Merge pull request #756 from leaph/check-ai-thinking-settings

feat(iflow): add model-specific thinking configs for GLM-4.7 and Mini…

.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)
@@ -27,4 +25,9 @@ config.yaml
 bin/*
 .claude/*
 .vscode/*
+.gemini/*
 .serena/*
+.agent/*
+.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.
 

.github/workflows/pr-test-build.yml

@@ -0,0 +1,23 @@
+name: pr-test-build
+
+on:
+  pull_request:
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Set up Go
+        uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: Build
+        run: |
+          go build -o test-output ./cmd/server
+          rm -f test-output

.gitignore

@@ -11,11 +11,15 @@ bin/*
 logs/*
 conv/*
 temp/*
+refs/*
+
+# Storage backends
 pgstore/*
 gitstore/*
 objectstore/*
+
+# Static assets
 static/*
-refs/*
 
 # Authentication data
 auths/*
@@ -30,7 +34,12 @@ GEMINI.md
 # Tooling metadata
 .vscode/*
 .claude/*
+.gemini/*
 .serena/*
+.agent/*
+.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
@@ -99,6 +114,10 @@ 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.
+
 > [!NOTE]  
 > If you developed a project based on CLIProxyAPI, please open a PR to add it to this list.
 

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
@@ -97,6 +113,10 @@ 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。
+
 > [!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。
 

cmd/server/main.go

@@ -405,7 +405,7 @@ func main() {
 	usage.SetStatisticsEnabled(cfg.UsageStatisticsEnabled)
 	coreauth.SetQuotaCooldownDisabled(cfg.DisableCooling)
 
-	if err = logging.ConfigureLogOutput(cfg.LoggingToFile); err != nil {
+	if err = logging.ConfigureLogOutput(cfg.LoggingToFile, cfg.LogsMaxTotalSizeMB); err != nil {
 		log.Errorf("failed to configure log output: %v", err)
 		return
 	}

config.example.yaml

@@ -25,6 +25,9 @@ remote-management:
   # Disable the bundled management control panel asset download and HTTP route when true.
   disable-control-panel: false
 
+  # GitHub repository for the management control panel. Accepts a repository URL or releases API URL.
+  panel-github-repository: "https://github.com/router-for-me/Cli-Proxy-API-Management-Center"
+
 # Authentication directory (supports ~ for home directory)
 auth-dir: "~/.cli-proxy-api"
 
@@ -36,15 +39,25 @@ api-keys:
 # 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
 
+# Maximum total size (MB) of log files under the logs directory. When exceeded, the oldest log
+# files are deleted until within the limit. Set to 0 to disable.
+logs-max-total-size-mb: 0
+
 # When false, disable in-memory usage statistics aggregation
 usage-statistics-enabled: false
 
 # Proxy URL. Supports socks5/http/https protocols. Example: socks5://user:pass@192.168.1.1:1080/
 proxy-url: ""
 
+# When true, unprefixed model requests only use credentials without a prefix (except when prefix == model name).
+force-model-prefix: false
+
 # Number of times to retry a request. Retries will occur if the HTTP response code is 403, 408, 500, 502, 503, or 504.
 request-retry: 3
 
@@ -56,12 +69,22 @@ 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
 
+# 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.
+
 # Gemini API keys
 # gemini-api-key:
 #   - api-key: "AIzaSy...01"
+#     prefix: "test" # optional: require calls like "test/gemini-3-pro-preview" to target this credential
 #     base-url: "https://generativelanguage.googleapis.com"
 #     headers:
 #       X-Custom-Header: "custom-value"
@@ -76,6 +99,7 @@ ws-auth: false
 # Codex API keys
 # codex-api-key:
 #   - api-key: "sk-atSM..."
+#     prefix: "test" # optional: require calls like "test/gpt-5-codex" to target this credential
 #     base-url: "https://www.example.com" # use the custom codex API endpoint
 #     headers:
 #       X-Custom-Header: "custom-value"
@@ -90,6 +114,7 @@ ws-auth: false
 # claude-api-key:
 #   - api-key: "sk-atSM..." # use the official claude API key, no need to set the base url
 #   - api-key: "sk-atSM..."
+#     prefix: "test" # optional: require calls like "test/claude-sonnet-latest" to target this credential
 #     base-url: "https://www.example.com" # use the custom claude API endpoint
 #     headers:
 #       X-Custom-Header: "custom-value"
@@ -106,6 +131,7 @@ ws-auth: false
 # OpenAI compatibility providers
 # openai-compatibility:
 #   - name: "openrouter" # The name of the provider; it will be used in the user agent and other places.
+#     prefix: "test" # optional: require calls like "test/kimi-k2" to target this provider's credentials
 #     base-url: "https://openrouter.ai/api/v1" # The base URL of the provider.
 #     headers:
 #       X-Custom-Header: "custom-value"
@@ -120,6 +146,7 @@ ws-auth: false
 # Vertex API keys (Vertex-compatible endpoints, use API key + base URL)
 # vertex-api-key:
 #   - api-key: "vk-123..."                        # x-goog-api-key header
+#     prefix: "test"                              # optional: require calls like "test/vertex-pro" to target this credential
 #     base-url: "https://example.com/api"         # e.g. https://zenmux.ai/api
 #     proxy-url: "socks5://proxy.example.com:1080" # optional per-key proxy override
 #     headers:

examples/custom-provider/main.go

@@ -23,13 +23,13 @@ import (
 	"time"
 
 	"github.com/gin-gonic/gin"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/api"
-	"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/sdk/api"
 	sdkAuth "github.com/router-for-me/CLIProxyAPI/v6/sdk/auth"
 	"github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy"
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	clipexec "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/logging"
 	sdktr "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
 )
 

go.mod

@@ -18,8 +18,8 @@ require (
 	github.com/tidwall/gjson v1.18.0
 	github.com/tidwall/sjson v1.2.5
 	github.com/tiktoken-go/tokenizer v0.7.0
-	golang.org/x/crypto v0.43.0
-	golang.org/x/net v0.46.0
+	golang.org/x/crypto v0.45.0
+	golang.org/x/net v0.47.0
 	golang.org/x/oauth2 v0.30.0
 	gopkg.in/natefinch/lumberjack.v2 v2.2.1
 	gopkg.in/yaml.v3 v3.0.1
@@ -68,9 +68,9 @@ require (
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.12 // indirect
 	golang.org/x/arch v0.8.0 // indirect
-	golang.org/x/sync v0.17.0 // indirect
-	golang.org/x/sys v0.37.0 // indirect
-	golang.org/x/text v0.30.0 // indirect
+	golang.org/x/sync v0.18.0 // indirect
+	golang.org/x/sys v0.38.0 // indirect
+	golang.org/x/text v0.31.0 // indirect
 	google.golang.org/protobuf v1.34.1 // indirect
 	gopkg.in/ini.v1 v1.67.0 // indirect
 )

go.sum

@@ -160,22 +160,22 @@ github.com/ugorji/go/codec v1.2.12/go.mod h1:UNopzCgEMSXjBc6AOMqYvWC1ktqTAfzJZUZ
 golang.org/x/arch v0.0.0-20210923205945-b76863e36670/go.mod h1:5om86z9Hs0C8fWVUuoMHwpExlXzs5Tkyp9hOrfG7pp8=
 golang.org/x/arch v0.8.0 h1:3wRIsP3pM4yUptoR96otTUOXI367OS0+c9eeRi9doIc=
 golang.org/x/arch v0.8.0/go.mod h1:FEVrYAQjsQXMVJ1nsMoVVXPZg6p2JE2mx8psSWTDQys=
-golang.org/x/crypto v0.43.0 h1:dduJYIi3A3KOfdGOHX8AVZ/jGiyPa3IbBozJ5kNuE04=
-golang.org/x/crypto v0.43.0/go.mod h1:BFbav4mRNlXJL4wNeejLpWxB7wMbc79PdRGhWKncxR0=
-golang.org/x/net v0.46.0 h1:giFlY12I07fugqwPuWJi68oOnpfqFnJIJzaIIm2JVV4=
-golang.org/x/net v0.46.0/go.mod h1:Q9BGdFy1y4nkUwiLvT5qtyhAnEHgnQ/zd8PfU6nc210=
+golang.org/x/crypto v0.45.0 h1:jMBrvKuj23MTlT0bQEOBcAE0mjg8mK9RXFhRH6nyF3Q=
+golang.org/x/crypto v0.45.0/go.mod h1:XTGrrkGJve7CYK7J8PEww4aY7gM3qMCElcJQ8n8JdX4=
+golang.org/x/net v0.47.0 h1:Mx+4dIFzqraBXUugkia1OOvlD6LemFo1ALMHjrXDOhY=
+golang.org/x/net v0.47.0/go.mod h1:/jNxtkgq5yWUGYkaZGqo27cfGZ1c5Nen03aYrrKpVRU=
 golang.org/x/oauth2 v0.30.0 h1:dnDm7JmhM45NNpd8FDDeLhK6FwqbOf4MLCM9zb1BOHI=
 golang.org/x/oauth2 v0.30.0/go.mod h1:B++QgG3ZKulg6sRPGD/mqlHQs5rB3Ml9erfeDY7xKlU=
-golang.org/x/sync v0.17.0 h1:l60nONMj9l5drqw6jlhIELNv9I0A4OFgRsG9k2oT9Ug=
-golang.org/x/sync v0.17.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
+golang.org/x/sync v0.18.0 h1:kr88TuHDroi+UVf+0hZnirlk8o8T+4MrK6mr60WkH/I=
+golang.org/x/sync v0.18.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
 golang.org/x/sys v0.0.0-20220715151400-c0bba94af5f8/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.37.0 h1:fdNQudmxPjkdUTPnLn5mdQv7Zwvbvpaxqs831goi9kQ=
-golang.org/x/sys v0.37.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
-golang.org/x/term v0.36.0 h1:zMPR+aF8gfksFprF/Nc/rd1wRS1EI6nDBGyWAvDzx2Q=
-golang.org/x/term v0.36.0/go.mod h1:Qu394IJq6V6dCBRgwqshf3mPF85AqzYEzofzRdZkWss=
-golang.org/x/text v0.30.0 h1:yznKA/E9zq54KzlzBEAWn1NXSQ8DIp/NYMy88xJjl4k=
-golang.org/x/text v0.30.0/go.mod h1:yDdHFIX9t+tORqspjENWgzaCVXgk0yYnYuSZ8UzzBVM=
+golang.org/x/sys v0.38.0 h1:3yZWxaJjBmCWXqhN1qh02AkOnCQ1poK6oF+a7xWL6Gc=
+golang.org/x/sys v0.38.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
+golang.org/x/term v0.37.0 h1:8EGAD0qCmHYZg6J17DvsMy9/wJ7/D/4pV/wfnld5lTU=
+golang.org/x/term v0.37.0/go.mod h1:5pB4lxRNYYVZuTLmy8oR2BH8dflOR+IbTYFD8fi3254=
+golang.org/x/text v0.31.0 h1:aC8ghyu4JhP8VojJ2lEHBnochRno1sgL6nEi9WGFGMM=
+golang.org/x/text v0.31.0/go.mod h1:tKRAlv61yKIjGGHX/4tP1LTbc13YSec1pxVEWXzfoeM=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543 h1:E7g+9GITq07hpfrRu66IVDexMakfv52eLZ2CXBWiKr4=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 google.golang.org/protobuf v1.34.1 h1:9ddQBjfCyZPOHPUiPxpYESBLc+T8P3E+Vo4IbKZgFWg=

internal/api/handlers/management/auth_files.go

@@ -36,10 +36,6 @@ import (
 	"golang.org/x/oauth2/google"
 )
 
-var (
-	oauthStatus = make(map[string]string)
-)
-
 var lastRefreshKeys = []string{"last_refresh", "lastRefresh", "last_refreshed_at", "lastRefreshedAt"}
 
 const (
@@ -201,6 +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
@@ -786,7 +795,10 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 		return
 	}
 
+	RegisterOAuthSession(state, "anthropic")
+
 	isWebUI := isWebUIRequest(c)
+	var forwarder *callbackForwarder
 	if isWebUI {
 		targetURL, errTarget := h.managementCallbackURL("/anthropic/callback")
 		if errTarget != nil {
@@ -794,7 +806,8 @@ 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
@@ -803,7 +816,7 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 
 	go func() {
 		if isWebUI {
-			defer stopCallbackForwarder(anthropicCallbackPort)
+			defer stopCallbackForwarderInstance(anthropicCallbackPort, forwarder)
 		}
 
 		// Helper: wait for callback file
@@ -811,8 +824,11 @@ 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) {
-					oauthStatus[state] = "Timeout waiting for OAuth callback"
+					SetOAuthSessionError(state, "Timeout waiting for OAuth callback")
 					return nil, fmt.Errorf("timeout waiting for OAuth callback")
 				}
 				data, errRead := os.ReadFile(path)
@@ -830,6 +846,9 @@ 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
@@ -837,13 +856,13 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 		if errStr := resultMap["error"]; errStr != "" {
 			oauthErr := claude.NewOAuthError(errStr, "", http.StatusBadRequest)
 			log.Error(claude.GetUserFriendlyMessage(oauthErr))
-			oauthStatus[state] = "Bad request"
+			SetOAuthSessionError(state, "Bad request")
 			return
 		}
 		if resultMap["state"] != state {
 			authErr := claude.NewAuthenticationError(claude.ErrInvalidState, fmt.Errorf("expected %s, got %s", state, resultMap["state"]))
 			log.Error(claude.GetUserFriendlyMessage(authErr))
-			oauthStatus[state] = "State code error"
+			SetOAuthSessionError(state, "State code error")
 			return
 		}
 
@@ -876,7 +895,7 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 		if errDo != nil {
 			authErr := claude.NewAuthenticationError(claude.ErrCodeExchangeFailed, errDo)
 			log.Errorf("Failed to exchange authorization code for tokens: %v", authErr)
-			oauthStatus[state] = "Failed to exchange authorization code for tokens"
+			SetOAuthSessionError(state, "Failed to exchange authorization code for tokens")
 			return
 		}
 		defer func() {
@@ -887,7 +906,7 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 		respBody, _ := io.ReadAll(resp.Body)
 		if resp.StatusCode != http.StatusOK {
 			log.Errorf("token exchange failed with status %d: %s", resp.StatusCode, string(respBody))
-			oauthStatus[state] = fmt.Sprintf("token exchange failed with status %d", resp.StatusCode)
+			SetOAuthSessionError(state, fmt.Sprintf("token exchange failed with status %d", resp.StatusCode))
 			return
 		}
 		var tResp struct {
@@ -900,7 +919,7 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 		}
 		if errU := json.Unmarshal(respBody, &tResp); errU != nil {
 			log.Errorf("failed to parse token response: %v", errU)
-			oauthStatus[state] = "Failed to parse token response"
+			SetOAuthSessionError(state, "Failed to parse token response")
 			return
 		}
 		bundle := &claude.ClaudeAuthBundle{
@@ -925,7 +944,7 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 		savedPath, errSave := h.saveTokenRecord(ctx, record)
 		if errSave != nil {
 			log.Errorf("Failed to save authentication tokens: %v", errSave)
-			oauthStatus[state] = "Failed to save authentication tokens"
+			SetOAuthSessionError(state, "Failed to save authentication tokens")
 			return
 		}
 
@@ -934,10 +953,10 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 			fmt.Println("API key obtained and saved")
 		}
 		fmt.Println("You can now use Claude services through this CLI")
-		delete(oauthStatus, state)
+		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("anthropic")
 	}()
 
-	oauthStatus[state] = ""
 	c.JSON(200, gin.H{"status": "ok", "url": authURL, "state": state})
 }
 
@@ -968,7 +987,10 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 	state := fmt.Sprintf("gem-%d", time.Now().UnixNano())
 	authURL := conf.AuthCodeURL(state, oauth2.AccessTypeOffline, oauth2.SetAuthURLParam("prompt", "consent"))
 
+	RegisterOAuthSession(state, "gemini")
+
 	isWebUI := isWebUIRequest(c)
+	var forwarder *callbackForwarder
 	if isWebUI {
 		targetURL, errTarget := h.managementCallbackURL("/google/callback")
 		if errTarget != nil {
@@ -976,7 +998,8 @@ 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
@@ -985,7 +1008,7 @@ 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
@@ -994,9 +1017,12 @@ 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")
-				oauthStatus[state] = "OAuth flow timed out"
+				SetOAuthSessionError(state, "OAuth flow timed out")
 				return
 			}
 			if data, errR := os.ReadFile(waitFile); errR == nil {
@@ -1005,13 +1031,13 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 				_ = os.Remove(waitFile)
 				if errStr := m["error"]; errStr != "" {
 					log.Errorf("Authentication failed: %s", errStr)
-					oauthStatus[state] = "Authentication failed"
+					SetOAuthSessionError(state, "Authentication failed")
 					return
 				}
 				authCode = m["code"]
 				if authCode == "" {
 					log.Errorf("Authentication failed: code not found")
-					oauthStatus[state] = "Authentication failed: code not found"
+					SetOAuthSessionError(state, "Authentication failed: code not found")
 					return
 				}
 				break
@@ -1023,7 +1049,7 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 		token, err := conf.Exchange(ctx, authCode)
 		if err != nil {
 			log.Errorf("Failed to exchange token: %v", err)
-			oauthStatus[state] = "Failed to exchange token"
+			SetOAuthSessionError(state, "Failed to exchange token")
 			return
 		}
 
@@ -1034,7 +1060,7 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 		req, errNewRequest := http.NewRequestWithContext(ctx, "GET", "https://www.googleapis.com/oauth2/v1/userinfo?alt=json", nil)
 		if errNewRequest != nil {
 			log.Errorf("Could not get user info: %v", errNewRequest)
-			oauthStatus[state] = "Could not get user info"
+			SetOAuthSessionError(state, "Could not get user info")
 			return
 		}
 		req.Header.Set("Content-Type", "application/json")
@@ -1043,7 +1069,7 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 		resp, errDo := authHTTPClient.Do(req)
 		if errDo != nil {
 			log.Errorf("Failed to execute request: %v", errDo)
-			oauthStatus[state] = "Failed to execute request"
+			SetOAuthSessionError(state, "Failed to execute request")
 			return
 		}
 		defer func() {
@@ -1055,7 +1081,7 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 		bodyBytes, _ := io.ReadAll(resp.Body)
 		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
 			log.Errorf("Get user info request failed with status %d: %s", resp.StatusCode, string(bodyBytes))
-			oauthStatus[state] = fmt.Sprintf("Get user info request failed with status %d", resp.StatusCode)
+			SetOAuthSessionError(state, fmt.Sprintf("Get user info request failed with status %d", resp.StatusCode))
 			return
 		}
 
@@ -1064,7 +1090,6 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			fmt.Printf("Authenticated user email: %s\n", email)
 		} else {
 			fmt.Println("Failed to get user email from token")
-			oauthStatus[state] = "Failed to get user email from token"
 		}
 
 		// Marshal/unmarshal oauth2.Token to generic map and enrich fields
@@ -1072,7 +1097,7 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 		jsonData, _ := json.Marshal(token)
 		if errUnmarshal := json.Unmarshal(jsonData, &ifToken); errUnmarshal != nil {
 			log.Errorf("Failed to unmarshal token: %v", errUnmarshal)
-			oauthStatus[state] = "Failed to unmarshal token"
+			SetOAuthSessionError(state, "Failed to unmarshal token")
 			return
 		}
 
@@ -1095,10 +1120,12 @@ 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)
-			oauthStatus[state] = "Failed to get authenticated client"
+			SetOAuthSessionError(state, "Failed to get authenticated client")
 			return
 		}
 		fmt.Println("Authentication successful.")
@@ -1108,12 +1135,12 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			projects, errAll := onboardAllGeminiProjects(ctx, gemClient, &ts)
 			if errAll != nil {
 				log.Errorf("Failed to complete Gemini CLI onboarding: %v", errAll)
-				oauthStatus[state] = "Failed to complete Gemini CLI onboarding"
+				SetOAuthSessionError(state, "Failed to complete Gemini CLI onboarding")
 				return
 			}
 			if errVerify := ensureGeminiProjectsEnabled(ctx, gemClient, projects); errVerify != nil {
 				log.Errorf("Failed to verify Cloud AI API status: %v", errVerify)
-				oauthStatus[state] = "Failed to verify Cloud AI API status"
+				SetOAuthSessionError(state, "Failed to verify Cloud AI API status")
 				return
 			}
 			ts.ProjectID = strings.Join(projects, ",")
@@ -1121,26 +1148,26 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 		} else {
 			if errEnsure := ensureGeminiProjectAndOnboard(ctx, gemClient, &ts, requestedProjectID); errEnsure != nil {
 				log.Errorf("Failed to complete Gemini CLI onboarding: %v", errEnsure)
-				oauthStatus[state] = "Failed to complete Gemini CLI onboarding"
+				SetOAuthSessionError(state, "Failed to complete Gemini CLI onboarding")
 				return
 			}
 
 			if strings.TrimSpace(ts.ProjectID) == "" {
 				log.Error("Onboarding did not return a project ID")
-				oauthStatus[state] = "Failed to resolve project ID"
+				SetOAuthSessionError(state, "Failed to resolve project ID")
 				return
 			}
 
 			isChecked, errCheck := checkCloudAPIIsEnabled(ctx, gemClient, ts.ProjectID)
 			if errCheck != nil {
 				log.Errorf("Failed to verify Cloud AI API status: %v", errCheck)
-				oauthStatus[state] = "Failed to verify Cloud AI API status"
+				SetOAuthSessionError(state, "Failed to verify Cloud AI API status")
 				return
 			}
 			ts.Checked = isChecked
 			if !isChecked {
 				log.Error("Cloud AI API is not enabled for the selected project")
-				oauthStatus[state] = "Cloud AI API not enabled"
+				SetOAuthSessionError(state, "Cloud AI API not enabled")
 				return
 			}
 		}
@@ -1163,15 +1190,15 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 		savedPath, errSave := h.saveTokenRecord(ctx, record)
 		if errSave != nil {
 			log.Errorf("Failed to save token to file: %v", errSave)
-			oauthStatus[state] = "Failed to save token to file"
+			SetOAuthSessionError(state, "Failed to save token to file")
 			return
 		}
 
-		delete(oauthStatus, state)
+		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("gemini")
 		fmt.Printf("You can now use Gemini CLI services through this CLI; token saved to %s\n", savedPath)
 	}()
 
-	oauthStatus[state] = ""
 	c.JSON(200, gin.H{"status": "ok", "url": authURL, "state": state})
 }
 
@@ -1207,7 +1234,10 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 		return
 	}
 
+	RegisterOAuthSession(state, "codex")
+
 	isWebUI := isWebUIRequest(c)
+	var forwarder *callbackForwarder
 	if isWebUI {
 		targetURL, errTarget := h.managementCallbackURL("/codex/callback")
 		if errTarget != nil {
@@ -1215,7 +1245,8 @@ 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
@@ -1224,7 +1255,7 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 
 	go func() {
 		if isWebUI {
-			defer stopCallbackForwarder(codexCallbackPort)
+			defer stopCallbackForwarderInstance(codexCallbackPort, forwarder)
 		}
 
 		// Wait for callback file
@@ -1232,10 +1263,13 @@ 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))
-				oauthStatus[state] = "Timeout waiting for OAuth callback"
+				SetOAuthSessionError(state, "Timeout waiting for OAuth callback")
 				return
 			}
 			if data, errR := os.ReadFile(waitFile); errR == nil {
@@ -1245,12 +1279,12 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 				if errStr := m["error"]; errStr != "" {
 					oauthErr := codex.NewOAuthError(errStr, "", http.StatusBadRequest)
 					log.Error(codex.GetUserFriendlyMessage(oauthErr))
-					oauthStatus[state] = "Bad Request"
+					SetOAuthSessionError(state, "Bad Request")
 					return
 				}
 				if m["state"] != state {
 					authErr := codex.NewAuthenticationError(codex.ErrInvalidState, fmt.Errorf("expected %s, got %s", state, m["state"]))
-					oauthStatus[state] = "State code error"
+					SetOAuthSessionError(state, "State code error")
 					log.Error(codex.GetUserFriendlyMessage(authErr))
 					return
 				}
@@ -1281,14 +1315,14 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 		resp, errDo := httpClient.Do(req)
 		if errDo != nil {
 			authErr := codex.NewAuthenticationError(codex.ErrCodeExchangeFailed, errDo)
-			oauthStatus[state] = "Failed to exchange authorization code for tokens"
+			SetOAuthSessionError(state, "Failed to exchange authorization code for tokens")
 			log.Errorf("Failed to exchange authorization code for tokens: %v", authErr)
 			return
 		}
 		defer func() { _ = resp.Body.Close() }()
 		respBody, _ := io.ReadAll(resp.Body)
 		if resp.StatusCode != http.StatusOK {
-			oauthStatus[state] = fmt.Sprintf("Token exchange failed with status %d", resp.StatusCode)
+			SetOAuthSessionError(state, fmt.Sprintf("Token exchange failed with status %d", resp.StatusCode))
 			log.Errorf("token exchange failed with status %d: %s", resp.StatusCode, string(respBody))
 			return
 		}
@@ -1299,7 +1333,7 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 			ExpiresIn    int    `json:"expires_in"`
 		}
 		if errU := json.Unmarshal(respBody, &tokenResp); errU != nil {
-			oauthStatus[state] = "Failed to parse token response"
+			SetOAuthSessionError(state, "Failed to parse token response")
 			log.Errorf("failed to parse token response: %v", errU)
 			return
 		}
@@ -1337,7 +1371,7 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 		}
 		savedPath, errSave := h.saveTokenRecord(ctx, record)
 		if errSave != nil {
-			oauthStatus[state] = "Failed to save authentication tokens"
+			SetOAuthSessionError(state, "Failed to save authentication tokens")
 			log.Errorf("Failed to save authentication tokens: %v", errSave)
 			return
 		}
@@ -1346,10 +1380,10 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 			fmt.Println("API key obtained and saved")
 		}
 		fmt.Println("You can now use Codex services through this CLI")
-		delete(oauthStatus, state)
+		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("codex")
 	}()
 
-	oauthStatus[state] = ""
 	c.JSON(200, gin.H{"status": "ok", "url": authURL, "state": state})
 }
 
@@ -1390,7 +1424,10 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 	params.Set("state", state)
 	authURL := "https://accounts.google.com/o/oauth2/v2/auth?" + params.Encode()
 
+	RegisterOAuthSession(state, "antigravity")
+
 	isWebUI := isWebUIRequest(c)
+	var forwarder *callbackForwarder
 	if isWebUI {
 		targetURL, errTarget := h.managementCallbackURL("/antigravity/callback")
 		if errTarget != nil {
@@ -1398,7 +1435,8 @@ 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
@@ -1407,16 +1445,19 @@ 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")
-				oauthStatus[state] = "OAuth flow timed out"
+				SetOAuthSessionError(state, "OAuth flow timed out")
 				return
 			}
 			if data, errReadFile := os.ReadFile(waitFile); errReadFile == nil {
@@ -1425,18 +1466,18 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 				_ = os.Remove(waitFile)
 				if errStr := strings.TrimSpace(payload["error"]); errStr != "" {
 					log.Errorf("Authentication failed: %s", errStr)
-					oauthStatus[state] = "Authentication failed"
+					SetOAuthSessionError(state, "Authentication failed")
 					return
 				}
 				if payloadState := strings.TrimSpace(payload["state"]); payloadState != "" && payloadState != state {
 					log.Errorf("Authentication failed: state mismatch")
-					oauthStatus[state] = "Authentication failed: state mismatch"
+					SetOAuthSessionError(state, "Authentication failed: state mismatch")
 					return
 				}
 				authCode = strings.TrimSpace(payload["code"])
 				if authCode == "" {
 					log.Error("Authentication failed: code not found")
-					oauthStatus[state] = "Authentication failed: code not found"
+					SetOAuthSessionError(state, "Authentication failed: code not found")
 					return
 				}
 				break
@@ -1455,7 +1496,7 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 		req, errNewRequest := http.NewRequestWithContext(ctx, http.MethodPost, "https://oauth2.googleapis.com/token", strings.NewReader(form.Encode()))
 		if errNewRequest != nil {
 			log.Errorf("Failed to build token request: %v", errNewRequest)
-			oauthStatus[state] = "Failed to build token request"
+			SetOAuthSessionError(state, "Failed to build token request")
 			return
 		}
 		req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
@@ -1463,7 +1504,7 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 		resp, errDo := httpClient.Do(req)
 		if errDo != nil {
 			log.Errorf("Failed to execute token request: %v", errDo)
-			oauthStatus[state] = "Failed to exchange token"
+			SetOAuthSessionError(state, "Failed to exchange token")
 			return
 		}
 		defer func() {
@@ -1475,7 +1516,7 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 		if resp.StatusCode < http.StatusOK || resp.StatusCode >= http.StatusMultipleChoices {
 			bodyBytes, _ := io.ReadAll(resp.Body)
 			log.Errorf("Antigravity token exchange failed with status %d: %s", resp.StatusCode, string(bodyBytes))
-			oauthStatus[state] = fmt.Sprintf("Token exchange failed: %d", resp.StatusCode)
+			SetOAuthSessionError(state, fmt.Sprintf("Token exchange failed: %d", resp.StatusCode))
 			return
 		}
 
@@ -1487,7 +1528,7 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 		}
 		if errDecode := json.NewDecoder(resp.Body).Decode(&tokenResp); errDecode != nil {
 			log.Errorf("Failed to parse token response: %v", errDecode)
-			oauthStatus[state] = "Failed to parse token response"
+			SetOAuthSessionError(state, "Failed to parse token response")
 			return
 		}
 
@@ -1496,7 +1537,7 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 			infoReq, errInfoReq := http.NewRequestWithContext(ctx, http.MethodGet, "https://www.googleapis.com/oauth2/v1/userinfo?alt=json", nil)
 			if errInfoReq != nil {
 				log.Errorf("Failed to build user info request: %v", errInfoReq)
-				oauthStatus[state] = "Failed to build user info request"
+				SetOAuthSessionError(state, "Failed to build user info request")
 				return
 			}
 			infoReq.Header.Set("Authorization", "Bearer "+tokenResp.AccessToken)
@@ -1504,7 +1545,7 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 			infoResp, errInfo := httpClient.Do(infoReq)
 			if errInfo != nil {
 				log.Errorf("Failed to execute user info request: %v", errInfo)
-				oauthStatus[state] = "Failed to execute user info request"
+				SetOAuthSessionError(state, "Failed to execute user info request")
 				return
 			}
 			defer func() {
@@ -1523,7 +1564,7 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 			} else {
 				bodyBytes, _ := io.ReadAll(infoResp.Body)
 				log.Errorf("User info request failed with status %d: %s", infoResp.StatusCode, string(bodyBytes))
-				oauthStatus[state] = fmt.Sprintf("User info request failed: %d", infoResp.StatusCode)
+				SetOAuthSessionError(state, fmt.Sprintf("User info request failed: %d", infoResp.StatusCode))
 				return
 			}
 		}
@@ -1571,11 +1612,12 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 		savedPath, errSave := h.saveTokenRecord(ctx, record)
 		if errSave != nil {
 			log.Errorf("Failed to save token to file: %v", errSave)
-			oauthStatus[state] = "Failed to save token to file"
+			SetOAuthSessionError(state, "Failed to save token to file")
 			return
 		}
 
-		delete(oauthStatus, state)
+		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("antigravity")
 		fmt.Printf("Authentication successful! Token saved to %s\n", savedPath)
 		if projectID != "" {
 			fmt.Printf("Using GCP project: %s\n", projectID)
@@ -1583,7 +1625,6 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 		fmt.Println("You can now use Antigravity services through this CLI")
 	}()
 
-	oauthStatus[state] = ""
 	c.JSON(200, gin.H{"status": "ok", "url": authURL, "state": state})
 }
 
@@ -1605,11 +1646,13 @@ func (h *Handler) RequestQwenToken(c *gin.Context) {
 	}
 	authURL := deviceFlow.VerificationURIComplete
 
+	RegisterOAuthSession(state, "qwen")
+
 	go func() {
 		fmt.Println("Waiting for authentication...")
 		tokenData, errPollForToken := qwenAuth.PollForToken(deviceFlow.DeviceCode, deviceFlow.CodeVerifier)
 		if errPollForToken != nil {
-			oauthStatus[state] = "Authentication failed"
+			SetOAuthSessionError(state, "Authentication failed")
 			fmt.Printf("Authentication failed: %v\n", errPollForToken)
 			return
 		}
@@ -1628,16 +1671,15 @@ func (h *Handler) RequestQwenToken(c *gin.Context) {
 		savedPath, errSave := h.saveTokenRecord(ctx, record)
 		if errSave != nil {
 			log.Errorf("Failed to save authentication tokens: %v", errSave)
-			oauthStatus[state] = "Failed to save authentication tokens"
+			SetOAuthSessionError(state, "Failed to save authentication tokens")
 			return
 		}
 
 		fmt.Printf("Authentication successful! Token saved to %s\n", savedPath)
 		fmt.Println("You can now use Qwen services through this CLI")
-		delete(oauthStatus, state)
+		CompleteOAuthSession(state)
 	}()
 
-	oauthStatus[state] = ""
 	c.JSON(200, gin.H{"status": "ok", "url": authURL, "state": state})
 }
 
@@ -1650,7 +1692,10 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 	authSvc := iflowauth.NewIFlowAuth(h.cfg)
 	authURL, redirectURI := authSvc.AuthorizationURL(state, iflowauth.CallbackPort)
 
+	RegisterOAuthSession(state, "iflow")
+
 	isWebUI := isWebUIRequest(c)
+	var forwarder *callbackForwarder
 	if isWebUI {
 		targetURL, errTarget := h.managementCallbackURL("/iflow/callback")
 		if errTarget != nil {
@@ -1658,7 +1703,8 @@ 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
@@ -1667,7 +1713,7 @@ 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...")
 
@@ -1675,8 +1721,11 @@ 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) {
-				oauthStatus[state] = "Authentication failed"
+				SetOAuthSessionError(state, "Authentication failed")
 				fmt.Println("Authentication failed: timeout waiting for callback")
 				return
 			}
@@ -1689,26 +1738,26 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 		}
 
 		if errStr := strings.TrimSpace(resultMap["error"]); errStr != "" {
-			oauthStatus[state] = "Authentication failed"
+			SetOAuthSessionError(state, "Authentication failed")
 			fmt.Printf("Authentication failed: %s\n", errStr)
 			return
 		}
 		if resultState := strings.TrimSpace(resultMap["state"]); resultState != state {
-			oauthStatus[state] = "Authentication failed"
+			SetOAuthSessionError(state, "Authentication failed")
 			fmt.Println("Authentication failed: state mismatch")
 			return
 		}
 
 		code := strings.TrimSpace(resultMap["code"])
 		if code == "" {
-			oauthStatus[state] = "Authentication failed"
+			SetOAuthSessionError(state, "Authentication failed")
 			fmt.Println("Authentication failed: code missing")
 			return
 		}
 
 		tokenData, errExchange := authSvc.ExchangeCodeForTokens(ctx, code, redirectURI)
 		if errExchange != nil {
-			oauthStatus[state] = "Authentication failed"
+			SetOAuthSessionError(state, "Authentication failed")
 			fmt.Printf("Authentication failed: %v\n", errExchange)
 			return
 		}
@@ -1730,7 +1779,7 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 
 		savedPath, errSave := h.saveTokenRecord(ctx, record)
 		if errSave != nil {
-			oauthStatus[state] = "Failed to save authentication tokens"
+			SetOAuthSessionError(state, "Failed to save authentication tokens")
 			log.Errorf("Failed to save authentication tokens: %v", errSave)
 			return
 		}
@@ -1740,10 +1789,10 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 			fmt.Println("API key obtained and saved")
 		}
 		fmt.Println("You can now use iFlow services through this CLI")
-		delete(oauthStatus, state)
+		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("iflow")
 	}()
 
-	oauthStatus[state] = ""
 	c.JSON(http.StatusOK, gin.H{"status": "ok", "url": authURL, "state": state})
 }
 
@@ -2179,16 +2228,24 @@ func checkCloudAPIIsEnabled(ctx context.Context, httpClient *http.Client, projec
 }
 
 func (h *Handler) GetAuthStatus(c *gin.Context) {
-	state := c.Query("state")
-	if err, ok := oauthStatus[state]; ok {
-		if err != "" {
-			c.JSON(200, gin.H{"status": "error", "error": err})
-		} else {
-			c.JSON(200, gin.H{"status": "wait"})
-			return
-		}
-	} else {
-		c.JSON(200, gin.H{"status": "ok"})
+	state := strings.TrimSpace(c.Query("state"))
+	if state == "" {
+		c.JSON(http.StatusOK, gin.H{"status": "ok"})
+		return
 	}
-	delete(oauthStatus, state)
+	if err := ValidateOAuthState(state); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"status": "error", "error": "invalid state"})
+		return
+	}
+
+	_, status, ok := GetOAuthSession(state)
+	if !ok {
+		c.JSON(http.StatusOK, gin.H{"status": "ok"})
+		return
+	}
+	if status != "" {
+		c.JSON(http.StatusOK, gin.H{"status": "error", "error": status})
+		return
+	}
+	c.JSON(http.StatusOK, gin.H{"status": "wait"})
 }

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))
@@ -563,66 +612,72 @@ 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"`
+		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.Headers != nil {
+		entry.Headers = config.NormalizeHeaders(*body.Value.Headers)
+	}
+	if body.Value.ExcludedModels != nil {
+		entry.ExcludedModels = config.NormalizeExcludedModels(*body.Value.ExcludedModels)
+	}
+	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))

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_callback.go

@@ -0,0 +1,100 @@
+package management
+
+import (
+	"errors"
+	"net/http"
+	"net/url"
+	"strings"
+
+	"github.com/gin-gonic/gin"
+)
+
+type oauthCallbackRequest struct {
+	Provider    string `json:"provider"`
+	RedirectURL string `json:"redirect_url"`
+	Code        string `json:"code"`
+	State       string `json:"state"`
+	Error       string `json:"error"`
+}
+
+func (h *Handler) PostOAuthCallback(c *gin.Context) {
+	if h == nil || h.cfg == nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"status": "error", "error": "handler not initialized"})
+		return
+	}
+
+	var req oauthCallbackRequest
+	if err := c.ShouldBindJSON(&req); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"status": "error", "error": "invalid body"})
+		return
+	}
+
+	canonicalProvider, err := NormalizeOAuthProvider(req.Provider)
+	if err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"status": "error", "error": "unsupported provider"})
+		return
+	}
+
+	state := strings.TrimSpace(req.State)
+	code := strings.TrimSpace(req.Code)
+	errMsg := strings.TrimSpace(req.Error)
+
+	if rawRedirect := strings.TrimSpace(req.RedirectURL); rawRedirect != "" {
+		u, errParse := url.Parse(rawRedirect)
+		if errParse != nil {
+			c.JSON(http.StatusBadRequest, gin.H{"status": "error", "error": "invalid redirect_url"})
+			return
+		}
+		q := u.Query()
+		if state == "" {
+			state = strings.TrimSpace(q.Get("state"))
+		}
+		if code == "" {
+			code = strings.TrimSpace(q.Get("code"))
+		}
+		if errMsg == "" {
+			errMsg = strings.TrimSpace(q.Get("error"))
+			if errMsg == "" {
+				errMsg = strings.TrimSpace(q.Get("error_description"))
+			}
+		}
+	}
+
+	if state == "" {
+		c.JSON(http.StatusBadRequest, gin.H{"status": "error", "error": "state is required"})
+		return
+	}
+	if err := ValidateOAuthState(state); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"status": "error", "error": "invalid state"})
+		return
+	}
+	if code == "" && errMsg == "" {
+		c.JSON(http.StatusBadRequest, gin.H{"status": "error", "error": "code or error is required"})
+		return
+	}
+
+	sessionProvider, sessionStatus, ok := GetOAuthSession(state)
+	if !ok {
+		c.JSON(http.StatusNotFound, gin.H{"status": "error", "error": "unknown or expired state"})
+		return
+	}
+	if sessionStatus != "" {
+		c.JSON(http.StatusConflict, gin.H{"status": "error", "error": "oauth flow is not pending"})
+		return
+	}
+	if !strings.EqualFold(sessionProvider, canonicalProvider) {
+		c.JSON(http.StatusBadRequest, gin.H{"status": "error", "error": "provider does not match state"})
+		return
+	}
+
+	if _, errWrite := WriteOAuthCallbackFileForPendingSession(h.cfg.AuthDir, canonicalProvider, state, code, errMsg); errWrite != nil {
+		if errors.Is(errWrite, errOAuthSessionNotPending) {
+			c.JSON(http.StatusConflict, gin.H{"status": "error", "error": "oauth flow is not pending"})
+			return
+		}
+		c.JSON(http.StatusInternalServerError, gin.H{"status": "error", "error": "failed to persist oauth callback"})
+		return
+	}
+
+	c.JSON(http.StatusOK, gin.H{"status": "ok"})
+}

internal/api/handlers/management/oauth_sessions.go

@@ -0,0 +1,283 @@
+package management
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"time"
+)
+
+const (
+	oauthSessionTTL     = 10 * time.Minute
+	maxOAuthStateLength = 128
+)
+
+var (
+	errInvalidOAuthState      = errors.New("invalid oauth state")
+	errUnsupportedOAuthFlow   = errors.New("unsupported oauth provider")
+	errOAuthSessionNotPending = errors.New("oauth session is not pending")
+)
+
+type oauthSession struct {
+	Provider  string
+	Status    string
+	CreatedAt time.Time
+	ExpiresAt time.Time
+}
+
+type oauthSessionStore struct {
+	mu       sync.RWMutex
+	ttl      time.Duration
+	sessions map[string]oauthSession
+}
+
+func newOAuthSessionStore(ttl time.Duration) *oauthSessionStore {
+	if ttl <= 0 {
+		ttl = oauthSessionTTL
+	}
+	return &oauthSessionStore{
+		ttl:      ttl,
+		sessions: make(map[string]oauthSession),
+	}
+}
+
+func (s *oauthSessionStore) purgeExpiredLocked(now time.Time) {
+	for state, session := range s.sessions {
+		if !session.ExpiresAt.IsZero() && now.After(session.ExpiresAt) {
+			delete(s.sessions, state)
+		}
+	}
+}
+
+func (s *oauthSessionStore) Register(state, provider string) {
+	state = strings.TrimSpace(state)
+	provider = strings.ToLower(strings.TrimSpace(provider))
+	if state == "" || provider == "" {
+		return
+	}
+	now := time.Now()
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	s.purgeExpiredLocked(now)
+	s.sessions[state] = oauthSession{
+		Provider:  provider,
+		Status:    "",
+		CreatedAt: now,
+		ExpiresAt: now.Add(s.ttl),
+	}
+}
+
+func (s *oauthSessionStore) SetError(state, message string) {
+	state = strings.TrimSpace(state)
+	message = strings.TrimSpace(message)
+	if state == "" {
+		return
+	}
+	if message == "" {
+		message = "Authentication failed"
+	}
+	now := time.Now()
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	s.purgeExpiredLocked(now)
+	session, ok := s.sessions[state]
+	if !ok {
+		return
+	}
+	session.Status = message
+	session.ExpiresAt = now.Add(s.ttl)
+	s.sessions[state] = session
+}
+
+func (s *oauthSessionStore) Complete(state string) {
+	state = strings.TrimSpace(state)
+	if state == "" {
+		return
+	}
+	now := time.Now()
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	s.purgeExpiredLocked(now)
+	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()
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	s.purgeExpiredLocked(now)
+	session, ok := s.sessions[state]
+	return session, ok
+}
+
+func (s *oauthSessionStore) IsPending(state, provider string) bool {
+	state = strings.TrimSpace(state)
+	provider = strings.ToLower(strings.TrimSpace(provider))
+	now := time.Now()
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	s.purgeExpiredLocked(now)
+	session, ok := s.sessions[state]
+	if !ok {
+		return false
+	}
+	if session.Status != "" {
+		return false
+	}
+	if provider == "" {
+		return true
+	}
+	return strings.EqualFold(session.Provider, provider)
+}
+
+var oauthSessions = newOAuthSessionStore(oauthSessionTTL)
+
+func RegisterOAuthSession(state, provider string) { oauthSessions.Register(state, provider) }
+
+func SetOAuthSessionError(state, message string) { oauthSessions.SetError(state, message) }
+
+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 {
+		return "", "", false
+	}
+	return session.Provider, session.Status, true
+}
+
+func IsOAuthSessionPending(state, provider string) bool {
+	return oauthSessions.IsPending(state, provider)
+}
+
+func ValidateOAuthState(state string) error {
+	trimmed := strings.TrimSpace(state)
+	if trimmed == "" {
+		return fmt.Errorf("%w: empty", errInvalidOAuthState)
+	}
+	if len(trimmed) > maxOAuthStateLength {
+		return fmt.Errorf("%w: too long", errInvalidOAuthState)
+	}
+	if strings.Contains(trimmed, "/") || strings.Contains(trimmed, "\\") {
+		return fmt.Errorf("%w: contains path separator", errInvalidOAuthState)
+	}
+	if strings.Contains(trimmed, "..") {
+		return fmt.Errorf("%w: contains '..'", errInvalidOAuthState)
+	}
+	for _, r := range trimmed {
+		switch {
+		case r >= 'a' && r <= 'z':
+		case r >= 'A' && r <= 'Z':
+		case r >= '0' && r <= '9':
+		case r == '-' || r == '_' || r == '.':
+		default:
+			return fmt.Errorf("%w: invalid character", errInvalidOAuthState)
+		}
+	}
+	return nil
+}
+
+func NormalizeOAuthProvider(provider string) (string, error) {
+	switch strings.ToLower(strings.TrimSpace(provider)) {
+	case "anthropic", "claude":
+		return "anthropic", nil
+	case "codex", "openai":
+		return "codex", nil
+	case "gemini", "google":
+		return "gemini", nil
+	case "iflow", "i-flow":
+		return "iflow", nil
+	case "antigravity", "anti-gravity":
+		return "antigravity", nil
+	case "qwen":
+		return "qwen", nil
+	default:
+		return "", errUnsupportedOAuthFlow
+	}
+}
+
+type oauthCallbackFilePayload struct {
+	Code  string `json:"code"`
+	State string `json:"state"`
+	Error string `json:"error"`
+}
+
+func WriteOAuthCallbackFile(authDir, provider, state, code, errorMessage string) (string, error) {
+	if strings.TrimSpace(authDir) == "" {
+		return "", fmt.Errorf("auth dir is empty")
+	}
+	canonicalProvider, err := NormalizeOAuthProvider(provider)
+	if err != nil {
+		return "", err
+	}
+	if err := ValidateOAuthState(state); err != nil {
+		return "", err
+	}
+
+	fileName := fmt.Sprintf(".oauth-%s-%s.oauth", canonicalProvider, state)
+	filePath := filepath.Join(authDir, fileName)
+	payload := oauthCallbackFilePayload{
+		Code:  strings.TrimSpace(code),
+		State: strings.TrimSpace(state),
+		Error: strings.TrimSpace(errorMessage),
+	}
+	data, err := json.Marshal(payload)
+	if err != nil {
+		return "", fmt.Errorf("marshal oauth callback payload: %w", err)
+	}
+	if err := os.WriteFile(filePath, data, 0o600); err != nil {
+		return "", fmt.Errorf("write oauth callback file: %w", err)
+	}
+	return filePath, nil
+}
+
+func WriteOAuthCallbackFileForPendingSession(authDir, provider, state, code, errorMessage string) (string, error) {
+	canonicalProvider, err := NormalizeOAuthProvider(provider)
+	if err != nil {
+		return "", err
+	}
+	if !IsOAuthSessionPending(state, canonicalProvider) {
+		return "", errOAuthSessionNotPending
+	}
+	return WriteOAuthCallbackFile(authDir, canonicalProvider, state, code, errorMessage)
+}

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

@@ -279,16 +279,23 @@ 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
 		}
 	}

internal/api/modules/amp/amp_test.go

@@ -146,6 +146,9 @@ func TestAmpModule_OnConfigUpdated_CacheInvalidation(t *testing.T) {
 	m := &AmpModule{enabled: true}
 	ms := NewMultiSourceSecretWithPath("", p, time.Minute)
 	m.secretSource = ms
+	m.lastConfig = &config.AmpCode{
+		UpstreamAPIKey: "old-key",
+	}
 
 	// Warm the cache
 	if _, err := ms.Get(context.Background()); err != nil {
@@ -157,7 +160,7 @@ func TestAmpModule_OnConfigUpdated_CacheInvalidation(t *testing.T) {
 	}
 
 	// Update config - should invalidate cache
-	if err := m.OnConfigUpdated(&config.Config{AmpCode: config.AmpCode{UpstreamURL: "http://x"}}); err != nil {
+	if err := m.OnConfigUpdated(&config.Config{AmpCode: config.AmpCode{UpstreamURL: "http://x", UpstreamAPIKey: "new-key"}}); err != nil {
 		t.Fatal(err)
 	}
 

internal/api/modules/amp/fallback_handlers.go

@@ -134,7 +134,43 @@ func (fh *FallbackHandler) WrapHandler(handler gin.HandlerFunc) gin.HandlerFunc
 		}
 
 		// Normalize model (handles dynamic thinking suffixes)
-		normalizedModel, _ := util.NormalizeThinkingModel(modelName)
+		normalizedModel, thinkingMetadata := util.NormalizeThinkingModel(modelName)
+		thinkingSuffix := ""
+		if thinkingMetadata != nil && strings.HasPrefix(modelName, normalizedModel) {
+			thinkingSuffix = modelName[len(normalizedModel):]
+		}
+
+		resolveMappedModel := func() (string, []string) {
+			if fh.modelMapper == nil {
+				return "", nil
+			}
+
+			mappedModel := fh.modelMapper.MapModel(modelName)
+			if mappedModel == "" {
+				mappedModel = fh.modelMapper.MapModel(normalizedModel)
+			}
+			mappedModel = strings.TrimSpace(mappedModel)
+			if mappedModel == "" {
+				return "", nil
+			}
+
+			// 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 {
+					mappedModel += thinkingSuffix
+				}
+			}
+
+			mappedBaseModel, _ := util.NormalizeThinkingModel(mappedModel)
+			mappedProviders := util.GetProviderName(mappedBaseModel)
+			if len(mappedProviders) == 0 {
+				return "", nil
+			}
+
+			return mappedModel, mappedProviders
+		}
 
 		// Track resolved model for logging (may change if mapping is applied)
 		resolvedModel := normalizedModel
@@ -147,21 +183,15 @@ func (fh *FallbackHandler) WrapHandler(handler gin.HandlerFunc) gin.HandlerFunc
 		if forceMappings {
 			// FORCE MODE: Check model mappings FIRST (takes precedence over local API keys)
 			// This allows users to route Amp requests to their preferred OAuth providers
-			if fh.modelMapper != nil {
-				if mappedModel := fh.modelMapper.MapModel(normalizedModel); mappedModel != "" {
-					// Mapping found - check if we have a provider for the mapped model
-					mappedProviders := util.GetProviderName(mappedModel)
-					if len(mappedProviders) > 0 {
-						// Mapping found and provider available - rewrite the model in request body
-						bodyBytes = rewriteModelInRequest(bodyBytes, mappedModel)
-						c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-						// Store mapped model in context for handlers that check it (like gemini bridge)
-						c.Set(MappedModelContextKey, mappedModel)
-						resolvedModel = mappedModel
-						usedMapping = true
-						providers = mappedProviders
-					}
-				}
+			if mappedModel, mappedProviders := resolveMappedModel(); mappedModel != "" {
+				// Mapping found and provider available - rewrite the model in request body
+				bodyBytes = rewriteModelInRequest(bodyBytes, mappedModel)
+				c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+				// Store mapped model in context for handlers that check it (like gemini bridge)
+				c.Set(MappedModelContextKey, mappedModel)
+				resolvedModel = mappedModel
+				usedMapping = true
+				providers = mappedProviders
 			}
 
 			// If no mapping applied, check for local providers
@@ -174,21 +204,15 @@ func (fh *FallbackHandler) WrapHandler(handler gin.HandlerFunc) gin.HandlerFunc
 
 			if len(providers) == 0 {
 				// No providers configured - check if we have a model mapping
-				if fh.modelMapper != nil {
-					if mappedModel := fh.modelMapper.MapModel(normalizedModel); mappedModel != "" {
-						// Mapping found - check if we have a provider for the mapped model
-						mappedProviders := util.GetProviderName(mappedModel)
-						if len(mappedProviders) > 0 {
-							// Mapping found and provider available - rewrite the model in request body
-							bodyBytes = rewriteModelInRequest(bodyBytes, mappedModel)
-							c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-							// Store mapped model in context for handlers that check it (like gemini bridge)
-							c.Set(MappedModelContextKey, mappedModel)
-							resolvedModel = mappedModel
-							usedMapping = true
-							providers = mappedProviders
-						}
-					}
+				if mappedModel, mappedProviders := resolveMappedModel(); mappedModel != "" {
+					// Mapping found and provider available - rewrite the model in request body
+					bodyBytes = rewriteModelInRequest(bodyBytes, mappedModel)
+					c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+					// Store mapped model in context for handlers that check it (like gemini bridge)
+					c.Set(MappedModelContextKey, mappedModel)
+					resolvedModel = mappedModel
+					usedMapping = true
+					providers = mappedProviders
 				}
 			}
 		}
@@ -222,14 +246,14 @@ func (fh *FallbackHandler) WrapHandler(handler gin.HandlerFunc) gin.HandlerFunc
 			// Log: Model was mapped to another model
 			log.Debugf("amp model mapping: request %s -> %s", normalizedModel, resolvedModel)
 			logAmpRouting(RouteTypeModelMapping, modelName, resolvedModel, providerName, requestPath)
-			rewriter := NewResponseRewriter(c.Writer, normalizedModel)
+			rewriter := NewResponseRewriter(c.Writer, modelName)
 			c.Writer = rewriter
 			// Filter Anthropic-Beta header only for local handling paths
 			filterAntropicBetaHeader(c)
 			c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
 			handler(c)
 			rewriter.Flush()
-			log.Debugf("amp model mapping: response %s -> %s", resolvedModel, normalizedModel)
+			log.Debugf("amp model mapping: response %s -> %s", resolvedModel, modelName)
 		} else if len(providers) > 0 {
 			// Log: Using local provider (free)
 			logAmpRouting(RouteTypeLocalProvider, modelName, resolvedModel, providerName, requestPath)

internal/api/modules/amp/fallback_handlers_test.go

@@ -0,0 +1,73 @@
+package amp
+
+import (
+	"bytes"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"net/http/httputil"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+)
+
+func TestFallbackHandler_ModelMapping_PreservesThinkingSuffixAndRewritesResponse(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient("test-client-amp-fallback", "codex", []*registry.ModelInfo{
+		{ID: "test/gpt-5.2", OwnedBy: "openai", Type: "codex"},
+	})
+	defer reg.UnregisterClient("test-client-amp-fallback")
+
+	mapper := NewModelMapper([]config.AmpModelMapping{
+		{From: "gpt-5.2", To: "test/gpt-5.2"},
+	})
+
+	fallback := NewFallbackHandlerWithMapper(func() *httputil.ReverseProxy { return nil }, mapper, nil)
+
+	handler := func(c *gin.Context) {
+		var req struct {
+			Model string `json:"model"`
+		}
+		if err := c.ShouldBindJSON(&req); err != nil {
+			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+			return
+		}
+
+		c.JSON(http.StatusOK, gin.H{
+			"model":      req.Model,
+			"seen_model": req.Model,
+		})
+	}
+
+	r := gin.New()
+	r.POST("/chat/completions", fallback.WrapHandler(handler))
+
+	reqBody := []byte(`{"model":"gpt-5.2(xhigh)"}`)
+	req := httptest.NewRequest(http.MethodPost, "/chat/completions", bytes.NewReader(reqBody))
+	req.Header.Set("Content-Type", "application/json")
+	w := httptest.NewRecorder()
+	r.ServeHTTP(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("Expected status 200, got %d", w.Code)
+	}
+
+	var resp struct {
+		Model     string `json:"model"`
+		SeenModel string `json:"seen_model"`
+	}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("Failed to parse response JSON: %v", err)
+	}
+
+	if resp.Model != "gpt-5.2(xhigh)" {
+		t.Errorf("Expected response model gpt-5.2(xhigh), got %s", resp.Model)
+	}
+	if resp.SeenModel != "test/gpt-5.2(xhigh)" {
+		t.Errorf("Expected handler to see test/gpt-5.2(xhigh), got %s", resp.SeenModel)
+	}
+}

internal/api/modules/amp/model_mapping.go

@@ -3,6 +3,7 @@
 package amp
 
 import (
+	"regexp"
 	"strings"
 	"sync"
 
@@ -26,13 +27,15 @@ 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
@@ -55,11 +58,23 @@ func (m *DefaultModelMapper) MapModel(requestedModel string) string {
 	// Check for direct mapping
 	targetModel, exists := m.mappings[normalizedRequest]
 	if !exists {
-		return ""
+		// Try regex mappings in order
+		base, _ := util.NormalizeThinkingModel(requestedModel)
+		for _, rm := range m.regexps {
+			if rm.re.MatchString(requestedModel) || (base != "" && rm.re.MatchString(base)) {
+				targetModel = rm.to
+				exists = true
+				break
+			}
+		}
+		if !exists {
+			return ""
+		}
 	}
 
 	// Verify target model has available providers
-	providers := util.GetProviderName(targetModel)
+	normalizedTarget, _ := util.NormalizeThinkingModel(targetModel)
+	providers := util.GetProviderName(normalizedTarget)
 	if len(providers) == 0 {
 		log.Debugf("amp model mapping: target model %s has no available providers, skipping mapping", targetModel)
 		return ""
@@ -77,6 +92,7 @@ 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)
@@ -87,16 +103,30 @@ 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).
@@ -110,3 +140,8 @@ 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

@@ -71,6 +71,25 @@ func TestModelMapper_MapModel_WithProvider(t *testing.T) {
 	}
 }
 
+func TestModelMapper_MapModel_TargetWithThinkingSuffix(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient("test-client-thinking", "codex", []*registry.ModelInfo{
+		{ID: "gpt-5.2", OwnedBy: "openai", Type: "codex"},
+	})
+	defer reg.UnregisterClient("test-client-thinking")
+
+	mappings := []config.AmpModelMapping{
+		{From: "gpt-5.2-alias", To: "gpt-5.2(xhigh)"},
+	}
+
+	mapper := NewModelMapper(mappings)
+
+	result := mapper.MapModel("gpt-5.2-alias")
+	if result != "gpt-5.2(xhigh)" {
+		t.Errorf("Expected gpt-5.2(xhigh), got %s", result)
+	}
+}
+
 func TestModelMapper_MapModel_CaseInsensitive(t *testing.T) {
 	reg := registry.GetGlobalRegistry()
 	reg.RegisterClient("test-client2", "claude", []*registry.ModelInfo{
@@ -184,3 +203,81 @@ 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$", To: "gemini-2.5-pro", Regex: true},
+	}
+
+	mapper := NewModelMapper(mappings)
+
+	// Incoming model has reasoning suffix but should match base via regex
+	result := mapper.MapModel("gpt-5(high)")
+	if result != "gemini-2.5-pro" {
+		t.Errorf("Expected gemini-2.5-pro, got %s", result)
+	}
+}
+
+func TestModelMapper_Regex_ExactPrecedence(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient("test-client-regex-2", "claude", []*registry.ModelInfo{
+		{ID: "claude-sonnet-4", OwnedBy: "anthropic", Type: "claude"},
+	})
+	reg.RegisterClient("test-client-regex-3", "gemini", []*registry.ModelInfo{
+		{ID: "gemini-2.5-pro", OwnedBy: "google", Type: "gemini"},
+	})
+	defer reg.UnregisterClient("test-client-regex-2")
+	defer reg.UnregisterClient("test-client-regex-3")
+
+	mappings := []config.AmpModelMapping{
+		{From: "gpt-5", To: "claude-sonnet-4"},                 // exact
+		{From: "^gpt-5.*$", 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)
+	}
+}

internal/api/modules/amp/routes.go

@@ -95,6 +95,20 @@ func (m *AmpModule) managementAvailabilityMiddleware() gin.HandlerFunc {
 	}
 }
 
+// wrapManagementAuth skips auth for selected management paths while keeping authentication elsewhere.
+func wrapManagementAuth(auth gin.HandlerFunc, prefixes ...string) gin.HandlerFunc {
+	return func(c *gin.Context) {
+		path := c.Request.URL.Path
+		for _, prefix := range prefixes {
+			if strings.HasPrefix(path, prefix) && (len(path) == len(prefix) || path[len(prefix)] == '/') {
+				c.Next()
+				return
+			}
+		}
+		auth(c)
+	}
+}
+
 // registerManagementRoutes registers Amp management proxy routes
 // These routes proxy through to the Amp control plane for OAuth, user management, etc.
 // Uses dynamic middleware and proxy getter for hot-reload support.
@@ -109,8 +123,10 @@ func (m *AmpModule) registerManagementRoutes(engine *gin.Engine, baseHandler *ha
 	ampAPI.Use(m.localhostOnlyMiddleware())
 
 	// Apply authentication middleware - requires valid API key in Authorization header
+	var authWithBypass gin.HandlerFunc
 	if auth != nil {
 		ampAPI.Use(auth)
+		authWithBypass = wrapManagementAuth(auth, "/threads", "/auth", "/docs", "/settings")
 	}
 
 	// Dynamic proxy handler that uses m.getProxy() for hot-reload support
@@ -156,10 +172,16 @@ func (m *AmpModule) registerManagementRoutes(engine *gin.Engine, baseHandler *ha
 	// Root-level routes that AMP CLI expects without /api prefix
 	// These need the same security middleware as the /api/* routes (dynamic for hot-reload)
 	rootMiddleware := []gin.HandlerFunc{m.managementAvailabilityMiddleware(), noCORSMiddleware(), m.localhostOnlyMiddleware()}
-	if auth != nil {
-		rootMiddleware = append(rootMiddleware, auth)
+	if authWithBypass != nil {
+		rootMiddleware = append(rootMiddleware, authWithBypass)
 	}
+	engine.GET("/threads", append(rootMiddleware, proxyHandler)...)
 	engine.GET("/threads/*path", append(rootMiddleware, proxyHandler)...)
+	engine.GET("/docs", append(rootMiddleware, proxyHandler)...)
+	engine.GET("/docs/*path", append(rootMiddleware, proxyHandler)...)
+	engine.GET("/settings", append(rootMiddleware, proxyHandler)...)
+	engine.GET("/settings/*path", append(rootMiddleware, proxyHandler)...)
+
 	engine.GET("/threads.rss", append(rootMiddleware, proxyHandler)...)
 	engine.GET("/news.rss", append(rootMiddleware, proxyHandler)...)
 
