feat(config): add payload filter rules to remove JSON paths

Introduce `Filter` rules in the payload configuration to remove specified JSON paths from the payload. Update related helper functions and add examples to `config.example.yaml`.

.github/workflows/docker-image.yml

@@ -10,13 +10,11 @@ env:
   DOCKERHUB_REPO: eceasy/cli-proxy-api
 
 jobs:
-  docker:
+  docker_amd64:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
         uses: actions/checkout@v4
-      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Login to DockerHub
@@ -29,18 +27,113 @@ jobs:
           echo VERSION=`git describe --tags --always --dirty` >> $GITHUB_ENV
           echo COMMIT=`git rev-parse --short HEAD` >> $GITHUB_ENV
           echo BUILD_DATE=`date -u +%Y-%m-%dT%H:%M:%SZ` >> $GITHUB_ENV
-      - name: Build and push
+      - name: Build and push (amd64)
         uses: docker/build-push-action@v6
         with:
           context: .
-          platforms: |
-            linux/amd64
-            linux/arm64
+          platforms: linux/amd64
           push: true
           build-args: |
             VERSION=${{ env.VERSION }}
             COMMIT=${{ env.COMMIT }}
             BUILD_DATE=${{ env.BUILD_DATE }}
           tags: |
-            ${{ env.DOCKERHUB_REPO }}:latest
-            ${{ env.DOCKERHUB_REPO }}:${{ env.VERSION }}
+            ${{ env.DOCKERHUB_REPO }}:latest-amd64
+            ${{ env.DOCKERHUB_REPO }}:${{ env.VERSION }}-amd64
+
+  docker_arm64:
+    runs-on: ubuntu-24.04-arm
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+      - name: Generate Build Metadata
+        run: |
+          echo VERSION=`git describe --tags --always --dirty` >> $GITHUB_ENV
+          echo COMMIT=`git rev-parse --short HEAD` >> $GITHUB_ENV
+          echo BUILD_DATE=`date -u +%Y-%m-%dT%H:%M:%SZ` >> $GITHUB_ENV
+      - name: Build and push (arm64)
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          platforms: linux/arm64
+          push: true
+          build-args: |
+            VERSION=${{ env.VERSION }}
+            COMMIT=${{ env.COMMIT }}
+            BUILD_DATE=${{ env.BUILD_DATE }}
+          tags: |
+            ${{ env.DOCKERHUB_REPO }}:latest-arm64
+            ${{ env.DOCKERHUB_REPO }}:${{ env.VERSION }}-arm64
+
+  docker_manifest:
+    runs-on: ubuntu-latest
+    needs:
+      - docker_amd64
+      - docker_arm64
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+      - name: Generate Build Metadata
+        run: |
+          echo VERSION=`git describe --tags --always --dirty` >> $GITHUB_ENV
+          echo COMMIT=`git rev-parse --short HEAD` >> $GITHUB_ENV
+          echo BUILD_DATE=`date -u +%Y-%m-%dT%H:%M:%SZ` >> $GITHUB_ENV
+      - name: Create and push multi-arch manifests
+        run: |
+          docker buildx imagetools create \
+            --tag "${DOCKERHUB_REPO}:latest" \
+            "${DOCKERHUB_REPO}:latest-amd64" \
+            "${DOCKERHUB_REPO}:latest-arm64"
+          docker buildx imagetools create \
+            --tag "${DOCKERHUB_REPO}:${VERSION}" \
+            "${DOCKERHUB_REPO}:${VERSION}-amd64" \
+            "${DOCKERHUB_REPO}:${VERSION}-arm64"
+      - name: Cleanup temporary tags
+        continue-on-error: true
+        env:
+          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_USERNAME }}
+          DOCKERHUB_TOKEN: ${{ secrets.DOCKERHUB_TOKEN }}
+        run: |
+          set -euo pipefail
+          namespace="${DOCKERHUB_REPO%%/*}"
+          repo_name="${DOCKERHUB_REPO#*/}"
+
+          token="$(
+            curl -fsSL \
+              -H 'Content-Type: application/json' \
+              -d "{\"username\":\"${DOCKERHUB_USERNAME}\",\"password\":\"${DOCKERHUB_TOKEN}\"}" \
+              'https://hub.docker.com/v2/users/login/' \
+              | python3 -c 'import json,sys; print(json.load(sys.stdin)["token"])'
+          )"
+
+          delete_tag() {
+            local tag="$1"
+            local url="https://hub.docker.com/v2/repositories/${namespace}/${repo_name}/tags/${tag}/"
+            local http_code
+            http_code="$(curl -sS -o /dev/null -w "%{http_code}" -X DELETE -H "Authorization: JWT ${token}" "${url}" || true)"
+            if [ "${http_code}" = "204" ] || [ "${http_code}" = "404" ]; then
+              echo "Docker Hub tag removed (or missing): ${DOCKERHUB_REPO}:${tag} (HTTP ${http_code})"
+              return 0
+            fi
+            echo "Docker Hub tag delete failed: ${DOCKERHUB_REPO}:${tag} (HTTP ${http_code})"
+            return 0
+          }
+
+          delete_tag "latest-amd64"
+          delete_tag "latest-arm64"
+          delete_tag "${VERSION}-amd64"
+          delete_tag "${VERSION}-arm64"

README.md

@@ -134,6 +134,10 @@ VSCode extension for quick switching between Claude Code models, featuring integ
 
 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.
 
+### [CPA-XXX Panel](https://github.com/ferretgeek/CPA-X)
+
+A lightweight web admin panel for CLIProxyAPI with health checks, resource monitoring, real-time logs, auto-update, request statistics and pricing display. Supports one-click installation and systemd service.
+
 > [!NOTE]  
 > If you developed a project based on CLIProxyAPI, please open a PR to add it to this list.
 

README_CN.md

@@ -133,6 +133,10 @@ CLI 封装器，用于通过 CLIProxyAPI OAuth 即时切换多个 Claude 账户
 
 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。
 
+### [CPA-XXX Panel](https://github.com/ferretgeek/CPA-X)
+
+面向 CLIProxyAPI 的 Web 管理面板，提供健康检查、资源监控、日志查看、自动更新、请求统计与定价展示，支持一键安装与 systemd 服务。
+
 > [!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。
 

config.example.yaml

@@ -285,24 +285,31 @@ oauth-model-alias:
 #   default: # Default rules only set parameters when they are missing in the payload.
 #     - 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
+#           protocol: "gemini" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex, antigravity
 #       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
+#           protocol: "gemini" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex, antigravity
 #       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
+#           protocol: "codex" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex, antigravity
 #       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
+#           protocol: "codex" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex, antigravity
 #       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\"}}}"
+#   filter: # Filter rules remove specified parameters from the payload.
+#     - 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, antigravity
+#       params: # JSON paths (gjson/sjson syntax) to remove from the payload
+#         - "generationConfig.thinkingConfig.thinkingBudget"
+#         - "generationConfig.responseJsonSchema"

go.mod

@@ -13,6 +13,7 @@ require (
 	github.com/joho/godotenv v1.5.1
 	github.com/klauspost/compress v1.17.4
 	github.com/minio/minio-go/v7 v7.0.66
+	github.com/refraction-networking/utls v1.8.2
 	github.com/sirupsen/logrus v1.9.3
 	github.com/skratchdot/open-golang v0.0.0-20200116055534-eef842397966
 	github.com/tidwall/gjson v1.18.0
@@ -21,7 +22,6 @@ require (
 	golang.org/x/crypto v0.45.0
 	golang.org/x/net v0.47.0
 	golang.org/x/oauth2 v0.30.0
-	golang.org/x/text v0.31.0
 	gopkg.in/natefinch/lumberjack.v2 v2.2.1
 	gopkg.in/yaml.v3 v3.0.1
 )
@@ -71,6 +71,7 @@ require (
 	golang.org/x/arch v0.8.0 // indirect
 	golang.org/x/sync v0.18.0 // indirect
 	golang.org/x/sys v0.38.0 // indirect
+	golang.org/x/text v0.31.0 // indirect
 	google.golang.org/protobuf v1.34.1 // indirect
 	gopkg.in/ini.v1 v1.67.0 // indirect
 )

go.sum

@@ -118,6 +118,8 @@ github.com/pjbgf/sha1cd v0.5.0 h1:a+UkboSi1znleCDUNT3M5YxjOnN1fz2FhN48FlwCxs0=
 github.com/pjbgf/sha1cd v0.5.0/go.mod h1:lhpGlyHLpQZoxMv8HcgXvZEhcGs0PG/vsZnEJ7H0iCM=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/refraction-networking/utls v1.8.2 h1:j4Q1gJj0xngdeH+Ox/qND11aEfhpgoEvV+S9iJ2IdQo=
+github.com/refraction-networking/utls v1.8.2/go.mod h1:jkSOEkLqn+S/jtpEHPOsVv/4V4EVnelwbMQl4vCWXAM=
 github.com/rogpeppe/go-internal v1.14.1 h1:UQB4HGPB6osV0SQTLymcB4TgvyWu6ZyliaW0tI/otEQ=
 github.com/rogpeppe/go-internal v1.14.1/go.mod h1:MaRKkUm5W0goXpeCfT7UZI6fk/L7L7so1lCWt35ZSgc=
 github.com/rs/xid v1.5.0 h1:mKX4bl4iPYJtEIxp6CYiUuLQ/8DYMoz0PUdtGgMFRVc=

internal/api/handlers/management/model_definitions.go

@@ -0,0 +1,33 @@
+package management
+
+import (
+	"net/http"
+	"strings"
+
+	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+)
+
+// GetStaticModelDefinitions returns static model metadata for a given channel.
+// Channel is provided via path param (:channel) or query param (?channel=...).
+func (h *Handler) GetStaticModelDefinitions(c *gin.Context) {
+	channel := strings.TrimSpace(c.Param("channel"))
+	if channel == "" {
+		channel = strings.TrimSpace(c.Query("channel"))
+	}
+	if channel == "" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "channel is required"})
+		return
+	}
+
+	models := registry.GetStaticModelDefinitionsByChannel(channel)
+	if models == nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "unknown channel", "channel": channel})
+		return
+	}
+
+	c.JSON(http.StatusOK, gin.H{
+		"channel": strings.ToLower(strings.TrimSpace(channel)),
+		"models":  models,
+	})
+}

internal/api/middleware/request_logging.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"net/http"
 	"strings"
+	"time"
 
 	"github.com/gin-gonic/gin"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/logging"
@@ -103,6 +104,7 @@ func captureRequestInfo(c *gin.Context) (*RequestInfo, error) {
 		Headers:   headers,
 		Body:      body,
 		RequestID: logging.GetGinRequestID(c),
+		Timestamp: time.Now(),
 	}, nil
 }
 

internal/api/middleware/response_writer.go

@@ -7,6 +7,7 @@ import (
 	"bytes"
 	"net/http"
 	"strings"
+	"time"
 
 	"github.com/gin-gonic/gin"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
@@ -20,22 +21,24 @@ type RequestInfo struct {
 	Headers   map[string][]string // Headers contains the request headers.
 	Body      []byte              // Body is the raw request body.
 	RequestID string              // RequestID is the unique identifier for the request.
+	Timestamp time.Time           // Timestamp is when the request was received.
 }
 
 // ResponseWriterWrapper wraps the standard gin.ResponseWriter to intercept and log response data.
 // It is designed to handle both standard and streaming responses, ensuring that logging operations do not block the client response.
 type ResponseWriterWrapper struct {
 	gin.ResponseWriter
-	body           *bytes.Buffer              // body is a buffer to store the response body for non-streaming responses.
-	isStreaming    bool                       // isStreaming indicates whether the response is a streaming type (e.g., text/event-stream).
-	streamWriter   logging.StreamingLogWriter // streamWriter is a writer for handling streaming log entries.
-	chunkChannel   chan []byte                // chunkChannel is a channel for asynchronously passing response chunks to the logger.
-	streamDone     chan struct{}              // streamDone signals when the streaming goroutine completes.
-	logger         logging.RequestLogger      // logger is the instance of the request logger service.
-	requestInfo    *RequestInfo               // requestInfo holds the details of the original request.
-	statusCode     int                        // statusCode stores the HTTP status code of the response.
-	headers        map[string][]string        // headers stores the response headers.
-	logOnErrorOnly bool                       // logOnErrorOnly enables logging only when an error response is detected.
+	body                *bytes.Buffer              // body is a buffer to store the response body for non-streaming responses.
+	isStreaming         bool                       // isStreaming indicates whether the response is a streaming type (e.g., text/event-stream).
+	streamWriter        logging.StreamingLogWriter // streamWriter is a writer for handling streaming log entries.
+	chunkChannel        chan []byte                // chunkChannel is a channel for asynchronously passing response chunks to the logger.
+	streamDone          chan struct{}              // streamDone signals when the streaming goroutine completes.
+	logger              logging.RequestLogger      // logger is the instance of the request logger service.
+	requestInfo         *RequestInfo               // requestInfo holds the details of the original request.
+	statusCode          int                        // statusCode stores the HTTP status code of the response.
+	headers             map[string][]string        // headers stores the response headers.
+	logOnErrorOnly      bool                       // logOnErrorOnly enables logging only when an error response is detected.
+	firstChunkTimestamp time.Time                  // firstChunkTimestamp captures TTFB for streaming responses.
 }
 
 // NewResponseWriterWrapper creates and initializes a new ResponseWriterWrapper.
@@ -73,6 +76,10 @@ func (w *ResponseWriterWrapper) Write(data []byte) (int, error) {
 
 	// THEN: Handle logging based on response type
 	if w.isStreaming && w.chunkChannel != nil {
+		// Capture TTFB on first chunk (synchronous, before async channel send)
+		if w.firstChunkTimestamp.IsZero() {
+			w.firstChunkTimestamp = time.Now()
+		}
 		// For streaming responses: Send to async logging channel (non-blocking)
 		select {
 		case w.chunkChannel <- append([]byte(nil), data...): // Non-blocking send with copy
@@ -117,6 +124,10 @@ func (w *ResponseWriterWrapper) WriteString(data string) (int, error) {
 
 	// THEN: Capture for logging
 	if w.isStreaming && w.chunkChannel != nil {
+		// Capture TTFB on first chunk (synchronous, before async channel send)
+		if w.firstChunkTimestamp.IsZero() {
+			w.firstChunkTimestamp = time.Now()
+		}
 		select {
 		case w.chunkChannel <- []byte(data):
 		default:
@@ -280,6 +291,8 @@ func (w *ResponseWriterWrapper) Finalize(c *gin.Context) error {
 			w.streamDone = nil
 		}
 
+		w.streamWriter.SetFirstChunkTimestamp(w.firstChunkTimestamp)
+
 		// Write API Request and Response to the streaming log before closing
 		apiRequest := w.extractAPIRequest(c)
 		if len(apiRequest) > 0 {
@@ -297,7 +310,7 @@ func (w *ResponseWriterWrapper) Finalize(c *gin.Context) error {
 		return nil
 	}
 
-	return w.logRequest(finalStatusCode, w.cloneHeaders(), w.body.Bytes(), w.extractAPIRequest(c), w.extractAPIResponse(c), slicesAPIResponseError, forceLog)
+	return w.logRequest(finalStatusCode, w.cloneHeaders(), w.body.Bytes(), w.extractAPIRequest(c), w.extractAPIResponse(c), w.extractAPIResponseTimestamp(c), slicesAPIResponseError, forceLog)
 }
 
 func (w *ResponseWriterWrapper) cloneHeaders() map[string][]string {
@@ -337,7 +350,18 @@ func (w *ResponseWriterWrapper) extractAPIResponse(c *gin.Context) []byte {
 	return data
 }
 
-func (w *ResponseWriterWrapper) logRequest(statusCode int, headers map[string][]string, body []byte, apiRequestBody, apiResponseBody []byte, apiResponseErrors []*interfaces.ErrorMessage, forceLog bool) error {
+func (w *ResponseWriterWrapper) extractAPIResponseTimestamp(c *gin.Context) time.Time {
+	ts, isExist := c.Get("API_RESPONSE_TIMESTAMP")
+	if !isExist {
+		return time.Time{}
+	}
+	if t, ok := ts.(time.Time); ok {
+		return t
+	}
+	return time.Time{}
+}
+
+func (w *ResponseWriterWrapper) logRequest(statusCode int, headers map[string][]string, body []byte, apiRequestBody, apiResponseBody []byte, apiResponseTimestamp time.Time, apiResponseErrors []*interfaces.ErrorMessage, forceLog bool) error {
 	if w.requestInfo == nil {
 		return nil
 	}
@@ -348,7 +372,7 @@ func (w *ResponseWriterWrapper) logRequest(statusCode int, headers map[string][]
 	}
 
 	if loggerWithOptions, ok := w.logger.(interface {
-		LogRequestWithOptions(string, string, map[string][]string, []byte, int, map[string][]string, []byte, []byte, []byte, []*interfaces.ErrorMessage, bool, string) error
+		LogRequestWithOptions(string, string, map[string][]string, []byte, int, map[string][]string, []byte, []byte, []byte, []*interfaces.ErrorMessage, bool, string, time.Time, time.Time) error
 	}); ok {
 		return loggerWithOptions.LogRequestWithOptions(
 			w.requestInfo.URL,
@@ -363,6 +387,8 @@ func (w *ResponseWriterWrapper) logRequest(statusCode int, headers map[string][]
 			apiResponseErrors,
 			forceLog,
 			w.requestInfo.RequestID,
+			w.requestInfo.Timestamp,
+			apiResponseTimestamp,
 		)
 	}
 
@@ -378,5 +404,7 @@ func (w *ResponseWriterWrapper) logRequest(statusCode int, headers map[string][]
 		apiResponseBody,
 		apiResponseErrors,
 		w.requestInfo.RequestID,
+		w.requestInfo.Timestamp,
+		apiResponseTimestamp,
 	)
 }

internal/api/server.go

@@ -12,6 +12,7 @@ import (
 	"net/http"
 	"os"
 	"path/filepath"
+	"reflect"
 	"strings"
 	"sync"
 	"sync/atomic"
@@ -325,6 +326,7 @@ func (s *Server) setupRoutes() {
 		v1.POST("/messages", claudeCodeHandlers.ClaudeMessages)
 		v1.POST("/messages/count_tokens", claudeCodeHandlers.ClaudeCountTokens)
 		v1.POST("/responses", openaiResponsesHandlers.Responses)
+		v1.POST("/responses/compact", openaiResponsesHandlers.Compact)
 	}
 
 	// Gemini compatible API routes
@@ -607,6 +609,7 @@ func (s *Server) registerManagementRoutes() {
 
 		mgmt.GET("/auth-files", s.mgmt.ListAuthFiles)
 		mgmt.GET("/auth-files/models", s.mgmt.GetAuthFileModels)
+		mgmt.GET("/model-definitions/:channel", s.mgmt.GetStaticModelDefinitions)
 		mgmt.GET("/auth-files/download", s.mgmt.DownloadAuthFile)
 		mgmt.POST("/auth-files", s.mgmt.UploadAuthFile)
 		mgmt.DELETE("/auth-files", s.mgmt.DeleteAuthFile)
@@ -989,14 +992,17 @@ func (s *Server) UpdateClients(cfg *config.Config) {
 		s.mgmt.SetAuthManager(s.handlers.AuthManager)
 	}
 
-	// Notify Amp module of config changes (for model mapping hot-reload)
-	if s.ampModule != nil {
-		log.Debugf("triggering amp module config update")
-		if err := s.ampModule.OnConfigUpdated(cfg); err != nil {
-			log.Errorf("failed to update Amp module config: %v", err)
+	// Notify Amp module only when Amp config has changed.
+	ampConfigChanged := oldCfg == nil || !reflect.DeepEqual(oldCfg.AmpCode, cfg.AmpCode)
+	if ampConfigChanged {
+		if s.ampModule != nil {
+			log.Debugf("triggering amp module config update")
+			if err := s.ampModule.OnConfigUpdated(cfg); err != nil {
+				log.Errorf("failed to update Amp module config: %v", err)
+			}
+		} else {
+			log.Warnf("amp module is nil, skipping config update")
 		}
-	} else {
-		log.Warnf("amp module is nil, skipping config update")
 	}
 
 	// Count client sources from configuration and auth store.

internal/auth/claude/anthropic_auth.go

@@ -14,7 +14,6 @@ import (
 	"time"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	log "github.com/sirupsen/logrus"
 )
 
@@ -51,7 +50,8 @@ type ClaudeAuth struct {
 }
 
 // NewClaudeAuth creates a new Anthropic authentication service.
-// It initializes the HTTP client with proxy settings from the configuration.
+// It initializes the HTTP client with a custom TLS transport that uses Firefox
+// fingerprint to bypass Cloudflare's TLS fingerprinting on Anthropic domains.
 //
 // Parameters:
 //   - cfg: The application configuration containing proxy settings
@@ -59,8 +59,10 @@ type ClaudeAuth struct {
 // Returns:
 //   - *ClaudeAuth: A new Claude authentication service instance
 func NewClaudeAuth(cfg *config.Config) *ClaudeAuth {
+	// Use custom HTTP client with Firefox TLS fingerprint to bypass
+	// Cloudflare's bot detection on Anthropic domains
 	return &ClaudeAuth{
-		httpClient: util.SetProxy(&cfg.SDKConfig, &http.Client{}),
+		httpClient: NewAnthropicHttpClient(&cfg.SDKConfig),
 	}
 }
 

internal/auth/claude/utls_transport.go

@@ -0,0 +1,165 @@
+// Package claude provides authentication functionality for Anthropic's Claude API.
+// This file implements a custom HTTP transport using utls to bypass TLS fingerprinting.
+package claude
+
+import (
+	"net/http"
+	"net/url"
+	"strings"
+	"sync"
+
+	tls "github.com/refraction-networking/utls"
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/net/http2"
+	"golang.org/x/net/proxy"
+)
+
+// utlsRoundTripper implements http.RoundTripper using utls with Firefox fingerprint
+// to bypass Cloudflare's TLS fingerprinting on Anthropic domains.
+type utlsRoundTripper struct {
+	// mu protects the connections map and pending map
+	mu sync.Mutex
+	// connections caches HTTP/2 client connections per host
+	connections map[string]*http2.ClientConn
+	// pending tracks hosts that are currently being connected to (prevents race condition)
+	pending map[string]*sync.Cond
+	// dialer is used to create network connections, supporting proxies
+	dialer proxy.Dialer
+}
+
+// newUtlsRoundTripper creates a new utls-based round tripper with optional proxy support
+func newUtlsRoundTripper(cfg *config.SDKConfig) *utlsRoundTripper {
+	var dialer proxy.Dialer = proxy.Direct
+	if cfg != nil && cfg.ProxyURL != "" {
+		proxyURL, err := url.Parse(cfg.ProxyURL)
+		if err != nil {
+			log.Errorf("failed to parse proxy URL %q: %v", cfg.ProxyURL, err)
+		} else {
+			pDialer, err := proxy.FromURL(proxyURL, proxy.Direct)
+			if err != nil {
+				log.Errorf("failed to create proxy dialer for %q: %v", cfg.ProxyURL, err)
+			} else {
+				dialer = pDialer
+			}
+		}
+	}
+
+	return &utlsRoundTripper{
+		connections: make(map[string]*http2.ClientConn),
+		pending:     make(map[string]*sync.Cond),
+		dialer:      dialer,
+	}
+}
+
+// getOrCreateConnection gets an existing connection or creates a new one.
+// It uses a per-host locking mechanism to prevent multiple goroutines from
+// creating connections to the same host simultaneously.
+func (t *utlsRoundTripper) getOrCreateConnection(host, addr string) (*http2.ClientConn, error) {
+	t.mu.Lock()
+
+	// Check if connection exists and is usable
+	if h2Conn, ok := t.connections[host]; ok && h2Conn.CanTakeNewRequest() {
+		t.mu.Unlock()
+		return h2Conn, nil
+	}
+
+	// Check if another goroutine is already creating a connection
+	if cond, ok := t.pending[host]; ok {
+		// Wait for the other goroutine to finish
+		cond.Wait()
+		// Check if connection is now available
+		if h2Conn, ok := t.connections[host]; ok && h2Conn.CanTakeNewRequest() {
+			t.mu.Unlock()
+			return h2Conn, nil
+		}
+		// Connection still not available, we'll create one
+	}
+
+	// Mark this host as pending
+	cond := sync.NewCond(&t.mu)
+	t.pending[host] = cond
+	t.mu.Unlock()
+
+	// Create connection outside the lock
+	h2Conn, err := t.createConnection(host, addr)
+
+	t.mu.Lock()
+	defer t.mu.Unlock()
+
+	// Remove pending marker and wake up waiting goroutines
+	delete(t.pending, host)
+	cond.Broadcast()
+
+	if err != nil {
+		return nil, err
+	}
+
+	// Store the new connection
+	t.connections[host] = h2Conn
+	return h2Conn, nil
+}
+
+// createConnection creates a new HTTP/2 connection with Firefox TLS fingerprint
+func (t *utlsRoundTripper) createConnection(host, addr string) (*http2.ClientConn, error) {
+	conn, err := t.dialer.Dial("tcp", addr)
+	if err != nil {
+		return nil, err
+	}
+
+	tlsConfig := &tls.Config{ServerName: host}
+	tlsConn := tls.UClient(conn, tlsConfig, tls.HelloFirefox_Auto)
+
+	if err := tlsConn.Handshake(); err != nil {
+		conn.Close()
+		return nil, err
+	}
+
+	tr := &http2.Transport{}
+	h2Conn, err := tr.NewClientConn(tlsConn)
+	if err != nil {
+		tlsConn.Close()
+		return nil, err
+	}
+
+	return h2Conn, nil
+}
+
+// RoundTrip implements http.RoundTripper
+func (t *utlsRoundTripper) RoundTrip(req *http.Request) (*http.Response, error) {
+	host := req.URL.Host
+	addr := host
+	if !strings.Contains(addr, ":") {
+		addr += ":443"
+	}
+
+	// Get hostname without port for TLS ServerName
+	hostname := req.URL.Hostname()
+
+	h2Conn, err := t.getOrCreateConnection(hostname, addr)
+	if err != nil {
+		return nil, err
+	}
+
+	resp, err := h2Conn.RoundTrip(req)
+	if err != nil {
+		// Connection failed, remove it from cache
+		t.mu.Lock()
+		if cached, ok := t.connections[hostname]; ok && cached == h2Conn {
+			delete(t.connections, hostname)
+		}
+		t.mu.Unlock()
+		return nil, err
+	}
+
+	return resp, nil
+}
+
+// NewAnthropicHttpClient creates an HTTP client that bypasses TLS fingerprinting
+// for Anthropic domains by using utls with Firefox fingerprint.
+// It accepts optional SDK configuration for proxy settings.
+func NewAnthropicHttpClient(cfg *config.SDKConfig) *http.Client {
+	return &http.Client{
+		Transport: newUtlsRoundTripper(cfg),
+	}
+}

internal/config/config.go

@@ -229,6 +229,16 @@ type PayloadConfig struct {
 	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"`
+	// Filter defines rules that remove parameters from the payload by JSON path.
+	Filter []PayloadFilterRule `yaml:"filter" json:"filter"`
+}
+
+// PayloadFilterRule describes a rule to remove specific JSON paths from matching model payloads.
+type PayloadFilterRule struct {
+	// Models lists model entries with name pattern and protocol constraint.
+	Models []PayloadModelRule `yaml:"models" json:"models"`
+	// Params lists JSON paths (gjson/sjson syntax) to remove from the payload.
+	Params []string `yaml:"params" json:"params"`
 }
 
 // PayloadRule describes a single rule targeting a list of models with parameter updates.
@@ -923,6 +933,7 @@ func SaveConfigPreserveComments(configFile string, cfg *Config) error {
 	removeLegacyGenerativeLanguageKeys(original.Content[0])
 
 	pruneMappingToGeneratedKeys(original.Content[0], generated.Content[0], "oauth-excluded-models")
+	pruneMappingToGeneratedKeys(original.Content[0], generated.Content[0], "oauth-model-alias")
 
 	// Merge generated into original in-place, preserving comments/order of existing nodes.
 	mergeMappingPreserve(original.Content[0], generated.Content[0])
@@ -1413,6 +1424,16 @@ func pruneMappingToGeneratedKeys(dstRoot, srcRoot *yaml.Node, key string) {
 	}
 	srcIdx := findMapKeyIndex(srcRoot, key)
 	if srcIdx < 0 {
+		// Keep an explicit empty mapping for oauth-model-alias when it was previously present.
+		//
+		// Rationale: LoadConfig runs MigrateOAuthModelAlias before unmarshalling. If the
+		// oauth-model-alias key is missing, migration will add the default antigravity aliases.
+		// When users delete the last channel from oauth-model-alias via the management API,
+		// we want that deletion to persist across hot reloads and restarts.
+		if key == "oauth-model-alias" {
+			dstRoot.Content[dstIdx+1] = &yaml.Node{Kind: yaml.MappingNode, Tag: "!!map"}
+			return
+		}
 		removeMapKey(dstRoot, key)
 		return
 	}

internal/logging/request_logger.go

@@ -44,10 +44,12 @@ type RequestLogger interface {
 	//   - apiRequest: The API request data
 	//   - apiResponse: The API response data
 	//   - requestID: Optional request ID for log file naming
+	//   - requestTimestamp: When the request was received
+	//   - apiResponseTimestamp: When the API response was received
 	//
 	// Returns:
 	//   - error: An error if logging fails, nil otherwise
-	LogRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, requestID string) error
+	LogRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, requestID string, requestTimestamp, apiResponseTimestamp time.Time) error
 
 	// LogStreamingRequest initiates logging for a streaming request and returns a writer for chunks.
 	//
@@ -109,6 +111,12 @@ type StreamingLogWriter interface {
 	//   - error: An error if writing fails, nil otherwise
 	WriteAPIResponse(apiResponse []byte) error
 
