Merge pull request #1099 from router-for-me/claude

refactor(claude): move max_tokens constraint enforcement to Apply method

README.md

@@ -118,9 +118,36 @@ Native macOS GUI for managing CLIProxyAPI: configure providers, model mappings,
 
 Native macOS menu bar app that unifies Claude, Gemini, OpenAI, Qwen, and Antigravity subscriptions with real-time quota tracking and smart auto-failover for AI coding tools like Claude Code, OpenCode, and Droid - no API keys needed.
 
+### [CodMate](https://github.com/loocor/CodMate)
+
+Native macOS SwiftUI app for managing CLI AI sessions (Codex, Claude Code, Gemini CLI) with unified provider management, Git review, project organization, global search, and terminal integration. Integrates CLIProxyAPI to provide OAuth authentication for Codex, Claude, Gemini, Antigravity, and Qwen Code, with built-in and third-party provider rerouting through a single proxy endpoint - no API keys needed for OAuth providers.
+
+### [ProxyPilot](https://github.com/Finesssee/ProxyPilot)
+
+Windows-native CLIProxyAPI fork with TUI, system tray, and multi-provider OAuth for AI coding tools - no API keys needed.
+
+### [Claude Proxy VSCode](https://github.com/uzhao/claude-proxy-vscode)
+
+VSCode extension for quick switching between Claude Code models, featuring integrated CLIProxyAPI as its backend with automatic background lifecycle management.
+
+### [ZeroLimit](https://github.com/0xtbug/zero-limit)
+
+Windows desktop app built with Tauri + React for monitoring AI coding assistant quotas via CLIProxyAPI. Track usage across Gemini, Claude, OpenAI Codex, and Antigravity accounts with real-time dashboard, system tray integration, and one-click proxy control - no API keys needed.
+
 > [!NOTE]  
 > If you developed a project based on CLIProxyAPI, please open a PR to add it to this list.
 
+## More choices
+
+Those projects are ports of CLIProxyAPI or inspired by it:
+
+### [9Router](https://github.com/decolua/9router)
+
+A Next.js implementation inspired by CLIProxyAPI, easy to install and use, built from scratch with format translation (OpenAI/Claude/Gemini/Ollama), combo system with auto-fallback, multi-account management with exponential backoff, a Next.js web dashboard, and support for CLI tools (Cursor, Claude Code, Cline, RooCode) - no API keys needed.
+
+> [!NOTE]  
+> If you have developed a port of CLIProxyAPI or a project inspired by it, please open a PR to add it to this list.
+
 ## License
 
 This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.

README_CN.md

@@ -117,9 +117,36 @@ CLI 封装器，用于通过 CLIProxyAPI OAuth 即时切换多个 Claude 账户
 
 原生 macOS 菜单栏应用，统一管理 Claude、Gemini、OpenAI、Qwen 和 Antigravity 订阅，提供实时配额追踪和智能自动故障转移，支持 Claude Code、OpenCode 和 Droid 等 AI 编程工具，无需 API 密钥。
 
+### [CodMate](https://github.com/loocor/CodMate)
+
+原生 macOS SwiftUI 应用，用于管理 CLI AI 会话（Claude Code、Codex、Gemini CLI），提供统一的提供商管理、Git 审查、项目组织、全局搜索和终端集成。集成 CLIProxyAPI 为 Codex、Claude、Gemini、Antigravity 和 Qwen Code 提供统一的 OAuth 认证，支持内置和第三方提供商通过单一代理端点重路由 - OAuth 提供商无需 API 密钥。
+
+### [ProxyPilot](https://github.com/Finesssee/ProxyPilot)
+
+原生 Windows CLIProxyAPI 分支，集成 TUI、系统托盘及多服务商 OAuth 认证，专为 AI 编程工具打造，无需 API 密钥。
+
+### [Claude Proxy VSCode](https://github.com/uzhao/claude-proxy-vscode)
+
+一款 VSCode 扩展，提供了在 VSCode 中快速切换 Claude Code 模型的功能，内置 CLIProxyAPI 作为其后端，支持后台自动启动和关闭。
+
+### [ZeroLimit](https://github.com/0xtbug/zero-limit)
+
+Windows 桌面应用，基于 Tauri + React 构建，用于通过 CLIProxyAPI 监控 AI 编程助手配额。支持跨 Gemini、Claude、OpenAI Codex 和 Antigravity 账户的使用量追踪，提供实时仪表盘、系统托盘集成和一键代理控制，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。
 
+## 更多选择
+
+以下项目是 CLIProxyAPI 的移植版或受其启发：
+
+### [9Router](https://github.com/decolua/9router)
+
+基于 Next.js 的实现，灵感来自 CLIProxyAPI，易于安装使用；自研格式转换（OpenAI/Claude/Gemini/Ollama）、组合系统与自动回退、多账户管理（指数退避）、Next.js Web 控制台，并支持 Cursor、Claude Code、Cline、RooCode 等 CLI 工具，无需 API 密钥。
+
+> [!NOTE]  
+> 如果你开发了 CLIProxyAPI 的移植或衍生项目，请提交 PR 将其添加到此列表中。
+
 ## 许可证
 
 此项目根据 MIT 许可证授权 - 有关详细信息，请参阅 [LICENSE](LICENSE) 文件。

cmd/server/main.go

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

config.example.yaml

@@ -77,11 +77,18 @@ routing:
 # When true, enable authentication for the WebSocket API (/v1/ws).
 ws-auth: false
 
+# When > 0, emit blank lines every N seconds for non-streaming responses to prevent idle timeouts.
+nonstream-keepalive-interval: 0
+
 # Streaming behavior (SSE keep-alives + safe bootstrap retries).
 # streaming:
 #   keepalive-seconds: 15   # Default: 0 (disabled). <= 0 disables keep-alives.
 #   bootstrap-retries: 1    # Default: 0 (disabled). Retries before first byte is sent.
 
+# When true, enable official Codex instructions injection for Codex API requests.
+# When false (default), CodexInstructionsForModel returns immediately without modification.
+codex-instructions-enabled: false
+
 # Gemini API keys
 # gemini-api-key:
 #   - api-key: "AIzaSy...01"
@@ -198,23 +205,37 @@ ws-auth: false
 #     - from: "claude-haiku-4-5-20251001"
 #       to: "gemini-2.5-flash"
 
-# Global OAuth model name mappings (per channel)
-# These mappings rename model IDs for both model listing and request routing.
+# Global OAuth model name aliases (per channel)
+# These aliases rename model IDs for both model listing and request routing.
 # Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow.
-# NOTE: Mappings do not apply to gemini-api-key, codex-api-key, claude-api-key, openai-compatibility, vertex-api-key, or ampcode.
-# oauth-model-mappings:
+# NOTE: Aliases do not apply to gemini-api-key, codex-api-key, claude-api-key, openai-compatibility, vertex-api-key, or ampcode.
+# You can repeat the same name with different aliases to expose multiple client model names.
+oauth-model-alias:
+  antigravity:
+    - name: "rev19-uic3-1p"
+      alias: "gemini-2.5-computer-use-preview-10-2025"
+    - name: "gemini-3-pro-image"
+      alias: "gemini-3-pro-image-preview"
+    - name: "gemini-3-pro-high"
+      alias: "gemini-3-pro-preview"
+    - name: "gemini-3-flash"
+      alias: "gemini-3-flash-preview"
+    - name: "claude-sonnet-4-5"
+      alias: "gemini-claude-sonnet-4-5"
+    - name: "claude-sonnet-4-5-thinking"
+      alias: "gemini-claude-sonnet-4-5-thinking"
+    - name: "claude-opus-4-5-thinking"
+      alias: "gemini-claude-opus-4-5-thinking"
 #   gemini-cli:
 #     - name: "gemini-2.5-pro"          # original model name under this channel
 #       alias: "g2.5p"                  # client-visible alias
+#       fork: true                      # when true, keep original and also add the alias as an extra model (default: false)
 #   vertex:
 #     - name: "gemini-2.5-pro"
 #       alias: "g2.5p"
 #   aistudio:
 #     - name: "gemini-2.5-pro"
 #       alias: "g2.5p"
-#   antigravity:
-#     - name: "gemini-3-pro-preview"
-#       alias: "g3p"
 #   claude:
 #     - name: "claude-sonnet-4-5-20250929"
 #       alias: "cs4.5"
@@ -258,9 +279,21 @@ ws-auth: false
 #           protocol: "gemini" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex
 #       params: # JSON path (gjson/sjson syntax) -> value
 #         "generationConfig.thinkingConfig.thinkingBudget": 32768
+#   default-raw: # Default raw rules set parameters using raw JSON when missing (must be valid JSON).
+#     - models:
+#         - name: "gemini-2.5-pro" # Supports wildcards (e.g., "gemini-*")
+#           protocol: "gemini" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex
+#       params: # JSON path (gjson/sjson syntax) -> raw JSON value (strings are used as-is, must be valid JSON)
+#         "generationConfig.responseJsonSchema": "{\"type\":\"object\",\"properties\":{\"answer\":{\"type\":\"string\"}}}"
 #   override: # Override rules always set parameters, overwriting any existing values.
 #     - models:
 #         - name: "gpt-*" # Supports wildcards (e.g., "gpt-*")
 #           protocol: "codex" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex
 #       params: # JSON path (gjson/sjson syntax) -> value
 #         "reasoning.effort": "high"
+#   override-raw: # Override raw rules always set parameters using raw JSON (must be valid JSON).
+#     - models:
+#         - name: "gpt-*" # Supports wildcards (e.g., "gpt-*")
+#           protocol: "codex" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex
+#       params: # JSON path (gjson/sjson syntax) -> raw JSON value (strings are used as-is, must be valid JSON)
+#         "response_format": "{\"type\":\"json_schema\",\"json_schema\":{\"name\":\"answer\",\"schema\":{\"type\":\"object\"}}}"

docker-build.sh

@@ -152,16 +152,16 @@ case "$choice" in
     # Build and start the services with a local-only image tag
     export CLI_PROXY_IMAGE="cli-proxy-api:local"
 
-    if [[ "${WITH_USAGE}" == "true" ]]; then
-      export_stats
-    fi
-
     echo "Building the Docker image..."
     docker compose build \
       --build-arg VERSION="${VERSION}" \
       --build-arg COMMIT="${COMMIT}" \
       --build-arg BUILD_DATE="${BUILD_DATE}"
 
+    if [[ "${WITH_USAGE}" == "true" ]]; then
+      export_stats
+    fi
+
     echo "Starting the services..."
     docker compose up -d --remove-orphans --pull never
 

docker-compose.yml

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

examples/custom-provider/main.go

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

examples/http-request/main.go

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

internal/api/handlers/management/api_tools.go

@@ -33,6 +33,13 @@ var geminiOAuthScopes = []string{
 	"https://www.googleapis.com/auth/userinfo.profile",
 }
 
+const (
+	antigravityOAuthClientID     = "1071006060591-tmhssin2h21lcre235vtolojh4g403ep.apps.googleusercontent.com"
+	antigravityOAuthClientSecret = "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"
+)
+
+var antigravityOAuthTokenURL = "https://oauth2.googleapis.com/token"
+
 type apiCallRequest struct {
 	AuthIndexSnake  *string           `json:"auth_index"`
 	AuthIndexCamel  *string           `json:"authIndex"`
@@ -251,6 +258,10 @@ func (h *Handler) resolveTokenForAuth(ctx context.Context, auth *coreauth.Auth)
 		token, errToken := h.refreshGeminiOAuthAccessToken(ctx, auth)
 		return token, errToken
 	}
+	if provider == "antigravity" {
+		token, errToken := h.refreshAntigravityOAuthAccessToken(ctx, auth)
+		return token, errToken
+	}
 
 	return tokenValueForAuth(auth), nil
 }
@@ -325,6 +336,161 @@ func (h *Handler) refreshGeminiOAuthAccessToken(ctx context.Context, auth *corea
 	return strings.TrimSpace(currentToken.AccessToken), nil
 }
 
+func (h *Handler) refreshAntigravityOAuthAccessToken(ctx context.Context, auth *coreauth.Auth) (string, error) {
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	if auth == nil {
+		return "", nil
+	}
+
+	metadata := auth.Metadata
+	if len(metadata) == 0 {
+		return "", fmt.Errorf("antigravity oauth metadata missing")
+	}
+
+	current := strings.TrimSpace(tokenValueFromMetadata(metadata))
+	if current != "" && !antigravityTokenNeedsRefresh(metadata) {
+		return current, nil
+	}
+
+	refreshToken := stringValue(metadata, "refresh_token")
+	if refreshToken == "" {
+		return "", fmt.Errorf("antigravity refresh token missing")
+	}
+
+	tokenURL := strings.TrimSpace(antigravityOAuthTokenURL)
+	if tokenURL == "" {
+		tokenURL = "https://oauth2.googleapis.com/token"
+	}
+	form := url.Values{}
+	form.Set("client_id", antigravityOAuthClientID)
+	form.Set("client_secret", antigravityOAuthClientSecret)
+	form.Set("grant_type", "refresh_token")
+	form.Set("refresh_token", refreshToken)
+
+	req, errReq := http.NewRequestWithContext(ctx, http.MethodPost, tokenURL, strings.NewReader(form.Encode()))
+	if errReq != nil {
+		return "", errReq
+	}
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+
+	httpClient := &http.Client{
+		Timeout:   defaultAPICallTimeout,
+		Transport: h.apiCallTransport(auth),
+	}
+	resp, errDo := httpClient.Do(req)
+	if errDo != nil {
+		return "", errDo
+	}
+	defer func() {
+		if errClose := resp.Body.Close(); errClose != nil {
+			log.Errorf("response body close error: %v", errClose)
+		}
+	}()
+
+	bodyBytes, errRead := io.ReadAll(resp.Body)
+	if errRead != nil {
+		return "", errRead
+	}
+	if resp.StatusCode < http.StatusOK || resp.StatusCode >= http.StatusMultipleChoices {
+		return "", fmt.Errorf("antigravity oauth token refresh failed: status %d: %s", resp.StatusCode, strings.TrimSpace(string(bodyBytes)))
+	}
+
+	var tokenResp struct {
+		AccessToken  string `json:"access_token"`
+		RefreshToken string `json:"refresh_token"`
+		ExpiresIn    int64  `json:"expires_in"`
+		TokenType    string `json:"token_type"`
+	}
+	if errUnmarshal := json.Unmarshal(bodyBytes, &tokenResp); errUnmarshal != nil {
+		return "", errUnmarshal
+	}
+
+	if strings.TrimSpace(tokenResp.AccessToken) == "" {
+		return "", fmt.Errorf("antigravity oauth token refresh returned empty access_token")
+	}
+
+	if auth.Metadata == nil {
+		auth.Metadata = make(map[string]any)
+	}
+	now := time.Now()
+	auth.Metadata["access_token"] = strings.TrimSpace(tokenResp.AccessToken)
+	if strings.TrimSpace(tokenResp.RefreshToken) != "" {
+		auth.Metadata["refresh_token"] = strings.TrimSpace(tokenResp.RefreshToken)
+	}
+	if tokenResp.ExpiresIn > 0 {
+		auth.Metadata["expires_in"] = tokenResp.ExpiresIn
+		auth.Metadata["timestamp"] = now.UnixMilli()
+		auth.Metadata["expired"] = now.Add(time.Duration(tokenResp.ExpiresIn) * time.Second).Format(time.RFC3339)
+	}
+	auth.Metadata["type"] = "antigravity"
+
+	if h != nil && h.authManager != nil {
+		auth.LastRefreshedAt = now
+		auth.UpdatedAt = now
+		_, _ = h.authManager.Update(ctx, auth)
+	}
+
+	return strings.TrimSpace(tokenResp.AccessToken), nil
+}
+
+func antigravityTokenNeedsRefresh(metadata map[string]any) bool {
+	// Refresh a bit early to avoid requests racing token expiry.
+	const skew = 30 * time.Second
+
+	if metadata == nil {
+		return true
+	}
+	if expStr, ok := metadata["expired"].(string); ok {
+		if ts, errParse := time.Parse(time.RFC3339, strings.TrimSpace(expStr)); errParse == nil {
+			return !ts.After(time.Now().Add(skew))
+		}
+	}
+	expiresIn := int64Value(metadata["expires_in"])
+	timestampMs := int64Value(metadata["timestamp"])
+	if expiresIn > 0 && timestampMs > 0 {
+		exp := time.UnixMilli(timestampMs).Add(time.Duration(expiresIn) * time.Second)
+		return !exp.After(time.Now().Add(skew))
+	}
+	return true
+}
+
+func int64Value(raw any) int64 {
+	switch typed := raw.(type) {
+	case int:
+		return int64(typed)
+	case int32:
+		return int64(typed)
+	case int64:
+		return typed
+	case uint:
+		return int64(typed)
+	case uint32:
+		return int64(typed)
+	case uint64:
+		if typed > uint64(^uint64(0)>>1) {
+			return 0
+		}
+		return int64(typed)
+	case float32:
+		return int64(typed)
+	case float64:
+		return int64(typed)
+	case json.Number:
+		if i, errParse := typed.Int64(); errParse == nil {
+			return i
+		}
+	case string:
+		if s := strings.TrimSpace(typed); s != "" {
+			if i, errParse := json.Number(s).Int64(); errParse == nil {
+				return i
+			}
+		}
+	}
+	return 0
+}
+
 func geminiOAuthMetadata(auth *coreauth.Auth) (map[string]any, func(map[string]any)) {
 	if auth == nil {
 		return nil, nil

internal/api/handlers/management/api_tools_test.go

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

internal/api/handlers/management/auth_files.go

@@ -460,6 +460,12 @@ func extractCodexIDTokenClaims(auth *coreauth.Auth) gin.H {
 	if v := strings.TrimSpace(claims.CodexAuthInfo.ChatgptPlanType); v != "" {
 		result["plan_type"] = v
 	}
+	if v := claims.CodexAuthInfo.ChatgptSubscriptionActiveStart; v != nil {
+		result["chatgpt_subscription_active_start"] = v
+	}
+	if v := claims.CodexAuthInfo.ChatgptSubscriptionActiveUntil; v != nil {
+		result["chatgpt_subscription_active_until"] = v
+	}
 
 	if len(result) == 0 {
 		return nil
@@ -1697,7 +1703,7 @@ func (h *Handler) RequestQwenToken(c *gin.Context) {
 		// Create token storage
 		tokenStorage := qwenAuth.CreateTokenStorage(tokenData)
 
-		tokenStorage.Email = fmt.Sprintf("qwen-%d", time.Now().UnixMilli())
+		tokenStorage.Email = fmt.Sprintf("%d", time.Now().UnixMilli())
 		record := &coreauth.Auth{
 			ID:       fmt.Sprintf("qwen-%s.json", tokenStorage.Email),
 			Provider: "qwen",
@@ -1802,7 +1808,7 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 		tokenStorage := authSvc.CreateTokenStorage(tokenData)
 		identifier := strings.TrimSpace(tokenStorage.Email)
 		if identifier == "" {
-			identifier = fmt.Sprintf("iflow-%d", time.Now().UnixMilli())
+			identifier = fmt.Sprintf("%d", time.Now().UnixMilli())
 			tokenStorage.Email = identifier
 		}
 		record := &coreauth.Auth{
@@ -1887,15 +1893,17 @@ func (h *Handler) RequestIFlowCookieToken(c *gin.Context) {
 	fileName := iflowauth.SanitizeIFlowFileName(email)
 	if fileName == "" {
 		fileName = fmt.Sprintf("iflow-%d", time.Now().UnixMilli())
+	} else {
+		fileName = fmt.Sprintf("iflow-%s", fileName)
 	}
 
 	tokenStorage.Email = email
 	timestamp := time.Now().Unix()
 
 	record := &coreauth.Auth{
-		ID:       fmt.Sprintf("iflow-%s-%d.json", fileName, timestamp),
+		ID:       fmt.Sprintf("%s-%d.json", fileName, timestamp),
 		Provider: "iflow",
-		FileName: fmt.Sprintf("iflow-%s-%d.json", fileName, timestamp),
+		FileName: fmt.Sprintf("%s-%d.json", fileName, timestamp),
 		Storage:  tokenStorage,
 		Metadata: map[string]any{
 			"email":        email,

internal/api/handlers/management/config_basic.go

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

internal/api/handlers/management/config_lists.go

@@ -487,6 +487,137 @@ func (h *Handler) DeleteOpenAICompat(c *gin.Context) {
 	c.JSON(400, gin.H{"error": "missing name or index"})
 }
 
+// vertex-api-key: []VertexCompatKey
+func (h *Handler) GetVertexCompatKeys(c *gin.Context) {
+	c.JSON(200, gin.H{"vertex-api-key": h.cfg.VertexCompatAPIKey})
+}
+func (h *Handler) PutVertexCompatKeys(c *gin.Context) {
+	data, err := c.GetRawData()
+	if err != nil {
+		c.JSON(400, gin.H{"error": "failed to read body"})
+		return
+	}
+	var arr []config.VertexCompatKey
+	if err = json.Unmarshal(data, &arr); err != nil {
+		var obj struct {
+			Items []config.VertexCompatKey `json:"items"`
+		}
+		if err2 := json.Unmarshal(data, &obj); err2 != nil || len(obj.Items) == 0 {
+			c.JSON(400, gin.H{"error": "invalid body"})
+			return
+		}
+		arr = obj.Items
+	}
+	for i := range arr {
+		normalizeVertexCompatKey(&arr[i])
+	}
+	h.cfg.VertexCompatAPIKey = arr
+	h.cfg.SanitizeVertexCompatKeys()
+	h.persist(c)
+}
+func (h *Handler) PatchVertexCompatKey(c *gin.Context) {
+	type vertexCompatPatch struct {
+		APIKey   *string                     `json:"api-key"`
+		Prefix   *string                     `json:"prefix"`
+		BaseURL  *string                     `json:"base-url"`
+		ProxyURL *string                     `json:"proxy-url"`
+		Headers  *map[string]string          `json:"headers"`
+		Models   *[]config.VertexCompatModel `json:"models"`
+	}
+	var body struct {
+		Index *int               `json:"index"`
+		Match *string            `json:"match"`
+		Value *vertexCompatPatch `json:"value"`
+	}
+	if errBindJSON := c.ShouldBindJSON(&body); errBindJSON != nil || body.Value == nil {
+		c.JSON(400, gin.H{"error": "invalid body"})
+		return
+	}
+	targetIndex := -1
+	if body.Index != nil && *body.Index >= 0 && *body.Index < len(h.cfg.VertexCompatAPIKey) {
+		targetIndex = *body.Index
+	}
+	if targetIndex == -1 && body.Match != nil {
+		match := strings.TrimSpace(*body.Match)
+		if match != "" {
+			for i := range h.cfg.VertexCompatAPIKey {
+				if h.cfg.VertexCompatAPIKey[i].APIKey == match {
+					targetIndex = i
+					break
+				}
+			}
+		}
+	}
+	if targetIndex == -1 {
+		c.JSON(404, gin.H{"error": "item not found"})
+		return
+	}
+
+	entry := h.cfg.VertexCompatAPIKey[targetIndex]
+	if body.Value.APIKey != nil {
+		trimmed := strings.TrimSpace(*body.Value.APIKey)
+		if trimmed == "" {
+			h.cfg.VertexCompatAPIKey = append(h.cfg.VertexCompatAPIKey[:targetIndex], h.cfg.VertexCompatAPIKey[targetIndex+1:]...)
+			h.cfg.SanitizeVertexCompatKeys()
+			h.persist(c)
+			return
+		}
+		entry.APIKey = trimmed
+	}
+	if body.Value.Prefix != nil {
+		entry.Prefix = strings.TrimSpace(*body.Value.Prefix)
+	}
+	if body.Value.BaseURL != nil {
+		trimmed := strings.TrimSpace(*body.Value.BaseURL)
+		if trimmed == "" {
+			h.cfg.VertexCompatAPIKey = append(h.cfg.VertexCompatAPIKey[:targetIndex], h.cfg.VertexCompatAPIKey[targetIndex+1:]...)
+			h.cfg.SanitizeVertexCompatKeys()
+			h.persist(c)
+			return
+		}
+		entry.BaseURL = trimmed
+	}
+	if body.Value.ProxyURL != nil {
+		entry.ProxyURL = strings.TrimSpace(*body.Value.ProxyURL)
+	}
+	if body.Value.Headers != nil {
+		entry.Headers = config.NormalizeHeaders(*body.Value.Headers)
+	}
+	if body.Value.Models != nil {
+		entry.Models = append([]config.VertexCompatModel(nil), (*body.Value.Models)...)
+	}
+	normalizeVertexCompatKey(&entry)
+	h.cfg.VertexCompatAPIKey[targetIndex] = entry
+	h.cfg.SanitizeVertexCompatKeys()
+	h.persist(c)
+}
+
+func (h *Handler) DeleteVertexCompatKey(c *gin.Context) {
+	if val := strings.TrimSpace(c.Query("api-key")); val != "" {
+		out := make([]config.VertexCompatKey, 0, len(h.cfg.VertexCompatAPIKey))
+		for _, v := range h.cfg.VertexCompatAPIKey {
+			if v.APIKey != val {
+				out = append(out, v)
+			}
+		}
+		h.cfg.VertexCompatAPIKey = out
+		h.cfg.SanitizeVertexCompatKeys()
+		h.persist(c)
+		return
+	}
+	if idxStr := c.Query("index"); idxStr != "" {
+		var idx int
+		_, errScan := fmt.Sscanf(idxStr, "%d", &idx)
+		if errScan == nil && idx >= 0 && idx < len(h.cfg.VertexCompatAPIKey) {
+			h.cfg.VertexCompatAPIKey = append(h.cfg.VertexCompatAPIKey[:idx], h.cfg.VertexCompatAPIKey[idx+1:]...)
+			h.cfg.SanitizeVertexCompatKeys()
+			h.persist(c)
+			return
+		}
+	}
+	c.JSON(400, gin.H{"error": "missing api-key or index"})
+}
+
 // oauth-excluded-models: map[string][]string
 func (h *Handler) GetOAuthExcludedModels(c *gin.Context) {
 	c.JSON(200, gin.H{"oauth-excluded-models": config.NormalizeOAuthExcludedModels(h.cfg.OAuthExcludedModels)})
@@ -572,6 +703,103 @@ func (h *Handler) DeleteOAuthExcludedModels(c *gin.Context) {
 	h.persist(c)
 }
 
+// oauth-model-alias: map[string][]OAuthModelAlias
+func (h *Handler) GetOAuthModelAlias(c *gin.Context) {
+	c.JSON(200, gin.H{"oauth-model-alias": sanitizedOAuthModelAlias(h.cfg.OAuthModelAlias)})
+}
+
+func (h *Handler) PutOAuthModelAlias(c *gin.Context) {
+	data, err := c.GetRawData()
+	if err != nil {
+		c.JSON(400, gin.H{"error": "failed to read body"})
+		return
+	}
+	var entries map[string][]config.OAuthModelAlias
+	if err = json.Unmarshal(data, &entries); err != nil {
+		var wrapper struct {
+			Items map[string][]config.OAuthModelAlias `json:"items"`
+		}
+		if err2 := json.Unmarshal(data, &wrapper); err2 != nil {
+			c.JSON(400, gin.H{"error": "invalid body"})
+			return
+		}
+		entries = wrapper.Items
+	}
+	h.cfg.OAuthModelAlias = sanitizedOAuthModelAlias(entries)
+	h.persist(c)
+}
+
+func (h *Handler) PatchOAuthModelAlias(c *gin.Context) {
+	var body struct {
+		Provider *string                  `json:"provider"`
+		Channel  *string                  `json:"channel"`
+		Aliases  []config.OAuthModelAlias `json:"aliases"`
+	}
+	if errBindJSON := c.ShouldBindJSON(&body); errBindJSON != nil {
+		c.JSON(400, gin.H{"error": "invalid body"})
+		return
+	}
+	channelRaw := ""
+	if body.Channel != nil {
+		channelRaw = *body.Channel
+	} else if body.Provider != nil {
+		channelRaw = *body.Provider
+	}
+	channel := strings.ToLower(strings.TrimSpace(channelRaw))
+	if channel == "" {
+		c.JSON(400, gin.H{"error": "invalid channel"})
+		return
+	}
+
+	normalizedMap := sanitizedOAuthModelAlias(map[string][]config.OAuthModelAlias{channel: body.Aliases})
+	normalized := normalizedMap[channel]
+	if len(normalized) == 0 {
+		if h.cfg.OAuthModelAlias == nil {
+			c.JSON(404, gin.H{"error": "channel not found"})
+			return
+		}
+		if _, ok := h.cfg.OAuthModelAlias[channel]; !ok {
+			c.JSON(404, gin.H{"error": "channel not found"})
+			return
+		}
+		delete(h.cfg.OAuthModelAlias, channel)
+		if len(h.cfg.OAuthModelAlias) == 0 {
+			h.cfg.OAuthModelAlias = nil
+		}
+		h.persist(c)
+		return
+	}
+	if h.cfg.OAuthModelAlias == nil {
+		h.cfg.OAuthModelAlias = make(map[string][]config.OAuthModelAlias)
+	}
+	h.cfg.OAuthModelAlias[channel] = normalized
+	h.persist(c)
+}
+
+func (h *Handler) DeleteOAuthModelAlias(c *gin.Context) {
+	channel := strings.ToLower(strings.TrimSpace(c.Query("channel")))
+	if channel == "" {
+		channel = strings.ToLower(strings.TrimSpace(c.Query("provider")))
+	}
+	if channel == "" {
+		c.JSON(400, gin.H{"error": "missing channel"})
+		return
+	}
+	if h.cfg.OAuthModelAlias == nil {
+		c.JSON(404, gin.H{"error": "channel not found"})
+		return
+	}
+	if _, ok := h.cfg.OAuthModelAlias[channel]; !ok {
+		c.JSON(404, gin.H{"error": "channel not found"})
+		return
+	}
+	delete(h.cfg.OAuthModelAlias, channel)
+	if len(h.cfg.OAuthModelAlias) == 0 {
+		h.cfg.OAuthModelAlias = nil
+	}
+	h.persist(c)
+}
+
 // codex-api-key: []CodexKey
 func (h *Handler) GetCodexKeys(c *gin.Context) {
 	c.JSON(200, gin.H{"codex-api-key": h.cfg.CodexKey})
@@ -789,6 +1017,53 @@ func normalizeCodexKey(entry *config.CodexKey) {
 	entry.Models = normalized
 }
 
+func normalizeVertexCompatKey(entry *config.VertexCompatKey) {
+	if entry == nil {
+		return
+	}
+	entry.APIKey = strings.TrimSpace(entry.APIKey)
+	entry.Prefix = strings.TrimSpace(entry.Prefix)
+	entry.BaseURL = strings.TrimSpace(entry.BaseURL)
+	entry.ProxyURL = strings.TrimSpace(entry.ProxyURL)
+	entry.Headers = config.NormalizeHeaders(entry.Headers)
+	if len(entry.Models) == 0 {
+		return
+	}
+	normalized := make([]config.VertexCompatModel, 0, len(entry.Models))
+	for i := range entry.Models {
+		model := entry.Models[i]
+		model.Name = strings.TrimSpace(model.Name)
+		model.Alias = strings.TrimSpace(model.Alias)
+		if model.Name == "" || model.Alias == "" {
+			continue
+		}
+		normalized = append(normalized, model)
+	}
+	entry.Models = normalized
+}
+
+func sanitizedOAuthModelAlias(entries map[string][]config.OAuthModelAlias) map[string][]config.OAuthModelAlias {
+	if len(entries) == 0 {
+		return nil
+	}
+	copied := make(map[string][]config.OAuthModelAlias, len(entries))
+	for channel, aliases := range entries {
+		if len(aliases) == 0 {
+			continue
+		}
+		copied[channel] = append([]config.OAuthModelAlias(nil), aliases...)
+	}
+	if len(copied) == 0 {
+		return nil
+	}
+	cfg := config.Config{OAuthModelAlias: copied}
+	cfg.SanitizeOAuthModelAlias()
+	if len(cfg.OAuthModelAlias) == 0 {
+		return nil
+	}
+	return cfg.OAuthModelAlias
+}
+
 // GetAmpCode returns the complete ampcode configuration.
 func (h *Handler) GetAmpCode(c *gin.Context) {
 	if h == nil || h.cfg == nil {

internal/api/handlers/management/handler.go

@@ -24,8 +24,15 @@ import (
 type attemptInfo struct {
 	count        int
 	blockedUntil time.Time
+	lastActivity time.Time // track last activity for cleanup
 }
 
+// attemptCleanupInterval controls how often stale IP entries are purged
+const attemptCleanupInterval = 1 * time.Hour
+
+// attemptMaxIdleTime controls how long an IP can be idle before cleanup
+const attemptMaxIdleTime = 2 * time.Hour
+
 // Handler aggregates config reference, persistence path and helpers.
 type Handler struct {
 	cfg                 *config.Config
@@ -47,7 +54,7 @@ func NewHandler(cfg *config.Config, configFilePath string, manager *coreauth.Man
 	envSecret, _ := os.LookupEnv("MANAGEMENT_PASSWORD")
 	envSecret = strings.TrimSpace(envSecret)
 
-	return &Handler{
+	h := &Handler{
 		cfg:                 cfg,
 		configFilePath:      configFilePath,
 		failedAttempts:      make(map[string]*attemptInfo),
@@ -57,6 +64,38 @@ func NewHandler(cfg *config.Config, configFilePath string, manager *coreauth.Man
 		allowRemoteOverride: envSecret != "",
 		envSecret:           envSecret,
 	}
+	h.startAttemptCleanup()
+	return h
+}
+
+// startAttemptCleanup launches a background goroutine that periodically
+// removes stale IP entries from failedAttempts to prevent memory leaks.
+func (h *Handler) startAttemptCleanup() {
+	go func() {
+		ticker := time.NewTicker(attemptCleanupInterval)
+		defer ticker.Stop()
+		for range ticker.C {
+			h.purgeStaleAttempts()
+		}
+	}()
+}
+
+// purgeStaleAttempts removes IP entries that have been idle beyond attemptMaxIdleTime
+// and whose ban (if any) has expired.
+func (h *Handler) purgeStaleAttempts() {
+	now := time.Now()
+	h.attemptsMu.Lock()
+	defer h.attemptsMu.Unlock()
+	for ip, ai := range h.failedAttempts {
+		// Skip if still banned
+		if !ai.blockedUntil.IsZero() && now.Before(ai.blockedUntil) {
+			continue
+		}
+		// Remove if idle too long
+		if now.Sub(ai.lastActivity) > attemptMaxIdleTime {
+			delete(h.failedAttempts, ip)
+		}
+	}
 }
 
 // NewHandler creates a new management handler instance.
@@ -149,6 +188,7 @@ func (h *Handler) Middleware() gin.HandlerFunc {
 					h.failedAttempts[clientIP] = aip
 				}
 				aip.count++
+				aip.lastActivity = time.Now()
 				if aip.count >= maxFailures {
 					aip.blockedUntil = time.Now().Add(banDuration)
 					aip.count = 0

internal/api/handlers/management/logs.go

@@ -13,7 +13,7 @@ import (
 	"time"
 
 	"github.com/gin-gonic/gin"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/logging"
 )
 
 const (
@@ -360,16 +360,7 @@ func (h *Handler) logDirectory() string {
 	if h.logDir != "" {
 		return h.logDir
 	}
-	if base := util.WritablePath(); base != "" {
-		return filepath.Join(base, "logs")
-	}
-	if h.configFilePath != "" {
-		dir := filepath.Dir(h.configFilePath)
-		if dir != "" && dir != "." {
-			return filepath.Join(dir, "logs")
-		}
-	}
-	return "logs"
+	return logging.ResolveLogDirectory(h.cfg)
 }
 
 func (h *Handler) collectLogFiles(dir string) ([]string, error) {

internal/api/modules/amp/fallback_handlers.go

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

internal/api/modules/amp/model_mapping.go

@@ -8,6 +8,7 @@ import (
 	"sync"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	log "github.com/sirupsen/logrus"
 )
@@ -44,6 +45,11 @@ func NewModelMapper(mappings []config.AmpModelMapping) *DefaultModelMapper {
 // MapModel checks if a mapping exists for the requested model and if the
 // target model has available local providers. Returns the mapped model name
 // or empty string if no valid mapping exists.
+//
+// If the requested model contains a thinking suffix (e.g., "g25p(8192)"),
+// the suffix is preserved in the returned model name (e.g., "gemini-2.5-pro(8192)").
+// However, if the mapping target already contains a suffix, the config suffix
+// takes priority over the user's suffix.
 func (m *DefaultModelMapper) MapModel(requestedModel string) string {
 	if requestedModel == "" {
 		return ""
@@ -52,16 +58,20 @@ func (m *DefaultModelMapper) MapModel(requestedModel string) string {
 	m.mu.RLock()
 	defer m.mu.RUnlock()
 
-	// Normalize the requested model for lookup
-	normalizedRequest := strings.ToLower(strings.TrimSpace(requestedModel))
+	// Extract thinking suffix from requested model using ParseSuffix
+	requestResult := thinking.ParseSuffix(requestedModel)
+	baseModel := requestResult.ModelName
 
-	// Check for direct mapping
-	targetModel, exists := m.mappings[normalizedRequest]
+	// Normalize the base model for lookup (case-insensitive)
+	normalizedBase := strings.ToLower(strings.TrimSpace(baseModel))
+
+	// Check for direct mapping using base model name
+	targetModel, exists := m.mappings[normalizedBase]
 	if !exists {
-		// Try regex mappings in order
-		base, _ := util.NormalizeThinkingModel(requestedModel)
+		// Try regex mappings in order using base model only
+		// (suffix is handled separately via ParseSuffix)
 		for _, rm := range m.regexps {
-			if rm.re.MatchString(requestedModel) || (base != "" && rm.re.MatchString(base)) {
+			if rm.re.MatchString(baseModel) {
 				targetModel = rm.to
 				exists = true
 				break
@@ -72,14 +82,28 @@ func (m *DefaultModelMapper) MapModel(requestedModel string) string {
 		}
 	}
 
-	// Verify target model has available providers
-	normalizedTarget, _ := util.NormalizeThinkingModel(targetModel)
-	providers := util.GetProviderName(normalizedTarget)
+	// Check if target model already has a thinking suffix (config priority)
+	targetResult := thinking.ParseSuffix(targetModel)
+
+	// Verify target model has available providers (use base model for lookup)
+	providers := util.GetProviderName(targetResult.ModelName)
 	if len(providers) == 0 {
 		log.Debugf("amp model mapping: target model %s has no available providers, skipping mapping", targetModel)
 		return ""
 	}
 
+	// Suffix handling: config suffix takes priority, otherwise preserve user suffix
+	if targetResult.HasSuffix {
+		// Config's "to" already contains a suffix - use it as-is (config priority)
+		return targetModel
+	}
+
+	// Preserve user's thinking suffix on the mapped model
+	// (skip empty suffixes to avoid returning "model()")
+	if requestResult.HasSuffix && requestResult.RawSuffix != "" {
+		return targetModel + "(" + requestResult.RawSuffix + ")"
+	}
+
 	// Note: Detailed routing log is handled by logAmpRouting in fallback_handlers.go
 	return targetModel
 }

internal/api/modules/amp/model_mapping_test.go

@@ -217,10 +217,10 @@ func TestModelMapper_Regex_MatchBaseWithoutParens(t *testing.T) {
 
 	mapper := NewModelMapper(mappings)
 
-	// Incoming model has reasoning suffix but should match base via regex
+	// Incoming model has reasoning suffix, regex matches base, suffix is preserved
 	result := mapper.MapModel("gpt-5(high)")
-	if result != "gemini-2.5-pro" {
-		t.Errorf("Expected gemini-2.5-pro, got %s", result)
+	if result != "gemini-2.5-pro(high)" {
+		t.Errorf("Expected gemini-2.5-pro(high), got %s", result)
 	}
 }
 
