refactor(executor, handlers): replace channel-based streams with StreamResult for consistency

- Updated `ExecuteStream` functions in executors to use `StreamResult` instead of channels. - Enhanced upstream header handling in OpenAI handlers. - Improved maintainability and alignment across executors and handlers.
Merge pull request #249 from router-for-me/plus
2026-03-12 08:43:58 +00:00 · 2026-02-19 22:07:14 +08:00 · 2026-02-19 21:58:42 +08:00 · 2026-02-19 21:58:04 +08:00 · 2026-02-19 21:49:44 +08:00 · 2026-02-19 21:33:23 +08:00
112 changed files with 13136 additions and 1709 deletions
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@@ -19,7 +19,7 @@ jobs:
      - run: git fetch --force --tags
      - uses: actions/setup-go@v4
        with:
-          go-version: '>=1.24.0'
+          go-version: '>=1.26.0'
          cache: true
      - name: Generate Build Metadata
        run: |
--- a/.gitignore
+++ b/.gitignore
@@ -3,10 +3,11 @@ cli-proxy-api
 cliproxy
 *.exe

+
 # Configuration
 config.yaml
 .env
-
+.mcp.json
 # Generated content
 bin/*
 logs/*
--- a/2
+++ b/2
@@ -1,4 +1,4 @@
-FROM golang:1.24-alpine AS builder
+FROM golang:1.26-alpine AS builder

 WORKDIR /app

--- a/cmd/server/main.go
+++ b/cmd/server/main.go
@@ -8,6 +8,7 @@ import (
 	"errors"
 	"flag"
 	"fmt"
+	"io"
 	"io/fs"
 	"net/url"
 	"os"
@@ -26,6 +27,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
 	"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"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/usage"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	sdkAuth "github.com/router-for-me/CLIProxyAPI/v6/sdk/auth"
@@ -72,6 +74,7 @@ func main() {
 	var codexLogin bool
 	var claudeLogin bool
 	var qwenLogin bool
+	var kiloLogin bool
 	var iflowLogin bool
 	var iflowCookie bool
 	var noBrowser bool
@@ -88,6 +91,8 @@ func main() {
 	var vertexImport string
 	var configPath string
 	var password string
+	var tuiMode bool
+	var standalone bool
 	var noIncognito bool
 	var useIncognito bool

@@ -96,6 +101,7 @@ func main() {
 	flag.BoolVar(&codexLogin, "codex-login", false, "Login to Codex using OAuth")
 	flag.BoolVar(&claudeLogin, "claude-login", false, "Login to Claude using OAuth")
 	flag.BoolVar(&qwenLogin, "qwen-login", false, "Login to Qwen using OAuth")
+	flag.BoolVar(&kiloLogin, "kilo-login", false, "Login to Kilo AI using device flow")
 	flag.BoolVar(&iflowLogin, "iflow-login", false, "Login to iFlow using OAuth")
 	flag.BoolVar(&iflowCookie, "iflow-cookie", false, "Login to iFlow using Cookie")
 	flag.BoolVar(&noBrowser, "no-browser", false, "Don't open browser automatically for OAuth")
@@ -114,6 +120,8 @@ func main() {
 	flag.StringVar(&configPath, "config", DefaultConfigPath, "Configure File Path")
 	flag.StringVar(&vertexImport, "vertex-import", "", "Import Vertex service account key JSON file")
 	flag.StringVar(&password, "password", "", "")
+	flag.BoolVar(&tuiMode, "tui", false, "Start with terminal management UI")
+	flag.BoolVar(&standalone, "standalone", false, "In TUI mode, start an embedded local server")

 	flag.CommandLine.Usage = func() {
 		out := flag.CommandLine.Output()
@@ -499,6 +507,8 @@ func main() {
 		cmd.DoClaudeLogin(cfg, options)
 	} else if qwenLogin {
 		cmd.DoQwenLogin(cfg, options)
+	} else if kiloLogin {
+		cmd.DoKiloLogin(cfg, options)
 	} else if iflowLogin {
 		cmd.DoIFlowLogin(cfg, options)
 	} else if iflowCookie {
@@ -536,15 +546,89 @@ func main() {
 			cmd.WaitForCloudDeploy()
 			return
 		}
-		// Start the main proxy service
-		managementasset.StartAutoUpdater(context.Background(), configFilePath)
+		if tuiMode {
+			if standalone {
+				// Standalone mode: start an embedded local server and connect TUI client to it.
+				managementasset.StartAutoUpdater(context.Background(), configFilePath)
+				hook := tui.NewLogHook(2000)
+				hook.SetFormatter(&logging.LogFormatter{})
+				log.AddHook(hook)

-		// 初始化并启动 Kiro token 后台刷新
-		if cfg.AuthDir != "" {
-			kiro.InitializeAndStart(cfg.AuthDir, cfg)
-			defer kiro.StopGlobalRefreshManager()
+				origStdout := os.Stdout
+				origStderr := os.Stderr
+				origLogOutput := log.StandardLogger().Out
+				log.SetOutput(io.Discard)
+
+				devNull, errOpenDevNull := os.Open(os.DevNull)
+				if errOpenDevNull == nil {
+					os.Stdout = devNull
+					os.Stderr = devNull
+				}
+
+				restoreIO := func() {
+					os.Stdout = origStdout
+					os.Stderr = origStderr
+					log.SetOutput(origLogOutput)
+					if devNull != nil {
+						_ = devNull.Close()
+					}
+				}
+
+				localMgmtPassword := fmt.Sprintf("tui-%d-%d", os.Getpid(), time.Now().UnixNano())
+				if password == "" {
+					password = localMgmtPassword
+				}
+
+				cancel, done := cmd.StartServiceBackground(cfg, configFilePath, password)
+
+				client := tui.NewClient(cfg.Port, password)
+				ready := false
+				backoff := 100 * time.Millisecond
+				for i := 0; i < 30; i++ {
+					if _, errGetConfig := client.GetConfig(); errGetConfig == nil {
+						ready = true
+						break
+					}
+					time.Sleep(backoff)
+					if backoff < time.Second {
+						backoff = time.Duration(float64(backoff) * 1.5)
+					}
+				}
+
+				if !ready {
+					restoreIO()
+					cancel()
+					<-done
+					fmt.Fprintf(os.Stderr, "TUI error: embedded server is not ready\n")
+					return
+				}
+
+				if errRun := tui.Run(cfg.Port, password, hook, origStdout); errRun != nil {
+					restoreIO()
+					fmt.Fprintf(os.Stderr, "TUI error: %v\n", errRun)
+				} else {
+					restoreIO()
+				}
+
+				cancel()
+				<-done
+			} else {
+				// Default TUI mode: pure management client.
+				// The proxy server must already be running.
+				if errRun := tui.Run(cfg.Port, password, nil, os.Stdout); errRun != nil {
+					fmt.Fprintf(os.Stderr, "TUI error: %v\n", errRun)
+				}
+			}
+		} else {
+      // Start the main proxy service
+      managementasset.StartAutoUpdater(context.Background(), configFilePath)
+
+      if cfg.AuthDir != "" {
+        kiro.InitializeAndStart(cfg.AuthDir, cfg)
+        defer kiro.StopGlobalRefreshManager()
+      }
+
+      cmd.StartService(cfg, configFilePath, password)
 		}
-
-		cmd.StartService(cfg, configFilePath, password)
 	}
 }
--- a/config.example.yaml
+++ b/config.example.yaml
@@ -1,6 +1,6 @@
 # Server host/interface to bind to. Default is empty ("") to bind all interfaces (IPv4 + IPv6).
 # Use "127.0.0.1" or "localhost" to restrict access to local machine only.
-host: ""
+host: ''

 # Server port
 port: 8317
@@ -8,8 +8,8 @@ port: 8317
 # TLS settings for HTTPS. When enabled, the server listens with the provided certificate and key.
 tls:
  enable: false
-  cert: ""
-  key: ""
+  cert: ''
+  key: ''

 # Management API settings
 remote-management:
@@ -20,22 +20,22 @@ remote-management:
  # Management key. If a plaintext value is provided here, it will be hashed on startup.
  # All management requests (even from localhost) require this key.
  # Leave empty to disable the Management API entirely (404 for all /v0/management routes).
-  secret-key: ""
+  secret-key: ''

  # Disable the bundled management control panel asset download and HTTP route when true.
  disable-control-panel: false

  # GitHub repository for the management control panel. Accepts a repository URL or releases API URL.
-  panel-github-repository: "https://github.com/router-for-me/Cli-Proxy-API-Management-Center"
+  panel-github-repository: 'https://github.com/router-for-me/Cli-Proxy-API-Management-Center'

 # Authentication directory (supports ~ for home directory)
-auth-dir: "~/.cli-proxy-api"
+auth-dir: '~/.cli-proxy-api'

 # API keys for authentication
 api-keys:
-  - "your-api-key-1"
-  - "your-api-key-2"
-  - "your-api-key-3"
+  - 'your-api-key-1'
+  - 'your-api-key-2'
+  - 'your-api-key-3'

 # Enable debug logging
 debug: false
@@ -43,7 +43,7 @@ debug: false
 # Enable pprof HTTP debug server (host:port). Keep it bound to localhost for safety.
 pprof:
  enable: false
-  addr: "127.0.0.1:8316"
+  addr: '127.0.0.1:8316'

 # When true, disable high-overhead HTTP middleware features to reduce per-request memory usage under high concurrency.
 commercial-mode: false
@@ -68,11 +68,15 @@ error-logs-max-files: 10
 usage-statistics-enabled: false

 # Proxy URL. Supports socks5/http/https protocols. Example: socks5://user:pass@192.168.1.1:1080/
-proxy-url: ""
+proxy-url: ''

 # When true, unprefixed model requests only use credentials without a prefix (except when prefix == model name).
 force-model-prefix: false

+# When true, forward filtered upstream response headers to downstream clients.
+# Default is false (disabled).
+passthrough-headers: false
+
 # Number of times to retry a request. Retries will occur if the HTTP response code is 403, 408, 500, 502, 503, or 504.
 request-retry: 3

@@ -86,7 +90,7 @@ quota-exceeded:

 # Routing strategy for selecting credentials when multiple match.
 routing:
-  strategy: "round-robin" # round-robin (default), fill-first
+  strategy: 'round-robin' # round-robin (default), fill-first

 # When true, enable authentication for the WebSocket API (/v1/ws).
 ws-auth: false
@@ -161,6 +165,14 @@ nonstream-keepalive-interval: 0
 #         - "API"
 #         - "proxy"

+# Default headers for Claude API requests. Update when Claude Code releases new versions.
+# These are used as fallbacks when the client does not send its own headers.
+# claude-header-defaults:
+#   user-agent: "claude-cli/2.1.44 (external, sdk-cli)"
+#   package-version: "0.74.0"
+#   runtime-version: "v24.3.0"
+#   timeout: "600"
+
 # Kiro (AWS CodeWhisperer) configuration
 # Note: Kiro API currently only operates in us-east-1 region
 #kiro:
@@ -171,6 +183,21 @@ nonstream-keepalive-interval: 0
 #    profile-arn: "arn:aws:codewhisperer:us-east-1:..."
 #    proxy-url: "socks5://proxy.example.com:1080" # optional: proxy override

+# Kilocode (OAuth-based code assistant)
+# Note: Kilocode uses OAuth device flow authentication.
+# Use the CLI command: ./server --kilo-login
+# This will save credentials to the auth directory (default: ~/.cli-proxy-api/)
+# oauth-model-alias:
+#   kilo:
+#     - name: "minimax/minimax-m2.5:free"
+#       alias: "minimax-m2.5"
+#     - name: "z-ai/glm-5:free"
+#       alias: "glm-5"
+# oauth-excluded-models:
+#   kilo:
+#     - "kilo-claude-opus-4-6"  # exclude specific models (exact match)
+#     - "*:free"                # wildcard matching suffix (e.g. all free models)
+
 # OpenAI compatibility providers
 # openai-compatibility:
 #   - name: "openrouter" # The name of the provider; it will be used in the user agent and other places.
--- a/examples/custom-provider/main.go
+++ b/examples/custom-provider/main.go
@@ -159,13 +159,13 @@ func (MyExecutor) CountTokens(context.Context, *coreauth.Auth, clipexec.Request,
 	return clipexec.Response{}, errors.New("count tokens not implemented")
 }

-func (MyExecutor) ExecuteStream(ctx context.Context, a *coreauth.Auth, req clipexec.Request, opts clipexec.Options) (<-chan clipexec.StreamChunk, error) {
+func (MyExecutor) ExecuteStream(ctx context.Context, a *coreauth.Auth, req clipexec.Request, opts clipexec.Options) (*clipexec.StreamResult, error) {
 	ch := make(chan clipexec.StreamChunk, 1)
 	go func() {
 		defer close(ch)
 		ch <- clipexec.StreamChunk{Payload: []byte("data: {\"ok\":true}\n\n")}
 	}()
-	return ch, nil
+	return &clipexec.StreamResult{Chunks: ch}, nil
 }

 func (MyExecutor) Refresh(ctx context.Context, a *coreauth.Auth) (*coreauth.Auth, error) {
--- a/examples/http-request/main.go
+++ b/examples/http-request/main.go
@@ -58,7 +58,7 @@ func (EchoExecutor) Execute(context.Context, *coreauth.Auth, clipexec.Request, c
 	return clipexec.Response{}, errors.New("echo executor: Execute not implemented")
 }

-func (EchoExecutor) ExecuteStream(context.Context, *coreauth.Auth, clipexec.Request, clipexec.Options) (<-chan clipexec.StreamChunk, error) {
+func (EchoExecutor) ExecuteStream(context.Context, *coreauth.Auth, clipexec.Request, clipexec.Options) (*clipexec.StreamResult, error) {
 	return nil, errors.New("echo executor: ExecuteStream not implemented")
 }

--- a/go.mod
+++ b/go.mod
@@ -1,9 +1,13 @@
 module github.com/router-for-me/CLIProxyAPI/v6

