Merge PR #424 from router-for-me/main

Fetch model catalog from network

.github/workflows/docker-image.yml

@@ -16,6 +16,8 @@ jobs:
     steps:
       - name: Checkout
         uses: actions/checkout@v4
+      - name: Refresh models catalog
+        run: curl -fsSL https://raw.githubusercontent.com/router-for-me/models/refs/heads/main/models.json -o internal/registry/models/models.json
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Login to DockerHub
@@ -47,6 +49,8 @@ jobs:
     steps:
       - name: Checkout
         uses: actions/checkout@v4
+      - name: Refresh models catalog
+        run: curl -fsSL https://raw.githubusercontent.com/router-for-me/models/refs/heads/main/models.json -o internal/registry/models/models.json
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Login to DockerHub

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

@@ -12,6 +12,8 @@ jobs:
     steps:
       - name: Checkout
         uses: actions/checkout@v4
+      - name: Refresh models catalog
+        run: curl -fsSL https://raw.githubusercontent.com/router-for-me/models/refs/heads/main/models.json -o internal/registry/models/models.json
       - name: Set up Go
         uses: actions/setup-go@v5
         with:

.github/workflows/release.yaml

@@ -16,6 +16,8 @@ jobs:
       - uses: actions/checkout@v4
         with:
           fetch-depth: 0
+      - name: Refresh models catalog
+        run: curl -fsSL https://raw.githubusercontent.com/router-for-me/models/refs/heads/main/models.json -o internal/registry/models/models.json
       - run: git fetch --force --tags
       - uses: actions/setup-go@v4
         with:

README.md

@@ -8,123 +8,6 @@ All third-party provider support is maintained by community contributors; CLIPro
 
 The Plus release stays in lockstep with the mainline features.
 
-## Differences from the Mainline
-
-[![z.ai](https://assets.router-for.me/english-5.png)](https://z.ai/subscribe?ic=8JVLJQFSKB)
-
-## New Features (Plus Enhanced)
-
-GLM CODING PLAN is a subscription service designed for AI coding, starting at just $10/month. It provides access to their flagship GLM-4.7 & （GLM-5 Only Available  for Pro Users）model across 10+ popular AI coding tools (Claude Code, Cline, Roo Code, etc.), offering developers top-tier, fast, and stable coding experiences.
-
-## Kiro Authentication
-
-### CLI Login
-
-> **Note:** Google/GitHub login is not available for third-party applications due to AWS Cognito restrictions.
-
-**AWS Builder ID** (recommended):
-
-```bash
-# Device code flow
-./CLIProxyAPI --kiro-aws-login
-
-# Authorization code flow
-./CLIProxyAPI --kiro-aws-authcode
-```
-
-**Import token from Kiro IDE:**
-
-```bash
-./CLIProxyAPI --kiro-import
-```
-
-To get a token from Kiro IDE:
-
-1. Open Kiro IDE and login with Google (or GitHub)
-2. Find the token file: `~/.kiro/kiro-auth-token.json`
-3. Run: `./CLIProxyAPI --kiro-import`
-
-**AWS IAM Identity Center (IDC):**
-
-```bash
-./CLIProxyAPI --kiro-idc-login --kiro-idc-start-url https://d-xxxxxxxxxx.awsapps.com/start
-
-# Specify region
-./CLIProxyAPI --kiro-idc-login --kiro-idc-start-url https://d-xxxxxxxxxx.awsapps.com/start --kiro-idc-region us-west-2
-```
-
-**Additional flags:**
-
-| Flag | Description |
-|------|-------------|
-| `--no-browser` | Don't open browser automatically, print URL instead |
-| `--no-incognito` | Use existing browser session (Kiro defaults to incognito). Useful for corporate SSO that requires an authenticated browser session |
-| `--kiro-idc-start-url` | IDC Start URL (required with `--kiro-idc-login`) |
-| `--kiro-idc-region` | IDC region (default: `us-east-1`) |
-| `--kiro-idc-flow` | IDC flow type: `authcode` (default) or `device` |
-
-### Web-based OAuth Login
-
-Access the Kiro OAuth web interface at:
-
-```
-http://your-server:8080/v0/oauth/kiro
-```
-
-This provides a browser-based OAuth flow for Kiro (AWS CodeWhisperer) authentication with:
-- AWS Builder ID login
-- AWS Identity Center (IDC) login
-- Token import from Kiro IDE
-
-## Quick Deployment with Docker
-
-### One-Command Deployment
-
-```bash
-# Create deployment directory
-mkdir -p ~/cli-proxy && cd ~/cli-proxy
-
-# Create docker-compose.yml
-cat > docker-compose.yml << 'EOF'
-services:
-  cli-proxy-api:
-    image: eceasy/cli-proxy-api-plus:latest
-    container_name: cli-proxy-api-plus
-    ports:
-      - "8317:8317"
-    volumes:
-      - ./config.yaml:/CLIProxyAPI/config.yaml
-      - ./auths:/root/.cli-proxy-api
-      - ./logs:/CLIProxyAPI/logs
-    restart: unless-stopped
-EOF
-
-# Download example config
-curl -o config.yaml https://raw.githubusercontent.com/router-for-me/CLIProxyAPIPlus/main/config.example.yaml
-
-# Pull and start
-docker compose pull && docker compose up -d
-```
-
-### Configuration
-
-Edit `config.yaml` before starting:
-
-```yaml
-# Basic configuration example
-server:
-  port: 8317
-
-# Add your provider configurations here
-```
-
-### Update to Latest Version
-
-```bash
-cd ~/cli-proxy
-docker compose pull && docker compose up -d
-```
-
 ## Contributing
 
 This project only accepts pull requests that relate to third-party provider support. Any pull requests unrelated to third-party provider support will be rejected.

README_CN.md

@@ -8,123 +8,6 @@
 
 该 Plus 版本的主线功能与主线功能强制同步。
 
-## 与主线版本版本差异
-
-[![bigmodel.cn](https://assets.router-for.me/chinese-5.png)](https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII)
-
-## 新增功能 (Plus 增强版)
-
-GLM CODING PLAN 是专为AI编码打造的订阅套餐，每月最低仅需20元，即可在十余款主流AI编码工具如 Claude Code、Cline、Roo Code 中畅享智谱旗舰模型GLM-4.7（受限于算力，目前仅限Pro用户开放），为开发者提供顶尖的编码体验。
-
-智谱AI为本产品提供了特别优惠，使用以下链接购买可以享受九折优惠：https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII
-
-### 命令行登录
-
-> **注意：** 由于 AWS Cognito 限制，Google/GitHub 登录不可用于第三方应用。
-
-**AWS Builder ID**（推荐）：
-
-```bash
-# 设备码流程
-./CLIProxyAPI --kiro-aws-login
-
-# 授权码流程
-./CLIProxyAPI --kiro-aws-authcode
-```
-
-**从 Kiro IDE 导入令牌：**
-
-```bash
-./CLIProxyAPI --kiro-import
-```
-
-获取令牌步骤：
-
-1. 打开 Kiro IDE，使用 Google（或 GitHub）登录
-2. 找到令牌文件：`~/.kiro/kiro-auth-token.json`
-3. 运行：`./CLIProxyAPI --kiro-import`
-
-**AWS IAM Identity Center (IDC)：**
-
-```bash
-./CLIProxyAPI --kiro-idc-login --kiro-idc-start-url https://d-xxxxxxxxxx.awsapps.com/start
-
-# 指定区域
-./CLIProxyAPI --kiro-idc-login --kiro-idc-start-url https://d-xxxxxxxxxx.awsapps.com/start --kiro-idc-region us-west-2
-```
-
-**附加参数：**
-
-| 参数 | 说明 |
-|------|------|
-| `--no-browser` | 不自动打开浏览器，打印 URL |
-| `--no-incognito` | 使用已有浏览器会话（Kiro 默认使用无痕模式），适用于需要已登录浏览器会话的企业 SSO 场景 |
-| `--kiro-idc-start-url` | IDC Start URL（`--kiro-idc-login` 必需） |
-| `--kiro-idc-region` | IDC 区域（默认：`us-east-1`） |
-| `--kiro-idc-flow` | IDC 流程类型：`authcode`（默认）或 `device` |
-
-### 网页端 OAuth 登录
-
-访问 Kiro OAuth 网页认证界面：
-
-```
-http://your-server:8080/v0/oauth/kiro
-```
-
-提供基于浏览器的 Kiro (AWS CodeWhisperer) OAuth 认证流程，支持：
-- AWS Builder ID 登录
-- AWS Identity Center (IDC) 登录
-- 从 Kiro IDE 导入令牌
-
-## Docker 快速部署
-
-### 一键部署
-
-```bash
-# 创建部署目录
-mkdir -p ~/cli-proxy && cd ~/cli-proxy
-
-# 创建 docker-compose.yml
-cat > docker-compose.yml << 'EOF'
-services:
-  cli-proxy-api:
-    image: eceasy/cli-proxy-api-plus:latest
-    container_name: cli-proxy-api-plus
-    ports:
-      - "8317:8317"
-    volumes:
-      - ./config.yaml:/CLIProxyAPI/config.yaml
-      - ./auths:/root/.cli-proxy-api
-      - ./logs:/CLIProxyAPI/logs
-    restart: unless-stopped
-EOF
-
-# 下载示例配置
-curl -o config.yaml https://raw.githubusercontent.com/router-for-me/CLIProxyAPIPlus/main/config.example.yaml
-
-# 拉取并启动
-docker compose pull && docker compose up -d
-```
-
-### 配置说明
-
-启动前请编辑 `config.yaml`：
-
-```yaml
-# 基本配置示例
-server:
-  port: 8317
-
-# 在此添加你的供应商配置
-```
-
-### 更新到最新版本
-
-```bash
-cd ~/cli-proxy
-docker compose pull && docker compose up -d
-```
-
 ## 贡献
 
 该项目仅接受第三方供应商支持的 Pull Request。任何非第三方供应商支持的 Pull Request 都将被拒绝。

cmd/server/main.go

@@ -25,6 +25,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/logging"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/managementasset"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/store"
 	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/translator"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/tui"
@@ -573,6 +574,7 @@ func main() {
 			if standalone {
 				// Standalone mode: start an embedded local server and connect TUI client to it.
 				managementasset.StartAutoUpdater(context.Background(), configFilePath)
+				registry.StartModelsUpdater(context.Background())
 				hook := tui.NewLogHook(2000)
 				hook.SetFormatter(&logging.LogFormatter{})
 				log.AddHook(hook)
@@ -643,15 +645,16 @@ func main() {
 				}
 			}
 		} else {
-      // Start the main proxy service
-      managementasset.StartAutoUpdater(context.Background(), configFilePath)
+			// Start the main proxy service
+			managementasset.StartAutoUpdater(context.Background(), configFilePath)
+			registry.StartModelsUpdater(context.Background())
 
-      if cfg.AuthDir != "" {
-        kiro.InitializeAndStart(cfg.AuthDir, cfg)
-        defer kiro.StopGlobalRefreshManager()
-      }
+			if cfg.AuthDir != "" {
+				kiro.InitializeAndStart(cfg.AuthDir, cfg)
+				defer kiro.StopGlobalRefreshManager()
+			}
 
-      cmd.StartService(cfg, configFilePath, password)
+			cmd.StartService(cfg, configFilePath, password)
 		}
 	}
 }

config.example.yaml

@@ -219,6 +219,17 @@ nonstream-keepalive-interval: 0
 #     models: # The models supported by the provider.
 #       - name: "moonshotai/kimi-k2:free" # The actual model name.
 #         alias: "kimi-k2" # The alias used in the API.
+#       # You may repeat the same alias to build an internal model pool.
+#       # The client still sees only one alias in the model list.
+#       # Requests to that alias will round-robin across the upstream names below,
+#       # and if the chosen upstream fails before producing output, the request will
+#       # continue with the next upstream model in the same alias pool.
+#       - name: "qwen3.5-plus"
+#         alias: "claude-opus-4.66"
+#       - name: "glm-5"
+#         alias: "claude-opus-4.66"
+#       - name: "kimi-k2.5"
+#         alias: "claude-opus-4.66"
 
 # Vertex API keys (Vertex-compatible endpoints, use API key + base URL)
 # vertex-api-key:
@@ -233,6 +244,9 @@ nonstream-keepalive-interval: 0
 #         alias: "vertex-flash"                   # client-visible alias
 #       - name: "gemini-2.5-pro"
 #         alias: "vertex-pro"
+#     excluded-models:                            # optional: models to exclude from listing
+#       - "imagen-3.0-generate-002"
+#       - "imagen-*"
 
 # Amp Integration
 # ampcode:

internal/api/handlers/management/auth_files.go

@@ -16,6 +16,7 @@ import (
 	"net/url"
 	"os"
 	"path/filepath"
+	"runtime"
 	"sort"
 	"strconv"
 	"strings"
@@ -698,17 +699,20 @@ func (h *Handler) authIDForPath(path string) string {
 	if path == "" {
 		return ""
 	}
-	if h == nil || h.cfg == nil {
-		return path
+	id := path
+	if h != nil && h.cfg != nil {
+		authDir := strings.TrimSpace(h.cfg.AuthDir)
+		if authDir != "" {
+			if rel, errRel := filepath.Rel(authDir, path); errRel == nil && rel != "" {
+				id = rel
+			}
+		}
 	}
-	authDir := strings.TrimSpace(h.cfg.AuthDir)
-	if authDir == "" {
-		return path
+	// On Windows, normalize ID casing to avoid duplicate auth entries caused by case-insensitive paths.
+	if runtime.GOOS == "windows" {
+		id = strings.ToLower(id)
 	}
-	if rel, err := filepath.Rel(authDir, path); err == nil && rel != "" {
-		return rel
-	}
-	return path
+	return id
 }
 
 func (h *Handler) registerAuthFromFile(ctx context.Context, path string, data []byte) error {
@@ -1308,12 +1312,12 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			projects, errAll := onboardAllGeminiProjects(ctx, gemClient, &ts)
 			if errAll != nil {
 				log.Errorf("Failed to complete Gemini CLI onboarding: %v", errAll)
-				SetOAuthSessionError(state, "Failed to complete Gemini CLI onboarding")
+				SetOAuthSessionError(state, fmt.Sprintf("Failed to complete Gemini CLI onboarding: %v", errAll))
 				return
 			}
 			if errVerify := ensureGeminiProjectsEnabled(ctx, gemClient, projects); errVerify != nil {
 				log.Errorf("Failed to verify Cloud AI API status: %v", errVerify)
-				SetOAuthSessionError(state, "Failed to verify Cloud AI API status")
+				SetOAuthSessionError(state, fmt.Sprintf("Failed to verify Cloud AI API status: %v", errVerify))
 				return
 			}
 			ts.ProjectID = strings.Join(projects, ",")
@@ -1322,7 +1326,7 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			ts.Auto = false
 			if errSetup := performGeminiCLISetup(ctx, gemClient, &ts, ""); errSetup != nil {
 				log.Errorf("Google One auto-discovery failed: %v", errSetup)
-				SetOAuthSessionError(state, "Google One auto-discovery failed")
+				SetOAuthSessionError(state, fmt.Sprintf("Google One auto-discovery failed: %v", errSetup))
 				return
 			}
 			if strings.TrimSpace(ts.ProjectID) == "" {
@@ -1333,19 +1337,19 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			isChecked, errCheck := checkCloudAPIIsEnabled(ctx, gemClient, ts.ProjectID)
 			if errCheck != nil {
 				log.Errorf("Failed to verify Cloud AI API status: %v", errCheck)
-				SetOAuthSessionError(state, "Failed to verify Cloud AI API status")
+				SetOAuthSessionError(state, fmt.Sprintf("Failed to verify Cloud AI API status: %v", errCheck))
 				return
 			}
 			ts.Checked = isChecked
 			if !isChecked {
 				log.Error("Cloud AI API is not enabled for the auto-discovered project")
-				SetOAuthSessionError(state, "Cloud AI API not enabled")
+				SetOAuthSessionError(state, fmt.Sprintf("Cloud AI API not enabled for project %s", ts.ProjectID))
 				return
 			}
 		} else {
 			if errEnsure := ensureGeminiProjectAndOnboard(ctx, gemClient, &ts, requestedProjectID); errEnsure != nil {
 				log.Errorf("Failed to complete Gemini CLI onboarding: %v", errEnsure)
-				SetOAuthSessionError(state, "Failed to complete Gemini CLI onboarding")
+				SetOAuthSessionError(state, fmt.Sprintf("Failed to complete Gemini CLI onboarding: %v", errEnsure))
 				return
 			}
 
@@ -1358,13 +1362,13 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			isChecked, errCheck := checkCloudAPIIsEnabled(ctx, gemClient, ts.ProjectID)
 			if errCheck != nil {
 				log.Errorf("Failed to verify Cloud AI API status: %v", errCheck)
-				SetOAuthSessionError(state, "Failed to verify Cloud AI API status")
+				SetOAuthSessionError(state, fmt.Sprintf("Failed to verify Cloud AI API status: %v", errCheck))
 				return
 			}
 			ts.Checked = isChecked
 			if !isChecked {
 				log.Error("Cloud AI API is not enabled for the selected project")
-				SetOAuthSessionError(state, "Cloud AI API not enabled")
+				SetOAuthSessionError(state, fmt.Sprintf("Cloud AI API not enabled for project %s", ts.ProjectID))
 				return
 			}
 		}

internal/api/handlers/management/config_lists.go

@@ -516,12 +516,13 @@ func (h *Handler) PutVertexCompatKeys(c *gin.Context) {
 }
 func (h *Handler) PatchVertexCompatKey(c *gin.Context) {
 	type vertexCompatPatch struct {
-		APIKey   *string                     `json:"api-key"`
-		Prefix   *string                     `json:"prefix"`
-		BaseURL  *string                     `json:"base-url"`
-		ProxyURL *string                     `json:"proxy-url"`
-		Headers  *map[string]string          `json:"headers"`
-		Models   *[]config.VertexCompatModel `json:"models"`
+		APIKey         *string                     `json:"api-key"`
+		Prefix         *string                     `json:"prefix"`
+		BaseURL        *string                     `json:"base-url"`
+		ProxyURL       *string                     `json:"proxy-url"`
+		Headers        *map[string]string          `json:"headers"`
+		Models         *[]config.VertexCompatModel `json:"models"`
+		ExcludedModels *[]string                   `json:"excluded-models"`
 	}
 	var body struct {
 		Index *int               `json:"index"`
@@ -585,6 +586,9 @@ func (h *Handler) PatchVertexCompatKey(c *gin.Context) {
 	if body.Value.Models != nil {
 		entry.Models = append([]config.VertexCompatModel(nil), (*body.Value.Models)...)
 	}
+	if body.Value.ExcludedModels != nil {
+		entry.ExcludedModels = config.NormalizeExcludedModels(*body.Value.ExcludedModels)
+	}
 	normalizeVertexCompatKey(&entry)
 	h.cfg.VertexCompatAPIKey[targetIndex] = entry
 	h.cfg.SanitizeVertexCompatKeys()
@@ -1029,6 +1033,7 @@ func normalizeVertexCompatKey(entry *config.VertexCompatKey) {
 	entry.BaseURL = strings.TrimSpace(entry.BaseURL)
 	entry.ProxyURL = strings.TrimSpace(entry.ProxyURL)
 	entry.Headers = config.NormalizeHeaders(entry.Headers)
+	entry.ExcludedModels = config.NormalizeExcludedModels(entry.ExcludedModels)
 	if len(entry.Models) == 0 {
 		return
 	}

internal/auth/copilot/copilot_auth.go

@@ -8,6 +8,8 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"net/url"
+	"strings"
 	"time"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
@@ -222,6 +224,97 @@ func (c *CopilotAuth) MakeAuthenticatedRequest(ctx context.Context, method, url
 	return req, nil
 }
 
+// CopilotModelEntry represents a single model entry returned by the Copilot /models API.
+type CopilotModelEntry struct {
+	ID           string         `json:"id"`
+	Object       string         `json:"object"`
+	Created      int64          `json:"created"`
+	OwnedBy      string         `json:"owned_by"`
+	Name         string         `json:"name,omitempty"`
+	Version      string         `json:"version,omitempty"`
+	Capabilities map[string]any `json:"capabilities,omitempty"`
+}
+
+// CopilotModelsResponse represents the response from the Copilot /models endpoint.
+type CopilotModelsResponse struct {
+	Data   []CopilotModelEntry `json:"data"`
+	Object string              `json:"object"`
+}
+
+// maxModelsResponseSize is the maximum allowed response size from the /models endpoint (2 MB).
+const maxModelsResponseSize = 2 * 1024 * 1024
+
+// allowedCopilotAPIHosts is the set of hosts that are considered safe for Copilot API requests.
+var allowedCopilotAPIHosts = map[string]bool{
+	"api.githubcopilot.com":          true,
+	"api.individual.githubcopilot.com": true,
+	"api.business.githubcopilot.com":   true,
+	"copilot-proxy.githubusercontent.com": true,
+}
+
+// ListModels fetches the list of available models from the Copilot API.
+// It requires a valid Copilot API token (not the GitHub access token).
+func (c *CopilotAuth) ListModels(ctx context.Context, apiToken *CopilotAPIToken) ([]CopilotModelEntry, error) {
+	if apiToken == nil || apiToken.Token == "" {
+		return nil, fmt.Errorf("copilot: api token is required for listing models")
+	}
+
+	// Build models URL, validating the endpoint host to prevent SSRF.
+	modelsURL := copilotAPIEndpoint + "/models"
+	if ep := strings.TrimRight(apiToken.Endpoints.API, "/"); ep != "" {
+		parsed, err := url.Parse(ep)
+		if err == nil && parsed.Scheme == "https" && allowedCopilotAPIHosts[parsed.Host] {
+			modelsURL = ep + "/models"
+		} else {
+			log.Warnf("copilot: ignoring untrusted API endpoint %q, using default", ep)
+		}
+	}
+
+	req, err := c.MakeAuthenticatedRequest(ctx, http.MethodGet, modelsURL, nil, apiToken)
+	if err != nil {
+		return nil, fmt.Errorf("copilot: failed to create models request: %w", err)
+	}
+
+	resp, err := c.httpClient.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("copilot: models request failed: %w", err)
+	}
+	defer func() {
+		if errClose := resp.Body.Close(); errClose != nil {
+			log.Errorf("copilot list models: close body error: %v", errClose)
+		}
+	}()
+
+	// Limit response body to prevent memory exhaustion.
+	limitedReader := io.LimitReader(resp.Body, maxModelsResponseSize)
+	bodyBytes, err := io.ReadAll(limitedReader)
+	if err != nil {
+		return nil, fmt.Errorf("copilot: failed to read models response: %w", err)
+	}
+
+	if !isHTTPSuccess(resp.StatusCode) {
+		return nil, fmt.Errorf("copilot: list models failed with status %d: %s", resp.StatusCode, string(bodyBytes))
+	}
+
+	var modelsResp CopilotModelsResponse
+	if err = json.Unmarshal(bodyBytes, &modelsResp); err != nil {
+		return nil, fmt.Errorf("copilot: failed to parse models response: %w", err)
+	}
+
+	return modelsResp.Data, nil
+}
+
+// ListModelsWithGitHubToken is a convenience method that exchanges a GitHub access token
+// for a Copilot API token and then fetches the available models.
+func (c *CopilotAuth) ListModelsWithGitHubToken(ctx context.Context, githubAccessToken string) ([]CopilotModelEntry, error) {
+	apiToken, err := c.GetCopilotAPIToken(ctx, githubAccessToken)
+	if err != nil {
+		return nil, fmt.Errorf("copilot: failed to get API token for model listing: %w", err)
+	}
+
+	return c.ListModels(ctx, apiToken)
+}
+
 // buildChatCompletionURL builds the URL for chat completions API.
 func buildChatCompletionURL() string {
 	return copilotAPIEndpoint + "/chat/completions"

internal/config/oauth_model_alias_defaults.go

@@ -1,5 +1,7 @@
 package config
 
+import "strings"
+
 // defaultKiroAliases returns default oauth-model-alias entries for Kiro.
 // These aliases expose standard Claude IDs for Kiro-prefixed upstream models.
 func defaultKiroAliases() []OAuthModelAlias {
@@ -35,3 +37,25 @@ func defaultGitHubCopilotAliases() []OAuthModelAlias {
 		{Name: "claude-sonnet-4.6", Alias: "claude-sonnet-4-6", Fork: true},
 	}
 }
+
+// GitHubCopilotAliasesFromModels generates oauth-model-alias entries from a dynamic
+// list of model IDs fetched from the Copilot API. It auto-creates aliases for
+// models whose ID contains a dot (e.g. "claude-opus-4.6" → "claude-opus-4-6"),
+// which is the pattern used by Claude models on Copilot.
+func GitHubCopilotAliasesFromModels(modelIDs []string) []OAuthModelAlias {
+	var aliases []OAuthModelAlias
+	seen := make(map[string]struct{})
+	for _, id := range modelIDs {
+		if !strings.Contains(id, ".") {
+			continue
+		}
+		hyphenID := strings.ReplaceAll(id, ".", "-")
+		key := id + "→" + hyphenID
+		if _, ok := seen[key]; ok {
+			continue
+		}
+		seen[key] = struct{}{}
+		aliases = append(aliases, OAuthModelAlias{Name: id, Alias: hyphenID, Fork: true})
+	}
+	return aliases
+}

internal/config/vertex_compat.go

@@ -34,6 +34,9 @@ type VertexCompatKey struct {
 
 	// Models defines the model configurations including aliases for routing.
 	Models []VertexCompatModel `yaml:"models,omitempty" json:"models,omitempty"`
+
+	// ExcludedModels lists model IDs that should be excluded for this provider.
+	ExcludedModels []string `yaml:"excluded-models,omitempty" json:"excluded-models,omitempty"`
 }
 
 func (k VertexCompatKey) GetAPIKey() string  { return k.APIKey }
@@ -74,6 +77,7 @@ func (cfg *Config) SanitizeVertexCompatKeys() {
 		}
 		entry.ProxyURL = strings.TrimSpace(entry.ProxyURL)
 		entry.Headers = NormalizeHeaders(entry.Headers)
+		entry.ExcludedModels = NormalizeExcludedModels(entry.ExcludedModels)
 
 		// Sanitize models: remove entries without valid alias
 		sanitizedModels := make([]VertexCompatModel, 0, len(entry.Models))

internal/registry/model_definitions.go

@@ -1,5 +1,5 @@
 // Package registry provides model definitions and lookup helpers for various AI providers.
-// Static model metadata is stored in model_definitions_static_data.go.
+// Static model metadata is loaded from the embedded models.json file and can be refreshed from network.
 package registry
 
 import (
@@ -7,6 +7,131 @@ import (
 	"strings"
 )
 
+// AntigravityModelConfig captures static antigravity model overrides, including
+// Thinking budget limits and provider max completion tokens.
+type AntigravityModelConfig struct {
+	Thinking            *ThinkingSupport `json:"thinking,omitempty"`
+	MaxCompletionTokens int              `json:"max_completion_tokens,omitempty"`
+}
+
+// staticModelsJSON mirrors the top-level structure of models.json.
+type staticModelsJSON struct {
+	Claude      []*ModelInfo                       `json:"claude"`
+	Gemini      []*ModelInfo                       `json:"gemini"`
+	Vertex      []*ModelInfo                       `json:"vertex"`
+	GeminiCLI   []*ModelInfo                       `json:"gemini-cli"`
+	AIStudio    []*ModelInfo                       `json:"aistudio"`
+	CodexFree   []*ModelInfo                       `json:"codex-free"`
+	CodexTeam   []*ModelInfo                       `json:"codex-team"`
+	CodexPlus   []*ModelInfo                       `json:"codex-plus"`
+	CodexPro    []*ModelInfo                       `json:"codex-pro"`
+	Qwen        []*ModelInfo                       `json:"qwen"`
+	IFlow       []*ModelInfo                       `json:"iflow"`
+	Kimi        []*ModelInfo                       `json:"kimi"`
+	Antigravity map[string]*AntigravityModelConfig `json:"antigravity"`
+}
+
+// GetClaudeModels returns the standard Claude model definitions.
+func GetClaudeModels() []*ModelInfo {
+	return cloneModelInfos(getModels().Claude)
+}
+
+// GetGeminiModels returns the standard Gemini model definitions.
+func GetGeminiModels() []*ModelInfo {
+	return cloneModelInfos(getModels().Gemini)
+}
+
+// GetGeminiVertexModels returns Gemini model definitions for Vertex AI.
+func GetGeminiVertexModels() []*ModelInfo {
+	return cloneModelInfos(getModels().Vertex)
+}
+
+// GetGeminiCLIModels returns Gemini model definitions for the Gemini CLI.
+func GetGeminiCLIModels() []*ModelInfo {
+	return cloneModelInfos(getModels().GeminiCLI)
+}
+
+// GetAIStudioModels returns model definitions for AI Studio.
+func GetAIStudioModels() []*ModelInfo {
+	return cloneModelInfos(getModels().AIStudio)
+}
+
+// GetCodexFreeModels returns model definitions for the Codex free plan tier.
+func GetCodexFreeModels() []*ModelInfo {
+	return cloneModelInfos(getModels().CodexFree)
+}
+
+// GetCodexTeamModels returns model definitions for the Codex team plan tier.
+func GetCodexTeamModels() []*ModelInfo {
+	return cloneModelInfos(getModels().CodexTeam)
+}
+
+// GetCodexPlusModels returns model definitions for the Codex plus plan tier.
+func GetCodexPlusModels() []*ModelInfo {
+	return cloneModelInfos(getModels().CodexPlus)
+}
+
+// GetCodexProModels returns model definitions for the Codex pro plan tier.
+func GetCodexProModels() []*ModelInfo {
+	return cloneModelInfos(getModels().CodexPro)
+}
+
+// GetQwenModels returns the standard Qwen model definitions.
+func GetQwenModels() []*ModelInfo {
+	return cloneModelInfos(getModels().Qwen)
+}
+
+// GetIFlowModels returns the standard iFlow model definitions.
+func GetIFlowModels() []*ModelInfo {
+	return cloneModelInfos(getModels().IFlow)
+}
+
+// GetKimiModels returns the standard Kimi (Moonshot AI) model definitions.
+func GetKimiModels() []*ModelInfo {
+	return cloneModelInfos(getModels().Kimi)
+}
+
+// GetAntigravityModelConfig returns static configuration for antigravity models.
+// Keys use upstream model names returned by the Antigravity models endpoint.
+func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
+	data := getModels()
+	if len(data.Antigravity) == 0 {
+		return nil
+	}
+	out := make(map[string]*AntigravityModelConfig, len(data.Antigravity))
+	for k, v := range data.Antigravity {
+		out[k] = cloneAntigravityModelConfig(v)
+	}
+	return out
+}
+
+func cloneAntigravityModelConfig(cfg *AntigravityModelConfig) *AntigravityModelConfig {
+	if cfg == nil {
+		return nil
+	}
+	copyConfig := *cfg
+	if cfg.Thinking != nil {
+		copyThinking := *cfg.Thinking
+		if len(cfg.Thinking.Levels) > 0 {
+			copyThinking.Levels = append([]string(nil), cfg.Thinking.Levels...)
+		}
+		copyConfig.Thinking = &copyThinking
+	}
+	return &copyConfig
+}
+
+// cloneModelInfos returns a shallow copy of the slice with each element deep-cloned.
+func cloneModelInfos(models []*ModelInfo) []*ModelInfo {
+	if len(models) == 0 {
+		return nil
+	}
+	out := make([]*ModelInfo, len(models))
+	for i, m := range models {
+		out[i] = cloneModelInfo(m)
+	}
+	return out
+}
+
 // GetStaticModelDefinitionsByChannel returns static model definitions for a given channel/provider.
 // It returns nil when the channel is unknown.
 //
@@ -23,7 +148,6 @@ import (
 //   - kiro
 //   - kilo
 //   - github-copilot
-//   - kiro
 //   - amazonq
 //   - antigravity (returns static overrides only)
 func GetStaticModelDefinitionsByChannel(channel string) []*ModelInfo {
@@ -40,7 +164,7 @@ func GetStaticModelDefinitionsByChannel(channel string) []*ModelInfo {
 	case "aistudio":
 		return GetAIStudioModels()
 	case "codex":
-		return GetOpenAIModels()
+		return GetCodexProModels()
 	case "qwen":
 		return GetQwenModels()
 	case "iflow":
@@ -90,16 +214,17 @@ func LookupStaticModelInfo(modelID string) *ModelInfo {
 		return nil
 	}
 
+	data := getModels()
 	allModels := [][]*ModelInfo{
-		GetClaudeModels(),
-		GetGeminiModels(),
-		GetGeminiVertexModels(),
-		GetGeminiCLIModels(),
-		GetAIStudioModels(),
-		GetOpenAIModels(),
-		GetQwenModels(),
-		GetIFlowModels(),
-		GetKimiModels(),
+		data.Claude,
+		data.Gemini,
+		data.Vertex,
+		data.GeminiCLI,
+		data.AIStudio,
+		data.CodexPro,
+		data.Qwen,
+		data.IFlow,
+		data.Kimi,
 		GetGitHubCopilotModels(),
 		GetKiroModels(),
 		GetKiloModels(),
@@ -108,13 +233,13 @@ func LookupStaticModelInfo(modelID string) *ModelInfo {
 	for _, models := range allModels {
 		for _, m := range models {
 			if m != nil && m.ID == modelID {
-				return m
+				return cloneModelInfo(m)
 			}
 		}
 	}
 
 	// Check Antigravity static config
-	if cfg := GetAntigravityModelConfig()[modelID]; cfg != nil {
+	if cfg := cloneAntigravityModelConfig(data.Antigravity[modelID]); cfg != nil {
 		return &ModelInfo{
 			ID:                  modelID,
 			Thinking:            cfg.Thinking,
@@ -152,6 +277,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			Description:         "OpenAI GPT-4.1 via GitHub Copilot",
 			ContextLength:       128000,
 			MaxCompletionTokens: 16384,
+			SupportedEndpoints:  []string{"/chat/completions", "/responses"},
 		},
 	}
 
@@ -166,6 +292,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			Description:         entry.Description,
 			ContextLength:       128000,
 			MaxCompletionTokens: 16384,
+			SupportedEndpoints:  []string{"/chat/completions", "/responses"},
 		})
 	}
 

internal/registry/model_registry.go

@@ -64,6 +64,11 @@ type ModelInfo struct {
 	UserDefined bool `json:"-"`
 }
 
+type availableModelsCacheEntry struct {
+	models    []map[string]any
+	expiresAt time.Time
+}
+
 // ThinkingSupport describes a model family's supported internal reasoning budget range.
 // Values are interpreted in provider-native token units.
 type ThinkingSupport struct {
@@ -118,6 +123,8 @@ type ModelRegistry struct {
 	clientProviders map[string]string
 	// mutex ensures thread-safe access to the registry
 	mutex *sync.RWMutex
+	// availableModelsCache stores per-handler snapshots for GetAvailableModels.
+	availableModelsCache map[string]availableModelsCacheEntry
 	// hook is an optional callback sink for model registration changes
 	hook ModelRegistryHook
 }
@@ -130,15 +137,28 @@ var registryOnce sync.Once
 func GetGlobalRegistry() *ModelRegistry {
 	registryOnce.Do(func() {
 		globalRegistry = &ModelRegistry{
-			models:           make(map[string]*ModelRegistration),
-			clientModels:     make(map[string][]string),
-			clientModelInfos: make(map[string]map[string]*ModelInfo),
-			clientProviders:  make(map[string]string),
-			mutex:            &sync.RWMutex{},
+			models:               make(map[string]*ModelRegistration),
+			clientModels:         make(map[string][]string),
+			clientModelInfos:     make(map[string]map[string]*ModelInfo),
+			clientProviders:      make(map[string]string),
+			availableModelsCache: make(map[string]availableModelsCacheEntry),
+			mutex:                &sync.RWMutex{},
 		}
 	})
 	return globalRegistry
 }
+func (r *ModelRegistry) ensureAvailableModelsCacheLocked() {
+	if r.availableModelsCache == nil {
+		r.availableModelsCache = make(map[string]availableModelsCacheEntry)
+	}
+}
+
+func (r *ModelRegistry) invalidateAvailableModelsCacheLocked() {
+	if len(r.availableModelsCache) == 0 {
+		return
+	}
+	clear(r.availableModelsCache)
+}
 
 // LookupModelInfo searches dynamic registry (provider-specific > global) then static definitions.
 func LookupModelInfo(modelID string, provider ...string) *ModelInfo {
@@ -153,9 +173,9 @@ func LookupModelInfo(modelID string, provider ...string) *ModelInfo {
 	}
 
 	if info := GetGlobalRegistry().GetModelInfo(modelID, p); info != nil {
-		return info
+		return cloneModelInfo(info)
 	}
-	return LookupStaticModelInfo(modelID)
+	return cloneModelInfo(LookupStaticModelInfo(modelID))
 }
 
 // SetHook sets an optional hook for observing model registration changes.
@@ -213,6 +233,7 @@ func (r *ModelRegistry) triggerModelsUnregistered(provider, clientID string) {
 func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models []*ModelInfo) {
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
 
 	provider := strings.ToLower(clientProvider)
 	uniqueModelIDs := make([]string, 0, len(models))
@@ -238,6 +259,7 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 		delete(r.clientModels, clientID)
 		delete(r.clientModelInfos, clientID)
 		delete(r.clientProviders, clientID)
+		r.invalidateAvailableModelsCacheLocked()
 		misc.LogCredentialSeparator()
 		return
 	}
@@ -265,6 +287,7 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 		} else {
 			delete(r.clientProviders, clientID)
 		}
+		r.invalidateAvailableModelsCacheLocked()
 		r.triggerModelsRegistered(provider, clientID, models)
 		log.Debugf("Registered client %s from provider %s with %d models", clientID, clientProvider, len(rawModelIDs))
 		misc.LogCredentialSeparator()
@@ -408,6 +431,7 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 		delete(r.clientProviders, clientID)
 	}
 
+	r.invalidateAvailableModelsCacheLocked()
 	r.triggerModelsRegistered(provider, clientID, models)
 	if len(added) == 0 && len(removed) == 0 && !providerChanged {
 		// Only metadata (e.g., display name) changed; skip separator when no log output.
@@ -511,6 +535,13 @@ func cloneModelInfo(model *ModelInfo) *ModelInfo {
 	if len(model.SupportedOutputModalities) > 0 {
 		copyModel.SupportedOutputModalities = append([]string(nil), model.SupportedOutputModalities...)
 	}
+	if model.Thinking != nil {
+		copyThinking := *model.Thinking
+		if len(model.Thinking.Levels) > 0 {
+			copyThinking.Levels = append([]string(nil), model.Thinking.Levels...)
+		}
+		copyModel.Thinking = &copyThinking
+	}
 	return &copyModel
 }
 
@@ -540,6 +571,7 @@ func (r *ModelRegistry) UnregisterClient(clientID string) {
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
 	r.unregisterClientInternal(clientID)
+	r.invalidateAvailableModelsCacheLocked()
 }
 
 // unregisterClientInternal performs the actual client unregistration (internal, no locking)
@@ -606,9 +638,12 @@ func (r *ModelRegistry) unregisterClientInternal(clientID string) {
 func (r *ModelRegistry) SetModelQuotaExceeded(clientID, modelID string) {
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
 
 	if registration, exists := r.models[modelID]; exists {
-		registration.QuotaExceededClients[clientID] = new(time.Now())
+		now := time.Now()
+		registration.QuotaExceededClients[clientID] = &now
+		r.invalidateAvailableModelsCacheLocked()
 		log.Debugf("Marked model %s as quota exceeded for client %s", modelID, clientID)
 	}
 }
@@ -620,9 +655,11 @@ func (r *ModelRegistry) SetModelQuotaExceeded(clientID, modelID string) {
 func (r *ModelRegistry) ClearModelQuotaExceeded(clientID, modelID string) {
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
 
 	if registration, exists := r.models[modelID]; exists {
 		delete(registration.QuotaExceededClients, clientID)
+		r.invalidateAvailableModelsCacheLocked()
 		// log.Debugf("Cleared quota exceeded status for model %s and client %s", modelID, clientID)
 	}
 }
@@ -638,6 +675,7 @@ func (r *ModelRegistry) SuspendClientModel(clientID, modelID, reason string) {
 	}
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
 
 	registration, exists := r.models[modelID]
 	if !exists || registration == nil {
@@ -651,6 +689,7 @@ func (r *ModelRegistry) SuspendClientModel(clientID, modelID, reason string) {
 	}
 	registration.SuspendedClients[clientID] = reason
 	registration.LastUpdated = time.Now()
+	r.invalidateAvailableModelsCacheLocked()
 	if reason != "" {
 		log.Debugf("Suspended client %s for model %s: %s", clientID, modelID, reason)
 	} else {
@@ -668,6 +707,7 @@ func (r *ModelRegistry) ResumeClientModel(clientID, modelID string) {
 	}
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
 
 	registration, exists := r.models[modelID]
 	if !exists || registration == nil || registration.SuspendedClients == nil {
@@ -678,6 +718,7 @@ func (r *ModelRegistry) ResumeClientModel(clientID, modelID string) {
 	}
 	delete(registration.SuspendedClients, clientID)
 	registration.LastUpdated = time.Now()
+	r.invalidateAvailableModelsCacheLocked()
 	log.Debugf("Resumed client %s for model %s", clientID, modelID)
 }
 
@@ -713,22 +754,52 @@ func (r *ModelRegistry) ClientSupportsModel(clientID, modelID string) bool {
 // Returns:
 //   - []map[string]any: List of available models in the requested format
 func (r *ModelRegistry) GetAvailableModels(handlerType string) []map[string]any {
-	r.mutex.RLock()
-	defer r.mutex.RUnlock()
+	now := time.Now()
 
-	models := make([]map[string]any, 0)
+	r.mutex.RLock()
+	if cache, ok := r.availableModelsCache[handlerType]; ok && (cache.expiresAt.IsZero() || now.Before(cache.expiresAt)) {
+		models := cloneModelMaps(cache.models)
+		r.mutex.RUnlock()
+		return models
+	}
+	r.mutex.RUnlock()
+
+	r.mutex.Lock()
+	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
+
+	if cache, ok := r.availableModelsCache[handlerType]; ok && (cache.expiresAt.IsZero() || now.Before(cache.expiresAt)) {
+		return cloneModelMaps(cache.models)
+	}
+
+	models, expiresAt := r.buildAvailableModelsLocked(handlerType, now)
+	r.availableModelsCache[handlerType] = availableModelsCacheEntry{
+		models:    cloneModelMaps(models),
+		expiresAt: expiresAt,
+	}
+
+	return models
+}
+
+func (r *ModelRegistry) buildAvailableModelsLocked(handlerType string, now time.Time) ([]map[string]any, time.Time) {
+	models := make([]map[string]any, 0, len(r.models))
 	quotaExpiredDuration := 5 * time.Minute
+	var expiresAt time.Time
 
 	for _, registration := range r.models {
-		// Check if model has any non-quota-exceeded clients
 		availableClients := registration.Count
-		now := time.Now()
 
-		// Count clients that have exceeded quota but haven't recovered yet
 		expiredClients := 0
 		for _, quotaTime := range registration.QuotaExceededClients {
-			if quotaTime != nil && now.Sub(*quotaTime) < quotaExpiredDuration {
+			if quotaTime == nil {
+				continue
+			}
+			recoveryAt := quotaTime.Add(quotaExpiredDuration)
+			if now.Before(recoveryAt) {
 				expiredClients++
+				if expiresAt.IsZero() || recoveryAt.Before(expiresAt) {
+					expiresAt = recoveryAt
+				}
 			}
 		}
 
@@ -749,7 +820,6 @@ func (r *ModelRegistry) GetAvailableModels(handlerType string) []map[string]any
 			effectiveClients = 0
 		}
 
-		// Include models that have available clients, or those solely cooling down.
 		if effectiveClients > 0 || (availableClients > 0 && (expiredClients > 0 || cooldownSuspended > 0) && otherSuspended == 0) {
 			model := r.convertModelToMap(registration.Info, handlerType)
 			if model != nil {
@@ -758,7 +828,44 @@ func (r *ModelRegistry) GetAvailableModels(handlerType string) []map[string]any
 		}
 	}
 
-	return models
+	return models, expiresAt
+}
+
+func cloneModelMaps(models []map[string]any) []map[string]any {
+	cloned := make([]map[string]any, 0, len(models))
+	for _, model := range models {
+		if model == nil {
+			cloned = append(cloned, nil)
+			continue
+		}
+		copyModel := make(map[string]any, len(model))
+		for key, value := range model {
+			copyModel[key] = cloneModelMapValue(value)
+		}
+		cloned = append(cloned, copyModel)
+	}
+	return cloned
+}
+
+func cloneModelMapValue(value any) any {
+	switch typed := value.(type) {
+	case map[string]any:
+		copyMap := make(map[string]any, len(typed))
+		for key, entry := range typed {
+			copyMap[key] = cloneModelMapValue(entry)
+		}
+		return copyMap
+	case []any:
+		copySlice := make([]any, len(typed))
+		for i, entry := range typed {
+			copySlice[i] = cloneModelMapValue(entry)
+		}
+		return copySlice
+	case []string:
+		return append([]string(nil), typed...)
+	default:
+		return value
+	}
 }
 