@@ -281,3 +281,95 @@ func TestModelMapper_Regex_CaseInsensitive(t *testing.T) {
 		t.Errorf("Expected claude-sonnet-4, got %s", result)
 	}
 }
+
+func TestModelMapper_SuffixPreservation(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+
+	// Register test models
+	reg.RegisterClient("test-client-suffix", "gemini", []*registry.ModelInfo{
+		{ID: "gemini-2.5-pro", OwnedBy: "google", Type: "gemini"},
+	})
+	reg.RegisterClient("test-client-suffix-2", "claude", []*registry.ModelInfo{
+		{ID: "claude-sonnet-4", OwnedBy: "anthropic", Type: "claude"},
+	})
+	defer reg.UnregisterClient("test-client-suffix")
+	defer reg.UnregisterClient("test-client-suffix-2")
+
+	tests := []struct {
+		name     string
+		mappings []config.AmpModelMapping
+		input    string
+		want     string
+	}{
+		{
+			name:     "numeric suffix preserved",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p(8192)",
+			want:     "gemini-2.5-pro(8192)",
+		},
+		{
+			name:     "level suffix preserved",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p(high)",
+			want:     "gemini-2.5-pro(high)",
+		},
+		{
+			name:     "no suffix unchanged",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p",
+			want:     "gemini-2.5-pro",
+		},
+		{
+			name:     "config suffix takes priority",
+			mappings: []config.AmpModelMapping{{From: "alias", To: "gemini-2.5-pro(medium)"}},
+			input:    "alias(high)",
+			want:     "gemini-2.5-pro(medium)",
+		},
+		{
+			name:     "regex with suffix preserved",
+			mappings: []config.AmpModelMapping{{From: "^g25.*", To: "gemini-2.5-pro", Regex: true}},
+			input:    "g25p(8192)",
+			want:     "gemini-2.5-pro(8192)",
+		},
+		{
+			name:     "auto suffix preserved",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p(auto)",
+			want:     "gemini-2.5-pro(auto)",
+		},
+		{
+			name:     "none suffix preserved",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p(none)",
+			want:     "gemini-2.5-pro(none)",
+		},
+		{
+			name:     "case insensitive base lookup with suffix",
+			mappings: []config.AmpModelMapping{{From: "G25P", To: "gemini-2.5-pro"}},
+			input:    "g25p(high)",
+			want:     "gemini-2.5-pro(high)",
+		},
+		{
+			name:     "empty suffix filtered out",
+			mappings: []config.AmpModelMapping{{From: "g25p", To: "gemini-2.5-pro"}},
+			input:    "g25p()",
+			want:     "gemini-2.5-pro",
+		},
+		{
+			name:     "incomplete suffix treated as no suffix",
+			mappings: []config.AmpModelMapping{{From: "g25p(high", To: "gemini-2.5-pro"}},
+			input:    "g25p(high",
+			want:     "gemini-2.5-pro",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			mapper := NewModelMapper(tt.mappings)
+			got := mapper.MapModel(tt.input)
+			if got != tt.want {
+				t.Errorf("MapModel(%q) = %q, want %q", tt.input, got, tt.want)
+			}
+		})
+	}
+}

internal/api/modules/amp/response_rewriter.go

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

internal/api/server.go

@@ -26,6 +26,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/logging"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/managementasset"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/usage"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	sdkaccess "github.com/router-for-me/CLIProxyAPI/v6/sdk/access"
@@ -254,15 +255,13 @@ func NewServer(cfg *config.Config, authManager *auth.Manager, accessManager *sdk
 	}
 	managementasset.SetCurrentConfig(cfg)
 	auth.SetQuotaCooldownDisabled(cfg.DisableCooling)
+	misc.SetCodexInstructionsEnabled(cfg.CodexInstructionsEnabled)
 	// Initialize management handler
 	s.mgmt = managementHandlers.NewHandler(cfg, configFilePath, authManager)
 	if optionState.localPassword != "" {
 		s.mgmt.SetLocalPassword(optionState.localPassword)
 	}
-	logDir := filepath.Join(s.currentPath, "logs")
-	if base := util.WritablePath(); base != "" {
-		logDir = filepath.Join(base, "logs")
-	}
+	logDir := logging.ResolveLogDirectory(cfg)
 	s.mgmt.SetLogDirectory(logDir)
 	s.localPassword = optionState.localPassword
 
@@ -492,6 +491,10 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.PUT("/logging-to-file", s.mgmt.PutLoggingToFile)
 		mgmt.PATCH("/logging-to-file", s.mgmt.PutLoggingToFile)
 
+		mgmt.GET("/logs-max-total-size-mb", s.mgmt.GetLogsMaxTotalSizeMB)
+		mgmt.PUT("/logs-max-total-size-mb", s.mgmt.PutLogsMaxTotalSizeMB)
+		mgmt.PATCH("/logs-max-total-size-mb", s.mgmt.PutLogsMaxTotalSizeMB)
+
 		mgmt.GET("/usage-statistics-enabled", s.mgmt.GetUsageStatisticsEnabled)
 		mgmt.PUT("/usage-statistics-enabled", s.mgmt.PutUsageStatisticsEnabled)
 		mgmt.PATCH("/usage-statistics-enabled", s.mgmt.PutUsageStatisticsEnabled)
@@ -564,6 +567,14 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.PUT("/max-retry-interval", s.mgmt.PutMaxRetryInterval)
 		mgmt.PATCH("/max-retry-interval", s.mgmt.PutMaxRetryInterval)
 
+		mgmt.GET("/force-model-prefix", s.mgmt.GetForceModelPrefix)
+		mgmt.PUT("/force-model-prefix", s.mgmt.PutForceModelPrefix)
+		mgmt.PATCH("/force-model-prefix", s.mgmt.PutForceModelPrefix)
+
+		mgmt.GET("/routing/strategy", s.mgmt.GetRoutingStrategy)
+		mgmt.PUT("/routing/strategy", s.mgmt.PutRoutingStrategy)
+		mgmt.PATCH("/routing/strategy", s.mgmt.PutRoutingStrategy)
+
 		mgmt.GET("/claude-api-key", s.mgmt.GetClaudeKeys)
 		mgmt.PUT("/claude-api-key", s.mgmt.PutClaudeKeys)
 		mgmt.PATCH("/claude-api-key", s.mgmt.PatchClaudeKey)
@@ -579,11 +590,21 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.PATCH("/openai-compatibility", s.mgmt.PatchOpenAICompat)
 		mgmt.DELETE("/openai-compatibility", s.mgmt.DeleteOpenAICompat)
 
+		mgmt.GET("/vertex-api-key", s.mgmt.GetVertexCompatKeys)
+		mgmt.PUT("/vertex-api-key", s.mgmt.PutVertexCompatKeys)
+		mgmt.PATCH("/vertex-api-key", s.mgmt.PatchVertexCompatKey)
+		mgmt.DELETE("/vertex-api-key", s.mgmt.DeleteVertexCompatKey)
+
 		mgmt.GET("/oauth-excluded-models", s.mgmt.GetOAuthExcludedModels)
 		mgmt.PUT("/oauth-excluded-models", s.mgmt.PutOAuthExcludedModels)
 		mgmt.PATCH("/oauth-excluded-models", s.mgmt.PatchOAuthExcludedModels)
 		mgmt.DELETE("/oauth-excluded-models", s.mgmt.DeleteOAuthExcludedModels)
 
+		mgmt.GET("/oauth-model-alias", s.mgmt.GetOAuthModelAlias)
+		mgmt.PUT("/oauth-model-alias", s.mgmt.PutOAuthModelAlias)
+		mgmt.PATCH("/oauth-model-alias", s.mgmt.PatchOAuthModelAlias)
+		mgmt.DELETE("/oauth-model-alias", s.mgmt.DeleteOAuthModelAlias)
+
 		mgmt.GET("/auth-files", s.mgmt.ListAuthFiles)
 		mgmt.GET("/auth-files/models", s.mgmt.GetAuthFileModels)
 		mgmt.GET("/auth-files/download", s.mgmt.DownloadAuthFile)
@@ -890,6 +911,16 @@ func (s *Server) UpdateClients(cfg *config.Config) {
 			log.Debugf("disable_cooling toggled to %t", cfg.DisableCooling)
 		}
 	}
+
+	if oldCfg == nil || oldCfg.CodexInstructionsEnabled != cfg.CodexInstructionsEnabled {
+		misc.SetCodexInstructionsEnabled(cfg.CodexInstructionsEnabled)
+		if oldCfg != nil {
+			log.Debugf("codex_instructions_enabled updated from %t to %t", oldCfg.CodexInstructionsEnabled, cfg.CodexInstructionsEnabled)
+		} else {
+			log.Debugf("codex_instructions_enabled toggled to %t", cfg.CodexInstructionsEnabled)
+		}
+	}
+
 	if s.handlers != nil && s.handlers.AuthManager != nil {
 		s.handlers.AuthManager.SetRetryConfig(cfg.RequestRetry, time.Duration(cfg.MaxRetryInterval)*time.Second)
 	}

internal/auth/gemini/gemini_auth.go

@@ -29,8 +29,9 @@ import (
 )
 
 const (
-	geminiOauthClientID     = "681255809395-oo8ft2oprdrnp9e3aqf6av3hmdib135j.apps.googleusercontent.com"
-	geminiOauthClientSecret = "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"
+	geminiOauthClientID       = "681255809395-oo8ft2oprdrnp9e3aqf6av3hmdib135j.apps.googleusercontent.com"
+	geminiOauthClientSecret   = "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"
+	geminiDefaultCallbackPort = 8085
 )
 
 var (
@@ -49,8 +50,9 @@ type GeminiAuth struct {
 
 // WebLoginOptions customizes the interactive OAuth flow.
 type WebLoginOptions struct {
-	NoBrowser bool
-	Prompt    func(string) (string, error)
+	NoBrowser    bool
+	CallbackPort int
+	Prompt       func(string) (string, error)
 }
 
 // NewGeminiAuth creates a new instance of GeminiAuth.
@@ -72,6 +74,12 @@ func NewGeminiAuth() *GeminiAuth {
 //   - *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, opts *WebLoginOptions) (*http.Client, error) {
+	callbackPort := geminiDefaultCallbackPort
+	if opts != nil && opts.CallbackPort > 0 {
+		callbackPort = opts.CallbackPort
+	}
+	callbackURL := fmt.Sprintf("http://localhost:%d/oauth2callback", callbackPort)
+
 	// Configure proxy settings for the HTTP client if a proxy URL is provided.
 	proxyURL, err := url.Parse(cfg.ProxyURL)
 	if err == nil {
@@ -106,7 +114,7 @@ func (g *GeminiAuth) GetAuthenticatedClient(ctx context.Context, ts *GeminiToken
 	conf := &oauth2.Config{
 		ClientID:     geminiOauthClientID,
 		ClientSecret: geminiOauthClientSecret,
-		RedirectURL:  "http://localhost:8085/oauth2callback", // This will be used by the local server.
+		RedirectURL:  callbackURL, // This will be used by the local server.
 		Scopes:       geminiOauthScopes,
 		Endpoint:     google.Endpoint,
 	}
@@ -218,14 +226,20 @@ func (g *GeminiAuth) createTokenStorage(ctx context.Context, config *oauth2.Conf
 //   - *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, opts *WebLoginOptions) (*oauth2.Token, error) {
+	callbackPort := geminiDefaultCallbackPort
+	if opts != nil && opts.CallbackPort > 0 {
+		callbackPort = opts.CallbackPort
+	}
+	callbackURL := fmt.Sprintf("http://localhost:%d/oauth2callback", callbackPort)
+
 	// Use a channel to pass the authorization code from the HTTP handler to the main function.
 	codeChan := make(chan string, 1)
 	errChan := make(chan error, 1)
 
 	// Create a new HTTP server with its own multiplexer.
 	mux := http.NewServeMux()
-	server := &http.Server{Addr: ":8085", Handler: mux}
-	config.RedirectURL = "http://localhost:8085/oauth2callback"
+	server := &http.Server{Addr: fmt.Sprintf(":%d", callbackPort), Handler: mux}
+	config.RedirectURL = callbackURL
 
 	mux.HandleFunc("/oauth2callback", func(w http.ResponseWriter, r *http.Request) {
 		if err := r.URL.Query().Get("error"); err != "" {
@@ -277,13 +291,13 @@ func (g *GeminiAuth) getTokenFromWeb(ctx context.Context, config *oauth2.Config,
 		// Check if browser is available
 		if !browser.IsAvailable() {
 			log.Warn("No browser available on this system")
-			util.PrintSSHTunnelInstructions(8085)
+			util.PrintSSHTunnelInstructions(callbackPort)
 			fmt.Printf("Please manually open this URL in your browser:\n\n%s\n", authURL)
 		} else {
 			if err := browser.OpenURL(authURL); err != nil {
 				authErr := codex.NewAuthenticationError(codex.ErrBrowserOpenFailed, err)
 				log.Warn(codex.GetUserFriendlyMessage(authErr))
-				util.PrintSSHTunnelInstructions(8085)
+				util.PrintSSHTunnelInstructions(callbackPort)
 				fmt.Printf("Please manually open this URL in your browser:\n\n%s\n", authURL)
 
 				// Log platform info for debugging
@@ -294,7 +308,7 @@ func (g *GeminiAuth) getTokenFromWeb(ctx context.Context, config *oauth2.Config,
 			}
 		}
 	} else {
-		util.PrintSSHTunnelInstructions(8085)
+		util.PrintSSHTunnelInstructions(callbackPort)
 		fmt.Printf("Please open this URL in your browser:\n\n%s\n", authURL)
 	}
 

internal/cache/signature_cache.go

@@ -3,7 +3,6 @@ package cache
 import (
 	"crypto/sha256"
 	"encoding/hex"
-	"sort"
 	"sync"
 	"time"
 )
@@ -16,21 +15,24 @@ type SignatureEntry struct {
 
 const (
 	// SignatureCacheTTL is how long signatures are valid
-	SignatureCacheTTL = 1 * time.Hour
-
-	// MaxEntriesPerSession limits memory usage per session
-	MaxEntriesPerSession = 100
+	SignatureCacheTTL = 3 * time.Hour
 
 	// SignatureTextHashLen is the length of the hash key (16 hex chars = 64-bit key space)
 	SignatureTextHashLen = 16
 
 	// MinValidSignatureLen is the minimum length for a signature to be considered valid
 	MinValidSignatureLen = 50
+
+	// SessionCleanupInterval controls how often stale sessions are purged
+	SessionCleanupInterval = 10 * time.Minute
 )
 
 // signatureCache stores signatures by sessionId -> textHash -> SignatureEntry
 var signatureCache sync.Map
 
+// sessionCleanupOnce ensures the background cleanup goroutine starts only once
+var sessionCleanupOnce sync.Once
+
 // sessionCache is the inner map type
 type sessionCache struct {
 	mu      sync.RWMutex
@@ -45,6 +47,9 @@ func hashText(text string) string {
 
 // getOrCreateSession gets or creates a session cache
 func getOrCreateSession(sessionID string) *sessionCache {
+	// Start background cleanup on first access
+	sessionCleanupOnce.Do(startSessionCleanup)
+
 	if val, ok := signatureCache.Load(sessionID); ok {
 		return val.(*sessionCache)
 	}
@@ -53,6 +58,40 @@ func getOrCreateSession(sessionID string) *sessionCache {
 	return actual.(*sessionCache)
 }
 
+// startSessionCleanup launches a background goroutine that periodically
+// removes sessions where all entries have expired.
+func startSessionCleanup() {
+	go func() {
+		ticker := time.NewTicker(SessionCleanupInterval)
+		defer ticker.Stop()
+		for range ticker.C {
+			purgeExpiredSessions()
+		}
+	}()
+}
+
+// purgeExpiredSessions removes sessions with no valid (non-expired) entries.
+func purgeExpiredSessions() {
+	now := time.Now()
+	signatureCache.Range(func(key, value any) bool {
+		sc := value.(*sessionCache)
+		sc.mu.Lock()
+		// Remove expired entries
+		for k, entry := range sc.entries {
+			if now.Sub(entry.Timestamp) > SignatureCacheTTL {
+				delete(sc.entries, k)
+			}
+		}
+		isEmpty := len(sc.entries) == 0
+		sc.mu.Unlock()
+		// Remove session if empty
+		if isEmpty {
+			signatureCache.Delete(key)
+		}
+		return true
+	})
+}
+
 // CacheSignature stores a thinking signature for a given session and text.
 // Used for Claude models that require signed thinking blocks in multi-turn conversations.
 func CacheSignature(sessionID, text, signature string) {
@@ -69,43 +108,6 @@ func CacheSignature(sessionID, text, signature string) {
 	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(),
@@ -127,22 +129,25 @@ func GetCachedSignature(sessionID, text string) string {
 
 	textHash := hashText(text)
 
-	sc.mu.RLock()
-	entry, exists := sc.entries[textHash]
-	sc.mu.RUnlock()
+	now := time.Now()
 
+	sc.mu.Lock()
+	entry, exists := sc.entries[textHash]
 	if !exists {
+		sc.mu.Unlock()
 		return ""
 	}
-
-	// Check if expired
-	if time.Since(entry.Timestamp) > SignatureCacheTTL {
-		sc.mu.Lock()
+	if now.Sub(entry.Timestamp) > SignatureCacheTTL {
 		delete(sc.entries, textHash)
 		sc.mu.Unlock()
 		return ""
 	}
 
+	// Refresh TTL on access (sliding expiration).
+	entry.Timestamp = now
+	sc.entries[textHash] = entry
+	sc.mu.Unlock()
+
 	return entry.Signature
 }
 

internal/cmd/anthropic_login.go

@@ -32,9 +32,10 @@ func DoClaudeLogin(cfg *config.Config, options *LoginOptions) {
 	manager := newAuthManager()
 
 	authOpts := &sdkAuth.LoginOptions{
-		NoBrowser: options.NoBrowser,
-		Metadata:  map[string]string{},
-		Prompt:    promptFn,
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       promptFn,
 	}
 
 	_, savedPath, err := manager.Login(context.Background(), "claude", cfg, authOpts)

internal/cmd/antigravity_login.go

@@ -22,9 +22,10 @@ func DoAntigravityLogin(cfg *config.Config, options *LoginOptions) {
 
 	manager := newAuthManager()
 	authOpts := &sdkAuth.LoginOptions{
-		NoBrowser: options.NoBrowser,
-		Metadata:  map[string]string{},
-		Prompt:    promptFn,
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       promptFn,
 	}
 
 	record, savedPath, err := manager.Login(context.Background(), "antigravity", cfg, authOpts)

internal/cmd/iflow_login.go

@@ -24,9 +24,10 @@ func DoIFlowLogin(cfg *config.Config, options *LoginOptions) {
 	}
 
 	authOpts := &sdkAuth.LoginOptions{
-		NoBrowser: options.NoBrowser,
-		Metadata:  map[string]string{},
-		Prompt:    promptFn,
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       promptFn,
 	}
 
 	_, savedPath, err := manager.Login(context.Background(), "iflow", cfg, authOpts)

internal/cmd/login.go

@@ -67,10 +67,11 @@ func DoLogin(cfg *config.Config, projectID string, options *LoginOptions) {
 	}
 
 	loginOpts := &sdkAuth.LoginOptions{
-		NoBrowser: options.NoBrowser,
-		ProjectID: trimmedProjectID,
-		Metadata:  map[string]string{},
-		Prompt:    callbackPrompt,
+		NoBrowser:    options.NoBrowser,
+		ProjectID:    trimmedProjectID,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       callbackPrompt,
 	}
 
 	authenticator := sdkAuth.NewGeminiAuthenticator()
@@ -88,8 +89,9 @@ func DoLogin(cfg *config.Config, projectID string, options *LoginOptions) {
 
 	geminiAuth := gemini.NewGeminiAuth()
 	httpClient, errClient := geminiAuth.GetAuthenticatedClient(ctx, storage, cfg, &gemini.WebLoginOptions{
-		NoBrowser: options.NoBrowser,
-		Prompt:    callbackPrompt,
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Prompt:       callbackPrompt,
 	})
 	if errClient != nil {
 		log.Errorf("Gemini authentication failed: %v", errClient)

internal/cmd/openai_login.go

@@ -19,6 +19,9 @@ type LoginOptions struct {
 	// NoBrowser indicates whether to skip opening the browser automatically.
 	NoBrowser bool
 
+	// CallbackPort overrides the local OAuth callback port when set (>0).
+	CallbackPort int
+
 	// Prompt allows the caller to provide interactive input when needed.
 	Prompt func(prompt string) (string, error)
 }
@@ -43,9 +46,10 @@ func DoCodexLogin(cfg *config.Config, options *LoginOptions) {
 	manager := newAuthManager()
 
 	authOpts := &sdkAuth.LoginOptions{
-		NoBrowser: options.NoBrowser,
-		Metadata:  map[string]string{},
-		Prompt:    promptFn,
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       promptFn,
 	}
 
 	_, savedPath, err := manager.Login(context.Background(), "codex", cfg, authOpts)

internal/cmd/qwen_login.go

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

internal/config/config.go

@@ -6,12 +6,14 @@ package config
 
 import (
 	"bytes"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"os"
 	"strings"
 	"syscall"
 
+	log "github.com/sirupsen/logrus"
 	"golang.org/x/crypto/bcrypt"
 	"gopkg.in/yaml.v3"
 )
@@ -69,6 +71,11 @@ type Config struct {
 	// WebsocketAuth enables or disables authentication for the WebSocket API.
 	WebsocketAuth bool `yaml:"ws-auth" json:"ws-auth"`
 
+	// CodexInstructionsEnabled controls whether official Codex instructions are injected.
+	// When false (default), CodexInstructionsForModel returns immediately without modification.
+	// When true, the original instruction injection logic is used.
+	CodexInstructionsEnabled bool `yaml:"codex-instructions-enabled" json:"codex-instructions-enabled"`
+
 	// GeminiKey defines Gemini API key configurations with optional routing overrides.
 	GeminiKey []GeminiKey `yaml:"gemini-api-key" json:"gemini-api-key"`
 
@@ -91,13 +98,13 @@ type Config struct {
 	// OAuthExcludedModels defines per-provider global model exclusions applied to OAuth/file-backed auth entries.
 	OAuthExcludedModels map[string][]string `yaml:"oauth-excluded-models,omitempty" json:"oauth-excluded-models,omitempty"`
 
-	// OAuthModelMappings defines global model name mappings for OAuth/file-backed auth channels.
-	// These mappings affect both model listing and model routing for supported channels:
+	// OAuthModelAlias defines global model name aliases for OAuth/file-backed auth channels.
+	// These aliases affect both model listing and model routing for supported channels:
 	// gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow.
 	//
 	// NOTE: This does not apply to existing per-credential model alias features under:
 	// gemini-api-key, codex-api-key, claude-api-key, openai-compatibility, vertex-api-key, and ampcode.
-	OAuthModelMappings map[string][]ModelNameMapping `yaml:"oauth-model-mappings,omitempty" json:"oauth-model-mappings,omitempty"`
+	OAuthModelAlias map[string][]OAuthModelAlias `yaml:"oauth-model-alias,omitempty" json:"oauth-model-alias,omitempty"`
 
 	// Payload defines default and override rules for provider payload parameters.
 	Payload PayloadConfig `yaml:"payload" json:"payload"`
@@ -145,11 +152,14 @@ type RoutingConfig struct {
 	Strategy string `yaml:"strategy,omitempty" json:"strategy,omitempty"`
 }
 
-// ModelNameMapping defines a model ID rename mapping for a specific channel.
-// It maps the original model name (Name) to the client-visible alias (Alias).
-type ModelNameMapping struct {
+// OAuthModelAlias defines a model ID alias for a specific channel.
+// It maps the upstream model name (Name) to the client-visible alias (Alias).
+// When Fork is true, the alias is added as an additional model in listings while
+// keeping the original model ID available.
+type OAuthModelAlias struct {
 	Name  string `yaml:"name" json:"name"`
 	Alias string `yaml:"alias" json:"alias"`
+	Fork  bool   `yaml:"fork,omitempty" json:"fork,omitempty"`
 }
 
 // AmpModelMapping defines a model name mapping for Amp CLI requests.
@@ -213,8 +223,12 @@ type AmpUpstreamAPIKeyEntry struct {
 type PayloadConfig struct {
 	// Default defines rules that only set parameters when they are missing in the payload.
 	Default []PayloadRule `yaml:"default" json:"default"`
+	// DefaultRaw defines rules that set raw JSON values only when they are missing.
+	DefaultRaw []PayloadRule `yaml:"default-raw" json:"default-raw"`
 	// Override defines rules that always set parameters, overwriting any existing values.
 	Override []PayloadRule `yaml:"override" json:"override"`
+	// OverrideRaw defines rules that always set raw JSON values, overwriting any existing values.
+	OverrideRaw []PayloadRule `yaml:"override-raw" json:"override-raw"`
 }
 
 // PayloadRule describes a single rule targeting a list of models with parameter updates.
@@ -222,6 +236,7 @@ type PayloadRule struct {
 	// Models lists model entries with name pattern and protocol constraint.
 	Models []PayloadModelRule `yaml:"models" json:"models"`
 	// Params maps JSON paths (gjson/sjson syntax) to values written into the payload.
+	// For *-raw rules, values are treated as raw JSON fragments (strings are used as-is).
 	Params map[string]any `yaml:"params" json:"params"`
 }
 
@@ -239,6 +254,10 @@ type ClaudeKey struct {
 	// APIKey is the authentication key for accessing Claude API services.
 	APIKey string `yaml:"api-key" json:"api-key"`
 
+	// Priority controls selection preference when multiple credentials match.
+	// Higher values are preferred; defaults to 0.
+	Priority int `yaml:"priority,omitempty" json:"priority,omitempty"`
+
 	// Prefix optionally namespaces models for this credential (e.g., "teamA/claude-sonnet-4").
 	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
 
@@ -259,6 +278,9 @@ type ClaudeKey struct {
 	ExcludedModels []string `yaml:"excluded-models,omitempty" json:"excluded-models,omitempty"`
 }
 
+func (k ClaudeKey) GetAPIKey() string  { return k.APIKey }
+func (k ClaudeKey) GetBaseURL() string { return k.BaseURL }
+
 // ClaudeModel describes a mapping between an alias and the actual upstream model name.
 type ClaudeModel struct {
 	// Name is the upstream model identifier used when issuing requests.
@@ -277,6 +299,10 @@ type CodexKey struct {
 	// APIKey is the authentication key for accessing Codex API services.
 	APIKey string `yaml:"api-key" json:"api-key"`
 
+	// Priority controls selection preference when multiple credentials match.
+	// Higher values are preferred; defaults to 0.
+	Priority int `yaml:"priority,omitempty" json:"priority,omitempty"`
+
 	// Prefix optionally namespaces models for this credential (e.g., "teamA/gpt-5-codex").
 	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
 
@@ -297,6 +323,9 @@ type CodexKey struct {
 	ExcludedModels []string `yaml:"excluded-models,omitempty" json:"excluded-models,omitempty"`
 }
 
+func (k CodexKey) GetAPIKey() string  { return k.APIKey }
+func (k CodexKey) GetBaseURL() string { return k.BaseURL }
+
 // CodexModel describes a mapping between an alias and the actual upstream model name.
 type CodexModel struct {
 	// Name is the upstream model identifier used when issuing requests.
@@ -315,6 +344,10 @@ type GeminiKey struct {
 	// APIKey is the authentication key for accessing Gemini API services.
 	APIKey string `yaml:"api-key" json:"api-key"`
 
+	// Priority controls selection preference when multiple credentials match.
+	// Higher values are preferred; defaults to 0.
+	Priority int `yaml:"priority,omitempty" json:"priority,omitempty"`
+
 	// Prefix optionally namespaces models for this credential (e.g., "teamA/gemini-3-pro-preview").
 	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
 
@@ -334,6 +367,9 @@ type GeminiKey struct {
 	ExcludedModels []string `yaml:"excluded-models,omitempty" json:"excluded-models,omitempty"`
 }
 
+func (k GeminiKey) GetAPIKey() string  { return k.APIKey }
+func (k GeminiKey) GetBaseURL() string { return k.BaseURL }
+
 // GeminiModel describes a mapping between an alias and the actual upstream model name.
 type GeminiModel struct {
 	// Name is the upstream model identifier used when issuing requests.
@@ -352,6 +388,10 @@ type OpenAICompatibility struct {
 	// Name is the identifier for this OpenAI compatibility configuration.
 	Name string `yaml:"name" json:"name"`
 
+	// Priority controls selection preference when multiple providers or credentials match.
+	// Higher values are preferred; defaults to 0.
+	Priority int `yaml:"priority,omitempty" json:"priority,omitempty"`
+
 	// Prefix optionally namespaces model aliases for this provider (e.g., "teamA/kimi-k2").
 	Prefix string `yaml:"prefix,omitempty" json:"prefix,omitempty"`
 
@@ -387,6 +427,9 @@ type OpenAICompatibilityModel struct {
 	Alias string `yaml:"alias" json:"alias"`
 }
 
+func (m OpenAICompatibilityModel) GetName() string  { return m.Name }
+func (m OpenAICompatibilityModel) GetAlias() string { return m.Alias }
+
 // LoadConfig reads a YAML configuration file from the given path,
 // unmarshals it into a Config struct, applies environment variable overrides,
 // and returns it.
@@ -405,6 +448,15 @@ func LoadConfig(configFile string) (*Config, error) {
 // If optional is true and the file is missing, it returns an empty Config.
 // If optional is true and the file is empty or invalid, it returns an empty Config.
 func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
+	// Perform oauth-model-alias migration before loading config.
+	// This migrates oauth-model-mappings to oauth-model-alias if needed.
+	if migrated, err := MigrateOAuthModelAlias(configFile); err != nil {
+		// Log warning but don't fail - config loading should still work
+		fmt.Printf("Warning: oauth-model-alias migration failed: %v\n", err)
+	} else if migrated {
+		fmt.Println("Migrated oauth-model-mappings to oauth-model-alias")
+	}
+
 	// Read the entire configuration file into memory.
 	data, err := os.ReadFile(configFile)
 	if err != nil {
@@ -497,8 +549,11 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 	// Normalize OAuth provider model exclusion map.
 	cfg.OAuthExcludedModels = NormalizeOAuthExcludedModels(cfg.OAuthExcludedModels)
 
-	// Normalize global OAuth model name mappings.
-	cfg.SanitizeOAuthModelMappings()
+	// Normalize global OAuth model name aliases.
+	cfg.SanitizeOAuthModelAlias()
+
+	// Validate raw payload rules and drop invalid entries.
+	cfg.SanitizePayloadRules()
 
 	if cfg.legacyMigrationPending {
 		fmt.Println("Detected legacy configuration keys, attempting to persist the normalized config...")
@@ -516,48 +571,97 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 	return &cfg, nil
 }
 
-// SanitizeOAuthModelMappings normalizes and deduplicates global OAuth model name mappings.
-// It trims whitespace, normalizes channel keys to lower-case, drops empty entries,
-// and ensures (From, To) pairs are unique within each channel.
-func (cfg *Config) SanitizeOAuthModelMappings() {
-	if cfg == nil || len(cfg.OAuthModelMappings) == 0 {
+// SanitizePayloadRules validates raw JSON payload rule params and drops invalid rules.
+func (cfg *Config) SanitizePayloadRules() {
+	if cfg == nil {
 		return
 	}
-	out := make(map[string][]ModelNameMapping, len(cfg.OAuthModelMappings))
-	for rawChannel, mappings := range cfg.OAuthModelMappings {
-		channel := strings.ToLower(strings.TrimSpace(rawChannel))
-		if channel == "" || len(mappings) == 0 {
+	cfg.Payload.DefaultRaw = sanitizePayloadRawRules(cfg.Payload.DefaultRaw, "default-raw")
+	cfg.Payload.OverrideRaw = sanitizePayloadRawRules(cfg.Payload.OverrideRaw, "override-raw")
+}
+
+func sanitizePayloadRawRules(rules []PayloadRule, section string) []PayloadRule {
+	if len(rules) == 0 {
+		return rules
+	}
+	out := make([]PayloadRule, 0, len(rules))
+	for i := range rules {
+		rule := rules[i]
+		if len(rule.Params) == 0 {
 			continue
 		}
-		seenName := make(map[string]struct{}, len(mappings))
-		seenAlias := make(map[string]struct{}, len(mappings))
-		clean := make([]ModelNameMapping, 0, len(mappings))
-		for _, mapping := range mappings {
-			name := strings.TrimSpace(mapping.Name)
-			alias := strings.TrimSpace(mapping.Alias)
+		invalid := false
+		for path, value := range rule.Params {
+			raw, ok := payloadRawString(value)
+			if !ok {
+				continue
+			}
+			trimmed := bytes.TrimSpace(raw)
+			if len(trimmed) == 0 || !json.Valid(trimmed) {
+				log.WithFields(log.Fields{
+					"section":    section,
+					"rule_index": i + 1,
+					"param":      path,
+				}).Warn("payload rule dropped: invalid raw JSON")
+				invalid = true
+				break
+			}
+		}
+		if invalid {
+			continue
+		}
+		out = append(out, rule)
+	}
+	return out
+}
+
+func payloadRawString(value any) ([]byte, bool) {
+	switch typed := value.(type) {
+	case string:
+		return []byte(typed), true
+	case []byte:
+		return typed, true
+	default:
+		return nil, false
+	}
+}
+
+// SanitizeOAuthModelAlias normalizes and deduplicates global OAuth model name aliases.
+// It trims whitespace, normalizes channel keys to lower-case, drops empty entries,
+// allows multiple aliases per upstream name, and ensures aliases are unique within each channel.
+func (cfg *Config) SanitizeOAuthModelAlias() {
+	if cfg == nil || len(cfg.OAuthModelAlias) == 0 {
+		return
+	}
+	out := make(map[string][]OAuthModelAlias, len(cfg.OAuthModelAlias))
+	for rawChannel, aliases := range cfg.OAuthModelAlias {
+		channel := strings.ToLower(strings.TrimSpace(rawChannel))
+		if channel == "" || len(aliases) == 0 {
+			continue
+		}
+		seenAlias := make(map[string]struct{}, len(aliases))
+		clean := make([]OAuthModelAlias, 0, len(aliases))
+		for _, entry := range aliases {
+			name := strings.TrimSpace(entry.Name)
+			alias := strings.TrimSpace(entry.Alias)
 			if name == "" || alias == "" {
 				continue
 			}
 			if strings.EqualFold(name, alias) {
 				continue
 			}
-			nameKey := strings.ToLower(name)
 			aliasKey := strings.ToLower(alias)
-			if _, ok := seenName[nameKey]; ok {
-				continue
-			}
 			if _, ok := seenAlias[aliasKey]; ok {
 				continue
 			}
-			seenName[nameKey] = struct{}{}
 			seenAlias[aliasKey] = struct{}{}
-			clean = append(clean, ModelNameMapping{Name: name, Alias: alias})
+			clean = append(clean, OAuthModelAlias{Name: name, Alias: alias, Fork: entry.Fork})
 		}
 		if len(clean) > 0 {
 			out[channel] = clean
 		}
 	}
-	cfg.OAuthModelMappings = out
+	cfg.OAuthModelAlias = out
 }
 
 // SanitizeOpenAICompatibility removes OpenAI-compatibility provider entries that are

internal/config/oauth_model_alias_migration.go

@@ -0,0 +1,275 @@
+package config
+
+import (
+	"os"
+	"strings"
+
+	"gopkg.in/yaml.v3"
+)
+
+// antigravityModelConversionTable maps old built-in aliases to actual model names
+// for the antigravity channel during migration.
+var antigravityModelConversionTable = map[string]string{
+	"gemini-2.5-computer-use-preview-10-2025": "rev19-uic3-1p",
+	"gemini-3-pro-image-preview":              "gemini-3-pro-image",
+	"gemini-3-pro-preview":                    "gemini-3-pro-high",
+	"gemini-3-flash-preview":                  "gemini-3-flash",
+	"gemini-claude-sonnet-4-5":                "claude-sonnet-4-5",
+	"gemini-claude-sonnet-4-5-thinking":       "claude-sonnet-4-5-thinking",
+	"gemini-claude-opus-4-5-thinking":         "claude-opus-4-5-thinking",
+}
+
+// defaultAntigravityAliases returns the default oauth-model-alias configuration
+// for the antigravity channel when neither field exists.
+func defaultAntigravityAliases() []OAuthModelAlias {
+	return []OAuthModelAlias{
+		{Name: "rev19-uic3-1p", Alias: "gemini-2.5-computer-use-preview-10-2025"},
+		{Name: "gemini-3-pro-image", Alias: "gemini-3-pro-image-preview"},
+		{Name: "gemini-3-pro-high", Alias: "gemini-3-pro-preview"},
+		{Name: "gemini-3-flash", Alias: "gemini-3-flash-preview"},
+		{Name: "claude-sonnet-4-5", Alias: "gemini-claude-sonnet-4-5"},
+		{Name: "claude-sonnet-4-5-thinking", Alias: "gemini-claude-sonnet-4-5-thinking"},
+		{Name: "claude-opus-4-5-thinking", Alias: "gemini-claude-opus-4-5-thinking"},
+	}
+}
+
+// MigrateOAuthModelAlias checks for and performs migration from oauth-model-mappings
+// to oauth-model-alias at startup. Returns true if migration was performed.
+//
+// Migration flow:
+// 1. Check if oauth-model-alias exists -> skip migration
+// 2. Check if oauth-model-mappings exists -> convert and migrate
+//   - For antigravity channel, convert old built-in aliases to actual model names
+//
+// 3. Neither exists -> add default antigravity config
+func MigrateOAuthModelAlias(configFile string) (bool, error) {
+	data, err := os.ReadFile(configFile)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return false, nil
+		}
+		return false, err
+	}
+	if len(data) == 0 {
+		return false, nil
+	}
+
+	// Parse YAML into node tree to preserve structure
+	var root yaml.Node
+	if err := yaml.Unmarshal(data, &root); err != nil {
+		return false, nil
+	}
+	if root.Kind != yaml.DocumentNode || len(root.Content) == 0 {
+		return false, nil
+	}
+	rootMap := root.Content[0]
+	if rootMap == nil || rootMap.Kind != yaml.MappingNode {
+		return false, nil
+	}
+
+	// Check if oauth-model-alias already exists
+	if findMapKeyIndex(rootMap, "oauth-model-alias") >= 0 {
+		return false, nil
+	}
+
+	// Check if oauth-model-mappings exists
+	oldIdx := findMapKeyIndex(rootMap, "oauth-model-mappings")
+	if oldIdx >= 0 {
+		// Migrate from old field
+		return migrateFromOldField(configFile, &root, rootMap, oldIdx)
+	}
+
+	// Neither field exists - add default antigravity config
+	return addDefaultAntigravityConfig(configFile, &root, rootMap)
+}
+
+// migrateFromOldField converts oauth-model-mappings to oauth-model-alias
+func migrateFromOldField(configFile string, root *yaml.Node, rootMap *yaml.Node, oldIdx int) (bool, error) {
+	if oldIdx+1 >= len(rootMap.Content) {
+		return false, nil
+	}
+	oldValue := rootMap.Content[oldIdx+1]
+	if oldValue == nil || oldValue.Kind != yaml.MappingNode {
+		return false, nil
+	}
+
+	// Parse the old aliases
+	oldAliases := parseOldAliasNode(oldValue)
+	if len(oldAliases) == 0 {
+		// Remove the old field and write
+		removeMapKeyByIndex(rootMap, oldIdx)
+		return writeYAMLNode(configFile, root)
+	}
+
+	// Convert model names for antigravity channel
+	newAliases := make(map[string][]OAuthModelAlias, len(oldAliases))
+	for channel, entries := range oldAliases {
+		converted := make([]OAuthModelAlias, 0, len(entries))
+		for _, entry := range entries {
+			newEntry := OAuthModelAlias{
+				Name:  entry.Name,
+				Alias: entry.Alias,
+				Fork:  entry.Fork,
+			}
+			// Convert model names for antigravity channel
+			if strings.EqualFold(channel, "antigravity") {
+				if actual, ok := antigravityModelConversionTable[entry.Name]; ok {
+					newEntry.Name = actual
+				}
+			}
+			converted = append(converted, newEntry)
+		}
+		newAliases[channel] = converted
+	}
+
+	// For antigravity channel, supplement missing default aliases
+	if antigravityEntries, exists := newAliases["antigravity"]; exists {
+		// Build a set of already configured model names (upstream names)
+		configuredModels := make(map[string]bool, len(antigravityEntries))
+		for _, entry := range antigravityEntries {
+			configuredModels[entry.Name] = true
+		}
+
+		// Add missing default aliases
+		for _, defaultAlias := range defaultAntigravityAliases() {
+			if !configuredModels[defaultAlias.Name] {
+				antigravityEntries = append(antigravityEntries, defaultAlias)
+			}
+		}
+		newAliases["antigravity"] = antigravityEntries
+	}
+
+	// Build new node
+	newNode := buildOAuthModelAliasNode(newAliases)
+
+	// Replace old key with new key and value
+	rootMap.Content[oldIdx].Value = "oauth-model-alias"
+	rootMap.Content[oldIdx+1] = newNode
+
+	return writeYAMLNode(configFile, root)
+}
+
+// addDefaultAntigravityConfig adds the default antigravity configuration
+func addDefaultAntigravityConfig(configFile string, root *yaml.Node, rootMap *yaml.Node) (bool, error) {
+	defaults := map[string][]OAuthModelAlias{
+		"antigravity": defaultAntigravityAliases(),
+	}
+	newNode := buildOAuthModelAliasNode(defaults)
+
+	// Add new key-value pair
+	keyNode := &yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "oauth-model-alias"}
+	rootMap.Content = append(rootMap.Content, keyNode, newNode)
+
+	return writeYAMLNode(configFile, root)
+}
+
+// parseOldAliasNode parses the old oauth-model-mappings node structure
+func parseOldAliasNode(node *yaml.Node) map[string][]OAuthModelAlias {
+	if node == nil || node.Kind != yaml.MappingNode {
+		return nil
+	}
+	result := make(map[string][]OAuthModelAlias)
+	for i := 0; i+1 < len(node.Content); i += 2 {
+		channelNode := node.Content[i]
+		entriesNode := node.Content[i+1]
+		if channelNode == nil || entriesNode == nil {
+			continue
+		}
+		channel := strings.ToLower(strings.TrimSpace(channelNode.Value))
+		if channel == "" || entriesNode.Kind != yaml.SequenceNode {
+			continue
+		}
+		entries := make([]OAuthModelAlias, 0, len(entriesNode.Content))
+		for _, entryNode := range entriesNode.Content {
+			if entryNode == nil || entryNode.Kind != yaml.MappingNode {
+				continue
+			}
+			entry := parseAliasEntry(entryNode)
+			if entry.Name != "" && entry.Alias != "" {
+				entries = append(entries, entry)
+			}
+		}
+		if len(entries) > 0 {
+			result[channel] = entries
+		}
+	}
+	return result
+}
+
+// parseAliasEntry parses a single alias entry node
+func parseAliasEntry(node *yaml.Node) OAuthModelAlias {
+	var entry OAuthModelAlias
+	for i := 0; i+1 < len(node.Content); i += 2 {
+		keyNode := node.Content[i]
+		valNode := node.Content[i+1]
+		if keyNode == nil || valNode == nil {
+			continue
+		}
+		switch strings.ToLower(strings.TrimSpace(keyNode.Value)) {
+		case "name":
+			entry.Name = strings.TrimSpace(valNode.Value)
+		case "alias":
+			entry.Alias = strings.TrimSpace(valNode.Value)
+		case "fork":
+			entry.Fork = strings.ToLower(strings.TrimSpace(valNode.Value)) == "true"
+		}
+	}
+	return entry
+}
+
+// buildOAuthModelAliasNode creates a YAML node for oauth-model-alias
+func buildOAuthModelAliasNode(aliases map[string][]OAuthModelAlias) *yaml.Node {
+	node := &yaml.Node{Kind: yaml.MappingNode, Tag: "!!map"}
+	for channel, entries := range aliases {
+		channelNode := &yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: channel}
+		entriesNode := &yaml.Node{Kind: yaml.SequenceNode, Tag: "!!seq"}
+		for _, entry := range entries {
+			entryNode := &yaml.Node{Kind: yaml.MappingNode, Tag: "!!map"}
+			entryNode.Content = append(entryNode.Content,
+				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "name"},
+				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: entry.Name},
+				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "alias"},
+				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: entry.Alias},
+			)
+			if entry.Fork {
+				entryNode.Content = append(entryNode.Content,
+					&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "fork"},
+					&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!bool", Value: "true"},
+				)
+			}
+			entriesNode.Content = append(entriesNode.Content, entryNode)
+		}
+		node.Content = append(node.Content, channelNode, entriesNode)
+	}
+	return node
+}
+
+// removeMapKeyByIndex removes a key-value pair from a mapping node by index
+func removeMapKeyByIndex(mapNode *yaml.Node, keyIdx int) {
+	if mapNode == nil || mapNode.Kind != yaml.MappingNode {
+		return
+	}
+	if keyIdx < 0 || keyIdx+1 >= len(mapNode.Content) {
+		return
+	}
+	mapNode.Content = append(mapNode.Content[:keyIdx], mapNode.Content[keyIdx+2:]...)
+}
+
+// writeYAMLNode writes the YAML node tree back to file
+func writeYAMLNode(configFile string, root *yaml.Node) (bool, error) {
+	f, err := os.Create(configFile)
+	if err != nil {
+		return false, err
+	}
+	defer f.Close()
+
+	enc := yaml.NewEncoder(f)
+	enc.SetIndent(2)
+	if err := enc.Encode(root); err != nil {
+		return false, err
+	}
+	if err := enc.Close(); err != nil {
+		return false, err
+	}
+	return true, nil
+}