@@ -267,7 +289,7 @@ func (m *AmpModule) registerProviderAliases(engine *gin.Engine, baseHandler *han
 	v1betaAmp := provider.Group("/v1beta")
 	{
 		v1betaAmp.GET("/models", geminiHandlers.GeminiModels)
-		v1betaAmp.POST("/models/:action", fallbackHandler.WrapHandler(geminiHandlers.GeminiHandler))
-		v1betaAmp.GET("/models/:action", geminiHandlers.GeminiGetHandler)
+		v1betaAmp.POST("/models/*action", fallbackHandler.WrapHandler(geminiHandlers.GeminiHandler))
+		v1betaAmp.GET("/models/*action", geminiHandlers.GeminiGetHandler)
 	}
 }

internal/api/modules/amp/routes_test.go

@@ -32,7 +32,9 @@ func TestRegisterManagementRoutes(t *testing.T) {
 	m.setProxy(proxy)
 
 	base := &handlers.BaseAPIHandler{}
-	m.registerManagementRoutes(r, base)
+	m.registerManagementRoutes(r, base, nil)
+	srv := httptest.NewServer(r)
+	defer srv.Close()
 
 	managementPaths := []struct {
 		path   string
@@ -63,11 +65,17 @@ func TestRegisterManagementRoutes(t *testing.T) {
 	for _, path := range managementPaths {
 		t.Run(path.path, func(t *testing.T) {
 			proxyCalled = false
-			req := httptest.NewRequest(path.method, path.path, nil)
-			w := httptest.NewRecorder()
-			r.ServeHTTP(w, req)
+			req, err := http.NewRequest(path.method, srv.URL+path.path, nil)
+			if err != nil {
+				t.Fatalf("failed to build request: %v", err)
+			}
+			resp, err := http.DefaultClient.Do(req)
+			if err != nil {
+				t.Fatalf("request failed: %v", err)
+			}
+			defer resp.Body.Close()
 
-			if w.Code == http.StatusNotFound {
+			if resp.StatusCode == http.StatusNotFound {
 				t.Fatalf("route %s not registered", path.path)
 			}
 			if !proxyCalled {

internal/api/server.go

@@ -209,13 +209,15 @@ 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
+			}
 		}
 	}
 
@@ -230,13 +232,9 @@ func NewServer(cfg *config.Config, authManager *auth.Manager, accessManager *sdk
 	envManagementSecret := envAdminPasswordSet && envAdminPassword != ""
 
 	// Create server instance
-	providerNames := make([]string, 0, len(cfg.OpenAICompatibility))
-	for _, p := range cfg.OpenAICompatibility {
-		providerNames = append(providerNames, p.Name)
-	}
 	s := &Server{
 		engine:              engine,
-		handlers:            handlers.NewBaseAPIHandlers(&cfg.SDKConfig, authManager, providerNames),
+		handlers:            handlers.NewBaseAPIHandlers(&cfg.SDKConfig, authManager),
 		cfg:                 cfg,
 		accessManager:       accessManager,
 		requestLogger:       requestLogger,
@@ -334,8 +332,8 @@ func (s *Server) setupRoutes() {
 	v1beta.Use(AuthMiddleware(s.accessManager))
 	{
 		v1beta.GET("/models", geminiHandlers.GeminiModels)
-		v1beta.POST("/models/:action", geminiHandlers.GeminiHandler)
-		v1beta.GET("/models/:action", geminiHandlers.GeminiGetHandler)
+		v1beta.POST("/models/*action", geminiHandlers.GeminiHandler)
+		v1beta.GET("/models/*action", geminiHandlers.GeminiGetHandler)
 	}
 
 	// Root endpoint
@@ -358,10 +356,11 @@ func (s *Server) setupRoutes() {
 		code := c.Query("code")
 		state := c.Query("state")
 		errStr := c.Query("error")
-		// Persist to a temporary file keyed by state
+		if errStr == "" {
+			errStr = c.Query("error_description")
+		}
 		if state != "" {
-			file := fmt.Sprintf("%s/.oauth-anthropic-%s.oauth", s.cfg.AuthDir, state)
-			_ = os.WriteFile(file, []byte(fmt.Sprintf(`{"code":"%s","state":"%s","error":"%s"}`, code, state, errStr)), 0o600)
+			_, _ = managementHandlers.WriteOAuthCallbackFileForPendingSession(s.cfg.AuthDir, "anthropic", state, code, errStr)
 		}
 		c.Header("Content-Type", "text/html; charset=utf-8")
 		c.String(http.StatusOK, oauthCallbackSuccessHTML)
@@ -371,9 +370,11 @@ func (s *Server) setupRoutes() {
 		code := c.Query("code")
 		state := c.Query("state")
 		errStr := c.Query("error")
+		if errStr == "" {
+			errStr = c.Query("error_description")
+		}
 		if state != "" {
-			file := fmt.Sprintf("%s/.oauth-codex-%s.oauth", s.cfg.AuthDir, state)
-			_ = os.WriteFile(file, []byte(fmt.Sprintf(`{"code":"%s","state":"%s","error":"%s"}`, code, state, errStr)), 0o600)
+			_, _ = managementHandlers.WriteOAuthCallbackFileForPendingSession(s.cfg.AuthDir, "codex", state, code, errStr)
 		}
 		c.Header("Content-Type", "text/html; charset=utf-8")
 		c.String(http.StatusOK, oauthCallbackSuccessHTML)
@@ -383,9 +384,11 @@ func (s *Server) setupRoutes() {
 		code := c.Query("code")
 		state := c.Query("state")
 		errStr := c.Query("error")
+		if errStr == "" {
+			errStr = c.Query("error_description")
+		}
 		if state != "" {
-			file := fmt.Sprintf("%s/.oauth-gemini-%s.oauth", s.cfg.AuthDir, state)
-			_ = os.WriteFile(file, []byte(fmt.Sprintf(`{"code":"%s","state":"%s","error":"%s"}`, code, state, errStr)), 0o600)
+			_, _ = managementHandlers.WriteOAuthCallbackFileForPendingSession(s.cfg.AuthDir, "gemini", state, code, errStr)
 		}
 		c.Header("Content-Type", "text/html; charset=utf-8")
 		c.String(http.StatusOK, oauthCallbackSuccessHTML)
@@ -395,9 +398,11 @@ func (s *Server) setupRoutes() {
 		code := c.Query("code")
 		state := c.Query("state")
 		errStr := c.Query("error")
+		if errStr == "" {
+			errStr = c.Query("error_description")
+		}
 		if state != "" {
-			file := fmt.Sprintf("%s/.oauth-iflow-%s.oauth", s.cfg.AuthDir, state)
-			_ = os.WriteFile(file, []byte(fmt.Sprintf(`{"code":"%s","state":"%s","error":"%s"}`, code, state, errStr)), 0o600)
+			_, _ = managementHandlers.WriteOAuthCallbackFileForPendingSession(s.cfg.AuthDir, "iflow", state, code, errStr)
 		}
 		c.Header("Content-Type", "text/html; charset=utf-8")
 		c.String(http.StatusOK, oauthCallbackSuccessHTML)
@@ -407,9 +412,11 @@ func (s *Server) setupRoutes() {
 		code := c.Query("code")
 		state := c.Query("state")
 		errStr := c.Query("error")
+		if errStr == "" {
+			errStr = c.Query("error_description")
+		}
 		if state != "" {
-			file := fmt.Sprintf("%s/.oauth-antigravity-%s.oauth", s.cfg.AuthDir, state)
-			_ = os.WriteFile(file, []byte(fmt.Sprintf(`{"code":"%s","state":"%s","error":"%s"}`, code, state, errStr)), 0o600)
+			_, _ = managementHandlers.WriteOAuthCallbackFileForPendingSession(s.cfg.AuthDir, "antigravity", state, code, errStr)
 		}
 		c.Header("Content-Type", "text/html; charset=utf-8")
 		c.String(http.StatusOK, oauthCallbackSuccessHTML)
@@ -469,6 +476,8 @@ 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)
@@ -513,6 +522,7 @@ 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)
@@ -581,6 +591,7 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.GET("/qwen-auth-url", s.mgmt.RequestQwenToken)
 		mgmt.GET("/iflow-auth-url", s.mgmt.RequestIFlowToken)
 		mgmt.POST("/iflow-auth-url", s.mgmt.RequestIFlowCookieToken)
+		mgmt.POST("/oauth-callback", s.mgmt.PostOAuthCallback)
 		mgmt.GET("/get-auth-status", s.mgmt.GetAuthStatus)
 	}
 }
@@ -609,7 +620,7 @@ func (s *Server) serveManagementControlPanel(c *gin.Context) {
 
 	if _, err := os.Stat(filePath); err != nil {
 		if os.IsNotExist(err) {
-			go managementasset.EnsureLatestManagementHTML(context.Background(), managementasset.StaticDir(s.configFilePath), cfg.ProxyURL)
+			go managementasset.EnsureLatestManagementHTML(context.Background(), managementasset.StaticDir(s.configFilePath), cfg.ProxyURL, cfg.RemoteManagement.PanelGitHubRepository)
 			c.AbortWithStatus(http.StatusNotFound)
 			return
 		}
@@ -838,11 +849,20 @@ func (s *Server) UpdateClients(cfg *config.Config) {
 		}
 	}
 
-	if oldCfg != nil && oldCfg.LoggingToFile != cfg.LoggingToFile {
-		if err := logging.ConfigureLogOutput(cfg.LoggingToFile); err != nil {
+	if oldCfg == nil || oldCfg.LoggingToFile != cfg.LoggingToFile || oldCfg.LogsMaxTotalSizeMB != cfg.LogsMaxTotalSizeMB {
+		if err := logging.ConfigureLogOutput(cfg.LoggingToFile, cfg.LogsMaxTotalSizeMB); err != nil {
 			log.Errorf("failed to reconfigure log output: %v", err)
 		} else {
-			log.Debugf("logging_to_file updated from %t to %t", oldCfg.LoggingToFile, cfg.LoggingToFile)
+			if oldCfg == nil {
+				log.Debug("log output configuration refreshed")
+			} else {
+				if oldCfg.LoggingToFile != cfg.LoggingToFile {
+					log.Debugf("logging_to_file updated from %t to %t", oldCfg.LoggingToFile, cfg.LoggingToFile)
+				}
+				if oldCfg.LogsMaxTotalSizeMB != cfg.LogsMaxTotalSizeMB {
+					log.Debugf("logs_max_total_size_mb updated from %d to %d", oldCfg.LogsMaxTotalSizeMB, cfg.LogsMaxTotalSizeMB)
+				}
+			}
 		}
 	}
 
@@ -919,17 +939,11 @@ func (s *Server) UpdateClients(cfg *config.Config) {
 	// Save YAML snapshot for next comparison
 	s.oldConfigYaml, _ = yaml.Marshal(cfg)
 
-	providerNames := make([]string, 0, len(cfg.OpenAICompatibility))
-	for _, p := range cfg.OpenAICompatibility {
-		providerNames = append(providerNames, p.Name)
-	}
-	s.handlers.OpenAICompatProviders = providerNames
-
 	s.handlers.UpdateClients(&cfg.SDKConfig)
 
 	if !cfg.RemoteManagement.DisableControlPanel {
 		staticDir := managementasset.StaticDir(s.configFilePath)
-		go managementasset.EnsureLatestManagementHTML(context.Background(), staticDir, cfg.ProxyURL)
+		go managementasset.EnsureLatestManagementHTML(context.Background(), staticDir, cfg.ProxyURL, cfg.RemoteManagement.PanelGitHubRepository)
 	}
 	if s.mgmt != nil {
 		s.mgmt.SetConfig(cfg)

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"
@@ -46,6 +47,12 @@ var (
 type GeminiAuth struct {
 }
 
+// WebLoginOptions customizes the interactive OAuth flow.
+type WebLoginOptions struct {
+	NoBrowser bool
+	Prompt    func(string) (string, error)
+}
+
 // NewGeminiAuth creates a new instance of GeminiAuth.
 func NewGeminiAuth() *GeminiAuth {
 	return &GeminiAuth{}
@@ -59,12 +66,12 @@ 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) {
 	// Configure proxy settings for the HTTP client if a proxy URL is provided.
 	proxyURL, err := url.Parse(cfg.ProxyURL)
 	if err == nil {
@@ -109,7 +116,7 @@ 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)
 		}
@@ -205,15 +212,15 @@ 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) {
 	// 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()
@@ -223,17 +230,26 @@ func (g *GeminiAuth) getTokenFromWeb(ctx context.Context, config *oauth2.Config,
 	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.
@@ -250,7 +266,12 @@ 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
@@ -281,13 +302,60 @@ 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 +1,164 @@
+package cache
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"sort"
+	"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 = 1 * time.Hour
+
+	// MaxEntriesPerSession limits memory usage per session
+	MaxEntriesPerSession = 100
+
+	// 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
+)
+
+// signatureCache stores signatures by sessionId -> textHash -> SignatureEntry
+var signatureCache sync.Map
+
+// 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 {
+	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)
+}
+
+// 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()
+
+	// Evict expired entries if at capacity
+	if len(sc.entries) >= MaxEntriesPerSession {
+		now := time.Now()
+		for key, entry := range sc.entries {
+			if now.Sub(entry.Timestamp) > SignatureCacheTTL {
+				delete(sc.entries, key)
+			}
+		}
+		// If still at capacity, remove oldest entries
+		if len(sc.entries) >= MaxEntriesPerSession {
+			// Find and remove oldest quarter
+			oldest := make([]struct {
+				key string
+				ts  time.Time
+			}, 0, len(sc.entries))
+			for key, entry := range sc.entries {
+				oldest = append(oldest, struct {
+					key string
+					ts  time.Time
+				}{key, entry.Timestamp})
+			}
+			// Sort by timestamp (oldest first) using sort.Slice
+			sort.Slice(oldest, func(i, j int) bool {
+				return oldest[i].ts.Before(oldest[j].ts)
+			})
+
+			toRemove := len(oldest) / 4
+			if toRemove < 1 {
+				toRemove = 1
+			}
+
+			for i := 0; i < toRemove; i++ {
+				delete(sc.entries, oldest[i].key)
+			}
+		}
+	}
+
+	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)
+
+	sc.mu.RLock()
+	entry, exists := sc.entries[textHash]
+	sc.mu.RUnlock()
+
+	if !exists {
+		return ""
+	}
+
+	// Check if expired
+	if time.Since(entry.Timestamp) > SignatureCacheTTL {
+		sc.mu.Lock()
+		delete(sc.entries, textHash)
+		sc.mu.Unlock()
+		return ""
+	}
+
+	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 +24,17 @@ 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,
+		Prompt:    promptFn,
 	}
 
 	_, savedPath, err := manager.Login(context.Background(), "claude", cfg, authOpts)

internal/cmd/antigravity_login.go

@@ -15,11 +15,16 @@ 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,
+		Prompt:    promptFn,
 	}
 
 	record, savedPath, err := manager.Login(context.Background(), "antigravity", cfg, authOpts)

internal/cmd/iflow_login.go

@@ -20,13 +20,7 @@ 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{

internal/cmd/login.go

@@ -55,11 +55,22 @@ 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),
+		ProjectID: trimmedProjectID,
 		Metadata:  map[string]string{},
-		Prompt:    options.Prompt,
+		Prompt:    callbackPrompt,
 	}
 
 	authenticator := sdkAuth.NewGeminiAuthenticator()
@@ -76,7 +87,10 @@ 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,
+		Prompt:    callbackPrompt,
+	})
 	if errClient != nil {
 		log.Errorf("Gemini authentication failed: %v", errClient)
 		return
@@ -90,12 +104,7 @@ 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

@@ -35,12 +35,17 @@ 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,
+		Prompt:    promptFn,
 	}
 
 	_, savedPath, err := manager.Login(context.Background(), "codex", cfg, authOpts)

internal/config/config.go

@@ -12,14 +12,15 @@ import (
 	"strings"
 	"syscall"
 
-	"github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
 	"golang.org/x/crypto/bcrypt"
 	"gopkg.in/yaml.v3"
 )
 
+const DefaultPanelGitHubRepository = "https://github.com/router-for-me/Cli-Proxy-API-Management-Center"
+
 // Config represents the application's configuration, loaded from a YAML file.
 type Config struct {
-	config.SDKConfig `yaml:",inline"`
+	SDKConfig `yaml:",inline"`
 	// Host is the network host/interface on which the API server will bind.
 	// Default is empty ("") to bind all interfaces (IPv4 + IPv6). Use "127.0.0.1" or "localhost" for local-only access.
 	Host string `yaml:"host" json:"-"`
@@ -38,9 +39,16 @@ 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"`
 
+	// LogsMaxTotalSizeMB limits the total size (in MB) of log files under the logs directory.
+	// When exceeded, the oldest log files are deleted until within the limit. Set to 0 to disable.
+	LogsMaxTotalSizeMB int `yaml:"logs-max-total-size-mb" json:"logs-max-total-size-mb"`
+
 	// UsageStatisticsEnabled toggles in-memory usage aggregation; when false, usage data is discarded.
 	UsageStatisticsEnabled bool `yaml:"usage-statistics-enabled" json:"usage-statistics-enabled"`
 
@@ -55,6 +63,9 @@ 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"`
 
@@ -104,6 +115,9 @@ type RemoteManagement struct {
 	SecretKey string `yaml:"secret-key"`
 	// DisableControlPanel skips serving and syncing the bundled management UI when true.
 	DisableControlPanel bool `yaml:"disable-control-panel"`
+	// PanelGitHubRepository overrides the GitHub repository used to fetch the management panel asset.
+	// Accepts either a repository URL (https://github.com/org/repo) or an API releases endpoint.
+	PanelGitHubRepository string `yaml:"panel-github-repository"`
 }
 
 // QuotaExceeded defines the behavior when API quota limits are exceeded.
@@ -116,6 +130,13 @@ 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"`
+}
+
 // 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.
@@ -126,6 +147,11 @@ 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,
@@ -182,6 +208,9 @@ type ClaudeKey struct {
 	// APIKey is the authentication key for accessing Claude API services.
 	APIKey string `yaml:"api-key" json:"api-key"`
 
+	// Prefix optionally namespaces models for this credential (e.g., "teamA/claude-sonnet-4").
+	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
+
 	// BaseURL is the base URL for the Claude API endpoint.
 	// If empty, the default Claude API URL will be used.
 	BaseURL string `yaml:"base-url" json:"base-url"`
@@ -214,6 +243,9 @@ type CodexKey struct {
 	// APIKey is the authentication key for accessing Codex API services.
 	APIKey string `yaml:"api-key" json:"api-key"`
 
+	// Prefix optionally namespaces models for this credential (e.g., "teamA/gpt-5-codex").
+	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
+
 	// BaseURL is the base URL for the Codex API endpoint.
 	// If empty, the default Codex API URL will be used.
 	BaseURL string `yaml:"base-url" json:"base-url"`
@@ -234,6 +266,9 @@ type GeminiKey struct {
 	// APIKey is the authentication key for accessing Gemini API services.
 	APIKey string `yaml:"api-key" json:"api-key"`
 
+	// Prefix optionally namespaces models for this credential (e.g., "teamA/gemini-3-pro-preview").
+	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
+
 	// BaseURL optionally overrides the Gemini API endpoint.
 	BaseURL string `yaml:"base-url,omitempty" json:"base-url,omitempty"`
 
@@ -253,6 +288,9 @@ type OpenAICompatibility struct {
 	// Name is the identifier for this OpenAI compatibility configuration.
 	Name string `yaml:"name" json:"name"`
 
+	// Prefix optionally namespaces model aliases for this provider (e.g., "teamA/kimi-k2").
+	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
+
 	// BaseURL is the base URL for the external OpenAI-compatible API endpoint.
 	BaseURL string `yaml:"base-url" json:"base-url"`
 
@@ -325,9 +363,11 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 	// Set defaults before unmarshal so that absent keys keep defaults.
 	cfg.Host = "" // Default empty: binds to all interfaces (IPv4 + IPv6)
 	cfg.LoggingToFile = false
+	cfg.LogsMaxTotalSizeMB = 0
 	cfg.UsageStatisticsEnabled = false
 	cfg.DisableCooling = false
 	cfg.AmpCode.RestrictManagementToLocalhost = false // Default to false: API key auth is sufficient
+	cfg.RemoteManagement.PanelGitHubRepository = DefaultPanelGitHubRepository
 	if err = yaml.Unmarshal(data, &cfg); err != nil {
 		if optional {
 			// In cloud deploy mode, if YAML parsing fails, return empty config instead of error.
@@ -363,6 +403,15 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 		_ = SaveConfigPreserveCommentsUpdateNestedScalar(configFile, []string{"remote-management", "secret-key"}, hashed)
 	}
 
+	cfg.RemoteManagement.PanelGitHubRepository = strings.TrimSpace(cfg.RemoteManagement.PanelGitHubRepository)
+	if cfg.RemoteManagement.PanelGitHubRepository == "" {
+		cfg.RemoteManagement.PanelGitHubRepository = DefaultPanelGitHubRepository
+	}
+
+	if cfg.LogsMaxTotalSizeMB < 0 {
+		cfg.LogsMaxTotalSizeMB = 0
+	}
+
 	// Sync request authentication providers with inline API keys for backwards compatibility.
 	syncInlineAccessProvider(&cfg)
 
@@ -411,6 +460,7 @@ func (cfg *Config) SanitizeOpenAICompatibility() {
 	for i := range cfg.OpenAICompatibility {
 		e := cfg.OpenAICompatibility[i]
 		e.Name = strings.TrimSpace(e.Name)
+		e.Prefix = normalizeModelPrefix(e.Prefix)
 		e.BaseURL = strings.TrimSpace(e.BaseURL)
 		e.Headers = NormalizeHeaders(e.Headers)
 		if e.BaseURL == "" {
@@ -431,6 +481,7 @@ func (cfg *Config) SanitizeCodexKeys() {
 	out := make([]CodexKey, 0, len(cfg.CodexKey))
 	for i := range cfg.CodexKey {
 		e := cfg.CodexKey[i]
+		e.Prefix = normalizeModelPrefix(e.Prefix)
 		e.BaseURL = strings.TrimSpace(e.BaseURL)
 		e.Headers = NormalizeHeaders(e.Headers)
 		e.ExcludedModels = NormalizeExcludedModels(e.ExcludedModels)
@@ -449,6 +500,7 @@ func (cfg *Config) SanitizeClaudeKeys() {
 	}
 	for i := range cfg.ClaudeKey {
 		entry := &cfg.ClaudeKey[i]
+		entry.Prefix = normalizeModelPrefix(entry.Prefix)
 		entry.Headers = NormalizeHeaders(entry.Headers)
 		entry.ExcludedModels = NormalizeExcludedModels(entry.ExcludedModels)
 	}
@@ -468,6 +520,7 @@ func (cfg *Config) SanitizeGeminiKeys() {
 		if entry.APIKey == "" {
 			continue
 		}
+		entry.Prefix = normalizeModelPrefix(entry.Prefix)
 		entry.BaseURL = strings.TrimSpace(entry.BaseURL)
 		entry.ProxyURL = strings.TrimSpace(entry.ProxyURL)
 		entry.Headers = NormalizeHeaders(entry.Headers)
@@ -481,6 +534,18 @@ func (cfg *Config) SanitizeGeminiKeys() {
 	cfg.GeminiKey = out
 }
 
+func normalizeModelPrefix(prefix string) string {
+	trimmed := strings.TrimSpace(prefix)
+	trimmed = strings.Trim(trimmed, "/")
+	if trimmed == "" {
+		return ""
+	}
+	if strings.Contains(trimmed, "/") {
+		return ""
+	}
+	return trimmed
+}
+
 func syncInlineAccessProvider(cfg *Config) {
 	if cfg == nil {
 		return
@@ -653,7 +718,7 @@ func sanitizeConfigForPersist(cfg *Config) *Config {
 	}
 	clone := *cfg
 	clone.SDKConfig = cfg.SDKConfig
-	clone.SDKConfig.Access = config.AccessConfig{}
+	clone.SDKConfig.Access = AccessConfig{}
 	return &clone
 }
 
@@ -752,8 +817,8 @@ 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
@@ -764,20 +829,19 @@ 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))
 		}
 	}
@@ -860,32 +924,49 @@ 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/sdk_config.go

@@ -0,0 +1,102 @@
+// Package config provides configuration management for the CLI Proxy API server.
+// It handles loading and parsing YAML configuration files, and provides structured
+// access to application settings including server port, authentication directory,
+// debug settings, proxy configuration, and API keys.
+package config
+
+// SDKConfig represents the application's configuration, loaded from a YAML file.
+type SDKConfig struct {
+	// ProxyURL is the URL of an optional proxy server to use for outbound requests.
+	ProxyURL string `yaml:"proxy-url" json:"proxy-url"`
+
+	// ForceModelPrefix requires explicit model prefixes (e.g., "teamA/gemini-3-pro-preview")
+	// to target prefixed credentials. When false, unprefixed model requests may use prefixed
+	// credentials as well.
+	ForceModelPrefix bool `yaml:"force-model-prefix" json:"force-model-prefix"`
+
+	// RequestLog enables or disables detailed request logging functionality.
+	RequestLog bool `yaml:"request-log" json:"request-log"`
+
+	// APIKeys is a list of keys for authenticating clients to this proxy server.
+	APIKeys []string `yaml:"api-keys" json:"api-keys"`
+
+	// 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"`
+}
+
+// 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.
+type AccessConfig struct {
+	// Providers lists configured authentication providers.
+	Providers []AccessProvider `yaml:"providers,omitempty" json:"providers,omitempty"`
+}
+
+// AccessProvider describes a request authentication provider entry.
+type AccessProvider struct {
+	// Name is the instance identifier for the provider.
+	Name string `yaml:"name" json:"name"`
+
+	// Type selects the provider implementation registered via the SDK.
+	Type string `yaml:"type" json:"type"`
+
+	// SDK optionally names a third-party SDK module providing this provider.
+	SDK string `yaml:"sdk,omitempty" json:"sdk,omitempty"`
+
+	// APIKeys lists inline keys for providers that require them.
+	APIKeys []string `yaml:"api-keys,omitempty" json:"api-keys,omitempty"`
+
+	// Config passes provider-specific options to the implementation.
+	Config map[string]any `yaml:"config,omitempty" json:"config,omitempty"`
+}
+
+const (
+	// AccessProviderTypeConfigAPIKey is the built-in provider validating inline API keys.
+	AccessProviderTypeConfigAPIKey = "config-api-key"
+
+	// DefaultAccessProviderName is applied when no provider name is supplied.
+	DefaultAccessProviderName = "config-inline"
+)
+
+// ConfigAPIKeyProvider returns the first inline API key provider if present.
+func (c *SDKConfig) ConfigAPIKeyProvider() *AccessProvider {
+	if c == nil {
+		return nil
+	}
+	for i := range c.Access.Providers {
+		if c.Access.Providers[i].Type == AccessProviderTypeConfigAPIKey {
+			if c.Access.Providers[i].Name == "" {
+				c.Access.Providers[i].Name = DefaultAccessProviderName
+			}
+			return &c.Access.Providers[i]
+		}
+	}
+	return nil
+}
+
+// MakeInlineAPIKeyProvider constructs an inline API key provider configuration.
+// It returns nil when no keys are supplied.
+func MakeInlineAPIKeyProvider(keys []string) *AccessProvider {
+	if len(keys) == 0 {
+		return nil
+	}
+	provider := &AccessProvider{
+		Name:    DefaultAccessProviderName,
+		Type:    AccessProviderTypeConfigAPIKey,
+		APIKeys: append([]string(nil), keys...),
+	}
+	return provider
+}

internal/config/vertex_compat.go

@@ -13,6 +13,9 @@ type VertexCompatKey struct {
 	// Maps to the x-goog-api-key header.
 	APIKey string `yaml:"api-key" json:"api-key"`
 
+	// Prefix optionally namespaces model aliases for this credential (e.g., "teamA/vertex-pro").
+	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
+
 	// BaseURL is the base URL for the Vertex-compatible API endpoint.
 	// The executor will append "/v1/publishers/google/models/{model}:action" to this.
 	// Example: "https://zenmux.ai/api" becomes "https://zenmux.ai/api/v1/publishers/google/models/..."
@@ -53,6 +56,7 @@ func (cfg *Config) SanitizeVertexCompatKeys() {
 		if entry.APIKey == "" {
 			continue
 		}
+		entry.Prefix = normalizeModelPrefix(entry.Prefix)
 		entry.BaseURL = strings.TrimSpace(entry.BaseURL)
 		if entry.BaseURL == "" {
 			// BaseURL is required for Vertex API key entries

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

@@ -24,7 +24,8 @@ 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{}
 
 // Format renders a single log entry with custom formatting.
@@ -39,11 +40,22 @@ 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)
+
 	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\n", timestamp, reqID, levelStr, filepath.Base(entry.Caller.File), entry.Caller.Line, message)
 	} else {
-		formatted = fmt.Sprintf("[%s] [%s] %s\n", timestamp, entry.Level, message)
+		formatted = fmt.Sprintf("[%s] [%s] [%s] %s\n", timestamp, reqID, levelStr, message)
 	}
 	buffer.WriteString(formatted)
 
@@ -72,39 +84,45 @@ func SetupBaseLogger() {
 }
 
 // ConfigureLogOutput switches the global log destination between rotating files and stdout.
-func ConfigureLogOutput(loggingToFile bool) error {
+// 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 {
 	SetupBaseLogger()
 
 	writerMu.Lock()
 	defer writerMu.Unlock()
 
+	logDir := "logs"
+	if base := util.WritablePath(); base != "" {
+		logDir = filepath.Join(base, "logs")
+	}
+
+	protectedPath := ""
 	if loggingToFile {
-		logDir := "logs"
-		if base := util.WritablePath(); base != "" {
-			logDir = filepath.Join(base, "logs")
-		}
 		if err := os.MkdirAll(logDir, 0o755); err != nil {
 			return fmt.Errorf("logging: failed to create log directory: %w", err)
 		}
 		if logWriter != nil {
 			_ = logWriter.Close()
 		}
+		protectedPath = filepath.Join(logDir, "main.log")
 		logWriter = &lumberjack.Logger{
-			Filename:   filepath.Join(logDir, "main.log"),
+			Filename:   protectedPath,
 			MaxSize:    10,
 			MaxBackups: 0,
 			MaxAge:     0,
 			Compress:   false,
 		}
 		log.SetOutput(logWriter)
-		return nil
+	} else {
+		if logWriter != nil {
+			_ = logWriter.Close()
+			logWriter = nil
+		}
+		log.SetOutput(os.Stdout)
 	}
 
-	if logWriter != nil {
-		_ = logWriter.Close()
-		logWriter = nil
-	}
-	log.SetOutput(os.Stdout)
+	configureLogDirCleanerLocked(logDir, logsMaxTotalSizeMB, protectedPath)
 	return nil
 }
 
@@ -112,6 +130,8 @@ func closeLogOutputs() {
 	writerMu.Lock()
 	defer writerMu.Unlock()
 
+	stopLogDirCleanerLocked()
+
 	if logWriter != nil {
 		_ = logWriter.Close()
 		logWriter = nil

internal/logging/log_dir_cleaner.go

@@ -0,0 +1,166 @@
+package logging
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+)
+
+const logDirCleanerInterval = time.Minute
+
+var logDirCleanerCancel context.CancelFunc
+
+func configureLogDirCleanerLocked(logDir string, maxTotalSizeMB int, protectedPath string) {
+	stopLogDirCleanerLocked()
+
+	if maxTotalSizeMB <= 0 {
+		return
+	}
+
+	maxBytes := int64(maxTotalSizeMB) * 1024 * 1024
+	if maxBytes <= 0 {
+		return
+	}
+
+	dir := strings.TrimSpace(logDir)
+	if dir == "" {
+		return
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	logDirCleanerCancel = cancel
+	go runLogDirCleaner(ctx, filepath.Clean(dir), maxBytes, strings.TrimSpace(protectedPath))
+}
+
+func stopLogDirCleanerLocked() {
+	if logDirCleanerCancel == nil {
+		return
+	}
+	logDirCleanerCancel()
+	logDirCleanerCancel = nil
+}
+
+func runLogDirCleaner(ctx context.Context, logDir string, maxBytes int64, protectedPath string) {
+	ticker := time.NewTicker(logDirCleanerInterval)
+	defer ticker.Stop()
+
+	cleanOnce := func() {
+		deleted, errClean := enforceLogDirSizeLimit(logDir, maxBytes, protectedPath)
+		if errClean != nil {
+			log.WithError(errClean).Warn("logging: failed to enforce log directory size limit")
+			return
+		}
+		if deleted > 0 {
+			log.Debugf("logging: removed %d old log file(s) to enforce log directory size limit", deleted)
+		}
+	}
+
+	cleanOnce()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+			cleanOnce()
+		}
+	}
+}
+
+func enforceLogDirSizeLimit(logDir string, maxBytes int64, protectedPath string) (int, error) {
+	if maxBytes <= 0 {
+		return 0, nil
+	}
+
+	dir := strings.TrimSpace(logDir)
+	if dir == "" {
+		return 0, nil
+	}
+	dir = filepath.Clean(dir)
+
+	entries, errRead := os.ReadDir(dir)
+	if errRead != nil {
+		if os.IsNotExist(errRead) {
+			return 0, nil
+		}
+		return 0, errRead
+	}
+
+	protected := strings.TrimSpace(protectedPath)
+	if protected != "" {
+		protected = filepath.Clean(protected)
+	}
+
+	type logFile struct {
+		path    string
+		size    int64
+		modTime time.Time
+	}
+
+	var (
+		files []logFile
+		total int64
+	)
+	for _, entry := range entries {
+		if entry.IsDir() {
+			continue
+		}
+		name := entry.Name()
+		if !isLogFileName(name) {
+			continue
+		}
+		info, errInfo := entry.Info()
+		if errInfo != nil {
+			continue
+		}
+		if !info.Mode().IsRegular() {
+			continue
+		}
+		path := filepath.Join(dir, name)
+		files = append(files, logFile{
+			path:    path,
+			size:    info.Size(),
+			modTime: info.ModTime(),
+		})
+		total += info.Size()
+	}
+
+	if total <= maxBytes {
+		return 0, nil
+	}
+
+	sort.Slice(files, func(i, j int) bool {
+		return files[i].modTime.Before(files[j].modTime)
+	})
+
+	deleted := 0
+	for _, file := range files {
+		if total <= maxBytes {
+			break
+		}
+		if protected != "" && filepath.Clean(file.path) == protected {
+			continue
+		}
+		if errRemove := os.Remove(file.path); errRemove != nil {
+			log.WithError(errRemove).Warnf("logging: failed to remove old log file: %s", filepath.Base(file.path))
+			continue
+		}
+		total -= file.size
+		deleted++
+	}
+
+	return deleted, nil
+}
+
+func isLogFileName(name string) bool {
+	trimmed := strings.TrimSpace(name)
+	if trimmed == "" {
+		return false
+	}
+	lower := strings.ToLower(trimmed)
+	return strings.HasSuffix(lower, ".log") || strings.HasSuffix(lower, ".log.gz")
+}

internal/logging/log_dir_cleaner_test.go

@@ -0,0 +1,70 @@
+package logging
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+func TestEnforceLogDirSizeLimitDeletesOldest(t *testing.T) {
+	dir := t.TempDir()
+
+	writeLogFile(t, filepath.Join(dir, "old.log"), 60, time.Unix(1, 0))
+	writeLogFile(t, filepath.Join(dir, "mid.log"), 60, time.Unix(2, 0))
+	protected := filepath.Join(dir, "main.log")
+	writeLogFile(t, protected, 60, time.Unix(3, 0))
+
+	deleted, err := enforceLogDirSizeLimit(dir, 120, protected)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if deleted != 1 {
+		t.Fatalf("expected 1 deleted file, got %d", deleted)
+	}
+
+	if _, err := os.Stat(filepath.Join(dir, "old.log")); !os.IsNotExist(err) {
+		t.Fatalf("expected old.log to be removed, stat error: %v", err)
+	}
+	if _, err := os.Stat(filepath.Join(dir, "mid.log")); err != nil {
+		t.Fatalf("expected mid.log to remain, stat error: %v", err)
+	}
+	if _, err := os.Stat(protected); err != nil {
+		t.Fatalf("expected protected main.log to remain, stat error: %v", err)
+	}
+}
+
+func TestEnforceLogDirSizeLimitSkipsProtected(t *testing.T) {
+	dir := t.TempDir()
+
+	protected := filepath.Join(dir, "main.log")
+	writeLogFile(t, protected, 200, time.Unix(1, 0))
+	writeLogFile(t, filepath.Join(dir, "other.log"), 50, time.Unix(2, 0))
+
+	deleted, err := enforceLogDirSizeLimit(dir, 100, protected)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if deleted != 1 {
+		t.Fatalf("expected 1 deleted file, got %d", deleted)
+	}
+
+	if _, err := os.Stat(protected); err != nil {
+		t.Fatalf("expected protected main.log to remain, stat error: %v", err)
+	}
+	if _, err := os.Stat(filepath.Join(dir, "other.log")); !os.IsNotExist(err) {
+		t.Fatalf("expected other.log to be removed, stat error: %v", err)
+	}
+}
+
+func writeLogFile(t *testing.T, path string, size int, modTime time.Time) {
+	t.Helper()
+
+	data := make([]byte, size)
+	if err := os.WriteFile(path, data, 0o644); err != nil {
+		t.Fatalf("write file: %v", err)
+	}
+	if err := os.Chtimes(path, modTime, modTime); err != nil {
+		t.Fatalf("set times: %v", err)
+	}
+}

internal/logging/request_logger.go

@@ -14,6 +14,7 @@ import (
 	"regexp"
 	"sort"
 	"strings"
+	"sync/atomic"
 	"time"
 
 	"github.com/andybalholm/brotli"
@@ -25,6 +26,8 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 )
 
+var requestLogID atomic.Uint64
+
 // RequestLogger defines the interface for logging HTTP requests and responses.
 // It provides methods for logging both regular and streaming HTTP request/response cycles.
 type RequestLogger interface {
@@ -40,10 +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.
 	//
@@ -52,11 +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.
 	//
@@ -174,20 +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
 	}
@@ -197,26 +203,59 @@ 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)
 
-	// Decompress response if needed
-	decompressedResponse, err := l.decompressResponse(responseHeaders, response)
-	if err != nil {
-		// If decompression fails, log the error but continue with original response
-		decompressedResponse = append(response, []byte(fmt.Sprintf("\n[DECOMPRESSION ERROR: %v]", err))...)
+	requestBodyPath, errTemp := l.writeRequestBodyTempFile(body)
+	if errTemp != nil {
+		log.WithError(errTemp).Warn("failed to create request body temp file, falling back to direct write")
+	}
+	if requestBodyPath != "" {
+		defer func() {
+			if errRemove := os.Remove(requestBodyPath); errRemove != nil {
+				log.WithError(errRemove).Warn("failed to remove request body temp file")
+			}
+		}()
 	}
 
-	// Create log content
-	content := l.formatLogContent(url, method, requestHeaders, body, apiRequest, apiResponse, decompressedResponse, statusCode, responseHeaders, apiResponseErrors)
+	responseToWrite, decompressErr := l.decompressResponse(responseHeaders, response)
+	if decompressErr != nil {
+		// If decompression fails, continue with original response and annotate the log output.
+		responseToWrite = response
+	}
 
-	// Write to file
-	if err = os.WriteFile(filePath, []byte(content), 0644); err != nil {
-		return fmt.Errorf("failed to write log file: %w", err)
+	logFile, errOpen := os.OpenFile(filePath, os.O_CREATE|os.O_WRONLY|os.O_TRUNC, 0644)
+	if errOpen != nil {
+		return fmt.Errorf("failed to create log file: %w", errOpen)
+	}
+
+	writeErr := l.writeNonStreamingLog(
+		logFile,
+		url,
+		method,
+		requestHeaders,
+		body,
+		requestBodyPath,
+		apiRequest,
+		apiResponse,
+		apiResponseErrors,
+		statusCode,
+		responseHeaders,
+		responseToWrite,
+		decompressErr,
+	)
+	if errClose := logFile.Close(); errClose != nil {
+		log.WithError(errClose).Warn("failed to close request log file")
+		if writeErr == nil {
+			return errClose
+		}
+	}
+	if writeErr != nil {
+		return fmt.Errorf("failed to write log file: %w", writeErr)
 	}
 
 	if force && !l.enabled {
@@ -235,11 +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
 	}
@@ -249,30 +289,42 @@ 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)
 
-	// Create and open file
-	file, err := os.Create(filePath)
-	if err != nil {
-		return nil, fmt.Errorf("failed to create log file: %w", err)
+	requestHeaders := make(map[string][]string, len(headers))
+	for key, values := range headers {
+		headerValues := make([]string, len(values))
+		copy(headerValues, values)
+		requestHeaders[key] = headerValues
 	}
 
-	// Write initial request information
-	requestInfo := l.formatRequestInfo(url, method, headers, body)
-	if _, err = file.WriteString(requestInfo); err != nil {
-		_ = file.Close()
-		return nil, fmt.Errorf("failed to write request info: %w", err)
+	requestBodyPath, errTemp := l.writeRequestBodyTempFile(body)
+	if errTemp != nil {
+		return nil, fmt.Errorf("failed to create request body temp file: %w", errTemp)
 	}
 
+	responseBodyFile, errCreate := os.CreateTemp(l.logsDir, "response-body-*.tmp")
+	if errCreate != nil {
+		_ = os.Remove(requestBodyPath)
+		return nil, fmt.Errorf("failed to create response body temp file: %w", errCreate)
+	}
+	responseBodyPath := responseBodyFile.Name()
+
 	// Create streaming writer
 	writer := &FileStreamingLogWriter{
-		file:           file,
-		chunkChan:      make(chan []byte, 100), // Buffered channel for async writes
-		closeChan:      make(chan struct{}),
-		errorChan:      make(chan error, 1),
-		bufferedChunks: &bytes.Buffer{},
+		logFilePath:      filePath,
+		url:              url,
+		method:           method,
+		timestamp:        time.Now(),
+		requestHeaders:   requestHeaders,
+		requestBodyPath:  requestBodyPath,
+		responseBodyPath: responseBodyPath,
+		responseBodyFile: responseBodyFile,
+		chunkChan:        make(chan []byte, 100), // Buffered channel for async writes
+		closeChan:        make(chan struct{}),
+		errorChan:        make(chan error, 1),
 	}
 
 	// Start async writer goroutine
@@ -282,8 +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.
@@ -298,13 +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, "?") {
@@ -320,10 +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")
 
-	return fmt.Sprintf("%s-%s.log", sanitized, timestamp)
+	// 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-%s.log", sanitized, timestamp, idPart)
 }
 
 // sanitizeForFilename replaces characters that are not safe for filenames.
@@ -405,6 +467,220 @@ func (l *FileRequestLogger) cleanupOldErrorLogs() error {
 	return nil
 }
 
+func (l *FileRequestLogger) writeRequestBodyTempFile(body []byte) (string, error) {
+	tmpFile, errCreate := os.CreateTemp(l.logsDir, "request-body-*.tmp")
+	if errCreate != nil {
+		return "", errCreate
+	}
+	tmpPath := tmpFile.Name()
+
+	if _, errCopy := io.Copy(tmpFile, bytes.NewReader(body)); errCopy != nil {
+		_ = tmpFile.Close()
+		_ = os.Remove(tmpPath)
+		return "", errCopy
+	}
+	if errClose := tmpFile.Close(); errClose != nil {
+		_ = os.Remove(tmpPath)
+		return "", errClose
+	}
+	return tmpPath, nil
+}
+
+func (l *FileRequestLogger) writeNonStreamingLog(
+	w io.Writer,
+	url, method string,
+	requestHeaders map[string][]string,
+	requestBody []byte,
+	requestBodyPath string,
+	apiRequest []byte,
+	apiResponse []byte,
+	apiResponseErrors []*interfaces.ErrorMessage,
+	statusCode int,
+	responseHeaders map[string][]string,
+	response []byte,
+	decompressErr error,
+) error {
+	if errWrite := writeRequestInfoWithBody(w, url, method, requestHeaders, requestBody, requestBodyPath, time.Now()); errWrite != nil {
+		return errWrite
+	}
+	if errWrite := writeAPISection(w, "=== API REQUEST ===\n", "=== API REQUEST", apiRequest); errWrite != nil {
+		return errWrite
+	}
+	if errWrite := writeAPIErrorResponses(w, apiResponseErrors); errWrite != nil {
+		return errWrite
+	}
+	if errWrite := writeAPISection(w, "=== API RESPONSE ===\n", "=== API RESPONSE", apiResponse); errWrite != nil {
+		return errWrite
+	}
+	return writeResponseSection(w, statusCode, true, responseHeaders, bytes.NewReader(response), decompressErr, true)
+}
+
+func writeRequestInfoWithBody(
+	w io.Writer,
+	url, method string,
+	headers map[string][]string,
+	body []byte,
+	bodyPath string,
+	timestamp time.Time,
+) error {
+	if _, errWrite := io.WriteString(w, "=== REQUEST INFO ===\n"); errWrite != nil {
+		return errWrite
+	}
+	if _, errWrite := io.WriteString(w, fmt.Sprintf("Version: %s\n", buildinfo.Version)); errWrite != nil {
+		return errWrite
+	}
+	if _, errWrite := io.WriteString(w, fmt.Sprintf("URL: %s\n", url)); errWrite != nil {
+		return errWrite
+	}
+	if _, errWrite := io.WriteString(w, fmt.Sprintf("Method: %s\n", method)); errWrite != nil {
+		return errWrite
+	}
+	if _, errWrite := io.WriteString(w, fmt.Sprintf("Timestamp: %s\n", timestamp.Format(time.RFC3339Nano))); errWrite != nil {
+		return errWrite
+	}
+	if _, errWrite := io.WriteString(w, "\n"); errWrite != nil {
+		return errWrite
+	}
+
+	if _, errWrite := io.WriteString(w, "=== HEADERS ===\n"); errWrite != nil {
+		return errWrite
+	}
+	for key, values := range headers {
+		for _, value := range values {
+			masked := util.MaskSensitiveHeaderValue(key, value)
+			if _, errWrite := io.WriteString(w, fmt.Sprintf("%s: %s\n", key, masked)); errWrite != nil {
+				return errWrite
+			}
+		}
+	}
+	if _, errWrite := io.WriteString(w, "\n"); errWrite != nil {
+		return errWrite
+	}
+
+	if _, errWrite := io.WriteString(w, "=== REQUEST BODY ===\n"); errWrite != nil {
+		return errWrite
+	}
+
+	if bodyPath != "" {
+		bodyFile, errOpen := os.Open(bodyPath)
+		if errOpen != nil {
+			return errOpen
+		}
+		if _, errCopy := io.Copy(w, bodyFile); errCopy != nil {
+			_ = bodyFile.Close()
+			return errCopy
+		}
+		if errClose := bodyFile.Close(); errClose != nil {
+			log.WithError(errClose).Warn("failed to close request body temp file")
+		}
+	} else if _, errWrite := w.Write(body); errWrite != nil {
+		return errWrite
+	}
+
+	if _, errWrite := io.WriteString(w, "\n\n"); errWrite != nil {
+		return errWrite
+	}
+	return nil
+}
+
+func writeAPISection(w io.Writer, sectionHeader string, sectionPrefix string, payload []byte) error {
+	if len(payload) == 0 {
+		return nil
+	}
+
+	if bytes.HasPrefix(payload, []byte(sectionPrefix)) {
+		if _, errWrite := w.Write(payload); errWrite != nil {
+			return errWrite
+		}
+		if !bytes.HasSuffix(payload, []byte("\n")) {
+			if _, errWrite := io.WriteString(w, "\n"); errWrite != nil {
+				return errWrite
+			}
+		}
+	} else {
+		if _, errWrite := io.WriteString(w, sectionHeader); errWrite != nil {
+			return errWrite
+		}
+		if _, errWrite := w.Write(payload); errWrite != nil {
+			return errWrite
+		}
+		if _, errWrite := io.WriteString(w, "\n"); errWrite != nil {
+			return errWrite
+		}
+	}
+
+	if _, errWrite := io.WriteString(w, "\n"); errWrite != nil {
+		return errWrite
+	}
+	return nil
+}
+
+func writeAPIErrorResponses(w io.Writer, apiResponseErrors []*interfaces.ErrorMessage) error {
+	for i := 0; i < len(apiResponseErrors); i++ {
+		if apiResponseErrors[i] == nil {
+			continue
+		}
+		if _, errWrite := io.WriteString(w, "=== API ERROR RESPONSE ===\n"); errWrite != nil {
+			return errWrite
+		}
+		if _, errWrite := io.WriteString(w, fmt.Sprintf("HTTP Status: %d\n", apiResponseErrors[i].StatusCode)); errWrite != nil {
+			return errWrite
+		}
+		if apiResponseErrors[i].Error != nil {
+			if _, errWrite := io.WriteString(w, apiResponseErrors[i].Error.Error()); errWrite != nil {
+				return errWrite
+			}
+		}
+		if _, errWrite := io.WriteString(w, "\n\n"); errWrite != nil {
+			return errWrite
+		}
+	}
+	return nil
+}
+
+func writeResponseSection(w io.Writer, statusCode int, statusWritten bool, responseHeaders map[string][]string, responseReader io.Reader, decompressErr error, trailingNewline bool) error {
+	if _, errWrite := io.WriteString(w, "=== RESPONSE ===\n"); errWrite != nil {
+		return errWrite
+	}
+	if statusWritten {
+		if _, errWrite := io.WriteString(w, fmt.Sprintf("Status: %d\n", statusCode)); errWrite != nil {
+			return errWrite
+		}
+	}
+
+	if responseHeaders != nil {
+		for key, values := range responseHeaders {
+			for _, value := range values {
+				if _, errWrite := io.WriteString(w, fmt.Sprintf("%s: %s\n", key, value)); errWrite != nil {
+					return errWrite
+				}
+			}
+		}
+	}
+
+	if _, errWrite := io.WriteString(w, "\n"); errWrite != nil {
+		return errWrite
+	}
+
+	if responseReader != nil {
+		if _, errCopy := io.Copy(w, responseReader); errCopy != nil {
+			return errCopy
+		}
+	}
+	if decompressErr != nil {
+		if _, errWrite := io.WriteString(w, fmt.Sprintf("\n[DECOMPRESSION ERROR: %v]", decompressErr)); errWrite != nil {
+			return errWrite
+		}
+	}
+
+	if trailingNewline {
+		if _, errWrite := io.WriteString(w, "\n"); errWrite != nil {
+			return errWrite
+		}
+	}
+	return nil
+}
+
 // formatLogContent creates the complete log content for non-streaming requests.
 //
 // Parameters:
@@ -648,13 +924,34 @@ func (l *FileRequestLogger) formatRequestInfo(url, method string, headers map[st
 }
 