+	// SetFirstChunkTimestamp sets the TTFB timestamp captured when first chunk was received.
+	//
+	// Parameters:
+	//   - timestamp: The time when first response chunk was received
+	SetFirstChunkTimestamp(timestamp time.Time)
+
 	// Close finalizes the log file and cleans up resources.
 	//
 	// Returns:
@@ -180,20 +188,22 @@ func (l *FileRequestLogger) SetEnabled(enabled bool) {
 //   - apiRequest: The API request data
 //   - apiResponse: The API response data
 //   - requestID: Optional request ID for log file naming
+//   - requestTimestamp: When the request was received
+//   - apiResponseTimestamp: When the API response was received
 //
 // Returns:
 //   - error: An error if logging fails, nil otherwise
-func (l *FileRequestLogger) LogRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, requestID string) error {
-	return l.logRequest(url, method, requestHeaders, body, statusCode, responseHeaders, response, apiRequest, apiResponse, apiResponseErrors, false, requestID)
+func (l *FileRequestLogger) LogRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, requestID string, requestTimestamp, apiResponseTimestamp time.Time) error {
+	return l.logRequest(url, method, requestHeaders, body, statusCode, responseHeaders, response, apiRequest, apiResponse, apiResponseErrors, false, requestID, requestTimestamp, apiResponseTimestamp)
 }
 
 // LogRequestWithOptions logs a request with optional forced logging behavior.
 // The force flag allows writing error logs even when regular request logging is disabled.
-func (l *FileRequestLogger) LogRequestWithOptions(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, force bool, requestID string) error {
-	return l.logRequest(url, method, requestHeaders, body, statusCode, responseHeaders, response, apiRequest, apiResponse, apiResponseErrors, force, requestID)
+func (l *FileRequestLogger) LogRequestWithOptions(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, force bool, requestID string, requestTimestamp, apiResponseTimestamp time.Time) error {
+	return l.logRequest(url, method, requestHeaders, body, statusCode, responseHeaders, response, apiRequest, apiResponse, apiResponseErrors, force, requestID, requestTimestamp, apiResponseTimestamp)
 }
 
-func (l *FileRequestLogger) logRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, force bool, requestID string) error {
+func (l *FileRequestLogger) logRequest(url, method string, requestHeaders map[string][]string, body []byte, statusCode int, responseHeaders map[string][]string, response, apiRequest, apiResponse []byte, apiResponseErrors []*interfaces.ErrorMessage, force bool, requestID string, requestTimestamp, apiResponseTimestamp time.Time) error {
 	if !l.enabled && !force {
 		return nil
 	}
@@ -247,6 +257,8 @@ func (l *FileRequestLogger) logRequest(url, method string, requestHeaders map[st
 		responseHeaders,
 		responseToWrite,
 		decompressErr,
+		requestTimestamp,
+		apiResponseTimestamp,
 	)
 	if errClose := logFile.Close(); errClose != nil {
 		log.WithError(errClose).Warn("failed to close request log file")
@@ -499,17 +511,22 @@ func (l *FileRequestLogger) writeNonStreamingLog(
 	responseHeaders map[string][]string,
 	response []byte,
 	decompressErr error,
+	requestTimestamp time.Time,
+	apiResponseTimestamp time.Time,
 ) error {
-	if errWrite := writeRequestInfoWithBody(w, url, method, requestHeaders, requestBody, requestBodyPath, time.Now()); errWrite != nil {
+	if requestTimestamp.IsZero() {
+		requestTimestamp = time.Now()
+	}
+	if errWrite := writeRequestInfoWithBody(w, url, method, requestHeaders, requestBody, requestBodyPath, requestTimestamp); errWrite != nil {
 		return errWrite
 	}
-	if errWrite := writeAPISection(w, "=== API REQUEST ===\n", "=== API REQUEST", apiRequest); errWrite != nil {
+	if errWrite := writeAPISection(w, "=== API REQUEST ===\n", "=== API REQUEST", apiRequest, time.Time{}); errWrite != nil {
 		return errWrite
 	}
 	if errWrite := writeAPIErrorResponses(w, apiResponseErrors); errWrite != nil {
 		return errWrite
 	}
-	if errWrite := writeAPISection(w, "=== API RESPONSE ===\n", "=== API RESPONSE", apiResponse); errWrite != nil {
+	if errWrite := writeAPISection(w, "=== API RESPONSE ===\n", "=== API RESPONSE", apiResponse, apiResponseTimestamp); errWrite != nil {
 		return errWrite
 	}
 	return writeResponseSection(w, statusCode, true, responseHeaders, bytes.NewReader(response), decompressErr, true)
@@ -583,7 +600,7 @@ func writeRequestInfoWithBody(
 	return nil
 }
 
-func writeAPISection(w io.Writer, sectionHeader string, sectionPrefix string, payload []byte) error {
+func writeAPISection(w io.Writer, sectionHeader string, sectionPrefix string, payload []byte, timestamp time.Time) error {
 	if len(payload) == 0 {
 		return nil
 	}
@@ -601,6 +618,11 @@ func writeAPISection(w io.Writer, sectionHeader string, sectionPrefix string, pa
 		if _, errWrite := io.WriteString(w, sectionHeader); errWrite != nil {
 			return errWrite
 		}
+		if !timestamp.IsZero() {
+			if _, errWrite := io.WriteString(w, fmt.Sprintf("Timestamp: %s\n", timestamp.Format(time.RFC3339Nano))); errWrite != nil {
+				return errWrite
+			}
+		}
 		if _, errWrite := w.Write(payload); errWrite != nil {
 			return errWrite
 		}
@@ -974,6 +996,9 @@ type FileStreamingLogWriter struct {
 
 	// apiResponse stores the upstream API response data.
 	apiResponse []byte
+
+	// apiResponseTimestamp captures when the API response was received.
+	apiResponseTimestamp time.Time
 }
 
 // WriteChunkAsync writes a response chunk asynchronously (non-blocking).
@@ -1053,6 +1078,12 @@ func (w *FileStreamingLogWriter) WriteAPIResponse(apiResponse []byte) error {
 	return nil
 }
 
+func (w *FileStreamingLogWriter) SetFirstChunkTimestamp(timestamp time.Time) {
+	if !timestamp.IsZero() {
+		w.apiResponseTimestamp = timestamp
+	}
+}
+
 // Close finalizes the log file and cleans up resources.
 // It writes all buffered data to the file in the correct order:
 // API REQUEST -> API RESPONSE -> RESPONSE (status, headers, body chunks)
@@ -1140,10 +1171,10 @@ func (w *FileStreamingLogWriter) writeFinalLog(logFile *os.File) error {
 	if errWrite := writeRequestInfoWithBody(logFile, w.url, w.method, w.requestHeaders, nil, w.requestBodyPath, w.timestamp); errWrite != nil {
 		return errWrite
 	}
-	if errWrite := writeAPISection(logFile, "=== API REQUEST ===\n", "=== API REQUEST", w.apiRequest); errWrite != nil {
+	if errWrite := writeAPISection(logFile, "=== API REQUEST ===\n", "=== API REQUEST", w.apiRequest, time.Time{}); errWrite != nil {
 		return errWrite
 	}
-	if errWrite := writeAPISection(logFile, "=== API RESPONSE ===\n", "=== API RESPONSE", w.apiResponse); errWrite != nil {
+	if errWrite := writeAPISection(logFile, "=== API RESPONSE ===\n", "=== API RESPONSE", w.apiResponse, w.apiResponseTimestamp); errWrite != nil {
 		return errWrite
 	}
 
@@ -1220,6 +1251,8 @@ func (w *NoOpStreamingLogWriter) WriteAPIResponse(_ []byte) error {
 	return nil
 }
 
+func (w *NoOpStreamingLogWriter) SetFirstChunkTimestamp(_ time.Time) {}
+
 // Close is a no-op implementation that does nothing and always returns nil.
 //
 // Returns:

internal/misc/codex_instructions.go

@@ -36,7 +36,7 @@ var opencodeCodexInstructions string
 
 const (
 	codexUserAgentKey  = "__cpa_user_agent"
-	userAgentOpenAISDK = "ai-sdk/openai/"
+	userAgentOpenAISDK = "opencode/"
 )
 
 func InjectCodexUserAgent(raw []byte, userAgent string) []byte {

internal/misc/opencode_codex_instructions.txt

@@ -1,318 +1,79 @@
-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.
+You are OpenCode, the best coding agent on the planet.
+
+You are an interactive CLI tool that helps users with software engineering tasks. Use the instructions below and the tools available to you to assist the user.
+
+## Editing constraints
+- Default to ASCII when editing or creating files. Only introduce non-ASCII or other Unicode characters when there is a clear justification and the file already uses them.
+- Only add comments if they are necessary to make a non-obvious block easier to understand.
+- Try to use apply_patch for single file edits, but it is fine to explore other options to make the edit if it does not work well. Do not use apply_patch for changes that are auto-generated (i.e. generating package.json or running a lint or format command like gofmt) or when scripting is more efficient (such as search and replacing a string across a codebase).
+
+## Tool usage
+- Prefer specialized tools over shell for file operations:
+  - Use Read to view files, Edit to modify files, and Write only when needed.
+  - Use Glob to find files by name and Grep to search file contents.
+- Use Bash for terminal operations (git, bun, builds, tests, running scripts).
+- Run tool calls in parallel when neither call needs the other’s output; otherwise run sequentially.
+
+## Git and workspace hygiene
+- You may be in a dirty git worktree.
+    * NEVER revert existing changes you did not make unless explicitly requested, since these changes were made by the user.
+    * If asked to make a commit or code edits and there are unrelated changes to your work or changes that you didn't make in those files, don't revert those changes.
+    * If the changes are in files you've touched recently, you should read carefully and understand how you can work with the changes rather than reverting them.
+    * If the changes are in unrelated files, just ignore them and don't revert them.
+- Do not amend commits unless explicitly requested.
+- **NEVER** use destructive commands like `git reset --hard` or `git checkout --` unless specifically requested or approved by the user.
+
+## Frontend tasks
+When doing frontend design tasks, avoid collapsing into bland, generic layouts.
+Aim for interfaces that feel intentional and deliberate.
+- Typography: Use expressive, purposeful fonts and avoid default stacks (Inter, Roboto, Arial, system).
+- Color & Look: Choose a clear visual direction; define CSS variables; avoid purple-on-white defaults. No purple bias or dark mode bias.
+- Motion: Use a few meaningful animations (page-load, staggered reveals) instead of generic micro-motions.
+- Background: Don't rely on flat, single-color backgrounds; use gradients, shapes, or subtle patterns to build atmosphere.
+- Overall: Avoid boilerplate layouts and interchangeable UI patterns. Vary themes, type families, and visual languages across outputs.
+- Ensure the page loads properly on both desktop and mobile.
+
+Exception: If working within an existing website or design system, preserve the established patterns, structure, and visual language.
 
 ## Presenting your work and final message
 
-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**
+- Default: be very concise; friendly coding teammate tone.
+- Default: do the work without asking questions. Treat short tasks as sufficient direction; infer missing details by reading the codebase and following existing conventions.
+- Questions: only ask when you are truly blocked after checking relevant context AND you cannot safely pick a reasonable default. This usually means one of:
+  * The request is ambiguous in a way that materially changes the result and you cannot disambiguate by reading the repo.
+  * The action is destructive/irreversible, touches production, or changes billing/security posture.
+  * You need a secret/credential/value that cannot be inferred (API key, account id, etc.).
+- If you must ask: do all non-blocked work first, then ask exactly one targeted question, include your recommended default, and state what would change based on the answer.
+- Never ask permission questions like "Should I proceed?" or "Do you want me to run tests?"; proceed with the most reasonable option and mention what you did.
+- For substantial work, summarize clearly; follow final‑answer formatting.
+- Skip heavy formatting for simple confirmations.
+- Don't dump large files you've written; reference paths only.
+- No "save/copy this file" - User is on the same machine.
+- Offer logical next steps (tests, commits, build) briefly; add verify steps if you couldn't do something.
+- For code changes:
+  * Lead with a quick explanation of the change, and then give more details on the context covering where and why a change was made. Do not start this explanation with "summary", just jump right in.
+  * If there are natural next steps the user may want to take, suggest them at the end of your response. Do not make suggestions if there are no natural next steps.
+  * When suggesting multiple options, use numeric lists for the suggestions so the user can quickly respond with a single number.
+- The user does not command execution outputs. When asked to show the output of a command (e.g. `git show`), relay the important details in your answer or summarize the key lines so the user understands the result.
 
-- 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.
+## Final answer structure and style guidelines
 
-**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:
+- Plain text; CLI handles styling. Use structure only when it helps scanability.
+- Headers: optional; short Title Case (1-3 words) wrapped in **…**; no blank line before the first bullet; add only if they truly help.
+- Bullets: use - ; merge related points; keep to one line when possible; 4–6 per list ordered by importance; keep phrasing consistent.
+- Monospace: backticks for commands/paths/env vars/code ids and inline examples; use for literal keyword bullets; never combine with **.
+- Code samples or multi-line snippets should be wrapped in fenced code blocks; include an info string as often as possible.
+- Structure: group related bullets; order sections general → specific → supporting; for subsections, start with a bolded keyword bullet, then items; match complexity to the task.
+- Tone: collaborative, concise, factual; present tense, active voice; self‑contained; no "above/below"; parallel wording.
+- Don'ts: no nested bullets/hierarchies; no ANSI codes; don't cram unrelated keywords; keep keyword lists short—wrap/reformat if long; avoid naming formatting styles in answers.
+- Adaptation: code explanations → precise, structured with code refs; simple tasks → lead with outcome; big changes → logical walkthrough + rationale + next actions; casual one-offs → plain sentences, no headers/bullets.
+- File References: When referencing files in your response follow the below rules:
   * Use inline code to make file paths clickable.
-  * Each reference should have a standalone path. Even if it's the same file.
+  * Each reference should have a stand alone path. Even if it's the same file.
   * Accepted: absolute, workspace‑relative, a/ or b/ diff prefixes, or bare filename/suffix.
-  * Line/column (1‑based, optional): :line[:column] or #Lline[Ccolumn] (column defaults to 1).
+  * Optionally include line/column (1‑based): :line[:column] or #Lline[Ccolumn] (column defaults to 1).
   * Do not use URIs like file://, vscode://, or https://.
   * Do not provide range of lines
   * Examples: src/app.ts, src/app.ts:42, b/server/index.js#L10, C:\repo\project\main.rs:12:5
-
-**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

@@ -1,848 +1,69 @@
-// Package registry provides model definitions for various AI service providers.
-// This file contains static model definitions that can be used by clients
-// when registering their supported models.
+// Package registry provides model definitions and lookup helpers for various AI providers.
+// Static model metadata is stored in model_definitions_static_data.go.
 package registry
 
-// GetClaudeModels returns the standard Claude model definitions
-func GetClaudeModels() []*ModelInfo {
-	return []*ModelInfo{
+import (
+	"sort"
+	"strings"
+)
 
-		{
-			ID:                  "claude-haiku-4-5-20251001",
-			Object:              "model",
-			Created:             1759276800, // 2025-10-01
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 4.5 Haiku",
-			ContextLength:       200000,
-			MaxCompletionTokens: 64000,
-			// Thinking: not supported for Haiku models
-		},
-		{
-			ID:                  "claude-sonnet-4-5-20250929",
-			Object:              "model",
-			Created:             1759104000, // 2025-09-29
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 4.5 Sonnet",
-			ContextLength:       200000,
-			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
-		},
-		{
-			ID:                  "claude-opus-4-5-20251101",
-			Object:              "model",
-			Created:             1761955200, // 2025-11-01
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 4.5 Opus",
-			Description:         "Premium model combining maximum intelligence with practical performance",
-			ContextLength:       200000,
-			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
-		},
-		{
-			ID:                  "claude-opus-4-1-20250805",
-			Object:              "model",
-			Created:             1722945600, // 2025-08-05
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 4.1 Opus",
-			ContextLength:       200000,
-			MaxCompletionTokens: 32000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
-		},
-		{
-			ID:                  "claude-opus-4-20250514",
-			Object:              "model",
-			Created:             1715644800, // 2025-05-14
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 4 Opus",
-			ContextLength:       200000,
-			MaxCompletionTokens: 32000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
-		},
-		{
-			ID:                  "claude-sonnet-4-20250514",
-			Object:              "model",
-			Created:             1715644800, // 2025-05-14
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 4 Sonnet",
-			ContextLength:       200000,
-			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
-		},
-		{
-			ID:                  "claude-3-7-sonnet-20250219",
-			Object:              "model",
-			Created:             1708300800, // 2025-02-19
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 3.7 Sonnet",
-			ContextLength:       128000,
-			MaxCompletionTokens: 8192,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
-		},
-		{
-			ID:                  "claude-3-5-haiku-20241022",
-			Object:              "model",
-			Created:             1729555200, // 2024-10-22
-			OwnedBy:             "anthropic",
-			Type:                "claude",
-			DisplayName:         "Claude 3.5 Haiku",
-			ContextLength:       128000,
-			MaxCompletionTokens: 8192,
-			// Thinking: not supported for Haiku models
-		},
-	}
-}
-
-// GetGeminiModels returns the standard Gemini model definitions
-func GetGeminiModels() []*ModelInfo {
-	return []*ModelInfo{
-		{
-			ID:                         "gemini-2.5-pro",
-			Object:                     "model",
-			Created:                    1750118400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-pro",
-			Version:                    "2.5",
-			DisplayName:                "Gemini 2.5 Pro",
-			Description:                "Stable release (June 17th, 2025) of Gemini 2.5 Pro",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-2.5-flash",
-			Object:                     "model",
-			Created:                    1750118400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-flash",
-			Version:                    "001",
-			DisplayName:                "Gemini 2.5 Flash",
-			Description:                "Stable version of Gemini 2.5 Flash, our mid-size multimodal model that supports up to 1 million tokens, released in June of 2025.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-2.5-flash-lite",
-			Object:                     "model",
-			Created:                    1753142400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-flash-lite",
-			Version:                    "2.5",
-			DisplayName:                "Gemini 2.5 Flash Lite",
-			Description:                "Our smallest and most cost effective model, built for at scale usage.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-3-pro-preview",
-			Object:                     "model",
-			Created:                    1737158400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-3-pro-preview",
-			Version:                    "3.0",
-			DisplayName:                "Gemini 3 Pro Preview",
-			Description:                "Gemini 3 Pro Preview",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
-		},
-		{
-			ID:                         "gemini-3-flash-preview",
-			Object:                     "model",
-			Created:                    1765929600,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-3-flash-preview",
-			Version:                    "3.0",
-			DisplayName:                "Gemini 3 Flash Preview",
-			Description:                "Gemini 3 Flash Preview",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
-		},
-		{
-			ID:                         "gemini-3-pro-image-preview",
-			Object:                     "model",
-			Created:                    1737158400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-3-pro-image-preview",
-			Version:                    "3.0",
-			DisplayName:                "Gemini 3 Pro Image Preview",
-			Description:                "Gemini 3 Pro Image Preview",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
-		},
-	}
-}
-
-func GetGeminiVertexModels() []*ModelInfo {
-	return []*ModelInfo{
-		{
-			ID:                         "gemini-2.5-pro",
-			Object:                     "model",
-			Created:                    1750118400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-pro",
-			Version:                    "2.5",
-			DisplayName:                "Gemini 2.5 Pro",
-			Description:                "Stable release (June 17th, 2025) of Gemini 2.5 Pro",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-2.5-flash",
-			Object:                     "model",
-			Created:                    1750118400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-flash",
-			Version:                    "001",
-			DisplayName:                "Gemini 2.5 Flash",
-			Description:                "Stable version of Gemini 2.5 Flash, our mid-size multimodal model that supports up to 1 million tokens, released in June of 2025.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-2.5-flash-lite",
-			Object:                     "model",
-			Created:                    1753142400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-flash-lite",
-			Version:                    "2.5",
-			DisplayName:                "Gemini 2.5 Flash Lite",
-			Description:                "Our smallest and most cost effective model, built for at scale usage.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-3-pro-preview",
-			Object:                     "model",
-			Created:                    1737158400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-3-pro-preview",
-			Version:                    "3.0",
-			DisplayName:                "Gemini 3 Pro Preview",
-			Description:                "Gemini 3 Pro Preview",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
-		},
-		{
-			ID:                         "gemini-3-flash-preview",
-			Object:                     "model",
-			Created:                    1765929600,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-3-flash-preview",
-			Version:                    "3.0",
-			DisplayName:                "Gemini 3 Flash Preview",
-			Description:                "Our most intelligent model built for speed, combining frontier intelligence with superior search and grounding.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
-		},
-		{
-			ID:                         "gemini-3-pro-image-preview",
-			Object:                     "model",
-			Created:                    1737158400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-3-pro-image-preview",
-			Version:                    "3.0",
-			DisplayName:                "Gemini 3 Pro Image Preview",
-			Description:                "Gemini 3 Pro Image Preview",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
-		},
-		// Imagen image generation models - use :predict action
-		{
-			ID:                         "imagen-4.0-generate-001",
-			Object:                     "model",
-			Created:                    1750000000,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/imagen-4.0-generate-001",
-			Version:                    "4.0",
-			DisplayName:                "Imagen 4.0 Generate",
-			Description:                "Imagen 4.0 image generation model",
-			SupportedGenerationMethods: []string{"predict"},
-		},
-		{
-			ID:                         "imagen-4.0-ultra-generate-001",
-			Object:                     "model",
-			Created:                    1750000000,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/imagen-4.0-ultra-generate-001",
-			Version:                    "4.0",
-			DisplayName:                "Imagen 4.0 Ultra Generate",
-			Description:                "Imagen 4.0 Ultra high-quality image generation model",
-			SupportedGenerationMethods: []string{"predict"},
-		},
-		{
-			ID:                         "imagen-3.0-generate-002",
-			Object:                     "model",
-			Created:                    1740000000,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/imagen-3.0-generate-002",
-			Version:                    "3.0",
-			DisplayName:                "Imagen 3.0 Generate",
-			Description:                "Imagen 3.0 image generation model",
-			SupportedGenerationMethods: []string{"predict"},
-		},
-		{
-			ID:                         "imagen-3.0-fast-generate-001",
-			Object:                     "model",
-			Created:                    1740000000,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/imagen-3.0-fast-generate-001",
-			Version:                    "3.0",
-			DisplayName:                "Imagen 3.0 Fast Generate",
-			Description:                "Imagen 3.0 fast image generation model",
-			SupportedGenerationMethods: []string{"predict"},
-		},
-		{
-			ID:                         "imagen-4.0-fast-generate-001",
-			Object:                     "model",
-			Created:                    1750000000,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/imagen-4.0-fast-generate-001",
-			Version:                    "4.0",
-			DisplayName:                "Imagen 4.0 Fast Generate",
-			Description:                "Imagen 4.0 fast image generation model",
-			SupportedGenerationMethods: []string{"predict"},
-		},
-	}
-}
-
-// GetGeminiCLIModels returns the standard Gemini model definitions
-func GetGeminiCLIModels() []*ModelInfo {
-	return []*ModelInfo{
-		{
-			ID:                         "gemini-2.5-pro",
-			Object:                     "model",
-			Created:                    1750118400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-pro",
-			Version:                    "2.5",
-			DisplayName:                "Gemini 2.5 Pro",
-			Description:                "Stable release (June 17th, 2025) of Gemini 2.5 Pro",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-2.5-flash",
-			Object:                     "model",
-			Created:                    1750118400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-flash",
-			Version:                    "001",
-			DisplayName:                "Gemini 2.5 Flash",
-			Description:                "Stable version of Gemini 2.5 Flash, our mid-size multimodal model that supports up to 1 million tokens, released in June of 2025.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-2.5-flash-lite",
-			Object:                     "model",
-			Created:                    1753142400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-flash-lite",
-			Version:                    "2.5",
-			DisplayName:                "Gemini 2.5 Flash Lite",
-			Description:                "Our smallest and most cost effective model, built for at scale usage.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-3-pro-preview",
-			Object:                     "model",
-			Created:                    1737158400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-3-pro-preview",
-			Version:                    "3.0",
-			DisplayName:                "Gemini 3 Pro Preview",
-			Description:                "Our most intelligent model with SOTA reasoning and multimodal understanding, and powerful agentic and vibe coding capabilities",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
-		},
-		{
-			ID:                         "gemini-3-flash-preview",
-			Object:                     "model",
-			Created:                    1765929600,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-3-flash-preview",
-			Version:                    "3.0",
-			DisplayName:                "Gemini 3 Flash Preview",
-			Description:                "Our most intelligent model built for speed, combining frontier intelligence with superior search and grounding.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
-		},
-	}
-}
-
-// GetAIStudioModels returns the Gemini model definitions for AI Studio integrations
-func GetAIStudioModels() []*ModelInfo {
-	return []*ModelInfo{
-		{
-			ID:                         "gemini-2.5-pro",
-			Object:                     "model",
-			Created:                    1750118400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-pro",
-			Version:                    "2.5",
-			DisplayName:                "Gemini 2.5 Pro",
-			Description:                "Stable release (June 17th, 2025) of Gemini 2.5 Pro",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-2.5-flash",
-			Object:                     "model",
-			Created:                    1750118400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-flash",
-			Version:                    "001",
-			DisplayName:                "Gemini 2.5 Flash",
-			Description:                "Stable version of Gemini 2.5 Flash, our mid-size multimodal model that supports up to 1 million tokens, released in June of 2025.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-2.5-flash-lite",
-			Object:                     "model",
-			Created:                    1753142400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-flash-lite",
-			Version:                    "2.5",
-			DisplayName:                "Gemini 2.5 Flash Lite",
-			Description:                "Our smallest and most cost effective model, built for at scale usage.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-3-pro-preview",
-			Object:                     "model",
-			Created:                    1737158400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-3-pro-preview",
-			Version:                    "3.0",
-			DisplayName:                "Gemini 3 Pro Preview",
-			Description:                "Gemini 3 Pro Preview",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-3-flash-preview",
-			Object:                     "model",
-			Created:                    1765929600,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-3-flash-preview",
-			Version:                    "3.0",
-			DisplayName:                "Gemini 3 Flash Preview",
-			Description:                "Our most intelligent model built for speed, combining frontier intelligence with superior search and grounding.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-pro-latest",
-			Object:                     "model",
-			Created:                    1750118400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-pro-latest",
-			Version:                    "2.5",
-			DisplayName:                "Gemini Pro Latest",
-			Description:                "Latest release of Gemini Pro",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-flash-latest",
-			Object:                     "model",
-			Created:                    1750118400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-flash-latest",
-			Version:                    "2.5",
-			DisplayName:                "Gemini Flash Latest",
-			Description:                "Latest release of Gemini Flash",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-flash-lite-latest",
-			Object:                     "model",
-			Created:                    1753142400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-flash-lite-latest",
-			Version:                    "2.5",
-			DisplayName:                "Gemini Flash-Lite Latest",
-			Description:                "Latest release of Gemini Flash-Lite",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           65536,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			Thinking:                   &ThinkingSupport{Min: 512, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
-		},
-		{
-			ID:                         "gemini-2.5-flash-image-preview",
-			Object:                     "model",
-			Created:                    1756166400,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-flash-image-preview",
-			Version:                    "2.5",
-			DisplayName:                "Gemini 2.5 Flash Image Preview",
-			Description:                "State-of-the-art image generation and editing model.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           8192,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			// image models don't support thinkingConfig; leave Thinking nil
-		},
-		{
-			ID:                         "gemini-2.5-flash-image",
-			Object:                     "model",
-			Created:                    1759363200,
-			OwnedBy:                    "google",
-			Type:                       "gemini",
-			Name:                       "models/gemini-2.5-flash-image",
-			Version:                    "2.5",
-			DisplayName:                "Gemini 2.5 Flash Image",
-			Description:                "State-of-the-art image generation and editing model.",
-			InputTokenLimit:            1048576,
-			OutputTokenLimit:           8192,
-			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
-			// image models don't support thinkingConfig; leave Thinking nil
-		},
-	}
-}
-
-// GetOpenAIModels returns the standard OpenAI model definitions
-func GetOpenAIModels() []*ModelInfo {
-	return []*ModelInfo{
-		{
-			ID:                  "gpt-5",
-			Object:              "model",
-			Created:             1754524800,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-08-07",
-			DisplayName:         "GPT 5",
-			Description:         "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-			Thinking:            &ThinkingSupport{Levels: []string{"minimal", "low", "medium", "high"}},
-		},
-		{
-			ID:                  "gpt-5-codex",
-			Object:              "model",
-			Created:             1757894400,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-09-15",
-			DisplayName:         "GPT 5 Codex",
-			Description:         "Stable version of GPT 5 Codex, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
-		},
-		{
-			ID:                  "gpt-5-codex-mini",
-			Object:              "model",
-			Created:             1762473600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5-2025-11-07",
-			DisplayName:         "GPT 5 Codex Mini",
-			Description:         "Stable version of GPT 5 Codex Mini: cheaper, faster, but less capable version of GPT 5 Codex.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
-		},
-		{
-			ID:                  "gpt-5.1",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5",
-			Description:         "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high"}},
-		},
-		{
-			ID:                  "gpt-5.1-codex",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5.1 Codex",
-			Description:         "Stable version of GPT 5.1 Codex, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
-		},
-		{
-			ID:                  "gpt-5.1-codex-mini",
-			Object:              "model",
-			Created:             1762905600,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-2025-11-12",
-			DisplayName:         "GPT 5.1 Codex Mini",
-			Description:         "Stable version of GPT 5.1 Codex Mini: cheaper, faster, but less capable version of GPT 5.1 Codex.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
-		},
-		{
-			ID:                  "gpt-5.1-codex-max",
-			Object:              "model",
-			Created:             1763424000,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.1-max",
-			DisplayName:         "GPT 5.1 Codex Max",
-			Description:         "Stable version of GPT 5.1 Codex Max",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
-		},
-		{
-			ID:                  "gpt-5.2",
-			Object:              "model",
-			Created:             1765440000,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.2",
-			DisplayName:         "GPT 5.2",
-			Description:         "Stable version of GPT 5.2",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high", "xhigh"}},
-		},
-		{
-			ID:                  "gpt-5.2-codex",
-			Object:              "model",
-			Created:             1765440000,
-			OwnedBy:             "openai",
-			Type:                "openai",
-			Version:             "gpt-5.2",
-			DisplayName:         "GPT 5.2 Codex",
-			Description:         "Stable version of GPT 5.2 Codex, The best model for coding and agentic tasks across domains.",
-			ContextLength:       400000,
-			MaxCompletionTokens: 128000,
-			SupportedParameters: []string{"tools"},
-			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
-		},
-	}
-}
-
-// GetQwenModels returns the standard Qwen model definitions
-func GetQwenModels() []*ModelInfo {
-	return []*ModelInfo{
-		{
-			ID:                  "qwen3-coder-plus",
-			Object:              "model",
-			Created:             1753228800,
-			OwnedBy:             "qwen",
-			Type:                "qwen",
-			Version:             "3.0",
-			DisplayName:         "Qwen3 Coder Plus",
-			Description:         "Advanced code generation and understanding model",
-			ContextLength:       32768,
-			MaxCompletionTokens: 8192,
-			SupportedParameters: []string{"temperature", "top_p", "max_tokens", "stream", "stop"},
-		},
-		{
-			ID:                  "qwen3-coder-flash",
-			Object:              "model",
-			Created:             1753228800,
-			OwnedBy:             "qwen",
-			Type:                "qwen",
-			Version:             "3.0",
-			DisplayName:         "Qwen3 Coder Flash",
-			Description:         "Fast code generation model",
-			ContextLength:       8192,
-			MaxCompletionTokens: 2048,
-			SupportedParameters: []string{"temperature", "top_p", "max_tokens", "stream", "stop"},
-		},
-		{
-			ID:                  "vision-model",
-			Object:              "model",
-			Created:             1758672000,
-			OwnedBy:             "qwen",
-			Type:                "qwen",
-			Version:             "3.0",
-			DisplayName:         "Qwen3 Vision Model",
-			Description:         "Vision model model",
-			ContextLength:       32768,
-			MaxCompletionTokens: 2048,
-			SupportedParameters: []string{"temperature", "top_p", "max_tokens", "stream", "stop"},
-		},
-	}
-}
-
-// iFlowThinkingSupport is a shared ThinkingSupport configuration for iFlow models
-// that support thinking mode via chat_template_kwargs.enable_thinking (boolean toggle).
-// Uses level-based configuration so standard normalization flows apply before conversion.
-var iFlowThinkingSupport = &ThinkingSupport{
-	Levels: []string{"none", "auto", "minimal", "low", "medium", "high", "xhigh"},
-}
-
-// GetIFlowModels returns supported models for iFlow OAuth accounts.
-func GetIFlowModels() []*ModelInfo {
-	entries := []struct {
-		ID          string
-		DisplayName string
-		Description string
-		Created     int64
-		Thinking    *ThinkingSupport
-	}{
-		{ID: "tstars2.0", DisplayName: "TStars-2.0", Description: "iFlow TStars-2.0 multimodal assistant", Created: 1746489600},
-		{ID: "qwen3-coder-plus", DisplayName: "Qwen3-Coder-Plus", Description: "Qwen3 Coder Plus code generation", Created: 1753228800},
-		{ID: "qwen3-max", DisplayName: "Qwen3-Max", Description: "Qwen3 flagship model", Created: 1758672000},
-		{ID: "qwen3-vl-plus", DisplayName: "Qwen3-VL-Plus", Description: "Qwen3 multimodal vision-language", Created: 1758672000},
-		{ID: "qwen3-max-preview", DisplayName: "Qwen3-Max-Preview", Description: "Qwen3 Max preview build", Created: 1757030400},
-		{ID: "kimi-k2-0905", DisplayName: "Kimi-K2-Instruct-0905", Description: "Moonshot Kimi K2 instruct 0905", Created: 1757030400},
-		{ID: "glm-4.6", DisplayName: "GLM-4.6", Description: "Zhipu GLM 4.6 general model", Created: 1759190400, Thinking: iFlowThinkingSupport},
-		{ID: "glm-4.7", DisplayName: "GLM-4.7", Description: "Zhipu GLM 4.7 general model", Created: 1766448000, Thinking: iFlowThinkingSupport},
-		{ID: "kimi-k2", DisplayName: "Kimi-K2", Description: "Moonshot Kimi K2 general model", Created: 1752192000},
-		{ID: "kimi-k2-thinking", DisplayName: "Kimi-K2-Thinking", Description: "Moonshot Kimi K2 thinking model", Created: 1762387200},
-		{ID: "deepseek-v3.2-chat", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2 Chat", Created: 1764576000},
-		{ID: "deepseek-v3.2-reasoner", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2 Reasoner", Created: 1764576000},
-		{ID: "deepseek-v3.2", DisplayName: "DeepSeek-V3.2-Exp", Description: "DeepSeek V3.2 experimental", Created: 1759104000},
-		{ID: "deepseek-v3.1", DisplayName: "DeepSeek-V3.1-Terminus", Description: "DeepSeek V3.1 Terminus", Created: 1756339200},
-		{ID: "deepseek-r1", DisplayName: "DeepSeek-R1", Description: "DeepSeek reasoning model R1", Created: 1737331200},
-		{ID: "deepseek-v3", DisplayName: "DeepSeek-V3-671B", Description: "DeepSeek V3 671B", Created: 1734307200},
-		{ID: "qwen3-32b", DisplayName: "Qwen3-32B", Description: "Qwen3 32B", Created: 1747094400},
-		{ID: "qwen3-235b-a22b-thinking-2507", DisplayName: "Qwen3-235B-A22B-Thinking", Description: "Qwen3 235B A22B Thinking (2507)", Created: 1753401600},
-		{ID: "qwen3-235b-a22b-instruct", DisplayName: "Qwen3-235B-A22B-Instruct", Description: "Qwen3 235B A22B Instruct", Created: 1753401600},
-		{ID: "qwen3-235b", DisplayName: "Qwen3-235B-A22B", Description: "Qwen3 235B A22B", Created: 1753401600},
-		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000, Thinking: 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 {
-		models = append(models, &ModelInfo{
-			ID:          entry.ID,
-			Object:      "model",
-			Created:     entry.Created,
-			OwnedBy:     "iflow",
-			Type:        "iflow",
-			DisplayName: entry.DisplayName,
-			Description: entry.Description,
-			Thinking:    entry.Thinking,
+// GetStaticModelDefinitionsByChannel returns static model definitions for a given channel/provider.
+// It returns nil when the channel is unknown.
+//
+// Supported channels:
+//   - claude
+//   - gemini
+//   - vertex
+//   - gemini-cli
+//   - aistudio
+//   - codex
+//   - qwen
+//   - iflow
+//   - antigravity (returns static overrides only)
+func GetStaticModelDefinitionsByChannel(channel string) []*ModelInfo {
+	key := strings.ToLower(strings.TrimSpace(channel))
+	switch key {
+	case "claude":
+		return GetClaudeModels()
+	case "gemini":
+		return GetGeminiModels()
+	case "vertex":
+		return GetGeminiVertexModels()
+	case "gemini-cli":
+		return GetGeminiCLIModels()
+	case "aistudio":
+		return GetAIStudioModels()
+	case "codex":
+		return GetOpenAIModels()
+	case "qwen":
+		return GetQwenModels()
+	case "iflow":
+		return GetIFlowModels()
+	case "antigravity":
+		cfg := GetAntigravityModelConfig()
+		if len(cfg) == 0 {
+			return nil
+		}
+		models := make([]*ModelInfo, 0, len(cfg))
+		for modelID, entry := range cfg {
+			if modelID == "" || entry == nil {
+				continue
+			}
+			models = append(models, &ModelInfo{
+				ID:                  modelID,
+				Object:              "model",
+				OwnedBy:             "antigravity",
+				Type:                "antigravity",
+				Thinking:            entry.Thinking,
+				MaxCompletionTokens: entry.MaxCompletionTokens,
+			})
+		}
+		sort.Slice(models, func(i, j int) bool {
+			return strings.ToLower(models[i].ID) < strings.ToLower(models[j].ID)
 		})
-	}
-	return models
-}
-
-// AntigravityModelConfig captures static antigravity model overrides, including
-// Thinking budget limits and provider max completion tokens.
-type AntigravityModelConfig struct {
-	Thinking            *ThinkingSupport
-	MaxCompletionTokens int
-}
-
-// GetAntigravityModelConfig returns static configuration for antigravity models.
-// 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}},
-		"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":     {},
+		return models
+	default:
+		return nil
 	}
 }
 