internal/config/oauth_model_alias_migration_test.go

@@ -0,0 +1,242 @@
+package config
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gopkg.in/yaml.v3"
+)
+
+func TestMigrateOAuthModelAlias_SkipsIfNewFieldExists(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	content := `oauth-model-alias:
+  gemini-cli:
+    - name: "gemini-2.5-pro"
+      alias: "g2.5p"
+`
+	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if migrated {
+		t.Fatal("expected no migration when oauth-model-alias already exists")
+	}
+
+	// Verify file unchanged
+	data, _ := os.ReadFile(configFile)
+	if !strings.Contains(string(data), "oauth-model-alias:") {
+		t.Fatal("file should still contain oauth-model-alias")
+	}
+}
+
+func TestMigrateOAuthModelAlias_MigratesOldField(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	content := `oauth-model-mappings:
+  gemini-cli:
+    - name: "gemini-2.5-pro"
+      alias: "g2.5p"
+      fork: true
+`
+	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !migrated {
+		t.Fatal("expected migration to occur")
+	}
+
+	// Verify new field exists and old field removed
+	data, _ := os.ReadFile(configFile)
+	if strings.Contains(string(data), "oauth-model-mappings:") {
+		t.Fatal("old field should be removed")
+	}
+	if !strings.Contains(string(data), "oauth-model-alias:") {
+		t.Fatal("new field should exist")
+	}
+
+	// Parse and verify structure
+	var root yaml.Node
+	if err := yaml.Unmarshal(data, &root); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestMigrateOAuthModelAlias_ConvertsAntigravityModels(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	// Use old model names that should be converted
+	content := `oauth-model-mappings:
+  antigravity:
+    - name: "gemini-2.5-computer-use-preview-10-2025"
+      alias: "computer-use"
+    - name: "gemini-3-pro-preview"
+      alias: "g3p"
+`
+	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !migrated {
+		t.Fatal("expected migration to occur")
+	}
+
+	// Verify model names were converted
+	data, _ := os.ReadFile(configFile)
+	content = string(data)
+	if !strings.Contains(content, "rev19-uic3-1p") {
+		t.Fatal("expected gemini-2.5-computer-use-preview-10-2025 to be converted to rev19-uic3-1p")
+	}
+	if !strings.Contains(content, "gemini-3-pro-high") {
+		t.Fatal("expected gemini-3-pro-preview to be converted to gemini-3-pro-high")
+	}
+
+	// Verify missing default aliases were supplemented
+	if !strings.Contains(content, "gemini-3-pro-image") {
+		t.Fatal("expected missing default alias gemini-3-pro-image to be added")
+	}
+	if !strings.Contains(content, "gemini-3-flash") {
+		t.Fatal("expected missing default alias gemini-3-flash to be added")
+	}
+	if !strings.Contains(content, "claude-sonnet-4-5") {
+		t.Fatal("expected missing default alias claude-sonnet-4-5 to be added")
+	}
+	if !strings.Contains(content, "claude-sonnet-4-5-thinking") {
+		t.Fatal("expected missing default alias claude-sonnet-4-5-thinking to be added")
+	}
+	if !strings.Contains(content, "claude-opus-4-5-thinking") {
+		t.Fatal("expected missing default alias claude-opus-4-5-thinking to be added")
+	}
+}
+
+func TestMigrateOAuthModelAlias_AddsDefaultIfNeitherExists(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	content := `debug: true
+port: 8080
+`
+	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !migrated {
+		t.Fatal("expected migration to add default config")
+	}
+
+	// Verify default antigravity config was added
+	data, _ := os.ReadFile(configFile)
+	content = string(data)
+	if !strings.Contains(content, "oauth-model-alias:") {
+		t.Fatal("expected oauth-model-alias to be added")
+	}
+	if !strings.Contains(content, "antigravity:") {
+		t.Fatal("expected antigravity channel to be added")
+	}
+	if !strings.Contains(content, "rev19-uic3-1p") {
+		t.Fatal("expected default antigravity aliases to include rev19-uic3-1p")
+	}
+}
+
+func TestMigrateOAuthModelAlias_PreservesOtherConfig(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	content := `debug: true
+port: 8080
+oauth-model-mappings:
+  gemini-cli:
+    - name: "test"
+      alias: "t"
+api-keys:
+  - "key1"
+  - "key2"
+`
+	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !migrated {
+		t.Fatal("expected migration to occur")
+	}
+
+	// Verify other config preserved
+	data, _ := os.ReadFile(configFile)
+	content = string(data)
+	if !strings.Contains(content, "debug: true") {
+		t.Fatal("expected debug field to be preserved")
+	}
+	if !strings.Contains(content, "port: 8080") {
+		t.Fatal("expected port field to be preserved")
+	}
+	if !strings.Contains(content, "api-keys:") {
+		t.Fatal("expected api-keys field to be preserved")
+	}
+}
+
+func TestMigrateOAuthModelAlias_NonexistentFile(t *testing.T) {
+	t.Parallel()
+
+	migrated, err := MigrateOAuthModelAlias("/nonexistent/path/config.yaml")
+	if err != nil {
+		t.Fatalf("unexpected error for nonexistent file: %v", err)
+	}
+	if migrated {
+		t.Fatal("expected no migration for nonexistent file")
+	}
+}
+
+func TestMigrateOAuthModelAlias_EmptyFile(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	configFile := filepath.Join(dir, "config.yaml")
+
+	if err := os.WriteFile(configFile, []byte(""), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	migrated, err := MigrateOAuthModelAlias(configFile)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if migrated {
+		t.Fatal("expected no migration for empty file")
+	}
+}

internal/config/oauth_model_alias_test.go

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

internal/config/sdk_config.go

@@ -25,6 +25,10 @@ type SDKConfig struct {
 
 	// Streaming configures server-side streaming behavior (keep-alives and safe bootstrap retries).
 	Streaming StreamingConfig `yaml:"streaming" json:"streaming"`
+
+	// NonStreamKeepAliveInterval controls how often blank lines are emitted for non-streaming responses.
+	// <= 0 disables keep-alives. Value is in seconds.
+	NonStreamKeepAliveInterval int `yaml:"nonstream-keepalive-interval,omitempty" json:"nonstream-keepalive-interval,omitempty"`
 }
 
 // StreamingConfig holds server streaming behavior configuration.

internal/config/vertex_compat.go

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

internal/logging/global_logger.go

@@ -29,6 +29,9 @@ var (
 // Format: [2025-12-23 20:14:04] [debug] [manager.go:524] | a1b2c3d4 | Use API key sk-9...0RHO for model gpt-5.2
 type LogFormatter struct{}
 
+// logFieldOrder defines the display order for common log fields.
+var logFieldOrder = []string{"provider", "model", "mode", "budget", "level", "original_mode", "original_value", "min", "max", "clamped_to", "error"}
+
 // Format renders a single log entry with custom formatting.
 func (m *LogFormatter) Format(entry *log.Entry) ([]byte, error) {
 	var buffer *bytes.Buffer
@@ -52,11 +55,25 @@ func (m *LogFormatter) Format(entry *log.Entry) ([]byte, error) {
 	}
 	levelStr := fmt.Sprintf("%-5s", level)
 
+	// Build fields string (only print fields in logFieldOrder)
+	var fieldsStr string
+	if len(entry.Data) > 0 {
+		var fields []string
+		for _, k := range logFieldOrder {
+			if v, ok := entry.Data[k]; ok {
+				fields = append(fields, fmt.Sprintf("%s=%v", k, v))
+			}
+		}
+		if len(fields) > 0 {
+			fieldsStr = " " + strings.Join(fields, " ")
+		}
+	}
+
 	var formatted string
 	if entry.Caller != nil {
-		formatted = fmt.Sprintf("[%s] [%s] [%s] [%s:%d] %s\n", timestamp, reqID, levelStr, filepath.Base(entry.Caller.File), entry.Caller.Line, message)
+		formatted = fmt.Sprintf("[%s] [%s] [%s] [%s:%d] %s%s\n", timestamp, reqID, levelStr, filepath.Base(entry.Caller.File), entry.Caller.Line, message, fieldsStr)
 	} else {
-		formatted = fmt.Sprintf("[%s] [%s] [%s] %s\n", timestamp, reqID, levelStr, message)
+		formatted = fmt.Sprintf("[%s] [%s] [%s] %s%s\n", timestamp, reqID, levelStr, message, fieldsStr)
 	}
 	buffer.WriteString(formatted)
 
@@ -104,6 +121,24 @@ func isDirWritable(dir string) bool {
 	return true
 }
 
+// ResolveLogDirectory determines the directory used for application logs.
+func ResolveLogDirectory(cfg *config.Config) string {
+	logDir := "logs"
+	if base := util.WritablePath(); base != "" {
+		return filepath.Join(base, "logs")
+	}
+	if cfg == nil {
+		return logDir
+	}
+	if !isDirWritable(logDir) {
+		authDir := strings.TrimSpace(cfg.AuthDir)
+		if authDir != "" {
+			logDir = filepath.Join(authDir, "logs")
+		}
+	}
+	return logDir
+}
+
 // ConfigureLogOutput switches the global log destination between rotating files and stdout.
 // When logsMaxTotalSizeMB > 0, a background cleaner removes the oldest log files in the logs directory
 // until the total size is within the limit.
@@ -113,12 +148,7 @@ func ConfigureLogOutput(cfg *config.Config) error {
 	writerMu.Lock()
 	defer writerMu.Unlock()
 
-	logDir := "logs"
-	if base := util.WritablePath(); base != "" {
-		logDir = filepath.Join(base, "logs")
-	} else if !isDirWritable(logDir) {
-		logDir = filepath.Join(cfg.AuthDir, "logs")
-	}
+	logDir := ResolveLogDirectory(cfg)
 
 	protectedPath := ""
 	if cfg.LoggingToFile {

internal/misc/codex_instructions.go

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

internal/misc/opencode_codex_instructions.txt

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

internal/registry/model_definitions.go

@@ -27,7 +27,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4.5 Sonnet",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-opus-4-5-20251101",
@@ -39,7 +39,7 @@ func GetClaudeModels() []*ModelInfo {
 			Description:         "Premium model combining maximum intelligence with practical performance",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-opus-4-1-20250805",
@@ -50,7 +50,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4.1 Opus",
 			ContextLength:       200000,
 			MaxCompletionTokens: 32000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-opus-4-20250514",
@@ -61,7 +61,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4 Opus",
 			ContextLength:       200000,
 			MaxCompletionTokens: 32000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-sonnet-4-20250514",
@@ -72,7 +72,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4 Sonnet",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-3-7-sonnet-20250219",
@@ -83,7 +83,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 3.7 Sonnet",
 			ContextLength:       128000,
 			MaxCompletionTokens: 8192,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-3-5-haiku-20241022",
@@ -432,7 +432,7 @@ func GetAIStudioModels() []*ModelInfo {
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
 		},
 		{
 			ID:                         "gemini-3-flash-preview",
@@ -447,7 +447,7 @@ func GetAIStudioModels() []*ModelInfo {
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
 		},
 		{
 			ID:                         "gemini-pro-latest",
@@ -742,6 +742,7 @@ func GetIFlowModels() []*ModelInfo {
 		{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: iFlowThinkingSupport},
 		{ID: "minimax-m2.1", DisplayName: "MiniMax-M2.1", Description: "MiniMax M2.1", Created: 1766448000, Thinking: iFlowThinkingSupport},
+		{ID: "iflow-rome-30ba3b", DisplayName: "iFlow-ROME", Description: "iFlow Rome 30BA3B model", Created: 1736899200},
 	}
 	models := make([]*ModelInfo, 0, len(entries))
 	for _, entry := range entries {
@@ -764,21 +765,23 @@ func GetIFlowModels() []*ModelInfo {
 type AntigravityModelConfig struct {
 	Thinking            *ThinkingSupport
 	MaxCompletionTokens int
-	Name                string
 }
 
 // GetAntigravityModelConfig returns static configuration for antigravity models.
-// Keys use the ALIASED model names (after modelName2Alias conversion) for direct lookup.
+// Keys use upstream model names returned by the Antigravity models endpoint.
 func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 	return map[string]*AntigravityModelConfig{
-		"gemini-2.5-flash":                        {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}, Name: "models/gemini-2.5-flash"},
-		"gemini-2.5-flash-lite":                   {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}, Name: "models/gemini-2.5-flash-lite"},
-		"gemini-2.5-computer-use-preview-10-2025": {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true}, Name: "models/gemini-2.5-computer-use-preview-10-2025"},
-		"gemini-3-pro-preview":                    {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}, Name: "models/gemini-3-pro-preview"},
-		"gemini-3-pro-image-preview":              {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}, Name: "models/gemini-3-pro-image-preview"},
-		"gemini-3-flash-preview":                  {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}, Name: "models/gemini-3-flash-preview"},
-		"gemini-claude-sonnet-4-5-thinking":       {Thinking: &ThinkingSupport{Min: 1024, Max: 200000, ZeroAllowed: false, DynamicAllowed: true}, MaxCompletionTokens: 64000},
-		"gemini-claude-opus-4-5-thinking":         {Thinking: &ThinkingSupport{Min: 1024, Max: 200000, ZeroAllowed: false, DynamicAllowed: true}, MaxCompletionTokens: 64000},
+		"gemini-2.5-flash":           {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}},
+		"gemini-2.5-flash-lite":      {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}},
+		"rev19-uic3-1p":              {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true}},
+		"gemini-3-pro-high":          {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
+		"gemini-3-pro-image":         {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
+		"gemini-3-flash":             {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}},
+		"claude-sonnet-4-5-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
+		"claude-opus-4-5-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
+		"claude-sonnet-4-5":          {MaxCompletionTokens: 64000},
+		"gpt-oss-120b-medium":        {},
+		"tab_flash_lite_preview":     {},
 	}
 }
 
@@ -788,6 +791,7 @@ func LookupStaticModelInfo(modelID string) *ModelInfo {
 	if modelID == "" {
 		return nil
 	}
+
 	allModels := [][]*ModelInfo{
 		GetClaudeModels(),
 		GetGeminiModels(),
@@ -805,5 +809,15 @@ func LookupStaticModelInfo(modelID string) *ModelInfo {
 			}
 		}
 	}
+
+	// Check Antigravity static config
+	if cfg := GetAntigravityModelConfig()[modelID]; cfg != nil {
+		return &ModelInfo{
+			ID:                  modelID,
+			Thinking:            cfg.Thinking,
+			MaxCompletionTokens: cfg.MaxCompletionTokens,
+		}
+	}
+
 	return nil
 }

internal/registry/model_registry.go

@@ -51,6 +51,11 @@ type ModelInfo struct {
 	// Thinking holds provider-specific reasoning/thinking budget capabilities.
 	// This is optional and currently used for Gemini thinking budget normalization.
 	Thinking *ThinkingSupport `json:"thinking,omitempty"`
+
+	// UserDefined indicates this model was defined through config file's models[]
+	// array (e.g., openai-compatibility.*.models[], *-api-key.models[]).
+	// UserDefined models have thinking configuration passed through without validation.
+	UserDefined bool `json:"-"`
 }
 
 // ThinkingSupport describes a model family's supported internal reasoning budget range.
@@ -127,6 +132,21 @@ func GetGlobalRegistry() *ModelRegistry {
 	return globalRegistry
 }
 
+// LookupModelInfo searches the dynamic registry first, then falls back to static model definitions.
+//
+// This helper exists because some code paths only have a model ID and still need Thinking and
+// max completion token metadata even when the dynamic registry hasn't been populated.
+func LookupModelInfo(modelID string) *ModelInfo {
+	modelID = strings.TrimSpace(modelID)
+	if modelID == "" {
+		return nil
+	}
+	if info := GetGlobalRegistry().GetModelInfo(modelID); info != nil {
+		return info
+	}
+	return LookupStaticModelInfo(modelID)
+}
+
 // SetHook sets an optional hook for observing model registration changes.
 func (r *ModelRegistry) SetHook(hook ModelRegistryHook) {
 	if r == nil {

internal/runtime/executor/aistudio_executor.go

@@ -8,12 +8,13 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"io"
 	"net/http"
 	"net/url"
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/wsrelay"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -50,9 +51,68 @@ func (e *AIStudioExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth)
 	return nil
 }
 
+// HttpRequest forwards an arbitrary HTTP request through the websocket relay.
+func (e *AIStudioExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("aistudio executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	if e.relay == nil {
+		return nil, fmt.Errorf("aistudio executor: ws relay is nil")
+	}
+	if auth == nil || auth.ID == "" {
+		return nil, fmt.Errorf("aistudio executor: missing auth")
+	}
+	httpReq := req.WithContext(ctx)
+	if httpReq.URL == nil || strings.TrimSpace(httpReq.URL.String()) == "" {
+		return nil, fmt.Errorf("aistudio executor: request URL is empty")
+	}
+
+	var body []byte
+	if httpReq.Body != nil {
+		b, errRead := io.ReadAll(httpReq.Body)
+		if errRead != nil {
+			return nil, errRead
+		}
+		body = b
+		httpReq.Body = io.NopCloser(bytes.NewReader(b))
+	}
+
+	wsReq := &wsrelay.HTTPRequest{
+		Method:  httpReq.Method,
+		URL:     httpReq.URL.String(),
+		Headers: httpReq.Header.Clone(),
+		Body:    body,
+	}
+	wsResp, errRelay := e.relay.NonStream(ctx, auth.ID, wsReq)
+	if errRelay != nil {
+		return nil, errRelay
+	}
+	if wsResp == nil {
+		return nil, fmt.Errorf("aistudio executor: ws response is nil")
+	}
+
+	statusText := http.StatusText(wsResp.Status)
+	if statusText == "" {
+		statusText = "Unknown"
+	}
+	resp := &http.Response{
+		StatusCode:    wsResp.Status,
+		Status:        fmt.Sprintf("%d %s", wsResp.Status, statusText),
+		Header:        wsResp.Headers.Clone(),
+		Body:          io.NopCloser(bytes.NewReader(wsResp.Body)),
+		ContentLength: int64(len(wsResp.Body)),
+		Request:       httpReq,
+	}
+	return resp, nil
+}
+
 // Execute performs a non-streaming request to the AI Studio API.
 func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	translatedReq, body, err := e.translateRequest(req, opts, false)
@@ -60,7 +120,7 @@ func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth,
 		return resp, err
 	}
 
-	endpoint := e.buildEndpoint(req.Model, body.action, opts.Alt)
+	endpoint := e.buildEndpoint(baseModel, body.action, opts.Alt)
 	wsReq := &wsrelay.HTTPRequest{
 		Method:  http.MethodPost,
 		URL:     endpoint,
@@ -107,7 +167,8 @@ func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth,
 
 // ExecuteStream performs a streaming request to the AI Studio API.
 func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	translatedReq, body, err := e.translateRequest(req, opts, true)
@@ -115,7 +176,7 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 		return nil, err
 	}
 
-	endpoint := e.buildEndpoint(req.Model, body.action, opts.Alt)
+	endpoint := e.buildEndpoint(baseModel, body.action, opts.Alt)
 	wsReq := &wsrelay.HTTPRequest{
 		Method:  http.MethodPost,
 		URL:     endpoint,
@@ -256,6 +317,7 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 
 // CountTokens counts tokens for the given request using the AI Studio API.
 func (e *AIStudioExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 	_, body, err := e.translateRequest(req, opts, false)
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
@@ -265,7 +327,7 @@ func (e *AIStudioExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.A
 	body.payload, _ = sjson.DeleteBytes(body.payload, "tools")
 	body.payload, _ = sjson.DeleteBytes(body.payload, "safetySettings")
 
-	endpoint := e.buildEndpoint(req.Model, "countTokens", "")
+	endpoint := e.buildEndpoint(baseModel, "countTokens", "")
 	wsReq := &wsrelay.HTTPRequest{
 		Method:  http.MethodPost,
 		URL:     endpoint,
@@ -321,22 +383,22 @@ type translatedPayload struct {
 }
 
 func (e *AIStudioExecutor) translateRequest(req cliproxyexecutor.Request, opts cliproxyexecutor.Options, stream bool) ([]byte, translatedPayload, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, stream)
-	payload := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
-	payload = ApplyThinkingMetadata(payload, req.Metadata, req.Model)
-	payload = util.ApplyGemini3ThinkingLevelFromMetadata(req.Model, req.Metadata, payload)
-	payload = util.ApplyDefaultThinkingIfNeeded(req.Model, payload)
-	payload = util.ConvertThinkingLevelToBudget(payload, req.Model, true)
-	payload = util.NormalizeGeminiThinkingBudget(req.Model, payload, true)
-	payload = util.StripThinkingConfigIfUnsupported(req.Model, payload)
-	payload = fixGeminiImageAspectRatio(req.Model, payload)
-	payload = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", payload, originalTranslated)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, stream)
+	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
+	payload, err := thinking.ApplyThinking(payload, req.Model, from.String(), to.String())
+	if err != nil {
+		return nil, translatedPayload{}, err
+	}
+	payload = fixGeminiImageAspectRatio(baseModel, payload)
+	payload = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", payload, originalTranslated)
 	payload, _ = sjson.DeleteBytes(payload, "generationConfig.maxOutputTokens")
 	payload, _ = sjson.DeleteBytes(payload, "generationConfig.responseMimeType")
 	payload, _ = sjson.DeleteBytes(payload, "generationConfig.responseJsonSchema")

internal/runtime/executor/antigravity_executor.go

@@ -10,6 +10,7 @@ import (
 	"crypto/sha256"
 	"encoding/binary"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"math/rand"
@@ -23,7 +24,9 @@ import (
 	"github.com/google/uuid"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	sdkAuth "github.com/router-for-me/CLIProxyAPI/v6/sdk/auth"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
@@ -45,6 +48,7 @@ const (
 	defaultAntigravityAgent        = "antigravity/1.104.0 darwin/arm64"
 	antigravityAuthType            = "antigravity"
 	refreshSkew                    = 3000 * time.Second
+	systemInstruction              = "You are Antigravity, a powerful agentic AI coding assistant designed by the Google Deepmind team working on Advanced Agentic Coding.You are pair programming with a USER to solve their coding task. The task may require creating a new codebase, modifying or debugging an existing codebase, or simply answering a question.**Absolute paths only****Proactiveness**"
 )
 
 var (
@@ -71,13 +75,44 @@ func NewAntigravityExecutor(cfg *config.Config) *AntigravityExecutor {
 // Identifier returns the executor identifier.
 func (e *AntigravityExecutor) Identifier() string { return antigravityAuthType }
 
-// PrepareRequest prepares the HTTP request for execution (no-op for Antigravity).
-func (e *AntigravityExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Antigravity credentials into the outgoing HTTP request.
+func (e *AntigravityExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	token, _, errToken := e.ensureAccessToken(req.Context(), auth)
+	if errToken != nil {
+		return errToken
+	}
+	if strings.TrimSpace(token) == "" {
+		return statusErr{code: http.StatusUnauthorized, msg: "missing access token"}
+	}
+	req.Header.Set("Authorization", "Bearer "+token)
+	return nil
+}
+
+// HttpRequest injects Antigravity credentials into the request and executes it.
+func (e *AntigravityExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("antigravity executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 // Execute performs a non-streaming request to the Antigravity API.
 func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	isClaude := strings.Contains(strings.ToLower(req.Model), "claude")
-	if isClaude {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+	isClaude := strings.Contains(strings.ToLower(baseModel), "claude")
+
+	if isClaude || strings.Contains(baseModel, "gemini-3-pro") {
 		return e.executeClaudeNonStream(ctx, auth, req, opts)
 	}
 
@@ -89,23 +124,25 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 		auth = updatedAuth
 	}
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("antigravity")
+
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, false)
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, 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, req.Metadata, translated)
-	translated = normalizeAntigravityThinking(req.Model, translated, isClaude)
-	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, "antigravity", "request", translated, originalTranslated)
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
+	if err != nil {
+		return resp, err
+	}
+
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated)
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
@@ -115,7 +152,7 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	var lastErr error
 
 	for idx, baseURL := range baseURLs {
-		httpReq, errReq := e.buildRequest(ctx, auth, token, req.Model, translated, false, opts.Alt, baseURL)
+		httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, translated, false, opts.Alt, baseURL)
 		if errReq != nil {
 			err = errReq
 			return resp, err
@@ -124,6 +161,9 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
 			recordAPIResponseError(ctx, e.cfg, errDo)
+			if errors.Is(errDo, context.Canceled) || errors.Is(errDo, context.DeadlineExceeded) {
+				return resp, errDo
+			}
 			lastStatus = 0
 			lastBody = nil
 			lastErr = errDo
@@ -156,7 +196,13 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 				log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 				continue
 			}
-			err = statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			if httpResp.StatusCode == http.StatusTooManyRequests {
+				if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+					sErr.retryAfter = retryAfter
+				}
+			}
+			err = sErr
 			return resp, err
 		}
 
@@ -170,7 +216,13 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 
 	switch {
 	case lastStatus != 0:
-		err = statusErr{code: lastStatus, msg: string(lastBody)}
+		sErr := statusErr{code: lastStatus, msg: string(lastBody)}
+		if lastStatus == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(lastBody); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		err = sErr
 	case lastErr != nil:
 		err = lastErr
 	default:
@@ -181,6 +233,8 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 
 // executeClaudeNonStream performs a claude non-streaming request to the Antigravity API.
 func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
 	if errToken != nil {
 		return resp, errToken
@@ -189,23 +243,25 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 		auth = updatedAuth
 	}
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("antigravity")
+
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, true)
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, 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, req.Metadata, translated)
-	translated = normalizeAntigravityThinking(req.Model, translated, true)
-	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, "antigravity", "request", translated, originalTranslated)
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
+	if err != nil {
+		return resp, err
+	}
+
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated)
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
@@ -215,7 +271,7 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	var lastErr error
 
 	for idx, baseURL := range baseURLs {
-		httpReq, errReq := e.buildRequest(ctx, auth, token, req.Model, translated, true, opts.Alt, baseURL)
+		httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, translated, true, opts.Alt, baseURL)
 		if errReq != nil {
 			err = errReq
 			return resp, err
@@ -224,6 +280,9 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
 			recordAPIResponseError(ctx, e.cfg, errDo)
+			if errors.Is(errDo, context.Canceled) || errors.Is(errDo, context.DeadlineExceeded) {
+				return resp, errDo
+			}
 			lastStatus = 0
 			lastBody = nil
 			lastErr = errDo
@@ -242,6 +301,14 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 			}
 			if errRead != nil {
 				recordAPIResponseError(ctx, e.cfg, errRead)
+				if errors.Is(errRead, context.Canceled) || errors.Is(errRead, context.DeadlineExceeded) {
+					err = errRead
+					return resp, err
+				}
+				if errCtx := ctx.Err(); errCtx != nil {
+					err = errCtx
+					return resp, err
+				}
 				lastStatus = 0
 				lastBody = nil
 				lastErr = errRead
@@ -260,7 +327,13 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 				log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 				continue
 			}
-			err = statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			if httpResp.StatusCode == http.StatusTooManyRequests {
+				if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+					sErr.retryAfter = retryAfter
+				}
+			}
+			err = sErr
 			return resp, err
 		}
 
@@ -325,7 +398,13 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 
 	switch {
 	case lastStatus != 0:
-		err = statusErr{code: lastStatus, msg: string(lastBody)}
+		sErr := statusErr{code: lastStatus, msg: string(lastBody)}
+		if lastStatus == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(lastBody); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		err = sErr
 	case lastErr != nil:
 		err = lastErr
 	default:
@@ -438,8 +517,8 @@ func (e *AntigravityExecutor) convertStreamToNonStream(stream []byte) []byte {
 		}
 		if usageResult := responseNode.Get("usageMetadata"); usageResult.Exists() {
 			usageRaw = usageResult.Raw
-		} else if usageResult := root.Get("usageMetadata"); usageResult.Exists() {
-			usageRaw = usageResult.Raw
+		} else if usageMetadataResult := root.Get("usageMetadata"); usageMetadataResult.Exists() {
+			usageRaw = usageMetadataResult.Raw
 		}
 
 		if partsResult := responseNode.Get("candidates.0.content.parts"); partsResult.IsArray() {
@@ -518,6 +597,8 @@ func (e *AntigravityExecutor) convertStreamToNonStream(stream []byte) []byte {
 
 // ExecuteStream performs a streaming request to the Antigravity API.
 func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	ctx = context.WithValue(ctx, "alt", "")
 
 	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
@@ -528,25 +609,25 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 		auth = updatedAuth
 	}
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
-	isClaude := strings.Contains(strings.ToLower(req.Model), "claude")
-
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("antigravity")
+
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, true)
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, 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, req.Metadata, translated)
-	translated = normalizeAntigravityThinking(req.Model, translated, isClaude)
-	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, "antigravity", "request", translated, originalTranslated)
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
+	if err != nil {
+		return nil, err
+	}
+
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated)
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
@@ -556,15 +637,17 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	var lastErr error
 
 	for idx, baseURL := range baseURLs {
-		httpReq, errReq := e.buildRequest(ctx, auth, token, req.Model, translated, true, opts.Alt, baseURL)
+		httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, translated, true, opts.Alt, baseURL)
 		if errReq != nil {
 			err = errReq
 			return nil, err
 		}
-
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
 			recordAPIResponseError(ctx, e.cfg, errDo)
+			if errors.Is(errDo, context.Canceled) || errors.Is(errDo, context.DeadlineExceeded) {
+				return nil, errDo
+			}
 			lastStatus = 0
 			lastBody = nil
 			lastErr = errDo
@@ -583,6 +666,14 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 			}
 			if errRead != nil {
 				recordAPIResponseError(ctx, e.cfg, errRead)
+				if errors.Is(errRead, context.Canceled) || errors.Is(errRead, context.DeadlineExceeded) {
+					err = errRead
+					return nil, err
+				}
+				if errCtx := ctx.Err(); errCtx != nil {
+					err = errCtx
+					return nil, err
+				}
 				lastStatus = 0
 				lastBody = nil
 				lastErr = errRead
@@ -601,7 +692,13 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 				log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 				continue
 			}
-			err = statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+			if httpResp.StatusCode == http.StatusTooManyRequests {
+				if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+					sErr.retryAfter = retryAfter
+				}
+			}
+			err = sErr
 			return nil, err
 		}
 
@@ -656,7 +753,13 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 
 	switch {
 	case lastStatus != 0:
-		err = statusErr{code: lastStatus, msg: string(lastBody)}
+		sErr := statusErr{code: lastStatus, msg: string(lastBody)}
+		if lastStatus == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(lastBody); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		err = sErr
 	case lastErr != nil:
 		err = lastErr
 	default:
@@ -679,6 +782,8 @@ func (e *AntigravityExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Au
 
 // CountTokens counts tokens for the given request using the Antigravity API.
 func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
 	if errToken != nil {
 		return cliproxyexecutor.Response{}, errToken
@@ -694,7 +799,17 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 	to := sdktranslator.FromString("antigravity")
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
 
-	isClaude := strings.Contains(strings.ToLower(req.Model), "claude")
+	// Prepare payload once (doesn't depend on baseURL)
+	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	payload, err := thinking.ApplyThinking(payload, req.Model, from.String(), to.String())
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
+
+	payload = deleteJSONField(payload, "project")
+	payload = deleteJSONField(payload, "model")
+	payload = deleteJSONField(payload, "request.safetySettings")
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
@@ -711,14 +826,6 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 	var lastErr error
 
 	for idx, baseURL := range baseURLs {
-		payload := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-		payload = ApplyThinkingMetadataCLI(payload, req.Metadata, req.Model)
-		payload = util.ApplyDefaultThinkingIfNeededCLI(req.Model, req.Metadata, payload)
-		payload = normalizeAntigravityThinking(req.Model, payload, isClaude)
-		payload = deleteJSONField(payload, "project")
-		payload = deleteJSONField(payload, "model")
-		payload = deleteJSONField(payload, "request.safetySettings")
-
 		base := strings.TrimSuffix(baseURL, "/")
 		if base == "" {
 			base = buildBaseURL(auth)
@@ -759,6 +866,9 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
 			recordAPIResponseError(ctx, e.cfg, errDo)
+			if errors.Is(errDo, context.Canceled) || errors.Is(errDo, context.DeadlineExceeded) {
+				return cliproxyexecutor.Response{}, errDo
+			}
 			lastStatus = 0
 			lastBody = nil
 			lastErr = errDo
@@ -793,12 +903,24 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 			log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 			continue
 		}
-		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+		sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+		if httpResp.StatusCode == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		return cliproxyexecutor.Response{}, sErr
 	}
 
 	switch {
 	case lastStatus != 0:
-		return cliproxyexecutor.Response{}, statusErr{code: lastStatus, msg: string(lastBody)}
+		sErr := statusErr{code: lastStatus, msg: string(lastBody)}
+		if lastStatus == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(lastBody); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		return cliproxyexecutor.Response{}, sErr
 	case lastErr != nil:
 		return cliproxyexecutor.Response{}, lastErr
 	default:
@@ -835,6 +957,9 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
+			if errors.Is(errDo, context.Canceled) || errors.Is(errDo, context.DeadlineExceeded) {
+				return nil
+			}
 			if idx+1 < len(baseURLs) {
 				log.Debugf("antigravity executor: models request error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 				continue
@@ -870,35 +995,37 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 		modelConfig := registry.GetAntigravityModelConfig()
 		models := make([]*registry.ModelInfo, 0, len(result.Map()))
 		for originalName := range result.Map() {
-			aliasName := modelName2Alias(originalName)
-			if aliasName != "" {
-				cfg := modelConfig[aliasName]
-				modelName := aliasName
-				if cfg != nil && cfg.Name != "" {
-					modelName = cfg.Name
-				}
-				modelInfo := &registry.ModelInfo{
-					ID:          aliasName,
-					Name:        modelName,
-					Description: aliasName,
-					DisplayName: aliasName,
-					Version:     aliasName,
-					Object:      "model",
-					Created:     now,
-					OwnedBy:     antigravityAuthType,
-					Type:        antigravityAuthType,
-				}
-				// Look up Thinking support from static config using alias name
-				if cfg != nil {
-					if cfg.Thinking != nil {
-						modelInfo.Thinking = cfg.Thinking
-					}
-					if cfg.MaxCompletionTokens > 0 {
-						modelInfo.MaxCompletionTokens = cfg.MaxCompletionTokens
-					}
-				}
-				models = append(models, modelInfo)
+			modelID := strings.TrimSpace(originalName)
+			if modelID == "" {
+				continue
 			}
+			switch modelID {
+			case "chat_20706", "chat_23310", "gemini-2.5-flash-thinking", "gemini-3-pro-low", "gemini-2.5-pro":
+				continue
+			}
+			modelCfg := modelConfig[modelID]
+			modelName := modelID
+			modelInfo := &registry.ModelInfo{
+				ID:          modelID,
+				Name:        modelName,
+				Description: modelID,
+				DisplayName: modelID,
+				Version:     modelID,
+				Object:      "model",
+				Created:     now,
+				OwnedBy:     antigravityAuthType,
+				Type:        antigravityAuthType,
+			}
+			// Look up Thinking support from static config using upstream model name.
+			if modelCfg != nil {
+				if modelCfg.Thinking != nil {
+					modelInfo.Thinking = modelCfg.Thinking
+				}
+				if modelCfg.MaxCompletionTokens > 0 {
+					modelInfo.MaxCompletionTokens = modelCfg.MaxCompletionTokens
+				}
+			}
+			models = append(models, modelInfo)
 		}
 		return models
 	}
@@ -914,7 +1041,13 @@ func (e *AntigravityExecutor) ensureAccessToken(ctx context.Context, auth *clipr
 	if accessToken != "" && expiry.After(time.Now().Add(refreshSkew)) {
 		return accessToken, nil, nil
 	}
-	updated, errRefresh := e.refreshToken(ctx, auth.Clone())
+	refreshCtx := context.Background()
+	if ctx != nil {
+		if rt, ok := ctx.Value("cliproxy.roundtripper").(http.RoundTripper); ok && rt != nil {
+			refreshCtx = context.WithValue(refreshCtx, "cliproxy.roundtripper", rt)
+		}
+	}
+	updated, errRefresh := e.refreshToken(refreshCtx, auth.Clone())
 	if errRefresh != nil {
 		return "", nil, errRefresh
 	}
@@ -961,7 +1094,13 @@ func (e *AntigravityExecutor) refreshToken(ctx context.Context, auth *cliproxyau
 	}
 
 	if httpResp.StatusCode < http.StatusOK || httpResp.StatusCode >= http.StatusMultipleChoices {
-		return auth, statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+		sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
+		if httpResp.StatusCode == http.StatusTooManyRequests {
+			if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+				sErr.retryAfter = retryAfter
+			}
+		}
+		return auth, sErr
 	}
 
 	var tokenResp struct {
@@ -982,12 +1121,49 @@ func (e *AntigravityExecutor) refreshToken(ctx context.Context, auth *cliproxyau
 		auth.Metadata["refresh_token"] = tokenResp.RefreshToken
 	}
 	auth.Metadata["expires_in"] = tokenResp.ExpiresIn
-	auth.Metadata["timestamp"] = time.Now().UnixMilli()
-	auth.Metadata["expired"] = time.Now().Add(time.Duration(tokenResp.ExpiresIn) * time.Second).Format(time.RFC3339)
+	now := time.Now()
+	auth.Metadata["timestamp"] = now.UnixMilli()
+	auth.Metadata["expired"] = now.Add(time.Duration(tokenResp.ExpiresIn) * time.Second).Format(time.RFC3339)
 	auth.Metadata["type"] = antigravityAuthType
+	if errProject := e.ensureAntigravityProjectID(ctx, auth, tokenResp.AccessToken); errProject != nil {
+		log.Warnf("antigravity executor: ensure project id failed: %v", errProject)
+	}
 	return auth, nil
 }
 