 // FileStreamingLogWriter implements StreamingLogWriter for file-based streaming logs.
-// It handles asynchronous writing of streaming response chunks to a file.
-// All data is buffered and written in the correct order when Close is called.
+// It spools streaming response chunks to a temporary file to avoid retaining large responses in memory.
+// The final log file is assembled when Close is called.
 type FileStreamingLogWriter struct {
-	// file is the file where log data is written.
-	file *os.File
+	// logFilePath is the final log file path.
+	logFilePath string
 
-	// chunkChan is a channel for receiving response chunks to buffer.
+	// url is the request URL (masked upstream in middleware).
+	url string
+
+	// method is the HTTP method.
+	method string
+
+	// timestamp is captured when the streaming log is initialized.
+	timestamp time.Time
+
+	// requestHeaders stores the request headers.
+	requestHeaders map[string][]string
+
+	// requestBodyPath is a temporary file path holding the request body.
+	requestBodyPath string
+
+	// responseBodyPath is a temporary file path holding the streaming response body.
+	responseBodyPath string
+
+	// responseBodyFile is the temp file where chunks are appended by the async writer.
+	responseBodyFile *os.File
+
+	// chunkChan is a channel for receiving response chunks to spool.
 	chunkChan chan []byte
 
 	// closeChan is a channel for signaling when the writer is closed.
@@ -663,9 +960,6 @@ type FileStreamingLogWriter struct {
 	// errorChan is a channel for reporting errors during writing.
 	errorChan chan error
 
-	// bufferedChunks stores the response chunks in order.
-	bufferedChunks *bytes.Buffer
-
 	// responseStatus stores the HTTP status code.
 	responseStatus int
 
@@ -770,85 +1064,115 @@ func (w *FileStreamingLogWriter) Close() error {
 		close(w.chunkChan)
 	}
 
-	// Wait for async writer to finish buffering chunks
+	// Wait for async writer to finish spooling chunks
 	if w.closeChan != nil {
 		<-w.closeChan
 		w.chunkChan = nil
 	}
 
-	if w.file == nil {
+	select {
+	case errWrite := <-w.errorChan:
+		w.cleanupTempFiles()
+		return errWrite
+	default:
+	}
+
+	if w.logFilePath == "" {
+		w.cleanupTempFiles()
 		return nil
 	}
 
-	// Write all content in the correct order
-	var content strings.Builder
-
-	// 1. Write API REQUEST section
-	if len(w.apiRequest) > 0 {
-		if bytes.HasPrefix(w.apiRequest, []byte("=== API REQUEST")) {
-			content.Write(w.apiRequest)
-			if !bytes.HasSuffix(w.apiRequest, []byte("\n")) {
-				content.WriteString("\n")
-			}
-		} else {
-			content.WriteString("=== API REQUEST ===\n")
-			content.Write(w.apiRequest)
-			content.WriteString("\n")
-		}
-		content.WriteString("\n")
+	logFile, errOpen := os.OpenFile(w.logFilePath, os.O_CREATE|os.O_WRONLY|os.O_TRUNC, 0644)
+	if errOpen != nil {
+		w.cleanupTempFiles()
+		return fmt.Errorf("failed to create log file: %w", errOpen)
 	}
 
-	// 2. Write API RESPONSE section
-	if len(w.apiResponse) > 0 {
-		if bytes.HasPrefix(w.apiResponse, []byte("=== API RESPONSE")) {
-			content.Write(w.apiResponse)
-			if !bytes.HasSuffix(w.apiResponse, []byte("\n")) {
-				content.WriteString("\n")
-			}
-		} else {
-			content.WriteString("=== API RESPONSE ===\n")
-			content.Write(w.apiResponse)
-			content.WriteString("\n")
-		}
-		content.WriteString("\n")
-	}
-
-	// 3. Write RESPONSE section (status, headers, buffered chunks)
-	content.WriteString("=== RESPONSE ===\n")
-	if w.statusWritten {
-		content.WriteString(fmt.Sprintf("Status: %d\n", w.responseStatus))
-	}
-
-	for key, values := range w.responseHeaders {
-		for _, value := range values {
-			content.WriteString(fmt.Sprintf("%s: %s\n", key, value))
+	writeErr := w.writeFinalLog(logFile)
+	if errClose := logFile.Close(); errClose != nil {
+		log.WithError(errClose).Warn("failed to close request log file")
+		if writeErr == nil {
+			writeErr = errClose
 		}
 	}
-	content.WriteString("\n")
 
-	// Write buffered response body chunks
-	if w.bufferedChunks != nil && w.bufferedChunks.Len() > 0 {
-		content.Write(w.bufferedChunks.Bytes())
-	}
-
-	// Write the complete content to file
-	if _, err := w.file.WriteString(content.String()); err != nil {
-		_ = w.file.Close()
-		return err
-	}
-
-	return w.file.Close()
+	w.cleanupTempFiles()
+	return writeErr
 }
 
 // asyncWriter runs in a goroutine to buffer chunks from the channel.
-// It continuously reads chunks from the channel and buffers them for later writing.
+// It continuously reads chunks from the channel and appends them to a temp file for later assembly.
 func (w *FileStreamingLogWriter) asyncWriter() {
 	defer close(w.closeChan)
 
 	for chunk := range w.chunkChan {
-		if w.bufferedChunks != nil {
-			w.bufferedChunks.Write(chunk)
+		if w.responseBodyFile == nil {
+			continue
 		}
+		if _, errWrite := w.responseBodyFile.Write(chunk); errWrite != nil {
+			select {
+			case w.errorChan <- errWrite:
+			default:
+			}
+			if errClose := w.responseBodyFile.Close(); errClose != nil {
+				select {
+				case w.errorChan <- errClose:
+				default:
+				}
+			}
+			w.responseBodyFile = nil
+		}
+	}
+
+	if w.responseBodyFile == nil {
+		return
+	}
+	if errClose := w.responseBodyFile.Close(); errClose != nil {
+		select {
+		case w.errorChan <- errClose:
+		default:
+		}
+	}
+	w.responseBodyFile = nil
+}
+
+func (w *FileStreamingLogWriter) writeFinalLog(logFile *os.File) error {
+	if errWrite := writeRequestInfoWithBody(logFile, w.url, w.method, w.requestHeaders, nil, w.requestBodyPath, w.timestamp); errWrite != nil {
+		return errWrite
+	}
+	if errWrite := writeAPISection(logFile, "=== API REQUEST ===\n", "=== API REQUEST", w.apiRequest); errWrite != nil {
+		return errWrite
+	}
+	if errWrite := writeAPISection(logFile, "=== API RESPONSE ===\n", "=== API RESPONSE", w.apiResponse); errWrite != nil {
+		return errWrite
+	}
+
+	responseBodyFile, errOpen := os.Open(w.responseBodyPath)
+	if errOpen != nil {
+		return errOpen
+	}
+	defer func() {
+		if errClose := responseBodyFile.Close(); errClose != nil {
+			log.WithError(errClose).Warn("failed to close response body temp file")
+		}
+	}()
+
+	return writeResponseSection(logFile, w.responseStatus, w.statusWritten, w.responseHeaders, responseBodyFile, nil, false)
+}
+
+func (w *FileStreamingLogWriter) cleanupTempFiles() {
+	if w.requestBodyPath != "" {
+		if errRemove := os.Remove(w.requestBodyPath); errRemove != nil {
+			log.WithError(errRemove).Warn("failed to remove request body temp file")
+		}
+		w.requestBodyPath = ""
+	}
+
+	if w.responseBodyPath != "" {
+		if errRemove := os.Remove(w.responseBodyPath); errRemove != nil {
+			log.WithError(errRemove).Warn("failed to remove response body temp file")
+		}
+		w.responseBodyPath = ""
 	}
 }
 

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

@@ -9,6 +9,7 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"net/url"
 	"os"
 	"path/filepath"
 	"strings"
@@ -23,10 +24,10 @@ import (
 )
 
 const (
-	managementReleaseURL = "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"
+	managementAssetName         = "management.html"
+	httpUserAgent               = "CLIProxyAPI-management-updater"
+	updateCheckInterval         = 3 * time.Hour
 )
 
 // ManagementFileName exposes the control panel asset filename.
@@ -97,7 +98,7 @@ func runAutoUpdater(ctx context.Context) {
 
 		configPath, _ := schedulerConfigPath.Load().(string)
 		staticDir := StaticDir(configPath)
-		EnsureLatestManagementHTML(ctx, staticDir, cfg.ProxyURL)
+		EnsureLatestManagementHTML(ctx, staticDir, cfg.ProxyURL, cfg.RemoteManagement.PanelGitHubRepository)
 	}
 
 	runOnce()
@@ -181,7 +182,7 @@ func FilePath(configFilePath string) string {
 // EnsureLatestManagementHTML checks the latest management.html asset and updates the local copy when needed.
 // The function is designed to run in a background goroutine and will never panic.
 // It enforces a 3-hour rate limit to avoid frequent checks on config/auth file changes.
-func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL string) {
+func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL string, panelRepository string) {
 	if ctx == nil {
 		ctx = context.Background()
 	}
@@ -214,6 +215,7 @@ func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL
 		return
 	}
 
+	releaseURL := resolveReleaseURL(panelRepository)
 	client := newHTTPClient(proxyURL)
 
 	localPath := filepath.Join(staticDir, managementAssetName)
@@ -225,7 +227,7 @@ func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL
 		localHash = ""
 	}
 
-	asset, remoteHash, err := fetchLatestAsset(ctx, client)
+	asset, remoteHash, err := fetchLatestAsset(ctx, client, releaseURL)
 	if err != nil {
 		log.WithError(err).Warn("failed to fetch latest management release information")
 		return
@@ -254,8 +256,44 @@ func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL
 	log.Infof("management asset updated successfully (hash=%s)", downloadedHash)
 }
 
-func fetchLatestAsset(ctx context.Context, client *http.Client) (*releaseAsset, string, error) {
-	req, err := http.NewRequestWithContext(ctx, http.MethodGet, managementReleaseURL, nil)
+func resolveReleaseURL(repo string) string {
+	repo = strings.TrimSpace(repo)
+	if repo == "" {
+		return defaultManagementReleaseURL
+	}
+
+	parsed, err := url.Parse(repo)
+	if err != nil || parsed.Host == "" {
+		return defaultManagementReleaseURL
+	}
+
+	host := strings.ToLower(parsed.Host)
+	parsed.Path = strings.TrimSuffix(parsed.Path, "/")
+
+	if host == "api.github.com" {
+		if !strings.HasSuffix(strings.ToLower(parsed.Path), "/releases/latest") {
+			parsed.Path = parsed.Path + "/releases/latest"
+		}
+		return parsed.String()
+	}
+
+	if host == "github.com" {
+		parts := strings.Split(strings.Trim(parsed.Path, "/"), "/")
+		if len(parts) >= 2 && parts[0] != "" && parts[1] != "" {
+			repoName := strings.TrimSuffix(parts[1], ".git")
+			return fmt.Sprintf("https://api.github.com/repos/%s/%s/releases/latest", parts[0], repoName)
+		}
+	}
+
+	return defaultManagementReleaseURL
+}
+
+func fetchLatestAsset(ctx context.Context, client *http.Client, releaseURL string) (*releaseAsset, string, error) {
+	if strings.TrimSpace(releaseURL) == "" {
+		releaseURL = defaultManagementReleaseURL
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, releaseURL, nil)
 	if err != nil {
 		return nil, "", fmt.Errorf("create release request: %w", err)
 	}

internal/misc/codex_instructions.go

@@ -20,6 +20,7 @@ func CodexInstructionsForModel(modelName, systemInstructions string) (bool, stri
 	lastCodexMaxPrompt := ""
 	last51Prompt := ""
 	last52Prompt := ""
+	last52CodexPrompt := ""
 	// lastReviewPrompt := ""
 	for _, entry := range entries {
 		content, _ := codexInstructionsDir.ReadFile("codex_instructions/" + entry.Name())
@@ -36,12 +37,16 @@ func CodexInstructionsForModel(modelName, systemInstructions string) (bool, stri
 			last51Prompt = string(content)
 		} else if strings.HasPrefix(entry.Name(), "gpt_5_2_prompt.md") {
 			last52Prompt = string(content)
+		} else if strings.HasPrefix(entry.Name(), "gpt-5.2-codex_prompt.md") {
+			last52CodexPrompt = string(content)
 		} else if strings.HasPrefix(entry.Name(), "review_prompt.md") {
 			// lastReviewPrompt = string(content)
 		}
 	}
 	if strings.Contains(modelName, "codex-max") {
 		return false, lastCodexMaxPrompt
+	} else if strings.Contains(modelName, "5.2-codex") {
+		return false, last52CodexPrompt
 	} else if strings.Contains(modelName, "codex") {
 		return false, lastCodexPrompt
 	} else if strings.Contains(modelName, "5.1") {

internal/misc/codex_instructions/gpt-5.2-codex_prompt.md-001-f084e5264b1b0ae9eb8c63c950c0953f40966fed

@@ -0,0 +1,117 @@
+You are Codex, based on GPT-5. You are running as a coding agent in the Codex CLI on a user's computer.
+
+## General
+
+- 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.)
+
+## Editing constraints
+
+- Default to ASCII when editing or creating files. Only introduce non-ASCII or other Unicode characters when there is a clear justification and the file already uses them.
+- Add succinct code comments that explain what is going on if code is not self-explanatory. You should not add comments like "Assigns the value to the variable", but a brief comment might be useful ahead of a complex code block that the user would otherwise have to spend time parsing out. Usage of these comments should be rare.
+- Try to use apply_patch for single file edits, but it is fine to explore other options to make the edit if it does not work well. Do not use apply_patch for changes that are auto-generated (i.e. generating package.json or running a lint or format command like gofmt) or when scripting is more efficient (such as search and replacing a string across a codebase).
+- You may be in a dirty git worktree.
+    * NEVER revert existing changes you did not make unless explicitly requested, since these changes were made by the user.
+    * If asked to make a commit or code edits and there are unrelated changes to your work or changes that you didn't make in those files, don't revert those changes.
+    * If the changes are in files you've touched recently, you should read carefully and understand how you can work with the changes rather than reverting them.
+    * If the changes are in unrelated files, just ignore them and don't revert them.
+- Do not amend a commit unless explicitly requested to do so.
+- While you are working, you might notice unexpected changes that you didn't make. If this happens, STOP IMMEDIATELY and ask the user how they would like to proceed.
+- **NEVER** use destructive commands like `git reset --hard` or `git checkout --` unless specifically requested or approved by the user.
+
+## Plan tool
+
+When using the planning tool:
+- Skip using the planning tool for straightforward tasks (roughly the easiest 25%).
+- Do not make single-step plans.
+- When you made a plan, update it after having performed one of the sub-tasks that you shared on the plan.
+
+## Codex CLI harness, sandboxing, and approvals
+
+The Codex CLI harness supports several different configurations for sandboxing and escalation approvals that the user can choose from.
+
+Filesystem sandboxing defines which files can be read or written. The options for `sandbox_mode` are:
+- **read-only**: The sandbox only permits reading files.
+- **workspace-write**: The sandbox permits reading files, and editing files in `cwd` and `writable_roots`. Editing files in other directories requires approval.
+- **danger-full-access**: No filesystem sandboxing - all commands are permitted.
+
+Network sandboxing defines whether network can be accessed without approval. Options for `network_access` are:
+- **restricted**: Requires approval
+- **enabled**: No approval needed
+
+Approvals are your mechanism to get user consent to run shell commands without the sandbox. Possible configuration options for `approval_policy` 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 paired 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 `approval_policy == 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 /var)
+- 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. ALWAYS proceed to use the `sandbox_permissions` and `justification` parameters - do not message the user before requesting approval for the command.
+- 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)
+
+When `sandbox_mode` 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 enabled, and approval on-failure.
+
+Although they introduce friction to the user because your work is paused until the user responds, you should leverage them when necessary to accomplish important work. If the completing the task requires escalated permissions, Do not let these settings or the sandbox deter you from attempting to accomplish the user's task unless it is set to "never", in which case never ask for approvals.
+
+When requesting approval to execute a command that will require escalated privileges:
+  - Provide the `sandbox_permissions` parameter with the value `"require_escalated"`
+  - Include a short, 1 sentence explanation for why you need escalated permissions in the justification parameter
+
+## Special user requests
+
+- If the user makes a simple request (such as asking for the time) which you can fulfill by running a terminal command (such as `date`), you should do so.
+- If the user asks for a "review", default to a code review mindset: prioritise identifying bugs, risks, behavioural regressions, and missing tests. Findings must be the primary focus of the response - keep summaries or overviews brief and only after enumerating the issues. Present findings first (ordered by severity with file/line references), follow with open questions or assumptions, and offer a change-summary only as a secondary detail. If no findings are discovered, state that explicitly and mention any residual risks or testing gaps.
+
+## Frontend tasks
+When doing frontend design tasks, avoid collapsing into "AI slop" or safe, average-looking layouts.
+Aim for interfaces that feel intentional, bold, and a bit surprising.
+- Typography: Use expressive, purposeful fonts and avoid default stacks (Inter, Roboto, Arial, system).
+- Color & Look: Choose a clear visual direction; define CSS variables; avoid purple-on-white defaults. No purple bias or dark mode bias.
+- Motion: Use a few meaningful animations (page-load, staggered reveals) instead of generic micro-motions.
+- Background: Don't rely on flat, single-color backgrounds; use gradients, shapes, or subtle patterns to build atmosphere.
+- Overall: Avoid boilerplate layouts and interchangeable UI patterns. Vary themes, type families, and visual languages across outputs.
+- Ensure the page loads properly on both desktop and mobile
+
+Exception: If working within an existing website or design system, preserve the established patterns, structure, and visual language.
+
+## Presenting your work and final message
+
+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.
+
+- Default: be very concise; friendly coding teammate tone.
+- Ask only when needed; suggest ideas; mirror the user's style.
+- For substantial work, summarize clearly; follow final‑answer formatting.
+- Skip heavy formatting for simple confirmations.
+- Don't dump large files you've written; reference paths only.
+- No "save/copy this file" - User is on the same machine.
+- Offer logical next steps (tests, commits, build) briefly; add verify steps if you couldn't do something.
+- For code changes:
+  * Lead with a quick explanation of the change, and then give more details on the context covering where and why a change was made. Do not start this explanation with "summary", just jump right in.
+  * If there are natural next steps the user may want to take, suggest them at the end of your response. Do not make suggestions if there are no natural next steps.
+  * When suggesting multiple options, use numeric lists for the suggestions so the user can quickly respond with a single number.
+- The user does not command execution outputs. When asked to show the output of a command (e.g. `git show`), relay the important details in your answer or summarize the key lines so the user understands the result.
+
+### Final answer structure and style guidelines
+
+- Plain text; CLI handles styling. Use structure only when it helps scanability.
+- Headers: optional; short Title Case (1-3 words) wrapped in **…**; no blank line before the first bullet; add only if they truly help.
+- Bullets: use - ; merge related points; keep to one line when possible; 4–6 per list ordered by importance; keep phrasing consistent.
+- Monospace: backticks for commands/paths/env vars/code ids and inline examples; use for literal keyword bullets; never combine with **.
+- Code samples or multi-line snippets should be wrapped in fenced code blocks; include an info string as often as possible.
+- Structure: group related bullets; order sections general → specific → supporting; for subsections, start with a bolded keyword bullet, then items; match complexity to the task.
+- Tone: collaborative, concise, factual; present tense, active voice; self‑contained; no "above/below"; parallel wording.
+- Don'ts: no nested bullets/hierarchies; no ANSI codes; don't cram unrelated keywords; keep keyword lists short—wrap/reformat if long; avoid naming formatting styles in answers.
+- Adaptation: code explanations → precise, structured with code refs; simple tasks → lead with outcome; big changes → logical walkthrough + rationale + next actions; casual one-offs → plain sentences, no headers/bullets.
+- File References: When referencing files in your response follow the below rules:
+  * Use inline code to make file paths clickable.
+  * Each reference should have a stand alone path. Even if it's the same file.
+  * Accepted: absolute, workspace‑relative, a/ or b/ diff prefixes, or bare filename/suffix.
+  * Optionally include line/column (1‑based): :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

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/registry/model_definitions.go

@@ -160,7 +160,22 @@ func GetGeminiModels() []*ModelInfo {
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
+		{
+			ID:                         "gemini-3-flash-preview",
+			Object:                     "model",
+			Created:                    1765929600,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-flash-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Flash Preview",
+			Description:                "Gemini 3 Flash Preview",
+			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"}},
 		},
 		{
 			ID:                         "gemini-3-pro-image-preview",
@@ -175,7 +190,7 @@ func GetGeminiModels() []*ModelInfo {
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
 		},
 	}
 }
@@ -240,7 +255,22 @@ func GetGeminiVertexModels() []*ModelInfo {
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
+		{
+			ID:                         "gemini-3-flash-preview",
+			Object:                     "model",
+			Created:                    1765929600,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-flash-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Flash Preview",
+			Description:                "Our most intelligent model built for speed, combining frontier intelligence with superior search and grounding.",
+			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"}},
 		},
 		{
 			ID:                         "gemini-3-pro-image-preview",
@@ -255,7 +285,7 @@ func GetGeminiVertexModels() []*ModelInfo {
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
 		},
 	}
 }
@@ -317,11 +347,26 @@ func GetGeminiCLIModels() []*ModelInfo {
 			Name:                       "models/gemini-3-pro-preview",
 			Version:                    "3.0",
 			DisplayName:                "Gemini 3 Pro Preview",
-			Description:                "Gemini 3 Pro Preview",
+			Description:                "Our most intelligent model with SOTA reasoning and multimodal understanding, and powerful agentic and vibe coding capabilities",
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
+		{
+			ID:                         "gemini-3-flash-preview",
+			Object:                     "model",
+			Created:                    1765929600,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-flash-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Flash Preview",
+			Description:                "Our most intelligent model built for speed, combining frontier intelligence with superior search and grounding.",
+			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"}},
 		},
 	}
 }
@@ -387,7 +432,22 @@ func GetAIStudioModels() []*ModelInfo {
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
+		{
+			ID:                         "gemini-3-flash-preview",
+			Object:                     "model",
+			Created:                    1765929600,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-flash-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Flash Preview",
+			Description:                "Our most intelligent model built for speed, combining frontier intelligence with superior search and grounding.",
+			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"}},
 		},
 		{
 			ID:                         "gemini-pro-latest",
@@ -582,6 +642,20 @@ func GetOpenAIModels() []*ModelInfo {
 			SupportedParameters: []string{"tools"},
 			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high", "xhigh"}},
 		},
+		{
+			ID:                  "gpt-5.2-codex",
+			Object:              "model",
+			Created:             1765440000,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5.2",
+			DisplayName:         "GPT 5.2 Codex",
+			Description:         "Stable version of GPT 5.2 Codex, The best model for coding and agentic tasks across domains.",
+			ContextLength:       400000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
+		},
 	}
 }
 
@@ -630,6 +704,13 @@ func GetQwenModels() []*ModelInfo {
 	}
 }
 
+// iFlowThinkingSupport is a shared ThinkingSupport configuration for iFlow models
+// that support thinking mode via chat_template_kwargs.enable_thinking (boolean toggle).
+// Uses level-based configuration so standard normalization flows apply before conversion.
+var iFlowThinkingSupport = &ThinkingSupport{
+	Levels: []string{"none", "auto", "minimal", "low", "medium", "high", "xhigh"},
+}
+
 // GetIFlowModels returns supported models for iFlow OAuth accounts.
 func GetIFlowModels() []*ModelInfo {
 	entries := []struct {
@@ -645,9 +726,10 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "qwen3-vl-plus", DisplayName: "Qwen3-VL-Plus", Description: "Qwen3 multimodal vision-language", Created: 1758672000},
 		{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},
+		{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, Thinking: &ThinkingSupport{Levels: []string{"low", "medium", "high"}}},
+		{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},
 		{ID: "deepseek-v3.2-reasoner", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2 Reasoner", Created: 1764576000},
 		{ID: "deepseek-v3.2", DisplayName: "DeepSeek-V3.2-Exp", Description: "DeepSeek V3.2 experimental", Created: 1759104000},
@@ -655,10 +737,11 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "deepseek-r1", DisplayName: "DeepSeek-R1", Description: "DeepSeek reasoning model R1", Created: 1737331200},
 		{ID: "deepseek-v3", DisplayName: "DeepSeek-V3-671B", Description: "DeepSeek V3 671B", Created: 1734307200},
 		{ID: "qwen3-32b", DisplayName: "Qwen3-32B", Description: "Qwen3 32B", Created: 1747094400},
-		{ID: "qwen3-235b-a22b-thinking-2507", DisplayName: "Qwen3-235B-A22B-Thinking", Description: "Qwen3 235B A22B Thinking (2507)", Created: 1753401600, Thinking: &ThinkingSupport{Levels: []string{"low", "medium", "high"}}},
+		{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, Thinking: &ThinkingSupport{Levels: []string{"low", "medium", "high"}}},
+		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000},
+		{ID: "minimax-m2.1", DisplayName: "MiniMax-M2.1", Description: "MiniMax M2.1", Created: 1766448000, Thinking: iFlowThinkingSupport},
 	}
 	models := make([]*ModelInfo, 0, len(entries))
 	for _, entry := range entries {
@@ -691,8 +774,9 @@ func GetAntigravityModelConfig() 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}, Name: "models/gemini-3-pro-preview"},
-		"gemini-3-pro-image-preview":              {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true}, Name: "models/gemini-3-pro-image-preview"},
+		"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},
 	}

internal/runtime/executor/aistudio_executor.go

@@ -323,9 +323,10 @@ func (e *AIStudioExecutor) translateRequest(req cliproxyexecutor.Request, opts c
 	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)
-	payload = util.NormalizeGeminiThinkingBudget(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)

internal/runtime/executor/antigravity_executor.go

@@ -7,6 +7,8 @@ import (
 	"bufio"
 	"bytes"
 	"context"
+	"crypto/sha256"
+	"encoding/binary"
 	"encoding/json"
 	"fmt"
 	"io"
@@ -15,6 +17,7 @@ import (
 	"net/url"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/google/uuid"
@@ -30,20 +33,24 @@ import (
 )
 
 const (
-	antigravityBaseURLDaily = "https://daily-cloudcode-pa.sandbox.googleapis.com"
-	// antigravityBaseURLAutopush     = "https://autopush-cloudcode-pa.sandbox.googleapis.com"
-	antigravityBaseURLProd  = "https://cloudcode-pa.googleapis.com"
-	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
 )
 
-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 {
@@ -69,6 +76,10 @@ func (e *AntigravityExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Au
 
 // 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") {
+		return e.executeClaudeNonStream(ctx, auth, req, opts)
+	}
+
 	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
 	if errToken != nil {
 		return resp, errToken
@@ -85,8 +96,10 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	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)
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
@@ -160,6 +173,338 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	return resp, err
 }
 
+// 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) {
+	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
+	if errToken != nil {
+		return resp, errToken
+	}
+	if updatedAuth != nil {
+		auth = updatedAuth
+	}
+
+	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, 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)
+
+	baseURLs := antigravityBaseURLFallbackOrder(auth)
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+
+	var lastStatus int
+	var lastBody []byte
+	var lastErr error
+
+	for idx, baseURL := range baseURLs {
+		httpReq, errReq := e.buildRequest(ctx, auth, token, req.Model, translated, true, opts.Alt, baseURL)
+		if errReq != nil {
+			err = errReq
+			return resp, err
+		}
+
+		httpResp, errDo := httpClient.Do(httpReq)
+		if errDo != nil {
+			recordAPIResponseError(ctx, e.cfg, errDo)
+			lastStatus = 0
+			lastBody = nil
+			lastErr = errDo
+			if idx+1 < len(baseURLs) {
+				log.Debugf("antigravity executor: request error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+				continue
+			}
+			err = errDo
+			return resp, err
+		}
+		recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+		if httpResp.StatusCode < http.StatusOK || httpResp.StatusCode >= http.StatusMultipleChoices {
+			bodyBytes, errRead := io.ReadAll(httpResp.Body)
+			if errClose := httpResp.Body.Close(); errClose != nil {
+				log.Errorf("antigravity executor: close response body error: %v", errClose)
+			}
+			if errRead != nil {
+				recordAPIResponseError(ctx, e.cfg, errRead)
+				lastStatus = 0
+				lastBody = nil
+				lastErr = errRead
+				if idx+1 < len(baseURLs) {
+					log.Debugf("antigravity executor: read error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+					continue
+				}
+				err = errRead
+				return resp, err
+			}
+			appendAPIResponseChunk(ctx, e.cfg, bodyBytes)
+			lastStatus = httpResp.StatusCode
+			lastBody = append([]byte(nil), bodyBytes...)
+			lastErr = nil
+			if httpResp.StatusCode == http.StatusTooManyRequests && idx+1 < len(baseURLs) {
+				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)}
+			return resp, err
+		}
+
+		out := make(chan cliproxyexecutor.StreamChunk)
+		go func(resp *http.Response) {
+			defer close(out)
+			defer func() {
+				if errClose := resp.Body.Close(); errClose != nil {
+					log.Errorf("antigravity executor: close response body error: %v", errClose)
+				}
+			}()
+			scanner := bufio.NewScanner(resp.Body)
+			scanner.Buffer(nil, streamScannerBuffer)
+			for scanner.Scan() {
+				line := scanner.Bytes()
+				appendAPIResponseChunk(ctx, e.cfg, line)
+
+				// Filter usage metadata for all models
+				// Only retain usage statistics in the terminal chunk
+				line = FilterSSEUsageMetadata(line)
+
+				payload := jsonPayload(line)
+				if payload == nil {
+					continue
+				}
+
+				if detail, ok := parseAntigravityStreamUsage(payload); ok {
+					reporter.publish(ctx, detail)
+				}
+
+				out <- cliproxyexecutor.StreamChunk{Payload: payload}
+			}
+			if errScan := scanner.Err(); errScan != nil {
+				recordAPIResponseError(ctx, e.cfg, errScan)
+				reporter.publishFailure(ctx)
+				out <- cliproxyexecutor.StreamChunk{Err: errScan}
+			} else {
+				reporter.ensurePublished(ctx)
+			}
+		}(httpResp)
+
+		var buffer bytes.Buffer
+		for chunk := range out {
+			if chunk.Err != nil {
+				return resp, chunk.Err
+			}
+			if len(chunk.Payload) > 0 {
+				_, _ = buffer.Write(chunk.Payload)
+				_, _ = buffer.Write([]byte("\n"))
+			}
+		}
+		resp = cliproxyexecutor.Response{Payload: e.convertStreamToNonStream(buffer.Bytes())}
+
+		reporter.publish(ctx, parseAntigravityUsage(resp.Payload))
+		var param any
+		converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, resp.Payload, &param)
+		resp = cliproxyexecutor.Response{Payload: []byte(converted)}
+		reporter.ensurePublished(ctx)
+
+		return resp, nil
+	}
+
+	switch {
+	case lastStatus != 0:
+		err = statusErr{code: lastStatus, msg: string(lastBody)}
+	case lastErr != nil:
+		err = lastErr
+	default:
+		err = statusErr{code: http.StatusServiceUnavailable, msg: "antigravity executor: no base url available"}
+	}
+	return resp, err
+}
+
+func (e *AntigravityExecutor) convertStreamToNonStream(stream []byte) []byte {
+	responseTemplate := ""
+	var traceID string
+	var finishReason string
+	var modelVersion string
+	var responseID string
+	var role string
+	var usageRaw string
+	parts := make([]map[string]interface{}, 0)
+	var pendingKind string
+	var pendingText strings.Builder
+	var pendingThoughtSig string
+
+	flushPending := func() {
+		if pendingKind == "" {
+			return
+		}
+		text := pendingText.String()
+		switch pendingKind {
+		case "text":
+			if strings.TrimSpace(text) == "" {
+				pendingKind = ""
+				pendingText.Reset()
+				pendingThoughtSig = ""
+				return
+			}
+			parts = append(parts, map[string]interface{}{"text": text})
+		case "thought":
+			if strings.TrimSpace(text) == "" && pendingThoughtSig == "" {
+				pendingKind = ""
+				pendingText.Reset()
+				pendingThoughtSig = ""
+				return
+			}
+			part := map[string]interface{}{"thought": true}
+			part["text"] = text
+			if pendingThoughtSig != "" {
+				part["thoughtSignature"] = pendingThoughtSig
+			}
+			parts = append(parts, part)
+		}
+		pendingKind = ""
+		pendingText.Reset()
+		pendingThoughtSig = ""
+	}
+
+	normalizePart := func(partResult gjson.Result) map[string]interface{} {
+		var m map[string]interface{}
+		_ = json.Unmarshal([]byte(partResult.Raw), &m)
+		if m == nil {
+			m = map[string]interface{}{}
+		}
+		sig := partResult.Get("thoughtSignature").String()
+		if sig == "" {
+			sig = partResult.Get("thought_signature").String()
+		}
+		if sig != "" {
+			m["thoughtSignature"] = sig
+			delete(m, "thought_signature")
+		}
+		if inlineData, ok := m["inline_data"]; ok {
+			m["inlineData"] = inlineData
+			delete(m, "inline_data")
+		}
+		return m
+	}
+
+	for _, line := range bytes.Split(stream, []byte("\n")) {
+		trimmed := bytes.TrimSpace(line)
+		if len(trimmed) == 0 || !gjson.ValidBytes(trimmed) {
+			continue
+		}
+
+		root := gjson.ParseBytes(trimmed)
+		responseNode := root.Get("response")
+		if !responseNode.Exists() {
+			if root.Get("candidates").Exists() {
+				responseNode = root
+			} else {
+				continue
+			}
+		}
+		responseTemplate = responseNode.Raw
+
+		if traceResult := root.Get("traceId"); traceResult.Exists() && traceResult.String() != "" {
+			traceID = traceResult.String()
+		}
+
+		if roleResult := responseNode.Get("candidates.0.content.role"); roleResult.Exists() {
+			role = roleResult.String()
+		}
+
+		if finishResult := responseNode.Get("candidates.0.finishReason"); finishResult.Exists() && finishResult.String() != "" {
+			finishReason = finishResult.String()
+		}
+
+		if modelResult := responseNode.Get("modelVersion"); modelResult.Exists() && modelResult.String() != "" {
+			modelVersion = modelResult.String()
+		}
+		if responseIDResult := responseNode.Get("responseId"); responseIDResult.Exists() && responseIDResult.String() != "" {
+			responseID = responseIDResult.String()
+		}
+		if usageResult := responseNode.Get("usageMetadata"); usageResult.Exists() {
+			usageRaw = usageResult.Raw
+		} else if usageResult := root.Get("usageMetadata"); usageResult.Exists() {
+			usageRaw = usageResult.Raw
+		}
+
+		if partsResult := responseNode.Get("candidates.0.content.parts"); partsResult.IsArray() {
+			for _, part := range partsResult.Array() {
+				hasFunctionCall := part.Get("functionCall").Exists()
+				hasInlineData := part.Get("inlineData").Exists() || part.Get("inline_data").Exists()
+				sig := part.Get("thoughtSignature").String()
+				if sig == "" {
+					sig = part.Get("thought_signature").String()
+				}
+				text := part.Get("text").String()
+				thought := part.Get("thought").Bool()
+
+				if hasFunctionCall || hasInlineData {
+					flushPending()
+					parts = append(parts, normalizePart(part))
+					continue
+				}
+
+				if thought || part.Get("text").Exists() {
+					kind := "text"
+					if thought {
+						kind = "thought"
+					}
+					if pendingKind != "" && pendingKind != kind {
+						flushPending()
+					}
+					pendingKind = kind
+					pendingText.WriteString(text)
+					if kind == "thought" && sig != "" {
+						pendingThoughtSig = sig
+					}
+					continue
+				}
+
+				flushPending()
+				parts = append(parts, normalizePart(part))
+			}
+		}
+	}
+	flushPending()
+
+	if responseTemplate == "" {
+		responseTemplate = `{"candidates":[{"content":{"role":"model","parts":[]}}]}`
+	}
+
+	partsJSON, _ := json.Marshal(parts)
+	responseTemplate, _ = sjson.SetRaw(responseTemplate, "candidates.0.content.parts", string(partsJSON))
+	if role != "" {
+		responseTemplate, _ = sjson.Set(responseTemplate, "candidates.0.content.role", role)
+	}
+	if finishReason != "" {
+		responseTemplate, _ = sjson.Set(responseTemplate, "candidates.0.finishReason", finishReason)
+	}
+	if modelVersion != "" {
+		responseTemplate, _ = sjson.Set(responseTemplate, "modelVersion", modelVersion)
+	}
+	if responseID != "" {
+		responseTemplate, _ = sjson.Set(responseTemplate, "responseId", responseID)
+	}
+	if usageRaw != "" {
+		responseTemplate, _ = sjson.SetRaw(responseTemplate, "usageMetadata", usageRaw)
+	} else if !gjson.Get(responseTemplate, "usageMetadata").Exists() {
+		responseTemplate, _ = sjson.Set(responseTemplate, "usageMetadata.promptTokenCount", 0)
+		responseTemplate, _ = sjson.Set(responseTemplate, "usageMetadata.candidatesTokenCount", 0)
+		responseTemplate, _ = sjson.Set(responseTemplate, "usageMetadata.totalTokenCount", 0)
+	}
+
+	output := `{"response":{},"traceId":""}`
+	output, _ = sjson.SetRaw(output, "response", responseTemplate)
+	if traceID != "" {
+		output, _ = sjson.Set(output, "traceId", traceID)
+	}
+	return []byte(output)
+}
+
 // 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) {
 	ctx = context.WithValue(ctx, "alt", "")
@@ -180,8 +525,10 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	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)
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
@@ -312,9 +659,131 @@ func (e *AntigravityExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Au
 	return updated, nil
 }
 
-// CountTokens counts tokens for the given request (not supported for Antigravity).
-func (e *AntigravityExecutor) CountTokens(context.Context, *cliproxyauth.Auth, cliproxyexecutor.Request, cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
-	return cliproxyexecutor.Response{}, statusErr{code: http.StatusNotImplemented, msg: "count tokens not supported"}
+// 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) {
+	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
+	if errToken != nil {
+		return cliproxyexecutor.Response{}, errToken
+	}
+	if updatedAuth != nil {
+		auth = updatedAuth
+	}
+	if strings.TrimSpace(token) == "" {
+		return cliproxyexecutor.Response{}, statusErr{code: http.StatusUnauthorized, msg: "missing access token"}
+	}
+
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("antigravity")
+	respCtx := context.WithValue(ctx, "alt", opts.Alt)
+
+	baseURLs := antigravityBaseURLFallbackOrder(auth)
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+
+	var authID, authLabel, authType, authValue string
+	if auth != nil {
+		authID = auth.ID
+		authLabel = auth.Label
+		authType, authValue = auth.AccountInfo()
+	}
+
+	var lastStatus int
+	var lastBody []byte
+	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)
+		}
+
+		var requestURL strings.Builder
+		requestURL.WriteString(base)
+		requestURL.WriteString(antigravityCountTokensPath)
+		if opts.Alt != "" {
+			requestURL.WriteString("?$alt=")
+			requestURL.WriteString(url.QueryEscape(opts.Alt))
+		}
+
+		httpReq, errReq := http.NewRequestWithContext(ctx, http.MethodPost, requestURL.String(), bytes.NewReader(payload))
+		if errReq != nil {
+			return cliproxyexecutor.Response{}, errReq
+		}
+		httpReq.Header.Set("Content-Type", "application/json")
+		httpReq.Header.Set("Authorization", "Bearer "+token)
+		httpReq.Header.Set("User-Agent", resolveUserAgent(auth))
+		httpReq.Header.Set("Accept", "application/json")
+		if host := resolveHost(base); host != "" {
+			httpReq.Host = host
+		}
+
+		recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
+			URL:       requestURL.String(),
+			Method:    http.MethodPost,
+			Headers:   httpReq.Header.Clone(),
+			Body:      payload,
+			Provider:  e.Identifier(),
+			AuthID:    authID,
+			AuthLabel: authLabel,
+			AuthType:  authType,
+			AuthValue: authValue,
+		})
+
+		httpResp, errDo := httpClient.Do(httpReq)
+		if errDo != nil {
+			recordAPIResponseError(ctx, e.cfg, errDo)
+			lastStatus = 0
+			lastBody = nil
+			lastErr = errDo
+			if idx+1 < len(baseURLs) {
+				log.Debugf("antigravity executor: request error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+				continue
+			}
+			return cliproxyexecutor.Response{}, errDo
+		}
+
+		recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+		bodyBytes, errRead := io.ReadAll(httpResp.Body)
+		if errClose := httpResp.Body.Close(); errClose != nil {
+			log.Errorf("antigravity executor: close response body error: %v", errClose)
+		}
+		if errRead != nil {
+			recordAPIResponseError(ctx, e.cfg, errRead)
+			return cliproxyexecutor.Response{}, errRead
+		}
+		appendAPIResponseChunk(ctx, e.cfg, bodyBytes)
+
+		if httpResp.StatusCode >= http.StatusOK && httpResp.StatusCode < http.StatusMultipleChoices {
+			count := gjson.GetBytes(bodyBytes, "totalTokens").Int()
+			translated := sdktranslator.TranslateTokenCount(respCtx, to, from, count, bodyBytes)
+			return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
+		}
+
+		lastStatus = httpResp.StatusCode
+		lastBody = append([]byte(nil), bodyBytes...)
+		lastErr = nil
+		if httpResp.StatusCode == http.StatusTooManyRequests && idx+1 < len(baseURLs) {
+			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)}
+	}
+
+	switch {
+	case lastStatus != 0:
+		return cliproxyexecutor.Response{}, statusErr{code: lastStatus, msg: string(lastBody)}
+	case lastErr != nil:
+		return cliproxyexecutor.Response{}, lastErr
+	default:
+		return cliproxyexecutor.Response{}, statusErr{code: http.StatusServiceUnavailable, msg: "antigravity executor: no base url available"}
+	}
 }
 
 // FetchAntigravityModels retrieves available models using the supplied auth.
@@ -545,27 +1014,9 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 			strJSON, _ = util.RenameKey(strJSON, p, p[:len(p)-len("parametersJsonSchema")]+"parameters")
 		}
 
-		strJSON = util.DeleteKey(strJSON, "$schema")
-		strJSON = util.DeleteKey(strJSON, "maxItems")
-		strJSON = util.DeleteKey(strJSON, "minItems")
-		strJSON = util.DeleteKey(strJSON, "minLength")
-		strJSON = util.DeleteKey(strJSON, "maxLength")
-		strJSON = util.DeleteKey(strJSON, "exclusiveMinimum")
-		strJSON = util.DeleteKey(strJSON, "exclusiveMaximum")
-		strJSON = util.DeleteKey(strJSON, "$ref")
-		strJSON = util.DeleteKey(strJSON, "$defs")
-
-		paths = make([]string, 0)
-		util.Walk(gjson.Parse(strJSON), "", "anyOf", &paths)
-		for _, p := range paths {
-			anyOf := gjson.Get(strJSON, p)
-			if anyOf.IsArray() {
-				anyOfItems := anyOf.Array()
-				if len(anyOfItems) > 0 {
-					strJSON, _ = sjson.SetRaw(strJSON, p[:len(p)-len(".anyOf")], anyOfItems[0].Raw)
-				}
-			}
-		}
+		// Use the centralized schema cleaner to handle unsupported keywords,
+		// const->enum conversion, and flattening of types/anyOf.
+		strJSON = util.CleanJSONSchemaForAntigravity(strJSON)
 
 		payload = []byte(strJSON)
 	}
@@ -705,7 +1156,7 @@ func antigravityBaseURLFallbackOrder(auth *cliproxyauth.Auth) []string {
 	}
 	return []string{
 		antigravityBaseURLDaily,
-		// antigravityBaseURLAutopush,
+		antigravitySandboxBaseURLDaily,
 		antigravityBaseURLProd,
 	}
 }
@@ -741,7 +1192,7 @@ 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")
@@ -777,15 +1228,36 @@ 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
 }
@@ -798,6 +1270,8 @@ func modelName2Alias(modelName string) string {
 		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":
@@ -819,6 +1293,8 @@ func alias2ModelName(modelName string) string {
 		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":

internal/runtime/executor/claude_executor.go

@@ -662,7 +662,14 @@ 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

internal/runtime/executor/gemini_cli_executor.go

@@ -79,6 +79,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 	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)
@@ -217,6 +218,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 	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)
@@ -418,6 +420,7 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 	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)
 		payload = deleteJSONField(payload, "project")
 		payload = deleteJSONField(payload, "model")
 		payload = deleteJSONField(payload, "request.safetySettings")

internal/runtime/executor/iflow_executor.go

@@ -66,6 +66,8 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
+	body = applyIFlowThinkingConfig(body)
+	body = preserveReasoningContentInMessages(body)
 	body = applyPayloadConfig(e.cfg, req.Model, body)
 
 	endpoint := strings.TrimSuffix(baseURL, "/") + iflowDefaultEndpoint
@@ -157,6 +159,8 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	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 {
@@ -442,3 +446,99 @@ func ensureToolsArray(body []byte) []byte {
 	}
 	return updated
 }
+
+// preserveReasoningContentInMessages ensures reasoning_content from assistant messages in the
+// conversation history is preserved when sending to iFlow models that support thinking.
+// This is critical for multi-turn conversations where the model needs to see its previous
+// reasoning to maintain coherent thought chains across tool calls and conversation turns.
+//
+// For GLM-4.7 and MiniMax-M2.1, the full assistant response (including reasoning) must be
+// appended back into message history before the next call.
+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-4.7") ||
+		strings.HasPrefix(model, "glm-4-7") ||
+		strings.HasPrefix(model, "minimax-m2.1") ||
+		strings.HasPrefix(model, "minimax-m2-1")
+
+	if !needsPreservation {
+		return body
+	}
+
+	messages := gjson.GetBytes(body, "messages")
+	if !messages.Exists() || !messages.IsArray() {
+		return body
+	}
+
+	// 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
+}
+
+// applyIFlowThinkingConfig converts normalized reasoning_effort to model-specific thinking configurations.
+// This should be called after NormalizeThinkingConfig has processed the payload.
+//
+// Model-specific handling:
+//   - GLM-4.7: Uses extra_body={"thinking": {"type": "enabled"}, "clear_thinking": false}
+//   - MiniMax-M2.1: Uses reasoning_split=true for OpenAI-style reasoning separation
+//   - Other iFlow models: Uses chat_template_kwargs.enable_thinking (boolean)
+func applyIFlowThinkingConfig(body []byte) []byte {
+	effort := gjson.GetBytes(body, "reasoning_effort")
+	model := strings.ToLower(gjson.GetBytes(body, "model").String())
+
+	// Check if thinking should be enabled
+	val := ""
+	if effort.Exists() {
+		val = strings.ToLower(strings.TrimSpace(effort.String()))
+	}
+	enableThinking := effort.Exists() && val != "none" && val != ""
+
+	// Remove reasoning_effort as we'll convert to model-specific format
+	if effort.Exists() {
+		body, _ = sjson.DeleteBytes(body, "reasoning_effort")
+	}
+
+	// GLM-4.7: Use extra_body with thinking config and clear_thinking: false
+	if strings.HasPrefix(model, "glm-4.7") || strings.HasPrefix(model, "glm-4-7") {
+		if enableThinking {
+			body, _ = sjson.SetBytes(body, "extra_body.thinking.type", "enabled")
+			body, _ = sjson.SetBytes(body, "extra_body.clear_thinking", false)
+		}
+		return body
+	}
+
+	// MiniMax-M2.1: Use reasoning_split=true for interleaved thinking
+	if strings.HasPrefix(model, "minimax-m2.1") || strings.HasPrefix(model, "minimax-m2-1") {
+		if enableThinking {
+			body, _ = sjson.SetBytes(body, "reasoning_split", true)
+		}
+		return body
+	}
+
+	// Other iFlow models (including GLM-4.6): Use chat_template_kwargs.enable_thinking
+	if effort.Exists() {
+		body, _ = sjson.SetBytes(body, "chat_template_kwargs.enable_thinking", enableThinking)
+	}
+
+	return body
+}

internal/runtime/executor/payload_helpers.go

@@ -72,13 +72,7 @@ func ApplyReasoningEffortMetadata(payload []byte, metadata map[string]any, model
 	// 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.OpenAIThinkingBudgetToEffort(baseModel, *budget); ok && effort != "" {
-				if *budget == 0 && effort == "none" && util.ModelUsesThinkingLevels(baseModel) {
-					if _, supported := util.NormalizeReasoningEffortLevel(baseModel, effort); !supported {
-						return StripThinkingFields(payload, false)
-					}
-				}
-
+			if effort, ok := util.ThinkingBudgetToEffort(baseModel, *budget); ok && effort != "" {
 				if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
 					return updated
 				}
@@ -273,7 +267,7 @@ func StripThinkingFields(payload []byte, effortOnly bool) []byte {
 		"reasoning.effort",
 	}
 	if !effortOnly {
-		fieldsToRemove = append([]string{"reasoning"}, fieldsToRemove...)
+		fieldsToRemove = append([]string{"reasoning", "thinking"}, fieldsToRemove...)
 	}
 	out := payload
 	for _, field := range fieldsToRemove {

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/translator/antigravity/claude/antigravity_claude_request.go

@@ -7,17 +7,40 @@ package claude
 
 import (
 	"bytes"
-	"encoding/json"
+	"crypto/sha256"
+	"encoding/hex"
 	"strings"
 
-	client "github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/cache"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	log "github.com/sirupsen/logrus"
 	"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
@@ -39,149 +62,329 @@ 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
-	var systemInstruction *client.Content
+	systemInstructionJSON := ""
+	hasSystemInstruction := false
 	systemResult := gjson.GetBytes(rawJSON, "system")
 	if systemResult.IsArray() {
 		systemResults := systemResult.Array()
-		systemInstruction = &client.Content{Role: "user", Parts: []client.Part{}}
+		systemInstructionJSON = `{"role":"user","parts":[]}`
 		for i := 0; i < len(systemResults); i++ {
 			systemPromptResult := systemResults[i]
 			systemTypePromptResult := systemPromptResult.Get("type")
 			if systemTypePromptResult.Type == gjson.String && systemTypePromptResult.String() == "text" {
 				systemPrompt := systemPromptResult.Get("text").String()
-				systemPart := client.Part{Text: systemPrompt}
-				systemInstruction.Parts = append(systemInstruction.Parts, systemPart)
+				partJSON := `{}`
+				if systemPrompt != "" {
+					partJSON, _ = sjson.Set(partJSON, "text", systemPrompt)
+				}
+				systemInstructionJSON, _ = sjson.SetRaw(systemInstructionJSON, "parts.-1", partJSON)
+				hasSystemInstruction = true
 			}
 		}
-		if len(systemInstruction.Parts) == 0 {
-			systemInstruction = nil
-		}
+	} else if systemResult.Type == gjson.String {
+		systemInstructionJSON = `{"role":"user","parts":[{"text":""}]}`
+		systemInstructionJSON, _ = sjson.Set(systemInstructionJSON, "parts.0.text", systemResult.String())
+		hasSystemInstruction = true
 	}
 
 	// contents
-	contents := make([]client.Content, 0)
+	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"
 			}
-			clientContent := client.Content{Role: role, Parts: []client.Part{}}
+			clientContentJSON := `{"role":"","parts":[]}`
+			clientContentJSON, _ = sjson.Set(clientContentJSON, "role", role)
 			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 := util.GetThinkingText(contentResult)
 						signatureResult := contentResult.Get("signature")
-						signature := geminiCLIClaudeThoughtSignature
-						if signatureResult.Exists() {
-							signature = signatureResult.String()
+						clientSignature := ""
+						if signatureResult.Exists() && signatureResult.String() != "" {
+							clientSignature = signatureResult.String()
 						}
-						clientContent.Parts = append(clientContent.Parts, client.Part{Text: prompt, Thought: true, ThoughtSignature: signature})
+
+						// 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 {
+							// TypeScript plugin approach: drop unsigned thinking blocks entirely
+							log.Debugf("Dropping unsigned thinking block (no valid signature)")
+							continue
+						}
+
+						// Valid signature, send as thought block
+						partJSON := `{}`
+						partJSON, _ = sjson.Set(partJSON, "thought", true)
+						if thinkingText != "" {
+							partJSON, _ = sjson.Set(partJSON, "text", thinkingText)
+						}
+						if signature != "" {
+							partJSON, _ = sjson.Set(partJSON, "thoughtSignature", signature)
+						}
+						clientContentJSON, _ = sjson.SetRaw(clientContentJSON, "parts.-1", partJSON)
 					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "text" {
 						prompt := contentResult.Get("text").String()
-						clientContent.Parts = append(clientContent.Parts, client.Part{Text: prompt})
+						partJSON := `{}`
+						if prompt != "" {
+							partJSON, _ = sjson.Set(partJSON, "text", prompt)
+						}
+						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.
+						// The TypeScript plugin removes unsigned thinking blocks instead of injecting dummies.
+
 						functionName := contentResult.Get("name").String()
-						functionArgs := contentResult.Get("input").String()
+						argsResult := contentResult.Get("input")
 						functionID := contentResult.Get("id").String()
-						var args map[string]any
-						if err := json.Unmarshal([]byte(functionArgs), &args); err == nil {
-							if strings.Contains(modelName, "claude") {
-								clientContent.Parts = append(clientContent.Parts, client.Part{
-									FunctionCall: &client.FunctionCall{ID: functionID, Name: functionName, Args: args},
-								})
-							} else {
-								clientContent.Parts = append(clientContent.Parts, client.Part{
-									FunctionCall:     &client.FunctionCall{ID: functionID, Name: functionName, Args: args},
-									ThoughtSignature: geminiCLIClaudeThoughtSignature,
-								})
+
+						// 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 != "" {
 							funcName := toolCallID
 							toolCallIDs := strings.Split(toolCallID, "-")
 							if len(toolCallIDs) > 1 {
-								funcName = strings.Join(toolCallIDs[0:len(toolCallIDs)-1], "-")
+								funcName = strings.Join(toolCallIDs[0:len(toolCallIDs)-2], "-")
 							}
-							responseData := contentResult.Get("content").Raw
-							functionResponse := client.FunctionResponse{ID: toolCallID, Name: funcName, Response: map[string]interface{}{"result": responseData}}
-							clientContent.Parts = append(clientContent.Parts, client.Part{FunctionResponse: &functionResponse})
+							functionResponseResult := contentResult.Get("content")
+
+							functionResponseJSON := `{}`
+							functionResponseJSON, _ = sjson.Set(functionResponseJSON, "id", toolCallID)
+							functionResponseJSON, _ = sjson.Set(functionResponseJSON, "name", funcName)
+
+							responseData := ""
+							if functionResponseResult.Type == gjson.String {
+								responseData = functionResponseResult.String()
+								functionResponseJSON, _ = sjson.Set(functionResponseJSON, "response.result", responseData)
+							} else if functionResponseResult.IsArray() {
+								frResults := functionResponseResult.Array()
+								if len(frResults) == 1 {
+									functionResponseJSON, _ = sjson.SetRaw(functionResponseJSON, "response.result", frResults[0].Raw)
+								} else {
+									functionResponseJSON, _ = sjson.SetRaw(functionResponseJSON, "response.result", functionResponseResult.Raw)
+								}
+
+							} else if functionResponseResult.IsObject() {
+								functionResponseJSON, _ = sjson.SetRaw(functionResponseJSON, "response.result", functionResponseResult.Raw)
+							} else {
+								functionResponseJSON, _ = sjson.SetRaw(functionResponseJSON, "response.result", functionResponseResult.Raw)
+							}
+
+							partJSON := `{}`
+							partJSON, _ = sjson.SetRaw(partJSON, "functionResponse", functionResponseJSON)
+							clientContentJSON, _ = sjson.SetRaw(clientContentJSON, "parts.-1", partJSON)
 						}
 					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "image" {
 						sourceResult := contentResult.Get("source")
 						if sourceResult.Get("type").String() == "base64" {
-							inlineData := &client.InlineData{
-								MimeType: sourceResult.Get("media_type").String(),
-								Data:     sourceResult.Get("data").String(),
+							inlineDataJSON := `{}`
+							if mimeType := sourceResult.Get("media_type").String(); mimeType != "" {
+								inlineDataJSON, _ = sjson.Set(inlineDataJSON, "mime_type", mimeType)
 							}
-							clientContent.Parts = append(clientContent.Parts, client.Part{InlineData: inlineData})
+							if data := sourceResult.Get("data").String(); data != "" {
+								inlineDataJSON, _ = sjson.Set(inlineDataJSON, "data", data)
+							}
+
+							partJSON := `{}`
+							partJSON, _ = sjson.SetRaw(partJSON, "inlineData", inlineDataJSON)
+							clientContentJSON, _ = sjson.SetRaw(clientContentJSON, "parts.-1", partJSON)
 						}
 					}
 				}
-				contents = append(contents, clientContent)
+
+				// 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 {
 				prompt := contentsResult.String()
-				contents = append(contents, client.Content{Role: role, Parts: []client.Part{{Text: prompt}}})
+				partJSON := `{}`
+				if prompt != "" {
+					partJSON, _ = sjson.Set(partJSON, "text", prompt)
+				}
+				clientContentJSON, _ = sjson.SetRaw(clientContentJSON, "parts.-1", partJSON)
+				contentsJSON, _ = sjson.SetRaw(contentsJSON, "-1", clientContentJSON)
+				hasContents = true
 			}
 		}
 	}
 