 // GetAvailableModelsByProvider returns models available for the given provider identifier.
@@ -874,11 +981,11 @@ func (r *ModelRegistry) GetAvailableModelsByProvider(provider string) []*ModelIn
 
 		if effectiveClients > 0 || (availableClients > 0 && (expiredClients > 0 || cooldownSuspended > 0) && otherSuspended == 0) {
 			if entry.info != nil {
-				result = append(result, entry.info)
+				result = append(result, cloneModelInfo(entry.info))
 				continue
 			}
 			if ok && registration != nil && registration.Info != nil {
-				result = append(result, registration.Info)
+				result = append(result, cloneModelInfo(registration.Info))
 			}
 		}
 	}
@@ -987,13 +1094,13 @@ func (r *ModelRegistry) GetModelInfo(modelID, provider string) *ModelInfo {
 			if reg.Providers != nil {
 				if count, ok := reg.Providers[provider]; ok && count > 0 {
 					if info, ok := reg.InfoByProvider[provider]; ok && info != nil {
-						return info
+						return cloneModelInfo(info)
 					}
 				}
 			}
 		}
 		// Fallback to global info (last registered)
-		return reg.Info
+		return cloneModelInfo(reg.Info)
 	}
 	return nil
 }
@@ -1033,7 +1140,7 @@ func (r *ModelRegistry) convertModelToMap(model *ModelInfo, handlerType string)
 			result["max_completion_tokens"] = model.MaxCompletionTokens
 		}
 		if len(model.SupportedParameters) > 0 {
-			result["supported_parameters"] = model.SupportedParameters
+			result["supported_parameters"] = append([]string(nil), model.SupportedParameters...)
 		}
 		if len(model.SupportedEndpoints) > 0 {
 			result["supported_endpoints"] = model.SupportedEndpoints
@@ -1094,13 +1201,13 @@ func (r *ModelRegistry) convertModelToMap(model *ModelInfo, handlerType string)
 			result["outputTokenLimit"] = model.OutputTokenLimit
 		}
 		if len(model.SupportedGenerationMethods) > 0 {
-			result["supportedGenerationMethods"] = model.SupportedGenerationMethods
+			result["supportedGenerationMethods"] = append([]string(nil), model.SupportedGenerationMethods...)
 		}
 		if len(model.SupportedInputModalities) > 0 {
-			result["supportedInputModalities"] = model.SupportedInputModalities
+			result["supportedInputModalities"] = append([]string(nil), model.SupportedInputModalities...)
 		}
 		if len(model.SupportedOutputModalities) > 0 {
-			result["supportedOutputModalities"] = model.SupportedOutputModalities
+			result["supportedOutputModalities"] = append([]string(nil), model.SupportedOutputModalities...)
 		}
 		return result
 
@@ -1130,15 +1237,20 @@ func (r *ModelRegistry) CleanupExpiredQuotas() {
 
 	now := time.Now()
 	quotaExpiredDuration := 5 * time.Minute
+	invalidated := false
 
 	for modelID, registration := range r.models {
 		for clientID, quotaTime := range registration.QuotaExceededClients {
 			if quotaTime != nil && now.Sub(*quotaTime) >= quotaExpiredDuration {
 				delete(registration.QuotaExceededClients, clientID)
+				invalidated = true
 				log.Debugf("Cleaned up expired quota tracking for model %s, client %s", modelID, clientID)
 			}
 		}
 	}
+	if invalidated {
+		r.invalidateAvailableModelsCacheLocked()
+	}
 }
 
 // GetFirstAvailableModel returns the first available model for the given handler type.
@@ -1152,8 +1264,6 @@ func (r *ModelRegistry) CleanupExpiredQuotas() {
 //   - string: The model ID of the first available model, or empty string if none available
 //   - error: An error if no models are available
 func (r *ModelRegistry) GetFirstAvailableModel(handlerType string) (string, error) {
-	r.mutex.RLock()
-	defer r.mutex.RUnlock()
 
 	// Get all available models for this handler type
 	models := r.GetAvailableModels(handlerType)
@@ -1213,13 +1323,13 @@ func (r *ModelRegistry) GetModelsForClient(clientID string) []*ModelInfo {
 		// Prefer client's own model info to preserve original type/owned_by
 		if clientInfos != nil {
 			if info, ok := clientInfos[modelID]; ok && info != nil {
-				result = append(result, info)
+				result = append(result, cloneModelInfo(info))
 				continue
 			}
 		}
 		// Fallback to global registry (for backwards compatibility)
 		if reg, ok := r.models[modelID]; ok && reg.Info != nil {
-			result = append(result, reg.Info)
+			result = append(result, cloneModelInfo(reg.Info))
 		}
 	}
 	return result

internal/registry/model_registry_cache_test.go

@@ -0,0 +1,54 @@
+package registry
+
+import "testing"
+
+func TestGetAvailableModelsReturnsClonedSnapshots(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "OpenAI", []*ModelInfo{{ID: "m1", OwnedBy: "team-a", DisplayName: "Model One"}})
+
+	first := r.GetAvailableModels("openai")
+	if len(first) != 1 {
+		t.Fatalf("expected 1 model, got %d", len(first))
+	}
+	first[0]["id"] = "mutated"
+	first[0]["display_name"] = "Mutated"
+
+	second := r.GetAvailableModels("openai")
+	if got := second[0]["id"]; got != "m1" {
+		t.Fatalf("expected cached snapshot to stay isolated, got id %v", got)
+	}
+	if got := second[0]["display_name"]; got != "Model One" {
+		t.Fatalf("expected cached snapshot to stay isolated, got display_name %v", got)
+	}
+}
+
+func TestGetAvailableModelsInvalidatesCacheOnRegistryChanges(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "OpenAI", []*ModelInfo{{ID: "m1", OwnedBy: "team-a", DisplayName: "Model One"}})
+
+	models := r.GetAvailableModels("openai")
+	if len(models) != 1 {
+		t.Fatalf("expected 1 model, got %d", len(models))
+	}
+	if got := models[0]["display_name"]; got != "Model One" {
+		t.Fatalf("expected initial display_name Model One, got %v", got)
+	}
+
+	r.RegisterClient("client-1", "OpenAI", []*ModelInfo{{ID: "m1", OwnedBy: "team-a", DisplayName: "Model One Updated"}})
+	models = r.GetAvailableModels("openai")
+	if got := models[0]["display_name"]; got != "Model One Updated" {
+		t.Fatalf("expected updated display_name after cache invalidation, got %v", got)
+	}
+
+	r.SuspendClientModel("client-1", "m1", "manual")
+	models = r.GetAvailableModels("openai")
+	if len(models) != 0 {
+		t.Fatalf("expected no available models after suspension, got %d", len(models))
+	}
+
+	r.ResumeClientModel("client-1", "m1")
+	models = r.GetAvailableModels("openai")
+	if len(models) != 1 {
+		t.Fatalf("expected model to reappear after resume, got %d", len(models))
+	}
+}

internal/registry/model_registry_safety_test.go

@@ -0,0 +1,149 @@
+package registry
+
+import (
+	"testing"
+	"time"
+)
+
+func TestGetModelInfoReturnsClone(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "gemini", []*ModelInfo{{
+		ID:          "m1",
+		DisplayName: "Model One",
+		Thinking:    &ThinkingSupport{Min: 1, Max: 2, Levels: []string{"low", "high"}},
+	}})
+
+	first := r.GetModelInfo("m1", "gemini")
+	if first == nil {
+		t.Fatal("expected model info")
+	}
+	first.DisplayName = "mutated"
+	first.Thinking.Levels[0] = "mutated"
+
+	second := r.GetModelInfo("m1", "gemini")
+	if second.DisplayName != "Model One" {
+		t.Fatalf("expected cloned display name, got %q", second.DisplayName)
+	}
+	if second.Thinking == nil || len(second.Thinking.Levels) == 0 || second.Thinking.Levels[0] != "low" {
+		t.Fatalf("expected cloned thinking levels, got %+v", second.Thinking)
+	}
+}
+
+func TestGetModelsForClientReturnsClones(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "gemini", []*ModelInfo{{
+		ID:          "m1",
+		DisplayName: "Model One",
+		Thinking:    &ThinkingSupport{Levels: []string{"low", "high"}},
+	}})
+
+	first := r.GetModelsForClient("client-1")
+	if len(first) != 1 || first[0] == nil {
+		t.Fatalf("expected one model, got %+v", first)
+	}
+	first[0].DisplayName = "mutated"
+	first[0].Thinking.Levels[0] = "mutated"
+
+	second := r.GetModelsForClient("client-1")
+	if len(second) != 1 || second[0] == nil {
+		t.Fatalf("expected one model on second fetch, got %+v", second)
+	}
+	if second[0].DisplayName != "Model One" {
+		t.Fatalf("expected cloned display name, got %q", second[0].DisplayName)
+	}
+	if second[0].Thinking == nil || len(second[0].Thinking.Levels) == 0 || second[0].Thinking.Levels[0] != "low" {
+		t.Fatalf("expected cloned thinking levels, got %+v", second[0].Thinking)
+	}
+}
+
+func TestGetAvailableModelsByProviderReturnsClones(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "gemini", []*ModelInfo{{
+		ID:          "m1",
+		DisplayName: "Model One",
+		Thinking:    &ThinkingSupport{Levels: []string{"low", "high"}},
+	}})
+
+	first := r.GetAvailableModelsByProvider("gemini")
+	if len(first) != 1 || first[0] == nil {
+		t.Fatalf("expected one model, got %+v", first)
+	}
+	first[0].DisplayName = "mutated"
+	first[0].Thinking.Levels[0] = "mutated"
+
+	second := r.GetAvailableModelsByProvider("gemini")
+	if len(second) != 1 || second[0] == nil {
+		t.Fatalf("expected one model on second fetch, got %+v", second)
+	}
+	if second[0].DisplayName != "Model One" {
+		t.Fatalf("expected cloned display name, got %q", second[0].DisplayName)
+	}
+	if second[0].Thinking == nil || len(second[0].Thinking.Levels) == 0 || second[0].Thinking.Levels[0] != "low" {
+		t.Fatalf("expected cloned thinking levels, got %+v", second[0].Thinking)
+	}
+}
+
+func TestCleanupExpiredQuotasInvalidatesAvailableModelsCache(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "openai", []*ModelInfo{{ID: "m1", Created: 1}})
+	r.SetModelQuotaExceeded("client-1", "m1")
+	if models := r.GetAvailableModels("openai"); len(models) != 1 {
+		t.Fatalf("expected cooldown model to remain listed before cleanup, got %d", len(models))
+	}
+
+	r.mutex.Lock()
+	quotaTime := time.Now().Add(-6 * time.Minute)
+	r.models["m1"].QuotaExceededClients["client-1"] = &quotaTime
+	r.mutex.Unlock()
+
+	r.CleanupExpiredQuotas()
+
+	if count := r.GetModelCount("m1"); count != 1 {
+		t.Fatalf("expected model count 1 after cleanup, got %d", count)
+	}
+	models := r.GetAvailableModels("openai")
+	if len(models) != 1 {
+		t.Fatalf("expected model to stay available after cleanup, got %d", len(models))
+	}
+	if got := models[0]["id"]; got != "m1" {
+		t.Fatalf("expected model id m1, got %v", got)
+	}
+}
+
+func TestGetAvailableModelsReturnsClonedSupportedParameters(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "openai", []*ModelInfo{{
+		ID:                  "m1",
+		DisplayName:         "Model One",
+		SupportedParameters: []string{"temperature", "top_p"},
+	}})
+
+	first := r.GetAvailableModels("openai")
+	if len(first) != 1 {
+		t.Fatalf("expected one model, got %d", len(first))
+	}
+	params, ok := first[0]["supported_parameters"].([]string)
+	if !ok || len(params) != 2 {
+		t.Fatalf("expected supported_parameters slice, got %#v", first[0]["supported_parameters"])
+	}
+	params[0] = "mutated"
+
+	second := r.GetAvailableModels("openai")
+	params, ok = second[0]["supported_parameters"].([]string)
+	if !ok || len(params) != 2 || params[0] != "temperature" {
+		t.Fatalf("expected cloned supported_parameters, got %#v", second[0]["supported_parameters"])
+	}
+}
+
+func TestLookupModelInfoReturnsCloneForStaticDefinitions(t *testing.T) {
+	first := LookupModelInfo("glm-4.6")
+	if first == nil || first.Thinking == nil || len(first.Thinking.Levels) == 0 {
+		t.Fatalf("expected static model with thinking levels, got %+v", first)
+	}
+	first.Thinking.Levels[0] = "mutated"
+
+	second := LookupModelInfo("glm-4.6")
+	if second == nil || second.Thinking == nil || len(second.Thinking.Levels) == 0 || second.Thinking.Levels[0] == "mutated" {
+		t.Fatalf("expected static lookup clone, got %+v", second)
+	}
+}

internal/registry/model_updater.go

@@ -0,0 +1,198 @@
+package registry
+
+import (
+	"context"
+	_ "embed"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"sync"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+)
+
+const (
+	modelsFetchTimeout = 30 * time.Second
+)
+
+var modelsURLs = []string{
+	"https://raw.githubusercontent.com/router-for-me/models/refs/heads/main/models.json",
+	"https://models.router-for.me/models.json",
+}
+
+//go:embed models/models.json
+var embeddedModelsJSON []byte
+
+type modelStore struct {
+	mu   sync.RWMutex
+	data *staticModelsJSON
+}
+
+var modelsCatalogStore = &modelStore{}
+
+var updaterOnce sync.Once
+
+func init() {
+	// Load embedded data as fallback on startup.
+	if err := loadModelsFromBytes(embeddedModelsJSON, "embed"); err != nil {
+		panic(fmt.Sprintf("registry: failed to parse embedded models.json: %v", err))
+	}
+}
+
+// StartModelsUpdater runs a one-time models refresh on startup.
+// It blocks until the startup fetch attempt finishes so service initialization
+// can wait for the refreshed catalog before registering auth-backed models.
+// Safe to call multiple times; only one refresh will run.
+func StartModelsUpdater(ctx context.Context) {
+	updaterOnce.Do(func() {
+		runModelsUpdater(ctx)
+	})
+}
+
+func runModelsUpdater(ctx context.Context) {
+	// Try network fetch once on startup, then stop.
+	// Periodic refresh is disabled - models are only refreshed at startup.
+	tryRefreshModels(ctx)
+}
+
+func tryRefreshModels(ctx context.Context) {
+	client := &http.Client{Timeout: modelsFetchTimeout}
+	for _, url := range modelsURLs {
+		reqCtx, cancel := context.WithTimeout(ctx, modelsFetchTimeout)
+		req, err := http.NewRequestWithContext(reqCtx, "GET", url, nil)
+		if err != nil {
+			cancel()
+			log.Debugf("models fetch request creation failed for %s: %v", url, err)
+			continue
+		}
+
+		resp, err := client.Do(req)
+		if err != nil {
+			cancel()
+			log.Debugf("models fetch failed from %s: %v", url, err)
+			continue
+		}
+
+		if resp.StatusCode != 200 {
+			resp.Body.Close()
+			cancel()
+			log.Debugf("models fetch returned %d from %s", resp.StatusCode, url)
+			continue
+		}
+
+		data, err := io.ReadAll(resp.Body)
+		resp.Body.Close()
+		cancel()
+
+		if err != nil {
+			log.Debugf("models fetch read error from %s: %v", url, err)
+			continue
+		}
+
+		if err := loadModelsFromBytes(data, url); err != nil {
+			log.Warnf("models parse failed from %s: %v", url, err)
+			continue
+		}
+
+		log.Infof("models updated from %s", url)
+		return
+	}
+	log.Warn("models refresh failed from all URLs, using current data")
+}
+
+func loadModelsFromBytes(data []byte, source string) error {
+	var parsed staticModelsJSON
+	if err := json.Unmarshal(data, &parsed); err != nil {
+		return fmt.Errorf("%s: decode models catalog: %w", source, err)
+	}
+	if err := validateModelsCatalog(&parsed); err != nil {
+		return fmt.Errorf("%s: validate models catalog: %w", source, err)
+	}
+
+	modelsCatalogStore.mu.Lock()
+	modelsCatalogStore.data = &parsed
+	modelsCatalogStore.mu.Unlock()
+	return nil
+}
+
+func getModels() *staticModelsJSON {
+	modelsCatalogStore.mu.RLock()
+	defer modelsCatalogStore.mu.RUnlock()
+	return modelsCatalogStore.data
+}
+
+func validateModelsCatalog(data *staticModelsJSON) error {
+	if data == nil {
+		return fmt.Errorf("catalog is nil")
+	}
+
+	requiredSections := []struct {
+		name   string
+		models []*ModelInfo
+	}{
+		{name: "claude", models: data.Claude},
+		{name: "gemini", models: data.Gemini},
+		{name: "vertex", models: data.Vertex},
+		{name: "gemini-cli", models: data.GeminiCLI},
+		{name: "aistudio", models: data.AIStudio},
+		{name: "codex-free", models: data.CodexFree},
+		{name: "codex-team", models: data.CodexTeam},
+		{name: "codex-plus", models: data.CodexPlus},
+		{name: "codex-pro", models: data.CodexPro},
+		{name: "qwen", models: data.Qwen},
+		{name: "iflow", models: data.IFlow},
+		{name: "kimi", models: data.Kimi},
+	}
+
+	for _, section := range requiredSections {
+		if err := validateModelSection(section.name, section.models); err != nil {
+			return err
+		}
+	}
+	if err := validateAntigravitySection(data.Antigravity); err != nil {
+		return err
+	}
+	return nil
+}
+
+func validateModelSection(section string, models []*ModelInfo) error {
+	if len(models) == 0 {
+		return fmt.Errorf("%s section is empty", section)
+	}
+
+	seen := make(map[string]struct{}, len(models))
+	for i, model := range models {
+		if model == nil {
+			return fmt.Errorf("%s[%d] is null", section, i)
+		}
+		modelID := strings.TrimSpace(model.ID)
+		if modelID == "" {
+			return fmt.Errorf("%s[%d] has empty id", section, i)
+		}
+		if _, exists := seen[modelID]; exists {
+			return fmt.Errorf("%s contains duplicate model id %q", section, modelID)
+		}
+		seen[modelID] = struct{}{}
+	}
+	return nil
+}
+
+func validateAntigravitySection(configs map[string]*AntigravityModelConfig) error {
+	if len(configs) == 0 {
+		return fmt.Errorf("antigravity section is empty")
+	}
+
+	for modelID, cfg := range configs {
+		trimmedID := strings.TrimSpace(modelID)
+		if trimmedID == "" {
+			return fmt.Errorf("antigravity contains empty model id")
+		}
+		if cfg == nil {
+			return fmt.Errorf("antigravity[%q] is null", trimmedID)
+		}
+	}
+	return nil
+}

internal/runtime/executor/antigravity_executor_buildrequest_test.go

@@ -59,6 +59,7 @@ func buildRequestBodyFromPayload(t *testing.T, modelName string) map[string]any
 										"properties": {
 											"mode": {
 												"type": "string",
+												"deprecated": true,
 												"enum": ["a", "b"],
 												"enumTitles": ["A", "B"]
 											}
@@ -156,4 +157,7 @@ func assertSchemaSanitizedAndPropertyPreserved(t *testing.T, params map[string]a
 	if _, ok := mode["enumTitles"]; ok {
 		t.Fatalf("enumTitles should be removed from nested schema")
 	}
+	if _, ok := mode["deprecated"]; ok {
+		t.Fatalf("deprecated should be removed from nested schema")
+	}
 }

internal/runtime/executor/claude_executor.go

@@ -187,17 +187,15 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	}
 	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		// Decompress error responses (e.g. gzip-compressed 400 errors from Anthropic API).
-		errBody := httpResp.Body
-		if ce := httpResp.Header.Get("Content-Encoding"); ce != "" {
-			var decErr error
-			errBody, decErr = decodeResponseBody(httpResp.Body, ce)
-			if decErr != nil {
-				recordAPIResponseError(ctx, e.cfg, decErr)
-				msg := fmt.Sprintf("failed to decode error response body (encoding=%s): %v", ce, decErr)
-				logWithRequestID(ctx).Warn(msg)
-				return resp, statusErr{code: httpResp.StatusCode, msg: msg}
-			}
+		// Decompress error responses — pass the Content-Encoding value (may be empty)
+		// and let decodeResponseBody handle both header-declared and magic-byte-detected
+		// compression.  This keeps error-path behaviour consistent with the success path.
+		errBody, decErr := decodeResponseBody(httpResp.Body, httpResp.Header.Get("Content-Encoding"))
+		if decErr != nil {
+			recordAPIResponseError(ctx, e.cfg, decErr)
+			msg := fmt.Sprintf("failed to decode error response body: %v", decErr)
+			logWithRequestID(ctx).Warn(msg)
+			return resp, statusErr{code: httpResp.StatusCode, msg: msg}
 		}
 		b, readErr := io.ReadAll(errBody)
 		if readErr != nil {
@@ -352,17 +350,15 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	}
 	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		// Decompress error responses (e.g. gzip-compressed 400 errors from Anthropic API).
-		errBody := httpResp.Body
-		if ce := httpResp.Header.Get("Content-Encoding"); ce != "" {
-			var decErr error
-			errBody, decErr = decodeResponseBody(httpResp.Body, ce)
-			if decErr != nil {
-				recordAPIResponseError(ctx, e.cfg, decErr)
-				msg := fmt.Sprintf("failed to decode error response body (encoding=%s): %v", ce, decErr)
-				logWithRequestID(ctx).Warn(msg)
-				return nil, statusErr{code: httpResp.StatusCode, msg: msg}
-			}
+		// Decompress error responses — pass the Content-Encoding value (may be empty)
+		// and let decodeResponseBody handle both header-declared and magic-byte-detected
+		// compression.  This keeps error-path behaviour consistent with the success path.
+		errBody, decErr := decodeResponseBody(httpResp.Body, httpResp.Header.Get("Content-Encoding"))
+		if decErr != nil {
+			recordAPIResponseError(ctx, e.cfg, decErr)
+			msg := fmt.Sprintf("failed to decode error response body: %v", decErr)
+			logWithRequestID(ctx).Warn(msg)
+			return nil, statusErr{code: httpResp.StatusCode, msg: msg}
 		}
 		b, readErr := io.ReadAll(errBody)
 		if readErr != nil {
@@ -521,17 +517,15 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	}
 	recordAPIResponseMetadata(ctx, e.cfg, resp.StatusCode, resp.Header.Clone())
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
-		// Decompress error responses (e.g. gzip-compressed 400 errors from Anthropic API).
-		errBody := resp.Body
-		if ce := resp.Header.Get("Content-Encoding"); ce != "" {
-			var decErr error
-			errBody, decErr = decodeResponseBody(resp.Body, ce)
-			if decErr != nil {
-				recordAPIResponseError(ctx, e.cfg, decErr)
-				msg := fmt.Sprintf("failed to decode error response body (encoding=%s): %v", ce, decErr)
-				logWithRequestID(ctx).Warn(msg)
-				return cliproxyexecutor.Response{}, statusErr{code: resp.StatusCode, msg: msg}
-			}
+		// Decompress error responses — pass the Content-Encoding value (may be empty)
+		// and let decodeResponseBody handle both header-declared and magic-byte-detected
+		// compression.  This keeps error-path behaviour consistent with the success path.
+		errBody, decErr := decodeResponseBody(resp.Body, resp.Header.Get("Content-Encoding"))
+		if decErr != nil {
+			recordAPIResponseError(ctx, e.cfg, decErr)
+			msg := fmt.Sprintf("failed to decode error response body: %v", decErr)
+			logWithRequestID(ctx).Warn(msg)
+			return cliproxyexecutor.Response{}, statusErr{code: resp.StatusCode, msg: msg}
 		}
 		b, readErr := io.ReadAll(errBody)
 		if readErr != nil {
@@ -662,12 +656,61 @@ func (c *compositeReadCloser) Close() error {
 	return firstErr
 }
 
+// peekableBody wraps a bufio.Reader around the original ReadCloser so that
+// magic bytes can be inspected without consuming them from the stream.
+type peekableBody struct {
+	*bufio.Reader
+	closer io.Closer
+}
+
+func (p *peekableBody) Close() error {
+	return p.closer.Close()
+}
+
 func decodeResponseBody(body io.ReadCloser, contentEncoding string) (io.ReadCloser, error) {
 	if body == nil {
 		return nil, fmt.Errorf("response body is nil")
 	}
 	if contentEncoding == "" {
-		return body, nil
+		// No Content-Encoding header.  Attempt best-effort magic-byte detection to
+		// handle misbehaving upstreams that compress without setting the header.
+		// Only gzip (1f 8b) and zstd (28 b5 2f fd) have reliable magic sequences;
+		// br and deflate have none and are left as-is.
+		// The bufio wrapper preserves unread bytes so callers always see the full
+		// stream regardless of whether decompression was applied.
+		pb := &peekableBody{Reader: bufio.NewReader(body), closer: body}
+		magic, peekErr := pb.Peek(4)
+		if peekErr == nil || (peekErr == io.EOF && len(magic) >= 2) {
+			switch {
+			case len(magic) >= 2 && magic[0] == 0x1f && magic[1] == 0x8b:
+				gzipReader, gzErr := gzip.NewReader(pb)
+				if gzErr != nil {
+					_ = pb.Close()
+					return nil, fmt.Errorf("magic-byte gzip: failed to create reader: %w", gzErr)
+				}
+				return &compositeReadCloser{
+					Reader: gzipReader,
+					closers: []func() error{
+						gzipReader.Close,
+						pb.Close,
+					},
+				}, nil
+			case len(magic) >= 4 && magic[0] == 0x28 && magic[1] == 0xb5 && magic[2] == 0x2f && magic[3] == 0xfd:
+				decoder, zdErr := zstd.NewReader(pb)
+				if zdErr != nil {
+					_ = pb.Close()
+					return nil, fmt.Errorf("magic-byte zstd: failed to create reader: %w", zdErr)
+				}
+				return &compositeReadCloser{
+					Reader: decoder,
+					closers: []func() error{
+						func() error { decoder.Close(); return nil },
+						pb.Close,
+					},
+				}, nil
+			}
+		}
+		return pb, nil
 	}
 	encodings := strings.Split(contentEncoding, ",")
 	for _, raw := range encodings {
@@ -844,11 +887,15 @@ func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string,
 		r.Header.Set("User-Agent", hdrDefault(hd.UserAgent, "claude-cli/2.1.63 (external, cli)"))
 	}
 	r.Header.Set("Connection", "keep-alive")
-	r.Header.Set("Accept-Encoding", "gzip, deflate, br, zstd")
 	if stream {
 		r.Header.Set("Accept", "text/event-stream")
+		// SSE streams must not be compressed: the downstream scanner reads
+		// line-delimited text and cannot parse compressed bytes.  Using
+		// "identity" tells the upstream to send an uncompressed stream.
+		r.Header.Set("Accept-Encoding", "identity")
 	} else {
 		r.Header.Set("Accept", "application/json")
+		r.Header.Set("Accept-Encoding", "gzip, deflate, br, zstd")
 	}
 	// Keep OS/Arch mapping dynamic (not configurable).
 	// They intentionally continue to derive from runtime.GOOS/runtime.GOARCH.
@@ -857,6 +904,12 @@ func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string,
 		attrs = auth.Attributes
 	}
 	util.ApplyCustomHeadersFromAttrs(r, attrs)
+	// Re-enforce Accept-Encoding: identity after ApplyCustomHeadersFromAttrs, which
+	// may override it with a user-configured value.  Compressed SSE breaks the line
+	// scanner regardless of user preference, so this is non-negotiable for streams.
+	if stream {
+		r.Header.Set("Accept-Encoding", "identity")
+	}
 }
 
 func claudeCreds(a *cliproxyauth.Auth) (apiKey, baseURL string) {
@@ -1213,6 +1266,10 @@ func checkSystemInstructionsWithMode(payload []byte, strictMode bool) []byte {
 			}
 			return true
 		})
+	} else if system.Type == gjson.String && system.String() != "" {
+		partJSON := `{"type":"text","cache_control":{"type":"ephemeral"}}`
+		partJSON, _ = sjson.Set(partJSON, "text", system.String())
+		result += "," + partJSON
 	}
 	result += "]"
 
@@ -1432,25 +1489,27 @@ func countCacheControlsMap(root map[string]any) int {
 	return count
 }
 
-func normalizeTTLForBlock(obj map[string]any, seen5m *bool) {
+func normalizeTTLForBlock(obj map[string]any, seen5m *bool) bool {
 	ccRaw, exists := obj["cache_control"]
 	if !exists {
-		return
+		return false
 	}
 	cc, ok := asObject(ccRaw)
 	if !ok {
 		*seen5m = true
-		return
+		return false
 	}
 	ttlRaw, ttlExists := cc["ttl"]
 	ttl, ttlIsString := ttlRaw.(string)
 	if !ttlExists || !ttlIsString || ttl != "1h" {
 		*seen5m = true
-		return
+		return false
 	}
 	if *seen5m {
 		delete(cc, "ttl")
+		return true
 	}
+	return false
 }
 
 func findLastCacheControlIndex(arr []any) int {
@@ -1546,11 +1605,14 @@ func normalizeCacheControlTTL(payload []byte) []byte {
 	}
 
 	seen5m := false
+	modified := false
 
 	if tools, ok := asArray(root["tools"]); ok {
 		for _, tool := range tools {
 			if obj, ok := asObject(tool); ok {
-				normalizeTTLForBlock(obj, &seen5m)
+				if normalizeTTLForBlock(obj, &seen5m) {
+					modified = true
+				}
 			}
 		}
 	}
@@ -1558,7 +1620,9 @@ func normalizeCacheControlTTL(payload []byte) []byte {
 	if system, ok := asArray(root["system"]); ok {
 		for _, item := range system {
 			if obj, ok := asObject(item); ok {
-				normalizeTTLForBlock(obj, &seen5m)
+				if normalizeTTLForBlock(obj, &seen5m) {
+					modified = true
+				}
 			}
 		}
 	}
@@ -1575,12 +1639,17 @@ func normalizeCacheControlTTL(payload []byte) []byte {
 			}
 			for _, item := range content {
 				if obj, ok := asObject(item); ok {
-					normalizeTTLForBlock(obj, &seen5m)
+					if normalizeTTLForBlock(obj, &seen5m) {
+						modified = true
+					}
 				}
 			}
 		}
 	}
 
+	if !modified {
+		return payload
+	}
 	return marshalPayloadObject(payload, root)
 }
 

internal/runtime/executor/claude_executor_test.go

@@ -2,6 +2,7 @@ package executor
 
 import (
 	"bytes"
+	"compress/gzip"
 	"context"
 	"io"
 	"net/http"
@@ -9,6 +10,7 @@ import (
 	"strings"
 	"testing"
 
+	"github.com/klauspost/compress/zstd"
 	"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"
@@ -367,6 +369,19 @@ func TestNormalizeCacheControlTTL_DowngradesLaterOneHourBlocks(t *testing.T) {
 	}
 }
 
+func TestNormalizeCacheControlTTL_PreservesOriginalBytesWhenNoChange(t *testing.T) {
+	// Payload where no TTL normalization is needed (all blocks use 1h with no
+	// preceding 5m block). The text intentionally contains HTML chars (<, >, &)
+	// that json.Marshal would escape to \u003c etc., altering byte identity.
+	payload := []byte(`{"tools":[{"name":"t1","cache_control":{"type":"ephemeral","ttl":"1h"}}],"system":[{"type":"text","text":"<system-reminder>foo & bar</system-reminder>","cache_control":{"type":"ephemeral","ttl":"1h"}}],"messages":[{"role":"user","content":[{"type":"text","text":"hello"}]}]}`)
+
+	out := normalizeCacheControlTTL(payload)
+
+	if !bytes.Equal(out, payload) {
+		t.Fatalf("normalizeCacheControlTTL altered bytes when no change was needed.\noriginal: %s\ngot:      %s", payload, out)
+	}
+}
+
 func TestEnforceCacheControlLimit_StripsNonLastToolBeforeMessages(t *testing.T) {
 	payload := []byte(`{
 		"tools": [
@@ -583,3 +598,469 @@ func testClaudeExecutorInvalidCompressedErrorBody(
 		t.Fatalf("expected status code 400, got: %v", err)
 	}
 }
+
+// TestClaudeExecutor_ExecuteStream_SetsIdentityAcceptEncoding verifies that streaming
+// requests use Accept-Encoding: identity so the upstream cannot respond with a
+// compressed SSE body that would silently break the line scanner.
+func TestClaudeExecutor_ExecuteStream_SetsIdentityAcceptEncoding(t *testing.T) {
+	var gotEncoding, gotAccept string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotEncoding = r.Header.Get("Accept-Encoding")
+		gotAccept = r.Header.Get("Accept")
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte("data: {\"type\":\"message_stop\"}\n\n"))
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	result, err := executor.ExecuteStream(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err != nil {
+		t.Fatalf("ExecuteStream error: %v", err)
+	}
+	for chunk := range result.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("unexpected chunk error: %v", chunk.Err)
+		}
+	}
+
+	if gotEncoding != "identity" {
+		t.Errorf("Accept-Encoding = %q, want %q", gotEncoding, "identity")
+	}
+	if gotAccept != "text/event-stream" {
+		t.Errorf("Accept = %q, want %q", gotAccept, "text/event-stream")
+	}
+}
+
+// TestClaudeExecutor_Execute_SetsCompressedAcceptEncoding verifies that non-streaming
+// requests keep the full accept-encoding to allow response compression (which
+// decodeResponseBody handles correctly).
+func TestClaudeExecutor_Execute_SetsCompressedAcceptEncoding(t *testing.T) {
+	var gotEncoding, gotAccept string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotEncoding = r.Header.Get("Accept-Encoding")
+		gotAccept = r.Header.Get("Accept")
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"id":"msg_1","type":"message","model":"claude-3-5-sonnet-20241022","role":"assistant","content":[{"type":"text","text":"hi"}],"usage":{"input_tokens":1,"output_tokens":1}}`))
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	_, err := executor.Execute(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err != nil {
+		t.Fatalf("Execute error: %v", err)
+	}
+
+	if gotEncoding != "gzip, deflate, br, zstd" {
+		t.Errorf("Accept-Encoding = %q, want %q", gotEncoding, "gzip, deflate, br, zstd")
+	}
+	if gotAccept != "application/json" {
+		t.Errorf("Accept = %q, want %q", gotAccept, "application/json")
+	}
+}
+
+// TestClaudeExecutor_ExecuteStream_GzipSuccessBodyDecoded verifies that a streaming
+// HTTP 200 response with Content-Encoding: gzip is correctly decompressed before
+// the line scanner runs, so SSE chunks are not silently dropped.
+func TestClaudeExecutor_ExecuteStream_GzipSuccessBodyDecoded(t *testing.T) {
+	var buf bytes.Buffer
+	gz := gzip.NewWriter(&buf)
+	_, _ = gz.Write([]byte("data: {\"type\":\"message_stop\"}\n"))
+	_ = gz.Close()
+	compressedBody := buf.Bytes()
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		w.Header().Set("Content-Encoding", "gzip")
+		_, _ = w.Write(compressedBody)
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	result, err := executor.ExecuteStream(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err != nil {
+		t.Fatalf("ExecuteStream error: %v", err)
+	}
+
+	var combined strings.Builder
+	for chunk := range result.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("chunk error: %v", chunk.Err)
+		}
+		combined.Write(chunk.Payload)
+	}
+
+	if combined.Len() == 0 {
+		t.Fatal("expected at least one chunk from gzip-encoded SSE body, got none (body was not decompressed)")
+	}
+	if !strings.Contains(combined.String(), "message_stop") {
+		t.Errorf("expected SSE content in chunks, got: %q", combined.String())
+	}
+}
+
+// TestDecodeResponseBody_MagicByteGzipNoHeader verifies that decodeResponseBody
+// detects gzip-compressed content via magic bytes even when Content-Encoding is absent.
+func TestDecodeResponseBody_MagicByteGzipNoHeader(t *testing.T) {
+	const plaintext = "data: {\"type\":\"message_stop\"}\n"
+
+	var buf bytes.Buffer
+	gz := gzip.NewWriter(&buf)
+	_, _ = gz.Write([]byte(plaintext))
+	_ = gz.Close()
+
+	rc := io.NopCloser(&buf)
+	decoded, err := decodeResponseBody(rc, "")
+	if err != nil {
+		t.Fatalf("decodeResponseBody error: %v", err)
+	}
+	defer decoded.Close()
+
+	got, err := io.ReadAll(decoded)
+	if err != nil {
+		t.Fatalf("ReadAll error: %v", err)
+	}
+	if string(got) != plaintext {
+		t.Errorf("decoded = %q, want %q", got, plaintext)
+	}
+}
+
+// TestDecodeResponseBody_PlainTextNoHeader verifies that decodeResponseBody returns
+// plain text untouched when Content-Encoding is absent and no magic bytes match.
+func TestDecodeResponseBody_PlainTextNoHeader(t *testing.T) {
+	const plaintext = "data: {\"type\":\"message_stop\"}\n"
+	rc := io.NopCloser(strings.NewReader(plaintext))
+	decoded, err := decodeResponseBody(rc, "")
+	if err != nil {
+		t.Fatalf("decodeResponseBody error: %v", err)
+	}
+	defer decoded.Close()
+
+	got, err := io.ReadAll(decoded)
+	if err != nil {
+		t.Fatalf("ReadAll error: %v", err)
+	}
+	if string(got) != plaintext {
+		t.Errorf("decoded = %q, want %q", got, plaintext)
+	}
+}
+
+// TestClaudeExecutor_ExecuteStream_GzipNoContentEncodingHeader verifies the full
+// pipeline: when the upstream returns a gzip-compressed SSE body WITHOUT setting
+// Content-Encoding (a misbehaving upstream), the magic-byte sniff in
+// decodeResponseBody still decompresses it, so chunks reach the caller.
+func TestClaudeExecutor_ExecuteStream_GzipNoContentEncodingHeader(t *testing.T) {
+	var buf bytes.Buffer
+	gz := gzip.NewWriter(&buf)
+	_, _ = gz.Write([]byte("data: {\"type\":\"message_stop\"}\n"))
+	_ = gz.Close()
+	compressedBody := buf.Bytes()
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		// Intentionally omit Content-Encoding to simulate misbehaving upstream.
+		_, _ = w.Write(compressedBody)
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	result, err := executor.ExecuteStream(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err != nil {
+		t.Fatalf("ExecuteStream error: %v", err)
+	}
+
+	var combined strings.Builder
+	for chunk := range result.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("chunk error: %v", chunk.Err)
+		}
+		combined.Write(chunk.Payload)
+	}
+
+	if combined.Len() == 0 {
+		t.Fatal("expected chunks from gzip body without Content-Encoding header, got none (magic-byte sniff failed)")
+	}
+	if !strings.Contains(combined.String(), "message_stop") {
+		t.Errorf("unexpected chunk content: %q", combined.String())
+	}
+}
+
+// TestClaudeExecutor_ExecuteStream_AcceptEncodingOverrideCannotBypassIdentity verifies
+// that injecting Accept-Encoding via auth.Attributes cannot override the stream
+// path's enforced identity encoding.
+func TestClaudeExecutor_ExecuteStream_AcceptEncodingOverrideCannotBypassIdentity(t *testing.T) {
+	var gotEncoding string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotEncoding = r.Header.Get("Accept-Encoding")
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte("data: {\"type\":\"message_stop\"}\n\n"))
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	// Inject Accept-Encoding via the custom header attribute mechanism.
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":                "key-123",
+		"base_url":               server.URL,
+		"header:Accept-Encoding": "gzip, deflate, br, zstd",
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	result, err := executor.ExecuteStream(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err != nil {
+		t.Fatalf("ExecuteStream error: %v", err)
+	}
+	for chunk := range result.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("unexpected chunk error: %v", chunk.Err)
+		}
+	}
+
+	if gotEncoding != "identity" {
+		t.Errorf("Accept-Encoding = %q; stream path must enforce identity regardless of auth.Attributes override", gotEncoding)
+	}
+}
+
+// TestDecodeResponseBody_MagicByteZstdNoHeader verifies that decodeResponseBody
+// detects zstd-compressed content via magic bytes (28 b5 2f fd) even when
+// Content-Encoding is absent.
+func TestDecodeResponseBody_MagicByteZstdNoHeader(t *testing.T) {
+	const plaintext = "data: {\"type\":\"message_stop\"}\n"
+
+	var buf bytes.Buffer
+	enc, err := zstd.NewWriter(&buf)
+	if err != nil {
+		t.Fatalf("zstd.NewWriter: %v", err)
+	}
+	_, _ = enc.Write([]byte(plaintext))
+	_ = enc.Close()
+
+	rc := io.NopCloser(&buf)
+	decoded, err := decodeResponseBody(rc, "")
+	if err != nil {
+		t.Fatalf("decodeResponseBody error: %v", err)
+	}
+	defer decoded.Close()
+
+	got, err := io.ReadAll(decoded)
+	if err != nil {
+		t.Fatalf("ReadAll error: %v", err)
+	}
+	if string(got) != plaintext {
+		t.Errorf("decoded = %q, want %q", got, plaintext)
+	}
+}
+
+// TestClaudeExecutor_Execute_GzipErrorBodyNoContentEncodingHeader verifies that the
+// error path (4xx) correctly decompresses a gzip body even when the upstream omits
+// the Content-Encoding header.  This closes the gap left by PR #1771, which only
+// fixed header-declared compression on the error path.
+func TestClaudeExecutor_Execute_GzipErrorBodyNoContentEncodingHeader(t *testing.T) {
+	const errJSON = `{"type":"error","error":{"type":"invalid_request_error","message":"test error"}}`
+
+	var buf bytes.Buffer
+	gz := gzip.NewWriter(&buf)
+	_, _ = gz.Write([]byte(errJSON))
+	_ = gz.Close()
+	compressedBody := buf.Bytes()
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		// Intentionally omit Content-Encoding to simulate misbehaving upstream.
+		w.WriteHeader(http.StatusBadRequest)
+		_, _ = w.Write(compressedBody)
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	_, err := executor.Execute(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err == nil {
+		t.Fatal("expected an error for 400 response, got nil")
+	}
+	if !strings.Contains(err.Error(), "test error") {
+		t.Errorf("error message should contain decompressed JSON, got: %q", err.Error())
+	}
+}
+
+// TestClaudeExecutor_ExecuteStream_GzipErrorBodyNoContentEncodingHeader verifies
+// the same for the streaming executor: 4xx gzip body without Content-Encoding is
+// decoded and the error message is readable.
+func TestClaudeExecutor_ExecuteStream_GzipErrorBodyNoContentEncodingHeader(t *testing.T) {
+	const errJSON = `{"type":"error","error":{"type":"invalid_request_error","message":"stream test error"}}`
+
+	var buf bytes.Buffer
+	gz := gzip.NewWriter(&buf)
+	_, _ = gz.Write([]byte(errJSON))
+	_ = gz.Close()
+	compressedBody := buf.Bytes()
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		// Intentionally omit Content-Encoding to simulate misbehaving upstream.
+		w.WriteHeader(http.StatusBadRequest)
+		_, _ = w.Write(compressedBody)
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	_, err := executor.ExecuteStream(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err == nil {
+		t.Fatal("expected an error for 400 response, got nil")
+	}
+	if !strings.Contains(err.Error(), "stream test error") {
+		t.Errorf("error message should contain decompressed JSON, got: %q", err.Error())
+	}
+}
+
+// Test case 1: String system prompt is preserved and converted to a content block
+func TestCheckSystemInstructionsWithMode_StringSystemPreserved(t *testing.T) {
+	payload := []byte(`{"system":"You are a helpful assistant.","messages":[{"role":"user","content":"hi"}]}`)
+
+	out := checkSystemInstructionsWithMode(payload, false)
+
+	system := gjson.GetBytes(out, "system")
+	if !system.IsArray() {
+		t.Fatalf("system should be an array, got %s", system.Type)
+	}
+
+	blocks := system.Array()
+	if len(blocks) != 3 {
+		t.Fatalf("expected 3 system blocks, got %d", len(blocks))
+	}
+
+	if !strings.HasPrefix(blocks[0].Get("text").String(), "x-anthropic-billing-header:") {
+		t.Fatalf("blocks[0] should be billing header, got %q", blocks[0].Get("text").String())
+	}
+	if blocks[1].Get("text").String() != "You are a Claude agent, built on Anthropic's Claude Agent SDK." {
+		t.Fatalf("blocks[1] should be agent block, got %q", blocks[1].Get("text").String())
+	}
+	if blocks[2].Get("text").String() != "You are a helpful assistant." {
+		t.Fatalf("blocks[2] should be user system prompt, got %q", blocks[2].Get("text").String())
+	}
+	if blocks[2].Get("cache_control.type").String() != "ephemeral" {
+		t.Fatalf("blocks[2] should have cache_control.type=ephemeral")
+	}
+}
+
+// Test case 2: Strict mode drops the string system prompt
+func TestCheckSystemInstructionsWithMode_StringSystemStrict(t *testing.T) {
+	payload := []byte(`{"system":"You are a helpful assistant.","messages":[{"role":"user","content":"hi"}]}`)
+
+	out := checkSystemInstructionsWithMode(payload, true)
+
+	blocks := gjson.GetBytes(out, "system").Array()
+	if len(blocks) != 2 {
+		t.Fatalf("strict mode should produce 2 blocks, got %d", len(blocks))
+	}
+}
+
+// Test case 3: Empty string system prompt does not produce a spurious block
+func TestCheckSystemInstructionsWithMode_EmptyStringSystemIgnored(t *testing.T) {
+	payload := []byte(`{"system":"","messages":[{"role":"user","content":"hi"}]}`)
+
+	out := checkSystemInstructionsWithMode(payload, false)
+
+	blocks := gjson.GetBytes(out, "system").Array()
+	if len(blocks) != 2 {
+		t.Fatalf("empty string system should produce 2 blocks, got %d", len(blocks))
+	}
+}
+
+// Test case 4: Array system prompt is unaffected by the string handling
+func TestCheckSystemInstructionsWithMode_ArraySystemStillWorks(t *testing.T) {
+	payload := []byte(`{"system":[{"type":"text","text":"Be concise."}],"messages":[{"role":"user","content":"hi"}]}`)
+
+	out := checkSystemInstructionsWithMode(payload, false)
+
+	blocks := gjson.GetBytes(out, "system").Array()
+	if len(blocks) != 3 {
+		t.Fatalf("expected 3 system blocks, got %d", len(blocks))
+	}
+	if blocks[2].Get("text").String() != "Be concise." {
+		t.Fatalf("blocks[2] should be user system prompt, got %q", blocks[2].Get("text").String())
+	}
+}
+
+// Test case 5: Special characters in string system prompt survive conversion
+func TestCheckSystemInstructionsWithMode_StringWithSpecialChars(t *testing.T) {
+	payload := []byte(`{"system":"Use <xml> tags & \"quotes\" in output.","messages":[{"role":"user","content":"hi"}]}`)
+
+	out := checkSystemInstructionsWithMode(payload, false)
+
+	blocks := gjson.GetBytes(out, "system").Array()
+	if len(blocks) != 3 {
+		t.Fatalf("expected 3 system blocks, got %d", len(blocks))
+	}
+	if blocks[2].Get("text").String() != `Use <xml> tags & "quotes" in output.` {
+		t.Fatalf("blocks[2] text mangled, got %q", blocks[2].Get("text").String())
+	}
+}

internal/runtime/executor/codex_executor.go

@@ -616,6 +616,10 @@ func (e *CodexExecutor) cacheHelper(ctx context.Context, from sdktranslator.Form
 		if promptCacheKey.Exists() {
 			cache.ID = promptCacheKey.String()
 		}
+	} else if from == "openai" {
+		if apiKey := strings.TrimSpace(apiKeyFromContext(ctx)); apiKey != "" {
+			cache.ID = uuid.NewSHA1(uuid.NameSpaceOID, []byte("cli-proxy-api:codex:prompt-cache:"+apiKey)).String()
+		}
 	}
 
 	if cache.ID != "" {

internal/runtime/executor/codex_executor_cache_test.go

@@ -0,0 +1,64 @@
+package executor
+
+import (
+	"context"
+	"io"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/uuid"
+	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 TestCodexExecutorCacheHelper_OpenAIChatCompletions_StablePromptCacheKeyFromAPIKey(t *testing.T) {
+	recorder := httptest.NewRecorder()
+	ginCtx, _ := gin.CreateTestContext(recorder)
+	ginCtx.Set("apiKey", "test-api-key")
+
+	ctx := context.WithValue(context.Background(), "gin", ginCtx)
+	executor := &CodexExecutor{}
+	rawJSON := []byte(`{"model":"gpt-5.3-codex","stream":true}`)
+	req := cliproxyexecutor.Request{
+		Model:   "gpt-5.3-codex",
+		Payload: []byte(`{"model":"gpt-5.3-codex"}`),
+	}
+	url := "https://example.com/responses"
+
+	httpReq, err := executor.cacheHelper(ctx, sdktranslator.FromString("openai"), url, req, rawJSON)
+	if err != nil {
+		t.Fatalf("cacheHelper error: %v", err)
+	}
+
+	body, errRead := io.ReadAll(httpReq.Body)
+	if errRead != nil {
+		t.Fatalf("read request body: %v", errRead)
+	}
+
+	expectedKey := uuid.NewSHA1(uuid.NameSpaceOID, []byte("cli-proxy-api:codex:prompt-cache:test-api-key")).String()
+	gotKey := gjson.GetBytes(body, "prompt_cache_key").String()
+	if gotKey != expectedKey {
+		t.Fatalf("prompt_cache_key = %q, want %q", gotKey, expectedKey)
+	}
+	if gotConversation := httpReq.Header.Get("Conversation_id"); gotConversation != expectedKey {
+		t.Fatalf("Conversation_id = %q, want %q", gotConversation, expectedKey)
+	}
+	if gotSession := httpReq.Header.Get("Session_id"); gotSession != expectedKey {
+		t.Fatalf("Session_id = %q, want %q", gotSession, expectedKey)
+	}
+
+	httpReq2, err := executor.cacheHelper(ctx, sdktranslator.FromString("openai"), url, req, rawJSON)
+	if err != nil {
+		t.Fatalf("cacheHelper error (second call): %v", err)
+	}
+	body2, errRead2 := io.ReadAll(httpReq2.Body)
+	if errRead2 != nil {
+		t.Fatalf("read request body (second call): %v", errRead2)
+	}
+	gotKey2 := gjson.GetBytes(body2, "prompt_cache_key").String()
+	if gotKey2 != expectedKey {
+		t.Fatalf("prompt_cache_key (second call) = %q, want %q", gotKey2, expectedKey)
+	}
+}

internal/runtime/executor/codex_websockets_executor.go

@@ -31,7 +31,7 @@ import (
 )
 
 const (
-	codexResponsesWebsocketBetaHeaderValue = "responses_websockets=2026-02-04"
+	codexResponsesWebsocketBetaHeaderValue = "responses_websockets=2026-02-06"
 	codexResponsesWebsocketIdleTimeout     = 5 * time.Minute
 	codexResponsesWebsocketHandshakeTO     = 30 * time.Second
 )
@@ -57,11 +57,6 @@ type codexWebsocketSession struct {
 	wsURL  string
 	authID string
 