+func (e *AntigravityExecutor) ensureAntigravityProjectID(ctx context.Context, auth *cliproxyauth.Auth, accessToken string) error {
+	if auth == nil {
+		return nil
+	}
+
+	if auth.Metadata["project_id"] != nil {
+		return nil
+	}
+
+	token := strings.TrimSpace(accessToken)
+	if token == "" {
+		token = metaStringValue(auth.Metadata, "access_token")
+	}
+	if token == "" {
+		return nil
+	}
+
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	projectID, errFetch := sdkAuth.FetchAntigravityProjectID(ctx, token, httpClient)
+	if errFetch != nil {
+		return errFetch
+	}
+	if strings.TrimSpace(projectID) == "" {
+		return nil
+	}
+	if auth.Metadata == nil {
+		auth.Metadata = make(map[string]any)
+	}
+	auth.Metadata["project_id"] = strings.TrimSpace(projectID)
+
+	return nil
+}
+
 func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyauth.Auth, token, modelName string, payload []byte, stream bool, alt, baseURL string) (*http.Request, error) {
 	if token == "" {
 		return nil, statusErr{code: http.StatusUnauthorized, msg: "missing access token"}
@@ -1024,7 +1200,7 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 		}
 	}
 	payload = geminiToAntigravity(modelName, payload, projectID)
-	payload, _ = sjson.SetBytes(payload, "model", alias2ModelName(modelName))
+	payload, _ = sjson.SetBytes(payload, "model", modelName)
 
 	if strings.Contains(modelName, "claude") {
 		strJSON := string(payload)
@@ -1041,6 +1217,19 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 		payload = []byte(strJSON)
 	}
 
+	if strings.Contains(modelName, "claude") || strings.Contains(modelName, "gemini-3-pro-high") {
+		systemInstructionPartsResult := gjson.GetBytes(payload, "request.systemInstruction.parts")
+		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.role", "user")
+		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.parts.0.text", systemInstruction)
+		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.parts.1.text", fmt.Sprintf("Please ignore following [ignore]%s[/ignore]", systemInstruction))
+
+		if systemInstructionPartsResult.Exists() && systemInstructionPartsResult.IsArray() {
+			for _, partResult := range systemInstructionPartsResult.Array() {
+				payload, _ = sjson.SetRawBytes(payload, "request.systemInstruction.parts.-1", []byte(partResult.Raw))
+			}
+		}
+	}
+
 	httpReq, errReq := http.NewRequestWithContext(ctx, http.MethodPost, requestURL.String(), bytes.NewReader(payload))
 	if errReq != nil {
 		return nil, errReq
@@ -1175,8 +1364,8 @@ func antigravityBaseURLFallbackOrder(auth *cliproxyauth.Auth) []string {
 		return []string{base}
 	}
 	return []string{
-		antigravityBaseURLDaily,
 		antigravitySandboxBaseURLDaily,
+		antigravityBaseURLDaily,
 		antigravityBaseURLProd,
 	}
 }
@@ -1204,6 +1393,7 @@ func resolveCustomAntigravityBaseURL(auth *cliproxyauth.Auth) string {
 func geminiToAntigravity(modelName string, payload []byte, projectID string) []byte {
 	template, _ := sjson.Set(string(payload), "model", modelName)
 	template, _ = sjson.Set(template, "userAgent", "antigravity")
+	template, _ = sjson.Set(template, "requestType", "agent")
 
 	// Use real project ID from auth if available, otherwise generate random (legacy fallback)
 	if projectID != "" {
@@ -1217,13 +1407,6 @@ func geminiToAntigravity(modelName string, payload []byte, projectID string) []b
 	template, _ = sjson.Delete(template, "request.safetySettings")
 	template, _ = sjson.Set(template, "request.toolConfig.functionCallingConfig.mode", "VALIDATED")
 
-	if !strings.HasPrefix(modelName, "gemini-3-") {
-		if thinkingLevel := gjson.Get(template, "request.generationConfig.thinkingConfig.thinkingLevel"); thinkingLevel.Exists() {
-			template, _ = sjson.Delete(template, "request.generationConfig.thinkingConfig.thinkingLevel")
-			template, _ = sjson.Set(template, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
-		}
-	}
-
 	if strings.Contains(modelName, "claude") {
 		gjson.Get(template, "request.tools").ForEach(func(key, tool gjson.Result) bool {
 			tool.Get("functionDeclarations").ForEach(func(funKey, funcDecl gjson.Result) bool {
@@ -1281,108 +1464,3 @@ func generateProjectID() string {
 	randomPart := strings.ToLower(uuid.NewString())[:5]
 	return adj + "-" + noun + "-" + randomPart
 }
-
-func modelName2Alias(modelName string) string {
-	switch modelName {
-	case "rev19-uic3-1p":
-		return "gemini-2.5-computer-use-preview-10-2025"
-	case "gemini-3-pro-image":
-		return "gemini-3-pro-image-preview"
-	case "gemini-3-pro-high":
-		return "gemini-3-pro-preview"
-	case "gemini-3-flash":
-		return "gemini-3-flash-preview"
-	case "claude-sonnet-4-5":
-		return "gemini-claude-sonnet-4-5"
-	case "claude-sonnet-4-5-thinking":
-		return "gemini-claude-sonnet-4-5-thinking"
-	case "claude-opus-4-5-thinking":
-		return "gemini-claude-opus-4-5-thinking"
-	case "chat_20706", "chat_23310", "gemini-2.5-flash-thinking", "gemini-3-pro-low", "gemini-2.5-pro":
-		return ""
-	default:
-		return modelName
-	}
-}
-
-func alias2ModelName(modelName string) string {
-	switch modelName {
-	case "gemini-2.5-computer-use-preview-10-2025":
-		return "rev19-uic3-1p"
-	case "gemini-3-pro-image-preview":
-		return "gemini-3-pro-image"
-	case "gemini-3-pro-preview":
-		return "gemini-3-pro-high"
-	case "gemini-3-flash-preview":
-		return "gemini-3-flash"
-	case "gemini-claude-sonnet-4-5":
-		return "claude-sonnet-4-5"
-	case "gemini-claude-sonnet-4-5-thinking":
-		return "claude-sonnet-4-5-thinking"
-	case "gemini-claude-opus-4-5-thinking":
-		return "claude-opus-4-5-thinking"
-	default:
-		return modelName
-	}
-}
-
-// normalizeAntigravityThinking clamps or removes thinking config based on model support.
-// For Claude models, it additionally ensures thinking budget < max_tokens.
-func normalizeAntigravityThinking(model string, payload []byte, isClaude bool) []byte {
-	payload = util.StripThinkingConfigIfUnsupported(model, payload)
-	if !util.ModelSupportsThinking(model) {
-		return payload
-	}
-	budget := gjson.GetBytes(payload, "request.generationConfig.thinkingConfig.thinkingBudget")
-	if !budget.Exists() {
-		return payload
-	}
-	raw := int(budget.Int())
-	normalized := util.NormalizeThinkingBudget(model, raw)
-
-	if isClaude {
-		effectiveMax, setDefaultMax := antigravityEffectiveMaxTokens(model, payload)
-		if effectiveMax > 0 && normalized >= effectiveMax {
-			normalized = effectiveMax - 1
-		}
-		minBudget := antigravityMinThinkingBudget(model)
-		if minBudget > 0 && normalized >= 0 && normalized < minBudget {
-			// Budget is below minimum, remove thinking config entirely
-			payload, _ = sjson.DeleteBytes(payload, "request.generationConfig.thinkingConfig")
-			return payload
-		}
-		if setDefaultMax {
-			if res, errSet := sjson.SetBytes(payload, "request.generationConfig.maxOutputTokens", effectiveMax); errSet == nil {
-				payload = res
-			}
-		}
-	}
-
-	updated, err := sjson.SetBytes(payload, "request.generationConfig.thinkingConfig.thinkingBudget", normalized)
-	if err != nil {
-		return payload
-	}
-	return updated
-}
-
-// antigravityEffectiveMaxTokens returns the max tokens to cap thinking:
-// prefer request-provided maxOutputTokens; otherwise fall back to model default.
-// The boolean indicates whether the value came from the model default (and thus should be written back).
-func antigravityEffectiveMaxTokens(model string, payload []byte) (max int, fromModel bool) {
-	if maxTok := gjson.GetBytes(payload, "request.generationConfig.maxOutputTokens"); maxTok.Exists() && maxTok.Int() > 0 {
-		return int(maxTok.Int()), false
-	}
-	if modelInfo := registry.GetGlobalRegistry().GetModelInfo(model); modelInfo != nil && modelInfo.MaxCompletionTokens > 0 {
-		return modelInfo.MaxCompletionTokens, true
-	}
-	return 0, false
-}
-
-// antigravityMinThinkingBudget returns the minimum thinking budget for a model.
-// Falls back to -1 if no model info is found.
-func antigravityMinThinkingBudget(model string) int {
-	if modelInfo := registry.GetGlobalRegistry().GetModelInfo(model); modelInfo != nil && modelInfo.Thinking != nil {
-		return modelInfo.Thinking.Min
-	}
-	return -1
-}

internal/runtime/executor/cache_helpers.go

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

internal/runtime/executor/claude_executor.go

@@ -17,7 +17,7 @@ import (
 	claudeauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/claude"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -35,24 +35,64 @@ type ClaudeExecutor struct {
 	cfg *config.Config
 }
 
+const claudeToolPrefix = "proxy_"
+
 func NewClaudeExecutor(cfg *config.Config) *ClaudeExecutor { return &ClaudeExecutor{cfg: cfg} }
 
 func (e *ClaudeExecutor) Identifier() string { return "claude" }
 
-func (e *ClaudeExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Claude credentials into the outgoing HTTP request.
+func (e *ClaudeExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	apiKey, _ := claudeCreds(auth)
+	if strings.TrimSpace(apiKey) == "" {
+		return nil
+	}
+	useAPIKey := auth != nil && auth.Attributes != nil && strings.TrimSpace(auth.Attributes["api_key"]) != ""
+	isAnthropicBase := req.URL != nil && strings.EqualFold(req.URL.Scheme, "https") && strings.EqualFold(req.URL.Host, "api.anthropic.com")
+	if isAnthropicBase && useAPIKey {
+		req.Header.Del("Authorization")
+		req.Header.Set("x-api-key", apiKey)
+	} else {
+		req.Header.Del("x-api-key")
+		req.Header.Set("Authorization", "Bearer "+apiKey)
+	}
+	var attrs map[string]string
+	if auth != nil {
+		attrs = auth.Attributes
+	}
+	util.ApplyCustomHeadersFromAttrs(req, attrs)
+	return nil
+}
+
+// HttpRequest injects Claude credentials into the request and executes it.
+func (e *ClaudeExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("claude executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	apiKey, baseURL := claudeCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	apiKey, baseURL := claudeCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://api.anthropic.com"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
-	model := req.Model
-	if override := e.resolveUpstreamModel(req.Model, auth); override != "" {
-		model = override
-	}
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("claude")
 	// Use streaming translation to preserve function calling, except for claude.
@@ -61,29 +101,34 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, model, originalPayload, stream)
-	body := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), stream)
-	body, _ = sjson.SetBytes(body, "model", model)
-	// Inject thinking config based on model metadata for thinking variants
-	body = e.injectThinkingConfig(model, req.Metadata, body)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, stream)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	if !strings.HasPrefix(model, "claude-3-5-haiku") {
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return resp, err
+	}
+
+	if !strings.HasPrefix(baseModel, "claude-3-5-haiku") {
 		body = checkSystemInstructions(body)
 	}
-	body = applyPayloadConfigWithRoot(e.cfg, model, to.String(), "", body, originalTranslated)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 
-	// Ensure max_tokens > thinking.budget_tokens when thinking is enabled
-	body = ensureMaxTokensForThinking(model, body)
-
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
+	bodyForTranslation := body
+	bodyForUpstream := body
+	if isClaudeOAuthToken(apiKey) {
+		bodyForUpstream = applyClaudeToolPrefix(body, claudeToolPrefix)
+	}
 
 	url := fmt.Sprintf("%s/v1/messages?beta=true", baseURL)
-	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(bodyForUpstream))
 	if err != nil {
 		return resp, err
 	}
@@ -98,7 +143,7 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 		URL:       url,
 		Method:    http.MethodPost,
 		Headers:   httpReq.Header.Clone(),
-		Body:      body,
+		Body:      bodyForUpstream,
 		Provider:  e.Identifier(),
 		AuthID:    authID,
 		AuthLabel: authLabel,
@@ -152,50 +197,66 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	} else {
 		reporter.publish(ctx, parseClaudeUsage(data))
 	}
+	if isClaudeOAuthToken(apiKey) {
+		data = stripClaudeToolPrefixFromResponse(data, claudeToolPrefix)
+	}
 	var param any
-	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
+	out := sdktranslator.TranslateNonStream(
+		ctx,
+		to,
+		from,
+		req.Model,
+		bytes.Clone(opts.OriginalRequest),
+		bodyForTranslation,
+		data,
+		&param,
+	)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
 }
 
 func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	apiKey, baseURL := claudeCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	apiKey, baseURL := claudeCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://api.anthropic.com"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("claude")
-	model := req.Model
-	if override := e.resolveUpstreamModel(req.Model, auth); override != "" {
-		model = override
-	}
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, model, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), true)
-	body, _ = sjson.SetBytes(body, "model", model)
-	// Inject thinking config based on model metadata for thinking variants
-	body = e.injectThinkingConfig(model, req.Metadata, body)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return nil, err
+	}
+
 	body = checkSystemInstructions(body)
-	body = applyPayloadConfigWithRoot(e.cfg, model, to.String(), "", body, originalTranslated)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 
-	// Ensure max_tokens > thinking.budget_tokens when thinking is enabled
-	body = ensureMaxTokensForThinking(model, body)
-
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
+	bodyForTranslation := body
+	bodyForUpstream := body
+	if isClaudeOAuthToken(apiKey) {
+		bodyForUpstream = applyClaudeToolPrefix(body, claudeToolPrefix)
+	}
 
 	url := fmt.Sprintf("%s/v1/messages?beta=true", baseURL)
-	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(bodyForUpstream))
 	if err != nil {
 		return nil, err
 	}
@@ -210,7 +271,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 		URL:       url,
 		Method:    http.MethodPost,
 		Headers:   httpReq.Header.Clone(),
-		Body:      body,
+		Body:      bodyForUpstream,
 		Provider:  e.Identifier(),
 		AuthID:    authID,
 		AuthLabel: authLabel,
@@ -263,6 +324,9 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 				if detail, ok := parseClaudeStreamUsage(line); ok {
 					reporter.publish(ctx, detail)
 				}
+				if isClaudeOAuthToken(apiKey) {
+					line = stripClaudeToolPrefixFromStreamLine(line, claudeToolPrefix)
+				}
 				// Forward the line as-is to preserve SSE format
 				cloned := make([]byte, len(line)+1)
 				copy(cloned, line)
@@ -287,7 +351,19 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 			if detail, ok := parseClaudeStreamUsage(line); ok {
 				reporter.publish(ctx, detail)
 			}
-			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(line), &param)
+			if isClaudeOAuthToken(apiKey) {
+				line = stripClaudeToolPrefixFromStreamLine(line, claudeToolPrefix)
+			}
+			chunks := sdktranslator.TranslateStream(
+				ctx,
+				to,
+				from,
+				req.Model,
+				bytes.Clone(opts.OriginalRequest),
+				bodyForTranslation,
+				bytes.Clone(line),
+				&param,
+			)
 			for i := range chunks {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
 			}
@@ -302,8 +378,9 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 }
 
 func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
-	apiKey, baseURL := claudeCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	apiKey, baseURL := claudeCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://api.anthropic.com"
 	}
@@ -312,20 +389,19 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	to := sdktranslator.FromString("claude")
 	// Use streaming translation to preserve function calling, except for claude.
 	stream := from != to
-	model := req.Model
-	if override := e.resolveUpstreamModel(req.Model, auth); override != "" {
-		model = override
-	}
-	body := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), stream)
-	body, _ = sjson.SetBytes(body, "model", model)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	if !strings.HasPrefix(model, "claude-3-5-haiku") {
+	if !strings.HasPrefix(baseModel, "claude-3-5-haiku") {
 		body = checkSystemInstructions(body)
 	}
 
 	// Extract betas from body and convert to header (for count_tokens too)
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
+	if isClaudeOAuthToken(apiKey) {
+		body = applyClaudeToolPrefix(body, claudeToolPrefix)
+	}
 
 	url := fmt.Sprintf("%s/v1/messages/count_tokens?beta=true", baseURL)
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
@@ -445,17 +521,6 @@ func extractAndRemoveBetas(body []byte) ([]string, []byte) {
 	return betas, body
 }
 
-// injectThinkingConfig adds thinking configuration based on metadata using the unified flow.
-// It uses util.ResolveClaudeThinkingConfig which internally calls ResolveThinkingConfigFromMetadata
-// and NormalizeThinkingBudget, ensuring consistency with other executors like Gemini.
-func (e *ClaudeExecutor) injectThinkingConfig(modelName string, metadata map[string]any, body []byte) []byte {
-	budget, ok := util.ResolveClaudeThinkingConfig(modelName, metadata)
-	if !ok {
-		return body
-	}
-	return util.ApplyClaudeThinkingConfig(body, budget)
-}
-
 // disableThinkingIfToolChoiceForced checks if tool_choice forces tool use and disables thinking.
 // Anthropic API does not allow thinking when tool_choice is set to "any" or a specific tool.
 // See: https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking#important-considerations
@@ -469,126 +534,6 @@ func disableThinkingIfToolChoiceForced(body []byte) []byte {
 	return body
 }
 
-// ensureMaxTokensForThinking ensures max_tokens > thinking.budget_tokens when thinking is enabled.
-// Anthropic API requires this constraint; violating it returns a 400 error.
-// This function should be called after all thinking configuration is finalized.
-// It looks up the model's MaxCompletionTokens from the registry to use as the cap.
-func ensureMaxTokensForThinking(modelName string, body []byte) []byte {
-	thinkingType := gjson.GetBytes(body, "thinking.type").String()
-	if thinkingType != "enabled" {
-		return body
-	}
-
-	budgetTokens := gjson.GetBytes(body, "thinking.budget_tokens").Int()
-	if budgetTokens <= 0 {
-		return body
-	}
-
-	maxTokens := gjson.GetBytes(body, "max_tokens").Int()
-
-	// Look up the model's max completion tokens from the registry
-	maxCompletionTokens := 0
-	if modelInfo := registry.GetGlobalRegistry().GetModelInfo(modelName); modelInfo != nil {
-		maxCompletionTokens = modelInfo.MaxCompletionTokens
-	}
-
-	// Fall back to budget + buffer if registry lookup fails or returns 0
-	const fallbackBuffer = 4000
-	requiredMaxTokens := budgetTokens + fallbackBuffer
-	if maxCompletionTokens > 0 {
-		requiredMaxTokens = int64(maxCompletionTokens)
-	}
-
-	if maxTokens < requiredMaxTokens {
-		body, _ = sjson.SetBytes(body, "max_tokens", requiredMaxTokens)
-	}
-	return body
-}
-
-func (e *ClaudeExecutor) resolveUpstreamModel(alias string, auth *cliproxyauth.Auth) string {
-	trimmed := strings.TrimSpace(alias)
-	if trimmed == "" {
-		return ""
-	}
-
-	entry := e.resolveClaudeConfig(auth)
-	if entry == nil {
-		return ""
-	}
-
-	normalizedModel, metadata := util.NormalizeThinkingModel(trimmed)
-
-	// Candidate names to match against configured aliases/names.
-	candidates := []string{strings.TrimSpace(normalizedModel)}
-	if !strings.EqualFold(normalizedModel, trimmed) {
-		candidates = append(candidates, trimmed)
-	}
-	if original := util.ResolveOriginalModel(normalizedModel, metadata); original != "" && !strings.EqualFold(original, normalizedModel) {
-		candidates = append(candidates, original)
-	}
-
-	for i := range entry.Models {
-		model := entry.Models[i]
-		name := strings.TrimSpace(model.Name)
-		modelAlias := strings.TrimSpace(model.Alias)
-
-		for _, candidate := range candidates {
-			if candidate == "" {
-				continue
-			}
-			if modelAlias != "" && strings.EqualFold(modelAlias, candidate) {
-				if name != "" {
-					return name
-				}
-				return candidate
-			}
-			if name != "" && strings.EqualFold(name, candidate) {
-				return name
-			}
-		}
-	}
-	return ""
-}
-
-func (e *ClaudeExecutor) resolveClaudeConfig(auth *cliproxyauth.Auth) *config.ClaudeKey {
-	if auth == nil || e.cfg == nil {
-		return nil
-	}
-	var attrKey, attrBase string
-	if auth.Attributes != nil {
-		attrKey = strings.TrimSpace(auth.Attributes["api_key"])
-		attrBase = strings.TrimSpace(auth.Attributes["base_url"])
-	}
-	for i := range e.cfg.ClaudeKey {
-		entry := &e.cfg.ClaudeKey[i]
-		cfgKey := strings.TrimSpace(entry.APIKey)
-		cfgBase := strings.TrimSpace(entry.BaseURL)
-		if attrKey != "" && attrBase != "" {
-			if strings.EqualFold(cfgKey, attrKey) && strings.EqualFold(cfgBase, attrBase) {
-				return entry
-			}
-			continue
-		}
-		if attrKey != "" && strings.EqualFold(cfgKey, attrKey) {
-			if cfgBase == "" || strings.EqualFold(cfgBase, attrBase) {
-				return entry
-			}
-		}
-		if attrKey == "" && attrBase != "" && strings.EqualFold(cfgBase, attrBase) {
-			return entry
-		}
-	}
-	if attrKey != "" {
-		for i := range e.cfg.ClaudeKey {
-			entry := &e.cfg.ClaudeKey[i]
-			if strings.EqualFold(strings.TrimSpace(entry.APIKey), attrKey) {
-				return entry
-			}
-		}
-	}
-	return nil
-}
-
 type compositeReadCloser struct {
 	io.Reader
 	closers []func() error
@@ -770,3 +715,107 @@ func checkSystemInstructions(payload []byte) []byte {
 	}
 	return payload
 }
+
+func isClaudeOAuthToken(apiKey string) bool {
+	return strings.Contains(apiKey, "sk-ant-oat")
+}
+
+func applyClaudeToolPrefix(body []byte, prefix string) []byte {
+	if prefix == "" {
+		return body
+	}
+
+	if tools := gjson.GetBytes(body, "tools"); tools.Exists() && tools.IsArray() {
+		tools.ForEach(func(index, tool gjson.Result) bool {
+			name := tool.Get("name").String()
+			if name == "" || strings.HasPrefix(name, prefix) {
+				return true
+			}
+			path := fmt.Sprintf("tools.%d.name", index.Int())
+			body, _ = sjson.SetBytes(body, path, prefix+name)
+			return true
+		})
+	}
+
+	if gjson.GetBytes(body, "tool_choice.type").String() == "tool" {
+		name := gjson.GetBytes(body, "tool_choice.name").String()
+		if name != "" && !strings.HasPrefix(name, prefix) {
+			body, _ = sjson.SetBytes(body, "tool_choice.name", prefix+name)
+		}
+	}
+
+	if messages := gjson.GetBytes(body, "messages"); messages.Exists() && messages.IsArray() {
+		messages.ForEach(func(msgIndex, msg gjson.Result) bool {
+			content := msg.Get("content")
+			if !content.Exists() || !content.IsArray() {
+				return true
+			}
+			content.ForEach(func(contentIndex, part gjson.Result) bool {
+				if part.Get("type").String() != "tool_use" {
+					return true
+				}
+				name := part.Get("name").String()
+				if name == "" || strings.HasPrefix(name, prefix) {
+					return true
+				}
+				path := fmt.Sprintf("messages.%d.content.%d.name", msgIndex.Int(), contentIndex.Int())
+				body, _ = sjson.SetBytes(body, path, prefix+name)
+				return true
+			})
+			return true
+		})
+	}
+
+	return body
+}
+
+func stripClaudeToolPrefixFromResponse(body []byte, prefix string) []byte {
+	if prefix == "" {
+		return body
+	}
+	content := gjson.GetBytes(body, "content")
+	if !content.Exists() || !content.IsArray() {
+		return body
+	}
+	content.ForEach(func(index, part gjson.Result) bool {
+		if part.Get("type").String() != "tool_use" {
+			return true
+		}
+		name := part.Get("name").String()
+		if !strings.HasPrefix(name, prefix) {
+			return true
+		}
+		path := fmt.Sprintf("content.%d.name", index.Int())
+		body, _ = sjson.SetBytes(body, path, strings.TrimPrefix(name, prefix))
+		return true
+	})
+	return body
+}
+
+func stripClaudeToolPrefixFromStreamLine(line []byte, prefix string) []byte {
+	if prefix == "" {
+		return line
+	}
+	payload := jsonPayload(line)
+	if len(payload) == 0 || !gjson.ValidBytes(payload) {
+		return line
+	}
+	contentBlock := gjson.GetBytes(payload, "content_block")
+	if !contentBlock.Exists() || contentBlock.Get("type").String() != "tool_use" {
+		return line
+	}
+	name := contentBlock.Get("name").String()
+	if !strings.HasPrefix(name, prefix) {
+		return line
+	}
+	updated, err := sjson.SetBytes(payload, "content_block.name", strings.TrimPrefix(name, prefix))
+	if err != nil {
+		return line
+	}
+
+	trimmed := bytes.TrimSpace(line)
+	if bytes.HasPrefix(trimmed, []byte("data:")) {
+		return append([]byte("data: "), updated...)
+	}
+	return updated
+}

internal/runtime/executor/claude_executor_test.go

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

internal/runtime/executor/codex_executor.go

@@ -13,6 +13,7 @@ import (
 	codexauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/codex"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -38,39 +39,77 @@ func NewCodexExecutor(cfg *config.Config) *CodexExecutor { return &CodexExecutor
 
 func (e *CodexExecutor) Identifier() string { return "codex" }
 
-func (e *CodexExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Codex credentials into the outgoing HTTP request.
+func (e *CodexExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	apiKey, _ := codexCreds(auth)
+	if strings.TrimSpace(apiKey) != "" {
+		req.Header.Set("Authorization", "Bearer "+apiKey)
+	}
+	var attrs map[string]string
+	if auth != nil {
+		attrs = auth.Attributes
+	}
+	util.ApplyCustomHeadersFromAttrs(req, attrs)
+	return nil
+}
+
+// HttpRequest injects Codex credentials into the request and executes it.
+func (e *CodexExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("codex executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	apiKey, baseURL := codexCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	apiKey, baseURL := codexCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://chatgpt.com/backend-api/codex"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
 
-	model := req.Model
-	if override := e.resolveUpstreamModel(req.Model, auth); override != "" {
-		model = override
-	}
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
+	userAgent := codexUserAgent(ctx)
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, model, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), false)
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, model, "reasoning.effort", false)
-	body = NormalizeThinkingConfig(body, model, false)
-	if errValidate := ValidateThinkingConfig(body, model); errValidate != nil {
-		return resp, errValidate
+	originalPayload = misc.InjectCodexUserAgent(originalPayload, userAgent)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := misc.InjectCodexUserAgent(bytes.Clone(req.Payload), userAgent)
+	body = sdktranslator.TranslateRequest(from, to, baseModel, body, false)
+	body = misc.StripCodexUserAgent(body)
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return resp, err
 	}
-	body = applyPayloadConfigWithRoot(e.cfg, model, to.String(), "", body, originalTranslated)
-	body, _ = sjson.SetBytes(body, "model", model)
+
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 	body, _ = sjson.SetBytes(body, "stream", true)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
+	body, _ = sjson.DeleteBytes(body, "prompt_cache_retention")
+	body, _ = sjson.DeleteBytes(body, "safety_identifier")
+	if !gjson.GetBytes(body, "instructions").Exists() {
+		body, _ = sjson.SetBytes(body, "instructions", "")
+	}
 
 	url := strings.TrimSuffix(baseURL, "/") + "/responses"
 	httpReq, err := e.cacheHelper(ctx, from, url, req, body)
@@ -137,7 +176,7 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 		}
 
 		var param any
-		out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, line, &param)
+		out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(originalPayload), body, line, &param)
 		resp = cliproxyexecutor.Response{Payload: []byte(out)}
 		return resp, nil
 	}
@@ -146,36 +185,42 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 }
 
 func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	apiKey, baseURL := codexCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	apiKey, baseURL := codexCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://chatgpt.com/backend-api/codex"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
 
-	model := req.Model
-	if override := e.resolveUpstreamModel(req.Model, auth); override != "" {
-		model = override
-	}
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
+	userAgent := codexUserAgent(ctx)
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, model, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), true)
+	originalPayload = misc.InjectCodexUserAgent(originalPayload, userAgent)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := misc.InjectCodexUserAgent(bytes.Clone(req.Payload), userAgent)
+	body = sdktranslator.TranslateRequest(from, to, baseModel, body, true)
+	body = misc.StripCodexUserAgent(body)
 
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, model, "reasoning.effort", false)
-	body = NormalizeThinkingConfig(body, model, false)
-	if errValidate := ValidateThinkingConfig(body, model); errValidate != nil {
-		return nil, errValidate
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return nil, err
 	}
-	body = applyPayloadConfigWithRoot(e.cfg, model, to.String(), "", body, originalTranslated)
+
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
-	body, _ = sjson.SetBytes(body, "model", model)
+	body, _ = sjson.DeleteBytes(body, "prompt_cache_retention")
+	body, _ = sjson.DeleteBytes(body, "safety_identifier")
+	body, _ = sjson.SetBytes(body, "model", baseModel)
+	if !gjson.GetBytes(body, "instructions").Exists() {
+		body, _ = sjson.SetBytes(body, "instructions", "")
+	}
 
 	url := strings.TrimSuffix(baseURL, "/") + "/responses"
 	httpReq, err := e.cacheHelper(ctx, from, url, req, body)
@@ -247,7 +292,7 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 				}
 			}
 
-			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(line), &param)
+			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(originalPayload), body, bytes.Clone(line), &param)
 			for i := range chunks {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
 			}
@@ -262,21 +307,30 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 }
 
 func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
-	model := req.Model
-	if override := e.resolveUpstreamModel(req.Model, auth); override != "" {
-		model = override
-	}
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
-	body := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), false)
+	userAgent := codexUserAgent(ctx)
+	body := misc.InjectCodexUserAgent(bytes.Clone(req.Payload), userAgent)
+	body = sdktranslator.TranslateRequest(from, to, baseModel, body, false)
+	body = misc.StripCodexUserAgent(body)
 
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, model, "reasoning.effort", false)
-	body, _ = sjson.SetBytes(body, "model", model)
+	body, err := thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
+
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
+	body, _ = sjson.DeleteBytes(body, "prompt_cache_retention")
+	body, _ = sjson.DeleteBytes(body, "safety_identifier")
 	body, _ = sjson.SetBytes(body, "stream", false)
+	if !gjson.GetBytes(body, "instructions").Exists() {
+		body, _ = sjson.SetBytes(body, "instructions", "")
+	}
 
-	enc, err := tokenizerForCodexModel(model)
+	enc, err := tokenizerForCodexModel(baseModel)
 	if err != nil {
 		return cliproxyexecutor.Response{}, fmt.Errorf("codex executor: tokenizer init failed: %w", err)
 	}
@@ -457,14 +511,14 @@ func (e *CodexExecutor) cacheHelper(ctx context.Context, from sdktranslator.Form
 	if from == "claude" {
 		userIDResult := gjson.GetBytes(req.Payload, "metadata.user_id")
 		if userIDResult.Exists() {
-			var hasKey bool
 			key := fmt.Sprintf("%s-%s", req.Model, userIDResult.String())
-			if cache, hasKey = codexCacheMap[key]; !hasKey || cache.Expire.Before(time.Now()) {
+			var ok bool
+			if cache, ok = getCodexCache(key); !ok {
 				cache = codexCache{
 					ID:     uuid.New().String(),
 					Expire: time.Now().Add(1 * time.Hour),
 				}
-				codexCacheMap[key] = cache
+				setCodexCache(key, cache)
 			}
 		}
 	} else if from == "openai-response" {
@@ -522,6 +576,16 @@ func applyCodexHeaders(r *http.Request, auth *cliproxyauth.Auth, token string) {
 	util.ApplyCustomHeadersFromAttrs(r, attrs)
 }
 
+func codexUserAgent(ctx context.Context) string {
+	if ctx == nil {
+		return ""
+	}
+	if ginCtx, ok := ctx.Value("gin").(*gin.Context); ok && ginCtx != nil && ginCtx.Request != nil {
+		return strings.TrimSpace(ginCtx.Request.UserAgent())
+	}
+	return ""
+}
+
 func codexCreds(a *cliproxyauth.Auth) (apiKey, baseURL string) {
 	if a == nil {
 		return "", ""
@@ -538,51 +602,6 @@ func codexCreds(a *cliproxyauth.Auth) (apiKey, baseURL string) {
 	return
 }
 
-func (e *CodexExecutor) resolveUpstreamModel(alias string, auth *cliproxyauth.Auth) string {
-	trimmed := strings.TrimSpace(alias)
-	if trimmed == "" {
-		return ""
-	}
-
-	entry := e.resolveCodexConfig(auth)
-	if entry == nil {
-		return ""
-	}
-
-	normalizedModel, metadata := util.NormalizeThinkingModel(trimmed)
-
-	// Candidate names to match against configured aliases/names.
-	candidates := []string{strings.TrimSpace(normalizedModel)}
-	if !strings.EqualFold(normalizedModel, trimmed) {
-		candidates = append(candidates, trimmed)
-	}
-	if original := util.ResolveOriginalModel(normalizedModel, metadata); original != "" && !strings.EqualFold(original, normalizedModel) {
-		candidates = append(candidates, original)
-	}
-
-	for i := range entry.Models {
-		model := entry.Models[i]
-		name := strings.TrimSpace(model.Name)
-		modelAlias := strings.TrimSpace(model.Alias)
-
-		for _, candidate := range candidates {
-			if candidate == "" {
-				continue
-			}
-			if modelAlias != "" && strings.EqualFold(modelAlias, candidate) {
-				if name != "" {
-					return name
-				}
-				return candidate
-			}
-			if name != "" && strings.EqualFold(name, candidate) {
-				return name
-			}
-		}
-	}
-	return ""
-}
-
 func (e *CodexExecutor) resolveCodexConfig(auth *cliproxyauth.Auth) *config.CodexKey {
 	if auth == nil || e.cfg == nil {
 		return nil

internal/runtime/executor/gemini_cli_executor.go

@@ -20,6 +20,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/runtime/geminicli"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -63,33 +64,72 @@ func NewGeminiCLIExecutor(cfg *config.Config) *GeminiCLIExecutor {
 // Identifier returns the executor identifier.
 func (e *GeminiCLIExecutor) Identifier() string { return "gemini-cli" }
 
-// PrepareRequest prepares the HTTP request for execution (no-op for Gemini CLI).
-func (e *GeminiCLIExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Gemini CLI credentials into the outgoing HTTP request.
+func (e *GeminiCLIExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	tokenSource, _, errSource := prepareGeminiCLITokenSource(req.Context(), e.cfg, auth)
+	if errSource != nil {
+		return errSource
+	}
+	tok, errTok := tokenSource.Token()
+	if errTok != nil {
+		return errTok
+	}
+	if strings.TrimSpace(tok.AccessToken) == "" {
+		return statusErr{code: http.StatusUnauthorized, msg: "missing access token"}
+	}
+	req.Header.Set("Authorization", "Bearer "+tok.AccessToken)
+	applyGeminiCLIHeaders(req)
+	return nil
+}
+
+// HttpRequest injects Gemini CLI credentials into the request and executes it.
+func (e *GeminiCLIExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("gemini-cli executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 // Execute performs a non-streaming request to the Gemini CLI API.
 func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	tokenSource, baseTokenData, err := prepareGeminiCLITokenSource(ctx, e.cfg, auth)
 	if err != nil {
 		return resp, err
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini-cli")
+
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, false)
-	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, req.Metadata, basePayload)
-	basePayload = util.NormalizeGeminiCLIThinkingBudget(req.Model, basePayload)
-	basePayload = util.StripThinkingConfigIfUnsupported(req.Model, basePayload)
-	basePayload = fixGeminiCLIImageAspectRatio(req.Model, basePayload)
-	basePayload = applyPayloadConfigWithRoot(e.cfg, req.Model, "gemini", "request", basePayload, originalTranslated)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, from.String(), to.String())
+	if err != nil {
+		return resp, err
+	}
+
+	basePayload = fixGeminiCLIImageAspectRatio(baseModel, basePayload)
+	basePayload = applyPayloadConfigWithRoot(e.cfg, baseModel, "gemini", "request", basePayload, originalTranslated)
 
 	action := "generateContent"
 	if req.Metadata != nil {
@@ -99,9 +139,9 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 	}
 
 	projectID := resolveGeminiProjectID(auth)
-	models := cliPreviewFallbackOrder(req.Model)
-	if len(models) == 0 || models[0] != req.Model {
-		models = append([]string{req.Model}, models...)
+	models := cliPreviewFallbackOrder(baseModel)
+	if len(models) == 0 || models[0] != baseModel {
+		models = append([]string{baseModel}, models...)
 	}
 
 	httpClient := newHTTPClient(ctx, e.cfg, auth, 0)
@@ -212,34 +252,39 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 
 // ExecuteStream performs a streaming request to the Gemini CLI API.
 func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	tokenSource, baseTokenData, err := prepareGeminiCLITokenSource(ctx, e.cfg, auth)
 	if err != nil {
 		return nil, err
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini-cli")
+
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, true)
-	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, req.Metadata, basePayload)
-	basePayload = util.NormalizeGeminiCLIThinkingBudget(req.Model, basePayload)
-	basePayload = util.StripThinkingConfigIfUnsupported(req.Model, basePayload)
-	basePayload = fixGeminiCLIImageAspectRatio(req.Model, basePayload)
-	basePayload = applyPayloadConfigWithRoot(e.cfg, req.Model, "gemini", "request", basePayload, originalTranslated)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+
+	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, from.String(), to.String())
+	if err != nil {
+		return nil, err
+	}
+
+	basePayload = fixGeminiCLIImageAspectRatio(baseModel, basePayload)
+	basePayload = applyPayloadConfigWithRoot(e.cfg, baseModel, "gemini", "request", basePayload, originalTranslated)
 
 	projectID := resolveGeminiProjectID(auth)
 
-	models := cliPreviewFallbackOrder(req.Model)
-	if len(models) == 0 || models[0] != req.Model {
-		models = append([]string{req.Model}, models...)
+	models := cliPreviewFallbackOrder(baseModel)
+	if len(models) == 0 || models[0] != baseModel {
+		models = append([]string{baseModel}, models...)
 	}
 
 	httpClient := newHTTPClient(ctx, e.cfg, auth, 0)
@@ -401,6 +446,8 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 
 // CountTokens counts tokens for the given request using the Gemini CLI API.
 func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	tokenSource, baseTokenData, err := prepareGeminiCLITokenSource(ctx, e.cfg, auth)
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
@@ -409,9 +456,9 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini-cli")
 