 	// tools
-	var tools []client.ToolDeclaration
+	toolsJSON := ""
+	toolDeclCount := 0
+	allowedToolKeys := []string{"name", "description", "behavior", "parameters", "parametersJsonSchema", "response", "responseJsonSchema"}
 	toolsResult := gjson.GetBytes(rawJSON, "tools")
 	if toolsResult.IsArray() {
-		tools = make([]client.ToolDeclaration, 1)
-		tools[0].FunctionDeclarations = make([]any, 0)
+		toolsJSON = `[{"functionDeclarations":[]}]`
 		toolsResults := toolsResult.Array()
 		for i := 0; i < len(toolsResults); i++ {
 			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")
-				var toolDeclaration any
-				if err := json.Unmarshal([]byte(tool), &toolDeclaration); err == nil {
-					tools[0].FunctionDeclarations = append(tools[0].FunctionDeclarations, toolDeclaration)
+				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++
 			}
 		}
-	} else {
-		tools = make([]client.ToolDeclaration, 0)
 	}
 
 	// Build output Gemini CLI request JSON
 	out := `{"model":"","request":{"contents":[]}}`
 	out, _ = sjson.Set(out, "model", modelName)
-	if systemInstruction != nil {
-		b, _ := json.Marshal(systemInstruction)
-		out, _ = sjson.SetRaw(out, "request.systemInstruction", string(b))
+
+	// 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 len(contents) > 0 {
-		b, _ := json.Marshal(contents)
-		out, _ = sjson.SetRaw(out, "request.contents", string(b))
+
+	if hasSystemInstruction {
+		out, _ = sjson.SetRaw(out, "request.systemInstruction", systemInstructionJSON)
 	}
-	if len(tools) > 0 && len(tools[0].FunctionDeclarations) > 0 {
-		b, _ := json.Marshal(tools)
-		out, _ = sjson.SetRaw(out, "request.tools", string(b))
+	if hasContents {
+		out, _ = sjson.SetRaw(out, "request.contents", contentsJSON)
+	}
+	if toolDeclCount > 0 {
+		out, _ = sjson.SetRaw(out, "request.tools", toolsJSON)
 	}
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled

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

@@ -9,12 +9,14 @@ package claude
 import (
 	"bytes"
 	"context"
-	"encoding/json"
 	"fmt"
 	"strings"
 	"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"
 )
@@ -33,9 +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.
@@ -63,6 +70,7 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 			HasFirstResponse: false,
 			ResponseType:     0,
 			ResponseIndex:    0,
+			SessionID:        deriveSessionID(originalRequestRawJSON),
 		}
 	}
 
@@ -91,6 +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())
@@ -120,11 +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)
@@ -153,6 +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")
@@ -251,7 +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()
@@ -303,6 +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
@@ -342,6 +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
@@ -350,24 +388,33 @@ func ConvertAntigravityResponseToClaudeNonStream(_ context.Context, _ string, or
 		}
 	}
 
-	response := map[string]interface{}{
-		"id":            root.Get("response.responseId").String(),
-		"type":          "message",
-		"role":          "assistant",
-		"model":         root.Get("response.modelVersion").String(),
-		"content":       []interface{}{},
-		"stop_reason":   nil,
-		"stop_sequence": nil,
-		"usage": map[string]interface{}{
-			"input_tokens":  promptTokens,
-			"output_tokens": outputTokens,
-		},
+	responseJSON := `{"id":"","type":"message","role":"assistant","model":"","content":null,"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}`
+	responseJSON, _ = sjson.Set(responseJSON, "id", root.Get("response.responseId").String())
+	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() {
+		if contentArrayInitialized {
+			return
+		}
+		responseJSON, _ = sjson.SetRaw(responseJSON, "content", "[]")
+		contentArrayInitialized = true
 	}
 
 	parts := root.Get("response.candidates.0.content.parts")
-	var contentBlocks []interface{}
 	textBuilder := strings.Builder{}
 	thinkingBuilder := strings.Builder{}
+	thinkingSignature := ""
 	toolIDCounter := 0
 	hasToolCall := false
 
@@ -375,28 +422,43 @@ func ConvertAntigravityResponseToClaudeNonStream(_ context.Context, _ string, or
 		if textBuilder.Len() == 0 {
 			return
 		}
-		contentBlocks = append(contentBlocks, map[string]interface{}{
-			"type": "text",
-			"text": textBuilder.String(),
-		})
+		ensureContentArray()
+		block := `{"type":"text","text":""}`
+		block, _ = sjson.Set(block, "text", textBuilder.String())
+		responseJSON, _ = sjson.SetRaw(responseJSON, "content.-1", block)
 		textBuilder.Reset()
 	}
 
 	flushThinking := func() {
-		if thinkingBuilder.Len() == 0 {
+		if thinkingBuilder.Len() == 0 && thinkingSignature == "" {
 			return
 		}
-		contentBlocks = append(contentBlocks, map[string]interface{}{
-			"type":     "thinking",
-			"thinking": thinkingBuilder.String(),
-		})
+		ensureContentArray()
+		block := `{"type":"thinking","thinking":""}`
+		block, _ = sjson.Set(block, "thinking", thinkingBuilder.String())
+		if thinkingSignature != "" {
+			block, _ = sjson.Set(block, "signature", thinkingSignature)
+		}
+		responseJSON, _ = sjson.SetRaw(responseJSON, "content.-1", block)
 		thinkingBuilder.Reset()
+		thinkingSignature = ""
 	}
 
 	if parts.IsArray() {
 		for _, part := range parts.Array() {
+			isThought := part.Get("thought").Bool()
+			if isThought {
+				sig := part.Get("thoughtSignature")
+				if !sig.Exists() {
+					sig = part.Get("thought_signature")
+				}
+				if sig.Exists() && sig.String() != "" {
+					thinkingSignature = sig.String()
+				}
+			}
+
 			if text := part.Get("text"); text.Exists() && text.String() != "" {
-				if part.Get("thought").Bool() {
+				if isThought {
 					flushText()
 					thinkingBuilder.WriteString(text.String())
 					continue
@@ -413,21 +475,16 @@ func ConvertAntigravityResponseToClaudeNonStream(_ context.Context, _ string, or
 
 				name := functionCall.Get("name").String()
 				toolIDCounter++
-				toolBlock := map[string]interface{}{
-					"type":  "tool_use",
-					"id":    fmt.Sprintf("tool_%d", toolIDCounter),
-					"name":  name,
-					"input": map[string]interface{}{},
+				toolBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
+				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) && args.IsObject() {
+					toolBlock, _ = sjson.SetRaw(toolBlock, "input", args.Raw)
 				}
 
-				if args := functionCall.Get("args"); args.Exists() {
-					var parsed interface{}
-					if err := json.Unmarshal([]byte(args.Raw), &parsed); err == nil {
-						toolBlock["input"] = parsed
-					}
-				}
-
-				contentBlocks = append(contentBlocks, toolBlock)
+				ensureContentArray()
+				responseJSON, _ = sjson.SetRaw(responseJSON, "content.-1", toolBlock)
 				continue
 			}
 		}
@@ -436,8 +493,6 @@ func ConvertAntigravityResponseToClaudeNonStream(_ context.Context, _ string, or
 	flushThinking()
 	flushText()
 
-	response["content"] = contentBlocks
-
 	stopReason := "end_turn"
 	if hasToolCall {
 		stopReason = "tool_use"
@@ -453,19 +508,15 @@ func ConvertAntigravityResponseToClaudeNonStream(_ context.Context, _ string, or
 			}
 		}
 	}
-	response["stop_reason"] = stopReason
+	responseJSON, _ = sjson.Set(responseJSON, "stop_reason", stopReason)
 
-	if usage := response["usage"].(map[string]interface{}); usage["input_tokens"] == int64(0) && usage["output_tokens"] == int64(0) {
+	if promptTokens == 0 && outputTokens == 0 {
 		if usageMeta := root.Get("response.usageMetadata"); !usageMeta.Exists() {
-			delete(response, "usage")
+			responseJSON, _ = sjson.Delete(responseJSON, "usage")
 		}
 	}
 
-	encoded, err := json.Marshal(response)
-	if err != nil {
-		return ""
-	}
-	return string(encoded)
+	return responseJSON
 }
 
 func ClaudeTokenCount(ctx context.Context, count int64) string {

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
 }

internal/translator/antigravity/gemini/antigravity_gemini_request_test.go

@@ -0,0 +1,129 @@
+package gemini
+
+import (
+	"fmt"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestConvertGeminiRequestToAntigravity_PreserveValidSignature(t *testing.T) {
+	// Valid signature on functionCall should be preserved
+	validSignature := "abc123validSignature1234567890123456789012345678901234567890"
+	inputJSON := []byte(fmt.Sprintf(`{
+		"model": "gemini-3-pro-preview",
+		"contents": [
+			{
+				"role": "model",
+				"parts": [
+					{"functionCall": {"name": "test_tool", "args": {}}, "thoughtSignature": "%s"}
+				]
+			}
+		]
+	}`, validSignature))
+
+	output := ConvertGeminiRequestToAntigravity("gemini-3-pro-preview", inputJSON, false)
+	outputStr := string(output)
+
+	// Check that valid thoughtSignature is preserved
+	parts := gjson.Get(outputStr, "request.contents.0.parts").Array()
+	if len(parts) != 1 {
+		t.Fatalf("Expected 1 part, got %d", len(parts))
+	}
+
+	sig := parts[0].Get("thoughtSignature").String()
+	if sig != validSignature {
+		t.Errorf("Expected thoughtSignature '%s', got '%s'", validSignature, sig)
+	}
+}
+
+func TestConvertGeminiRequestToAntigravity_AddSkipSentinelToFunctionCall(t *testing.T) {
+	// functionCall without signature should get skip_thought_signature_validator
+	inputJSON := []byte(`{
+		"model": "gemini-3-pro-preview",
+		"contents": [
+			{
+				"role": "model",
+				"parts": [
+					{"functionCall": {"name": "test_tool", "args": {}}}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertGeminiRequestToAntigravity("gemini-3-pro-preview", inputJSON, false)
+	outputStr := string(output)
+
+	// Check that skip_thought_signature_validator is added to functionCall
+	sig := gjson.Get(outputStr, "request.contents.0.parts.0.thoughtSignature").String()
+	expectedSig := "skip_thought_signature_validator"
+	if sig != expectedSig {
+		t.Errorf("Expected skip sentinel '%s', got '%s'", expectedSig, sig)
+	}
+}
+
+func TestConvertGeminiRequestToAntigravity_RemoveThinkingBlocks(t *testing.T) {
+	// Thinking blocks should be removed entirely for Gemini
+	validSignature := "abc123validSignature1234567890123456789012345678901234567890"
+	inputJSON := []byte(fmt.Sprintf(`{
+		"model": "gemini-3-pro-preview",
+		"contents": [
+			{
+				"role": "model",
+				"parts": [
+					{"thought": true, "text": "Thinking...", "thoughtSignature": "%s"},
+					{"text": "Here is my response"}
+				]
+			}
+		]
+	}`, validSignature))
+
+	output := ConvertGeminiRequestToAntigravity("gemini-3-pro-preview", inputJSON, false)
+	outputStr := string(output)
+
+	// Check that thinking block is removed
+	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 for Gemini")
+	}
+	if parts[0].Get("text").String() != "Here is my response" {
+		t.Errorf("Expected text 'Here is my response', got '%s'", parts[0].Get("text").String())
+	}
+}
+
+func TestConvertGeminiRequestToAntigravity_ParallelFunctionCalls(t *testing.T) {
+	// Multiple functionCalls should all get skip_thought_signature_validator
+	inputJSON := []byte(`{
+		"model": "gemini-3-pro-preview",
+		"contents": [
+			{
+				"role": "model",
+				"parts": [
+					{"functionCall": {"name": "tool_one", "args": {"a": "1"}}},
+					{"functionCall": {"name": "tool_two", "args": {"b": "2"}}}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertGeminiRequestToAntigravity("gemini-3-pro-preview", inputJSON, false)
+	outputStr := string(output)
+
+	parts := gjson.Get(outputStr, "request.contents.0.parts").Array()
+	if len(parts) != 2 {
+		t.Fatalf("Expected 2 parts, got %d", len(parts))
+	}
+
+	expectedSig := "skip_thought_signature_validator"
+	for i, part := range parts {
+		sig := part.Get("thoughtSignature").String()
+		if sig != expectedSig {
+			t.Errorf("Part %d: Expected '%s', got '%s'", i, expectedSig, sig)
+		}
+	}
+}

internal/translator/antigravity/openai/chat-completions/antigravity_openai_request.go

@@ -39,8 +39,23 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 	// Note: OpenAI official fields take precedence over extra_body.google.thinking_config
 	re := gjson.GetBytes(rawJSON, "reasoning_effort")
 	hasOfficialThinking := re.Exists()
-	if hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
-		out = util.ApplyReasoningEffortToGeminiCLI(out, re.String())
+	if hasOfficialThinking && util.ModelSupportsThinking(modelName) {
+		effort := strings.ToLower(strings.TrimSpace(re.String()))
+		if util.IsGemini3Model(modelName) {
+			switch effort {
+			case "none":
+				out, _ = sjson.DeleteBytes(out, "request.generationConfig.thinkingConfig")
+			case "auto":
+				includeThoughts := true
+				out = util.ApplyGeminiCLIThinkingLevel(out, "", &includeThoughts)
+			default:
+				if level, ok := util.ValidateGemini3ThinkingLevel(modelName, effort); ok {
+					out = util.ApplyGeminiCLIThinkingLevel(out, level, nil)
+				}
+			}
+		} else if !util.ModelUsesThinkingLevels(modelName) {
+			out = util.ApplyReasoningEffortToGeminiCLI(out, effort)
+		}
 	}
 
 	// Cherry Studio extension extra_body.google.thinking_config (effective only when official fields are absent)
@@ -177,6 +192,14 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 				} else if content.IsObject() && content.Get("type").String() == "text" {
 					out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
 					out, _ = sjson.SetBytes(out, "request.systemInstruction.parts.0.text", content.Get("text").String())
+				} else if content.IsArray() {
+					contents := content.Array()
+					if len(contents) > 0 {
+						out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
+						for j := 0; j < len(contents); j++ {
+							out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", j), contents[j].Get("text").String())
+						}
+					}
 				}
 			} else if role == "user" || (role == "system" && len(arr) == 1) {
 				// Build single user content node to avoid splitting into multiple contents
@@ -222,64 +245,89 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 				}
 				out, _ = sjson.SetRawBytes(out, "request.contents.-1", node)
 			} else if role == "assistant" {
-				if content.Type == gjson.String {
-					// Assistant text -> single model content
-					node := []byte(`{"role":"model","parts":[{"text":""}]}`)
-					node, _ = sjson.SetBytes(node, "parts.0.text", content.String())
-					out, _ = sjson.SetRawBytes(out, "request.contents.-1", node)
-				} else if !content.Exists() || content.Type == gjson.Null {
-					// Tool calls -> single model content with functionCall parts
-					tcs := m.Get("tool_calls")
-					if tcs.IsArray() {
-						node := []byte(`{"role":"model","parts":[]}`)
-						p := 0
-						fIDs := make([]string, 0)
-						for _, tc := range tcs.Array() {
-							if tc.Get("type").String() != "function" {
-								continue
-							}
-							fid := tc.Get("id").String()
-							fname := tc.Get("function.name").String()
-							fargs := tc.Get("function.arguments").String()
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".functionCall.id", fid)
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".functionCall.name", fname)
-							node, _ = sjson.SetRawBytes(node, "parts."+itoa(p)+".functionCall.args", []byte(fargs))
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiCLIFunctionThoughtSignature)
+				node := []byte(`{"role":"model","parts":[]}`)
+				p := 0
+				if content.Type == gjson.String && content.String() != "" {
+					node, _ = sjson.SetBytes(node, "parts.-1.text", content.String())
+					p++
+				} else if content.IsArray() {
+					// Assistant multimodal content (e.g. text + image) -> single model content with parts
+					for _, item := range content.Array() {
+						switch item.Get("type").String() {
+						case "text":
 							p++
-							if fid != "" {
-								fIDs = append(fIDs, fid)
-							}
-						}
-						out, _ = sjson.SetRawBytes(out, "request.contents.-1", node)
-
-						// Append a single tool content combining name + response per function
-						toolNode := []byte(`{"role":"user","parts":[]}`)
-						pp := 0
-						for _, fid := range fIDs {
-							if name, ok := tcID2Name[fid]; ok {
-								toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.id", fid)
-								toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.name", name)
-								resp := toolResponses[fid]
-								if resp == "" {
-									resp = "{}"
+						case "image_url":
+							// If the assistant returned an inline data URL, preserve it for history fidelity.
+							imageURL := item.Get("image_url.url").String()
+							if len(imageURL) > 5 { // expect data:...
+								pieces := strings.SplitN(imageURL[5:], ";", 2)
+								if len(pieces) == 2 && len(pieces[1]) > 7 {
+									mime := pieces[0]
+									data := pieces[1][7:]
+									node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".inlineData.mime_type", mime)
+									node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".inlineData.data", data)
+									p++
 								}
-								// Handle non-JSON output gracefully (matches dev branch approach)
-								if resp != "null" {
-									parsed := gjson.Parse(resp)
-									if parsed.Type == gjson.JSON {
-										toolNode, _ = sjson.SetRawBytes(toolNode, "parts."+itoa(pp)+".functionResponse.response.result", []byte(parsed.Raw))
-									} else {
-										toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.response.result", resp)
-									}
-								}
-								pp++
 							}
 						}
-						if pp > 0 {
-							out, _ = sjson.SetRawBytes(out, "request.contents.-1", toolNode)
-						}
 					}
 				}
+
+				// Tool calls -> single model content with functionCall parts
+				tcs := m.Get("tool_calls")
+				if tcs.IsArray() {
+					fIDs := make([]string, 0)
+					for _, tc := range tcs.Array() {
+						if tc.Get("type").String() != "function" {
+							continue
+						}
+						fid := tc.Get("id").String()
+						fname := tc.Get("function.name").String()
+						fargs := tc.Get("function.arguments").String()
+						node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".functionCall.id", fid)
+						node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".functionCall.name", fname)
+						if gjson.Valid(fargs) {
+							node, _ = sjson.SetRawBytes(node, "parts."+itoa(p)+".functionCall.args", []byte(fargs))
+						} else {
+							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".functionCall.args.params", []byte(fargs))
+						}
+						node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiCLIFunctionThoughtSignature)
+						p++
+						if fid != "" {
+							fIDs = append(fIDs, fid)
+						}
+					}
+					out, _ = sjson.SetRawBytes(out, "request.contents.-1", node)
+
+					// Append a single tool content combining name + response per function
+					toolNode := []byte(`{"role":"user","parts":[]}`)
+					pp := 0
+					for _, fid := range fIDs {
+						if name, ok := tcID2Name[fid]; ok {
+							toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.id", fid)
+							toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.name", name)
+							resp := toolResponses[fid]
+							if resp == "" {
+								resp = "{}"
+							}
+							// Handle non-JSON output gracefully (matches dev branch approach)
+							if resp != "null" {
+								parsed := gjson.Parse(resp)
+								if parsed.Type == gjson.JSON {
+									toolNode, _ = sjson.SetRawBytes(toolNode, "parts."+itoa(pp)+".functionResponse.response.result", []byte(parsed.Raw))
+								} else {
+									toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.response.result", resp)
+								}
+							}
+							pp++
+						}
+					}
+					if pp > 0 {
+						out, _ = sjson.SetRawBytes(out, "request.contents.-1", toolNode)
+					}
+				} else {
+					out, _ = sjson.SetRawBytes(out, "request.contents.-1", node)
+				}
 			}
 		}
 	}
@@ -305,7 +353,7 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 								log.Warnf("Failed to set default schema type for tool '%s': %v", fn.Get("name").String(), errSet)
 								continue
 							}
-							fnRaw, errSet = sjson.Set(fnRaw, "parametersJsonSchema.properties", map[string]interface{}{})
+							fnRaw, errSet = sjson.SetRaw(fnRaw, "parametersJsonSchema.properties", `{}`)
 							if errSet != nil {
 								log.Warnf("Failed to set default schema properties for tool '%s': %v", fn.Get("name").String(), errSet)
 								continue
@@ -320,7 +368,7 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 							log.Warnf("Failed to set default schema type for tool '%s': %v", fn.Get("name").String(), errSet)
 							continue
 						}
-						fnRaw, errSet = sjson.Set(fnRaw, "parametersJsonSchema.properties", map[string]interface{}{})
+						fnRaw, errSet = sjson.SetRaw(fnRaw, "parametersJsonSchema.properties", `{}`)
 						if errSet != nil {
 							log.Warnf("Failed to set default schema properties for tool '%s': %v", fn.Get("name").String(), errSet)
 							continue
@@ -361,18 +409,3 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 
 // itoa converts int to string without strconv import for few usages.
 func itoa(i int) string { return fmt.Sprintf("%d", i) }
-
-// quoteIfNeeded ensures a string is valid JSON value (quotes plain text), pass-through for JSON objects/arrays.
-func quoteIfNeeded(s string) string {
-	s = strings.TrimSpace(s)
-	if s == "" {
-		return "\"\""
-	}
-	if len(s) > 0 && (s[0] == '{' || s[0] == '[') {
-		return s
-	}
-	// escape quotes minimally
-	s = strings.ReplaceAll(s, "\\", "\\\\")
-	s = strings.ReplaceAll(s, "\"", "\\\"")
-	return "\"" + s + "\""
-}

internal/translator/antigravity/openai/chat-completions/antigravity_openai_response.go

@@ -8,12 +8,13 @@ package chat_completions
 import (
 	"bytes"
 	"context"
-	"encoding/json"
 	"fmt"
 	"strings"
 	"sync/atomic"
 	"time"
 
+	log "github.com/sirupsen/logrus"
+
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/openai/chat-completions"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -86,18 +87,27 @@ func ConvertAntigravityResponseToOpenAI(_ context.Context, _ string, originalReq
 
 	// Extract and set usage metadata (token counts).
 	if usageResult := gjson.GetBytes(rawJSON, "response.usageMetadata"); usageResult.Exists() {
+		cachedTokenCount := usageResult.Get("cachedContentTokenCount").Int()
 		if candidatesTokenCountResult := usageResult.Get("candidatesTokenCount"); candidatesTokenCountResult.Exists() {
 			template, _ = sjson.Set(template, "usage.completion_tokens", candidatesTokenCountResult.Int())
 		}
 		if totalTokenCountResult := usageResult.Get("totalTokenCount"); totalTokenCountResult.Exists() {
 			template, _ = sjson.Set(template, "usage.total_tokens", totalTokenCountResult.Int())
 		}
-		promptTokenCount := usageResult.Get("promptTokenCount").Int()
+		promptTokenCount := usageResult.Get("promptTokenCount").Int() - cachedTokenCount
 		thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
 		template, _ = sjson.Set(template, "usage.prompt_tokens", promptTokenCount+thoughtsTokenCount)
 		if thoughtsTokenCount > 0 {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", thoughtsTokenCount)
 		}
+		// Include cached token count if present (indicates prompt caching is working)
+		if cachedTokenCount > 0 {
+			var err error
+			template, err = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cachedTokenCount)
+			if err != nil {
+				log.Warnf("antigravity openai response: failed to set cached_tokens: %v", err)
+			}
+		}
 	}
 
 	// Process the main content part of the response.
@@ -171,21 +181,16 @@ func ConvertAntigravityResponseToOpenAI(_ context.Context, _ string, originalReq
 					mimeType = "image/png"
 				}
 				imageURL := fmt.Sprintf("data:%s;base64,%s", mimeType, data)
-				imagePayload, err := json.Marshal(map[string]any{
-					"type": "image_url",
-					"image_url": map[string]string{
-						"url": imageURL,
-					},
-				})
-				if err != nil {
-					continue
-				}
 				imagesResult := gjson.Get(template, "choices.0.delta.images")
 				if !imagesResult.Exists() || !imagesResult.IsArray() {
 					template, _ = sjson.SetRaw(template, "choices.0.delta.images", `[]`)
 				}
+				imageIndex := len(gjson.Get(template, "choices.0.delta.images").Array())
+				imagePayload := `{"type":"image_url","image_url":{"url":""}}`
+				imagePayload, _ = sjson.Set(imagePayload, "index", imageIndex)
+				imagePayload, _ = sjson.Set(imagePayload, "image_url.url", imageURL)
 				template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
-				template, _ = sjson.SetRaw(template, "choices.0.delta.images.-1", string(imagePayload))
+				template, _ = sjson.SetRaw(template, "choices.0.delta.images.-1", imagePayload)
 			}
 		}
 	}

internal/translator/claude/gemini/claude_gemini_request.go

@@ -194,7 +194,7 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 						if name := fc.Get("name"); name.Exists() {
 							toolUse, _ = sjson.Set(toolUse, "name", name.String())
 						}
-						if args := fc.Get("args"); args.Exists() {
+						if args := fc.Get("args"); args.Exists() && args.IsObject() {
 							toolUse, _ = sjson.SetRaw(toolUse, "input", args.Raw)
 						}
 						msg, _ = sjson.SetRaw(msg, "content.-1", toolUse)
@@ -314,11 +314,11 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 			if mode := funcCalling.Get("mode"); mode.Exists() {
 				switch mode.String() {
 				case "AUTO":
-					out, _ = sjson.Set(out, "tool_choice", map[string]interface{}{"type": "auto"})
+					out, _ = sjson.SetRaw(out, "tool_choice", `{"type":"auto"}`)
 				case "NONE":
-					out, _ = sjson.Set(out, "tool_choice", map[string]interface{}{"type": "none"})
+					out, _ = sjson.SetRaw(out, "tool_choice", `{"type":"none"}`)
 				case "ANY":
-					out, _ = sjson.Set(out, "tool_choice", map[string]interface{}{"type": "any"})
+					out, _ = sjson.SetRaw(out, "tool_choice", `{"type":"any"}`)
 				}
 			}
 		}

internal/translator/claude/gemini/claude_gemini_response.go

@@ -263,51 +263,6 @@ func ConvertClaudeResponseToGemini(_ context.Context, modelName string, original
 	}
 }
 
-// convertArrayToJSON converts []interface{} to JSON array string
-func convertArrayToJSON(arr []interface{}) string {
-	result := "[]"
-	for _, item := range arr {
-		switch itemData := item.(type) {
-		case map[string]interface{}:
-			itemJSON := convertMapToJSON(itemData)
-			result, _ = sjson.SetRaw(result, "-1", itemJSON)
-		case string:
-			result, _ = sjson.Set(result, "-1", itemData)
-		case bool:
-			result, _ = sjson.Set(result, "-1", itemData)
-		case float64, int, int64:
-			result, _ = sjson.Set(result, "-1", itemData)
-		default:
-			result, _ = sjson.Set(result, "-1", itemData)
-		}
-	}
-	return result
-}
-
-// convertMapToJSON converts map[string]interface{} to JSON object string
-func convertMapToJSON(m map[string]interface{}) string {
-	result := "{}"
-	for key, value := range m {
-		switch val := value.(type) {
-		case map[string]interface{}:
-			nestedJSON := convertMapToJSON(val)
-			result, _ = sjson.SetRaw(result, key, nestedJSON)
-		case []interface{}:
-			arrayJSON := convertArrayToJSON(val)
-			result, _ = sjson.SetRaw(result, key, arrayJSON)
-		case string:
-			result, _ = sjson.Set(result, key, val)
-		case bool:
-			result, _ = sjson.Set(result, key, val)
-		case float64, int, int64:
-			result, _ = sjson.Set(result, key, val)
-		default:
-			result, _ = sjson.Set(result, key, val)
-		}
-	}
-	return result
-}
-
 // ConvertClaudeResponseToGeminiNonStream converts a non-streaming Claude Code response to a non-streaming Gemini response.
 // This function processes the complete Claude Code response and transforms it into a single Gemini-compatible
 // JSON response. It handles message content, tool calls, reasoning content, and usage metadata, combining all
@@ -356,8 +311,8 @@ func ConvertClaudeResponseToGeminiNonStream(_ context.Context, modelName string,
 	}
 
 	// Process each streaming event and collect parts
-	var allParts []interface{}
-	var finalUsage map[string]interface{}
+	var allParts []string
+	var finalUsageJSON string
 	var responseID string
 	var createdAt int64
 
@@ -407,16 +362,14 @@ func ConvertClaudeResponseToGeminiNonStream(_ context.Context, modelName string,
 					if text := delta.Get("text"); text.Exists() && text.String() != "" {
 						partJSON := `{"text":""}`
 						partJSON, _ = sjson.Set(partJSON, "text", text.String())
-						part := gjson.Parse(partJSON).Value().(map[string]interface{})
-						allParts = append(allParts, part)
+						allParts = append(allParts, partJSON)
 					}
 				case "thinking_delta":
 					// Process reasoning/thinking content
 					if text := delta.Get("thinking"); text.Exists() && text.String() != "" {
 						partJSON := `{"thought":true,"text":""}`
 						partJSON, _ = sjson.Set(partJSON, "text", text.String())
-						part := gjson.Parse(partJSON).Value().(map[string]interface{})
-						allParts = append(allParts, part)
+						allParts = append(allParts, partJSON)
 					}
 				case "input_json_delta":
 					// accumulate args partial_json for this index
@@ -456,9 +409,7 @@ func ConvertClaudeResponseToGeminiNonStream(_ context.Context, modelName string,
 				if argsTrim != "" {
 					functionCallJSON, _ = sjson.SetRaw(functionCallJSON, "functionCall.args", argsTrim)
 				}
-				// Parse back to interface{} for allParts
-				functionCall := gjson.Parse(functionCallJSON).Value().(map[string]interface{})
-				allParts = append(allParts, functionCall)
+				allParts = append(allParts, functionCallJSON)
 				// cleanup used state for this index
 				if newParam.ToolUseArgs != nil {
 					delete(newParam.ToolUseArgs, idx)
@@ -501,8 +452,7 @@ func ConvertClaudeResponseToGeminiNonStream(_ context.Context, modelName string,
 				// Set traffic type (required by Gemini API)
 				usageJSON, _ = sjson.Set(usageJSON, "trafficType", "PROVISIONED_THROUGHPUT")
 
-				// Convert to map[string]interface{} using gjson
-				finalUsage = gjson.Parse(usageJSON).Value().(map[string]interface{})
+				finalUsageJSON = usageJSON
 			}
 		}
 	}
@@ -520,12 +470,16 @@ func ConvertClaudeResponseToGeminiNonStream(_ context.Context, modelName string,
 
 	// Set the consolidated parts array
 	if len(consolidatedParts) > 0 {
-		template, _ = sjson.SetRaw(template, "candidates.0.content.parts", convertToJSONString(consolidatedParts))
+		partsJSON := "[]"
+		for _, partJSON := range consolidatedParts {
+			partsJSON, _ = sjson.SetRaw(partsJSON, "-1", partJSON)
+		}
+		template, _ = sjson.SetRaw(template, "candidates.0.content.parts", partsJSON)
 	}
 
 	// Set usage metadata
-	if finalUsage != nil {
-		template, _ = sjson.SetRaw(template, "usageMetadata", convertToJSONString(finalUsage))
+	if finalUsageJSON != "" {
+		template, _ = sjson.SetRaw(template, "usageMetadata", finalUsageJSON)
 	}
 
 	return template
@@ -539,12 +493,12 @@ func GeminiTokenCount(ctx context.Context, count int64) string {
 // This function processes the parts array to combine adjacent text elements and thinking elements
 // into single consolidated parts, which results in a more readable and efficient response structure.
 // Tool calls and other non-text parts are preserved as separate elements.
-func consolidateParts(parts []interface{}) []interface{} {
+func consolidateParts(parts []string) []string {
 	if len(parts) == 0 {
 		return parts
 	}
 
-	var consolidated []interface{}
+	var consolidated []string
 	var currentTextPart strings.Builder
 	var currentThoughtPart strings.Builder
 	var hasText, hasThought bool
@@ -554,8 +508,7 @@ func consolidateParts(parts []interface{}) []interface{} {
 		if hasText && currentTextPart.Len() > 0 {
 			textPartJSON := `{"text":""}`
 			textPartJSON, _ = sjson.Set(textPartJSON, "text", currentTextPart.String())
-			textPart := gjson.Parse(textPartJSON).Value().(map[string]interface{})
-			consolidated = append(consolidated, textPart)
+			consolidated = append(consolidated, textPartJSON)
 			currentTextPart.Reset()
 			hasText = false
 		}
@@ -566,42 +519,42 @@ func consolidateParts(parts []interface{}) []interface{} {
 		if hasThought && currentThoughtPart.Len() > 0 {
 			thoughtPartJSON := `{"thought":true,"text":""}`
 			thoughtPartJSON, _ = sjson.Set(thoughtPartJSON, "text", currentThoughtPart.String())
-			thoughtPart := gjson.Parse(thoughtPartJSON).Value().(map[string]interface{})
-			consolidated = append(consolidated, thoughtPart)
+			consolidated = append(consolidated, thoughtPartJSON)
 			currentThoughtPart.Reset()
 			hasThought = false
 		}
 	}
 
-	for _, part := range parts {
-		partMap, ok := part.(map[string]interface{})
-		if !ok {
+	for _, partJSON := range parts {
+		part := gjson.Parse(partJSON)
+		if !part.Exists() || !part.IsObject() {
 			// Flush any pending parts and add this non-text part
 			flushText()
 			flushThought()
-			consolidated = append(consolidated, part)
+			consolidated = append(consolidated, partJSON)
 			continue
 		}
 
-		if thought, isThought := partMap["thought"]; isThought && thought == true {
+		thought := part.Get("thought")
+		if thought.Exists() && thought.Type == gjson.True {
 			// This is a thinking part - flush any pending text first
 			flushText() // Flush any pending text first
 
-			if text, hasTextContent := partMap["text"].(string); hasTextContent {
-				currentThoughtPart.WriteString(text)
+			if text := part.Get("text"); text.Exists() && text.Type == gjson.String {
+				currentThoughtPart.WriteString(text.String())
 				hasThought = true
 			}
-		} else if text, hasTextContent := partMap["text"].(string); hasTextContent {
+		} else if text := part.Get("text"); text.Exists() && text.Type == gjson.String {
 			// This is a regular text part - flush any pending thought first
 			flushThought() // Flush any pending thought first
 
-			currentTextPart.WriteString(text)
+			currentTextPart.WriteString(text.String())
 			hasText = true
 		} else {
 			// This is some other type of part (like function call) - flush both text and thought
 			flushText()
 			flushThought()
-			consolidated = append(consolidated, part)
+			consolidated = append(consolidated, partJSON)
 		}
 	}
 
@@ -611,20 +564,3 @@ func consolidateParts(parts []interface{}) []interface{} {
 
 	return consolidated
 }
-
-// convertToJSONString converts interface{} to JSON string using sjson/gjson.
-// This function provides a consistent way to serialize different data types to JSON strings
-// for inclusion in the Gemini API response structure.
-func convertToJSONString(v interface{}) string {
-	switch val := v.(type) {
-	case []interface{}:
-		return convertArrayToJSON(val)
-	case map[string]interface{}:
-		return convertMapToJSON(val)
-	default:
-		// For simple types, create a temporary JSON and extract the value
-		temp := `{"temp":null}`
-		temp, _ = sjson.Set(temp, "temp", val)
-		return gjson.Get(temp, "temp").Raw
-	}
-}

internal/translator/claude/openai/chat-completions/claude_openai_request.go

@@ -10,7 +10,6 @@ import (
 	"crypto/rand"
 	"crypto/sha256"
 	"encoding/hex"
-	"encoding/json"
 	"fmt"
 	"math/big"
 	"strings"
@@ -137,9 +136,6 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 	out, _ = sjson.Set(out, "stream", stream)
 
 	// Process messages and transform them to Claude Code format
-	var anthropicMessages []interface{}
-	var toolCallIDs []string // Track tool call IDs for matching with tool results
-
 	if messages := root.Get("messages"); messages.Exists() && messages.IsArray() {
 		messages.ForEach(func(_, message gjson.Result) bool {
 			role := message.Get("role").String()
@@ -152,33 +148,23 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 					role = "user"
 				}
 
-				msg := map[string]interface{}{
-					"role":    role,
-					"content": []interface{}{},
-				}
+				msg := `{"role":"","content":[]}`
+				msg, _ = sjson.Set(msg, "role", role)
 
 				// Handle content based on its type (string or array)
 				if contentResult.Exists() && contentResult.Type == gjson.String && contentResult.String() != "" {
-					// Simple text content conversion
-					msg["content"] = []interface{}{
-						map[string]interface{}{
-							"type": "text",
-							"text": contentResult.String(),
-						},
-					}
+					part := `{"type":"text","text":""}`
+					part, _ = sjson.Set(part, "text", contentResult.String())
+					msg, _ = sjson.SetRaw(msg, "content.-1", part)
 				} else if contentResult.Exists() && contentResult.IsArray() {
-					// Array of content parts processing
-					var contentParts []interface{}
 					contentResult.ForEach(func(_, part gjson.Result) bool {
 						partType := part.Get("type").String()
 
 						switch partType {
 						case "text":
-							// Text part conversion
-							contentParts = append(contentParts, map[string]interface{}{
-								"type": "text",
-								"text": part.Get("text").String(),
-							})
+							textPart := `{"type":"text","text":""}`
+							textPart, _ = sjson.Set(textPart, "text", part.Get("text").String())
+							msg, _ = sjson.SetRaw(msg, "content.-1", textPart)
 
 						case "image_url":
 							// Convert OpenAI image format to Claude Code format
@@ -191,132 +177,95 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 									mediaType := strings.TrimPrefix(mediaTypePart, "data:")
 									data := parts[1]
 
-									contentParts = append(contentParts, map[string]interface{}{
-										"type": "image",
-										"source": map[string]interface{}{
-											"type":       "base64",
-											"media_type": mediaType,
-											"data":       data,
-										},
-									})
+									imagePart := `{"type":"image","source":{"type":"base64","media_type":"","data":""}}`
+									imagePart, _ = sjson.Set(imagePart, "source.media_type", mediaType)
+									imagePart, _ = sjson.Set(imagePart, "source.data", data)
+									msg, _ = sjson.SetRaw(msg, "content.-1", imagePart)
 								}
 							}
 						}
 						return true
 					})
-					if len(contentParts) > 0 {
-						msg["content"] = contentParts
-					}
-				} else {
-					// Initialize empty content array for tool calls
-					msg["content"] = []interface{}{}
 				}
 
 				// Handle tool calls (for assistant messages)
 				if toolCalls := message.Get("tool_calls"); toolCalls.Exists() && toolCalls.IsArray() && role == "assistant" {
-					var contentParts []interface{}
-
-					// Add existing text content if any
-					if existingContent, ok := msg["content"].([]interface{}); ok {
-						contentParts = existingContent
-					}
-
 					toolCalls.ForEach(func(_, toolCall gjson.Result) bool {
 						if toolCall.Get("type").String() == "function" {
 							toolCallID := toolCall.Get("id").String()
 							if toolCallID == "" {
 								toolCallID = genToolCallID()
 							}
-							toolCallIDs = append(toolCallIDs, toolCallID)
 
 							function := toolCall.Get("function")
-							toolUse := map[string]interface{}{
-								"type": "tool_use",
-								"id":   toolCallID,
-								"name": function.Get("name").String(),
-							}
+							toolUse := `{"type":"tool_use","id":"","name":"","input":{}}`
+							toolUse, _ = sjson.Set(toolUse, "id", toolCallID)
+							toolUse, _ = sjson.Set(toolUse, "name", function.Get("name").String())
 
 							// Parse arguments for the tool call
 							if args := function.Get("arguments"); args.Exists() {
 								argsStr := args.String()
-								if argsStr != "" {
-									var argsMap map[string]interface{}
-									if err := json.Unmarshal([]byte(argsStr), &argsMap); err == nil {
-										toolUse["input"] = argsMap
+								if argsStr != "" && gjson.Valid(argsStr) {
+									argsJSON := gjson.Parse(argsStr)
+									if argsJSON.IsObject() {
+										toolUse, _ = sjson.SetRaw(toolUse, "input", argsJSON.Raw)
 									} else {
-										toolUse["input"] = map[string]interface{}{}
+										toolUse, _ = sjson.SetRaw(toolUse, "input", "{}")
 									}
 								} else {
-									toolUse["input"] = map[string]interface{}{}
+									toolUse, _ = sjson.SetRaw(toolUse, "input", "{}")
 								}
 							} else {
-								toolUse["input"] = map[string]interface{}{}
+								toolUse, _ = sjson.SetRaw(toolUse, "input", "{}")
 							}
 
-							contentParts = append(contentParts, toolUse)
+							msg, _ = sjson.SetRaw(msg, "content.-1", toolUse)
 						}
 						return true
 					})
-					msg["content"] = contentParts
 				}
 
-				anthropicMessages = append(anthropicMessages, msg)
+				out, _ = sjson.SetRaw(out, "messages.-1", msg)
 
 			case "tool":
 				// Handle tool result messages conversion
 				toolCallID := message.Get("tool_call_id").String()
 				content := message.Get("content").String()
 
-				// Create tool result message in Claude Code format
-				msg := map[string]interface{}{
-					"role": "user",
-					"content": []interface{}{
-						map[string]interface{}{
-							"type":        "tool_result",
-							"tool_use_id": toolCallID,
-							"content":     content,
-						},
-					},
-				}
-
-				anthropicMessages = append(anthropicMessages, msg)
+				msg := `{"role":"user","content":[{"type":"tool_result","tool_use_id":"","content":""}]}`
+				msg, _ = sjson.Set(msg, "content.0.tool_use_id", toolCallID)
+				msg, _ = sjson.Set(msg, "content.0.content", content)
+				out, _ = sjson.SetRaw(out, "messages.-1", msg)
 			}
 			return true
 		})
 	}
 
-	// Set messages in the output template
-	if len(anthropicMessages) > 0 {
-		messagesJSON, _ := json.Marshal(anthropicMessages)
-		out, _ = sjson.SetRaw(out, "messages", string(messagesJSON))
-	}
-
 	// Tools mapping: OpenAI tools -> Claude Code tools
 	if tools := root.Get("tools"); tools.Exists() && tools.IsArray() && len(tools.Array()) > 0 {
-		var anthropicTools []interface{}
+		hasAnthropicTools := false
 		tools.ForEach(func(_, tool gjson.Result) bool {
 			if tool.Get("type").String() == "function" {
 				function := tool.Get("function")
-				anthropicTool := map[string]interface{}{
-					"name":        function.Get("name").String(),
-					"description": function.Get("description").String(),
-				}
+				anthropicTool := `{"name":"","description":""}`
+				anthropicTool, _ = sjson.Set(anthropicTool, "name", function.Get("name").String())
+				anthropicTool, _ = sjson.Set(anthropicTool, "description", function.Get("description").String())
 
 				// Convert parameters schema for the tool
 				if parameters := function.Get("parameters"); parameters.Exists() {
-					anthropicTool["input_schema"] = parameters.Value()
-				} else if parameters = function.Get("parametersJsonSchema"); parameters.Exists() {
-					anthropicTool["input_schema"] = parameters.Value()
+					anthropicTool, _ = sjson.SetRaw(anthropicTool, "input_schema", parameters.Raw)
+				} else if parameters := function.Get("parametersJsonSchema"); parameters.Exists() {
+					anthropicTool, _ = sjson.SetRaw(anthropicTool, "input_schema", parameters.Raw)
 				}
 
-				anthropicTools = append(anthropicTools, anthropicTool)
+				out, _ = sjson.SetRaw(out, "tools.-1", anthropicTool)
+				hasAnthropicTools = true
 			}
 			return true
 		})
 
-		if len(anthropicTools) > 0 {
-			toolsJSON, _ := json.Marshal(anthropicTools)
-			out, _ = sjson.SetRaw(out, "tools", string(toolsJSON))
+		if !hasAnthropicTools {
+			out, _ = sjson.Delete(out, "tools")
 		}
 	}
 
@@ -329,18 +278,17 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 			case "none":
 				// Don't set tool_choice, Claude Code will not use tools
 			case "auto":
-				out, _ = sjson.Set(out, "tool_choice", map[string]interface{}{"type": "auto"})
+				out, _ = sjson.SetRaw(out, "tool_choice", `{"type":"auto"}`)
 			case "required":
-				out, _ = sjson.Set(out, "tool_choice", map[string]interface{}{"type": "any"})
+				out, _ = sjson.SetRaw(out, "tool_choice", `{"type":"any"}`)
 			}
 		case gjson.JSON:
 			// Specific tool choice mapping
 			if toolChoice.Get("type").String() == "function" {
 				functionName := toolChoice.Get("function.name").String()
-				out, _ = sjson.Set(out, "tool_choice", map[string]interface{}{
-					"type": "tool",
-					"name": functionName,
-				})
+				toolChoiceJSON := `{"type":"tool","name":""}`
+				toolChoiceJSON, _ = sjson.Set(toolChoiceJSON, "name", functionName)
+				out, _ = sjson.SetRaw(out, "tool_choice", toolChoiceJSON)
 			}
 		default:
 		}

internal/translator/claude/openai/chat-completions/claude_openai_response.go

@@ -8,7 +8,7 @@ package chat_completions
 import (
 	"bytes"
 	"context"
-	"encoding/json"
+	"fmt"
 	"strings"
 	"time"
 
@@ -178,18 +178,11 @@ func ConvertClaudeResponseToOpenAI(_ context.Context, modelName string, original
 				if arguments == "" {
 					arguments = "{}"
 				}
-
-				toolCall := map[string]interface{}{
-					"index": index,
-					"id":    accumulator.ID,
-					"type":  "function",
-					"function": map[string]interface{}{
-						"name":      accumulator.Name,
-						"arguments": arguments,
-					},
-				}
-
-				template, _ = sjson.Set(template, "choices.0.delta.tool_calls", []interface{}{toolCall})
+				template, _ = sjson.Set(template, "choices.0.delta.tool_calls.0.index", index)
+				template, _ = sjson.Set(template, "choices.0.delta.tool_calls.0.id", accumulator.ID)
+				template, _ = sjson.Set(template, "choices.0.delta.tool_calls.0.type", "function")
+				template, _ = sjson.Set(template, "choices.0.delta.tool_calls.0.function.name", accumulator.Name)
+				template, _ = sjson.Set(template, "choices.0.delta.tool_calls.0.function.arguments", arguments)
 
 				// Clean up the accumulator for this index
 				delete((*param).(*ConvertAnthropicResponseToOpenAIParams).ToolCallsAccumulator, index)
@@ -210,12 +203,14 @@ func ConvertClaudeResponseToOpenAI(_ context.Context, modelName string, original
 
 		// Handle usage information for token counts
 		if usage := root.Get("usage"); usage.Exists() {
-			usageObj := map[string]interface{}{
-				"prompt_tokens":     usage.Get("input_tokens").Int(),
-				"completion_tokens": usage.Get("output_tokens").Int(),
-				"total_tokens":      usage.Get("input_tokens").Int() + usage.Get("output_tokens").Int(),
-			}
-			template, _ = sjson.Set(template, "usage", usageObj)
+			inputTokens := usage.Get("input_tokens").Int()
+			outputTokens := usage.Get("output_tokens").Int()
+			cacheReadInputTokens := usage.Get("cache_read_input_tokens").Int()
+			cacheCreationInputTokens := usage.Get("cache_creation_input_tokens").Int()
+			template, _ = sjson.Set(template, "usage.prompt_tokens", inputTokens+cacheCreationInputTokens)
+			template, _ = sjson.Set(template, "usage.completion_tokens", outputTokens)
+			template, _ = sjson.Set(template, "usage.total_tokens", inputTokens+outputTokens)
+			template, _ = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cacheReadInputTokens)
 		}
 		return []string{template}
 
@@ -230,14 +225,10 @@ func ConvertClaudeResponseToOpenAI(_ context.Context, modelName string, original
 	case "error":
 		// Error event - format and return error response
 		if errorData := root.Get("error"); errorData.Exists() {
-			errorResponse := map[string]interface{}{
-				"error": map[string]interface{}{
-					"message": errorData.Get("message").String(),
-					"type":    errorData.Get("type").String(),
-				},
-			}
-			errorJSON, _ := json.Marshal(errorResponse)
-			return []string{string(errorJSON)}
+			errorJSON := `{"error":{"message":"","type":""}}`
+			errorJSON, _ = sjson.Set(errorJSON, "error.message", errorData.Get("message").String())
+			errorJSON, _ = sjson.Set(errorJSON, "error.type", errorData.Get("type").String())
+			return []string{errorJSON}
 		}
 		return []string{}
 
@@ -293,15 +284,10 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 	var messageID string
 	var model string
 	var createdAt int64
-	var inputTokens, outputTokens int64
-	var reasoningTokens int64
 	var stopReason string
 	var contentParts []string
 	var reasoningParts []string
-	// Use map to track tool calls by index for proper merging
-	toolCallsMap := make(map[int]map[string]interface{})
-	// Track tool call arguments accumulation
-	toolCallArgsMap := make(map[int]strings.Builder)
+	toolCallsAccumulator := make(map[int]*ToolCallAccumulator)
 
 	for _, chunk := range chunks {
 		root := gjson.ParseBytes(chunk)
@@ -314,9 +300,6 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 				messageID = message.Get("id").String()
 				model = message.Get("model").String()
 				createdAt = time.Now().Unix()
-				if usage := message.Get("usage"); usage.Exists() {
-					inputTokens = usage.Get("input_tokens").Int()
-				}
 			}
 
 		case "content_block_start":
@@ -327,18 +310,12 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 					// Start of thinking/reasoning content - skip for now as it's handled in delta
 					continue
 				} else if blockType == "tool_use" {
-					// Initialize tool call tracking for this index
+					// Initialize tool call accumulator for this index
 					index := int(root.Get("index").Int())
-					toolCallsMap[index] = map[string]interface{}{
-						"id":   contentBlock.Get("id").String(),
-						"type": "function",
-						"function": map[string]interface{}{
-							"name":      contentBlock.Get("name").String(),
-							"arguments": "",
-						},
+					toolCallsAccumulator[index] = &ToolCallAccumulator{
+						ID:   contentBlock.Get("id").String(),
+						Name: contentBlock.Get("name").String(),
 					}
-					// Initialize arguments builder for this tool call
-					toolCallArgsMap[index] = strings.Builder{}
 				}
 			}
 
@@ -361,9 +338,8 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 					// Accumulate tool call arguments
 					if partialJSON := delta.Get("partial_json"); partialJSON.Exists() {
 						index := int(root.Get("index").Int())
-						if builder, exists := toolCallArgsMap[index]; exists {
-							builder.WriteString(partialJSON.String())
-							toolCallArgsMap[index] = builder
+						if accumulator, exists := toolCallsAccumulator[index]; exists {
+							accumulator.Arguments.WriteString(partialJSON.String())
 						}
 					}
 				}
@@ -372,14 +348,9 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 		case "content_block_stop":
 			// Finalize tool call arguments for this index when content block ends
 			index := int(root.Get("index").Int())
-			if toolCall, exists := toolCallsMap[index]; exists {
-				if builder, argsExists := toolCallArgsMap[index]; argsExists {
-					// Set the accumulated arguments for the tool call
-					arguments := builder.String()
-					if arguments == "" {
-						arguments = "{}"
-					}
-					toolCall["function"].(map[string]interface{})["arguments"] = arguments
+			if accumulator, exists := toolCallsAccumulator[index]; exists {
+				if accumulator.Arguments.Len() == 0 {
+					accumulator.Arguments.WriteString("{}")
 				}
 			}
 
@@ -391,11 +362,14 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 				}
 			}
 			if usage := root.Get("usage"); usage.Exists() {
-				outputTokens = usage.Get("output_tokens").Int()
-				// Estimate reasoning tokens from accumulated thinking content
-				if len(reasoningParts) > 0 {
-					reasoningTokens = int64(len(strings.Join(reasoningParts, "")) / 4) // Rough estimation
-				}
+				inputTokens := usage.Get("input_tokens").Int()
+				outputTokens := usage.Get("output_tokens").Int()
+				cacheReadInputTokens := usage.Get("cache_read_input_tokens").Int()
+				cacheCreationInputTokens := usage.Get("cache_creation_input_tokens").Int()
+				out, _ = sjson.Set(out, "usage.prompt_tokens", inputTokens+cacheCreationInputTokens)
+				out, _ = sjson.Set(out, "usage.completion_tokens", outputTokens)
+				out, _ = sjson.Set(out, "usage.total_tokens", inputTokens+outputTokens)
+				out, _ = sjson.Set(out, "usage.prompt_tokens_details.cached_tokens", cacheReadInputTokens)
 			}
 		}
 	}
@@ -417,24 +391,35 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 	}
 
 	// Set tool calls if any were accumulated during processing
-	if len(toolCallsMap) > 0 {
-		// Convert tool calls map to array, preserving order by index
-		var toolCallsArray []interface{}
-		// Find the maximum index to determine the range
+	if len(toolCallsAccumulator) > 0 {
+		toolCallsCount := 0
 		maxIndex := -1
-		for index := range toolCallsMap {
+		for index := range toolCallsAccumulator {
 			if index > maxIndex {
 				maxIndex = index
 			}
 		}
-		// Iterate through all possible indices up to maxIndex
+
 		for i := 0; i <= maxIndex; i++ {
-			if toolCall, exists := toolCallsMap[i]; exists {
-				toolCallsArray = append(toolCallsArray, toolCall)
+			accumulator, exists := toolCallsAccumulator[i]
+			if !exists {
+				continue
 			}
+
+			arguments := accumulator.Arguments.String()
+
+			idPath := fmt.Sprintf("choices.0.message.tool_calls.%d.id", toolCallsCount)
+			typePath := fmt.Sprintf("choices.0.message.tool_calls.%d.type", toolCallsCount)
+			namePath := fmt.Sprintf("choices.0.message.tool_calls.%d.function.name", toolCallsCount)
+			argumentsPath := fmt.Sprintf("choices.0.message.tool_calls.%d.function.arguments", toolCallsCount)
+
+			out, _ = sjson.Set(out, idPath, accumulator.ID)
+			out, _ = sjson.Set(out, typePath, "function")
+			out, _ = sjson.Set(out, namePath, accumulator.Name)
+			out, _ = sjson.Set(out, argumentsPath, arguments)
+			toolCallsCount++
 		}
-		if len(toolCallsArray) > 0 {
-			out, _ = sjson.Set(out, "choices.0.message.tool_calls", toolCallsArray)
+		if toolCallsCount > 0 {
 			out, _ = sjson.Set(out, "choices.0.finish_reason", "tool_calls")
 		} else {
 			out, _ = sjson.Set(out, "choices.0.finish_reason", mapAnthropicStopReasonToOpenAI(stopReason))
@@ -443,16 +428,5 @@ func ConvertClaudeResponseToOpenAINonStream(_ context.Context, _ string, origina
 		out, _ = sjson.Set(out, "choices.0.finish_reason", mapAnthropicStopReasonToOpenAI(stopReason))
 	}
 