-	// connCreateSent tracks whether a `response.create` message has been successfully sent
-	// on the current websocket connection. The upstream expects the first message on each
-	// connection to be `response.create`.
-	connCreateSent bool
-
 	writeMu sync.Mutex
 
 	activeMu     sync.Mutex
@@ -212,13 +207,7 @@ func (e *CodexWebsocketsExecutor) Execute(ctx context.Context, auth *cliproxyaut
 		defer sess.reqMu.Unlock()
 	}
 
-	allowAppend := true
-	if sess != nil {
-		sess.connMu.Lock()
-		allowAppend = sess.connCreateSent
-		sess.connMu.Unlock()
-	}
-	wsReqBody := buildCodexWebsocketRequestBody(body, allowAppend)
+	wsReqBody := buildCodexWebsocketRequestBody(body)
 	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 		URL:       wsURL,
 		Method:    "WEBSOCKET",
@@ -280,10 +269,7 @@ func (e *CodexWebsocketsExecutor) Execute(ctx context.Context, auth *cliproxyaut
 			// execution session.
 			connRetry, _, errDialRetry := e.ensureUpstreamConn(ctx, auth, sess, authID, wsURL, wsHeaders)
 			if errDialRetry == nil && connRetry != nil {
-				sess.connMu.Lock()
-				allowAppend = sess.connCreateSent
-				sess.connMu.Unlock()
-				wsReqBodyRetry := buildCodexWebsocketRequestBody(body, allowAppend)
+				wsReqBodyRetry := buildCodexWebsocketRequestBody(body)
 				recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 					URL:       wsURL,
 					Method:    "WEBSOCKET",
@@ -312,7 +298,6 @@ func (e *CodexWebsocketsExecutor) Execute(ctx context.Context, auth *cliproxyaut
 			return resp, errSend
 		}
 	}
-	markCodexWebsocketCreateSent(sess, conn, wsReqBody)
 
 	for {
 		if ctx != nil && ctx.Err() != nil {
@@ -403,26 +388,20 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 	wsHeaders = applyCodexWebsocketHeaders(ctx, wsHeaders, auth, apiKey)
 
 	var authID, authLabel, authType, authValue string
-	if auth != nil {
-		authID = auth.ID
-		authLabel = auth.Label
-		authType, authValue = auth.AccountInfo()
-	}
+	authID = auth.ID
+	authLabel = auth.Label
+	authType, authValue = auth.AccountInfo()
 
 	executionSessionID := executionSessionIDFromOptions(opts)
 	var sess *codexWebsocketSession
 	if executionSessionID != "" {
 		sess = e.getOrCreateSession(executionSessionID)
-		sess.reqMu.Lock()
+		if sess != nil {
+			sess.reqMu.Lock()
+		}
 	}
 
-	allowAppend := true
-	if sess != nil {
-		sess.connMu.Lock()
-		allowAppend = sess.connCreateSent
-		sess.connMu.Unlock()
-	}
-	wsReqBody := buildCodexWebsocketRequestBody(body, allowAppend)
+	wsReqBody := buildCodexWebsocketRequestBody(body)
 	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 		URL:       wsURL,
 		Method:    "WEBSOCKET",
@@ -483,10 +462,7 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 				sess.reqMu.Unlock()
 				return nil, errDialRetry
 			}
-			sess.connMu.Lock()
-			allowAppend = sess.connCreateSent
-			sess.connMu.Unlock()
-			wsReqBodyRetry := buildCodexWebsocketRequestBody(body, allowAppend)
+			wsReqBodyRetry := buildCodexWebsocketRequestBody(body)
 			recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 				URL:       wsURL,
 				Method:    "WEBSOCKET",
@@ -515,7 +491,6 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 			return nil, errSend
 		}
 	}
-	markCodexWebsocketCreateSent(sess, conn, wsReqBody)
 
 	out := make(chan cliproxyexecutor.StreamChunk)
 	go func() {
@@ -657,31 +632,14 @@ func writeCodexWebsocketMessage(sess *codexWebsocketSession, conn *websocket.Con
 	return conn.WriteMessage(websocket.TextMessage, payload)
 }
 
-func buildCodexWebsocketRequestBody(body []byte, allowAppend bool) []byte {
+func buildCodexWebsocketRequestBody(body []byte) []byte {
 	if len(body) == 0 {
 		return nil
 	}
 
-	// Codex CLI websocket v2 uses `response.create` with `previous_response_id` for incremental turns.
-	// The upstream ChatGPT Codex websocket currently rejects that with close 1008 (policy violation).
-	// Fall back to v1 `response.append` semantics on the same websocket connection to keep the session alive.
-	//
-	// NOTE: The upstream expects the first websocket event on each connection to be `response.create`,
-	// so we only use `response.append` after we have initialized the current connection.
-	if allowAppend {
-		if prev := strings.TrimSpace(gjson.GetBytes(body, "previous_response_id").String()); prev != "" {
-			inputNode := gjson.GetBytes(body, "input")
-			wsReqBody := []byte(`{}`)
-			wsReqBody, _ = sjson.SetBytes(wsReqBody, "type", "response.append")
-			if inputNode.Exists() && inputNode.IsArray() && strings.TrimSpace(inputNode.Raw) != "" {
-				wsReqBody, _ = sjson.SetRawBytes(wsReqBody, "input", []byte(inputNode.Raw))
-				return wsReqBody
-			}
-			wsReqBody, _ = sjson.SetRawBytes(wsReqBody, "input", []byte("[]"))
-			return wsReqBody
-		}
-	}
-
+	// Match codex-rs websocket v2 semantics: every request is `response.create`.
+	// Incremental follow-up turns continue on the same websocket using
+	// `previous_response_id` + incremental `input`, not `response.append`.
 	wsReqBody, errSet := sjson.SetBytes(bytes.Clone(body), "type", "response.create")
 	if errSet == nil && len(wsReqBody) > 0 {
 		return wsReqBody
@@ -725,21 +683,6 @@ func readCodexWebsocketMessage(ctx context.Context, sess *codexWebsocketSession,
 	}
 }
 
-func markCodexWebsocketCreateSent(sess *codexWebsocketSession, conn *websocket.Conn, payload []byte) {
-	if sess == nil || conn == nil || len(payload) == 0 {
-		return
-	}
-	if strings.TrimSpace(gjson.GetBytes(payload, "type").String()) != "response.create" {
-		return
-	}
-
-	sess.connMu.Lock()
-	if sess.conn == conn {
-		sess.connCreateSent = true
-	}
-	sess.connMu.Unlock()
-}
-
 func newProxyAwareWebsocketDialer(cfg *config.Config, auth *cliproxyauth.Auth) *websocket.Dialer {
 	dialer := &websocket.Dialer{
 		Proxy:             http.ProxyFromEnvironment,
@@ -1017,36 +960,6 @@ func closeHTTPResponseBody(resp *http.Response, logPrefix string) {
 	}
 }
 
-func closeOnContextDone(ctx context.Context, conn *websocket.Conn) chan struct{} {
-	done := make(chan struct{})
-	if ctx == nil || conn == nil {
-		return done
-	}
-	go func() {
-		select {
-		case <-done:
-		case <-ctx.Done():
-			_ = conn.Close()
-		}
-	}()
-	return done
-}
-
-func cancelReadOnContextDone(ctx context.Context, conn *websocket.Conn) chan struct{} {
-	done := make(chan struct{})
-	if ctx == nil || conn == nil {
-		return done
-	}
-	go func() {
-		select {
-		case <-done:
-		case <-ctx.Done():
-			_ = conn.SetReadDeadline(time.Now())
-		}
-	}()
-	return done
-}
-
 func executionSessionIDFromOptions(opts cliproxyexecutor.Options) string {
 	if len(opts.Metadata) == 0 {
 		return ""
@@ -1120,7 +1033,6 @@ func (e *CodexWebsocketsExecutor) ensureUpstreamConn(ctx context.Context, auth *
 	sess.conn = conn
 	sess.wsURL = wsURL
 	sess.authID = authID
-	sess.connCreateSent = false
 	sess.readerConn = conn
 	sess.connMu.Unlock()
 
@@ -1206,7 +1118,6 @@ func (e *CodexWebsocketsExecutor) invalidateUpstreamConn(sess *codexWebsocketSes
 		return
 	}
 	sess.conn = nil
-	sess.connCreateSent = false
 	if sess.readerConn == conn {
 		sess.readerConn = nil
 	}
@@ -1273,7 +1184,6 @@ func (e *CodexWebsocketsExecutor) closeExecutionSession(sess *codexWebsocketSess
 	authID := sess.authID
 	wsURL := sess.wsURL
 	sess.conn = nil
-	sess.connCreateSent = false
 	if sess.readerConn == conn {
 		sess.readerConn = nil
 	}

internal/runtime/executor/codex_websockets_executor_test.go

@@ -0,0 +1,36 @@
+package executor
+
+import (
+	"context"
+	"net/http"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestBuildCodexWebsocketRequestBodyPreservesPreviousResponseID(t *testing.T) {
+	body := []byte(`{"model":"gpt-5-codex","previous_response_id":"resp-1","input":[{"type":"message","id":"msg-1"}]}`)
+
+	wsReqBody := buildCodexWebsocketRequestBody(body)
+
+	if got := gjson.GetBytes(wsReqBody, "type").String(); got != "response.create" {
+		t.Fatalf("type = %s, want response.create", got)
+	}
+	if got := gjson.GetBytes(wsReqBody, "previous_response_id").String(); got != "resp-1" {
+		t.Fatalf("previous_response_id = %s, want resp-1", got)
+	}
+	if gjson.GetBytes(wsReqBody, "input.0.id").String() != "msg-1" {
+		t.Fatalf("input item id mismatch")
+	}
+	if got := gjson.GetBytes(wsReqBody, "type").String(); got == "response.append" {
+		t.Fatalf("unexpected websocket request type: %s", got)
+	}
+}
+
+func TestApplyCodexWebsocketHeadersDefaultsToCurrentResponsesBeta(t *testing.T) {
+	headers := applyCodexWebsocketHeaders(context.Background(), http.Header{}, nil, "")
+
+	if got := headers.Get("OpenAI-Beta"); got != codexResponsesWebsocketBetaHeaderValue {
+		t.Fatalf("OpenAI-Beta = %s, want %s", got, codexResponsesWebsocketBetaHeaderValue)
+	}
+}

internal/runtime/executor/gemini_vertex_executor.go

@@ -460,7 +460,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 
 	// For API key auth, use simpler URL format without project/location
 	if baseURL == "" {
-		baseURL = "https://generativelanguage.googleapis.com"
+		baseURL = "https://aiplatform.googleapis.com"
 	}
 	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, action)
 	if opts.Alt != "" && action != "countTokens" {
@@ -683,7 +683,7 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	action := getVertexAction(baseModel, true)
 	// For API key auth, use simpler URL format without project/location
 	if baseURL == "" {
-		baseURL = "https://generativelanguage.googleapis.com"
+		baseURL = "https://aiplatform.googleapis.com"
 	}
 	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, action)
 	// Imagen models don't support streaming, skip SSE params
@@ -883,7 +883,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 
 	// For API key auth, use simpler URL format without project/location
 	if baseURL == "" {
-		baseURL = "https://generativelanguage.googleapis.com"
+		baseURL = "https://aiplatform.googleapis.com"
 	}
 	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, "countTokens")
 

internal/runtime/executor/github_copilot_executor.go

@@ -14,6 +14,7 @@ import (
 	"github.com/google/uuid"
 	copilotauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/copilot"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -521,9 +522,9 @@ func detectLastConversationRole(body []byte) string {
 			}
 
 			switch item.Get("type").String() {
-			case "function_call", "function_call_arguments":
+			case "function_call", "function_call_arguments", "computer_call":
 				return "assistant"
-			case "function_call_output", "function_call_response", "tool_result":
+			case "function_call_output", "function_call_response", "tool_result", "computer_call_output":
 				return "tool"
 			}
 		}
@@ -831,6 +832,10 @@ func normalizeGitHubCopilotResponsesTools(body []byte) []byte {
 		if tools.IsArray() {
 			for _, tool := range tools.Array() {
 				toolType := tool.Get("type").String()
+				if isGitHubCopilotResponsesBuiltinTool(toolType) {
+					filtered, _ = sjson.SetRaw(filtered, "-1", tool.Raw)
+					continue
+				}
 				// Accept OpenAI format (type="function") and Claude format
 				// (no type field, but has top-level name + input_schema).
 				if toolType != "" && toolType != "function" {
@@ -878,6 +883,10 @@ func normalizeGitHubCopilotResponsesTools(body []byte) []byte {
 	}
 	if toolChoice.Type == gjson.JSON {
 		choiceType := toolChoice.Get("type").String()
+		if isGitHubCopilotResponsesBuiltinTool(choiceType) {
+			body, _ = sjson.SetRawBytes(body, "tool_choice", []byte(toolChoice.Raw))
+			return body
+		}
 		if choiceType == "function" {
 			name := toolChoice.Get("name").String()
 			if name == "" {
@@ -895,6 +904,15 @@ func normalizeGitHubCopilotResponsesTools(body []byte) []byte {
 	return body
 }
 
+func isGitHubCopilotResponsesBuiltinTool(toolType string) bool {
+	switch strings.TrimSpace(toolType) {
+	case "computer", "computer_use_preview":
+		return true
+	default:
+		return false
+	}
+}
+
 func collectTextFromNode(node gjson.Result) string {
 	if !node.Exists() {
 		return ""
@@ -1264,3 +1282,99 @@ func translateGitHubCopilotResponsesStreamToClaude(line []byte, param *any) []st
 func isHTTPSuccess(statusCode int) bool {
 	return statusCode >= 200 && statusCode < 300
 }
+
+const (
+	// defaultCopilotContextLength is the default context window for unknown Copilot models.
+	defaultCopilotContextLength = 128000
+	// defaultCopilotMaxCompletionTokens is the default max output tokens for unknown Copilot models.
+	defaultCopilotMaxCompletionTokens = 16384
+)
+
+// FetchGitHubCopilotModels dynamically fetches available models from the GitHub Copilot API.
+// It exchanges the GitHub access token stored in auth.Metadata for a Copilot API token,
+// then queries the /models endpoint. Falls back to the static registry on any failure.
+func FetchGitHubCopilotModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *config.Config) []*registry.ModelInfo {
+	if auth == nil {
+		log.Debug("github-copilot: auth is nil, using static models")
+		return registry.GetGitHubCopilotModels()
+	}
+
+	accessToken := metaStringValue(auth.Metadata, "access_token")
+	if accessToken == "" {
+		log.Debug("github-copilot: no access_token in auth metadata, using static models")
+		return registry.GetGitHubCopilotModels()
+	}
+
+	copilotAuth := copilotauth.NewCopilotAuth(cfg)
+
+	entries, err := copilotAuth.ListModelsWithGitHubToken(ctx, accessToken)
+	if err != nil {
+		log.Warnf("github-copilot: failed to fetch dynamic models: %v, using static models", err)
+		return registry.GetGitHubCopilotModels()
+	}
+
+	if len(entries) == 0 {
+		log.Debug("github-copilot: API returned no models, using static models")
+		return registry.GetGitHubCopilotModels()
+	}
+
+	// Build a lookup from the static definitions so we can enrich dynamic entries
+	// with known context lengths, thinking support, etc.
+	staticMap := make(map[string]*registry.ModelInfo)
+	for _, m := range registry.GetGitHubCopilotModels() {
+		staticMap[m.ID] = m
+	}
+
+	now := time.Now().Unix()
+	models := make([]*registry.ModelInfo, 0, len(entries))
+	seen := make(map[string]struct{}, len(entries))
+	for _, entry := range entries {
+		if entry.ID == "" {
+			continue
+		}
+		// Deduplicate model IDs to avoid incorrect reference counting.
+		if _, dup := seen[entry.ID]; dup {
+			continue
+		}
+		seen[entry.ID] = struct{}{}
+
+		m := &registry.ModelInfo{
+			ID:      entry.ID,
+			Object:  "model",
+			Created: now,
+			OwnedBy: "github-copilot",
+			Type:    "github-copilot",
+		}
+
+		if entry.Created > 0 {
+			m.Created = entry.Created
+		}
+		if entry.Name != "" {
+			m.DisplayName = entry.Name
+		} else {
+			m.DisplayName = entry.ID
+		}
+
+		// Merge known metadata from the static fallback list
+		if static, ok := staticMap[entry.ID]; ok {
+			if m.DisplayName == entry.ID && static.DisplayName != "" {
+				m.DisplayName = static.DisplayName
+			}
+			m.Description = static.Description
+			m.ContextLength = static.ContextLength
+			m.MaxCompletionTokens = static.MaxCompletionTokens
+			m.SupportedEndpoints = static.SupportedEndpoints
+			m.Thinking = static.Thinking
+		} else {
+			// Sensible defaults for models not in the static list
+			m.Description = entry.ID + " via GitHub Copilot"
+			m.ContextLength = defaultCopilotContextLength
+			m.MaxCompletionTokens = defaultCopilotMaxCompletionTokens
+		}
+
+		models = append(models, m)
+	}
+
+	log.Infof("github-copilot: fetched %d models from API", len(models))
+	return models
+}

internal/runtime/executor/kiro_executor.go

@@ -2458,7 +2458,6 @@ func (e *KiroExecutor) streamToChannel(ctx context.Context, body io.Reader, out
 	reader := bufio.NewReaderSize(body, 20*1024*1024) // 20MB buffer to match other providers
 	var totalUsage usage.Detail
 	var hasToolUses bool          // Track if any tool uses were emitted
-	var hasTruncatedTools bool    // Track if any tool uses were truncated
 	var upstreamStopReason string // Track stop_reason from upstream events
 
 	// Tool use state tracking for input buffering and deduplication
@@ -3286,59 +3285,9 @@ func (e *KiroExecutor) streamToChannel(ctx context.Context, body io.Reader, out
 
 			// Emit completed tool uses
 			for _, tu := range completedToolUses {
-				// Check if this tool was truncated - emit with SOFT_LIMIT_REACHED marker
+				// Skip truncated tools - don't emit fake marker tool_use
 				if tu.IsTruncated {
-					hasTruncatedTools = true
-					log.Infof("kiro: streamToChannel emitting truncated tool with SOFT_LIMIT_REACHED: %s (ID: %s)", tu.Name, tu.ToolUseID)
-
-					// Close text block if open
-					if isTextBlockOpen && contentBlockIndex >= 0 {
-						blockStop := kiroclaude.BuildClaudeContentBlockStopEvent(contentBlockIndex)
-						sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
-						for _, chunk := range sseData {
-							if chunk != "" {
-								out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
-							}
-						}
-						isTextBlockOpen = false
-					}
-
-					contentBlockIndex++
-
-					// Emit tool_use with SOFT_LIMIT_REACHED marker input
-					blockStart := kiroclaude.BuildClaudeContentBlockStartEvent(contentBlockIndex, "tool_use", tu.ToolUseID, tu.Name)
-					sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStart, &translatorParam)
-					for _, chunk := range sseData {
-						if chunk != "" {
-							out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
-						}
-					}
-
-					// Build SOFT_LIMIT_REACHED marker input
-					markerInput := map[string]interface{}{
-						"_status":  "SOFT_LIMIT_REACHED",
-						"_message": "Tool output was truncated. Split content into smaller chunks (max 300 lines). Due to potential model hallucination, you MUST re-fetch the current working directory and generate the correct file_path.",
-					}
-
-					markerJSON, _ := json.Marshal(markerInput)
-					inputDelta := kiroclaude.BuildClaudeInputJsonDeltaEvent(string(markerJSON), contentBlockIndex)
-					sseData = sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, inputDelta, &translatorParam)
-					for _, chunk := range sseData {
-						if chunk != "" {
-							out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
-						}
-					}
-
-					// Close tool_use block
-					blockStop := kiroclaude.BuildClaudeContentBlockStopEvent(contentBlockIndex)
-					sseData = sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
-					for _, chunk := range sseData {
-						if chunk != "" {
-							out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
-						}
-					}
-
-					hasToolUses = true // Keep this so stop_reason = tool_use
+					log.Warnf("kiro: streamToChannel skipping truncated tool: %s (ID: %s)", tu.Name, tu.ToolUseID)
 					continue
 				}
 
@@ -3640,12 +3589,7 @@ func (e *KiroExecutor) streamToChannel(ctx context.Context, body io.Reader, out
 	}
 
 	// Determine stop reason: prefer upstream, then detect tool_use, default to end_turn
-	// SOFT_LIMIT_REACHED: Keep stop_reason = "tool_use" so Claude continues the loop
 	stopReason := upstreamStopReason
-	if hasTruncatedTools {
-		// Log that we're using SOFT_LIMIT_REACHED approach
-		log.Infof("kiro: streamToChannel using SOFT_LIMIT_REACHED - keeping stop_reason=tool_use for truncated tools")
-	}
 	if stopReason == "" {
 		if hasToolUses {
 			stopReason = "tool_use"

internal/thinking/apply.go

@@ -257,7 +257,10 @@ func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromForma
 	if suffixResult.HasSuffix {
 		config = parseSuffixToConfig(suffixResult.RawSuffix, toFormat, modelID)
 	} else {
-		config = extractThinkingConfig(body, toFormat)
+		config = extractThinkingConfig(body, fromFormat)
+		if !hasThinkingConfig(config) && fromFormat != toFormat {
+			config = extractThinkingConfig(body, toFormat)
+		}
 	}
 
 	if !hasThinkingConfig(config) {
@@ -293,6 +296,9 @@ func normalizeUserDefinedConfig(config ThinkingConfig, fromFormat, toFormat stri
 	if config.Mode != ModeLevel {
 		return config
 	}
+	if toFormat == "claude" {
+		return config
+	}
 	if !isBudgetCapableProvider(toFormat) {
 		return config
 	}

internal/thinking/apply_user_defined_test.go

@@ -0,0 +1,55 @@
+package thinking_test
+
+import (
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/claude"
+	"github.com/tidwall/gjson"
+)
+
+func TestApplyThinking_UserDefinedClaudePreservesAdaptiveLevel(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	clientID := "test-user-defined-claude-" + t.Name()
+	modelID := "custom-claude-4-6"
+	reg.RegisterClient(clientID, "claude", []*registry.ModelInfo{{ID: modelID, UserDefined: true}})
+	t.Cleanup(func() {
+		reg.UnregisterClient(clientID)
+	})
+
+	tests := []struct {
+		name  string
+		model string
+		body  []byte
+	}{
+		{
+			name:  "claude adaptive effort body",
+			model: modelID,
+			body:  []byte(`{"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`),
+		},
+		{
+			name:  "suffix level",
+			model: modelID + "(high)",
+			body:  []byte(`{}`),
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			out, err := thinking.ApplyThinking(tt.body, tt.model, "openai", "claude", "claude")
+			if err != nil {
+				t.Fatalf("ApplyThinking() error = %v", err)
+			}
+			if got := gjson.GetBytes(out, "thinking.type").String(); got != "adaptive" {
+				t.Fatalf("thinking.type = %q, want %q, body=%s", got, "adaptive", string(out))
+			}
+			if got := gjson.GetBytes(out, "output_config.effort").String(); got != "high" {
+				t.Fatalf("output_config.effort = %q, want %q, body=%s", got, "high", string(out))
+			}
+			if gjson.GetBytes(out, "thinking.budget_tokens").Exists() {
+				t.Fatalf("thinking.budget_tokens should be removed, body=%s", string(out))
+			}
+		})
+	}
+}

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -431,6 +431,33 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 		out, _ = sjson.SetRaw(out, "request.tools", toolsJSON)
 	}
 
+	// tool_choice
+	toolChoiceResult := gjson.GetBytes(rawJSON, "tool_choice")
+	if toolChoiceResult.Exists() {
+		toolChoiceType := ""
+		toolChoiceName := ""
+		if toolChoiceResult.IsObject() {
+			toolChoiceType = toolChoiceResult.Get("type").String()
+			toolChoiceName = toolChoiceResult.Get("name").String()
+		} else if toolChoiceResult.Type == gjson.String {
+			toolChoiceType = toolChoiceResult.String()
+		}
+
+		switch toolChoiceType {
+		case "auto":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "AUTO")
+		case "none":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "NONE")
+		case "any":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "ANY")
+		case "tool":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "ANY")
+			if toolChoiceName != "" {
+				out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.allowedFunctionNames", []string{toolChoiceName})
+			}
+		}
+	}
+
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
 	if t := gjson.GetBytes(rawJSON, "thinking"); enableThoughtTranslate && t.Exists() && t.IsObject() {
 		switch t.Get("type").String() {
@@ -441,9 +468,19 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		case "adaptive", "auto":
-			// Keep adaptive/auto as a high level sentinel; ApplyThinking resolves it
-			// to model-specific max capability.
-			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			// For adaptive thinking:
+			// - If output_config.effort is explicitly present, pass through as thinkingLevel.
+			// - Otherwise, treat it as "enabled with target-model maximum" and emit high.
+			// ApplyThinking handles clamping to target model's supported levels.
+			effort := ""
+			if v := gjson.GetBytes(rawJSON, "output_config.effort"); v.Exists() && v.Type == gjson.String {
+				effort = strings.ToLower(strings.TrimSpace(v.String()))
+			}
+			if effort != "" {
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", effort)
+			} else {
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			}
 			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}

internal/translator/antigravity/claude/antigravity_claude_request_test.go

@@ -193,6 +193,42 @@ func TestConvertClaudeRequestToAntigravity_ToolDeclarations(t *testing.T) {
 	}
 }
 
+func TestConvertClaudeRequestToAntigravity_ToolChoice_SpecificTool(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "gemini-3-flash-preview",
+		"messages": [
+			{
+				"role": "user",
+				"content": [
+					{"type": "text", "text": "hi"}
+				]
+			}
+		],
+		"tools": [
+			{
+				"name": "json",
+				"description": "A JSON tool",
+				"input_schema": {
+					"type": "object",
+					"properties": {}
+				}
+			}
+		],
+		"tool_choice": {"type": "tool", "name": "json"}
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("gemini-3-flash-preview", inputJSON, false)
+	outputStr := string(output)
+
+	if got := gjson.Get(outputStr, "request.toolConfig.functionCallingConfig.mode").String(); got != "ANY" {
+		t.Fatalf("Expected toolConfig.functionCallingConfig.mode 'ANY', got '%s'", got)
+	}
+	allowed := gjson.Get(outputStr, "request.toolConfig.functionCallingConfig.allowedFunctionNames").Array()
+	if len(allowed) != 1 || allowed[0].String() != "json" {
+		t.Fatalf("Expected allowedFunctionNames ['json'], got %s", gjson.Get(outputStr, "request.toolConfig.functionCallingConfig.allowedFunctionNames").Raw)
+	}
+}
+
 func TestConvertClaudeRequestToAntigravity_ToolUse(t *testing.T) {
 	inputJSON := []byte(`{
 		"model": "claude-3-5-sonnet-20240620",

internal/translator/antigravity/claude/antigravity_claude_response.go

@@ -15,6 +15,7 @@ import (
 	"time"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/cache"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	log "github.com/sirupsen/logrus"
 
 	"github.com/tidwall/gjson"
@@ -256,7 +257,7 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 
 				// Create the tool use block with unique ID and function details
 				data := fmt.Sprintf(`{"type":"content_block_start","index":%d,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`, params.ResponseIndex)
-				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1)))
+				data, _ = sjson.Set(data, "content_block.id", util.SanitizeClaudeToolID(fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1))))
 				data, _ = sjson.Set(data, "content_block.name", fcName)
 				output = output + fmt.Sprintf("data: %s\n\n\n", data)
 

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

@@ -212,6 +212,33 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 							} else {
 								log.Warnf("Unknown file name extension '%s' in user message, skip", ext)
 							}
+						case "input_audio":
+							audioData := item.Get("input_audio.data").String()
+							audioFormat := item.Get("input_audio.format").String()
+							if audioData != "" {
+								audioMimeMap := map[string]string{
+									"mp3":       "audio/mpeg",
+									"wav":       "audio/wav",
+									"ogg":       "audio/ogg",
+									"flac":      "audio/flac",
+									"aac":       "audio/aac",
+									"webm":      "audio/webm",
+									"pcm16":     "audio/pcm",
+									"g711_ulaw": "audio/basic",
+									"g711_alaw": "audio/basic",
+								}
+								mimeType := "audio/wav"
+								if audioFormat != "" {
+									if mapped, ok := audioMimeMap[audioFormat]; ok {
+										mimeType = mapped
+									} else {
+										mimeType = "audio/" + audioFormat
+									}
+								}
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".inlineData.mime_type", mimeType)
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".inlineData.data", audioData)
+								p++
+							}
 						}
 					}
 				}

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

@@ -203,46 +203,9 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 					msg, _ = sjson.SetRaw(msg, "content.-1", part)
 				} else if contentResult.Exists() && contentResult.IsArray() {
 					contentResult.ForEach(func(_, part gjson.Result) bool {
-						partType := part.Get("type").String()
-
-						switch partType {
-						case "text":
-							textPart := `{"type":"text","text":""}`
-							textPart, _ = sjson.Set(textPart, "text", part.Get("text").String())
-							msg, _ = sjson.SetRaw(msg, "content.-1", textPart)
-
-						case "image_url":
-							// Convert OpenAI image format to Claude Code format
-							imageURL := part.Get("image_url.url").String()
-							if strings.HasPrefix(imageURL, "data:") {
-								// Extract base64 data and media type from data URL
-								parts := strings.Split(imageURL, ",")
-								if len(parts) == 2 {
-									mediaTypePart := strings.Split(parts[0], ";")[0]
-									mediaType := strings.TrimPrefix(mediaTypePart, "data:")
-									data := parts[1]
-
-									imagePart := `{"type":"image","source":{"type":"base64","media_type":"","data":""}}`
-									imagePart, _ = sjson.Set(imagePart, "source.media_type", mediaType)
-									imagePart, _ = sjson.Set(imagePart, "source.data", data)
-									msg, _ = sjson.SetRaw(msg, "content.-1", imagePart)
-								}
-							}
-
-						case "file":
-							fileData := part.Get("file.file_data").String()
-							if strings.HasPrefix(fileData, "data:") {
-								semicolonIdx := strings.Index(fileData, ";")
-								commaIdx := strings.Index(fileData, ",")
-								if semicolonIdx != -1 && commaIdx != -1 && commaIdx > semicolonIdx {
-									mediaType := strings.TrimPrefix(fileData[:semicolonIdx], "data:")
-									data := fileData[commaIdx+1:]
-									docPart := `{"type":"document","source":{"type":"base64","media_type":"","data":""}}`
-									docPart, _ = sjson.Set(docPart, "source.media_type", mediaType)
-									docPart, _ = sjson.Set(docPart, "source.data", data)
-									msg, _ = sjson.SetRaw(msg, "content.-1", docPart)
-								}
-							}
+						claudePart := convertOpenAIContentPartToClaudePart(part)
+						if claudePart != "" {
+							msg, _ = sjson.SetRaw(msg, "content.-1", claudePart)
 						}
 						return true
 					})
@@ -291,11 +254,16 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 			case "tool":
 				// Handle tool result messages conversion
 				toolCallID := message.Get("tool_call_id").String()
-				content := message.Get("content").String()
+				toolContentResult := message.Get("content")
 
 				msg := `{"role":"user","content":[{"type":"tool_result","tool_use_id":"","content":""}]}`
 				msg, _ = sjson.Set(msg, "content.0.tool_use_id", toolCallID)
-				msg, _ = sjson.Set(msg, "content.0.content", content)
+				toolResultContent, toolResultContentRaw := convertOpenAIToolResultContent(toolContentResult)
+				if toolResultContentRaw {
+					msg, _ = sjson.SetRaw(msg, "content.0.content", toolResultContent)
+				} else {
+					msg, _ = sjson.Set(msg, "content.0.content", toolResultContent)
+				}
 				out, _ = sjson.SetRaw(out, "messages.-1", msg)
 				messageIndex++
 			}
@@ -358,3 +326,110 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 
 	return []byte(out)
 }
+
+func convertOpenAIContentPartToClaudePart(part gjson.Result) string {
+	switch part.Get("type").String() {
+	case "text":
+		textPart := `{"type":"text","text":""}`
+		textPart, _ = sjson.Set(textPart, "text", part.Get("text").String())
+		return textPart
+
+	case "image_url":
+		return convertOpenAIImageURLToClaudePart(part.Get("image_url.url").String())
+
+	case "file":
+		fileData := part.Get("file.file_data").String()
+		if strings.HasPrefix(fileData, "data:") {
+			semicolonIdx := strings.Index(fileData, ";")
+			commaIdx := strings.Index(fileData, ",")
+			if semicolonIdx != -1 && commaIdx != -1 && commaIdx > semicolonIdx {
+				mediaType := strings.TrimPrefix(fileData[:semicolonIdx], "data:")
+				data := fileData[commaIdx+1:]
+				docPart := `{"type":"document","source":{"type":"base64","media_type":"","data":""}}`
+				docPart, _ = sjson.Set(docPart, "source.media_type", mediaType)
+				docPart, _ = sjson.Set(docPart, "source.data", data)
+				return docPart
+			}
+		}
+	}
+
+	return ""
+}
+
+func convertOpenAIImageURLToClaudePart(imageURL string) string {
+	if imageURL == "" {
+		return ""
+	}
+
+	if strings.HasPrefix(imageURL, "data:") {
+		parts := strings.SplitN(imageURL, ",", 2)
+		if len(parts) != 2 {
+			return ""
+		}
+
+		mediaTypePart := strings.SplitN(parts[0], ";", 2)[0]
+		mediaType := strings.TrimPrefix(mediaTypePart, "data:")
+		if mediaType == "" {
+			mediaType = "application/octet-stream"
+		}
+
+		imagePart := `{"type":"image","source":{"type":"base64","media_type":"","data":""}}`
+		imagePart, _ = sjson.Set(imagePart, "source.media_type", mediaType)
+		imagePart, _ = sjson.Set(imagePart, "source.data", parts[1])
+		return imagePart
+	}
+
+	imagePart := `{"type":"image","source":{"type":"url","url":""}}`
+	imagePart, _ = sjson.Set(imagePart, "source.url", imageURL)
+	return imagePart
+}
+
+func convertOpenAIToolResultContent(content gjson.Result) (string, bool) {
+	if !content.Exists() {
+		return "", false
+	}
+
+	if content.Type == gjson.String {
+		return content.String(), false
+	}
+
+	if content.IsArray() {
+		claudeContent := "[]"
+		partCount := 0
+
+		content.ForEach(func(_, part gjson.Result) bool {
+			if part.Type == gjson.String {
+				textPart := `{"type":"text","text":""}`
+				textPart, _ = sjson.Set(textPart, "text", part.String())
+				claudeContent, _ = sjson.SetRaw(claudeContent, "-1", textPart)
+				partCount++
+				return true
+			}
+
+			claudePart := convertOpenAIContentPartToClaudePart(part)
+			if claudePart != "" {
+				claudeContent, _ = sjson.SetRaw(claudeContent, "-1", claudePart)
+				partCount++
+			}
+			return true
+		})
+
+		if partCount > 0 || len(content.Array()) == 0 {
+			return claudeContent, true
+		}
+
+		return content.Raw, false
+	}
+
+	if content.IsObject() {
+		claudePart := convertOpenAIContentPartToClaudePart(content)
+		if claudePart != "" {
+			claudeContent := "[]"
+			claudeContent, _ = sjson.SetRaw(claudeContent, "-1", claudePart)
+			return claudeContent, true
+		}
+		return content.Raw, false
+	}
+
+	return content.Raw, false
+}

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

@@ -0,0 +1,137 @@
+package chat_completions
+
+import (
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestConvertOpenAIRequestToClaude_ToolResultTextAndBase64Image(t *testing.T) {
+	inputJSON := `{
+		"model": "gpt-4.1",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": "",
+				"tool_calls": [
+					{
+						"id": "call_1",
+						"type": "function",
+						"function": {
+							"name": "do_work",
+							"arguments": "{\"a\":1}"
+						}
+					}
+				]
+			},
+			{
+				"role": "tool",
+				"tool_call_id": "call_1",
+				"content": [
+					{"type": "text", "text": "tool ok"},
+					{
+						"type": "image_url",
+						"image_url": {
+							"url": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUg=="
+						}
+					}
+				]
+			}
+		]
+	}`
+
+	result := ConvertOpenAIRequestToClaude("claude-sonnet-4-5", []byte(inputJSON), false)
+	resultJSON := gjson.ParseBytes(result)
+	messages := resultJSON.Get("messages").Array()
+
+	if len(messages) != 2 {
+		t.Fatalf("Expected 2 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	}
+
+	toolResult := messages[1].Get("content.0")
+	if got := toolResult.Get("type").String(); got != "tool_result" {
+		t.Fatalf("Expected content[0].type %q, got %q", "tool_result", got)
+	}
+	if got := toolResult.Get("tool_use_id").String(); got != "call_1" {
+		t.Fatalf("Expected tool_use_id %q, got %q", "call_1", got)
+	}
+
+	toolContent := toolResult.Get("content")
+	if !toolContent.IsArray() {
+		t.Fatalf("Expected tool_result content array, got %s", toolContent.Raw)
+	}
+	if got := toolContent.Get("0.type").String(); got != "text" {
+		t.Fatalf("Expected first tool_result part type %q, got %q", "text", got)
+	}
+	if got := toolContent.Get("0.text").String(); got != "tool ok" {
+		t.Fatalf("Expected first tool_result part text %q, got %q", "tool ok", got)
+	}
+	if got := toolContent.Get("1.type").String(); got != "image" {
+		t.Fatalf("Expected second tool_result part type %q, got %q", "image", got)
+	}
+	if got := toolContent.Get("1.source.type").String(); got != "base64" {
+		t.Fatalf("Expected image source type %q, got %q", "base64", got)
+	}
+	if got := toolContent.Get("1.source.media_type").String(); got != "image/png" {
+		t.Fatalf("Expected image media type %q, got %q", "image/png", got)
+	}
+	if got := toolContent.Get("1.source.data").String(); got != "iVBORw0KGgoAAAANSUhEUg==" {
+		t.Fatalf("Unexpected base64 image data: %q", got)
+	}
+}
+
+func TestConvertOpenAIRequestToClaude_ToolResultURLImageOnly(t *testing.T) {
+	inputJSON := `{
+		"model": "gpt-4.1",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": "",
+				"tool_calls": [
+					{
+						"id": "call_1",
+						"type": "function",
+						"function": {
+							"name": "do_work",
+							"arguments": "{\"a\":1}"
+						}
+					}
+				]
+			},
+			{
+				"role": "tool",
+				"tool_call_id": "call_1",
+				"content": [
+					{
+						"type": "image_url",
+						"image_url": {
+							"url": "https://example.com/tool.png"
+						}
+					}
+				]
+			}
+		]
+	}`
+
+	result := ConvertOpenAIRequestToClaude("claude-sonnet-4-5", []byte(inputJSON), false)
+	resultJSON := gjson.ParseBytes(result)
+	messages := resultJSON.Get("messages").Array()
+
+	if len(messages) != 2 {
+		t.Fatalf("Expected 2 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	}
+
+	toolContent := messages[1].Get("content.0.content")
+	if !toolContent.IsArray() {
+		t.Fatalf("Expected tool_result content array, got %s", toolContent.Raw)
+	}
+	if got := toolContent.Get("0.type").String(); got != "image" {
+		t.Fatalf("Expected tool_result part type %q, got %q", "image", got)
+	}
+	if got := toolContent.Get("0.source.type").String(); got != "url" {
+		t.Fatalf("Expected image source type %q, got %q", "url", got)
+	}
+	if got := toolContent.Get("0.source.url").String(); got != "https://example.com/tool.png" {
+		t.Fatalf("Unexpected image URL: %q", got)
+	}
+}

internal/translator/codex/claude/codex_claude_request.go

@@ -255,6 +255,8 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 			tool, _ = sjson.SetRaw(tool, "parameters", normalizeToolParameters(toolResult.Get("input_schema").Raw))
 			tool, _ = sjson.Delete(tool, "input_schema")
 			tool, _ = sjson.Delete(tool, "parameters.$schema")
+			tool, _ = sjson.Delete(tool, "cache_control")
+			tool, _ = sjson.Delete(tool, "defer_loading")
 			tool, _ = sjson.Set(tool, "strict", false)
 			template, _ = sjson.SetRaw(template, "tools.-1", tool)
 		}

internal/translator/codex/claude/codex_claude_response.go

@@ -12,6 +12,7 @@ import (
 	"fmt"
 	"strings"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -141,7 +142,7 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 			(*param).(*ConvertCodexResponseToClaudeParams).HasReceivedArgumentsDelta = false
 			template = `{"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`
 			template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
-			template, _ = sjson.Set(template, "content_block.id", itemResult.Get("call_id").String())
+			template, _ = sjson.Set(template, "content_block.id", util.SanitizeClaudeToolID(itemResult.Get("call_id").String()))
 			{
 				// Restore original tool name if shortened
 				name := itemResult.Get("name").String()
@@ -310,7 +311,7 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 				}
 
 				toolBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
-				toolBlock, _ = sjson.Set(toolBlock, "id", item.Get("call_id").String())
+				toolBlock, _ = sjson.Set(toolBlock, "id", util.SanitizeClaudeToolID(item.Get("call_id").String()))
 				toolBlock, _ = sjson.Set(toolBlock, "name", name)
 				inputRaw := "{}"
 				if argsStr := item.Get("arguments").String(); argsStr != "" && gjson.Valid(argsStr) {

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

@@ -74,8 +74,13 @@ func ConvertCodexResponseToOpenAI(_ context.Context, modelName string, originalR
 	}
 
 	// Extract and set the model version.
+	cachedModel := (*param).(*ConvertCliToOpenAIParams).Model
 	if modelResult := gjson.GetBytes(rawJSON, "model"); modelResult.Exists() {
 		template, _ = sjson.Set(template, "model", modelResult.String())
+	} else if cachedModel != "" {
+		template, _ = sjson.Set(template, "model", cachedModel)
+	} else if modelName != "" {
+		template, _ = sjson.Set(template, "model", modelName)
 	}
 
 	template, _ = sjson.Set(template, "created", (*param).(*ConvertCliToOpenAIParams).CreatedAt)

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

@@ -0,0 +1,47 @@
+package chat_completions
+
+import (
+	"context"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestConvertCodexResponseToOpenAI_StreamSetsModelFromResponseCreated(t *testing.T) {
+	ctx := context.Background()
+	var param any
+
+	modelName := "gpt-5.3-codex"
+
+	out := ConvertCodexResponseToOpenAI(ctx, modelName, nil, nil, []byte(`data: {"type":"response.created","response":{"id":"resp_123","created_at":1700000000,"model":"gpt-5.3-codex"}}`), &param)
+	if len(out) != 0 {
+		t.Fatalf("expected no output for response.created, got %d chunks", len(out))
+	}
+
+	out = ConvertCodexResponseToOpenAI(ctx, modelName, nil, nil, []byte(`data: {"type":"response.output_text.delta","delta":"hello"}`), &param)
+	if len(out) != 1 {
+		t.Fatalf("expected 1 chunk, got %d", len(out))
+	}
+
+	gotModel := gjson.Get(out[0], "model").String()
+	if gotModel != modelName {
+		t.Fatalf("expected model %q, got %q", modelName, gotModel)
+	}
+}
+
+func TestConvertCodexResponseToOpenAI_FirstChunkUsesRequestModelName(t *testing.T) {
+	ctx := context.Background()
+	var param any
+
+	modelName := "gpt-5.3-codex"
+
+	out := ConvertCodexResponseToOpenAI(ctx, modelName, nil, nil, []byte(`data: {"type":"response.output_text.delta","delta":"hello"}`), &param)
+	if len(out) != 1 {
+		t.Fatalf("expected 1 chunk, got %d", len(out))
+	}
+
+	gotModel := gjson.Get(out[0], "model").String()
+	if gotModel != modelName {
+		t.Fatalf("expected model %q, got %q", modelName, gotModel)
+	}
+}

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

@@ -25,7 +25,12 @@ func ConvertOpenAIResponsesRequestToCodex(modelName string, inputRawJSON []byte,
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "max_completion_tokens")
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "temperature")
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "top_p")
-	rawJSON, _ = sjson.DeleteBytes(rawJSON, "service_tier")
+	if v := gjson.GetBytes(rawJSON, "service_tier"); v.Exists() {
+		if v.String() != "priority" {
+			rawJSON, _ = sjson.DeleteBytes(rawJSON, "service_tier")
+		}
+	}
+
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "truncation")
 	rawJSON = applyResponsesCompactionCompatibility(rawJSON)
 

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

@@ -6,24 +6,14 @@ import (
 	"fmt"
 
 	"github.com/tidwall/gjson"
-	"github.com/tidwall/sjson"
 )
 
 // ConvertCodexResponseToOpenAIResponses converts OpenAI Chat Completions streaming chunks
 // to OpenAI Responses SSE events (response.*).
 