-	models := cliPreviewFallbackOrder(req.Model)
-	if len(models) == 0 || models[0] != req.Model {
-		models = append([]string{req.Model}, models...)
+	models := cliPreviewFallbackOrder(baseModel)
+	if len(models) == 0 || models[0] != baseModel {
+		models = append([]string{baseModel}, models...)
 	}
 
 	httpClient := newHTTPClient(ctx, e.cfg, auth, 0)
@@ -429,15 +476,18 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 
 	// The loop variable attemptModel is only used as the concrete model id sent to the upstream
 	// Gemini CLI endpoint when iterating fallback variants.
-	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)
+	for range models {
+		payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+		payload, err = thinking.ApplyThinking(payload, req.Model, from.String(), to.String())
+		if err != nil {
+			return cliproxyexecutor.Response{}, err
+		}
+
 		payload = deleteJSONField(payload, "project")
 		payload = deleteJSONField(payload, "model")
 		payload = deleteJSONField(payload, "request.safetySettings")
-		payload = util.StripThinkingConfigIfUnsupported(req.Model, payload)
-		payload = fixGeminiCLIImageAspectRatio(req.Model, payload)
+		payload = fixGeminiCLIImageAspectRatio(baseModel, payload)
 
 		tok, errTok := tokenSource.Token()
 		if errTok != nil {

internal/runtime/executor/gemini_executor.go

@@ -13,6 +13,7 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -55,8 +56,38 @@ func NewGeminiExecutor(cfg *config.Config) *GeminiExecutor {
 // Identifier returns the executor identifier.
 func (e *GeminiExecutor) Identifier() string { return "gemini" }
 
-// PrepareRequest prepares the HTTP request for execution (no-op for Gemini).
-func (e *GeminiExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Gemini credentials into the outgoing HTTP request.
+func (e *GeminiExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	apiKey, bearer := geminiCreds(auth)
+	if apiKey != "" {
+		req.Header.Set("x-goog-api-key", apiKey)
+		req.Header.Del("Authorization")
+	} else if bearer != "" {
+		req.Header.Set("Authorization", "Bearer "+bearer)
+		req.Header.Del("x-goog-api-key")
+	}
+	applyGeminiHeaders(req, auth)
+	return nil
+}
+
+// HttpRequest injects Gemini credentials into the request and executes it.
+func (e *GeminiExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("gemini executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 // Execute performs a non-streaming request to the Gemini API.
 // It translates the request to Gemini format, sends it to the API, and translates
@@ -72,16 +103,13 @@ func (e *GeminiExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) e
 //   - cliproxyexecutor.Response: The response from the API
 //   - error: An error if the request fails
 func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	apiKey, bearer := geminiCreds(auth)
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
-	model := req.Model
-	if override := e.resolveUpstreamModel(model, auth); override != "" {
-		model = override
-	}
-
 	// Official Gemini API via API key or OAuth bearer
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
@@ -89,15 +117,17 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, model, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), false)
-	body = ApplyThinkingMetadata(body, req.Metadata, model)
-	body = util.ApplyDefaultThinkingIfNeeded(model, body)
-	body = util.NormalizeGeminiThinkingBudget(model, body)
-	body = util.StripThinkingConfigIfUnsupported(model, body)
-	body = fixGeminiImageAspectRatio(model, body)
-	body = applyPayloadConfigWithRoot(e.cfg, model, to.String(), "", body, originalTranslated)
-	body, _ = sjson.SetBytes(body, "model", model)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return resp, err
+	}
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	action := "generateContent"
 	if req.Metadata != nil {
@@ -106,7 +136,7 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 		}
 	}
 	baseURL := resolveGeminiBaseURL(auth)
-	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, model, action)
+	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, baseModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -176,34 +206,33 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 
 // ExecuteStream performs a streaming request to the Gemini API.
 func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	apiKey, bearer := geminiCreds(auth)
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
-	model := req.Model
-	if override := e.resolveUpstreamModel(model, auth); override != "" {
-		model = override
-	}
-
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, model, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), true)
-	body = ApplyThinkingMetadata(body, req.Metadata, model)
-	body = util.ApplyDefaultThinkingIfNeeded(model, body)
-	body = util.NormalizeGeminiThinkingBudget(model, body)
-	body = util.StripThinkingConfigIfUnsupported(model, body)
-	body = fixGeminiImageAspectRatio(model, body)
-	body = applyPayloadConfigWithRoot(e.cfg, model, to.String(), "", body, originalTranslated)
-	body, _ = sjson.SetBytes(body, "model", model)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return nil, err
+	}
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	baseURL := resolveGeminiBaseURL(auth)
-	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, model, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, baseModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
@@ -301,27 +330,28 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 
 // CountTokens counts tokens for the given request using the Gemini API.
 func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
-	apiKey, bearer := geminiCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
-	model := req.Model
-	if override := e.resolveUpstreamModel(model, auth); override != "" {
-		model = override
-	}
+	apiKey, bearer := geminiCreds(auth)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	translatedReq := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), false)
-	translatedReq = ApplyThinkingMetadata(translatedReq, req.Metadata, model)
-	translatedReq = util.StripThinkingConfigIfUnsupported(model, translatedReq)
-	translatedReq = fixGeminiImageAspectRatio(model, translatedReq)
+	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String())
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
+
+	translatedReq = fixGeminiImageAspectRatio(baseModel, translatedReq)
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "safetySettings")
-	translatedReq, _ = sjson.SetBytes(translatedReq, "model", model)
+	translatedReq, _ = sjson.SetBytes(translatedReq, "model", baseModel)
 
 	baseURL := resolveGeminiBaseURL(auth)
-	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, model, "countTokens")
+	url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, baseModel, "countTokens")
 
 	requestBody := bytes.NewReader(translatedReq)
 
@@ -420,51 +450,6 @@ func resolveGeminiBaseURL(auth *cliproxyauth.Auth) string {
 	return base
 }
 
-func (e *GeminiExecutor) resolveUpstreamModel(alias string, auth *cliproxyauth.Auth) string {
-	trimmed := strings.TrimSpace(alias)
-	if trimmed == "" {
-		return ""
-	}
-
-	entry := e.resolveGeminiConfig(auth)
-	if entry == nil {
-		return ""
-	}
-
-	normalizedModel, metadata := util.NormalizeThinkingModel(trimmed)
-
-	// Candidate names to match against configured aliases/names.
-	candidates := []string{strings.TrimSpace(normalizedModel)}
-	if !strings.EqualFold(normalizedModel, trimmed) {
-		candidates = append(candidates, trimmed)
-	}
-	if original := util.ResolveOriginalModel(normalizedModel, metadata); original != "" && !strings.EqualFold(original, normalizedModel) {
-		candidates = append(candidates, original)
-	}
-
-	for i := range entry.Models {
-		model := entry.Models[i]
-		name := strings.TrimSpace(model.Name)
-		modelAlias := strings.TrimSpace(model.Alias)
-
-		for _, candidate := range candidates {
-			if candidate == "" {
-				continue
-			}
-			if modelAlias != "" && strings.EqualFold(modelAlias, candidate) {
-				if name != "" {
-					return name
-				}
-				return candidate
-			}
-			if name != "" && strings.EqualFold(name, candidate) {
-				return name
-			}
-		}
-	}
-	return ""
-}
-
 func (e *GeminiExecutor) resolveGeminiConfig(auth *cliproxyauth.Auth) *config.GeminiKey {
 	if auth == nil || e.cfg == nil {
 		return nil

internal/runtime/executor/gemini_vertex_executor.go

@@ -15,7 +15,7 @@ import (
 
 	vertexauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/vertex"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
@@ -50,11 +50,49 @@ func NewGeminiVertexExecutor(cfg *config.Config) *GeminiVertexExecutor {
 // Identifier returns the executor identifier.
 func (e *GeminiVertexExecutor) Identifier() string { return "vertex" }
 
-// PrepareRequest prepares the HTTP request for execution (no-op for Vertex).
-func (e *GeminiVertexExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error {
+// PrepareRequest injects Vertex credentials into the outgoing HTTP request.
+func (e *GeminiVertexExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	apiKey, _ := vertexAPICreds(auth)
+	if strings.TrimSpace(apiKey) != "" {
+		req.Header.Set("x-goog-api-key", apiKey)
+		req.Header.Del("Authorization")
+		return nil
+	}
+	_, _, saJSON, errCreds := vertexCreds(auth)
+	if errCreds != nil {
+		return errCreds
+	}
+	token, errToken := vertexAccessToken(req.Context(), e.cfg, auth, saJSON)
+	if errToken != nil {
+		return errToken
+	}
+	if strings.TrimSpace(token) == "" {
+		return statusErr{code: http.StatusUnauthorized, msg: "missing access token"}
+	}
+	req.Header.Set("Authorization", "Bearer "+token)
+	req.Header.Del("x-goog-api-key")
 	return nil
 }
 
+// HttpRequest injects Vertex credentials into the request and executes it.
+func (e *GeminiVertexExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("vertex executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
+
 // Execute performs a non-streaming request to the Vertex AI API.
 func (e *GeminiVertexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
 	// Try API key authentication first
@@ -117,30 +155,29 @@ func (e *GeminiVertexExecutor) Refresh(_ context.Context, auth *cliproxyauth.Aut
 // executeWithServiceAccount handles authentication using service account credentials.
 // This method contains the original service account authentication logic.
 func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (resp cliproxyexecutor.Response, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
+
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		body = util.ApplyGeminiThinkingConfig(body, budgetOverride, includeOverride)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return resp, err
 	}
-	body = util.ApplyDefaultThinkingIfNeeded(req.Model, body)
-	body = util.NormalizeGeminiThinkingBudget(req.Model, body)
-	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
-	body = fixGeminiImageAspectRatio(req.Model, body)
-	body = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", body, originalTranslated)
-	body, _ = sjson.SetBytes(body, "model", req.Model)
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	action := "generateContent"
 	if req.Metadata != nil {
@@ -149,7 +186,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 		}
 	}
 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, action)
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, baseModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -220,35 +257,29 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 
 // executeWithAPIKey handles authentication using API key credentials.
 func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (resp cliproxyexecutor.Response, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
-	model := req.Model
-	if override := e.resolveUpstreamModel(req.Model, auth); override != "" {
-		model = override
-	}
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
+
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, model, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, req.Metadata); ok && util.ModelSupportsThinking(model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		body = util.ApplyGeminiThinkingConfig(body, budgetOverride, includeOverride)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return resp, err
 	}
-	body = util.ApplyDefaultThinkingIfNeeded(model, body)
-	body = util.NormalizeGeminiThinkingBudget(model, body)
-	body = util.StripThinkingConfigIfUnsupported(model, body)
-	body = fixGeminiImageAspectRatio(model, body)
-	body = applyPayloadConfigWithRoot(e.cfg, model, to.String(), "", body, originalTranslated)
-	body, _ = sjson.SetBytes(body, "model", model)
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	action := "generateContent"
 	if req.Metadata != nil {
@@ -261,7 +292,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	if baseURL == "" {
 		baseURL = "https://generativelanguage.googleapis.com"
 	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, model, action)
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, action)
 	if opts.Alt != "" && action != "countTokens" {
 		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
 	}
@@ -329,33 +360,32 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 
 // executeStreamWithServiceAccount handles streaming authentication using service account credentials.
 func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
+
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		body = util.ApplyGeminiThinkingConfig(body, budgetOverride, includeOverride)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return nil, err
 	}
-	body = util.ApplyDefaultThinkingIfNeeded(req.Model, body)
-	body = util.NormalizeGeminiThinkingBudget(req.Model, body)
-	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
-	body = fixGeminiImageAspectRatio(req.Model, body)
-	body = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", body, originalTranslated)
-	body, _ = sjson.SetBytes(body, "model", req.Model)
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, baseModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
@@ -449,41 +479,35 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 
 // executeStreamWithAPIKey handles streaming authentication using API key credentials.
 func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
-	model := req.Model
-	if override := e.resolveUpstreamModel(req.Model, auth); override != "" {
-		model = override
-	}
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
+
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, model, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), true)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, req.Metadata); ok && util.ModelSupportsThinking(model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		body = util.ApplyGeminiThinkingConfig(body, budgetOverride, includeOverride)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return nil, err
 	}
-	body = util.ApplyDefaultThinkingIfNeeded(model, body)
-	body = util.NormalizeGeminiThinkingBudget(model, body)
-	body = util.StripThinkingConfigIfUnsupported(model, body)
-	body = fixGeminiImageAspectRatio(model, body)
-	body = applyPayloadConfigWithRoot(e.cfg, model, to.String(), "", body, originalTranslated)
-	body, _ = sjson.SetBytes(body, "model", model)
+
+	body = fixGeminiImageAspectRatio(baseModel, body)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	// For API key auth, use simpler URL format without project/location
 	if baseURL == "" {
 		baseURL = "https://generativelanguage.googleapis.com"
 	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, model, "streamGenerateContent")
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, "streamGenerateContent")
 	if opts.Alt == "" {
 		url = url + "?alt=sse"
 	} else {
@@ -574,26 +598,27 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 
 // countTokensWithServiceAccount counts tokens using service account credentials.
 func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		translatedReq = util.ApplyGeminiThinkingConfig(translatedReq, budgetOverride, includeOverride)
+
+	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String())
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
 	}
-	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
-	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
-	translatedReq, _ = sjson.SetBytes(translatedReq, "model", req.Model)
+
+	translatedReq = fixGeminiImageAspectRatio(baseModel, translatedReq)
+	translatedReq, _ = sjson.SetBytes(translatedReq, "model", baseModel)
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "safetySettings")
 
 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, "countTokens")
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, baseModel, "countTokens")
 
 	httpReq, errNewReq := http.NewRequestWithContext(respCtx, http.MethodPost, url, bytes.NewReader(translatedReq))
 	if errNewReq != nil {
@@ -650,10 +675,6 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 		return cliproxyexecutor.Response{}, errRead
 	}
 	appendAPIResponseChunk(ctx, e.cfg, data)
-	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
-		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(data)}
-	}
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
 	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
@@ -661,24 +682,20 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 
 // countTokensWithAPIKey handles token counting using API key credentials.
 func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (cliproxyexecutor.Response, error) {
-	model := req.Model
-	if override := e.resolveUpstreamModel(req.Model, auth); override != "" {
-		model = override
-	}
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	translatedReq := sdktranslator.TranslateRequest(from, to, model, bytes.Clone(req.Payload), false)
-	if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, req.Metadata); ok && util.ModelSupportsThinking(model) {
-		if budgetOverride != nil {
-			norm := util.NormalizeThinkingBudget(model, *budgetOverride)
-			budgetOverride = &norm
-		}
-		translatedReq = util.ApplyGeminiThinkingConfig(translatedReq, budgetOverride, includeOverride)
+
+	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String())
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
 	}
-	translatedReq = util.StripThinkingConfigIfUnsupported(model, translatedReq)
-	translatedReq = fixGeminiImageAspectRatio(model, translatedReq)
-	translatedReq, _ = sjson.SetBytes(translatedReq, "model", model)
+
+	translatedReq = fixGeminiImageAspectRatio(baseModel, translatedReq)
+	translatedReq, _ = sjson.SetBytes(translatedReq, "model", baseModel)
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
 	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
@@ -688,7 +705,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 	if baseURL == "" {
 		baseURL = "https://generativelanguage.googleapis.com"
 	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, model, "countTokens")
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, "countTokens")
 
 	httpReq, errNewReq := http.NewRequestWithContext(respCtx, http.MethodPost, url, bytes.NewReader(translatedReq))
 	if errNewReq != nil {
@@ -742,10 +759,6 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 		return cliproxyexecutor.Response{}, errRead
 	}
 	appendAPIResponseChunk(ctx, e.cfg, data)
-	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
-		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(data)}
-	}
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
 	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
@@ -832,53 +845,6 @@ func vertexAccessToken(ctx context.Context, cfg *config.Config, auth *cliproxyau
 	return tok.AccessToken, nil
 }
 
-// resolveUpstreamModel resolves the upstream model name from vertex-api-key configuration.
-// It matches the requested model alias against configured models and returns the actual upstream name.
-func (e *GeminiVertexExecutor) resolveUpstreamModel(alias string, auth *cliproxyauth.Auth) string {
-	trimmed := strings.TrimSpace(alias)
-	if trimmed == "" {
-		return ""
-	}
-
-	entry := e.resolveVertexConfig(auth)
-	if entry == nil {
-		return ""
-	}
-
-	normalizedModel, metadata := util.NormalizeThinkingModel(trimmed)
-
-	// Candidate names to match against configured aliases/names.
-	candidates := []string{strings.TrimSpace(normalizedModel)}
-	if !strings.EqualFold(normalizedModel, trimmed) {
-		candidates = append(candidates, trimmed)
-	}
-	if original := util.ResolveOriginalModel(normalizedModel, metadata); original != "" && !strings.EqualFold(original, normalizedModel) {
-		candidates = append(candidates, original)
-	}
-
-	for i := range entry.Models {
-		model := entry.Models[i]
-		name := strings.TrimSpace(model.Name)
-		modelAlias := strings.TrimSpace(model.Alias)
-
-		for _, candidate := range candidates {
-			if candidate == "" {
-				continue
-			}
-			if modelAlias != "" && strings.EqualFold(modelAlias, candidate) {
-				if name != "" {
-					return name
-				}
-				return candidate
-			}
-			if name != "" && strings.EqualFold(name, candidate) {
-				return name
-			}
-		}
-	}
-	return ""
-}
-
 // resolveVertexConfig finds the matching vertex-api-key configuration entry for the given auth.
 func (e *GeminiVertexExecutor) resolveVertexConfig(auth *cliproxyauth.Auth) *config.VertexCompatKey {
 	if auth == nil || e.cfg == nil {

internal/runtime/executor/iflow_executor.go

@@ -12,6 +12,7 @@ import (
 
 	iflowauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/iflow"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -37,11 +38,38 @@ func NewIFlowExecutor(cfg *config.Config) *IFlowExecutor { return &IFlowExecutor
 // Identifier returns the provider key.
 func (e *IFlowExecutor) Identifier() string { return "iflow" }
 
-// PrepareRequest implements ProviderExecutor but requires no preprocessing.
-func (e *IFlowExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects iFlow credentials into the outgoing HTTP request.
+func (e *IFlowExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	apiKey, _ := iflowCreds(auth)
+	if strings.TrimSpace(apiKey) != "" {
+		req.Header.Set("Authorization", "Bearer "+apiKey)
+	}
+	return nil
+}
+
+// HttpRequest injects iFlow credentials into the request and executes it.
+func (e *IFlowExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("iflow executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 // Execute performs a non-streaming chat completion request.
 func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	apiKey, baseURL := iflowCreds(auth)
 	if strings.TrimSpace(apiKey) == "" {
 		err = fmt.Errorf("iflow executor: missing api key")
@@ -51,7 +79,7 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 		baseURL = iflowauth.DefaultAPIBaseURL
 	}
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
@@ -60,17 +88,17 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
-	body, _ = sjson.SetBytes(body, "model", req.Model)
-	body = NormalizeThinkingConfig(body, req.Model, false)
-	if errValidate := ValidateThinkingConfig(body, req.Model); errValidate != nil {
-		return resp, errValidate
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "iflow")
+	if err != nil {
+		return resp, err
 	}
-	body = applyIFlowThinkingConfig(body)
+
 	body = preserveReasoningContentInMessages(body)
-	body = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", body, originalTranslated)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	endpoint := strings.TrimSuffix(baseURL, "/") + iflowDefaultEndpoint
 
@@ -129,6 +157,8 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	reporter.ensurePublished(ctx)
 
 	var param any
+	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
+	// the original model name in the response for client compatibility.
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
@@ -136,6 +166,8 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 
 // ExecuteStream performs a streaming chat completion request.
 func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	apiKey, baseURL := iflowCreds(auth)
 	if strings.TrimSpace(apiKey) == "" {
 		err = fmt.Errorf("iflow executor: missing api key")
@@ -145,7 +177,7 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 		baseURL = iflowauth.DefaultAPIBaseURL
 	}
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
@@ -154,23 +186,22 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
-	body, _ = sjson.SetBytes(body, "model", req.Model)
-	body = NormalizeThinkingConfig(body, req.Model, false)
-	if errValidate := ValidateThinkingConfig(body, req.Model); errValidate != nil {
-		return nil, errValidate
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "iflow")
+	if err != nil {
+		return nil, err
 	}
-	body = applyIFlowThinkingConfig(body)
+
 	body = preserveReasoningContentInMessages(body)
 	// Ensure tools array exists to avoid provider quirks similar to Qwen's behaviour.
 	toolsResult := gjson.GetBytes(body, "tools")
 	if toolsResult.Exists() && toolsResult.IsArray() && len(toolsResult.Array()) == 0 {
 		body = ensureToolsArray(body)
 	}
-	body = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", body, originalTranslated)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	endpoint := strings.TrimSuffix(baseURL, "/") + iflowDefaultEndpoint
 
@@ -253,11 +284,13 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 }
 
 func (e *IFlowExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	enc, err := tokenizerForModel(req.Model)
+	enc, err := tokenizerForModel(baseModel)
 	if err != nil {
 		return cliproxyexecutor.Response{}, fmt.Errorf("iflow executor: tokenizer init failed: %w", err)
 	}
@@ -495,41 +528,3 @@ func preserveReasoningContentInMessages(body []byte) []byte {
 
 	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-4.6/4.7: Uses chat_template_kwargs.enable_thinking (boolean) and chat_template_kwargs.clear_thinking=false
-//   - MiniMax M2/M2.1: Uses reasoning_split=true for OpenAI-style reasoning separation
-func applyIFlowThinkingConfig(body []byte) []byte {
-	effort := gjson.GetBytes(body, "reasoning_effort")
-	if !effort.Exists() {
-		return body
-	}
-
-	model := strings.ToLower(gjson.GetBytes(body, "model").String())
-	val := strings.ToLower(strings.TrimSpace(effort.String()))
-	enableThinking := val != "none" && val != ""
-
-	// Remove reasoning_effort as we'll convert to model-specific format
-	body, _ = sjson.DeleteBytes(body, "reasoning_effort")
-	body, _ = sjson.DeleteBytes(body, "thinking")
-
-	// GLM-4.6/4.7: Use chat_template_kwargs
-	if strings.HasPrefix(model, "glm-4") {
-		body, _ = sjson.SetBytes(body, "chat_template_kwargs.enable_thinking", enableThinking)
-		if enableThinking {
-			body, _ = sjson.SetBytes(body, "chat_template_kwargs.clear_thinking", false)
-		}
-		return body
-	}
-
-	// MiniMax M2/M2.1: Use reasoning_split
-	if strings.HasPrefix(model, "minimax-m2") {
-		body, _ = sjson.SetBytes(body, "reasoning_split", enableThinking)
-		return body
-	}
-
-	return body
-}

internal/runtime/executor/iflow_executor_test.go

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

internal/runtime/executor/logging_helpers.go

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

internal/runtime/executor/openai_compat_executor.go

@@ -11,6 +11,7 @@ import (
 	"time"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -35,13 +36,43 @@ func NewOpenAICompatExecutor(provider string, cfg *config.Config) *OpenAICompatE
 // Identifier implements cliproxyauth.ProviderExecutor.
 func (e *OpenAICompatExecutor) Identifier() string { return e.provider }
 
-// PrepareRequest is a no-op for now (credentials are added via headers at execution time).
-func (e *OpenAICompatExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error {
+// PrepareRequest injects OpenAI-compatible credentials into the outgoing HTTP request.
+func (e *OpenAICompatExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	_, apiKey := e.resolveCredentials(auth)
+	if strings.TrimSpace(apiKey) != "" {
+		req.Header.Set("Authorization", "Bearer "+apiKey)
+	}
+	var attrs map[string]string
+	if auth != nil {
+		attrs = auth.Attributes
+	}
+	util.ApplyCustomHeadersFromAttrs(req, attrs)
 	return nil
 }
 
+// HttpRequest injects OpenAI-compatible credentials into the request and executes it.
+func (e *OpenAICompatExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("openai compat executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
+
 func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	baseURL, apiKey := e.resolveCredentials(auth)
@@ -57,18 +88,13 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, opts.Stream)
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), opts.Stream)
-	modelOverride := e.resolveUpstreamModel(req.Model, auth)
-	if modelOverride != "" {
-		translated = e.overrideModel(translated, modelOverride)
-	}
-	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated, originalTranslated)
-	allowCompat := e.allowCompatReasoningEffort(req.Model, auth)
-	translated = ApplyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
-	translated = NormalizeThinkingConfig(translated, req.Model, allowCompat)
-	if errValidate := ValidateThinkingConfig(translated, req.Model); errValidate != nil {
-		return resp, errValidate
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, opts.Stream)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), opts.Stream)
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated)
+
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
+	if err != nil {
+		return resp, err
 	}
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
@@ -140,7 +166,9 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 }
 
 func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	baseURL, apiKey := e.resolveCredentials(auth)
@@ -148,24 +176,20 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 		err = statusErr{code: http.StatusUnauthorized, msg: "missing provider baseURL"}
 		return nil, err
 	}
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, true)
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	modelOverride := e.resolveUpstreamModel(req.Model, auth)
-	if modelOverride != "" {
-		translated = e.overrideModel(translated, modelOverride)
-	}
-	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated, originalTranslated)
-	allowCompat := e.allowCompatReasoningEffort(req.Model, auth)
-	translated = ApplyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
-	translated = NormalizeThinkingConfig(translated, req.Model, allowCompat)
-	if errValidate := ValidateThinkingConfig(translated, req.Model); errValidate != nil {
-		return nil, errValidate
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated)
+
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
+	if err != nil {
+		return nil, err
 	}
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
@@ -241,6 +265,11 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 			if len(line) == 0 {
 				continue
 			}
+
+			if !bytes.HasPrefix(line, []byte("data:")) {
+				continue
+			}
+
 			// OpenAI-compatible streams are SSE: lines typically prefixed with "data: ".
 			// Pass through translator; it yields one or more chunks for the target schema.
 			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, bytes.Clone(line), &param)
@@ -260,14 +289,17 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 }
 
 func (e *OpenAICompatExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	modelForCounting := req.Model
-	if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
-		translated = e.overrideModel(translated, modelOverride)
-		modelForCounting = modelOverride
+	modelForCounting := baseModel
+
+	translated, err := thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
 	}
 
 	enc, err := tokenizerForModel(modelForCounting)
@@ -303,53 +335,6 @@ func (e *OpenAICompatExecutor) resolveCredentials(auth *cliproxyauth.Auth) (base
 	return
 }
 
-func (e *OpenAICompatExecutor) resolveUpstreamModel(alias string, auth *cliproxyauth.Auth) string {
-	if alias == "" || auth == nil || e.cfg == nil {
-		return ""
-	}
-	compat := e.resolveCompatConfig(auth)
-	if compat == nil {
-		return ""
-	}
-	for i := range compat.Models {
-		model := compat.Models[i]
-		if model.Alias != "" {
-			if strings.EqualFold(model.Alias, alias) {
-				if model.Name != "" {
-					return model.Name
-				}
-				return alias
-			}
-			continue
-		}
-		if strings.EqualFold(model.Name, alias) {
-			return model.Name
-		}
-	}
-	return ""
-}
-
-func (e *OpenAICompatExecutor) allowCompatReasoningEffort(model string, auth *cliproxyauth.Auth) bool {
-	trimmed := strings.TrimSpace(model)
-	if trimmed == "" || e == nil || e.cfg == nil {
-		return false
-	}
-	compat := e.resolveCompatConfig(auth)
-	if compat == nil || len(compat.Models) == 0 {
-		return false
-	}
-	for i := range compat.Models {
-		entry := compat.Models[i]
-		if strings.EqualFold(strings.TrimSpace(entry.Alias), trimmed) {
-			return true
-		}
-		if strings.EqualFold(strings.TrimSpace(entry.Name), trimmed) {
-			return true
-		}
-	}
-	return false
-}
-
 func (e *OpenAICompatExecutor) resolveCompatConfig(auth *cliproxyauth.Auth) *config.OpenAICompatibility {
 	if auth == nil || e.cfg == nil {
 		return nil

internal/runtime/executor/payload_helpers.go

@@ -1,109 +1,14 @@
 package executor
 
 import (
-	"fmt"
-	"net/http"
+	"encoding/json"
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
 
-// ApplyThinkingMetadata applies thinking config from model suffix metadata (e.g., (high), (8192))
-// for standard Gemini format payloads. It normalizes the budget when the model supports thinking.
-func ApplyThinkingMetadata(payload []byte, metadata map[string]any, model string) []byte {
-	// Use the alias from metadata if available, as it's registered in the global registry
-	// with thinking metadata; the upstream model name may not be registered.
-	lookupModel := util.ResolveOriginalModel(model, metadata)
-
-	// Determine which model to use for thinking support check.
-	// If the alias (lookupModel) is not in the registry, fall back to the upstream model.
-	thinkingModel := lookupModel
-	if !util.ModelSupportsThinking(lookupModel) && util.ModelSupportsThinking(model) {
-		thinkingModel = model
-	}
-
-	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(thinkingModel, metadata)
-	if !ok || (budgetOverride == nil && includeOverride == nil) {
-		return payload
-	}
-	if !util.ModelSupportsThinking(thinkingModel) {
-		return payload
-	}
-	if budgetOverride != nil {
-		norm := util.NormalizeThinkingBudget(thinkingModel, *budgetOverride)
-		budgetOverride = &norm
-	}
-	return util.ApplyGeminiThinkingConfig(payload, budgetOverride, includeOverride)
-}
-
-// ApplyThinkingMetadataCLI applies thinking config from model suffix metadata (e.g., (high), (8192))
-// for Gemini CLI format payloads (nested under "request"). It normalizes the budget when the model supports thinking.
-func ApplyThinkingMetadataCLI(payload []byte, metadata map[string]any, model string) []byte {
-	// Use the alias from metadata if available, as it's registered in the global registry
-	// with thinking metadata; the upstream model name may not be registered.
-	lookupModel := util.ResolveOriginalModel(model, metadata)
-
-	// Determine which model to use for thinking support check.
-	// If the alias (lookupModel) is not in the registry, fall back to the upstream model.
-	thinkingModel := lookupModel
-	if !util.ModelSupportsThinking(lookupModel) && util.ModelSupportsThinking(model) {
-		thinkingModel = model
-	}
-
-	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(thinkingModel, metadata)
-	if !ok || (budgetOverride == nil && includeOverride == nil) {
-		return payload
-	}
-	if !util.ModelSupportsThinking(thinkingModel) {
-		return payload
-	}
-	if budgetOverride != nil {
-		norm := util.NormalizeThinkingBudget(thinkingModel, *budgetOverride)
-		budgetOverride = &norm
-	}
-	return util.ApplyGeminiCLIThinkingConfig(payload, budgetOverride, includeOverride)
-}
-
-// ApplyReasoningEffortMetadata applies reasoning effort overrides from metadata to the given JSON path.
-// Metadata values take precedence over any existing field when the model supports thinking, intentionally
-// overwriting caller-provided values to honor suffix/default metadata priority.
-func ApplyReasoningEffortMetadata(payload []byte, metadata map[string]any, model, field string, allowCompat bool) []byte {
-	if len(metadata) == 0 {
-		return payload
-	}
-	if field == "" {
-		return payload
-	}
-	baseModel := util.ResolveOriginalModel(model, metadata)
-	if baseModel == "" {
-		baseModel = model
-	}
-	if !util.ModelSupportsThinking(baseModel) && !allowCompat {
-		return payload
-	}
-	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
-		if util.ModelUsesThinkingLevels(baseModel) || allowCompat {
-			if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
-				return updated
-			}
-		}
-	}
-	// Fallback: numeric thinking_budget suffix for level-based (OpenAI-style) models.
-	if util.ModelUsesThinkingLevels(baseModel) || allowCompat {
-		if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
-			if effort, ok := util.ThinkingBudgetToEffort(baseModel, *budget); ok && effort != "" {
-				if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
-					return updated
-				}
-			}
-		}
-	}
-	return payload
-}
-
 // applyPayloadConfigWithRoot behaves like applyPayloadConfig but treats all parameter
 // paths as relative to the provided root path (for example, "request" for Gemini CLI)
 // and restricts matches to the given protocol when supplied. Defaults are checked
@@ -113,13 +18,14 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 		return payload
 	}
 	rules := cfg.Payload
-	if len(rules.Default) == 0 && len(rules.Override) == 0 {
+	if len(rules.Default) == 0 && len(rules.DefaultRaw) == 0 && len(rules.Override) == 0 && len(rules.OverrideRaw) == 0 {
 		return payload
 	}
 	model = strings.TrimSpace(model)
 	if model == "" {
 		return payload
 	}
+	candidates := payloadModelCandidates(cfg, model, protocol)
 	out := payload
 	source := original
 	if len(source) == 0 {
@@ -129,7 +35,7 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 	// Apply default rules: first write wins per field across all matching rules.
 	for i := range rules.Default {
 		rule := &rules.Default[i]
-		if !payloadRuleMatchesModel(rule, model, protocol) {
+		if !payloadRuleMatchesModels(rule, protocol, candidates) {
 			continue
 		}
 		for path, value := range rule.Params {
@@ -151,10 +57,39 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 			appliedDefaults[fullPath] = struct{}{}
 		}
 	}
+	// Apply default raw rules: first write wins per field across all matching rules.
+	for i := range rules.DefaultRaw {
+		rule := &rules.DefaultRaw[i]
+		if !payloadRuleMatchesModels(rule, protocol, candidates) {
+			continue
+		}
+		for path, value := range rule.Params {
+			fullPath := buildPayloadPath(root, path)
+			if fullPath == "" {
+				continue
+			}
+			if gjson.GetBytes(source, fullPath).Exists() {
+				continue
+			}
+			if _, ok := appliedDefaults[fullPath]; ok {
+				continue
+			}
+			rawValue, ok := payloadRawValue(value)
+			if !ok {
+				continue
+			}
+			updated, errSet := sjson.SetRawBytes(out, fullPath, rawValue)
+			if errSet != nil {
+				continue
+			}
+			out = updated
+			appliedDefaults[fullPath] = struct{}{}
+		}
+	}
 	// Apply override rules: last write wins per field across all matching rules.
 	for i := range rules.Override {
 		rule := &rules.Override[i]
-		if !payloadRuleMatchesModel(rule, model, protocol) {
+		if !payloadRuleMatchesModels(rule, protocol, candidates) {
 			continue
 		}
 		for path, value := range rule.Params {
@@ -169,9 +104,43 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 			out = updated
 		}
 	}
+	// Apply override raw rules: last write wins per field across all matching rules.
+	for i := range rules.OverrideRaw {
+		rule := &rules.OverrideRaw[i]
+		if !payloadRuleMatchesModels(rule, protocol, candidates) {
+			continue
+		}
+		for path, value := range rule.Params {
+			fullPath := buildPayloadPath(root, path)
+			if fullPath == "" {
+				continue
+			}
+			rawValue, ok := payloadRawValue(value)
+			if !ok {
+				continue
+			}
+			updated, errSet := sjson.SetRawBytes(out, fullPath, rawValue)
+			if errSet != nil {
+				continue
+			}
+			out = updated
+		}
+	}
 	return out
 }
 
+func payloadRuleMatchesModels(rule *config.PayloadRule, protocol string, models []string) bool {
+	if rule == nil || len(models) == 0 {
+		return false
+	}
+	for _, model := range models {
+		if payloadRuleMatchesModel(rule, model, protocol) {
+			return true
+		}
+	}
+	return false
+}
+
 func payloadRuleMatchesModel(rule *config.PayloadRule, model, protocol string) bool {
 	if rule == nil {
 		return false
@@ -194,6 +163,65 @@ func payloadRuleMatchesModel(rule *config.PayloadRule, model, protocol string) b
 	return false
 }
 
+func payloadModelCandidates(cfg *config.Config, model, protocol string) []string {
+	model = strings.TrimSpace(model)
+	if model == "" {
+		return nil
+	}
+	candidates := []string{model}
+	if cfg == nil {
+		return candidates
+	}
+	aliases := payloadModelAliases(cfg, model, protocol)
+	if len(aliases) == 0 {
+		return candidates
+	}
+	seen := map[string]struct{}{strings.ToLower(model): struct{}{}}
+	for _, alias := range aliases {
+		alias = strings.TrimSpace(alias)
+		if alias == "" {
+			continue
+		}
+		key := strings.ToLower(alias)
+		if _, ok := seen[key]; ok {
+			continue
+		}
+		seen[key] = struct{}{}
+		candidates = append(candidates, alias)
+	}
+	return candidates
+}
+
+func payloadModelAliases(cfg *config.Config, model, protocol string) []string {
+	if cfg == nil {
+		return nil
+	}
+	model = strings.TrimSpace(model)
+	if model == "" {
+		return nil
+	}
+	channel := strings.ToLower(strings.TrimSpace(protocol))
+	if channel == "" {
+		return nil
+	}
+	entries := cfg.OAuthModelAlias[channel]
+	if len(entries) == 0 {
+		return nil
+	}
+	aliases := make([]string, 0, 2)
+	for _, entry := range entries {
+		if !strings.EqualFold(strings.TrimSpace(entry.Name), model) {
+			continue
+		}
+		alias := strings.TrimSpace(entry.Alias)
+		if alias == "" {
+			continue
+		}
+		aliases = append(aliases, alias)
+	}
+	return aliases
+}
+
 // buildPayloadPath combines an optional root path with a relative parameter path.
 // When root is empty, the parameter path is used as-is. When root is non-empty,
 // the parameter path is treated as relative to root.
@@ -212,6 +240,24 @@ func buildPayloadPath(root, path string) string {
 	return r + "." + p
 }
 