-	// Set usage information including prompt tokens, completion tokens, and total tokens
-	totalTokens := inputTokens + outputTokens
-	out, _ = sjson.Set(out, "usage.prompt_tokens", inputTokens)
-	out, _ = sjson.Set(out, "usage.completion_tokens", outputTokens)
-	out, _ = sjson.Set(out, "usage.total_tokens", totalTokens)
-
-	// Add reasoning tokens to usage details if any reasoning content was processed
-	if reasoningTokens > 0 {
-		out, _ = sjson.Set(out, "usage.completion_tokens_details.reasoning_tokens", reasoningTokens)
-	}
-
 	return out
 }

internal/translator/claude/openai/responses/claude_openai-responses_request.go

@@ -114,13 +114,16 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 					var builder strings.Builder
 					if parts := item.Get("content"); parts.Exists() && parts.IsArray() {
 						parts.ForEach(func(_, part gjson.Result) bool {
-							text := part.Get("text").String()
+							textResult := part.Get("text")
+							text := textResult.String()
 							if builder.Len() > 0 && text != "" {
 								builder.WriteByte('\n')
 							}
 							builder.WriteString(text)
 							return true
 						})
+					} else if parts.Type == gjson.String {
+						builder.WriteString(parts.String())
 					}
 					instructionsText = builder.String()
 					if instructionsText != "" {
@@ -207,6 +210,8 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 						}
 						return true
 					})
+				} else if parts.Type == gjson.String {
+					textAggregate.WriteString(parts.String())
 				}
 
 				// Fallback to given role if content types not decisive
@@ -254,7 +259,10 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 				toolUse, _ = sjson.Set(toolUse, "id", callID)
 				toolUse, _ = sjson.Set(toolUse, "name", name)
 				if argsStr != "" && gjson.Valid(argsStr) {
-					toolUse, _ = sjson.SetRaw(toolUse, "input", argsStr)
+					argsJSON := gjson.Parse(argsStr)
+					if argsJSON.IsObject() {
+						toolUse, _ = sjson.SetRaw(toolUse, "input", argsJSON.Raw)
+					}
 				}
 
 				asst := `{"role":"assistant","content":[]}`
@@ -309,16 +317,18 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 		case gjson.String:
 			switch toolChoice.String() {
 			case "auto":
-				out, _ = sjson.Set(out, "tool_choice", map[string]interface{}{"type": "auto"})
+				out, _ = sjson.SetRaw(out, "tool_choice", `{"type":"auto"}`)
 			case "none":
 				// Leave unset; implies no tools
 			case "required":
-				out, _ = sjson.Set(out, "tool_choice", map[string]interface{}{"type": "any"})
+				out, _ = sjson.SetRaw(out, "tool_choice", `{"type":"any"}`)
 			}
 		case gjson.JSON:
 			if toolChoice.Get("type").String() == "function" {
 				fn := toolChoice.Get("function.name").String()
-				out, _ = sjson.Set(out, "tool_choice", map[string]interface{}{"type": "tool", "name": fn})
+				toolChoiceJSON := `{"name":"","type":"tool"}`
+				toolChoiceJSON, _ = sjson.Set(toolChoiceJSON, "name", fn)
+				out, _ = sjson.SetRaw(out, "tool_choice", toolChoiceJSON)
 			}
 		default:
 

internal/translator/claude/openai/responses/claude_openai-responses_response.go

@@ -95,7 +95,7 @@ func ConvertClaudeResponseToOpenAIResponses(ctx context.Context, modelName strin
 				}
 			}
 			// response.created
-			created := `{"type":"response.created","sequence_number":0,"response":{"id":"","object":"response","created_at":0,"status":"in_progress","background":false,"error":null,"instructions":""}}`
+			created := `{"type":"response.created","sequence_number":0,"response":{"id":"","object":"response","created_at":0,"status":"in_progress","background":false,"error":null,"output":[]}}`
 			created, _ = sjson.Set(created, "sequence_number", nextSeq())
 			created, _ = sjson.Set(created, "response.id", st.ResponseID)
 			created, _ = sjson.Set(created, "response.created_at", st.CreatedAt)
@@ -197,11 +197,11 @@ func ConvertClaudeResponseToOpenAIResponses(ctx context.Context, modelName strin
 			if st.ReasoningActive {
 				if t := d.Get("thinking"); t.Exists() {
 					st.ReasoningBuf.WriteString(t.String())
-					msg := `{"type":"response.reasoning_summary_text.delta","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"text":""}`
+					msg := `{"type":"response.reasoning_summary_text.delta","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"delta":""}`
 					msg, _ = sjson.Set(msg, "sequence_number", nextSeq())
 					msg, _ = sjson.Set(msg, "item_id", st.ReasoningItemID)
 					msg, _ = sjson.Set(msg, "output_index", st.ReasoningIndex)
-					msg, _ = sjson.Set(msg, "text", t.String())
+					msg, _ = sjson.Set(msg, "delta", t.String())
 					out = append(out, emitEvent("response.reasoning_summary_text.delta", msg))
 				}
 			}
@@ -344,31 +344,20 @@ func ConvertClaudeResponseToOpenAIResponses(ctx context.Context, modelName strin
 		}
 
 		// Build response.output from aggregated state
-		var outputs []interface{}
+		outputsWrapper := `{"arr":[]}`
 		// reasoning item (if any)
 		if st.ReasoningBuf.Len() > 0 || st.ReasoningPartAdded {
-			r := map[string]interface{}{
-				"id":      st.ReasoningItemID,
-				"type":    "reasoning",
-				"summary": []interface{}{map[string]interface{}{"type": "summary_text", "text": st.ReasoningBuf.String()}},
-			}
-			outputs = append(outputs, r)
+			item := `{"id":"","type":"reasoning","summary":[{"type":"summary_text","text":""}]}`
+			item, _ = sjson.Set(item, "id", st.ReasoningItemID)
+			item, _ = sjson.Set(item, "summary.0.text", st.ReasoningBuf.String())
+			outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 		}
 		// assistant message item (if any text)
 		if st.TextBuf.Len() > 0 || st.InTextBlock || st.CurrentMsgID != "" {
-			m := map[string]interface{}{
-				"id":     st.CurrentMsgID,
-				"type":   "message",
-				"status": "completed",
-				"content": []interface{}{map[string]interface{}{
-					"type":        "output_text",
-					"annotations": []interface{}{},
-					"logprobs":    []interface{}{},
-					"text":        st.TextBuf.String(),
-				}},
-				"role": "assistant",
-			}
-			outputs = append(outputs, m)
+			item := `{"id":"","type":"message","status":"completed","content":[{"type":"output_text","annotations":[],"logprobs":[],"text":""}],"role":"assistant"}`
+			item, _ = sjson.Set(item, "id", st.CurrentMsgID)
+			item, _ = sjson.Set(item, "content.0.text", st.TextBuf.String())
+			outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 		}
 		// function_call items (in ascending index order for determinism)
 		if len(st.FuncArgsBuf) > 0 {
@@ -395,19 +384,16 @@ func ConvertClaudeResponseToOpenAIResponses(ctx context.Context, modelName strin
 				if callID == "" && st.CurrentFCID != "" {
 					callID = st.CurrentFCID
 				}
-				item := map[string]interface{}{
-					"id":        fmt.Sprintf("fc_%s", callID),
-					"type":      "function_call",
-					"status":    "completed",
-					"arguments": args,
-					"call_id":   callID,
-					"name":      name,
-				}
-				outputs = append(outputs, item)
+				item := `{"id":"","type":"function_call","status":"completed","arguments":"","call_id":"","name":""}`
+				item, _ = sjson.Set(item, "id", fmt.Sprintf("fc_%s", callID))
+				item, _ = sjson.Set(item, "arguments", args)
+				item, _ = sjson.Set(item, "call_id", callID)
+				item, _ = sjson.Set(item, "name", name)
+				outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 			}
 		}
-		if len(outputs) > 0 {
-			completed, _ = sjson.Set(completed, "response.output", outputs)
+		if gjson.Get(outputsWrapper, "arr.#").Int() > 0 {
+			completed, _ = sjson.SetRaw(completed, "response.output", gjson.Get(outputsWrapper, "arr").Raw)
 		}
 
 		reasoningTokens := int64(0)
@@ -628,27 +614,18 @@ func ConvertClaudeResponseToOpenAIResponsesNonStream(_ context.Context, _ string
 	}
 
 	// Build output array
-	var outputs []interface{}
+	outputsWrapper := `{"arr":[]}`
 	if reasoningBuf.Len() > 0 {
-		outputs = append(outputs, map[string]interface{}{
-			"id":      reasoningItemID,
-			"type":    "reasoning",
-			"summary": []interface{}{map[string]interface{}{"type": "summary_text", "text": reasoningBuf.String()}},
-		})
+		item := `{"id":"","type":"reasoning","summary":[{"type":"summary_text","text":""}]}`
+		item, _ = sjson.Set(item, "id", reasoningItemID)
+		item, _ = sjson.Set(item, "summary.0.text", reasoningBuf.String())
+		outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 	}
 	if currentMsgID != "" || textBuf.Len() > 0 {
-		outputs = append(outputs, map[string]interface{}{
-			"id":     currentMsgID,
-			"type":   "message",
-			"status": "completed",
-			"content": []interface{}{map[string]interface{}{
-				"type":        "output_text",
-				"annotations": []interface{}{},
-				"logprobs":    []interface{}{},
-				"text":        textBuf.String(),
-			}},
-			"role": "assistant",
-		})
+		item := `{"id":"","type":"message","status":"completed","content":[{"type":"output_text","annotations":[],"logprobs":[],"text":""}],"role":"assistant"}`
+		item, _ = sjson.Set(item, "id", currentMsgID)
+		item, _ = sjson.Set(item, "content.0.text", textBuf.String())
+		outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 	}
 	if len(toolCalls) > 0 {
 		// Preserve index order
@@ -669,18 +646,16 @@ func ConvertClaudeResponseToOpenAIResponsesNonStream(_ context.Context, _ string
 			if args == "" {
 				args = "{}"
 			}
-			outputs = append(outputs, map[string]interface{}{
-				"id":        fmt.Sprintf("fc_%s", st.id),
-				"type":      "function_call",
-				"status":    "completed",
-				"arguments": args,
-				"call_id":   st.id,
-				"name":      st.name,
-			})
+			item := `{"id":"","type":"function_call","status":"completed","arguments":"","call_id":"","name":""}`
+			item, _ = sjson.Set(item, "id", fmt.Sprintf("fc_%s", st.id))
+			item, _ = sjson.Set(item, "arguments", args)
+			item, _ = sjson.Set(item, "call_id", st.id)
+			item, _ = sjson.Set(item, "name", st.name)
+			outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 		}
 	}
-	if len(outputs) > 0 {
-		out, _ = sjson.Set(out, "output", outputs)
+	if gjson.Get(outputsWrapper, "arr.#").Int() > 0 {
+		out, _ = sjson.SetRaw(out, "output", gjson.Get(outputsWrapper, "arr").Raw)
 	}
 
 	// Usage

internal/translator/codex/claude/codex_claude_request.go

@@ -219,15 +219,20 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	// Convert thinking.budget_tokens to reasoning.effort for level-based models
 	reasoningEffort := "medium" // default
 	if thinking := rootResult.Get("thinking"); thinking.Exists() && thinking.IsObject() {
-		if thinking.Get("type").String() == "enabled" {
+		switch thinking.Get("type").String() {
+		case "enabled":
 			if util.ModelUsesThinkingLevels(modelName) {
 				if budgetTokens := thinking.Get("budget_tokens"); budgetTokens.Exists() {
 					budget := int(budgetTokens.Int())
-					if effort, ok := util.OpenAIThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
+					if effort, ok := util.ThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
 						reasoningEffort = effort
 					}
 				}
 			}
+		case "disabled":
+			if effort, ok := util.ThinkingBudgetToEffort(modelName, 0); ok && effort != "" {
+				reasoningEffort = effort
+			}
 		}
 	}
 	template, _ = sjson.Set(template, "reasoning.effort", reasoningEffort)

internal/translator/codex/claude/codex_claude_response.go

@@ -9,7 +9,6 @@ package claude
 import (
 	"bytes"
 	"context"
-	"encoding/json"
 	"fmt"
 	"strings"
 
@@ -191,21 +190,12 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 		return ""
 	}
 
-	response := map[string]interface{}{
-		"id":            responseData.Get("id").String(),
-		"type":          "message",
-		"role":          "assistant",
-		"model":         responseData.Get("model").String(),
-		"content":       []interface{}{},
-		"stop_reason":   nil,
-		"stop_sequence": nil,
-		"usage": map[string]interface{}{
-			"input_tokens":  responseData.Get("usage.input_tokens").Int(),
-			"output_tokens": responseData.Get("usage.output_tokens").Int(),
-		},
-	}
+	out := `{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}`
+	out, _ = sjson.Set(out, "id", responseData.Get("id").String())
+	out, _ = sjson.Set(out, "model", responseData.Get("model").String())
+	out, _ = sjson.Set(out, "usage.input_tokens", responseData.Get("usage.input_tokens").Int())
+	out, _ = sjson.Set(out, "usage.output_tokens", responseData.Get("usage.output_tokens").Int())
 
-	var contentBlocks []interface{}
 	hasToolCall := false
 
 	if output := responseData.Get("output"); output.Exists() && output.IsArray() {
@@ -244,10 +234,9 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 					}
 				}
 				if thinkingBuilder.Len() > 0 {
-					contentBlocks = append(contentBlocks, map[string]interface{}{
-						"type":     "thinking",
-						"thinking": thinkingBuilder.String(),
-					})
+					block := `{"type":"thinking","thinking":""}`
+					block, _ = sjson.Set(block, "thinking", thinkingBuilder.String())
+					out, _ = sjson.SetRaw(out, "content.-1", block)
 				}
 			case "message":
 				if content := item.Get("content"); content.Exists() {
@@ -256,10 +245,9 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 							if part.Get("type").String() == "output_text" {
 								text := part.Get("text").String()
 								if text != "" {
-									contentBlocks = append(contentBlocks, map[string]interface{}{
-										"type": "text",
-										"text": text,
-									})
+									block := `{"type":"text","text":""}`
+									block, _ = sjson.Set(block, "text", text)
+									out, _ = sjson.SetRaw(out, "content.-1", block)
 								}
 							}
 							return true
@@ -267,10 +255,9 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 					} else {
 						text := content.String()
 						if text != "" {
-							contentBlocks = append(contentBlocks, map[string]interface{}{
-								"type": "text",
-								"text": text,
-							})
+							block := `{"type":"text","text":""}`
+							block, _ = sjson.Set(block, "text", text)
+							out, _ = sjson.SetRaw(out, "content.-1", block)
 						}
 					}
 				}
@@ -281,54 +268,41 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 					name = original
 				}
 
-				toolBlock := map[string]interface{}{
-					"type":  "tool_use",
-					"id":    item.Get("call_id").String(),
-					"name":  name,
-					"input": map[string]interface{}{},
-				}
-
-				if argsStr := item.Get("arguments").String(); argsStr != "" {
-					var args interface{}
-					if err := json.Unmarshal([]byte(argsStr), &args); err == nil {
-						toolBlock["input"] = args
+				toolBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
+				toolBlock, _ = sjson.Set(toolBlock, "id", item.Get("call_id").String())
+				toolBlock, _ = sjson.Set(toolBlock, "name", name)
+				inputRaw := "{}"
+				if argsStr := item.Get("arguments").String(); argsStr != "" && gjson.Valid(argsStr) {
+					argsJSON := gjson.Parse(argsStr)
+					if argsJSON.IsObject() {
+						inputRaw = argsJSON.Raw
 					}
 				}
-
-				contentBlocks = append(contentBlocks, toolBlock)
+				toolBlock, _ = sjson.SetRaw(toolBlock, "input", inputRaw)
+				out, _ = sjson.SetRaw(out, "content.-1", toolBlock)
 			}
 			return true
 		})
 	}
 
-	if len(contentBlocks) > 0 {
-		response["content"] = contentBlocks
-	}
-
 	if stopReason := responseData.Get("stop_reason"); stopReason.Exists() && stopReason.String() != "" {
-		response["stop_reason"] = stopReason.String()
+		out, _ = sjson.Set(out, "stop_reason", stopReason.String())
 	} else if hasToolCall {
-		response["stop_reason"] = "tool_use"
+		out, _ = sjson.Set(out, "stop_reason", "tool_use")
 	} else {
-		response["stop_reason"] = "end_turn"
+		out, _ = sjson.Set(out, "stop_reason", "end_turn")
 	}
 
 	if stopSequence := responseData.Get("stop_sequence"); stopSequence.Exists() && stopSequence.String() != "" {
-		response["stop_sequence"] = stopSequence.Value()
+		out, _ = sjson.SetRaw(out, "stop_sequence", stopSequence.Raw)
 	}
 
 	if responseData.Get("usage.input_tokens").Exists() || responseData.Get("usage.output_tokens").Exists() {
-		response["usage"] = map[string]interface{}{
-			"input_tokens":  responseData.Get("usage.input_tokens").Int(),
-			"output_tokens": responseData.Get("usage.output_tokens").Int(),
-		}
+		out, _ = sjson.Set(out, "usage.input_tokens", responseData.Get("usage.input_tokens").Int())
+		out, _ = sjson.Set(out, "usage.output_tokens", responseData.Get("usage.output_tokens").Int())
 	}
 
-	responseJSON, err := json.Marshal(response)
-	if err != nil {
-		return ""
-	}
-	return string(responseJSON)
+	return out
 }
 
 // buildReverseMapFromClaudeOriginalShortToOriginal builds a map[short]original from original Claude request tools.

internal/translator/codex/gemini/codex_gemini_request.go

@@ -253,7 +253,7 @@ func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 			if util.ModelUsesThinkingLevels(modelName) {
 				if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
 					budget := int(thinkingBudget.Int())
-					if effort, ok := util.OpenAIThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
+					if effort, ok := util.ThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
 						reasoningEffort = effort
 					}
 				}

internal/translator/codex/gemini/codex_gemini_response.go

@@ -7,7 +7,6 @@ package gemini
 import (
 	"bytes"
 	"context"
-	"encoding/json"
 	"fmt"
 	"time"
 
@@ -190,19 +189,19 @@ func ConvertCodexResponseToGeminiNonStream(_ context.Context, modelName string,
 		}
 
 		// Process output content to build parts array
-		var parts []interface{}
 		hasToolCall := false
-		var pendingFunctionCalls []interface{}
+		var pendingFunctionCalls []string
 
 		flushPendingFunctionCalls := func() {
-			if len(pendingFunctionCalls) > 0 {
-				// Add all pending function calls as individual parts
-				// This maintains the original Gemini API format while ensuring consecutive calls are grouped together
-				for _, fc := range pendingFunctionCalls {
-					parts = append(parts, fc)
-				}
-				pendingFunctionCalls = nil
+			if len(pendingFunctionCalls) == 0 {
+				return
 			}
+			// Add all pending function calls as individual parts
+			// This maintains the original Gemini API format while ensuring consecutive calls are grouped together
+			for _, fc := range pendingFunctionCalls {
+				template, _ = sjson.SetRaw(template, "candidates.0.content.parts.-1", fc)
+			}
+			pendingFunctionCalls = nil
 		}
 
 		if output := responseData.Get("output"); output.Exists() && output.IsArray() {
@@ -216,11 +215,9 @@ func ConvertCodexResponseToGeminiNonStream(_ context.Context, modelName string,
 
 					// Add thinking content
 					if content := value.Get("content"); content.Exists() {
-						part := map[string]interface{}{
-							"thought": true,
-							"text":    content.String(),
-						}
-						parts = append(parts, part)
+						part := `{"text":"","thought":true}`
+						part, _ = sjson.Set(part, "text", content.String())
+						template, _ = sjson.SetRaw(template, "candidates.0.content.parts.-1", part)
 					}
 
 				case "message":
@@ -232,10 +229,9 @@ func ConvertCodexResponseToGeminiNonStream(_ context.Context, modelName string,
 						content.ForEach(func(_, contentItem gjson.Result) bool {
 							if contentItem.Get("type").String() == "output_text" {
 								if text := contentItem.Get("text"); text.Exists() {
-									part := map[string]interface{}{
-										"text": text.String(),
-									}
-									parts = append(parts, part)
+									part := `{"text":""}`
+									part, _ = sjson.Set(part, "text", text.String())
+									template, _ = sjson.SetRaw(template, "candidates.0.content.parts.-1", part)
 								}
 							}
 							return true
@@ -245,28 +241,21 @@ func ConvertCodexResponseToGeminiNonStream(_ context.Context, modelName string,
 				case "function_call":
 					// Collect function call for potential merging with consecutive ones
 					hasToolCall = true
-					functionCall := map[string]interface{}{
-						"functionCall": map[string]interface{}{
-							"name": func() string {
-								n := value.Get("name").String()
-								rev := buildReverseMapFromGeminiOriginal(originalRequestRawJSON)
-								if orig, ok := rev[n]; ok {
-									return orig
-								}
-								return n
-							}(),
-							"args": map[string]interface{}{},
-						},
+					functionCall := `{"functionCall":{"args":{},"name":""}}`
+					{
+						n := value.Get("name").String()
+						rev := buildReverseMapFromGeminiOriginal(originalRequestRawJSON)
+						if orig, ok := rev[n]; ok {
+							n = orig
+						}
+						functionCall, _ = sjson.Set(functionCall, "functionCall.name", n)
 					}
 
 					// Parse and set arguments
 					if argsStr := value.Get("arguments").String(); argsStr != "" {
 						argsResult := gjson.Parse(argsStr)
 						if argsResult.IsObject() {
-							var args map[string]interface{}
-							if err := json.Unmarshal([]byte(argsStr), &args); err == nil {
-								functionCall["functionCall"].(map[string]interface{})["args"] = args
-							}
+							functionCall, _ = sjson.SetRaw(functionCall, "functionCall.args", argsStr)
 						}
 					}
 
@@ -279,11 +268,6 @@ func ConvertCodexResponseToGeminiNonStream(_ context.Context, modelName string,
 			flushPendingFunctionCalls()
 		}
 
-		// Set the parts array
-		if len(parts) > 0 {
-			template, _ = sjson.SetRaw(template, "candidates.0.content.parts", mustMarshalJSON(parts))
-		}
-
 		// Set finish reason based on whether there were tool calls
 		if hasToolCall {
 			template, _ = sjson.Set(template, "candidates.0.finishReason", "STOP")
@@ -323,15 +307,6 @@ func buildReverseMapFromGeminiOriginal(original []byte) map[string]string {
 	return rev
 }
 
-// mustMarshalJSON marshals a value to JSON, panicking on error.
-func mustMarshalJSON(v interface{}) string {
-	data, err := json.Marshal(v)
-	if err != nil {
-		return ""
-	}
-	return string(data)
-}
-
 func GeminiTokenCount(ctx context.Context, count int64) string {
 	return fmt.Sprintf(`{"totalTokens":%d,"promptTokensDetails":[{"modality":"TEXT","tokenCount":%d}]}`, count, count)
 }

internal/translator/gemini-cli/claude/gemini-cli_claude_request.go

@@ -7,10 +7,8 @@ package claude
 
 import (
 	"bytes"
-	"encoding/json"
 	"strings"
 
-	client "github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
@@ -41,92 +39,102 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 	rawJSON := bytes.Clone(inputRawJSON)
 	rawJSON = bytes.Replace(rawJSON, []byte(`"url":{"type":"string","format":"uri",`), []byte(`"url":{"type":"string",`), -1)
 
+	// Build output Gemini CLI request JSON
+	out := `{"model":"","request":{"contents":[]}}`
+	out, _ = sjson.Set(out, "model", modelName)
+
 	// system instruction
-	var systemInstruction *client.Content
-	systemResult := gjson.GetBytes(rawJSON, "system")
-	if systemResult.IsArray() {
-		systemResults := systemResult.Array()
-		systemInstruction = &client.Content{Role: "user", Parts: []client.Part{}}
-		for i := 0; i < len(systemResults); i++ {
-			systemPromptResult := systemResults[i]
-			systemTypePromptResult := systemPromptResult.Get("type")
-			if systemTypePromptResult.Type == gjson.String && systemTypePromptResult.String() == "text" {
-				systemPrompt := systemPromptResult.Get("text").String()
-				systemPart := client.Part{Text: systemPrompt}
-				systemInstruction.Parts = append(systemInstruction.Parts, systemPart)
+	if systemResult := gjson.GetBytes(rawJSON, "system"); systemResult.IsArray() {
+		systemInstruction := `{"role":"user","parts":[]}`
+		hasSystemParts := false
+		systemResult.ForEach(func(_, systemPromptResult gjson.Result) bool {
+			if systemPromptResult.Get("type").String() == "text" {
+				textResult := systemPromptResult.Get("text")
+				if textResult.Type == gjson.String {
+					part := `{"text":""}`
+					part, _ = sjson.Set(part, "text", textResult.String())
+					systemInstruction, _ = sjson.SetRaw(systemInstruction, "parts.-1", part)
+					hasSystemParts = true
+				}
 			}
+			return true
+		})
+		if hasSystemParts {
+			out, _ = sjson.SetRaw(out, "request.systemInstruction", systemInstruction)
 		}
-		if len(systemInstruction.Parts) == 0 {
-			systemInstruction = nil
-		}
+	} else if systemResult.Type == gjson.String {
+		out, _ = sjson.Set(out, "request.systemInstruction.parts.-1.text", systemResult.String())
 	}
 
 	// contents
-	contents := make([]client.Content, 0)
-	messagesResult := gjson.GetBytes(rawJSON, "messages")
-	if messagesResult.IsArray() {
-		messageResults := messagesResult.Array()
-		for i := 0; i < len(messageResults); i++ {
-			messageResult := messageResults[i]
+	if messagesResult := gjson.GetBytes(rawJSON, "messages"); messagesResult.IsArray() {
+		messagesResult.ForEach(func(_, messageResult gjson.Result) bool {
 			roleResult := messageResult.Get("role")
 			if roleResult.Type != gjson.String {
-				continue
+				return true
 			}
 			role := roleResult.String()
 			if role == "assistant" {
 				role = "model"
 			}
-			clientContent := client.Content{Role: role, Parts: []client.Part{}}
+
+			contentJSON := `{"role":"","parts":[]}`
+			contentJSON, _ = sjson.Set(contentJSON, "role", role)
+
 			contentsResult := messageResult.Get("content")
 			if contentsResult.IsArray() {
-				contentResults := contentsResult.Array()
-				for j := 0; j < len(contentResults); j++ {
-					contentResult := contentResults[j]
-					contentTypeResult := contentResult.Get("type")
-					if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "text" {
-						prompt := contentResult.Get("text").String()
-						clientContent.Parts = append(clientContent.Parts, client.Part{Text: prompt})
-					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "tool_use" {
+				contentsResult.ForEach(func(_, contentResult gjson.Result) bool {
+					switch contentResult.Get("type").String() {
+					case "text":
+						part := `{"text":""}`
+						part, _ = sjson.Set(part, "text", contentResult.Get("text").String())
+						contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
+
+					case "tool_use":
 						functionName := contentResult.Get("name").String()
 						functionArgs := contentResult.Get("input").String()
-						var args map[string]any
-						if err := json.Unmarshal([]byte(functionArgs), &args); err == nil {
-							clientContent.Parts = append(clientContent.Parts, client.Part{
-								FunctionCall:     &client.FunctionCall{Name: functionName, Args: args},
-								ThoughtSignature: geminiCLIClaudeThoughtSignature,
-							})
+						argsResult := gjson.Parse(functionArgs)
+						if argsResult.IsObject() && gjson.Valid(functionArgs) {
+							part := `{"thoughtSignature":"","functionCall":{"name":"","args":{}}}`
+							part, _ = sjson.Set(part, "thoughtSignature", geminiCLIClaudeThoughtSignature)
+							part, _ = sjson.Set(part, "functionCall.name", functionName)
+							part, _ = sjson.SetRaw(part, "functionCall.args", functionArgs)
+							contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
 						}
-					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "tool_result" {
+
+					case "tool_result":
 						toolCallID := contentResult.Get("tool_use_id").String()
-						if toolCallID != "" {
-							funcName := toolCallID
-							toolCallIDs := strings.Split(toolCallID, "-")
-							if len(toolCallIDs) > 1 {
-								funcName = strings.Join(toolCallIDs[0:len(toolCallIDs)-1], "-")
-							}
-							responseData := contentResult.Get("content").Raw
-							functionResponse := client.FunctionResponse{Name: funcName, Response: map[string]interface{}{"result": responseData}}
-							clientContent.Parts = append(clientContent.Parts, client.Part{FunctionResponse: &functionResponse})
+						if toolCallID == "" {
+							return true
 						}
+						funcName := toolCallID
+						toolCallIDs := strings.Split(toolCallID, "-")
+						if len(toolCallIDs) > 1 {
+							funcName = strings.Join(toolCallIDs[0:len(toolCallIDs)-1], "-")
+						}
+						responseData := contentResult.Get("content").Raw
+						part := `{"functionResponse":{"name":"","response":{"result":""}}}`
+						part, _ = sjson.Set(part, "functionResponse.name", funcName)
+						part, _ = sjson.Set(part, "functionResponse.response.result", responseData)
+						contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
 					}
-				}
-				contents = append(contents, clientContent)
+					return true
+				})
+				out, _ = sjson.SetRaw(out, "request.contents.-1", contentJSON)
 			} else if contentsResult.Type == gjson.String {
-				prompt := contentsResult.String()
-				contents = append(contents, client.Content{Role: role, Parts: []client.Part{{Text: prompt}}})
+				part := `{"text":""}`
+				part, _ = sjson.Set(part, "text", contentsResult.String())
+				contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
+				out, _ = sjson.SetRaw(out, "request.contents.-1", contentJSON)
 			}
-		}
+			return true
+		})
 	}
 
 	// tools
-	var tools []client.ToolDeclaration
-	toolsResult := gjson.GetBytes(rawJSON, "tools")
-	if toolsResult.IsArray() {
-		tools = make([]client.ToolDeclaration, 1)
-		tools[0].FunctionDeclarations = make([]any, 0)
-		toolsResults := toolsResult.Array()
-		for i := 0; i < len(toolsResults); i++ {
-			toolResult := toolsResults[i]
+	if toolsResult := gjson.GetBytes(rawJSON, "tools"); toolsResult.IsArray() {
+		hasTools := false
+		toolsResult.ForEach(func(_, toolResult gjson.Result) bool {
 			inputSchemaResult := toolResult.Get("input_schema")
 			if inputSchemaResult.Exists() && inputSchemaResult.IsObject() {
 				inputSchema := inputSchemaResult.Raw
@@ -134,30 +142,21 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 				tool, _ = sjson.SetRaw(tool, "parametersJsonSchema", inputSchema)
 				tool, _ = sjson.Delete(tool, "strict")
 				tool, _ = sjson.Delete(tool, "input_examples")
-				var toolDeclaration any
-				if err := json.Unmarshal([]byte(tool), &toolDeclaration); err == nil {
-					tools[0].FunctionDeclarations = append(tools[0].FunctionDeclarations, toolDeclaration)
+				tool, _ = sjson.Delete(tool, "type")
+				tool, _ = sjson.Delete(tool, "cache_control")
+				if gjson.Valid(tool) && gjson.Parse(tool).IsObject() {
+					if !hasTools {
+						out, _ = sjson.SetRaw(out, "request.tools", `[{"functionDeclarations":[]}]`)
+						hasTools = true
+					}
+					out, _ = sjson.SetRaw(out, "request.tools.0.functionDeclarations.-1", tool)
 				}
 			}
+			return true
+		})
+		if !hasTools {
+			out, _ = sjson.Delete(out, "request.tools")
 		}
-	} else {
-		tools = make([]client.ToolDeclaration, 0)
-	}
-
-	// Build output Gemini CLI request JSON
-	out := `{"model":"","request":{"contents":[]}}`
-	out, _ = sjson.Set(out, "model", modelName)
-	if systemInstruction != nil {
-		b, _ := json.Marshal(systemInstruction)
-		out, _ = sjson.SetRaw(out, "request.systemInstruction", string(b))
-	}
-	if len(contents) > 0 {
-		b, _ := json.Marshal(contents)
-		out, _ = sjson.SetRaw(out, "request.contents", string(b))
-	}
-	if len(tools) > 0 && len(tools[0].FunctionDeclarations) > 0 {
-		b, _ := json.Marshal(tools)
-		out, _ = sjson.SetRaw(out, "request.tools", string(b))
 	}
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled

internal/translator/gemini-cli/claude/gemini-cli_claude_response.go

@@ -9,7 +9,6 @@ package claude
 import (
 	"bytes"
 	"context"
-	"encoding/json"
 	"fmt"
 	"strings"
 	"sync/atomic"
@@ -276,22 +275,16 @@ func ConvertGeminiCLIResponseToClaudeNonStream(_ context.Context, _ string, orig
 
 	root := gjson.ParseBytes(rawJSON)
 
-	response := map[string]interface{}{
-		"id":            root.Get("response.responseId").String(),
-		"type":          "message",
-		"role":          "assistant",
-		"model":         root.Get("response.modelVersion").String(),
-		"content":       []interface{}{},
-		"stop_reason":   nil,
-		"stop_sequence": nil,
-		"usage": map[string]interface{}{
-			"input_tokens":  root.Get("response.usageMetadata.promptTokenCount").Int(),
-			"output_tokens": root.Get("response.usageMetadata.candidatesTokenCount").Int() + root.Get("response.usageMetadata.thoughtsTokenCount").Int(),
-		},
-	}
+	out := `{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}`
+	out, _ = sjson.Set(out, "id", root.Get("response.responseId").String())
+	out, _ = sjson.Set(out, "model", root.Get("response.modelVersion").String())
+
+	inputTokens := root.Get("response.usageMetadata.promptTokenCount").Int()
+	outputTokens := root.Get("response.usageMetadata.candidatesTokenCount").Int() + root.Get("response.usageMetadata.thoughtsTokenCount").Int()
+	out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+	out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
 
 	parts := root.Get("response.candidates.0.content.parts")
-	var contentBlocks []interface{}
 	textBuilder := strings.Builder{}
 	thinkingBuilder := strings.Builder{}
 	toolIDCounter := 0
@@ -301,10 +294,9 @@ func ConvertGeminiCLIResponseToClaudeNonStream(_ context.Context, _ string, orig
 		if textBuilder.Len() == 0 {
 			return
 		}
-		contentBlocks = append(contentBlocks, map[string]interface{}{
-			"type": "text",
-			"text": textBuilder.String(),
-		})
+		block := `{"type":"text","text":""}`
+		block, _ = sjson.Set(block, "text", textBuilder.String())
+		out, _ = sjson.SetRaw(out, "content.-1", block)
 		textBuilder.Reset()
 	}
 
@@ -312,10 +304,9 @@ func ConvertGeminiCLIResponseToClaudeNonStream(_ context.Context, _ string, orig
 		if thinkingBuilder.Len() == 0 {
 			return
 		}
-		contentBlocks = append(contentBlocks, map[string]interface{}{
-			"type":     "thinking",
-			"thinking": thinkingBuilder.String(),
-		})
+		block := `{"type":"thinking","thinking":""}`
+		block, _ = sjson.Set(block, "thinking", thinkingBuilder.String())
+		out, _ = sjson.SetRaw(out, "content.-1", block)
 		thinkingBuilder.Reset()
 	}
 
@@ -339,21 +330,15 @@ func ConvertGeminiCLIResponseToClaudeNonStream(_ context.Context, _ string, orig
 
 				name := functionCall.Get("name").String()
 				toolIDCounter++
-				toolBlock := map[string]interface{}{
-					"type":  "tool_use",
-					"id":    fmt.Sprintf("tool_%d", toolIDCounter),
-					"name":  name,
-					"input": map[string]interface{}{},
+				toolBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
+				toolBlock, _ = sjson.Set(toolBlock, "id", fmt.Sprintf("tool_%d", toolIDCounter))
+				toolBlock, _ = sjson.Set(toolBlock, "name", name)
+				inputRaw := "{}"
+				if args := functionCall.Get("args"); args.Exists() && gjson.Valid(args.Raw) && args.IsObject() {
+					inputRaw = args.Raw
 				}
-
-				if args := functionCall.Get("args"); args.Exists() {
-					var parsed interface{}
-					if err := json.Unmarshal([]byte(args.Raw), &parsed); err == nil {
-						toolBlock["input"] = parsed
-					}
-				}
-
-				contentBlocks = append(contentBlocks, toolBlock)
+				toolBlock, _ = sjson.SetRaw(toolBlock, "input", inputRaw)
+				out, _ = sjson.SetRaw(out, "content.-1", toolBlock)
 				continue
 			}
 		}
@@ -362,8 +347,6 @@ func ConvertGeminiCLIResponseToClaudeNonStream(_ context.Context, _ string, orig
 	flushThinking()
 	flushText()
 
-	response["content"] = contentBlocks
-
 	stopReason := "end_turn"
 	if hasToolCall {
 		stopReason = "tool_use"
@@ -379,19 +362,13 @@ func ConvertGeminiCLIResponseToClaudeNonStream(_ context.Context, _ string, orig
 			}
 		}
 	}
-	response["stop_reason"] = stopReason
+	out, _ = sjson.Set(out, "stop_reason", stopReason)
 
-	if usage := response["usage"].(map[string]interface{}); usage["input_tokens"] == int64(0) && usage["output_tokens"] == int64(0) {
-		if usageMeta := root.Get("response.usageMetadata"); !usageMeta.Exists() {
-			delete(response, "usage")
-		}
+	if inputTokens == int64(0) && outputTokens == int64(0) && !root.Get("response.usageMetadata").Exists() {
+		out, _ = sjson.Delete(out, "usage")
 	}
 
-	encoded, err := json.Marshal(response)
-	if err != nil {
-		return ""
-	}
-	return string(encoded)
+	return out
 }
 
 func ClaudeTokenCount(ctx context.Context, count int64) string {

internal/translator/gemini-cli/gemini/gemini-cli_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"
@@ -117,8 +116,6 @@ func ConvertGeminiRequestToGeminiCLI(_ string, inputRawJSON []byte, _ bool) []by
 
 // FunctionCallGroup represents a group of function calls and their responses
 type FunctionCallGroup struct {
-	ModelContent    map[string]interface{}
-	FunctionCalls   []gjson.Result
 	ResponsesNeeded int
 }
 
@@ -146,7 +143,7 @@ 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
 
@@ -178,23 +175,17 @@ 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 {
-						var responseMap map[string]interface{}
-						errUnmarshal := json.Unmarshal([]byte(response.Raw), &responseMap)
-						if errUnmarshal != nil {
-							log.Warnf("failed to unmarshal function response: %v\n", errUnmarshal)
+						if !response.IsObject() {
+							log.Warnf("failed to parse function response")
 							continue
 						}
-						responseParts = append(responseParts, responseMap)
+						functionResponseContent, _ = sjson.SetRaw(functionResponseContent, "parts.-1", response.Raw)
 					}
 
-					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
@@ -208,50 +199,42 @@ 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
@@ -263,31 +246,24 @@ 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 {
-				var responseMap map[string]interface{}
-				errUnmarshal := json.Unmarshal([]byte(response.Raw), &responseMap)
-				if errUnmarshal != nil {
-					log.Warnf("failed to unmarshal function response: %v\n", errUnmarshal)
+				if !response.IsObject() {
+					log.Warnf("failed to parse function response")
 					continue
 				}
-				responseParts = append(responseParts, responseMap)
+				functionResponseContent, _ = sjson.SetRaw(functionResponseContent, "parts.-1", response.Raw)
 			}
 
-			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
 }

internal/translator/gemini-cli/openai/chat-completions/gemini-cli_openai_request.go

@@ -160,6 +160,14 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 				} else if content.IsObject() && content.Get("type").String() == "text" {
 					out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
 					out, _ = sjson.SetBytes(out, "request.systemInstruction.parts.0.text", content.Get("text").String())
+				} else if content.IsArray() {
+					contents := content.Array()
+					if len(contents) > 0 {
+						out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
+						for j := 0; j < len(contents); j++ {
+							out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", j), contents[j].Get("text").String())
+						}
+					}
 				}
 			} else if role == "user" || (role == "system" && len(arr) == 1) {
 				// Build single user content node to avoid splitting into multiple contents
@@ -205,54 +213,77 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 				}
 				out, _ = sjson.SetRawBytes(out, "request.contents.-1", node)
 			} else if role == "assistant" {
+				p := 0
+				node := []byte(`{"role":"model","parts":[]}`)
 				if content.Type == gjson.String {
 					// Assistant text -> single model content
-					node := []byte(`{"role":"model","parts":[{"text":""}]}`)
-					node, _ = sjson.SetBytes(node, "parts.0.text", content.String())
-					out, _ = sjson.SetRawBytes(out, "request.contents.-1", node)
-				} else if !content.Exists() || content.Type == gjson.Null {
-					// Tool calls -> single model content with functionCall parts
-					tcs := m.Get("tool_calls")
-					if tcs.IsArray() {
-						node := []byte(`{"role":"model","parts":[]}`)
-						p := 0
-						fIDs := make([]string, 0)
-						for _, tc := range tcs.Array() {
-							if tc.Get("type").String() != "function" {
-								continue
-							}
-							fid := tc.Get("id").String()
-							fname := tc.Get("function.name").String()
-							fargs := tc.Get("function.arguments").String()
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".functionCall.name", fname)
-							node, _ = sjson.SetRawBytes(node, "parts."+itoa(p)+".functionCall.args", []byte(fargs))
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiCLIFunctionThoughtSignature)
+					node, _ = sjson.SetBytes(node, "parts.-1.text", content.String())
+					p++
+				} else if content.IsArray() {
+					// Assistant multimodal content (e.g. text + image) -> single model content with parts
+					for _, item := range content.Array() {
+						switch item.Get("type").String() {
+						case "text":
+							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", item.Get("text").String())
 							p++
-							if fid != "" {
-								fIDs = append(fIDs, fid)
-							}
-						}
-						out, _ = sjson.SetRawBytes(out, "request.contents.-1", node)
-
-						// Append a single tool content combining name + response per function
-						toolNode := []byte(`{"role":"tool","parts":[]}`)
-						pp := 0
-						for _, fid := range fIDs {
-							if name, ok := tcID2Name[fid]; ok {
-								toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.name", name)
-								resp := toolResponses[fid]
-								if resp == "" {
-									resp = "{}"
+						case "image_url":
+							// If the assistant returned an inline data URL, preserve it for history fidelity.
+							imageURL := item.Get("image_url.url").String()
+							if len(imageURL) > 5 { // expect data:...
+								pieces := strings.SplitN(imageURL[5:], ";", 2)
+								if len(pieces) == 2 && len(pieces[1]) > 7 {
+									mime := pieces[0]
+									data := pieces[1][7:]
+									node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".inlineData.mime_type", mime)
+									node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".inlineData.data", data)
+									p++
 								}
-								toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.response.result", []byte(resp))
-								pp++
 							}
 						}
-						if pp > 0 {
-							out, _ = sjson.SetRawBytes(out, "request.contents.-1", toolNode)
-						}
 					}
 				}
+
+				// Tool calls -> single model content with functionCall parts
+				tcs := m.Get("tool_calls")
+				if tcs.IsArray() {
+					fIDs := make([]string, 0)
+					for _, tc := range tcs.Array() {
+						if tc.Get("type").String() != "function" {
+							continue
+						}
+						fid := tc.Get("id").String()
+						fname := tc.Get("function.name").String()
+						fargs := tc.Get("function.arguments").String()
+						node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".functionCall.name", fname)
+						node, _ = sjson.SetRawBytes(node, "parts."+itoa(p)+".functionCall.args", []byte(fargs))
+						node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiCLIFunctionThoughtSignature)
+						p++
+						if fid != "" {
+							fIDs = append(fIDs, fid)
+						}
+					}
+					out, _ = sjson.SetRawBytes(out, "request.contents.-1", node)
+
+					// Append a single tool content combining name + response per function
+					toolNode := []byte(`{"role":"user","parts":[]}`)
+					pp := 0
+					for _, fid := range fIDs {
+						if name, ok := tcID2Name[fid]; ok {
+							toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.name", name)
+							resp := toolResponses[fid]
+							if resp == "" {
+								resp = "{}"
+							}
+							toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.response.result", []byte(resp))
+							pp++
+						}
+					}
+					if pp > 0 {
+						out, _ = sjson.SetRawBytes(out, "request.contents.-1", toolNode)
+					}
+				} else {
+					out, _ = sjson.SetRawBytes(out, "request.contents.-1", node)
+				}
 			}
 		}
 	}
@@ -278,7 +309,7 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 								log.Warnf("Failed to set default schema type for tool '%s': %v", fn.Get("name").String(), errSet)
 								continue
 							}
-							fnRaw, errSet = sjson.Set(fnRaw, "parametersJsonSchema.properties", map[string]interface{}{})
+							fnRaw, errSet = sjson.SetRaw(fnRaw, "parametersJsonSchema.properties", `{}`)
 							if errSet != nil {
 								log.Warnf("Failed to set default schema properties for tool '%s': %v", fn.Get("name").String(), errSet)
 								continue
@@ -293,7 +324,7 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 							log.Warnf("Failed to set default schema type for tool '%s': %v", fn.Get("name").String(), errSet)
 							continue
 						}
-						fnRaw, errSet = sjson.Set(fnRaw, "parametersJsonSchema.properties", map[string]interface{}{})
+						fnRaw, errSet = sjson.SetRaw(fnRaw, "parametersJsonSchema.properties", `{}`)
 						if errSet != nil {
 							log.Warnf("Failed to set default schema properties for tool '%s': %v", fn.Get("name").String(), errSet)
 							continue
@@ -334,18 +365,3 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 
 // itoa converts int to string without strconv import for few usages.
 func itoa(i int) string { return fmt.Sprintf("%d", i) }
-
-// quoteIfNeeded ensures a string is valid JSON value (quotes plain text), pass-through for JSON objects/arrays.
-func quoteIfNeeded(s string) string {
-	s = strings.TrimSpace(s)
-	if s == "" {
-		return "\"\""
-	}
-	if len(s) > 0 && (s[0] == '{' || s[0] == '[') {
-		return s
-	}
-	// escape quotes minimally
-	s = strings.ReplaceAll(s, "\\", "\\\\")
-	s = strings.ReplaceAll(s, "\"", "\\\"")
-	return "\"" + s + "\""
-}

internal/translator/gemini-cli/openai/chat-completions/gemini-cli_openai_response.go

@@ -8,7 +8,6 @@ package chat_completions
 import (
 	"bytes"
 	"context"
-	"encoding/json"
 	"fmt"
 	"strings"
 	"sync/atomic"
@@ -171,21 +170,16 @@ func ConvertCliResponseToOpenAI(_ context.Context, _ string, originalRequestRawJ
 					mimeType = "image/png"
 				}
 				imageURL := fmt.Sprintf("data:%s;base64,%s", mimeType, data)
-				imagePayload, err := json.Marshal(map[string]any{
-					"type": "image_url",
-					"image_url": map[string]string{
-						"url": imageURL,
-					},
-				})
-				if err != nil {
-					continue
-				}
 				imagesResult := gjson.Get(template, "choices.0.delta.images")
 				if !imagesResult.Exists() || !imagesResult.IsArray() {
 					template, _ = sjson.SetRaw(template, "choices.0.delta.images", `[]`)
 				}
+				imageIndex := len(gjson.Get(template, "choices.0.delta.images").Array())
+				imagePayload := `{"type":"image_url","image_url":{"url":""}}`
+				imagePayload, _ = sjson.Set(imagePayload, "index", imageIndex)
+				imagePayload, _ = sjson.Set(imagePayload, "image_url.url", imageURL)
 				template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
-				template, _ = sjson.SetRaw(template, "choices.0.delta.images.-1", string(imagePayload))
+				template, _ = sjson.SetRaw(template, "choices.0.delta.images.-1", imagePayload)
 			}
 		}
 	}

internal/translator/gemini/claude/gemini_claude_request.go

@@ -7,10 +7,8 @@ package claude
 
 import (
 	"bytes"
-	"encoding/json"
 	"strings"
 
-	client "github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
@@ -34,92 +32,102 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 	rawJSON := bytes.Clone(inputRawJSON)
 	rawJSON = bytes.Replace(rawJSON, []byte(`"url":{"type":"string","format":"uri",`), []byte(`"url":{"type":"string",`), -1)
 
+	// Build output Gemini CLI request JSON
+	out := `{"contents":[]}`
+	out, _ = sjson.Set(out, "model", modelName)
+
 	// system instruction
-	var systemInstruction *client.Content
-	systemResult := gjson.GetBytes(rawJSON, "system")
-	if systemResult.IsArray() {
-		systemResults := systemResult.Array()
-		systemInstruction = &client.Content{Role: "user", Parts: []client.Part{}}
-		for i := 0; i < len(systemResults); i++ {
-			systemPromptResult := systemResults[i]
-			systemTypePromptResult := systemPromptResult.Get("type")
-			if systemTypePromptResult.Type == gjson.String && systemTypePromptResult.String() == "text" {
-				systemPrompt := systemPromptResult.Get("text").String()
-				systemPart := client.Part{Text: systemPrompt}
-				systemInstruction.Parts = append(systemInstruction.Parts, systemPart)
+	if systemResult := gjson.GetBytes(rawJSON, "system"); systemResult.IsArray() {
+		systemInstruction := `{"role":"user","parts":[]}`
+		hasSystemParts := false
+		systemResult.ForEach(func(_, systemPromptResult gjson.Result) bool {
+			if systemPromptResult.Get("type").String() == "text" {
+				textResult := systemPromptResult.Get("text")
+				if textResult.Type == gjson.String {
+					part := `{"text":""}`
+					part, _ = sjson.Set(part, "text", textResult.String())
+					systemInstruction, _ = sjson.SetRaw(systemInstruction, "parts.-1", part)
+					hasSystemParts = true
+				}
 			}
+			return true
+		})
+		if hasSystemParts {
+			out, _ = sjson.SetRaw(out, "system_instruction", systemInstruction)
 		}
-		if len(systemInstruction.Parts) == 0 {
-			systemInstruction = nil
-		}
+	} else if systemResult.Type == gjson.String {
+		out, _ = sjson.Set(out, "request.system_instruction.parts.-1.text", systemResult.String())
 	}
 
 	// contents
-	contents := make([]client.Content, 0)
-	messagesResult := gjson.GetBytes(rawJSON, "messages")
-	if messagesResult.IsArray() {
-		messageResults := messagesResult.Array()
-		for i := 0; i < len(messageResults); i++ {
-			messageResult := messageResults[i]
+	if messagesResult := gjson.GetBytes(rawJSON, "messages"); messagesResult.IsArray() {
+		messagesResult.ForEach(func(_, messageResult gjson.Result) bool {
 			roleResult := messageResult.Get("role")
 			if roleResult.Type != gjson.String {
-				continue
+				return true
 			}
 			role := roleResult.String()
 			if role == "assistant" {
 				role = "model"
 			}
-			clientContent := client.Content{Role: role, Parts: []client.Part{}}
+
+			contentJSON := `{"role":"","parts":[]}`
+			contentJSON, _ = sjson.Set(contentJSON, "role", role)
+
 			contentsResult := messageResult.Get("content")
 			if contentsResult.IsArray() {
-				contentResults := contentsResult.Array()
-				for j := 0; j < len(contentResults); j++ {
-					contentResult := contentResults[j]
-					contentTypeResult := contentResult.Get("type")
-					if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "text" {
-						prompt := contentResult.Get("text").String()
-						clientContent.Parts = append(clientContent.Parts, client.Part{Text: prompt})
-					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "tool_use" {
+				contentsResult.ForEach(func(_, contentResult gjson.Result) bool {
+					switch contentResult.Get("type").String() {
+					case "text":
+						part := `{"text":""}`
+						part, _ = sjson.Set(part, "text", contentResult.Get("text").String())
+						contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
+
+					case "tool_use":
 						functionName := contentResult.Get("name").String()
 						functionArgs := contentResult.Get("input").String()
-						var args map[string]any
-						if err := json.Unmarshal([]byte(functionArgs), &args); err == nil {
-							clientContent.Parts = append(clientContent.Parts, client.Part{
-								FunctionCall:     &client.FunctionCall{Name: functionName, Args: args},
-								ThoughtSignature: geminiClaudeThoughtSignature,
-							})
+						argsResult := gjson.Parse(functionArgs)
+						if argsResult.IsObject() && gjson.Valid(functionArgs) {
+							part := `{"thoughtSignature":"","functionCall":{"name":"","args":{}}}`
+							part, _ = sjson.Set(part, "thoughtSignature", geminiClaudeThoughtSignature)
+							part, _ = sjson.Set(part, "functionCall.name", functionName)
+							part, _ = sjson.SetRaw(part, "functionCall.args", functionArgs)
+							contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
 						}
-					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "tool_result" {
+
+					case "tool_result":
 						toolCallID := contentResult.Get("tool_use_id").String()
-						if toolCallID != "" {
-							funcName := toolCallID
-							toolCallIDs := strings.Split(toolCallID, "-")
-							if len(toolCallIDs) > 1 {
-								funcName = strings.Join(toolCallIDs[0:len(toolCallIDs)-1], "-")
-							}
-							responseData := contentResult.Get("content").Raw
-							functionResponse := client.FunctionResponse{Name: funcName, Response: map[string]interface{}{"result": responseData}}
-							clientContent.Parts = append(clientContent.Parts, client.Part{FunctionResponse: &functionResponse})
+						if toolCallID == "" {
+							return true
 						}
+						funcName := toolCallID
+						toolCallIDs := strings.Split(toolCallID, "-")
+						if len(toolCallIDs) > 1 {
+							funcName = strings.Join(toolCallIDs[0:len(toolCallIDs)-1], "-")
+						}
+						responseData := contentResult.Get("content").Raw
+						part := `{"functionResponse":{"name":"","response":{"result":""}}}`
+						part, _ = sjson.Set(part, "functionResponse.name", funcName)
+						part, _ = sjson.Set(part, "functionResponse.response.result", responseData)
+						contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
 					}
-				}
-				contents = append(contents, clientContent)
+					return true
+				})
+				out, _ = sjson.SetRaw(out, "contents.-1", contentJSON)
 			} else if contentsResult.Type == gjson.String {
-				prompt := contentsResult.String()
-				contents = append(contents, client.Content{Role: role, Parts: []client.Part{{Text: prompt}}})
+				part := `{"text":""}`
+				part, _ = sjson.Set(part, "text", contentsResult.String())
+				contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
+				out, _ = sjson.SetRaw(out, "contents.-1", contentJSON)
 			}
-		}
+			return true
+		})
 	}
 
 	// tools
-	var tools []client.ToolDeclaration
-	toolsResult := gjson.GetBytes(rawJSON, "tools")
-	if toolsResult.IsArray() {
-		tools = make([]client.ToolDeclaration, 1)
-		tools[0].FunctionDeclarations = make([]any, 0)
-		toolsResults := toolsResult.Array()
-		for i := 0; i < len(toolsResults); i++ {
-			toolResult := toolsResults[i]
+	if toolsResult := gjson.GetBytes(rawJSON, "tools"); toolsResult.IsArray() {
+		hasTools := false
+		toolsResult.ForEach(func(_, toolResult gjson.Result) bool {
 			inputSchemaResult := toolResult.Get("input_schema")
 			if inputSchemaResult.Exists() && inputSchemaResult.IsObject() {
 				inputSchema := inputSchemaResult.Raw
@@ -127,30 +135,21 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 				tool, _ = sjson.SetRaw(tool, "parametersJsonSchema", inputSchema)
 				tool, _ = sjson.Delete(tool, "strict")
 				tool, _ = sjson.Delete(tool, "input_examples")
-				var toolDeclaration any
-				if err := json.Unmarshal([]byte(tool), &toolDeclaration); err == nil {
-					tools[0].FunctionDeclarations = append(tools[0].FunctionDeclarations, toolDeclaration)
+				tool, _ = sjson.Delete(tool, "type")
+				tool, _ = sjson.Delete(tool, "cache_control")
+				if gjson.Valid(tool) && gjson.Parse(tool).IsObject() {
+					if !hasTools {
+						out, _ = sjson.SetRaw(out, "tools", `[{"functionDeclarations":[]}]`)
+						hasTools = true
+					}
+					out, _ = sjson.SetRaw(out, "tools.0.functionDeclarations.-1", tool)
 				}
 			}
+			return true
+		})
+		if !hasTools {
+			out, _ = sjson.Delete(out, "tools")
 		}
-	} else {
-		tools = make([]client.ToolDeclaration, 0)
-	}
-
-	// Build output Gemini CLI request JSON
-	out := `{"contents":[]}`
-	out, _ = sjson.Set(out, "model", modelName)
-	if systemInstruction != nil {
-		b, _ := json.Marshal(systemInstruction)
-		out, _ = sjson.SetRaw(out, "system_instruction", string(b))
-	}
-	if len(contents) > 0 {
-		b, _ := json.Marshal(contents)
-		out, _ = sjson.SetRaw(out, "contents", string(b))
-	}
-	if len(tools) > 0 && len(tools[0].FunctionDeclarations) > 0 {
-		b, _ := json.Marshal(tools)
-		out, _ = sjson.SetRaw(out, "tools", string(b))
 	}
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when enabled

internal/translator/gemini/claude/gemini_claude_response.go

@@ -9,7 +9,6 @@ package claude
 import (
 	"bytes"
 	"context"
-	"encoding/json"
 	"fmt"
 	"strings"
 	"sync/atomic"
@@ -179,6 +178,18 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 				usedTool = true
 				fcName := functionCallResult.Get("name").String()
 