internal/registry/model_definitions_static_data.go

@@ -0,0 +1,846 @@
+// Package registry provides model definitions for various AI service providers.
+// This file stores the static model metadata catalog.
+package registry
+
+// GetClaudeModels returns the standard Claude model definitions
+func GetClaudeModels() []*ModelInfo {
+	return []*ModelInfo{
+
+		{
+			ID:                  "claude-haiku-4-5-20251001",
+			Object:              "model",
+			Created:             1759276800, // 2025-10-01
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4.5 Haiku",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			// Thinking: not supported for Haiku models
+		},
+		{
+			ID:                  "claude-sonnet-4-5-20250929",
+			Object:              "model",
+			Created:             1759104000, // 2025-09-29
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4.5 Sonnet",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
+		},
+		{
+			ID:                  "claude-opus-4-5-20251101",
+			Object:              "model",
+			Created:             1761955200, // 2025-11-01
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4.5 Opus",
+			Description:         "Premium model combining maximum intelligence with practical performance",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
+		},
+		{
+			ID:                  "claude-opus-4-1-20250805",
+			Object:              "model",
+			Created:             1722945600, // 2025-08-05
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4.1 Opus",
+			ContextLength:       200000,
+			MaxCompletionTokens: 32000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
+		},
+		{
+			ID:                  "claude-opus-4-20250514",
+			Object:              "model",
+			Created:             1715644800, // 2025-05-14
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4 Opus",
+			ContextLength:       200000,
+			MaxCompletionTokens: 32000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
+		},
+		{
+			ID:                  "claude-sonnet-4-20250514",
+			Object:              "model",
+			Created:             1715644800, // 2025-05-14
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4 Sonnet",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
+		},
+		{
+			ID:                  "claude-3-7-sonnet-20250219",
+			Object:              "model",
+			Created:             1708300800, // 2025-02-19
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 3.7 Sonnet",
+			ContextLength:       128000,
+			MaxCompletionTokens: 8192,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: false, DynamicAllowed: false},
+		},
+		{
+			ID:                  "claude-3-5-haiku-20241022",
+			Object:              "model",
+			Created:             1729555200, // 2024-10-22
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 3.5 Haiku",
+			ContextLength:       128000,
+			MaxCompletionTokens: 8192,
+			// Thinking: not supported for Haiku models
+		},
+	}
+}
+
+// GetGeminiModels returns the standard Gemini model definitions
+func GetGeminiModels() []*ModelInfo {
+	return []*ModelInfo{
+		{
+			ID:                         "gemini-2.5-pro",
+			Object:                     "model",
+			Created:                    1750118400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-pro",
+			Version:                    "2.5",
+			DisplayName:                "Gemini 2.5 Pro",
+			Description:                "Stable release (June 17th, 2025) of Gemini 2.5 Pro",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-2.5-flash",
+			Object:                     "model",
+			Created:                    1750118400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-flash",
+			Version:                    "001",
+			DisplayName:                "Gemini 2.5 Flash",
+			Description:                "Stable version of Gemini 2.5 Flash, our mid-size multimodal model that supports up to 1 million tokens, released in June of 2025.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-2.5-flash-lite",
+			Object:                     "model",
+			Created:                    1753142400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-flash-lite",
+			Version:                    "2.5",
+			DisplayName:                "Gemini 2.5 Flash Lite",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-3-pro-preview",
+			Object:                     "model",
+			Created:                    1737158400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-pro-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Pro Preview",
+			Description:                "Gemini 3 Pro Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
+		{
+			ID:                         "gemini-3-flash-preview",
+			Object:                     "model",
+			Created:                    1765929600,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-flash-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Flash Preview",
+			Description:                "Gemini 3 Flash Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
+		},
+		{
+			ID:                         "gemini-3-pro-image-preview",
+			Object:                     "model",
+			Created:                    1737158400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-pro-image-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Pro Image Preview",
+			Description:                "Gemini 3 Pro Image Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
+	}
+}
+
+func GetGeminiVertexModels() []*ModelInfo {
+	return []*ModelInfo{
+		{
+			ID:                         "gemini-2.5-pro",
+			Object:                     "model",
+			Created:                    1750118400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-pro",
+			Version:                    "2.5",
+			DisplayName:                "Gemini 2.5 Pro",
+			Description:                "Stable release (June 17th, 2025) of Gemini 2.5 Pro",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-2.5-flash",
+			Object:                     "model",
+			Created:                    1750118400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-flash",
+			Version:                    "001",
+			DisplayName:                "Gemini 2.5 Flash",
+			Description:                "Stable version of Gemini 2.5 Flash, our mid-size multimodal model that supports up to 1 million tokens, released in June of 2025.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-2.5-flash-lite",
+			Object:                     "model",
+			Created:                    1753142400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-flash-lite",
+			Version:                    "2.5",
+			DisplayName:                "Gemini 2.5 Flash Lite",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-3-pro-preview",
+			Object:                     "model",
+			Created:                    1737158400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-pro-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Pro Preview",
+			Description:                "Gemini 3 Pro Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
+		{
+			ID:                         "gemini-3-flash-preview",
+			Object:                     "model",
+			Created:                    1765929600,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-flash-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Flash Preview",
+			Description:                "Our most intelligent model built for speed, combining frontier intelligence with superior search and grounding.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
+		},
+		{
+			ID:                         "gemini-3-pro-image-preview",
+			Object:                     "model",
+			Created:                    1737158400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-pro-image-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Pro Image Preview",
+			Description:                "Gemini 3 Pro Image Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
+		// Imagen image generation models - use :predict action
+		{
+			ID:                         "imagen-4.0-generate-001",
+			Object:                     "model",
+			Created:                    1750000000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/imagen-4.0-generate-001",
+			Version:                    "4.0",
+			DisplayName:                "Imagen 4.0 Generate",
+			Description:                "Imagen 4.0 image generation model",
+			SupportedGenerationMethods: []string{"predict"},
+		},
+		{
+			ID:                         "imagen-4.0-ultra-generate-001",
+			Object:                     "model",
+			Created:                    1750000000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/imagen-4.0-ultra-generate-001",
+			Version:                    "4.0",
+			DisplayName:                "Imagen 4.0 Ultra Generate",
+			Description:                "Imagen 4.0 Ultra high-quality image generation model",
+			SupportedGenerationMethods: []string{"predict"},
+		},
+		{
+			ID:                         "imagen-3.0-generate-002",
+			Object:                     "model",
+			Created:                    1740000000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/imagen-3.0-generate-002",
+			Version:                    "3.0",
+			DisplayName:                "Imagen 3.0 Generate",
+			Description:                "Imagen 3.0 image generation model",
+			SupportedGenerationMethods: []string{"predict"},
+		},
+		{
+			ID:                         "imagen-3.0-fast-generate-001",
+			Object:                     "model",
+			Created:                    1740000000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/imagen-3.0-fast-generate-001",
+			Version:                    "3.0",
+			DisplayName:                "Imagen 3.0 Fast Generate",
+			Description:                "Imagen 3.0 fast image generation model",
+			SupportedGenerationMethods: []string{"predict"},
+		},
+		{
+			ID:                         "imagen-4.0-fast-generate-001",
+			Object:                     "model",
+			Created:                    1750000000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/imagen-4.0-fast-generate-001",
+			Version:                    "4.0",
+			DisplayName:                "Imagen 4.0 Fast Generate",
+			Description:                "Imagen 4.0 fast image generation model",
+			SupportedGenerationMethods: []string{"predict"},
+		},
+	}
+}
+
+// GetGeminiCLIModels returns the standard Gemini model definitions
+func GetGeminiCLIModels() []*ModelInfo {
+	return []*ModelInfo{
+		{
+			ID:                         "gemini-2.5-pro",
+			Object:                     "model",
+			Created:                    1750118400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-pro",
+			Version:                    "2.5",
+			DisplayName:                "Gemini 2.5 Pro",
+			Description:                "Stable release (June 17th, 2025) of Gemini 2.5 Pro",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-2.5-flash",
+			Object:                     "model",
+			Created:                    1750118400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-flash",
+			Version:                    "001",
+			DisplayName:                "Gemini 2.5 Flash",
+			Description:                "Stable version of Gemini 2.5 Flash, our mid-size multimodal model that supports up to 1 million tokens, released in June of 2025.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-2.5-flash-lite",
+			Object:                     "model",
+			Created:                    1753142400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-flash-lite",
+			Version:                    "2.5",
+			DisplayName:                "Gemini 2.5 Flash Lite",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-3-pro-preview",
+			Object:                     "model",
+			Created:                    1737158400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-pro-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Pro Preview",
+			Description:                "Our most intelligent model with SOTA reasoning and multimodal understanding, and powerful agentic and vibe coding capabilities",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
+		{
+			ID:                         "gemini-3-flash-preview",
+			Object:                     "model",
+			Created:                    1765929600,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-flash-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Flash Preview",
+			Description:                "Our most intelligent model built for speed, combining frontier intelligence with superior search and grounding.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
+		},
+	}
+}
+
+// GetAIStudioModels returns the Gemini model definitions for AI Studio integrations
+func GetAIStudioModels() []*ModelInfo {
+	return []*ModelInfo{
+		{
+			ID:                         "gemini-2.5-pro",
+			Object:                     "model",
+			Created:                    1750118400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-pro",
+			Version:                    "2.5",
+			DisplayName:                "Gemini 2.5 Pro",
+			Description:                "Stable release (June 17th, 2025) of Gemini 2.5 Pro",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-2.5-flash",
+			Object:                     "model",
+			Created:                    1750118400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-flash",
+			Version:                    "001",
+			DisplayName:                "Gemini 2.5 Flash",
+			Description:                "Stable version of Gemini 2.5 Flash, our mid-size multimodal model that supports up to 1 million tokens, released in June of 2025.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-2.5-flash-lite",
+			Object:                     "model",
+			Created:                    1753142400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-flash-lite",
+			Version:                    "2.5",
+			DisplayName:                "Gemini 2.5 Flash Lite",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-3-pro-preview",
+			Object:                     "model",
+			Created:                    1737158400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-pro-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Pro Preview",
+			Description:                "Gemini 3 Pro Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-3-flash-preview",
+			Object:                     "model",
+			Created:                    1765929600,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3-flash-preview",
+			Version:                    "3.0",
+			DisplayName:                "Gemini 3 Flash Preview",
+			Description:                "Our most intelligent model built for speed, combining frontier intelligence with superior search and grounding.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-pro-latest",
+			Object:                     "model",
+			Created:                    1750118400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-pro-latest",
+			Version:                    "2.5",
+			DisplayName:                "Gemini Pro Latest",
+			Description:                "Latest release of Gemini Pro",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-flash-latest",
+			Object:                     "model",
+			Created:                    1750118400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-flash-latest",
+			Version:                    "2.5",
+			DisplayName:                "Gemini Flash Latest",
+			Description:                "Latest release of Gemini Flash",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                         "gemini-flash-lite-latest",
+			Object:                     "model",
+			Created:                    1753142400,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-flash-lite-latest",
+			Version:                    "2.5",
+			DisplayName:                "Gemini Flash-Lite Latest",
+			Description:                "Latest release of Gemini Flash-Lite",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 512, Max: 24576, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		// {
+		// 	ID:                         "gemini-2.5-flash-image-preview",
+		// 	Object:                     "model",
+		// 	Created:                    1756166400,
+		// 	OwnedBy:                    "google",
+		// 	Type:                       "gemini",
+		// 	Name:                       "models/gemini-2.5-flash-image-preview",
+		// 	Version:                    "2.5",
+		// 	DisplayName:                "Gemini 2.5 Flash Image Preview",
+		// 	Description:                "State-of-the-art image generation and editing model.",
+		// 	InputTokenLimit:            1048576,
+		// 	OutputTokenLimit:           8192,
+		// 	SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+		// 	// image models don't support thinkingConfig; leave Thinking nil
+		// },
+		{
+			ID:                         "gemini-2.5-flash-image",
+			Object:                     "model",
+			Created:                    1759363200,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-2.5-flash-image",
+			Version:                    "2.5",
+			DisplayName:                "Gemini 2.5 Flash Image",
+			Description:                "State-of-the-art image generation and editing model.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           8192,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			// image models don't support thinkingConfig; leave Thinking nil
+		},
+	}
+}
+
+// GetOpenAIModels returns the standard OpenAI model definitions
+func GetOpenAIModels() []*ModelInfo {
+	return []*ModelInfo{
+		{
+			ID:                  "gpt-5",
+			Object:              "model",
+			Created:             1754524800,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5-2025-08-07",
+			DisplayName:         "GPT 5",
+			Description:         "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
+			ContextLength:       400000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"minimal", "low", "medium", "high"}},
+		},
+		{
+			ID:                  "gpt-5-codex",
+			Object:              "model",
+			Created:             1757894400,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5-2025-09-15",
+			DisplayName:         "GPT 5 Codex",
+			Description:         "Stable version of GPT 5 Codex, The best model for coding and agentic tasks across domains.",
+			ContextLength:       400000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
+		},
+		{
+			ID:                  "gpt-5-codex-mini",
+			Object:              "model",
+			Created:             1762473600,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5-2025-11-07",
+			DisplayName:         "GPT 5 Codex Mini",
+			Description:         "Stable version of GPT 5 Codex Mini: cheaper, faster, but less capable version of GPT 5 Codex.",
+			ContextLength:       400000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
+		},
+		{
+			ID:                  "gpt-5.1",
+			Object:              "model",
+			Created:             1762905600,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5.1-2025-11-12",
+			DisplayName:         "GPT 5",
+			Description:         "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
+			ContextLength:       400000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high"}},
+		},
+		{
+			ID:                  "gpt-5.1-codex",
+			Object:              "model",
+			Created:             1762905600,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5.1-2025-11-12",
+			DisplayName:         "GPT 5.1 Codex",
+			Description:         "Stable version of GPT 5.1 Codex, The best model for coding and agentic tasks across domains.",
+			ContextLength:       400000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
+		},
+		{
+			ID:                  "gpt-5.1-codex-mini",
+			Object:              "model",
+			Created:             1762905600,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5.1-2025-11-12",
+			DisplayName:         "GPT 5.1 Codex Mini",
+			Description:         "Stable version of GPT 5.1 Codex Mini: cheaper, faster, but less capable version of GPT 5.1 Codex.",
+			ContextLength:       400000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
+		},
+		{
+			ID:                  "gpt-5.1-codex-max",
+			Object:              "model",
+			Created:             1763424000,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5.1-max",
+			DisplayName:         "GPT 5.1 Codex Max",
+			Description:         "Stable version of GPT 5.1 Codex Max",
+			ContextLength:       400000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
+		},
+		{
+			ID:                  "gpt-5.2",
+			Object:              "model",
+			Created:             1765440000,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5.2",
+			DisplayName:         "GPT 5.2",
+			Description:         "Stable version of GPT 5.2",
+			ContextLength:       400000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high", "xhigh"}},
+		},
+		{
+			ID:                  "gpt-5.2-codex",
+			Object:              "model",
+			Created:             1765440000,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5.2",
+			DisplayName:         "GPT 5.2 Codex",
+			Description:         "Stable version of GPT 5.2 Codex, The best model for coding and agentic tasks across domains.",
+			ContextLength:       400000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
+		},
+	}
+}
+
+// GetQwenModels returns the standard Qwen model definitions
+func GetQwenModels() []*ModelInfo {
+	return []*ModelInfo{
+		{
+			ID:                  "qwen3-coder-plus",
+			Object:              "model",
+			Created:             1753228800,
+			OwnedBy:             "qwen",
+			Type:                "qwen",
+			Version:             "3.0",
+			DisplayName:         "Qwen3 Coder Plus",
+			Description:         "Advanced code generation and understanding model",
+			ContextLength:       32768,
+			MaxCompletionTokens: 8192,
+			SupportedParameters: []string{"temperature", "top_p", "max_tokens", "stream", "stop"},
+		},
+		{
+			ID:                  "qwen3-coder-flash",
+			Object:              "model",
+			Created:             1753228800,
+			OwnedBy:             "qwen",
+			Type:                "qwen",
+			Version:             "3.0",
+			DisplayName:         "Qwen3 Coder Flash",
+			Description:         "Fast code generation model",
+			ContextLength:       8192,
+			MaxCompletionTokens: 2048,
+			SupportedParameters: []string{"temperature", "top_p", "max_tokens", "stream", "stop"},
+		},
+		{
+			ID:                  "vision-model",
+			Object:              "model",
+			Created:             1758672000,
+			OwnedBy:             "qwen",
+			Type:                "qwen",
+			Version:             "3.0",
+			DisplayName:         "Qwen3 Vision Model",
+			Description:         "Vision model model",
+			ContextLength:       32768,
+			MaxCompletionTokens: 2048,
+			SupportedParameters: []string{"temperature", "top_p", "max_tokens", "stream", "stop"},
+		},
+	}
+}
+
+// iFlowThinkingSupport is a shared ThinkingSupport configuration for iFlow models
+// that support thinking mode via chat_template_kwargs.enable_thinking (boolean toggle).
+// Uses level-based configuration so standard normalization flows apply before conversion.
+var iFlowThinkingSupport = &ThinkingSupport{
+	Levels: []string{"none", "auto", "minimal", "low", "medium", "high", "xhigh"},
+}
+
+// GetIFlowModels returns supported models for iFlow OAuth accounts.
+func GetIFlowModels() []*ModelInfo {
+	entries := []struct {
+		ID          string
+		DisplayName string
+		Description string
+		Created     int64
+		Thinking    *ThinkingSupport
+	}{
+		{ID: "tstars2.0", DisplayName: "TStars-2.0", Description: "iFlow TStars-2.0 multimodal assistant", Created: 1746489600},
+		{ID: "qwen3-coder-plus", DisplayName: "Qwen3-Coder-Plus", Description: "Qwen3 Coder Plus code generation", Created: 1753228800},
+		{ID: "qwen3-max", DisplayName: "Qwen3-Max", Description: "Qwen3 flagship model", Created: 1758672000},
+		{ID: "qwen3-vl-plus", DisplayName: "Qwen3-VL-Plus", Description: "Qwen3 multimodal vision-language", Created: 1758672000},
+		{ID: "qwen3-max-preview", DisplayName: "Qwen3-Max-Preview", Description: "Qwen3 Max preview build", Created: 1757030400, Thinking: iFlowThinkingSupport},
+		{ID: "kimi-k2-0905", DisplayName: "Kimi-K2-Instruct-0905", Description: "Moonshot Kimi K2 instruct 0905", Created: 1757030400},
+		{ID: "glm-4.6", DisplayName: "GLM-4.6", Description: "Zhipu GLM 4.6 general model", Created: 1759190400, Thinking: iFlowThinkingSupport},
+		{ID: "glm-4.7", DisplayName: "GLM-4.7", Description: "Zhipu GLM 4.7 general model", Created: 1766448000, Thinking: iFlowThinkingSupport},
+		{ID: "kimi-k2", DisplayName: "Kimi-K2", Description: "Moonshot Kimi K2 general model", Created: 1752192000},
+		{ID: "kimi-k2-thinking", DisplayName: "Kimi-K2-Thinking", Description: "Moonshot Kimi K2 thinking model", Created: 1762387200},
+		{ID: "deepseek-v3.2-chat", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2 Chat", Created: 1764576000},
+		{ID: "deepseek-v3.2-reasoner", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2 Reasoner", Created: 1764576000},
+		{ID: "deepseek-v3.2", DisplayName: "DeepSeek-V3.2-Exp", Description: "DeepSeek V3.2 experimental", Created: 1759104000, Thinking: iFlowThinkingSupport},
+		{ID: "deepseek-v3.1", DisplayName: "DeepSeek-V3.1-Terminus", Description: "DeepSeek V3.1 Terminus", Created: 1756339200, Thinking: iFlowThinkingSupport},
+		{ID: "deepseek-r1", DisplayName: "DeepSeek-R1", Description: "DeepSeek reasoning model R1", Created: 1737331200},
+		{ID: "deepseek-v3", DisplayName: "DeepSeek-V3-671B", Description: "DeepSeek V3 671B", Created: 1734307200},
+		{ID: "qwen3-32b", DisplayName: "Qwen3-32B", Description: "Qwen3 32B", Created: 1747094400},
+		{ID: "qwen3-235b-a22b-thinking-2507", DisplayName: "Qwen3-235B-A22B-Thinking", Description: "Qwen3 235B A22B Thinking (2507)", Created: 1753401600},
+		{ID: "qwen3-235b-a22b-instruct", DisplayName: "Qwen3-235B-A22B-Instruct", Description: "Qwen3 235B A22B Instruct", Created: 1753401600},
+		{ID: "qwen3-235b", DisplayName: "Qwen3-235B-A22B", Description: "Qwen3 235B A22B", Created: 1753401600},
+		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000, Thinking: 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 {
+		models = append(models, &ModelInfo{
+			ID:          entry.ID,
+			Object:      "model",
+			Created:     entry.Created,
+			OwnedBy:     "iflow",
+			Type:        "iflow",
+			DisplayName: entry.DisplayName,
+			Description: entry.Description,
+			Thinking:    entry.Thinking,
+		})
+	}
+	return models
+}
+
+// AntigravityModelConfig captures static antigravity model overrides, including
+// Thinking budget limits and provider max completion tokens.
+type AntigravityModelConfig struct {
+	Thinking            *ThinkingSupport
+	MaxCompletionTokens int
+}
+
+// GetAntigravityModelConfig returns static configuration for antigravity models.
+// Keys use upstream model names returned by the Antigravity models endpoint.
+func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
+	return map[string]*AntigravityModelConfig{
+		// "rev19-uic3-1p":              {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true}},
+		"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}},
+		"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":     {},
+	}
+}

internal/runtime/executor/aistudio_executor.go

@@ -111,6 +111,9 @@ func (e *AIStudioExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.A
 