-func ConvertCodexResponseToOpenAIResponses(ctx context.Context, modelName string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, param *any) []string {
+func ConvertCodexResponseToOpenAIResponses(_ context.Context, _ string, _, _, rawJSON []byte, _ *any) []string {
 	if bytes.HasPrefix(rawJSON, []byte("data:")) {
 		rawJSON = bytes.TrimSpace(rawJSON[5:])
-		if typeResult := gjson.GetBytes(rawJSON, "type"); typeResult.Exists() {
-			typeStr := typeResult.String()
-			if typeStr == "response.created" || typeStr == "response.in_progress" || typeStr == "response.completed" {
-				if gjson.GetBytes(rawJSON, "response.instructions").Exists() {
-					instructions := gjson.GetBytes(originalRequestRawJSON, "instructions").String()
-					rawJSON, _ = sjson.SetBytes(rawJSON, "response.instructions", instructions)
-				}
-			}
-		}
 		out := fmt.Sprintf("data: %s", string(rawJSON))
 		return []string{out}
 	}
@@ -32,17 +22,12 @@ func ConvertCodexResponseToOpenAIResponses(ctx context.Context, modelName string
 
 // ConvertCodexResponseToOpenAIResponsesNonStream builds a single Responses JSON
 // from a non-streaming OpenAI Chat Completions response.
-func ConvertCodexResponseToOpenAIResponsesNonStream(_ context.Context, modelName string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, _ *any) string {
+func ConvertCodexResponseToOpenAIResponsesNonStream(_ context.Context, _ string, _, _, rawJSON []byte, _ *any) string {
 	rootResult := gjson.ParseBytes(rawJSON)
 	// Verify this is a response.completed event
 	if rootResult.Get("type").String() != "response.completed" {
 		return ""
 	}
 	responseResult := rootResult.Get("response")
-	template := responseResult.Raw
-	if responseResult.Get("instructions").Exists() {
-		instructions := gjson.GetBytes(originalRequestRawJSON, "instructions").String()
-		template, _ = sjson.Set(template, "instructions", instructions)
-	}
-	return template
+	return responseResult.Raw
 }

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

@@ -156,6 +156,7 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 				tool, _ = sjson.Delete(tool, "input_examples")
 				tool, _ = sjson.Delete(tool, "type")
 				tool, _ = sjson.Delete(tool, "cache_control")
+				tool, _ = sjson.Delete(tool, "defer_loading")
 				if gjson.Valid(tool) && gjson.Parse(tool).IsObject() {
 					if !hasTools {
 						out, _ = sjson.SetRaw(out, "request.tools", `[{"functionDeclarations":[]}]`)
@@ -171,6 +172,33 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 		}
 	}
 
+	// tool_choice
+	toolChoiceResult := gjson.GetBytes(rawJSON, "tool_choice")
+	if toolChoiceResult.Exists() {
+		toolChoiceType := ""
+		toolChoiceName := ""
+		if toolChoiceResult.IsObject() {
+			toolChoiceType = toolChoiceResult.Get("type").String()
+			toolChoiceName = toolChoiceResult.Get("name").String()
+		} else if toolChoiceResult.Type == gjson.String {
+			toolChoiceType = toolChoiceResult.String()
+		}
+
+		switch toolChoiceType {
+		case "auto":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "AUTO")
+		case "none":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "NONE")
+		case "any":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "ANY")
+		case "tool":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "ANY")
+			if toolChoiceName != "" {
+				out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.allowedFunctionNames", []string{toolChoiceName})
+			}
+		}
+	}
+
 	// Map Anthropic thinking -> Gemini CLI thinkingConfig when enabled
 	// Translator only does format conversion, ApplyThinking handles model capability validation.
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {

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

@@ -0,0 +1,42 @@
+package claude
+
+import (
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestConvertClaudeRequestToCLI_ToolChoice_SpecificTool(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "gemini-3-flash-preview",
+		"messages": [
+			{
+				"role": "user",
+				"content": [
+					{"type": "text", "text": "hi"}
+				]
+			}
+		],
+		"tools": [
+			{
+				"name": "json",
+				"description": "A JSON tool",
+				"input_schema": {
+					"type": "object",
+					"properties": {}
+				}
+			}
+		],
+		"tool_choice": {"type": "tool", "name": "json"}
+	}`)
+
+	output := ConvertClaudeRequestToCLI("gemini-3-flash-preview", inputJSON, false)
+
+	if got := gjson.GetBytes(output, "request.toolConfig.functionCallingConfig.mode").String(); got != "ANY" {
+		t.Fatalf("Expected request.toolConfig.functionCallingConfig.mode 'ANY', got '%s'", got)
+	}
+	allowed := gjson.GetBytes(output, "request.toolConfig.functionCallingConfig.allowedFunctionNames").Array()
+	if len(allowed) != 1 || allowed[0].String() != "json" {
+		t.Fatalf("Expected allowedFunctionNames ['json'], got %s", gjson.GetBytes(output, "request.toolConfig.functionCallingConfig.allowedFunctionNames").Raw)
+	}
+}

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

@@ -14,6 +14,7 @@ import (
 	"sync/atomic"
 	"time"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -209,7 +210,7 @@ func ConvertGeminiCLIResponseToClaude(_ context.Context, _ string, originalReque
 
 				// Create the tool use block with unique ID and function details
 				data := fmt.Sprintf(`{"type":"content_block_start","index":%d,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`, (*param).(*Params).ResponseIndex)
-				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1)))
+				data, _ = sjson.Set(data, "content_block.id", util.SanitizeClaudeToolID(fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1))))
 				data, _ = sjson.Set(data, "content_block.name", fcName)
 				output = output + fmt.Sprintf("data: %s\n\n\n", data)
 

internal/translator/gemini/claude/gemini_claude_request.go

@@ -85,6 +85,11 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 
 					case "tool_use":
 						functionName := contentResult.Get("name").String()
+						if toolUseID := contentResult.Get("id").String(); toolUseID != "" {
+							if derived := toolNameFromClaudeToolUseID(toolUseID); derived != "" {
+								functionName = derived
+							}
+						}
 						functionArgs := contentResult.Get("input").String()
 						argsResult := gjson.Parse(functionArgs)
 						if argsResult.IsObject() && gjson.Valid(functionArgs) {
@@ -100,10 +105,9 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 						if toolCallID == "" {
 							return true
 						}
-						funcName := toolCallID
-						toolCallIDs := strings.Split(toolCallID, "-")
-						if len(toolCallIDs) > 1 {
-							funcName = strings.Join(toolCallIDs[0:len(toolCallIDs)-1], "-")
+						funcName := toolNameFromClaudeToolUseID(toolCallID)
+						if funcName == "" {
+							funcName = toolCallID
 						}
 						responseData := contentResult.Get("content").Raw
 						part := `{"functionResponse":{"name":"","response":{"result":""}}}`
@@ -137,6 +141,7 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 				tool, _ = sjson.Delete(tool, "input_examples")
 				tool, _ = sjson.Delete(tool, "type")
 				tool, _ = sjson.Delete(tool, "cache_control")
+				tool, _ = sjson.Delete(tool, "defer_loading")
 				if gjson.Valid(tool) && gjson.Parse(tool).IsObject() {
 					if !hasTools {
 						out, _ = sjson.SetRaw(out, "tools", `[{"functionDeclarations":[]}]`)
@@ -152,6 +157,33 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 		}
 	}
 
+	// tool_choice
+	toolChoiceResult := gjson.GetBytes(rawJSON, "tool_choice")
+	if toolChoiceResult.Exists() {
+		toolChoiceType := ""
+		toolChoiceName := ""
+		if toolChoiceResult.IsObject() {
+			toolChoiceType = toolChoiceResult.Get("type").String()
+			toolChoiceName = toolChoiceResult.Get("name").String()
+		} else if toolChoiceResult.Type == gjson.String {
+			toolChoiceType = toolChoiceResult.String()
+		}
+
+		switch toolChoiceType {
+		case "auto":
+			out, _ = sjson.Set(out, "toolConfig.functionCallingConfig.mode", "AUTO")
+		case "none":
+			out, _ = sjson.Set(out, "toolConfig.functionCallingConfig.mode", "NONE")
+		case "any":
+			out, _ = sjson.Set(out, "toolConfig.functionCallingConfig.mode", "ANY")
+		case "tool":
+			out, _ = sjson.Set(out, "toolConfig.functionCallingConfig.mode", "ANY")
+			if toolChoiceName != "" {
+				out, _ = sjson.Set(out, "toolConfig.functionCallingConfig.allowedFunctionNames", []string{toolChoiceName})
+			}
+		}
+	}
+
 	// Map Anthropic thinking -> Gemini thinking config when enabled
 	// Translator only does format conversion, ApplyThinking handles model capability validation.
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
@@ -202,3 +234,11 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 
 	return result
 }
+
+func toolNameFromClaudeToolUseID(toolUseID string) string {
+	parts := strings.Split(toolUseID, "-")
+	if len(parts) <= 1 {
+		return ""
+	}
+	return strings.Join(parts[0:len(parts)-1], "-")
+}

internal/translator/gemini/claude/gemini_claude_request_test.go

@@ -0,0 +1,42 @@
+package claude
+
+import (
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestConvertClaudeRequestToGemini_ToolChoice_SpecificTool(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "gemini-3-flash-preview",
+		"messages": [
+			{
+				"role": "user",
+				"content": [
+					{"type": "text", "text": "hi"}
+				]
+			}
+		],
+		"tools": [
+			{
+				"name": "json",
+				"description": "A JSON tool",
+				"input_schema": {
+					"type": "object",
+					"properties": {}
+				}
+			}
+		],
+		"tool_choice": {"type": "tool", "name": "json"}
+	}`)
+
+	output := ConvertClaudeRequestToGemini("gemini-3-flash-preview", inputJSON, false)
+
+	if got := gjson.GetBytes(output, "toolConfig.functionCallingConfig.mode").String(); got != "ANY" {
+		t.Fatalf("Expected toolConfig.functionCallingConfig.mode 'ANY', got '%s'", got)
+	}
+	allowed := gjson.GetBytes(output, "toolConfig.functionCallingConfig.allowedFunctionNames").Array()
+	if len(allowed) != 1 || allowed[0].String() != "json" {
+		t.Fatalf("Expected allowedFunctionNames ['json'], got %s", gjson.GetBytes(output, "toolConfig.functionCallingConfig.allowedFunctionNames").Raw)
+	}
+}

internal/translator/gemini/claude/gemini_claude_response.go

@@ -12,8 +12,8 @@ import (
 	"fmt"
 	"strings"
 	"sync/atomic"
-	"time"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -25,6 +25,8 @@ type Params struct {
 	ResponseType     int
 	ResponseIndex    int
 	HasContent       bool // Tracks whether any content (text, thinking, or tool use) has been output
+	ToolNameMap      map[string]string
+	SawToolCall      bool
 }
 
 // toolUseIDCounter provides a process-wide unique counter for tool use identifiers.
@@ -53,6 +55,8 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 			HasFirstResponse: false,
 			ResponseType:     0,
 			ResponseIndex:    0,
+			ToolNameMap:      util.ToolNameMapFromClaudeRequest(originalRequestRawJSON),
+			SawToolCall:      false,
 		}
 	}
 
@@ -66,8 +70,6 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 		return []string{}
 	}
 
-	// Track whether tools are being used in this response chunk
-	usedTool := false
 	output := ""
 
 	// Initialize the streaming session with a message_start event
@@ -175,12 +177,13 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 			} else if functionCallResult.Exists() {
 				// Handle function/tool calls from the AI model
 				// This processes tool usage requests and formats them for Claude API compatibility
-				usedTool = true
-				fcName := functionCallResult.Get("name").String()
+				(*param).(*Params).SawToolCall = true
+				upstreamToolName := functionCallResult.Get("name").String()
+				clientToolName := util.MapToolName((*param).(*Params).ToolNameMap, upstreamToolName)
 
 				// FIX: Handle streaming split/delta where name might be empty in subsequent chunks.
 				// If we are already in tool use mode and name is empty, treat as continuation (delta).
-				if (*param).(*Params).ResponseType == 3 && fcName == "" {
+				if (*param).(*Params).ResponseType == 3 && upstreamToolName == "" {
 					if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
 						output = output + "event: content_block_delta\n"
 						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"input_json_delta","partial_json":""}}`, (*param).(*Params).ResponseIndex), "delta.partial_json", fcArgsResult.Raw)
@@ -221,8 +224,8 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 
 				// Create the tool use block with unique ID and function details
 				data := fmt.Sprintf(`{"type":"content_block_start","index":%d,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`, (*param).(*Params).ResponseIndex)
-				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1)))
-				data, _ = sjson.Set(data, "content_block.name", fcName)
+				data, _ = sjson.Set(data, "content_block.id", util.SanitizeClaudeToolID(fmt.Sprintf("%s-%d", upstreamToolName, atomic.AddUint64(&toolUseIDCounter, 1))))
+				data, _ = sjson.Set(data, "content_block.name", clientToolName)
 				output = output + fmt.Sprintf("data: %s\n\n\n", data)
 
 				if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
@@ -249,7 +252,7 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 				output = output + `data: `
 
 				template := `{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
-				if usedTool {
+				if (*param).(*Params).SawToolCall {
 					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}}`
@@ -278,10 +281,10 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 // Returns:
 //   - string: A Claude-compatible JSON response.
 func ConvertGeminiResponseToClaudeNonStream(_ context.Context, _ string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, _ *any) string {
-	_ = originalRequestRawJSON
 	_ = requestRawJSON
 
 	root := gjson.ParseBytes(rawJSON)
+	toolNameMap := util.ToolNameMapFromClaudeRequest(originalRequestRawJSON)
 
 	out := `{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}`
 	out, _ = sjson.Set(out, "id", root.Get("responseId").String())
@@ -336,11 +339,12 @@ func ConvertGeminiResponseToClaudeNonStream(_ context.Context, _ string, origina
 				flushText()
 				hasToolCall = true
 
-				name := functionCall.Get("name").String()
+				upstreamToolName := functionCall.Get("name").String()
+				clientToolName := util.MapToolName(toolNameMap, upstreamToolName)
 				toolIDCounter++
 				toolBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
-				toolBlock, _ = sjson.Set(toolBlock, "id", fmt.Sprintf("tool_%d", toolIDCounter))
-				toolBlock, _ = sjson.Set(toolBlock, "name", name)
+				toolBlock, _ = sjson.Set(toolBlock, "id", util.SanitizeClaudeToolID(fmt.Sprintf("%s-%d", upstreamToolName, toolIDCounter)))
+				toolBlock, _ = sjson.Set(toolBlock, "name", clientToolName)
 				inputRaw := "{}"
 				if args := functionCall.Get("args"); args.Exists() && gjson.Valid(args.Raw) && args.IsObject() {
 					inputRaw = args.Raw

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

@@ -147,21 +147,21 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 			content := m.Get("content")
 
 			if (role == "system" || role == "developer") && len(arr) > 1 {
-				// system -> system_instruction as a user message style
+				// system -> systemInstruction as a user message style
 				if content.Type == gjson.String {
-					out, _ = sjson.SetBytes(out, "system_instruction.role", "user")
-					out, _ = sjson.SetBytes(out, fmt.Sprintf("system_instruction.parts.%d.text", systemPartIndex), content.String())
+					out, _ = sjson.SetBytes(out, "systemInstruction.role", "user")
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("systemInstruction.parts.%d.text", systemPartIndex), content.String())
 					systemPartIndex++
 				} else if content.IsObject() && content.Get("type").String() == "text" {
-					out, _ = sjson.SetBytes(out, "system_instruction.role", "user")
-					out, _ = sjson.SetBytes(out, fmt.Sprintf("system_instruction.parts.%d.text", systemPartIndex), content.Get("text").String())
+					out, _ = sjson.SetBytes(out, "systemInstruction.role", "user")
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("systemInstruction.parts.%d.text", systemPartIndex), content.Get("text").String())
 					systemPartIndex++
 				} else if content.IsArray() {
 					contents := content.Array()
 					if len(contents) > 0 {
-						out, _ = sjson.SetBytes(out, "system_instruction.role", "user")
+						out, _ = sjson.SetBytes(out, "systemInstruction.role", "user")
 						for j := 0; j < len(contents); j++ {
-							out, _ = sjson.SetBytes(out, fmt.Sprintf("system_instruction.parts.%d.text", systemPartIndex), contents[j].Get("text").String())
+							out, _ = sjson.SetBytes(out, fmt.Sprintf("systemInstruction.parts.%d.text", systemPartIndex), contents[j].Get("text").String())
 							systemPartIndex++
 						}
 					}

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

@@ -26,7 +26,7 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 	if instructions := root.Get("instructions"); instructions.Exists() {
 		systemInstr := `{"parts":[{"text":""}]}`
 		systemInstr, _ = sjson.Set(systemInstr, "parts.0.text", instructions.String())
-		out, _ = sjson.SetRaw(out, "system_instruction", systemInstr)
+		out, _ = sjson.SetRaw(out, "systemInstruction", systemInstr)
 	}
 
 	// Convert input messages to Gemini contents format
@@ -119,7 +119,7 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 				if strings.EqualFold(itemRole, "system") {
 					if contentArray := item.Get("content"); contentArray.Exists() {
 						systemInstr := ""
-						if systemInstructionResult := gjson.Get(out, "system_instruction"); systemInstructionResult.Exists() {
+						if systemInstructionResult := gjson.Get(out, "systemInstruction"); systemInstructionResult.Exists() {
 							systemInstr = systemInstructionResult.Raw
 						} else {
 							systemInstr = `{"parts":[]}`
@@ -140,7 +140,7 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 						}
 
 						if systemInstr != `{"parts":[]}` {
-							out, _ = sjson.SetRaw(out, "system_instruction", systemInstr)
+							out, _ = sjson.SetRaw(out, "systemInstruction", systemInstr)
 						}
 					}
 					continue
@@ -237,6 +237,33 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 									partJSON, _ = sjson.Set(partJSON, "inline_data.data", data)
 								}
 							}
+						case "input_audio":
+							audioData := contentItem.Get("data").String()
+							audioFormat := contentItem.Get("format").String()
+							if audioData != "" {
+								audioMimeMap := map[string]string{
+									"mp3":       "audio/mpeg",
+									"wav":       "audio/wav",
+									"ogg":       "audio/ogg",
+									"flac":      "audio/flac",
+									"aac":       "audio/aac",
+									"webm":      "audio/webm",
+									"pcm16":     "audio/pcm",
+									"g711_ulaw": "audio/basic",
+									"g711_alaw": "audio/basic",
+								}
+								mimeType := "audio/wav"
+								if audioFormat != "" {
+									if mapped, ok := audioMimeMap[audioFormat]; ok {
+										mimeType = mapped
+									} else {
+										mimeType = "audio/" + audioFormat
+									}
+								}
+								partJSON = `{"inline_data":{"mime_type":"","data":""}}`
+								partJSON, _ = sjson.Set(partJSON, "inline_data.mime_type", mimeType)
+								partJSON, _ = sjson.Set(partJSON, "inline_data.data", audioData)
+							}
 						}
 
 						if partJSON != "" {
@@ -354,22 +381,7 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 					funcDecl, _ = sjson.Set(funcDecl, "description", desc.String())
 				}
 				if params := tool.Get("parameters"); params.Exists() {
-					// Convert parameter types from OpenAI format to Gemini format
-					cleaned := params.Raw
-					// Convert type values to uppercase for Gemini
-					paramsResult := gjson.Parse(cleaned)
-					if properties := paramsResult.Get("properties"); properties.Exists() {
-						properties.ForEach(func(key, value gjson.Result) bool {
-							if propType := value.Get("type"); propType.Exists() {
-								upperType := strings.ToUpper(propType.String())
-								cleaned, _ = sjson.Set(cleaned, "properties."+key.String()+".type", upperType)
-							}
-							return true
-						})
-					}
-					// Set the overall type to OBJECT
-					cleaned, _ = sjson.Set(cleaned, "type", "OBJECT")
-					funcDecl, _ = sjson.SetRaw(funcDecl, "parametersJsonSchema", cleaned)
+					funcDecl, _ = sjson.SetRaw(funcDecl, "parametersJsonSchema", params.Raw)
 				}
 
 				geminiTools, _ = sjson.SetRaw(geminiTools, "0.functionDeclarations.-1", funcDecl)

internal/translator/kiro/claude/kiro_claude_request.go

@@ -605,10 +605,6 @@ func convertClaudeToolsToKiro(tools gjson.Result) []KiroToolWrapper {
 		})
 	}
 
-	// Apply dynamic compression if total tools size exceeds threshold
-	// This prevents 500 errors when Claude Code sends too many tools
-	kiroTools = compressToolsIfNeeded(kiroTools)
-
 	return kiroTools
 }
 
@@ -858,34 +854,7 @@ func BuildUserMessageStruct(msg gjson.Result, modelID, origin string) (KiroUserI
 
 				var textContents []KiroTextContent
 
-				// Check if this tool_result contains error from our SOFT_LIMIT_REACHED tool_use
-				// The client will return an error when trying to execute a tool with marker input
-				resultStr := resultContent.String()
-				isSoftLimitError := strings.Contains(resultStr, "SOFT_LIMIT_REACHED") ||
-					strings.Contains(resultStr, "_status") ||
-					strings.Contains(resultStr, "truncated") ||
-					strings.Contains(resultStr, "missing required") ||
-					strings.Contains(resultStr, "invalid input") ||
-					strings.Contains(resultStr, "Error writing file")
-
-				if isError && isSoftLimitError {
-					// Replace error content with SOFT_LIMIT_REACHED guidance
-					log.Infof("kiro: detected SOFT_LIMIT_REACHED in tool_result for %s, replacing with guidance", toolUseID)
-					softLimitMsg := `SOFT_LIMIT_REACHED
-
-Your previous tool call was incomplete due to API output size limits.
-The content was PARTIALLY transmitted but NOT executed.
-
-REQUIRED ACTION:
-1. Split your content into smaller chunks (max 300 lines per call)
-2. For file writes: Create file with first chunk, then use append for remaining
-3. Do NOT regenerate content you already attempted - continue from where you stopped
-
-STATUS: This is NOT an error. Continue with smaller chunks.`
-					textContents = append(textContents, KiroTextContent{Text: softLimitMsg})
-					// Mark as SUCCESS so Claude doesn't treat it as a failure
-					isError = false
-				} else if resultContent.IsArray() {
+				if resultContent.IsArray() {
 					for _, item := range resultContent.Array() {
 						if item.Get("type").String() == "text" {
 							textContents = append(textContents, KiroTextContent{Text: item.Get("text").String()})

internal/translator/kiro/claude/kiro_claude_response.go

@@ -55,39 +55,18 @@ func BuildClaudeResponse(content string, toolUses []KiroToolUse, model string, u
 		}
 	}
 
-	// Add tool_use blocks - emit truncated tools with SOFT_LIMIT_REACHED marker
-	hasTruncatedTools := false
+	// Add tool_use blocks - skip truncated tools and log warning
 	for _, toolUse := range toolUses {
 		if toolUse.IsTruncated && toolUse.TruncationInfo != nil {
-			// Emit tool_use with SOFT_LIMIT_REACHED marker input
-			hasTruncatedTools = true
-			log.Infof("kiro: buildClaudeResponse emitting truncated tool with SOFT_LIMIT_REACHED: %s (ID: %s)", toolUse.Name, toolUse.ToolUseID)
-
-			markerInput := map[string]interface{}{
-				"_status":  "SOFT_LIMIT_REACHED",
-				"_message": "Tool output was truncated. Split content into smaller chunks (max 300 lines). Due to potential model hallucination, you MUST re-fetch the current working directory and generate the correct file_path.",
-			}
-
-			contentBlocks = append(contentBlocks, map[string]interface{}{
-				"type":  "tool_use",
-				"id":    toolUse.ToolUseID,
-				"name":  toolUse.Name,
-				"input": markerInput,
-			})
-		} else {
-			// Normal tool use
-			contentBlocks = append(contentBlocks, map[string]interface{}{
-				"type":  "tool_use",
-				"id":    toolUse.ToolUseID,
-				"name":  toolUse.Name,
-				"input": toolUse.Input,
-			})
+			log.Warnf("kiro: buildClaudeResponse skipping truncated tool: %s (ID: %s)", toolUse.Name, toolUse.ToolUseID)
+			continue
 		}
-	}
-
-	// Log if we used SOFT_LIMIT_REACHED
-	if hasTruncatedTools {
-		log.Infof("kiro: buildClaudeResponse using SOFT_LIMIT_REACHED - keeping stop_reason=tool_use")
+		contentBlocks = append(contentBlocks, map[string]interface{}{
+			"type":  "tool_use",
+			"id":    toolUse.ToolUseID,
+			"name":  toolUse.Name,
+			"input": toolUse.Input,
+		})
 	}
 
 	// Ensure at least one content block (Claude API requires non-empty content)

internal/translator/kiro/claude/kiro_claude_stream_parser.go

@@ -192,8 +192,8 @@ func AnalyzeBufferedStream(chunks [][]byte) BufferedStreamResult {
 							if idx, ok := event["index"].(float64); ok {
 								currentToolIndex = int(idx)
 							}
-							// Capture tool use ID for toolResults handshake
-							if id, ok := cb["id"].(string); ok {
+							// Capture tool use ID only for web_search toolResults handshake
+							if id, ok := cb["id"].(string); ok && (currentToolName == "web_search" || currentToolName == "remote_web_search") {
 								result.WebSearchToolUseId = id
 							}
 							toolInputBuilder.Reset()

internal/translator/kiro/claude/tool_compression.go

@@ -1,191 +0,0 @@
-// Package claude provides tool compression functionality for Kiro translator.
-// This file implements dynamic tool compression to reduce tool payload size
-// when it exceeds the target threshold, preventing 500 errors from Kiro API.
-package claude
-
-import (
-	"encoding/json"
-	"unicode/utf8"
-
-	kirocommon "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/common"
-	log "github.com/sirupsen/logrus"
-)
-
-// calculateToolsSize calculates the JSON serialized size of the tools list.
-// Returns the size in bytes.
-func calculateToolsSize(tools []KiroToolWrapper) int {
-	if len(tools) == 0 {
-		return 0
-	}
-	data, err := json.Marshal(tools)
-	if err != nil {
-		log.Warnf("kiro: failed to marshal tools for size calculation: %v", err)
-		return 0
-	}
-	return len(data)
-}
-
-// simplifyInputSchema simplifies the input_schema by keeping only essential fields:
-// type, enum, required. Recursively processes nested properties.
-func simplifyInputSchema(schema interface{}) interface{} {
-	if schema == nil {
-		return nil
-	}
-
-	schemaMap, ok := schema.(map[string]interface{})
-	if !ok {
-		return schema
-	}
-
-	simplified := make(map[string]interface{})
-
-	// Keep essential fields
-	if t, ok := schemaMap["type"]; ok {
-		simplified["type"] = t
-	}
-	if enum, ok := schemaMap["enum"]; ok {
-		simplified["enum"] = enum
-	}
-	if required, ok := schemaMap["required"]; ok {
-		simplified["required"] = required
-	}
-
-	// Recursively process properties
-	if properties, ok := schemaMap["properties"].(map[string]interface{}); ok {
-		simplifiedProps := make(map[string]interface{})
-		for key, value := range properties {
-			simplifiedProps[key] = simplifyInputSchema(value)
-		}
-		simplified["properties"] = simplifiedProps
-	}
-
-	// Process items for array types
-	if items, ok := schemaMap["items"]; ok {
-		simplified["items"] = simplifyInputSchema(items)
-	}
-
-	// Process additionalProperties if present
-	if additionalProps, ok := schemaMap["additionalProperties"]; ok {
-		simplified["additionalProperties"] = simplifyInputSchema(additionalProps)
-	}
-
-	// Process anyOf, oneOf, allOf
-	for _, key := range []string{"anyOf", "oneOf", "allOf"} {
-		if arr, ok := schemaMap[key].([]interface{}); ok {
-			simplifiedArr := make([]interface{}, len(arr))
-			for i, item := range arr {
-				simplifiedArr[i] = simplifyInputSchema(item)
-			}
-			simplified[key] = simplifiedArr
-		}
-	}
-
-	return simplified
-}
-
-// compressToolDescription compresses a description to the target length.
-// Ensures the result is at least MinToolDescriptionLength characters.
-// Uses UTF-8 safe truncation.
-func compressToolDescription(description string, targetLength int) string {
-	if targetLength < kirocommon.MinToolDescriptionLength {
-		targetLength = kirocommon.MinToolDescriptionLength
-	}
-
-	if len(description) <= targetLength {
-		return description
-	}
-
-	// Find a safe truncation point (UTF-8 boundary)
-	truncLen := targetLength - 3 // Leave room for "..."
-
-	// Ensure we don't cut in the middle of a UTF-8 character
-	for truncLen > 0 && !utf8.RuneStart(description[truncLen]) {
-		truncLen--
-	}
-
-	if truncLen <= 0 {
-		return description[:kirocommon.MinToolDescriptionLength]
-	}
-
-	return description[:truncLen] + "..."
-}
-
-// compressToolsIfNeeded compresses tools if their total size exceeds the target threshold.
-// Compression strategy:
-// 1. First, check if compression is needed (size > ToolCompressionTargetSize)
-// 2. Step 1: Simplify input_schema (keep only type/enum/required)
-// 3. Step 2: Proportionally compress descriptions (minimum MinToolDescriptionLength chars)
-// Returns the compressed tools list.
-func compressToolsIfNeeded(tools []KiroToolWrapper) []KiroToolWrapper {
-	if len(tools) == 0 {
-		return tools
-	}
-
-	originalSize := calculateToolsSize(tools)
-	if originalSize <= kirocommon.ToolCompressionTargetSize {
-		log.Debugf("kiro: tools size %d bytes is within target %d bytes, no compression needed",
-			originalSize, kirocommon.ToolCompressionTargetSize)
-		return tools
-	}
-
-	log.Infof("kiro: tools size %d bytes exceeds target %d bytes, starting compression",
-		originalSize, kirocommon.ToolCompressionTargetSize)
-
-	// Create a copy of tools to avoid modifying the original
-	compressedTools := make([]KiroToolWrapper, len(tools))
-	for i, tool := range tools {
-		compressedTools[i] = KiroToolWrapper{
-			ToolSpecification: KiroToolSpecification{
-				Name:        tool.ToolSpecification.Name,
-				Description: tool.ToolSpecification.Description,
-				InputSchema: KiroInputSchema{JSON: tool.ToolSpecification.InputSchema.JSON},
-			},
-		}
-	}
-
-	// Step 1: Simplify input_schema
-	for i := range compressedTools {
-		compressedTools[i].ToolSpecification.InputSchema.JSON =
-			simplifyInputSchema(compressedTools[i].ToolSpecification.InputSchema.JSON)
-	}
-
-	sizeAfterSchemaSimplification := calculateToolsSize(compressedTools)
-	log.Debugf("kiro: size after schema simplification: %d bytes (reduced by %d bytes)",
-		sizeAfterSchemaSimplification, originalSize-sizeAfterSchemaSimplification)
-
-	// Check if we're within target after schema simplification
-	if sizeAfterSchemaSimplification <= kirocommon.ToolCompressionTargetSize {
-		log.Infof("kiro: compression complete after schema simplification, final size: %d bytes",
-			sizeAfterSchemaSimplification)
-		return compressedTools
-	}
-
-	// Step 2: Compress descriptions proportionally
-	sizeToReduce := float64(sizeAfterSchemaSimplification - kirocommon.ToolCompressionTargetSize)
-	var totalDescLen float64
-	for _, tool := range compressedTools {
-		totalDescLen += float64(len(tool.ToolSpecification.Description))
-	}
-
-	if totalDescLen > 0 {
-		// Assume size reduction comes primarily from descriptions.
-		keepRatio := 1.0 - (sizeToReduce / totalDescLen)
-		if keepRatio > 1.0 {
-			keepRatio = 1.0
-		} else if keepRatio < 0 {
-			keepRatio = 0
-		}
-
-		for i := range compressedTools {
-			desc := compressedTools[i].ToolSpecification.Description
-			targetLen := int(float64(len(desc)) * keepRatio)
-			compressedTools[i].ToolSpecification.Description = compressToolDescription(desc, targetLen)
-		}
-	}
-
-	finalSize := calculateToolsSize(compressedTools)
-	log.Infof("kiro: compression complete, original: %d bytes, final: %d bytes (%.1f%% reduction)",
-		originalSize, finalSize, float64(originalSize-finalSize)/float64(originalSize)*100)
-
-	return compressedTools
-}

internal/translator/kiro/claude/truncation_detector.go

@@ -84,13 +84,18 @@ func DetectTruncation(toolName, toolUseID, rawInput string, parsedInput map[stri
 		ParsedFields: make(map[string]string),
 	}
 
-	// Scenario 1: Empty input buffer - no data received at all
+	// Scenario 1: Empty input buffer - only flag as truncation if tool has required fields
+	// Many tools (e.g. TaskList, TaskGet) have no required params, so empty input is valid
 	if strings.TrimSpace(rawInput) == "" {
-		info.IsTruncated = true
-		info.TruncationType = TruncationTypeEmptyInput
-		info.ErrorMessage = "Tool input was completely empty - API response may have been truncated before tool parameters were transmitted"
-		log.Warnf("kiro: truncation detected [%s] for tool %s (ID: %s): empty input buffer",
-			info.TruncationType, toolName, toolUseID)
+		if _, hasRequirements := RequiredFieldsByTool[toolName]; hasRequirements {
+			info.IsTruncated = true
+			info.TruncationType = TruncationTypeEmptyInput
+			info.ErrorMessage = "Tool input was completely empty - API response may have been truncated before tool parameters were transmitted"
+			log.Warnf("kiro: truncation detected [%s] for tool %s (ID: %s): empty input buffer",
+				info.TruncationType, toolName, toolUseID)
+			return info
+		}
+		log.Debugf("kiro: empty input for tool %s (ID: %s) - no required fields, treating as valid", toolName, toolUseID)
 		return info
 	}
 
@@ -342,7 +347,7 @@ func buildTruncationErrorMessage(toolName, truncationType string, parsedFields m
 	}
 
 	sb.WriteString(" Received ")
-	sb.WriteString(string(rune(len(rawInput))))
+	sb.WriteString(formatInt(len(rawInput)))
 	sb.WriteString(" bytes. Please retry with smaller content chunks.")
 
 	return sb.String()

internal/translator/kiro/common/constants.go

@@ -6,14 +6,6 @@ const (
 	// Kiro API limit is 10240 bytes, leave room for "..."
 	KiroMaxToolDescLen = 10237
 
-	// ToolCompressionTargetSize is the target total size for compressed tools (20KB).
-	// If tools exceed this size, compression will be applied.
-	ToolCompressionTargetSize = 20 * 1024 // 20KB
-
-	// MinToolDescriptionLength is the minimum description length after compression.
-	// Descriptions will not be shortened below this length.
-	MinToolDescriptionLength = 50
-
 	// ThinkingStartTag is the start tag for thinking blocks in responses.
 	ThinkingStartTag = "<thinking>"
 

internal/translator/openai/claude/openai_claude_request.go

@@ -183,7 +183,12 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 						// Collect tool_result to emit after the main message (ensures tool results follow tool_calls)
 						toolResultJSON := `{"role":"tool","tool_call_id":"","content":""}`
 						toolResultJSON, _ = sjson.Set(toolResultJSON, "tool_call_id", part.Get("tool_use_id").String())
-						toolResultJSON, _ = sjson.Set(toolResultJSON, "content", convertClaudeToolResultContentToString(part.Get("content")))
+						toolResultContent, toolResultContentRaw := convertClaudeToolResultContent(part.Get("content"))
+						if toolResultContentRaw {
+							toolResultJSON, _ = sjson.SetRaw(toolResultJSON, "content", toolResultContent)
+						} else {
+							toolResultJSON, _ = sjson.Set(toolResultJSON, "content", toolResultContent)
+						}
 						toolResults = append(toolResults, toolResultJSON)
 					}
 					return true
@@ -374,21 +379,41 @@ func convertClaudeContentPart(part gjson.Result) (string, bool) {
 	}
 }
 
-func convertClaudeToolResultContentToString(content gjson.Result) string {
+func convertClaudeToolResultContent(content gjson.Result) (string, bool) {
 	if !content.Exists() {
-		return ""
+		return "", false
 	}
 
 	if content.Type == gjson.String {
-		return content.String()
+		return content.String(), false
 	}
 
 	if content.IsArray() {
 		var parts []string
+		contentJSON := "[]"
+		hasImagePart := false
 		content.ForEach(func(_, item gjson.Result) bool {
 			switch {
 			case item.Type == gjson.String:
-				parts = append(parts, item.String())
+				text := item.String()
+				parts = append(parts, text)
+				textContent := `{"type":"text","text":""}`
+				textContent, _ = sjson.Set(textContent, "text", text)
+				contentJSON, _ = sjson.SetRaw(contentJSON, "-1", textContent)
+			case item.IsObject() && item.Get("type").String() == "text":
+				text := item.Get("text").String()
+				parts = append(parts, text)
+				textContent := `{"type":"text","text":""}`
+				textContent, _ = sjson.Set(textContent, "text", text)
+				contentJSON, _ = sjson.SetRaw(contentJSON, "-1", textContent)
+			case item.IsObject() && item.Get("type").String() == "image":
+				contentItem, ok := convertClaudeContentPart(item)
+				if ok {
+					contentJSON, _ = sjson.SetRaw(contentJSON, "-1", contentItem)
+					hasImagePart = true
+				} else {
+					parts = append(parts, item.Raw)
+				}
 			case item.IsObject() && item.Get("text").Exists() && item.Get("text").Type == gjson.String:
 				parts = append(parts, item.Get("text").String())
 			default:
@@ -397,19 +422,31 @@ func convertClaudeToolResultContentToString(content gjson.Result) string {
 			return true
 		})
 
+		if hasImagePart {
+			return contentJSON, true
+		}
+
 		joined := strings.Join(parts, "\n\n")
 		if strings.TrimSpace(joined) != "" {
-			return joined
+			return joined, false
 		}
-		return content.Raw
+		return content.Raw, false
 	}
 
 	if content.IsObject() {
-		if text := content.Get("text"); text.Exists() && text.Type == gjson.String {
-			return text.String()
+		if content.Get("type").String() == "image" {
+			contentItem, ok := convertClaudeContentPart(content)
+			if ok {
+				contentJSON := "[]"
+				contentJSON, _ = sjson.SetRaw(contentJSON, "-1", contentItem)
+				return contentJSON, true
+			}
 		}
-		return content.Raw
+		if text := content.Get("text"); text.Exists() && text.Type == gjson.String {
+			return text.String(), false
+		}
+		return content.Raw, false
 	}
 
-	return content.Raw
+	return content.Raw, false
 }

internal/translator/openai/claude/openai_claude_request_test.go

@@ -488,6 +488,114 @@ func TestConvertClaudeRequestToOpenAI_ToolResultObjectContent(t *testing.T) {
 	}
 }
 
+func TestConvertClaudeRequestToOpenAI_ToolResultTextAndImageContent(t *testing.T) {
+	inputJSON := `{
+		"model": "claude-3-opus",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{"type": "tool_use", "id": "call_1", "name": "do_work", "input": {"a": 1}}
+				]
+			},
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "tool_result",
+						"tool_use_id": "call_1",
+						"content": [
+							{"type": "text", "text": "tool ok"},
+							{
+								"type": "image",
+								"source": {
+									"type": "base64",
+									"media_type": "image/png",
+									"data": "iVBORw0KGgoAAAANSUhEUg=="
+								}
+							}
+						]
+					}
+				]
+			}
+		]
+	}`
+
+	result := ConvertClaudeRequestToOpenAI("test-model", []byte(inputJSON), false)
+	resultJSON := gjson.ParseBytes(result)
+	messages := resultJSON.Get("messages").Array()
+
+	if len(messages) != 2 {
+		t.Fatalf("Expected 2 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	}
+
+	toolContent := messages[1].Get("content")
+	if !toolContent.IsArray() {
+		t.Fatalf("Expected tool content array, got %s", toolContent.Raw)
+	}
+	if got := toolContent.Get("0.type").String(); got != "text" {
+		t.Fatalf("Expected first tool content type %q, got %q", "text", got)
+	}
+	if got := toolContent.Get("0.text").String(); got != "tool ok" {
+		t.Fatalf("Expected first tool content text %q, got %q", "tool ok", got)
+	}
+	if got := toolContent.Get("1.type").String(); got != "image_url" {
+		t.Fatalf("Expected second tool content type %q, got %q", "image_url", got)
+	}
+	if got := toolContent.Get("1.image_url.url").String(); got != "data:image/png;base64,iVBORw0KGgoAAAANSUhEUg==" {
+		t.Fatalf("Unexpected image_url: %q", got)
+	}
+}
+
+func TestConvertClaudeRequestToOpenAI_ToolResultURLImageOnly(t *testing.T) {
+	inputJSON := `{
+		"model": "claude-3-opus",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{"type": "tool_use", "id": "call_1", "name": "do_work", "input": {"a": 1}}
+				]
+			},
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "tool_result",
+						"tool_use_id": "call_1",
+						"content": {
+							"type": "image",
+							"source": {
+								"type": "url",
+								"url": "https://example.com/tool.png"
+							}
+						}
+					}
+				]
+			}
+		]
+	}`
+
+	result := ConvertClaudeRequestToOpenAI("test-model", []byte(inputJSON), false)
+	resultJSON := gjson.ParseBytes(result)
+	messages := resultJSON.Get("messages").Array()
+
+	if len(messages) != 2 {
+		t.Fatalf("Expected 2 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	}
+
+	toolContent := messages[1].Get("content")
+	if !toolContent.IsArray() {
+		t.Fatalf("Expected tool content array, got %s", toolContent.Raw)
+	}
+	if got := toolContent.Get("0.type").String(); got != "image_url" {
+		t.Fatalf("Expected tool content type %q, got %q", "image_url", got)
+	}
+	if got := toolContent.Get("0.image_url.url").String(); got != "https://example.com/tool.png" {
+		t.Fatalf("Unexpected image_url: %q", got)
+	}
+}
+
 func TestConvertClaudeRequestToOpenAI_AssistantTextToolUseTextOrder(t *testing.T) {
 	inputJSON := `{
 		"model": "claude-3-opus",

internal/translator/openai/claude/openai_claude_response.go

@@ -22,9 +22,11 @@ var (
 
 // ConvertOpenAIResponseToAnthropicParams holds parameters for response conversion
 type ConvertOpenAIResponseToAnthropicParams struct {
-	MessageID string
-	Model     string
-	CreatedAt int64
+	MessageID   string
+	Model       string
+	CreatedAt   int64
+	ToolNameMap map[string]string
+	SawToolCall bool
 	// Content accumulator for streaming
 	ContentAccumulator strings.Builder
 	// Tool calls accumulator for streaming
@@ -78,6 +80,8 @@ func ConvertOpenAIResponseToClaude(_ context.Context, _ string, originalRequestR
 			MessageID:                   "",
 			Model:                       "",
 			CreatedAt:                   0,
+			ToolNameMap:                 nil,
+			SawToolCall:                 false,
 			ContentAccumulator:          strings.Builder{},
 			ToolCallsAccumulator:        nil,
 			TextContentBlockStarted:     false,
@@ -97,6 +101,10 @@ func ConvertOpenAIResponseToClaude(_ context.Context, _ string, originalRequestR
 	}
 	rawJSON = bytes.TrimSpace(rawJSON[5:])
 