-go 1.24.0
+go 1.26.0

 require (
 	github.com/andybalholm/brotli v1.0.6
+	github.com/atotto/clipboard v0.1.4
+	github.com/charmbracelet/bubbles v1.0.0
+	github.com/charmbracelet/bubbletea v1.3.10
+	github.com/charmbracelet/lipgloss v1.1.0
 	github.com/fsnotify/fsnotify v1.9.0
 	github.com/fxamacker/cbor/v2 v2.9.0
 	github.com/gin-gonic/gin v1.10.1
@@ -33,8 +37,16 @@ require (
 	cloud.google.com/go/compute/metadata v0.3.0 // indirect
 	github.com/Microsoft/go-winio v0.6.2 // indirect
 	github.com/ProtonMail/go-crypto v1.3.0 // indirect
+	github.com/aymanbagabas/go-osc52/v2 v2.0.1 // indirect
 	github.com/bytedance/sonic v1.11.6 // indirect
 	github.com/bytedance/sonic/loader v0.1.1 // indirect
+	github.com/charmbracelet/colorprofile v0.4.1 // indirect
+	github.com/charmbracelet/x/ansi v0.11.6 // indirect
+	github.com/charmbracelet/x/cellbuf v0.0.15 // indirect
+	github.com/charmbracelet/x/term v0.2.2 // indirect
+	github.com/clipperhouse/displaywidth v0.9.0 // indirect
+	github.com/clipperhouse/stringish v0.1.1 // indirect
+	github.com/clipperhouse/uax29/v2 v2.5.0 // indirect
 	github.com/cloudflare/circl v1.6.1 // indirect
 	github.com/cloudwego/base64x v0.1.4 // indirect
 	github.com/cloudwego/iasm v0.2.0 // indirect
@@ -42,6 +54,7 @@ require (
 	github.com/dlclark/regexp2 v1.11.5 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/emirpasic/gods v1.18.1 // indirect
+	github.com/erikgeiser/coninput v0.0.0-20211004153227-1c3628e74d0f // indirect
 	github.com/gabriel-vasile/mimetype v1.4.3 // indirect
 	github.com/gin-contrib/sse v0.1.0 // indirect
 	github.com/go-git/gcfg/v2 v2.0.2 // indirect
@@ -58,19 +71,27 @@ require (
 	github.com/kevinburke/ssh_config v1.4.0 // indirect
 	github.com/klauspost/cpuid/v2 v2.3.0 // indirect
 	github.com/leodido/go-urn v1.4.0 // indirect
+	github.com/lucasb-eyer/go-colorful v1.3.0 // indirect
 	github.com/mattn/go-isatty v0.0.20 // indirect
+	github.com/mattn/go-localereader v0.0.1 // indirect
+	github.com/mattn/go-runewidth v0.0.19 // indirect
 	github.com/minio/md5-simd v1.1.2 // indirect
 	github.com/minio/sha256-simd v1.0.1 // indirect
 	github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd // indirect
 	github.com/modern-go/reflect2 v1.0.2 // indirect
+	github.com/muesli/ansi v0.0.0-20230316100256-276c6243b2f6 // indirect
+	github.com/muesli/cancelreader v0.2.2 // indirect
+	github.com/muesli/termenv v0.16.0 // indirect
 	github.com/pelletier/go-toml/v2 v2.2.2 // indirect
 	github.com/pjbgf/sha1cd v0.5.0 // indirect
+	github.com/rivo/uniseg v0.4.7 // indirect
 	github.com/rs/xid v1.5.0 // indirect
 	github.com/sergi/go-diff v1.4.0 // indirect
 	github.com/tidwall/match v1.1.1 // indirect
 	github.com/tidwall/pretty v1.2.0 // indirect
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.12 // indirect
+	github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e // indirect
 	github.com/x448/float16 v0.8.4 // indirect
 	golang.org/x/arch v0.8.0 // indirect
 	golang.org/x/sys v0.38.0 // indirect
--- a/go.sum
+++ b/go.sum
@@ -10,10 +10,34 @@ github.com/anmitsu/go-shlex v0.0.0-20200514113438-38f4b401e2be h1:9AeTilPcZAjCFI
 github.com/anmitsu/go-shlex v0.0.0-20200514113438-38f4b401e2be/go.mod h1:ySMOLuWl6zY27l47sB3qLNK6tF2fkHG55UZxx8oIVo4=
 github.com/armon/go-socks5 v0.0.0-20160902184237-e75332964ef5 h1:0CwZNZbxp69SHPdPJAN/hZIm0C4OItdklCFmMRWYpio=
 github.com/armon/go-socks5 v0.0.0-20160902184237-e75332964ef5/go.mod h1:wHh0iHkYZB8zMSxRWpUBQtwG5a7fFgvEO+odwuTv2gs=
+github.com/atotto/clipboard v0.1.4 h1:EH0zSVneZPSuFR11BlR9YppQTVDbh5+16AmcJi4g1z4=
+github.com/atotto/clipboard v0.1.4/go.mod h1:ZY9tmq7sm5xIbd9bOK4onWV4S6X0u6GY7Vn0Yu86PYI=
+github.com/aymanbagabas/go-osc52/v2 v2.0.1 h1:HwpRHbFMcZLEVr42D4p7XBqjyuxQH5SMiErDT4WkJ2k=
+github.com/aymanbagabas/go-osc52/v2 v2.0.1/go.mod h1:uYgXzlJ7ZpABp8OJ+exZzJJhRNQ2ASbcXHWsFqH8hp8=
 github.com/bytedance/sonic v1.11.6 h1:oUp34TzMlL+OY1OUWxHqsdkgC/Zfc85zGqw9siXjrc0=
 github.com/bytedance/sonic v1.11.6/go.mod h1:LysEHSvpvDySVdC2f87zGWf6CIKJcAvqab1ZaiQtds4=
 github.com/bytedance/sonic/loader v0.1.1 h1:c+e5Pt1k/cy5wMveRDyk2X4B9hF4g7an8N3zCYjJFNM=
 github.com/bytedance/sonic/loader v0.1.1/go.mod h1:ncP89zfokxS5LZrJxl5z0UJcsk4M4yY2JpfqGeCtNLU=
+github.com/charmbracelet/bubbles v1.0.0 h1:12J8/ak/uCZEMQ6KU7pcfwceyjLlWsDLAxB5fXonfvc=
+github.com/charmbracelet/bubbles v1.0.0/go.mod h1:9d/Zd5GdnauMI5ivUIVisuEm3ave1XwXtD1ckyV6r3E=
+github.com/charmbracelet/bubbletea v1.3.10 h1:otUDHWMMzQSB0Pkc87rm691KZ3SWa4KUlvF9nRvCICw=
+github.com/charmbracelet/bubbletea v1.3.10/go.mod h1:ORQfo0fk8U+po9VaNvnV95UPWA1BitP1E0N6xJPlHr4=
+github.com/charmbracelet/colorprofile v0.4.1 h1:a1lO03qTrSIRaK8c3JRxJDZOvhvIeSco3ej+ngLk1kk=
+github.com/charmbracelet/colorprofile v0.4.1/go.mod h1:U1d9Dljmdf9DLegaJ0nGZNJvoXAhayhmidOdcBwAvKk=
+github.com/charmbracelet/lipgloss v1.1.0 h1:vYXsiLHVkK7fp74RkV7b2kq9+zDLoEU4MZoFqR/noCY=
+github.com/charmbracelet/lipgloss v1.1.0/go.mod h1:/6Q8FR2o+kj8rz4Dq0zQc3vYf7X+B0binUUBwA0aL30=
+github.com/charmbracelet/x/ansi v0.11.6 h1:GhV21SiDz/45W9AnV2R61xZMRri5NlLnl6CVF7ihZW8=
+github.com/charmbracelet/x/ansi v0.11.6/go.mod h1:2JNYLgQUsyqaiLovhU2Rv/pb8r6ydXKS3NIttu3VGZQ=
+github.com/charmbracelet/x/cellbuf v0.0.15 h1:ur3pZy0o6z/R7EylET877CBxaiE1Sp1GMxoFPAIztPI=
+github.com/charmbracelet/x/cellbuf v0.0.15/go.mod h1:J1YVbR7MUuEGIFPCaaZ96KDl5NoS0DAWkskup+mOY+Q=
+github.com/charmbracelet/x/term v0.2.2 h1:xVRT/S2ZcKdhhOuSP4t5cLi5o+JxklsoEObBSgfgZRk=
+github.com/charmbracelet/x/term v0.2.2/go.mod h1:kF8CY5RddLWrsgVwpw4kAa6TESp6EB5y3uxGLeCqzAI=
+github.com/clipperhouse/displaywidth v0.9.0 h1:Qb4KOhYwRiN3viMv1v/3cTBlz3AcAZX3+y9OLhMtAtA=
+github.com/clipperhouse/displaywidth v0.9.0/go.mod h1:aCAAqTlh4GIVkhQnJpbL0T/WfcrJXHcj8C0yjYcjOZA=
+github.com/clipperhouse/stringish v0.1.1 h1:+NSqMOr3GR6k1FdRhhnXrLfztGzuG+VuFDfatpWHKCs=
+github.com/clipperhouse/stringish v0.1.1/go.mod h1:v/WhFtE1q0ovMta2+m+UbpZ+2/HEXNWYXQgCt4hdOzA=
+github.com/clipperhouse/uax29/v2 v2.5.0 h1:x7T0T4eTHDONxFJsL94uKNKPHrclyFI0lm7+w94cO8U=
+github.com/clipperhouse/uax29/v2 v2.5.0/go.mod h1:Wn1g7MK6OoeDT0vL+Q0SQLDz/KpfsVRgg6W7ihQeh4g=
 github.com/cloudflare/circl v1.6.1 h1:zqIqSPIndyBh1bjLVVDHMPpVKqp8Su/V+6MeDzzQBQ0=
 github.com/cloudflare/circl v1.6.1/go.mod h1:uddAzsPgqdMAYatqJ0lsjX1oECcQLIlRpzZh3pJrofs=
 github.com/cloudwego/base64x v0.1.4 h1:jwCgWpFanWmN8xoIUHa2rtzmkd5J2plF/dnLS6Xd/0Y=
@@ -33,6 +57,8 @@ github.com/elazarl/goproxy v1.7.2 h1:Y2o6urb7Eule09PjlhQRGNsqRfPmYI3KKQLFpCAV3+o
 github.com/elazarl/goproxy v1.7.2/go.mod h1:82vkLNir0ALaW14Rc399OTTjyNREgmdL2cVoIbS6XaE=
 github.com/emirpasic/gods v1.18.1 h1:FXtiHYKDGKCW2KzwZKx0iC0PQmdlorYgdFG9jPXJ1Bc=
 github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FMNAnJvWQ=
+github.com/erikgeiser/coninput v0.0.0-20211004153227-1c3628e74d0f h1:Y/CXytFA4m6baUTXGLOoWe4PQhGxaX0KpnayAqC48p4=
+github.com/erikgeiser/coninput v0.0.0-20211004153227-1c3628e74d0f/go.mod h1:vw97MGsxSvLiUE2X8qFplwetxpGLQrlU1Q9AUEIzCaM=
 github.com/fsnotify/fsnotify v1.9.0 h1:2Ml+OJNzbYCTzsxtv8vKSFD9PbJjmhYF14k/jKC7S9k=
 github.com/fsnotify/fsnotify v1.9.0/go.mod h1:8jBTzvmWwFyi3Pb8djgCCO5IBqzKJ/Jwo8TRcHyHii0=
 github.com/fxamacker/cbor/v2 v2.9.0 h1:NpKPmjDBgUfBms6tr6JZkTHtfFGcMKsw3eGcmD/sapM=
@@ -101,8 +127,14 @@ github.com/kr/text v0.1.0 h1:45sCR5RtlFHMR4UwH9sdQ5TC8v0qDQCHnXt+kaKSTVE=
 github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI=
 github.com/leodido/go-urn v1.4.0 h1:WT9HwE9SGECu3lg4d/dIA+jxlljEa1/ffXKmRjqdmIQ=
 github.com/leodido/go-urn v1.4.0/go.mod h1:bvxc+MVxLKB4z00jd1z+Dvzr47oO32F/QSNjSBOlFxI=
+github.com/lucasb-eyer/go-colorful v1.3.0 h1:2/yBRLdWBZKrf7gB40FoiKfAWYQ0lqNcbuQwVHXptag=
+github.com/lucasb-eyer/go-colorful v1.3.0/go.mod h1:R4dSotOR9KMtayYi1e77YzuveK+i7ruzyGqttikkLy0=
 github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWEY=
 github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
+github.com/mattn/go-localereader v0.0.1 h1:ygSAOl7ZXTx4RdPYinUpg6W99U8jWvWi9Ye2JC/oIi4=
+github.com/mattn/go-localereader v0.0.1/go.mod h1:8fBrzywKY7BI3czFoHkuzRoWE9C+EiG4R1k4Cjx5p88=
+github.com/mattn/go-runewidth v0.0.19 h1:v++JhqYnZuu5jSKrk9RbgF5v4CGUjqRfBm05byFGLdw=
+github.com/mattn/go-runewidth v0.0.19/go.mod h1:XBkDxAl56ILZc9knddidhrOlY5R/pDhgLpndooCuJAs=
 github.com/minio/md5-simd v1.1.2 h1:Gdi1DZK69+ZVMoNHRXJyNcxrMA4dSxoYHZSQbirFg34=
 github.com/minio/md5-simd v1.1.2/go.mod h1:MzdKDxYpY2BT9XQFocsiZf/NKVtR7nkE4RoEpN+20RM=
 github.com/minio/minio-go/v7 v7.0.66 h1:bnTOXOHjOqv/gcMuiVbN9o2ngRItvqE774dG9nq0Dzw=
@@ -114,6 +146,12 @@ github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd h1:TRLaZ9cD/w
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
 github.com/modern-go/reflect2 v1.0.2 h1:xBagoLtFs94CBntxluKeaWgTMpvLxC4ur3nMaC9Gz0M=
 github.com/modern-go/reflect2 v1.0.2/go.mod h1:yWuevngMOJpCy52FWWMvUC8ws7m/LJsjYzDa0/r8luk=
+github.com/muesli/ansi v0.0.0-20230316100256-276c6243b2f6 h1:ZK8zHtRHOkbHy6Mmr5D264iyp3TiX5OmNcI5cIARiQI=
+github.com/muesli/ansi v0.0.0-20230316100256-276c6243b2f6/go.mod h1:CJlz5H+gyd6CUWT45Oy4q24RdLyn7Md9Vj2/ldJBSIo=
+github.com/muesli/cancelreader v0.2.2 h1:3I4Kt4BQjOR54NavqnDogx/MIoWBFa0StPA8ELUXHmA=
+github.com/muesli/cancelreader v0.2.2/go.mod h1:3XuTXfFS2VjM+HTLZY9Ak0l6eUKfijIfMUZ4EgX0QYo=
+github.com/muesli/termenv v0.16.0 h1:S5AlUN9dENB57rsbnkPyfdGuWIlkmzJjbFf0Tf5FWUc=
+github.com/muesli/termenv v0.16.0/go.mod h1:ZRfOIKPFDYQoDFF4Olj7/QJbW60Ol/kL1pU3VfY/Cnk=
 github.com/pelletier/go-toml/v2 v2.2.2 h1:aYUidT7k73Pcl9nb2gScu7NSrKCSHIDE89b3+6Wq+LM=
 github.com/pelletier/go-toml/v2 v2.2.2/go.mod h1:1t835xjRzz80PqgE6HHgN2JOsmgYu/h4qDAS4n929Rs=
 github.com/pjbgf/sha1cd v0.5.0 h1:a+UkboSi1znleCDUNT3M5YxjOnN1fz2FhN48FlwCxs0=
@@ -124,6 +162,8 @@ github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZb
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/refraction-networking/utls v1.8.2 h1:j4Q1gJj0xngdeH+Ox/qND11aEfhpgoEvV+S9iJ2IdQo=
 github.com/refraction-networking/utls v1.8.2/go.mod h1:jkSOEkLqn+S/jtpEHPOsVv/4V4EVnelwbMQl4vCWXAM=
+github.com/rivo/uniseg v0.4.7 h1:WUdvkW8uEhrYfLC4ZzdpI2ztxP1I582+49Oc5Mq64VQ=
+github.com/rivo/uniseg v0.4.7/go.mod h1:FN3SvrM+Zdj16jyLfmOkMNblXMcoc8DfTHruCPUcx88=
 github.com/rogpeppe/go-internal v1.14.1 h1:UQB4HGPB6osV0SQTLymcB4TgvyWu6ZyliaW0tI/otEQ=
 github.com/rogpeppe/go-internal v1.14.1/go.mod h1:MaRKkUm5W0goXpeCfT7UZI6fk/L7L7so1lCWt35ZSgc=
 github.com/rs/xid v1.5.0 h1:mKX4bl4iPYJtEIxp6CYiUuLQ/8DYMoz0PUdtGgMFRVc=
@@ -161,6 +201,8 @@ github.com/twitchyliquid64/golang-asm v0.15.1 h1:SU5vSMR7hnwNxj24w34ZyCi/FmDZTkS
 github.com/twitchyliquid64/golang-asm v0.15.1/go.mod h1:a1lVb/DtPvCB8fslRZhAngC2+aY1QWCk3Cedj/Gdt08=
 github.com/ugorji/go/codec v1.2.12 h1:9LC83zGrHhuUA9l16C9AHXAqEV/2wBQ4nkvumAE65EE=
 github.com/ugorji/go/codec v1.2.12/go.mod h1:UNopzCgEMSXjBc6AOMqYvWC1ktqTAfzJZUZgYf6w6lg=
+github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e h1:JVG44RsyaB9T2KIHavMF/ppJZNG9ZpyihvCd0w101no=
+github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e/go.mod h1:RbqR21r5mrJuqunuUZ/Dhy/avygyECGrLceyNeo4LiM=
 github.com/x448/float16 v0.8.4 h1:qLwI1I70+NjRFUR3zs1JPUCgaCXSh3SW62uAKT1mSBM=
 github.com/x448/float16 v0.8.4/go.mod h1:14CWIYCyZA/cWjXOioeEpHeN/83MdbZDRQHoFcYsOfg=
 golang.org/x/arch v0.0.0-20210923205945-b76863e36670/go.mod h1:5om86z9Hs0C8fWVUuoMHwpExlXzs5Tkyp9hOrfG7pp8=
@@ -168,12 +210,15 @@ golang.org/x/arch v0.8.0 h1:3wRIsP3pM4yUptoR96otTUOXI367OS0+c9eeRi9doIc=
 golang.org/x/arch v0.8.0/go.mod h1:FEVrYAQjsQXMVJ1nsMoVVXPZg6p2JE2mx8psSWTDQys=
 golang.org/x/crypto v0.45.0 h1:jMBrvKuj23MTlT0bQEOBcAE0mjg8mK9RXFhRH6nyF3Q=
 golang.org/x/crypto v0.45.0/go.mod h1:XTGrrkGJve7CYK7J8PEww4aY7gM3qMCElcJQ8n8JdX4=
+golang.org/x/exp v0.0.0-20231006140011-7918f672742d h1:jtJma62tbqLibJ5sFQz8bKtEM8rJBtfilJ2qTU199MI=
+golang.org/x/exp v0.0.0-20231006140011-7918f672742d/go.mod h1:ldy0pHrwJyGW56pPQzzkH36rKxoZW1tw7ZJpeKx+hdo=
 golang.org/x/net v0.47.0 h1:Mx+4dIFzqraBXUugkia1OOvlD6LemFo1ALMHjrXDOhY=
 golang.org/x/net v0.47.0/go.mod h1:/jNxtkgq5yWUGYkaZGqo27cfGZ1c5Nen03aYrrKpVRU=
 golang.org/x/oauth2 v0.30.0 h1:dnDm7JmhM45NNpd8FDDeLhK6FwqbOf4MLCM9zb1BOHI=
 golang.org/x/oauth2 v0.30.0/go.mod h1:B++QgG3ZKulg6sRPGD/mqlHQs5rB3Ml9erfeDY7xKlU=
 golang.org/x/sync v0.18.0 h1:kr88TuHDroi+UVf+0hZnirlk8o8T+4MrK6mr60WkH/I=
 golang.org/x/sync v0.18.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
+golang.org/x/sys v0.0.0-20210809222454-d867a43fc93e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20220715151400-c0bba94af5f8/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.1.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
--- a/internal/api/handlers/management/api_tools.go
+++ b/internal/api/handlers/management/api_tools.go
@@ -1,6 +1,7 @@
 package management

 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"fmt"
@@ -189,9 +190,21 @@ func (h *Handler) APICall(c *gin.Context) {
 		reqHeaders[key] = strings.ReplaceAll(value, "$TOKEN$", token)
 	}

+	// When caller indicates CBOR in request headers, convert JSON string payload to CBOR bytes.
+	useCBORPayload := headerContainsValue(reqHeaders, "Content-Type", "application/cbor")
+
 	var requestBody io.Reader
 	if body.Data != "" {
-		requestBody = strings.NewReader(body.Data)
+		if useCBORPayload {
+			cborPayload, errEncode := encodeJSONStringToCBOR(body.Data)
+			if errEncode != nil {
+				c.JSON(http.StatusBadRequest, gin.H{"error": "invalid json data for cbor content-type"})
+				return
+			}
+			requestBody = bytes.NewReader(cborPayload)
+		} else {
+			requestBody = strings.NewReader(body.Data)
+		}
 	}

 	req, errNewRequest := http.NewRequestWithContext(c.Request.Context(), method, urlStr, requestBody)
@@ -234,10 +247,18 @@ func (h *Handler) APICall(c *gin.Context) {
 		return
 	}