 // 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) {
+	if opts.Alt == "responses/compact" {
+		return resp, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
@@ -167,6 +170,9 @@ 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) {
+	if opts.Alt == "responses/compact" {
+		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)

internal/runtime/executor/antigravity_executor.go

@@ -109,6 +109,9 @@ func (e *AntigravityExecutor) HttpRequest(ctx context.Context, auth *cliproxyaut
 
 // Execute performs a non-streaming request to the Antigravity API.
 func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	if opts.Alt == "responses/compact" {
+		return resp, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 	isClaude := strings.Contains(strings.ToLower(baseModel), "claude")
 
@@ -148,87 +151,108 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
 
-	var lastStatus int
-	var lastBody []byte
-	var lastErr error
+	attempts := antigravityRetryAttempts(auth, e.cfg)
 
-	for idx, baseURL := range baseURLs {
-		httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, translated, false, opts.Alt, baseURL)
-		if errReq != nil {
-			err = errReq
-			return resp, err
-		}
+attemptLoop:
+	for attempt := 0; attempt < attempts; attempt++ {
+		var lastStatus int
+		var lastBody []byte
+		var lastErr error
 
-		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
+		for idx, baseURL := range baseURLs {
+			httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, translated, false, opts.Alt, baseURL)
+			if errReq != nil {
+				err = errReq
+				return resp, err
 			}
-			lastStatus = 0
-			lastBody = nil
-			lastErr = errDo
-			if idx+1 < len(baseURLs) {
-				log.Debugf("antigravity executor: request error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
-				continue
+
+			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
+				if idx+1 < len(baseURLs) {
+					log.Debugf("antigravity executor: request error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+					continue
+				}
+				err = errDo
+				return resp, err
 			}
-			err = errDo
-			return resp, err
+
+			recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+			bodyBytes, errRead := io.ReadAll(httpResp.Body)
+			if errClose := httpResp.Body.Close(); errClose != nil {
+				log.Errorf("antigravity executor: close response body error: %v", errClose)
+			}
+			if errRead != nil {
+				recordAPIResponseError(ctx, e.cfg, errRead)
+				err = errRead
+				return resp, err
+			}
+			appendAPIResponseChunk(ctx, e.cfg, bodyBytes)
+
+			if httpResp.StatusCode < http.StatusOK || httpResp.StatusCode >= http.StatusMultipleChoices {
+				log.Debugf("antigravity executor: upstream error status: %d, body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), bodyBytes))
+				lastStatus = httpResp.StatusCode
+				lastBody = append([]byte(nil), bodyBytes...)
+				lastErr = nil
+				if httpResp.StatusCode == http.StatusTooManyRequests && idx+1 < len(baseURLs) {
+					log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+					continue
+				}
+				if antigravityShouldRetryNoCapacity(httpResp.StatusCode, bodyBytes) {
+					if idx+1 < len(baseURLs) {
+						log.Debugf("antigravity executor: no capacity on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+						continue
+					}
+					if attempt+1 < attempts {
+						delay := antigravityNoCapacityRetryDelay(attempt)
+						log.Debugf("antigravity executor: no capacity for model %s, retrying in %s (attempt %d/%d)", baseModel, delay, attempt+1, attempts)
+						if errWait := antigravityWait(ctx, delay); errWait != nil {
+							return resp, errWait
+						}
+						continue attemptLoop
+					}
+				}
+				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
+			}
+
+			reporter.publish(ctx, parseAntigravityUsage(bodyBytes))
+			var param any
+			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, bodyBytes, &param)
+			resp = cliproxyexecutor.Response{Payload: []byte(converted)}
+			reporter.ensurePublished(ctx)
+			return resp, nil
 		}
 
-		recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
-		bodyBytes, errRead := io.ReadAll(httpResp.Body)
-		if errClose := httpResp.Body.Close(); errClose != nil {
-			log.Errorf("antigravity executor: close response body error: %v", errClose)
-		}
-		if errRead != nil {
-			recordAPIResponseError(ctx, e.cfg, errRead)
-			err = errRead
-			return resp, err
-		}
-		appendAPIResponseChunk(ctx, e.cfg, bodyBytes)
-
-		if httpResp.StatusCode < http.StatusOK || httpResp.StatusCode >= http.StatusMultipleChoices {
-			log.Debugf("antigravity executor: upstream error status: %d, body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), bodyBytes))
-			lastStatus = httpResp.StatusCode
-			lastBody = append([]byte(nil), bodyBytes...)
-			lastErr = nil
-			if httpResp.StatusCode == http.StatusTooManyRequests && idx+1 < len(baseURLs) {
-				log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
-				continue
-			}
-			sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
-			if httpResp.StatusCode == http.StatusTooManyRequests {
-				if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+		switch {
+		case lastStatus != 0:
+			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
-			return resp, err
+		case lastErr != nil:
+			err = lastErr
+		default:
+			err = statusErr{code: http.StatusServiceUnavailable, msg: "antigravity executor: no base url available"}
 		}
-
-		reporter.publish(ctx, parseAntigravityUsage(bodyBytes))
-		var param any
-		converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, bodyBytes, &param)
-		resp = cliproxyexecutor.Response{Payload: []byte(converted)}
-		reporter.ensurePublished(ctx)
-		return resp, nil
+		return resp, err
 	}
 
-	switch {
-	case lastStatus != 0:
-		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:
-		err = statusErr{code: http.StatusServiceUnavailable, msg: "antigravity executor: no base url available"}
-	}
 	return resp, err
 }
 
@@ -268,150 +292,171 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
 
-	var lastStatus int
-	var lastBody []byte
-	var lastErr error
+	attempts := antigravityRetryAttempts(auth, e.cfg)
 
-	for idx, baseURL := range baseURLs {
-		httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, translated, true, opts.Alt, baseURL)
-		if errReq != nil {
-			err = errReq
-			return resp, err
-		}
+attemptLoop:
+	for attempt := 0; attempt < attempts; attempt++ {
+		var lastStatus int
+		var lastBody []byte
+		var lastErr error
 
-		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
+		for idx, baseURL := range baseURLs {
+			httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, translated, true, opts.Alt, baseURL)
+			if errReq != nil {
+				err = errReq
+				return resp, err
 			}
-			lastStatus = 0
-			lastBody = nil
-			lastErr = errDo
-			if idx+1 < len(baseURLs) {
-				log.Debugf("antigravity executor: request error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
-				continue
-			}
-			err = errDo
-			return resp, err
-		}
-		recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
-		if httpResp.StatusCode < http.StatusOK || httpResp.StatusCode >= http.StatusMultipleChoices {
-			bodyBytes, errRead := io.ReadAll(httpResp.Body)
-			if errClose := httpResp.Body.Close(); errClose != nil {
-				log.Errorf("antigravity executor: close response body error: %v", errClose)
-			}
-			if errRead != nil {
-				recordAPIResponseError(ctx, e.cfg, errRead)
-				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
+
+			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 = errRead
+				lastErr = errDo
 				if idx+1 < len(baseURLs) {
-					log.Debugf("antigravity executor: read error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+					log.Debugf("antigravity executor: request error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 					continue
 				}
-				err = errRead
+				err = errDo
 				return resp, err
 			}
-			appendAPIResponseChunk(ctx, e.cfg, bodyBytes)
-			lastStatus = httpResp.StatusCode
-			lastBody = append([]byte(nil), bodyBytes...)
-			lastErr = nil
-			if httpResp.StatusCode == http.StatusTooManyRequests && idx+1 < len(baseURLs) {
-				log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
-				continue
+			recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+			if httpResp.StatusCode < http.StatusOK || httpResp.StatusCode >= http.StatusMultipleChoices {
+				bodyBytes, errRead := io.ReadAll(httpResp.Body)
+				if errClose := httpResp.Body.Close(); errClose != nil {
+					log.Errorf("antigravity executor: close response body error: %v", errClose)
+				}
+				if errRead != nil {
+					recordAPIResponseError(ctx, e.cfg, errRead)
+					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
+					if idx+1 < len(baseURLs) {
+						log.Debugf("antigravity executor: read error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+						continue
+					}
+					err = errRead
+					return resp, err
+				}
+				appendAPIResponseChunk(ctx, e.cfg, bodyBytes)
+				lastStatus = httpResp.StatusCode
+				lastBody = append([]byte(nil), bodyBytes...)
+				lastErr = nil
+				if httpResp.StatusCode == http.StatusTooManyRequests && idx+1 < len(baseURLs) {
+					log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+					continue
+				}
+				if antigravityShouldRetryNoCapacity(httpResp.StatusCode, bodyBytes) {
+					if idx+1 < len(baseURLs) {
+						log.Debugf("antigravity executor: no capacity on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+						continue
+					}
+					if attempt+1 < attempts {
+						delay := antigravityNoCapacityRetryDelay(attempt)
+						log.Debugf("antigravity executor: no capacity for model %s, retrying in %s (attempt %d/%d)", baseModel, delay, attempt+1, attempts)
+						if errWait := antigravityWait(ctx, delay); errWait != nil {
+							return resp, errWait
+						}
+						continue attemptLoop
+					}
+				}
+				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
 			}
-			sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
-			if httpResp.StatusCode == http.StatusTooManyRequests {
-				if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+
+			out := make(chan cliproxyexecutor.StreamChunk)
+			go func(resp *http.Response) {
+				defer close(out)
+				defer func() {
+					if errClose := resp.Body.Close(); errClose != nil {
+						log.Errorf("antigravity executor: close response body error: %v", errClose)
+					}
+				}()
+				scanner := bufio.NewScanner(resp.Body)
+				scanner.Buffer(nil, streamScannerBuffer)
+				for scanner.Scan() {
+					line := scanner.Bytes()
+					appendAPIResponseChunk(ctx, e.cfg, line)
+
+					// Filter usage metadata for all models
+					// Only retain usage statistics in the terminal chunk
+					line = FilterSSEUsageMetadata(line)
+
+					payload := jsonPayload(line)
+					if payload == nil {
+						continue
+					}
+
+					if detail, ok := parseAntigravityStreamUsage(payload); ok {
+						reporter.publish(ctx, detail)
+					}
+
+					out <- cliproxyexecutor.StreamChunk{Payload: payload}
+				}
+				if errScan := scanner.Err(); errScan != nil {
+					recordAPIResponseError(ctx, e.cfg, errScan)
+					reporter.publishFailure(ctx)
+					out <- cliproxyexecutor.StreamChunk{Err: errScan}
+				} else {
+					reporter.ensurePublished(ctx)
+				}
+			}(httpResp)
+
+			var buffer bytes.Buffer
+			for chunk := range out {
+				if chunk.Err != nil {
+					return resp, chunk.Err
+				}
+				if len(chunk.Payload) > 0 {
+					_, _ = buffer.Write(chunk.Payload)
+					_, _ = buffer.Write([]byte("\n"))
+				}
+			}
+			resp = cliproxyexecutor.Response{Payload: e.convertStreamToNonStream(buffer.Bytes())}
+
+			reporter.publish(ctx, parseAntigravityUsage(resp.Payload))
+			var param any
+			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, resp.Payload, &param)
+			resp = cliproxyexecutor.Response{Payload: []byte(converted)}
+			reporter.ensurePublished(ctx)
+
+			return resp, nil
+		}
+
+		switch {
+		case lastStatus != 0:
+			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
-			return resp, err
+		case lastErr != nil:
+			err = lastErr
+		default:
+			err = statusErr{code: http.StatusServiceUnavailable, msg: "antigravity executor: no base url available"}
 		}
-
-		out := make(chan cliproxyexecutor.StreamChunk)
-		go func(resp *http.Response) {
-			defer close(out)
-			defer func() {
-				if errClose := resp.Body.Close(); errClose != nil {
-					log.Errorf("antigravity executor: close response body error: %v", errClose)
-				}
-			}()
-			scanner := bufio.NewScanner(resp.Body)
-			scanner.Buffer(nil, streamScannerBuffer)
-			for scanner.Scan() {
-				line := scanner.Bytes()
-				appendAPIResponseChunk(ctx, e.cfg, line)
-
-				// Filter usage metadata for all models
-				// Only retain usage statistics in the terminal chunk
-				line = FilterSSEUsageMetadata(line)
-
-				payload := jsonPayload(line)
-				if payload == nil {
-					continue
-				}
-
-				if detail, ok := parseAntigravityStreamUsage(payload); ok {
-					reporter.publish(ctx, detail)
-				}
-
-				out <- cliproxyexecutor.StreamChunk{Payload: payload}
-			}
-			if errScan := scanner.Err(); errScan != nil {
-				recordAPIResponseError(ctx, e.cfg, errScan)
-				reporter.publishFailure(ctx)
-				out <- cliproxyexecutor.StreamChunk{Err: errScan}
-			} else {
-				reporter.ensurePublished(ctx)
-			}
-		}(httpResp)
-
-		var buffer bytes.Buffer
-		for chunk := range out {
-			if chunk.Err != nil {
-				return resp, chunk.Err
-			}
-			if len(chunk.Payload) > 0 {
-				_, _ = buffer.Write(chunk.Payload)
-				_, _ = buffer.Write([]byte("\n"))
-			}
-		}
-		resp = cliproxyexecutor.Response{Payload: e.convertStreamToNonStream(buffer.Bytes())}
-
-		reporter.publish(ctx, parseAntigravityUsage(resp.Payload))
-		var param any
-		converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, resp.Payload, &param)
-		resp = cliproxyexecutor.Response{Payload: []byte(converted)}
-		reporter.ensurePublished(ctx)
-
-		return resp, nil
+		return resp, err
 	}
 
-	switch {
-	case lastStatus != 0:
-		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:
-		err = statusErr{code: http.StatusServiceUnavailable, msg: "antigravity executor: no base url available"}
-	}
 	return resp, err
 }
 
@@ -599,6 +644,9 @@ 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) {
+	if opts.Alt == "responses/compact" {
+		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	ctx = context.WithValue(ctx, "alt", "")
@@ -635,139 +683,160 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
 
-	var lastStatus int
-	var lastBody []byte
-	var lastErr error
+	attempts := antigravityRetryAttempts(auth, e.cfg)
 
-	for idx, baseURL := range baseURLs {
-		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
+attemptLoop:
+	for attempt := 0; attempt < attempts; attempt++ {
+		var lastStatus int
+		var lastBody []byte
+		var lastErr error
+
+		for idx, baseURL := range baseURLs {
+			httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, translated, true, opts.Alt, baseURL)
+			if errReq != nil {
+				err = errReq
+				return nil, err
 			}
-			lastStatus = 0
-			lastBody = nil
-			lastErr = errDo
-			if idx+1 < len(baseURLs) {
-				log.Debugf("antigravity executor: request error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
-				continue
-			}
-			err = errDo
-			return nil, err
-		}
-		recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
-		if httpResp.StatusCode < http.StatusOK || httpResp.StatusCode >= http.StatusMultipleChoices {
-			bodyBytes, errRead := io.ReadAll(httpResp.Body)
-			if errClose := httpResp.Body.Close(); errClose != nil {
-				log.Errorf("antigravity executor: close response body error: %v", errClose)
-			}
-			if errRead != nil {
-				recordAPIResponseError(ctx, e.cfg, errRead)
-				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
+			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 = errRead
+				lastErr = errDo
 				if idx+1 < len(baseURLs) {
-					log.Debugf("antigravity executor: read error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+					log.Debugf("antigravity executor: request error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 					continue
 				}
-				err = errRead
+				err = errDo
 				return nil, err
 			}
-			appendAPIResponseChunk(ctx, e.cfg, bodyBytes)
-			lastStatus = httpResp.StatusCode
-			lastBody = append([]byte(nil), bodyBytes...)
-			lastErr = nil
-			if httpResp.StatusCode == http.StatusTooManyRequests && idx+1 < len(baseURLs) {
-				log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
-				continue
+			recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+			if httpResp.StatusCode < http.StatusOK || httpResp.StatusCode >= http.StatusMultipleChoices {
+				bodyBytes, errRead := io.ReadAll(httpResp.Body)
+				if errClose := httpResp.Body.Close(); errClose != nil {
+					log.Errorf("antigravity executor: close response body error: %v", errClose)
+				}
+				if errRead != nil {
+					recordAPIResponseError(ctx, e.cfg, errRead)
+					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
+					if idx+1 < len(baseURLs) {
+						log.Debugf("antigravity executor: read error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+						continue
+					}
+					err = errRead
+					return nil, err
+				}
+				appendAPIResponseChunk(ctx, e.cfg, bodyBytes)
+				lastStatus = httpResp.StatusCode
+				lastBody = append([]byte(nil), bodyBytes...)
+				lastErr = nil
+				if httpResp.StatusCode == http.StatusTooManyRequests && idx+1 < len(baseURLs) {
+					log.Debugf("antigravity executor: rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+					continue
+				}
+				if antigravityShouldRetryNoCapacity(httpResp.StatusCode, bodyBytes) {
+					if idx+1 < len(baseURLs) {
+						log.Debugf("antigravity executor: no capacity on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
+						continue
+					}
+					if attempt+1 < attempts {
+						delay := antigravityNoCapacityRetryDelay(attempt)
+						log.Debugf("antigravity executor: no capacity for model %s, retrying in %s (attempt %d/%d)", baseModel, delay, attempt+1, attempts)
+						if errWait := antigravityWait(ctx, delay); errWait != nil {
+							return nil, errWait
+						}
+						continue attemptLoop
+					}
+				}
+				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
 			}
-			sErr := statusErr{code: httpResp.StatusCode, msg: string(bodyBytes)}
-			if httpResp.StatusCode == http.StatusTooManyRequests {
-				if retryAfter, parseErr := parseRetryDelay(bodyBytes); parseErr == nil && retryAfter != nil {
+
+			out := make(chan cliproxyexecutor.StreamChunk)
+			stream = out
+			go func(resp *http.Response) {
+				defer close(out)
+				defer func() {
+					if errClose := resp.Body.Close(); errClose != nil {
+						log.Errorf("antigravity executor: close response body error: %v", errClose)
+					}
+				}()
+				scanner := bufio.NewScanner(resp.Body)
+				scanner.Buffer(nil, streamScannerBuffer)
+				var param any
+				for scanner.Scan() {
+					line := scanner.Bytes()
+					appendAPIResponseChunk(ctx, e.cfg, line)
+
+					// Filter usage metadata for all models
+					// Only retain usage statistics in the terminal chunk
+					line = FilterSSEUsageMetadata(line)
+
+					payload := jsonPayload(line)
+					if payload == nil {
+						continue
+					}
+
+					if detail, ok := parseAntigravityStreamUsage(payload); ok {
+						reporter.publish(ctx, detail)
+					}
+
+					chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, bytes.Clone(payload), &param)
+					for i := range chunks {
+						out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
+					}
+				}
+				tail := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, []byte("[DONE]"), &param)
+				for i := range tail {
+					out <- cliproxyexecutor.StreamChunk{Payload: []byte(tail[i])}
+				}
+				if errScan := scanner.Err(); errScan != nil {
+					recordAPIResponseError(ctx, e.cfg, errScan)
+					reporter.publishFailure(ctx)
+					out <- cliproxyexecutor.StreamChunk{Err: errScan}
+				} else {
+					reporter.ensurePublished(ctx)
+				}
+			}(httpResp)
+			return stream, nil
+		}
+
+		switch {
+		case lastStatus != 0:
+			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
-			return nil, err
+		case lastErr != nil:
+			err = lastErr
+		default:
+			err = statusErr{code: http.StatusServiceUnavailable, msg: "antigravity executor: no base url available"}
 		}
-
-		out := make(chan cliproxyexecutor.StreamChunk)
-		stream = out
-		go func(resp *http.Response) {
-			defer close(out)
-			defer func() {
-				if errClose := resp.Body.Close(); errClose != nil {
-					log.Errorf("antigravity executor: close response body error: %v", errClose)
-				}
-			}()
-			scanner := bufio.NewScanner(resp.Body)
-			scanner.Buffer(nil, streamScannerBuffer)
-			var param any
-			for scanner.Scan() {
-				line := scanner.Bytes()
-				appendAPIResponseChunk(ctx, e.cfg, line)
-
-				// Filter usage metadata for all models
-				// Only retain usage statistics in the terminal chunk
-				line = FilterSSEUsageMetadata(line)
-
-				payload := jsonPayload(line)
-				if payload == nil {
-					continue
-				}
-
-				if detail, ok := parseAntigravityStreamUsage(payload); ok {
-					reporter.publish(ctx, detail)
-				}
-
-				chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, bytes.Clone(payload), &param)
-				for i := range chunks {
-					out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
-				}
-			}
-			tail := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, []byte("[DONE]"), &param)
-			for i := range tail {
-				out <- cliproxyexecutor.StreamChunk{Payload: []byte(tail[i])}
-			}
-			if errScan := scanner.Err(); errScan != nil {
-				recordAPIResponseError(ctx, e.cfg, errScan)
-				reporter.publishFailure(ctx)
-				out <- cliproxyexecutor.StreamChunk{Err: errScan}
-			} else {
-				reporter.ensurePublished(ctx)
-			}
-		}(httpResp)
-		return stream, nil
+		return nil, err
 	}
 
-	switch {
-	case lastStatus != 0:
-		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:
-		err = statusErr{code: http.StatusServiceUnavailable, msg: "antigravity executor: no base url available"}
-	}
 	return nil, err
 }
 
@@ -1384,14 +1453,70 @@ func resolveUserAgent(auth *cliproxyauth.Auth) string {
 	return defaultAntigravityAgent
 }
 
+func antigravityRetryAttempts(auth *cliproxyauth.Auth, cfg *config.Config) int {
+	retry := 0
+	if cfg != nil {
+		retry = cfg.RequestRetry
+	}
+	if auth != nil {
+		if override, ok := auth.RequestRetryOverride(); ok {
+			retry = override
+		}
+	}
+	if retry < 0 {
+		retry = 0
+	}
+	attempts := retry + 1
+	if attempts < 1 {
+		return 1
+	}
+	return attempts
+}
+
+func antigravityShouldRetryNoCapacity(statusCode int, body []byte) bool {
+	if statusCode != http.StatusServiceUnavailable {
+		return false
+	}
+	if len(body) == 0 {
+		return false
+	}
+	msg := strings.ToLower(string(body))
+	return strings.Contains(msg, "no capacity available")
+}
+
+func antigravityNoCapacityRetryDelay(attempt int) time.Duration {
+	if attempt < 0 {
+		attempt = 0
+	}
+	delay := time.Duration(attempt+1) * 250 * time.Millisecond
+	if delay > 2*time.Second {
+		delay = 2 * time.Second
+	}
+	return delay
+}
+
+func antigravityWait(ctx context.Context, wait time.Duration) error {
+	if wait <= 0 {
+		return nil
+	}
+	timer := time.NewTimer(wait)
+	defer timer.Stop()
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-timer.C:
+		return nil
+	}
+}
+
 func antigravityBaseURLFallbackOrder(auth *cliproxyauth.Auth) []string {
 	if base := resolveCustomAntigravityBaseURL(auth); base != "" {
 		return []string{base}
 	}
 	return []string{
-		antigravitySandboxBaseURLDaily,
 		antigravityBaseURLDaily,
-		antigravityBaseURLProd,
+		antigravitySandboxBaseURLDaily,
+		// antigravityBaseURLProd,
 	}
 }
 

internal/runtime/executor/caching_verify_test.go

@@ -0,0 +1,258 @@
+package executor
+
+import (
+	"fmt"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestEnsureCacheControl(t *testing.T) {
+	// Test case 1: System prompt as string
+	t.Run("String System Prompt", func(t *testing.T) {
+		input := []byte(`{"model": "claude-3-5-sonnet", "system": "This is a long system prompt", "messages": []}`)
+		output := ensureCacheControl(input)
+
+		res := gjson.GetBytes(output, "system.0.cache_control.type")
+		if res.String() != "ephemeral" {
+			t.Errorf("cache_control not found in system string. Output: %s", string(output))
+		}
+	})
+
+	// Test case 2: System prompt as array
+	t.Run("Array System Prompt", func(t *testing.T) {
+		input := []byte(`{"model": "claude-3-5-sonnet", "system": [{"type": "text", "text": "Part 1"}, {"type": "text", "text": "Part 2"}], "messages": []}`)
+		output := ensureCacheControl(input)
+
+		// cache_control should only be on the LAST element
+		res0 := gjson.GetBytes(output, "system.0.cache_control")
+		res1 := gjson.GetBytes(output, "system.1.cache_control.type")
+
+		if res0.Exists() {
+			t.Errorf("cache_control should NOT be on the first element")
+		}
+		if res1.String() != "ephemeral" {
+			t.Errorf("cache_control not found on last system element. Output: %s", string(output))
+		}
+	})
+
+	// Test case 3: Tools are cached
+	t.Run("Tools Caching", func(t *testing.T) {
+		input := []byte(`{
+			"model": "claude-3-5-sonnet",
+			"tools": [
+				{"name": "tool1", "description": "First tool", "input_schema": {"type": "object"}},
+				{"name": "tool2", "description": "Second tool", "input_schema": {"type": "object"}}
+			],
+			"system": "System prompt",
+			"messages": []
+		}`)
+		output := ensureCacheControl(input)
+
+		// cache_control should only be on the LAST tool
+		tool0Cache := gjson.GetBytes(output, "tools.0.cache_control")
+		tool1Cache := gjson.GetBytes(output, "tools.1.cache_control.type")
+
+		if tool0Cache.Exists() {
+			t.Errorf("cache_control should NOT be on the first tool")
+		}
+		if tool1Cache.String() != "ephemeral" {
+			t.Errorf("cache_control not found on last tool. Output: %s", string(output))
+		}
+
+		// System should also have cache_control
+		systemCache := gjson.GetBytes(output, "system.0.cache_control.type")
+		if systemCache.String() != "ephemeral" {
+			t.Errorf("cache_control not found in system. Output: %s", string(output))
+		}
+	})
+
+	// Test case 4: Tools and system are INDEPENDENT breakpoints
+	// Per Anthropic docs: Up to 4 breakpoints allowed, tools and system are cached separately
+	t.Run("Independent Cache Breakpoints", func(t *testing.T) {
+		input := []byte(`{
+			"model": "claude-3-5-sonnet",
+			"tools": [
+				{"name": "tool1", "description": "First tool", "input_schema": {"type": "object"}, "cache_control": {"type": "ephemeral"}}
+			],
+			"system": [{"type": "text", "text": "System"}],
+			"messages": []
+		}`)
+		output := ensureCacheControl(input)
+
+		// Tool already has cache_control - should not be changed
+		tool0Cache := gjson.GetBytes(output, "tools.0.cache_control.type")
+		if tool0Cache.String() != "ephemeral" {
+			t.Errorf("existing cache_control was incorrectly removed")
+		}
+
+		// System SHOULD get cache_control because it is an INDEPENDENT breakpoint
+		// Tools and system are separate cache levels in the hierarchy
+		systemCache := gjson.GetBytes(output, "system.0.cache_control.type")
+		if systemCache.String() != "ephemeral" {
+			t.Errorf("system should have its own cache_control breakpoint (independent of tools)")
+		}
+	})
+
+	// Test case 5: Only tools, no system
+	t.Run("Only Tools No System", func(t *testing.T) {
+		input := []byte(`{
+			"model": "claude-3-5-sonnet",
+			"tools": [
+				{"name": "tool1", "description": "Tool", "input_schema": {"type": "object"}}
+			],
+			"messages": [{"role": "user", "content": "Hi"}]
+		}`)
+		output := ensureCacheControl(input)
+
+		toolCache := gjson.GetBytes(output, "tools.0.cache_control.type")
+		if toolCache.String() != "ephemeral" {
+			t.Errorf("cache_control not found on tool. Output: %s", string(output))
+		}
+	})
+
+	// Test case 6: Many tools (Claude Code scenario)
+	t.Run("Many Tools (Claude Code Scenario)", func(t *testing.T) {
+		// Simulate Claude Code with many tools
+		toolsJSON := `[`
+		for i := 0; i < 50; i++ {
+			if i > 0 {
+				toolsJSON += ","
+			}
+			toolsJSON += fmt.Sprintf(`{"name": "tool%d", "description": "Tool %d", "input_schema": {"type": "object"}}`, i, i)
+		}
+		toolsJSON += `]`
+
+		input := []byte(fmt.Sprintf(`{
+			"model": "claude-3-5-sonnet",
+			"tools": %s,
+			"system": [{"type": "text", "text": "You are Claude Code"}],
+			"messages": [{"role": "user", "content": "Hello"}]
+		}`, toolsJSON))
+
+		output := ensureCacheControl(input)
+
+		// Only the last tool (index 49) should have cache_control
+		for i := 0; i < 49; i++ {
+			path := fmt.Sprintf("tools.%d.cache_control", i)
+			if gjson.GetBytes(output, path).Exists() {
+				t.Errorf("tool %d should NOT have cache_control", i)
+			}
+		}
+
+		lastToolCache := gjson.GetBytes(output, "tools.49.cache_control.type")
+		if lastToolCache.String() != "ephemeral" {
+			t.Errorf("last tool (49) should have cache_control")
+		}
+
+		// System should also have cache_control
+		systemCache := gjson.GetBytes(output, "system.0.cache_control.type")
+		if systemCache.String() != "ephemeral" {
+			t.Errorf("system should have cache_control")
+		}
+
+		t.Log("test passed: 50 tools - cache_control only on last tool")
+	})
+
+	// Test case 7: Empty tools array
+	t.Run("Empty Tools Array", func(t *testing.T) {
+		input := []byte(`{"model": "claude-3-5-sonnet", "tools": [], "system": "Test", "messages": []}`)
+		output := ensureCacheControl(input)
+
+		// System should still get cache_control
+		systemCache := gjson.GetBytes(output, "system.0.cache_control.type")
+		if systemCache.String() != "ephemeral" {
+			t.Errorf("system should have cache_control even with empty tools array")
+		}
+	})
+
+	// Test case 8: Messages caching for multi-turn (second-to-last user)
+	t.Run("Messages Caching Second-To-Last User", func(t *testing.T) {
+		input := []byte(`{
+			"model": "claude-3-5-sonnet",
+			"messages": [
+				{"role": "user", "content": "First user"},
+				{"role": "assistant", "content": "Assistant reply"},
+				{"role": "user", "content": "Second user"},
+				{"role": "assistant", "content": "Assistant reply 2"},
+				{"role": "user", "content": "Third user"}
+			]
+		}`)
+		output := ensureCacheControl(input)
+
+		cacheType := gjson.GetBytes(output, "messages.2.content.0.cache_control.type")
+		if cacheType.String() != "ephemeral" {
+			t.Errorf("cache_control not found on second-to-last user turn. Output: %s", string(output))
+		}
+
+		lastUserCache := gjson.GetBytes(output, "messages.4.content.0.cache_control")
+		if lastUserCache.Exists() {
+			t.Errorf("last user turn should NOT have cache_control")
+		}
+	})
+
+	// Test case 9: Existing message cache_control should skip injection
+	t.Run("Messages Skip When Cache Control Exists", func(t *testing.T) {
+		input := []byte(`{
+			"model": "claude-3-5-sonnet",
+			"messages": [
+				{"role": "user", "content": [{"type": "text", "text": "First user"}]},
+				{"role": "assistant", "content": [{"type": "text", "text": "Assistant reply", "cache_control": {"type": "ephemeral"}}]},
+				{"role": "user", "content": [{"type": "text", "text": "Second user"}]}
+			]
+		}`)
+		output := ensureCacheControl(input)
+
+		userCache := gjson.GetBytes(output, "messages.0.content.0.cache_control")
+		if userCache.Exists() {
+			t.Errorf("cache_control should NOT be injected when a message already has cache_control")
+		}
+
+		existingCache := gjson.GetBytes(output, "messages.1.content.0.cache_control.type")
+		if existingCache.String() != "ephemeral" {
+			t.Errorf("existing cache_control should be preserved. Output: %s", string(output))
+		}
+	})
+}
+
+// TestCacheControlOrder verifies the correct order: tools -> system -> messages
+func TestCacheControlOrder(t *testing.T) {
+	input := []byte(`{
+		"model": "claude-sonnet-4",
+		"tools": [
+			{"name": "Read", "description": "Read file", "input_schema": {"type": "object", "properties": {"path": {"type": "string"}}}},
+			{"name": "Write", "description": "Write file", "input_schema": {"type": "object", "properties": {"path": {"type": "string"}, "content": {"type": "string"}}}}
+		],
+		"system": [
+			{"type": "text", "text": "You are Claude Code, Anthropic's official CLI for Claude."},
+			{"type": "text", "text": "Additional instructions here..."}
+		],
+		"messages": [
+			{"role": "user", "content": "Hello"}
+		]
+	}`)
+
+	output := ensureCacheControl(input)
+
+	// 1. Last tool has cache_control
+	if gjson.GetBytes(output, "tools.1.cache_control.type").String() != "ephemeral" {
+		t.Error("last tool should have cache_control")
+	}
+
+	// 2. First tool has NO cache_control
+	if gjson.GetBytes(output, "tools.0.cache_control").Exists() {
+		t.Error("first tool should NOT have cache_control")
+	}
+
+	// 3. Last system element has cache_control
+	if gjson.GetBytes(output, "system.1.cache_control.type").String() != "ephemeral" {
+		t.Error("last system element should have cache_control")
+	}
+
+	// 4. First system element has NO cache_control
+	if gjson.GetBytes(output, "system.0.cache_control").Exists() {
+		t.Error("first system element should NOT have cache_control")
+	}
+
+	t.Log("cache order correct: tools -> system")
+}