+				// FIX: Handle streaming split/delta where name might be empty in subsequent chunks.
+				// If we are already in tool use mode and name is empty, treat as continuation (delta).
+				if (*param).(*Params).ResponseType == 3 && fcName == "" {
+					if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
+						output = output + "event: content_block_delta\n"
+						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"input_json_delta","partial_json":""}}`, (*param).(*Params).ResponseIndex), "delta.partial_json", fcArgsResult.Raw)
+						output = output + fmt.Sprintf("data: %s\n\n\n", data)
+					}
+					// Continue to next part without closing/opening logic
+					continue
+				}
+
 				// Handle state transitions when switching to function calls
 				// Close any existing function call block first
 				if (*param).(*Params).ResponseType == 3 {
@@ -270,22 +281,16 @@ func ConvertGeminiResponseToClaudeNonStream(_ context.Context, _ string, origina
 
 	root := gjson.ParseBytes(rawJSON)
 
-	response := map[string]interface{}{
-		"id":            root.Get("responseId").String(),
-		"type":          "message",
-		"role":          "assistant",
-		"model":         root.Get("modelVersion").String(),
-		"content":       []interface{}{},
-		"stop_reason":   nil,
-		"stop_sequence": nil,
-		"usage": map[string]interface{}{
-			"input_tokens":  root.Get("usageMetadata.promptTokenCount").Int(),
-			"output_tokens": root.Get("usageMetadata.candidatesTokenCount").Int() + root.Get("usageMetadata.thoughtsTokenCount").Int(),
-		},
-	}
+	out := `{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}`
+	out, _ = sjson.Set(out, "id", root.Get("responseId").String())
+	out, _ = sjson.Set(out, "model", root.Get("modelVersion").String())
+
+	inputTokens := root.Get("usageMetadata.promptTokenCount").Int()
+	outputTokens := root.Get("usageMetadata.candidatesTokenCount").Int() + root.Get("usageMetadata.thoughtsTokenCount").Int()
+	out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+	out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
 
 	parts := root.Get("candidates.0.content.parts")
-	var contentBlocks []interface{}
 	textBuilder := strings.Builder{}
 	thinkingBuilder := strings.Builder{}
 	toolIDCounter := 0
@@ -295,10 +300,9 @@ func ConvertGeminiResponseToClaudeNonStream(_ context.Context, _ string, origina
 		if textBuilder.Len() == 0 {
 			return
 		}
-		contentBlocks = append(contentBlocks, map[string]interface{}{
-			"type": "text",
-			"text": textBuilder.String(),
-		})
+		block := `{"type":"text","text":""}`
+		block, _ = sjson.Set(block, "text", textBuilder.String())
+		out, _ = sjson.SetRaw(out, "content.-1", block)
 		textBuilder.Reset()
 	}
 
@@ -306,10 +310,9 @@ func ConvertGeminiResponseToClaudeNonStream(_ context.Context, _ string, origina
 		if thinkingBuilder.Len() == 0 {
 			return
 		}
-		contentBlocks = append(contentBlocks, map[string]interface{}{
-			"type":     "thinking",
-			"thinking": thinkingBuilder.String(),
-		})
+		block := `{"type":"thinking","thinking":""}`
+		block, _ = sjson.Set(block, "thinking", thinkingBuilder.String())
+		out, _ = sjson.SetRaw(out, "content.-1", block)
 		thinkingBuilder.Reset()
 	}
 
@@ -333,21 +336,15 @@ func ConvertGeminiResponseToClaudeNonStream(_ context.Context, _ string, origina
 
 				name := functionCall.Get("name").String()
 				toolIDCounter++
-				toolBlock := map[string]interface{}{
-					"type":  "tool_use",
-					"id":    fmt.Sprintf("tool_%d", toolIDCounter),
-					"name":  name,
-					"input": map[string]interface{}{},
+				toolBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
+				toolBlock, _ = sjson.Set(toolBlock, "id", fmt.Sprintf("tool_%d", toolIDCounter))
+				toolBlock, _ = sjson.Set(toolBlock, "name", name)
+				inputRaw := "{}"
+				if args := functionCall.Get("args"); args.Exists() && gjson.Valid(args.Raw) && args.IsObject() {
+					inputRaw = args.Raw
 				}
-
-				if args := functionCall.Get("args"); args.Exists() {
-					var parsed interface{}
-					if err := json.Unmarshal([]byte(args.Raw), &parsed); err == nil {
-						toolBlock["input"] = parsed
-					}
-				}
-
-				contentBlocks = append(contentBlocks, toolBlock)
+				toolBlock, _ = sjson.SetRaw(toolBlock, "input", inputRaw)
+				out, _ = sjson.SetRaw(out, "content.-1", toolBlock)
 				continue
 			}
 		}
@@ -356,8 +353,6 @@ func ConvertGeminiResponseToClaudeNonStream(_ context.Context, _ string, origina
 	flushThinking()
 	flushText()
 
-	response["content"] = contentBlocks
-
 	stopReason := "end_turn"
 	if hasToolCall {
 		stopReason = "tool_use"
@@ -373,19 +368,13 @@ func ConvertGeminiResponseToClaudeNonStream(_ context.Context, _ string, origina
 			}
 		}
 	}
-	response["stop_reason"] = stopReason
+	out, _ = sjson.Set(out, "stop_reason", stopReason)
 
-	if usage := response["usage"].(map[string]interface{}); usage["input_tokens"] == int64(0) && usage["output_tokens"] == int64(0) {
-		if usageMeta := root.Get("usageMetadata"); !usageMeta.Exists() {
-			delete(response, "usage")
-		}
+	if inputTokens == int64(0) && outputTokens == int64(0) && !root.Get("usageMetadata").Exists() {
+		out, _ = sjson.Delete(out, "usage")
 	}
 
-	encoded, err := json.Marshal(response)
-	if err != nil {
-		return ""
-	}
-	return string(encoded)
+	return out
 }
 
 func ClaudeTokenCount(ctx context.Context, count int64) string {

internal/translator/gemini/openai/chat-completions/gemini_openai_request.go

@@ -37,12 +37,28 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 
 	// Reasoning effort -> thinkingBudget/include_thoughts
 	// Note: OpenAI official fields take precedence over extra_body.google.thinking_config
-	// Only convert for models that use numeric budgets (not discrete levels) to avoid
-	// incorrectly applying thinkingBudget for level-based models like gpt-5.
+	// Only apply numeric budgets for models that use budgets (not discrete levels) to avoid
+	// incorrectly applying thinkingBudget for level-based models like gpt-5. Gemini 3 models
+	// use thinkingLevel/includeThoughts instead.
 	re := gjson.GetBytes(rawJSON, "reasoning_effort")
 	hasOfficialThinking := re.Exists()
-	if hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
-		out = util.ApplyReasoningEffortToGemini(out, re.String())
+	if hasOfficialThinking && util.ModelSupportsThinking(modelName) {
+		effort := strings.ToLower(strings.TrimSpace(re.String()))
+		if util.IsGemini3Model(modelName) {
+			switch effort {
+			case "none":
+				out, _ = sjson.DeleteBytes(out, "generationConfig.thinkingConfig")
+			case "auto":
+				includeThoughts := true
+				out = util.ApplyGeminiThinkingLevel(out, "", &includeThoughts)
+			default:
+				if level, ok := util.ValidateGemini3ThinkingLevel(modelName, effort); ok {
+					out = util.ApplyGeminiThinkingLevel(out, level, nil)
+				}
+			}
+		} else if !util.ModelUsesThinkingLevels(modelName) {
+			out = util.ApplyReasoningEffortToGemini(out, effort)
+		}
 	}
 
 	// Cherry Studio extension extra_body.google.thinking_config (effective only when official fields are absent)
@@ -162,6 +178,14 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 				} else if content.IsObject() && content.Get("type").String() == "text" {
 					out, _ = sjson.SetBytes(out, "system_instruction.role", "user")
 					out, _ = sjson.SetBytes(out, "system_instruction.parts.0.text", content.Get("text").String())
+				} else if content.IsArray() {
+					contents := content.Array()
+					if len(contents) > 0 {
+						out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
+						for j := 0; j < len(contents); j++ {
+							out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", j), contents[j].Get("text").String())
+						}
+					}
 				}
 			} else if role == "user" || (role == "system" && len(arr) == 1) {
 				// Build single user content node to avoid splitting into multiple contents
@@ -207,19 +231,17 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 				}
 				out, _ = sjson.SetRawBytes(out, "contents.-1", node)
 			} else if role == "assistant" {
+				node := []byte(`{"role":"model","parts":[]}`)
+				p := 0
 				if content.Type == gjson.String {
 					// Assistant text -> single model content
-					node := []byte(`{"role":"model","parts":[{"text":""}]}`)
-					node, _ = sjson.SetBytes(node, "parts.0.text", content.String())
-					out, _ = sjson.SetRawBytes(out, "contents.-1", node)
+					node, _ = sjson.SetBytes(node, "parts.-1.text", content.String())
+					p++
 				} else if content.IsArray() {
 					// Assistant multimodal content (e.g. text + image) -> single model content with parts
-					node := []byte(`{"role":"model","parts":[]}`)
-					p := 0
 					for _, item := range content.Array() {
 						switch item.Get("type").String() {
 						case "text":
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", item.Get("text").String())
 							p++
 						case "image_url":
 							// If the assistant returned an inline data URL, preserve it for history fidelity.
@@ -236,49 +258,48 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 							}
 						}
 					}
-					out, _ = sjson.SetRawBytes(out, "contents.-1", node)
-				} else if !content.Exists() || content.Type == gjson.Null {
-					// Tool calls -> single model content with functionCall parts
-					tcs := m.Get("tool_calls")
-					if tcs.IsArray() {
-						node := []byte(`{"role":"model","parts":[]}`)
-						p := 0
-						fIDs := make([]string, 0)
-						for _, tc := range tcs.Array() {
-							if tc.Get("type").String() != "function" {
-								continue
-							}
-							fid := tc.Get("id").String()
-							fname := tc.Get("function.name").String()
-							fargs := tc.Get("function.arguments").String()
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".functionCall.name", fname)
-							node, _ = sjson.SetRawBytes(node, "parts."+itoa(p)+".functionCall.args", []byte(fargs))
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiFunctionThoughtSignature)
-							p++
-							if fid != "" {
-								fIDs = append(fIDs, fid)
-							}
-						}
-						out, _ = sjson.SetRawBytes(out, "contents.-1", node)
+				}
 
-						// Append a single tool content combining name + response per function
-						toolNode := []byte(`{"role":"tool","parts":[]}`)
-						pp := 0
-						for _, fid := range fIDs {
-							if name, ok := tcID2Name[fid]; ok {
-								toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.name", name)
-								resp := toolResponses[fid]
-								if resp == "" {
-									resp = "{}"
-								}
-								toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.response.result", []byte(resp))
-								pp++
-							}
+				// Tool calls -> single model content with functionCall parts
+				tcs := m.Get("tool_calls")
+				if tcs.IsArray() {
+					fIDs := make([]string, 0)
+					for _, tc := range tcs.Array() {
+						if tc.Get("type").String() != "function" {
+							continue
 						}
-						if pp > 0 {
-							out, _ = sjson.SetRawBytes(out, "contents.-1", toolNode)
+						fid := tc.Get("id").String()
+						fname := tc.Get("function.name").String()
+						fargs := tc.Get("function.arguments").String()
+						node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".functionCall.name", fname)
+						node, _ = sjson.SetRawBytes(node, "parts."+itoa(p)+".functionCall.args", []byte(fargs))
+						node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiFunctionThoughtSignature)
+						p++
+						if fid != "" {
+							fIDs = append(fIDs, fid)
 						}
 					}
+					out, _ = sjson.SetRawBytes(out, "contents.-1", node)
+
+					// Append a single tool content combining name + response per function
+					toolNode := []byte(`{"role":"user","parts":[]}`)
+					pp := 0
+					for _, fid := range fIDs {
+						if name, ok := tcID2Name[fid]; ok {
+							toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.name", name)
+							resp := toolResponses[fid]
+							if resp == "" {
+								resp = "{}"
+							}
+							toolNode, _ = sjson.SetBytes(toolNode, "parts."+itoa(pp)+".functionResponse.response.result", []byte(resp))
+							pp++
+						}
+					}
+					if pp > 0 {
+						out, _ = sjson.SetRawBytes(out, "contents.-1", toolNode)
+					}
+				} else {
+					out, _ = sjson.SetRawBytes(out, "contents.-1", node)
 				}
 			}
 		}
@@ -305,7 +326,7 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 								log.Warnf("Failed to set default schema type for tool '%s': %v", fn.Get("name").String(), errSet)
 								continue
 							}
-							fnRaw, errSet = sjson.Set(fnRaw, "parametersJsonSchema.properties", map[string]interface{}{})
+							fnRaw, errSet = sjson.SetRaw(fnRaw, "parametersJsonSchema.properties", `{}`)
 							if errSet != nil {
 								log.Warnf("Failed to set default schema properties for tool '%s': %v", fn.Get("name").String(), errSet)
 								continue
@@ -320,7 +341,7 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 							log.Warnf("Failed to set default schema type for tool '%s': %v", fn.Get("name").String(), errSet)
 							continue
 						}
-						fnRaw, errSet = sjson.Set(fnRaw, "parametersJsonSchema.properties", map[string]interface{}{})
+						fnRaw, errSet = sjson.SetRaw(fnRaw, "parametersJsonSchema.properties", `{}`)
 						if errSet != nil {
 							log.Warnf("Failed to set default schema properties for tool '%s': %v", fn.Get("name").String(), errSet)
 							continue
@@ -363,18 +384,3 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 
 // itoa converts int to string without strconv import for few usages.
 func itoa(i int) string { return fmt.Sprintf("%d", i) }
-
-// quoteIfNeeded ensures a string is valid JSON value (quotes plain text), pass-through for JSON objects/arrays.
-func quoteIfNeeded(s string) string {
-	s = strings.TrimSpace(s)
-	if s == "" {
-		return "\"\""
-	}
-	if len(s) > 0 && (s[0] == '{' || s[0] == '[') {
-		return s
-	}
-	// escape quotes minimally
-	s = strings.ReplaceAll(s, "\\", "\\\\")
-	s = strings.ReplaceAll(s, "\"", "\\\"")
-	return "\"" + s + "\""
-}

internal/translator/gemini/openai/chat-completions/gemini_openai_response.go

@@ -8,12 +8,12 @@ package chat_completions
 import (
 	"bytes"
 	"context"
-	"encoding/json"
 	"fmt"
 	"strings"
 	"sync/atomic"
 	"time"
 
+	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -89,18 +89,27 @@ func ConvertGeminiResponseToOpenAI(_ context.Context, _ string, originalRequestR
 
 	// Extract and set usage metadata (token counts).
 	if usageResult := gjson.GetBytes(rawJSON, "usageMetadata"); usageResult.Exists() {
+		cachedTokenCount := usageResult.Get("cachedContentTokenCount").Int()
 		if candidatesTokenCountResult := usageResult.Get("candidatesTokenCount"); candidatesTokenCountResult.Exists() {
 			template, _ = sjson.Set(template, "usage.completion_tokens", candidatesTokenCountResult.Int())
 		}
 		if totalTokenCountResult := usageResult.Get("totalTokenCount"); totalTokenCountResult.Exists() {
 			template, _ = sjson.Set(template, "usage.total_tokens", totalTokenCountResult.Int())
 		}
-		promptTokenCount := usageResult.Get("promptTokenCount").Int()
+		promptTokenCount := usageResult.Get("promptTokenCount").Int() - cachedTokenCount
 		thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
 		template, _ = sjson.Set(template, "usage.prompt_tokens", promptTokenCount+thoughtsTokenCount)
 		if thoughtsTokenCount > 0 {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", thoughtsTokenCount)
 		}
+		// Include cached token count if present (indicates prompt caching is working)
+		if cachedTokenCount > 0 {
+			var err error
+			template, err = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cachedTokenCount)
+			if err != nil {
+				log.Warnf("gemini openai response: failed to set cached_tokens in streaming: %v", err)
+			}
+		}
 	}
 
 	// Process the main content part of the response.
@@ -173,21 +182,16 @@ func ConvertGeminiResponseToOpenAI(_ context.Context, _ string, originalRequestR
 					mimeType = "image/png"
 				}
 				imageURL := fmt.Sprintf("data:%s;base64,%s", mimeType, data)
-				imagePayload, err := json.Marshal(map[string]any{
-					"type": "image_url",
-					"image_url": map[string]string{
-						"url": imageURL,
-					},
-				})
-				if err != nil {
-					continue
-				}
 				imagesResult := gjson.Get(template, "choices.0.delta.images")
 				if !imagesResult.Exists() || !imagesResult.IsArray() {
 					template, _ = sjson.SetRaw(template, "choices.0.delta.images", `[]`)
 				}
+				imageIndex := len(gjson.Get(template, "choices.0.delta.images").Array())
+				imagePayload := `{"type":"image_url","image_url":{"url":""}}`
+				imagePayload, _ = sjson.Set(imagePayload, "index", imageIndex)
+				imagePayload, _ = sjson.Set(imagePayload, "image_url.url", imageURL)
 				template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
-				template, _ = sjson.SetRaw(template, "choices.0.delta.images.-1", string(imagePayload))
+				template, _ = sjson.SetRaw(template, "choices.0.delta.images.-1", imagePayload)
 			}
 		}
 	}
@@ -248,10 +252,19 @@ func ConvertGeminiResponseToOpenAINonStream(_ context.Context, _ string, origina
 		}
 		promptTokenCount := usageResult.Get("promptTokenCount").Int()
 		thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
+		cachedTokenCount := usageResult.Get("cachedContentTokenCount").Int()
 		template, _ = sjson.Set(template, "usage.prompt_tokens", promptTokenCount+thoughtsTokenCount)
 		if thoughtsTokenCount > 0 {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", thoughtsTokenCount)
 		}
+		// Include cached token count if present (indicates prompt caching is working)
+		if cachedTokenCount > 0 {
+			var err error
+			template, err = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cachedTokenCount)
+			if err != nil {
+				log.Warnf("gemini openai response: failed to set cached_tokens in non-streaming: %v", err)
+			}
+		}
 	}
 
 	// Process the main content part of the response.
@@ -305,21 +318,16 @@ func ConvertGeminiResponseToOpenAINonStream(_ context.Context, _ string, origina
 					mimeType = "image/png"
 				}
 				imageURL := fmt.Sprintf("data:%s;base64,%s", mimeType, data)
-				imagePayload, err := json.Marshal(map[string]any{
-					"type": "image_url",
-					"image_url": map[string]string{
-						"url": imageURL,
-					},
-				})
-				if err != nil {
-					continue
-				}
 				imagesResult := gjson.Get(template, "choices.0.message.images")
 				if !imagesResult.Exists() || !imagesResult.IsArray() {
 					template, _ = sjson.SetRaw(template, "choices.0.message.images", `[]`)
 				}
+				imageIndex := len(gjson.Get(template, "choices.0.message.images").Array())
+				imagePayload := `{"type":"image_url","image_url":{"url":""}}`
+				imagePayload, _ = sjson.Set(imagePayload, "index", imageIndex)
+				imagePayload, _ = sjson.Set(imagePayload, "image_url.url", imageURL)
 				template, _ = sjson.Set(template, "choices.0.message.role", "assistant")
-				template, _ = sjson.SetRaw(template, "choices.0.message.images.-1", string(imagePayload))
+				template, _ = sjson.SetRaw(template, "choices.0.message.images.-1", imagePayload)
 			}
 		}
 	}

internal/translator/gemini/openai/responses/gemini_openai-responses_response.go

@@ -117,7 +117,7 @@ func ConvertGeminiResponseToOpenAIResponses(_ context.Context, modelName string,
 			st.CreatedAt = time.Now().Unix()
 		}
 
-		created := `{"type":"response.created","sequence_number":0,"response":{"id":"","object":"response","created_at":0,"status":"in_progress","background":false,"error":null}}`
+		created := `{"type":"response.created","sequence_number":0,"response":{"id":"","object":"response","created_at":0,"status":"in_progress","background":false,"error":null,"output":[]}}`
 		created, _ = sjson.Set(created, "sequence_number", nextSeq())
 		created, _ = sjson.Set(created, "response.id", st.ResponseID)
 		created, _ = sjson.Set(created, "response.created_at", st.CreatedAt)
@@ -160,11 +160,11 @@ func ConvertGeminiResponseToOpenAIResponses(_ context.Context, modelName string,
 				}
 				if t := part.Get("text"); t.Exists() && t.String() != "" {
 					st.ReasoningBuf.WriteString(t.String())
-					msg := `{"type":"response.reasoning_summary_text.delta","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"text":""}`
+					msg := `{"type":"response.reasoning_summary_text.delta","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"delta":""}`
 					msg, _ = sjson.Set(msg, "sequence_number", nextSeq())
 					msg, _ = sjson.Set(msg, "item_id", st.ReasoningItemID)
 					msg, _ = sjson.Set(msg, "output_index", st.ReasoningIndex)
-					msg, _ = sjson.Set(msg, "text", t.String())
+					msg, _ = sjson.Set(msg, "delta", t.String())
 					out = append(out, emitEvent("response.reasoning_summary_text.delta", msg))
 				}
 				return true
@@ -377,27 +377,18 @@ func ConvertGeminiResponseToOpenAIResponses(_ context.Context, modelName string,
 		}
 
 		// Compose outputs in encountered order: reasoning, message, function_calls
-		var outputs []interface{}
+		outputsWrapper := `{"arr":[]}`
 		if st.ReasoningOpened {
-			outputs = append(outputs, map[string]interface{}{
-				"id":      st.ReasoningItemID,
-				"type":    "reasoning",
-				"summary": []interface{}{map[string]interface{}{"type": "summary_text", "text": st.ReasoningBuf.String()}},
-			})
+			item := `{"id":"","type":"reasoning","summary":[{"type":"summary_text","text":""}]}`
+			item, _ = sjson.Set(item, "id", st.ReasoningItemID)
+			item, _ = sjson.Set(item, "summary.0.text", st.ReasoningBuf.String())
+			outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 		}
 		if st.MsgOpened {
-			outputs = append(outputs, map[string]interface{}{
-				"id":     st.CurrentMsgID,
-				"type":   "message",
-				"status": "completed",
-				"content": []interface{}{map[string]interface{}{
-					"type":        "output_text",
-					"annotations": []interface{}{},
-					"logprobs":    []interface{}{},
-					"text":        st.TextBuf.String(),
-				}},
-				"role": "assistant",
-			})
+			item := `{"id":"","type":"message","status":"completed","content":[{"type":"output_text","annotations":[],"logprobs":[],"text":""}],"role":"assistant"}`
+			item, _ = sjson.Set(item, "id", st.CurrentMsgID)
+			item, _ = sjson.Set(item, "content.0.text", st.TextBuf.String())
+			outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 		}
 		if len(st.FuncArgsBuf) > 0 {
 			idxs := make([]int, 0, len(st.FuncArgsBuf))
@@ -416,18 +407,16 @@ func ConvertGeminiResponseToOpenAIResponses(_ context.Context, modelName string,
 				if b := st.FuncArgsBuf[idx]; b != nil {
 					args = b.String()
 				}
-				outputs = append(outputs, map[string]interface{}{
-					"id":        fmt.Sprintf("fc_%s", st.FuncCallIDs[idx]),
-					"type":      "function_call",
-					"status":    "completed",
-					"arguments": args,
-					"call_id":   st.FuncCallIDs[idx],
-					"name":      st.FuncNames[idx],
-				})
+				item := `{"id":"","type":"function_call","status":"completed","arguments":"","call_id":"","name":""}`
+				item, _ = sjson.Set(item, "id", fmt.Sprintf("fc_%s", st.FuncCallIDs[idx]))
+				item, _ = sjson.Set(item, "arguments", args)
+				item, _ = sjson.Set(item, "call_id", st.FuncCallIDs[idx])
+				item, _ = sjson.Set(item, "name", st.FuncNames[idx])
+				outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 			}
 		}
-		if len(outputs) > 0 {
-			completed, _ = sjson.Set(completed, "response.output", outputs)
+		if gjson.Get(outputsWrapper, "arr.#").Int() > 0 {
+			completed, _ = sjson.SetRaw(completed, "response.output", gjson.Get(outputsWrapper, "arr").Raw)
 		}
 
 		// usage mapping
@@ -558,11 +547,24 @@ func ConvertGeminiResponseToOpenAIResponsesNonStream(_ context.Context, _ string
 	}
 
 	// Build outputs from candidates[0].content.parts
-	var outputs []interface{}
 	var reasoningText strings.Builder
 	var reasoningEncrypted string
 	var messageText strings.Builder
 	var haveMessage bool
+
+	haveOutput := false
+	ensureOutput := func() {
+		if haveOutput {
+			return
+		}
+		resp, _ = sjson.SetRaw(resp, "output", "[]")
+		haveOutput = true
+	}
+	appendOutput := func(itemJSON string) {
+		ensureOutput()
+		resp, _ = sjson.SetRaw(resp, "output.-1", itemJSON)
+	}
+
 	if parts := root.Get("candidates.0.content.parts"); parts.Exists() && parts.IsArray() {
 		parts.ForEach(func(_, p gjson.Result) bool {
 			if p.Get("thought").Bool() {
@@ -583,19 +585,16 @@ func ConvertGeminiResponseToOpenAIResponsesNonStream(_ context.Context, _ string
 				name := fc.Get("name").String()
 				args := fc.Get("args")
 				callID := fmt.Sprintf("call_%x_%d", time.Now().UnixNano(), atomic.AddUint64(&funcCallIDCounter, 1))
-				outputs = append(outputs, map[string]interface{}{
-					"id":     fmt.Sprintf("fc_%s", callID),
-					"type":   "function_call",
-					"status": "completed",
-					"arguments": func() string {
-						if args.Exists() {
-							return args.Raw
-						}
-						return ""
-					}(),
-					"call_id": callID,
-					"name":    name,
-				})
+				itemJSON := `{"id":"","type":"function_call","status":"completed","arguments":"","call_id":"","name":""}`
+				itemJSON, _ = sjson.Set(itemJSON, "id", fmt.Sprintf("fc_%s", callID))
+				itemJSON, _ = sjson.Set(itemJSON, "call_id", callID)
+				itemJSON, _ = sjson.Set(itemJSON, "name", name)
+				argsStr := ""
+				if args.Exists() {
+					argsStr = args.Raw
+				}
+				itemJSON, _ = sjson.Set(itemJSON, "arguments", argsStr)
+				appendOutput(itemJSON)
 				return true
 			}
 			return true
@@ -605,42 +604,24 @@ func ConvertGeminiResponseToOpenAIResponsesNonStream(_ context.Context, _ string
 	// Reasoning output item
 	if reasoningText.Len() > 0 || reasoningEncrypted != "" {
 		rid := strings.TrimPrefix(id, "resp_")
-		item := map[string]interface{}{
-			"id":                fmt.Sprintf("rs_%s", rid),
-			"type":              "reasoning",
-			"encrypted_content": reasoningEncrypted,
-		}
-		var summaries []interface{}
+		itemJSON := `{"id":"","type":"reasoning","encrypted_content":""}`
+		itemJSON, _ = sjson.Set(itemJSON, "id", fmt.Sprintf("rs_%s", rid))
+		itemJSON, _ = sjson.Set(itemJSON, "encrypted_content", reasoningEncrypted)
 		if reasoningText.Len() > 0 {
-			summaries = append(summaries, map[string]interface{}{
-				"type": "summary_text",
-				"text": reasoningText.String(),
-			})
+			summaryJSON := `{"type":"summary_text","text":""}`
+			summaryJSON, _ = sjson.Set(summaryJSON, "text", reasoningText.String())
+			itemJSON, _ = sjson.SetRaw(itemJSON, "summary", "[]")
+			itemJSON, _ = sjson.SetRaw(itemJSON, "summary.-1", summaryJSON)
 		}
-		if summaries != nil {
-			item["summary"] = summaries
-		}
-		outputs = append(outputs, item)
+		appendOutput(itemJSON)
 	}
 
 	// Assistant message output item
 	if haveMessage {
-		outputs = append(outputs, map[string]interface{}{
-			"id":     fmt.Sprintf("msg_%s_0", strings.TrimPrefix(id, "resp_")),
-			"type":   "message",
-			"status": "completed",
-			"content": []interface{}{map[string]interface{}{
-				"type":        "output_text",
-				"annotations": []interface{}{},
-				"logprobs":    []interface{}{},
-				"text":        messageText.String(),
-			}},
-			"role": "assistant",
-		})
-	}
-
-	if len(outputs) > 0 {
-		resp, _ = sjson.Set(resp, "output", outputs)
+		itemJSON := `{"id":"","type":"message","status":"completed","content":[{"type":"output_text","annotations":[],"logprobs":[],"text":""}],"role":"assistant"}`
+		itemJSON, _ = sjson.Set(itemJSON, "id", fmt.Sprintf("msg_%s_0", strings.TrimPrefix(id, "resp_")))
+		itemJSON, _ = sjson.Set(itemJSON, "content.0.text", messageText.String())
+		appendOutput(itemJSON)
 	}
 
 	// usage mapping

internal/translator/openai/claude/openai_claude_request.go

@@ -7,7 +7,6 @@ package claude
 
 import (
 	"bytes"
-	"encoding/json"
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
@@ -63,10 +62,22 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 
 	// Thinking: Convert Claude thinking.budget_tokens to OpenAI reasoning_effort
 	if thinking := root.Get("thinking"); thinking.Exists() && thinking.IsObject() {
-		if thinkingType := thinking.Get("type"); thinkingType.Exists() && thinkingType.String() == "enabled" {
-			if budgetTokens := thinking.Get("budget_tokens"); budgetTokens.Exists() {
-				budget := int(budgetTokens.Int())
-				if effort, ok := util.OpenAIThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
+		if thinkingType := thinking.Get("type"); thinkingType.Exists() {
+			switch thinkingType.String() {
+			case "enabled":
+				if budgetTokens := thinking.Get("budget_tokens"); budgetTokens.Exists() {
+					budget := int(budgetTokens.Int())
+					if effort, ok := util.ThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
+						out, _ = sjson.Set(out, "reasoning_effort", effort)
+					}
+				} else {
+					// No budget_tokens specified, default to "auto" for enabled thinking
+					if effort, ok := util.ThinkingBudgetToEffort(modelName, -1); ok && effort != "" {
+						out, _ = sjson.Set(out, "reasoning_effort", effort)
+					}
+				}
+			case "disabled":
+				if effort, ok := util.ThinkingBudgetToEffort(modelName, 0); ok && effort != "" {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)
 				}
 			}
@@ -126,11 +137,7 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 
 						// Convert input to arguments JSON string
 						if input := part.Get("input"); input.Exists() {
-							if inputJSON, err := json.Marshal(input.Value()); err == nil {
-								toolCallJSON, _ = sjson.Set(toolCallJSON, "function.arguments", string(inputJSON))
-							} else {
-								toolCallJSON, _ = sjson.Set(toolCallJSON, "function.arguments", "{}")
-							}
+							toolCallJSON, _ = sjson.Set(toolCallJSON, "function.arguments", input.Raw)
 						} else {
 							toolCallJSON, _ = sjson.Set(toolCallJSON, "function.arguments", "{}")
 						}
@@ -179,8 +186,7 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 				// Emit tool calls in a separate assistant message
 				if role == "assistant" && len(toolCalls) > 0 {
 					toolCallMsgJSON := `{"role":"assistant","tool_calls":[]}`
-					toolCallsJSON, _ := json.Marshal(toolCalls)
-					toolCallMsgJSON, _ = sjson.SetRaw(toolCallMsgJSON, "tool_calls", string(toolCallsJSON))
+					toolCallMsgJSON, _ = sjson.Set(toolCallMsgJSON, "tool_calls", toolCalls)
 					messagesJSON, _ = sjson.Set(messagesJSON, "-1", gjson.Parse(toolCallMsgJSON).Value())
 				}
 

internal/translator/openai/claude/openai_claude_response.go

@@ -8,7 +8,6 @@ package claude
 import (
 	"bytes"
 	"context"
-	"encoding/json"
 	"fmt"
 	"strings"
 
@@ -128,28 +127,15 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 		param.CreatedAt = root.Get("created").Int()
 	}
 
-	// Check if this is the first chunk (has role)
+	// Emit message_start on the very first chunk, regardless of whether it has a role field.
+	// Some providers (like Copilot) may send tool_calls in the first chunk without a role field.
 	if delta := root.Get("choices.0.delta"); delta.Exists() {
-		if role := delta.Get("role"); role.Exists() && role.String() == "assistant" && !param.MessageStarted {
+		if !param.MessageStarted {
 			// Send message_start event
-			messageStart := map[string]interface{}{
-				"type": "message_start",
-				"message": map[string]interface{}{
-					"id":            param.MessageID,
-					"type":          "message",
-					"role":          "assistant",
-					"model":         param.Model,
-					"content":       []interface{}{},
-					"stop_reason":   nil,
-					"stop_sequence": nil,
-					"usage": map[string]interface{}{
-						"input_tokens":  0,
-						"output_tokens": 0,
-					},
-				},
-			}
-			messageStartJSON, _ := json.Marshal(messageStart)
-			results = append(results, "event: message_start\ndata: "+string(messageStartJSON)+"\n\n")
+			messageStartJSON := `{"type":"message_start","message":{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}}`
+			messageStartJSON, _ = sjson.Set(messageStartJSON, "message.id", param.MessageID)
+			messageStartJSON, _ = sjson.Set(messageStartJSON, "message.model", param.Model)
+			results = append(results, "event: message_start\ndata: "+messageStartJSON+"\n\n")
 			param.MessageStarted = true
 
 			// Don't send content_block_start for text here - wait for actual content
@@ -167,29 +153,16 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 						param.ThinkingContentBlockIndex = param.NextContentBlockIndex
 						param.NextContentBlockIndex++
 					}
-					contentBlockStart := map[string]interface{}{
-						"type":  "content_block_start",
-						"index": param.ThinkingContentBlockIndex,
-						"content_block": map[string]interface{}{
-							"type":     "thinking",
-							"thinking": "",
-						},
-					}
-					contentBlockStartJSON, _ := json.Marshal(contentBlockStart)
-					results = append(results, "event: content_block_start\ndata: "+string(contentBlockStartJSON)+"\n\n")
+					contentBlockStartJSON := `{"type":"content_block_start","index":0,"content_block":{"type":"thinking","thinking":""}}`
+					contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "index", param.ThinkingContentBlockIndex)
+					results = append(results, "event: content_block_start\ndata: "+contentBlockStartJSON+"\n\n")
 					param.ThinkingContentBlockStarted = true
 				}
 
-				thinkingDelta := map[string]interface{}{
-					"type":  "content_block_delta",
-					"index": param.ThinkingContentBlockIndex,
-					"delta": map[string]interface{}{
-						"type":     "thinking_delta",
-						"thinking": reasoningText,
-					},
-				}
-				thinkingDeltaJSON, _ := json.Marshal(thinkingDelta)
-				results = append(results, "event: content_block_delta\ndata: "+string(thinkingDeltaJSON)+"\n\n")
+				thinkingDeltaJSON := `{"type":"content_block_delta","index":0,"delta":{"type":"thinking_delta","thinking":""}}`
+				thinkingDeltaJSON, _ = sjson.Set(thinkingDeltaJSON, "index", param.ThinkingContentBlockIndex)
+				thinkingDeltaJSON, _ = sjson.Set(thinkingDeltaJSON, "delta.thinking", reasoningText)
+				results = append(results, "event: content_block_delta\ndata: "+thinkingDeltaJSON+"\n\n")
 			}
 		}
 
@@ -202,29 +175,16 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 					param.TextContentBlockIndex = param.NextContentBlockIndex
 					param.NextContentBlockIndex++
 				}
-				contentBlockStart := map[string]interface{}{
-					"type":  "content_block_start",
-					"index": param.TextContentBlockIndex,
-					"content_block": map[string]interface{}{
-						"type": "text",
-						"text": "",
-					},
-				}
-				contentBlockStartJSON, _ := json.Marshal(contentBlockStart)
-				results = append(results, "event: content_block_start\ndata: "+string(contentBlockStartJSON)+"\n\n")
+				contentBlockStartJSON := `{"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}`
+				contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "index", param.TextContentBlockIndex)
+				results = append(results, "event: content_block_start\ndata: "+contentBlockStartJSON+"\n\n")
 				param.TextContentBlockStarted = true
 			}
 
-			contentDelta := map[string]interface{}{
-				"type":  "content_block_delta",
-				"index": param.TextContentBlockIndex,
-				"delta": map[string]interface{}{
-					"type": "text_delta",
-					"text": content.String(),
-				},
-			}
-			contentDeltaJSON, _ := json.Marshal(contentDelta)
-			results = append(results, "event: content_block_delta\ndata: "+string(contentDeltaJSON)+"\n\n")
+			contentDeltaJSON := `{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":""}}`
+			contentDeltaJSON, _ = sjson.Set(contentDeltaJSON, "index", param.TextContentBlockIndex)
+			contentDeltaJSON, _ = sjson.Set(contentDeltaJSON, "delta.text", content.String())
+			results = append(results, "event: content_block_delta\ndata: "+contentDeltaJSON+"\n\n")
 
 			// Accumulate content
 			param.ContentAccumulator.WriteString(content.String())
@@ -262,18 +222,11 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 						stopTextContentBlock(param, &results)
 
 						// Send content_block_start for tool_use
-						contentBlockStart := map[string]interface{}{
-							"type":  "content_block_start",
-							"index": blockIndex,
-							"content_block": map[string]interface{}{
-								"type":  "tool_use",
-								"id":    accumulator.ID,
-								"name":  accumulator.Name,
-								"input": map[string]interface{}{},
-							},
-						}
-						contentBlockStartJSON, _ := json.Marshal(contentBlockStart)
-						results = append(results, "event: content_block_start\ndata: "+string(contentBlockStartJSON)+"\n\n")
+						contentBlockStartJSON := `{"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`
+						contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "index", blockIndex)
+						contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "content_block.id", accumulator.ID)
+						contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "content_block.name", accumulator.Name)
+						results = append(results, "event: content_block_start\ndata: "+contentBlockStartJSON+"\n\n")
 					}
 
 					// Handle function arguments
@@ -297,12 +250,9 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 
 		// Send content_block_stop for thinking content if needed
 		if param.ThinkingContentBlockStarted {
-			contentBlockStop := map[string]interface{}{
-				"type":  "content_block_stop",
-				"index": param.ThinkingContentBlockIndex,
-			}
-			contentBlockStopJSON, _ := json.Marshal(contentBlockStop)
-			results = append(results, "event: content_block_stop\ndata: "+string(contentBlockStopJSON)+"\n\n")
+			contentBlockStopJSON := `{"type":"content_block_stop","index":0}`
+			contentBlockStopJSON, _ = sjson.Set(contentBlockStopJSON, "index", param.ThinkingContentBlockIndex)
+			results = append(results, "event: content_block_stop\ndata: "+contentBlockStopJSON+"\n\n")
 			param.ThinkingContentBlockStarted = false
 			param.ThinkingContentBlockIndex = -1
 		}
@@ -318,24 +268,15 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 
 				// Send complete input_json_delta with all accumulated arguments
 				if accumulator.Arguments.Len() > 0 {
-					inputDelta := map[string]interface{}{
-						"type":  "content_block_delta",
-						"index": blockIndex,
-						"delta": map[string]interface{}{
-							"type":         "input_json_delta",
-							"partial_json": util.FixJSON(accumulator.Arguments.String()),
-						},
-					}
-					inputDeltaJSON, _ := json.Marshal(inputDelta)
-					results = append(results, "event: content_block_delta\ndata: "+string(inputDeltaJSON)+"\n\n")
+					inputDeltaJSON := `{"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""}}`
+					inputDeltaJSON, _ = sjson.Set(inputDeltaJSON, "index", blockIndex)
+					inputDeltaJSON, _ = sjson.Set(inputDeltaJSON, "delta.partial_json", util.FixJSON(accumulator.Arguments.String()))
+					results = append(results, "event: content_block_delta\ndata: "+inputDeltaJSON+"\n\n")
 				}
 
-				contentBlockStop := map[string]interface{}{
-					"type":  "content_block_stop",
-					"index": blockIndex,
-				}
-				contentBlockStopJSON, _ := json.Marshal(contentBlockStop)
-				results = append(results, "event: content_block_stop\ndata: "+string(contentBlockStopJSON)+"\n\n")
+				contentBlockStopJSON := `{"type":"content_block_stop","index":0}`
+				contentBlockStopJSON, _ = sjson.Set(contentBlockStopJSON, "index", blockIndex)
+				results = append(results, "event: content_block_stop\ndata: "+contentBlockStopJSON+"\n\n")
 				delete(param.ToolCallBlockIndexes, index)
 			}
 			param.ContentBlocksStopped = true
@@ -360,20 +301,11 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 			}
 		}
 		// Send message_delta with usage
-		messageDelta := map[string]interface{}{
-			"type": "message_delta",
-			"delta": map[string]interface{}{
-				"stop_reason":   mapOpenAIFinishReasonToAnthropic(param.FinishReason),
-				"stop_sequence": nil,
-			},
-			"usage": map[string]interface{}{
-				"input_tokens":  inputTokens,
-				"output_tokens": outputTokens,
-			},
-		}
-
-		messageDeltaJSON, _ := json.Marshal(messageDelta)
-		results = append(results, "event: message_delta\ndata: "+string(messageDeltaJSON)+"\n\n")
+		messageDeltaJSON := `{"type":"message_delta","delta":{"stop_reason":"","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+		messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(param.FinishReason))
+		messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.input_tokens", inputTokens)
+		messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.output_tokens", outputTokens)
+		results = append(results, "event: message_delta\ndata: "+messageDeltaJSON+"\n\n")
 		param.MessageDeltaSent = true
 
 		emitMessageStopIfNeeded(param, &results)
@@ -389,12 +321,9 @@ func convertOpenAIDoneToAnthropic(param *ConvertOpenAIResponseToAnthropicParams)
 
 	// Ensure all content blocks are stopped before final events
 	if param.ThinkingContentBlockStarted {
-		contentBlockStop := map[string]interface{}{
-			"type":  "content_block_stop",
-			"index": param.ThinkingContentBlockIndex,
-		}
-		contentBlockStopJSON, _ := json.Marshal(contentBlockStop)
-		results = append(results, "event: content_block_stop\ndata: "+string(contentBlockStopJSON)+"\n\n")
+		contentBlockStopJSON := `{"type":"content_block_stop","index":0}`
+		contentBlockStopJSON, _ = sjson.Set(contentBlockStopJSON, "index", param.ThinkingContentBlockIndex)
+		results = append(results, "event: content_block_stop\ndata: "+contentBlockStopJSON+"\n\n")
 		param.ThinkingContentBlockStarted = false
 		param.ThinkingContentBlockIndex = -1
 	}
@@ -407,24 +336,15 @@ func convertOpenAIDoneToAnthropic(param *ConvertOpenAIResponseToAnthropicParams)
 			blockIndex := param.toolContentBlockIndex(index)
 
 			if accumulator.Arguments.Len() > 0 {
-				inputDelta := map[string]interface{}{
-					"type":  "content_block_delta",
-					"index": blockIndex,
-					"delta": map[string]interface{}{
-						"type":         "input_json_delta",
-						"partial_json": util.FixJSON(accumulator.Arguments.String()),
-					},
-				}
-				inputDeltaJSON, _ := json.Marshal(inputDelta)
-				results = append(results, "event: content_block_delta\ndata: "+string(inputDeltaJSON)+"\n\n")
+				inputDeltaJSON := `{"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""}}`
+				inputDeltaJSON, _ = sjson.Set(inputDeltaJSON, "index", blockIndex)
+				inputDeltaJSON, _ = sjson.Set(inputDeltaJSON, "delta.partial_json", util.FixJSON(accumulator.Arguments.String()))
+				results = append(results, "event: content_block_delta\ndata: "+inputDeltaJSON+"\n\n")
 			}
 
-			contentBlockStop := map[string]interface{}{
-				"type":  "content_block_stop",
-				"index": blockIndex,
-			}
-			contentBlockStopJSON, _ := json.Marshal(contentBlockStop)
-			results = append(results, "event: content_block_stop\ndata: "+string(contentBlockStopJSON)+"\n\n")
+			contentBlockStopJSON := `{"type":"content_block_stop","index":0}`
+			contentBlockStopJSON, _ = sjson.Set(contentBlockStopJSON, "index", blockIndex)
+			results = append(results, "event: content_block_stop\ndata: "+contentBlockStopJSON+"\n\n")
 			delete(param.ToolCallBlockIndexes, index)
 		}
 		param.ContentBlocksStopped = true
@@ -432,16 +352,9 @@ func convertOpenAIDoneToAnthropic(param *ConvertOpenAIResponseToAnthropicParams)
 
 	// If we haven't sent message_delta yet (no usage info was received), send it now
 	if param.FinishReason != "" && !param.MessageDeltaSent {
-		messageDelta := map[string]interface{}{
-			"type": "message_delta",
-			"delta": map[string]interface{}{
-				"stop_reason":   mapOpenAIFinishReasonToAnthropic(param.FinishReason),
-				"stop_sequence": nil,
-			},
-		}
-
-		messageDeltaJSON, _ := json.Marshal(messageDelta)
-		results = append(results, "event: message_delta\ndata: "+string(messageDeltaJSON)+"\n\n")
+		messageDeltaJSON := `{"type":"message_delta","delta":{"stop_reason":"","stop_sequence":null}}`
+		messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(param.FinishReason))
+		results = append(results, "event: message_delta\ndata: "+messageDeltaJSON+"\n\n")
 		param.MessageDeltaSent = true
 	}
 
@@ -454,105 +367,73 @@ func convertOpenAIDoneToAnthropic(param *ConvertOpenAIResponseToAnthropicParams)
 func convertOpenAINonStreamingToAnthropic(rawJSON []byte) []string {
 	root := gjson.ParseBytes(rawJSON)
 
-	// Build Anthropic response
-	response := map[string]interface{}{
-		"id":            root.Get("id").String(),
-		"type":          "message",
-		"role":          "assistant",
-		"model":         root.Get("model").String(),
-		"content":       []interface{}{},
-		"stop_reason":   nil,
-		"stop_sequence": nil,
-		"usage": map[string]interface{}{
-			"input_tokens":  0,
-			"output_tokens": 0,
-		},
-	}
+	out := `{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}`
+	out, _ = sjson.Set(out, "id", root.Get("id").String())
+	out, _ = sjson.Set(out, "model", root.Get("model").String())
 
 	// Process message content and tool calls
-	var contentBlocks []interface{}
-
-	if choices := root.Get("choices"); choices.Exists() && choices.IsArray() {
+	if choices := root.Get("choices"); choices.Exists() && choices.IsArray() && len(choices.Array()) > 0 {
 		choice := choices.Array()[0] // Take first choice
-		reasoningNode := choice.Get("message.reasoning_content")
-		allReasoning := collectOpenAIReasoningTexts(reasoningNode)
 
-		for _, reasoningText := range allReasoning {
+		reasoningNode := choice.Get("message.reasoning_content")
+		for _, reasoningText := range collectOpenAIReasoningTexts(reasoningNode) {
 			if reasoningText == "" {
 				continue
 			}
-			contentBlocks = append(contentBlocks, map[string]interface{}{
-				"type":     "thinking",
-				"thinking": reasoningText,
-			})
+			block := `{"type":"thinking","thinking":""}`
+			block, _ = sjson.Set(block, "thinking", reasoningText)
+			out, _ = sjson.SetRaw(out, "content.-1", block)
 		}
 
 		// Handle text content
 		if content := choice.Get("message.content"); content.Exists() && content.String() != "" {
-			textBlock := map[string]interface{}{
-				"type": "text",
-				"text": content.String(),
-			}
-			contentBlocks = append(contentBlocks, textBlock)
+			block := `{"type":"text","text":""}`
+			block, _ = sjson.Set(block, "text", content.String())
+			out, _ = sjson.SetRaw(out, "content.-1", block)
 		}
 
 		// Handle tool calls
 		if toolCalls := choice.Get("message.tool_calls"); toolCalls.Exists() && toolCalls.IsArray() {
 			toolCalls.ForEach(func(_, toolCall gjson.Result) bool {
-				toolUseBlock := map[string]interface{}{
-					"type": "tool_use",
-					"id":   toolCall.Get("id").String(),
-					"name": toolCall.Get("function.name").String(),
-				}
+				toolUseBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
+				toolUseBlock, _ = sjson.Set(toolUseBlock, "id", toolCall.Get("id").String())
+				toolUseBlock, _ = sjson.Set(toolUseBlock, "name", toolCall.Get("function.name").String())
 
-				// Parse arguments
-				argsStr := toolCall.Get("function.arguments").String()
-				argsStr = util.FixJSON(argsStr)
-				if argsStr != "" {
-					var args interface{}
-					if err := json.Unmarshal([]byte(argsStr), &args); err == nil {
-						toolUseBlock["input"] = args
+				argsStr := util.FixJSON(toolCall.Get("function.arguments").String())
+				if argsStr != "" && gjson.Valid(argsStr) {
+					argsJSON := gjson.Parse(argsStr)
+					if argsJSON.IsObject() {
+						toolUseBlock, _ = sjson.SetRaw(toolUseBlock, "input", argsJSON.Raw)
 					} else {
-						toolUseBlock["input"] = map[string]interface{}{}
+						toolUseBlock, _ = sjson.SetRaw(toolUseBlock, "input", "{}")
 					}
 				} else {
-					toolUseBlock["input"] = map[string]interface{}{}
+					toolUseBlock, _ = sjson.SetRaw(toolUseBlock, "input", "{}")
 				}
 
-				contentBlocks = append(contentBlocks, toolUseBlock)
+				out, _ = sjson.SetRaw(out, "content.-1", toolUseBlock)
 				return true
 			})
 		}
 
 		// Set stop reason
 		if finishReason := choice.Get("finish_reason"); finishReason.Exists() {
-			response["stop_reason"] = mapOpenAIFinishReasonToAnthropic(finishReason.String())
+			out, _ = sjson.Set(out, "stop_reason", mapOpenAIFinishReasonToAnthropic(finishReason.String()))
 		}
 	}
 
-	response["content"] = contentBlocks
-
 	// Set usage information
 	if usage := root.Get("usage"); usage.Exists() {
-		response["usage"] = map[string]interface{}{
-			"input_tokens":  usage.Get("prompt_tokens").Int(),
-			"output_tokens": usage.Get("completion_tokens").Int(),
-			"reasoning_tokens": func() int64 {
-				if v := usage.Get("completion_tokens_details.reasoning_tokens"); v.Exists() {
-					return v.Int()
-				}
-				return 0
-			}(),
-		}
-	} else {
-		response["usage"] = map[string]interface{}{
-			"input_tokens":  0,
-			"output_tokens": 0,
+		out, _ = sjson.Set(out, "usage.input_tokens", usage.Get("prompt_tokens").Int())
+		out, _ = sjson.Set(out, "usage.output_tokens", usage.Get("completion_tokens").Int())
+		reasoningTokens := int64(0)
+		if v := usage.Get("completion_tokens_details.reasoning_tokens"); v.Exists() {
+			reasoningTokens = v.Int()
 		}
+		out, _ = sjson.Set(out, "usage.reasoning_tokens", reasoningTokens)
 	}
 
-	responseJSON, _ := json.Marshal(response)
-	return []string{string(responseJSON)}
+	return []string{out}
 }
 
 // mapOpenAIFinishReasonToAnthropic maps OpenAI finish reasons to Anthropic equivalents
@@ -619,12 +500,9 @@ func stopThinkingContentBlock(param *ConvertOpenAIResponseToAnthropicParams, res
 	if !param.ThinkingContentBlockStarted {
 		return
 	}
-	contentBlockStop := map[string]interface{}{
-		"type":  "content_block_stop",
-		"index": param.ThinkingContentBlockIndex,
-	}
-	contentBlockStopJSON, _ := json.Marshal(contentBlockStop)
-	*results = append(*results, "event: content_block_stop\ndata: "+string(contentBlockStopJSON)+"\n\n")
+	contentBlockStopJSON := `{"type":"content_block_stop","index":0}`
+	contentBlockStopJSON, _ = sjson.Set(contentBlockStopJSON, "index", param.ThinkingContentBlockIndex)
+	*results = append(*results, "event: content_block_stop\ndata: "+contentBlockStopJSON+"\n\n")
 	param.ThinkingContentBlockStarted = false
 	param.ThinkingContentBlockIndex = -1
 }
@@ -641,12 +519,9 @@ func stopTextContentBlock(param *ConvertOpenAIResponseToAnthropicParams, results
 	if !param.TextContentBlockStarted {
 		return
 	}
-	contentBlockStop := map[string]interface{}{
-		"type":  "content_block_stop",
-		"index": param.TextContentBlockIndex,
-	}
-	contentBlockStopJSON, _ := json.Marshal(contentBlockStop)
-	*results = append(*results, "event: content_block_stop\ndata: "+string(contentBlockStopJSON)+"\n\n")
+	contentBlockStopJSON := `{"type":"content_block_stop","index":0}`
+	contentBlockStopJSON, _ = sjson.Set(contentBlockStopJSON, "index", param.TextContentBlockIndex)
+	*results = append(*results, "event: content_block_stop\ndata: "+contentBlockStopJSON+"\n\n")
 	param.TextContentBlockStarted = false
 	param.TextContentBlockIndex = -1
 }
@@ -666,29 +541,19 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 	_ = requestRawJSON
 
 	root := gjson.ParseBytes(rawJSON)
+	out := `{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}`
+	out, _ = sjson.Set(out, "id", root.Get("id").String())
+	out, _ = sjson.Set(out, "model", root.Get("model").String())
 
-	response := map[string]interface{}{
-		"id":            root.Get("id").String(),
-		"type":          "message",
-		"role":          "assistant",
-		"model":         root.Get("model").String(),
-		"content":       []interface{}{},
-		"stop_reason":   nil,
-		"stop_sequence": nil,
-		"usage": map[string]interface{}{
-			"input_tokens":  0,
-			"output_tokens": 0,
-		},
-	}
-
-	contentBlocks := make([]interface{}, 0)
 	hasToolCall := false
+	stopReasonSet := false
 
 	if choices := root.Get("choices"); choices.Exists() && choices.IsArray() && len(choices.Array()) > 0 {
 		choice := choices.Array()[0]
 
 		if finishReason := choice.Get("finish_reason"); finishReason.Exists() {
-			response["stop_reason"] = mapOpenAIFinishReasonToAnthropic(finishReason.String())
+			out, _ = sjson.Set(out, "stop_reason", mapOpenAIFinishReasonToAnthropic(finishReason.String()))
+			stopReasonSet = true
 		}
 
 		if message := choice.Get("message"); message.Exists() {
@@ -701,10 +566,9 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 						if textBuilder.Len() == 0 {
 							return
 						}
-						contentBlocks = append(contentBlocks, map[string]interface{}{
-							"type": "text",
-							"text": textBuilder.String(),
-						})
+						block := `{"type":"text","text":""}`
+						block, _ = sjson.Set(block, "text", textBuilder.String())
+						out, _ = sjson.SetRaw(out, "content.-1", block)
 						textBuilder.Reset()
 					}
 