+	if (*param).(*ConvertOpenAIResponseToAnthropicParams).ToolNameMap == nil {
+		(*param).(*ConvertOpenAIResponseToAnthropicParams).ToolNameMap = util.ToolNameMapFromClaudeRequest(originalRequestRawJSON)
+	}
+
 	// Check if this is the [DONE] marker
 	rawStr := strings.TrimSpace(string(rawJSON))
 	if rawStr == "[DONE]" {
@@ -111,6 +119,16 @@ func ConvertOpenAIResponseToClaude(_ context.Context, _ string, originalRequestR
 	}
 }
 
+func effectiveOpenAIFinishReason(param *ConvertOpenAIResponseToAnthropicParams) string {
+	if param == nil {
+		return ""
+	}
+	if param.SawToolCall {
+		return "tool_calls"
+	}
+	return param.FinishReason
+}
+
 // convertOpenAIStreamingChunkToAnthropic converts OpenAI streaming chunk to Anthropic streaming events
 func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAIResponseToAnthropicParams) []string {
 	root := gjson.ParseBytes(rawJSON)
@@ -197,6 +215,7 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 			}
 
 			toolCalls.ForEach(func(_, toolCall gjson.Result) bool {
+				param.SawToolCall = true
 				index := int(toolCall.Get("index").Int())
 				blockIndex := param.toolContentBlockIndex(index)
 
@@ -215,7 +234,7 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 				// Handle function name
 				if function := toolCall.Get("function"); function.Exists() {
 					if name := function.Get("name"); name.Exists() {
-						accumulator.Name = name.String()
+						accumulator.Name = util.MapToolName(param.ToolNameMap, name.String())
 
 						stopThinkingContentBlock(param, &results)
 
@@ -224,7 +243,7 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 						// Send content_block_start for tool_use
 						contentBlockStartJSON := `{"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`
 						contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "index", blockIndex)
-						contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "content_block.id", accumulator.ID)
+						contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "content_block.id", util.SanitizeClaudeToolID(accumulator.ID))
 						contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "content_block.name", accumulator.Name)
 						results = append(results, "event: content_block_start\ndata: "+contentBlockStartJSON+"\n\n")
 					}
@@ -246,7 +265,11 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 	// Handle finish_reason (but don't send message_delta/message_stop yet)
 	if finishReason := root.Get("choices.0.finish_reason"); finishReason.Exists() && finishReason.String() != "" {
 		reason := finishReason.String()
-		param.FinishReason = reason
+		if param.SawToolCall {
+			param.FinishReason = "tool_calls"
+		} else {
+			param.FinishReason = reason
+		}
 
 		// Send content_block_stop for thinking content if needed
 		if param.ThinkingContentBlockStarted {
@@ -294,7 +317,7 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 			inputTokens, outputTokens, cachedTokens = extractOpenAIUsage(usage)
 			// Send message_delta with usage
 			messageDeltaJSON := `{"type":"message_delta","delta":{"stop_reason":"","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
-			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(param.FinishReason))
+			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(effectiveOpenAIFinishReason(param)))
 			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.input_tokens", inputTokens)
 			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.output_tokens", outputTokens)
 			if cachedTokens > 0 {
@@ -348,7 +371,7 @@ 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},"usage":{"input_tokens":0,"output_tokens":0}}`
-		messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(param.FinishReason))
+		messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(effectiveOpenAIFinishReason(param)))
 		results = append(results, "event: message_delta\ndata: "+messageDeltaJSON+"\n\n")
 		param.MessageDeltaSent = true
 	}
@@ -391,7 +414,7 @@ func convertOpenAINonStreamingToAnthropic(rawJSON []byte) []string {
 		if toolCalls := choice.Get("message.tool_calls"); toolCalls.Exists() && toolCalls.IsArray() {
 			toolCalls.ForEach(func(_, toolCall gjson.Result) bool {
 				toolUseBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
-				toolUseBlock, _ = sjson.Set(toolUseBlock, "id", toolCall.Get("id").String())
+				toolUseBlock, _ = sjson.Set(toolUseBlock, "id", util.SanitizeClaudeToolID(toolCall.Get("id").String()))
 				toolUseBlock, _ = sjson.Set(toolUseBlock, "name", toolCall.Get("function.name").String())
 
 				argsStr := util.FixJSON(toolCall.Get("function.arguments").String())
@@ -531,10 +554,10 @@ func stopTextContentBlock(param *ConvertOpenAIResponseToAnthropicParams, results
 // Returns:
 //   - string: An Anthropic-compatible JSON response.
 func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, _ *any) string {
-	_ = originalRequestRawJSON
 	_ = requestRawJSON
 
 	root := gjson.ParseBytes(rawJSON)
+	toolNameMap := util.ToolNameMapFromClaudeRequest(originalRequestRawJSON)
 	out := `{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}`
 	out, _ = sjson.Set(out, "id", root.Get("id").String())
 	out, _ = sjson.Set(out, "model", root.Get("model").String())
@@ -589,8 +612,8 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 								toolCalls.ForEach(func(_, tc gjson.Result) bool {
 									hasToolCall = true
 									toolUse := `{"type":"tool_use","id":"","name":"","input":{}}`
-									toolUse, _ = sjson.Set(toolUse, "id", tc.Get("id").String())
-									toolUse, _ = sjson.Set(toolUse, "name", tc.Get("function.name").String())
+									toolUse, _ = sjson.Set(toolUse, "id", util.SanitizeClaudeToolID(tc.Get("id").String()))
+									toolUse, _ = sjson.Set(toolUse, "name", util.MapToolName(toolNameMap, tc.Get("function.name").String()))
 
 									argsStr := util.FixJSON(tc.Get("function.arguments").String())
 									if argsStr != "" && gjson.Valid(argsStr) {
@@ -646,8 +669,8 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 				toolCalls.ForEach(func(_, toolCall gjson.Result) bool {
 					hasToolCall = true
 					toolUseBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
-					toolUseBlock, _ = sjson.Set(toolUseBlock, "id", toolCall.Get("id").String())
-					toolUseBlock, _ = sjson.Set(toolUseBlock, "name", toolCall.Get("function.name").String())
+					toolUseBlock, _ = sjson.Set(toolUseBlock, "id", util.SanitizeClaudeToolID(toolCall.Get("id").String()))
+					toolUseBlock, _ = sjson.Set(toolUseBlock, "name", util.MapToolName(toolNameMap, toolCall.Get("function.name").String()))
 
 					argsStr := util.FixJSON(toolCall.Get("function.arguments").String())
 					if argsStr != "" && gjson.Valid(argsStr) {

internal/util/claude_tool_id.go

@@ -0,0 +1,24 @@
+package util
+
+import (
+	"fmt"
+	"regexp"
+	"sync/atomic"
+	"time"
+)
+
+var (
+	claudeToolUseIDSanitizer = regexp.MustCompile(`[^a-zA-Z0-9_-]`)
+	claudeToolUseIDCounter   uint64
+)
+
+// SanitizeClaudeToolID ensures the given id conforms to Claude's
+// tool_use.id regex ^[a-zA-Z0-9_-]+$.  Non-conforming characters are
+// replaced with '_'; an empty result gets a generated fallback.
+func SanitizeClaudeToolID(id string) string {
+	s := claudeToolUseIDSanitizer.ReplaceAllString(id, "_")
+	if s == "" {
+		s = fmt.Sprintf("toolu_%d_%d", time.Now().UnixNano(), atomic.AddUint64(&claudeToolUseIDCounter, 1))
+	}
+	return s
+}

internal/util/gemini_schema.go

@@ -430,7 +430,7 @@ func removeUnsupportedKeywords(jsonStr string) string {
 	keywords := append(unsupportedConstraints,
 		"$schema", "$defs", "definitions", "const", "$ref", "$id", "additionalProperties",
 		"propertyNames", "patternProperties", // Gemini doesn't support these schema keywords
-		"enumTitles", "prefill", // Claude/OpenCode schema metadata fields unsupported by Gemini
+		"enumTitles", "prefill", "deprecated", // Schema metadata fields unsupported by Gemini
 	)
 
 	deletePaths := make([]string, 0)

internal/util/translator.go

@@ -6,6 +6,7 @@ package util
 import (
 	"bytes"
 	"fmt"
+	"strings"
 
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -219,3 +220,54 @@ func FixJSON(input string) string {
 
 	return out.String()
 }
+
+func CanonicalToolName(name string) string {
+	canonical := strings.TrimSpace(name)
+	canonical = strings.TrimLeft(canonical, "_")
+	return strings.ToLower(canonical)
+}
+
+// ToolNameMapFromClaudeRequest returns a canonical-name -> original-name map extracted from a Claude request.
+// It is used to restore exact tool name casing for clients that require strict tool name matching (e.g. Claude Code).
+func ToolNameMapFromClaudeRequest(rawJSON []byte) map[string]string {
+	if len(rawJSON) == 0 || !gjson.ValidBytes(rawJSON) {
+		return nil
+	}
+
+	tools := gjson.GetBytes(rawJSON, "tools")
+	if !tools.Exists() || !tools.IsArray() {
+		return nil
+	}
+
+	toolResults := tools.Array()
+	out := make(map[string]string, len(toolResults))
+	tools.ForEach(func(_, tool gjson.Result) bool {
+		name := strings.TrimSpace(tool.Get("name").String())
+		if name == "" {
+			return true
+		}
+		key := CanonicalToolName(name)
+		if key == "" {
+			return true
+		}
+		if _, exists := out[key]; !exists {
+			out[key] = name
+		}
+		return true
+	})
+
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
+func MapToolName(toolNameMap map[string]string, name string) string {
+	if name == "" || toolNameMap == nil {
+		return name
+	}
+	if mapped, ok := toolNameMap[CanonicalToolName(name)]; ok && mapped != "" {
+		return mapped
+	}
+	return name
+}

internal/watcher/clients.go

@@ -17,6 +17,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/watcher/diff"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/watcher/synthesizer"
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	log "github.com/sirupsen/logrus"
 )
@@ -75,6 +76,7 @@ func (w *Watcher) reloadClients(rescanAuth bool, affectedOAuthProviders []string
 
 		w.lastAuthHashes = make(map[string]string)
 		w.lastAuthContents = make(map[string]*coreauth.Auth)
+		w.fileAuthsByPath = make(map[string]map[string]*coreauth.Auth)
 		if resolvedAuthDir, errResolveAuthDir := util.ResolveAuthDir(cfg.AuthDir); errResolveAuthDir != nil {
 			log.Errorf("failed to resolve auth directory for hash cache: %v", errResolveAuthDir)
 		} else if resolvedAuthDir != "" {
@@ -92,6 +94,17 @@ func (w *Watcher) reloadClients(rescanAuth bool, affectedOAuthProviders []string
 						if errParse := json.Unmarshal(data, &auth); errParse == nil {
 							w.lastAuthContents[normalizedPath] = &auth
 						}
+						ctx := &synthesizer.SynthesisContext{
+							Config:      cfg,
+							AuthDir:     resolvedAuthDir,
+							Now:         time.Now(),
+							IDGenerator: synthesizer.NewStableIDGenerator(),
+						}
+						if generated := synthesizer.SynthesizeAuthFile(ctx, path, data); len(generated) > 0 {
+							if pathAuths := authSliceToMap(generated); len(pathAuths) > 0 {
+								w.fileAuthsByPath[normalizedPath] = pathAuths
+							}
+						}
 					}
 				}
 				return nil
@@ -143,13 +156,14 @@ func (w *Watcher) addOrUpdateClient(path string) {
 	}
 
 	w.clientsMutex.Lock()
-
-	cfg := w.config
-	if cfg == nil {
+	if w.config == nil {
 		log.Error("config is nil, cannot add or update client")
 		w.clientsMutex.Unlock()
 		return
 	}
+	if w.fileAuthsByPath == nil {
+		w.fileAuthsByPath = make(map[string]map[string]*coreauth.Auth)
+	}
 	if prev, ok := w.lastAuthHashes[normalized]; ok && prev == curHash {
 		log.Debugf("auth file unchanged (hash match), skipping reload: %s", filepath.Base(path))
 		w.clientsMutex.Unlock()
@@ -177,34 +191,86 @@ func (w *Watcher) addOrUpdateClient(path string) {
 	}
 	w.lastAuthContents[normalized] = &newAuth
 
-	w.clientsMutex.Unlock() // Unlock before the callback
-
-	w.refreshAuthState(false)
-
-	if w.reloadCallback != nil {
-		log.Debugf("triggering server update callback after add/update")
-		w.reloadCallback(cfg)
+	oldByID := make(map[string]*coreauth.Auth, len(w.fileAuthsByPath[normalized]))
+	for id, a := range w.fileAuthsByPath[normalized] {
+		oldByID[id] = a
 	}
+
+	// Build synthesized auth entries for this single file only.
+	sctx := &synthesizer.SynthesisContext{
+		Config:      w.config,
+		AuthDir:     w.authDir,
+		Now:         time.Now(),
+		IDGenerator: synthesizer.NewStableIDGenerator(),
+	}
+	generated := synthesizer.SynthesizeAuthFile(sctx, path, data)
+	newByID := authSliceToMap(generated)
+	if len(newByID) > 0 {
+		w.fileAuthsByPath[normalized] = newByID
+	} else {
+		delete(w.fileAuthsByPath, normalized)
+	}
+	updates := w.computePerPathUpdatesLocked(oldByID, newByID)
+	w.clientsMutex.Unlock()
+
 	w.persistAuthAsync(fmt.Sprintf("Sync auth %s", filepath.Base(path)), path)
+	w.dispatchAuthUpdates(updates)
 }
 
 func (w *Watcher) removeClient(path string) {
 	normalized := w.normalizeAuthPath(path)
 	w.clientsMutex.Lock()
-
-	cfg := w.config
+	oldByID := make(map[string]*coreauth.Auth, len(w.fileAuthsByPath[normalized]))
+	for id, a := range w.fileAuthsByPath[normalized] {
+		oldByID[id] = a
+	}
 	delete(w.lastAuthHashes, normalized)
 	delete(w.lastAuthContents, normalized)
+	delete(w.fileAuthsByPath, normalized)
 
-	w.clientsMutex.Unlock() // Release the lock before the callback
+	updates := w.computePerPathUpdatesLocked(oldByID, map[string]*coreauth.Auth{})
+	w.clientsMutex.Unlock()
 
-	w.refreshAuthState(false)
-
-	if w.reloadCallback != nil {
-		log.Debugf("triggering server update callback after removal")
-		w.reloadCallback(cfg)
-	}
 	w.persistAuthAsync(fmt.Sprintf("Remove auth %s", filepath.Base(path)), path)
+	w.dispatchAuthUpdates(updates)
+}
+
+func (w *Watcher) computePerPathUpdatesLocked(oldByID, newByID map[string]*coreauth.Auth) []AuthUpdate {
+	if w.currentAuths == nil {
+		w.currentAuths = make(map[string]*coreauth.Auth)
+	}
+	updates := make([]AuthUpdate, 0, len(oldByID)+len(newByID))
+	for id, newAuth := range newByID {
+		existing, ok := w.currentAuths[id]
+		if !ok {
+			w.currentAuths[id] = newAuth.Clone()
+			updates = append(updates, AuthUpdate{Action: AuthUpdateActionAdd, ID: id, Auth: newAuth.Clone()})
+			continue
+		}
+		if !authEqual(existing, newAuth) {
+			w.currentAuths[id] = newAuth.Clone()
+			updates = append(updates, AuthUpdate{Action: AuthUpdateActionModify, ID: id, Auth: newAuth.Clone()})
+		}
+	}
+	for id := range oldByID {
+		if _, stillExists := newByID[id]; stillExists {
+			continue
+		}
+		delete(w.currentAuths, id)
+		updates = append(updates, AuthUpdate{Action: AuthUpdateActionDelete, ID: id})
+	}
+	return updates
+}
+
+func authSliceToMap(auths []*coreauth.Auth) map[string]*coreauth.Auth {
+	byID := make(map[string]*coreauth.Auth, len(auths))
+	for _, a := range auths {
+		if a == nil || strings.TrimSpace(a.ID) == "" {
+			continue
+		}
+		byID[a.ID] = a
+	}
+	return byID
 }
 
 func (w *Watcher) loadFileClients(cfg *config.Config) int {
@@ -303,3 +369,79 @@ func (w *Watcher) persistAuthAsync(message string, paths ...string) {
 		}
 	}()
 }
+
+func (w *Watcher) stopServerUpdateTimer() {
+	w.serverUpdateMu.Lock()
+	defer w.serverUpdateMu.Unlock()
+	if w.serverUpdateTimer != nil {
+		w.serverUpdateTimer.Stop()
+		w.serverUpdateTimer = nil
+	}
+	w.serverUpdatePend = false
+}
+
+func (w *Watcher) triggerServerUpdate(cfg *config.Config) {
+	if w == nil || w.reloadCallback == nil || cfg == nil {
+		return
+	}
+	if w.stopped.Load() {
+		return
+	}
+
+	now := time.Now()
+
+	w.serverUpdateMu.Lock()
+	if w.serverUpdateLast.IsZero() || now.Sub(w.serverUpdateLast) >= serverUpdateDebounce {
+		w.serverUpdateLast = now
+		if w.serverUpdateTimer != nil {
+			w.serverUpdateTimer.Stop()
+			w.serverUpdateTimer = nil
+		}
+		w.serverUpdatePend = false
+		w.serverUpdateMu.Unlock()
+		w.reloadCallback(cfg)
+		return
+	}
+
+	if w.serverUpdatePend {
+		w.serverUpdateMu.Unlock()
+		return
+	}
+
+	delay := serverUpdateDebounce - now.Sub(w.serverUpdateLast)
+	if delay < 10*time.Millisecond {
+		delay = 10 * time.Millisecond
+	}
+	w.serverUpdatePend = true
+	if w.serverUpdateTimer != nil {
+		w.serverUpdateTimer.Stop()
+		w.serverUpdateTimer = nil
+	}
+	var timer *time.Timer
+	timer = time.AfterFunc(delay, func() {
+		if w.stopped.Load() {
+			return
+		}
+		w.clientsMutex.RLock()
+		latestCfg := w.config
+		w.clientsMutex.RUnlock()
+
+		w.serverUpdateMu.Lock()
+		if w.serverUpdateTimer != timer || !w.serverUpdatePend {
+			w.serverUpdateMu.Unlock()
+			return
+		}
+		w.serverUpdateTimer = nil
+		w.serverUpdatePend = false
+		if latestCfg == nil || w.reloadCallback == nil || w.stopped.Load() {
+			w.serverUpdateMu.Unlock()
+			return
+		}
+
+		w.serverUpdateLast = time.Now()
+		w.serverUpdateMu.Unlock()
+		w.reloadCallback(latestCfg)
+	})
+	w.serverUpdateTimer = timer
+	w.serverUpdateMu.Unlock()
+}

internal/watcher/diff/config_diff.go

@@ -304,6 +304,11 @@ func BuildConfigChangeDetails(oldCfg, newCfg *config.Config) []string {
 			if oldModels.hash != newModels.hash {
 				changes = append(changes, fmt.Sprintf("vertex[%d].models: updated (%d -> %d entries)", i, oldModels.count, newModels.count))
 			}
+			oldExcluded := SummarizeExcludedModels(o.ExcludedModels)
+			newExcluded := SummarizeExcludedModels(n.ExcludedModels)
+			if oldExcluded.hash != newExcluded.hash {
+				changes = append(changes, fmt.Sprintf("vertex[%d].excluded-models: updated (%d -> %d entries)", i, oldExcluded.count, newExcluded.count))
+			}
 			if !equalStringMap(o.Headers, n.Headers) {
 				changes = append(changes, fmt.Sprintf("vertex[%d].headers: updated", i))
 			}

internal/watcher/dispatcher.go

@@ -14,6 +14,8 @@ import (
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 )
 
+var snapshotCoreAuthsFunc = snapshotCoreAuths
+
 func (w *Watcher) setAuthUpdateQueue(queue chan<- AuthUpdate) {
 	w.clientsMutex.Lock()
 	defer w.clientsMutex.Unlock()
@@ -76,7 +78,11 @@ func (w *Watcher) dispatchRuntimeAuthUpdate(update AuthUpdate) bool {
 }
 
 func (w *Watcher) refreshAuthState(force bool) {
-	auths := w.SnapshotCoreAuths()
+	w.clientsMutex.RLock()
+	cfg := w.config
+	authDir := w.authDir
+	w.clientsMutex.RUnlock()
+	auths := snapshotCoreAuthsFunc(cfg, authDir)
 	w.clientsMutex.Lock()
 	if len(w.runtimeAuths) > 0 {
 		for _, a := range w.runtimeAuths {

internal/watcher/synthesizer/config.go

@@ -319,7 +319,7 @@ func (s *ConfigSynthesizer) synthesizeVertexCompat(ctx *SynthesisContext) []*cor
 			CreatedAt:  now,
 			UpdatedAt:  now,
 		}
-		ApplyAuthExcludedModelsMeta(a, cfg, nil, "apikey")
+		ApplyAuthExcludedModelsMeta(a, cfg, compat.ExcludedModels, "apikey")
 		out = append(out, a)
 	}
 	return out

internal/watcher/synthesizer/file.go

@@ -5,10 +5,12 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"runtime"
 	"strconv"
 	"strings"
 	"time"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/codex"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/runtime/geminicli"
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 )
@@ -35,9 +37,6 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 		return out, nil
 	}
 
-	now := ctx.Now
-	cfg := ctx.Config
-
 	for _, e := range entries {
 		if e.IsDir() {
 			continue
@@ -51,95 +50,130 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 		if errRead != nil || len(data) == 0 {
 			continue
 		}
-		var metadata map[string]any
-		if errUnmarshal := json.Unmarshal(data, &metadata); errUnmarshal != nil {
+		auths := synthesizeFileAuths(ctx, full, data)
+		if len(auths) == 0 {
 			continue
 		}
-		t, _ := metadata["type"].(string)
-		if t == "" {
-			continue
-		}
-		provider := strings.ToLower(t)
-		if provider == "gemini" {
-			provider = "gemini-cli"
-		}
-		label := provider
-		if email, _ := metadata["email"].(string); email != "" {
-			label = email
-		}
-		// Use relative path under authDir as ID to stay consistent with the file-based token store
-		id := full
-		if rel, errRel := filepath.Rel(ctx.AuthDir, full); errRel == nil && rel != "" {
-			id = rel
-		}
-
-		proxyURL := ""
-		if p, ok := metadata["proxy_url"].(string); ok {
-			proxyURL = p
-		}
-
-		prefix := ""
-		if rawPrefix, ok := metadata["prefix"].(string); ok {
-			trimmed := strings.TrimSpace(rawPrefix)
-			trimmed = strings.Trim(trimmed, "/")
-			if trimmed != "" && !strings.Contains(trimmed, "/") {
-				prefix = trimmed
-			}
-		}
-
-		disabled, _ := metadata["disabled"].(bool)
-		status := coreauth.StatusActive
-		if disabled {
-			status = coreauth.StatusDisabled
-		}
-
-		// Read per-account excluded models from the OAuth JSON file
-		perAccountExcluded := extractExcludedModelsFromMetadata(metadata)
-
-		a := &coreauth.Auth{
-			ID:       id,
-			Provider: provider,
-			Label:    label,
-			Prefix:   prefix,
-			Status:   status,
-			Disabled: disabled,
-			Attributes: map[string]string{
-				"source": full,
-				"path":   full,
-			},
-			ProxyURL:  proxyURL,
-			Metadata:  metadata,
-			CreatedAt: now,
-			UpdatedAt: now,
-		}
-		// Read priority from auth file
-		if rawPriority, ok := metadata["priority"]; ok {
-			switch v := rawPriority.(type) {
-			case float64:
-				a.Attributes["priority"] = strconv.Itoa(int(v))
-			case string:
-				priority := strings.TrimSpace(v)
-				if _, errAtoi := strconv.Atoi(priority); errAtoi == nil {
-					a.Attributes["priority"] = priority
-				}
-			}
-		}
-		ApplyAuthExcludedModelsMeta(a, cfg, perAccountExcluded, "oauth")
-		if provider == "gemini-cli" {
-			if virtuals := SynthesizeGeminiVirtualAuths(a, metadata, now); len(virtuals) > 0 {
-				for _, v := range virtuals {
-					ApplyAuthExcludedModelsMeta(v, cfg, perAccountExcluded, "oauth")
-				}
-				out = append(out, a)
-				out = append(out, virtuals...)
-				continue
-			}
-		}
-		out = append(out, a)
+		out = append(out, auths...)
 	}
 	return out, nil
 }
 
+// SynthesizeAuthFile generates Auth entries for one auth JSON file payload.
+// It shares exactly the same mapping behavior as FileSynthesizer.Synthesize.
+func SynthesizeAuthFile(ctx *SynthesisContext, fullPath string, data []byte) []*coreauth.Auth {
+	return synthesizeFileAuths(ctx, fullPath, data)
+}
+
+func synthesizeFileAuths(ctx *SynthesisContext, fullPath string, data []byte) []*coreauth.Auth {
+	if ctx == nil || len(data) == 0 {
+		return nil
+	}
+	now := ctx.Now
+	cfg := ctx.Config
+	var metadata map[string]any
+	if errUnmarshal := json.Unmarshal(data, &metadata); errUnmarshal != nil {
+		return nil
+	}
+	t, _ := metadata["type"].(string)
+	if t == "" {
+		return nil
+	}
+	provider := strings.ToLower(t)
+	if provider == "gemini" {
+		provider = "gemini-cli"
+	}
+	label := provider
+	if email, _ := metadata["email"].(string); email != "" {
+		label = email
+	}
+	// Use relative path under authDir as ID to stay consistent with the file-based token store.
+	id := fullPath
+	if strings.TrimSpace(ctx.AuthDir) != "" {
+		if rel, errRel := filepath.Rel(ctx.AuthDir, fullPath); errRel == nil && rel != "" {
+			id = rel
+		}
+	}
+	if runtime.GOOS == "windows" {
+		id = strings.ToLower(id)
+	}
+
+	proxyURL := ""
+	if p, ok := metadata["proxy_url"].(string); ok {
+		proxyURL = p
+	}
+
+	prefix := ""
+	if rawPrefix, ok := metadata["prefix"].(string); ok {
+		trimmed := strings.TrimSpace(rawPrefix)
+		trimmed = strings.Trim(trimmed, "/")
+		if trimmed != "" && !strings.Contains(trimmed, "/") {
+			prefix = trimmed
+		}
+	}
+
+	disabled, _ := metadata["disabled"].(bool)
+	status := coreauth.StatusActive
+	if disabled {
+		status = coreauth.StatusDisabled
+	}
+
+	// Read per-account excluded models from the OAuth JSON file.
+	perAccountExcluded := extractExcludedModelsFromMetadata(metadata)
+
+	a := &coreauth.Auth{
+		ID:       id,
+		Provider: provider,
+		Label:    label,
+		Prefix:   prefix,
+		Status:   status,
+		Disabled: disabled,
+		Attributes: map[string]string{
+			"source": fullPath,
+			"path":   fullPath,
+		},
+		ProxyURL:  proxyURL,
+		Metadata:  metadata,
+		CreatedAt: now,
+		UpdatedAt: now,
+	}
+	// Read priority from auth file.
+	if rawPriority, ok := metadata["priority"]; ok {
+		switch v := rawPriority.(type) {
+		case float64:
+			a.Attributes["priority"] = strconv.Itoa(int(v))
+		case string:
+			priority := strings.TrimSpace(v)
+			if _, errAtoi := strconv.Atoi(priority); errAtoi == nil {
+				a.Attributes["priority"] = priority
+			}
+		}
+	}
+	ApplyAuthExcludedModelsMeta(a, cfg, perAccountExcluded, "oauth")
+	// For codex auth files, extract plan_type from the JWT id_token.
+	if provider == "codex" {
+		if idTokenRaw, ok := metadata["id_token"].(string); ok && strings.TrimSpace(idTokenRaw) != "" {
+			if claims, errParse := codex.ParseJWTToken(idTokenRaw); errParse == nil && claims != nil {
+				if pt := strings.TrimSpace(claims.CodexAuthInfo.ChatgptPlanType); pt != "" {
+					a.Attributes["plan_type"] = pt
+				}
+			}
+		}
+	}
+	if provider == "gemini-cli" {
+		if virtuals := SynthesizeGeminiVirtualAuths(a, metadata, now); len(virtuals) > 0 {
+			for _, v := range virtuals {
+				ApplyAuthExcludedModelsMeta(v, cfg, perAccountExcluded, "oauth")
+			}
+			out := make([]*coreauth.Auth, 0, 1+len(virtuals))
+			out = append(out, a)
+			out = append(out, virtuals...)
+			return out
+		}
+	}
+	return []*coreauth.Auth{a}
+}
+
 // SynthesizeGeminiVirtualAuths creates virtual Auth entries for multi-project Gemini credentials.
 // It disables the primary auth and creates one virtual auth per project.
 func SynthesizeGeminiVirtualAuths(primary *coreauth.Auth, metadata map[string]any, now time.Time) []*coreauth.Auth {

internal/watcher/watcher.go

@@ -6,6 +6,7 @@ import (
 	"context"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"github.com/fsnotify/fsnotify"
@@ -35,10 +36,16 @@ type Watcher struct {
 	clientsMutex      sync.RWMutex
 	configReloadMu    sync.Mutex
 	configReloadTimer *time.Timer
+	serverUpdateMu    sync.Mutex
+	serverUpdateTimer *time.Timer
+	serverUpdateLast  time.Time
+	serverUpdatePend  bool
+	stopped           atomic.Bool
 	reloadCallback    func(*config.Config)
 	watcher           *fsnotify.Watcher
 	lastAuthHashes    map[string]string
 	lastAuthContents  map[string]*coreauth.Auth
+	fileAuthsByPath   map[string]map[string]*coreauth.Auth
 	lastRemoveTimes   map[string]time.Time
 	lastConfigHash    string
 	authQueue         chan<- AuthUpdate
@@ -76,6 +83,7 @@ const (
 	replaceCheckDelay        = 50 * time.Millisecond
 	configReloadDebounce     = 150 * time.Millisecond
 	authRemoveDebounceWindow = 1 * time.Second
+	serverUpdateDebounce     = 1 * time.Second
 )
 
 // NewWatcher creates a new file watcher instance
@@ -85,11 +93,12 @@ func NewWatcher(configPath, authDir string, reloadCallback func(*config.Config))
 		return nil, errNewWatcher
 	}
 	w := &Watcher{
-		configPath:     configPath,
-		authDir:        authDir,
-		reloadCallback: reloadCallback,
-		watcher:        watcher,
-		lastAuthHashes: make(map[string]string),
+		configPath:      configPath,
+		authDir:         authDir,
+		reloadCallback:  reloadCallback,
+		watcher:         watcher,
+		lastAuthHashes:  make(map[string]string),
+		fileAuthsByPath: make(map[string]map[string]*coreauth.Auth),
 	}
 	w.dispatchCond = sync.NewCond(&w.dispatchMu)
 	if store := sdkAuth.GetTokenStore(); store != nil {
@@ -114,8 +123,10 @@ func (w *Watcher) Start(ctx context.Context) error {
 
 // Stop stops the file watcher
 func (w *Watcher) Stop() error {
+	w.stopped.Store(true)
 	w.stopDispatch()
 	w.stopConfigReloadTimer()
+	w.stopServerUpdateTimer()
 	return w.watcher.Close()
 }
 

internal/watcher/watcher_test.go

@@ -406,8 +406,8 @@ func TestAddOrUpdateClientTriggersReloadAndHash(t *testing.T) {
 
 	w.addOrUpdateClient(authFile)
 
-	if got := atomic.LoadInt32(&reloads); got != 1 {
-		t.Fatalf("expected reload callback once, got %d", got)
+	if got := atomic.LoadInt32(&reloads); got != 0 {
+		t.Fatalf("expected no reload callback for auth update, got %d", got)
 	}
 	// Use normalizeAuthPath to match how addOrUpdateClient stores the key
 	normalized := w.normalizeAuthPath(authFile)
@@ -436,8 +436,150 @@ func TestRemoveClientRemovesHash(t *testing.T) {
 	if _, ok := w.lastAuthHashes[w.normalizeAuthPath(authFile)]; ok {
 		t.Fatal("expected hash to be removed after deletion")
 	}
+	if got := atomic.LoadInt32(&reloads); got != 0 {
+		t.Fatalf("expected no reload callback for auth removal, got %d", got)
+	}
+}
+
+func TestAuthFileEventsDoNotInvokeSnapshotCoreAuths(t *testing.T) {
+	tmpDir := t.TempDir()
+	authFile := filepath.Join(tmpDir, "sample.json")
+	if err := os.WriteFile(authFile, []byte(`{"type":"codex","email":"u@example.com"}`), 0o644); err != nil {
+		t.Fatalf("failed to create auth file: %v", err)
+	}
+
+	origSnapshot := snapshotCoreAuthsFunc
+	var snapshotCalls int32
+	snapshotCoreAuthsFunc = func(cfg *config.Config, authDir string) []*coreauth.Auth {
+		atomic.AddInt32(&snapshotCalls, 1)
+		return origSnapshot(cfg, authDir)
+	}
+	defer func() { snapshotCoreAuthsFunc = origSnapshot }()
+
+	w := &Watcher{
+		authDir:          tmpDir,
+		lastAuthHashes:   make(map[string]string),
+		lastAuthContents: make(map[string]*coreauth.Auth),
+		fileAuthsByPath:  make(map[string]map[string]*coreauth.Auth),
+	}
+	w.SetConfig(&config.Config{AuthDir: tmpDir})
+
+	w.addOrUpdateClient(authFile)
+	w.removeClient(authFile)
+
+	if got := atomic.LoadInt32(&snapshotCalls); got != 0 {
+		t.Fatalf("expected auth file events to avoid full snapshot, got %d calls", got)
+	}
+}
+
+func TestAuthSliceToMap(t *testing.T) {
+	t.Parallel()
+
+	valid1 := &coreauth.Auth{ID: "a"}
+	valid2 := &coreauth.Auth{ID: "b"}
+	dupOld := &coreauth.Auth{ID: "dup", Label: "old"}
+	dupNew := &coreauth.Auth{ID: "dup", Label: "new"}
+	empty := &coreauth.Auth{ID: "  "}
+
+	tests := []struct {
+		name string
+		in   []*coreauth.Auth
+		want map[string]*coreauth.Auth
+	}{
+		{
+			name: "nil input",
+			in:   nil,
+			want: map[string]*coreauth.Auth{},
+		},
+		{
+			name: "empty input",
+			in:   []*coreauth.Auth{},
+			want: map[string]*coreauth.Auth{},
+		},
+		{
+			name: "filters invalid auths",
+			in:   []*coreauth.Auth{nil, empty},
+			want: map[string]*coreauth.Auth{},
+		},
+		{
+			name: "keeps valid auths",
+			in:   []*coreauth.Auth{valid1, nil, valid2},
+			want: map[string]*coreauth.Auth{"a": valid1, "b": valid2},
+		},
+		{
+			name: "last duplicate wins",
+			in:   []*coreauth.Auth{dupOld, dupNew},
+			want: map[string]*coreauth.Auth{"dup": dupNew},
+		},
+	}
+
+	for _, tc := range tests {
+		tc := tc
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+			got := authSliceToMap(tc.in)
+			if len(tc.want) == 0 {
+				if got == nil {
+					t.Fatal("expected empty map, got nil")
+				}
+				if len(got) != 0 {
+					t.Fatalf("expected empty map, got %#v", got)
+				}
+				return
+			}
+			if len(got) != len(tc.want) {
+				t.Fatalf("unexpected map length: got %d, want %d", len(got), len(tc.want))
+			}
+			for id, wantAuth := range tc.want {
+				gotAuth, ok := got[id]
+				if !ok {
+					t.Fatalf("missing id %q in result map", id)
+				}
+				if !authEqual(gotAuth, wantAuth) {
+					t.Fatalf("unexpected auth for id %q: got %#v, want %#v", id, gotAuth, wantAuth)
+				}
+			}
+		})
+	}
+}
+
+func TestTriggerServerUpdateCancelsPendingTimerOnImmediate(t *testing.T) {
+	tmpDir := t.TempDir()
+	cfg := &config.Config{AuthDir: tmpDir}
+
+	var reloads int32
+	w := &Watcher{
+		reloadCallback: func(*config.Config) {
+			atomic.AddInt32(&reloads, 1)
+		},
+	}
+	w.SetConfig(cfg)
+
+	w.serverUpdateMu.Lock()
+	w.serverUpdateLast = time.Now().Add(-(serverUpdateDebounce - 100*time.Millisecond))
+	w.serverUpdateMu.Unlock()
+	w.triggerServerUpdate(cfg)
+
+	if got := atomic.LoadInt32(&reloads); got != 0 {
+		t.Fatalf("expected no immediate reload, got %d", got)
+	}
+
+	w.serverUpdateMu.Lock()
+	if !w.serverUpdatePend || w.serverUpdateTimer == nil {
+		w.serverUpdateMu.Unlock()
+		t.Fatal("expected a pending server update timer")
+	}
+	w.serverUpdateLast = time.Now().Add(-(serverUpdateDebounce + 10*time.Millisecond))
+	w.serverUpdateMu.Unlock()
+
+	w.triggerServerUpdate(cfg)
 	if got := atomic.LoadInt32(&reloads); got != 1 {
-		t.Fatalf("expected reload callback once, got %d", got)
+		t.Fatalf("expected immediate reload once, got %d", got)
+	}
+
+	time.Sleep(250 * time.Millisecond)
+	if got := atomic.LoadInt32(&reloads); got != 1 {
+		t.Fatalf("expected pending timer to be cancelled, got %d reloads", got)
 	}
 }
 
@@ -655,8 +797,8 @@ func TestHandleEventRemovesAuthFile(t *testing.T) {
 
 	w.handleEvent(fsnotify.Event{Name: authFile, Op: fsnotify.Remove})
 
-	if atomic.LoadInt32(&reloads) != 1 {
-		t.Fatalf("expected reload callback once, got %d", reloads)
+	if atomic.LoadInt32(&reloads) != 0 {
+		t.Fatalf("expected no reload callback for auth removal, got %d", reloads)
 	}
 	if _, ok := w.lastAuthHashes[w.normalizeAuthPath(authFile)]; ok {
 		t.Fatal("expected hash entry to be removed")
@@ -853,8 +995,8 @@ func TestHandleEventAuthWriteTriggersUpdate(t *testing.T) {
 	w.SetConfig(&config.Config{AuthDir: authDir})
 
 	w.handleEvent(fsnotify.Event{Name: authFile, Op: fsnotify.Write})
-	if atomic.LoadInt32(&reloads) != 1 {
-		t.Fatalf("expected auth write to trigger reload callback, got %d", reloads)
+	if atomic.LoadInt32(&reloads) != 0 {
+		t.Fatalf("expected auth write to avoid global reload callback, got %d", reloads)
 	}
 }
 
@@ -950,8 +1092,8 @@ func TestHandleEventAtomicReplaceChangedTriggersUpdate(t *testing.T) {
 	w.lastAuthHashes[w.normalizeAuthPath(authFile)] = hexString(oldSum[:])
 
 	w.handleEvent(fsnotify.Event{Name: authFile, Op: fsnotify.Rename})
-	if atomic.LoadInt32(&reloads) != 1 {
-		t.Fatalf("expected changed atomic replace to trigger update, got %d", reloads)
+	if atomic.LoadInt32(&reloads) != 0 {
+		t.Fatalf("expected changed atomic replace to avoid global reload, got %d", reloads)
 	}
 }
 
@@ -1005,8 +1147,8 @@ func TestHandleEventRemoveKnownFileDeletes(t *testing.T) {
 	w.lastAuthHashes[w.normalizeAuthPath(authFile)] = "hash"
 
 	w.handleEvent(fsnotify.Event{Name: authFile, Op: fsnotify.Remove})
-	if atomic.LoadInt32(&reloads) != 1 {
-		t.Fatalf("expected known remove to trigger reload, got %d", reloads)
+	if atomic.LoadInt32(&reloads) != 0 {
+		t.Fatalf("expected known remove to avoid global reload, got %d", reloads)
 	}
 	if _, ok := w.lastAuthHashes[w.normalizeAuthPath(authFile)]; ok {
 		t.Fatal("expected known auth hash to be deleted")

sdk/api/handlers/openai/openai_responses_websocket.go

@@ -14,7 +14,11 @@ import (
 	"github.com/google/uuid"
 	"github.com/gorilla/websocket"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/router-for-me/CLIProxyAPI/v6/sdk/api/handlers"
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
@@ -26,11 +30,12 @@ const (
 	wsRequestTypeAppend  = "response.append"
 	wsEventTypeError     = "error"
 	wsEventTypeCompleted = "response.completed"
-	wsEventTypeDone      = "response.done"
 	wsDoneMarker         = "[DONE]"
 	wsTurnStateHeader    = "x-codex-turn-state"
 	wsRequestBodyKey     = "REQUEST_BODY_OVERRIDE"
 	wsPayloadLogMaxSize  = 2048
+	wsBodyLogMaxSize     = 64 * 1024
+	wsBodyLogTruncated   = "\n[websocket log truncated]\n"
 )
 
 var responsesWebsocketUpgrader = websocket.Upgrader{
@@ -101,11 +106,17 @@ func (h *OpenAIResponsesAPIHandler) ResponsesWebsocket(c *gin.Context) {
 		// )
 		appendWebsocketEvent(&wsBodyLog, "request", payload)
 
-		allowIncrementalInputWithPreviousResponseID := websocketUpstreamSupportsIncrementalInput(nil, nil)
+		allowIncrementalInputWithPreviousResponseID := false
 		if pinnedAuthID != "" && h != nil && h.AuthManager != nil {
 			if pinnedAuth, ok := h.AuthManager.GetByID(pinnedAuthID); ok && pinnedAuth != nil {
 				allowIncrementalInputWithPreviousResponseID = websocketUpstreamSupportsIncrementalInput(pinnedAuth.Attributes, pinnedAuth.Metadata)
 			}
+		} else {
+			requestModelName := strings.TrimSpace(gjson.GetBytes(payload, "model").String())
+			if requestModelName == "" {
+				requestModelName = strings.TrimSpace(gjson.GetBytes(lastRequest, "model").String())
+			}
+			allowIncrementalInputWithPreviousResponseID = h.websocketUpstreamSupportsIncrementalInputForModel(requestModelName)
 		}
 
 		var requestJSON []byte
@@ -140,6 +151,22 @@ func (h *OpenAIResponsesAPIHandler) ResponsesWebsocket(c *gin.Context) {
 			}
 			continue
 		}
+		if shouldHandleResponsesWebsocketPrewarmLocally(payload, lastRequest, allowIncrementalInputWithPreviousResponseID) {
+			if updated, errDelete := sjson.DeleteBytes(requestJSON, "generate"); errDelete == nil {
+				requestJSON = updated
+			}
+			if updated, errDelete := sjson.DeleteBytes(updatedLastRequest, "generate"); errDelete == nil {
+				updatedLastRequest = updated
+			}
+			lastRequest = updatedLastRequest
+			lastResponseOutput = []byte("[]")
+			if errWrite := writeResponsesWebsocketSyntheticPrewarm(c, conn, requestJSON, &wsBodyLog, passthroughSessionID); errWrite != nil {
+				wsTerminateErr = errWrite
+				appendWebsocketEvent(&wsBodyLog, "disconnect", []byte(errWrite.Error()))
+				return
+			}
+			continue
+		}
 		lastRequest = updatedLastRequest
 
 		modelName := gjson.GetBytes(requestJSON, "model").String()
@@ -340,6 +367,192 @@ func websocketUpstreamSupportsIncrementalInput(attributes map[string]string, met
 	return false
 }
 