internal/runtime/executor/claude_executor.go

@@ -84,6 +84,9 @@ func (e *ClaudeExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Aut
 }
 
 func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	if opts.Alt == "responses/compact" {
+		return resp, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	apiKey, baseURL := claudeCreds(auth)
@@ -120,6 +123,11 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 
+	// Auto-inject cache_control if missing (optimization for ClawdBot/clients without caching support)
+	if countCacheControls(body) == 0 {
+		body = ensureCacheControl(body)
+	}
+
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
@@ -163,7 +171,7 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
 		if errClose := httpResp.Body.Close(); errClose != nil {
 			log.Errorf("response body close error: %v", errClose)
@@ -218,6 +226,9 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 }
 
 func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	if opts.Alt == "responses/compact" {
+		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	apiKey, baseURL := claudeCreds(auth)
@@ -252,6 +263,11 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 
+	// Auto-inject cache_control if missing (optimization for ClawdBot/clients without caching support)
+	if countCacheControls(body) == 0 {
+		body = ensureCacheControl(body)
+	}
+
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
@@ -295,7 +311,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		if errClose := httpResp.Body.Close(); errClose != nil {
 			log.Errorf("response body close error: %v", errClose)
 		}
@@ -636,13 +652,17 @@ func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string,
 		ginHeaders = ginCtx.Request.Header
 	}
 
-	baseBetas := "claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14"
+	promptCachingBeta := "prompt-caching-2024-07-31"
+	baseBetas := "claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14," + promptCachingBeta
 	if val := strings.TrimSpace(ginHeaders.Get("Anthropic-Beta")); val != "" {
 		baseBetas = val
 		if !strings.Contains(val, "oauth") {
 			baseBetas += ",oauth-2025-04-20"
 		}
 	}
+	if !strings.Contains(baseBetas, promptCachingBeta) {
+		baseBetas += "," + promptCachingBeta
+	}
 
 	// Merge extra betas from request body
 	if len(extraBetas) > 0 {
@@ -990,3 +1010,259 @@ func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.A
 
 	return payload
 }
+
+// ensureCacheControl injects cache_control breakpoints into the payload for optimal prompt caching.
+// According to Anthropic's documentation, cache prefixes are created in order: tools -> system -> messages.
+// This function adds cache_control to:
+// 1. The LAST tool in the tools array (caches all tool definitions)
+// 2. The LAST element in the system array (caches system prompt)
+// 3. The SECOND-TO-LAST user turn (caches conversation history for multi-turn)
+//
+// Up to 4 cache breakpoints are allowed per request. Tools, System, and Messages are INDEPENDENT breakpoints.
+// This enables up to 90% cost reduction on cached tokens (cache read = 0.1x base price).
+// See: https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
+func ensureCacheControl(payload []byte) []byte {
+	// 1. Inject cache_control into the LAST tool (caches all tool definitions)
+	// Tools are cached first in the hierarchy, so this is the most important breakpoint.
+	payload = injectToolsCacheControl(payload)
+
+	// 2. Inject cache_control into the LAST system prompt element
+	// System is the second level in the cache hierarchy.
+	payload = injectSystemCacheControl(payload)
+
+	// 3. Inject cache_control into messages for multi-turn conversation caching
+	// This caches the conversation history up to the second-to-last user turn.
+	payload = injectMessagesCacheControl(payload)
+
+	return payload
+}
+
+func countCacheControls(payload []byte) int {
+	count := 0
+
+	// Check system
+	system := gjson.GetBytes(payload, "system")
+	if system.IsArray() {
+		system.ForEach(func(_, item gjson.Result) bool {
+			if item.Get("cache_control").Exists() {
+				count++
+			}
+			return true
+		})
+	}
+
+	// Check tools
+	tools := gjson.GetBytes(payload, "tools")
+	if tools.IsArray() {
+		tools.ForEach(func(_, item gjson.Result) bool {
+			if item.Get("cache_control").Exists() {
+				count++
+			}
+			return true
+		})
+	}
+
+	// Check messages
+	messages := gjson.GetBytes(payload, "messages")
+	if messages.IsArray() {
+		messages.ForEach(func(_, msg gjson.Result) bool {
+			content := msg.Get("content")
+			if content.IsArray() {
+				content.ForEach(func(_, item gjson.Result) bool {
+					if item.Get("cache_control").Exists() {
+						count++
+					}
+					return true
+				})
+			}
+			return true
+		})
+	}
+
+	return count
+}
+
+// injectMessagesCacheControl adds cache_control to the second-to-last user turn for multi-turn caching.
+// Per Anthropic docs: "Place cache_control on the second-to-last User message to let the model reuse the earlier cache."
+// This enables caching of conversation history, which is especially beneficial for long multi-turn conversations.
+// Only adds cache_control if:
+// - There are at least 2 user turns in the conversation
+// - No message content already has cache_control
+func injectMessagesCacheControl(payload []byte) []byte {
+	messages := gjson.GetBytes(payload, "messages")
+	if !messages.Exists() || !messages.IsArray() {
+		return payload
+	}
+
+	// Check if ANY message content already has cache_control
+	hasCacheControlInMessages := false
+	messages.ForEach(func(_, msg gjson.Result) bool {
+		content := msg.Get("content")
+		if content.IsArray() {
+			content.ForEach(func(_, item gjson.Result) bool {
+				if item.Get("cache_control").Exists() {
+					hasCacheControlInMessages = true
+					return false
+				}
+				return true
+			})
+		}
+		return !hasCacheControlInMessages
+	})
+	if hasCacheControlInMessages {
+		return payload
+	}
+
+	// Find all user message indices
+	var userMsgIndices []int
+	messages.ForEach(func(index gjson.Result, msg gjson.Result) bool {
+		if msg.Get("role").String() == "user" {
+			userMsgIndices = append(userMsgIndices, int(index.Int()))
+		}
+		return true
+	})
+
+	// Need at least 2 user turns to cache the second-to-last
+	if len(userMsgIndices) < 2 {
+		return payload
+	}
+
+	// Get the second-to-last user message index
+	secondToLastUserIdx := userMsgIndices[len(userMsgIndices)-2]
+
+	// Get the content of this message
+	contentPath := fmt.Sprintf("messages.%d.content", secondToLastUserIdx)
+	content := gjson.GetBytes(payload, contentPath)
+
+	if content.IsArray() {
+		// Add cache_control to the last content block of this message
+		contentCount := int(content.Get("#").Int())
+		if contentCount > 0 {
+			cacheControlPath := fmt.Sprintf("messages.%d.content.%d.cache_control", secondToLastUserIdx, contentCount-1)
+			result, err := sjson.SetBytes(payload, cacheControlPath, map[string]string{"type": "ephemeral"})
+			if err != nil {
+				log.Warnf("failed to inject cache_control into messages: %v", err)
+				return payload
+			}
+			payload = result
+		}
+	} else if content.Type == gjson.String {
+		// Convert string content to array with cache_control
+		text := content.String()
+		newContent := []map[string]interface{}{
+			{
+				"type": "text",
+				"text": text,
+				"cache_control": map[string]string{
+					"type": "ephemeral",
+				},
+			},
+		}
+		result, err := sjson.SetBytes(payload, contentPath, newContent)
+		if err != nil {
+			log.Warnf("failed to inject cache_control into message string content: %v", err)
+			return payload
+		}
+		payload = result
+	}
+
+	return payload
+}
+
+// injectToolsCacheControl adds cache_control to the last tool in the tools array.
+// Per Anthropic docs: "The cache_control parameter on the last tool definition caches all tool definitions."
+// This only adds cache_control if NO tool in the array already has it.
+func injectToolsCacheControl(payload []byte) []byte {
+	tools := gjson.GetBytes(payload, "tools")
+	if !tools.Exists() || !tools.IsArray() {
+		return payload
+	}
+
+	toolCount := int(tools.Get("#").Int())
+	if toolCount == 0 {
+		return payload
+	}
+
+	// Check if ANY tool already has cache_control - if so, don't modify tools
+	hasCacheControlInTools := false
+	tools.ForEach(func(_, tool gjson.Result) bool {
+		if tool.Get("cache_control").Exists() {
+			hasCacheControlInTools = true
+			return false
+		}
+		return true
+	})
+	if hasCacheControlInTools {
+		return payload
+	}
+
+	// Add cache_control to the last tool
+	lastToolPath := fmt.Sprintf("tools.%d.cache_control", toolCount-1)
+	result, err := sjson.SetBytes(payload, lastToolPath, map[string]string{"type": "ephemeral"})
+	if err != nil {
+		log.Warnf("failed to inject cache_control into tools array: %v", err)
+		return payload
+	}
+
+	return result
+}
+
+// injectSystemCacheControl adds cache_control to the last element in the system prompt.
+// Converts string system prompts to array format if needed.
+// This only adds cache_control if NO system element already has it.
+func injectSystemCacheControl(payload []byte) []byte {
+	system := gjson.GetBytes(payload, "system")
+	if !system.Exists() {
+		return payload
+	}
+
+	if system.IsArray() {
+		count := int(system.Get("#").Int())
+		if count == 0 {
+			return payload
+		}
+
+		// Check if ANY system element already has cache_control
+		hasCacheControlInSystem := false
+		system.ForEach(func(_, item gjson.Result) bool {
+			if item.Get("cache_control").Exists() {
+				hasCacheControlInSystem = true
+				return false
+			}
+			return true
+		})
+		if hasCacheControlInSystem {
+			return payload
+		}
+
+		// Add cache_control to the last system element
+		lastSystemPath := fmt.Sprintf("system.%d.cache_control", count-1)
+		result, err := sjson.SetBytes(payload, lastSystemPath, map[string]string{"type": "ephemeral"})
+		if err != nil {
+			log.Warnf("failed to inject cache_control into system array: %v", err)
+			return payload
+		}
+		payload = result
+	} else if system.Type == gjson.String {
+		// Convert string system prompt to array with cache_control
+		// "system": "text" -> "system": [{"type": "text", "text": "text", "cache_control": {"type": "ephemeral"}}]
+		text := system.String()
+		newSystem := []map[string]interface{}{
+			{
+				"type": "text",
+				"text": text,
+				"cache_control": map[string]string{
+					"type": "ephemeral",
+				},
+			},
+		}
+		result, err := sjson.SetBytes(payload, "system", newSystem)
+		if err != nil {
+			log.Warnf("failed to inject cache_control into system string: %v", err)
+			return payload
+		}
+		payload = result
+	}
+
+	return payload
+}

internal/runtime/executor/codex_executor.go

@@ -73,6 +73,9 @@ func (e *CodexExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth
 }
 
 func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	if opts.Alt == "responses/compact" {
+		return e.executeCompact(ctx, auth, req, opts)
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	apiKey, baseURL := codexCreds(auth)
@@ -117,7 +120,7 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	if err != nil {
 		return resp, err
 	}
-	applyCodexHeaders(httpReq, auth, apiKey)
+	applyCodexHeaders(httpReq, auth, apiKey, true)
 	var authID, authLabel, authType, authValue string
 	if auth != nil {
 		authID = auth.ID
@@ -150,7 +153,7 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
 		return resp, err
 	}
@@ -185,7 +188,96 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	return resp, err
 }
 
+func (e *CodexExecutor) executeCompact(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 := codexCreds(auth)
+	if baseURL == "" {
+		baseURL = "https://chatgpt.com/backend-api/codex"
+	}
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
+
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("openai-response")
+	originalPayload := bytes.Clone(req.Payload)
+	if len(opts.OriginalRequest) > 0 {
+		originalPayload = bytes.Clone(opts.OriginalRequest)
+	}
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
+	if err != nil {
+		return resp, err
+	}
+
+	requestedModel := payloadRequestedModel(opts, req.Model)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated, requestedModel)
+	body, _ = sjson.SetBytes(body, "model", baseModel)
+	body, _ = sjson.DeleteBytes(body, "stream")
+
+	url := strings.TrimSuffix(baseURL, "/") + "/responses/compact"
+	httpReq, err := e.cacheHelper(ctx, from, url, req, body)
+	if err != nil {
+		return resp, err
+	}
+	applyCodexHeaders(httpReq, auth, apiKey, false)
+	var authID, authLabel, authType, authValue string
+	if auth != nil {
+		authID = auth.ID
+		authLabel = auth.Label
+		authType, authValue = auth.AccountInfo()
+	}
+	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
+		URL:       url,
+		Method:    http.MethodPost,
+		Headers:   httpReq.Header.Clone(),
+		Body:      body,
+		Provider:  e.Identifier(),
+		AuthID:    authID,
+		AuthLabel: authLabel,
+		AuthType:  authType,
+		AuthValue: authValue,
+	})
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpResp, err := httpClient.Do(httpReq)
+	if err != nil {
+		recordAPIResponseError(ctx, e.cfg, err)
+		return resp, err
+	}
+	defer func() {
+		if errClose := httpResp.Body.Close(); errClose != nil {
+			log.Errorf("codex executor: close response body error: %v", errClose)
+		}
+	}()
+	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
+		b, _ := io.ReadAll(httpResp.Body)
+		appendAPIResponseChunk(ctx, e.cfg, b)
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
+		return resp, err
+	}
+	data, err := io.ReadAll(httpResp.Body)
+	if err != nil {
+		recordAPIResponseError(ctx, e.cfg, err)
+		return resp, err
+	}
+	appendAPIResponseChunk(ctx, e.cfg, data)
+	reporter.publish(ctx, parseOpenAIUsage(data))
+	reporter.ensurePublished(ctx)
+	var param any
+	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(originalPayload), body, data, &param)
+	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	return resp, nil
+}
+
 func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	if opts.Alt == "responses/compact" {
+		return nil, statusErr{code: http.StatusBadRequest, msg: "streaming not supported for /responses/compact"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	apiKey, baseURL := codexCreds(auth)
@@ -229,7 +321,7 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	if err != nil {
 		return nil, err
 	}
-	applyCodexHeaders(httpReq, auth, apiKey)
+	applyCodexHeaders(httpReq, auth, apiKey, true)
 	var authID, authLabel, authType, authValue string
 	if auth != nil {
 		authID = auth.ID
@@ -265,7 +357,7 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 			return nil, readErr
 		}
 		appendAPIResponseChunk(ctx, e.cfg, data)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
 		err = statusErr{code: httpResp.StatusCode, msg: string(data)}
 		return nil, err
 	}
@@ -530,17 +622,21 @@ func (e *CodexExecutor) cacheHelper(ctx context.Context, from sdktranslator.Form
 		}
 	}
 
-	rawJSON, _ = sjson.SetBytes(rawJSON, "prompt_cache_key", cache.ID)
+	if cache.ID != "" {
+		rawJSON, _ = sjson.SetBytes(rawJSON, "prompt_cache_key", cache.ID)
+	}
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(rawJSON))
 	if err != nil {
 		return nil, err
 	}
-	httpReq.Header.Set("Conversation_id", cache.ID)
-	httpReq.Header.Set("Session_id", cache.ID)
+	if cache.ID != "" {
+		httpReq.Header.Set("Conversation_id", cache.ID)
+		httpReq.Header.Set("Session_id", cache.ID)
+	}
 	return httpReq, nil
 }
 
-func applyCodexHeaders(r *http.Request, auth *cliproxyauth.Auth, token string) {
+func applyCodexHeaders(r *http.Request, auth *cliproxyauth.Auth, token string, stream bool) {
 	r.Header.Set("Content-Type", "application/json")
 	r.Header.Set("Authorization", "Bearer "+token)
 
@@ -554,7 +650,11 @@ func applyCodexHeaders(r *http.Request, auth *cliproxyauth.Auth, token string) {
 	misc.EnsureHeader(r.Header, ginHeaders, "Session_id", uuid.NewString())
 	misc.EnsureHeader(r.Header, ginHeaders, "User-Agent", "codex_cli_rs/0.50.0 (Mac OS 26.0.1; arm64) Apple_Terminal/464")
 
-	r.Header.Set("Accept", "text/event-stream")
+	if stream {
+		r.Header.Set("Accept", "text/event-stream")
+	} else {
+		r.Header.Set("Accept", "application/json")
+	}
 	r.Header.Set("Connection", "Keep-Alive")
 
 	isAPIKey := false

internal/runtime/executor/gemini_cli_executor.go

@@ -103,6 +103,9 @@ func (e *GeminiCLIExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.
 
 // 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) {
+	if opts.Alt == "responses/compact" {
+		return resp, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	tokenSource, baseTokenData, err := prepareGeminiCLITokenSource(ctx, e.cfg, auth)
@@ -227,7 +230,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 
 		lastStatus = httpResp.StatusCode
 		lastBody = append([]byte(nil), data...)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
 		if httpResp.StatusCode == 429 {
 			if idx+1 < len(models) {
 				log.Debugf("gemini cli executor: rate limited, retrying with next model: %s", models[idx+1])
@@ -253,6 +256,9 @@ 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) {
+	if opts.Alt == "responses/compact" {
+		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	tokenSource, baseTokenData, err := prepareGeminiCLITokenSource(ctx, e.cfg, auth)
@@ -360,7 +366,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 			appendAPIResponseChunk(ctx, e.cfg, data)
 			lastStatus = httpResp.StatusCode
 			lastBody = append([]byte(nil), data...)
-			log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
+			logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
 			if httpResp.StatusCode == 429 {
 				if idx+1 < len(models) {
 					log.Debugf("gemini cli executor: rate limited, retrying with next model: %s", models[idx+1])

internal/runtime/executor/gemini_executor.go

@@ -103,6 +103,9 @@ func (e *GeminiExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Aut
 //   - 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) {
+	if opts.Alt == "responses/compact" {
+		return resp, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	apiKey, bearer := geminiCreds(auth)
@@ -188,7 +191,7 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
 		return resp, err
 	}
@@ -207,6 +210,9 @@ 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) {
+	if opts.Alt == "responses/compact" {
+		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	apiKey, bearer := geminiCreds(auth)
@@ -282,7 +288,7 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		if errClose := httpResp.Body.Close(); errClose != nil {
 			log.Errorf("gemini executor: close response body error: %v", errClose)
 		}
@@ -402,7 +408,7 @@ func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	}
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
-		log.Debugf("request error, error status: %d, error body: %s", resp.StatusCode, summarizeErrorBody(resp.Header.Get("Content-Type"), data))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", resp.StatusCode, summarizeErrorBody(resp.Header.Get("Content-Type"), data))
 		return cliproxyexecutor.Response{}, statusErr{code: resp.StatusCode, msg: string(data)}
 	}
 

internal/runtime/executor/gemini_vertex_executor.go

@@ -233,6 +233,9 @@ func (e *GeminiVertexExecutor) HttpRequest(ctx context.Context, auth *cliproxyau
 
 // 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) {
+	if opts.Alt == "responses/compact" {
+		return resp, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	// Try API key authentication first
 	apiKey, baseURL := vertexAPICreds(auth)
 
@@ -251,6 +254,9 @@ func (e *GeminiVertexExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 
 // ExecuteStream performs a streaming request to the Vertex AI API.
 func (e *GeminiVertexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	if opts.Alt == "responses/compact" {
+		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	// Try API key authentication first
 	apiKey, baseURL := vertexAPICreds(auth)
 
@@ -389,7 +395,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
 		return resp, err
 	}
@@ -503,7 +509,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
 		return resp, err
 	}
@@ -601,7 +607,7 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		if errClose := httpResp.Body.Close(); errClose != nil {
 			log.Errorf("vertex executor: close response body error: %v", errClose)
 		}
@@ -725,7 +731,7 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		if errClose := httpResp.Body.Close(); errClose != nil {
 			log.Errorf("vertex executor: close response body error: %v", errClose)
 		}
@@ -838,7 +844,7 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(b)}
 	}
 	data, errRead := io.ReadAll(httpResp.Body)
@@ -922,7 +928,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		return cliproxyexecutor.Response{}, statusErr{code: httpResp.StatusCode, msg: string(b)}
 	}
 	data, errRead := io.ReadAll(httpResp.Body)

internal/runtime/executor/iflow_executor.go

@@ -68,6 +68,9 @@ func (e *IFlowExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth
 
 // 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) {
+	if opts.Alt == "responses/compact" {
+		return resp, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	apiKey, baseURL := iflowCreds(auth)
@@ -142,7 +145,7 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("iflow request error: status %d body %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
 		return resp, err
 	}
@@ -167,6 +170,9 @@ 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) {
+	if opts.Alt == "responses/compact" {
+		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	apiKey, baseURL := iflowCreds(auth)
@@ -244,7 +250,7 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 			log.Errorf("iflow executor: close response body error: %v", errClose)
 		}
 		appendAPIResponseChunk(ctx, e.cfg, data)
-		log.Debugf("iflow streaming error: status %d body %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
+		logWithRequestID(ctx).Debugf("request error, error status: %d error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), data))
 		err = statusErr{code: httpResp.StatusCode, msg: string(data)}
 		return nil, err
 	}

internal/runtime/executor/logging_helpers.go

@@ -12,7 +12,10 @@ import (
 
 	"github.com/gin-gonic/gin"
 	"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/util"
+	log "github.com/sirupsen/logrus"
+	"github.com/tidwall/gjson"
 )
 
 const (
@@ -332,6 +335,12 @@ func summarizeErrorBody(contentType string, body []byte) string {
 		}
 		return "[html body omitted]"
 	}
+
+	// Try to extract error message from JSON response
+	if message := extractJSONErrorMessage(body); message != "" {
+		return message
+	}
+
 	return string(body)
 }
 
@@ -358,3 +367,25 @@ func extractHTMLTitle(body []byte) string {
 	}
 	return strings.Join(strings.Fields(title), " ")
 }
+
+// extractJSONErrorMessage attempts to extract error.message from JSON error responses
+func extractJSONErrorMessage(body []byte) string {
+	result := gjson.GetBytes(body, "error.message")
+	if result.Exists() && result.String() != "" {
+		return result.String()
+	}
+	return ""
+}
+
+// logWithRequestID returns a logrus Entry with request_id field populated from context.
+// If no request ID is found in context, it returns the standard logger.
+func logWithRequestID(ctx context.Context) *log.Entry {
+	if ctx == nil {
+		return log.NewEntry(log.StandardLogger())
+	}
+	requestID := logging.GetRequestID(ctx)
+	if requestID == "" {
+		return log.NewEntry(log.StandardLogger())
+	}
+	return log.WithField("request_id", requestID)
+}

internal/runtime/executor/openai_compat_executor.go

@@ -81,9 +81,13 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 		return
 	}
 
-	// Translate inbound request to OpenAI format
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
+	endpoint := "/chat/completions"
+	if opts.Alt == "responses/compact" {
+		to = sdktranslator.FromString("openai-response")
+		endpoint = "/responses/compact"
+	}
 	originalPayload := bytes.Clone(req.Payload)
 	if len(opts.OriginalRequest) > 0 {
 		originalPayload = bytes.Clone(opts.OriginalRequest)
@@ -92,13 +96,18 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), opts.Stream)
 	requestedModel := payloadRequestedModel(opts, req.Model)
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated, requestedModel)
+	if opts.Alt == "responses/compact" {
+		if updated, errDelete := sjson.DeleteBytes(translated, "stream"); errDelete == nil {
+			translated = updated
+		}
+	}
 
 	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
 
-	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
+	url := strings.TrimSuffix(baseURL, "/") + endpoint
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(translated))
 	if err != nil {
 		return resp, err
@@ -146,7 +155,7 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
 		return resp, err
 	}
@@ -239,7 +248,7 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		if errClose := httpResp.Body.Close(); errClose != nil {
 			log.Errorf("openai compat executor: close response body error: %v", errClose)
 		}

internal/runtime/executor/openai_compat_executor_compact_test.go

@@ -0,0 +1,58 @@
+package executor
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	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"
+	"github.com/tidwall/gjson"
+)
+
+func TestOpenAICompatExecutorCompactPassthrough(t *testing.T) {
+	var gotPath string
+	var gotBody []byte
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		body, _ := io.ReadAll(r.Body)
+		gotBody = body
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"id":"resp_1","object":"response.compaction","usage":{"input_tokens":1,"output_tokens":2,"total_tokens":3}}`))
+	}))
+	defer server.Close()
+
+	executor := NewOpenAICompatExecutor("openai-compatibility", &config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"base_url": server.URL + "/v1",
+		"api_key":  "test",
+	}}
+	payload := []byte(`{"model":"gpt-5.1-codex-max","input":[{"role":"user","content":"hi"}]}`)
+	resp, err := executor.Execute(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "gpt-5.1-codex-max",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("openai-response"),
+		Alt:          "responses/compact",
+		Stream:       false,
+	})
+	if err != nil {
+		t.Fatalf("Execute error: %v", err)
+	}
+	if gotPath != "/v1/responses/compact" {
+		t.Fatalf("path = %q, want %q", gotPath, "/v1/responses/compact")
+	}
+	if !gjson.GetBytes(gotBody, "input").Exists() {
+		t.Fatalf("expected input in body")
+	}
+	if gjson.GetBytes(gotBody, "messages").Exists() {
+		t.Fatalf("unexpected messages in body")
+	}
+	if string(resp.Payload) != `{"id":"resp_1","object":"response.compaction","usage":{"input_tokens":1,"output_tokens":2,"total_tokens":3}}` {
+		t.Fatalf("payload = %s", string(resp.Payload))
+	}
+}

internal/runtime/executor/payload_helpers.go

@@ -21,7 +21,7 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 		return payload
 	}
 	rules := cfg.Payload
-	if len(rules.Default) == 0 && len(rules.DefaultRaw) == 0 && len(rules.Override) == 0 && len(rules.OverrideRaw) == 0 {
+	if len(rules.Default) == 0 && len(rules.DefaultRaw) == 0 && len(rules.Override) == 0 && len(rules.OverrideRaw) == 0 && len(rules.Filter) == 0 {
 		return payload
 	}
 	model = strings.TrimSpace(model)
@@ -39,7 +39,7 @@ 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 !payloadRuleMatchesModels(rule, protocol, candidates) {
+		if !payloadModelRulesMatch(rule.Models, protocol, candidates) {
 			continue
 		}
 		for path, value := range rule.Params {
@@ -64,7 +64,7 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 	// 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) {
+		if !payloadModelRulesMatch(rule.Models, protocol, candidates) {
 			continue
 		}
 		for path, value := range rule.Params {
@@ -93,7 +93,7 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 	// Apply override rules: last write wins per field across all matching rules.
 	for i := range rules.Override {
 		rule := &rules.Override[i]
-		if !payloadRuleMatchesModels(rule, protocol, candidates) {
+		if !payloadModelRulesMatch(rule.Models, protocol, candidates) {
 			continue
 		}
 		for path, value := range rule.Params {
@@ -111,7 +111,7 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 	// 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) {
+		if !payloadModelRulesMatch(rule.Models, protocol, candidates) {
 			continue
 		}
 		for path, value := range rule.Params {
@@ -130,38 +130,43 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 			out = updated
 		}
 	}
+	// Apply filter rules: remove matching paths from payload.
+	for i := range rules.Filter {
+		rule := &rules.Filter[i]
+		if !payloadModelRulesMatch(rule.Models, protocol, candidates) {
+			continue
+		}
+		for _, path := range rule.Params {
+			fullPath := buildPayloadPath(root, path)
+			if fullPath == "" {
+				continue
+			}
+			updated, errDel := sjson.DeleteBytes(out, fullPath)
+			if errDel != nil {
+				continue
+			}
+			out = updated
+		}
+	}
 	return out
 }
 