+func payloadRawValue(value any) ([]byte, bool) {
+	if value == nil {
+		return nil, false
+	}
+	switch typed := value.(type) {
+	case string:
+		return []byte(typed), true
+	case []byte:
+		return typed, true
+	default:
+		raw, errMarshal := json.Marshal(typed)
+		if errMarshal != nil {
+			return nil, false
+		}
+		return raw, true
+	}
+}
+
 // matchModelPattern performs simple wildcard matching where '*' matches zero or more characters.
 // Examples:
 //
@@ -256,102 +302,3 @@ func matchModelPattern(pattern, model string) bool {
 	}
 	return pi == len(pattern)
 }
-
-// NormalizeThinkingConfig normalizes thinking-related fields in the payload
-// based on model capabilities. For models without thinking support, it strips
-// reasoning fields. For models with level-based thinking, it validates and
-// normalizes the reasoning effort level. For models with numeric budget thinking,
-// it strips the effort string fields.
-func NormalizeThinkingConfig(payload []byte, model string, allowCompat bool) []byte {
-	if len(payload) == 0 || model == "" {
-		return payload
-	}
-
-	if !util.ModelSupportsThinking(model) {
-		if allowCompat {
-			return payload
-		}
-		return StripThinkingFields(payload, false)
-	}
-
-	if util.ModelUsesThinkingLevels(model) {
-		return NormalizeReasoningEffortLevel(payload, model)
-	}
-
-	// Model supports thinking but uses numeric budgets, not levels.
-	// Strip effort string fields since they are not applicable.
-	return StripThinkingFields(payload, true)
-}
-
-// StripThinkingFields removes thinking-related fields from the payload for
-// models that do not support thinking. If effortOnly is true, only removes
-// effort string fields (for models using numeric budgets).
-func StripThinkingFields(payload []byte, effortOnly bool) []byte {
-	fieldsToRemove := []string{
-		"reasoning_effort",
-		"reasoning.effort",
-	}
-	if !effortOnly {
-		fieldsToRemove = append([]string{"reasoning", "thinking"}, fieldsToRemove...)
-	}
-	out := payload
-	for _, field := range fieldsToRemove {
-		if gjson.GetBytes(out, field).Exists() {
-			out, _ = sjson.DeleteBytes(out, field)
-		}
-	}
-	return out
-}
-
-// NormalizeReasoningEffortLevel validates and normalizes the reasoning_effort
-// or reasoning.effort field for level-based thinking models.
-func NormalizeReasoningEffortLevel(payload []byte, model string) []byte {
-	out := payload
-
-	if effort := gjson.GetBytes(out, "reasoning_effort"); effort.Exists() {
-		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
-			out, _ = sjson.SetBytes(out, "reasoning_effort", normalized)
-		}
-	}
-
-	if effort := gjson.GetBytes(out, "reasoning.effort"); effort.Exists() {
-		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
-			out, _ = sjson.SetBytes(out, "reasoning.effort", normalized)
-		}
-	}
-
-	return out
-}
-
-// ValidateThinkingConfig checks for unsupported reasoning levels on level-based models.
-// Returns a statusErr with 400 when an unsupported level is supplied to avoid silently
-// downgrading requests.
-func ValidateThinkingConfig(payload []byte, model string) error {
-	if len(payload) == 0 || model == "" {
-		return nil
-	}
-	if !util.ModelSupportsThinking(model) || !util.ModelUsesThinkingLevels(model) {
-		return nil
-	}
-
-	levels := util.GetModelThinkingLevels(model)
-	checkField := func(path string) error {
-		if effort := gjson.GetBytes(payload, path); effort.Exists() {
-			if _, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); !ok {
-				return statusErr{
-					code: http.StatusBadRequest,
-					msg:  fmt.Sprintf("unsupported reasoning effort level %q for model %s (supported: %s)", effort.String(), model, strings.Join(levels, ", ")),
-				}
-			}
-		}
-		return nil
-	}
-
-	if err := checkField("reasoning_effort"); err != nil {
-		return err
-	}
-	if err := checkField("reasoning.effort"); err != nil {
-		return err
-	}
-	return nil
-}

internal/runtime/executor/qwen_executor.go

@@ -12,6 +12,7 @@ import (
 
 	qwenauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/qwen"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
@@ -36,15 +37,43 @@ func NewQwenExecutor(cfg *config.Config) *QwenExecutor { return &QwenExecutor{cf
 
 func (e *QwenExecutor) Identifier() string { return "qwen" }
 
-func (e *QwenExecutor) PrepareRequest(_ *http.Request, _ *cliproxyauth.Auth) error { return nil }
+// PrepareRequest injects Qwen credentials into the outgoing HTTP request.
+func (e *QwenExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	token, _ := qwenCreds(auth)
+	if strings.TrimSpace(token) != "" {
+		req.Header.Set("Authorization", "Bearer "+token)
+	}
+	return nil
+}
+
+// HttpRequest injects Qwen credentials into the request and executes it.
+func (e *QwenExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("qwen executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
 
 func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	token, baseURL := qwenCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	token, baseURL := qwenCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://portal.qwen.ai/v1"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
@@ -53,15 +82,16 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
-	body, _ = sjson.SetBytes(body, "model", req.Model)
-	body = NormalizeThinkingConfig(body, req.Model, false)
-	if errValidate := ValidateThinkingConfig(body, req.Model); errValidate != nil {
-		return resp, errValidate
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return resp, err
 	}
-	body = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", body, originalTranslated)
+
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
@@ -114,18 +144,22 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	reporter.publish(ctx, parseOpenAIUsage(data))
 	var param any
+	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
+	// the original model name in the response for client compatibility.
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
 }
 
 func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	token, baseURL := qwenCreds(auth)
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
+	token, baseURL := qwenCreds(auth)
 	if baseURL == "" {
 		baseURL = "https://portal.qwen.ai/v1"
 	}
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
 	from := opts.SourceFormat
@@ -134,15 +168,15 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
 	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
-	body, _ = sjson.SetBytes(body, "model", req.Model)
-	body = NormalizeThinkingConfig(body, req.Model, false)
-	if errValidate := ValidateThinkingConfig(body, req.Model); errValidate != nil {
-		return nil, errValidate
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
+	if err != nil {
+		return nil, err
 	}
+
 	toolsResult := gjson.GetBytes(body, "tools")
 	// I'm addressing the Qwen3 "poisoning" issue, which is caused by the model needing a tool to be defined. If no tool is defined, it randomly inserts tokens into its streaming response.
 	// This will have no real consequences. It's just to scare Qwen3.
@@ -150,7 +184,7 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		body, _ = sjson.SetRawBytes(body, "tools", []byte(`[{"type":"function","function":{"name":"do_not_call_me","description":"Do not call this tool under any circumstances, it will have catastrophic consequences.","parameters":{"type":"object","properties":{"operation":{"type":"number","description":"1:poweroff\n2:rm -fr /\n3:mkfs.ext4 /dev/sda1"}},"required":["operation"]}}}]`))
 	}
 	body, _ = sjson.SetBytes(body, "stream_options.include_usage", true)
-	body = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", body, originalTranslated)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
@@ -230,13 +264,15 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 }
 
 func (e *QwenExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
 	modelName := gjson.GetBytes(body, "model").String()
 	if strings.TrimSpace(modelName) == "" {
-		modelName = req.Model
+		modelName = baseModel
 	}
 
 	enc, err := tokenizerForModel(modelName)

internal/runtime/executor/qwen_executor_test.go

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

internal/runtime/executor/thinking_providers.go

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

internal/thinking/apply.go

@@ -0,0 +1,481 @@
+// Package thinking provides unified thinking configuration processing.
+package thinking
+
+import (
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	log "github.com/sirupsen/logrus"
+	"github.com/tidwall/gjson"
+)
+
+// providerAppliers maps provider names to their ProviderApplier implementations.
+var providerAppliers = map[string]ProviderApplier{
+	"gemini":      nil,
+	"gemini-cli":  nil,
+	"claude":      nil,
+	"openai":      nil,
+	"codex":       nil,
+	"iflow":       nil,
+	"antigravity": nil,
+}
+
+// GetProviderApplier returns the ProviderApplier for the given provider name.
+// Returns nil if the provider is not registered.
+func GetProviderApplier(provider string) ProviderApplier {
+	return providerAppliers[provider]
+}
+
+// RegisterProvider registers a provider applier by name.
+func RegisterProvider(name string, applier ProviderApplier) {
+	providerAppliers[name] = applier
+}
+
+// IsUserDefinedModel reports whether the model is a user-defined model that should
+// have thinking configuration passed through without validation.
+//
+// User-defined models are configured via config file's models[] array
+// (e.g., openai-compatibility.*.models[], *-api-key.models[]). These models
+// are marked with UserDefined=true at registration time.
+//
+// User-defined models should have their thinking configuration applied directly,
+// letting the upstream service validate the configuration.
+func IsUserDefinedModel(modelInfo *registry.ModelInfo) bool {
+	if modelInfo == nil {
+		return true
+	}
+	return modelInfo.UserDefined
+}
+
+// ApplyThinking applies thinking configuration to a request body.
+//
+// This is the unified entry point for all providers. It follows the processing
+// order defined in FR25: route check → model capability query → config extraction
+// → validation → application.
+//
+// Suffix Priority: When the model name includes a thinking suffix (e.g., "gemini-2.5-pro(8192)"),
+// the suffix configuration takes priority over any thinking parameters in the request body.
+// This enables users to override thinking settings via the model name without modifying their
+// request payload.
+//
+// Parameters:
+//   - body: Original request body JSON
+//   - model: Model name, optionally with thinking suffix (e.g., "claude-sonnet-4-5(16384)")
+//   - fromFormat: Source request format (e.g., openai, codex, gemini)
+//   - toFormat: Target provider format for the request body (gemini, gemini-cli, antigravity, claude, openai, codex, iflow)
+//
+// Returns:
+//   - Modified request body JSON with thinking configuration applied
+//   - Error if validation fails (ThinkingError). On error, the original body
+//     is returned (not nil) to enable defensive programming patterns.
+//
+// Passthrough behavior (returns original body without error):
+//   - Unknown provider (not in providerAppliers map)
+//   - modelInfo.Thinking is nil (model doesn't support thinking)
+//
+// Note: Unknown models (modelInfo is nil) are treated as user-defined models: we skip
+// validation and still apply the thinking config so the upstream can validate it.
+//
+// Example:
+//
+//	// With suffix - suffix config takes priority
+//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro(8192)", "gemini", "gemini")
+//
+//	// Without suffix - uses body config
+//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro", "gemini", "gemini")
+func ApplyThinking(body []byte, model string, fromFormat string, toFormat string) ([]byte, error) {
+	providerFormat := strings.ToLower(strings.TrimSpace(toFormat))
+	fromFormat = strings.ToLower(strings.TrimSpace(fromFormat))
+	if fromFormat == "" {
+		fromFormat = providerFormat
+	}
+	// 1. Route check: Get provider applier
+	applier := GetProviderApplier(providerFormat)
+	if applier == nil {
+		log.WithFields(log.Fields{
+			"provider": providerFormat,
+			"model":    model,
+		}).Debug("thinking: unknown provider, passthrough |")
+		return body, nil
+	}
+
+	// 2. Parse suffix and get modelInfo
+	suffixResult := ParseSuffix(model)
+	baseModel := suffixResult.ModelName
+	modelInfo := registry.LookupModelInfo(baseModel)
+
+	// 3. Model capability check
+	// Unknown models are treated as user-defined so thinking config can still be applied.
+	// The upstream service is responsible for validating the configuration.
+	if IsUserDefinedModel(modelInfo) {
+		return applyUserDefinedModel(body, modelInfo, fromFormat, providerFormat, suffixResult)
+	}
+	if modelInfo.Thinking == nil {
+		config := extractThinkingConfig(body, providerFormat)
+		if hasThinkingConfig(config) {
+			log.WithFields(log.Fields{
+				"model":    baseModel,
+				"provider": providerFormat,
+			}).Debug("thinking: model does not support thinking, stripping config |")
+			return StripThinkingConfig(body, providerFormat), nil
+		}
+		log.WithFields(log.Fields{
+			"provider": providerFormat,
+			"model":    baseModel,
+		}).Debug("thinking: model does not support thinking, passthrough |")
+		return body, nil
+	}
+
+	// 4. Get config: suffix priority over body
+	var config ThinkingConfig
+	if suffixResult.HasSuffix {
+		config = parseSuffixToConfig(suffixResult.RawSuffix, providerFormat, model)
+		log.WithFields(log.Fields{
+			"provider": providerFormat,
+			"model":    model,
+			"mode":     config.Mode,
+			"budget":   config.Budget,
+			"level":    config.Level,
+		}).Debug("thinking: config from model suffix |")
+	} else {
+		config = extractThinkingConfig(body, providerFormat)
+		if hasThinkingConfig(config) {
+			log.WithFields(log.Fields{
+				"provider": providerFormat,
+				"model":    modelInfo.ID,
+				"mode":     config.Mode,
+				"budget":   config.Budget,
+				"level":    config.Level,
+			}).Debug("thinking: original config from request |")
+		}
+	}
+
+	if !hasThinkingConfig(config) {
+		log.WithFields(log.Fields{
+			"provider": providerFormat,
+			"model":    modelInfo.ID,
+		}).Debug("thinking: no config found, passthrough |")
+		return body, nil
+	}
+
+	// 5. Validate and normalize configuration
+	validated, err := ValidateConfig(config, modelInfo, fromFormat, providerFormat, suffixResult.HasSuffix)
+	if err != nil {
+		log.WithFields(log.Fields{
+			"provider": providerFormat,
+			"model":    modelInfo.ID,
+			"error":    err.Error(),
+		}).Warn("thinking: validation failed |")
+		// Return original body on validation failure (defensive programming).
+		// This ensures callers who ignore the error won't receive nil body.
+		// The upstream service will decide how to handle the unmodified request.
+		return body, err
+	}
+
+	// Defensive check: ValidateConfig should never return (nil, nil)
+	if validated == nil {
+		log.WithFields(log.Fields{
+			"provider": providerFormat,
+			"model":    modelInfo.ID,
+		}).Warn("thinking: ValidateConfig returned nil config without error, passthrough |")
+		return body, nil
+	}
+
+	log.WithFields(log.Fields{
+		"provider": providerFormat,
+		"model":    modelInfo.ID,
+		"mode":     validated.Mode,
+		"budget":   validated.Budget,
+		"level":    validated.Level,
+	}).Debug("thinking: processed config to apply |")
+
+	// 6. Apply configuration using provider-specific applier
+	return applier.Apply(body, *validated, modelInfo)
+}
+
+// parseSuffixToConfig converts a raw suffix string to ThinkingConfig.
+//
+// Parsing priority:
+//  1. Special values: "none" → ModeNone, "auto"/"-1" → ModeAuto
+//  2. Level names: "minimal", "low", "medium", "high", "xhigh" → ModeLevel
+//  3. Numeric values: positive integers → ModeBudget, 0 → ModeNone
+//
+// If none of the above match, returns empty ThinkingConfig (treated as no config).
+func parseSuffixToConfig(rawSuffix, provider, model string) ThinkingConfig {
+	// 1. Try special values first (none, auto, -1)
+	if mode, ok := ParseSpecialSuffix(rawSuffix); ok {
+		switch mode {
+		case ModeNone:
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		case ModeAuto:
+			return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+		}
+	}
+
+	// 2. Try level parsing (minimal, low, medium, high, xhigh)
+	if level, ok := ParseLevelSuffix(rawSuffix); ok {
+		return ThinkingConfig{Mode: ModeLevel, Level: level}
+	}
+
+	// 3. Try numeric parsing
+	if budget, ok := ParseNumericSuffix(rawSuffix); ok {
+		if budget == 0 {
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		}
+		return ThinkingConfig{Mode: ModeBudget, Budget: budget}
+	}
+
+	// Unknown suffix format - return empty config
+	log.WithFields(log.Fields{
+		"provider":   provider,
+		"model":      model,
+		"raw_suffix": rawSuffix,
+	}).Debug("thinking: unknown suffix format, treating as no config |")
+	return ThinkingConfig{}
+}
+
+// applyUserDefinedModel applies thinking configuration for user-defined models
+// without ThinkingSupport validation.
+func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromFormat, toFormat string, suffixResult SuffixResult) ([]byte, error) {
+	// Get model ID for logging
+	modelID := ""
+	if modelInfo != nil {
+		modelID = modelInfo.ID
+	} else {
+		modelID = suffixResult.ModelName
+	}
+
+	// Get config: suffix priority over body
+	var config ThinkingConfig
+	if suffixResult.HasSuffix {
+		config = parseSuffixToConfig(suffixResult.RawSuffix, toFormat, modelID)
+	} else {
+		config = extractThinkingConfig(body, toFormat)
+	}
+
+	if !hasThinkingConfig(config) {
+		log.WithFields(log.Fields{
+			"model":    modelID,
+			"provider": toFormat,
+		}).Debug("thinking: user-defined model, passthrough (no config) |")
+		return body, nil
+	}
+
+	applier := GetProviderApplier(toFormat)
+	if applier == nil {
+		log.WithFields(log.Fields{
+			"model":    modelID,
+			"provider": toFormat,
+		}).Debug("thinking: user-defined model, passthrough (unknown provider) |")
+		return body, nil
+	}
+
+	log.WithFields(log.Fields{
+		"provider": toFormat,
+		"model":    modelID,
+		"mode":     config.Mode,
+		"budget":   config.Budget,
+		"level":    config.Level,
+	}).Debug("thinking: applying config for user-defined model (skip validation)")
+
+	config = normalizeUserDefinedConfig(config, fromFormat, toFormat)
+	return applier.Apply(body, config, modelInfo)
+}
+
+func normalizeUserDefinedConfig(config ThinkingConfig, fromFormat, toFormat string) ThinkingConfig {
+	if config.Mode != ModeLevel {
+		return config
+	}
+	if !isBudgetBasedProvider(toFormat) || !isLevelBasedProvider(fromFormat) {
+		return config
+	}
+	budget, ok := ConvertLevelToBudget(string(config.Level))
+	if !ok {
+		return config
+	}
+	config.Mode = ModeBudget
+	config.Budget = budget
+	config.Level = ""
+	return config
+}
+
+// extractThinkingConfig extracts provider-specific thinking config from request body.
+func extractThinkingConfig(body []byte, provider string) ThinkingConfig {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		return ThinkingConfig{}
+	}
+
+	switch provider {
+	case "claude":
+		return extractClaudeConfig(body)
+	case "gemini", "gemini-cli", "antigravity":
+		return extractGeminiConfig(body, provider)
+	case "openai":
+		return extractOpenAIConfig(body)
+	case "codex":
+		return extractCodexConfig(body)
+	case "iflow":
+		config := extractIFlowConfig(body)
+		if hasThinkingConfig(config) {
+			return config
+		}
+		return extractOpenAIConfig(body)
+	default:
+		return ThinkingConfig{}
+	}
+}
+
+func hasThinkingConfig(config ThinkingConfig) bool {
+	return config.Mode != ModeBudget || config.Budget != 0 || config.Level != ""
+}
+
+// extractClaudeConfig extracts thinking configuration from Claude format request body.
+//
+// Claude API format:
+//   - thinking.type: "enabled" or "disabled"
+//   - thinking.budget_tokens: integer (-1=auto, 0=disabled, >0=budget)
+//
+// Priority: thinking.type="disabled" takes precedence over budget_tokens.
+// When type="enabled" without budget_tokens, returns ModeAuto to indicate
+// the user wants thinking enabled but didn't specify a budget.
+func extractClaudeConfig(body []byte) ThinkingConfig {
+	thinkingType := gjson.GetBytes(body, "thinking.type").String()
+	if thinkingType == "disabled" {
+		return ThinkingConfig{Mode: ModeNone, Budget: 0}
+	}
+
+	// Check budget_tokens
+	if budget := gjson.GetBytes(body, "thinking.budget_tokens"); budget.Exists() {
+		value := int(budget.Int())
+		switch value {
+		case 0:
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		case -1:
+			return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+		default:
+			return ThinkingConfig{Mode: ModeBudget, Budget: value}
+		}
+	}
+
+	// If type="enabled" but no budget_tokens, treat as auto (user wants thinking but no budget specified)
+	if thinkingType == "enabled" {
+		return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+	}
+
+	return ThinkingConfig{}
+}
+
+// extractGeminiConfig extracts thinking configuration from Gemini format request body.
+//
+// Gemini API format:
+//   - generationConfig.thinkingConfig.thinkingLevel: "none", "auto", or level name (Gemini 3)
+//   - generationConfig.thinkingConfig.thinkingBudget: integer (Gemini 2.5)
+//
+// For gemini-cli and antigravity providers, the path is prefixed with "request.".
+//
+// Priority: thinkingLevel is checked first (Gemini 3 format), then thinkingBudget (Gemini 2.5 format).
+// This allows newer Gemini 3 level-based configs to take precedence.
+func extractGeminiConfig(body []byte, provider string) ThinkingConfig {
+	prefix := "generationConfig.thinkingConfig"
+	if provider == "gemini-cli" || provider == "antigravity" {
+		prefix = "request.generationConfig.thinkingConfig"
+	}
+
+	// Check thinkingLevel first (Gemini 3 format takes precedence)
+	if level := gjson.GetBytes(body, prefix+".thinkingLevel"); level.Exists() {
+		value := level.String()
+		switch value {
+		case "none":
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		case "auto":
+			return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+		default:
+			return ThinkingConfig{Mode: ModeLevel, Level: ThinkingLevel(value)}
+		}
+	}
+
+	// Check thinkingBudget (Gemini 2.5 format)
+	if budget := gjson.GetBytes(body, prefix+".thinkingBudget"); budget.Exists() {
+		value := int(budget.Int())
+		switch value {
+		case 0:
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		case -1:
+			return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+		default:
+			return ThinkingConfig{Mode: ModeBudget, Budget: value}
+		}
+	}
+
+	return ThinkingConfig{}
+}
+
+// extractOpenAIConfig extracts thinking configuration from OpenAI format request body.
+//
+// OpenAI API format:
+//   - reasoning_effort: "none", "low", "medium", "high" (discrete levels)
+//
+// OpenAI uses level-based thinking configuration only, no numeric budget support.
+// The "none" value is treated specially to return ModeNone.
+func extractOpenAIConfig(body []byte) ThinkingConfig {
+	// Check reasoning_effort (OpenAI Chat Completions format)
+	if effort := gjson.GetBytes(body, "reasoning_effort"); effort.Exists() {
+		value := effort.String()
+		if value == "none" {
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		}
+		return ThinkingConfig{Mode: ModeLevel, Level: ThinkingLevel(value)}
+	}
+
+	return ThinkingConfig{}
+}
+
+// extractCodexConfig extracts thinking configuration from Codex format request body.
+//
+// Codex API format (OpenAI Responses API):
+//   - reasoning.effort: "none", "low", "medium", "high"
+//
+// This is similar to OpenAI but uses nested field "reasoning.effort" instead of "reasoning_effort".
+func extractCodexConfig(body []byte) ThinkingConfig {
+	// Check reasoning.effort (Codex / OpenAI Responses API format)
+	if effort := gjson.GetBytes(body, "reasoning.effort"); effort.Exists() {
+		value := effort.String()
+		if value == "none" {
+			return ThinkingConfig{Mode: ModeNone, Budget: 0}
+		}
+		return ThinkingConfig{Mode: ModeLevel, Level: ThinkingLevel(value)}
+	}
+
+	return ThinkingConfig{}
+}
+
+// extractIFlowConfig extracts thinking configuration from iFlow format request body.
+//
+// iFlow API format (supports multiple model families):
+//   - GLM format: chat_template_kwargs.enable_thinking (boolean)
+//   - MiniMax format: reasoning_split (boolean)
+//
+// Returns ModeBudget with Budget=1 as a sentinel value indicating "enabled".
+// The actual budget/configuration is determined by the iFlow applier based on model capabilities.
+// Budget=1 is used because iFlow models don't use numeric budgets; they only support on/off.
+func extractIFlowConfig(body []byte) ThinkingConfig {
+	// GLM format: chat_template_kwargs.enable_thinking
+	if enabled := gjson.GetBytes(body, "chat_template_kwargs.enable_thinking"); enabled.Exists() {
+		if enabled.Bool() {
+			// Budget=1 is a sentinel meaning "enabled" (iFlow doesn't use numeric budgets)
+			return ThinkingConfig{Mode: ModeBudget, Budget: 1}
+		}
+		return ThinkingConfig{Mode: ModeNone, Budget: 0}
+	}
+
+	// MiniMax format: reasoning_split
+	if split := gjson.GetBytes(body, "reasoning_split"); split.Exists() {
+		if split.Bool() {
+			// Budget=1 is a sentinel meaning "enabled" (iFlow doesn't use numeric budgets)
+			return ThinkingConfig{Mode: ModeBudget, Budget: 1}
+		}
+		return ThinkingConfig{Mode: ModeNone, Budget: 0}
+	}
+
+	return ThinkingConfig{}
+}

internal/thinking/convert.go

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

internal/thinking/errors.go

@@ -0,0 +1,82 @@
+// Package thinking provides unified thinking configuration processing logic.
+package thinking
+
+import "net/http"
+
+// ErrorCode represents the type of thinking configuration error.
+type ErrorCode string
+
+// Error codes for thinking configuration processing.
+const (
+	// ErrInvalidSuffix indicates the suffix format cannot be parsed.
+	// Example: "model(abc" (missing closing parenthesis)
+	ErrInvalidSuffix ErrorCode = "INVALID_SUFFIX"
+
+	// ErrUnknownLevel indicates the level value is not in the valid list.
+	// Example: "model(ultra)" where "ultra" is not a valid level
+	ErrUnknownLevel ErrorCode = "UNKNOWN_LEVEL"
+
+	// ErrThinkingNotSupported indicates the model does not support thinking.
+	// Example: claude-haiku-4-5 does not have thinking capability
+	ErrThinkingNotSupported ErrorCode = "THINKING_NOT_SUPPORTED"
+
+	// ErrLevelNotSupported indicates the model does not support level mode.
+	// Example: using level with a budget-only model
+	ErrLevelNotSupported ErrorCode = "LEVEL_NOT_SUPPORTED"
+
+	// ErrBudgetOutOfRange indicates the budget value is outside model range.
+	// Example: budget 64000 exceeds max 20000
+	ErrBudgetOutOfRange ErrorCode = "BUDGET_OUT_OF_RANGE"
+
+	// ErrProviderMismatch indicates the provider does not match the model.
+	// Example: applying Claude format to a Gemini model
+	ErrProviderMismatch ErrorCode = "PROVIDER_MISMATCH"
+)
+
+// ThinkingError represents an error that occurred during thinking configuration processing.
+//
+// This error type provides structured information about the error, including:
+//   - Code: A machine-readable error code for programmatic handling
+//   - Message: A human-readable description of the error
+//   - Model: The model name related to the error (optional)
+//   - Details: Additional context information (optional)
+type ThinkingError struct {
+	// Code is the machine-readable error code
+	Code ErrorCode
+	// Message is the human-readable error description.
+	// Should be lowercase, no trailing period, with context if applicable.
+	Message string
+	// Model is the model name related to this error (optional)
+	Model string
+	// Details contains additional context information (optional)
+	Details map[string]interface{}
+}
+
+// Error implements the error interface.
+// Returns the message directly without code prefix.
+// Use Code field for programmatic error handling.
+func (e *ThinkingError) Error() string {
+	return e.Message
+}
+
+// NewThinkingError creates a new ThinkingError with the given code and message.
+func NewThinkingError(code ErrorCode, message string) *ThinkingError {
+	return &ThinkingError{
+		Code:    code,
+		Message: message,
+	}
+}
+
+// NewThinkingErrorWithModel creates a new ThinkingError with model context.
+func NewThinkingErrorWithModel(code ErrorCode, message, model string) *ThinkingError {
+	return &ThinkingError{
+		Code:    code,
+		Message: message,
+		Model:   model,
+	}
+}
+
+// StatusCode implements a portable status code interface for HTTP handlers.
+func (e *ThinkingError) StatusCode() int {
+	return http.StatusBadRequest
+}

internal/thinking/provider/antigravity/apply.go

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

internal/thinking/provider/claude/apply.go

@@ -0,0 +1,166 @@
+// Package claude implements thinking configuration scaffolding for Claude models.
+//
+// Claude models use the thinking.budget_tokens format with values in the range
+// 1024-128000. Some Claude models support ZeroAllowed (sonnet-4-5, opus-4-5),
+// while older models do not.
+// See: _bmad-output/planning-artifacts/architecture.md#Epic-6
+package claude
+
+import (
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier implements thinking.ProviderApplier for Claude models.
+// This applier is stateless and holds no configuration.
+type Applier struct{}
+
+// NewApplier creates a new Claude thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("claude", NewApplier())
+}
+
+// Apply applies thinking configuration to Claude request body.
+//
+// IMPORTANT: This method expects config to be pre-validated by thinking.ValidateConfig.
+// ValidateConfig handles:
+//   - Mode conversion (Level→Budget, Auto→Budget)
+//   - Budget clamping to model range
+//   - ZeroAllowed constraint enforcement
+//
+// Apply only processes ModeBudget and ModeNone; other modes are passed through unchanged.
+//
+// Expected output format when enabled:
+//
+//	{
+//	  "thinking": {
+//	    "type": "enabled",
+//	    "budget_tokens": 16384
+//	  }
+//	}
+//
+// Expected output format when disabled:
+//
+//	{
+//	  "thinking": {
+//	    "type": "disabled"
+//	  }
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return applyCompatibleClaude(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	// Only process ModeBudget and ModeNone; other modes pass through
+	// (caller should use ValidateConfig first to normalize modes)
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	// Budget is expected to be pre-validated by ValidateConfig (clamped, ZeroAllowed enforced)
+	// Decide enabled/disabled based on budget value
+	if config.Budget == 0 {
+		result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		return result, nil
+	}
+
+	result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+	result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+
+	// Ensure max_tokens > thinking.budget_tokens (Anthropic API constraint)
+	result = a.normalizeClaudeBudget(result, config.Budget, modelInfo)
+	return result, nil
+}
+
+// normalizeClaudeBudget applies Claude-specific constraints to ensure max_tokens > budget_tokens.
+// Anthropic API requires this constraint; violating it returns a 400 error.
+func (a *Applier) normalizeClaudeBudget(body []byte, budgetTokens int, modelInfo *registry.ModelInfo) []byte {
+	if budgetTokens <= 0 {
+		return body
+	}
+
+	// Ensure the request satisfies Claude constraints:
+	//  1) Determine effective max_tokens (request overrides model default)
+	//  2) If budget_tokens >= max_tokens, reduce budget_tokens to max_tokens-1
+	//  3) If the adjusted budget falls below the model minimum, leave the request unchanged
+	//  4) If max_tokens came from model default, write it back into the request
+
+	effectiveMax, setDefaultMax := a.effectiveMaxTokens(body, modelInfo)
+	if setDefaultMax && effectiveMax > 0 {
+		body, _ = sjson.SetBytes(body, "max_tokens", effectiveMax)
+	}
+
+	// Compute the budget we would apply after enforcing budget_tokens < max_tokens.
+	adjustedBudget := budgetTokens
+	if effectiveMax > 0 && adjustedBudget >= effectiveMax {
+		adjustedBudget = effectiveMax - 1
+	}
+
+	minBudget := 0
+	if modelInfo != nil && modelInfo.Thinking != nil {
+		minBudget = modelInfo.Thinking.Min
+	}
+	if minBudget > 0 && adjustedBudget > 0 && adjustedBudget < minBudget {
+		// If enforcing the max_tokens constraint would push the budget below the model minimum,
+		// leave the request unchanged.
+		return body
+	}
+
+	if adjustedBudget != budgetTokens {
+		body, _ = sjson.SetBytes(body, "thinking.budget_tokens", adjustedBudget)
+	}
+
+	return body
+}
+
+// effectiveMaxTokens returns the max tokens to cap thinking:
+// prefer request-provided max_tokens; otherwise fall back to model default.
+// The boolean indicates whether the value came from the model default (and thus should be written back).
+func (a *Applier) effectiveMaxTokens(body []byte, modelInfo *registry.ModelInfo) (max int, fromModel bool) {
+	if maxTok := gjson.GetBytes(body, "max_tokens"); maxTok.Exists() && maxTok.Int() > 0 {
+		return int(maxTok.Int()), false
+	}
+	if modelInfo != nil && modelInfo.MaxCompletionTokens > 0 {
+		return modelInfo.MaxCompletionTokens, true
+	}
+	return 0, false
+}
+
+func applyCompatibleClaude(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	switch config.Mode {
+	case thinking.ModeNone:
+		result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		return result, nil
+	case thinking.ModeAuto:
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		return result, nil
+	default:
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+		return result, nil
+	}
+}

internal/thinking/provider/codex/apply.go

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

internal/thinking/provider/gemini/apply.go

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

internal/thinking/provider/geminicli/apply.go

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

internal/thinking/provider/iflow/apply.go

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

internal/thinking/provider/openai/apply.go

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

internal/thinking/strip.go

@@ -0,0 +1,58 @@
+// Package thinking provides unified thinking configuration processing.
+package thinking
+
+import (
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// StripThinkingConfig removes thinking configuration fields from request body.
+//
+// This function is used when a model doesn't support thinking but the request
+// contains thinking configuration. The configuration is silently removed to
+// prevent upstream API errors.
+//
+// Parameters:
+//   - body: Original request body JSON
+//   - provider: Provider name (determines which fields to strip)
+//
+// Returns:
+//   - Modified request body JSON with thinking configuration removed
+//   - Original body is returned unchanged if:
+//   - body is empty or invalid JSON
+//   - provider is unknown
+//   - no thinking configuration found
+func StripThinkingConfig(body []byte, provider string) []byte {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		return body
+	}
+
+	var paths []string
+	switch provider {
+	case "claude":
+		paths = []string{"thinking"}
+	case "gemini":
+		paths = []string{"generationConfig.thinkingConfig"}
+	case "gemini-cli", "antigravity":
+		paths = []string{"request.generationConfig.thinkingConfig"}
+	case "openai":
+		paths = []string{"reasoning_effort"}
+	case "codex":
+		paths = []string{"reasoning.effort"}
+	case "iflow":
+		paths = []string{
+			"chat_template_kwargs.enable_thinking",
+			"chat_template_kwargs.clear_thinking",
+			"reasoning_split",
+			"reasoning_effort",
+		}
+	default:
+		return body
+	}
+
+	result := body
+	for _, path := range paths {
+		result, _ = sjson.DeleteBytes(result, path)
+	}
+	return result
+}

internal/thinking/suffix.go

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

internal/thinking/text.go

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

internal/thinking/types.go

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

internal/thinking/validate.go

@@ -0,0 +1,378 @@
+// Package thinking provides unified thinking configuration processing logic.
+package thinking
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	log "github.com/sirupsen/logrus"
+)
+
+// ValidateConfig validates a thinking configuration against model capabilities.
+//
+// This function performs comprehensive validation:
+//   - Checks if the model supports thinking
+//   - Auto-converts between Budget and Level formats based on model capability
+//   - Validates that requested level is in the model's supported levels list
+//   - Clamps budget values to model's allowed range
+//   - When converting Budget -> Level for level-only models, clamps the derived standard level to the nearest supported level
+//     (special values none/auto are preserved)
+//   - When config comes from a model suffix, strict budget validation is disabled (we clamp instead of error)
+//
+// Parameters:
+//   - config: The thinking configuration to validate
+//   - support: Model's ThinkingSupport properties (nil means no thinking support)
+//   - fromFormat: Source provider format (used to determine strict validation rules)
+//   - toFormat: Target provider format
+//   - fromSuffix: Whether config was sourced from model suffix
+//
+// Returns:
+//   - Normalized ThinkingConfig with clamped values
+//   - ThinkingError if validation fails (ErrThinkingNotSupported, ErrLevelNotSupported, etc.)
+//
+// Auto-conversion behavior:
+//   - Budget-only model + Level config → Level converted to Budget
+//   - Level-only model + Budget config → Budget converted to Level
+//   - Hybrid model → preserve original format
+func ValidateConfig(config ThinkingConfig, modelInfo *registry.ModelInfo, fromFormat, toFormat string, fromSuffix bool) (*ThinkingConfig, error) {
+	fromFormat, toFormat = strings.ToLower(strings.TrimSpace(fromFormat)), strings.ToLower(strings.TrimSpace(toFormat))
+	model := "unknown"
+	support := (*registry.ThinkingSupport)(nil)
+	if modelInfo != nil {
+		if modelInfo.ID != "" {
+			model = modelInfo.ID
+		}
+		support = modelInfo.Thinking
+	}
+
+	if support == nil {
+		if config.Mode != ModeNone {
+			return nil, NewThinkingErrorWithModel(ErrThinkingNotSupported, "thinking not supported for this model", model)
+		}
+		return &config, nil
+	}
+
+	allowClampUnsupported := isBudgetBasedProvider(fromFormat) && isLevelBasedProvider(toFormat)
+	strictBudget := !fromSuffix && fromFormat != "" && isSameProviderFamily(fromFormat, toFormat)
+	budgetDerivedFromLevel := false
+
+	capability := detectModelCapability(modelInfo)
+	switch capability {
+	case CapabilityBudgetOnly:
+		if config.Mode == ModeLevel {
+			if config.Level == LevelAuto {
+				break
+			}
+			budget, ok := ConvertLevelToBudget(string(config.Level))
+			if !ok {
+				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("unknown level: %s", config.Level))
+			}
+			config.Mode = ModeBudget
+			config.Budget = budget
+			config.Level = ""
+			budgetDerivedFromLevel = true
+		}
+	case CapabilityLevelOnly:
+		if config.Mode == ModeBudget {
+			level, ok := ConvertBudgetToLevel(config.Budget)
+			if !ok {
+				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("budget %d cannot be converted to a valid level", config.Budget))
+			}
+			// When converting Budget -> Level for level-only models, clamp the derived standard level
+			// to the nearest supported level. Special values (none/auto) are preserved.
+			config.Mode = ModeLevel
+			config.Level = clampLevel(ThinkingLevel(level), modelInfo, toFormat)
+			config.Budget = 0
+		}
+	case CapabilityHybrid:
+	}
+
+	if config.Mode == ModeLevel && config.Level == LevelNone {
+		config.Mode = ModeNone
+		config.Budget = 0
+		config.Level = ""
+	}
+	if config.Mode == ModeLevel && config.Level == LevelAuto {
+		config.Mode = ModeAuto
+		config.Budget = -1
+		config.Level = ""
+	}
+	if config.Mode == ModeBudget && config.Budget == 0 {
+		config.Mode = ModeNone
+		config.Level = ""
+	}
+
+	if len(support.Levels) > 0 && config.Mode == ModeLevel {
+		if !isLevelSupported(string(config.Level), support.Levels) {
+			if allowClampUnsupported {
+				config.Level = clampLevel(config.Level, modelInfo, toFormat)
+			}
+			if !isLevelSupported(string(config.Level), support.Levels) {
+				// User explicitly specified an unsupported level - return error
+				// (budget-derived levels may be clamped based on source format)
+				validLevels := normalizeLevels(support.Levels)
+				message := fmt.Sprintf("level %q not supported, valid levels: %s", strings.ToLower(string(config.Level)), strings.Join(validLevels, ", "))
+				return nil, NewThinkingError(ErrLevelNotSupported, message)
+			}
+		}
+	}
+
+	if strictBudget && config.Mode == ModeBudget && !budgetDerivedFromLevel {
+		min, max := support.Min, support.Max
+		if min != 0 || max != 0 {
+			if config.Budget < min || config.Budget > max || (config.Budget == 0 && !support.ZeroAllowed) {
+				message := fmt.Sprintf("budget %d out of range [%d,%d]", config.Budget, min, max)
+				return nil, NewThinkingError(ErrBudgetOutOfRange, message)
+			}
+		}
+	}
+
+	// Convert ModeAuto to mid-range if dynamic not allowed
+	if config.Mode == ModeAuto && !support.DynamicAllowed {
+		config = convertAutoToMidRange(config, support, toFormat, model)
+	}
+
+	if config.Mode == ModeNone && toFormat == "claude" {
+		// Claude supports explicit disable via thinking.type="disabled".
+		// Keep Budget=0 so applier can omit budget_tokens.
+		config.Budget = 0
+		config.Level = ""
+	} else {
+		switch config.Mode {
+		case ModeBudget, ModeAuto, ModeNone:
+			config.Budget = clampBudget(config.Budget, modelInfo, toFormat)
+		}
+
+		// ModeNone with clamped Budget > 0: set Level to lowest for Level-only/Hybrid models
+		// This ensures Apply layer doesn't need to access support.Levels
+		if config.Mode == ModeNone && config.Budget > 0 && len(support.Levels) > 0 {
+			config.Level = ThinkingLevel(support.Levels[0])
+		}
+	}
+
+	return &config, nil
+}
+
+// convertAutoToMidRange converts ModeAuto to a mid-range value when dynamic is not allowed.
+//
+// This function handles the case where a model does not support dynamic/auto thinking.
+// The auto mode is silently converted to a fixed value based on model capability:
+//   - Level-only models: convert to ModeLevel with LevelMedium
+//   - Budget models: convert to ModeBudget with mid = (Min + Max) / 2
+//
+// Logging:
+//   - Debug level when conversion occurs
+//   - Fields: original_mode, clamped_to, reason
+func convertAutoToMidRange(config ThinkingConfig, support *registry.ThinkingSupport, provider, model string) ThinkingConfig {
+	// For level-only models (has Levels but no Min/Max range), use ModeLevel with medium
+	if len(support.Levels) > 0 && support.Min == 0 && support.Max == 0 {
+		config.Mode = ModeLevel
+		config.Level = LevelMedium
+		config.Budget = 0
+		log.WithFields(log.Fields{
+			"provider":      provider,
+			"model":         model,
+			"original_mode": "auto",
+			"clamped_to":    string(LevelMedium),
+		}).Debug("thinking: mode converted, dynamic not allowed, using medium level |")
+		return config
+	}
+
+	// For budget models, use mid-range budget
+	mid := (support.Min + support.Max) / 2
+	if mid <= 0 && support.ZeroAllowed {
+		config.Mode = ModeNone
+		config.Budget = 0
+	} else if mid <= 0 {
+		config.Mode = ModeBudget
+		config.Budget = support.Min
+	} else {
+		config.Mode = ModeBudget
+		config.Budget = mid
+	}
+	log.WithFields(log.Fields{
+		"provider":      provider,
+		"model":         model,
+		"original_mode": "auto",
+		"clamped_to":    config.Budget,
+	}).Debug("thinking: mode converted, dynamic not allowed |")
+	return config
+}
+
+// standardLevelOrder defines the canonical ordering of thinking levels from lowest to highest.
+var standardLevelOrder = []ThinkingLevel{LevelMinimal, LevelLow, LevelMedium, LevelHigh, LevelXHigh}
+
+// clampLevel clamps the given level to the nearest supported level.
+// On tie, prefers the lower level.
+func clampLevel(level ThinkingLevel, modelInfo *registry.ModelInfo, provider string) ThinkingLevel {
+	model := "unknown"
+	var supported []string
+	if modelInfo != nil {
+		if modelInfo.ID != "" {
+			model = modelInfo.ID
+		}
+		if modelInfo.Thinking != nil {
+			supported = modelInfo.Thinking.Levels
+		}
+	}
+
+	if len(supported) == 0 || isLevelSupported(string(level), supported) {
+		return level
+	}
+
+	pos := levelIndex(string(level))
+	if pos == -1 {
+		return level
+	}
+	bestIdx, bestDist := -1, len(standardLevelOrder)+1
+
+	for _, s := range supported {
+		if idx := levelIndex(strings.TrimSpace(s)); idx != -1 {
+			if dist := abs(pos - idx); dist < bestDist || (dist == bestDist && idx < bestIdx) {
+				bestIdx, bestDist = idx, dist
+			}
+		}
+	}
+
+	if bestIdx >= 0 {
+		clamped := standardLevelOrder[bestIdx]
+		log.WithFields(log.Fields{
+			"provider":       provider,
+			"model":          model,
+			"original_value": string(level),
+			"clamped_to":     string(clamped),
+		}).Debug("thinking: level clamped |")
+		return clamped
+	}
+	return level
+}
+
+// clampBudget clamps a budget value to the model's supported range.
+func clampBudget(value int, modelInfo *registry.ModelInfo, provider string) int {
+	model := "unknown"
+	support := (*registry.ThinkingSupport)(nil)
+	if modelInfo != nil {
+		if modelInfo.ID != "" {
+			model = modelInfo.ID
+		}
+		support = modelInfo.Thinking
+	}
+	if support == nil {
+		return value
+	}
+
+	// Auto value (-1) passes through without clamping.
+	if value == -1 {
+		return value
+	}
+
+	min, max := support.Min, support.Max
+	if value == 0 && !support.ZeroAllowed {
+		log.WithFields(log.Fields{
+			"provider":       provider,
+			"model":          model,
+			"original_value": value,
+			"clamped_to":     min,
+			"min":            min,
+			"max":            max,
+		}).Warn("thinking: budget zero not allowed |")
+		return min
+	}
+
+	// Some models are level-only and do not define numeric budget ranges.
+	if min == 0 && max == 0 {
+		return value
+	}
+
+	if value < min {
+		if value == 0 && support.ZeroAllowed {
+			return 0
+		}
+		logClamp(provider, model, value, min, min, max)
+		return min
+	}
+	if value > max {
+		logClamp(provider, model, value, max, min, max)
+		return max
+	}
+	return value
+}
+
+func isLevelSupported(level string, supported []string) bool {
+	for _, s := range supported {
+		if strings.EqualFold(level, strings.TrimSpace(s)) {
+			return true
+		}
+	}
+	return false
+}
+
+func levelIndex(level string) int {
+	for i, l := range standardLevelOrder {
+		if strings.EqualFold(level, string(l)) {
+			return i
+		}
+	}
+	return -1
+}
+
+func normalizeLevels(levels []string) []string {
+	out := make([]string, len(levels))
+	for i, l := range levels {
+		out[i] = strings.ToLower(strings.TrimSpace(l))
+	}
+	return out
+}
+
+func isBudgetBasedProvider(provider string) bool {
+	switch provider {
+	case "gemini", "gemini-cli", "antigravity", "claude":
+		return true
+	default:
+		return false
+	}
+}
+
+func isLevelBasedProvider(provider string) bool {
+	switch provider {
+	case "openai", "openai-response", "codex":
+		return true
+	default:
+		return false
+	}
+}
+
+func isGeminiFamily(provider string) bool {
+	switch provider {
+	case "gemini", "gemini-cli", "antigravity":
+		return true
+	default:
+		return false
+	}
+}
+
+func isSameProviderFamily(from, to string) bool {
+	if from == to {
+		return true
+	}
+	return isGeminiFamily(from) && isGeminiFamily(to)
+}
+
+func abs(x int) int {
+	if x < 0 {
+		return -x
+	}
+	return x
+}
+
+func logClamp(provider, model string, original, clampedTo, min, max int) {
+	log.WithFields(log.Fields{
+		"provider":       provider,
+		"model":          model,
+		"original_value": original,
+		"min":            min,
+		"max":            max,
+		"clamped_to":     clampedTo,
+	}).Debug("thinking: budget clamped |")
+}