+func (h *OpenAIResponsesAPIHandler) websocketUpstreamSupportsIncrementalInputForModel(modelName string) bool {
+	if h == nil || h.AuthManager == nil {
+		return false
+	}
+
+	resolvedModelName := modelName
+	initialSuffix := thinking.ParseSuffix(modelName)
+	if initialSuffix.ModelName == "auto" {
+		resolvedBase := util.ResolveAutoModel(initialSuffix.ModelName)
+		if initialSuffix.HasSuffix {
+			resolvedModelName = fmt.Sprintf("%s(%s)", resolvedBase, initialSuffix.RawSuffix)
+		} else {
+			resolvedModelName = resolvedBase
+		}
+	} else {
+		resolvedModelName = util.ResolveAutoModel(modelName)
+	}
+
+	parsed := thinking.ParseSuffix(resolvedModelName)
+	baseModel := strings.TrimSpace(parsed.ModelName)
+	providers := util.GetProviderName(baseModel)
+	if len(providers) == 0 && baseModel != resolvedModelName {
+		providers = util.GetProviderName(resolvedModelName)
+	}
+	if len(providers) == 0 {
+		return false
+	}
+
+	providerSet := make(map[string]struct{}, len(providers))
+	for i := 0; i < len(providers); i++ {
+		providerKey := strings.TrimSpace(strings.ToLower(providers[i]))
+		if providerKey == "" {
+			continue
+		}
+		providerSet[providerKey] = struct{}{}
+	}
+	if len(providerSet) == 0 {
+		return false
+	}
+
+	modelKey := baseModel
+	if modelKey == "" {
+		modelKey = strings.TrimSpace(resolvedModelName)
+	}
+	registryRef := registry.GetGlobalRegistry()
+	now := time.Now()
+	auths := h.AuthManager.List()
+	for i := 0; i < len(auths); i++ {
+		auth := auths[i]
+		if auth == nil {
+			continue
+		}
+		providerKey := strings.TrimSpace(strings.ToLower(auth.Provider))
+		if _, ok := providerSet[providerKey]; !ok {
+			continue
+		}
+		if modelKey != "" && registryRef != nil && !registryRef.ClientSupportsModel(auth.ID, modelKey) {
+			continue
+		}
+		if !responsesWebsocketAuthAvailableForModel(auth, modelKey, now) {
+			continue
+		}
+		if websocketUpstreamSupportsIncrementalInput(auth.Attributes, auth.Metadata) {
+			return true
+		}
+	}
+	return false
+}
+
+func responsesWebsocketAuthAvailableForModel(auth *coreauth.Auth, modelName string, now time.Time) bool {
+	if auth == nil {
+		return false
+	}
+	if auth.Disabled || auth.Status == coreauth.StatusDisabled {
+		return false
+	}
+	if modelName != "" && len(auth.ModelStates) > 0 {
+		state, ok := auth.ModelStates[modelName]
+		if (!ok || state == nil) && modelName != "" {
+			baseModel := strings.TrimSpace(thinking.ParseSuffix(modelName).ModelName)
+			if baseModel != "" && baseModel != modelName {
+				state, ok = auth.ModelStates[baseModel]
+			}
+		}
+		if ok && state != nil {
+			if state.Status == coreauth.StatusDisabled {
+				return false
+			}
+			if state.Unavailable && !state.NextRetryAfter.IsZero() && state.NextRetryAfter.After(now) {
+				return false
+			}
+			return true
+		}
+	}
+	if auth.Unavailable && !auth.NextRetryAfter.IsZero() && auth.NextRetryAfter.After(now) {
+		return false
+	}
+	return true
+}
+
+func shouldHandleResponsesWebsocketPrewarmLocally(rawJSON []byte, lastRequest []byte, allowIncrementalInputWithPreviousResponseID bool) bool {
+	if allowIncrementalInputWithPreviousResponseID || len(lastRequest) != 0 {
+		return false
+	}
+	if strings.TrimSpace(gjson.GetBytes(rawJSON, "type").String()) != wsRequestTypeCreate {
+		return false
+	}
+	generateResult := gjson.GetBytes(rawJSON, "generate")
+	return generateResult.Exists() && !generateResult.Bool()
+}
+
+func writeResponsesWebsocketSyntheticPrewarm(
+	c *gin.Context,
+	conn *websocket.Conn,
+	requestJSON []byte,
+	wsBodyLog *strings.Builder,
+	sessionID string,
+) error {
+	payloads, errPayloads := syntheticResponsesWebsocketPrewarmPayloads(requestJSON)
+	if errPayloads != nil {
+		return errPayloads
+	}
+	for i := 0; i < len(payloads); i++ {
+		markAPIResponseTimestamp(c)
+		appendWebsocketEvent(wsBodyLog, "response", payloads[i])
+		// log.Infof(
+		// 	"responses websocket: downstream_out id=%s type=%d event=%s payload=%s",
+		// 	sessionID,
+		// 	websocket.TextMessage,
+		// 	websocketPayloadEventType(payloads[i]),
+		// 	websocketPayloadPreview(payloads[i]),
+		// )
+		if errWrite := conn.WriteMessage(websocket.TextMessage, payloads[i]); errWrite != nil {
+			log.Warnf(
+				"responses websocket: downstream_out write failed id=%s event=%s error=%v",
+				sessionID,
+				websocketPayloadEventType(payloads[i]),
+				errWrite,
+			)
+			return errWrite
+		}
+	}
+	return nil
+}
+
+func syntheticResponsesWebsocketPrewarmPayloads(requestJSON []byte) ([][]byte, error) {
+	responseID := "resp_prewarm_" + uuid.NewString()
+	createdAt := time.Now().Unix()
+	modelName := strings.TrimSpace(gjson.GetBytes(requestJSON, "model").String())
+
+	createdPayload := []byte(`{"type":"response.created","sequence_number":0,"response":{"id":"","object":"response","created_at":0,"status":"in_progress","background":false,"error":null,"output":[]}}`)
+	var errSet error
+	createdPayload, errSet = sjson.SetBytes(createdPayload, "response.id", responseID)
+	if errSet != nil {
+		return nil, errSet
+	}
+	createdPayload, errSet = sjson.SetBytes(createdPayload, "response.created_at", createdAt)
+	if errSet != nil {
+		return nil, errSet
+	}
+	if modelName != "" {
+		createdPayload, errSet = sjson.SetBytes(createdPayload, "response.model", modelName)
+		if errSet != nil {
+			return nil, errSet
+		}
+	}
+
+	completedPayload := []byte(`{"type":"response.completed","sequence_number":1,"response":{"id":"","object":"response","created_at":0,"status":"completed","background":false,"error":null,"output":[],"usage":{"input_tokens":0,"output_tokens":0,"total_tokens":0}}}`)
+	completedPayload, errSet = sjson.SetBytes(completedPayload, "response.id", responseID)
+	if errSet != nil {
+		return nil, errSet
+	}
+	completedPayload, errSet = sjson.SetBytes(completedPayload, "response.created_at", createdAt)
+	if errSet != nil {
+		return nil, errSet
+	}
+	if modelName != "" {
+		completedPayload, errSet = sjson.SetBytes(completedPayload, "response.model", modelName)
+		if errSet != nil {
+			return nil, errSet
+		}
+	}
+
+	return [][]byte{createdPayload, completedPayload}, nil
+}
+
 func mergeJSONArrayRaw(existingRaw, appendRaw string) (string, error) {
 	existingRaw = strings.TrimSpace(existingRaw)
 	appendRaw = strings.TrimSpace(appendRaw)
@@ -469,9 +682,6 @@ func (h *OpenAIResponsesAPIHandler) forwardResponsesWebsocket(
 			for i := range payloads {
 				eventType := gjson.GetBytes(payloads[i], "type").String()
 				if eventType == wsEventTypeCompleted {
-					// log.Infof("replace %s with %s", wsEventTypeCompleted, wsEventTypeDone)
-					payloads[i], _ = sjson.SetBytes(payloads[i], "type", wsEventTypeDone)
-
 					completed = true
 					completedOutput = responseCompletedOutputFromPayload(payloads[i])
 				}
@@ -554,65 +764,134 @@ func writeResponsesWebsocketError(conn *websocket.Conn, errMsg *interfaces.Error
 	}
 
 	body := handlers.BuildErrorResponseBody(status, errText)
-	payload := map[string]any{
-		"type":   wsEventTypeError,
-		"status": status,
+	payload := []byte(`{}`)
+	var errSet error
+	payload, errSet = sjson.SetBytes(payload, "type", wsEventTypeError)
+	if errSet != nil {
+		return nil, errSet
+	}
+	payload, errSet = sjson.SetBytes(payload, "status", status)
+	if errSet != nil {
+		return nil, errSet
 	}
 
 	if errMsg != nil && errMsg.Addon != nil {
-		headers := map[string]any{}
+		headers := []byte(`{}`)
+		hasHeaders := false
 		for key, values := range errMsg.Addon {
 			if len(values) == 0 {
 				continue
 			}
-			headers[key] = values[0]
+			headerPath := strings.ReplaceAll(strings.ReplaceAll(key, `\\`, `\\\\`), ".", `\\.`)
+			headers, errSet = sjson.SetBytes(headers, headerPath, values[0])
+			if errSet != nil {
+				return nil, errSet
+			}
+			hasHeaders = true
 		}
-		if len(headers) > 0 {
-			payload["headers"] = headers
-		}
-	}
-
-	if len(body) > 0 && json.Valid(body) {
-		var decoded map[string]any
-		if errDecode := json.Unmarshal(body, &decoded); errDecode == nil {
-			if inner, ok := decoded["error"]; ok {
-				payload["error"] = inner
-			} else {
-				payload["error"] = decoded
+		if hasHeaders {
+			payload, errSet = sjson.SetRawBytes(payload, "headers", headers)
+			if errSet != nil {
+				return nil, errSet
 			}
 		}
 	}
 
-	if _, ok := payload["error"]; !ok {
-		payload["error"] = map[string]any{
-			"type":    "server_error",
-			"message": errText,
+	if len(body) > 0 && json.Valid(body) {
+		errorNode := gjson.GetBytes(body, "error")
+		if errorNode.Exists() {
+			payload, errSet = sjson.SetRawBytes(payload, "error", []byte(errorNode.Raw))
+		} else {
+			payload, errSet = sjson.SetRawBytes(payload, "error", body)
+		}
+		if errSet != nil {
+			return nil, errSet
 		}
 	}
 
-	data, err := json.Marshal(payload)
-	if err != nil {
-		return nil, err
+	if !gjson.GetBytes(payload, "error").Exists() {
+		payload, errSet = sjson.SetBytes(payload, "error.type", "server_error")
+		if errSet != nil {
+			return nil, errSet
+		}
+		payload, errSet = sjson.SetBytes(payload, "error.message", errText)
+		if errSet != nil {
+			return nil, errSet
+		}
 	}
-	return data, conn.WriteMessage(websocket.TextMessage, data)
+
+	return payload, conn.WriteMessage(websocket.TextMessage, payload)
 }
 
 func appendWebsocketEvent(builder *strings.Builder, eventType string, payload []byte) {
 	if builder == nil {
 		return
 	}
+	if builder.Len() >= wsBodyLogMaxSize {
+		return
+	}
 	trimmedPayload := bytes.TrimSpace(payload)
 	if len(trimmedPayload) == 0 {
 		return
 	}
 	if builder.Len() > 0 {
-		builder.WriteString("\n")
+		if !appendWebsocketLogString(builder, "\n") {
+			return
+		}
 	}
-	builder.WriteString("websocket.")
-	builder.WriteString(eventType)
-	builder.WriteString("\n")
-	builder.Write(trimmedPayload)
-	builder.WriteString("\n")
+	if !appendWebsocketLogString(builder, "websocket.") {
+		return
+	}
+	if !appendWebsocketLogString(builder, eventType) {
+		return
+	}
+	if !appendWebsocketLogString(builder, "\n") {
+		return
+	}
+	if !appendWebsocketLogBytes(builder, trimmedPayload, len(wsBodyLogTruncated)) {
+		appendWebsocketLogString(builder, wsBodyLogTruncated)
+		return
+	}
+	appendWebsocketLogString(builder, "\n")
+}
+
+func appendWebsocketLogString(builder *strings.Builder, value string) bool {
+	if builder == nil {
+		return false
+	}
+	remaining := wsBodyLogMaxSize - builder.Len()
+	if remaining <= 0 {
+		return false
+	}
+	if len(value) <= remaining {
+		builder.WriteString(value)
+		return true
+	}
+	builder.WriteString(value[:remaining])
+	return false
+}
+
+func appendWebsocketLogBytes(builder *strings.Builder, value []byte, reserveForSuffix int) bool {
+	if builder == nil {
+		return false
+	}
+	remaining := wsBodyLogMaxSize - builder.Len()
+	if remaining <= 0 {
+		return false
+	}
+	if len(value) <= remaining {
+		builder.Write(value)
+		return true
+	}
+	limit := remaining - reserveForSuffix
+	if limit < 0 {
+		limit = 0
+	}
+	if limit > len(value) {
+		limit = len(value)
+	}
+	builder.Write(value[:limit])
+	return false
 }
 
 func websocketPayloadEventType(payload []byte) string {

sdk/api/handlers/openai/openai_responses_websocket_test.go

@@ -2,15 +2,57 @@ package openai
 
 import (
 	"bytes"
+	"context"
+	"errors"
+	"fmt"
 	"net/http"
 	"net/http/httptest"
 	"strings"
 	"testing"
 
 	"github.com/gin-gonic/gin"
+	"github.com/gorilla/websocket"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	"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"
 	"github.com/tidwall/gjson"
 )
 
+type websocketCaptureExecutor struct {
+	streamCalls int
+	payloads    [][]byte
+}
+
+func (e *websocketCaptureExecutor) Identifier() string { return "test-provider" }
+
+func (e *websocketCaptureExecutor) Execute(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
+	return coreexecutor.Response{}, errors.New("not implemented")
+}
+
+func (e *websocketCaptureExecutor) ExecuteStream(_ context.Context, _ *coreauth.Auth, req coreexecutor.Request, _ coreexecutor.Options) (*coreexecutor.StreamResult, error) {
+	e.streamCalls++
+	e.payloads = append(e.payloads, bytes.Clone(req.Payload))
+	chunks := make(chan coreexecutor.StreamChunk, 1)
+	chunks <- coreexecutor.StreamChunk{Payload: []byte(`{"type":"response.completed","response":{"id":"resp-upstream","output":[{"type":"message","id":"out-1"}]}}`)}
+	close(chunks)
+	return &coreexecutor.StreamResult{Chunks: chunks}, nil
+}
+
+func (e *websocketCaptureExecutor) Refresh(_ context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
+	return auth, nil
+}
+
+func (e *websocketCaptureExecutor) CountTokens(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
+	return coreexecutor.Response{}, errors.New("not implemented")
+}
+
+func (e *websocketCaptureExecutor) HttpRequest(context.Context, *coreauth.Auth, *http.Request) (*http.Response, error) {
+	return nil, errors.New("not implemented")
+}
+
 func TestNormalizeResponsesWebsocketRequestCreate(t *testing.T) {
 	raw := []byte(`{"type":"response.create","model":"test-model","stream":false,"input":[{"type":"message","id":"msg-1"}]}`)
 
@@ -224,6 +266,33 @@ func TestAppendWebsocketEvent(t *testing.T) {
 	}
 }
 
+func TestAppendWebsocketEventTruncatesAtLimit(t *testing.T) {
+	var builder strings.Builder
+	payload := bytes.Repeat([]byte("x"), wsBodyLogMaxSize)
+
+	appendWebsocketEvent(&builder, "request", payload)
+
+	got := builder.String()
+	if len(got) > wsBodyLogMaxSize {
+		t.Fatalf("body log len = %d, want <= %d", len(got), wsBodyLogMaxSize)
+	}
+	if !strings.Contains(got, wsBodyLogTruncated) {
+		t.Fatalf("expected truncation marker in body log")
+	}
+}
+
+func TestAppendWebsocketEventNoGrowthAfterLimit(t *testing.T) {
+	var builder strings.Builder
+	appendWebsocketEvent(&builder, "request", bytes.Repeat([]byte("x"), wsBodyLogMaxSize))
+	initial := builder.String()
+
+	appendWebsocketEvent(&builder, "response", []byte(`{"type":"response.completed"}`))
+
+	if builder.String() != initial {
+		t.Fatalf("builder grew after reaching limit")
+	}
+}
+
 func TestSetWebsocketRequestBody(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 	recorder := httptest.NewRecorder()
@@ -247,3 +316,206 @@ func TestSetWebsocketRequestBody(t *testing.T) {
 		t.Fatalf("request body = %q, want %q", string(bodyBytes), "event body")
 	}
 }
+
+func TestForwardResponsesWebsocketPreservesCompletedEvent(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	serverErrCh := make(chan error, 1)
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		conn, err := responsesWebsocketUpgrader.Upgrade(w, r, nil)
+		if err != nil {
+			serverErrCh <- err
+			return
+		}
+		defer func() {
+			errClose := conn.Close()
+			if errClose != nil {
+				serverErrCh <- errClose
+			}
+		}()
+
+		ctx, _ := gin.CreateTestContext(httptest.NewRecorder())
+		ctx.Request = r
+
+		data := make(chan []byte, 1)
+		errCh := make(chan *interfaces.ErrorMessage)
+		data <- []byte("data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp-1\",\"output\":[{\"type\":\"message\",\"id\":\"out-1\"}]}}\n\n")
+		close(data)
+		close(errCh)
+
+		var bodyLog strings.Builder
+		completedOutput, err := (*OpenAIResponsesAPIHandler)(nil).forwardResponsesWebsocket(
+			ctx,
+			conn,
+			func(...interface{}) {},
+			data,
+			errCh,
+			&bodyLog,
+			"session-1",
+		)
+		if err != nil {
+			serverErrCh <- err
+			return
+		}
+		if gjson.GetBytes(completedOutput, "0.id").String() != "out-1" {
+			serverErrCh <- errors.New("completed output not captured")
+			return
+		}
+		serverErrCh <- nil
+	}))
+	defer server.Close()
+
+	wsURL := "ws" + strings.TrimPrefix(server.URL, "http")
+	conn, _, err := websocket.DefaultDialer.Dial(wsURL, nil)
+	if err != nil {
+		t.Fatalf("dial websocket: %v", err)
+	}
+	defer func() {
+		errClose := conn.Close()
+		if errClose != nil {
+			t.Fatalf("close websocket: %v", errClose)
+		}
+	}()
+
+	_, payload, errReadMessage := conn.ReadMessage()
+	if errReadMessage != nil {
+		t.Fatalf("read websocket message: %v", errReadMessage)
+	}
+	if gjson.GetBytes(payload, "type").String() != wsEventTypeCompleted {
+		t.Fatalf("payload type = %s, want %s", gjson.GetBytes(payload, "type").String(), wsEventTypeCompleted)
+	}
+	if strings.Contains(string(payload), "response.done") {
+		t.Fatalf("payload unexpectedly rewrote completed event: %s", payload)
+	}
+
+	if errServer := <-serverErrCh; errServer != nil {
+		t.Fatalf("server error: %v", errServer)
+	}
+}
+
+func TestWebsocketUpstreamSupportsIncrementalInputForModel(t *testing.T) {
+	manager := coreauth.NewManager(nil, nil, nil)
+	auth := &coreauth.Auth{
+		ID:         "auth-ws",
+		Provider:   "test-provider",
+		Status:     coreauth.StatusActive,
+		Attributes: map[string]string{"websockets": "true"},
+	}
+	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)
+	if !h.websocketUpstreamSupportsIncrementalInputForModel("test-model") {
+		t.Fatalf("expected websocket-capable upstream for test-model")
+	}
+}
+
+func TestResponsesWebsocketPrewarmHandledLocallyForSSEUpstream(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	executor := &websocketCaptureExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+	auth := &coreauth.Auth{ID: "auth-sse", 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.GET("/v1/responses/ws", h.ResponsesWebsocket)
+
+	server := httptest.NewServer(router)
+	defer server.Close()
+
+	wsURL := "ws" + strings.TrimPrefix(server.URL, "http") + "/v1/responses/ws"
+	conn, _, err := websocket.DefaultDialer.Dial(wsURL, nil)
+	if err != nil {
+		t.Fatalf("dial websocket: %v", err)
+	}
+	defer func() {
+		errClose := conn.Close()
+		if errClose != nil {
+			t.Fatalf("close websocket: %v", errClose)
+		}
+	}()
+
+	errWrite := conn.WriteMessage(websocket.TextMessage, []byte(`{"type":"response.create","model":"test-model","generate":false}`))
+	if errWrite != nil {
+		t.Fatalf("write prewarm websocket message: %v", errWrite)
+	}
+
+	_, createdPayload, errReadMessage := conn.ReadMessage()
+	if errReadMessage != nil {
+		t.Fatalf("read prewarm created message: %v", errReadMessage)
+	}
+	if gjson.GetBytes(createdPayload, "type").String() != "response.created" {
+		t.Fatalf("created payload type = %s, want response.created", gjson.GetBytes(createdPayload, "type").String())
+	}
+	prewarmResponseID := gjson.GetBytes(createdPayload, "response.id").String()
+	if prewarmResponseID == "" {
+		t.Fatalf("prewarm response id is empty")
+	}
+	if executor.streamCalls != 0 {
+		t.Fatalf("stream calls after prewarm = %d, want 0", executor.streamCalls)
+	}
+
+	_, completedPayload, errReadMessage := conn.ReadMessage()
+	if errReadMessage != nil {
+		t.Fatalf("read prewarm completed message: %v", errReadMessage)
+	}
+	if gjson.GetBytes(completedPayload, "type").String() != wsEventTypeCompleted {
+		t.Fatalf("completed payload type = %s, want %s", gjson.GetBytes(completedPayload, "type").String(), wsEventTypeCompleted)
+	}
+	if gjson.GetBytes(completedPayload, "response.id").String() != prewarmResponseID {
+		t.Fatalf("completed response id = %s, want %s", gjson.GetBytes(completedPayload, "response.id").String(), prewarmResponseID)
+	}
+	if gjson.GetBytes(completedPayload, "response.usage.total_tokens").Int() != 0 {
+		t.Fatalf("prewarm total tokens = %d, want 0", gjson.GetBytes(completedPayload, "response.usage.total_tokens").Int())
+	}
+
+	secondRequest := fmt.Sprintf(`{"type":"response.create","previous_response_id":%q,"input":[{"type":"message","id":"msg-1"}]}`, prewarmResponseID)
+	errWrite = conn.WriteMessage(websocket.TextMessage, []byte(secondRequest))
+	if errWrite != nil {
+		t.Fatalf("write follow-up websocket message: %v", errWrite)
+	}
+
+	_, upstreamPayload, errReadMessage := conn.ReadMessage()
+	if errReadMessage != nil {
+		t.Fatalf("read upstream completed message: %v", errReadMessage)
+	}
+	if gjson.GetBytes(upstreamPayload, "type").String() != wsEventTypeCompleted {
+		t.Fatalf("upstream payload type = %s, want %s", gjson.GetBytes(upstreamPayload, "type").String(), wsEventTypeCompleted)
+	}
+	if executor.streamCalls != 1 {
+		t.Fatalf("stream calls after follow-up = %d, want 1", executor.streamCalls)
+	}
+	if len(executor.payloads) != 1 {
+		t.Fatalf("captured upstream payloads = %d, want 1", len(executor.payloads))
+	}
+	forwarded := executor.payloads[0]
+	if gjson.GetBytes(forwarded, "previous_response_id").Exists() {
+		t.Fatalf("previous_response_id leaked upstream: %s", forwarded)
+	}
+	if gjson.GetBytes(forwarded, "generate").Exists() {
+		t.Fatalf("generate leaked upstream: %s", forwarded)
+	}
+	if gjson.GetBytes(forwarded, "model").String() != "test-model" {
+		t.Fatalf("forwarded model = %s, want test-model", gjson.GetBytes(forwarded, "model").String())
+	}
+	input := gjson.GetBytes(forwarded, "input").Array()
+	if len(input) != 1 || input[0].Get("id").String() != "msg-1" {
+		t.Fatalf("unexpected forwarded input: %s", forwarded)
+	}
+}

sdk/auth/codex_device.go

@@ -287,5 +287,8 @@ func (a *CodexAuthenticator) buildAuthRecord(authSvc *codex.CodexAuth, authBundl
 		FileName: fileName,
 		Storage:  tokenStorage,
 		Metadata: metadata,
+		Attributes: map[string]string{
+			"plan_type": planType,
+		},
 	}, nil
 }

sdk/auth/filestore.go

@@ -10,6 +10,7 @@ import (
 	"net/url"
 	"os"
 	"path/filepath"
+	"runtime"
 	"strings"
 	"sync"
 	"time"
@@ -266,14 +267,17 @@ func (s *FileTokenStore) readAuthFile(path, baseDir string) (*cliproxyauth.Auth,
 }
 
 func (s *FileTokenStore) idFor(path, baseDir string) string {
-	if baseDir == "" {
-		return path
+	id := path
+	if baseDir != "" {
+		if rel, errRel := filepath.Rel(baseDir, path); errRel == nil && rel != "" {
+			id = rel
+		}
 	}
-	rel, err := filepath.Rel(baseDir, path)
-	if err != nil {
-		return path
+	// On Windows, normalize ID casing to avoid duplicate auth entries caused by case-insensitive paths.
+	if runtime.GOOS == "windows" {
+		id = strings.ToLower(id)
 	}
-	return rel
+	return id
 }
 
 func (s *FileTokenStore) resolveAuthPath(auth *cliproxyauth.Auth) (string, error) {

sdk/cliproxy/auth/conductor.go

@@ -134,6 +134,7 @@ type Manager struct {
 	hook      Hook
 	mu        sync.RWMutex
 	auths     map[string]*Auth
+	scheduler *authScheduler
 	// providerOffsets tracks per-model provider rotation state for multi-provider routing.
 	providerOffsets map[string]int
 
@@ -149,6 +150,9 @@ type Manager struct {
 	// Keyed by auth.ID, value is alias(lower) -> upstream model (including suffix).
 	apiKeyModelAlias atomic.Value
 
+	// modelPoolOffsets tracks per-auth alias pool rotation state.
+	modelPoolOffsets map[string]int
+
 	// runtimeConfig stores the latest application config for request-time decisions.
 	// It is initialized in NewManager; never Load() before first Store().
 	runtimeConfig atomic.Value
@@ -176,14 +180,59 @@ func NewManager(store Store, selector Selector, hook Hook) *Manager {
 		hook:             hook,
 		auths:            make(map[string]*Auth),
 		providerOffsets:  make(map[string]int),
+		modelPoolOffsets: make(map[string]int),
 		refreshSemaphore: make(chan struct{}, refreshMaxConcurrency),
 	}
 	// atomic.Value requires non-nil initial value.
 	manager.runtimeConfig.Store(&internalconfig.Config{})
 	manager.apiKeyModelAlias.Store(apiKeyModelAliasTable(nil))
+	manager.scheduler = newAuthScheduler(selector)
 	return manager
 }
 
+func isBuiltInSelector(selector Selector) bool {
+	switch selector.(type) {
+	case *RoundRobinSelector, *FillFirstSelector:
+		return true
+	default:
+		return false
+	}
+}
+
+func (m *Manager) syncSchedulerFromSnapshot(auths []*Auth) {
+	if m == nil || m.scheduler == nil {
+		return
+	}
+	m.scheduler.rebuild(auths)
+}
+
+func (m *Manager) syncScheduler() {
+	if m == nil || m.scheduler == nil {
+		return
+	}
+	m.syncSchedulerFromSnapshot(m.snapshotAuths())
+}
+
+// RefreshSchedulerEntry re-upserts a single auth into the scheduler so that its
+// supportedModelSet is rebuilt from the current global model registry state.
+// This must be called after models have been registered for a newly added auth,
+// because the initial scheduler.upsertAuth during Register/Update runs before
+// registerModelsForAuth and therefore snapshots an empty model set.
+func (m *Manager) RefreshSchedulerEntry(authID string) {
+	if m == nil || m.scheduler == nil || authID == "" {
+		return
+	}
+	m.mu.RLock()
+	auth, ok := m.auths[authID]
+	if !ok || auth == nil {
+		m.mu.RUnlock()
+		return
+	}
+	snapshot := auth.Clone()
+	m.mu.RUnlock()
+	m.scheduler.upsertAuth(snapshot)
+}
+
 func (m *Manager) SetSelector(selector Selector) {
 	if m == nil {
 		return
@@ -194,6 +243,10 @@ func (m *Manager) SetSelector(selector Selector) {
 	m.mu.Lock()
 	m.selector = selector
 	m.mu.Unlock()
+	if m.scheduler != nil {
+		m.scheduler.setSelector(selector)
+		m.syncScheduler()
+	}
 }
 
 // SetStore swaps the underlying persistence store.
@@ -251,16 +304,323 @@ func (m *Manager) lookupAPIKeyUpstreamModel(authID, requestedModel string) strin
 	if resolved == "" {
 		return ""
 	}
-	// Preserve thinking suffix from the client's requested model unless config already has one.
-	requestResult := thinking.ParseSuffix(requestedModel)
-	if thinking.ParseSuffix(resolved).HasSuffix {
-		return resolved
-	}
-	if requestResult.HasSuffix && requestResult.RawSuffix != "" {
-		return resolved + "(" + requestResult.RawSuffix + ")"
-	}
-	return resolved
+	return preserveRequestedModelSuffix(requestedModel, resolved)
+}
 
+func isAPIKeyAuth(auth *Auth) bool {
+	if auth == nil {
+		return false
+	}
+	kind, _ := auth.AccountInfo()
+	return strings.EqualFold(strings.TrimSpace(kind), "api_key")
+}
+
+func isOpenAICompatAPIKeyAuth(auth *Auth) bool {
+	if !isAPIKeyAuth(auth) {
+		return false
+	}
+	if strings.EqualFold(strings.TrimSpace(auth.Provider), "openai-compatibility") {
+		return true
+	}
+	if auth.Attributes == nil {
+		return false
+	}
+	return strings.TrimSpace(auth.Attributes["compat_name"]) != ""
+}
+
+func openAICompatProviderKey(auth *Auth) string {
+	if auth == nil {
+		return ""
+	}
+	if auth.Attributes != nil {
+		if providerKey := strings.TrimSpace(auth.Attributes["provider_key"]); providerKey != "" {
+			return strings.ToLower(providerKey)
+		}
+		if compatName := strings.TrimSpace(auth.Attributes["compat_name"]); compatName != "" {
+			return strings.ToLower(compatName)
+		}
+	}
+	return strings.ToLower(strings.TrimSpace(auth.Provider))
+}
+
+func openAICompatModelPoolKey(auth *Auth, requestedModel string) string {
+	base := strings.TrimSpace(thinking.ParseSuffix(requestedModel).ModelName)
+	if base == "" {
+		base = strings.TrimSpace(requestedModel)
+	}
+	return strings.ToLower(strings.TrimSpace(auth.ID)) + "|" + openAICompatProviderKey(auth) + "|" + strings.ToLower(base)
+}
+
+func (m *Manager) nextModelPoolOffset(key string, size int) int {
+	if m == nil || size <= 1 {
+		return 0
+	}
+	key = strings.TrimSpace(key)
+	if key == "" {
+		return 0
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if m.modelPoolOffsets == nil {
+		m.modelPoolOffsets = make(map[string]int)
+	}
+	offset := m.modelPoolOffsets[key]
+	if offset >= 2_147_483_640 {
+		offset = 0
+	}
+	m.modelPoolOffsets[key] = offset + 1
+	if size <= 0 {
+		return 0
+	}
+	return offset % size
+}
+
+func rotateStrings(values []string, offset int) []string {
+	if len(values) <= 1 {
+		return values
+	}
+	if offset <= 0 {
+		out := make([]string, len(values))
+		copy(out, values)
+		return out
+	}
+	offset = offset % len(values)
+	out := make([]string, 0, len(values))
+	out = append(out, values[offset:]...)
+	out = append(out, values[:offset]...)
+	return out
+}
+
+func (m *Manager) resolveOpenAICompatUpstreamModelPool(auth *Auth, requestedModel string) []string {
+	if m == nil || !isOpenAICompatAPIKeyAuth(auth) {
+		return nil
+	}
+	requestedModel = strings.TrimSpace(requestedModel)
+	if requestedModel == "" {
+		return nil
+	}
+	cfg, _ := m.runtimeConfig.Load().(*internalconfig.Config)
+	if cfg == nil {
+		cfg = &internalconfig.Config{}
+	}
+	providerKey := ""
+	compatName := ""
+	if auth.Attributes != nil {
+		providerKey = strings.TrimSpace(auth.Attributes["provider_key"])
+		compatName = strings.TrimSpace(auth.Attributes["compat_name"])
+	}
+	entry := resolveOpenAICompatConfig(cfg, providerKey, compatName, auth.Provider)
+	if entry == nil {
+		return nil
+	}
+	return resolveModelAliasPoolFromConfigModels(requestedModel, asModelAliasEntries(entry.Models))
+}
+
+func preserveRequestedModelSuffix(requestedModel, resolved string) string {
+	return preserveResolvedModelSuffix(resolved, thinking.ParseSuffix(requestedModel))
+}
+
+func (m *Manager) executionModelCandidates(auth *Auth, routeModel string) []string {
+	return m.prepareExecutionModels(auth, routeModel)
+}
+
+func (m *Manager) prepareExecutionModels(auth *Auth, routeModel string) []string {
+	requestedModel := rewriteModelForAuth(routeModel, auth)
+	requestedModel = m.applyOAuthModelAlias(auth, requestedModel)
+	if pool := m.resolveOpenAICompatUpstreamModelPool(auth, requestedModel); len(pool) > 0 {
+		if len(pool) == 1 {
+			return pool
+		}
+		offset := m.nextModelPoolOffset(openAICompatModelPoolKey(auth, requestedModel), len(pool))
+		return rotateStrings(pool, offset)
+	}
+	resolved := m.applyAPIKeyModelAlias(auth, requestedModel)
+	if strings.TrimSpace(resolved) == "" {
+		resolved = requestedModel
+	}
+	return []string{resolved}
+}
+
+func discardStreamChunks(ch <-chan cliproxyexecutor.StreamChunk) {
+	if ch == nil {
+		return
+	}
+	go func() {
+		for range ch {
+		}
+	}()
+}
+
+func readStreamBootstrap(ctx context.Context, ch <-chan cliproxyexecutor.StreamChunk) ([]cliproxyexecutor.StreamChunk, bool, error) {
+	if ch == nil {
+		return nil, true, nil
+	}
+	buffered := make([]cliproxyexecutor.StreamChunk, 0, 1)
+	for {
+		var (
+			chunk cliproxyexecutor.StreamChunk
+			ok    bool
+		)
+		if ctx != nil {
+			select {
+			case <-ctx.Done():
+				return nil, false, ctx.Err()
+			case chunk, ok = <-ch:
+			}
+		} else {
+			chunk, ok = <-ch
+		}
+		if !ok {
+			return buffered, true, nil
+		}
+		if chunk.Err != nil {
+			return nil, false, chunk.Err
+		}
+		buffered = append(buffered, chunk)
+		if len(chunk.Payload) > 0 {
+			return buffered, false, nil
+		}
+	}
+}
+
+func (m *Manager) wrapStreamResult(ctx context.Context, auth *Auth, provider, routeModel string, headers http.Header, buffered []cliproxyexecutor.StreamChunk, remaining <-chan cliproxyexecutor.StreamChunk) *cliproxyexecutor.StreamResult {
+	out := make(chan cliproxyexecutor.StreamChunk)
+	go func() {
+		defer close(out)
+		var failed bool
+		forward := true
+		emit := func(chunk cliproxyexecutor.StreamChunk) bool {
+			if chunk.Err != nil && !failed {
+				failed = true
+				rerr := &Error{Message: chunk.Err.Error()}
+				if se, ok := errors.AsType[cliproxyexecutor.StatusError](chunk.Err); ok && se != nil {
+					rerr.HTTPStatus = se.StatusCode()
+				}
+				m.MarkResult(ctx, Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr})
+			}
+			if !forward {
+				return false
+			}
+			if ctx == nil {
+				out <- chunk
+				return true
+			}
+			select {
+			case <-ctx.Done():
+				forward = false
+				return false
+			case out <- chunk:
+				return true
+			}
+		}
+		for _, chunk := range buffered {
+			if ok := emit(chunk); !ok {
+				discardStreamChunks(remaining)
+				return
+			}
+		}
+		for chunk := range remaining {
+			if ok := emit(chunk); !ok {
+				discardStreamChunks(remaining)
+				return
+			}
+		}
+		if !failed {
+			m.MarkResult(ctx, Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: true})
+		}
+	}()
+	return &cliproxyexecutor.StreamResult{Headers: headers, Chunks: out}
+}
+
+func (m *Manager) executeStreamWithModelPool(ctx context.Context, executor ProviderExecutor, auth *Auth, provider string, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, routeModel string) (*cliproxyexecutor.StreamResult, error) {
+	if executor == nil {
+		return nil, &Error{Code: "executor_not_found", Message: "executor not registered"}
+	}
+	execModels := m.prepareExecutionModels(auth, routeModel)
+	var lastErr error
+	for idx, execModel := range execModels {
+		execReq := req
+		execReq.Model = execModel
+		streamResult, errStream := executor.ExecuteStream(ctx, auth, execReq, opts)
+		if errStream != nil {
+			if errCtx := ctx.Err(); errCtx != nil {
+				return nil, errCtx
+			}
+			rerr := &Error{Message: errStream.Error()}
+			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errStream); ok && se != nil {
+				rerr.HTTPStatus = se.StatusCode()
+			}
+			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr}
+			result.RetryAfter = retryAfterFromError(errStream)
+			m.MarkResult(ctx, result)
+			if isRequestInvalidError(errStream) {
+				return nil, errStream
+			}
+			lastErr = errStream
+			continue
+		}
+
+		buffered, closed, bootstrapErr := readStreamBootstrap(ctx, streamResult.Chunks)
+		if bootstrapErr != nil {
+			if errCtx := ctx.Err(); errCtx != nil {
+				discardStreamChunks(streamResult.Chunks)
+				return nil, errCtx
+			}
+			if isRequestInvalidError(bootstrapErr) {
+				rerr := &Error{Message: bootstrapErr.Error()}
+				if se, ok := errors.AsType[cliproxyexecutor.StatusError](bootstrapErr); ok && se != nil {
+					rerr.HTTPStatus = se.StatusCode()
+				}
+				result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr}
+				result.RetryAfter = retryAfterFromError(bootstrapErr)
+				m.MarkResult(ctx, result)
+				discardStreamChunks(streamResult.Chunks)
+				return nil, bootstrapErr
+			}
+			if idx < len(execModels)-1 {
+				rerr := &Error{Message: bootstrapErr.Error()}
+				if se, ok := errors.AsType[cliproxyexecutor.StatusError](bootstrapErr); ok && se != nil {
+					rerr.HTTPStatus = se.StatusCode()
+				}
+				result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr}
+				result.RetryAfter = retryAfterFromError(bootstrapErr)
+				m.MarkResult(ctx, result)
+				discardStreamChunks(streamResult.Chunks)
+				lastErr = bootstrapErr
+				continue
+			}
+			errCh := make(chan cliproxyexecutor.StreamChunk, 1)
+			errCh <- cliproxyexecutor.StreamChunk{Err: bootstrapErr}
+			close(errCh)
+			return m.wrapStreamResult(ctx, auth.Clone(), provider, routeModel, streamResult.Headers, nil, errCh), nil
+		}
+
+		if closed && len(buffered) == 0 {
+			emptyErr := &Error{Code: "empty_stream", Message: "upstream stream closed before first payload", Retryable: true}
+			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: emptyErr}
+			m.MarkResult(ctx, result)
+			if idx < len(execModels)-1 {
+				lastErr = emptyErr
+				continue
+			}
+			errCh := make(chan cliproxyexecutor.StreamChunk, 1)
+			errCh <- cliproxyexecutor.StreamChunk{Err: emptyErr}
+			close(errCh)
+			return m.wrapStreamResult(ctx, auth.Clone(), provider, routeModel, streamResult.Headers, nil, errCh), nil
+		}
+
+		remaining := streamResult.Chunks
+		if closed {
+			closedCh := make(chan cliproxyexecutor.StreamChunk)
+			close(closedCh)
+			remaining = closedCh
+		}
+		return m.wrapStreamResult(ctx, auth.Clone(), provider, routeModel, streamResult.Headers, buffered, remaining), nil
+	}
+	if lastErr == nil {
+		lastErr = &Error{Code: "auth_not_found", Message: "no upstream model available"}
+	}
+	return nil, lastErr
 }
 
 func (m *Manager) rebuildAPIKeyModelAliasFromRuntimeConfig() {
@@ -448,10 +808,14 @@ func (m *Manager) Register(ctx context.Context, auth *Auth) (*Auth, error) {
 		auth.ID = uuid.NewString()
 	}
 	auth.EnsureIndex()
+	authClone := auth.Clone()
 	m.mu.Lock()
-	m.auths[auth.ID] = auth.Clone()
+	m.auths[auth.ID] = authClone
 	m.mu.Unlock()
 	m.rebuildAPIKeyModelAliasFromRuntimeConfig()
+	if m.scheduler != nil {
+		m.scheduler.upsertAuth(authClone)
+	}
 	_ = m.persist(ctx, auth)
 	m.hook.OnAuthRegistered(ctx, auth.Clone())
 	return auth.Clone(), nil
@@ -463,14 +827,23 @@ func (m *Manager) Update(ctx context.Context, auth *Auth) (*Auth, error) {
 		return nil, nil
 	}
 	m.mu.Lock()
-	if existing, ok := m.auths[auth.ID]; ok && existing != nil && !auth.indexAssigned && auth.Index == "" {
-		auth.Index = existing.Index
-		auth.indexAssigned = existing.indexAssigned
+	if existing, ok := m.auths[auth.ID]; ok && existing != nil {
+		if !auth.indexAssigned && auth.Index == "" {
+			auth.Index = existing.Index
+			auth.indexAssigned = existing.indexAssigned
+		}
+		if len(auth.ModelStates) == 0 && len(existing.ModelStates) > 0 {
+			auth.ModelStates = existing.ModelStates
+		}
 	}
 	auth.EnsureIndex()
-	m.auths[auth.ID] = auth.Clone()
+	authClone := auth.Clone()
+	m.auths[auth.ID] = authClone
 	m.mu.Unlock()
 	m.rebuildAPIKeyModelAliasFromRuntimeConfig()
+	if m.scheduler != nil {
+		m.scheduler.upsertAuth(authClone)
+	}
 	_ = m.persist(ctx, auth)
 	m.hook.OnAuthUpdated(ctx, auth.Clone())
 	return auth.Clone(), nil
@@ -479,12 +852,13 @@ func (m *Manager) Update(ctx context.Context, auth *Auth) (*Auth, error) {
 // Load resets manager state from the backing store.
 func (m *Manager) Load(ctx context.Context) error {
 	m.mu.Lock()
-	defer m.mu.Unlock()
 	if m.store == nil {
+		m.mu.Unlock()
 		return nil
 	}
 	items, err := m.store.List(ctx)
 	if err != nil {
+		m.mu.Unlock()
 		return err
 	}
 	m.auths = make(map[string]*Auth, len(items))
@@ -500,6 +874,8 @@ func (m *Manager) Load(ctx context.Context) error {
 		cfg = &internalconfig.Config{}
 	}
 	m.rebuildAPIKeyModelAliasLocked(cfg)
+	m.mu.Unlock()
+	m.syncScheduler()
 	return nil
 }
 
@@ -629,32 +1005,42 @@ func (m *Manager) executeMixedOnce(ctx context.Context, providers []string, req
 			execCtx = context.WithValue(execCtx, roundTripperContextKey{}, rt)
 			execCtx = context.WithValue(execCtx, "cliproxy.roundtripper", rt)
 		}
-		execReq := req
-		execReq.Model = rewriteModelForAuth(routeModel, auth)
-		execReq.Model = m.applyOAuthModelAlias(auth, execReq.Model)
-		execReq.Model = m.applyAPIKeyModelAlias(auth, execReq.Model)
-		resp, errExec := executor.Execute(execCtx, auth, execReq, opts)
-		result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: errExec == nil}
-		if errExec != nil {
-			if errCtx := execCtx.Err(); errCtx != nil {
-				return cliproxyexecutor.Response{}, errCtx
-			}
-			result.Error = &Error{Message: errExec.Error()}
-			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
-				result.Error.HTTPStatus = se.StatusCode()
-			}
-			if ra := retryAfterFromError(errExec); ra != nil {
-				result.RetryAfter = ra
+
+		models := m.prepareExecutionModels(auth, routeModel)
+		var authErr error
+		for _, upstreamModel := range models {
+			execReq := req
+			execReq.Model = upstreamModel
+			resp, errExec := executor.Execute(execCtx, auth, execReq, opts)
+			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: errExec == nil}
+			if errExec != nil {
+				if errCtx := execCtx.Err(); errCtx != nil {
+					return cliproxyexecutor.Response{}, errCtx
+				}
+				result.Error = &Error{Message: errExec.Error()}
+				if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
+					result.Error.HTTPStatus = se.StatusCode()
+				}
+				if ra := retryAfterFromError(errExec); ra != nil {
+					result.RetryAfter = ra
+				}
+				m.MarkResult(execCtx, result)
+				if isRequestInvalidError(errExec) {
+					return cliproxyexecutor.Response{}, errExec
+				}
+				authErr = errExec
+				continue
 			}
 			m.MarkResult(execCtx, result)
-			if isRequestInvalidError(errExec) {
-				return cliproxyexecutor.Response{}, errExec
+			return resp, nil
+		}
+		if authErr != nil {
+			if isRequestInvalidError(authErr) {
+				return cliproxyexecutor.Response{}, authErr
 			}
-			lastErr = errExec
+			lastErr = authErr
 			continue
 		}
-		m.MarkResult(execCtx, result)
-		return resp, nil
 	}
 }
 
@@ -691,32 +1077,42 @@ func (m *Manager) executeCountMixedOnce(ctx context.Context, providers []string,
 			execCtx = context.WithValue(execCtx, roundTripperContextKey{}, rt)
 			execCtx = context.WithValue(execCtx, "cliproxy.roundtripper", rt)
 		}
-		execReq := req
-		execReq.Model = rewriteModelForAuth(routeModel, auth)
-		execReq.Model = m.applyOAuthModelAlias(auth, execReq.Model)
-		execReq.Model = m.applyAPIKeyModelAlias(auth, execReq.Model)
-		resp, errExec := executor.CountTokens(execCtx, auth, execReq, opts)
-		result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: errExec == nil}
-		if errExec != nil {
-			if errCtx := execCtx.Err(); errCtx != nil {
-				return cliproxyexecutor.Response{}, errCtx
-			}
-			result.Error = &Error{Message: errExec.Error()}
-			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
-				result.Error.HTTPStatus = se.StatusCode()
-			}
-			if ra := retryAfterFromError(errExec); ra != nil {
-				result.RetryAfter = ra
+
+		models := m.prepareExecutionModels(auth, routeModel)
+		var authErr error
+		for _, upstreamModel := range models {
+			execReq := req
+			execReq.Model = upstreamModel
+			resp, errExec := executor.CountTokens(execCtx, auth, execReq, opts)
+			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: errExec == nil}
+			if errExec != nil {
+				if errCtx := execCtx.Err(); errCtx != nil {
+					return cliproxyexecutor.Response{}, errCtx
+				}
+				result.Error = &Error{Message: errExec.Error()}
+				if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
+					result.Error.HTTPStatus = se.StatusCode()
+				}
+				if ra := retryAfterFromError(errExec); ra != nil {
+					result.RetryAfter = ra
+				}
+				m.hook.OnResult(execCtx, result)
+				if isRequestInvalidError(errExec) {
+					return cliproxyexecutor.Response{}, errExec
+				}
+				authErr = errExec
+				continue
 			}
 			m.hook.OnResult(execCtx, result)
-			if isRequestInvalidError(errExec) {
-				return cliproxyexecutor.Response{}, errExec
+			return resp, nil
+		}
+		if authErr != nil {
+			if isRequestInvalidError(authErr) {
+				return cliproxyexecutor.Response{}, authErr
 			}
-			lastErr = errExec
+			lastErr = authErr
 			continue
 		}
-		m.hook.OnResult(execCtx, result)
-		return resp, nil
 	}
 }
 