+	// For CBOR upstream responses, decode into plain text or JSON string before returning.
+	responseBodyText := string(respBody)
+	if headerContainsValue(reqHeaders, "Accept", "application/cbor") || strings.Contains(strings.ToLower(resp.Header.Get("Content-Type")), "application/cbor") {
+		if decodedBody, errDecode := decodeCBORBodyToTextOrJSON(respBody); errDecode == nil {
+			responseBodyText = decodedBody
+		}
+	}
+
 	response := apiCallResponse{
 		StatusCode: resp.StatusCode,
 		Header:     resp.Header,
-		Body:       string(respBody),
+		Body:       responseBodyText,
 	}

 	// If this is a GitHub Copilot token endpoint response, try to enrich with quota information
@@ -747,6 +768,83 @@ func buildProxyTransport(proxyStr string) *http.Transport {
 	return nil
 }

+// headerContainsValue checks whether a header map contains a target value (case-insensitive key and value).
+func headerContainsValue(headers map[string]string, targetKey, targetValue string) bool {
+	if len(headers) == 0 {
+		return false
+	}
+	for key, value := range headers {
+		if !strings.EqualFold(strings.TrimSpace(key), strings.TrimSpace(targetKey)) {
+			continue
+		}
+		if strings.Contains(strings.ToLower(value), strings.ToLower(strings.TrimSpace(targetValue))) {
+			return true
+		}
+	}
+	return false
+}
+
+// encodeJSONStringToCBOR converts a JSON string payload into CBOR bytes.
+func encodeJSONStringToCBOR(jsonString string) ([]byte, error) {
+	var payload any
+	if errUnmarshal := json.Unmarshal([]byte(jsonString), &payload); errUnmarshal != nil {
+		return nil, errUnmarshal
+	}
+	return cbor.Marshal(payload)
+}
+
+// decodeCBORBodyToTextOrJSON decodes CBOR bytes to plain text (for string payloads) or JSON string.
+func decodeCBORBodyToTextOrJSON(raw []byte) (string, error) {
+	if len(raw) == 0 {
+		return "", nil
+	}
+
+	var payload any
+	if errUnmarshal := cbor.Unmarshal(raw, &payload); errUnmarshal != nil {
+		return "", errUnmarshal
+	}
+
+	jsonCompatible := cborValueToJSONCompatible(payload)
+	switch typed := jsonCompatible.(type) {
+	case string:
+		return typed, nil
+	case []byte:
+		return string(typed), nil
+	default:
+		jsonBytes, errMarshal := json.Marshal(jsonCompatible)
+		if errMarshal != nil {
+			return "", errMarshal
+		}
+		return string(jsonBytes), nil
+	}
+}
+
+// cborValueToJSONCompatible recursively converts CBOR-decoded values into JSON-marshalable values.
+func cborValueToJSONCompatible(value any) any {
+	switch typed := value.(type) {
+	case map[any]any:
+		out := make(map[string]any, len(typed))
+		for key, item := range typed {
+			out[fmt.Sprint(key)] = cborValueToJSONCompatible(item)
+		}
+		return out
+	case map[string]any:
+		out := make(map[string]any, len(typed))
+		for key, item := range typed {
+			out[key] = cborValueToJSONCompatible(item)
+		}
+		return out
+	case []any:
+		out := make([]any, len(typed))
+		for i, item := range typed {
+			out[i] = cborValueToJSONCompatible(item)
+		}
+		return out
+	default:
+		return typed
+	}
+}
+
 // QuotaDetail represents quota information for a specific resource type
 type QuotaDetail struct {
 	Entitlement      float64 `json:"entitlement"`
--- a/internal/api/handlers/management/auth_files.go
+++ b/internal/api/handlers/management/auth_files.go
@@ -29,6 +29,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/copilot"
 	geminiAuth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/gemini"
 	iflowauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/iflow"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/kilo"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/kimi"
 	kiroauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/kiro"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/qwen"
@@ -813,6 +814,87 @@ func (h *Handler) PatchAuthFileStatus(c *gin.Context) {
 	c.JSON(http.StatusOK, gin.H{"status": "ok", "disabled": *req.Disabled})
 }

+// PatchAuthFileFields updates editable fields (prefix, proxy_url, priority) of an auth file.
+func (h *Handler) PatchAuthFileFields(c *gin.Context) {
+	if h.authManager == nil {
+		c.JSON(http.StatusServiceUnavailable, gin.H{"error": "core auth manager unavailable"})
+		return
+	}
+
+	var req struct {
+		Name     string  `json:"name"`
+		Prefix   *string `json:"prefix"`
+		ProxyURL *string `json:"proxy_url"`
+		Priority *int    `json:"priority"`
+	}
+	if err := c.ShouldBindJSON(&req); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid request body"})
+		return
+	}
+
+	name := strings.TrimSpace(req.Name)
+	if name == "" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "name is required"})
+		return
+	}
+
+	ctx := c.Request.Context()
+
+	// Find auth by name or ID
+	var targetAuth *coreauth.Auth
+	if auth, ok := h.authManager.GetByID(name); ok {
+		targetAuth = auth
+	} else {
+		auths := h.authManager.List()
+		for _, auth := range auths {
+			if auth.FileName == name {
+				targetAuth = auth
+				break
+			}
+		}
+	}
+
+	if targetAuth == nil {
+		c.JSON(http.StatusNotFound, gin.H{"error": "auth file not found"})
+		return
+	}
+
+	changed := false
+	if req.Prefix != nil {
+		targetAuth.Prefix = *req.Prefix
+		changed = true
+	}
+	if req.ProxyURL != nil {
+		targetAuth.ProxyURL = *req.ProxyURL
+		changed = true
+	}
+	if req.Priority != nil {
+		if targetAuth.Metadata == nil {
+			targetAuth.Metadata = make(map[string]any)
+		}
+		if *req.Priority == 0 {
+			delete(targetAuth.Metadata, "priority")
+		} else {
+			targetAuth.Metadata["priority"] = *req.Priority
+		}
+		changed = true
+	}
+
+	if !changed {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "no fields to update"})
+		return
+	}
+
+	targetAuth.UpdatedAt = time.Now()
+
+	if _, err := h.authManager.Update(ctx, targetAuth); err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": fmt.Sprintf("failed to update auth: %v", err)})
+		return
+	}
+
+	c.JSON(http.StatusOK, gin.H{"status": "ok"})
+}
+
 func (h *Handler) disableAuth(ctx context.Context, id string) {
 	if h == nil || h.authManager == nil {
 		return
@@ -1193,6 +1275,30 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			}
 			ts.ProjectID = strings.Join(projects, ",")
 			ts.Checked = true
+		} else if strings.EqualFold(requestedProjectID, "GOOGLE_ONE") {
+			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")
+				return
+			}
+			if strings.TrimSpace(ts.ProjectID) == "" {
+				log.Error("Google One auto-discovery returned empty project ID")
+				SetOAuthSessionError(state, "Google One auto-discovery returned empty project ID")
+				return
+			}
+			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")
+				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")
+				return
+			}
 		} else {
 			if errEnsure := ensureGeminiProjectAndOnboard(ctx, gemClient, &ts, requestedProjectID); errEnsure != nil {
 				log.Errorf("Failed to complete Gemini CLI onboarding: %v", errEnsure)
@@ -2124,7 +2230,48 @@ func performGeminiCLISetup(ctx context.Context, httpClient *http.Client, storage
 		}
 	}
 	if projectID == "" {
-		return &projectSelectionRequiredError{}
+		// Auto-discovery: try onboardUser without specifying a project
+		// to let Google auto-provision one (matches Gemini CLI headless behavior
+		// and Antigravity's FetchProjectID pattern).
+		autoOnboardReq := map[string]any{
+			"tierId":   tierID,
+			"metadata": metadata,
+		}
+
+		autoCtx, autoCancel := context.WithTimeout(ctx, 30*time.Second)
+		defer autoCancel()
+		for attempt := 1; ; attempt++ {
+			var onboardResp map[string]any
+			if errOnboard := callGeminiCLI(autoCtx, httpClient, "onboardUser", autoOnboardReq, &onboardResp); errOnboard != nil {
+				return fmt.Errorf("auto-discovery onboardUser: %w", errOnboard)
+			}
+
+			if done, okDone := onboardResp["done"].(bool); okDone && done {
+				if resp, okResp := onboardResp["response"].(map[string]any); okResp {
+					switch v := resp["cloudaicompanionProject"].(type) {
+					case string:
+						projectID = strings.TrimSpace(v)
+					case map[string]any:
+						if id, okID := v["id"].(string); okID {
+							projectID = strings.TrimSpace(id)
+						}
+					}
+				}
+				break
+			}
+
+			log.Debugf("Auto-discovery: onboarding in progress, attempt %d...", attempt)
+			select {
+			case <-autoCtx.Done():
+				return &projectSelectionRequiredError{}
+			case <-time.After(2 * time.Second):
+			}
+		}
+
+		if projectID == "" {
+			return &projectSelectionRequiredError{}
+		}
+		log.Infof("Auto-discovered project ID via onboarding: %s", projectID)
 	}

 	onboardReqBody := map[string]any{
@@ -2668,3 +2815,88 @@ func generateKiroPKCE() (verifier, challenge string, err error) {

 	return verifier, challenge, nil
 }
+
+func (h *Handler) RequestKiloToken(c *gin.Context) {
+	ctx := context.Background()
+
+	fmt.Println("Initializing Kilo authentication...")
+
+	state := fmt.Sprintf("kil-%d", time.Now().UnixNano())
+	kilocodeAuth := kilo.NewKiloAuth()
+
+	resp, err := kilocodeAuth.InitiateDeviceFlow(ctx)
+	if err != nil {
+		log.Errorf("Failed to initiate device flow: %v", err)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to initiate device flow"})
+		return
+	}
+
+	RegisterOAuthSession(state, "kilo")
+
+	go func() {
+		fmt.Printf("Please visit %s and enter code: %s\n", resp.VerificationURL, resp.Code)
+
+		status, err := kilocodeAuth.PollForToken(ctx, resp.Code)
+		if err != nil {
+			SetOAuthSessionError(state, "Authentication failed")
+			fmt.Printf("Authentication failed: %v\n", err)
+			return
+		}
+
+		profile, err := kilocodeAuth.GetProfile(ctx, status.Token)
+		if err != nil {
+			log.Warnf("Failed to fetch profile: %v", err)
+			profile = &kilo.Profile{Email: status.UserEmail}
+		}
+
+		var orgID string
+		if len(profile.Orgs) > 0 {
+			orgID = profile.Orgs[0].ID
+		}
+
+		defaults, err := kilocodeAuth.GetDefaults(ctx, status.Token, orgID)
+		if err != nil {
+			defaults = &kilo.Defaults{}
+		}
+
+		ts := &kilo.KiloTokenStorage{
+			Token:          status.Token,
+			OrganizationID: orgID,
+			Model:          defaults.Model,
+			Email:          status.UserEmail,
+			Type:           "kilo",
+		}
+
+		fileName := kilo.CredentialFileName(status.UserEmail)
+		record := &coreauth.Auth{
+			ID:       fileName,
+			Provider: "kilo",
+			FileName: fileName,
+			Storage:  ts,
+			Metadata: map[string]any{
+				"email":           status.UserEmail,
+				"organization_id": orgID,
+				"model":          defaults.Model,
+			},
+		}
+
+		savedPath, errSave := h.saveTokenRecord(ctx, record)
+		if errSave != nil {
+			log.Errorf("Failed to save authentication tokens: %v", errSave)
+			SetOAuthSessionError(state, "Failed to save authentication tokens")
+			return
+		}
+
+		fmt.Printf("Authentication successful! Token saved to %s\n", savedPath)
+		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("kilo")
+	}()
+
+	c.JSON(200, gin.H{
+		"status":           "ok",
+		"url":              resp.VerificationURL,
+		"state":            state,
+		"user_code":        resp.Code,
+		"verification_uri": resp.VerificationURL,
+	})
+}
--- a/internal/api/handlers/management/config_basic.go
+++ b/internal/api/handlers/management/config_basic.go
@@ -28,8 +28,7 @@ func (h *Handler) GetConfig(c *gin.Context) {
 		c.JSON(200, gin.H{})
 		return
 	}
-	cfgCopy := *h.cfg
-	c.JSON(200, &cfgCopy)
+	c.JSON(200, new(*h.cfg))
 }

 type releaseInfo struct {
--- a/internal/api/handlers/management/config_lists.go
+++ b/internal/api/handlers/management/config_lists.go
@@ -796,10 +796,10 @@ func (h *Handler) DeleteOAuthModelAlias(c *gin.Context) {
 		c.JSON(404, gin.H{"error": "channel not found"})
 		return
 	}
-	delete(h.cfg.OAuthModelAlias, channel)
-	if len(h.cfg.OAuthModelAlias) == 0 {
-		h.cfg.OAuthModelAlias = nil
-	}
+	// Set to nil instead of deleting the key so that the "explicitly disabled"
+	// marker survives config reload and prevents SanitizeOAuthModelAlias from
+	// re-injecting default aliases (fixes #222).
+	h.cfg.OAuthModelAlias[channel] = nil
 	h.persist(c)
 }

--- a/internal/api/middleware/request_logging.go
+++ b/internal/api/middleware/request_logging.go
@@ -15,10 +15,12 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 )

+const maxErrorOnlyCapturedRequestBodyBytes int64 = 1 << 20 // 1 MiB
+
 // RequestLoggingMiddleware creates a Gin middleware that logs HTTP requests and responses.
 // It captures detailed information about the request and response, including headers and body,
-// and uses the provided RequestLogger to record this data. When logging is disabled in the
-// logger, it still captures data so that upstream errors can be persisted.
+// and uses the provided RequestLogger to record this data. When full request logging is disabled,
+// body capture is limited to small known-size payloads to avoid large per-request memory spikes.
 func RequestLoggingMiddleware(logger logging.RequestLogger) gin.HandlerFunc {
 	return func(c *gin.Context) {
 		if logger == nil {
@@ -26,7 +28,7 @@ func RequestLoggingMiddleware(logger logging.RequestLogger) gin.HandlerFunc {
 			return
 		}

-		if c.Request.Method == http.MethodGet {
+		if shouldSkipMethodForRequestLogging(c.Request) {
 			c.Next()
 			return
 		}
@@ -37,8 +39,10 @@ func RequestLoggingMiddleware(logger logging.RequestLogger) gin.HandlerFunc {
 			return
 		}

+		loggerEnabled := logger.IsEnabled()
+
 		// Capture request information
-		requestInfo, err := captureRequestInfo(c)
+		requestInfo, err := captureRequestInfo(c, shouldCaptureRequestBody(loggerEnabled, c.Request))
 		if err != nil {
 			// Log error but continue processing
 			// In a real implementation, you might want to use a proper logger here
@@ -48,7 +52,7 @@ func RequestLoggingMiddleware(logger logging.RequestLogger) gin.HandlerFunc {

 		// Create response writer wrapper
 		wrapper := NewResponseWriterWrapper(c.Writer, logger, requestInfo)
-		if !logger.IsEnabled() {
+		if !loggerEnabled {
 			wrapper.logOnErrorOnly = true
 		}
 		c.Writer = wrapper
@@ -64,10 +68,47 @@ func RequestLoggingMiddleware(logger logging.RequestLogger) gin.HandlerFunc {
 	}
 }

+func shouldSkipMethodForRequestLogging(req *http.Request) bool {
+	if req == nil {
+		return true
+	}
+	if req.Method != http.MethodGet {
+		return false
+	}
+	return !isResponsesWebsocketUpgrade(req)
+}
+
+func isResponsesWebsocketUpgrade(req *http.Request) bool {
+	if req == nil || req.URL == nil {
+		return false
+	}
+	if req.URL.Path != "/v1/responses" {
+		return false
+	}
+	return strings.EqualFold(strings.TrimSpace(req.Header.Get("Upgrade")), "websocket")
+}
+
+func shouldCaptureRequestBody(loggerEnabled bool, req *http.Request) bool {
+	if loggerEnabled {
+		return true
+	}
+	if req == nil || req.Body == nil {
+		return false
+	}
+	contentType := strings.ToLower(strings.TrimSpace(req.Header.Get("Content-Type")))
+	if strings.HasPrefix(contentType, "multipart/form-data") {
+		return false
+	}
+	if req.ContentLength <= 0 {
+		return false
+	}
+	return req.ContentLength <= maxErrorOnlyCapturedRequestBodyBytes
+}
+
 // captureRequestInfo extracts relevant information from the incoming HTTP request.
 // It captures the URL, method, headers, and body. The request body is read and then
 // restored so that it can be processed by subsequent handlers.
-func captureRequestInfo(c *gin.Context) (*RequestInfo, error) {
+func captureRequestInfo(c *gin.Context, captureBody bool) (*RequestInfo, error) {
 	// Capture URL with sensitive query parameters masked
 	maskedQuery := util.MaskSensitiveQuery(c.Request.URL.RawQuery)
 	url := c.Request.URL.Path
@@ -86,7 +127,7 @@ func captureRequestInfo(c *gin.Context) (*RequestInfo, error) {