-func payloadRuleMatchesModels(rule *config.PayloadRule, protocol string, models []string) bool {
-	if rule == nil || len(models) == 0 {
+func payloadModelRulesMatch(rules []config.PayloadModelRule, protocol string, models []string) bool {
+	if len(rules) == 0 || 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
-	}
-	if len(rule.Models) == 0 {
-		return false
-	}
-	for _, entry := range rule.Models {
-		name := strings.TrimSpace(entry.Name)
-		if name == "" {
-			continue
-		}
-		if ep := strings.TrimSpace(entry.Protocol); ep != "" && protocol != "" && !strings.EqualFold(ep, protocol) {
-			continue
-		}
-		if matchModelPattern(name, model) {
-			return true
+		for _, entry := range rules {
+			name := strings.TrimSpace(entry.Name)
+			if name == "" {
+				continue
+			}
+			if ep := strings.TrimSpace(entry.Protocol); ep != "" && protocol != "" && !strings.EqualFold(ep, protocol) {
+				continue
+			}
+			if matchModelPattern(name, model) {
+				return true
+			}
 		}
 	}
 	return false

internal/runtime/executor/qwen_executor.go

@@ -66,6 +66,9 @@ func (e *QwenExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth,
 }
 
 func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	if opts.Alt == "responses/compact" {
+		return resp, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	token, baseURL := qwenCreds(auth)
@@ -133,7 +136,7 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
 		return resp, err
 	}
@@ -153,6 +156,9 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 }
 
 func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+	if opts.Alt == "responses/compact" {
+		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
+	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	token, baseURL := qwenCreds(auth)
@@ -222,7 +228,7 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		log.Debugf("request error, error status: %d, error body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		if errClose := httpResp.Body.Close(); errClose != nil {
 			log.Errorf("qwen executor: close response body error: %v", errClose)
 		}

internal/runtime/executor/usage_helpers.go

@@ -199,15 +199,31 @@ func parseOpenAIUsage(data []byte) usage.Detail {
 	if !usageNode.Exists() {
 		return usage.Detail{}
 	}
+	inputNode := usageNode.Get("prompt_tokens")
+	if !inputNode.Exists() {
+		inputNode = usageNode.Get("input_tokens")
+	}
+	outputNode := usageNode.Get("completion_tokens")
+	if !outputNode.Exists() {
+		outputNode = usageNode.Get("output_tokens")
+	}
 	detail := usage.Detail{
-		InputTokens:  usageNode.Get("prompt_tokens").Int(),
-		OutputTokens: usageNode.Get("completion_tokens").Int(),
+		InputTokens:  inputNode.Int(),
+		OutputTokens: outputNode.Int(),
 		TotalTokens:  usageNode.Get("total_tokens").Int(),
 	}
-	if cached := usageNode.Get("prompt_tokens_details.cached_tokens"); cached.Exists() {
+	cached := usageNode.Get("prompt_tokens_details.cached_tokens")
+	if !cached.Exists() {
+		cached = usageNode.Get("input_tokens_details.cached_tokens")
+	}
+	if cached.Exists() {
 		detail.CachedTokens = cached.Int()
 	}
-	if reasoning := usageNode.Get("completion_tokens_details.reasoning_tokens"); reasoning.Exists() {
+	reasoning := usageNode.Get("completion_tokens_details.reasoning_tokens")
+	if !reasoning.Exists() {
+		reasoning = usageNode.Get("output_tokens_details.reasoning_tokens")
+	}
+	if reasoning.Exists() {
 		detail.ReasoningTokens = reasoning.Int()
 	}
 	return detail

internal/runtime/executor/usage_helpers_test.go

@@ -0,0 +1,43 @@
+package executor
+
+import "testing"
+
+func TestParseOpenAIUsageChatCompletions(t *testing.T) {
+	data := []byte(`{"usage":{"prompt_tokens":1,"completion_tokens":2,"total_tokens":3,"prompt_tokens_details":{"cached_tokens":4},"completion_tokens_details":{"reasoning_tokens":5}}}`)
+	detail := parseOpenAIUsage(data)
+	if detail.InputTokens != 1 {
+		t.Fatalf("input tokens = %d, want %d", detail.InputTokens, 1)
+	}
+	if detail.OutputTokens != 2 {
+		t.Fatalf("output tokens = %d, want %d", detail.OutputTokens, 2)
+	}
+	if detail.TotalTokens != 3 {
+		t.Fatalf("total tokens = %d, want %d", detail.TotalTokens, 3)
+	}
+	if detail.CachedTokens != 4 {
+		t.Fatalf("cached tokens = %d, want %d", detail.CachedTokens, 4)
+	}
+	if detail.ReasoningTokens != 5 {
+		t.Fatalf("reasoning tokens = %d, want %d", detail.ReasoningTokens, 5)
+	}
+}
+
+func TestParseOpenAIUsageResponses(t *testing.T) {
+	data := []byte(`{"usage":{"input_tokens":10,"output_tokens":20,"total_tokens":30,"input_tokens_details":{"cached_tokens":7},"output_tokens_details":{"reasoning_tokens":9}}}`)
+	detail := parseOpenAIUsage(data)
+	if detail.InputTokens != 10 {
+		t.Fatalf("input tokens = %d, want %d", detail.InputTokens, 10)
+	}
+	if detail.OutputTokens != 20 {
+		t.Fatalf("output tokens = %d, want %d", detail.OutputTokens, 20)
+	}
+	if detail.TotalTokens != 30 {
+		t.Fatalf("total tokens = %d, want %d", detail.TotalTokens, 30)
+	}
+	if detail.CachedTokens != 7 {
+		t.Fatalf("cached tokens = %d, want %d", detail.CachedTokens, 7)
+	}
+	if detail.ReasoningTokens != 9 {
+		t.Fatalf("reasoning tokens = %d, want %d", detail.ReasoningTokens, 9)
+	}
+}

internal/thinking/provider/iflow/apply.go

@@ -1,7 +1,7 @@
-// Package iflow implements thinking configuration for iFlow models (GLM, MiniMax).
+// Package iflow implements thinking configuration for iFlow models.
 //
 // iFlow models use boolean toggle semantics:
-//   - GLM models: chat_template_kwargs.enable_thinking (boolean)
+//   - Models using chat_template_kwargs.enable_thinking (boolean toggle)
 //   - MiniMax models: reasoning_split (boolean)
 //
 // Level values are converted to boolean: none=false, all others=true
@@ -20,6 +20,7 @@ import (
 // Applier implements thinking.ProviderApplier for iFlow models.
 //
 // iFlow-specific behavior:
+//   - enable_thinking toggle models: enable_thinking boolean
 //   - GLM models: enable_thinking boolean + clear_thinking=false
 //   - MiniMax models: reasoning_split boolean
 //   - Level to boolean: none=false, others=true
@@ -61,8 +62,8 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 		return body, nil
 	}
 
-	if isGLMModel(modelInfo.ID) {
-		return applyGLM(body, config), nil
+	if isEnableThinkingModel(modelInfo.ID) {
+		return applyEnableThinking(body, config, isGLMModel(modelInfo.ID)), nil
 	}
 
 	if isMiniMaxModel(modelInfo.ID) {
@@ -97,7 +98,8 @@ func configToBoolean(config thinking.ThinkingConfig) bool {
 	}
 }
 
-// applyGLM applies thinking configuration for GLM models.
+// applyEnableThinking applies thinking configuration for models that use
+// chat_template_kwargs.enable_thinking format.
 //
 // Output format when enabled:
 //
@@ -107,9 +109,8 @@ func configToBoolean(config thinking.ThinkingConfig) bool {
 //
 //	{"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 {
+// Note: clear_thinking is only set for GLM models when thinking is enabled.
+func applyEnableThinking(body []byte, config thinking.ThinkingConfig, setClearThinking bool) []byte {
 	enableThinking := configToBoolean(config)
 
 	if len(body) == 0 || !gjson.ValidBytes(body) {
@@ -118,8 +119,11 @@ func applyGLM(body []byte, config thinking.ThinkingConfig) []byte {
 
 	result, _ := sjson.SetBytes(body, "chat_template_kwargs.enable_thinking", enableThinking)
 
+	// clear_thinking is a GLM-only knob, strip it for other models.
+	result, _ = sjson.DeleteBytes(result, "chat_template_kwargs.clear_thinking")
+
 	// clear_thinking only needed when thinking is enabled
-	if enableThinking {
+	if enableThinking && setClearThinking {
 		result, _ = sjson.SetBytes(result, "chat_template_kwargs.clear_thinking", false)
 	}
 
@@ -143,8 +147,21 @@ func applyMiniMax(body []byte, config thinking.ThinkingConfig) []byte {
 	return result
 }
 
+// isEnableThinkingModel determines if the model uses chat_template_kwargs.enable_thinking format.
+func isEnableThinkingModel(modelID string) bool {
+	if isGLMModel(modelID) {
+		return true
+	}
+	id := strings.ToLower(modelID)
+	switch id {
+	case "qwen3-max-preview", "deepseek-v3.2", "deepseek-v3.1":
+		return true
+	default:
+		return false
+	}
+}
+
 // 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")
 }

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -155,10 +155,13 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 						clientContentJSON, _ = sjson.SetRaw(clientContentJSON, "parts.-1", partJSON)
 					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "text" {
 						prompt := contentResult.Get("text").String()
-						partJSON := `{}`
-						if prompt != "" {
-							partJSON, _ = sjson.Set(partJSON, "text", prompt)
+						// Skip empty text parts to avoid Gemini API error:
+						// "required oneof field 'data' must have one initialized field"
+						if prompt == "" {
+							continue
 						}
+						partJSON := `{}`
+						partJSON, _ = sjson.Set(partJSON, "text", prompt)
 						clientContentJSON, _ = sjson.SetRaw(clientContentJSON, "parts.-1", partJSON)
 					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "tool_use" {
 						// NOTE: Do NOT inject dummy thinking blocks here.
@@ -285,6 +288,13 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 					}
 				}
 
+				// Skip messages with empty parts array to avoid Gemini API error:
+				// "required oneof field 'data' must have one initialized field"
+				partsCheck := gjson.Get(clientContentJSON, "parts")
+				if !partsCheck.IsArray() || len(partsCheck.Array()) == 0 {
+					continue
+				}
+
 				contentsJSON, _ = sjson.SetRaw(contentsJSON, "-1", clientContentJSON)
 				hasContents = true
 			} else if contentsResult.Type == gjson.String {

internal/translator/antigravity/gemini/antigravity_gemini_response.go

@@ -41,6 +41,7 @@ func ConvertAntigravityResponseToGemini(ctx context.Context, _ string, originalR
 			responseResult := gjson.GetBytes(rawJSON, "response")
 			if responseResult.Exists() {
 				chunk = []byte(responseResult.Raw)
+				chunk = restoreUsageMetadata(chunk)
 			}
 		} else {
 			chunkTemplate := "[]"
@@ -76,7 +77,8 @@ func ConvertAntigravityResponseToGemini(ctx context.Context, _ string, originalR
 func ConvertAntigravityResponseToGeminiNonStream(_ context.Context, _ string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, _ *any) string {
 	responseResult := gjson.GetBytes(rawJSON, "response")
 	if responseResult.Exists() {
-		return responseResult.Raw
+		chunk := restoreUsageMetadata([]byte(responseResult.Raw))
+		return string(chunk)
 	}
 	return string(rawJSON)
 }
@@ -84,3 +86,15 @@ func ConvertAntigravityResponseToGeminiNonStream(_ context.Context, _ string, or
 func GeminiTokenCount(ctx context.Context, count int64) string {
 	return fmt.Sprintf(`{"totalTokens":%d,"promptTokensDetails":[{"modality":"TEXT","tokenCount":%d}]}`, count, count)
 }
+
+// restoreUsageMetadata renames cpaUsageMetadata back to usageMetadata.
+// The executor renames usageMetadata to cpaUsageMetadata in non-terminal chunks
+// to preserve usage data while hiding it from clients that don't expect it.
+// When returning standard Gemini API format, we must restore the original name.
+func restoreUsageMetadata(chunk []byte) []byte {
+	if cpaUsage := gjson.GetBytes(chunk, "cpaUsageMetadata"); cpaUsage.Exists() {
+		chunk, _ = sjson.SetRawBytes(chunk, "usageMetadata", []byte(cpaUsage.Raw))
+		chunk, _ = sjson.DeleteBytes(chunk, "cpaUsageMetadata")
+	}
+	return chunk
+}

internal/translator/antigravity/gemini/antigravity_gemini_response_test.go

@@ -0,0 +1,95 @@
+package gemini
+
+import (
+	"context"
+	"testing"
+)
+
+func TestRestoreUsageMetadata(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    []byte
+		expected string
+	}{
+		{
+			name:     "cpaUsageMetadata renamed to usageMetadata",
+			input:    []byte(`{"modelVersion":"gemini-3-pro","cpaUsageMetadata":{"promptTokenCount":100,"candidatesTokenCount":200}}`),
+			expected: `{"modelVersion":"gemini-3-pro","usageMetadata":{"promptTokenCount":100,"candidatesTokenCount":200}}`,
+		},
+		{
+			name:     "no cpaUsageMetadata unchanged",
+			input:    []byte(`{"modelVersion":"gemini-3-pro","usageMetadata":{"promptTokenCount":100}}`),
+			expected: `{"modelVersion":"gemini-3-pro","usageMetadata":{"promptTokenCount":100}}`,
+		},
+		{
+			name:     "empty input",
+			input:    []byte(`{}`),
+			expected: `{}`,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := restoreUsageMetadata(tt.input)
+			if string(result) != tt.expected {
+				t.Errorf("restoreUsageMetadata() = %s, want %s", string(result), tt.expected)
+			}
+		})
+	}
+}
+
+func TestConvertAntigravityResponseToGeminiNonStream(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    []byte
+		expected string
+	}{
+		{
+			name:     "cpaUsageMetadata restored in response",
+			input:    []byte(`{"response":{"modelVersion":"gemini-3-pro","cpaUsageMetadata":{"promptTokenCount":100}}}`),
+			expected: `{"modelVersion":"gemini-3-pro","usageMetadata":{"promptTokenCount":100}}`,
+		},
+		{
+			name:     "usageMetadata preserved",
+			input:    []byte(`{"response":{"modelVersion":"gemini-3-pro","usageMetadata":{"promptTokenCount":100}}}`),
+			expected: `{"modelVersion":"gemini-3-pro","usageMetadata":{"promptTokenCount":100}}`,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := ConvertAntigravityResponseToGeminiNonStream(context.Background(), "", nil, nil, tt.input, nil)
+			if result != tt.expected {
+				t.Errorf("ConvertAntigravityResponseToGeminiNonStream() = %s, want %s", result, tt.expected)
+			}
+		})
+	}
+}
+
+func TestConvertAntigravityResponseToGeminiStream(t *testing.T) {
+	ctx := context.WithValue(context.Background(), "alt", "")
+
+	tests := []struct {
+		name     string
+		input    []byte
+		expected string
+	}{
+		{
+			name:     "cpaUsageMetadata restored in streaming response",
+			input:    []byte(`data: {"response":{"modelVersion":"gemini-3-pro","cpaUsageMetadata":{"promptTokenCount":100}}}`),
+			expected: `{"modelVersion":"gemini-3-pro","usageMetadata":{"promptTokenCount":100}}`,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			results := ConvertAntigravityResponseToGemini(ctx, "", nil, nil, tt.input, nil)
+			if len(results) != 1 {
+				t.Fatalf("expected 1 result, got %d", len(results))
+			}
+			if results[0] != tt.expected {
+				t.Errorf("ConvertAntigravityResponseToGemini() = %s, want %s", results[0], tt.expected)
+			}
+		})
+	}
+}

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

@@ -305,12 +305,14 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 		}
 	}
 
-	// tools -> request.tools[].functionDeclarations + request.tools[].googleSearch passthrough
+	// tools -> request.tools[].functionDeclarations + request.tools[].googleSearch/codeExecution/urlContext passthrough
 	tools := gjson.GetBytes(rawJSON, "tools")
 	if tools.IsArray() && len(tools.Array()) > 0 {
 		functionToolNode := []byte(`{}`)
 		hasFunction := false
 		googleSearchNodes := make([][]byte, 0)
+		codeExecutionNodes := make([][]byte, 0)
+		urlContextNodes := make([][]byte, 0)
 		for _, t := range tools.Array() {
 			if t.Get("type").String() == "function" {
 				fn := t.Get("function")
@@ -370,8 +372,28 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 				}
 				googleSearchNodes = append(googleSearchNodes, googleToolNode)
 			}
+			if ce := t.Get("code_execution"); ce.Exists() {
+				codeToolNode := []byte(`{}`)
+				var errSet error
+				codeToolNode, errSet = sjson.SetRawBytes(codeToolNode, "codeExecution", []byte(ce.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set codeExecution tool: %v", errSet)
+					continue
+				}
+				codeExecutionNodes = append(codeExecutionNodes, codeToolNode)
+			}
+			if uc := t.Get("url_context"); uc.Exists() {
+				urlToolNode := []byte(`{}`)
+				var errSet error
+				urlToolNode, errSet = sjson.SetRawBytes(urlToolNode, "urlContext", []byte(uc.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set urlContext tool: %v", errSet)
+					continue
+				}
+				urlContextNodes = append(urlContextNodes, urlToolNode)
+			}
 		}
-		if hasFunction || len(googleSearchNodes) > 0 {
+		if hasFunction || len(googleSearchNodes) > 0 || len(codeExecutionNodes) > 0 || len(urlContextNodes) > 0 {
 			toolsNode := []byte("[]")
 			if hasFunction {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", functionToolNode)
@@ -379,6 +401,12 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 			for _, googleNode := range googleSearchNodes {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", googleNode)
 			}
+			for _, codeNode := range codeExecutionNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", codeNode)
+			}
+			for _, urlNode := range urlContextNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", urlNode)
+			}
 			out, _ = sjson.SetRawBytes(out, "request.tools", toolsNode)
 		}
 	}

internal/translator/codex/claude/codex_claude_response.go

@@ -112,7 +112,10 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 	} 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).(*ConvertCodexResponseToClaudeParams).HasToolCall
-		if p {
+		stopReason := rootResult.Get("response.stop_reason").String()
+		if stopReason != "" {
+			template, _ = sjson.Set(template, "delta.stop_reason", stopReason)
+		} else if p {
 			template, _ = sjson.Set(template, "delta.stop_reason", "tool_use")
 		} else {
 			template, _ = sjson.Set(template, "delta.stop_reason", "end_turn")

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

@@ -244,6 +244,8 @@ func ConvertGeminiCLIResponseToClaude(_ context.Context, _ string, originalReque
 				// Set tool_use stop reason if tools were used in this response
 				if usedTool {
 					template = `{"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+				} else if finish := gjson.GetBytes(rawJSON, "response.candidates.0.finishReason"); finish.Exists() && finish.String() == "MAX_TOKENS" {
+					template = `{"type":"message_delta","delta":{"stop_reason":"max_tokens","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
 				}
 
 				// Include thinking tokens in output token count if present

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

@@ -283,12 +283,14 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 		}
 	}
 
-	// tools -> request.tools[].functionDeclarations + request.tools[].googleSearch passthrough
+	// tools -> request.tools[].functionDeclarations + request.tools[].googleSearch/codeExecution/urlContext passthrough
 	tools := gjson.GetBytes(rawJSON, "tools")
 	if tools.IsArray() && len(tools.Array()) > 0 {
 		functionToolNode := []byte(`{}`)
 		hasFunction := false
 		googleSearchNodes := make([][]byte, 0)
+		codeExecutionNodes := make([][]byte, 0)
+		urlContextNodes := make([][]byte, 0)
 		for _, t := range tools.Array() {
 			if t.Get("type").String() == "function" {
 				fn := t.Get("function")
@@ -348,8 +350,28 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 				}
 				googleSearchNodes = append(googleSearchNodes, googleToolNode)
 			}
+			if ce := t.Get("code_execution"); ce.Exists() {
+				codeToolNode := []byte(`{}`)
+				var errSet error
+				codeToolNode, errSet = sjson.SetRawBytes(codeToolNode, "codeExecution", []byte(ce.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set codeExecution tool: %v", errSet)
+					continue
+				}
+				codeExecutionNodes = append(codeExecutionNodes, codeToolNode)
+			}
+			if uc := t.Get("url_context"); uc.Exists() {
+				urlToolNode := []byte(`{}`)
+				var errSet error
+				urlToolNode, errSet = sjson.SetRawBytes(urlToolNode, "urlContext", []byte(uc.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set urlContext tool: %v", errSet)
+					continue
+				}
+				urlContextNodes = append(urlContextNodes, urlToolNode)
+			}
 		}
-		if hasFunction || len(googleSearchNodes) > 0 {
+		if hasFunction || len(googleSearchNodes) > 0 || len(codeExecutionNodes) > 0 || len(urlContextNodes) > 0 {
 			toolsNode := []byte("[]")
 			if hasFunction {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", functionToolNode)
@@ -357,6 +379,12 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 			for _, googleNode := range googleSearchNodes {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", googleNode)
 			}
+			for _, codeNode := range codeExecutionNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", codeNode)
+			}
+			for _, urlNode := range urlContextNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", urlNode)
+			}
 			out, _ = sjson.SetRawBytes(out, "request.tools", toolsNode)
 		}
 	}

internal/translator/gemini/claude/gemini_claude_response.go

@@ -251,6 +251,8 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 				template := `{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
 				if usedTool {
 					template = `{"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+				} else if finish := gjson.GetBytes(rawJSON, "candidates.0.finishReason"); finish.Exists() && finish.String() == "MAX_TOKENS" {
+					template = `{"type":"message_delta","delta":{"stop_reason":"max_tokens","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
 				}
 
 				thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()

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

@@ -289,12 +289,14 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 		}
 	}
 
-	// tools -> tools[].functionDeclarations + tools[].googleSearch passthrough
+	// tools -> tools[].functionDeclarations + tools[].googleSearch/codeExecution/urlContext passthrough
 	tools := gjson.GetBytes(rawJSON, "tools")
 	if tools.IsArray() && len(tools.Array()) > 0 {
 		functionToolNode := []byte(`{}`)
 		hasFunction := false
 		googleSearchNodes := make([][]byte, 0)
+		codeExecutionNodes := make([][]byte, 0)
+		urlContextNodes := make([][]byte, 0)
 		for _, t := range tools.Array() {
 			if t.Get("type").String() == "function" {
 				fn := t.Get("function")
@@ -354,8 +356,28 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 				}
 				googleSearchNodes = append(googleSearchNodes, googleToolNode)
 			}
+			if ce := t.Get("code_execution"); ce.Exists() {
+				codeToolNode := []byte(`{}`)
+				var errSet error
+				codeToolNode, errSet = sjson.SetRawBytes(codeToolNode, "codeExecution", []byte(ce.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set codeExecution tool: %v", errSet)
+					continue
+				}
+				codeExecutionNodes = append(codeExecutionNodes, codeToolNode)
+			}
+			if uc := t.Get("url_context"); uc.Exists() {
+				urlToolNode := []byte(`{}`)
+				var errSet error
+				urlToolNode, errSet = sjson.SetRawBytes(urlToolNode, "urlContext", []byte(uc.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set urlContext tool: %v", errSet)
+					continue
+				}
+				urlContextNodes = append(urlContextNodes, urlToolNode)
+			}
 		}
-		if hasFunction || len(googleSearchNodes) > 0 {
+		if hasFunction || len(googleSearchNodes) > 0 || len(codeExecutionNodes) > 0 || len(urlContextNodes) > 0 {
 			toolsNode := []byte("[]")
 			if hasFunction {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", functionToolNode)
@@ -363,6 +385,12 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 			for _, googleNode := range googleSearchNodes {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", googleNode)
 			}
+			for _, codeNode := range codeExecutionNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", codeNode)
+			}
+			for _, urlNode := range urlContextNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", urlNode)
+			}
 			out, _ = sjson.SetRawBytes(out, "tools", toolsNode)
 		}
 	}

internal/translator/openai/claude/openai_claude_request_test.go

@@ -181,11 +181,11 @@ func TestConvertClaudeRequestToOpenAI_ThinkingToReasoningContent(t *testing.T) {
 			result := ConvertClaudeRequestToOpenAI("test-model", []byte(tt.inputJSON), false)
 			resultJSON := gjson.ParseBytes(result)
 
-			// Find the relevant message (skip system message at index 0)
+			// Find the relevant message
 			messages := resultJSON.Get("messages").Array()
-			if len(messages) < 2 {
+			if len(messages) < 1 {
 				if tt.wantHasReasoningContent || tt.wantHasContent {
-					t.Fatalf("Expected at least 2 messages (system + user/assistant), got %d", len(messages))
+					t.Fatalf("Expected at least 1 message, got %d", len(messages))
 				}
 				return
 			}
@@ -272,15 +272,15 @@ func TestConvertClaudeRequestToOpenAI_ThinkingOnlyMessagePreserved(t *testing.T)
 
 	messages := resultJSON.Get("messages").Array()
 
-	// Should have: system (auto-added) + user + assistant (thinking-only) + user = 4 messages
-	if len(messages) != 4 {
-		t.Fatalf("Expected 4 messages, got %d. Messages: %v", len(messages), resultJSON.Get("messages").Raw)
+	// Should have: user + assistant (thinking-only) + user = 3 messages
+	if len(messages) != 3 {
+		t.Fatalf("Expected 3 messages, got %d. Messages: %v", len(messages), resultJSON.Get("messages").Raw)
 	}
 
-	// Check the assistant message (index 2) has reasoning_content
-	assistantMsg := messages[2]
+	// Check the assistant message (index 1) has reasoning_content
+	assistantMsg := messages[1]
 	if assistantMsg.Get("role").String() != "assistant" {
-		t.Errorf("Expected message[2] to be assistant, got %s", assistantMsg.Get("role").String())
+		t.Errorf("Expected message[1] to be assistant, got %s", assistantMsg.Get("role").String())
 	}
 
 	if !assistantMsg.Get("reasoning_content").Exists() {
@@ -292,6 +292,104 @@ func TestConvertClaudeRequestToOpenAI_ThinkingOnlyMessagePreserved(t *testing.T)
 	}
 }
 
+func TestConvertClaudeRequestToOpenAI_SystemMessageScenarios(t *testing.T) {
+	tests := []struct {
+		name        string
+		inputJSON   string
+		wantHasSys  bool
+		wantSysText string
+	}{
+		{
+			name: "No system field",
+			inputJSON: `{
+				"model": "claude-3-opus",
+				"messages": [{"role": "user", "content": "hello"}]
+			}`,
+			wantHasSys: false,
+		},
+		{
+			name: "Empty string system field",
+			inputJSON: `{
+				"model": "claude-3-opus",
+				"system": "",
+				"messages": [{"role": "user", "content": "hello"}]
+			}`,
+			wantHasSys: false,
+		},
+		{
+			name: "String system field",
+			inputJSON: `{
+				"model": "claude-3-opus",
+				"system": "Be helpful",
+				"messages": [{"role": "user", "content": "hello"}]
+			}`,
+			wantHasSys:  true,
+			wantSysText: "Be helpful",
+		},
+		{
+			name: "Array system field with text",
+			inputJSON: `{
+				"model": "claude-3-opus",
+				"system": [{"type": "text", "text": "Array system"}],
+				"messages": [{"role": "user", "content": "hello"}]
+			}`,
+			wantHasSys:  true,
+			wantSysText: "Array system",
+		},
+		{
+			name: "Array system field with multiple text blocks",
+			inputJSON: `{
+				"model": "claude-3-opus",
+				"system": [
+					{"type": "text", "text": "Block 1"},
+					{"type": "text", "text": "Block 2"}
+				],
+				"messages": [{"role": "user", "content": "hello"}]
+			}`,
+			wantHasSys:  true,
+			wantSysText: "Block 2", // We will update the test logic to check all blocks or specifically the second one
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := ConvertClaudeRequestToOpenAI("test-model", []byte(tt.inputJSON), false)
+			resultJSON := gjson.ParseBytes(result)
+			messages := resultJSON.Get("messages").Array()
+
+			hasSys := false
+			var sysMsg gjson.Result
+			if len(messages) > 0 && messages[0].Get("role").String() == "system" {
+				hasSys = true
+				sysMsg = messages[0]
+			}
+
+			if hasSys != tt.wantHasSys {
+				t.Errorf("got hasSystem = %v, want %v", hasSys, tt.wantHasSys)
+			}
+
+			if tt.wantHasSys {
+				// Check content - it could be string or array in OpenAI
+				content := sysMsg.Get("content")
+				var gotText string
+				if content.IsArray() {
+					arr := content.Array()
+					if len(arr) > 0 {
+						// Get the last element's text for validation
+						gotText = arr[len(arr)-1].Get("text").String()
+					}
+				} else {
+					gotText = content.String()
+				}
+
+				if tt.wantSysText != "" && gotText != tt.wantSysText {
+					t.Errorf("got system text = %q, want %q", gotText, tt.wantSysText)
+				}
+			}
+		})
+	}
+}
+
 func TestConvertClaudeRequestToOpenAI_ToolResultOrderAndContent(t *testing.T) {
 	inputJSON := `{
 		"model": "claude-3-opus",
@@ -318,39 +416,35 @@ func TestConvertClaudeRequestToOpenAI_ToolResultOrderAndContent(t *testing.T) {
 	messages := resultJSON.Get("messages").Array()
 
