{}
-				errUnmarshal := json.Unmarshal([]byte(value.Raw), &contentMap)
-				if errUnmarshal != nil {
-					log.Warnf("failed to unmarshal content: %v\n", errUnmarshal)
+				if !value.IsObject() {
+					log.Warnf("failed to parse content")
 					return true
 				}
-				newContents = append(newContents, contentMap)
+				contentsWrapper, _ = sjson.SetRaw(contentsWrapper, "contents.-1", value.Raw)
 			}
 		} else {
 			// Non-model content (user, etc.)
-			var contentMap map[string]interface{}
-			errUnmarshal := json.Unmarshal([]byte(value.Raw), &contentMap)
-			if errUnmarshal != nil {
-				log.Warnf("failed to unmarshal content: %v\n", errUnmarshal)
+			if !value.IsObject() {
+				log.Warnf("failed to parse content")
 				return true
 			}
-			newContents = append(newContents, contentMap)
+			contentsWrapper, _ = sjson.SetRaw(contentsWrapper, "contents.-1", value.Raw)
 		}
 
 		return true
@@ -289,25 +287,23 @@ func fixCLIToolResponse(input string) (string, error) {
 			groupResponses := collectedResponses[:group.ResponsesNeeded]
 			collectedResponses = collectedResponses[group.ResponsesNeeded:]
 
-			var responseParts []interface{}
+			functionResponseContent := `{"parts":[],"role":"function"}`
 			for _, response := range groupResponses {
-				responseParts = append(responseParts, parseFunctionResponse(response))
+				partRaw := parseFunctionResponseRaw(response)
+				if partRaw != "" {
+					functionResponseContent, _ = sjson.SetRaw(functionResponseContent, "parts.-1", partRaw)
+				}
 			}
 
-			if len(responseParts) > 0 {
-				functionResponseContent := map[string]interface{}{
-					"parts": responseParts,
-					"role":  "function",
-				}
-				newContents = append(newContents, functionResponseContent)
+			if gjson.Get(functionResponseContent, "parts.#").Int() > 0 {
+				contentsWrapper, _ = sjson.SetRaw(contentsWrapper, "contents.-1", functionResponseContent)
 			}
 		}
 	}
 
 	// Update the original JSON with the new contents
 	result := input
-	newContentsJSON, _ := json.Marshal(newContents)
-	result, _ = sjson.Set(result, "request.contents", json.RawMessage(newContentsJSON))
+	result, _ = sjson.SetRaw(result, "request.contents", gjson.Get(contentsWrapper, "contents").Raw)
 
 	return result, nil
 }