 	// Capture request body
 	var body []byte
-	if c.Request.Body != nil {
+	if captureBody && c.Request.Body != nil {
 		// Read the body
 		bodyBytes, err := io.ReadAll(c.Request.Body)
 		if err != nil {
--- a/internal/api/middleware/request_logging_test.go
+++ b/internal/api/middleware/request_logging_test.go
@@ -0,0 +1,138 @@
+package middleware
+
+import (
+	"io"
+	"net/http"
+	"net/url"
+	"strings"
+	"testing"
+)
+
+func TestShouldSkipMethodForRequestLogging(t *testing.T) {
+	tests := []struct {
+		name string
+		req  *http.Request
+		skip bool
+	}{
+		{
+			name: "nil request",
+			req:  nil,
+			skip: true,
+		},
+		{
+			name: "post request should not skip",
+			req: &http.Request{
+				Method: http.MethodPost,
+				URL:    &url.URL{Path: "/v1/responses"},
+			},
+			skip: false,
+		},
+		{
+			name: "plain get should skip",
+			req: &http.Request{
+				Method: http.MethodGet,
+				URL:    &url.URL{Path: "/v1/models"},
+				Header: http.Header{},
+			},
+			skip: true,
+		},
+		{
+			name: "responses websocket upgrade should not skip",
+			req: &http.Request{
+				Method: http.MethodGet,
+				URL:    &url.URL{Path: "/v1/responses"},
+				Header: http.Header{"Upgrade": []string{"websocket"}},
+			},
+			skip: false,
+		},
+		{
+			name: "responses get without upgrade should skip",
+			req: &http.Request{
+				Method: http.MethodGet,
+				URL:    &url.URL{Path: "/v1/responses"},
+				Header: http.Header{},
+			},
+			skip: true,
+		},
+	}
+
+	for i := range tests {
+		got := shouldSkipMethodForRequestLogging(tests[i].req)
+		if got != tests[i].skip {
+			t.Fatalf("%s: got skip=%t, want %t", tests[i].name, got, tests[i].skip)
+		}
+	}
+}
+
+func TestShouldCaptureRequestBody(t *testing.T) {
+	tests := []struct {
+		name          string
+		loggerEnabled bool
+		req           *http.Request
+		want          bool
+	}{
+		{
+			name:          "logger enabled always captures",
+			loggerEnabled: true,
+			req: &http.Request{
+				Body:          io.NopCloser(strings.NewReader("{}")),
+				ContentLength: -1,
+				Header:        http.Header{"Content-Type": []string{"application/json"}},
+			},
+			want: true,
+		},
+		{
+			name:          "nil request",
+			loggerEnabled: false,
+			req:           nil,
+			want:          false,
+		},
+		{
+			name:          "small known size json in error-only mode",
+			loggerEnabled: false,
+			req: &http.Request{
+				Body:          io.NopCloser(strings.NewReader("{}")),
+				ContentLength: 2,
+				Header:        http.Header{"Content-Type": []string{"application/json"}},
+			},
+			want: true,
+		},
+		{
+			name:          "large known size skipped in error-only mode",
+			loggerEnabled: false,
+			req: &http.Request{
+				Body:          io.NopCloser(strings.NewReader("x")),
+				ContentLength: maxErrorOnlyCapturedRequestBodyBytes + 1,
+				Header:        http.Header{"Content-Type": []string{"application/json"}},
+			},
+			want: false,
+		},
+		{
+			name:          "unknown size skipped in error-only mode",
+			loggerEnabled: false,
+			req: &http.Request{
+				Body:          io.NopCloser(strings.NewReader("x")),
+				ContentLength: -1,
+				Header:        http.Header{"Content-Type": []string{"application/json"}},
+			},
+			want: false,
+		},
+		{
+			name:          "multipart skipped in error-only mode",
+			loggerEnabled: false,
+			req: &http.Request{
+				Body:          io.NopCloser(strings.NewReader("x")),
+				ContentLength: 1,
+				Header:        http.Header{"Content-Type": []string{"multipart/form-data; boundary=abc"}},
+			},
+			want: false,
+		},
+	}
+
+	for i := range tests {
+		got := shouldCaptureRequestBody(tests[i].loggerEnabled, tests[i].req)
+		if got != tests[i].want {
+			t.Fatalf("%s: got %t, want %t", tests[i].name, got, tests[i].want)
+		}
+	}
+}
--- a/internal/api/middleware/response_writer.go
+++ b/internal/api/middleware/response_writer.go
@@ -14,6 +14,8 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/logging"
 )

+const requestBodyOverrideContextKey = "REQUEST_BODY_OVERRIDE"
+
 // RequestInfo holds essential details of an incoming HTTP request for logging purposes.
 type RequestInfo struct {
 	URL       string              // URL is the request URL.
@@ -223,8 +225,8 @@ func (w *ResponseWriterWrapper) detectStreaming(contentType string) bool {

 	// Only fall back to request payload hints when Content-Type is not set yet.
 	if w.requestInfo != nil && len(w.requestInfo.Body) > 0 {
-		bodyStr := string(w.requestInfo.Body)
-		return strings.Contains(bodyStr, `"stream": true`) || strings.Contains(bodyStr, `"stream":true`)
+		return bytes.Contains(w.requestInfo.Body, []byte(`"stream": true`)) ||
+			bytes.Contains(w.requestInfo.Body, []byte(`"stream":true`))
 	}

 	return false
@@ -310,7 +312,7 @@ func (w *ResponseWriterWrapper) Finalize(c *gin.Context) error {
 		return nil
 	}

-	return w.logRequest(finalStatusCode, w.cloneHeaders(), w.body.Bytes(), w.extractAPIRequest(c), w.extractAPIResponse(c), w.extractAPIResponseTimestamp(c), slicesAPIResponseError, forceLog)
+	return w.logRequest(w.extractRequestBody(c), finalStatusCode, w.cloneHeaders(), w.body.Bytes(), w.extractAPIRequest(c), w.extractAPIResponse(c), w.extractAPIResponseTimestamp(c), slicesAPIResponseError, forceLog)
 }

 func (w *ResponseWriterWrapper) cloneHeaders() map[string][]string {
@@ -361,16 +363,32 @@ func (w *ResponseWriterWrapper) extractAPIResponseTimestamp(c *gin.Context) time
 	return time.Time{}
 }

-func (w *ResponseWriterWrapper) logRequest(statusCode int, headers map[string][]string, body []byte, apiRequestBody, apiResponseBody []byte, apiResponseTimestamp time.Time, apiResponseErrors []*interfaces.ErrorMessage, forceLog bool) error {
+func (w *ResponseWriterWrapper) extractRequestBody(c *gin.Context) []byte {
+	if c != nil {
+		if bodyOverride, isExist := c.Get(requestBodyOverrideContextKey); isExist {
+			switch value := bodyOverride.(type) {
+			case []byte:
+				if len(value) > 0 {
+					return bytes.Clone(value)
+				}
+			case string:
+				if strings.TrimSpace(value) != "" {
+					return []byte(value)
+				}
+			}
+		}
+	}
+	if w.requestInfo != nil && len(w.requestInfo.Body) > 0 {
+		return w.requestInfo.Body
+	}
+	return nil
+}
+
+func (w *ResponseWriterWrapper) logRequest(requestBody []byte, statusCode int, headers map[string][]string, body []byte, apiRequestBody, apiResponseBody []byte, apiResponseTimestamp time.Time, apiResponseErrors []*interfaces.ErrorMessage, forceLog bool) error {
 	if w.requestInfo == nil {
 		return nil
 	}

-	var requestBody []byte
-	if len(w.requestInfo.Body) > 0 {
-		requestBody = w.requestInfo.Body
-	}
-
 	if loggerWithOptions, ok := w.logger.(interface {
 		LogRequestWithOptions(string, string, map[string][]string, []byte, int, map[string][]string, []byte, []byte, []byte, []*interfaces.ErrorMessage, bool, string, time.Time, time.Time) error
 	}); ok {
--- a/internal/api/middleware/response_writer_test.go
+++ b/internal/api/middleware/response_writer_test.go
@@ -0,0 +1,43 @@
+package middleware
+
+import (
+	"net/http/httptest"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+)
+
+func TestExtractRequestBodyPrefersOverride(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+
+	wrapper := &ResponseWriterWrapper{
+		requestInfo: &RequestInfo{Body: []byte("original-body")},
+	}
+
+	body := wrapper.extractRequestBody(c)
+	if string(body) != "original-body" {
+		t.Fatalf("request body = %q, want %q", string(body), "original-body")
+	}
+
+	c.Set(requestBodyOverrideContextKey, []byte("override-body"))
+	body = wrapper.extractRequestBody(c)
+	if string(body) != "override-body" {
+		t.Fatalf("request body = %q, want %q", string(body), "override-body")
+	}
+}
+
+func TestExtractRequestBodySupportsStringOverride(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+
+	wrapper := &ResponseWriterWrapper{}
+	c.Set(requestBodyOverrideContextKey, "override-as-string")
+
+	body := wrapper.extractRequestBody(c)
+	if string(body) != "override-as-string" {
+		t.Fatalf("request body = %q, want %q", string(body), "override-as-string")
+	}
+}
--- a/internal/api/modules/amp/amp.go
+++ b/internal/api/modules/amp/amp.go
@@ -127,8 +127,7 @@ func (m *AmpModule) Register(ctx modules.Context) error {
 		m.modelMapper = NewModelMapper(settings.ModelMappings)

 		// Store initial config for partial reload comparison
-		settingsCopy := settings
-		m.lastConfig = &settingsCopy
+		m.lastConfig = new(settings)

 		// Initialize localhost restriction setting (hot-reloadable)
 		m.setRestrictToLocalhost(settings.RestrictManagementToLocalhost)
--- a/internal/api/server.go
+++ b/internal/api/server.go
@@ -285,8 +285,9 @@ func NewServer(cfg *config.Config, authManager *auth.Manager, accessManager *sdk
 		optionState.routerConfigurator(engine, s.handlers, cfg)
 	}

-	// Register management routes when configuration or environment secrets are available.
-	hasManagementSecret := cfg.RemoteManagement.SecretKey != "" || envManagementSecret
+	// Register management routes when configuration or environment secrets are available,
+	// or when a local management password is provided (e.g. TUI mode).
+	hasManagementSecret := cfg.RemoteManagement.SecretKey != "" || envManagementSecret || s.localPassword != ""
 	s.managementRoutesEnabled.Store(hasManagementSecret)
 	if hasManagementSecret {
 		s.registerManagementRoutes()
@@ -329,6 +330,7 @@ func (s *Server) setupRoutes() {
 		v1.POST("/completions", openaiHandlers.Completions)
 		v1.POST("/messages", claudeCodeHandlers.ClaudeMessages)
 		v1.POST("/messages/count_tokens", claudeCodeHandlers.ClaudeCountTokens)
+		v1.GET("/responses", openaiResponsesHandlers.ResponsesWebsocket)
 		v1.POST("/responses", openaiResponsesHandlers.Responses)
 		v1.POST("/responses/compact", openaiResponsesHandlers.Compact)
 	}
@@ -642,6 +644,7 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.POST("/auth-files", s.mgmt.UploadAuthFile)
 		mgmt.DELETE("/auth-files", s.mgmt.DeleteAuthFile)
 		mgmt.PATCH("/auth-files/status", s.mgmt.PatchAuthFileStatus)
+		mgmt.PATCH("/auth-files/fields", s.mgmt.PatchAuthFileFields)
 		mgmt.POST("/vertex/import", s.mgmt.ImportVertexCredential)

 		mgmt.GET("/anthropic-auth-url", s.mgmt.RequestAnthropicToken)
@@ -649,6 +652,7 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.GET("/gemini-cli-auth-url", s.mgmt.RequestGeminiCLIToken)
 		mgmt.GET("/antigravity-auth-url", s.mgmt.RequestAntigravityToken)
 		mgmt.GET("/qwen-auth-url", s.mgmt.RequestQwenToken)
+		mgmt.GET("/kilo-auth-url", s.mgmt.RequestKiloToken)
 		mgmt.GET("/kimi-auth-url", s.mgmt.RequestKimiToken)
 		mgmt.GET("/iflow-auth-url", s.mgmt.RequestIFlowToken)
 		mgmt.POST("/iflow-auth-url", s.mgmt.RequestIFlowCookieToken)
--- a/internal/auth/kilo/kilo_auth.go
+++ b/internal/auth/kilo/kilo_auth.go
@@ -0,0 +1,168 @@
+// Package kilo provides authentication and token management functionality
+// for Kilo AI services.
+package kilo
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"time"
+)
+
+const (
+	// BaseURL is the base URL for the Kilo AI API.
+	BaseURL = "https://api.kilo.ai/api"
+)
+
+// DeviceAuthResponse represents the response from initiating device flow.
+type DeviceAuthResponse struct {
+	Code            string `json:"code"`
+	VerificationURL string `json:"verificationUrl"`
+	ExpiresIn       int    `json:"expiresIn"`
+}
+
+// DeviceStatusResponse represents the response when polling for device flow status.
+type DeviceStatusResponse struct {
+	Status    string `json:"status"`
+	Token     string `json:"token"`
+	UserEmail string `json:"userEmail"`
+}
+
+// Profile represents the user profile from Kilo AI.
+type Profile struct {
+	Email string         `json:"email"`
+	Orgs  []Organization `json:"organizations"`
+}
+
+// Organization represents a Kilo AI organization.
+type Organization struct {
+	ID   string `json:"id"`
+	Name string `json:"name"`
+}
+
+// Defaults represents default settings for an organization or user.
+type Defaults struct {
+	Model string `json:"model"`
+}
+
+// KiloAuth provides methods for handling the Kilo AI authentication flow.
+type KiloAuth struct {
+	client *http.Client
+}
+
+// NewKiloAuth creates a new instance of KiloAuth.
+func NewKiloAuth() *KiloAuth {
+	return &KiloAuth{
+		client: &http.Client{Timeout: 30 * time.Second},
+	}
+}
+
+// InitiateDeviceFlow starts the device authentication flow.
+func (k *KiloAuth) InitiateDeviceFlow(ctx context.Context) (*DeviceAuthResponse, error) {
+	resp, err := k.client.Post(BaseURL+"/device-auth/codes", "application/json", nil)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusCreated && resp.StatusCode != http.StatusOK {
+		return nil, fmt.Errorf("failed to initiate device flow: status %d", resp.StatusCode)
+	}
+
+	var data DeviceAuthResponse
+	if err := json.NewDecoder(resp.Body).Decode(&data); err != nil {
+		return nil, err
+	}
+	return &data, nil
+}
+
+// PollForToken polls for the device flow completion.
+func (k *KiloAuth) PollForToken(ctx context.Context, code string) (*DeviceStatusResponse, error) {
+	ticker := time.NewTicker(5 * time.Second)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		case <-ticker.C:
+			resp, err := k.client.Get(BaseURL + "/device-auth/codes/" + code)
+			if err != nil {
+				return nil, err
+			}
+			defer resp.Body.Close()
+
+			var data DeviceStatusResponse
+			if err := json.NewDecoder(resp.Body).Decode(&data); err != nil {
+				return nil, err
+			}
+
+			switch data.Status {
+			case "approved":
+				return &data, nil
+			case "denied", "expired":
+				return nil, fmt.Errorf("device flow %s", data.Status)
+			case "pending":
+				continue
+			default:
+				return nil, fmt.Errorf("unknown status: %s", data.Status)
+			}
+		}
+	}
+}
+
+// GetProfile fetches the user's profile.
+func (k *KiloAuth) GetProfile(ctx context.Context, token string) (*Profile, error) {
+	req, err := http.NewRequestWithContext(ctx, "GET", BaseURL+"/profile", nil)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create get profile request: %w", err)
+	}
+	req.Header.Set("Authorization", "Bearer "+token)
+
+	resp, err := k.client.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusOK {
+		return nil, fmt.Errorf("failed to get profile: status %d", resp.StatusCode)
+	}
+
+	var profile Profile
+	if err := json.NewDecoder(resp.Body).Decode(&profile); err != nil {
+		return nil, err
+	}
+	return &profile, nil
+}
+
+// GetDefaults fetches default settings for an organization.
+func (k *KiloAuth) GetDefaults(ctx context.Context, token, orgID string) (*Defaults, error) {
+	url := BaseURL + "/defaults"
+	if orgID != "" {
+		url = BaseURL + "/organizations/" + orgID + "/defaults"
+	}
+
+	req, err := http.NewRequestWithContext(ctx, "GET", url, nil)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create get defaults request: %w", err)
+	}
+	req.Header.Set("Authorization", "Bearer "+token)
+
+	resp, err := k.client.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusOK {
+		return nil, fmt.Errorf("failed to get defaults: status %d", resp.StatusCode)
+	}
+
+	var defaults Defaults
+	if err := json.NewDecoder(resp.Body).Decode(&defaults); err != nil {
+		return nil, err
+	}
+	return &defaults, nil
+}
--- a/internal/auth/kilo/kilo_token.go
+++ b/internal/auth/kilo/kilo_token.go
@@ -0,0 +1,60 @@
+// Package kilo provides authentication and token management functionality
+// for Kilo AI services.
+package kilo
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
+	log "github.com/sirupsen/logrus"
+)
+
+// KiloTokenStorage stores token information for Kilo AI authentication.
+type KiloTokenStorage struct {
+	// Token is the Kilo access token.
+	Token string `json:"kilocodeToken"`
+
+	// OrganizationID is the Kilo organization ID.
+	OrganizationID string `json:"kilocodeOrganizationId"`
+
+	// Model is the default model to use.
+	Model string `json:"kilocodeModel"`
+
+	// Email is the email address of the authenticated user.
+	Email string `json:"email"`
+
+	// Type indicates the authentication provider type, always "kilo" for this storage.
+	Type string `json:"type"`
+}
+
+// SaveTokenToFile serializes the Kilo token storage to a JSON file.
+func (ts *KiloTokenStorage) SaveTokenToFile(authFilePath string) error {
+	misc.LogSavingCredentials(authFilePath)
+	ts.Type = "kilo"
+	if err := os.MkdirAll(filepath.Dir(authFilePath), 0700); err != nil {
+		return fmt.Errorf("failed to create directory: %v", err)
+	}
+
+	f, err := os.Create(authFilePath)
+	if err != nil {
+		return fmt.Errorf("failed to create token file: %w", err)
+	}
+	defer func() {
+		if errClose := f.Close(); errClose != nil {
+			log.Errorf("failed to close file: %v", errClose)
+		}
+	}()
+
+	if err = json.NewEncoder(f).Encode(ts); err != nil {
+		return fmt.Errorf("failed to write token to file: %w", err)
+	}
+	return nil
+}
+
+// CredentialFileName returns the filename used to persist Kilo credentials.
+func CredentialFileName(email string) string {
+	return fmt.Sprintf("kilo-%s.json", email)
+}
--- a/internal/cmd/anthropic_login.go
+++ b/internal/cmd/anthropic_login.go
@@ -40,8 +40,7 @@ func DoClaudeLogin(cfg *config.Config, options *LoginOptions) {