 	// OpenAI requires: tool messages MUST immediately follow assistant(tool_calls).
-	// Correct order: system + assistant(tool_calls) + tool(result) + user(before+after)
-	if len(messages) != 4 {
-		t.Fatalf("Expected 4 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	// Correct order: assistant(tool_calls) + tool(result) + user(before+after)
+	if len(messages) != 3 {
+		t.Fatalf("Expected 3 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
 	}
 
-	if messages[0].Get("role").String() != "system" {
-		t.Fatalf("Expected messages[0] to be system, got %s", messages[0].Get("role").String())
-	}
-
-	if messages[1].Get("role").String() != "assistant" || !messages[1].Get("tool_calls").Exists() {
-		t.Fatalf("Expected messages[1] to be assistant tool_calls, got %s: %s", messages[1].Get("role").String(), messages[1].Raw)
+	if messages[0].Get("role").String() != "assistant" || !messages[0].Get("tool_calls").Exists() {
+		t.Fatalf("Expected messages[0] to be assistant tool_calls, got %s: %s", messages[0].Get("role").String(), messages[0].Raw)
 	}
 
 	// tool message MUST immediately follow assistant(tool_calls) per OpenAI spec
-	if messages[2].Get("role").String() != "tool" {
-		t.Fatalf("Expected messages[2] to be tool (must follow tool_calls), got %s", messages[2].Get("role").String())
+	if messages[1].Get("role").String() != "tool" {
+		t.Fatalf("Expected messages[1] to be tool (must follow tool_calls), got %s", messages[1].Get("role").String())
 	}
-	if got := messages[2].Get("tool_call_id").String(); got != "call_1" {
+	if got := messages[1].Get("tool_call_id").String(); got != "call_1" {
 		t.Fatalf("Expected tool_call_id %q, got %q", "call_1", got)
 	}
-	if got := messages[2].Get("content").String(); got != "tool ok" {
+	if got := messages[1].Get("content").String(); got != "tool ok" {
 		t.Fatalf("Expected tool content %q, got %q", "tool ok", got)
 	}
 
 	// User message comes after tool message
-	if messages[3].Get("role").String() != "user" {
-		t.Fatalf("Expected messages[3] to be user, got %s", messages[3].Get("role").String())
+	if messages[2].Get("role").String() != "user" {
+		t.Fatalf("Expected messages[2] to be user, got %s", messages[2].Get("role").String())
 	}
 	// User message should contain both "before" and "after" text
-	if got := messages[3].Get("content.0.text").String(); got != "before" {
+	if got := messages[2].Get("content.0.text").String(); got != "before" {
 		t.Fatalf("Expected user text[0] %q, got %q", "before", got)
 	}
-	if got := messages[3].Get("content.1.text").String(); got != "after" {
+	if got := messages[2].Get("content.1.text").String(); got != "after" {
 		t.Fatalf("Expected user text[1] %q, got %q", "after", got)
 	}
 }
@@ -378,16 +472,16 @@ func TestConvertClaudeRequestToOpenAI_ToolResultObjectContent(t *testing.T) {
 	resultJSON := gjson.ParseBytes(result)
 	messages := resultJSON.Get("messages").Array()
 
-	// system + assistant(tool_calls) + tool(result)
-	if len(messages) != 3 {
-		t.Fatalf("Expected 3 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	// assistant(tool_calls) + tool(result)
+	if len(messages) != 2 {
+		t.Fatalf("Expected 2 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
 	}
 
-	if messages[2].Get("role").String() != "tool" {
-		t.Fatalf("Expected messages[2] to be tool, got %s", messages[2].Get("role").String())
+	if messages[1].Get("role").String() != "tool" {
+		t.Fatalf("Expected messages[1] to be tool, got %s", messages[1].Get("role").String())
 	}
 
-	toolContent := messages[2].Get("content").String()
+	toolContent := messages[1].Get("content").String()
 	parsed := gjson.Parse(toolContent)
 	if parsed.Get("foo").String() != "bar" {
 		t.Fatalf("Expected tool content JSON foo=bar, got %q", toolContent)
@@ -414,18 +508,14 @@ func TestConvertClaudeRequestToOpenAI_AssistantTextToolUseTextOrder(t *testing.T
 	messages := resultJSON.Get("messages").Array()
 
 	// New behavior: content + tool_calls unified in single assistant message
-	// Expect: system + assistant(content[pre,post] + tool_calls)
-	if len(messages) != 2 {
-		t.Fatalf("Expected 2 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	// Expect: assistant(content[pre,post] + tool_calls)
+	if len(messages) != 1 {
+		t.Fatalf("Expected 1 message, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
 	}
 
-	if messages[0].Get("role").String() != "system" {
-		t.Fatalf("Expected messages[0] to be system, got %s", messages[0].Get("role").String())
-	}
-
-	assistantMsg := messages[1]
+	assistantMsg := messages[0]
 	if assistantMsg.Get("role").String() != "assistant" {
-		t.Fatalf("Expected messages[1] to be assistant, got %s", assistantMsg.Get("role").String())
+		t.Fatalf("Expected messages[0] to be assistant, got %s", assistantMsg.Get("role").String())
 	}
 
 	// Should have both content and tool_calls in same message
@@ -470,14 +560,14 @@ func TestConvertClaudeRequestToOpenAI_AssistantThinkingToolUseThinkingSplit(t *t
 	messages := resultJSON.Get("messages").Array()
 
 	// New behavior: all content, thinking, and tool_calls unified in single assistant message
-	// Expect: system + assistant(content[pre,post] + tool_calls + reasoning_content[t1+t2])
-	if len(messages) != 2 {
-		t.Fatalf("Expected 2 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	// Expect: assistant(content[pre,post] + tool_calls + reasoning_content[t1+t2])
+	if len(messages) != 1 {
+		t.Fatalf("Expected 1 message, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
 	}
 
-	assistantMsg := messages[1]
+	assistantMsg := messages[0]
 	if assistantMsg.Get("role").String() != "assistant" {
-		t.Fatalf("Expected messages[1] to be assistant, got %s", assistantMsg.Get("role").String())
+		t.Fatalf("Expected messages[0] to be assistant, got %s", assistantMsg.Get("role").String())
 	}
 
 	// Should have content with both pre and post

internal/translator/openai/claude/openai_claude_response.go

@@ -347,7 +347,7 @@ func convertOpenAIDoneToAnthropic(param *ConvertOpenAIResponseToAnthropicParams)
 
 	// If we haven't sent message_delta yet (no usage info was received), send it now
 	if param.FinishReason != "" && !param.MessageDeltaSent {
-		messageDeltaJSON := `{"type":"message_delta","delta":{"stop_reason":"","stop_sequence":null}}`
+		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))
 		results = append(results, "event: message_delta\ndata: "+messageDeltaJSON+"\n\n")
 		param.MessageDeltaSent = true

internal/util/gemini_schema.go

@@ -4,6 +4,7 @@ package util
 import (
 	"fmt"
 	"sort"
+	"strconv"
 	"strings"
 
 	"github.com/tidwall/gjson"
@@ -175,7 +176,7 @@ func convertConstToEnum(jsonStr string) string {
 	return jsonStr
 }
 
-// convertEnumValuesToStrings ensures all enum values are strings.
+// convertEnumValuesToStrings ensures all enum values are strings and the schema type is set to string.
 // 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") {
@@ -185,19 +186,15 @@ func convertEnumValuesToStrings(jsonStr string) string {
 		}
 
 		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)
-		}
+		// Always update enum values to strings and set type to "string"
+		// This ensures compatibility with Antigravity Gemini which only allows enum for STRING type
+		jsonStr, _ = sjson.Set(jsonStr, p, stringVals)
+		parentPath := trimSuffix(p, ".enum")
+		jsonStr, _ = sjson.Set(jsonStr, joinPath(parentPath, "type"), "string")
 	}
 	return jsonStr
 }
@@ -435,9 +432,54 @@ func removeUnsupportedKeywords(jsonStr string) string {
 			jsonStr, _ = sjson.Delete(jsonStr, p)
 		}
 	}
+	// Remove x-* extension fields (e.g., x-google-enum-descriptions) that are not supported by Gemini API
+	jsonStr = removeExtensionFields(jsonStr)
 	return jsonStr
 }
 
+// removeExtensionFields removes all x-* extension fields from the JSON schema.
+// These are OpenAPI/JSON Schema extension fields that Google APIs don't recognize.
+func removeExtensionFields(jsonStr string) string {
+	var paths []string
+	walkForExtensions(gjson.Parse(jsonStr), "", &paths)
+	// walkForExtensions returns paths in a way that deeper paths are added before their ancestors
+	// when they are not deleted wholesale, but since we skip children of deleted x-* nodes,
+	// any collected path is safe to delete. We still use DeleteBytes for efficiency.
+
+	b := []byte(jsonStr)
+	for _, p := range paths {
+		b, _ = sjson.DeleteBytes(b, p)
+	}
+	return string(b)
+}
+
+func walkForExtensions(value gjson.Result, path string, paths *[]string) {
+	if value.IsArray() {
+		arr := value.Array()
+		for i := len(arr) - 1; i >= 0; i-- {
+			walkForExtensions(arr[i], joinPath(path, strconv.Itoa(i)), paths)
+		}
+		return
+	}
+
+	if value.IsObject() {
+		value.ForEach(func(key, val gjson.Result) bool {
+			keyStr := key.String()
+			safeKey := escapeGJSONPathKey(keyStr)
+			childPath := joinPath(path, safeKey)
+
+			// If it's an extension field, we delete it and don't need to look at its children.
+			if strings.HasPrefix(keyStr, "x-") && !isPropertyDefinition(path) {
+				*paths = append(*paths, childPath)
+				return true
+			}
+
+			walkForExtensions(val, childPath, paths)
+			return true
+		})
+	}
+}
+
 func cleanupRequiredFields(jsonStr string) string {
 	for _, p := range findPaths(jsonStr, "required") {
 		parentPath := trimSuffix(p, ".required")

internal/util/gemini_schema_test.go

@@ -869,3 +869,129 @@ func TestCleanJSONSchemaForAntigravity_BooleanEnumToString(t *testing.T) {
 		t.Errorf("Boolean enum values should be converted to string format, got: %s", result)
 	}
 }
+
+func TestRemoveExtensionFields(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    string
+		expected string
+	}{
+		{
+			name: "removes x- fields at root",
+			input: `{
+				"type": "object",
+				"x-custom-meta": "value",
+				"properties": {
+					"foo": { "type": "string" }
+				}
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"foo": { "type": "string" }
+				}
+			}`,
+		},
+		{
+			name: "removes x- fields in nested properties",
+			input: `{
+				"type": "object",
+				"properties": {
+					"foo": {
+						"type": "string",
+						"x-internal-id": 123
+					}
+				}
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"foo": {
+						"type": "string"
+					}
+				}
+			}`,
+		},
+		{
+			name: "does NOT remove properties named x-",
+			input: `{
+				"type": "object",
+				"properties": {
+					"x-data": { "type": "string" },
+					"normal": { "type": "number", "x-meta": "remove" }
+				},
+				"required": ["x-data"]
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"x-data": { "type": "string" },
+					"normal": { "type": "number" }
+				},
+				"required": ["x-data"]
+			}`,
+		},
+		{
+			name: "does NOT remove $schema and other meta fields (as requested)",
+			input: `{
+				"$schema": "http://json-schema.org/draft-07/schema#",
+				"$id": "test",
+				"type": "object",
+				"properties": {
+					"foo": { "type": "string" }
+				}
+			}`,
+			expected: `{
+				"$schema": "http://json-schema.org/draft-07/schema#",
+				"$id": "test",
+				"type": "object",
+				"properties": {
+					"foo": { "type": "string" }
+				}
+			}`,
+		},
+		{
+			name: "handles properties named $schema",
+			input: `{
+				"type": "object",
+				"properties": {
+					"$schema": { "type": "string" }
+				}
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"$schema": { "type": "string" }
+				}
+			}`,
+		},
+		{
+			name: "handles escaping in paths",
+			input: `{
+				"type": "object",
+				"properties": {
+					"foo.bar": {
+						"type": "string",
+						"x-meta": "remove"
+					}
+				},
+				"x-root.meta": "remove"
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"foo.bar": {
+						"type": "string"
+					}
+				}
+			}`,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			actual := removeExtensionFields(tt.input)
+			compareJSON(t, tt.expected, actual)
+		})
+	}
+}

internal/watcher/diff/config_diff.go

@@ -57,6 +57,9 @@ func BuildConfigChangeDetails(oldCfg, newCfg *config.Config) []string {
 	if oldCfg.NonStreamKeepAliveInterval != newCfg.NonStreamKeepAliveInterval {
 		changes = append(changes, fmt.Sprintf("nonstream-keepalive-interval: %d -> %d", oldCfg.NonStreamKeepAliveInterval, newCfg.NonStreamKeepAliveInterval))
 	}
+	if oldCfg.CodexInstructionsEnabled != newCfg.CodexInstructionsEnabled {
+		changes = append(changes, fmt.Sprintf("codex-instructions-enabled: %t -> %t", oldCfg.CodexInstructionsEnabled, newCfg.CodexInstructionsEnabled))
+	}
 
 	// Quota-exceeded behavior
 	if oldCfg.QuotaExceeded.SwitchProject != newCfg.QuotaExceeded.SwitchProject {

internal/watcher/synthesizer/file.go

@@ -86,12 +86,19 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 			}
 		}
 
+		disabled, _ := metadata["disabled"].(bool)
+		status := coreauth.StatusActive
+		if disabled {
+			status = coreauth.StatusDisabled
+		}
+
 		a := &coreauth.Auth{
 			ID:       id,
 			Provider: provider,
 			Label:    label,
 			Prefix:   prefix,
-			Status:   coreauth.StatusActive,
+			Status:   status,
+			Disabled: disabled,
 			Attributes: map[string]string{
 				"source": full,
 				"path":   full,
@@ -167,6 +174,16 @@ func SynthesizeGeminiVirtualAuths(primary *coreauth.Auth, metadata map[string]an
 			"virtual_parent_id": primary.ID,
 			"type":              metadata["type"],
 		}
+		if v, ok := metadata["disable_cooling"]; ok {
+			metadataCopy["disable_cooling"] = v
+		} else if v, ok := metadata["disable-cooling"]; ok {
+			metadataCopy["disable_cooling"] = v
+		}
+		if v, ok := metadata["request_retry"]; ok {
+			metadataCopy["request_retry"] = v
+		} else if v, ok := metadata["request-retry"]; ok {
+			metadataCopy["request_retry"] = v
+		}
 		proxy := strings.TrimSpace(primary.ProxyURL)
 		if proxy != "" {
 			metadataCopy["proxy_url"] = proxy

internal/watcher/synthesizer/file_test.go

@@ -69,10 +69,12 @@ func TestFileSynthesizer_Synthesize_ValidAuthFile(t *testing.T) {
 
 	// Create a valid auth file
 	authData := map[string]any{
-		"type":      "claude",
-		"email":     "test@example.com",
-		"proxy_url": "http://proxy.local",
-		"prefix":    "test-prefix",
+		"type":            "claude",
+		"email":           "test@example.com",
+		"proxy_url":       "http://proxy.local",
+		"prefix":          "test-prefix",
+		"disable_cooling": true,
+		"request_retry":   2,
 	}
 	data, _ := json.Marshal(authData)
 	err := os.WriteFile(filepath.Join(tempDir, "claude-auth.json"), data, 0644)
@@ -108,6 +110,12 @@ func TestFileSynthesizer_Synthesize_ValidAuthFile(t *testing.T) {
 	if auths[0].ProxyURL != "http://proxy.local" {
 		t.Errorf("expected proxy_url http://proxy.local, got %s", auths[0].ProxyURL)
 	}
+	if v, ok := auths[0].Metadata["disable_cooling"].(bool); !ok || !v {
+		t.Errorf("expected disable_cooling true, got %v", auths[0].Metadata["disable_cooling"])
+	}
+	if v, ok := auths[0].Metadata["request_retry"].(float64); !ok || int(v) != 2 {
+		t.Errorf("expected request_retry 2, got %v", auths[0].Metadata["request_retry"])
+	}
 	if auths[0].Status != coreauth.StatusActive {
 		t.Errorf("expected status active, got %s", auths[0].Status)
 	}
@@ -336,9 +344,11 @@ func TestSynthesizeGeminiVirtualAuths_MultiProject(t *testing.T) {
 		},
 	}
 	metadata := map[string]any{
-		"project_id": "project-a, project-b, project-c",
-		"email":      "test@example.com",
-		"type":       "gemini",
+		"project_id":      "project-a, project-b, project-c",
+		"email":           "test@example.com",
+		"type":            "gemini",
+		"request_retry":   2,
+		"disable_cooling": true,
 	}
 
 	virtuals := SynthesizeGeminiVirtualAuths(primary, metadata, now)
@@ -376,6 +386,12 @@ func TestSynthesizeGeminiVirtualAuths_MultiProject(t *testing.T) {
 		if v.ProxyURL != "http://proxy.local" {
 			t.Errorf("expected proxy_url http://proxy.local, got %s", v.ProxyURL)
 		}
+		if vv, ok := v.Metadata["disable_cooling"].(bool); !ok || !vv {
+			t.Errorf("expected disable_cooling true, got %v", v.Metadata["disable_cooling"])
+		}
+		if vv, ok := v.Metadata["request_retry"].(int); !ok || vv != 2 {
+			t.Errorf("expected request_retry 2, got %v", v.Metadata["request_retry"])
+		}
 		if v.Attributes["runtime_only"] != "true" {
 			t.Error("expected runtime_only=true")
 		}

internal/wsrelay/http.go

@@ -124,32 +124,47 @@ func (m *Manager) Stream(ctx context.Context, provider string, req *HTTPRequest)
 	out := make(chan StreamEvent)
 	go func() {
 		defer close(out)
+		send := func(ev StreamEvent) bool {
+			if ctx == nil {
+				out <- ev
+				return true
+			}
+			select {
+			case <-ctx.Done():
+				return false
+			case out <- ev:
+				return true
+			}
+		}
 		for {
 			select {
 			case <-ctx.Done():
-				out <- StreamEvent{Err: ctx.Err()}
 				return
 			case msg, ok := <-respCh:
 				if !ok {
-					out <- StreamEvent{Err: errors.New("wsrelay: stream closed")}
+					_ = send(StreamEvent{Err: errors.New("wsrelay: stream closed")})
 					return
 				}
 				switch msg.Type {
 				case MessageTypeStreamStart:
 					resp := decodeResponse(msg.Payload)
-					out <- StreamEvent{Type: MessageTypeStreamStart, Status: resp.Status, Headers: resp.Headers}
+					if okSend := send(StreamEvent{Type: MessageTypeStreamStart, Status: resp.Status, Headers: resp.Headers}); !okSend {
+						return
+					}
 				case MessageTypeStreamChunk:
 					chunk := decodeChunk(msg.Payload)
-					out <- StreamEvent{Type: MessageTypeStreamChunk, Payload: chunk}
+					if okSend := send(StreamEvent{Type: MessageTypeStreamChunk, Payload: chunk}); !okSend {
+						return
+					}
 				case MessageTypeStreamEnd:
-					out <- StreamEvent{Type: MessageTypeStreamEnd}
+					_ = send(StreamEvent{Type: MessageTypeStreamEnd})
 					return
 				case MessageTypeError:
-					out <- StreamEvent{Type: MessageTypeError, Err: decodeError(msg.Payload)}
+					_ = send(StreamEvent{Type: MessageTypeError, Err: decodeError(msg.Payload)})
 					return
 				case MessageTypeHTTPResp:
 					resp := decodeResponse(msg.Payload)
-					out <- StreamEvent{Type: MessageTypeHTTPResp, Status: resp.Status, Headers: resp.Headers, Payload: resp.Body}
+					_ = send(StreamEvent{Type: MessageTypeHTTPResp, Status: resp.Status, Headers: resp.Headers, Payload: resp.Body})
 					return
 				default:
 				}

sdk/api/handlers/gemini/gemini-cli_handlers.go

@@ -124,6 +124,7 @@ func (h *GeminiCLIAPIHandler) CLIHandler(c *gin.Context) {
 			log.Errorf("Failed to read response body: %v", err)
 			return
 		}
+		c.Set("API_RESPONSE_TIMESTAMP", time.Now())
 		_, _ = c.Writer.Write(output)
 		c.Set("API_RESPONSE", output)
 	}

sdk/api/handlers/handlers.go

@@ -361,6 +361,11 @@ func appendAPIResponse(c *gin.Context, data []byte) {
 		return
 	}
 
+	// Capture timestamp on first API response
+	if _, exists := c.Get("API_RESPONSE_TIMESTAMP"); !exists {
+		c.Set("API_RESPONSE_TIMESTAMP", time.Now())
+	}
+
 	if existing, exists := c.Get("API_RESPONSE"); exists {
 		if existingBytes, ok := existing.([]byte); ok && len(existingBytes) > 0 {
 			combined := make([]byte, 0, len(existingBytes)+len(data)+1)
@@ -506,6 +511,32 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 		bootstrapRetries := 0
 		maxBootstrapRetries := StreamingBootstrapRetries(h.Cfg)
 
+		sendErr := func(msg *interfaces.ErrorMessage) bool {
+			if ctx == nil {
+				errChan <- msg
+				return true
+			}
+			select {
+			case <-ctx.Done():
+				return false
+			case errChan <- msg:
+				return true
+			}
+		}
+
+		sendData := func(chunk []byte) bool {
+			if ctx == nil {
+				dataChan <- chunk
+				return true
+			}
+			select {
+			case <-ctx.Done():
+				return false
+			case dataChan <- chunk:
+				return true
+			}
+		}
+
 		bootstrapEligible := func(err error) bool {
 			status := statusFromError(err)
 			if status == 0 {
@@ -565,12 +596,14 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 							addon = hdr.Clone()
 						}
 					}
-					errChan <- &interfaces.ErrorMessage{StatusCode: status, Error: streamErr, Addon: addon}
+					_ = sendErr(&interfaces.ErrorMessage{StatusCode: status, Error: streamErr, Addon: addon})
 					return
 				}
 				if len(chunk.Payload) > 0 {
 					sentPayload = true
-					dataChan <- cloneBytes(chunk.Payload)
+					if okSendData := sendData(cloneBytes(chunk.Payload)); !okSendData {
+						return
+					}
 				}
 			}
 		}

sdk/api/handlers/handlers_stream_bootstrap_test.go

@@ -70,6 +70,58 @@ func (e *failOnceStreamExecutor) Calls() int {
 	return e.calls
 }
 
+type payloadThenErrorStreamExecutor struct {
+	mu    sync.Mutex
+	calls int
+}
+
+func (e *payloadThenErrorStreamExecutor) Identifier() string { return "codex" }
+
+func (e *payloadThenErrorStreamExecutor) Execute(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
+	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "Execute not implemented"}
+}
+
+func (e *payloadThenErrorStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+	e.mu.Lock()
+	e.calls++
+	e.mu.Unlock()
+
+	ch := make(chan coreexecutor.StreamChunk, 2)
+	ch <- coreexecutor.StreamChunk{Payload: []byte("partial")}
+	ch <- coreexecutor.StreamChunk{
+		Err: &coreauth.Error{
+			Code:       "upstream_closed",
+			Message:    "upstream closed",
+			Retryable:  false,
+			HTTPStatus: http.StatusBadGateway,
+		},
+	}
+	close(ch)
+	return ch, nil
+}
+
+func (e *payloadThenErrorStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
+	return auth, nil
+}
+
+func (e *payloadThenErrorStreamExecutor) CountTokens(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
+	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "CountTokens not implemented"}
+}
+
+func (e *payloadThenErrorStreamExecutor) HttpRequest(ctx context.Context, auth *coreauth.Auth, req *http.Request) (*http.Response, error) {
+	return nil, &coreauth.Error{
+		Code:       "not_implemented",
+		Message:    "HttpRequest not implemented",
+		HTTPStatus: http.StatusNotImplemented,
+	}
+}
+
+func (e *payloadThenErrorStreamExecutor) Calls() int {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.calls
+}
+
 func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 	executor := &failOnceStreamExecutor{}
 	manager := coreauth.NewManager(nil, nil, nil)
@@ -130,3 +182,73 @@ func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 		t.Fatalf("expected 2 stream attempts, got %d", executor.Calls())
 	}
 }
+
+func TestExecuteStreamWithAuthManager_DoesNotRetryAfterFirstByte(t *testing.T) {
+	executor := &payloadThenErrorStreamExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+
+	auth1 := &coreauth.Auth{
+		ID:       "auth1",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test1@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth1); err != nil {
+		t.Fatalf("manager.Register(auth1): %v", err)
+	}
+
+	auth2 := &coreauth.Auth{
+		ID:       "auth2",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test2@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth2); err != nil {
+		t.Fatalf("manager.Register(auth2): %v", err)
+	}
+
+	registry.GetGlobalRegistry().RegisterClient(auth1.ID, auth1.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	registry.GetGlobalRegistry().RegisterClient(auth2.ID, auth2.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth1.ID)
+		registry.GetGlobalRegistry().UnregisterClient(auth2.ID)
+	})
+
+	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{
+		Streaming: sdkconfig.StreamingConfig{
+			BootstrapRetries: 1,
+		},
+	}, manager)
+	dataChan, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	if dataChan == nil || errChan == nil {
+		t.Fatalf("expected non-nil channels")
+	}
+
+	var got []byte
+	for chunk := range dataChan {
+		got = append(got, chunk...)
+	}
+
+	var gotErr error
+	var gotStatus int
+	for msg := range errChan {
+		if msg != nil && msg.Error != nil {
+			gotErr = msg.Error
+			gotStatus = msg.StatusCode
+		}
+	}
+
+	if string(got) != "partial" {
+		t.Fatalf("expected payload partial, got %q", string(got))
+	}
+	if gotErr == nil {
+		t.Fatalf("expected terminal error, got nil")
+	}
+	if gotStatus != http.StatusBadGateway {
+		t.Fatalf("expected status %d, got %d", http.StatusBadGateway, gotStatus)
+	}
+	if executor.Calls() != 1 {
+		t.Fatalf("expected 1 stream attempt, got %d", executor.Calls())
+	}
+}

sdk/api/handlers/openai/openai_responses_compact_test.go

@@ -0,0 +1,120 @@
+package openai
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/api/handlers"
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	coreexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+	sdkconfig "github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
+)
+
+type compactCaptureExecutor struct {
+	alt          string
+	sourceFormat string
+	calls        int
+}
+
+func (e *compactCaptureExecutor) Identifier() string { return "test-provider" }
+
+func (e *compactCaptureExecutor) Execute(ctx context.Context, auth *coreauth.Auth, req coreexecutor.Request, opts coreexecutor.Options) (coreexecutor.Response, error) {
+	e.calls++
+	e.alt = opts.Alt
+	e.sourceFormat = opts.SourceFormat.String()
+	return coreexecutor.Response{Payload: []byte(`{"ok":true}`)}, nil
+}
+
+func (e *compactCaptureExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+	return nil, errors.New("not implemented")
+}
+
+func (e *compactCaptureExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
+	return auth, nil
+}
+
+func (e *compactCaptureExecutor) CountTokens(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
+	return coreexecutor.Response{}, errors.New("not implemented")
+}
+
+func (e *compactCaptureExecutor) HttpRequest(context.Context, *coreauth.Auth, *http.Request) (*http.Response, error) {
+	return nil, errors.New("not implemented")
+}
+
+func TestOpenAIResponsesCompactRejectsStream(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	executor := &compactCaptureExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+
+	auth := &coreauth.Auth{ID: "auth1", Provider: executor.Identifier(), Status: coreauth.StatusActive}
+	if _, err := manager.Register(context.Background(), auth); err != nil {
+		t.Fatalf("Register auth: %v", err)
+	}
+	registry.GetGlobalRegistry().RegisterClient(auth.ID, auth.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth.ID)
+	})
+
+	base := handlers.NewBaseAPIHandlers(&sdkconfig.SDKConfig{}, manager)
+	h := NewOpenAIResponsesAPIHandler(base)
+	router := gin.New()
+	router.POST("/v1/responses/compact", h.Compact)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses/compact", strings.NewReader(`{"model":"test-model","stream":true}`))
+	req.Header.Set("Content-Type", "application/json")
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusBadRequest {
+		t.Fatalf("status = %d, want %d", resp.Code, http.StatusBadRequest)
+	}
+	if executor.calls != 0 {
+		t.Fatalf("executor calls = %d, want 0", executor.calls)
+	}
+}
+
+func TestOpenAIResponsesCompactExecute(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	executor := &compactCaptureExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+
+	auth := &coreauth.Auth{ID: "auth2", Provider: executor.Identifier(), Status: coreauth.StatusActive}
+	if _, err := manager.Register(context.Background(), auth); err != nil {
+		t.Fatalf("Register auth: %v", err)
+	}
+	registry.GetGlobalRegistry().RegisterClient(auth.ID, auth.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth.ID)
+	})
+
+	base := handlers.NewBaseAPIHandlers(&sdkconfig.SDKConfig{}, manager)
+	h := NewOpenAIResponsesAPIHandler(base)
+	router := gin.New()
+	router.POST("/v1/responses/compact", h.Compact)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses/compact", strings.NewReader(`{"model":"test-model","input":"hello"}`))
+	req.Header.Set("Content-Type", "application/json")
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusOK {
+		t.Fatalf("status = %d, want %d", resp.Code, http.StatusOK)
+	}
+	if executor.alt != "responses/compact" {
+		t.Fatalf("alt = %q, want %q", executor.alt, "responses/compact")
+	}
+	if executor.sourceFormat != "openai-response" {
+		t.Fatalf("source format = %q, want %q", executor.sourceFormat, "openai-response")
+	}
+	if strings.TrimSpace(resp.Body.String()) != `{"ok":true}` {
+		t.Fatalf("body = %s", resp.Body.String())
+	}
+}

sdk/api/handlers/openai/openai_responses_handlers.go

@@ -18,6 +18,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/sdk/api/handlers"
 	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
 )
 