@@ -712,16 +576,14 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 						if thinkingBuilder.Len() == 0 {
 							return
 						}
-						contentBlocks = append(contentBlocks, map[string]interface{}{
-							"type":     "thinking",
-							"thinking": thinkingBuilder.String(),
-						})
+						block := `{"type":"thinking","thinking":""}`
+						block, _ = sjson.Set(block, "thinking", thinkingBuilder.String())
+						out, _ = sjson.SetRaw(out, "content.-1", block)
 						thinkingBuilder.Reset()
 					}
 
 					for _, item := range contentResult.Array() {
-						typeStr := item.Get("type").String()
-						switch typeStr {
+						switch item.Get("type").String() {
 						case "text":
 							flushThinking()
 							textBuilder.WriteString(item.Get("text").String())
@@ -732,25 +594,23 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 							if toolCalls.IsArray() {
 								toolCalls.ForEach(func(_, tc gjson.Result) bool {
 									hasToolCall = true
-									toolUse := map[string]interface{}{
-										"type": "tool_use",
-										"id":   tc.Get("id").String(),
-										"name": tc.Get("function.name").String(),
-									}
+									toolUse := `{"type":"tool_use","id":"","name":"","input":{}}`
+									toolUse, _ = sjson.Set(toolUse, "id", tc.Get("id").String())
+									toolUse, _ = sjson.Set(toolUse, "name", tc.Get("function.name").String())
 
 									argsStr := util.FixJSON(tc.Get("function.arguments").String())
-									if argsStr != "" {
-										var parsed interface{}
-										if err := json.Unmarshal([]byte(argsStr), &parsed); err == nil {
-											toolUse["input"] = parsed
+									if argsStr != "" && gjson.Valid(argsStr) {
+										argsJSON := gjson.Parse(argsStr)
+										if argsJSON.IsObject() {
+											toolUse, _ = sjson.SetRaw(toolUse, "input", argsJSON.Raw)
 										} else {
-											toolUse["input"] = map[string]interface{}{}
+											toolUse, _ = sjson.SetRaw(toolUse, "input", "{}")
 										}
 									} else {
-										toolUse["input"] = map[string]interface{}{}
+										toolUse, _ = sjson.SetRaw(toolUse, "input", "{}")
 									}
 
-									contentBlocks = append(contentBlocks, toolUse)
+									out, _ = sjson.SetRaw(out, "content.-1", toolUse)
 									return true
 								})
 							}
@@ -770,10 +630,9 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 				} else if contentResult.Type == gjson.String {
 					textContent := contentResult.String()
 					if textContent != "" {
-						contentBlocks = append(contentBlocks, map[string]interface{}{
-							"type": "text",
-							"text": textContent,
-						})
+						block := `{"type":"text","text":""}`
+						block, _ = sjson.Set(block, "text", textContent)
+						out, _ = sjson.SetRaw(out, "content.-1", block)
 					}
 				}
 			}
@@ -783,81 +642,52 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 					if reasoningText == "" {
 						continue
 					}
-					contentBlocks = append(contentBlocks, map[string]interface{}{
-						"type":     "thinking",
-						"thinking": reasoningText,
-					})
+					block := `{"type":"thinking","thinking":""}`
+					block, _ = sjson.Set(block, "thinking", reasoningText)
+					out, _ = sjson.SetRaw(out, "content.-1", block)
 				}
 			}
 
 			if toolCalls := message.Get("tool_calls"); toolCalls.Exists() && toolCalls.IsArray() {
 				toolCalls.ForEach(func(_, toolCall gjson.Result) bool {
 					hasToolCall = true
-					toolUseBlock := map[string]interface{}{
-						"type": "tool_use",
-						"id":   toolCall.Get("id").String(),
-						"name": toolCall.Get("function.name").String(),
-					}
+					toolUseBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
+					toolUseBlock, _ = sjson.Set(toolUseBlock, "id", toolCall.Get("id").String())
+					toolUseBlock, _ = sjson.Set(toolUseBlock, "name", toolCall.Get("function.name").String())
 
-					argsStr := toolCall.Get("function.arguments").String()
-					argsStr = util.FixJSON(argsStr)
-					if argsStr != "" {
-						var args interface{}
-						if err := json.Unmarshal([]byte(argsStr), &args); err == nil {
-							toolUseBlock["input"] = args
+					argsStr := util.FixJSON(toolCall.Get("function.arguments").String())
+					if argsStr != "" && gjson.Valid(argsStr) {
+						argsJSON := gjson.Parse(argsStr)
+						if argsJSON.IsObject() {
+							toolUseBlock, _ = sjson.SetRaw(toolUseBlock, "input", argsJSON.Raw)
 						} else {
-							toolUseBlock["input"] = map[string]interface{}{}
+							toolUseBlock, _ = sjson.SetRaw(toolUseBlock, "input", "{}")
 						}
 					} else {
-						toolUseBlock["input"] = map[string]interface{}{}
+						toolUseBlock, _ = sjson.SetRaw(toolUseBlock, "input", "{}")
 					}
 
-					contentBlocks = append(contentBlocks, toolUseBlock)
+					out, _ = sjson.SetRaw(out, "content.-1", toolUseBlock)
 					return true
 				})
 			}
 		}
 	}
 
-	response["content"] = contentBlocks
-
 	if respUsage := root.Get("usage"); respUsage.Exists() {
-		usageJSON := `{}`
-		usageJSON, _ = sjson.Set(usageJSON, "input_tokens", respUsage.Get("prompt_tokens").Int())
-		usageJSON, _ = sjson.Set(usageJSON, "output_tokens", respUsage.Get("completion_tokens").Int())
-		parsedUsage := gjson.Parse(usageJSON).Value().(map[string]interface{})
-		response["usage"] = parsedUsage
-	} else {
-		response["usage"] = `{"input_tokens":0,"output_tokens":0}`
+		out, _ = sjson.Set(out, "usage.input_tokens", respUsage.Get("prompt_tokens").Int())
+		out, _ = sjson.Set(out, "usage.output_tokens", respUsage.Get("completion_tokens").Int())
 	}
 
-	if response["stop_reason"] == nil {
+	if !stopReasonSet {
 		if hasToolCall {
-			response["stop_reason"] = "tool_use"
+			out, _ = sjson.Set(out, "stop_reason", "tool_use")
 		} else {
-			response["stop_reason"] = "end_turn"
+			out, _ = sjson.Set(out, "stop_reason", "end_turn")
 		}
 	}
 
-	if !hasToolCall {
-		if toolBlocks := response["content"].([]interface{}); len(toolBlocks) > 0 {
-			for _, block := range toolBlocks {
-				if m, ok := block.(map[string]interface{}); ok && m["type"] == "tool_use" {
-					hasToolCall = true
-					break
-				}
-			}
-		}
-		if hasToolCall {
-			response["stop_reason"] = "tool_use"
-		}
-	}
-
-	responseJSON, err := json.Marshal(response)
-	if err != nil {
-		return ""
-	}
-	return string(responseJSON)
+	return out
 }
 
 func ClaudeTokenCount(ctx context.Context, count int64) string {

internal/translator/openai/gemini/openai_gemini_request.go

@@ -8,7 +8,6 @@ package gemini
 import (
 	"bytes"
 	"crypto/rand"
-	"encoding/json"
 	"fmt"
 	"math/big"
 	"strings"
@@ -83,7 +82,7 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
 			if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
 				budget := int(thinkingBudget.Int())
-				if effort, ok := util.OpenAIThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
+				if effort, ok := util.ThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)
 				}
 			}
@@ -94,7 +93,6 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 	out, _ = sjson.Set(out, "stream", stream)
 
 	// Process contents (Gemini messages) -> OpenAI messages
-	var openAIMessages []interface{}
 	var toolCallIDs []string // Track tool call IDs for matching with tool results
 
 	// System instruction -> OpenAI system message
@@ -105,22 +103,17 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 	}
 	if systemInstruction.Exists() {
 		parts := systemInstruction.Get("parts")
-		msg := map[string]interface{}{
-			"role":    "system",
-			"content": []interface{}{},
-		}
-
-		var aggregatedParts []interface{}
+		msg := `{"role":"system","content":[]}`
+		hasContent := false
 
 		if parts.Exists() && parts.IsArray() {
 			parts.ForEach(func(_, part gjson.Result) bool {
 				// Handle text parts
 				if text := part.Get("text"); text.Exists() {
-					formattedText := text.String()
-					aggregatedParts = append(aggregatedParts, map[string]interface{}{
-						"type": "text",
-						"text": formattedText,
-					})
+					contentPart := `{"type":"text","text":""}`
+					contentPart, _ = sjson.Set(contentPart, "text", text.String())
+					msg, _ = sjson.SetRaw(msg, "content.-1", contentPart)
+					hasContent = true
 				}
 
 				// Handle inline data (e.g., images)
@@ -132,20 +125,17 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 					data := inlineData.Get("data").String()
 					imageURL := fmt.Sprintf("data:%s;base64,%s", mimeType, data)
 
-					aggregatedParts = append(aggregatedParts, map[string]interface{}{
-						"type": "image_url",
-						"image_url": map[string]interface{}{
-							"url": imageURL,
-						},
-					})
+					contentPart := `{"type":"image_url","image_url":{"url":""}}`
+					contentPart, _ = sjson.Set(contentPart, "image_url.url", imageURL)
+					msg, _ = sjson.SetRaw(msg, "content.-1", contentPart)
+					hasContent = true
 				}
 				return true
 			})
 		}
 
-		if len(aggregatedParts) > 0 {
-			msg["content"] = aggregatedParts
-			openAIMessages = append(openAIMessages, msg)
+		if hasContent {
+			out, _ = sjson.SetRaw(out, "messages.-1", msg)
 		}
 	}
 
@@ -159,16 +149,15 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 				role = "assistant"
 			}
 
-			// Create OpenAI message
-			msg := map[string]interface{}{
-				"role":    role,
-				"content": "",
-			}
+			msg := `{"role":"","content":""}`
+			msg, _ = sjson.Set(msg, "role", role)
 
 			var textBuilder strings.Builder
-			var aggregatedParts []interface{}
+			contentWrapper := `{"arr":[]}`
+			contentPartsCount := 0
 			onlyTextContent := true
-			var toolCalls []interface{}
+			toolCallsWrapper := `{"arr":[]}`
+			toolCallsCount := 0
 
 			if parts.Exists() && parts.IsArray() {
 				parts.ForEach(func(_, part gjson.Result) bool {
@@ -176,10 +165,10 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 					if text := part.Get("text"); text.Exists() {
 						formattedText := text.String()
 						textBuilder.WriteString(formattedText)
-						aggregatedParts = append(aggregatedParts, map[string]interface{}{
-							"type": "text",
-							"text": formattedText,
-						})
+						contentPart := `{"type":"text","text":""}`
+						contentPart, _ = sjson.Set(contentPart, "text", formattedText)
+						contentWrapper, _ = sjson.SetRaw(contentWrapper, "arr.-1", contentPart)
+						contentPartsCount++
 					}
 
 					// Handle inline data (e.g., images)
@@ -193,12 +182,10 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 						data := inlineData.Get("data").String()
 						imageURL := fmt.Sprintf("data:%s;base64,%s", mimeType, data)
 
-						aggregatedParts = append(aggregatedParts, map[string]interface{}{
-							"type": "image_url",
-							"image_url": map[string]interface{}{
-								"url": imageURL,
-							},
-						})
+						contentPart := `{"type":"image_url","image_url":{"url":""}}`
+						contentPart, _ = sjson.Set(contentPart, "image_url.url", imageURL)
+						contentWrapper, _ = sjson.SetRaw(contentWrapper, "arr.-1", contentPart)
+						contentPartsCount++
 					}
 
 					// Handle function calls (Gemini) -> tool calls (OpenAI)
@@ -206,44 +193,32 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 						toolCallID := genToolCallID()
 						toolCallIDs = append(toolCallIDs, toolCallID)
 
-						toolCall := map[string]interface{}{
-							"id":   toolCallID,
-							"type": "function",
-							"function": map[string]interface{}{
-								"name": functionCall.Get("name").String(),
-							},
-						}
+						toolCall := `{"id":"","type":"function","function":{"name":"","arguments":""}}`
+						toolCall, _ = sjson.Set(toolCall, "id", toolCallID)
+						toolCall, _ = sjson.Set(toolCall, "function.name", functionCall.Get("name").String())
 
 						// Convert args to arguments JSON string
 						if args := functionCall.Get("args"); args.Exists() {
-							argsJSON, _ := json.Marshal(args.Value())
-							toolCall["function"].(map[string]interface{})["arguments"] = string(argsJSON)
+							toolCall, _ = sjson.Set(toolCall, "function.arguments", args.Raw)
 						} else {
-							toolCall["function"].(map[string]interface{})["arguments"] = "{}"
+							toolCall, _ = sjson.Set(toolCall, "function.arguments", "{}")
 						}
 
-						toolCalls = append(toolCalls, toolCall)
+						toolCallsWrapper, _ = sjson.SetRaw(toolCallsWrapper, "arr.-1", toolCall)
+						toolCallsCount++
 					}
 
 					// Handle function responses (Gemini) -> tool role messages (OpenAI)
 					if functionResponse := part.Get("functionResponse"); functionResponse.Exists() {
 						// Create tool message for function response
-						toolMsg := map[string]interface{}{
-							"role":         "tool",
-							"tool_call_id": "", // Will be set based on context
-							"content":      "",
-						}
+						toolMsg := `{"role":"tool","tool_call_id":"","content":""}`
 
 						// Convert response.content to JSON string
 						if response := functionResponse.Get("response"); response.Exists() {
-							if content = response.Get("content"); content.Exists() {
-								// Use the content field from the response
-								contentJSON, _ := json.Marshal(content.Value())
-								toolMsg["content"] = string(contentJSON)
+							if contentField := response.Get("content"); contentField.Exists() {
+								toolMsg, _ = sjson.Set(toolMsg, "content", contentField.Raw)
 							} else {
-								// Fallback to entire response
-								responseJSON, _ := json.Marshal(response.Value())
-								toolMsg["content"] = string(responseJSON)
+								toolMsg, _ = sjson.Set(toolMsg, "content", response.Raw)
 							}
 						}
 
@@ -252,13 +227,13 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 						if len(toolCallIDs) > 0 {
 							// Use the last tool call ID (simple matching by function name)
 							// In a real implementation, you might want more sophisticated matching
-							toolMsg["tool_call_id"] = toolCallIDs[len(toolCallIDs)-1]
+							toolMsg, _ = sjson.Set(toolMsg, "tool_call_id", toolCallIDs[len(toolCallIDs)-1])
 						} else {
 							// Generate a tool call ID if none available
-							toolMsg["tool_call_id"] = genToolCallID()
+							toolMsg, _ = sjson.Set(toolMsg, "tool_call_id", genToolCallID())
 						}
 
-						openAIMessages = append(openAIMessages, toolMsg)
+						out, _ = sjson.SetRaw(out, "messages.-1", toolMsg)
 					}
 
 					return true
@@ -266,170 +241,46 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 			}
 
 			// Set content
-			if len(aggregatedParts) > 0 {
+			if contentPartsCount > 0 {
 				if onlyTextContent {
-					msg["content"] = textBuilder.String()
+					msg, _ = sjson.Set(msg, "content", textBuilder.String())
 				} else {
-					msg["content"] = aggregatedParts
+					msg, _ = sjson.SetRaw(msg, "content", gjson.Get(contentWrapper, "arr").Raw)
 				}
 			}
 
 			// Set tool calls if any
-			if len(toolCalls) > 0 {
-				msg["tool_calls"] = toolCalls
+			if toolCallsCount > 0 {
+				msg, _ = sjson.SetRaw(msg, "tool_calls", gjson.Get(toolCallsWrapper, "arr").Raw)
 			}
 
-			openAIMessages = append(openAIMessages, msg)
-
-			// switch role {
-			// case "user", "model":
-			// 	// Convert role: model -> assistant
-			// 	if role == "model" {
-			// 		role = "assistant"
-			// 	}
-			//
-			// 	// Create OpenAI message
-			// 	msg := map[string]interface{}{
-			// 		"role":    role,
-			// 		"content": "",
-			// 	}
-			//
-			// 	var contentParts []string
-			// 	var toolCalls []interface{}
-			//
-			// 	if parts.Exists() && parts.IsArray() {
-			// 		parts.ForEach(func(_, part gjson.Result) bool {
-			// 			// Handle text parts
-			// 			if text := part.Get("text"); text.Exists() {
-			// 				contentParts = append(contentParts, text.String())
-			// 			}
-			//
-			// 			// Handle function calls (Gemini) -> tool calls (OpenAI)
-			// 			if functionCall := part.Get("functionCall"); functionCall.Exists() {
-			// 				toolCallID := genToolCallID()
-			// 				toolCallIDs = append(toolCallIDs, toolCallID)
-			//
-			// 				toolCall := map[string]interface{}{
-			// 					"id":   toolCallID,
-			// 					"type": "function",
-			// 					"function": map[string]interface{}{
-			// 						"name": functionCall.Get("name").String(),
-			// 					},
-			// 				}
-			//
-			// 				// Convert args to arguments JSON string
-			// 				if args := functionCall.Get("args"); args.Exists() {
-			// 					argsJSON, _ := json.Marshal(args.Value())
-			// 					toolCall["function"].(map[string]interface{})["arguments"] = string(argsJSON)
-			// 				} else {
-			// 					toolCall["function"].(map[string]interface{})["arguments"] = "{}"
-			// 				}
-			//
-			// 				toolCalls = append(toolCalls, toolCall)
-			// 			}
-			//
-			// 			return true
-			// 		})
-			// 	}
-			//
-			// 	// Set content
-			// 	if len(contentParts) > 0 {
-			// 		msg["content"] = strings.Join(contentParts, "")
-			// 	}
-			//
-			// 	// Set tool calls if any
-			// 	if len(toolCalls) > 0 {
-			// 		msg["tool_calls"] = toolCalls
-			// 	}
-			//
-			// 	openAIMessages = append(openAIMessages, msg)
-			//
-			// case "function":
-			// 	// Handle Gemini function role -> OpenAI tool role
-			// 	if parts.Exists() && parts.IsArray() {
-			// 		parts.ForEach(func(_, part gjson.Result) bool {
-			// 			// Handle function responses (Gemini) -> tool role messages (OpenAI)
-			// 			if functionResponse := part.Get("functionResponse"); functionResponse.Exists() {
-			// 				// Create tool message for function response
-			// 				toolMsg := map[string]interface{}{
-			// 					"role":         "tool",
-			// 					"tool_call_id": "", // Will be set based on context
-			// 					"content":      "",
-			// 				}
-			//
-			// 				// Convert response.content to JSON string
-			// 				if response := functionResponse.Get("response"); response.Exists() {
-			// 					if content = response.Get("content"); content.Exists() {
-			// 						// Use the content field from the response
-			// 						contentJSON, _ := json.Marshal(content.Value())
-			// 						toolMsg["content"] = string(contentJSON)
-			// 					} else {
-			// 						// Fallback to entire response
-			// 						responseJSON, _ := json.Marshal(response.Value())
-			// 						toolMsg["content"] = string(responseJSON)
-			// 					}
-			// 				}
-			//
-			// 				// Try to match with previous tool call ID
-			// 				_ = functionResponse.Get("name").String() // functionName not used for now
-			// 				if len(toolCallIDs) > 0 {
-			// 					// Use the last tool call ID (simple matching by function name)
-			// 					// In a real implementation, you might want more sophisticated matching
-			// 					toolMsg["tool_call_id"] = toolCallIDs[len(toolCallIDs)-1]
-			// 				} else {
-			// 					// Generate a tool call ID if none available
-			// 					toolMsg["tool_call_id"] = genToolCallID()
-			// 				}
-			//
-			// 				openAIMessages = append(openAIMessages, toolMsg)
-			// 			}
-			//
-			// 			return true
-			// 		})
-			// 	}
-			// }
+			out, _ = sjson.SetRaw(out, "messages.-1", msg)
 			return true
 		})
 	}
 
-	// Set messages
-	if len(openAIMessages) > 0 {
-		messagesJSON, _ := json.Marshal(openAIMessages)
-		out, _ = sjson.SetRaw(out, "messages", string(messagesJSON))
-	}
-
 	// Tools mapping: Gemini tools -> OpenAI tools
 	if tools := root.Get("tools"); tools.Exists() && tools.IsArray() {
-		var openAITools []interface{}
 		tools.ForEach(func(_, tool gjson.Result) bool {
 			if functionDeclarations := tool.Get("functionDeclarations"); functionDeclarations.Exists() && functionDeclarations.IsArray() {
 				functionDeclarations.ForEach(func(_, funcDecl gjson.Result) bool {
-					openAITool := map[string]interface{}{
-						"type": "function",
-						"function": map[string]interface{}{
-							"name":        funcDecl.Get("name").String(),
-							"description": funcDecl.Get("description").String(),
-						},
-					}
+					openAITool := `{"type":"function","function":{"name":"","description":""}}`
+					openAITool, _ = sjson.Set(openAITool, "function.name", funcDecl.Get("name").String())
+					openAITool, _ = sjson.Set(openAITool, "function.description", funcDecl.Get("description").String())
 
 					// Convert parameters schema
 					if parameters := funcDecl.Get("parameters"); parameters.Exists() {
-						openAITool["function"].(map[string]interface{})["parameters"] = parameters.Value()
-					} else if parameters = funcDecl.Get("parametersJsonSchema"); parameters.Exists() {
-						openAITool["function"].(map[string]interface{})["parameters"] = parameters.Value()
+						openAITool, _ = sjson.SetRaw(openAITool, "function.parameters", parameters.Raw)
+					} else if parameters := funcDecl.Get("parametersJsonSchema"); parameters.Exists() {
+						openAITool, _ = sjson.SetRaw(openAITool, "function.parameters", parameters.Raw)
 					}
 
-					openAITools = append(openAITools, openAITool)
+					out, _ = sjson.SetRaw(out, "tools.-1", openAITool)
 					return true
 				})
 			}
 			return true
 		})
-
-		if len(openAITools) > 0 {
-			toolsJSON, _ := json.Marshal(openAITools)
-			out, _ = sjson.SetRaw(out, "tools", string(toolsJSON))
-		}
 	}
 
 	// Tool choice mapping (Gemini doesn't have direct equivalent, but we can handle it)

internal/translator/openai/gemini/openai_gemini_response.go

@@ -8,7 +8,6 @@ package gemini
 import (
 	"bytes"
 	"context"
-	"encoding/json"
 	"fmt"
 	"strconv"
 	"strings"
@@ -84,15 +83,12 @@ func ConvertOpenAIResponseToGemini(_ context.Context, _ string, originalRequestR
 					template, _ = sjson.Set(template, "model", model.String())
 				}
 
-				usageObj := map[string]interface{}{
-					"promptTokenCount":     usage.Get("prompt_tokens").Int(),
-					"candidatesTokenCount": usage.Get("completion_tokens").Int(),
-					"totalTokenCount":      usage.Get("total_tokens").Int(),
-				}
+				template, _ = sjson.Set(template, "usageMetadata.promptTokenCount", usage.Get("prompt_tokens").Int())
+				template, _ = sjson.Set(template, "usageMetadata.candidatesTokenCount", usage.Get("completion_tokens").Int())
+				template, _ = sjson.Set(template, "usageMetadata.totalTokenCount", usage.Get("total_tokens").Int())
 				if reasoningTokens := reasoningTokensFromUsage(usage); reasoningTokens > 0 {
-					usageObj["thoughtsTokenCount"] = reasoningTokens
+					template, _ = sjson.Set(template, "usageMetadata.thoughtsTokenCount", reasoningTokens)
 				}
-				template, _ = sjson.Set(template, "usageMetadata", usageObj)
 				return []string{template}
 			}
 			return []string{}
@@ -133,13 +129,8 @@ func ConvertOpenAIResponseToGemini(_ context.Context, _ string, originalRequestR
 						continue
 					}
 					reasoningTemplate := baseTemplate
-					parts := []interface{}{
-						map[string]interface{}{
-							"thought": true,
-							"text":    reasoningText,
-						},
-					}
-					reasoningTemplate, _ = sjson.Set(reasoningTemplate, "candidates.0.content.parts", parts)
+					reasoningTemplate, _ = sjson.Set(reasoningTemplate, "candidates.0.content.parts.0.thought", true)
+					reasoningTemplate, _ = sjson.Set(reasoningTemplate, "candidates.0.content.parts.0.text", reasoningText)
 					chunkOutputs = append(chunkOutputs, reasoningTemplate)
 				}
 			}
@@ -150,13 +141,8 @@ func ConvertOpenAIResponseToGemini(_ context.Context, _ string, originalRequestR
 				(*param).(*ConvertOpenAIResponseToGeminiParams).ContentAccumulator.WriteString(contentText)
 
 				// Create text part for this delta
-				parts := []interface{}{
-					map[string]interface{}{
-						"text": contentText,
-					},
-				}
 				contentTemplate := baseTemplate
-				contentTemplate, _ = sjson.Set(contentTemplate, "candidates.0.content.parts", parts)
+				contentTemplate, _ = sjson.Set(contentTemplate, "candidates.0.content.parts.0.text", contentText)
 				chunkOutputs = append(chunkOutputs, contentTemplate)
 			}
 
@@ -225,24 +211,13 @@ func ConvertOpenAIResponseToGemini(_ context.Context, _ string, originalRequestR
 
 				// If we have accumulated tool calls, output them now
 				if len((*param).(*ConvertOpenAIResponseToGeminiParams).ToolCallsAccumulator) > 0 {
-					var parts []interface{}
+					partIndex := 0
 					for _, accumulator := range (*param).(*ConvertOpenAIResponseToGeminiParams).ToolCallsAccumulator {
-						argsStr := accumulator.Arguments.String()
-						var argsMap map[string]interface{}
-
-						argsMap = parseArgsToMap(argsStr)
-
-						functionCallPart := map[string]interface{}{
-							"functionCall": map[string]interface{}{
-								"name": accumulator.Name,
-								"args": argsMap,
-							},
-						}
-						parts = append(parts, functionCallPart)
-					}
-
-					if len(parts) > 0 {
-						template, _ = sjson.Set(template, "candidates.0.content.parts", parts)
+						namePath := fmt.Sprintf("candidates.0.content.parts.%d.functionCall.name", partIndex)
+						argsPath := fmt.Sprintf("candidates.0.content.parts.%d.functionCall.args", partIndex)
+						template, _ = sjson.Set(template, namePath, accumulator.Name)
+						template, _ = sjson.SetRaw(template, argsPath, parseArgsToObjectRaw(accumulator.Arguments.String()))
+						partIndex++
 					}
 
 					// Clear accumulators
@@ -255,15 +230,12 @@ func ConvertOpenAIResponseToGemini(_ context.Context, _ string, originalRequestR
 
 			// Handle usage information
 			if usage := root.Get("usage"); usage.Exists() {
-				usageObj := map[string]interface{}{
-					"promptTokenCount":     usage.Get("prompt_tokens").Int(),
-					"candidatesTokenCount": usage.Get("completion_tokens").Int(),
-					"totalTokenCount":      usage.Get("total_tokens").Int(),
-				}
+				template, _ = sjson.Set(template, "usageMetadata.promptTokenCount", usage.Get("prompt_tokens").Int())
+				template, _ = sjson.Set(template, "usageMetadata.candidatesTokenCount", usage.Get("completion_tokens").Int())
+				template, _ = sjson.Set(template, "usageMetadata.totalTokenCount", usage.Get("total_tokens").Int())
 				if reasoningTokens := reasoningTokensFromUsage(usage); reasoningTokens > 0 {
-					usageObj["thoughtsTokenCount"] = reasoningTokens
+					template, _ = sjson.Set(template, "usageMetadata.thoughtsTokenCount", reasoningTokens)
 				}
-				template, _ = sjson.Set(template, "usageMetadata", usageObj)
 				results = append(results, template)
 				return true
 			}
@@ -291,46 +263,54 @@ func mapOpenAIFinishReasonToGemini(openAIReason string) string {
 	}
 }
 
-// parseArgsToMap safely parses a JSON string of function arguments into a map.
-// It returns an empty map if the input is empty or cannot be parsed as a JSON object.
-func parseArgsToMap(argsStr string) map[string]interface{} {
+// parseArgsToObjectRaw safely parses a JSON string of function arguments into an object JSON string.
+// It returns "{}" if the input is empty or cannot be parsed as a JSON object.
+func parseArgsToObjectRaw(argsStr string) string {
 	trimmed := strings.TrimSpace(argsStr)
 	if trimmed == "" || trimmed == "{}" {
-		return map[string]interface{}{}
+		return "{}"
 	}
 
 	// First try strict JSON
-	var out map[string]interface{}
-	if errUnmarshal := json.Unmarshal([]byte(trimmed), &out); errUnmarshal == nil {
-		return out
+	if gjson.Valid(trimmed) {
+		strict := gjson.Parse(trimmed)
+		if strict.IsObject() {
+			return strict.Raw
+		}
 	}
 
 	// Tolerant parse: handle streams where values are barewords (e.g., 北京, celsius)
-	tolerant := tolerantParseJSONMap(trimmed)
-	if len(tolerant) > 0 {
+	tolerant := tolerantParseJSONObjectRaw(trimmed)
+	if tolerant != "{}" {
 		return tolerant
 	}
 
 	// Fallback: return empty object when parsing fails
-	return map[string]interface{}{}
+	return "{}"
 }
 
-// tolerantParseJSONMap attempts to parse a JSON-like object string into a map, tolerating
+func escapeSjsonPathKey(key string) string {
+	key = strings.ReplaceAll(key, `\`, `\\`)
+	key = strings.ReplaceAll(key, `.`, `\.`)
+	return key
+}
+
+// tolerantParseJSONObjectRaw attempts to parse a JSON-like object string into a JSON object string, tolerating
 // bareword values (unquoted strings) commonly seen during streamed tool calls.
 // Example input: {"location": 北京, "unit": celsius}
-func tolerantParseJSONMap(s string) map[string]interface{} {
+func tolerantParseJSONObjectRaw(s string) string {
 	// Ensure we operate within the outermost braces if present
 	start := strings.Index(s, "{")
 	end := strings.LastIndex(s, "}")
 	if start == -1 || end == -1 || start >= end {
-		return map[string]interface{}{}
+		return "{}"
 	}
 	content := s[start+1 : end]
 
 	runes := []rune(content)
 	n := len(runes)
 	i := 0
-	result := make(map[string]interface{})
+	result := "{}"
 
 	for i < n {
 		// Skip whitespace and commas
@@ -356,6 +336,7 @@ func tolerantParseJSONMap(s string) map[string]interface{} {
 			break
 		}
 		keyName := jsonStringTokenToRawString(keyToken)
+		sjsonKey := escapeSjsonPathKey(keyName)
 		i = nextIdx
 
 		// Skip whitespace
@@ -375,17 +356,16 @@ func tolerantParseJSONMap(s string) map[string]interface{} {
 		}
 
 		// Parse value (string, number, object/array, bareword)
-		var value interface{}
 		switch runes[i] {
 		case '"':
 			// JSON string
 			valToken, ni := parseJSONStringRunes(runes, i)
 			if ni == -1 {
 				// Malformed; treat as empty string
-				value = ""
+				result, _ = sjson.Set(result, sjsonKey, "")
 				i = n
 			} else {
-				value = jsonStringTokenToRawString(valToken)
+				result, _ = sjson.Set(result, sjsonKey, jsonStringTokenToRawString(valToken))
 				i = ni
 			}
 		case '{', '[':
@@ -394,11 +374,10 @@ func tolerantParseJSONMap(s string) map[string]interface{} {
 			if ni == -1 {
 				i = n
 			} else {
-				var anyVal interface{}
-				if errUnmarshal := json.Unmarshal([]byte(seg), &anyVal); errUnmarshal == nil {
-					value = anyVal
+				if gjson.Valid(seg) {
+					result, _ = sjson.SetRaw(result, sjsonKey, seg)
 				} else {
-					value = seg
+					result, _ = sjson.Set(result, sjsonKey, seg)
 				}
 				i = ni
 			}
@@ -411,21 +390,19 @@ func tolerantParseJSONMap(s string) map[string]interface{} {
 			token := strings.TrimSpace(string(runes[i:j]))
 			// Interpret common JSON atoms and numbers; otherwise treat as string
 			if token == "true" {
-				value = true
+				result, _ = sjson.Set(result, sjsonKey, true)
 			} else if token == "false" {
-				value = false
+				result, _ = sjson.Set(result, sjsonKey, false)
 			} else if token == "null" {
-				value = nil
+				result, _ = sjson.Set(result, sjsonKey, nil)
 			} else if numVal, ok := tryParseNumber(token); ok {
-				value = numVal
+				result, _ = sjson.Set(result, sjsonKey, numVal)
 			} else {
-				value = token
+				result, _ = sjson.Set(result, sjsonKey, token)
 			}
 			i = j
 		}
 
-		result[keyName] = value
-
 		// Skip trailing whitespace and optional comma before next pair
 		for i < n && (runes[i] == ' ' || runes[i] == '\n' || runes[i] == '\r' || runes[i] == '\t') {
 			i++
@@ -463,9 +440,9 @@ func parseJSONStringRunes(runes []rune, start int) (string, int) {
 
 // jsonStringTokenToRawString converts a JSON string token (including quotes) to a raw Go string value.
 func jsonStringTokenToRawString(token string) string {
-	var s string
-	if errUnmarshal := json.Unmarshal([]byte(token), &s); errUnmarshal == nil {
-		return s
+	r := gjson.Parse(token)
+	if r.Type == gjson.String {
+		return r.String()
 	}
 	// Fallback: strip surrounding quotes if present
 	if len(token) >= 2 && token[0] == '"' && token[len(token)-1] == '"' {
@@ -579,7 +556,7 @@ func ConvertOpenAIResponseToGeminiNonStream(_ context.Context, _ string, origina
 				}
 			}
 
-			var parts []interface{}
+			partIndex := 0
 
 			// Handle reasoning content before visible text
 			if reasoning := message.Get("reasoning_content"); reasoning.Exists() {
@@ -587,18 +564,16 @@ func ConvertOpenAIResponseToGeminiNonStream(_ context.Context, _ string, origina
 					if reasoningText == "" {
 						continue
 					}
-					parts = append(parts, map[string]interface{}{
-						"thought": true,
-						"text":    reasoningText,
-					})
+					out, _ = sjson.Set(out, fmt.Sprintf("candidates.0.content.parts.%d.thought", partIndex), true)
+					out, _ = sjson.Set(out, fmt.Sprintf("candidates.0.content.parts.%d.text", partIndex), reasoningText)
+					partIndex++
 				}
 			}
 
 			// Handle content first
 			if content := message.Get("content"); content.Exists() && content.String() != "" {
-				parts = append(parts, map[string]interface{}{
-					"text": content.String(),
-				})
+				out, _ = sjson.Set(out, fmt.Sprintf("candidates.0.content.parts.%d.text", partIndex), content.String())
+				partIndex++
 			}
 
 			// Handle tool calls
@@ -609,27 +584,16 @@ func ConvertOpenAIResponseToGeminiNonStream(_ context.Context, _ string, origina
 						functionName := function.Get("name").String()
 						functionArgs := function.Get("arguments").String()
 
-						// Parse arguments
-						var argsMap map[string]interface{}
-						argsMap = parseArgsToMap(functionArgs)
-
-						functionCallPart := map[string]interface{}{
-							"functionCall": map[string]interface{}{
-								"name": functionName,
-								"args": argsMap,
-							},
-						}
-						parts = append(parts, functionCallPart)
+						namePath := fmt.Sprintf("candidates.0.content.parts.%d.functionCall.name", partIndex)
+						argsPath := fmt.Sprintf("candidates.0.content.parts.%d.functionCall.args", partIndex)
+						out, _ = sjson.Set(out, namePath, functionName)
+						out, _ = sjson.SetRaw(out, argsPath, parseArgsToObjectRaw(functionArgs))
+						partIndex++
 					}
 					return true
 				})
 			}
 
-			// Set parts
-			if len(parts) > 0 {
-				out, _ = sjson.Set(out, "candidates.0.content.parts", parts)
-			}
-
 			// Handle finish reason
 			if finishReason := choice.Get("finish_reason"); finishReason.Exists() {
 				geminiFinishReason := mapOpenAIFinishReasonToGemini(finishReason.String())
@@ -645,15 +609,12 @@ func ConvertOpenAIResponseToGeminiNonStream(_ context.Context, _ string, origina
 
 	// Handle usage information
 	if usage := root.Get("usage"); usage.Exists() {
-		usageObj := map[string]interface{}{
-			"promptTokenCount":     usage.Get("prompt_tokens").Int(),
-			"candidatesTokenCount": usage.Get("completion_tokens").Int(),
-			"totalTokenCount":      usage.Get("total_tokens").Int(),
-		}
+		out, _ = sjson.Set(out, "usageMetadata.promptTokenCount", usage.Get("prompt_tokens").Int())
+		out, _ = sjson.Set(out, "usageMetadata.candidatesTokenCount", usage.Get("completion_tokens").Int())
+		out, _ = sjson.Set(out, "usageMetadata.totalTokenCount", usage.Get("total_tokens").Int())
 		if reasoningTokens := reasoningTokensFromUsage(usage); reasoningTokens > 0 {
-			usageObj["thoughtsTokenCount"] = reasoningTokens
+			out, _ = sjson.Set(out, "usageMetadata.thoughtsTokenCount", reasoningTokens)
 		}
-		out, _ = sjson.Set(out, "usageMetadata", usageObj)
 	}
 
 	return out

internal/translator/openai/openai/responses/openai_openai-responses_response.go

@@ -143,7 +143,7 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 		st.ReasoningTokens = 0
 		st.UsageSeen = false
 		// response.created
-		created := `{"type":"response.created","sequence_number":0,"response":{"id":"","object":"response","created_at":0,"status":"in_progress","background":false,"error":null}}`
+		created := `{"type":"response.created","sequence_number":0,"response":{"id":"","object":"response","created_at":0,"status":"in_progress","background":false,"error":null,"output":[]}}`
 		created, _ = sjson.Set(created, "sequence_number", nextSeq())
 		created, _ = sjson.Set(created, "response.id", st.ResponseID)
 		created, _ = sjson.Set(created, "response.created_at", st.Created)
@@ -216,11 +216,11 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 					}
 					// Append incremental text to reasoning buffer
 					st.ReasoningBuf.WriteString(rc.String())
-					msg := `{"type":"response.reasoning_summary_text.delta","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"text":""}`
+					msg := `{"type":"response.reasoning_summary_text.delta","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"delta":""}`
 					msg, _ = sjson.Set(msg, "sequence_number", nextSeq())
 					msg, _ = sjson.Set(msg, "item_id", st.ReasoningID)
 					msg, _ = sjson.Set(msg, "output_index", st.ReasoningIndex)
-					msg, _ = sjson.Set(msg, "text", rc.String())
+					msg, _ = sjson.Set(msg, "delta", rc.String())
 					out = append(out, emitRespEvent("response.reasoning_summary_text.delta", msg))
 				}
 
@@ -484,16 +484,12 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 					}
 				}
 				// Build response.output using aggregated buffers
-				var outputs []interface{}
+				outputsWrapper := `{"arr":[]}`
 				if st.ReasoningBuf.Len() > 0 {
-					outputs = append(outputs, map[string]interface{}{
-						"id":   st.ReasoningID,
-						"type": "reasoning",
-						"summary": []interface{}{map[string]interface{}{
-							"type": "summary_text",
-							"text": st.ReasoningBuf.String(),
-						}},
-					})
+					item := `{"id":"","type":"reasoning","summary":[{"type":"summary_text","text":""}]}`
+					item, _ = sjson.Set(item, "id", st.ReasoningID)
+					item, _ = sjson.Set(item, "summary.0.text", st.ReasoningBuf.String())
+					outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 				}
 				// Append message items in ascending index order
 				if len(st.MsgItemAdded) > 0 {
@@ -513,18 +509,10 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 						if b := st.MsgTextBuf[i]; b != nil {
 							txt = b.String()
 						}
-						outputs = append(outputs, map[string]interface{}{
-							"id":     fmt.Sprintf("msg_%s_%d", st.ResponseID, i),
-							"type":   "message",
-							"status": "completed",
-							"content": []interface{}{map[string]interface{}{
-								"type":        "output_text",
-								"annotations": []interface{}{},
-								"logprobs":    []interface{}{},
-								"text":        txt,
-							}},
-							"role": "assistant",
-						})
+						item := `{"id":"","type":"message","status":"completed","content":[{"type":"output_text","annotations":[],"logprobs":[],"text":""}],"role":"assistant"}`
+						item, _ = sjson.Set(item, "id", fmt.Sprintf("msg_%s_%d", st.ResponseID, i))
+						item, _ = sjson.Set(item, "content.0.text", txt)
+						outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 					}
 				}
 				if len(st.FuncArgsBuf) > 0 {
@@ -547,18 +535,16 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 						}
 						callID := st.FuncCallIDs[i]
 						name := st.FuncNames[i]
-						outputs = append(outputs, map[string]interface{}{
-							"id":        fmt.Sprintf("fc_%s", callID),
-							"type":      "function_call",
-							"status":    "completed",
-							"arguments": args,
-							"call_id":   callID,
-							"name":      name,
-						})
+						item := `{"id":"","type":"function_call","status":"completed","arguments":"","call_id":"","name":""}`
+						item, _ = sjson.Set(item, "id", fmt.Sprintf("fc_%s", callID))
+						item, _ = sjson.Set(item, "arguments", args)
+						item, _ = sjson.Set(item, "call_id", callID)
+						item, _ = sjson.Set(item, "name", name)
+						outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 					}
 				}
-				if len(outputs) > 0 {
-					completed, _ = sjson.Set(completed, "response.output", outputs)
+				if gjson.Get(outputsWrapper, "arr.#").Int() > 0 {
+					completed, _ = sjson.SetRaw(completed, "response.output", gjson.Get(outputsWrapper, "arr").Raw)
 				}
 				if st.UsageSeen {
 					completed, _ = sjson.Set(completed, "response.usage.input_tokens", st.PromptTokens)
@@ -681,7 +667,7 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponsesNonStream(_ context.Co
 	}
 
 	// Build output list from choices[...]
-	var outputs []interface{}
+	outputsWrapper := `{"arr":[]}`
 	// Detect and capture reasoning content if present
 	rcText := gjson.GetBytes(rawJSON, "choices.0.message.reasoning_content").String()
 	includeReasoning := rcText != ""
@@ -693,21 +679,14 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponsesNonStream(_ context.Co
 		if strings.HasPrefix(rid, "resp_") {
 			rid = strings.TrimPrefix(rid, "resp_")
 		}
-		reasoningItem := map[string]interface{}{
-			"id":                fmt.Sprintf("rs_%s", rid),
-			"type":              "reasoning",
-			"encrypted_content": "",
-		}
 		// Prefer summary_text from reasoning_content; encrypted_content is optional
-		var summaries []interface{}
+		reasoningItem := `{"id":"","type":"reasoning","encrypted_content":"","summary":[]}`
+		reasoningItem, _ = sjson.Set(reasoningItem, "id", fmt.Sprintf("rs_%s", rid))
 		if rcText != "" {
-			summaries = append(summaries, map[string]interface{}{
-				"type": "summary_text",
-				"text": rcText,
-			})
+			reasoningItem, _ = sjson.Set(reasoningItem, "summary.0.type", "summary_text")
+			reasoningItem, _ = sjson.Set(reasoningItem, "summary.0.text", rcText)
 		}
-		reasoningItem["summary"] = summaries
-		outputs = append(outputs, reasoningItem)
+		outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", reasoningItem)
 	}
 
 	if choices := root.Get("choices"); choices.Exists() && choices.IsArray() {
@@ -716,18 +695,10 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponsesNonStream(_ context.Co
 			if msg.Exists() {
 				// Text message part
 				if c := msg.Get("content"); c.Exists() && c.String() != "" {
-					outputs = append(outputs, map[string]interface{}{
-						"id":     fmt.Sprintf("msg_%s_%d", id, int(choice.Get("index").Int())),
-						"type":   "message",
-						"status": "completed",
-						"content": []interface{}{map[string]interface{}{
-							"type":        "output_text",
-							"annotations": []interface{}{},
-							"logprobs":    []interface{}{},
-							"text":        c.String(),
-						}},
-						"role": "assistant",
-					})
+					item := `{"id":"","type":"message","status":"completed","content":[{"type":"output_text","annotations":[],"logprobs":[],"text":""}],"role":"assistant"}`
+					item, _ = sjson.Set(item, "id", fmt.Sprintf("msg_%s_%d", id, int(choice.Get("index").Int())))
+					item, _ = sjson.Set(item, "content.0.text", c.String())
+					outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 				}
 
 				// Function/tool calls
@@ -736,14 +707,12 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponsesNonStream(_ context.Co
 						callID := tc.Get("id").String()
 						name := tc.Get("function.name").String()
 						args := tc.Get("function.arguments").String()
-						outputs = append(outputs, map[string]interface{}{
-							"id":        fmt.Sprintf("fc_%s", callID),
-							"type":      "function_call",
-							"status":    "completed",
-							"arguments": args,
-							"call_id":   callID,
-							"name":      name,
-						})
+						item := `{"id":"","type":"function_call","status":"completed","arguments":"","call_id":"","name":""}`
+						item, _ = sjson.Set(item, "id", fmt.Sprintf("fc_%s", callID))
+						item, _ = sjson.Set(item, "arguments", args)
+						item, _ = sjson.Set(item, "call_id", callID)
+						item, _ = sjson.Set(item, "name", name)
+						outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
 						return true
 					})
 				}
@@ -751,8 +720,8 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponsesNonStream(_ context.Co
 			return true
 		})
 	}
-	if len(outputs) > 0 {
-		resp, _ = sjson.Set(resp, "output", outputs)
+	if gjson.Get(outputsWrapper, "arr.#").Int() > 0 {
+		resp, _ = sjson.SetRaw(resp, "output", gjson.Get(outputsWrapper, "arr").Raw)
 	}
 
 	// usage mapping

internal/usage/logger_plugin.go

@@ -6,6 +6,7 @@ package usage
 import (
 	"context"
 	"fmt"
+	"strings"
 	"sync"
 	"sync/atomic"
 	"time"
@@ -90,7 +91,7 @@ type modelStats struct {
 type RequestDetail struct {
 	Timestamp time.Time  `json:"timestamp"`
 	Source    string     `json:"source"`
-	AuthIndex uint64     `json:"auth_index"`
+	AuthIndex string     `json:"auth_index"`
 	Tokens    TokenStats `json:"tokens"`
 	Failed    bool       `json:"failed"`
 }
@@ -281,6 +282,118 @@ func (s *RequestStatistics) Snapshot() StatisticsSnapshot {
 	return result
 }
 
+type MergeResult struct {
+	Added   int64 `json:"added"`
+	Skipped int64 `json:"skipped"`
+}
+
+// MergeSnapshot merges an exported statistics snapshot into the current store.
+// Existing data is preserved and duplicate request details are skipped.
+func (s *RequestStatistics) MergeSnapshot(snapshot StatisticsSnapshot) MergeResult {
+	result := MergeResult{}
+	if s == nil {
+		return result
+	}
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	seen := make(map[string]struct{})
+	for apiName, stats := range s.apis {
+		if stats == nil {
+			continue
+		}
+		for modelName, modelStatsValue := range stats.Models {
+			if modelStatsValue == nil {
+				continue
+			}
+			for _, detail := range modelStatsValue.Details {
+				seen[dedupKey(apiName, modelName, detail)] = struct{}{}
+			}
+		}
+	}
+
+	for apiName, apiSnapshot := range snapshot.APIs {
+		apiName = strings.TrimSpace(apiName)
+		if apiName == "" {
+			continue
+		}
+		stats, ok := s.apis[apiName]
+		if !ok || stats == nil {
+			stats = &apiStats{Models: make(map[string]*modelStats)}
+			s.apis[apiName] = stats
+		} else if stats.Models == nil {
+			stats.Models = make(map[string]*modelStats)
+		}
+		for modelName, modelSnapshot := range apiSnapshot.Models {
+			modelName = strings.TrimSpace(modelName)
+			if modelName == "" {
+				modelName = "unknown"
+			}
+			for _, detail := range modelSnapshot.Details {
+				detail.Tokens = normaliseTokenStats(detail.Tokens)
+				if detail.Timestamp.IsZero() {
+					detail.Timestamp = time.Now()
+				}
+				key := dedupKey(apiName, modelName, detail)
+				if _, exists := seen[key]; exists {
+					result.Skipped++
+					continue
+				}
+				seen[key] = struct{}{}
+				s.recordImported(apiName, modelName, stats, detail)
+				result.Added++
+			}
+		}
+	}
+
+	return result
+}
+
+func (s *RequestStatistics) recordImported(apiName, modelName string, stats *apiStats, detail RequestDetail) {
+	totalTokens := detail.Tokens.TotalTokens
+	if totalTokens < 0 {
+		totalTokens = 0
+	}
+
+	s.totalRequests++
+	if detail.Failed {
+		s.failureCount++
+	} else {
+		s.successCount++
+	}
+	s.totalTokens += totalTokens
+
+	s.updateAPIStats(stats, modelName, detail)
+
+	dayKey := detail.Timestamp.Format("2006-01-02")
+	hourKey := detail.Timestamp.Hour()
+
+	s.requestsByDay[dayKey]++
+	s.requestsByHour[hourKey]++
+	s.tokensByDay[dayKey] += totalTokens
+	s.tokensByHour[hourKey] += totalTokens
+}
+
+func dedupKey(apiName, modelName string, detail RequestDetail) string {
+	timestamp := detail.Timestamp.UTC().Format(time.RFC3339Nano)
+	tokens := normaliseTokenStats(detail.Tokens)
+	return fmt.Sprintf(
+		"%s|%s|%s|%s|%s|%t|%d|%d|%d|%d|%d",
+		apiName,
+		modelName,
+		timestamp,
+		detail.Source,
+		detail.AuthIndex,
+		detail.Failed,
+		tokens.InputTokens,
+		tokens.OutputTokens,
+		tokens.ReasoningTokens,
+		tokens.CachedTokens,
+		tokens.TotalTokens,
+	)
+}
+
 func resolveAPIIdentifier(ctx context.Context, record coreusage.Record) string {
 	if ctx != nil {
 		if ginCtx, ok := ctx.Value("gin").(*gin.Context); ok && ginCtx != nil {
@@ -340,6 +453,16 @@ func normaliseDetail(detail coreusage.Detail) TokenStats {
 	return tokens
 }
 