 	_, savedPath, err := manager.Login(context.Background(), "claude", cfg, authOpts)
 	if err != nil {
-		var authErr *claude.AuthenticationError
-		if errors.As(err, &authErr) {
+		if authErr, ok := errors.AsType[*claude.AuthenticationError](err); ok {
 			log.Error(claude.GetUserFriendlyMessage(authErr))
 			if authErr.Type == claude.ErrPortInUse.Type {
 				os.Exit(claude.ErrPortInUse.Code)
--- a/internal/cmd/auth_manager.go
+++ b/internal/cmd/auth_manager.go
@@ -22,6 +22,7 @@ func newAuthManager() *sdkAuth.Manager {
 		sdkAuth.NewKimiAuthenticator(),
 		sdkAuth.NewKiroAuthenticator(),
 		sdkAuth.NewGitHubCopilotAuthenticator(),
+		sdkAuth.NewKiloAuthenticator(),
 	)
 	return manager
 }
--- a/internal/cmd/iflow_login.go
+++ b/internal/cmd/iflow_login.go
@@ -32,8 +32,7 @@ func DoIFlowLogin(cfg *config.Config, options *LoginOptions) {

 	_, savedPath, err := manager.Login(context.Background(), "iflow", cfg, authOpts)
 	if err != nil {
-		var emailErr *sdkAuth.EmailRequiredError
-		if errors.As(err, &emailErr) {
+		if emailErr, ok := errors.AsType[*sdkAuth.EmailRequiredError](err); ok {
 			log.Error(emailErr.Error())
 			return
 		}
--- a/internal/cmd/kilo_login.go
+++ b/internal/cmd/kilo_login.go
@@ -0,0 +1,54 @@
+package cmd
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	sdkAuth "github.com/router-for-me/CLIProxyAPI/v6/sdk/auth"
+)
+
+// DoKiloLogin handles the Kilo device flow using the shared authentication manager.
+// It initiates the device-based authentication process for Kilo AI services and saves
+// the authentication tokens to the configured auth directory.
+//
+// Parameters:
+//   - cfg: The application configuration
+//   - options: Login options including browser behavior and prompts
+func DoKiloLogin(cfg *config.Config, options *LoginOptions) {
+	if options == nil {
+		options = &LoginOptions{}
+	}
+
+	manager := newAuthManager()
+
+	promptFn := options.Prompt
+	if promptFn == nil {
+		promptFn = func(prompt string) (string, error) {
+			fmt.Print(prompt)
+			var value string
+			fmt.Scanln(&value)
+			return strings.TrimSpace(value), nil
+		}
+	}
+
+	authOpts := &sdkAuth.LoginOptions{
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Metadata:     map[string]string{},
+		Prompt:       promptFn,
+	}
+
+	_, savedPath, err := manager.Login(context.Background(), "kilo", cfg, authOpts)
+	if err != nil {
+		fmt.Printf("Kilo authentication failed: %v\n", err)
+		return
+	}
+
+	if savedPath != "" {
+		fmt.Printf("Authentication saved to %s\n", savedPath)
+	}
+
+	fmt.Println("Kilo authentication successful!")
+}
--- a/internal/cmd/login.go
+++ b/internal/cmd/login.go
@@ -100,49 +100,74 @@ func DoLogin(cfg *config.Config, projectID string, options *LoginOptions) {

 	log.Info("Authentication successful.")

-	projects, errProjects := fetchGCPProjects(ctx, httpClient)
-	if errProjects != nil {
-		log.Errorf("Failed to get project list: %v", errProjects)
-		return
+	var activatedProjects []string
+
+	useGoogleOne := false
+	if trimmedProjectID == "" && promptFn != nil {
+		fmt.Println("\nSelect login mode:")
+		fmt.Println("  1. Code Assist  (GCP project, manual selection)")
+		fmt.Println("  2. Google One   (personal account, auto-discover project)")
+		choice, errPrompt := promptFn("Enter choice [1/2] (default: 1): ")
+		if errPrompt == nil && strings.TrimSpace(choice) == "2" {
+			useGoogleOne = true
+		}
 	}

-	selectedProjectID := promptForProjectSelection(projects, trimmedProjectID, promptFn)
-	projectSelections, errSelection := resolveProjectSelections(selectedProjectID, projects)
-	if errSelection != nil {
-		log.Errorf("Invalid project selection: %v", errSelection)
-		return
-	}
-	if len(projectSelections) == 0 {
-		log.Error("No project selected; aborting login.")
-		return
-	}
-
-	activatedProjects := make([]string, 0, len(projectSelections))
-	seenProjects := make(map[string]bool)
-	for _, candidateID := range projectSelections {
-		log.Infof("Activating project %s", candidateID)
-		if errSetup := performGeminiCLISetup(ctx, httpClient, storage, candidateID); errSetup != nil {
-			var projectErr *projectSelectionRequiredError
-			if errors.As(errSetup, &projectErr) {
-				log.Error("Failed to start user onboarding: A project ID is required.")
-				showProjectSelectionHelp(storage.Email, projects)
-				return
-			}
-			log.Errorf("Failed to complete user setup: %v", errSetup)
+	if useGoogleOne {
+		log.Info("Google One mode: auto-discovering project...")
+		if errSetup := performGeminiCLISetup(ctx, httpClient, storage, ""); errSetup != nil {
+			log.Errorf("Google One auto-discovery failed: %v", errSetup)
 			return
 		}
-		finalID := strings.TrimSpace(storage.ProjectID)
-		if finalID == "" {
-			finalID = candidateID
+		autoProject := strings.TrimSpace(storage.ProjectID)
+		if autoProject == "" {
+			log.Error("Google One auto-discovery returned empty project ID")
+			return
+		}
+		log.Infof("Auto-discovered project: %s", autoProject)
+		activatedProjects = []string{autoProject}
+	} else {
+		projects, errProjects := fetchGCPProjects(ctx, httpClient)
+		if errProjects != nil {
+			log.Errorf("Failed to get project list: %v", errProjects)
+			return
 		}

-		// Skip duplicates
-		if seenProjects[finalID] {
-			log.Infof("Project %s already activated, skipping", finalID)
-			continue
+		selectedProjectID := promptForProjectSelection(projects, trimmedProjectID, promptFn)
+		projectSelections, errSelection := resolveProjectSelections(selectedProjectID, projects)
+		if errSelection != nil {
+			log.Errorf("Invalid project selection: %v", errSelection)
+			return
+		}
+		if len(projectSelections) == 0 {
+			log.Error("No project selected; aborting login.")
+			return
+		}
+
+		seenProjects := make(map[string]bool)
+		for _, candidateID := range projectSelections {
+			log.Infof("Activating project %s", candidateID)
+			if errSetup := performGeminiCLISetup(ctx, httpClient, storage, candidateID); errSetup != nil {
+				if _, ok := errors.AsType[*projectSelectionRequiredError](errSetup); ok {
+					log.Error("Failed to start user onboarding: A project ID is required.")
+					showProjectSelectionHelp(storage.Email, projects)
+					return
+				}
+				log.Errorf("Failed to complete user setup: %v", errSetup)
+				return
+			}
+			finalID := strings.TrimSpace(storage.ProjectID)
+			if finalID == "" {
+				finalID = candidateID
+			}
+
+			if seenProjects[finalID] {
+				log.Infof("Project %s already activated, skipping", finalID)
+				continue
+			}
+			seenProjects[finalID] = true
+			activatedProjects = append(activatedProjects, finalID)
 		}
-		seenProjects[finalID] = true
-		activatedProjects = append(activatedProjects, finalID)
 	}

 	storage.Auto = false
@@ -235,7 +260,48 @@ func performGeminiCLISetup(ctx context.Context, httpClient *http.Client, storage
 		}
 	}
 	if projectID == "" {
-		return &projectSelectionRequiredError{}
+		// Auto-discovery: try onboardUser without specifying a project
+		// to let Google auto-provision one (matches Gemini CLI headless behavior
+		// and Antigravity's FetchProjectID pattern).
+		autoOnboardReq := map[string]any{
+			"tierId":   tierID,
+			"metadata": metadata,
+		}
+
+		autoCtx, autoCancel := context.WithTimeout(ctx, 30*time.Second)
+		defer autoCancel()
+		for attempt := 1; ; attempt++ {
+			var onboardResp map[string]any
+			if errOnboard := callGeminiCLI(autoCtx, httpClient, "onboardUser", autoOnboardReq, &onboardResp); errOnboard != nil {
+				return fmt.Errorf("auto-discovery onboardUser: %w", errOnboard)
+			}
+
+			if done, okDone := onboardResp["done"].(bool); okDone && done {
+				if resp, okResp := onboardResp["response"].(map[string]any); okResp {
+					switch v := resp["cloudaicompanionProject"].(type) {
+					case string:
+						projectID = strings.TrimSpace(v)
+					case map[string]any:
+						if id, okID := v["id"].(string); okID {
+							projectID = strings.TrimSpace(id)
+						}
+					}
+				}
+				break
+			}
+
+			log.Debugf("Auto-discovery: onboarding in progress, attempt %d...", attempt)
+			select {
+			case <-autoCtx.Done():
+				return &projectSelectionRequiredError{}
+			case <-time.After(2 * time.Second):
+			}
+		}
+
+		if projectID == "" {
+			return &projectSelectionRequiredError{}
+		}
+		log.Infof("Auto-discovered project ID via onboarding: %s", projectID)
 	}

 	onboardReqBody := map[string]any{
@@ -617,7 +683,7 @@ func updateAuthRecord(record *cliproxyauth.Auth, storage *gemini.GeminiTokenStor
 		return
 	}

-	finalName := gemini.CredentialFileName(storage.Email, storage.ProjectID, false)
+	finalName := gemini.CredentialFileName(storage.Email, storage.ProjectID, true)

 	if record.Metadata == nil {
 		record.Metadata = make(map[string]any)
--- a/internal/cmd/openai_login.go
+++ b/internal/cmd/openai_login.go
@@ -54,8 +54,7 @@ func DoCodexLogin(cfg *config.Config, options *LoginOptions) {

 	_, savedPath, err := manager.Login(context.Background(), "codex", cfg, authOpts)
 	if err != nil {
-		var authErr *codex.AuthenticationError
-		if errors.As(err, &authErr) {
+		if authErr, ok := errors.AsType[*codex.AuthenticationError](err); ok {
 			log.Error(codex.GetUserFriendlyMessage(authErr))
 			if authErr.Type == codex.ErrPortInUse.Type {
 				os.Exit(codex.ErrPortInUse.Code)
--- a/internal/cmd/qwen_login.go
+++ b/internal/cmd/qwen_login.go
@@ -44,8 +44,7 @@ func DoQwenLogin(cfg *config.Config, options *LoginOptions) {

 	_, savedPath, err := manager.Login(context.Background(), "qwen", cfg, authOpts)
 	if err != nil {
-		var emailErr *sdkAuth.EmailRequiredError
-		if errors.As(err, &emailErr) {
+		if emailErr, ok := errors.AsType[*sdkAuth.EmailRequiredError](err); ok {
 			log.Error(emailErr.Error())
 			return
 		}
--- a/internal/cmd/run.go
+++ b/internal/cmd/run.go
@@ -55,6 +55,34 @@ func StartService(cfg *config.Config, configPath string, localPassword string) {
 	}
 }

+// StartServiceBackground starts the proxy service in a background goroutine
+// and returns a cancel function for shutdown and a done channel.
+func StartServiceBackground(cfg *config.Config, configPath string, localPassword string) (cancel func(), done <-chan struct{}) {
+	builder := cliproxy.NewBuilder().
+		WithConfig(cfg).
+		WithConfigPath(configPath).
+		WithLocalManagementPassword(localPassword)
+
+	ctx, cancelFn := context.WithCancel(context.Background())
+	doneCh := make(chan struct{})
+
+	service, err := builder.Build()
+	if err != nil {
+		log.Errorf("failed to build proxy service: %v", err)
+		close(doneCh)
+		return cancelFn, doneCh
+	}
+
+	go func() {
+		defer close(doneCh)
+		if err := service.Run(ctx); err != nil && !errors.Is(err, context.Canceled) {
+			log.Errorf("proxy service exited with error: %v", err)
+		}
+	}()
+
+	return cancelFn, doneCh
+}
+
 // WaitForCloudDeploy waits indefinitely for shutdown signals in cloud deploy mode
 // when no configuration file is available.
 func WaitForCloudDeploy() {
--- a/internal/config/config.go
+++ b/internal/config/config.go
@@ -97,6 +97,10 @@ type Config struct {
 	// ClaudeKey defines a list of Claude API key configurations as specified in the YAML configuration file.
 	ClaudeKey []ClaudeKey `yaml:"claude-api-key" json:"claude-api-key"`

+	// ClaudeHeaderDefaults configures default header values for Claude API requests.
+	// These are used as fallbacks when the client does not send its own headers.
+	ClaudeHeaderDefaults ClaudeHeaderDefaults `yaml:"claude-header-defaults" json:"claude-header-defaults"`
+
 	// OpenAICompatibility defines OpenAI API compatibility configurations for external providers.
 	OpenAICompatibility []OpenAICompatibility `yaml:"openai-compatibility" json:"openai-compatibility"`

@@ -130,6 +134,15 @@ type Config struct {
 	legacyMigrationPending bool `yaml:"-" json:"-"`
 }

+// ClaudeHeaderDefaults configures default header values injected into Claude API requests
+// when the client does not send them. Update these when Claude Code releases a new version.
+type ClaudeHeaderDefaults struct {
+	UserAgent      string `yaml:"user-agent" json:"user-agent"`
+	PackageVersion string `yaml:"package-version" json:"package-version"`
+	RuntimeVersion string `yaml:"runtime-version" json:"runtime-version"`
+	Timeout        string `yaml:"timeout" json:"timeout"`
+}
+
 // TLSConfig holds HTTPS server settings.
 type TLSConfig struct {
 	// Enable toggles HTTPS server mode.
@@ -368,6 +381,9 @@ type CodexKey struct {
 	// If empty, the default Codex API URL will be used.
 	BaseURL string `yaml:"base-url" json:"base-url"`

+	// Websockets enables the Responses API websocket transport for this credential.
+	Websockets bool `yaml:"websockets,omitempty" json:"websockets,omitempty"`
+
 	// ProxyURL overrides the global proxy setting for this API key if provided.
 	ProxyURL string `yaml:"proxy-url" json:"proxy-url"`

@@ -767,7 +783,13 @@ func (cfg *Config) SanitizeOAuthModelAlias() {
 	out := make(map[string][]OAuthModelAlias, len(cfg.OAuthModelAlias))
 	for rawChannel, aliases := range cfg.OAuthModelAlias {
 		channel := strings.ToLower(strings.TrimSpace(rawChannel))
-		if channel == "" || len(aliases) == 0 {
+		if channel == "" {
+			continue
+		}
+		// Preserve channels that were explicitly set to empty/nil – they act
+		// as "disabled" markers so default injection won't re-add them (#222).
+		if len(aliases) == 0 {
+			out[channel] = nil
 			continue
 		}
 		seenAlias := make(map[string]struct{}, len(aliases))
--- a/internal/config/oauth_model_alias_test.go
+++ b/internal/config/oauth_model_alias_test.go
@@ -128,6 +128,50 @@ func TestSanitizeOAuthModelAlias_DoesNotOverrideUserKiroAliases(t *testing.T) {
 	}
 }