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -12,9 +12,9 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/cache"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
-	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -123,7 +123,7 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 					contentTypeResult := contentResult.Get("type")
 					if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "thinking" {
 						// Use GetThinkingText to handle wrapped thinking objects
-						thinkingText := util.GetThinkingText(contentResult)
+						thinkingText := thinking.GetThinkingText(contentResult)
 						signatureResult := contentResult.Get("signature")
 						clientSignature := ""
 						if signatureResult.Exists() && signatureResult.String() != "" {
@@ -136,14 +136,14 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 						if sessionID != "" && thinkingText != "" {
 							if cachedSig := cache.GetCachedSignature(sessionID, thinkingText); cachedSig != "" {
 								signature = cachedSig
-								log.Debugf("Using cached signature for thinking block")
+								// 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")
+							// log.Debugf("Using client-provided signature for thinking block")
 						}
 
 						// Store for subsequent tool_use in the same message
@@ -158,8 +158,7 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 						// 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)")
+							// log.Debugf("Dropping unsigned thinking block (no valid signature)")
 							continue
 						}
 
@@ -183,7 +182,6 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 					} 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()
 						argsResult := contentResult.Get("input")
@@ -388,12 +386,12 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 	}
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
-	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() && util.ModelSupportsThinking(modelName) {
+	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
 		if t.Get("type").String() == "enabled" {
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		}
 	}

internal/translator/antigravity/claude/antigravity_claude_request_test.go

@@ -343,8 +343,8 @@ func TestConvertClaudeRequestToAntigravity_ThinkingConfig(t *testing.T) {
 		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")
+		if !thinkingConfig.Get("includeThoughts").Bool() {
+			t.Error("includeThoughts should be true")
 		}
 	} else {
 		t.Log("thinkingConfig not present - model may not be registered in test registry")

internal/translator/antigravity/claude/antigravity_claude_response.go

@@ -136,11 +136,11 @@ 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")
+						// 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())
+							// log.Debugf("Cached signature for thinking block (sessionID=%s, textLen=%d)", params.SessionID, params.CurrentThinkingText.Len())
 							params.CurrentThinkingText.Reset()
 						}
 

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

@@ -35,66 +35,19 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 	// Model
 	out, _ = sjson.SetBytes(out, "model", modelName)
 
-	// Reasoning effort -> thinkingBudget/include_thoughts
-	// Note: OpenAI official fields take precedence over extra_body.google.thinking_config
+	// Apply thinking configuration: convert OpenAI reasoning_effort to Gemini CLI thinkingConfig.
+	// Inline translation-only mapping; capability checks happen later in ApplyThinking.
 	re := gjson.GetBytes(rawJSON, "reasoning_effort")
-	hasOfficialThinking := re.Exists()
-	if hasOfficialThinking && util.ModelSupportsThinking(modelName) {
+	if re.Exists() {
 		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)
-	// Only apply for models that use numeric budgets, not discrete levels.
-	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
-		if tc := gjson.GetBytes(rawJSON, "extra_body.google.thinking_config"); tc.Exists() && tc.IsObject() {
-			var setBudget bool
-			var budget int
-
-			if v := tc.Get("thinkingBudget"); v.Exists() {
-				budget = int(v.Int())
-				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-				setBudget = true
-			} else if v := tc.Get("thinking_budget"); v.Exists() {
-				budget = int(v.Int())
-				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-				setBudget = true
-			}
-
-			if v := tc.Get("includeThoughts"); v.Exists() {
-				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", v.Bool())
-			} else if v := tc.Get("include_thoughts"); v.Exists() {
-				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", v.Bool())
-			} else if setBudget && budget != 0 {
-				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-			}
-		}
-	}
-
-	// Claude/Anthropic API format: thinking.type == "enabled" with budget_tokens
-	// This allows Claude Code and other Claude API clients to pass thinking configuration
-	if !gjson.GetBytes(out, "request.generationConfig.thinkingConfig").Exists() && util.ModelSupportsThinking(modelName) {
-		if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
-			if t.Get("type").String() == "enabled" {
-				if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
-					budget := int(b.Int())
-					out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-					out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-				}
+		if effort != "" {
+			thinkingPath := "request.generationConfig.thinkingConfig"
+			if effort == "auto" {
+				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingBudget", -1)
+				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", true)
+			} else {
+				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingLevel", effort)
+				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", effort != "none")
 			}
 		}
 	}
@@ -179,29 +132,33 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 			}
 		}
 
+		systemPartIndex := 0
 		for i := 0; i < len(arr); i++ {
 			m := arr[i]
 			role := m.Get("role").String()
 			content := m.Get("content")
 
-			if role == "system" && len(arr) > 1 {
+			if (role == "system" || role == "developer") && len(arr) > 1 {
 				// system -> request.systemInstruction as a user message style
 				if content.Type == gjson.String {
 					out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
-					out, _ = sjson.SetBytes(out, "request.systemInstruction.parts.0.text", content.String())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), content.String())
+					systemPartIndex++
 				} 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())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), content.Get("text").String())
+					systemPartIndex++
 				} 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())
+							out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), contents[j].Get("text").String())
+							systemPartIndex++
 						}
 					}
 				}
-			} else if role == "user" || (role == "system" && len(arr) == 1) {
+			} else if role == "user" || ((role == "system" || role == "developer") && len(arr) == 1) {
 				// Build single user content node to avoid splitting into multiple contents
 				node := []byte(`{"role":"user","parts":[]}`)
 				if content.Type == gjson.String {
@@ -212,7 +169,10 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 					for _, item := range items {
 						switch item.Get("type").String() {
 						case "text":
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", item.Get("text").String())
+							text := item.Get("text").String()
+							if text != "" {
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", text)
+							}
 							p++
 						case "image_url":
 							imageURL := item.Get("image_url.url").String()
@@ -223,6 +183,7 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 									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)
+									node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiCLIFunctionThoughtSignature)
 									p++
 								}
 							}
@@ -255,6 +216,10 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 					for _, item := range content.Array() {
 						switch item.Get("type").String() {
 						case "text":
+							text := item.Get("text").String()
+							if text != "" {
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", text)
+							}
 							p++
 						case "image_url":
 							// If the assistant returned an inline data URL, preserve it for history fidelity.
@@ -266,6 +231,7 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 									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)
+									node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiCLIFunctionThoughtSignature)
 									p++
 								}
 							}

internal/translator/claude/gemini/claude_gemini_request.go

@@ -15,6 +15,7 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -114,15 +115,40 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 			}
 		}
 		// Include thoughts configuration for reasoning process visibility
-		// Only apply for models that support thinking and use numeric budgets, not discrete levels.
-		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
-			// Check for thinkingBudget first - if present, enable thinking with budget
-			if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() && thinkingBudget.Int() > 0 {
+		// Translator only does format conversion, ApplyThinking handles model capability validation.
+		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
+			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+				level := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
+				switch level {
+				case "":
+				case "none":
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				case "auto":
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				default:
+					if budget, ok := thinking.ConvertLevelToBudget(level); ok {
+						out, _ = sjson.Set(out, "thinking.type", "enabled")
+						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					}
+				}
+			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+				budget := int(thinkingBudget.Int())
+				switch budget {
+				case 0:
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				case -1:
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				default:
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+				}
+			} else if includeThoughts := thinkingConfig.Get("includeThoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
 				out, _ = sjson.Set(out, "thinking.type", "enabled")
-				normalizedBudget := util.NormalizeThinkingBudget(modelName, int(thinkingBudget.Int()))
-				out, _ = sjson.Set(out, "thinking.budget_tokens", normalizedBudget)
 			} else if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
-				// Fallback to include_thoughts if no budget specified
 				out, _ = sjson.Set(out, "thinking.type", "enabled")
 			}
 		}

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

@@ -15,7 +15,7 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -65,10 +65,11 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 
 	root := gjson.ParseBytes(rawJSON)
 
-	if v := root.Get("reasoning_effort"); v.Exists() && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
+	// Convert OpenAI reasoning_effort to Claude thinking config.
+	if v := root.Get("reasoning_effort"); v.Exists() {
 		effort := strings.ToLower(strings.TrimSpace(v.String()))
 		if effort != "" {
-			budget, ok := util.ThinkingEffortToBudget(modelName, effort)
+			budget, ok := thinking.ConvertLevelToBudget(effort)
 			if ok {
 				switch budget {
 				case 0:
@@ -137,17 +138,35 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 
 	// Process messages and transform them to Claude Code format
 	if messages := root.Get("messages"); messages.Exists() && messages.IsArray() {
+		messageIndex := 0
+		systemMessageIndex := -1
 		messages.ForEach(func(_, message gjson.Result) bool {
 			role := message.Get("role").String()
 			contentResult := message.Get("content")
 
 			switch role {
-			case "system", "user", "assistant":
-				// Create Claude Code message with appropriate role mapping
-				if role == "system" {
-					role = "user"
+			case "system":
+				if systemMessageIndex == -1 {
+					systemMsg := `{"role":"user","content":[]}`
+					out, _ = sjson.SetRaw(out, "messages.-1", systemMsg)
+					systemMessageIndex = messageIndex
+					messageIndex++
 				}
-
+				if contentResult.Exists() && contentResult.Type == gjson.String && contentResult.String() != "" {
+					textPart := `{"type":"text","text":""}`
+					textPart, _ = sjson.Set(textPart, "text", contentResult.String())
+					out, _ = sjson.SetRaw(out, fmt.Sprintf("messages.%d.content.-1", systemMessageIndex), textPart)
+				} else if contentResult.Exists() && contentResult.IsArray() {
+					contentResult.ForEach(func(_, part gjson.Result) bool {
+						if part.Get("type").String() == "text" {
+							textPart := `{"type":"text","text":""}`
+							textPart, _ = sjson.Set(textPart, "text", part.Get("text").String())
+							out, _ = sjson.SetRaw(out, fmt.Sprintf("messages.%d.content.-1", systemMessageIndex), textPart)
+						}
+						return true
+					})
+				}
+			case "user", "assistant":
 				msg := `{"role":"","content":[]}`
 				msg, _ = sjson.Set(msg, "role", role)
 
@@ -226,6 +245,7 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 				}
 
 				out, _ = sjson.SetRaw(out, "messages.-1", msg)
+				messageIndex++
 
 			case "tool":
 				// Handle tool result messages conversion
@@ -236,6 +256,7 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 				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)
+				messageIndex++
 			}
 			return true
 		})

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

@@ -10,7 +10,7 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -53,10 +53,11 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 
 	root := gjson.ParseBytes(rawJSON)
 
-	if v := root.Get("reasoning.effort"); v.Exists() && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
+	// Convert OpenAI Responses reasoning.effort to Claude thinking config.
+	if v := root.Get("reasoning.effort"); v.Exists() {
 		effort := strings.ToLower(strings.TrimSpace(v.String()))
 		if effort != "" {
-			budget, ok := util.ThinkingEffortToBudget(modelName, effort)
+			budget, ok := thinking.ConvertLevelToBudget(effort)
 			if ok {
 				switch budget {
 				case 0:

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

@@ -40,6 +40,16 @@ type claudeToResponsesState struct {
 
 var dataTag = []byte("data:")
 
+func pickRequestJSON(originalRequestRawJSON, requestRawJSON []byte) []byte {
+	if len(originalRequestRawJSON) > 0 && gjson.ValidBytes(originalRequestRawJSON) {
+		return originalRequestRawJSON
+	}
+	if len(requestRawJSON) > 0 && gjson.ValidBytes(requestRawJSON) {
+		return requestRawJSON
+	}
+	return nil
+}
+
 func emitEvent(event string, payload string) string {
 	return fmt.Sprintf("event: %s\ndata: %s", event, payload)
 }
@@ -241,6 +251,7 @@ func ConvertClaudeResponseToOpenAIResponses(ctx context.Context, modelName strin
 			itemDone, _ = sjson.Set(itemDone, "item.id", fmt.Sprintf("fc_%s", st.CurrentFCID))
 			itemDone, _ = sjson.Set(itemDone, "item.arguments", args)
 			itemDone, _ = sjson.Set(itemDone, "item.call_id", st.CurrentFCID)
+			itemDone, _ = sjson.Set(itemDone, "item.name", st.FuncNames[idx])
 			out = append(out, emitEvent("response.output_item.done", itemDone))
 			st.InFuncBlock = false
 		} else if st.ReasoningActive {
@@ -279,8 +290,9 @@ func ConvertClaudeResponseToOpenAIResponses(ctx context.Context, modelName strin
 		completed, _ = sjson.Set(completed, "response.created_at", st.CreatedAt)
 		// Inject original request fields into response as per docs/response.completed.json
 
-		if requestRawJSON != nil {
-			req := gjson.ParseBytes(requestRawJSON)
+		reqBytes := pickRequestJSON(originalRequestRawJSON, requestRawJSON)
+		if len(reqBytes) > 0 {
+			req := gjson.ParseBytes(reqBytes)
 			if v := req.Get("instructions"); v.Exists() {
 				completed, _ = sjson.Set(completed, "response.instructions", v.String())
 			}
@@ -549,8 +561,9 @@ func ConvertClaudeResponseToOpenAIResponsesNonStream(_ context.Context, _ string
 	out, _ = sjson.Set(out, "created_at", createdAt)
 
 	// Inject request echo fields as top-level (similar to streaming variant)
-	if requestRawJSON != nil {
-		req := gjson.ParseBytes(requestRawJSON)
+	reqBytes := pickRequestJSON(originalRequestRawJSON, requestRawJSON)
+	if len(reqBytes) > 0 {
+		req := gjson.ParseBytes(reqBytes)
 		if v := req.Get("instructions"); v.Exists() {
 			out, _ = sjson.Set(out, "instructions", v.String())
 		}

internal/translator/codex/claude/codex_claude_request.go

@@ -12,7 +12,7 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -37,10 +37,11 @@ import (
 //   - []byte: The transformed request data in internal client format
 func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool) []byte {
 	rawJSON := bytes.Clone(inputRawJSON)
+	userAgent := misc.ExtractCodexUserAgent(rawJSON)
 
 	template := `{"model":"","instructions":"","input":[]}`
 
-	_, instructions := misc.CodexInstructionsForModel(modelName, "")
+	_, instructions := misc.CodexInstructionsForModel(modelName, "", userAgent)
 	template, _ = sjson.Set(template, "instructions", instructions)
 
 	rootResult := gjson.ParseBytes(rawJSON)
@@ -50,7 +51,7 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	systemsResult := rootResult.Get("system")
 	if systemsResult.IsArray() {
 		systemResults := systemsResult.Array()
-		message := `{"type":"message","role":"user","content":[]}`
+		message := `{"type":"message","role":"developer","content":[]}`
 		for i := 0; i < len(systemResults); i++ {
 			systemResult := systemResults[i]
 			systemTypeResult := systemResult.Get("type")
@@ -216,21 +217,19 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	// Add additional configuration parameters for the Codex API.
 	template, _ = sjson.Set(template, "parallel_tool_calls", true)
 
-	// Convert thinking.budget_tokens to reasoning.effort for level-based models
-	reasoningEffort := "medium" // default
-	if thinking := rootResult.Get("thinking"); thinking.Exists() && thinking.IsObject() {
-		switch thinking.Get("type").String() {
+	// Convert thinking.budget_tokens to reasoning.effort.
+	reasoningEffort := "medium"
+	if thinkingConfig := rootResult.Get("thinking"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
+		switch thinkingConfig.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.ThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
-						reasoningEffort = effort
-					}
+			if budgetTokens := thinkingConfig.Get("budget_tokens"); budgetTokens.Exists() {
+				budget := int(budgetTokens.Int())
+				if effort, ok := thinking.ConvertBudgetToLevel(budget); ok && effort != "" {
+					reasoningEffort = effort
 				}
 			}
 		case "disabled":
-			if effort, ok := util.ThinkingBudgetToEffort(modelName, 0); ok && effort != "" {
+			if effort, ok := thinking.ConvertBudgetToLevel(0); ok && effort != "" {
 				reasoningEffort = effort
 			}
 		}
@@ -242,21 +241,23 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	template, _ = sjson.Set(template, "include", []string{"reasoning.encrypted_content"})
 
 	// Add a first message to ignore system instructions and ensure proper execution.
-	inputResult := gjson.Get(template, "input")
-	if inputResult.Exists() && inputResult.IsArray() {
-		inputResults := inputResult.Array()
-		newInput := "[]"
-		for i := 0; i < len(inputResults); i++ {
-			if i == 0 {
-				firstText := inputResults[i].Get("content.0.text")
-				firstInstructions := "EXECUTE ACCORDING TO THE FOLLOWING INSTRUCTIONS!!!"
-				if firstText.Exists() && firstText.String() != firstInstructions {
-					newInput, _ = sjson.SetRaw(newInput, "-1", `{"type":"message","role":"user","content":[{"type":"input_text","text":"EXECUTE ACCORDING TO THE FOLLOWING INSTRUCTIONS!!!"}]}`)
+	if misc.GetCodexInstructionsEnabled() {
+		inputResult := gjson.Get(template, "input")
+		if inputResult.Exists() && inputResult.IsArray() {
+			inputResults := inputResult.Array()
+			newInput := "[]"
+			for i := 0; i < len(inputResults); i++ {
+				if i == 0 {
+					firstText := inputResults[i].Get("content.0.text")
+					firstInstructions := "EXECUTE ACCORDING TO THE FOLLOWING INSTRUCTIONS!!!"
+					if firstText.Exists() && firstText.String() != firstInstructions {
+						newInput, _ = sjson.SetRaw(newInput, "-1", `{"type":"message","role":"user","content":[{"type":"input_text","text":"EXECUTE ACCORDING TO THE FOLLOWING INSTRUCTIONS!!!"}]}`)
+					}
 				}
+				newInput, _ = sjson.SetRaw(newInput, "-1", inputResults[i].Raw)
 			}
-			newInput, _ = sjson.SetRaw(newInput, "-1", inputResults[i].Raw)
+			template, _ = sjson.SetRaw(template, "input", newInput)
 		}
-		template, _ = sjson.SetRaw(template, "input", newInput)
 	}
 
 	return []byte(template)

internal/translator/codex/claude/codex_claude_response.go

@@ -20,6 +20,12 @@ var (
 	dataTag = []byte("data:")
 )
 
+// ConvertCodexResponseToClaudeParams holds parameters for response conversion.
+type ConvertCodexResponseToClaudeParams struct {
+	HasToolCall bool
+	BlockIndex  int
+}
+
 // ConvertCodexResponseToClaude performs sophisticated streaming response format conversion.
 // This function implements a complex state machine that translates Codex API responses
 // into Claude Code-compatible Server-Sent Events (SSE) format. It manages different response types
@@ -38,8 +44,10 @@ var (
 //   - []string: A slice of strings, each containing a Claude Code-compatible JSON response
 func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, param *any) []string {
 	if *param == nil {
-		hasToolCall := false
-		*param = &hasToolCall
+		*param = &ConvertCodexResponseToClaudeParams{
+			HasToolCall: false,
+			BlockIndex:  0,
+		}
 	}
 
 	// log.Debugf("rawJSON: %s", string(rawJSON))
@@ -62,52 +70,59 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 		output += fmt.Sprintf("data: %s\n\n", template)
 	} else if typeStr == "response.reasoning_summary_part.added" {
 		template = `{"type":"content_block_start","index":0,"content_block":{"type":"thinking","thinking":""}}`
-		template, _ = sjson.Set(template, "index", rootResult.Get("output_index").Int())
+		template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
 
 		output = "event: content_block_start\n"
 		output += fmt.Sprintf("data: %s\n\n", template)
 	} else if typeStr == "response.reasoning_summary_text.delta" {
 		template = `{"type":"content_block_delta","index":0,"delta":{"type":"thinking_delta","thinking":""}}`
-		template, _ = sjson.Set(template, "index", rootResult.Get("output_index").Int())
+		template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
 		template, _ = sjson.Set(template, "delta.thinking", rootResult.Get("delta").String())
 
 		output = "event: content_block_delta\n"
 		output += fmt.Sprintf("data: %s\n\n", template)
 	} else if typeStr == "response.reasoning_summary_part.done" {
 		template = `{"type":"content_block_stop","index":0}`
-		template, _ = sjson.Set(template, "index", rootResult.Get("output_index").Int())
+		template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
+		(*param).(*ConvertCodexResponseToClaudeParams).BlockIndex++
 
 		output = "event: content_block_stop\n"
 		output += fmt.Sprintf("data: %s\n\n", template)
+
 	} else if typeStr == "response.content_part.added" {
 		template = `{"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}`
-		template, _ = sjson.Set(template, "index", rootResult.Get("output_index").Int())
+		template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
 
 		output = "event: content_block_start\n"
 		output += fmt.Sprintf("data: %s\n\n", template)
 	} else if typeStr == "response.output_text.delta" {
 		template = `{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":""}}`
-		template, _ = sjson.Set(template, "index", rootResult.Get("output_index").Int())
+		template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
 		template, _ = sjson.Set(template, "delta.text", rootResult.Get("delta").String())
 
 		output = "event: content_block_delta\n"
 		output += fmt.Sprintf("data: %s\n\n", template)
 	} else if typeStr == "response.content_part.done" {
 		template = `{"type":"content_block_stop","index":0}`
-		template, _ = sjson.Set(template, "index", rootResult.Get("output_index").Int())
+		template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
+		(*param).(*ConvertCodexResponseToClaudeParams).BlockIndex++
 
 		output = "event: content_block_stop\n"
 		output += fmt.Sprintf("data: %s\n\n", template)
 	} else if typeStr == "response.completed" {
 		template = `{"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
-		p := (*param).(*bool)
-		if *p {
+		p := (*param).(*ConvertCodexResponseToClaudeParams).HasToolCall
+		if p {
 			template, _ = sjson.Set(template, "delta.stop_reason", "tool_use")
 		} else {
 			template, _ = sjson.Set(template, "delta.stop_reason", "end_turn")
 		}
-		template, _ = sjson.Set(template, "usage.input_tokens", rootResult.Get("response.usage.input_tokens").Int())
-		template, _ = sjson.Set(template, "usage.output_tokens", rootResult.Get("response.usage.output_tokens").Int())
+		inputTokens, outputTokens, cachedTokens := extractResponsesUsage(rootResult.Get("response.usage"))
+		template, _ = sjson.Set(template, "usage.input_tokens", inputTokens)
+		template, _ = sjson.Set(template, "usage.output_tokens", outputTokens)
+		if cachedTokens > 0 {
+			template, _ = sjson.Set(template, "usage.cache_read_input_tokens", cachedTokens)
+		}
 
 		output = "event: message_delta\n"
 		output += fmt.Sprintf("data: %s\n\n", template)
@@ -118,10 +133,9 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 		itemResult := rootResult.Get("item")
 		itemType := itemResult.Get("type").String()
 		if itemType == "function_call" {
-			p := true
-			*param = &p
+			(*param).(*ConvertCodexResponseToClaudeParams).HasToolCall = true
 			template = `{"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`
-			template, _ = sjson.Set(template, "index", rootResult.Get("output_index").Int())
+			template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
 			template, _ = sjson.Set(template, "content_block.id", itemResult.Get("call_id").String())
 			{
 				// Restore original tool name if shortened
@@ -137,7 +151,7 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 			output += fmt.Sprintf("data: %s\n\n", template)
 
 			template = `{"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""}}`
-			template, _ = sjson.Set(template, "index", rootResult.Get("output_index").Int())
+			template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
 
 			output += "event: content_block_delta\n"
 			output += fmt.Sprintf("data: %s\n\n", template)
@@ -147,14 +161,15 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 		itemType := itemResult.Get("type").String()
 		if itemType == "function_call" {
 			template = `{"type":"content_block_stop","index":0}`
-			template, _ = sjson.Set(template, "index", rootResult.Get("output_index").Int())
+			template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
+			(*param).(*ConvertCodexResponseToClaudeParams).BlockIndex++
 
 			output = "event: content_block_stop\n"
 			output += fmt.Sprintf("data: %s\n\n", template)
 		}
 	} else if typeStr == "response.function_call_arguments.delta" {
 		template = `{"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""}}`
-		template, _ = sjson.Set(template, "index", rootResult.Get("output_index").Int())
+		template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
 		template, _ = sjson.Set(template, "delta.partial_json", rootResult.Get("delta").String())
 
 		output += "event: content_block_delta\n"
@@ -193,8 +208,12 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 	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())
+	inputTokens, outputTokens, cachedTokens := extractResponsesUsage(responseData.Get("usage"))
+	out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+	out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
+	if cachedTokens > 0 {
+		out, _ = sjson.Set(out, "usage.cache_read_input_tokens", cachedTokens)
+	}
 
 	hasToolCall := false
 
@@ -297,12 +316,27 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 		out, _ = sjson.SetRaw(out, "stop_sequence", stopSequence.Raw)
 	}
 
-	if responseData.Get("usage.input_tokens").Exists() || responseData.Get("usage.output_tokens").Exists() {
-		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())
+	return out
+}
+
+func extractResponsesUsage(usage gjson.Result) (int64, int64, int64) {
+	if !usage.Exists() || usage.Type == gjson.Null {
+		return 0, 0, 0
 	}
 
-	return out
+	inputTokens := usage.Get("input_tokens").Int()
+	outputTokens := usage.Get("output_tokens").Int()
+	cachedTokens := usage.Get("input_tokens_details.cached_tokens").Int()
+
+	if cachedTokens > 0 {
+		if inputTokens >= cachedTokens {
+			inputTokens -= cachedTokens
+		} else {
+			inputTokens = 0
+		}
+	}
+
+	return inputTokens, outputTokens, cachedTokens
 }
 
 // buildReverseMapFromClaudeOriginalShortToOriginal builds a map[short]original from original Claude request tools.

internal/translator/codex/gemini/codex_gemini_request.go

@@ -14,6 +14,7 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -38,11 +39,12 @@ import (
 //   - []byte: The transformed request data in Codex API format
 func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool) []byte {
 	rawJSON := bytes.Clone(inputRawJSON)
+	userAgent := misc.ExtractCodexUserAgent(rawJSON)
 	// Base template
 	out := `{"model":"","instructions":"","input":[]}`
 
 	// Inject standard Codex instructions
-	_, instructions := misc.CodexInstructionsForModel(modelName, "")
+	_, instructions := misc.CodexInstructionsForModel(modelName, "", userAgent)
 	out, _ = sjson.Set(out, "instructions", instructions)
 
 	root := gjson.ParseBytes(rawJSON)
@@ -92,7 +94,7 @@ func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	// System instruction -> as a user message with input_text parts
 	sysParts := root.Get("system_instruction.parts")
 	if sysParts.IsArray() {
-		msg := `{"type":"message","role":"user","content":[]}`
+		msg := `{"type":"message","role":"developer","content":[]}`
 		arr := sysParts.Array()
 		for i := 0; i < len(arr); i++ {
 			p := arr[i]
@@ -246,21 +248,28 @@ func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	// Fixed flags aligning with Codex expectations
 	out, _ = sjson.Set(out, "parallel_tool_calls", true)
 
-	// Convert thinkingBudget to reasoning.effort for level-based models
-	reasoningEffort := "medium" // default
+	// Convert Gemini thinkingConfig to Codex reasoning.effort.
+	effortSet := false
 	if genConfig := root.Get("generationConfig"); genConfig.Exists() {
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			if util.ModelUsesThinkingLevels(modelName) {
-				if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
-					budget := int(thinkingBudget.Int())
-					if effort, ok := util.ThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
-						reasoningEffort = effort
-					}
+			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+				effort := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
+				if effort != "" {
+					out, _ = sjson.Set(out, "reasoning.effort", effort)
+					effortSet = true
+				}
+			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+				if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
+					out, _ = sjson.Set(out, "reasoning.effort", effort)
+					effortSet = true
 				}
 			}
 		}
 	}
-	out, _ = sjson.Set(out, "reasoning.effort", reasoningEffort)
+	if !effortSet {
+		// No thinking config, set default effort
+		out, _ = sjson.Set(out, "reasoning.effort", "medium")
+	}
 	out, _ = sjson.Set(out, "reasoning.summary", "auto")
 	out, _ = sjson.Set(out, "stream", true)
 	out, _ = sjson.Set(out, "store", false)

internal/translator/codex/openai/chat-completions/codex_openai_request.go

@@ -31,8 +31,9 @@ import (
 //   - []byte: The transformed request data in OpenAI Responses API format
 func ConvertOpenAIRequestToCodex(modelName string, inputRawJSON []byte, stream bool) []byte {
 	rawJSON := bytes.Clone(inputRawJSON)
+	userAgent := misc.ExtractCodexUserAgent(rawJSON)
 	// Start with empty JSON object
-	out := `{}`
+	out := `{"instructions":""}`
 
 	// Stream must be set to true
 	out, _ = sjson.Set(out, "stream", stream)
@@ -96,8 +97,10 @@ func ConvertOpenAIRequestToCodex(modelName string, inputRawJSON []byte, stream b
 
 	// Extract system instructions from first system message (string or text object)
 	messages := gjson.GetBytes(rawJSON, "messages")
-	_, instructions := misc.CodexInstructionsForModel(modelName, "")
-	out, _ = sjson.Set(out, "instructions", instructions)
+	_, instructions := misc.CodexInstructionsForModel(modelName, "", userAgent)
+	if misc.GetCodexInstructionsEnabled() {
+		out, _ = sjson.Set(out, "instructions", instructions)
+	}
 	// if messages.IsArray() {
 	// 	arr := messages.Array()
 	// 	for i := 0; i < len(arr); i++ {
@@ -140,7 +143,7 @@ func ConvertOpenAIRequestToCodex(modelName string, inputRawJSON []byte, stream b
 				msg := `{}`
 				msg, _ = sjson.Set(msg, "type", "message")
 				if role == "system" {
-					msg, _ = sjson.Set(msg, "role", "user")
+					msg, _ = sjson.Set(msg, "role", "developer")
 				} else {
 					msg, _ = sjson.Set(msg, "role", role)
 				}
@@ -275,7 +278,15 @@ func ConvertOpenAIRequestToCodex(modelName string, inputRawJSON []byte, stream b
 		arr := tools.Array()
 		for i := 0; i < len(arr); i++ {
 			t := arr[i]
-			if t.Get("type").String() == "function" {
+			toolType := t.Get("type").String()
+			// Pass through built-in tools (e.g. {"type":"web_search"}) directly for the Responses API.
+			// Only "function" needs structural conversion because Chat Completions nests details under "function".
+			if toolType != "" && toolType != "function" && t.IsObject() {
+				out, _ = sjson.SetRaw(out, "tools.-1", t.Raw)
+				continue
+			}
+
+			if toolType == "function" {
 				item := `{}`
 				item, _ = sjson.Set(item, "type", "function")
 				fn := t.Get("function")
@@ -304,6 +315,37 @@ func ConvertOpenAIRequestToCodex(modelName string, inputRawJSON []byte, stream b
 		}
 	}
 
+	// Map tool_choice when present.
+	// Chat Completions: "tool_choice" can be a string ("auto"/"none") or an object (e.g. {"type":"function","function":{"name":"..."}}).
+	// Responses API: keep built-in tool choices as-is; flatten function choice to {"type":"function","name":"..."}.
+	if tc := gjson.GetBytes(rawJSON, "tool_choice"); tc.Exists() {
+		switch {
+		case tc.Type == gjson.String:
+			out, _ = sjson.Set(out, "tool_choice", tc.String())
+		case tc.IsObject():
+			tcType := tc.Get("type").String()
+			if tcType == "function" {
+				name := tc.Get("function.name").String()
+				if name != "" {
+					if short, ok := originalToolNameMap[name]; ok {
+						name = short
+					} else {
+						name = shortenNameIfNeeded(name)
+					}
+				}
+				choice := `{}`
+				choice, _ = sjson.Set(choice, "type", "function")
+				if name != "" {
+					choice, _ = sjson.Set(choice, "name", name)
+				}
+				out, _ = sjson.SetRaw(out, "tool_choice", choice)
+			} else if tcType != "" {
+				// Built-in tool choices (e.g. {"type":"web_search"}) are already Responses-compatible.
+				out, _ = sjson.SetRaw(out, "tool_choice", tc.Raw)
+			}
+		}
+	}
+
 	out, _ = sjson.Set(out, "store", false)
 	return []byte(out)
 }

internal/translator/codex/openai/responses/codex_openai-responses_request.go

@@ -12,6 +12,8 @@ import (
 
 func ConvertOpenAIResponsesRequestToCodex(modelName string, inputRawJSON []byte, _ bool) []byte {
 	rawJSON := bytes.Clone(inputRawJSON)
+	userAgent := misc.ExtractCodexUserAgent(rawJSON)
+	rawJSON = misc.StripCodexUserAgent(rawJSON)
 
 	rawJSON, _ = sjson.SetBytes(rawJSON, "stream", true)
 	rawJSON, _ = sjson.SetBytes(rawJSON, "store", false)
@@ -32,7 +34,7 @@ func ConvertOpenAIResponsesRequestToCodex(modelName string, inputRawJSON []byte,
 		originalInstructionsText = originalInstructionsResult.String()
 	}
 
-	hasOfficialInstructions, instructions := misc.CodexInstructionsForModel(modelName, originalInstructionsResult.String())
+	hasOfficialInstructions, instructions := misc.CodexInstructionsForModel(modelName, originalInstructionsResult.String(), userAgent)
 
 	inputResult := gjson.GetBytes(rawJSON, "input")
 	var inputResults []gjson.Result
@@ -72,6 +74,11 @@ func ConvertOpenAIResponsesRequestToCodex(modelName string, inputRawJSON []byte,
 	}
 
 	if hasOfficialInstructions {
+		newInput := "[]"
+		for _, item := range inputResults {
+			newInput, _ = sjson.SetRaw(newInput, "-1", item.Raw)
+		}
+		rawJSON, _ = sjson.SetRawBytes(rawJSON, "input", []byte(newInput))
 		return rawJSON
 	}
 	// log.Debugf("instructions not matched, %s\n", originalInstructions)

internal/translator/codex/openai/responses/codex_openai-responses_response.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"fmt"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -18,7 +19,10 @@ func ConvertCodexResponseToOpenAIResponses(ctx context.Context, modelName string
 		if typeResult := gjson.GetBytes(rawJSON, "type"); typeResult.Exists() {
 			typeStr := typeResult.String()
 			if typeStr == "response.created" || typeStr == "response.in_progress" || typeStr == "response.completed" {
-				rawJSON, _ = sjson.SetBytes(rawJSON, "response.instructions", gjson.GetBytes(originalRequestRawJSON, "instructions").String())
+				if gjson.GetBytes(rawJSON, "response.instructions").Exists() {
+					instructions := selectInstructions(originalRequestRawJSON, requestRawJSON)
+					rawJSON, _ = sjson.SetBytes(rawJSON, "response.instructions", instructions)
+				}
 			}
 		}
 		out := fmt.Sprintf("data: %s", string(rawJSON))
@@ -37,6 +41,16 @@ func ConvertCodexResponseToOpenAIResponsesNonStream(_ context.Context, modelName
 	}
 	responseResult := rootResult.Get("response")
 	template := responseResult.Raw
-	template, _ = sjson.Set(template, "instructions", gjson.GetBytes(originalRequestRawJSON, "instructions").String())
+	if responseResult.Get("instructions").Exists() {
+		template, _ = sjson.Set(template, "instructions", selectInstructions(originalRequestRawJSON, requestRawJSON))
+	}
 	return template
 }
+
+func selectInstructions(originalRequestRawJSON, requestRawJSON []byte) string {
+	userAgent := misc.ExtractCodexUserAgent(originalRequestRawJSON)
+	if misc.IsOpenCodeUserAgent(userAgent) {
+		return gjson.GetBytes(requestRawJSON, "instructions").String()
+	}
+	return gjson.GetBytes(originalRequestRawJSON, "instructions").String()
+}

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

@@ -10,7 +10,6 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -160,12 +159,12 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 	}
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
-	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() && util.ModelSupportsThinking(modelName) {
+	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
 		if t.Get("type").String() == "enabled" {
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		}
 	}

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

@@ -35,37 +35,19 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 	// Model
 	out, _ = sjson.SetBytes(out, "model", modelName)
 
-	// Reasoning effort -> thinkingBudget/include_thoughts
-	// Note: OpenAI official fields take precedence over extra_body.google.thinking_config
+	// Apply thinking configuration: convert OpenAI reasoning_effort to Gemini CLI thinkingConfig.
+	// Inline translation-only mapping; capability checks happen later in ApplyThinking.
 	re := gjson.GetBytes(rawJSON, "reasoning_effort")
-	hasOfficialThinking := re.Exists()
-	if hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
-		out = util.ApplyReasoningEffortToGeminiCLI(out, re.String())
-	}
-
-	// Cherry Studio extension extra_body.google.thinking_config (effective only when official fields are absent)
-	// Only apply for models that use numeric budgets, not discrete levels.
-	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
-		if tc := gjson.GetBytes(rawJSON, "extra_body.google.thinking_config"); tc.Exists() && tc.IsObject() {
-			var setBudget bool
-			var budget int
-
-			if v := tc.Get("thinkingBudget"); v.Exists() {
-				budget = int(v.Int())
-				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-				setBudget = true
-			} else if v := tc.Get("thinking_budget"); v.Exists() {
-				budget = int(v.Int())
-				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-				setBudget = true
-			}
-
-			if v := tc.Get("includeThoughts"); v.Exists() {
-				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", v.Bool())
-			} else if v := tc.Get("include_thoughts"); v.Exists() {
-				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", v.Bool())
-			} else if setBudget && budget != 0 {
-				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
+	if re.Exists() {
+		effort := strings.ToLower(strings.TrimSpace(re.String()))
+		if effort != "" {
+			thinkingPath := "request.generationConfig.thinkingConfig"
+			if effort == "auto" {
+				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingBudget", -1)
+				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", true)
+			} else {
+				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingLevel", effort)
+				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", effort != "none")
 			}
 		}
 	}
@@ -147,29 +129,33 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 			}
 		}
 