+func normaliseTokenStats(tokens TokenStats) TokenStats {
+	if tokens.TotalTokens == 0 {
+		tokens.TotalTokens = tokens.InputTokens + tokens.OutputTokens + tokens.ReasoningTokens
+	}
+	if tokens.TotalTokens == 0 {
+		tokens.TotalTokens = tokens.InputTokens + tokens.OutputTokens + tokens.ReasoningTokens + tokens.CachedTokens
+	}
+	return tokens
+}
+
 func formatHour(hour int) string {
 	if hour < 0 {
 		hour = 0

internal/util/claude_model.go

@@ -0,0 +1,10 @@
+package util
+
+import "strings"
+
+// IsClaudeThinkingModel checks if the model is a Claude thinking model
+// that requires the interleaved-thinking beta header.
+func IsClaudeThinkingModel(model string) bool {
+	lower := strings.ToLower(model)
+	return strings.Contains(lower, "claude") && strings.Contains(lower, "thinking")
+}

internal/util/claude_model_test.go

@@ -0,0 +1,41 @@
+package util
+
+import "testing"
+
+func TestIsClaudeThinkingModel(t *testing.T) {
+	tests := []struct {
+		name     string
+		model    string
+		expected bool
+	}{
+		// Claude thinking models - should return true
+		{"claude-sonnet-4-5-thinking", "claude-sonnet-4-5-thinking", true},
+		{"claude-opus-4-5-thinking", "claude-opus-4-5-thinking", true},
+		{"Claude-Sonnet-Thinking uppercase", "Claude-Sonnet-4-5-Thinking", true},
+		{"claude thinking mixed case", "Claude-THINKING-Model", true},
+
+		// Non-thinking Claude models - should return false
+		{"claude-sonnet-4-5 (no thinking)", "claude-sonnet-4-5", false},
+		{"claude-opus-4-5 (no thinking)", "claude-opus-4-5", false},
+		{"claude-3-5-sonnet", "claude-3-5-sonnet-20240620", false},
+
+		// Non-Claude models - should return false
+		{"gemini-3-pro-preview", "gemini-3-pro-preview", false},
+		{"gemini-thinking model", "gemini-3-pro-thinking", false}, // not Claude
+		{"gpt-4o", "gpt-4o", false},
+		{"empty string", "", false},
+
+		// Edge cases
+		{"thinking without claude", "thinking-model", false},
+		{"claude without thinking", "claude-model", false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := IsClaudeThinkingModel(tt.model)
+			if result != tt.expected {
+				t.Errorf("IsClaudeThinkingModel(%q) = %v, expected %v", tt.model, result, tt.expected)
+			}
+		})
+	}
+}

internal/util/gemini_schema.go

@@ -0,0 +1,547 @@
+// Package util provides utility functions for the CLI Proxy API server.
+package util
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+var gjsonPathKeyReplacer = strings.NewReplacer(".", "\\.", "*", "\\*", "?", "\\?")
+
+// CleanJSONSchemaForAntigravity transforms a JSON schema to be compatible with Antigravity API.
+// It handles unsupported keywords, type flattening, and schema simplification while preserving
+// semantic information as description hints.
+func CleanJSONSchemaForAntigravity(jsonStr string) string {
+	// Phase 1: Convert and add hints
+	jsonStr = convertRefsToHints(jsonStr)
+	jsonStr = convertConstToEnum(jsonStr)
+	jsonStr = addEnumHints(jsonStr)
+	jsonStr = addAdditionalPropertiesHints(jsonStr)
+	jsonStr = moveConstraintsToDescription(jsonStr)
+
+	// Phase 2: Flatten complex structures
+	jsonStr = mergeAllOf(jsonStr)
+	jsonStr = flattenAnyOfOneOf(jsonStr)
+	jsonStr = flattenTypeArrays(jsonStr)
+
+	// Phase 3: Cleanup
+	jsonStr = removeUnsupportedKeywords(jsonStr)
+	jsonStr = cleanupRequiredFields(jsonStr)
+
+	// Phase 4: Add placeholder for empty object schemas (Claude VALIDATED mode requirement)
+	jsonStr = addEmptySchemaPlaceholder(jsonStr)
+
+	return jsonStr
+}
+
+// convertRefsToHints converts $ref to description hints (Lazy Hint strategy).
+func convertRefsToHints(jsonStr string) string {
+	paths := findPaths(jsonStr, "$ref")
+	sortByDepth(paths)
+
+	for _, p := range paths {
+		refVal := gjson.Get(jsonStr, p).String()
+		defName := refVal
+		if idx := strings.LastIndex(refVal, "/"); idx >= 0 {
+			defName = refVal[idx+1:]
+		}
+
+		parentPath := trimSuffix(p, ".$ref")
+		hint := fmt.Sprintf("See: %s", defName)
+		if existing := gjson.Get(jsonStr, descriptionPath(parentPath)).String(); existing != "" {
+			hint = fmt.Sprintf("%s (%s)", existing, hint)
+		}
+
+		replacement := `{"type":"object","description":""}`
+		replacement, _ = sjson.Set(replacement, "description", hint)
+		jsonStr = setRawAt(jsonStr, parentPath, replacement)
+	}
+	return jsonStr
+}
+
+func convertConstToEnum(jsonStr string) string {
+	for _, p := range findPaths(jsonStr, "const") {
+		val := gjson.Get(jsonStr, p)
+		if !val.Exists() {
+			continue
+		}
+		enumPath := trimSuffix(p, ".const") + ".enum"
+		if !gjson.Get(jsonStr, enumPath).Exists() {
+			jsonStr, _ = sjson.Set(jsonStr, enumPath, []interface{}{val.Value()})
+		}
+	}
+	return jsonStr
+}
+
+func addEnumHints(jsonStr string) string {
+	for _, p := range findPaths(jsonStr, "enum") {
+		arr := gjson.Get(jsonStr, p)
+		if !arr.IsArray() {
+			continue
+		}
+		items := arr.Array()
+		if len(items) <= 1 || len(items) > 10 {
+			continue
+		}
+
+		var vals []string
+		for _, item := range items {
+			vals = append(vals, item.String())
+		}
+		jsonStr = appendHint(jsonStr, trimSuffix(p, ".enum"), "Allowed: "+strings.Join(vals, ", "))
+	}
+	return jsonStr
+}
+
+func addAdditionalPropertiesHints(jsonStr string) string {
+	for _, p := range findPaths(jsonStr, "additionalProperties") {
+		if gjson.Get(jsonStr, p).Type == gjson.False {
+			jsonStr = appendHint(jsonStr, trimSuffix(p, ".additionalProperties"), "No extra properties allowed")
+		}
+	}
+	return jsonStr
+}
+
+var unsupportedConstraints = []string{
+	"minLength", "maxLength", "exclusiveMinimum", "exclusiveMaximum",
+	"pattern", "minItems", "maxItems", "format",
+	"default", "examples", // Claude rejects these in VALIDATED mode
+}
+
+func moveConstraintsToDescription(jsonStr string) string {
+	for _, key := range unsupportedConstraints {
+		for _, p := range findPaths(jsonStr, key) {
+			val := gjson.Get(jsonStr, p)
+			if !val.Exists() || val.IsObject() || val.IsArray() {
+				continue
+			}
+			parentPath := trimSuffix(p, "."+key)
+			if isPropertyDefinition(parentPath) {
+				continue
+			}
+			jsonStr = appendHint(jsonStr, parentPath, fmt.Sprintf("%s: %s", key, val.String()))
+		}
+	}
+	return jsonStr
+}
+
+func mergeAllOf(jsonStr string) string {
+	paths := findPaths(jsonStr, "allOf")
+	sortByDepth(paths)
+
+	for _, p := range paths {
+		allOf := gjson.Get(jsonStr, p)
+		if !allOf.IsArray() {
+			continue
+		}
+		parentPath := trimSuffix(p, ".allOf")
+
+		for _, item := range allOf.Array() {
+			if props := item.Get("properties"); props.IsObject() {
+				props.ForEach(func(key, value gjson.Result) bool {
+					destPath := joinPath(parentPath, "properties."+escapeGJSONPathKey(key.String()))
+					jsonStr, _ = sjson.SetRaw(jsonStr, destPath, value.Raw)
+					return true
+				})
+			}
+			if req := item.Get("required"); req.IsArray() {
+				reqPath := joinPath(parentPath, "required")
+				current := getStrings(jsonStr, reqPath)
+				for _, r := range req.Array() {
+					if s := r.String(); !contains(current, s) {
+						current = append(current, s)
+					}
+				}
+				jsonStr, _ = sjson.Set(jsonStr, reqPath, current)
+			}
+		}
+		jsonStr, _ = sjson.Delete(jsonStr, p)
+	}
+	return jsonStr
+}
+
+func flattenAnyOfOneOf(jsonStr string) string {
+	for _, key := range []string{"anyOf", "oneOf"} {
+		paths := findPaths(jsonStr, key)
+		sortByDepth(paths)
+
+		for _, p := range paths {
+			arr := gjson.Get(jsonStr, p)
+			if !arr.IsArray() || len(arr.Array()) == 0 {
+				continue
+			}
+
+			parentPath := trimSuffix(p, "."+key)
+			parentDesc := gjson.Get(jsonStr, descriptionPath(parentPath)).String()
+
+			items := arr.Array()
+			bestIdx, allTypes := selectBest(items)
+			selected := items[bestIdx].Raw
+
+			if parentDesc != "" {
+				selected = mergeDescriptionRaw(selected, parentDesc)
+			}
+
+			if len(allTypes) > 1 {
+				hint := "Accepts: " + strings.Join(allTypes, " | ")
+				selected = appendHintRaw(selected, hint)
+			}
+
+			jsonStr = setRawAt(jsonStr, parentPath, selected)
+		}
+	}
+	return jsonStr
+}
+
+func selectBest(items []gjson.Result) (bestIdx int, types []string) {
+	bestScore := -1
+	for i, item := range items {
+		t := item.Get("type").String()
+		score := 0
+
+		switch {
+		case t == "object" || item.Get("properties").Exists():
+			score, t = 3, orDefault(t, "object")
+		case t == "array" || item.Get("items").Exists():
+			score, t = 2, orDefault(t, "array")
+		case t != "" && t != "null":
+			score = 1
+		default:
+			t = orDefault(t, "null")
+		}
+
+		if t != "" {
+			types = append(types, t)
+		}
+		if score > bestScore {
+			bestScore, bestIdx = score, i
+		}
+	}
+	return
+}
+
+func flattenTypeArrays(jsonStr string) string {
+	paths := findPaths(jsonStr, "type")
+	sortByDepth(paths)
+
+	nullableFields := make(map[string][]string)
+
+	for _, p := range paths {
+		res := gjson.Get(jsonStr, p)
+		if !res.IsArray() || len(res.Array()) == 0 {
+			continue
+		}
+
+		hasNull := false
+		var nonNullTypes []string
+		for _, item := range res.Array() {
+			s := item.String()
+			if s == "null" {
+				hasNull = true
+			} else if s != "" {
+				nonNullTypes = append(nonNullTypes, s)
+			}
+		}
+
+		firstType := "string"
+		if len(nonNullTypes) > 0 {
+			firstType = nonNullTypes[0]
+		}
+
+		jsonStr, _ = sjson.Set(jsonStr, p, firstType)
+
+		parentPath := trimSuffix(p, ".type")
+		if len(nonNullTypes) > 1 {
+			hint := "Accepts: " + strings.Join(nonNullTypes, " | ")
+			jsonStr = appendHint(jsonStr, parentPath, hint)
+		}
+
+		if hasNull {
+			parts := splitGJSONPath(p)
+			if len(parts) >= 3 && parts[len(parts)-3] == "properties" {
+				fieldNameEscaped := parts[len(parts)-2]
+				fieldName := unescapeGJSONPathKey(fieldNameEscaped)
+				objectPath := strings.Join(parts[:len(parts)-3], ".")
+				nullableFields[objectPath] = append(nullableFields[objectPath], fieldName)
+
+				propPath := joinPath(objectPath, "properties."+fieldNameEscaped)
+				jsonStr = appendHint(jsonStr, propPath, "(nullable)")
+			}
+		}
+	}
+
+	for objectPath, fields := range nullableFields {
+		reqPath := joinPath(objectPath, "required")
+		req := gjson.Get(jsonStr, reqPath)
+		if !req.IsArray() {
+			continue
+		}
+
+		var filtered []string
+		for _, r := range req.Array() {
+			if !contains(fields, r.String()) {
+				filtered = append(filtered, r.String())
+			}
+		}
+
+		if len(filtered) == 0 {
+			jsonStr, _ = sjson.Delete(jsonStr, reqPath)
+		} else {
+			jsonStr, _ = sjson.Set(jsonStr, reqPath, filtered)
+		}
+	}
+	return jsonStr
+}
+
+func removeUnsupportedKeywords(jsonStr string) string {
+	keywords := append(unsupportedConstraints,
+		"$schema", "$defs", "definitions", "const", "$ref", "additionalProperties",
+		"propertyNames", // Gemini doesn't support property name validation
+	)
+	for _, key := range keywords {
+		for _, p := range findPaths(jsonStr, key) {
+			if isPropertyDefinition(trimSuffix(p, "."+key)) {
+				continue
+			}
+			jsonStr, _ = sjson.Delete(jsonStr, p)
+		}
+	}
+	return jsonStr
+}
+
+func cleanupRequiredFields(jsonStr string) string {
+	for _, p := range findPaths(jsonStr, "required") {
+		parentPath := trimSuffix(p, ".required")
+		propsPath := joinPath(parentPath, "properties")
+
+		req := gjson.Get(jsonStr, p)
+		props := gjson.Get(jsonStr, propsPath)
+		if !req.IsArray() || !props.IsObject() {
+			continue
+		}
+
+		var valid []string
+		for _, r := range req.Array() {
+			key := r.String()
+			if props.Get(escapeGJSONPathKey(key)).Exists() {
+				valid = append(valid, key)
+			}
+		}
+
+		if len(valid) != len(req.Array()) {
+			if len(valid) == 0 {
+				jsonStr, _ = sjson.Delete(jsonStr, p)
+			} else {
+				jsonStr, _ = sjson.Set(jsonStr, p, valid)
+			}
+		}
+	}
+	return jsonStr
+}
+
+// addEmptySchemaPlaceholder adds a placeholder "reason" property to empty object schemas.
+// Claude VALIDATED mode requires at least one property in tool schemas.
+func addEmptySchemaPlaceholder(jsonStr string) string {
+	// Find all "type" fields
+	paths := findPaths(jsonStr, "type")
+
+	// Process from deepest to shallowest (to handle nested objects properly)
+	sortByDepth(paths)
+
+	for _, p := range paths {
+		typeVal := gjson.Get(jsonStr, p)
+		if typeVal.String() != "object" {
+			continue
+		}
+
+		// Get the parent path (the object containing "type")
+		parentPath := trimSuffix(p, ".type")
+
+		// Check if properties exists and is empty or missing
+		propsPath := joinPath(parentPath, "properties")
+		propsVal := gjson.Get(jsonStr, propsPath)
+
+		needsPlaceholder := false
+		if !propsVal.Exists() {
+			// No properties field at all
+			needsPlaceholder = true
+		} else if propsVal.IsObject() && len(propsVal.Map()) == 0 {
+			// Empty properties object
+			needsPlaceholder = true
+		}
+
+		if needsPlaceholder {
+			// Add placeholder "reason" property
+			reasonPath := joinPath(propsPath, "reason")
+			jsonStr, _ = sjson.Set(jsonStr, reasonPath+".type", "string")
+			jsonStr, _ = sjson.Set(jsonStr, reasonPath+".description", "Brief explanation of why you are calling this tool")
+
+			// Add to required array
+			reqPath := joinPath(parentPath, "required")
+			jsonStr, _ = sjson.Set(jsonStr, reqPath, []string{"reason"})
+		}
+	}
+
+	return jsonStr
+}
+
+// --- Helpers ---
+
+func findPaths(jsonStr, field string) []string {
+	var paths []string
+	Walk(gjson.Parse(jsonStr), "", field, &paths)
+	return paths
+}
+
+func sortByDepth(paths []string) {
+	sort.Slice(paths, func(i, j int) bool { return len(paths[i]) > len(paths[j]) })
+}
+
+func trimSuffix(path, suffix string) string {
+	if path == strings.TrimPrefix(suffix, ".") {
+		return ""
+	}
+	return strings.TrimSuffix(path, suffix)
+}
+
+func joinPath(base, suffix string) string {
+	if base == "" {
+		return suffix
+	}
+	return base + "." + suffix
+}
+
+func setRawAt(jsonStr, path, value string) string {
+	if path == "" {
+		return value
+	}
+	result, _ := sjson.SetRaw(jsonStr, path, value)
+	return result
+}
+
+func isPropertyDefinition(path string) bool {
+	return path == "properties" || strings.HasSuffix(path, ".properties")
+}
+
+func descriptionPath(parentPath string) string {
+	if parentPath == "" || parentPath == "@this" {
+		return "description"
+	}
+	return parentPath + ".description"
+}
+
+func appendHint(jsonStr, parentPath, hint string) string {
+	descPath := parentPath + ".description"
+	if parentPath == "" || parentPath == "@this" {
+		descPath = "description"
+	}
+	existing := gjson.Get(jsonStr, descPath).String()
+	if existing != "" {
+		hint = fmt.Sprintf("%s (%s)", existing, hint)
+	}
+	jsonStr, _ = sjson.Set(jsonStr, descPath, hint)
+	return jsonStr
+}
+
+func appendHintRaw(jsonRaw, hint string) string {
+	existing := gjson.Get(jsonRaw, "description").String()
+	if existing != "" {
+		hint = fmt.Sprintf("%s (%s)", existing, hint)
+	}
+	jsonRaw, _ = sjson.Set(jsonRaw, "description", hint)
+	return jsonRaw
+}
+
+func getStrings(jsonStr, path string) []string {
+	var result []string
+	if arr := gjson.Get(jsonStr, path); arr.IsArray() {
+		for _, r := range arr.Array() {
+			result = append(result, r.String())
+		}
+	}
+	return result
+}
+
+func contains(slice []string, item string) bool {
+	for _, s := range slice {
+		if s == item {
+			return true
+		}
+	}
+	return false
+}
+
+func orDefault(val, def string) string {
+	if val == "" {
+		return def
+	}
+	return val
+}
+
+func escapeGJSONPathKey(key string) string {
+	return gjsonPathKeyReplacer.Replace(key)
+}
+
+func unescapeGJSONPathKey(key string) string {
+	if !strings.Contains(key, "\\") {
+		return key
+	}
+	var b strings.Builder
+	b.Grow(len(key))
+	for i := 0; i < len(key); i++ {
+		if key[i] == '\\' && i+1 < len(key) {
+			i++
+			b.WriteByte(key[i])
+			continue
+		}
+		b.WriteByte(key[i])
+	}
+	return b.String()
+}
+
+func splitGJSONPath(path string) []string {
+	if path == "" {
+		return nil
+	}
+
+	parts := make([]string, 0, strings.Count(path, ".")+1)
+	var b strings.Builder
+	b.Grow(len(path))
+
+	for i := 0; i < len(path); i++ {
+		c := path[i]
+		if c == '\\' && i+1 < len(path) {
+			b.WriteByte('\\')
+			i++
+			b.WriteByte(path[i])
+			continue
+		}
+		if c == '.' {
+			parts = append(parts, b.String())
+			b.Reset()
+			continue
+		}
+		b.WriteByte(c)
+	}
+	parts = append(parts, b.String())
+	return parts
+}
+
+func mergeDescriptionRaw(schemaRaw, parentDesc string) string {
+	childDesc := gjson.Get(schemaRaw, "description").String()
+	switch {
+	case childDesc == "":
+		schemaRaw, _ = sjson.Set(schemaRaw, "description", parentDesc)
+		return schemaRaw
+	case childDesc == parentDesc:
+		return schemaRaw
+	default:
+		combined := fmt.Sprintf("%s (%s)", parentDesc, childDesc)
+		schemaRaw, _ = sjson.Set(schemaRaw, "description", combined)
+		return schemaRaw
+	}
+}

internal/util/gemini_schema_test.go

@@ -0,0 +1,883 @@
+package util
+
+import (
+	"encoding/json"
+	"reflect"
+	"strings"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestCleanJSONSchemaForAntigravity_ConstToEnum(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"kind": {
+				"type": "string",
+				"const": "InsightVizNode"
+			}
+		}
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"kind": {
+				"type": "string",
+				"enum": ["InsightVizNode"]
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+}
+
+func TestCleanJSONSchemaForAntigravity_TypeFlattening_Nullable(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"name": {
+				"type": ["string", "null"]
+			},
+			"other": {
+				"type": "string"
+			}
+		},
+		"required": ["name", "other"]
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"name": {
+				"type": "string",
+				"description": "(nullable)"
+			},
+			"other": {
+				"type": "string"
+			}
+		},
+		"required": ["other"]
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+}
+
+func TestCleanJSONSchemaForAntigravity_ConstraintsToDescription(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"tags": {
+				"type": "array",
+				"description": "List of tags",
+				"minItems": 1
+			},
+			"name": {
+				"type": "string",
+				"description": "User name",
+				"minLength": 3
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// minItems should be REMOVED and moved to description
+	if strings.Contains(result, `"minItems"`) {
+		t.Errorf("minItems keyword should be removed")
+	}
+	if !strings.Contains(result, "minItems: 1") {
+		t.Errorf("minItems hint missing in description")
+	}
+
+	// minLength should be moved to description
+	if !strings.Contains(result, "minLength: 3") {
+		t.Errorf("minLength hint missing in description")
+	}
+	if strings.Contains(result, `"minLength":`) || strings.Contains(result, `"minLength" :`) {
+		t.Errorf("minLength keyword should be removed")
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_AnyOfFlattening_SmartSelection(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"query": {
+				"anyOf": [
+					{ "type": "null" },
+					{
+						"type": "object",
+						"properties": {
+							"kind": { "type": "string" }
+						}
+					}
+				]
+			}
+		}
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"query": {
+				"type": "object",
+				"description": "Accepts: null | object",
+				"properties": {
+					"kind": { "type": "string" }
+				}
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+}
+
+func TestCleanJSONSchemaForAntigravity_OneOfFlattening(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"config": {
+				"oneOf": [
+					{ "type": "string" },
+					{ "type": "integer" }
+				]
+			}
+		}
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"config": {
+				"type": "string",
+				"description": "Accepts: string | integer"
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+}
+
+func TestCleanJSONSchemaForAntigravity_AllOfMerging(t *testing.T) {
+	input := `{
+		"type": "object",
+		"allOf": [
+			{
+				"properties": {
+					"a": { "type": "string" }
+				},
+				"required": ["a"]
+			},
+			{
+				"properties": {
+					"b": { "type": "integer" }
+				},
+				"required": ["b"]
+			}
+		]
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"a": { "type": "string" },
+			"b": { "type": "integer" }
+		},
+		"required": ["a", "b"]
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+}
+
+func TestCleanJSONSchemaForAntigravity_RefHandling(t *testing.T) {
+	input := `{
+		"definitions": {
+			"User": {
+				"type": "object",
+				"properties": {
+					"name": { "type": "string" }
+				}
+			}
+		},
+		"type": "object",
+		"properties": {
+			"customer": { "$ref": "#/definitions/User" }
+		}
+	}`
+
+	// After $ref is converted to placeholder object, empty schema placeholder is also added
+	expected := `{
+		"type": "object",
+		"properties": {
+			"customer": {
+				"type": "object",
+				"description": "See: User",
+				"properties": {
+					"reason": {
+						"type": "string",
+						"description": "Brief explanation of why you are calling this tool"
+					}
+				},
+				"required": ["reason"]
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+}
+
+func TestCleanJSONSchemaForAntigravity_RefHandling_DescriptionEscaping(t *testing.T) {
+	input := `{
+		"definitions": {
+			"User": {
+				"type": "object",
+				"properties": {
+					"name": { "type": "string" }
+				}
+			}
+		},
+		"type": "object",
+		"properties": {
+			"customer": {
+				"description": "He said \"hi\"\\nsecond line",
+				"$ref": "#/definitions/User"
+			}
+		}
+	}`
+
+	// After $ref is converted, empty schema placeholder is also added
+	expected := `{
+		"type": "object",
+		"properties": {
+			"customer": {
+				"type": "object",
+				"description": "He said \"hi\"\\nsecond line (See: User)",
+				"properties": {
+					"reason": {
+						"type": "string",
+						"description": "Brief explanation of why you are calling this tool"
+					}
+				},
+				"required": ["reason"]
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+}
+
+func TestCleanJSONSchemaForAntigravity_CyclicRefDefaults(t *testing.T) {
+	input := `{
+		"definitions": {
+			"Node": {
+				"type": "object",
+				"properties": {
+					"child": { "$ref": "#/definitions/Node" }
+				}
+			}
+		},
+		"$ref": "#/definitions/Node"
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	var resMap map[string]interface{}
+	json.Unmarshal([]byte(result), &resMap)
+
+	if resMap["type"] != "object" {
+		t.Errorf("Expected type: object, got: %v", resMap["type"])
+	}
+
+	desc, ok := resMap["description"].(string)
+	if !ok || !strings.Contains(desc, "Node") {
+		t.Errorf("Expected description hint containing 'Node', got: %v", resMap["description"])
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_RequiredCleanup(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"a": {"type": "string"},
+			"b": {"type": "string"}
+		},
+		"required": ["a", "b", "c"]
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"a": {"type": "string"},
+			"b": {"type": "string"}
+		},
+		"required": ["a", "b"]
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+}
+
+func TestCleanJSONSchemaForAntigravity_AllOfMerging_DotKeys(t *testing.T) {
+	input := `{
+		"type": "object",
+		"allOf": [
+			{
+				"properties": {
+					"my.param": { "type": "string" }
+				},
+				"required": ["my.param"]
+			},
+			{
+				"properties": {
+					"b": { "type": "integer" }
+				},
+				"required": ["b"]
+			}
+		]
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"my.param": { "type": "string" },
+			"b": { "type": "integer" }
+		},
+		"required": ["my.param", "b"]
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+}
+
+func TestCleanJSONSchemaForAntigravity_PropertyNameCollision(t *testing.T) {
+	// A tool has an argument named "pattern" - should NOT be treated as a constraint
+	input := `{
+		"type": "object",
+		"properties": {
+			"pattern": {
+				"type": "string",
+				"description": "The regex pattern"
+			}
+		},
+		"required": ["pattern"]
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"pattern": {
+				"type": "string",
+				"description": "The regex pattern"
+			}
+		},
+		"required": ["pattern"]
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+
+	var resMap map[string]interface{}
+	json.Unmarshal([]byte(result), &resMap)
+	props, _ := resMap["properties"].(map[string]interface{})
+	if _, ok := props["description"]; ok {
+		t.Errorf("Invalid 'description' property injected into properties map")
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_DotKeys(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"my.param": {
+				"type": "string",
+				"$ref": "#/definitions/MyType"
+			}
+		},
+		"definitions": {
+			"MyType": { "type": "string" }
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	var resMap map[string]interface{}
+	if err := json.Unmarshal([]byte(result), &resMap); err != nil {
+		t.Fatalf("Failed to unmarshal result: %v", err)
+	}
+
+	props, ok := resMap["properties"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("properties missing")
+	}
+
+	if val, ok := props["my.param"]; !ok {
+		t.Fatalf("Key 'my.param' is missing. Result: %s", result)
+	} else {
+		valMap, _ := val.(map[string]interface{})
+		if _, hasRef := valMap["$ref"]; hasRef {
+			t.Errorf("Key 'my.param' still contains $ref")
+		}
+		if _, ok := props["my"]; ok {
+			t.Errorf("Artifact key 'my' created by sjson splitting")
+		}
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_AnyOfAlternativeHints(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"value": {
+				"anyOf": [
+					{ "type": "string" },
+					{ "type": "integer" },
+					{ "type": "null" }
+				]
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	if !strings.Contains(result, "Accepts:") {
+		t.Errorf("Expected alternative types hint, got: %s", result)
+	}
+	if !strings.Contains(result, "string") || !strings.Contains(result, "integer") {
+		t.Errorf("Expected all alternative types in hint, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_NullableHint(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"name": {
+				"type": ["string", "null"],
+				"description": "User name"
+			}
+		},
+		"required": ["name"]
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	if !strings.Contains(result, "(nullable)") {
+		t.Errorf("Expected nullable hint, got: %s", result)
+	}
+	if !strings.Contains(result, "User name") {
+		t.Errorf("Expected original description to be preserved, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_TypeFlattening_Nullable_DotKey(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"my.param": {
+				"type": ["string", "null"]
+			},
+			"other": {
+				"type": "string"
+			}
+		},
+		"required": ["my.param", "other"]
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"my.param": {
+				"type": "string",
+				"description": "(nullable)"
+			},
+			"other": {
+				"type": "string"
+			}
+		},
+		"required": ["other"]
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+}
+
+func TestCleanJSONSchemaForAntigravity_EnumHint(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"status": {
+				"type": "string",
+				"enum": ["active", "inactive", "pending"],
+				"description": "Current status"
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	if !strings.Contains(result, "Allowed:") {
+		t.Errorf("Expected enum values hint, got: %s", result)
+	}
+	if !strings.Contains(result, "active") || !strings.Contains(result, "inactive") {
+		t.Errorf("Expected enum values in hint, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_AdditionalPropertiesHint(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"name": { "type": "string" }
+		},
+		"additionalProperties": false
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	if !strings.Contains(result, "No extra properties allowed") {
+		t.Errorf("Expected additionalProperties hint, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_AnyOfFlattening_PreservesDescription(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"config": {
+				"description": "Parent desc",
+				"anyOf": [
+					{ "type": "string", "description": "Child desc" },
+					{ "type": "integer" }
+				]
+			}
+		}
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"config": {
+				"type": "string",
+				"description": "Parent desc (Child desc) (Accepts: string | integer)"
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+	compareJSON(t, expected, result)
+}
+
+func TestCleanJSONSchemaForAntigravity_SingleEnumNoHint(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"kind": {
+				"type": "string",
+				"enum": ["fixed"]
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	if strings.Contains(result, "Allowed:") {
+		t.Errorf("Single value enum should not add Allowed hint, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_MultipleNonNullTypes(t *testing.T) {
+	input := `{
+		"type": "object",
+		"properties": {
+			"value": {
+				"type": ["string", "integer", "boolean"]
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	if !strings.Contains(result, "Accepts:") {
+		t.Errorf("Expected multiple types hint, got: %s", result)
+	}
+	if !strings.Contains(result, "string") || !strings.Contains(result, "integer") || !strings.Contains(result, "boolean") {
+		t.Errorf("Expected all types in hint, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForGemini_PropertyNamesRemoval(t *testing.T) {
+	// propertyNames is used to validate object property names (e.g., must match a pattern)
+	// Gemini doesn't support this keyword and will reject requests containing it
+	input := `{
+		"type": "object",
+		"properties": {
+			"metadata": {
+				"type": "object",
+				"propertyNames": {
+					"pattern": "^[a-zA-Z_][a-zA-Z0-9_]*$"
+				},
+				"additionalProperties": {
+					"type": "string"
+				}
+			}
+		}
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"metadata": {
+				"type": "object"
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForGemini(input)
+	compareJSON(t, expected, result)
+
+	// Verify propertyNames is completely removed
+	if strings.Contains(result, "propertyNames") {
+		t.Errorf("propertyNames keyword should be removed, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForGemini_PropertyNamesRemoval_Nested(t *testing.T) {
+	// Test deeply nested propertyNames (as seen in real Claude tool schemas)
+	input := `{
+		"type": "object",
+		"properties": {
+			"items": {
+				"type": "array",
+				"items": {
+					"type": "object",
+					"properties": {
+						"config": {
+							"type": "object",
+							"propertyNames": {
+								"type": "string"
+							}
+						}
+					}
+				}
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForGemini(input)
+
+	if strings.Contains(result, "propertyNames") {
+		t.Errorf("Nested propertyNames should be removed, got: %s", result)
+	}
+}
+
+func compareJSON(t *testing.T, expectedJSON, actualJSON string) {
+	var expMap, actMap map[string]interface{}
+	errExp := json.Unmarshal([]byte(expectedJSON), &expMap)
+	errAct := json.Unmarshal([]byte(actualJSON), &actMap)
+
+	if errExp != nil || errAct != nil {
+		t.Fatalf("JSON Unmarshal error. Exp: %v, Act: %v", errExp, errAct)
+	}
+
+	if !reflect.DeepEqual(expMap, actMap) {
+		expBytes, _ := json.MarshalIndent(expMap, "", "  ")
+		actBytes, _ := json.MarshalIndent(actMap, "", "  ")
+		t.Errorf("JSON mismatch:\nExpected:\n%s\n\nActual:\n%s", string(expBytes), string(actBytes))
+	}
+}
+
+// ============================================================================
+// Empty Schema Placeholder Tests
+// ============================================================================
+
+func TestCleanJSONSchemaForAntigravity_EmptySchemaPlaceholder(t *testing.T) {
+	// Empty object schema with no properties should get a placeholder
+	input := `{
+		"type": "object"
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// Should have placeholder property added
+	if !strings.Contains(result, `"reason"`) {
+		t.Errorf("Empty schema should have 'reason' placeholder property, got: %s", result)
+	}
+	if !strings.Contains(result, `"required"`) {
+		t.Errorf("Empty schema should have 'required' with 'reason', got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_EmptyPropertiesPlaceholder(t *testing.T) {
+	// Object with empty properties object
+	input := `{
+		"type": "object",
+		"properties": {}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// Should have placeholder property added
+	if !strings.Contains(result, `"reason"`) {
+		t.Errorf("Empty properties should have 'reason' placeholder, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_NonEmptySchemaUnchanged(t *testing.T) {
+	// Schema with properties should NOT get placeholder
+	input := `{
+		"type": "object",
+		"properties": {
+			"name": {"type": "string"}
+		},
+		"required": ["name"]
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// Should NOT have placeholder property
+	if strings.Contains(result, `"reason"`) {
+		t.Errorf("Non-empty schema should NOT have 'reason' placeholder, got: %s", result)
+	}
+	// Original properties should be preserved
+	if !strings.Contains(result, `"name"`) {
+		t.Errorf("Original property 'name' should be preserved, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_NestedEmptySchema(t *testing.T) {
+	// Nested empty object in items should also get placeholder
+	input := `{
+		"type": "object",
+		"properties": {
+			"items": {
+				"type": "array",
+				"items": {
+					"type": "object"
+				}
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// Nested empty object should also get placeholder
+	// Check that the nested object has a reason property
+	parsed := gjson.Parse(result)
+	nestedProps := parsed.Get("properties.items.items.properties")
+	if !nestedProps.Exists() || !nestedProps.Get("reason").Exists() {
+		t.Errorf("Nested empty object should have 'reason' placeholder, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_EmptySchemaWithDescription(t *testing.T) {
+	// Empty schema with description should preserve description and add placeholder
+	input := `{
+		"type": "object",
+		"description": "An empty object"
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// Should have both description and placeholder
+	if !strings.Contains(result, `"An empty object"`) {
+		t.Errorf("Description should be preserved, got: %s", result)
+	}
+	if !strings.Contains(result, `"reason"`) {
+		t.Errorf("Empty schema should have 'reason' placeholder, got: %s", result)
+	}
+}
+
+// ============================================================================
+// Format field handling (ad-hoc patch removal)
+// ============================================================================
+
+func TestCleanJSONSchemaForAntigravity_FormatFieldRemoval(t *testing.T) {
+	// format:"uri" should be removed and added as hint
+	input := `{
+		"type": "object",
+		"properties": {
+			"url": {
+				"type": "string",
+				"format": "uri",
+				"description": "A URL"
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// format should be removed
+	if strings.Contains(result, `"format"`) {
+		t.Errorf("format field should be removed, got: %s", result)
+	}
+	// hint should be added to description
+	if !strings.Contains(result, "format: uri") {
+		t.Errorf("format hint should be added to description, got: %s", result)
+	}
+	// original description should be preserved
+	if !strings.Contains(result, "A URL") {
+		t.Errorf("Original description should be preserved, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_FormatFieldNoDescription(t *testing.T) {
+	// format without description should create description with hint
+	input := `{
+		"type": "object",
+		"properties": {
+			"email": {
+				"type": "string",
+				"format": "email"
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// format should be removed
+	if strings.Contains(result, `"format"`) {
+		t.Errorf("format field should be removed, got: %s", result)
+	}
+	// hint should be added
+	if !strings.Contains(result, "format: email") {
+		t.Errorf("format hint should be added, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_MultipleFormats(t *testing.T) {
+	// Multiple format fields should all be handled
+	input := `{
+		"type": "object",
+		"properties": {
+			"url": {"type": "string", "format": "uri"},
+			"email": {"type": "string", "format": "email"},
+			"date": {"type": "string", "format": "date-time"}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// All format fields should be removed
+	if strings.Contains(result, `"format"`) {
+		t.Errorf("All format fields should be removed, got: %s", result)
+	}
+	// All hints should be added
+	if !strings.Contains(result, "format: uri") {
+		t.Errorf("uri format hint should be added, got: %s", result)
+	}
+	if !strings.Contains(result, "format: email") {
+		t.Errorf("email format hint should be added, got: %s", result)
+	}
+	if !strings.Contains(result, "format: date-time") {
+		t.Errorf("date-time format hint should be added, got: %s", result)
+	}
+}

internal/util/gemini_thinking.go

@@ -1,6 +1,7 @@
 package util
 
 import (
+	"regexp"
 	"strings"
 
 	"github.com/tidwall/gjson"
@@ -13,6 +14,44 @@ const (
 	GeminiOriginalModelMetadataKey   = "gemini_original_model"
 )
 
+// Gemini model family detection patterns
+var (
+	gemini3Pattern      = regexp.MustCompile(`(?i)^gemini[_-]?3[_-]`)
+	gemini3ProPattern   = regexp.MustCompile(`(?i)^gemini[_-]?3[_-]pro`)
+	gemini3FlashPattern = regexp.MustCompile(`(?i)^gemini[_-]?3[_-]flash`)
+	gemini25Pattern     = regexp.MustCompile(`(?i)^gemini[_-]?2\.5[_-]`)
+)
+
+// IsGemini3Model returns true if the model is a Gemini 3 family model.
+// Gemini 3 models should use thinkingLevel (string) instead of thinkingBudget (number).
+func IsGemini3Model(model string) bool {
+	return gemini3Pattern.MatchString(model)
+}
+
+// IsGemini3ProModel returns true if the model is a Gemini 3 Pro variant.
+// Gemini 3 Pro supports thinkingLevel: "low", "high" (default: "high")
+func IsGemini3ProModel(model string) bool {
+	return gemini3ProPattern.MatchString(model)
+}
+
+// IsGemini3FlashModel returns true if the model is a Gemini 3 Flash variant.
+// Gemini 3 Flash supports thinkingLevel: "minimal", "low", "medium", "high" (default: "high")
+func IsGemini3FlashModel(model string) bool {
+	return gemini3FlashPattern.MatchString(model)
+}
+
+// IsGemini25Model returns true if the model is a Gemini 2.5 family model.
+// Gemini 2.5 models should use thinkingBudget (number).
+func IsGemini25Model(model string) bool {
+	return gemini25Pattern.MatchString(model)
+}
+
+// Gemini3ProThinkingLevels are the valid thinkingLevel values for Gemini 3 Pro models.
+var Gemini3ProThinkingLevels = []string{"low", "high"}
+
+// Gemini3FlashThinkingLevels are the valid thinkingLevel values for Gemini 3 Flash models.
+var Gemini3FlashThinkingLevels = []string{"minimal", "low", "medium", "high"}
+
 func ApplyGeminiThinkingConfig(body []byte, budget *int, includeThoughts *bool) []byte {
 	if budget == nil && includeThoughts == nil {
 		return body
@@ -69,10 +108,153 @@ func ApplyGeminiCLIThinkingConfig(body []byte, budget *int, includeThoughts *boo
 	return updated
 }
 
+// ApplyGeminiThinkingLevel applies thinkingLevel config for Gemini 3 models.
+// For standard Gemini API format (generationConfig.thinkingConfig path).
+// Per Google's documentation, Gemini 3 models should use thinkingLevel instead of thinkingBudget.
+func ApplyGeminiThinkingLevel(body []byte, level string, includeThoughts *bool) []byte {
+	if level == "" && includeThoughts == nil {
+		return body
+	}
+	updated := body
+	if level != "" {
+		valuePath := "generationConfig.thinkingConfig.thinkingLevel"
+		rewritten, err := sjson.SetBytes(updated, valuePath, level)
+		if err == nil {
+			updated = rewritten
+		}
+	}
+	// Default to including thoughts when a level is set but no explicit include flag is provided.
+	incl := includeThoughts
+	if incl == nil && level != "" {
+		defaultInclude := true
+		incl = &defaultInclude
+	}
+	if incl != nil {
+		valuePath := "generationConfig.thinkingConfig.includeThoughts"
+		rewritten, err := sjson.SetBytes(updated, valuePath, *incl)
+		if err == nil {
+			updated = rewritten
+		}
+	}
+	if it := gjson.GetBytes(body, "generationConfig.thinkingConfig.include_thoughts"); it.Exists() {
+		updated, _ = sjson.DeleteBytes(updated, "generationConfig.thinkingConfig.include_thoughts")
+	}
+	if tb := gjson.GetBytes(body, "generationConfig.thinkingConfig.thinkingBudget"); tb.Exists() {
+		updated, _ = sjson.DeleteBytes(updated, "generationConfig.thinkingConfig.thinkingBudget")
+	}
+	return updated
+}
+
+// ApplyGeminiCLIThinkingLevel applies thinkingLevel config for Gemini 3 models.
+// For Gemini CLI API format (request.generationConfig.thinkingConfig path).
+// Per Google's documentation, Gemini 3 models should use thinkingLevel instead of thinkingBudget.
+func ApplyGeminiCLIThinkingLevel(body []byte, level string, includeThoughts *bool) []byte {
+	if level == "" && includeThoughts == nil {
+		return body
+	}
+	updated := body
+	if level != "" {
+		valuePath := "request.generationConfig.thinkingConfig.thinkingLevel"
+		rewritten, err := sjson.SetBytes(updated, valuePath, level)
+		if err == nil {
+			updated = rewritten
+		}
+	}
+	// Default to including thoughts when a level is set but no explicit include flag is provided.
+	incl := includeThoughts
+	if incl == nil && level != "" {
+		defaultInclude := true
+		incl = &defaultInclude
+	}
+	if incl != nil {
+		valuePath := "request.generationConfig.thinkingConfig.includeThoughts"
+		rewritten, err := sjson.SetBytes(updated, valuePath, *incl)
+		if err == nil {
+			updated = rewritten
+		}
+	}
+	if it := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.include_thoughts"); it.Exists() {
+		updated, _ = sjson.DeleteBytes(updated, "request.generationConfig.thinkingConfig.include_thoughts")
+	}
+	if tb := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.thinkingBudget"); tb.Exists() {
+		updated, _ = sjson.DeleteBytes(updated, "request.generationConfig.thinkingConfig.thinkingBudget")
+	}
+	return updated
+}
+
+// ValidateGemini3ThinkingLevel validates that the thinkingLevel is valid for the Gemini 3 model variant.
+// Returns the validated level (normalized to lowercase) and true if valid, or empty string and false if invalid.
+func ValidateGemini3ThinkingLevel(model, level string) (string, bool) {
+	if level == "" {
+		return "", false
+	}
+	normalized := strings.ToLower(strings.TrimSpace(level))
+
+	var validLevels []string
+	if IsGemini3ProModel(model) {
+		validLevels = Gemini3ProThinkingLevels
+	} else if IsGemini3FlashModel(model) {
+		validLevels = Gemini3FlashThinkingLevels
+	} else if IsGemini3Model(model) {
+		// Unknown Gemini 3 variant - allow all levels as fallback
+		validLevels = Gemini3FlashThinkingLevels
+	} else {
+		return "", false
+	}
+
+	for _, valid := range validLevels {
+		if normalized == valid {
+			return normalized, true
+		}
+	}
+	return "", false
+}
+
+// ThinkingBudgetToGemini3Level converts a thinkingBudget to a thinkingLevel for Gemini 3 models.
+// This provides backward compatibility when thinkingBudget is provided for Gemini 3 models.
+// Returns the appropriate thinkingLevel and true if conversion is possible.
+func ThinkingBudgetToGemini3Level(model string, budget int) (string, bool) {
+	if !IsGemini3Model(model) {
+		return "", false
+	}
+
+	// Map budget to level based on Google's documentation
+	// Gemini 3 Pro: "low", "high" (default: "high")
+	// Gemini 3 Flash: "minimal", "low", "medium", "high" (default: "high")
+	switch {
+	case budget == -1:
+		// Dynamic budget maps to "high" (API default)
+		return "high", true
+	case budget == 0:
+		// Zero budget - Gemini 3 doesn't support disabling thinking
+		// Map to lowest available level
+		if IsGemini3FlashModel(model) {
+			return "minimal", true
+		}
+		return "low", true
+	case budget > 0 && budget <= 512:
+		if IsGemini3FlashModel(model) {
+			return "minimal", true
+		}
+		return "low", true
+	case budget <= 1024:
+		return "low", true
+	case budget <= 8192:
+		if IsGemini3FlashModel(model) {
+			return "medium", true
+		}
+		return "low", true // Pro doesn't have medium, use low
+	default:
+		return "high", true
+	}
+}
+
 // modelsWithDefaultThinking lists models that should have thinking enabled by default
 // when no explicit thinkingConfig is provided.
 var modelsWithDefaultThinking = map[string]bool{
-	"gemini-3-pro-preview": true,
+	"gemini-3-pro-preview":       true,
+	"gemini-3-pro-image-preview": true,
+	// "gemini-3-flash-preview":     true,
 }
 
 // ModelHasDefaultThinking returns true if the model should have thinking enabled by default.
@@ -83,6 +265,7 @@ func ModelHasDefaultThinking(model string) bool {
 // ApplyDefaultThinkingIfNeeded injects default thinkingConfig for models that require it.
 // For standard Gemini API format (generationConfig.thinkingConfig path).
 // Returns the modified body if thinkingConfig was added, otherwise returns the original.
+// For Gemini 3 models, uses thinkingLevel instead of thinkingBudget per Google's documentation.
 func ApplyDefaultThinkingIfNeeded(model string, body []byte) []byte {
 	if !ModelHasDefaultThinking(model) {
 		return body
@@ -90,14 +273,59 @@ func ApplyDefaultThinkingIfNeeded(model string, body []byte) []byte {
 	if gjson.GetBytes(body, "generationConfig.thinkingConfig").Exists() {
 		return body
 	}
+	// Gemini 3 models use thinkingLevel instead of thinkingBudget
+	if IsGemini3Model(model) {
+		// Don't set a default - let the API use its dynamic default ("high")
+		// Only set includeThoughts
+		updated, _ := sjson.SetBytes(body, "generationConfig.thinkingConfig.includeThoughts", true)
+		return updated
+	}
+	// Gemini 2.5 and other models use thinkingBudget
 	updated, _ := sjson.SetBytes(body, "generationConfig.thinkingConfig.thinkingBudget", -1)
 	updated, _ = sjson.SetBytes(updated, "generationConfig.thinkingConfig.include_thoughts", true)
 	return updated
 }
 
+// ApplyGemini3ThinkingLevelFromMetadata applies thinkingLevel from metadata for Gemini 3 models.
+// For standard Gemini API format (generationConfig.thinkingConfig path).
+// This handles the case where reasoning_effort is specified via model name suffix (e.g., model(minimal)).
+func ApplyGemini3ThinkingLevelFromMetadata(model string, metadata map[string]any, body []byte) []byte {
+	if !IsGemini3Model(model) {
+		return body
+	}
+	effort, ok := ReasoningEffortFromMetadata(metadata)
+	if !ok || effort == "" {
+		return body
+	}
+	// Validate and apply the thinkingLevel
+	if level, valid := ValidateGemini3ThinkingLevel(model, effort); valid {
+		return ApplyGeminiThinkingLevel(body, level, nil)
+	}
+	return body
+}
+
+// ApplyGemini3ThinkingLevelFromMetadataCLI applies thinkingLevel from metadata for Gemini 3 models.
+// For Gemini CLI API format (request.generationConfig.thinkingConfig path).
+// This handles the case where reasoning_effort is specified via model name suffix (e.g., model(minimal)).
+func ApplyGemini3ThinkingLevelFromMetadataCLI(model string, metadata map[string]any, body []byte) []byte {
+	if !IsGemini3Model(model) {
+		return body
+	}
+	effort, ok := ReasoningEffortFromMetadata(metadata)
+	if !ok || effort == "" {
+		return body
+	}
+	// Validate and apply the thinkingLevel
+	if level, valid := ValidateGemini3ThinkingLevel(model, effort); valid {
+		return ApplyGeminiCLIThinkingLevel(body, level, nil)
+	}
+	return body
+}
+
 // ApplyDefaultThinkingIfNeededCLI injects default thinkingConfig for models that require it.
 // For Gemini CLI API format (request.generationConfig.thinkingConfig path).
 // Returns the modified body if thinkingConfig was added, otherwise returns the original.
+// For Gemini 3 models, uses thinkingLevel instead of thinkingBudget per Google's documentation.
 func ApplyDefaultThinkingIfNeededCLI(model string, body []byte) []byte {
 	if !ModelHasDefaultThinking(model) {
 		return body
@@ -105,6 +333,14 @@ func ApplyDefaultThinkingIfNeededCLI(model string, body []byte) []byte {
 	if gjson.GetBytes(body, "request.generationConfig.thinkingConfig").Exists() {
 		return body
 	}
+	// Gemini 3 models use thinkingLevel instead of thinkingBudget
+	if IsGemini3Model(model) {
+		// Don't set a default - let the API use its dynamic default ("high")
+		// Only set includeThoughts
+		updated, _ := sjson.SetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts", true)
+		return updated
+	}
+	// Gemini 2.5 and other models use thinkingBudget
 	updated, _ := sjson.SetBytes(body, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
 	updated, _ = sjson.SetBytes(updated, "request.generationConfig.thinkingConfig.include_thoughts", true)
 	return updated
@@ -128,12 +364,31 @@ func StripThinkingConfigIfUnsupported(model string, body []byte) []byte {
 
 // NormalizeGeminiThinkingBudget normalizes the thinkingBudget value in a standard Gemini
 // request body (generationConfig.thinkingConfig.thinkingBudget path).
-func NormalizeGeminiThinkingBudget(model string, body []byte) []byte {
+// For Gemini 3 models, converts thinkingBudget to thinkingLevel per Google's documentation,
+// unless skipGemini3Check is provided and true.
+func NormalizeGeminiThinkingBudget(model string, body []byte, skipGemini3Check ...bool) []byte {
 	const budgetPath = "generationConfig.thinkingConfig.thinkingBudget"
+	const levelPath = "generationConfig.thinkingConfig.thinkingLevel"
+
 	budget := gjson.GetBytes(body, budgetPath)
 	if !budget.Exists() {
 		return body
 	}
+
+	// For Gemini 3 models, convert thinkingBudget to thinkingLevel
+	skipGemini3 := len(skipGemini3Check) > 0 && skipGemini3Check[0]
+	if IsGemini3Model(model) && !skipGemini3 {
+		if level, ok := ThinkingBudgetToGemini3Level(model, int(budget.Int())); ok {
+			updated, _ := sjson.SetBytes(body, levelPath, level)
+			updated, _ = sjson.DeleteBytes(updated, budgetPath)
+			return updated
+		}
+		// If conversion fails, just remove the budget (let API use default)
+		updated, _ := sjson.DeleteBytes(body, budgetPath)
+		return updated
+	}
+
+	// For Gemini 2.5 and other models, normalize the budget value
 	normalized := NormalizeThinkingBudget(model, int(budget.Int()))
 	updated, _ := sjson.SetBytes(body, budgetPath, normalized)
 	return updated
@@ -141,12 +396,31 @@ func NormalizeGeminiThinkingBudget(model string, body []byte) []byte {
 
 // NormalizeGeminiCLIThinkingBudget normalizes the thinkingBudget value in a Gemini CLI
 // request body (request.generationConfig.thinkingConfig.thinkingBudget path).
-func NormalizeGeminiCLIThinkingBudget(model string, body []byte) []byte {
+// For Gemini 3 models, converts thinkingBudget to thinkingLevel per Google's documentation,
+// unless skipGemini3Check is provided and true.
+func NormalizeGeminiCLIThinkingBudget(model string, body []byte, skipGemini3Check ...bool) []byte {
 	const budgetPath = "request.generationConfig.thinkingConfig.thinkingBudget"
+	const levelPath = "request.generationConfig.thinkingConfig.thinkingLevel"
+
 	budget := gjson.GetBytes(body, budgetPath)
 	if !budget.Exists() {
 		return body
 	}
+
+	// For Gemini 3 models, convert thinkingBudget to thinkingLevel
+	skipGemini3 := len(skipGemini3Check) > 0 && skipGemini3Check[0]
+	if IsGemini3Model(model) && !skipGemini3 {
+		if level, ok := ThinkingBudgetToGemini3Level(model, int(budget.Int())); ok {
+			updated, _ := sjson.SetBytes(body, levelPath, level)
+			updated, _ = sjson.DeleteBytes(updated, budgetPath)
+			return updated
+		}
+		// If conversion fails, just remove the budget (let API use default)
+		updated, _ := sjson.DeleteBytes(body, budgetPath)
+		return updated
+	}
+
+	// For Gemini 2.5 and other models, normalize the budget value
 	normalized := NormalizeThinkingBudget(model, int(budget.Int()))
 	updated, _ := sjson.SetBytes(body, budgetPath, normalized)
 	return updated
@@ -218,44 +492,74 @@ func ApplyReasoningEffortToGeminiCLI(body []byte, effort string) []byte {
 }
 
 // ConvertThinkingLevelToBudget checks for "generationConfig.thinkingConfig.thinkingLevel"
-// and converts it to "thinkingBudget".
-// "high" -> 32768
-// "low" -> 128
-// It removes "thinkingLevel" after conversion.
-func ConvertThinkingLevelToBudget(body []byte) []byte {
+// and converts it to "thinkingBudget" for Gemini 2.5 models.
+// For Gemini 3 models, preserves thinkingLevel unless skipGemini3Check is provided and true.
+// Mappings for Gemini 2.5:
+//   - "high" -> 32768
+//   - "medium" -> 8192
+//   - "low" -> 1024
+//   - "minimal" -> 512
+//
+// It removes "thinkingLevel" after conversion (for Gemini 2.5 only).
+func ConvertThinkingLevelToBudget(body []byte, model string, skipGemini3Check ...bool) []byte {
 	levelPath := "generationConfig.thinkingConfig.thinkingLevel"
 	res := gjson.GetBytes(body, levelPath)
 	if !res.Exists() {
 		return body
 	}
 
-	level := strings.ToLower(res.String())
-	var budget int
-	switch level {
-	case "high":
-		budget = 32768
-	case "low":
-		budget = 128
-	default:
-		// If unknown level, we might just leave it or default.
-		// User only specified high and low. We'll assume we shouldn't touch it if it's something else,
-		// or maybe we should just remove the invalid level?
-		// For safety adhering to strict instructions: "If high... if low...".
-		// If it's something else, the upstream might fail anyway if we leave it,
-		// but let's just delete the level if we processed it.
-		// Actually, let's check if we need to do anything for other values.
-		// For now, only handle high/low.
+	// For Gemini 3 models, preserve thinkingLevel unless explicitly skipped
+	skipGemini3 := len(skipGemini3Check) > 0 && skipGemini3Check[0]
+	if IsGemini3Model(model) && !skipGemini3 {
 		return body
 	}
 
-	// Set budget
+	budget, ok := ThinkingLevelToBudget(res.String())
+	if !ok {
+		updated, _ := sjson.DeleteBytes(body, levelPath)
+		return updated
+	}
+
 	budgetPath := "generationConfig.thinkingConfig.thinkingBudget"
 	updated, err := sjson.SetBytes(body, budgetPath, budget)
 	if err != nil {
 		return body
 	}
 
-	// Remove level
+	updated, err = sjson.DeleteBytes(updated, levelPath)
+	if err != nil {
+		return body
+	}
+	return updated
+}
+
+// ConvertThinkingLevelToBudgetCLI checks for "request.generationConfig.thinkingConfig.thinkingLevel"
+// and converts it to "thinkingBudget" for Gemini 2.5 models.
+// For Gemini 3 models, preserves thinkingLevel as-is (does not convert).
+func ConvertThinkingLevelToBudgetCLI(body []byte, model string) []byte {
+	levelPath := "request.generationConfig.thinkingConfig.thinkingLevel"
+	res := gjson.GetBytes(body, levelPath)
+	if !res.Exists() {
+		return body
+	}
+
+	// For Gemini 3 models, preserve thinkingLevel - don't convert to budget
+	if IsGemini3Model(model) {
+		return body
+	}
+
+	budget, ok := ThinkingLevelToBudget(res.String())
+	if !ok {
+		updated, _ := sjson.DeleteBytes(body, levelPath)
+		return updated
+	}
+
+	budgetPath := "request.generationConfig.thinkingConfig.thinkingBudget"
+	updated, err := sjson.SetBytes(body, budgetPath, budget)
+	if err != nil {
+		return body
+	}
+
 	updated, err = sjson.DeleteBytes(updated, levelPath)
 	if err != nil {
 		return body

internal/util/openai_thinking.go

@@ -1,37 +0,0 @@
-package util
-
-// OpenAIThinkingBudgetToEffort maps a numeric thinking budget (tokens)
-// into an OpenAI-style reasoning effort level for level-based models.
-//
-// Ranges:
-//   - 0            -> "none"
-//   - -1           -> "auto"
-//   - 1..1024      -> "low"
-//   - 1025..8192   -> "medium"
-//   - 8193..24576  -> "high"
-//   - 24577..      -> highest supported level for the model (defaults to "xhigh")
-//
-// Negative values other than -1 are treated as unsupported.
-func OpenAIThinkingBudgetToEffort(model string, budget int) (string, bool) {
-	switch {
-	case budget == -1:
-		return "auto", true
-	case budget < -1:
-		return "", false
-	case budget == 0:
-		return "none", true
-	case budget > 0 && budget <= 1024:
-		return "low", true
-	case budget <= 8192:
-		return "medium", true
-	case budget <= 24576:
-		return "high", true
-	case budget > 24576:
-		if levels := GetModelThinkingLevels(model); len(levels) > 0 {
-			return levels[len(levels)-1], true
-		}
-		return "xhigh", true
-	default:
-		return "", false
-	}
-}

internal/util/thinking.go

@@ -118,3 +118,111 @@ func IsOpenAICompatibilityModel(model string) bool {
 	}
 	return strings.EqualFold(strings.TrimSpace(info.Type), "openai-compatibility")
 }
+
+// ThinkingEffortToBudget maps a reasoning effort level to a numeric thinking budget (tokens),
+// clamping the result to the model's supported range.
+//
+// Mappings (values are normalized to model's supported range):
+//   - "none"    -> 0
+//   - "auto"    -> -1
+//   - "minimal" -> 512
+//   - "low"     -> 1024
+//   - "medium"  -> 8192
+//   - "high"    -> 24576
+//   - "xhigh"   -> 32768
+//
+// Returns false when the effort level is empty or unsupported.
+func ThinkingEffortToBudget(model, effort string) (int, bool) {
+	if effort == "" {
+		return 0, false
+	}
+	normalized, ok := NormalizeReasoningEffortLevel(model, effort)
+	if !ok {
+		normalized = strings.ToLower(strings.TrimSpace(effort))
+	}
+	switch normalized {
+	case "none":
+		return 0, true
+	case "auto":
+		return NormalizeThinkingBudget(model, -1), true
+	case "minimal":
+		return NormalizeThinkingBudget(model, 512), true
+	case "low":
+		return NormalizeThinkingBudget(model, 1024), true
+	case "medium":
+		return NormalizeThinkingBudget(model, 8192), true
+	case "high":
+		return NormalizeThinkingBudget(model, 24576), true
+	case "xhigh":
+		return NormalizeThinkingBudget(model, 32768), true
+	default:
+		return 0, false
+	}
+}
+
+// ThinkingLevelToBudget maps a Gemini thinkingLevel to a numeric thinking budget (tokens).
+//
+// Mappings:
+//   - "minimal" -> 512
+//   - "low"     -> 1024
+//   - "medium"  -> 8192
+//   - "high"    -> 32768
+//
+// Returns false when the level is empty or unsupported.
+func ThinkingLevelToBudget(level string) (int, bool) {
+	if level == "" {
+		return 0, false
+	}
+	normalized := strings.ToLower(strings.TrimSpace(level))
+	switch normalized {
+	case "minimal":
+		return 512, true
+	case "low":
+		return 1024, true
+	case "medium":
+		return 8192, true
+	case "high":
+		return 32768, true
+	default:
+		return 0, false
+	}
+}
+
+// ThinkingBudgetToEffort maps a numeric thinking budget (tokens)
+// to a reasoning effort level for level-based models.
+//
+// Mappings:
+//   - 0            -> "none" (or lowest supported level if model doesn't support "none")
+//   - -1           -> "auto"
+//   - 1..1024      -> "low"
+//   - 1025..8192   -> "medium"
+//   - 8193..24576  -> "high"
+//   - 24577..      -> highest supported level for the model (defaults to "xhigh")
+//
+// Returns false when the budget is unsupported (negative values other than -1).
+func ThinkingBudgetToEffort(model string, budget int) (string, bool) {
+	switch {
+	case budget == -1:
+		return "auto", true
+	case budget < -1:
+		return "", false
+	case budget == 0:
+		if levels := GetModelThinkingLevels(model); len(levels) > 0 {
+			return levels[0], true
+		}
+		return "none", true
+	case budget > 0 && budget <= 1024:
+		return "low", true
+	case budget <= 8192:
+		return "medium", true
+	case budget <= 24576:
+		return "high", true
+	case budget > 24576:
+		if levels := GetModelThinkingLevels(model); len(levels) > 0 {
+			return levels[len(levels)-1], true
+		}
+		return "xhigh", true
+	default:
+		return "", false
+	}
+}