+func TestSanitizeOAuthModelAlias_DoesNotReinjectAfterExplicitDeletion(t *testing.T) {
+	// When user explicitly deletes kiro aliases (key exists with nil value),
+	// defaults should NOT be re-injected on subsequent sanitize calls (#222).
+	cfg := &Config{
+		OAuthModelAlias: map[string][]OAuthModelAlias{
+			"kiro":  nil, // explicitly deleted
+			"codex": {{Name: "gpt-5", Alias: "g5"}},
+		},
+	}
+
+	cfg.SanitizeOAuthModelAlias()
+
+	kiroAliases := cfg.OAuthModelAlias["kiro"]
+	if len(kiroAliases) != 0 {
+		t.Fatalf("expected kiro aliases to remain empty after explicit deletion, got %d aliases", len(kiroAliases))
+	}
+	// The key itself must still be present to prevent re-injection on next reload
+	if _, exists := cfg.OAuthModelAlias["kiro"]; !exists {
+		t.Fatal("expected kiro key to be preserved as nil marker after sanitization")
+	}
+	// Other channels should be unaffected
+	if len(cfg.OAuthModelAlias["codex"]) != 1 {
+		t.Fatal("expected codex aliases to be preserved")
+	}
+}
+
+func TestSanitizeOAuthModelAlias_DoesNotReinjectAfterExplicitDeletionEmpty(t *testing.T) {
+	// Same as above but with empty slice instead of nil (PUT with empty body).
+	cfg := &Config{
+		OAuthModelAlias: map[string][]OAuthModelAlias{
+			"kiro": {}, // explicitly set to empty
+		},
+	}
+
+	cfg.SanitizeOAuthModelAlias()
+
+	if len(cfg.OAuthModelAlias["kiro"]) != 0 {
+		t.Fatalf("expected kiro aliases to remain empty, got %d aliases", len(cfg.OAuthModelAlias["kiro"]))
+	}
+	if _, exists := cfg.OAuthModelAlias["kiro"]; !exists {
+		t.Fatal("expected kiro key to be preserved")
+	}
+}
+
 func TestSanitizeOAuthModelAlias_InjectsDefaultKiroWhenEmpty(t *testing.T) {
 	// When OAuthModelAlias is nil, kiro defaults should still be injected
 	cfg := &Config{}
--- a/internal/config/sdk_config.go
+++ b/internal/config/sdk_config.go
@@ -20,6 +20,10 @@ type SDKConfig struct {
 	// APIKeys is a list of keys for authenticating clients to this proxy server.
 	APIKeys []string `yaml:"api-keys" json:"api-keys"`

+	// PassthroughHeaders controls whether upstream response headers are forwarded to downstream clients.
+	// Default is false (disabled).
+	PassthroughHeaders bool `yaml:"passthrough-headers" json:"passthrough-headers"`
+
 	// Streaming configures server-side streaming behavior (keep-alives and safe bootstrap retries).
 	Streaming StreamingConfig `yaml:"streaming" json:"streaming"`

--- a/internal/constant/constant.go
+++ b/internal/constant/constant.go
@@ -27,4 +27,7 @@ const (

 	// Kiro represents the AWS CodeWhisperer (Kiro) provider identifier.
 	Kiro = "kiro"
+
+	// Kilo represents the Kilo AI provider identifier.
+	Kilo = "kilo"
 )
--- a/internal/registry/kilo_models.go
+++ b/internal/registry/kilo_models.go
@@ -0,0 +1,21 @@
+// Package registry provides model definitions for various AI service providers.
+package registry
+
+// GetKiloModels returns the Kilo model definitions
+func GetKiloModels() []*ModelInfo {
+	return []*ModelInfo{
+		// --- Base Models ---
+		{
+			ID:                  "kilo/auto",
+			Object:              "model",
+			Created:             1732752000,
+			OwnedBy:             "kilo",
+			Type:                "kilo",
+			DisplayName:         "Kilo Auto",
+			Description:         "Automatic model selection by Kilo",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
+		},
+	}
+}
--- a/internal/registry/model_definitions.go
+++ b/internal/registry/model_definitions.go
@@ -19,7 +19,9 @@ import (
 //   - codex
 //   - qwen
 //   - iflow
+//   - kimi
 //   - kiro
+//   - kilo
 //   - github-copilot
 //   - kiro
 //   - amazonq
@@ -43,10 +45,14 @@ func GetStaticModelDefinitionsByChannel(channel string) []*ModelInfo {
 		return GetQwenModels()
 	case "iflow":
 		return GetIFlowModels()
+	case "kimi":
+		return GetKimiModels()
 	case "github-copilot":
 		return GetGitHubCopilotModels()
 	case "kiro":
 		return GetKiroModels()
+	case "kilo":
+		return GetKiloModels()
 	case "amazonq":
 		return GetAmazonQModels()
 	case "antigravity":
@@ -93,8 +99,10 @@ func LookupStaticModelInfo(modelID string) *ModelInfo {
 		GetOpenAIModels(),
 		GetQwenModels(),
 		GetIFlowModels(),
+		GetKimiModels(),
 		GetGitHubCopilotModels(),
 		GetKiroModels(),
+		GetKiloModels(),
 		GetAmazonQModels(),
 	}
 	for _, models := range allModels {
@@ -144,6 +152,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       200000,
 			MaxCompletionTokens: 32768,
 			SupportedEndpoints:  []string{"/chat/completions", "/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5-mini",
@@ -156,6 +165,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       128000,
 			MaxCompletionTokens: 16384,
 			SupportedEndpoints:  []string{"/chat/completions", "/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5-codex",
@@ -168,6 +178,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       200000,
 			MaxCompletionTokens: 32768,
 			SupportedEndpoints:  []string{"/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5.1",
@@ -180,6 +191,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       200000,
 			MaxCompletionTokens: 32768,
 			SupportedEndpoints:  []string{"/chat/completions", "/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5.1-codex",
@@ -192,6 +204,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       200000,
 			MaxCompletionTokens: 32768,
 			SupportedEndpoints:  []string{"/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5.1-codex-mini",
@@ -204,6 +217,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       128000,
 			MaxCompletionTokens: 16384,
 			SupportedEndpoints:  []string{"/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5.1-codex-max",
@@ -216,6 +230,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       200000,
 			MaxCompletionTokens: 32768,
 			SupportedEndpoints:  []string{"/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high", "xhigh"}},
 		},
 		{
 			ID:                  "gpt-5.2",
@@ -228,6 +243,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       200000,
 			MaxCompletionTokens: 32768,
 			SupportedEndpoints:  []string{"/chat/completions", "/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high", "xhigh"}},
 		},
 		{
 			ID:                  "gpt-5.2-codex",
@@ -240,6 +256,20 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       200000,
 			MaxCompletionTokens: 32768,
 			SupportedEndpoints:  []string{"/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high", "xhigh"}},
+		},
+		{
+			ID:                  "gpt-5.3-codex",
+			Object:              "model",
+			Created:             now,
+			OwnedBy:             "github-copilot",
+			Type:                "github-copilot",
+			DisplayName:         "GPT-5.3 Codex",
+			Description:         "OpenAI GPT-5.3 Codex via GitHub Copilot",
+			ContextLength:       200000,
+			MaxCompletionTokens: 32768,
+			SupportedEndpoints:  []string{"/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high", "xhigh"}},
 		},
 		{
 			ID:                  "claude-haiku-4.5",
@@ -313,6 +343,18 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			MaxCompletionTokens: 64000,
 			SupportedEndpoints:  []string{"/chat/completions"},
 		},
+		{
+			ID:                  "claude-sonnet-4.6",
+			Object:              "model",
+			Created:             now,
+			OwnedBy:             "github-copilot",
+			Type:                "github-copilot",
+			DisplayName:         "Claude Sonnet 4.6",
+			Description:         "Anthropic Claude Sonnet 4.6 via GitHub Copilot",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			SupportedEndpoints:  []string{"/chat/completions"},
+		},
 		{
 			ID:                  "gemini-2.5-pro",
 			Object:              "model",
@@ -400,6 +442,18 @@ func GetKiroModels() []*ModelInfo {
 			MaxCompletionTokens: 64000,
 			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
 		},
+		{
+			ID:                  "kiro-claude-sonnet-4-6",
+			Object:              "model",
+			Created:             1739836800, // 2025-02-18
+			OwnedBy:             "aws",
+			Type:                "kiro",
+			DisplayName:         "Kiro Claude Sonnet 4.6",
+			Description:         "Claude Sonnet 4.6 via Kiro (1.3x credit)",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
+		},
 		{
 			ID:                  "kiro-claude-opus-4-5",
 			Object:              "model",
@@ -448,6 +502,87 @@ func GetKiroModels() []*ModelInfo {
 			MaxCompletionTokens: 64000,
 			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
 		},
+		// --- 第三方模型 (通过 Kiro 接入) ---
+		{
+			ID:                  "kiro-deepseek-3-2",
+			Object:              "model",
+			Created:             1732752000,
+			OwnedBy:             "aws",
+			Type:                "kiro",
+			DisplayName:         "Kiro DeepSeek 3.2",
+			Description:         "DeepSeek 3.2 via Kiro",
+			ContextLength:       128000,
+			MaxCompletionTokens: 32768,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                  "kiro-minimax-m2-1",
+			Object:              "model",
+			Created:             1732752000,
+			OwnedBy:             "aws",
+			Type:                "kiro",
+			DisplayName:         "Kiro MiniMax M2.1",
+			Description:         "MiniMax M2.1 via Kiro",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                  "kiro-qwen3-coder-next",
+			Object:              "model",
+			Created:             1732752000,
+			OwnedBy:             "aws",
+			Type:                "kiro",
+			DisplayName:         "Kiro Qwen3 Coder Next",
+			Description:         "Qwen3 Coder Next via Kiro",
+			ContextLength:       128000,
+			MaxCompletionTokens: 32768,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                  "kiro-gpt-4o",
+			Object:              "model",
+			Created:             1732752000,
+			OwnedBy:             "aws",
+			Type:                "kiro",
+			DisplayName:         "Kiro GPT-4o",
+			Description:         "OpenAI GPT-4o via Kiro",
+			ContextLength:       128000,
+			MaxCompletionTokens: 16384,
+		},
+		{
+			ID:                  "kiro-gpt-4",
+			Object:              "model",
+			Created:             1732752000,
+			OwnedBy:             "aws",
+			Type:                "kiro",
+			DisplayName:         "Kiro GPT-4",
+			Description:         "OpenAI GPT-4 via Kiro",
+			ContextLength:       128000,
+			MaxCompletionTokens: 8192,
+		},
+		{
+			ID:                  "kiro-gpt-4-turbo",
+			Object:              "model",
+			Created:             1732752000,
+			OwnedBy:             "aws",
+			Type:                "kiro",
+			DisplayName:         "Kiro GPT-4 Turbo",
+			Description:         "OpenAI GPT-4 Turbo via Kiro",
+			ContextLength:       128000,
+			MaxCompletionTokens: 16384,
+		},
+		{
+			ID:                  "kiro-gpt-3-5-turbo",
+			Object:              "model",
+			Created:             1732752000,
+			OwnedBy:             "aws",
+			Type:                "kiro",
+			DisplayName:         "Kiro GPT-3.5 Turbo",
+			Description:         "OpenAI GPT-3.5 Turbo via Kiro",
+			ContextLength:       16384,
+			MaxCompletionTokens: 4096,
+		},
 		// --- Agentic Variants (Optimized for coding agents with chunked writes) ---
 		{
 			ID:                  "kiro-claude-opus-4-6-agentic",
@@ -461,6 +596,18 @@ func GetKiroModels() []*ModelInfo {
 			MaxCompletionTokens: 64000,
 			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
 		},
+		{
+			ID:                  "kiro-claude-sonnet-4-6-agentic",
+			Object:              "model",
+			Created:             1739836800, // 2025-02-18
+			OwnedBy:             "aws",
+			Type:                "kiro",
+			DisplayName:         "Kiro Claude Sonnet 4.6 (Agentic)",
+			Description:         "Claude Sonnet 4.6 optimized for coding agents (chunked writes)",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
+		},
 		{
 			ID:                  "kiro-claude-opus-4-5-agentic",
 			Object:              "model",
--- a/internal/registry/model_definitions_static_data.go
+++ b/internal/registry/model_definitions_static_data.go
@@ -28,6 +28,17 @@ func GetClaudeModels() []*ModelInfo {
 			MaxCompletionTokens: 64000,
 			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
 		},
+		{
+			ID:                  "claude-sonnet-4-6",
+			Object:              "model",
+			Created:             1771372800, // 2026-02-17
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4.6 Sonnet",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
+		},
 		{
 			ID:                  "claude-opus-4-6",
 			Object:              "model",
@@ -40,6 +51,18 @@ func GetClaudeModels() []*ModelInfo {
 			MaxCompletionTokens: 128000,
 			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
 		},
+		{
+			ID:                  "claude-sonnet-4-6",
+			Object:              "model",
+			Created:             1771286400, // 2026-02-17
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4.6 Sonnet",
+			Description:         "Best combination of speed and intelligence",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
+		},
 		{
 			ID:                  "claude-opus-4-5-20251101",
 			Object:              "model",
@@ -742,6 +765,20 @@ func GetOpenAIModels() []*ModelInfo {
 			SupportedParameters: []string{"tools"},
 			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
 		},
+		{
+			ID:                  "gpt-5.3-codex-spark",
+			Object:              "model",
+			Created:             1770912000,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5.3",
+			DisplayName:         "GPT 5.3 Codex Spark",
+			Description:         "Ultra-fast coding model.",
+			ContextLength:       128000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
+		},
 	}
 }

@@ -774,6 +811,19 @@ func GetQwenModels() []*ModelInfo {
 			MaxCompletionTokens: 2048,
 			SupportedParameters: []string{"temperature", "top_p", "max_tokens", "stream", "stop"},
 		},
+		{
+			ID:                  "coder-model",
+			Object:              "model",
+			Created:             1771171200,
+			OwnedBy:             "qwen",
+			Type:                "qwen",
+			Version:             "3.5",
+			DisplayName:         "Qwen 3.5 Plus",
+			Description:         "efficient hybrid model with leading coding performance",
+			ContextLength:       1048576,
+			MaxCompletionTokens: 65536,
+			SupportedParameters: []string{"temperature", "top_p", "max_tokens", "stream", "stop"},
+		},
 		{
 			ID:                  "vision-model",
 			Object:              "model",
@@ -829,6 +879,7 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "qwen3-235b", DisplayName: "Qwen3-235B-A22B", Description: "Qwen3 235B A22B", Created: 1753401600},
 		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000, Thinking: iFlowThinkingSupport},
 		{ID: "minimax-m2.1", DisplayName: "MiniMax-M2.1", Description: "MiniMax M2.1", Created: 1766448000, Thinking: iFlowThinkingSupport},
+		{ID: "minimax-m2.5", DisplayName: "MiniMax-M2.5", Description: "MiniMax M2.5", Created: 1770825600, Thinking: iFlowThinkingSupport},
 		{ID: "iflow-rome-30ba3b", DisplayName: "iFlow-ROME", Description: "iFlow Rome 30BA3B model", Created: 1736899200},
 		{ID: "kimi-k2.5", DisplayName: "Kimi-K2.5", Description: "Moonshot Kimi K2.5", Created: 1769443200, Thinking: iFlowThinkingSupport},
 	}
@@ -865,10 +916,12 @@ func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 		"gemini-3-pro-high":          {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
 		"gemini-3-pro-image":         {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
 		"gemini-3-flash":             {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}},
-		"claude-sonnet-4-5-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-opus-4-5-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-opus-4-6-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-sonnet-4-5":          {MaxCompletionTokens: 64000},
+		"claude-sonnet-4-5-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
+		"claude-sonnet-4-6":          {MaxCompletionTokens: 64000},
+		"claude-sonnet-4-6-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"gpt-oss-120b-medium":        {},
 		"tab_flash_lite_preview":     {},
 	}
--- a/internal/registry/model_registry.go
+++ b/internal/registry/model_registry.go
@@ -601,8 +601,7 @@ func (r *ModelRegistry) SetModelQuotaExceeded(clientID, modelID string) {
 	defer r.mutex.Unlock()

 	if registration, exists := r.models[modelID]; exists {
-		now := time.Now()
-		registration.QuotaExceededClients[clientID] = &now
+		registration.QuotaExceededClients[clientID] = new(time.Now())
 		log.Debugf("Marked model %s as quota exceeded for client %s", modelID, clientID)
 	}
 }
--- a/internal/runtime/executor/aistudio_executor.go
+++ b/internal/runtime/executor/aistudio_executor.go
@@ -164,12 +164,12 @@ func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth,
 	reporter.publish(ctx, parseGeminiUsage(wsResp.Body))
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, body.toFormat, opts.SourceFormat, req.Model, opts.OriginalRequest, translatedReq, wsResp.Body, &param)
-	resp = cliproxyexecutor.Response{Payload: ensureColonSpacedJSON([]byte(out))}
+	resp = cliproxyexecutor.Response{Payload: ensureColonSpacedJSON([]byte(out)), Headers: wsResp.Headers.Clone()}
 	return resp, nil
 }

 // ExecuteStream performs a streaming request to the AI Studio API.
-func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -254,7 +254,6 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 		return nil, statusErr{code: firstEvent.Status, msg: body.String()}
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func(first wsrelay.StreamEvent) {
 		defer close(out)
 		var param any
@@ -318,7 +317,7 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 			}
 		}
 	}(firstEvent)
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: firstEvent.Headers.Clone(), Chunks: out}, nil
 }

 // CountTokens counts tokens for the given request using the AI Studio API.
--- a/internal/runtime/executor/antigravity_executor.go
+++ b/internal/runtime/executor/antigravity_executor.go
@@ -232,7 +232,7 @@ attemptLoop:
 			reporter.publish(ctx, parseAntigravityUsage(bodyBytes))
 			var param any
 			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, bodyBytes, &param)
-			resp = cliproxyexecutor.Response{Payload: []byte(converted)}
+			resp = cliproxyexecutor.Response{Payload: []byte(converted), Headers: httpResp.Header.Clone()}
 			reporter.ensurePublished(ctx)
 			return resp, nil
 		}
@@ -436,7 +436,7 @@ attemptLoop:
 			reporter.publish(ctx, parseAntigravityUsage(resp.Payload))
 			var param any
 			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, resp.Payload, &param)
-			resp = cliproxyexecutor.Response{Payload: []byte(converted)}
+			resp = cliproxyexecutor.Response{Payload: []byte(converted), Headers: httpResp.Header.Clone()}
 			reporter.ensurePublished(ctx)

 			return resp, nil
@@ -645,7 +645,7 @@ func (e *AntigravityExecutor) convertStreamToNonStream(stream []byte) []byte {
 }