@@ -753,63 +1149,18 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 			execCtx = context.WithValue(execCtx, roundTripperContextKey{}, rt)
 			execCtx = context.WithValue(execCtx, "cliproxy.roundtripper", rt)
 		}
-		execReq := req
-		execReq.Model = rewriteModelForAuth(routeModel, auth)
-		execReq.Model = m.applyOAuthModelAlias(auth, execReq.Model)
-		execReq.Model = m.applyAPIKeyModelAlias(auth, execReq.Model)
-		streamResult, errStream := executor.ExecuteStream(execCtx, auth, execReq, opts)
+		streamResult, errStream := m.executeStreamWithModelPool(execCtx, executor, auth, provider, req, opts, routeModel)
 		if errStream != nil {
 			if errCtx := execCtx.Err(); errCtx != nil {
 				return nil, errCtx
 			}
-			rerr := &Error{Message: errStream.Error()}
-			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errStream); ok && se != nil {
-				rerr.HTTPStatus = se.StatusCode()
-			}
-			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr}
-			result.RetryAfter = retryAfterFromError(errStream)
-			m.MarkResult(execCtx, result)
 			if isRequestInvalidError(errStream) {
 				return nil, errStream
 			}
 			lastErr = errStream
 			continue
 		}
-		out := make(chan cliproxyexecutor.StreamChunk)
-		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
-					rerr := &Error{Message: chunk.Err.Error()}
-					if se, ok := errors.AsType[cliproxyexecutor.StatusError](chunk.Err); ok && se != nil {
-						rerr.HTTPStatus = se.StatusCode()
-					}
-					m.MarkResult(streamCtx, Result{AuthID: streamAuth.ID, Provider: streamProvider, Model: routeModel, Success: false, Error: rerr})
-				}
-				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})
-			}
-		}(execCtx, auth.Clone(), provider, streamResult.Chunks)
-		return &cliproxyexecutor.StreamResult{
-			Headers: streamResult.Headers,
-			Chunks:  out,
-		}, nil
+		return streamResult, nil
 	}
 }
 
@@ -1240,6 +1591,7 @@ func (m *Manager) MarkResult(ctx context.Context, result Result) {
 	suspendReason := ""
 	clearModelQuota := false
 	setModelQuota := false
+	var authSnapshot *Auth
 
 	m.mu.Lock()
 	if auth, ok := m.auths[result.AuthID]; ok && auth != nil {
@@ -1333,8 +1685,12 @@ func (m *Manager) MarkResult(ctx context.Context, result Result) {
 		}
 
 		_ = m.persist(ctx, auth)
+		authSnapshot = auth.Clone()
 	}
 	m.mu.Unlock()
+	if m.scheduler != nil && authSnapshot != nil {
+		m.scheduler.upsertAuth(authSnapshot)
+	}
 
 	if clearModelQuota && result.Model != "" {
 		registry.GetGlobalRegistry().ClearModelQuotaExceeded(result.AuthID, result.Model)
@@ -1528,18 +1884,22 @@ func statusCodeFromResult(err *Error) int {
 }
 
 // isRequestInvalidError returns true if the error represents a client request
-// error that should not be retried. Specifically, it checks for 400 Bad Request
-// with "invalid_request_error" in the message, indicating the request itself is
-// malformed and switching to a different auth will not help.
+// error that should not be retried. Specifically, it treats 400 responses with
+// "invalid_request_error" and all 422 responses as request-shape failures,
+// where switching auths or pooled upstream models will not help.
 func isRequestInvalidError(err error) bool {
 	if err == nil {
 		return false
 	}
 	status := statusCodeFromError(err)
-	if status != http.StatusBadRequest {
+	switch status {
+	case http.StatusBadRequest:
+		return strings.Contains(err.Error(), "invalid_request_error")
+	case http.StatusUnprocessableEntity:
+		return true
+	default:
 		return false
 	}
-	return strings.Contains(err.Error(), "invalid_request_error")
 }
 
 func applyAuthFailureState(auth *Auth, resultErr *Error, retryAfter *time.Duration, now time.Time) {
@@ -1687,7 +2047,29 @@ func (m *Manager) CloseExecutionSession(sessionID string) {
 	}
 }
 
-func (m *Manager) pickNext(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, error) {
+func (m *Manager) useSchedulerFastPath() bool {
+	if m == nil || m.scheduler == nil {
+		return false
+	}
+	return isBuiltInSelector(m.selector)
+}
+
+func shouldRetrySchedulerPick(err error) bool {
+	if err == nil {
+		return false
+	}
+	var cooldownErr *modelCooldownError
+	if errors.As(err, &cooldownErr) {
+		return true
+	}
+	var authErr *Error
+	if !errors.As(err, &authErr) || authErr == nil {
+		return false
+	}
+	return authErr.Code == "auth_not_found" || authErr.Code == "auth_unavailable"
+}
+
+func (m *Manager) pickNextLegacy(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, error) {
 	pinnedAuthID := pinnedAuthIDFromMetadata(opts.Metadata)
 
 	m.mu.RLock()
@@ -1747,7 +2129,38 @@ func (m *Manager) pickNext(ctx context.Context, provider, model string, opts cli
 	return authCopy, executor, nil
 }
 
-func (m *Manager) pickNextMixed(ctx context.Context, providers []string, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, string, error) {
+func (m *Manager) pickNext(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, error) {
+	if !m.useSchedulerFastPath() {
+		return m.pickNextLegacy(ctx, provider, model, opts, tried)
+	}
+	executor, okExecutor := m.Executor(provider)
+	if !okExecutor {
+		return nil, nil, &Error{Code: "executor_not_found", Message: "executor not registered"}
+	}
+	selected, errPick := m.scheduler.pickSingle(ctx, provider, model, opts, tried)
+	if errPick != nil && model != "" && shouldRetrySchedulerPick(errPick) {
+		m.syncScheduler()
+		selected, errPick = m.scheduler.pickSingle(ctx, provider, model, opts, tried)
+	}
+	if errPick != nil {
+		return nil, nil, errPick
+	}
+	if selected == nil {
+		return nil, nil, &Error{Code: "auth_not_found", Message: "selector returned no auth"}
+	}
+	authCopy := selected.Clone()
+	if !selected.indexAssigned {
+		m.mu.Lock()
+		if current := m.auths[authCopy.ID]; current != nil && !current.indexAssigned {
+			current.EnsureIndex()
+			authCopy = current.Clone()
+		}
+		m.mu.Unlock()
+	}
+	return authCopy, executor, nil
+}
+
+func (m *Manager) pickNextMixedLegacy(ctx context.Context, providers []string, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, string, error) {
 	pinnedAuthID := pinnedAuthIDFromMetadata(opts.Metadata)
 
 	providerSet := make(map[string]struct{}, len(providers))
@@ -1830,6 +2243,58 @@ func (m *Manager) pickNextMixed(ctx context.Context, providers []string, model s
 	return authCopy, executor, providerKey, nil
 }
 
+func (m *Manager) pickNextMixed(ctx context.Context, providers []string, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, string, error) {
+	if !m.useSchedulerFastPath() {
+		return m.pickNextMixedLegacy(ctx, providers, model, opts, tried)
+	}
+
+	eligibleProviders := make([]string, 0, len(providers))
+	seenProviders := make(map[string]struct{}, len(providers))
+	for _, provider := range providers {
+		providerKey := strings.TrimSpace(strings.ToLower(provider))
+		if providerKey == "" {
+			continue
+		}
+		if _, seen := seenProviders[providerKey]; seen {
+			continue
+		}
+		if _, okExecutor := m.Executor(providerKey); !okExecutor {
+			continue
+		}
+		seenProviders[providerKey] = struct{}{}
+		eligibleProviders = append(eligibleProviders, providerKey)
+	}
+	if len(eligibleProviders) == 0 {
+		return nil, nil, "", &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+
+	selected, providerKey, errPick := m.scheduler.pickMixed(ctx, eligibleProviders, model, opts, tried)
+	if errPick != nil && model != "" && shouldRetrySchedulerPick(errPick) {
+		m.syncScheduler()
+		selected, providerKey, errPick = m.scheduler.pickMixed(ctx, eligibleProviders, model, opts, tried)
+	}
+	if errPick != nil {
+		return nil, nil, "", errPick
+	}
+	if selected == nil {
+		return nil, nil, "", &Error{Code: "auth_not_found", Message: "selector returned no auth"}
+	}
+	executor, okExecutor := m.Executor(providerKey)
+	if !okExecutor {
+		return nil, nil, "", &Error{Code: "executor_not_found", Message: "executor not registered"}
+	}
+	authCopy := selected.Clone()
+	if !selected.indexAssigned {
+		m.mu.Lock()
+		if current := m.auths[authCopy.ID]; current != nil && !current.indexAssigned {
+			current.EnsureIndex()
+			authCopy = current.Clone()
+		}
+		m.mu.Unlock()
+	}
+	return authCopy, executor, providerKey, nil
+}
+
 func (m *Manager) persist(ctx context.Context, auth *Auth) error {
 	if m.store == nil || auth == nil {
 		return nil
@@ -2181,6 +2646,9 @@ func (m *Manager) refreshAuth(ctx context.Context, id string) {
 			current.NextRefreshAfter = now.Add(refreshFailureBackoff)
 			current.LastError = &Error{Message: err.Error()}
 			m.auths[id] = current
+			if m.scheduler != nil {
+				m.scheduler.upsertAuth(current.Clone())
+			}
 		}
 		m.mu.Unlock()
 		return

sdk/cliproxy/auth/conductor_scheduler_refresh_test.go

@@ -0,0 +1,163 @@
+package auth
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+)
+
+type schedulerProviderTestExecutor struct {
+	provider string
+}
+
+func (e schedulerProviderTestExecutor) Identifier() string { return e.provider }
+
+func (e schedulerProviderTestExecutor) Execute(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (e schedulerProviderTestExecutor) ExecuteStream(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
+	return nil, nil
+}
+
+func (e schedulerProviderTestExecutor) Refresh(ctx context.Context, auth *Auth) (*Auth, error) {
+	return auth, nil
+}
+
+func (e schedulerProviderTestExecutor) CountTokens(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (e schedulerProviderTestExecutor) HttpRequest(ctx context.Context, auth *Auth, req *http.Request) (*http.Response, error) {
+	return nil, nil
+}
+
+func TestManager_RefreshSchedulerEntry_RebuildsSupportedModelSetAfterModelRegistration(t *testing.T) {
+	ctx := context.Background()
+
+	testCases := []struct {
+		name  string
+		prime func(*Manager, *Auth) error
+	}{
+		{
+			name: "register",
+			prime: func(manager *Manager, auth *Auth) error {
+				_, errRegister := manager.Register(ctx, auth)
+				return errRegister
+			},
+		},
+		{
+			name: "update",
+			prime: func(manager *Manager, auth *Auth) error {
+				_, errRegister := manager.Register(ctx, auth)
+				if errRegister != nil {
+					return errRegister
+				}
+				updated := auth.Clone()
+				updated.Metadata = map[string]any{"updated": true}
+				_, errUpdate := manager.Update(ctx, updated)
+				return errUpdate
+			},
+		},
+	}
+
+	for _, testCase := range testCases {
+		testCase := testCase
+		t.Run(testCase.name, func(t *testing.T) {
+			manager := NewManager(nil, &RoundRobinSelector{}, nil)
+			auth := &Auth{
+				ID:       "refresh-entry-" + testCase.name,
+				Provider: "gemini",
+			}
+			if errPrime := testCase.prime(manager, auth); errPrime != nil {
+				t.Fatalf("prime auth %s: %v", testCase.name, errPrime)
+			}
+
+			registerSchedulerModels(t, "gemini", "scheduler-refresh-model", auth.ID)
+
+			got, errPick := manager.scheduler.pickSingle(ctx, "gemini", "scheduler-refresh-model", cliproxyexecutor.Options{}, nil)
+			var authErr *Error
+			if !errors.As(errPick, &authErr) || authErr == nil {
+				t.Fatalf("pickSingle() before refresh error = %v, want auth_not_found", errPick)
+			}
+			if authErr.Code != "auth_not_found" {
+				t.Fatalf("pickSingle() before refresh code = %q, want %q", authErr.Code, "auth_not_found")
+			}
+			if got != nil {
+				t.Fatalf("pickSingle() before refresh auth = %v, want nil", got)
+			}
+
+			manager.RefreshSchedulerEntry(auth.ID)
+
+			got, errPick = manager.scheduler.pickSingle(ctx, "gemini", "scheduler-refresh-model", cliproxyexecutor.Options{}, nil)
+			if errPick != nil {
+				t.Fatalf("pickSingle() after refresh error = %v", errPick)
+			}
+			if got == nil || got.ID != auth.ID {
+				t.Fatalf("pickSingle() after refresh auth = %v, want %q", got, auth.ID)
+			}
+		})
+	}
+}
+
+func TestManager_PickNext_RebuildsSchedulerAfterModelCooldownError(t *testing.T) {
+	ctx := context.Background()
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	manager.RegisterExecutor(schedulerProviderTestExecutor{provider: "gemini"})
+
+	registerSchedulerModels(t, "gemini", "scheduler-cooldown-rebuild-model", "cooldown-stale-old")
+
+	oldAuth := &Auth{
+		ID:       "cooldown-stale-old",
+		Provider: "gemini",
+	}
+	if _, errRegister := manager.Register(ctx, oldAuth); errRegister != nil {
+		t.Fatalf("register old auth: %v", errRegister)
+	}
+
+	manager.MarkResult(ctx, Result{
+		AuthID:   oldAuth.ID,
+		Provider: "gemini",
+		Model:    "scheduler-cooldown-rebuild-model",
+		Success:  false,
+		Error:    &Error{HTTPStatus: http.StatusTooManyRequests, Message: "quota"},
+	})
+
+	newAuth := &Auth{
+		ID:       "cooldown-stale-new",
+		Provider: "gemini",
+	}
+	if _, errRegister := manager.Register(ctx, newAuth); errRegister != nil {
+		t.Fatalf("register new auth: %v", errRegister)
+	}
+
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient(newAuth.ID, "gemini", []*registry.ModelInfo{{ID: "scheduler-cooldown-rebuild-model"}})
+	t.Cleanup(func() {
+		reg.UnregisterClient(newAuth.ID)
+	})
+
+	got, errPick := manager.scheduler.pickSingle(ctx, "gemini", "scheduler-cooldown-rebuild-model", cliproxyexecutor.Options{}, nil)
+	var cooldownErr *modelCooldownError
+	if !errors.As(errPick, &cooldownErr) {
+		t.Fatalf("pickSingle() before sync error = %v, want modelCooldownError", errPick)
+	}
+	if got != nil {
+		t.Fatalf("pickSingle() before sync auth = %v, want nil", got)
+	}
+
+	got, executor, errPick := manager.pickNext(ctx, "gemini", "scheduler-cooldown-rebuild-model", cliproxyexecutor.Options{}, nil)
+	if errPick != nil {
+		t.Fatalf("pickNext() error = %v", errPick)
+	}
+	if executor == nil {
+		t.Fatal("pickNext() executor = nil")
+	}
+	if got == nil || got.ID != newAuth.ID {
+		t.Fatalf("pickNext() auth = %v, want %q", got, newAuth.ID)
+	}
+}

sdk/cliproxy/auth/conductor_update_test.go

@@ -0,0 +1,49 @@
+package auth
+
+import (
+	"context"
+	"testing"
+)
+
+func TestManager_Update_PreservesModelStates(t *testing.T) {
+	m := NewManager(nil, nil, nil)
+
+	model := "test-model"
+	backoffLevel := 7
+
+	if _, errRegister := m.Register(context.Background(), &Auth{
+		ID:       "auth-1",
+		Provider: "claude",
+		Metadata: map[string]any{"k": "v"},
+		ModelStates: map[string]*ModelState{
+			model: {
+				Quota: QuotaState{BackoffLevel: backoffLevel},
+			},
+		},
+	}); errRegister != nil {
+		t.Fatalf("register auth: %v", errRegister)
+	}
+
+	if _, errUpdate := m.Update(context.Background(), &Auth{
+		ID:       "auth-1",
+		Provider: "claude",
+		Metadata: map[string]any{"k": "v2"},
+	}); errUpdate != nil {
+		t.Fatalf("update auth: %v", errUpdate)
+	}
+
+	updated, ok := m.GetByID("auth-1")
+	if !ok || updated == nil {
+		t.Fatalf("expected auth to be present")
+	}
+	if len(updated.ModelStates) == 0 {
+		t.Fatalf("expected ModelStates to be preserved")
+	}
+	state := updated.ModelStates[model]
+	if state == nil {
+		t.Fatalf("expected model state to be present")
+	}
+	if state.Quota.BackoffLevel != backoffLevel {
+		t.Fatalf("expected BackoffLevel to be %d, got %d", backoffLevel, state.Quota.BackoffLevel)
+	}
+}

sdk/cliproxy/auth/oauth_model_alias.go

@@ -80,54 +80,98 @@ func (m *Manager) applyOAuthModelAlias(auth *Auth, requestedModel string) string
 	return upstreamModel
 }
 
-func resolveModelAliasFromConfigModels(requestedModel string, models []modelAliasEntry) string {
+func modelAliasLookupCandidates(requestedModel string) (thinking.SuffixResult, []string) {
 	requestedModel = strings.TrimSpace(requestedModel)
 	if requestedModel == "" {
-		return ""
+		return thinking.SuffixResult{}, nil
 	}
-	if len(models) == 0 {
-		return ""
-	}
-
 	requestResult := thinking.ParseSuffix(requestedModel)
 	base := requestResult.ModelName
+	if base == "" {
+		base = requestedModel
+	}
 	candidates := []string{base}
 	if base != requestedModel {
 		candidates = append(candidates, requestedModel)
 	}
+	return requestResult, candidates
+}
 
-	preserveSuffix := func(resolved string) string {
-		resolved = strings.TrimSpace(resolved)
-		if resolved == "" {
-			return ""
-		}
-		if thinking.ParseSuffix(resolved).HasSuffix {
-			return resolved
-		}
-		if requestResult.HasSuffix && requestResult.RawSuffix != "" {
-			return resolved + "(" + requestResult.RawSuffix + ")"
-		}
+func preserveResolvedModelSuffix(resolved string, requestResult thinking.SuffixResult) string {
+	resolved = strings.TrimSpace(resolved)
+	if resolved == "" {
+		return ""
+	}
+	if thinking.ParseSuffix(resolved).HasSuffix {
 		return resolved
 	}
+	if requestResult.HasSuffix && requestResult.RawSuffix != "" {
+		return resolved + "(" + requestResult.RawSuffix + ")"
+	}
+	return resolved
+}
 
+func resolveModelAliasPoolFromConfigModels(requestedModel string, models []modelAliasEntry) []string {
+	requestedModel = strings.TrimSpace(requestedModel)
+	if requestedModel == "" {
+		return nil
+	}
+	if len(models) == 0 {
+		return nil
+	}
+
+	requestResult, candidates := modelAliasLookupCandidates(requestedModel)
+	if len(candidates) == 0 {
+		return nil
+	}
+
+	out := make([]string, 0)
+	seen := make(map[string]struct{})
 	for i := range models {
 		name := strings.TrimSpace(models[i].GetName())
 		alias := strings.TrimSpace(models[i].GetAlias())
 		for _, candidate := range candidates {
-			if candidate == "" {
+			if candidate == "" || alias == "" || !strings.EqualFold(alias, candidate) {
 				continue
 			}
-			if alias != "" && strings.EqualFold(alias, candidate) {
-				if name != "" {
-					return preserveSuffix(name)
-				}
-				return preserveSuffix(candidate)
+			resolved := candidate
+			if name != "" {
+				resolved = name
 			}
-			if name != "" && strings.EqualFold(name, candidate) {
-				return preserveSuffix(name)
+			resolved = preserveResolvedModelSuffix(resolved, requestResult)
+			key := strings.ToLower(strings.TrimSpace(resolved))
+			if key == "" {
+				break
 			}
+			if _, exists := seen[key]; exists {
+				break
+			}
+			seen[key] = struct{}{}
+			out = append(out, resolved)
+			break
 		}
 	}
+	if len(out) > 0 {
+		return out
+	}
+
+	for i := range models {
+		name := strings.TrimSpace(models[i].GetName())
+		for _, candidate := range candidates {
+			if candidate == "" || name == "" || !strings.EqualFold(name, candidate) {
+				continue
+			}
+			return []string{preserveResolvedModelSuffix(name, requestResult)}
+		}
+	}
+	return nil
+}
+
+func resolveModelAliasFromConfigModels(requestedModel string, models []modelAliasEntry) string {
+	resolved := resolveModelAliasPoolFromConfigModels(requestedModel, models)
+	if len(resolved) > 0 {
+		return resolved[0]
+	}
 	return ""
 }
 

sdk/cliproxy/auth/openai_compat_pool_test.go

@@ -0,0 +1,419 @@
+package auth
+
+import (
+	"context"
+	"net/http"
+	"sync"
+	"testing"
+
+	internalconfig "github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+)
+
+type openAICompatPoolExecutor struct {
+	id string
+
+	mu                sync.Mutex
+	executeModels     []string
+	countModels       []string
+	streamModels      []string
+	executeErrors     map[string]error
+	countErrors       map[string]error
+	streamFirstErrors map[string]error
+	streamPayloads    map[string][]cliproxyexecutor.StreamChunk
+}
+
+func (e *openAICompatPoolExecutor) Identifier() string { return e.id }
+
+func (e *openAICompatPoolExecutor) Execute(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	_ = ctx
+	_ = auth
+	_ = opts
+	e.mu.Lock()
+	e.executeModels = append(e.executeModels, req.Model)
+	err := e.executeErrors[req.Model]
+	e.mu.Unlock()
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
+	return cliproxyexecutor.Response{Payload: []byte(req.Model)}, nil
+}
+
+func (e *openAICompatPoolExecutor) ExecuteStream(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
+	_ = ctx
+	_ = auth
+	_ = opts
+	e.mu.Lock()
+	e.streamModels = append(e.streamModels, req.Model)
+	err := e.streamFirstErrors[req.Model]
+	payloadChunks, hasCustomChunks := e.streamPayloads[req.Model]
+	chunks := append([]cliproxyexecutor.StreamChunk(nil), payloadChunks...)
+	e.mu.Unlock()
+	ch := make(chan cliproxyexecutor.StreamChunk, max(1, len(chunks)))
+	if err != nil {
+		ch <- cliproxyexecutor.StreamChunk{Err: err}
+		close(ch)
+		return &cliproxyexecutor.StreamResult{Headers: http.Header{"X-Model": {req.Model}}, Chunks: ch}, nil
+	}
+	if !hasCustomChunks {
+		ch <- cliproxyexecutor.StreamChunk{Payload: []byte(req.Model)}
+	} else {
+		for _, chunk := range chunks {
+			ch <- chunk
+		}
+	}
+	close(ch)
+	return &cliproxyexecutor.StreamResult{Headers: http.Header{"X-Model": {req.Model}}, Chunks: ch}, nil
+}
+
+func (e *openAICompatPoolExecutor) Refresh(_ context.Context, auth *Auth) (*Auth, error) {
+	return auth, nil
+}
+
+func (e *openAICompatPoolExecutor) CountTokens(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	_ = ctx
+	_ = auth
+	_ = opts
+	e.mu.Lock()
+	e.countModels = append(e.countModels, req.Model)
+	err := e.countErrors[req.Model]
+	e.mu.Unlock()
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
+	return cliproxyexecutor.Response{Payload: []byte(req.Model)}, nil
+}
+
+func (e *openAICompatPoolExecutor) HttpRequest(ctx context.Context, auth *Auth, req *http.Request) (*http.Response, error) {
+	_ = ctx
+	_ = auth
+	_ = req
+	return nil, &Error{HTTPStatus: http.StatusNotImplemented, Message: "HttpRequest not implemented"}
+}
+
+func (e *openAICompatPoolExecutor) ExecuteModels() []string {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	out := make([]string, len(e.executeModels))
+	copy(out, e.executeModels)
+	return out
+}
+
+func (e *openAICompatPoolExecutor) CountModels() []string {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	out := make([]string, len(e.countModels))
+	copy(out, e.countModels)
+	return out
+}
+
+func (e *openAICompatPoolExecutor) StreamModels() []string {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	out := make([]string, len(e.streamModels))
+	copy(out, e.streamModels)
+	return out
+}
+
+func newOpenAICompatPoolTestManager(t *testing.T, alias string, models []internalconfig.OpenAICompatibilityModel, executor *openAICompatPoolExecutor) *Manager {
+	t.Helper()
+	cfg := &internalconfig.Config{
+		OpenAICompatibility: []internalconfig.OpenAICompatibility{{
+			Name:   "pool",
+			Models: models,
+		}},
+	}
+	m := NewManager(nil, nil, nil)
+	m.SetConfig(cfg)
+	if executor == nil {
+		executor = &openAICompatPoolExecutor{id: "pool"}
+	}
+	m.RegisterExecutor(executor)
+
+	auth := &Auth{
+		ID:       "pool-auth-" + t.Name(),
+		Provider: "pool",
+		Status:   StatusActive,
+		Attributes: map[string]string{
+			"api_key":      "test-key",
+			"compat_name":  "pool",
+			"provider_key": "pool",
+		},
+	}
+	if _, err := m.Register(context.Background(), auth); err != nil {
+		t.Fatalf("register auth: %v", err)
+	}
+
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient(auth.ID, "pool", []*registry.ModelInfo{{ID: alias}})
+	t.Cleanup(func() {
+		reg.UnregisterClient(auth.ID)
+	})
+	return m
+}
+
+func TestManagerExecuteCount_OpenAICompatAliasPoolStopsOnInvalidRequest(t *testing.T) {
+	alias := "claude-opus-4.66"
+	invalidErr := &Error{HTTPStatus: http.StatusUnprocessableEntity, Message: "unprocessable entity"}
+	executor := &openAICompatPoolExecutor{
+		id:          "pool",
+		countErrors: map[string]error{"qwen3.5-plus": invalidErr},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	_, err := m.ExecuteCount(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err == nil || err.Error() != invalidErr.Error() {
+		t.Fatalf("execute count error = %v, want %v", err, invalidErr)
+	}
+	got := executor.CountModels()
+	if len(got) != 1 || got[0] != "qwen3.5-plus" {
+		t.Fatalf("count calls = %v, want only first invalid model", got)
+	}
+}
+func TestResolveModelAliasPoolFromConfigModels(t *testing.T) {
+	models := []modelAliasEntry{
+		internalconfig.OpenAICompatibilityModel{Name: "qwen3.5-plus", Alias: "claude-opus-4.66"},
+		internalconfig.OpenAICompatibilityModel{Name: "glm-5", Alias: "claude-opus-4.66"},
+		internalconfig.OpenAICompatibilityModel{Name: "kimi-k2.5", Alias: "claude-opus-4.66"},
+	}
+	got := resolveModelAliasPoolFromConfigModels("claude-opus-4.66(8192)", models)
+	want := []string{"qwen3.5-plus(8192)", "glm-5(8192)", "kimi-k2.5(8192)"}
+	if len(got) != len(want) {
+		t.Fatalf("pool len = %d, want %d (%v)", len(got), len(want), got)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("pool[%d] = %q, want %q", i, got[i], want[i])
+		}
+	}
+}
+
+func TestManagerExecute_OpenAICompatAliasPoolRotatesWithinAuth(t *testing.T) {
+	alias := "claude-opus-4.66"
+	executor := &openAICompatPoolExecutor{id: "pool"}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	for i := 0; i < 3; i++ {
+		resp, err := m.Execute(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+		if err != nil {
+			t.Fatalf("execute %d: %v", i, err)
+		}
+		if len(resp.Payload) == 0 {
+			t.Fatalf("execute %d returned empty payload", i)
+		}
+	}
+
+	got := executor.ExecuteModels()
+	want := []string{"qwen3.5-plus", "glm-5", "qwen3.5-plus"}
+	if len(got) != len(want) {
+		t.Fatalf("execute calls = %v, want %v", got, want)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("execute call %d model = %q, want %q", i, got[i], want[i])
+		}
+	}
+}
+
+func TestManagerExecute_OpenAICompatAliasPoolStopsOnBadRequest(t *testing.T) {
+	alias := "claude-opus-4.66"
+	invalidErr := &Error{HTTPStatus: http.StatusBadRequest, Message: "invalid_request_error: malformed payload"}
+	executor := &openAICompatPoolExecutor{
+		id:            "pool",
+		executeErrors: map[string]error{"qwen3.5-plus": invalidErr},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	_, err := m.Execute(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err == nil || err.Error() != invalidErr.Error() {
+		t.Fatalf("execute error = %v, want %v", err, invalidErr)
+	}
+	got := executor.ExecuteModels()
+	if len(got) != 1 || got[0] != "qwen3.5-plus" {
+		t.Fatalf("execute calls = %v, want only first invalid model", got)
+	}
+}
+func TestManagerExecute_OpenAICompatAliasPoolFallsBackWithinSameAuth(t *testing.T) {
+	alias := "claude-opus-4.66"
+	executor := &openAICompatPoolExecutor{
+		id:            "pool",
+		executeErrors: map[string]error{"qwen3.5-plus": &Error{HTTPStatus: http.StatusTooManyRequests, Message: "quota"}},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	resp, err := m.Execute(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err != nil {
+		t.Fatalf("execute: %v", err)
+	}
+	if string(resp.Payload) != "glm-5" {
+		t.Fatalf("payload = %q, want %q", string(resp.Payload), "glm-5")
+	}
+	got := executor.ExecuteModels()
+	want := []string{"qwen3.5-plus", "glm-5"}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("execute call %d model = %q, want %q", i, got[i], want[i])
+		}
+	}
+}
+
+func TestManagerExecuteStream_OpenAICompatAliasPoolRetriesOnEmptyBootstrap(t *testing.T) {
+	alias := "claude-opus-4.66"
+	executor := &openAICompatPoolExecutor{
+		id: "pool",
+		streamPayloads: map[string][]cliproxyexecutor.StreamChunk{
+			"qwen3.5-plus": {},
+		},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	streamResult, err := m.ExecuteStream(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err != nil {
+		t.Fatalf("execute stream: %v", err)
+	}
+	var payload []byte
+	for chunk := range streamResult.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("unexpected stream error: %v", chunk.Err)
+		}
+		payload = append(payload, chunk.Payload...)
+	}
+	if string(payload) != "glm-5" {
+		t.Fatalf("payload = %q, want %q", string(payload), "glm-5")
+	}
+	got := executor.StreamModels()
+	want := []string{"qwen3.5-plus", "glm-5"}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("stream call %d model = %q, want %q", i, got[i], want[i])
+		}
+	}
+}
+
+func TestManagerExecuteStream_OpenAICompatAliasPoolFallsBackBeforeFirstByte(t *testing.T) {
+	alias := "claude-opus-4.66"
+	executor := &openAICompatPoolExecutor{
+		id:                "pool",
+		streamFirstErrors: map[string]error{"qwen3.5-plus": &Error{HTTPStatus: http.StatusTooManyRequests, Message: "quota"}},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	streamResult, err := m.ExecuteStream(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err != nil {
+		t.Fatalf("execute stream: %v", err)
+	}
+	var payload []byte
+	for chunk := range streamResult.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("unexpected stream error: %v", chunk.Err)
+		}
+		payload = append(payload, chunk.Payload...)
+	}
+	if string(payload) != "glm-5" {
+		t.Fatalf("payload = %q, want %q", string(payload), "glm-5")
+	}
+	got := executor.StreamModels()
+	want := []string{"qwen3.5-plus", "glm-5"}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("stream call %d model = %q, want %q", i, got[i], want[i])
+		}
+	}
+	if gotHeader := streamResult.Headers.Get("X-Model"); gotHeader != "glm-5" {
+		t.Fatalf("header X-Model = %q, want %q", gotHeader, "glm-5")
+	}
+}
+
+func TestManagerExecuteStream_OpenAICompatAliasPoolStopsOnInvalidRequest(t *testing.T) {
+	alias := "claude-opus-4.66"
+	invalidErr := &Error{HTTPStatus: http.StatusUnprocessableEntity, Message: "unprocessable entity"}
+	executor := &openAICompatPoolExecutor{
+		id:                "pool",
+		streamFirstErrors: map[string]error{"qwen3.5-plus": invalidErr},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	_, err := m.ExecuteStream(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err == nil || err.Error() != invalidErr.Error() {
+		t.Fatalf("execute stream error = %v, want %v", err, invalidErr)
+	}
+	got := executor.StreamModels()
+	if len(got) != 1 || got[0] != "qwen3.5-plus" {
+		t.Fatalf("stream calls = %v, want only first invalid model", got)
+	}
+}
+func TestManagerExecuteCount_OpenAICompatAliasPoolRotatesWithinAuth(t *testing.T) {
+	alias := "claude-opus-4.66"
+	executor := &openAICompatPoolExecutor{id: "pool"}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	for i := 0; i < 2; i++ {
+		resp, err := m.ExecuteCount(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+		if err != nil {
+			t.Fatalf("execute count %d: %v", i, err)
+		}
+		if len(resp.Payload) == 0 {
+			t.Fatalf("execute count %d returned empty payload", i)
+		}
+	}
+
+	got := executor.CountModels()
+	want := []string{"qwen3.5-plus", "glm-5"}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("count call %d model = %q, want %q", i, got[i], want[i])
+		}
+	}
+}
+
+func TestManagerExecuteStream_OpenAICompatAliasPoolStopsOnInvalidBootstrap(t *testing.T) {
+	alias := "claude-opus-4.66"
+	invalidErr := &Error{HTTPStatus: http.StatusBadRequest, Message: "invalid_request_error: malformed payload"}
+	executor := &openAICompatPoolExecutor{
+		id:                "pool",
+		streamFirstErrors: map[string]error{"qwen3.5-plus": invalidErr},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	streamResult, err := m.ExecuteStream(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err == nil {
+		t.Fatal("expected invalid request error")
+	}
+	if err != invalidErr {
+		t.Fatalf("error = %v, want %v", err, invalidErr)
+	}
+	if streamResult != nil {
+		t.Fatalf("streamResult = %#v, want nil on invalid bootstrap", streamResult)
+	}
+	if got := executor.StreamModels(); len(got) != 1 || got[0] != "qwen3.5-plus" {
+		t.Fatalf("stream calls = %v, want only first upstream model", got)
+	}
+}