+		systemPartIndex := 0
 		for i := 0; i < len(arr); i++ {
 			m := arr[i]
 			role := m.Get("role").String()
 			content := m.Get("content")
 
-			if role == "system" && len(arr) > 1 {
+			if (role == "system" || role == "developer") && len(arr) > 1 {
 				// system -> request.systemInstruction as a user message style
 				if content.Type == gjson.String {
 					out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
-					out, _ = sjson.SetBytes(out, "request.systemInstruction.parts.0.text", content.String())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), content.String())
+					systemPartIndex++
 				} 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())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), content.Get("text").String())
+					systemPartIndex++
 				} 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())
+							out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), contents[j].Get("text").String())
+							systemPartIndex++
 						}
 					}
 				}
-			} else if role == "user" || (role == "system" && len(arr) == 1) {
+			} else if role == "user" || ((role == "system" || role == "developer") && len(arr) == 1) {
 				// Build single user content node to avoid splitting into multiple contents
 				node := []byte(`{"role":"user","parts":[]}`)
 				if content.Type == gjson.String {
@@ -191,6 +177,7 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 									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)
+									node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiCLIFunctionThoughtSignature)
 									p++
 								}
 							}
@@ -236,6 +223,7 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 									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)
+									node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiCLIFunctionThoughtSignature)
 									p++
 								}
 							}

internal/translator/gemini/claude/gemini_claude_request.go

@@ -10,7 +10,6 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -153,13 +152,13 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 	}
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when enabled
-	// Only apply for models that use numeric budgets, not discrete levels.
-	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
+	// Translator only does format conversion, ApplyThinking handles model capability validation.
+	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
 		if t.Get("type").String() == "enabled" {
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
 				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
-				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", true)
+				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		}
 	}

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

@@ -35,55 +35,19 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 	// Model
 	out, _ = sjson.SetBytes(out, "model", modelName)
 
-	// Reasoning effort -> thinkingBudget/include_thoughts
-	// Note: OpenAI official fields take precedence over extra_body.google.thinking_config
-	// 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.
+	// Apply thinking configuration: convert OpenAI reasoning_effort to Gemini thinkingConfig.
+	// Inline translation-only mapping; capability checks happen later in ApplyThinking.
 	re := gjson.GetBytes(rawJSON, "reasoning_effort")
-	hasOfficialThinking := re.Exists()
-	if hasOfficialThinking && util.ModelSupportsThinking(modelName) {
+	if re.Exists() {
 		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)
-	// Only apply for models that use numeric budgets, not discrete levels.
-	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
-		if tc := gjson.GetBytes(rawJSON, "extra_body.google.thinking_config"); tc.Exists() && tc.IsObject() {
-			var setBudget bool
-			var budget int
-
-			if v := tc.Get("thinkingBudget"); v.Exists() {
-				budget = int(v.Int())
-				out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
-				setBudget = true
-			} else if v := tc.Get("thinking_budget"); v.Exists() {
-				budget = int(v.Int())
-				out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
-				setBudget = true
-			}
-
-			if v := tc.Get("includeThoughts"); v.Exists() {
-				out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.include_thoughts", v.Bool())
-			} else if v := tc.Get("include_thoughts"); v.Exists() {
-				out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.include_thoughts", v.Bool())
-			} else if setBudget && budget != 0 {
-				out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.include_thoughts", true)
+		if effort != "" {
+			thinkingPath := "generationConfig.thinkingConfig"
+			if effort == "auto" {
+				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingBudget", -1)
+				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", true)
+			} else {
+				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingLevel", effort)
+				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", effort != "none")
 			}
 		}
 	}
@@ -165,29 +129,33 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 			}
 		}
 
+		systemPartIndex := 0
 		for i := 0; i < len(arr); i++ {
 			m := arr[i]
 			role := m.Get("role").String()
 			content := m.Get("content")
 
-			if role == "system" && len(arr) > 1 {
+			if (role == "system" || role == "developer") && len(arr) > 1 {
 				// system -> system_instruction as a user message style
 				if content.Type == gjson.String {
 					out, _ = sjson.SetBytes(out, "system_instruction.role", "user")
-					out, _ = sjson.SetBytes(out, "system_instruction.parts.0.text", content.String())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("system_instruction.parts.%d.text", systemPartIndex), content.String())
+					systemPartIndex++
 				} 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())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("system_instruction.parts.%d.text", systemPartIndex), content.Get("text").String())
+					systemPartIndex++
 				} else if content.IsArray() {
 					contents := content.Array()
 					if len(contents) > 0 {
-						out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
+						out, _ = sjson.SetBytes(out, "system_instruction.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())
+							out, _ = sjson.SetBytes(out, fmt.Sprintf("system_instruction.parts.%d.text", systemPartIndex), contents[j].Get("text").String())
+							systemPartIndex++
 						}
 					}
 				}
-			} else if role == "user" || (role == "system" && len(arr) == 1) {
+			} else if role == "user" || ((role == "system" || role == "developer") && len(arr) == 1) {
 				// Build single user content node to avoid splitting into multiple contents
 				node := []byte(`{"role":"user","parts":[]}`)
 				if content.Type == gjson.String {
@@ -198,7 +166,10 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 					for _, item := range items {
 						switch item.Get("type").String() {
 						case "text":
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", item.Get("text").String())
+							text := item.Get("text").String()
+							if text != "" {
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", text)
+							}
 							p++
 						case "image_url":
 							imageURL := item.Get("image_url.url").String()
@@ -209,6 +180,7 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 									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)
+									node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiFunctionThoughtSignature)
 									p++
 								}
 							}
@@ -242,6 +214,10 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 					for _, item := range content.Array() {
 						switch item.Get("type").String() {
 						case "text":
+							text := item.Get("text").String()
+							if text != "" {
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", text)
+							}
 							p++
 						case "image_url":
 							// If the assistant returned an inline data URL, preserve it for history fidelity.
@@ -253,6 +229,7 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 									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)
+									node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".thoughtSignature", geminiFunctionThoughtSignature)
 									p++
 								}
 							}

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

@@ -5,7 +5,6 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -388,31 +387,19 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 		out, _ = sjson.Set(out, "generationConfig.stopSequences", sequences)
 	}
 
-	// OpenAI official reasoning fields take precedence
-	// Only convert for models that use numeric budgets (not discrete levels).
-	hasOfficialThinking := root.Get("reasoning.effort").Exists()
-	if hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
-		reasoningEffort := root.Get("reasoning.effort")
-		out = string(util.ApplyReasoningEffortToGemini([]byte(out), reasoningEffort.String()))
-	}
-
-	// Cherry Studio extension (applies only when official fields are missing)
-	// Only apply for models that use numeric budgets, not discrete levels.
-	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
-		if tc := root.Get("extra_body.google.thinking_config"); tc.Exists() && tc.IsObject() {
-			var setBudget bool
-			var budget int
-			if v := tc.Get("thinking_budget"); v.Exists() {
-				budget = int(v.Int())
-				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
-				setBudget = true
-			}
-			if v := tc.Get("include_thoughts"); v.Exists() {
-				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", v.Bool())
-			} else if setBudget {
-				if budget != 0 {
-					out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", true)
-				}
+	// Apply thinking configuration: convert OpenAI Responses API reasoning.effort to Gemini thinkingConfig.
+	// Inline translation-only mapping; capability checks happen later in ApplyThinking.
+	re := root.Get("reasoning.effort")
+	if re.Exists() {
+		effort := strings.ToLower(strings.TrimSpace(re.String()))
+		if effort != "" {
+			thinkingPath := "generationConfig.thinkingConfig"
+			if effort == "auto" {
+				out, _ = sjson.Set(out, thinkingPath+".thinkingBudget", -1)
+				out, _ = sjson.Set(out, thinkingPath+".includeThoughts", true)
+			} else {
+				out, _ = sjson.Set(out, thinkingPath+".thinkingLevel", effort)
+				out, _ = sjson.Set(out, thinkingPath+".includeThoughts", effort != "none")
 			}
 		}
 	}

internal/translator/openai/claude/openai_claude_request.go

@@ -9,7 +9,7 @@ import (
 	"bytes"
 	"strings"
 
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -61,23 +61,23 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 	out, _ = sjson.Set(out, "stream", 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() {
+	if thinkingConfig := root.Get("thinking"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
+		if thinkingType := thinkingConfig.Get("type"); thinkingType.Exists() {
 			switch thinkingType.String() {
 			case "enabled":
-				if budgetTokens := thinking.Get("budget_tokens"); budgetTokens.Exists() {
+				if budgetTokens := thinkingConfig.Get("budget_tokens"); budgetTokens.Exists() {
 					budget := int(budgetTokens.Int())
-					if effort, ok := util.ThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
+					if effort, ok := thinking.ConvertBudgetToLevel(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 != "" {
+					if effort, ok := thinking.ConvertBudgetToLevel(-1); ok && effort != "" {
 						out, _ = sjson.Set(out, "reasoning_effort", effort)
 					}
 				}
 			case "disabled":
-				if effort, ok := util.ThinkingBudgetToEffort(modelName, 0); ok && effort != "" {
+				if effort, ok := thinking.ConvertBudgetToLevel(0); ok && effort != "" {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)
 				}
 			}
@@ -88,7 +88,7 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 	var messagesJSON = "[]"
 
 	// Handle system message first
-	systemMsgJSON := `{"role":"system","content":[{"type":"text","text":"Use ANY tool, the parameters MUST accord with RFC 8259 (The JavaScript Object Notation (JSON) Data Interchange Format), the keys and value MUST be enclosed in double quotes."}]}`
+	systemMsgJSON := `{"role":"system","content":[]}`
 	if system := root.Get("system"); system.Exists() {
 		if system.Type == gjson.String {
 			if system.String() != "" {
@@ -129,7 +129,7 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 					case "thinking":
 						// Only map thinking to reasoning_content for assistant messages (security: prevent injection)
 						if role == "assistant" {
-							thinkingText := util.GetThinkingText(part)
+							thinkingText := thinking.GetThinkingText(part)
 							// Skip empty or whitespace-only thinking
 							if strings.TrimSpace(thinkingText) != "" {
 								reasoningParts = append(reasoningParts, thinkingText)

internal/translator/openai/claude/openai_claude_response.go

@@ -289,27 +289,22 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 	// Only process if usage has actual values (not null)
 	if param.FinishReason != "" {
 		usage := root.Get("usage")
-		var inputTokens, outputTokens int64
+		var inputTokens, outputTokens, cachedTokens int64
 		if usage.Exists() && usage.Type != gjson.Null {
-			// Check if usage has actual token counts
-			promptTokens := usage.Get("prompt_tokens")
-			completionTokens := usage.Get("completion_tokens")
-
-			if promptTokens.Exists() && completionTokens.Exists() {
-				inputTokens = promptTokens.Int()
-				outputTokens = completionTokens.Int()
+			inputTokens, outputTokens, cachedTokens = extractOpenAIUsage(usage)
+			// Send message_delta with usage
+			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)
+			if cachedTokens > 0 {
+				messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.cache_read_input_tokens", cachedTokens)
 			}
+			results = append(results, "event: message_delta\ndata: "+messageDeltaJSON+"\n\n")
+			param.MessageDeltaSent = true
+
+			emitMessageStopIfNeeded(param, &results)
 		}
-		// Send message_delta with usage
-		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)
-
 	}
 
 	return results
@@ -424,13 +419,12 @@ func convertOpenAINonStreamingToAnthropic(rawJSON []byte) []string {
 
 	// Set usage information
 	if usage := root.Get("usage"); usage.Exists() {
-		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()
+		inputTokens, outputTokens, cachedTokens := extractOpenAIUsage(usage)
+		out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+		out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
+		if cachedTokens > 0 {
+			out, _ = sjson.Set(out, "usage.cache_read_input_tokens", cachedTokens)
 		}
-		out, _ = sjson.Set(out, "usage.reasoning_tokens", reasoningTokens)
 	}
 
 	return []string{out}
@@ -675,8 +669,12 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 	}
 
 	if respUsage := root.Get("usage"); respUsage.Exists() {
-		out, _ = sjson.Set(out, "usage.input_tokens", respUsage.Get("prompt_tokens").Int())
-		out, _ = sjson.Set(out, "usage.output_tokens", respUsage.Get("completion_tokens").Int())
+		inputTokens, outputTokens, cachedTokens := extractOpenAIUsage(respUsage)
+		out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+		out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
+		if cachedTokens > 0 {
+			out, _ = sjson.Set(out, "usage.cache_read_input_tokens", cachedTokens)
+		}
 	}
 
 	if !stopReasonSet {
@@ -693,3 +691,23 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 func ClaudeTokenCount(ctx context.Context, count int64) string {
 	return fmt.Sprintf(`{"input_tokens":%d}`, count)
 }
+
+func extractOpenAIUsage(usage gjson.Result) (int64, int64, int64) {
+	if !usage.Exists() || usage.Type == gjson.Null {
+		return 0, 0, 0
+	}
+
+	inputTokens := usage.Get("prompt_tokens").Int()
+	outputTokens := usage.Get("completion_tokens").Int()
+	cachedTokens := usage.Get("prompt_tokens_details.cached_tokens").Int()
+
+	if cachedTokens > 0 {
+		if inputTokens >= cachedTokens {
+			inputTokens -= cachedTokens
+		} else {
+			inputTokens = 0
+		}
+	}
+
+	return inputTokens, outputTokens, cachedTokens
+}

internal/translator/openai/gemini/openai_gemini_request.go

@@ -12,7 +12,7 @@ import (
 	"math/big"
 	"strings"
 
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -77,12 +77,15 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 			}
 		}
 
-		// Convert thinkingBudget to reasoning_effort
-		// Always perform conversion to support allowCompat models that may not be in registry
+		// Map Gemini thinkingConfig to OpenAI reasoning_effort.
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
-				budget := int(thinkingBudget.Int())
-				if effort, ok := util.ThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
+			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+				effort := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
+				if effort != "" {
+					out, _ = sjson.Set(out, "reasoning_effort", effort)
+				}
+			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+				if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)
 				}
 			}

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

@@ -163,6 +163,14 @@ func ConvertOpenAIResponsesRequestToOpenAIChatCompletions(modelName string, inpu
 		var chatCompletionsTools []interface{}
 
 		tools.ForEach(func(_, tool gjson.Result) bool {
+			// Built-in tools (e.g. {"type":"web_search"}) are already compatible with the Chat Completions schema.
+			// Only function tools need structural conversion because Chat Completions nests details under "function".
+			toolType := tool.Get("type").String()
+			if toolType != "" && toolType != "function" && tool.IsObject() {
+				chatCompletionsTools = append(chatCompletionsTools, tool.Value())
+				return true
+			}
+
 			chatTool := `{"type":"function","function":{}}`
 
 			// Convert tool structure from responses format to chat completions format

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

@@ -12,6 +12,10 @@ import (
 	"github.com/tidwall/sjson"
 )
 
+type oaiToResponsesStateReasoning struct {
+	ReasoningID   string
+	ReasoningData string
+}
 type oaiToResponsesState struct {
 	Seq            int
 	ResponseID     string
@@ -23,6 +27,7 @@ type oaiToResponsesState struct {
 	// Per-output message text buffers by index
 	MsgTextBuf   map[int]*strings.Builder
 	ReasoningBuf strings.Builder
+	Reasonings   []oaiToResponsesStateReasoning
 	FuncArgsBuf  map[int]*strings.Builder // index -> args
 	FuncNames    map[int]string           // index -> name
 	FuncCallIDs  map[int]string           // index -> call_id
@@ -63,6 +68,7 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 			MsgItemDone:     make(map[int]bool),
 			FuncArgsDone:    make(map[int]bool),
 			FuncItemDone:    make(map[int]bool),
+			Reasonings:      make([]oaiToResponsesStateReasoning, 0),
 		}
 	}
 	st := (*param).(*oaiToResponsesState)
@@ -157,6 +163,31 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 		st.Started = true
 	}
 
+	stopReasoning := func(text string) {
+		// Emit reasoning done events
+		textDone := `{"type":"response.reasoning_summary_text.done","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"text":""}`
+		textDone, _ = sjson.Set(textDone, "sequence_number", nextSeq())
+		textDone, _ = sjson.Set(textDone, "item_id", st.ReasoningID)
+		textDone, _ = sjson.Set(textDone, "output_index", st.ReasoningIndex)
+		textDone, _ = sjson.Set(textDone, "text", text)
+		out = append(out, emitRespEvent("response.reasoning_summary_text.done", textDone))
+		partDone := `{"type":"response.reasoning_summary_part.done","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"part":{"type":"summary_text","text":""}}`
+		partDone, _ = sjson.Set(partDone, "sequence_number", nextSeq())
+		partDone, _ = sjson.Set(partDone, "item_id", st.ReasoningID)
+		partDone, _ = sjson.Set(partDone, "output_index", st.ReasoningIndex)
+		partDone, _ = sjson.Set(partDone, "part.text", text)
+		out = append(out, emitRespEvent("response.reasoning_summary_part.done", partDone))
+		outputItemDone := `{"type":"response.output_item.done","item":{"id":"","type":"reasoning","encrypted_content":"","summary":[{"type":"summary_text","text":""}]},"output_index":0,"sequence_number":0}`
+		outputItemDone, _ = sjson.Set(outputItemDone, "sequence_number", nextSeq())
+		outputItemDone, _ = sjson.Set(outputItemDone, "item.id", st.ReasoningID)
+		outputItemDone, _ = sjson.Set(outputItemDone, "output_index", st.ReasoningIndex)
+		outputItemDone, _ = sjson.Set(outputItemDone, "item.summary.text", text)
+		out = append(out, emitRespEvent("response.output_item.done", outputItemDone))
+
+		st.Reasonings = append(st.Reasonings, oaiToResponsesStateReasoning{ReasoningID: st.ReasoningID, ReasoningData: text})
+		st.ReasoningID = ""
+	}
+
 	// choices[].delta content / tool_calls / reasoning_content
 	if choices := root.Get("choices"); choices.Exists() && choices.IsArray() {
 		choices.ForEach(func(_, choice gjson.Result) bool {
@@ -165,6 +196,10 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 			if delta.Exists() {
 				if c := delta.Get("content"); c.Exists() && c.String() != "" {
 					// Ensure the message item and its first content part are announced before any text deltas
+					if st.ReasoningID != "" {
+						stopReasoning(st.ReasoningBuf.String())
+						st.ReasoningBuf.Reset()
+					}
 					if !st.MsgItemAdded[idx] {
 						item := `{"type":"response.output_item.added","sequence_number":0,"output_index":0,"item":{"id":"","type":"message","status":"in_progress","content":[],"role":"assistant"}}`
 						item, _ = sjson.Set(item, "sequence_number", nextSeq())
@@ -226,6 +261,10 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 
 				// tool calls
 				if tcs := delta.Get("tool_calls"); tcs.Exists() && tcs.IsArray() {
+					if st.ReasoningID != "" {
+						stopReasoning(st.ReasoningBuf.String())
+						st.ReasoningBuf.Reset()
+					}
 					// Before emitting any function events, if a message is open for this index,
 					// close its text/content to match Codex expected ordering.
 					if st.MsgItemAdded[idx] && !st.MsgItemDone[idx] {
@@ -361,17 +400,8 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 				}
 
 				if st.ReasoningID != "" {
-					// Emit reasoning done events
-					textDone := `{"type":"response.reasoning_summary_text.done","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"text":""}`
-					textDone, _ = sjson.Set(textDone, "sequence_number", nextSeq())
-					textDone, _ = sjson.Set(textDone, "item_id", st.ReasoningID)
-					textDone, _ = sjson.Set(textDone, "output_index", st.ReasoningIndex)
-					out = append(out, emitRespEvent("response.reasoning_summary_text.done", textDone))
-					partDone := `{"type":"response.reasoning_summary_part.done","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"part":{"type":"summary_text","text":""}}`
-					partDone, _ = sjson.Set(partDone, "sequence_number", nextSeq())
-					partDone, _ = sjson.Set(partDone, "item_id", st.ReasoningID)
-					partDone, _ = sjson.Set(partDone, "output_index", st.ReasoningIndex)
-					out = append(out, emitRespEvent("response.reasoning_summary_part.done", partDone))
+					stopReasoning(st.ReasoningBuf.String())
+					st.ReasoningBuf.Reset()
 				}
 
 				// Emit function call done events for any active function calls
@@ -485,11 +515,13 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 				}
 				// Build response.output using aggregated buffers
 				outputsWrapper := `{"arr":[]}`
-				if st.ReasoningBuf.Len() > 0 {
-					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)
+				if len(st.Reasonings) > 0 {
+					for _, r := range st.Reasonings {
+						item := `{"id":"","type":"reasoning","summary":[{"type":"summary_text","text":""}]}`
+						item, _ = sjson.Set(item, "id", r.ReasoningID)
+						item, _ = sjson.Set(item, "summary.0.text", r.ReasoningData)
+						outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
+					}
 				}
 				// Append message items in ascending index order
 				if len(st.MsgItemAdded) > 0 {

internal/util/claude_thinking.go

@@ -1,49 +0,0 @@
-package util
-
-import (
-	"github.com/tidwall/gjson"
-	"github.com/tidwall/sjson"
-)
-
-// ApplyClaudeThinkingConfig applies thinking configuration to a Claude API request payload.
-// It sets the thinking.type to "enabled" and thinking.budget_tokens to the specified budget.
-// If budget is nil or the payload already has thinking config, it returns the payload unchanged.
-func ApplyClaudeThinkingConfig(body []byte, budget *int) []byte {
-	if budget == nil {
-		return body
-	}
-	if gjson.GetBytes(body, "thinking").Exists() {
-		return body
-	}
-	if *budget <= 0 {
-		return body
-	}
-	updated := body
-	updated, _ = sjson.SetBytes(updated, "thinking.type", "enabled")
-	updated, _ = sjson.SetBytes(updated, "thinking.budget_tokens", *budget)
-	return updated
-}
-
-// ResolveClaudeThinkingConfig resolves thinking configuration from metadata for Claude models.
-// It uses the unified ResolveThinkingConfigFromMetadata and normalizes the budget.
-// Returns the normalized budget (nil if thinking should not be enabled) and whether it matched.
-func ResolveClaudeThinkingConfig(modelName string, metadata map[string]any) (*int, bool) {
-	if !ModelSupportsThinking(modelName) {
-		return nil, false
-	}
-	budget, include, matched := ResolveThinkingConfigFromMetadata(modelName, metadata)
-	if !matched {
-		return nil, false
-	}
-	if include != nil && !*include {
-		return nil, true
-	}
-	if budget == nil {
-		return nil, true
-	}
-	normalized := NormalizeThinkingBudget(modelName, *budget)
-	if normalized <= 0 {
-		return nil, true
-	}
-	return &normalized, true
-}

internal/util/gemini_schema.go

@@ -19,6 +19,7 @@ func CleanJSONSchemaForAntigravity(jsonStr string) string {
 	// Phase 1: Convert and add hints
 	jsonStr = convertRefsToHints(jsonStr)
 	jsonStr = convertConstToEnum(jsonStr)
+	jsonStr = convertEnumValuesToStrings(jsonStr)
 	jsonStr = addEnumHints(jsonStr)
 	jsonStr = addAdditionalPropertiesHints(jsonStr)
 	jsonStr = moveConstraintsToDescription(jsonStr)
@@ -77,6 +78,33 @@ func convertConstToEnum(jsonStr string) string {
 	return jsonStr
 }
 
+// convertEnumValuesToStrings ensures all enum values are strings.
+// Gemini API requires enum values to be of type string, not numbers or booleans.
+func convertEnumValuesToStrings(jsonStr string) string {
+	for _, p := range findPaths(jsonStr, "enum") {
+		arr := gjson.Get(jsonStr, p)
+		if !arr.IsArray() {
+			continue
+		}
+
+		var stringVals []string
+		needsConversion := false
+		for _, item := range arr.Array() {
+			// Check if any value is not a string
+			if item.Type != gjson.String {
+				needsConversion = true
+			}
+			stringVals = append(stringVals, item.String())
+		}
+
+		// Only update if we found non-string values
+		if needsConversion {
+			jsonStr, _ = sjson.Set(jsonStr, p, stringVals)
+		}
+	}
+	return jsonStr
+}
+
 func addEnumHints(jsonStr string) string {
 	for _, p := range findPaths(jsonStr, "enum") {
 		arr := gjson.Get(jsonStr, p)

internal/util/gemini_schema_test.go

@@ -818,3 +818,54 @@ func TestCleanJSONSchemaForAntigravity_MultipleFormats(t *testing.T) {
 		t.Errorf("date-time format hint should be added, got: %s", result)
 	}
 }
+
+func TestCleanJSONSchemaForAntigravity_NumericEnumToString(t *testing.T) {
+	// Gemini API requires enum values to be strings, not numbers
+	input := `{
+		"type": "object",
+		"properties": {
+			"priority": {"type": "integer", "enum": [0, 1, 2]},
+			"level": {"type": "number", "enum": [1.5, 2.5, 3.5]},
+			"status": {"type": "string", "enum": ["active", "inactive"]}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// Numeric enum values should be converted to strings
+	if strings.Contains(result, `"enum":[0,1,2]`) {
+		t.Errorf("Integer enum values should be converted to strings, got: %s", result)
+	}
+	if strings.Contains(result, `"enum":[1.5,2.5,3.5]`) {
+		t.Errorf("Float enum values should be converted to strings, got: %s", result)
+	}
+	// Should contain string versions
+	if !strings.Contains(result, `"0"`) || !strings.Contains(result, `"1"`) || !strings.Contains(result, `"2"`) {
+		t.Errorf("Integer enum values should be converted to string format, got: %s", result)
+	}
+	// String enum values should remain unchanged
+	if !strings.Contains(result, `"active"`) || !strings.Contains(result, `"inactive"`) {
+		t.Errorf("String enum values should remain unchanged, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_BooleanEnumToString(t *testing.T) {
+	// Boolean enum values should also be converted to strings
+	input := `{
+		"type": "object",
+		"properties": {
+			"enabled": {"type": "boolean", "enum": [true, false]}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// Boolean enum values should be converted to strings
+	if strings.Contains(result, `"enum":[true,false]`) {
+		t.Errorf("Boolean enum values should be converted to strings, got: %s", result)
+	}
+	// Should contain string versions "true" and "false"
+	if !strings.Contains(result, `"true"`) || !strings.Contains(result, `"false"`) {
+		t.Errorf("Boolean enum values should be converted to string format, got: %s", result)
+	}
+}

internal/util/gemini_thinking.go

@@ -1,611 +0,0 @@
-package util
-
-import (
-	"regexp"
-	"strings"
-
-	"github.com/tidwall/gjson"
-	"github.com/tidwall/sjson"
-)
-
-const (
-	GeminiThinkingBudgetMetadataKey  = "gemini_thinking_budget"
-	GeminiIncludeThoughtsMetadataKey = "gemini_include_thoughts"
-	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
-	}
-	updated := body
-	if budget != nil {
-		valuePath := "generationConfig.thinkingConfig.thinkingBudget"
-		rewritten, err := sjson.SetBytes(updated, valuePath, *budget)
-		if err == nil {
-			updated = rewritten
-		}
-	}
-	// Default to including thoughts when a budget override is present but no explicit include flag is provided.
-	incl := includeThoughts
-	if incl == nil && budget != nil && *budget != 0 {
-		defaultInclude := true
-		incl = &defaultInclude
-	}
-	if incl != nil {
-		valuePath := "generationConfig.thinkingConfig.include_thoughts"
-		rewritten, err := sjson.SetBytes(updated, valuePath, *incl)
-		if err == nil {
-			updated = rewritten
-		}
-	}
-	return updated
-}
-
-func ApplyGeminiCLIThinkingConfig(body []byte, budget *int, includeThoughts *bool) []byte {
-	if budget == nil && includeThoughts == nil {
-		return body
-	}
-	updated := body
-	if budget != nil {
-		valuePath := "request.generationConfig.thinkingConfig.thinkingBudget"
-		rewritten, err := sjson.SetBytes(updated, valuePath, *budget)
-		if err == nil {
-			updated = rewritten
-		}
-	}
-	// Default to including thoughts when a budget override is present but no explicit include flag is provided.
-	incl := includeThoughts
-	if incl == nil && budget != nil && *budget != 0 {
-		defaultInclude := true
-		incl = &defaultInclude
-	}
-	if incl != nil {
-		valuePath := "request.generationConfig.thinkingConfig.include_thoughts"
-		rewritten, err := sjson.SetBytes(updated, valuePath, *incl)
-		if err == nil {
-			updated = rewritten
-		}
-	}
-	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.
-// Note: Gemini 3 models are NOT included here because per Google's official documentation:
-//   - thinkingLevel defaults to "high" (dynamic thinking)
-//   - includeThoughts defaults to false
-//
-// We should not override these API defaults; let users explicitly configure if needed.
-var modelsWithDefaultThinking = map[string]bool{
-	// "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.
-func ModelHasDefaultThinking(model string) bool {
-	return modelsWithDefaultThinking[model]
-}
-
-// 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
-	}
-	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))
-// or numeric budget suffix (e.g., model(1000)) which gets converted to a thinkingLevel.
-func ApplyGemini3ThinkingLevelFromMetadata(model string, metadata map[string]any, body []byte) []byte {
-	// Use the alias from metadata if available for model type detection
-	lookupModel := ResolveOriginalModel(model, metadata)
-	if !IsGemini3Model(lookupModel) && !IsGemini3Model(model) {
-		return body
-	}
-
-	// Determine which model to use for validation
-	checkModel := model
-	if IsGemini3Model(lookupModel) {
-		checkModel = lookupModel
-	}
-
-	// First try to get effort string from metadata
-	effort, ok := ReasoningEffortFromMetadata(metadata)
-	if ok && effort != "" {
-		if level, valid := ValidateGemini3ThinkingLevel(checkModel, effort); valid {
-			return ApplyGeminiThinkingLevel(body, level, nil)
-		}
-	}
-
-	// Fallback: check for numeric budget and convert to thinkingLevel
-	budget, _, _, matched := ThinkingFromMetadata(metadata)
-	if matched && budget != nil {
-		if level, valid := ThinkingBudgetToGemini3Level(checkModel, *budget); 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))
-// or numeric budget suffix (e.g., model(1000)) which gets converted to a thinkingLevel.
-func ApplyGemini3ThinkingLevelFromMetadataCLI(model string, metadata map[string]any, body []byte) []byte {
-	// Use the alias from metadata if available for model type detection
-	lookupModel := ResolveOriginalModel(model, metadata)
-	if !IsGemini3Model(lookupModel) && !IsGemini3Model(model) {
-		return body
-	}
-
-	// Determine which model to use for validation
-	checkModel := model
-	if IsGemini3Model(lookupModel) {
-		checkModel = lookupModel
-	}
-
-	// First try to get effort string from metadata
-	effort, ok := ReasoningEffortFromMetadata(metadata)
-	if ok && effort != "" {
-		if level, valid := ValidateGemini3ThinkingLevel(checkModel, effort); valid {
-			return ApplyGeminiCLIThinkingLevel(body, level, nil)
-		}
-	}
-
-	// Fallback: check for numeric budget and convert to thinkingLevel
-	budget, _, _, matched := ThinkingFromMetadata(metadata)
-	if matched && budget != nil {
-		if level, valid := ThinkingBudgetToGemini3Level(checkModel, *budget); 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, metadata map[string]any, body []byte) []byte {
-	// Use the alias from metadata if available for model property lookup
-	lookupModel := ResolveOriginalModel(model, metadata)
-	if !ModelHasDefaultThinking(lookupModel) && !ModelHasDefaultThinking(model) {
-		return body
-	}
-	if gjson.GetBytes(body, "request.generationConfig.thinkingConfig").Exists() {
-		return body
-	}
-	// Gemini 3 models use thinkingLevel instead of thinkingBudget
-	if IsGemini3Model(lookupModel) || 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
-}
-
-// StripThinkingConfigIfUnsupported removes thinkingConfig from the request body
-// when the target model does not advertise Thinking capability. It cleans both
-// standard Gemini and Gemini CLI JSON envelopes. This acts as a final safety net
-// in case upstream injected thinking for an unsupported model.
-func StripThinkingConfigIfUnsupported(model string, body []byte) []byte {
-	if ModelSupportsThinking(model) || len(body) == 0 {
-		return body
-	}
-	updated := body
-	// Gemini CLI path
-	updated, _ = sjson.DeleteBytes(updated, "request.generationConfig.thinkingConfig")
-	// Standard Gemini path
-	updated, _ = sjson.DeleteBytes(updated, "generationConfig.thinkingConfig")
-	return updated
-}
-
-// NormalizeGeminiThinkingBudget normalizes the thinkingBudget value in a standard Gemini
-// request body (generationConfig.thinkingConfig.thinkingBudget path).
-// 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
-}
-
-// NormalizeGeminiCLIThinkingBudget normalizes the thinkingBudget value in a Gemini CLI
-// request body (request.generationConfig.thinkingConfig.thinkingBudget path).
-// 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
-}
-
-// ReasoningEffortBudgetMapping defines the thinkingBudget values for each reasoning effort level.
-var ReasoningEffortBudgetMapping = map[string]int{
-	"none":    0,
-	"auto":    -1,
-	"minimal": 512,
-	"low":     1024,
-	"medium":  8192,
-	"high":    24576,
-	"xhigh":   32768,
-}
-
-// ApplyReasoningEffortToGemini applies OpenAI reasoning_effort to Gemini thinkingConfig
-// for standard Gemini API format (generationConfig.thinkingConfig path).
-// Returns the modified body with thinkingBudget and include_thoughts set.
-func ApplyReasoningEffortToGemini(body []byte, effort string) []byte {
-	normalized := strings.ToLower(strings.TrimSpace(effort))
-	if normalized == "" {
-		return body
-	}
-
-	budgetPath := "generationConfig.thinkingConfig.thinkingBudget"
-	includePath := "generationConfig.thinkingConfig.include_thoughts"
-
-	if normalized == "none" {
-		body, _ = sjson.DeleteBytes(body, "generationConfig.thinkingConfig")
-		return body
-	}
-
-	budget, ok := ReasoningEffortBudgetMapping[normalized]
-	if !ok {
-		return body
-	}
-
-	body, _ = sjson.SetBytes(body, budgetPath, budget)
-	body, _ = sjson.SetBytes(body, includePath, true)
-	return body
-}
-
-// ApplyReasoningEffortToGeminiCLI applies OpenAI reasoning_effort to Gemini CLI thinkingConfig
-// for Gemini CLI API format (request.generationConfig.thinkingConfig path).
-// Returns the modified body with thinkingBudget and include_thoughts set.
-func ApplyReasoningEffortToGeminiCLI(body []byte, effort string) []byte {
-	normalized := strings.ToLower(strings.TrimSpace(effort))
-	if normalized == "" {
-		return body
-	}
-
-	budgetPath := "request.generationConfig.thinkingConfig.thinkingBudget"
-	includePath := "request.generationConfig.thinkingConfig.include_thoughts"
-
-	if normalized == "none" {
-		body, _ = sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig")
-		return body
-	}
-
-	budget, ok := ReasoningEffortBudgetMapping[normalized]
-	if !ok {
-		return body
-	}
-
-	body, _ = sjson.SetBytes(body, budgetPath, budget)
-	body, _ = sjson.SetBytes(body, includePath, true)
-	return body
-}
-
-// ConvertThinkingLevelToBudget checks for "generationConfig.thinkingConfig.thinkingLevel"
-// 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
-	}
-
-	// For Gemini 3 models, preserve thinkingLevel unless explicitly skipped
-	skipGemini3 := len(skipGemini3Check) > 0 && skipGemini3Check[0]
-	if IsGemini3Model(model) && !skipGemini3 {
-		return body
-	}
-
-	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
-	}
-
-	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
-	}
-	return updated
-}

internal/util/thinking.go

@@ -1,245 +0,0 @@
-package util
-
-import (
-	"strings"
-
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
-)
-
-// ModelSupportsThinking reports whether the given model has Thinking capability
-// according to the model registry metadata (provider-agnostic).
-func ModelSupportsThinking(model string) bool {
-	if model == "" {
-		return false
-	}
-	// First check the global dynamic registry
-	if info := registry.GetGlobalRegistry().GetModelInfo(model); info != nil {
-		return info.Thinking != nil
-	}
-	// Fallback: check static model definitions
-	if info := registry.LookupStaticModelInfo(model); info != nil {
-		return info.Thinking != nil
-	}
-	// Fallback: check Antigravity static config
-	if cfg := registry.GetAntigravityModelConfig()[model]; cfg != nil {
-		return cfg.Thinking != nil
-	}
-	return false
-}
-
-// NormalizeThinkingBudget clamps the requested thinking budget to the
-// supported range for the specified model using registry metadata only.
-// If the model is unknown or has no Thinking metadata, returns the original budget.
-// For dynamic (-1), returns -1 if DynamicAllowed; otherwise approximates mid-range
-// or min (0 if zero is allowed and mid <= 0).
-func NormalizeThinkingBudget(model string, budget int) int {
-	if budget == -1 { // dynamic
-		if found, minBudget, maxBudget, zeroAllowed, dynamicAllowed := thinkingRangeFromRegistry(model); found {
-			if dynamicAllowed {
-				return -1
-			}
-			mid := (minBudget + maxBudget) / 2
-			if mid <= 0 && zeroAllowed {
-				return 0
-			}
-			if mid <= 0 {
-				return minBudget
-			}
-			return mid
-		}
-		return -1
-	}
-	if found, minBudget, maxBudget, zeroAllowed, _ := thinkingRangeFromRegistry(model); found {
-		if budget == 0 {
-			if zeroAllowed {
-				return 0
-			}
-			return minBudget
-		}
-		if budget < minBudget {
-			return minBudget
-		}
-		if budget > maxBudget {
-			return maxBudget
-		}
-		return budget
-	}
-	return budget
-}
-
-// thinkingRangeFromRegistry attempts to read thinking ranges from the model registry.
-func thinkingRangeFromRegistry(model string) (found bool, min int, max int, zeroAllowed bool, dynamicAllowed bool) {
-	if model == "" {
-		return false, 0, 0, false, false
-	}
-	// First check global dynamic registry
-	if info := registry.GetGlobalRegistry().GetModelInfo(model); info != nil && info.Thinking != nil {
-		return true, info.Thinking.Min, info.Thinking.Max, info.Thinking.ZeroAllowed, info.Thinking.DynamicAllowed
-	}
-	// Fallback: check static model definitions
-	if info := registry.LookupStaticModelInfo(model); info != nil && info.Thinking != nil {
-		return true, info.Thinking.Min, info.Thinking.Max, info.Thinking.ZeroAllowed, info.Thinking.DynamicAllowed
-	}
-	// Fallback: check Antigravity static config
-	if cfg := registry.GetAntigravityModelConfig()[model]; cfg != nil && cfg.Thinking != nil {
-		return true, cfg.Thinking.Min, cfg.Thinking.Max, cfg.Thinking.ZeroAllowed, cfg.Thinking.DynamicAllowed
-	}
-	return false, 0, 0, false, false
-}
-
-// GetModelThinkingLevels returns the discrete reasoning effort levels for the model.
-// Returns nil if the model has no thinking support or no levels defined.
-func GetModelThinkingLevels(model string) []string {
-	if model == "" {
-		return nil
-	}
-	info := registry.GetGlobalRegistry().GetModelInfo(model)
-	if info == nil || info.Thinking == nil {
-		return nil
-	}
-	return info.Thinking.Levels
-}
-
-// ModelUsesThinkingLevels reports whether the model uses discrete reasoning
-// effort levels instead of numeric budgets.
-func ModelUsesThinkingLevels(model string) bool {
-	levels := GetModelThinkingLevels(model)
-	return len(levels) > 0
-}
-
-// NormalizeReasoningEffortLevel validates and normalizes a reasoning effort
-// level for the given model. Returns false when the level is not supported.
-func NormalizeReasoningEffortLevel(model, effort string) (string, bool) {
-	levels := GetModelThinkingLevels(model)
-	if len(levels) == 0 {
-		return "", false
-	}
-	loweredEffort := strings.ToLower(strings.TrimSpace(effort))
-	for _, lvl := range levels {
-		if strings.ToLower(lvl) == loweredEffort {
-			return lvl, true
-		}
-	}
-	return "", false
-}
-
-// IsOpenAICompatibilityModel reports whether the model is registered as an OpenAI-compatibility model.
-// These models may not advertise Thinking metadata in the registry.
-func IsOpenAICompatibilityModel(model string) bool {
-	if model == "" {
-		return false
-	}
-	info := registry.GetGlobalRegistry().GetModelInfo(model)
-	if info == nil {
-		return false
-	}
-	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
-	}
-}