 // ExecuteStream performs a streaming request to the Antigravity API.
-func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -775,7 +775,6 @@ attemptLoop:
 			}

 			out := make(chan cliproxyexecutor.StreamChunk)
-			stream = out
 			go func(resp *http.Response) {
 				defer close(out)
 				defer func() {
@@ -820,7 +819,7 @@ attemptLoop:
 					reporter.ensurePublished(ctx)
 				}
 			}(httpResp)
-			return stream, nil
+			return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 		}

 		switch {
@@ -968,7 +967,7 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 		if httpResp.StatusCode >= http.StatusOK && httpResp.StatusCode < http.StatusMultipleChoices {
 			count := gjson.GetBytes(bodyBytes, "totalTokens").Int()
 			translated := sdktranslator.TranslateTokenCount(respCtx, to, from, count, bodyBytes)
-			return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
+			return cliproxyexecutor.Response{Payload: []byte(translated), Headers: httpResp.Header.Clone()}, nil
 		}

 		lastStatus = httpResp.StatusCode
@@ -1007,7 +1006,12 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *config.Config) []*registry.ModelInfo {
 	exec := &AntigravityExecutor{cfg: cfg}
 	token, updatedAuth, errToken := exec.ensureAccessToken(ctx, auth)
-	if errToken != nil || token == "" {
+	if errToken != nil {
+		log.Warnf("antigravity executor: fetch models failed for %s: token error: %v", auth.ID, errToken)
+		return nil
+	}
+	if token == "" {
+		log.Warnf("antigravity executor: fetch models failed for %s: got empty token", auth.ID)
 		return nil
 	}
 	if updatedAuth != nil {
@@ -1021,6 +1025,7 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 		modelsURL := baseURL + antigravityModelsPath
 		httpReq, errReq := http.NewRequestWithContext(ctx, http.MethodPost, modelsURL, bytes.NewReader([]byte(`{}`)))
 		if errReq != nil {
+			log.Warnf("antigravity executor: fetch models failed for %s: create request error: %v", auth.ID, errReq)
 			return nil
 		}
 		httpReq.Header.Set("Content-Type", "application/json")
@@ -1033,12 +1038,14 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
 			if errors.Is(errDo, context.Canceled) || errors.Is(errDo, context.DeadlineExceeded) {
+				log.Warnf("antigravity executor: fetch models failed for %s: context canceled: %v", auth.ID, errDo)
 				return nil
 			}
 			if idx+1 < len(baseURLs) {
 				log.Debugf("antigravity executor: models request error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 				continue
 			}
+			log.Warnf("antigravity executor: fetch models failed for %s: request error: %v", auth.ID, errDo)
 			return nil
 		}

@@ -1051,6 +1058,7 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 				log.Debugf("antigravity executor: models read error on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 				continue
 			}
+			log.Warnf("antigravity executor: fetch models failed for %s: read body error: %v", auth.ID, errRead)
 			return nil
 		}
 		if httpResp.StatusCode < http.StatusOK || httpResp.StatusCode >= http.StatusMultipleChoices {
@@ -1058,11 +1066,13 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 				log.Debugf("antigravity executor: models request rate limited on base url %s, retrying with fallback base url: %s", baseURL, baseURLs[idx+1])
 				continue
 			}
+			log.Warnf("antigravity executor: fetch models failed for %s: unexpected status %d, body: %s", auth.ID, httpResp.StatusCode, string(bodyBytes))
 			return nil
 		}

 		result := gjson.GetBytes(bodyBytes, "models")
 		if !result.Exists() {
+			log.Warnf("antigravity executor: fetch models failed for %s: no models field in response, body: %s", auth.ID, string(bodyBytes))
 			return nil
 		}

--- a/internal/runtime/executor/claude_executor.go
+++ b/internal/runtime/executor/claude_executor.go
@@ -9,6 +9,7 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"runtime"
 	"strings"
 	"time"

@@ -134,7 +135,7 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	extraBetas, body = extractAndRemoveBetas(body)
 	bodyForTranslation := body
 	bodyForUpstream := body
-	if isClaudeOAuthToken(apiKey) {
+	if isClaudeOAuthToken(apiKey) && !auth.ToolPrefixDisabled() {
 		bodyForUpstream = applyClaudeToolPrefix(body, claudeToolPrefix)
 	}

@@ -143,7 +144,7 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	if err != nil {
 		return resp, err
 	}
-	applyClaudeHeaders(httpReq, auth, apiKey, false, extraBetas)
+	applyClaudeHeaders(httpReq, auth, apiKey, false, extraBetas, e.cfg)
 	var authID, authLabel, authType, authValue string
 	if auth != nil {
 		authID = auth.ID
@@ -208,7 +209,7 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	} else {
 		reporter.publish(ctx, parseClaudeUsage(data))
 	}
-	if isClaudeOAuthToken(apiKey) {
+	if isClaudeOAuthToken(apiKey) && !auth.ToolPrefixDisabled() {
 		data = stripClaudeToolPrefixFromResponse(data, claudeToolPrefix)
 	}
 	var param any
@@ -222,11 +223,11 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 		data,
 		&param,
 	)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }

-func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -275,7 +276,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	extraBetas, body = extractAndRemoveBetas(body)
 	bodyForTranslation := body
 	bodyForUpstream := body
-	if isClaudeOAuthToken(apiKey) {
+	if isClaudeOAuthToken(apiKey) && !auth.ToolPrefixDisabled() {
 		bodyForUpstream = applyClaudeToolPrefix(body, claudeToolPrefix)
 	}

@@ -284,7 +285,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	if err != nil {
 		return nil, err
 	}
-	applyClaudeHeaders(httpReq, auth, apiKey, true, extraBetas)
+	applyClaudeHeaders(httpReq, auth, apiKey, true, extraBetas, e.cfg)
 	var authID, authLabel, authType, authValue string
 	if auth != nil {
 		authID = auth.ID
@@ -329,7 +330,6 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -348,7 +348,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 				if detail, ok := parseClaudeStreamUsage(line); ok {
 					reporter.publish(ctx, detail)
 				}
-				if isClaudeOAuthToken(apiKey) {
+				if isClaudeOAuthToken(apiKey) && !auth.ToolPrefixDisabled() {
 					line = stripClaudeToolPrefixFromStreamLine(line, claudeToolPrefix)
 				}
 				// Forward the line as-is to preserve SSE format
@@ -375,7 +375,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 			if detail, ok := parseClaudeStreamUsage(line); ok {
 				reporter.publish(ctx, detail)
 			}
-			if isClaudeOAuthToken(apiKey) {
+			if isClaudeOAuthToken(apiKey) && !auth.ToolPrefixDisabled() {
 				line = stripClaudeToolPrefixFromStreamLine(line, claudeToolPrefix)
 			}
 			chunks := sdktranslator.TranslateStream(
@@ -398,7 +398,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }

 func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
@@ -423,7 +423,7 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	// Extract betas from body and convert to header (for count_tokens too)
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
-	if isClaudeOAuthToken(apiKey) {
+	if isClaudeOAuthToken(apiKey) && !auth.ToolPrefixDisabled() {
 		body = applyClaudeToolPrefix(body, claudeToolPrefix)
 	}

@@ -432,7 +432,7 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}
-	applyClaudeHeaders(httpReq, auth, apiKey, false, extraBetas)
+	applyClaudeHeaders(httpReq, auth, apiKey, false, extraBetas, e.cfg)
 	var authID, authLabel, authType, authValue string
 	if auth != nil {
 		authID = auth.ID
@@ -487,7 +487,7 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	count := gjson.GetBytes(data, "input_tokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(out), Headers: resp.Header.Clone()}, nil
 }

 func (e *ClaudeExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
@@ -638,7 +638,49 @@ func decodeResponseBody(body io.ReadCloser, contentEncoding string) (io.ReadClos
 	return body, nil
 }

-func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string, stream bool, extraBetas []string) {
+// mapStainlessOS maps runtime.GOOS to Stainless SDK OS names.
+func mapStainlessOS() string {
+	switch runtime.GOOS {
+	case "darwin":
+		return "MacOS"
+	case "windows":
+		return "Windows"
+	case "linux":
+		return "Linux"
+	case "freebsd":
+		return "FreeBSD"
+	default:
+		return "Other::" + runtime.GOOS
+	}
+}
+
+// mapStainlessArch maps runtime.GOARCH to Stainless SDK architecture names.
+func mapStainlessArch() string {
+	switch runtime.GOARCH {
+	case "amd64":
+		return "x64"
+	case "arm64":
+		return "arm64"
+	case "386":
+		return "x86"
+	default:
+		return "other::" + runtime.GOARCH
+	}
+}
+
+func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string, stream bool, extraBetas []string, cfg *config.Config) {
+	hdrDefault := func(cfgVal, fallback string) string {
+		if cfgVal != "" {
+			return cfgVal
+		}
+		return fallback
+	}
+
+	var hd config.ClaudeHeaderDefaults
+	if cfg != nil {
+		hd = cfg.ClaudeHeaderDefaults
+	}
+
 	useAPIKey := auth != nil && auth.Attributes != nil && strings.TrimSpace(auth.Attributes["api_key"]) != ""
 	isAnthropicBase := r.URL != nil && strings.EqualFold(r.URL.Scheme, "https") && strings.EqualFold(r.URL.Host, "api.anthropic.com")
 	if isAnthropicBase && useAPIKey {
@@ -685,16 +727,17 @@ func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string,
 	misc.EnsureHeader(r.Header, ginHeaders, "Anthropic-Version", "2023-06-01")
 	misc.EnsureHeader(r.Header, ginHeaders, "Anthropic-Dangerous-Direct-Browser-Access", "true")
 	misc.EnsureHeader(r.Header, ginHeaders, "X-App", "cli")
+	// Values below match Claude Code 2.1.44 / @anthropic-ai/sdk 0.74.0 (captured 2026-02-17).
 	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Helper-Method", "stream")
 	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Retry-Count", "0")
-	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Runtime-Version", "v24.3.0")
-	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Package-Version", "0.55.1")
+	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Runtime-Version", hdrDefault(hd.RuntimeVersion, "v24.3.0"))
+	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Package-Version", hdrDefault(hd.PackageVersion, "0.74.0"))
 	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Runtime", "node")
 	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Lang", "js")
-	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Arch", "arm64")
-	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Os", "MacOS")
-	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Timeout", "60")
-	misc.EnsureHeader(r.Header, ginHeaders, "User-Agent", "claude-cli/1.0.83 (external, cli)")
+	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Arch", mapStainlessArch())
+	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Os", mapStainlessOS())
+	misc.EnsureHeader(r.Header, ginHeaders, "X-Stainless-Timeout", hdrDefault(hd.Timeout, "600"))
+	misc.EnsureHeader(r.Header, ginHeaders, "User-Agent", hdrDefault(hd.UserAgent, "claude-cli/2.1.44 (external, sdk-cli)"))
 	r.Header.Set("Connection", "keep-alive")
 	r.Header.Set("Accept-Encoding", "gzip, deflate, br, zstd")
 	if stream {
@@ -702,6 +745,8 @@ func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string,
 	} else {
 		r.Header.Set("Accept", "application/json")
 	}
+	// Keep OS/Arch mapping dynamic (not configurable).
+	// They intentionally continue to derive from runtime.GOOS/runtime.GOARCH.
 	var attrs map[string]string
 	if auth != nil {
 		attrs = auth.Attributes
@@ -753,11 +798,21 @@ func applyClaudeToolPrefix(body []byte, prefix string) []byte {
 		return body
 	}

+	// Collect built-in tool names (those with a non-empty "type" field) so we can
+	// skip them consistently in both tools and message history.
+	builtinTools := map[string]bool{}
+	for _, name := range []string{"web_search", "code_execution", "text_editor", "computer"} {
+		builtinTools[name] = true
+	}
+
 	if tools := gjson.GetBytes(body, "tools"); tools.Exists() && tools.IsArray() {
 		tools.ForEach(func(index, tool gjson.Result) bool {
 			// Skip built-in tools (web_search, code_execution, etc.) which have
 			// a "type" field and require their name to remain unchanged.
 			if tool.Get("type").Exists() && tool.Get("type").String() != "" {
+				if n := tool.Get("name").String(); n != "" {
+					builtinTools[n] = true
+				}
 				return true
 			}
 			name := tool.Get("name").String()
@@ -772,7 +827,7 @@ func applyClaudeToolPrefix(body []byte, prefix string) []byte {

 	if gjson.GetBytes(body, "tool_choice.type").String() == "tool" {
 		name := gjson.GetBytes(body, "tool_choice.name").String()
-		if name != "" && !strings.HasPrefix(name, prefix) {
+		if name != "" && !strings.HasPrefix(name, prefix) && !builtinTools[name] {
 			body, _ = sjson.SetBytes(body, "tool_choice.name", prefix+name)
 		}
 	}
@@ -784,15 +839,38 @@ func applyClaudeToolPrefix(body []byte, prefix string) []byte {
 				return true
 			}
 			content.ForEach(func(contentIndex, part gjson.Result) bool {
-				if part.Get("type").String() != "tool_use" {
-					return true
+				partType := part.Get("type").String()
+				switch partType {
+				case "tool_use":
+					name := part.Get("name").String()
+					if name == "" || strings.HasPrefix(name, prefix) || builtinTools[name] {
+						return true
+					}
+					path := fmt.Sprintf("messages.%d.content.%d.name", msgIndex.Int(), contentIndex.Int())
+					body, _ = sjson.SetBytes(body, path, prefix+name)
+				case "tool_reference":
+					toolName := part.Get("tool_name").String()
+					if toolName == "" || strings.HasPrefix(toolName, prefix) || builtinTools[toolName] {
+						return true
+					}
+					path := fmt.Sprintf("messages.%d.content.%d.tool_name", msgIndex.Int(), contentIndex.Int())
+					body, _ = sjson.SetBytes(body, path, prefix+toolName)
+				case "tool_result":
+					// Handle nested tool_reference blocks inside tool_result.content[]
+					nestedContent := part.Get("content")
+					if nestedContent.Exists() && nestedContent.IsArray() {
+						nestedContent.ForEach(func(nestedIndex, nestedPart gjson.Result) bool {
+							if nestedPart.Get("type").String() == "tool_reference" {
+								nestedToolName := nestedPart.Get("tool_name").String()
+								if nestedToolName != "" && !strings.HasPrefix(nestedToolName, prefix) && !builtinTools[nestedToolName] {
+									nestedPath := fmt.Sprintf("messages.%d.content.%d.content.%d.tool_name", msgIndex.Int(), contentIndex.Int(), nestedIndex.Int())
+									body, _ = sjson.SetBytes(body, nestedPath, prefix+nestedToolName)
+								}
+							}
+							return true
+						})
+					}
 				}
-				name := part.Get("name").String()
-				if name == "" || strings.HasPrefix(name, prefix) {
-					return true
-				}
-				path := fmt.Sprintf("messages.%d.content.%d.name", msgIndex.Int(), contentIndex.Int())
-				body, _ = sjson.SetBytes(body, path, prefix+name)
 				return true
 			})
 			return true
@@ -811,15 +889,38 @@ func stripClaudeToolPrefixFromResponse(body []byte, prefix string) []byte {
 		return body
 	}
 	content.ForEach(func(index, part gjson.Result) bool {
-		if part.Get("type").String() != "tool_use" {
-			return true
+		partType := part.Get("type").String()
+		switch partType {
+		case "tool_use":
+			name := part.Get("name").String()
+			if !strings.HasPrefix(name, prefix) {
+				return true
+			}
+			path := fmt.Sprintf("content.%d.name", index.Int())
+			body, _ = sjson.SetBytes(body, path, strings.TrimPrefix(name, prefix))
+		case "tool_reference":
+			toolName := part.Get("tool_name").String()
+			if !strings.HasPrefix(toolName, prefix) {
+				return true
+			}
+			path := fmt.Sprintf("content.%d.tool_name", index.Int())
+			body, _ = sjson.SetBytes(body, path, strings.TrimPrefix(toolName, prefix))
+		case "tool_result":
+			// Handle nested tool_reference blocks inside tool_result.content[]
+			nestedContent := part.Get("content")
+			if nestedContent.Exists() && nestedContent.IsArray() {
+				nestedContent.ForEach(func(nestedIndex, nestedPart gjson.Result) bool {
+					if nestedPart.Get("type").String() == "tool_reference" {
+						nestedToolName := nestedPart.Get("tool_name").String()
+						if strings.HasPrefix(nestedToolName, prefix) {
+							nestedPath := fmt.Sprintf("content.%d.content.%d.tool_name", index.Int(), nestedIndex.Int())
+							body, _ = sjson.SetBytes(body, nestedPath, strings.TrimPrefix(nestedToolName, prefix))
+						}
+					}
+					return true
+				})
+			}
 		}
-		name := part.Get("name").String()
-		if !strings.HasPrefix(name, prefix) {
-			return true
-		}
-		path := fmt.Sprintf("content.%d.name", index.Int())
-		body, _ = sjson.SetBytes(body, path, strings.TrimPrefix(name, prefix))
 		return true
 	})
 	return body
@@ -834,15 +935,34 @@ func stripClaudeToolPrefixFromStreamLine(line []byte, prefix string) []byte {
 		return line
 	}
 	contentBlock := gjson.GetBytes(payload, "content_block")
-	if !contentBlock.Exists() || contentBlock.Get("type").String() != "tool_use" {
+	if !contentBlock.Exists() {
 		return line
 	}
-	name := contentBlock.Get("name").String()
-	if !strings.HasPrefix(name, prefix) {
-		return line
-	}
-	updated, err := sjson.SetBytes(payload, "content_block.name", strings.TrimPrefix(name, prefix))
-	if err != nil {
+
+	blockType := contentBlock.Get("type").String()
+	var updated []byte
+	var err error
+
+	switch blockType {
+	case "tool_use":
+		name := contentBlock.Get("name").String()
+		if !strings.HasPrefix(name, prefix) {
+			return line
+		}
+		updated, err = sjson.SetBytes(payload, "content_block.name", strings.TrimPrefix(name, prefix))
+		if err != nil {
+			return line
+		}
+	case "tool_reference":
+		toolName := contentBlock.Get("tool_name").String()
+		if !strings.HasPrefix(toolName, prefix) {
+			return line
+		}
+		updated, err = sjson.SetBytes(payload, "content_block.tool_name", strings.TrimPrefix(toolName, prefix))
+		if err != nil {
+			return line
+		}
+	default:
 		return line
 	}