sdk/cliproxy/auth/scheduler.go

@@ -0,0 +1,904 @@
+package auth
+
+import (
+	"context"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+)
+
+// schedulerStrategy identifies which built-in routing semantics the scheduler should apply.
+type schedulerStrategy int
+
+const (
+	schedulerStrategyCustom schedulerStrategy = iota
+	schedulerStrategyRoundRobin
+	schedulerStrategyFillFirst
+)
+
+// scheduledState describes how an auth currently participates in a model shard.
+type scheduledState int
+
+const (
+	scheduledStateReady scheduledState = iota
+	scheduledStateCooldown
+	scheduledStateBlocked
+	scheduledStateDisabled
+)
+
+// authScheduler keeps the incremental provider/model scheduling state used by Manager.
+type authScheduler struct {
+	mu            sync.Mutex
+	strategy      schedulerStrategy
+	providers     map[string]*providerScheduler
+	authProviders map[string]string
+	mixedCursors  map[string]int
+}
+
+// providerScheduler stores auth metadata and model shards for a single provider.
+type providerScheduler struct {
+	providerKey string
+	auths       map[string]*scheduledAuthMeta
+	modelShards map[string]*modelScheduler
+}
+
+// scheduledAuthMeta stores the immutable scheduling fields derived from an auth snapshot.
+type scheduledAuthMeta struct {
+	auth              *Auth
+	providerKey       string
+	priority          int
+	virtualParent     string
+	websocketEnabled  bool
+	supportedModelSet map[string]struct{}
+}
+
+// modelScheduler tracks ready and blocked auths for one provider/model combination.
+type modelScheduler struct {
+	modelKey        string
+	entries         map[string]*scheduledAuth
+	priorityOrder   []int
+	readyByPriority map[int]*readyBucket
+	blocked         cooldownQueue
+}
+
+// scheduledAuth stores the runtime scheduling state for a single auth inside a model shard.
+type scheduledAuth struct {
+	meta        *scheduledAuthMeta
+	auth        *Auth
+	state       scheduledState
+	nextRetryAt time.Time
+}
+
+// readyBucket keeps the ready views for one priority level.
+type readyBucket struct {
+	all readyView
+	ws  readyView
+}
+
+// readyView holds the selection order for flat or grouped round-robin traversal.
+type readyView struct {
+	flat         []*scheduledAuth
+	cursor       int
+	parentOrder  []string
+	parentCursor int
+	children     map[string]*childBucket
+}
+
+// childBucket keeps the per-parent rotation state for grouped Gemini virtual auths.
+type childBucket struct {
+	items  []*scheduledAuth
+	cursor int
+}
+
+// cooldownQueue is the blocked auth collection ordered by next retry time during rebuilds.
+type cooldownQueue []*scheduledAuth
+
+// newAuthScheduler constructs an empty scheduler configured for the supplied selector strategy.
+func newAuthScheduler(selector Selector) *authScheduler {
+	return &authScheduler{
+		strategy:      selectorStrategy(selector),
+		providers:     make(map[string]*providerScheduler),
+		authProviders: make(map[string]string),
+		mixedCursors:  make(map[string]int),
+	}
+}
+
+// selectorStrategy maps a selector implementation to the scheduler semantics it should emulate.
+func selectorStrategy(selector Selector) schedulerStrategy {
+	switch selector.(type) {
+	case *FillFirstSelector:
+		return schedulerStrategyFillFirst
+	case nil, *RoundRobinSelector:
+		return schedulerStrategyRoundRobin
+	default:
+		return schedulerStrategyCustom
+	}
+}
+
+// setSelector updates the active built-in strategy and resets mixed-provider cursors.
+func (s *authScheduler) setSelector(selector Selector) {
+	if s == nil {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.strategy = selectorStrategy(selector)
+	clear(s.mixedCursors)
+}
+
+// rebuild recreates the complete scheduler state from an auth snapshot.
+func (s *authScheduler) rebuild(auths []*Auth) {
+	if s == nil {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.providers = make(map[string]*providerScheduler)
+	s.authProviders = make(map[string]string)
+	s.mixedCursors = make(map[string]int)
+	now := time.Now()
+	for _, auth := range auths {
+		s.upsertAuthLocked(auth, now)
+	}
+}
+
+// upsertAuth incrementally synchronizes one auth into the scheduler.
+func (s *authScheduler) upsertAuth(auth *Auth) {
+	if s == nil {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.upsertAuthLocked(auth, time.Now())
+}
+
+// removeAuth deletes one auth from every scheduler shard that references it.
+func (s *authScheduler) removeAuth(authID string) {
+	if s == nil {
+		return
+	}
+	authID = strings.TrimSpace(authID)
+	if authID == "" {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.removeAuthLocked(authID)
+}
+
+// pickSingle returns the next auth for a single provider/model request using scheduler state.
+func (s *authScheduler) pickSingle(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, error) {
+	if s == nil {
+		return nil, &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	providerKey := strings.ToLower(strings.TrimSpace(provider))
+	modelKey := canonicalModelKey(model)
+	pinnedAuthID := pinnedAuthIDFromMetadata(opts.Metadata)
+	preferWebsocket := cliproxyexecutor.DownstreamWebsocket(ctx) && providerKey == "codex" && pinnedAuthID == ""
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	providerState := s.providers[providerKey]
+	if providerState == nil {
+		return nil, &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	shard := providerState.ensureModelLocked(modelKey, time.Now())
+	if shard == nil {
+		return nil, &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	predicate := func(entry *scheduledAuth) bool {
+		if entry == nil || entry.auth == nil {
+			return false
+		}
+		if pinnedAuthID != "" && entry.auth.ID != pinnedAuthID {
+			return false
+		}
+		if len(tried) > 0 {
+			if _, ok := tried[entry.auth.ID]; ok {
+				return false
+			}
+		}
+		return true
+	}
+	if picked := shard.pickReadyLocked(preferWebsocket, s.strategy, predicate); picked != nil {
+		return picked, nil
+	}
+	return nil, shard.unavailableErrorLocked(provider, model, predicate)
+}
+
+// pickMixed returns the next auth and provider for a mixed-provider request.
+func (s *authScheduler) pickMixed(ctx context.Context, providers []string, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, string, error) {
+	if s == nil {
+		return nil, "", &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	normalized := normalizeProviderKeys(providers)
+	if len(normalized) == 0 {
+		return nil, "", &Error{Code: "provider_not_found", Message: "no provider supplied"}
+	}
+	pinnedAuthID := pinnedAuthIDFromMetadata(opts.Metadata)
+	modelKey := canonicalModelKey(model)
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if pinnedAuthID != "" {
+		providerKey := s.authProviders[pinnedAuthID]
+		if providerKey == "" || !containsProvider(normalized, providerKey) {
+			return nil, "", &Error{Code: "auth_not_found", Message: "no auth available"}
+		}
+		providerState := s.providers[providerKey]
+		if providerState == nil {
+			return nil, "", &Error{Code: "auth_not_found", Message: "no auth available"}
+		}
+		shard := providerState.ensureModelLocked(modelKey, time.Now())
+		predicate := func(entry *scheduledAuth) bool {
+			if entry == nil || entry.auth == nil || entry.auth.ID != pinnedAuthID {
+				return false
+			}
+			if len(tried) == 0 {
+				return true
+			}
+			_, ok := tried[pinnedAuthID]
+			return !ok
+		}
+		if picked := shard.pickReadyLocked(false, s.strategy, predicate); picked != nil {
+			return picked, providerKey, nil
+		}
+		return nil, "", shard.unavailableErrorLocked("mixed", model, predicate)
+	}
+
+	predicate := triedPredicate(tried)
+	candidateShards := make([]*modelScheduler, len(normalized))
+	bestPriority := 0
+	hasCandidate := false
+	now := time.Now()
+	for providerIndex, providerKey := range normalized {
+		providerState := s.providers[providerKey]
+		if providerState == nil {
+			continue
+		}
+		shard := providerState.ensureModelLocked(modelKey, now)
+		candidateShards[providerIndex] = shard
+		if shard == nil {
+			continue
+		}
+		priorityReady, okPriority := shard.highestReadyPriorityLocked(false, predicate)
+		if !okPriority {
+			continue
+		}
+		if !hasCandidate || priorityReady > bestPriority {
+			bestPriority = priorityReady
+			hasCandidate = true
+		}
+	}
+	if !hasCandidate {
+		return nil, "", s.mixedUnavailableErrorLocked(normalized, model, tried)
+	}
+
+	if s.strategy == schedulerStrategyFillFirst {
+		for providerIndex, providerKey := range normalized {
+			shard := candidateShards[providerIndex]
+			if shard == nil {
+				continue
+			}
+			picked := shard.pickReadyAtPriorityLocked(false, bestPriority, s.strategy, predicate)
+			if picked != nil {
+				return picked, providerKey, nil
+			}
+		}
+		return nil, "", s.mixedUnavailableErrorLocked(normalized, model, tried)
+	}
+
+	cursorKey := strings.Join(normalized, ",") + ":" + modelKey
+	start := 0
+	if len(normalized) > 0 {
+		start = s.mixedCursors[cursorKey] % len(normalized)
+	}
+	for offset := 0; offset < len(normalized); offset++ {
+		providerIndex := (start + offset) % len(normalized)
+		providerKey := normalized[providerIndex]
+		shard := candidateShards[providerIndex]
+		if shard == nil {
+			continue
+		}
+		picked := shard.pickReadyAtPriorityLocked(false, bestPriority, schedulerStrategyRoundRobin, predicate)
+		if picked == nil {
+			continue
+		}
+		s.mixedCursors[cursorKey] = providerIndex + 1
+		return picked, providerKey, nil
+	}
+	return nil, "", s.mixedUnavailableErrorLocked(normalized, model, tried)
+}
+
+// mixedUnavailableErrorLocked synthesizes the mixed-provider cooldown or unavailable error.
+func (s *authScheduler) mixedUnavailableErrorLocked(providers []string, model string, tried map[string]struct{}) error {
+	now := time.Now()
+	total := 0
+	cooldownCount := 0
+	earliest := time.Time{}
+	for _, providerKey := range providers {
+		providerState := s.providers[providerKey]
+		if providerState == nil {
+			continue
+		}
+		shard := providerState.ensureModelLocked(canonicalModelKey(model), now)
+		if shard == nil {
+			continue
+		}
+		localTotal, localCooldownCount, localEarliest := shard.availabilitySummaryLocked(triedPredicate(tried))
+		total += localTotal
+		cooldownCount += localCooldownCount
+		if !localEarliest.IsZero() && (earliest.IsZero() || localEarliest.Before(earliest)) {
+			earliest = localEarliest
+		}
+	}
+	if total == 0 {
+		return &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	if cooldownCount == total && !earliest.IsZero() {
+		resetIn := earliest.Sub(now)
+		if resetIn < 0 {
+			resetIn = 0
+		}
+		return newModelCooldownError(model, "", resetIn)
+	}
+	return &Error{Code: "auth_unavailable", Message: "no auth available"}
+}
+
+// triedPredicate builds a filter that excludes auths already attempted for the current request.
+func triedPredicate(tried map[string]struct{}) func(*scheduledAuth) bool {
+	if len(tried) == 0 {
+		return func(entry *scheduledAuth) bool { return entry != nil && entry.auth != nil }
+	}
+	return func(entry *scheduledAuth) bool {
+		if entry == nil || entry.auth == nil {
+			return false
+		}
+		_, ok := tried[entry.auth.ID]
+		return !ok
+	}
+}
+
+// normalizeProviderKeys lowercases, trims, and de-duplicates provider keys while preserving order.
+func normalizeProviderKeys(providers []string) []string {
+	seen := make(map[string]struct{}, len(providers))
+	out := make([]string, 0, len(providers))
+	for _, provider := range providers {
+		providerKey := strings.ToLower(strings.TrimSpace(provider))
+		if providerKey == "" {
+			continue
+		}
+		if _, ok := seen[providerKey]; ok {
+			continue
+		}
+		seen[providerKey] = struct{}{}
+		out = append(out, providerKey)
+	}
+	return out
+}
+
+// containsProvider reports whether provider is present in the normalized provider list.
+func containsProvider(providers []string, provider string) bool {
+	for _, candidate := range providers {
+		if candidate == provider {
+			return true
+		}
+	}
+	return false
+}
+
+// upsertAuthLocked updates one auth in-place while the scheduler mutex is held.
+func (s *authScheduler) upsertAuthLocked(auth *Auth, now time.Time) {
+	if auth == nil {
+		return
+	}
+	authID := strings.TrimSpace(auth.ID)
+	providerKey := strings.ToLower(strings.TrimSpace(auth.Provider))
+	if authID == "" || providerKey == "" || auth.Disabled {
+		s.removeAuthLocked(authID)
+		return
+	}
+	if previousProvider := s.authProviders[authID]; previousProvider != "" && previousProvider != providerKey {
+		if previousState := s.providers[previousProvider]; previousState != nil {
+			previousState.removeAuthLocked(authID)
+		}
+	}
+	meta := buildScheduledAuthMeta(auth)
+	s.authProviders[authID] = providerKey
+	s.ensureProviderLocked(providerKey).upsertAuthLocked(meta, now)
+}
+
+// removeAuthLocked removes one auth from the scheduler while the scheduler mutex is held.
+func (s *authScheduler) removeAuthLocked(authID string) {
+	if authID == "" {
+		return
+	}
+	if providerKey := s.authProviders[authID]; providerKey != "" {
+		if providerState := s.providers[providerKey]; providerState != nil {
+			providerState.removeAuthLocked(authID)
+		}
+		delete(s.authProviders, authID)
+	}
+}
+
+// ensureProviderLocked returns the provider scheduler for providerKey, creating it when needed.
+func (s *authScheduler) ensureProviderLocked(providerKey string) *providerScheduler {
+	if s.providers == nil {
+		s.providers = make(map[string]*providerScheduler)
+	}
+	providerState := s.providers[providerKey]
+	if providerState == nil {
+		providerState = &providerScheduler{
+			providerKey: providerKey,
+			auths:       make(map[string]*scheduledAuthMeta),
+			modelShards: make(map[string]*modelScheduler),
+		}
+		s.providers[providerKey] = providerState
+	}
+	return providerState
+}
+
+// buildScheduledAuthMeta extracts the scheduling metadata needed for shard bookkeeping.
+func buildScheduledAuthMeta(auth *Auth) *scheduledAuthMeta {
+	providerKey := strings.ToLower(strings.TrimSpace(auth.Provider))
+	virtualParent := ""
+	if auth.Attributes != nil {
+		virtualParent = strings.TrimSpace(auth.Attributes["gemini_virtual_parent"])
+	}
+	return &scheduledAuthMeta{
+		auth:              auth,
+		providerKey:       providerKey,
+		priority:          authPriority(auth),
+		virtualParent:     virtualParent,
+		websocketEnabled:  authWebsocketsEnabled(auth),
+		supportedModelSet: supportedModelSetForAuth(auth.ID),
+	}
+}
+
+// supportedModelSetForAuth snapshots the registry models currently registered for an auth.
+func supportedModelSetForAuth(authID string) map[string]struct{} {
+	authID = strings.TrimSpace(authID)
+	if authID == "" {
+		return nil
+	}
+	models := registry.GetGlobalRegistry().GetModelsForClient(authID)
+	if len(models) == 0 {
+		return nil
+	}
+	set := make(map[string]struct{}, len(models))
+	for _, model := range models {
+		if model == nil {
+			continue
+		}
+		modelKey := canonicalModelKey(model.ID)
+		if modelKey == "" {
+			continue
+		}
+		set[modelKey] = struct{}{}
+	}
+	return set
+}
+
+// upsertAuthLocked updates every existing model shard that can reference the auth metadata.
+func (p *providerScheduler) upsertAuthLocked(meta *scheduledAuthMeta, now time.Time) {
+	if p == nil || meta == nil || meta.auth == nil {
+		return
+	}
+	p.auths[meta.auth.ID] = meta
+	for modelKey, shard := range p.modelShards {
+		if shard == nil {
+			continue
+		}
+		if !meta.supportsModel(modelKey) {
+			shard.removeEntryLocked(meta.auth.ID)
+			continue
+		}
+		shard.upsertEntryLocked(meta, now)
+	}
+}
+
+// removeAuthLocked removes an auth from all model shards owned by the provider scheduler.
+func (p *providerScheduler) removeAuthLocked(authID string) {
+	if p == nil || authID == "" {
+		return
+	}
+	delete(p.auths, authID)
+	for _, shard := range p.modelShards {
+		if shard != nil {
+			shard.removeEntryLocked(authID)
+		}
+	}
+}
+
+// ensureModelLocked returns the shard for modelKey, building it lazily from provider auths.
+func (p *providerScheduler) ensureModelLocked(modelKey string, now time.Time) *modelScheduler {
+	if p == nil {
+		return nil
+	}
+	modelKey = canonicalModelKey(modelKey)
+	if shard, ok := p.modelShards[modelKey]; ok && shard != nil {
+		shard.promoteExpiredLocked(now)
+		return shard
+	}
+	shard := &modelScheduler{
+		modelKey:        modelKey,
+		entries:         make(map[string]*scheduledAuth),
+		readyByPriority: make(map[int]*readyBucket),
+	}
+	for _, meta := range p.auths {
+		if meta == nil || !meta.supportsModel(modelKey) {
+			continue
+		}
+		shard.upsertEntryLocked(meta, now)
+	}
+	p.modelShards[modelKey] = shard
+	return shard
+}
+
+// supportsModel reports whether the auth metadata currently supports modelKey.
+func (m *scheduledAuthMeta) supportsModel(modelKey string) bool {
+	modelKey = canonicalModelKey(modelKey)
+	if modelKey == "" {
+		return true
+	}
+	if len(m.supportedModelSet) == 0 {
+		return false
+	}
+	_, ok := m.supportedModelSet[modelKey]
+	return ok
+}
+
+// upsertEntryLocked updates or inserts one auth entry and rebuilds indexes when ordering changes.
+func (m *modelScheduler) upsertEntryLocked(meta *scheduledAuthMeta, now time.Time) {
+	if m == nil || meta == nil || meta.auth == nil {
+		return
+	}
+	entry, ok := m.entries[meta.auth.ID]
+	if !ok || entry == nil {
+		entry = &scheduledAuth{}
+		m.entries[meta.auth.ID] = entry
+	}
+	previousState := entry.state
+	previousNextRetryAt := entry.nextRetryAt
+	previousPriority := 0
+	previousParent := ""
+	previousWebsocketEnabled := false
+	if entry.meta != nil {
+		previousPriority = entry.meta.priority
+		previousParent = entry.meta.virtualParent
+		previousWebsocketEnabled = entry.meta.websocketEnabled
+	}
+
+	entry.meta = meta
+	entry.auth = meta.auth
+	entry.nextRetryAt = time.Time{}
+	blocked, reason, next := isAuthBlockedForModel(meta.auth, m.modelKey, now)
+	switch {
+	case !blocked:
+		entry.state = scheduledStateReady
+	case reason == blockReasonCooldown:
+		entry.state = scheduledStateCooldown
+		entry.nextRetryAt = next
+	case reason == blockReasonDisabled:
+		entry.state = scheduledStateDisabled
+	default:
+		entry.state = scheduledStateBlocked
+		entry.nextRetryAt = next
+	}
+
+	if ok && previousState == entry.state && previousNextRetryAt.Equal(entry.nextRetryAt) && previousPriority == meta.priority && previousParent == meta.virtualParent && previousWebsocketEnabled == meta.websocketEnabled {
+		return
+	}
+	m.rebuildIndexesLocked()
+}
+
+// removeEntryLocked deletes one auth entry and rebuilds the shard indexes if needed.
+func (m *modelScheduler) removeEntryLocked(authID string) {
+	if m == nil || authID == "" {
+		return
+	}
+	if _, ok := m.entries[authID]; !ok {
+		return
+	}
+	delete(m.entries, authID)
+	m.rebuildIndexesLocked()
+}
+
+// promoteExpiredLocked reevaluates blocked auths whose retry time has elapsed.
+func (m *modelScheduler) promoteExpiredLocked(now time.Time) {
+	if m == nil || len(m.blocked) == 0 {
+		return
+	}
+	changed := false
+	for _, entry := range m.blocked {
+		if entry == nil || entry.auth == nil {
+			continue
+		}
+		if entry.nextRetryAt.IsZero() || entry.nextRetryAt.After(now) {
+			continue
+		}
+		blocked, reason, next := isAuthBlockedForModel(entry.auth, m.modelKey, now)
+		switch {
+		case !blocked:
+			entry.state = scheduledStateReady
+			entry.nextRetryAt = time.Time{}
+		case reason == blockReasonCooldown:
+			entry.state = scheduledStateCooldown
+			entry.nextRetryAt = next
+		case reason == blockReasonDisabled:
+			entry.state = scheduledStateDisabled
+			entry.nextRetryAt = time.Time{}
+		default:
+			entry.state = scheduledStateBlocked
+			entry.nextRetryAt = next
+		}
+		changed = true
+	}
+	if changed {
+		m.rebuildIndexesLocked()
+	}
+}
+
+// pickReadyLocked selects the next ready auth from the highest available priority bucket.
+func (m *modelScheduler) pickReadyLocked(preferWebsocket bool, strategy schedulerStrategy, predicate func(*scheduledAuth) bool) *Auth {
+	if m == nil {
+		return nil
+	}
+	m.promoteExpiredLocked(time.Now())
+	priorityReady, okPriority := m.highestReadyPriorityLocked(preferWebsocket, predicate)
+	if !okPriority {
+		return nil
+	}
+	return m.pickReadyAtPriorityLocked(preferWebsocket, priorityReady, strategy, predicate)
+}
+
+// highestReadyPriorityLocked returns the highest priority bucket that still has a matching ready auth.
+// The caller must ensure expired entries are already promoted when needed.
+func (m *modelScheduler) highestReadyPriorityLocked(preferWebsocket bool, predicate func(*scheduledAuth) bool) (int, bool) {
+	if m == nil {
+		return 0, false
+	}
+	for _, priority := range m.priorityOrder {
+		bucket := m.readyByPriority[priority]
+		if bucket == nil {
+			continue
+		}
+		view := &bucket.all
+		if preferWebsocket && len(bucket.ws.flat) > 0 {
+			view = &bucket.ws
+		}
+		if view.pickFirst(predicate) != nil {
+			return priority, true
+		}
+	}
+	return 0, false
+}
+
+// pickReadyAtPriorityLocked selects the next ready auth from a specific priority bucket.
+// The caller must ensure expired entries are already promoted when needed.
+func (m *modelScheduler) pickReadyAtPriorityLocked(preferWebsocket bool, priority int, strategy schedulerStrategy, predicate func(*scheduledAuth) bool) *Auth {
+	if m == nil {
+		return nil
+	}
+	bucket := m.readyByPriority[priority]
+	if bucket == nil {
+		return nil
+	}
+	view := &bucket.all
+	if preferWebsocket && len(bucket.ws.flat) > 0 {
+		view = &bucket.ws
+	}
+	var picked *scheduledAuth
+	if strategy == schedulerStrategyFillFirst {
+		picked = view.pickFirst(predicate)
+	} else {
+		picked = view.pickRoundRobin(predicate)
+	}
+	if picked == nil || picked.auth == nil {
+		return nil
+	}
+	return picked.auth
+}
+
+// unavailableErrorLocked returns the correct unavailable or cooldown error for the shard.
+func (m *modelScheduler) unavailableErrorLocked(provider, model string, predicate func(*scheduledAuth) bool) error {
+	now := time.Now()
+	total, cooldownCount, earliest := m.availabilitySummaryLocked(predicate)
+	if total == 0 {
+		return &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	if cooldownCount == total && !earliest.IsZero() {
+		providerForError := provider
+		if providerForError == "mixed" {
+			providerForError = ""
+		}
+		resetIn := earliest.Sub(now)
+		if resetIn < 0 {
+			resetIn = 0
+		}
+		return newModelCooldownError(model, providerForError, resetIn)
+	}
+	return &Error{Code: "auth_unavailable", Message: "no auth available"}
+}
+
+// availabilitySummaryLocked summarizes total candidates, cooldown count, and earliest retry time.
+func (m *modelScheduler) availabilitySummaryLocked(predicate func(*scheduledAuth) bool) (int, int, time.Time) {
+	if m == nil {
+		return 0, 0, time.Time{}
+	}
+	total := 0
+	cooldownCount := 0
+	earliest := time.Time{}
+	for _, entry := range m.entries {
+		if predicate != nil && !predicate(entry) {
+			continue
+		}
+		total++
+		if entry == nil || entry.auth == nil {
+			continue
+		}
+		if entry.state != scheduledStateCooldown {
+			continue
+		}
+		cooldownCount++
+		if !entry.nextRetryAt.IsZero() && (earliest.IsZero() || entry.nextRetryAt.Before(earliest)) {
+			earliest = entry.nextRetryAt
+		}
+	}
+	return total, cooldownCount, earliest
+}
+
+// rebuildIndexesLocked reconstructs ready and blocked views from the current entry map.
+func (m *modelScheduler) rebuildIndexesLocked() {
+	m.readyByPriority = make(map[int]*readyBucket)
+	m.priorityOrder = m.priorityOrder[:0]
+	m.blocked = m.blocked[:0]
+	priorityBuckets := make(map[int][]*scheduledAuth)
+	for _, entry := range m.entries {
+		if entry == nil || entry.auth == nil {
+			continue
+		}
+		switch entry.state {
+		case scheduledStateReady:
+			priority := entry.meta.priority
+			priorityBuckets[priority] = append(priorityBuckets[priority], entry)
+		case scheduledStateCooldown, scheduledStateBlocked:
+			m.blocked = append(m.blocked, entry)
+		}
+	}
+	for priority, entries := range priorityBuckets {
+		sort.Slice(entries, func(i, j int) bool {
+			return entries[i].auth.ID < entries[j].auth.ID
+		})
+		m.readyByPriority[priority] = buildReadyBucket(entries)
+		m.priorityOrder = append(m.priorityOrder, priority)
+	}
+	sort.Slice(m.priorityOrder, func(i, j int) bool {
+		return m.priorityOrder[i] > m.priorityOrder[j]
+	})
+	sort.Slice(m.blocked, func(i, j int) bool {
+		left := m.blocked[i]
+		right := m.blocked[j]
+		if left == nil || right == nil {
+			return left != nil
+		}
+		if left.nextRetryAt.Equal(right.nextRetryAt) {
+			return left.auth.ID < right.auth.ID
+		}
+		if left.nextRetryAt.IsZero() {
+			return false
+		}
+		if right.nextRetryAt.IsZero() {
+			return true
+		}
+		return left.nextRetryAt.Before(right.nextRetryAt)
+	})
+}
+
+// buildReadyBucket prepares the general and websocket-only ready views for one priority bucket.
+func buildReadyBucket(entries []*scheduledAuth) *readyBucket {
+	bucket := &readyBucket{}
+	bucket.all = buildReadyView(entries)
+	wsEntries := make([]*scheduledAuth, 0, len(entries))
+	for _, entry := range entries {
+		if entry != nil && entry.meta != nil && entry.meta.websocketEnabled {
+			wsEntries = append(wsEntries, entry)
+		}
+	}
+	bucket.ws = buildReadyView(wsEntries)
+	return bucket
+}
+
+// buildReadyView creates either a flat view or a grouped parent/child view for rotation.
+func buildReadyView(entries []*scheduledAuth) readyView {
+	view := readyView{flat: append([]*scheduledAuth(nil), entries...)}
+	if len(entries) == 0 {
+		return view
+	}
+	groups := make(map[string][]*scheduledAuth)
+	for _, entry := range entries {
+		if entry == nil || entry.meta == nil || entry.meta.virtualParent == "" {
+			return view
+		}
+		groups[entry.meta.virtualParent] = append(groups[entry.meta.virtualParent], entry)
+	}
+	if len(groups) <= 1 {
+		return view
+	}
+	view.children = make(map[string]*childBucket, len(groups))
+	view.parentOrder = make([]string, 0, len(groups))
+	for parent := range groups {
+		view.parentOrder = append(view.parentOrder, parent)
+	}
+	sort.Strings(view.parentOrder)
+	for _, parent := range view.parentOrder {
+		view.children[parent] = &childBucket{items: append([]*scheduledAuth(nil), groups[parent]...)}
+	}
+	return view
+}
+
+// pickFirst returns the first ready entry that satisfies predicate without advancing cursors.
+func (v *readyView) pickFirst(predicate func(*scheduledAuth) bool) *scheduledAuth {
+	for _, entry := range v.flat {
+		if predicate == nil || predicate(entry) {
+			return entry
+		}
+	}
+	return nil
+}
+
+// pickRoundRobin returns the next ready entry using flat or grouped round-robin traversal.
+func (v *readyView) pickRoundRobin(predicate func(*scheduledAuth) bool) *scheduledAuth {
+	if len(v.parentOrder) > 1 && len(v.children) > 0 {
+		return v.pickGroupedRoundRobin(predicate)
+	}
+	if len(v.flat) == 0 {
+		return nil
+	}
+	start := 0
+	if len(v.flat) > 0 {
+		start = v.cursor % len(v.flat)
+	}
+	for offset := 0; offset < len(v.flat); offset++ {
+		index := (start + offset) % len(v.flat)
+		entry := v.flat[index]
+		if predicate != nil && !predicate(entry) {
+			continue
+		}
+		v.cursor = index + 1
+		return entry
+	}
+	return nil
+}
+
+// pickGroupedRoundRobin rotates across parents first and then within the selected parent.
+func (v *readyView) pickGroupedRoundRobin(predicate func(*scheduledAuth) bool) *scheduledAuth {
+	start := 0
+	if len(v.parentOrder) > 0 {
+		start = v.parentCursor % len(v.parentOrder)
+	}
+	for offset := 0; offset < len(v.parentOrder); offset++ {
+		parentIndex := (start + offset) % len(v.parentOrder)
+		parent := v.parentOrder[parentIndex]
+		child := v.children[parent]
+		if child == nil || len(child.items) == 0 {
+			continue
+		}
+		itemStart := child.cursor % len(child.items)
+		for itemOffset := 0; itemOffset < len(child.items); itemOffset++ {
+			itemIndex := (itemStart + itemOffset) % len(child.items)
+			entry := child.items[itemIndex]
+			if predicate != nil && !predicate(entry) {
+				continue
+			}
+			child.cursor = itemIndex + 1
+			v.parentCursor = parentIndex + 1
+			return entry
+		}
+	}
+	return nil
+}

sdk/cliproxy/auth/scheduler_benchmark_test.go

@@ -0,0 +1,216 @@
+package auth
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+)
+
+type schedulerBenchmarkExecutor struct {
+	id string
+}
+
+func (e schedulerBenchmarkExecutor) Identifier() string { return e.id }
+
+func (e schedulerBenchmarkExecutor) Execute(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (e schedulerBenchmarkExecutor) ExecuteStream(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
+	return nil, nil
+}
+
+func (e schedulerBenchmarkExecutor) Refresh(ctx context.Context, auth *Auth) (*Auth, error) {
+	return auth, nil
+}
+
+func (e schedulerBenchmarkExecutor) CountTokens(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (e schedulerBenchmarkExecutor) HttpRequest(ctx context.Context, auth *Auth, req *http.Request) (*http.Response, error) {
+	return nil, nil
+}
+
+func benchmarkManagerSetup(b *testing.B, total int, mixed bool, withPriority bool) (*Manager, []string, string) {
+	b.Helper()
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	providers := []string{"gemini"}
+	manager.executors["gemini"] = schedulerBenchmarkExecutor{id: "gemini"}
+	if mixed {
+		providers = []string{"gemini", "claude"}
+		manager.executors["claude"] = schedulerBenchmarkExecutor{id: "claude"}
+	}
+
+	reg := registry.GetGlobalRegistry()
+	model := "bench-model"
+	for index := 0; index < total; index++ {
+		provider := providers[0]
+		if mixed && index%2 == 1 {
+			provider = providers[1]
+		}
+		auth := &Auth{ID: fmt.Sprintf("bench-%s-%04d", provider, index), Provider: provider}
+		if withPriority {
+			priority := "0"
+			if index%2 == 0 {
+				priority = "10"
+			}
+			auth.Attributes = map[string]string{"priority": priority}
+		}
+		_, errRegister := manager.Register(context.Background(), auth)
+		if errRegister != nil {
+			b.Fatalf("Register(%s) error = %v", auth.ID, errRegister)
+		}
+		reg.RegisterClient(auth.ID, provider, []*registry.ModelInfo{{ID: model}})
+	}
+	manager.syncScheduler()
+	b.Cleanup(func() {
+		for index := 0; index < total; index++ {
+			provider := providers[0]
+			if mixed && index%2 == 1 {
+				provider = providers[1]
+			}
+			reg.UnregisterClient(fmt.Sprintf("bench-%s-%04d", provider, index))
+		}
+	})
+
+	return manager, providers, model
+}
+
+func BenchmarkManagerPickNext500(b *testing.B) {
+	manager, _, model := benchmarkManagerSetup(b, 500, false, false)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, errWarm := manager.pickNext(ctx, "gemini", model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNext error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, exec, errPick := manager.pickNext(ctx, "gemini", model, opts, tried)
+		if errPick != nil || auth == nil || exec == nil {
+			b.Fatalf("pickNext failed: auth=%v exec=%v err=%v", auth, exec, errPick)
+		}
+	}
+}
+
+func BenchmarkManagerPickNext1000(b *testing.B) {
+	manager, _, model := benchmarkManagerSetup(b, 1000, false, false)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, errWarm := manager.pickNext(ctx, "gemini", model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNext error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, exec, errPick := manager.pickNext(ctx, "gemini", model, opts, tried)
+		if errPick != nil || auth == nil || exec == nil {
+			b.Fatalf("pickNext failed: auth=%v exec=%v err=%v", auth, exec, errPick)
+		}
+	}
+}
+
+func BenchmarkManagerPickNextPriority500(b *testing.B) {
+	manager, _, model := benchmarkManagerSetup(b, 500, false, true)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, errWarm := manager.pickNext(ctx, "gemini", model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNext error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, exec, errPick := manager.pickNext(ctx, "gemini", model, opts, tried)
+		if errPick != nil || auth == nil || exec == nil {
+			b.Fatalf("pickNext failed: auth=%v exec=%v err=%v", auth, exec, errPick)
+		}
+	}
+}
+
+func BenchmarkManagerPickNextPriority1000(b *testing.B) {
+	manager, _, model := benchmarkManagerSetup(b, 1000, false, true)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, errWarm := manager.pickNext(ctx, "gemini", model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNext error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, exec, errPick := manager.pickNext(ctx, "gemini", model, opts, tried)
+		if errPick != nil || auth == nil || exec == nil {
+			b.Fatalf("pickNext failed: auth=%v exec=%v err=%v", auth, exec, errPick)
+		}
+	}
+}
+
+func BenchmarkManagerPickNextMixed500(b *testing.B) {
+	manager, providers, model := benchmarkManagerSetup(b, 500, true, false)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, _, errWarm := manager.pickNextMixed(ctx, providers, model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNextMixed error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, exec, provider, errPick := manager.pickNextMixed(ctx, providers, model, opts, tried)
+		if errPick != nil || auth == nil || exec == nil || provider == "" {
+			b.Fatalf("pickNextMixed failed: auth=%v exec=%v provider=%q err=%v", auth, exec, provider, errPick)
+		}
+	}
+}
+
+func BenchmarkManagerPickNextMixedPriority500(b *testing.B) {
+	manager, providers, model := benchmarkManagerSetup(b, 500, true, true)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, _, errWarm := manager.pickNextMixed(ctx, providers, model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNextMixed error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, exec, provider, errPick := manager.pickNextMixed(ctx, providers, model, opts, tried)
+		if errPick != nil || auth == nil || exec == nil || provider == "" {
+			b.Fatalf("pickNextMixed failed: auth=%v exec=%v provider=%q err=%v", auth, exec, provider, errPick)
+		}
+	}
+}
+
+func BenchmarkManagerPickNextAndMarkResult1000(b *testing.B) {
+	manager, _, model := benchmarkManagerSetup(b, 1000, false, false)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, errWarm := manager.pickNext(ctx, "gemini", model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNext error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, _, errPick := manager.pickNext(ctx, "gemini", model, opts, tried)
+		if errPick != nil || auth == nil {
+			b.Fatalf("pickNext failed: auth=%v err=%v", auth, errPick)
+		}
+		manager.MarkResult(ctx, Result{AuthID: auth.ID, Provider: "gemini", Model: model, Success: true})
+	}
+}

sdk/cliproxy/auth/scheduler_test.go

@@ -0,0 +1,503 @@
+package auth
+
+import (
+	"context"
+	"net/http"
+	"testing"
+	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+)
+
+type schedulerTestExecutor struct{}
+
+func (schedulerTestExecutor) Identifier() string { return "test" }
+
+func (schedulerTestExecutor) Execute(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (schedulerTestExecutor) ExecuteStream(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
+	return nil, nil
+}
+
+func (schedulerTestExecutor) Refresh(ctx context.Context, auth *Auth) (*Auth, error) {
+	return auth, nil
+}
+
+func (schedulerTestExecutor) CountTokens(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (schedulerTestExecutor) HttpRequest(ctx context.Context, auth *Auth, req *http.Request) (*http.Response, error) {
+	return nil, nil
+}
+
+type trackingSelector struct {
+	calls      int
+	lastAuthID []string
+}
+
+func (s *trackingSelector) Pick(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, auths []*Auth) (*Auth, error) {
+	s.calls++
+	s.lastAuthID = s.lastAuthID[:0]
+	for _, auth := range auths {
+		s.lastAuthID = append(s.lastAuthID, auth.ID)
+	}
+	if len(auths) == 0 {
+		return nil, nil
+	}
+	return auths[len(auths)-1], nil
+}
+
+func newSchedulerForTest(selector Selector, auths ...*Auth) *authScheduler {
+	scheduler := newAuthScheduler(selector)
+	scheduler.rebuild(auths)
+	return scheduler
+}
+
+func registerSchedulerModels(t *testing.T, provider string, model string, authIDs ...string) {
+	t.Helper()
+	reg := registry.GetGlobalRegistry()
+	for _, authID := range authIDs {
+		reg.RegisterClient(authID, provider, []*registry.ModelInfo{{ID: model}})
+	}
+	t.Cleanup(func() {
+		for _, authID := range authIDs {
+			reg.UnregisterClient(authID)
+		}
+	})
+}
+
+func TestSchedulerPick_RoundRobinHighestPriority(t *testing.T) {
+	t.Parallel()
+
+	scheduler := newSchedulerForTest(
+		&RoundRobinSelector{},
+		&Auth{ID: "low", Provider: "gemini", Attributes: map[string]string{"priority": "0"}},
+		&Auth{ID: "high-b", Provider: "gemini", Attributes: map[string]string{"priority": "10"}},
+		&Auth{ID: "high-a", Provider: "gemini", Attributes: map[string]string{"priority": "10"}},
+	)
+
+	want := []string{"high-a", "high-b", "high-a"}
+	for index, wantID := range want {
+		got, errPick := scheduler.pickSingle(context.Background(), "gemini", "", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickSingle() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickSingle() #%d auth = nil", index)
+		}
+		if got.ID != wantID {
+			t.Fatalf("pickSingle() #%d auth.ID = %q, want %q", index, got.ID, wantID)
+		}
+	}
+}
+
+func TestSchedulerPick_FillFirstSticksToFirstReady(t *testing.T) {
+	t.Parallel()
+
+	scheduler := newSchedulerForTest(
+		&FillFirstSelector{},
+		&Auth{ID: "b", Provider: "gemini"},
+		&Auth{ID: "a", Provider: "gemini"},
+		&Auth{ID: "c", Provider: "gemini"},
+	)
+
+	for index := 0; index < 3; index++ {
+		got, errPick := scheduler.pickSingle(context.Background(), "gemini", "", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickSingle() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickSingle() #%d auth = nil", index)
+		}
+		if got.ID != "a" {
+			t.Fatalf("pickSingle() #%d auth.ID = %q, want %q", index, got.ID, "a")
+		}
+	}
+}
+
+func TestSchedulerPick_PromotesExpiredCooldownBeforePick(t *testing.T) {
+	t.Parallel()
+
+	model := "gemini-2.5-pro"
+	registerSchedulerModels(t, "gemini", model, "cooldown-expired")
+	scheduler := newSchedulerForTest(
+		&RoundRobinSelector{},
+		&Auth{
+			ID:       "cooldown-expired",
+			Provider: "gemini",
+			ModelStates: map[string]*ModelState{
+				model: {
+					Status:         StatusError,
+					Unavailable:    true,
+					NextRetryAfter: time.Now().Add(-1 * time.Second),
+				},
+			},
+		},
+	)
+
+	got, errPick := scheduler.pickSingle(context.Background(), "gemini", model, cliproxyexecutor.Options{}, nil)
+	if errPick != nil {
+		t.Fatalf("pickSingle() error = %v", errPick)
+	}
+	if got == nil {
+		t.Fatalf("pickSingle() auth = nil")
+	}
+	if got.ID != "cooldown-expired" {
+		t.Fatalf("pickSingle() auth.ID = %q, want %q", got.ID, "cooldown-expired")
+	}
+}
+
+func TestSchedulerPick_GeminiVirtualParentUsesTwoLevelRotation(t *testing.T) {
+	t.Parallel()
+
+	registerSchedulerModels(t, "gemini-cli", "gemini-2.5-pro", "cred-a::proj-1", "cred-a::proj-2", "cred-b::proj-1", "cred-b::proj-2")
+	scheduler := newSchedulerForTest(
+		&RoundRobinSelector{},
+		&Auth{ID: "cred-a::proj-1", Provider: "gemini-cli", Attributes: map[string]string{"gemini_virtual_parent": "cred-a"}},
+		&Auth{ID: "cred-a::proj-2", Provider: "gemini-cli", Attributes: map[string]string{"gemini_virtual_parent": "cred-a"}},
+		&Auth{ID: "cred-b::proj-1", Provider: "gemini-cli", Attributes: map[string]string{"gemini_virtual_parent": "cred-b"}},
+		&Auth{ID: "cred-b::proj-2", Provider: "gemini-cli", Attributes: map[string]string{"gemini_virtual_parent": "cred-b"}},
+	)
+
+	wantParents := []string{"cred-a", "cred-b", "cred-a", "cred-b"}
+	wantIDs := []string{"cred-a::proj-1", "cred-b::proj-1", "cred-a::proj-2", "cred-b::proj-2"}
+	for index := range wantIDs {
+		got, errPick := scheduler.pickSingle(context.Background(), "gemini-cli", "gemini-2.5-pro", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickSingle() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickSingle() #%d auth = nil", index)
+		}
+		if got.ID != wantIDs[index] {
+			t.Fatalf("pickSingle() #%d auth.ID = %q, want %q", index, got.ID, wantIDs[index])
+		}
+		if got.Attributes["gemini_virtual_parent"] != wantParents[index] {
+			t.Fatalf("pickSingle() #%d parent = %q, want %q", index, got.Attributes["gemini_virtual_parent"], wantParents[index])
+		}
+	}
+}
+
+func TestSchedulerPick_CodexWebsocketPrefersWebsocketEnabledSubset(t *testing.T) {
+	t.Parallel()
+
+	scheduler := newSchedulerForTest(
+		&RoundRobinSelector{},
+		&Auth{ID: "codex-http", Provider: "codex"},
+		&Auth{ID: "codex-ws-a", Provider: "codex", Attributes: map[string]string{"websockets": "true"}},
+		&Auth{ID: "codex-ws-b", Provider: "codex", Attributes: map[string]string{"websockets": "true"}},
+	)
+
+	ctx := cliproxyexecutor.WithDownstreamWebsocket(context.Background())
+	want := []string{"codex-ws-a", "codex-ws-b", "codex-ws-a"}
+	for index, wantID := range want {
+		got, errPick := scheduler.pickSingle(ctx, "codex", "", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickSingle() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickSingle() #%d auth = nil", index)
+		}
+		if got.ID != wantID {
+			t.Fatalf("pickSingle() #%d auth.ID = %q, want %q", index, got.ID, wantID)
+		}
+	}
+}
+
+func TestSchedulerPick_MixedProvidersUsesProviderRotationOverReadyCandidates(t *testing.T) {
+	t.Parallel()
+
+	scheduler := newSchedulerForTest(
+		&RoundRobinSelector{},
+		&Auth{ID: "gemini-a", Provider: "gemini"},
+		&Auth{ID: "gemini-b", Provider: "gemini"},
+		&Auth{ID: "claude-a", Provider: "claude"},
+	)
+
+	wantProviders := []string{"gemini", "claude", "gemini", "claude"}
+	wantIDs := []string{"gemini-a", "claude-a", "gemini-b", "claude-a"}
+	for index := range wantProviders {
+		got, provider, errPick := scheduler.pickMixed(context.Background(), []string{"gemini", "claude"}, "", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickMixed() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickMixed() #%d auth = nil", index)
+		}
+		if provider != wantProviders[index] {
+			t.Fatalf("pickMixed() #%d provider = %q, want %q", index, provider, wantProviders[index])
+		}
+		if got.ID != wantIDs[index] {
+			t.Fatalf("pickMixed() #%d auth.ID = %q, want %q", index, got.ID, wantIDs[index])
+		}
+	}
+}
+
+func TestSchedulerPick_MixedProvidersPrefersHighestPriorityTier(t *testing.T) {
+	t.Parallel()
+
+	model := "gpt-default"
+	registerSchedulerModels(t, "provider-low", model, "low")
+	registerSchedulerModels(t, "provider-high-a", model, "high-a")
+	registerSchedulerModels(t, "provider-high-b", model, "high-b")
+
+	scheduler := newSchedulerForTest(
+		&RoundRobinSelector{},
+		&Auth{ID: "low", Provider: "provider-low", Attributes: map[string]string{"priority": "4"}},
+		&Auth{ID: "high-a", Provider: "provider-high-a", Attributes: map[string]string{"priority": "7"}},
+		&Auth{ID: "high-b", Provider: "provider-high-b", Attributes: map[string]string{"priority": "7"}},
+	)
+
+	providers := []string{"provider-low", "provider-high-a", "provider-high-b"}
+	wantProviders := []string{"provider-high-a", "provider-high-b", "provider-high-a", "provider-high-b"}
+	wantIDs := []string{"high-a", "high-b", "high-a", "high-b"}
+	for index := range wantProviders {
+		got, provider, errPick := scheduler.pickMixed(context.Background(), providers, model, cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickMixed() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickMixed() #%d auth = nil", index)
+		}
+		if provider != wantProviders[index] {
+			t.Fatalf("pickMixed() #%d provider = %q, want %q", index, provider, wantProviders[index])
+		}
+		if got.ID != wantIDs[index] {
+			t.Fatalf("pickMixed() #%d auth.ID = %q, want %q", index, got.ID, wantIDs[index])
+		}
+	}
+}
+
+func TestManager_PickNextMixed_UsesProviderRotationBeforeCredentialRotation(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	manager.executors["gemini"] = schedulerTestExecutor{}
+	manager.executors["claude"] = schedulerTestExecutor{}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "gemini-a", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(gemini-a) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "gemini-b", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(gemini-b) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "claude-a", Provider: "claude"}); errRegister != nil {
+		t.Fatalf("Register(claude-a) error = %v", errRegister)
+	}
+
+	wantProviders := []string{"gemini", "claude", "gemini", "claude"}
+	wantIDs := []string{"gemini-a", "claude-a", "gemini-b", "claude-a"}
+	for index := range wantProviders {
+		got, _, provider, errPick := manager.pickNextMixed(context.Background(), []string{"gemini", "claude"}, "", cliproxyexecutor.Options{}, map[string]struct{}{})
+		if errPick != nil {
+			t.Fatalf("pickNextMixed() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickNextMixed() #%d auth = nil", index)
+		}
+		if provider != wantProviders[index] {
+			t.Fatalf("pickNextMixed() #%d provider = %q, want %q", index, provider, wantProviders[index])
+		}
+		if got.ID != wantIDs[index] {
+			t.Fatalf("pickNextMixed() #%d auth.ID = %q, want %q", index, got.ID, wantIDs[index])
+		}
+	}
+}
+
+func TestManagerCustomSelector_FallsBackToLegacyPath(t *testing.T) {
+	t.Parallel()
+
+	selector := &trackingSelector{}
+	manager := NewManager(nil, selector, nil)
+	manager.executors["gemini"] = schedulerTestExecutor{}
+	manager.auths["auth-a"] = &Auth{ID: "auth-a", Provider: "gemini"}
+	manager.auths["auth-b"] = &Auth{ID: "auth-b", Provider: "gemini"}
+
+	got, _, errPick := manager.pickNext(context.Background(), "gemini", "", cliproxyexecutor.Options{}, map[string]struct{}{})
+	if errPick != nil {
+		t.Fatalf("pickNext() error = %v", errPick)
+	}
+	if got == nil {
+		t.Fatalf("pickNext() auth = nil")
+	}
+	if selector.calls != 1 {
+		t.Fatalf("selector.calls = %d, want %d", selector.calls, 1)
+	}
+	if len(selector.lastAuthID) != 2 {
+		t.Fatalf("len(selector.lastAuthID) = %d, want %d", len(selector.lastAuthID), 2)
+	}
+	if got.ID != selector.lastAuthID[len(selector.lastAuthID)-1] {
+		t.Fatalf("pickNext() auth.ID = %q, want selector-picked %q", got.ID, selector.lastAuthID[len(selector.lastAuthID)-1])
+	}
+}
+
+func TestManager_InitializesSchedulerForBuiltInSelector(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	if manager.scheduler == nil {
+		t.Fatalf("manager.scheduler = nil")
+	}
+	if manager.scheduler.strategy != schedulerStrategyRoundRobin {
+		t.Fatalf("manager.scheduler.strategy = %v, want %v", manager.scheduler.strategy, schedulerStrategyRoundRobin)
+	}
+
+	manager.SetSelector(&FillFirstSelector{})
+	if manager.scheduler.strategy != schedulerStrategyFillFirst {
+		t.Fatalf("manager.scheduler.strategy = %v, want %v", manager.scheduler.strategy, schedulerStrategyFillFirst)
+	}
+}
+
+func TestManager_SchedulerTracksRegisterAndUpdate(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "auth-b", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(auth-b) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "auth-a", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(auth-a) error = %v", errRegister)
+	}
+
+	got, errPick := manager.scheduler.pickSingle(context.Background(), "gemini", "", cliproxyexecutor.Options{}, nil)
+	if errPick != nil {
+		t.Fatalf("scheduler.pickSingle() error = %v", errPick)
+	}
+	if got == nil || got.ID != "auth-a" {
+		t.Fatalf("scheduler.pickSingle() auth = %v, want auth-a", got)
+	}
+
+	if _, errUpdate := manager.Update(context.Background(), &Auth{ID: "auth-a", Provider: "gemini", Disabled: true}); errUpdate != nil {
+		t.Fatalf("Update(auth-a) error = %v", errUpdate)
+	}
+
+	got, errPick = manager.scheduler.pickSingle(context.Background(), "gemini", "", cliproxyexecutor.Options{}, nil)
+	if errPick != nil {
+		t.Fatalf("scheduler.pickSingle() after update error = %v", errPick)
+	}
+	if got == nil || got.ID != "auth-b" {
+		t.Fatalf("scheduler.pickSingle() after update auth = %v, want auth-b", got)
+	}
+}
+
+func TestManager_PickNextMixed_UsesSchedulerRotation(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	manager.executors["gemini"] = schedulerTestExecutor{}
+	manager.executors["claude"] = schedulerTestExecutor{}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "gemini-a", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(gemini-a) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "gemini-b", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(gemini-b) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "claude-a", Provider: "claude"}); errRegister != nil {
+		t.Fatalf("Register(claude-a) error = %v", errRegister)
+	}
+
+	wantProviders := []string{"gemini", "claude", "gemini", "claude"}
+	wantIDs := []string{"gemini-a", "claude-a", "gemini-b", "claude-a"}
+	for index := range wantProviders {
+		got, _, provider, errPick := manager.pickNextMixed(context.Background(), []string{"gemini", "claude"}, "", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickNextMixed() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickNextMixed() #%d auth = nil", index)
+		}
+		if provider != wantProviders[index] {
+			t.Fatalf("pickNextMixed() #%d provider = %q, want %q", index, provider, wantProviders[index])
+		}
+		if got.ID != wantIDs[index] {
+			t.Fatalf("pickNextMixed() #%d auth.ID = %q, want %q", index, got.ID, wantIDs[index])
+		}
+	}
+}
+
+func TestManager_PickNextMixed_SkipsProvidersWithoutExecutors(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	manager.executors["claude"] = schedulerTestExecutor{}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "gemini-a", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(gemini-a) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "claude-a", Provider: "claude"}); errRegister != nil {
+		t.Fatalf("Register(claude-a) error = %v", errRegister)
+	}
+
+	got, _, provider, errPick := manager.pickNextMixed(context.Background(), []string{"gemini", "claude"}, "", cliproxyexecutor.Options{}, nil)
+	if errPick != nil {
+		t.Fatalf("pickNextMixed() error = %v", errPick)
+	}
+	if got == nil {
+		t.Fatalf("pickNextMixed() auth = nil")
+	}
+	if provider != "claude" {
+		t.Fatalf("pickNextMixed() provider = %q, want %q", provider, "claude")
+	}
+	if got.ID != "claude-a" {
+		t.Fatalf("pickNextMixed() auth.ID = %q, want %q", got.ID, "claude-a")
+	}
+}
+
+func TestManager_SchedulerTracksMarkResultCooldownAndRecovery(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient("auth-a", "gemini", []*registry.ModelInfo{{ID: "test-model"}})
+	reg.RegisterClient("auth-b", "gemini", []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		reg.UnregisterClient("auth-a")
+		reg.UnregisterClient("auth-b")
+	})
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "auth-a", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(auth-a) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "auth-b", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(auth-b) error = %v", errRegister)
+	}
+
+	manager.MarkResult(context.Background(), Result{
+		AuthID:   "auth-a",
+		Provider: "gemini",
+		Model:    "test-model",
+		Success:  false,
+		Error:    &Error{HTTPStatus: 429, Message: "quota"},
+	})
+
+	got, errPick := manager.scheduler.pickSingle(context.Background(), "gemini", "test-model", cliproxyexecutor.Options{}, nil)
+	if errPick != nil {
+		t.Fatalf("scheduler.pickSingle() after cooldown error = %v", errPick)
+	}
+	if got == nil || got.ID != "auth-b" {
+		t.Fatalf("scheduler.pickSingle() after cooldown auth = %v, want auth-b", got)
+	}
+
+	manager.MarkResult(context.Background(), Result{
+		AuthID:   "auth-a",
+		Provider: "gemini",
+		Model:    "test-model",
+		Success:  true,
+	})
+
+	seen := make(map[string]struct{}, 2)
+	for index := 0; index < 2; index++ {
+		got, errPick = manager.scheduler.pickSingle(context.Background(), "gemini", "test-model", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("scheduler.pickSingle() after recovery #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("scheduler.pickSingle() after recovery #%d auth = nil", index)
+		}
+		seen[got.ID] = struct{}{}
+	}
+	if len(seen) != 2 {
+		t.Fatalf("len(seen) = %d, want %d", len(seen), 2)
+	}
+}

sdk/cliproxy/service.go

@@ -301,6 +301,9 @@ func (s *Service) applyCoreAuthAddOrUpdate(ctx context.Context, auth *coreauth.A
 		auth.CreatedAt = existing.CreatedAt
 		auth.LastRefreshedAt = existing.LastRefreshedAt
 		auth.NextRefreshAfter = existing.NextRefreshAfter
+		if len(auth.ModelStates) == 0 && len(existing.ModelStates) > 0 {
+			auth.ModelStates = existing.ModelStates
+		}
 		op = "update"
 		_, err = s.coreManager.Update(ctx, auth)
 	} else {
@@ -320,6 +323,12 @@ func (s *Service) applyCoreAuthAddOrUpdate(ctx context.Context, auth *coreauth.A
 	// This operation may block on network calls, but the auth configuration
 	// is already effective at this point.
 	s.registerModelsForAuth(auth)
+
+	// Refresh the scheduler entry so that the auth's supportedModelSet is rebuilt
+	// from the now-populated global model registry. Without this, newly added auths
+	// have an empty supportedModelSet (because Register/Update upserts into the
+	// scheduler before registerModelsForAuth runs) and are invisible to the scheduler.
+	s.coreManager.RefreshSchedulerEntry(auth.ID)
 }
 
 func (s *Service) applyCoreAuthRemoval(ctx context.Context, id string) {
@@ -817,10 +826,13 @@ func (s *Service) registerModelsForAuth(a *coreauth.Auth) {
 	case "vertex":
 		// Vertex AI Gemini supports the same model identifiers as Gemini.
 		models = registry.GetGeminiVertexModels()
-		if authKind == "apikey" {
-			if entry := s.resolveConfigVertexCompatKey(a); entry != nil && len(entry.Models) > 0 {
+		if entry := s.resolveConfigVertexCompatKey(a); entry != nil {
+			if len(entry.Models) > 0 {
 				models = buildVertexCompatConfigModels(entry)
 			}
+			if authKind == "apikey" {
+				excluded = entry.ExcludedModels
+			}
 		}
 		models = applyExcludedModels(models, excluded)
 	case "gemini-cli":
@@ -846,7 +858,22 @@ func (s *Service) registerModelsForAuth(a *coreauth.Auth) {
 		}
 		models = applyExcludedModels(models, excluded)
 	case "codex":
-		models = registry.GetOpenAIModels()
+		codexPlanType := ""
+		if a.Attributes != nil {
+			codexPlanType = strings.TrimSpace(a.Attributes["plan_type"])
+		}
+		switch strings.ToLower(codexPlanType) {
+		case "pro":
+			models = registry.GetCodexProModels()
+		case "plus":
+			models = registry.GetCodexPlusModels()
+		case "team":
+			models = registry.GetCodexTeamModels()
+		case "free":
+			models = registry.GetCodexFreeModels()
+		default:
+			models = registry.GetCodexProModels()
+		}
 		if entry := s.resolveConfigCodexKey(a); entry != nil {
 			if len(entry.Models) > 0 {
 				models = buildCodexConfigModels(entry)
@@ -866,7 +893,9 @@ func (s *Service) registerModelsForAuth(a *coreauth.Auth) {
 		models = registry.GetKimiModels()
     models = applyExcludedModels(models, excluded)
 	case "github-copilot":
-		models = registry.GetGitHubCopilotModels()
+		ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second)
+		defer cancel()
+		models = executor.FetchGitHubCopilotModels(ctx, a, s.cfg)
 		models = applyExcludedModels(models, excluded)
 	case "kiro":
 		models = s.fetchKiroModels(a)