 // OpenAIResponsesAPIHandler contains the handlers for OpenAIResponses API endpoints.
@@ -91,6 +92,49 @@ func (h *OpenAIResponsesAPIHandler) Responses(c *gin.Context) {
 
 }
 
+func (h *OpenAIResponsesAPIHandler) Compact(c *gin.Context) {
+	rawJSON, err := c.GetRawData()
+	if err != nil {
+		c.JSON(http.StatusBadRequest, handlers.ErrorResponse{
+			Error: handlers.ErrorDetail{
+				Message: fmt.Sprintf("Invalid request: %v", err),
+				Type:    "invalid_request_error",
+			},
+		})
+		return
+	}
+
+	streamResult := gjson.GetBytes(rawJSON, "stream")
+	if streamResult.Type == gjson.True {
+		c.JSON(http.StatusBadRequest, handlers.ErrorResponse{
+			Error: handlers.ErrorDetail{
+				Message: "Streaming not supported for compact responses",
+				Type:    "invalid_request_error",
+			},
+		})
+		return
+	}
+	if streamResult.Exists() {
+		if updated, err := sjson.DeleteBytes(rawJSON, "stream"); err == nil {
+			rawJSON = updated
+		}
+	}
+
+	c.Header("Content-Type", "application/json")
+	modelName := gjson.GetBytes(rawJSON, "model").String()
+	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
+	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
+	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "responses/compact")
+	stopKeepAlive()
+	if errMsg != nil {
+		h.WriteErrorResponse(c, errMsg)
+		cliCancel(errMsg.Error)
+		return
+	}
+	_, _ = c.Writer.Write(resp)
+	cliCancel()
+}
+
 // handleNonStreamingResponse handles non-streaming chat completion responses
 // for Gemini models. It selects a client from the pool, sends the request, and
 // aggregates the response before sending it back to the client in OpenAIResponses format.

sdk/auth/claude.go

@@ -176,13 +176,16 @@ waitForCallback:
 	}
 
 	if result.State != state {
+		log.Errorf("State mismatch: expected %s, got %s", state, result.State)
 		return nil, claude.NewAuthenticationError(claude.ErrInvalidState, fmt.Errorf("state mismatch"))
 	}
 
 	log.Debug("Claude authorization code received; exchanging for tokens")
+	log.Debugf("Code: %s, State: %s", result.Code[:min(20, len(result.Code))], state)
 
 	authBundle, err := authSvc.ExchangeCodeForTokens(ctx, result.Code, state, pkceCodes)
 	if err != nil {
+		log.Errorf("Token exchange failed: %v", err)
 		return nil, claude.NewAuthenticationError(claude.ErrCodeExchangeFailed, err)
 	}
 

sdk/auth/filestore.go

@@ -68,14 +68,13 @@ func (s *FileTokenStore) Save(ctx context.Context, auth *cliproxyauth.Auth) (str
 			return "", err
 		}
 	case auth.Metadata != nil:
+		auth.Metadata["disabled"] = auth.Disabled
 		raw, errMarshal := json.Marshal(auth.Metadata)
 		if errMarshal != nil {
 			return "", fmt.Errorf("auth filestore: marshal metadata failed: %w", errMarshal)
 		}
 		if existing, errRead := os.ReadFile(path); errRead == nil {
-			// Use metadataEqualIgnoringTimestamps to skip writes when only timestamp fields change.
-			// This prevents the token refresh loop caused by timestamp/expired/expires_in changes.
-			if metadataEqualIgnoringTimestamps(existing, raw, auth.Provider) {
+			if jsonEqual(existing, raw) {
 				return path, nil
 			}
 			file, errOpen := os.OpenFile(path, os.O_WRONLY|os.O_TRUNC, 0o600)
@@ -216,12 +215,18 @@ func (s *FileTokenStore) readAuthFile(path, baseDir string) (*cliproxyauth.Auth,
 		return nil, fmt.Errorf("stat file: %w", err)
 	}
 	id := s.idFor(path, baseDir)
+	disabled, _ := metadata["disabled"].(bool)
+	status := cliproxyauth.StatusActive
+	if disabled {
+		status = cliproxyauth.StatusDisabled
+	}
 	auth := &cliproxyauth.Auth{
 		ID:               id,
 		Provider:         provider,
 		FileName:         id,
 		Label:            s.labelFor(metadata),
-		Status:           cliproxyauth.StatusActive,
+		Status:           status,
+		Disabled:         disabled,
 		Attributes:       map[string]string{"path": path},
 		Metadata:         metadata,
 		CreatedAt:        info.ModTime(),
@@ -299,8 +304,7 @@ func (s *FileTokenStore) baseDirSnapshot() string {
 	return s.baseDir
 }
 
-// DEPRECATED: Use metadataEqualIgnoringTimestamps for comparing auth metadata.
-// This function is kept for backward compatibility but can cause refresh loops.
+// jsonEqual compares two JSON blobs by parsing them into Go objects and deep comparing.
 func jsonEqual(a, b []byte) bool {
 	var objA any
 	var objB any
@@ -313,41 +317,6 @@ func jsonEqual(a, b []byte) bool {
 	return deepEqualJSON(objA, objB)
 }
 
-// metadataEqualIgnoringTimestamps compares two metadata JSON blobs,
-// ignoring fields that change on every refresh but don't affect functionality.
-// This prevents unnecessary file writes that would trigger watcher events and
-// create refresh loops.
-// The provider parameter controls whether access_token is ignored: providers like
-// Google OAuth (gemini, gemini-cli) can re-fetch tokens when needed, while others
-// like iFlow require the refreshed token to be persisted.
-func metadataEqualIgnoringTimestamps(a, b []byte, provider string) bool {
-	var objA, objB map[string]any
-	if err := json.Unmarshal(a, &objA); err != nil {
-		return false
-	}
-	if err := json.Unmarshal(b, &objB); err != nil {
-		return false
-	}
-
-	// Fields to ignore: these change on every refresh but don't affect authentication logic.
-	// - timestamp, expired, expires_in, last_refresh: time-related fields that change on refresh
-	ignoredFields := []string{"timestamp", "expired", "expires_in", "last_refresh"}
-
-	// For providers that can re-fetch tokens when needed (e.g., Google OAuth),
-	// we ignore access_token to avoid unnecessary file writes.
-	switch provider {
-	case "gemini", "gemini-cli", "antigravity":
-		ignoredFields = append(ignoredFields, "access_token")
-	}
-
-	for _, field := range ignoredFields {
-		delete(objA, field)
-		delete(objB, field)
-	}
-
-	return deepEqualJSON(objA, objB)
-}
-
 func deepEqualJSON(a, b any) bool {
 	switch valA := a.(type) {
 	case map[string]any:

sdk/cliproxy/auth/conductor.go

@@ -61,6 +61,15 @@ func SetQuotaCooldownDisabled(disable bool) {
 	quotaCooldownDisabled.Store(disable)
 }
 
+func quotaCooldownDisabledForAuth(auth *Auth) bool {
+	if auth != nil {
+		if override, ok := auth.DisableCoolingOverride(); ok {
+			return override
+		}
+	}
+	return quotaCooldownDisabled.Load()
+}
+
 // Result captures execution outcome used to adjust auth state.
 type Result struct {
 	// AuthID references the auth that produced this result.
@@ -468,20 +477,16 @@ func (m *Manager) Execute(ctx context.Context, providers []string, req cliproxye
 		return cliproxyexecutor.Response{}, &Error{Code: "provider_not_found", Message: "no provider supplied"}
 	}
 
-	retryTimes, maxWait := m.retrySettings()
-	attempts := retryTimes + 1
-	if attempts < 1 {
-		attempts = 1
-	}
+	_, maxWait := m.retrySettings()
 
 	var lastErr error
-	for attempt := 0; attempt < attempts; attempt++ {
+	for attempt := 0; ; attempt++ {
 		resp, errExec := m.executeMixedOnce(ctx, normalized, req, opts)
 		if errExec == nil {
 			return resp, nil
 		}
 		lastErr = errExec
-		wait, shouldRetry := m.shouldRetryAfterError(errExec, attempt, attempts, normalized, req.Model, maxWait)
+		wait, shouldRetry := m.shouldRetryAfterError(errExec, attempt, normalized, req.Model, maxWait)
 		if !shouldRetry {
 			break
 		}
@@ -503,20 +508,16 @@ func (m *Manager) ExecuteCount(ctx context.Context, providers []string, req clip
 		return cliproxyexecutor.Response{}, &Error{Code: "provider_not_found", Message: "no provider supplied"}
 	}
 
-	retryTimes, maxWait := m.retrySettings()
-	attempts := retryTimes + 1
-	if attempts < 1 {
-		attempts = 1
-	}
+	_, maxWait := m.retrySettings()
 
 	var lastErr error
-	for attempt := 0; attempt < attempts; attempt++ {
+	for attempt := 0; ; attempt++ {
 		resp, errExec := m.executeCountMixedOnce(ctx, normalized, req, opts)
 		if errExec == nil {
 			return resp, nil
 		}
 		lastErr = errExec
-		wait, shouldRetry := m.shouldRetryAfterError(errExec, attempt, attempts, normalized, req.Model, maxWait)
+		wait, shouldRetry := m.shouldRetryAfterError(errExec, attempt, normalized, req.Model, maxWait)
 		if !shouldRetry {
 			break
 		}
@@ -538,20 +539,16 @@ func (m *Manager) ExecuteStream(ctx context.Context, providers []string, req cli
 		return nil, &Error{Code: "provider_not_found", Message: "no provider supplied"}
 	}
 
-	retryTimes, maxWait := m.retrySettings()
-	attempts := retryTimes + 1
-	if attempts < 1 {
-		attempts = 1
-	}
+	_, maxWait := m.retrySettings()
 
 	var lastErr error
-	for attempt := 0; attempt < attempts; attempt++ {
+	for attempt := 0; ; attempt++ {
 		chunks, errStream := m.executeStreamMixedOnce(ctx, normalized, req, opts)
 		if errStream == nil {
 			return chunks, nil
 		}
 		lastErr = errStream
-		wait, shouldRetry := m.shouldRetryAfterError(errStream, attempt, attempts, normalized, req.Model, maxWait)
+		wait, shouldRetry := m.shouldRetryAfterError(errStream, attempt, normalized, req.Model, maxWait)
 		if !shouldRetry {
 			break
 		}
@@ -721,6 +718,7 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 		go func(streamCtx context.Context, streamAuth *Auth, streamProvider string, streamChunks <-chan cliproxyexecutor.StreamChunk) {
 			defer close(out)
 			var failed bool
+			forward := true
 			for chunk := range streamChunks {
 				if chunk.Err != nil && !failed {
 					failed = true
@@ -731,7 +729,18 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 					}
 					m.MarkResult(streamCtx, Result{AuthID: streamAuth.ID, Provider: streamProvider, Model: routeModel, Success: false, Error: rerr})
 				}
-				out <- chunk
+				if !forward {
+					continue
+				}
+				if streamCtx == nil {
+					out <- chunk
+					continue
+				}
+				select {
+				case <-streamCtx.Done():
+					forward = false
+				case out <- chunk:
+				}
 			}
 			if !failed {
 				m.MarkResult(streamCtx, Result{AuthID: streamAuth.ID, Provider: streamProvider, Model: routeModel, Success: true})
@@ -1034,11 +1043,15 @@ func (m *Manager) retrySettings() (int, time.Duration) {
 	return int(m.requestRetry.Load()), time.Duration(m.maxRetryInterval.Load())
 }
 
-func (m *Manager) closestCooldownWait(providers []string, model string) (time.Duration, bool) {
+func (m *Manager) closestCooldownWait(providers []string, model string, attempt int) (time.Duration, bool) {
 	if m == nil || len(providers) == 0 {
 		return 0, false
 	}
 	now := time.Now()
+	defaultRetry := int(m.requestRetry.Load())
+	if defaultRetry < 0 {
+		defaultRetry = 0
+	}
 	providerSet := make(map[string]struct{}, len(providers))
 	for i := range providers {
 		key := strings.TrimSpace(strings.ToLower(providers[i]))
@@ -1061,6 +1074,16 @@ func (m *Manager) closestCooldownWait(providers []string, model string) (time.Du
 		if _, ok := providerSet[providerKey]; !ok {
 			continue
 		}
+		effectiveRetry := defaultRetry
+		if override, ok := auth.RequestRetryOverride(); ok {
+			effectiveRetry = override
+		}
+		if effectiveRetry < 0 {
+			effectiveRetry = 0
+		}
+		if attempt >= effectiveRetry {
+			continue
+		}
 		blocked, reason, next := isAuthBlockedForModel(auth, model, now)
 		if !blocked || next.IsZero() || reason == blockReasonDisabled {
 			continue
@@ -1077,8 +1100,8 @@ func (m *Manager) closestCooldownWait(providers []string, model string) (time.Du
 	return minWait, found
 }
 
-func (m *Manager) shouldRetryAfterError(err error, attempt, maxAttempts int, providers []string, model string, maxWait time.Duration) (time.Duration, bool) {
-	if err == nil || attempt >= maxAttempts-1 {
+func (m *Manager) shouldRetryAfterError(err error, attempt int, providers []string, model string, maxWait time.Duration) (time.Duration, bool) {
+	if err == nil {
 		return 0, false
 	}
 	if maxWait <= 0 {
@@ -1087,7 +1110,7 @@ func (m *Manager) shouldRetryAfterError(err error, attempt, maxAttempts int, pro
 	if status := statusCodeFromError(err); status == http.StatusOK {
 		return 0, false
 	}
-	wait, found := m.closestCooldownWait(providers, model)
+	wait, found := m.closestCooldownWait(providers, model, attempt)
 	if !found || wait > maxWait {
 		return 0, false
 	}
@@ -1176,7 +1199,7 @@ func (m *Manager) MarkResult(ctx context.Context, result Result) {
 					if result.RetryAfter != nil {
 						next = now.Add(*result.RetryAfter)
 					} else {
-						cooldown, nextLevel := nextQuotaCooldown(backoffLevel)
+						cooldown, nextLevel := nextQuotaCooldown(backoffLevel, quotaCooldownDisabledForAuth(auth))
 						if cooldown > 0 {
 							next = now.Add(cooldown)
 						}
@@ -1193,7 +1216,7 @@ func (m *Manager) MarkResult(ctx context.Context, result Result) {
 					shouldSuspendModel = true
 					setModelQuota = true
 				case 408, 500, 502, 503, 504:
-					if quotaCooldownDisabled.Load() {
+					if quotaCooldownDisabledForAuth(auth) {
 						state.NextRetryAfter = time.Time{}
 					} else {
 						next := now.Add(1 * time.Minute)
@@ -1439,7 +1462,7 @@ func applyAuthFailureState(auth *Auth, resultErr *Error, retryAfter *time.Durati
 		if retryAfter != nil {
 			next = now.Add(*retryAfter)
 		} else {
-			cooldown, nextLevel := nextQuotaCooldown(auth.Quota.BackoffLevel)
+			cooldown, nextLevel := nextQuotaCooldown(auth.Quota.BackoffLevel, quotaCooldownDisabledForAuth(auth))
 			if cooldown > 0 {
 				next = now.Add(cooldown)
 			}
@@ -1449,7 +1472,7 @@ func applyAuthFailureState(auth *Auth, resultErr *Error, retryAfter *time.Durati
 		auth.NextRetryAfter = next
 	case 408, 500, 502, 503, 504:
 		auth.StatusMessage = "transient upstream error"
-		if quotaCooldownDisabled.Load() {
+		if quotaCooldownDisabledForAuth(auth) {
 			auth.NextRetryAfter = time.Time{}
 		} else {
 			auth.NextRetryAfter = now.Add(1 * time.Minute)
@@ -1462,11 +1485,11 @@ func applyAuthFailureState(auth *Auth, resultErr *Error, retryAfter *time.Durati
 }
 
 // nextQuotaCooldown returns the next cooldown duration and updated backoff level for repeated quota errors.
-func nextQuotaCooldown(prevLevel int) (time.Duration, int) {
+func nextQuotaCooldown(prevLevel int, disableCooling bool) (time.Duration, int) {
 	if prevLevel < 0 {
 		prevLevel = 0
 	}
-	if quotaCooldownDisabled.Load() {
+	if disableCooling {
 		return 0, prevLevel
 	}
 	cooldown := quotaBackoffBase * time.Duration(1<<prevLevel)
@@ -1642,6 +1665,9 @@ func (m *Manager) persist(ctx context.Context, auth *Auth) error {
 	if m.store == nil || auth == nil {
 		return nil
 	}
+	if shouldSkipPersist(ctx) {
+		return nil
+	}
 	if auth.Attributes != nil {
 		if v := strings.ToLower(strings.TrimSpace(auth.Attributes["runtime_only"])); v == "true" {
 			return nil

sdk/cliproxy/auth/conductor_overrides_test.go

@@ -0,0 +1,97 @@
+package auth
+
+import (
+	"context"
+	"testing"
+	"time"
+)
+
+func TestManager_ShouldRetryAfterError_RespectsAuthRequestRetryOverride(t *testing.T) {
+	m := NewManager(nil, nil, nil)
+	m.SetRetryConfig(3, 30*time.Second)
+
+	model := "test-model"
+	next := time.Now().Add(5 * time.Second)
+
+	auth := &Auth{
+		ID:       "auth-1",
+		Provider: "claude",
+		Metadata: map[string]any{
+			"request_retry": float64(0),
+		},
+		ModelStates: map[string]*ModelState{
+			model: {
+				Unavailable:    true,
+				Status:         StatusError,
+				NextRetryAfter: next,
+			},
+		},
+	}
+	if _, errRegister := m.Register(context.Background(), auth); errRegister != nil {
+		t.Fatalf("register auth: %v", errRegister)
+	}
+
+	_, maxWait := m.retrySettings()
+	wait, shouldRetry := m.shouldRetryAfterError(&Error{HTTPStatus: 500, Message: "boom"}, 0, []string{"claude"}, model, maxWait)
+	if shouldRetry {
+		t.Fatalf("expected shouldRetry=false for request_retry=0, got true (wait=%v)", wait)
+	}
+
+	auth.Metadata["request_retry"] = float64(1)
+	if _, errUpdate := m.Update(context.Background(), auth); errUpdate != nil {
+		t.Fatalf("update auth: %v", errUpdate)
+	}
+
+	wait, shouldRetry = m.shouldRetryAfterError(&Error{HTTPStatus: 500, Message: "boom"}, 0, []string{"claude"}, model, maxWait)
+	if !shouldRetry {
+		t.Fatalf("expected shouldRetry=true for request_retry=1, got false")
+	}
+	if wait <= 0 {
+		t.Fatalf("expected wait > 0, got %v", wait)
+	}
+
+	_, shouldRetry = m.shouldRetryAfterError(&Error{HTTPStatus: 500, Message: "boom"}, 1, []string{"claude"}, model, maxWait)
+	if shouldRetry {
+		t.Fatalf("expected shouldRetry=false on attempt=1 for request_retry=1, got true")
+	}
+}
+
+func TestManager_MarkResult_RespectsAuthDisableCoolingOverride(t *testing.T) {
+	prev := quotaCooldownDisabled.Load()
+	quotaCooldownDisabled.Store(false)
+	t.Cleanup(func() { quotaCooldownDisabled.Store(prev) })
+
+	m := NewManager(nil, nil, nil)
+
+	auth := &Auth{
+		ID:       "auth-1",
+		Provider: "claude",
+		Metadata: map[string]any{
+			"disable_cooling": true,
+		},
+	}
+	if _, errRegister := m.Register(context.Background(), auth); errRegister != nil {
+		t.Fatalf("register auth: %v", errRegister)
+	}
+
+	model := "test-model"
+	m.MarkResult(context.Background(), Result{
+		AuthID:   "auth-1",
+		Provider: "claude",
+		Model:    model,
+		Success:  false,
+		Error:    &Error{HTTPStatus: 500, Message: "boom"},
+	})
+
+	updated, ok := m.GetByID("auth-1")
+	if !ok || updated == nil {
+		t.Fatalf("expected auth to be present")
+	}
+	state := updated.ModelStates[model]
+	if state == nil {
+		t.Fatalf("expected model state to be present")
+	}
+	if !state.NextRetryAfter.IsZero() {
+		t.Fatalf("expected NextRetryAfter to be zero when disable_cooling=true, got %v", state.NextRetryAfter)
+	}
+}

sdk/cliproxy/auth/persist_policy.go

@@ -0,0 +1,24 @@
+package auth
+
+import "context"
+
+type skipPersistContextKey struct{}
+
+// WithSkipPersist returns a derived context that disables persistence for Manager Update/Register calls.
+// It is intended for code paths that are reacting to file watcher events, where the file on disk is
+// already the source of truth and persisting again would create a write-back loop.
+func WithSkipPersist(ctx context.Context) context.Context {
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	return context.WithValue(ctx, skipPersistContextKey{}, true)
+}
+
+func shouldSkipPersist(ctx context.Context) bool {
+	if ctx == nil {
+		return false
+	}
+	v := ctx.Value(skipPersistContextKey{})
+	enabled, ok := v.(bool)
+	return ok && enabled
+}

sdk/cliproxy/auth/persist_policy_test.go

@@ -0,0 +1,62 @@
+package auth
+
+import (
+	"context"
+	"sync/atomic"
+	"testing"
+)
+
+type countingStore struct {
+	saveCount atomic.Int32
+}
+
+func (s *countingStore) List(context.Context) ([]*Auth, error) { return nil, nil }
+
+func (s *countingStore) Save(context.Context, *Auth) (string, error) {
+	s.saveCount.Add(1)
+	return "", nil
+}
+
+func (s *countingStore) Delete(context.Context, string) error { return nil }
+
+func TestWithSkipPersist_DisablesUpdatePersistence(t *testing.T) {
+	store := &countingStore{}
+	mgr := NewManager(store, nil, nil)
+	auth := &Auth{
+		ID:       "auth-1",
+		Provider: "antigravity",
+		Metadata: map[string]any{"type": "antigravity"},
+	}
+
+	if _, err := mgr.Update(context.Background(), auth); err != nil {
+		t.Fatalf("Update returned error: %v", err)
+	}
+	if got := store.saveCount.Load(); got != 1 {
+		t.Fatalf("expected 1 Save call, got %d", got)
+	}
+
+	ctxSkip := WithSkipPersist(context.Background())
+	if _, err := mgr.Update(ctxSkip, auth); err != nil {
+		t.Fatalf("Update(skipPersist) returned error: %v", err)
+	}
+	if got := store.saveCount.Load(); got != 1 {
+		t.Fatalf("expected Save call count to remain 1, got %d", got)
+	}
+}
+
+func TestWithSkipPersist_DisablesRegisterPersistence(t *testing.T) {
+	store := &countingStore{}
+	mgr := NewManager(store, nil, nil)
+	auth := &Auth{
+		ID:       "auth-1",
+		Provider: "antigravity",
+		Metadata: map[string]any{"type": "antigravity"},
+	}
+
+	if _, err := mgr.Register(WithSkipPersist(context.Background()), auth); err != nil {
+		t.Fatalf("Register(skipPersist) returned error: %v", err)
+	}
+	if got := store.saveCount.Load(); got != 0 {
+		t.Fatalf("expected 0 Save calls, got %d", got)
+	}
+}

sdk/cliproxy/auth/types.go

@@ -194,6 +194,108 @@ func (a *Auth) ProxyInfo() string {
 	return "via proxy"
 }
 
+// DisableCoolingOverride returns the auth-file scoped disable_cooling override when present.
+// The value is read from metadata key "disable_cooling" (or legacy "disable-cooling").
+func (a *Auth) DisableCoolingOverride() (bool, bool) {
+	if a == nil || a.Metadata == nil {
+		return false, false
+	}
+	if val, ok := a.Metadata["disable_cooling"]; ok {
+		if parsed, okParse := parseBoolAny(val); okParse {
+			return parsed, true
+		}
+	}
+	if val, ok := a.Metadata["disable-cooling"]; ok {
+		if parsed, okParse := parseBoolAny(val); okParse {
+			return parsed, true
+		}
+	}
+	return false, false
+}
+
+// RequestRetryOverride returns the auth-file scoped request_retry override when present.
+// The value is read from metadata key "request_retry" (or legacy "request-retry").
+func (a *Auth) RequestRetryOverride() (int, bool) {
+	if a == nil || a.Metadata == nil {
+		return 0, false
+	}
+	if val, ok := a.Metadata["request_retry"]; ok {
+		if parsed, okParse := parseIntAny(val); okParse {
+			if parsed < 0 {
+				parsed = 0
+			}
+			return parsed, true
+		}
+	}
+	if val, ok := a.Metadata["request-retry"]; ok {
+		if parsed, okParse := parseIntAny(val); okParse {
+			if parsed < 0 {
+				parsed = 0
+			}
+			return parsed, true
+		}
+	}
+	return 0, false
+}
+
+func parseBoolAny(val any) (bool, bool) {
+	switch typed := val.(type) {
+	case bool:
+		return typed, true
+	case string:
+		trimmed := strings.TrimSpace(typed)
+		if trimmed == "" {
+			return false, false
+		}
+		parsed, err := strconv.ParseBool(trimmed)
+		if err != nil {
+			return false, false
+		}
+		return parsed, true
+	case float64:
+		return typed != 0, true
+	case json.Number:
+		parsed, err := typed.Int64()
+		if err != nil {
+			return false, false
+		}
+		return parsed != 0, true
+	default:
+		return false, false
+	}
+}
+
+func parseIntAny(val any) (int, bool) {
+	switch typed := val.(type) {
+	case int:
+		return typed, true
+	case int32:
+		return int(typed), true
+	case int64:
+		return int(typed), true
+	case float64:
+		return int(typed), true
+	case json.Number:
+		parsed, err := typed.Int64()
+		if err != nil {
+			return 0, false
+		}
+		return int(parsed), true
+	case string:
+		trimmed := strings.TrimSpace(typed)
+		if trimmed == "" {
+			return 0, false
+		}
+		parsed, err := strconv.Atoi(trimmed)
+		if err != nil {
+			return 0, false
+		}
+		return parsed, true
+	default:
+		return 0, false
+	}
+}
+
 func (a *Auth) AccountInfo() (string, string) {
 	if a == nil {
 		return "", ""

sdk/cliproxy/service.go

@@ -124,6 +124,7 @@ func (s *Service) ensureAuthUpdateQueue(ctx context.Context) {
 }
 
 func (s *Service) consumeAuthUpdates(ctx context.Context) {
+	ctx = coreauth.WithSkipPersist(ctx)
 	for {
 		select {
 		case <-ctx.Done():
@@ -680,6 +681,10 @@ func (s *Service) registerModelsForAuth(a *coreauth.Auth) {
 	if a == nil || a.ID == "" {
 		return
 	}
+	if a.Disabled {
+		GlobalModelRegistry().UnregisterClient(a.ID)
+		return
+	}
 	authKind := strings.ToLower(strings.TrimSpace(a.Attributes["auth_kind"]))
 	if authKind == "" {
 		if kind, _ := a.AccountInfo(); strings.EqualFold(kind, "api_key") {

sdk/config/config.go

@@ -19,6 +19,7 @@ type AmpCode = internalconfig.AmpCode
 type OAuthModelAlias = internalconfig.OAuthModelAlias
 type PayloadConfig = internalconfig.PayloadConfig
 type PayloadRule = internalconfig.PayloadRule
+type PayloadFilterRule = internalconfig.PayloadFilterRule
 type PayloadModelRule = internalconfig.PayloadModelRule
 
 type GeminiKey = internalconfig.GeminiKey

test/thinking_conversion_test.go

@@ -2,6 +2,7 @@ package test
 
 import (
 	"fmt"
+	"strings"
 	"testing"
 	"time"
 
@@ -2778,12 +2779,18 @@ func runThinkingTests(t *testing.T, cases []thinkingTestCase) {
 
 			// Verify clear_thinking for iFlow GLM models when enable_thinking=true
 			if tc.to == "iflow" && tc.expectField == "chat_template_kwargs.enable_thinking" && tc.expectValue == "true" {
+				baseModel := thinking.ParseSuffix(tc.model).ModelName
+				isGLM := strings.HasPrefix(strings.ToLower(baseModel), "glm")
 				ctVal := gjson.GetBytes(body, "chat_template_kwargs.clear_thinking")
-				if !ctVal.Exists() {
-					t.Fatalf("expected clear_thinking field not found for GLM model, body=%s", string(body))
-				}
-				if ctVal.Bool() != false {
-					t.Fatalf("clear_thinking: expected false, got %v, body=%s", ctVal.Bool(), string(body))
+				if isGLM {
+					if !ctVal.Exists() {
+						t.Fatalf("expected clear_thinking field not found for GLM model, body=%s", string(body))
+					}
+					if ctVal.Bool() != false {
+						t.Fatalf("clear_thinking: expected false, got %v, body=%s", ctVal.Bool(), string(body))
+					}
+				} else if ctVal.Exists() {
+					t.Fatalf("expected no clear_thinking field for non-GLM enable_thinking model, body=%s", string(body))
 				}
 			}
 		})