--- a/internal/runtime/executor/claude_executor_test.go
+++ b/internal/runtime/executor/claude_executor_test.go
@@ -25,6 +25,18 @@ func TestApplyClaudeToolPrefix(t *testing.T) {
 	}
 }

+func TestApplyClaudeToolPrefix_WithToolReference(t *testing.T) {
+	input := []byte(`{"tools":[{"name":"alpha"}],"messages":[{"role":"user","content":[{"type":"tool_reference","tool_name":"beta"},{"type":"tool_reference","tool_name":"proxy_gamma"}]}]}`)
+	out := applyClaudeToolPrefix(input, "proxy_")
+
+	if got := gjson.GetBytes(out, "messages.0.content.0.tool_name").String(); got != "proxy_beta" {
+		t.Fatalf("messages.0.content.0.tool_name = %q, want %q", got, "proxy_beta")
+	}
+	if got := gjson.GetBytes(out, "messages.0.content.1.tool_name").String(); got != "proxy_gamma" {
+		t.Fatalf("messages.0.content.1.tool_name = %q, want %q", got, "proxy_gamma")
+	}
+}
+
 func TestApplyClaudeToolPrefix_SkipsBuiltinTools(t *testing.T) {
 	input := []byte(`{"tools":[{"type":"web_search_20250305","name":"web_search"},{"name":"my_custom_tool","input_schema":{"type":"object"}}]}`)
 	out := applyClaudeToolPrefix(input, "proxy_")
@@ -37,6 +49,97 @@ func TestApplyClaudeToolPrefix_SkipsBuiltinTools(t *testing.T) {
 	}
 }

+func TestApplyClaudeToolPrefix_BuiltinToolSkipped(t *testing.T) {
+	body := []byte(`{
+		"tools": [
+			{"type": "web_search_20250305", "name": "web_search", "max_uses": 5},
+			{"name": "Read"}
+		],
+		"messages": [
+			{"role": "user", "content": [
+				{"type": "tool_use", "name": "web_search", "id": "ws1", "input": {}},
+				{"type": "tool_use", "name": "Read", "id": "r1", "input": {}}
+			]}
+		]
+	}`)
+	out := applyClaudeToolPrefix(body, "proxy_")
+
+	if got := gjson.GetBytes(out, "tools.0.name").String(); got != "web_search" {
+		t.Fatalf("tools.0.name = %q, want %q", got, "web_search")
+	}
+	if got := gjson.GetBytes(out, "messages.0.content.0.name").String(); got != "web_search" {
+		t.Fatalf("messages.0.content.0.name = %q, want %q", got, "web_search")
+	}
+	if got := gjson.GetBytes(out, "tools.1.name").String(); got != "proxy_Read" {
+		t.Fatalf("tools.1.name = %q, want %q", got, "proxy_Read")
+	}
+	if got := gjson.GetBytes(out, "messages.0.content.1.name").String(); got != "proxy_Read" {
+		t.Fatalf("messages.0.content.1.name = %q, want %q", got, "proxy_Read")
+	}
+}
+
+func TestApplyClaudeToolPrefix_KnownBuiltinInHistoryOnly(t *testing.T) {
+	body := []byte(`{
+		"tools": [
+			{"name": "Read"}
+		],
+		"messages": [
+			{"role": "user", "content": [
+				{"type": "tool_use", "name": "web_search", "id": "ws1", "input": {}}
+			]}
+		]
+	}`)
+	out := applyClaudeToolPrefix(body, "proxy_")
+
+	if got := gjson.GetBytes(out, "messages.0.content.0.name").String(); got != "web_search" {
+		t.Fatalf("messages.0.content.0.name = %q, want %q", got, "web_search")
+	}
+	if got := gjson.GetBytes(out, "tools.0.name").String(); got != "proxy_Read" {
+		t.Fatalf("tools.0.name = %q, want %q", got, "proxy_Read")
+	}
+}
+
+func TestApplyClaudeToolPrefix_CustomToolsPrefixed(t *testing.T) {
+	body := []byte(`{
+		"tools": [{"name": "Read"}, {"name": "Write"}],
+		"messages": [
+			{"role": "user", "content": [
+				{"type": "tool_use", "name": "Read", "id": "r1", "input": {}},
+				{"type": "tool_use", "name": "Write", "id": "w1", "input": {}}
+			]}
+		]
+	}`)
+	out := applyClaudeToolPrefix(body, "proxy_")
+
+	if got := gjson.GetBytes(out, "tools.0.name").String(); got != "proxy_Read" {
+		t.Fatalf("tools.0.name = %q, want %q", got, "proxy_Read")
+	}
+	if got := gjson.GetBytes(out, "tools.1.name").String(); got != "proxy_Write" {
+		t.Fatalf("tools.1.name = %q, want %q", got, "proxy_Write")
+	}
+	if got := gjson.GetBytes(out, "messages.0.content.0.name").String(); got != "proxy_Read" {
+		t.Fatalf("messages.0.content.0.name = %q, want %q", got, "proxy_Read")
+	}
+	if got := gjson.GetBytes(out, "messages.0.content.1.name").String(); got != "proxy_Write" {
+		t.Fatalf("messages.0.content.1.name = %q, want %q", got, "proxy_Write")
+	}
+}
+
+func TestApplyClaudeToolPrefix_ToolChoiceBuiltin(t *testing.T) {
+	body := []byte(`{
+		"tools": [
+			{"type": "web_search_20250305", "name": "web_search"},
+			{"name": "Read"}
+		],
+		"tool_choice": {"type": "tool", "name": "web_search"}
+	}`)
+	out := applyClaudeToolPrefix(body, "proxy_")
+
+	if got := gjson.GetBytes(out, "tool_choice.name").String(); got != "web_search" {
+		t.Fatalf("tool_choice.name = %q, want %q", got, "web_search")
+	}
+}
+
 func TestStripClaudeToolPrefixFromResponse(t *testing.T) {
 	input := []byte(`{"content":[{"type":"tool_use","name":"proxy_alpha","id":"t1","input":{}},{"type":"tool_use","name":"bravo","id":"t2","input":{}}]}`)
 	out := stripClaudeToolPrefixFromResponse(input, "proxy_")
@@ -49,6 +152,18 @@ func TestStripClaudeToolPrefixFromResponse(t *testing.T) {
 	}
 }

+func TestStripClaudeToolPrefixFromResponse_WithToolReference(t *testing.T) {
+	input := []byte(`{"content":[{"type":"tool_reference","tool_name":"proxy_alpha"},{"type":"tool_reference","tool_name":"bravo"}]}`)
+	out := stripClaudeToolPrefixFromResponse(input, "proxy_")
+
+	if got := gjson.GetBytes(out, "content.0.tool_name").String(); got != "alpha" {
+		t.Fatalf("content.0.tool_name = %q, want %q", got, "alpha")
+	}
+	if got := gjson.GetBytes(out, "content.1.tool_name").String(); got != "bravo" {
+		t.Fatalf("content.1.tool_name = %q, want %q", got, "bravo")
+	}
+}
+
 func TestStripClaudeToolPrefixFromStreamLine(t *testing.T) {
 	line := []byte(`data: {"type":"content_block_start","content_block":{"type":"tool_use","name":"proxy_alpha","id":"t1"},"index":0}`)
 	out := stripClaudeToolPrefixFromStreamLine(line, "proxy_")
@@ -61,3 +176,53 @@ func TestStripClaudeToolPrefixFromStreamLine(t *testing.T) {
 		t.Fatalf("content_block.name = %q, want %q", got, "alpha")
 	}
 }
+
+func TestStripClaudeToolPrefixFromStreamLine_WithToolReference(t *testing.T) {
+	line := []byte(`data: {"type":"content_block_start","content_block":{"type":"tool_reference","tool_name":"proxy_beta"},"index":0}`)
+	out := stripClaudeToolPrefixFromStreamLine(line, "proxy_")
+
+	payload := bytes.TrimSpace(out)
+	if bytes.HasPrefix(payload, []byte("data:")) {
+		payload = bytes.TrimSpace(payload[len("data:"):])
+	}
+	if got := gjson.GetBytes(payload, "content_block.tool_name").String(); got != "beta" {
+		t.Fatalf("content_block.tool_name = %q, want %q", got, "beta")
+	}
+}
+
+func TestApplyClaudeToolPrefix_NestedToolReference(t *testing.T) {
+	input := []byte(`{"messages":[{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_123","content":[{"type":"tool_reference","tool_name":"mcp__nia__manage_resource"}]}]}]}`)
+	out := applyClaudeToolPrefix(input, "proxy_")
+	got := gjson.GetBytes(out, "messages.0.content.0.content.0.tool_name").String()
+	if got != "proxy_mcp__nia__manage_resource" {
+		t.Fatalf("nested tool_reference tool_name = %q, want %q", got, "proxy_mcp__nia__manage_resource")
+	}
+}
+
+func TestStripClaudeToolPrefixFromResponse_NestedToolReference(t *testing.T) {
+	input := []byte(`{"content":[{"type":"tool_result","tool_use_id":"toolu_123","content":[{"type":"tool_reference","tool_name":"proxy_mcp__nia__manage_resource"}]}]}`)
+	out := stripClaudeToolPrefixFromResponse(input, "proxy_")
+	got := gjson.GetBytes(out, "content.0.content.0.tool_name").String()
+	if got != "mcp__nia__manage_resource" {
+		t.Fatalf("nested tool_reference tool_name = %q, want %q", got, "mcp__nia__manage_resource")
+	}
+}
+
+func TestApplyClaudeToolPrefix_NestedToolReferenceWithStringContent(t *testing.T) {
+	// tool_result.content can be a string - should not be processed
+	input := []byte(`{"messages":[{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_123","content":"plain string result"}]}]}`)
+	out := applyClaudeToolPrefix(input, "proxy_")
+	got := gjson.GetBytes(out, "messages.0.content.0.content").String()
+	if got != "plain string result" {
+		t.Fatalf("string content should remain unchanged = %q", got)
+	}
+}
+
+func TestApplyClaudeToolPrefix_SkipsBuiltinToolReference(t *testing.T) {
+	input := []byte(`{"tools":[{"type":"web_search_20250305","name":"web_search"}],"messages":[{"role":"user","content":[{"type":"tool_result","tool_use_id":"t1","content":[{"type":"tool_reference","tool_name":"web_search"}]}]}]}`)
+	out := applyClaudeToolPrefix(input, "proxy_")
+	got := gjson.GetBytes(out, "messages.0.content.0.content.0.tool_name").String()
+	if got != "web_search" {
+		t.Fatalf("built-in tool_reference should not be prefixed, got %q", got)
+	}
+}
--- a/internal/runtime/executor/codex_executor.go
+++ b/internal/runtime/executor/codex_executor.go
@@ -28,8 +28,8 @@ import (
 )

 const (
-	codexClientVersion = "0.98.0"
-	codexUserAgent     = "codex_cli_rs/0.98.0 (Mac OS 26.0.1; arm64) Apple_Terminal/464"
+	codexClientVersion = "0.101.0"
+	codexUserAgent     = "codex_cli_rs/0.101.0 (Mac OS 26.0.1; arm64) Apple_Terminal/464"
 )

 var dataTag = []byte("data:")
@@ -183,7 +183,7 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re

 		var param any
 		out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, originalPayload, body, line, &param)
-		resp = cliproxyexecutor.Response{Payload: []byte(out)}
+		resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 		return resp, nil
 	}
 	err = statusErr{code: 408, msg: "stream error: stream disconnected before completion: stream closed before response.completed"}
@@ -273,11 +273,11 @@ func (e *CodexExecutor) executeCompact(ctx context.Context, auth *cliproxyauth.A
 	reporter.ensurePublished(ctx)
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, originalPayload, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }

-func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusBadRequest, msg: "streaming not supported for /responses/compact"}
 	}
@@ -362,7 +362,6 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -397,7 +396,7 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }

 func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
@@ -643,7 +642,6 @@ func applyCodexHeaders(r *http.Request, auth *cliproxyauth.Auth, token string, s
 	}

 	misc.EnsureHeader(r.Header, ginHeaders, "Version", codexClientVersion)
-	misc.EnsureHeader(r.Header, ginHeaders, "Openai-Beta", "responses=experimental")
 	misc.EnsureHeader(r.Header, ginHeaders, "Session_id", uuid.NewString())
 	misc.EnsureHeader(r.Header, ginHeaders, "User-Agent", codexUserAgent)

--- a/internal/runtime/executor/codex_websockets_executor.go
+++ b/internal/runtime/executor/codex_websockets_executor.go
--- a/internal/runtime/executor/gemini_cli_executor.go
+++ b/internal/runtime/executor/gemini_cli_executor.go
@@ -225,7 +225,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 			reporter.publish(ctx, parseGeminiCLIUsage(data))
 			var param any
 			out := sdktranslator.TranslateNonStream(respCtx, to, from, attemptModel, opts.OriginalRequest, payload, data, &param)
-			resp = cliproxyexecutor.Response{Payload: []byte(out)}
+			resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 			return resp, nil
 		}

@@ -256,7 +256,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 }

 // ExecuteStream performs a streaming request to the Gemini CLI API.
-func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -382,7 +382,6 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 		}

 		out := make(chan cliproxyexecutor.StreamChunk)
-		stream = out
 		go func(resp *http.Response, reqBody []byte, attemptModel string) {
 			defer close(out)
 			defer func() {
@@ -441,7 +440,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 			}
 		}(httpResp, append([]byte(nil), payload...), attemptModel)

-		return stream, nil
+		return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 	}

 	if len(lastBody) > 0 {
@@ -546,7 +545,7 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 		if resp.StatusCode >= 200 && resp.StatusCode < 300 {
 			count := gjson.GetBytes(data, "totalTokens").Int()
 			translated := sdktranslator.TranslateTokenCount(respCtx, to, from, count, data)
-			return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
+			return cliproxyexecutor.Response{Payload: []byte(translated), Headers: resp.Header.Clone()}, nil
 		}
 		lastStatus = resp.StatusCode
 		lastBody = append([]byte(nil), data...)
@@ -899,8 +898,7 @@ func parseRetryDelay(errorBody []byte) (*time.Duration, error) {
 		if matches := re.FindStringSubmatch(message); len(matches) > 1 {
 			seconds, err := strconv.Atoi(matches[1])
 			if err == nil {
-				duration := time.Duration(seconds) * time.Second
-				return &duration, nil
+				return new(time.Duration(seconds) * time.Second), nil
 			}
 		}
 	}
--- a/internal/runtime/executor/gemini_executor.go
+++ b/internal/runtime/executor/gemini_executor.go
@@ -205,12 +205,12 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	reporter.publish(ctx, parseGeminiUsage(data))
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }

 // ExecuteStream performs a streaming request to the Gemini API.
-func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -298,7 +298,6 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -335,7 +334,7 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }

 // CountTokens counts tokens for the given request using the Gemini API.
@@ -416,7 +415,7 @@ func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut

 	count := gjson.GetBytes(data, "totalTokens").Int()
 	translated := sdktranslator.TranslateTokenCount(respCtx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(translated), Headers: resp.Header.Clone()}, nil
 }

 // Refresh refreshes the authentication credentials (no-op for Gemini API key).
--- a/internal/runtime/executor/gemini_vertex_executor.go
+++ b/internal/runtime/executor/gemini_vertex_executor.go
@@ -253,7 +253,7 @@ func (e *GeminiVertexExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 }

 // ExecuteStream performs a streaming request to the Vertex AI API.
-func (e *GeminiVertexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiVertexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -419,7 +419,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	to := sdktranslator.FromString("gemini")
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }

@@ -524,12 +524,12 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	reporter.publish(ctx, parseGeminiUsage(data))
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }

 // executeStreamWithServiceAccount handles streaming authentication using service account credentials.
-func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (_ *cliproxyexecutor.StreamResult, err error) {
 	baseModel := thinking.ParseSuffix(req.Model).ModelName

 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
@@ -618,7 +618,6 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	}

 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -650,11 +649,11 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }

 // executeStreamWithAPIKey handles streaming authentication using API key credentials.
-func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (_ *cliproxyexecutor.StreamResult, err error) {
 	baseModel := thinking.ParseSuffix(req.Model).ModelName

 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
@@ -743,7 +742,6 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	}

 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -775,7 +773,7 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }

 // countTokensWithServiceAccount counts tokens using service account credentials.
@@ -859,7 +857,7 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}, nil
 }

 // countTokensWithAPIKey handles token counting using API key credentials.
@@ -943,7 +941,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}, nil
 }

 // vertexCreds extracts project, location and raw service account JSON from auth metadata.
--- a/internal/runtime/executor/github_copilot_executor.go
+++ b/internal/runtime/executor/github_copilot_executor.go
@@ -39,7 +39,8 @@ const (
 	copilotEditorVersion = "vscode/1.107.0"
 	copilotPluginVersion = "copilot-chat/0.35.0"
 	copilotIntegrationID = "vscode-chat"
-	copilotOpenAIIntent  = "conversation-edits"
+	copilotOpenAIIntent  = "conversation-panel"
+	copilotGitHubAPIVer  = "2025-04-01"
 )

 // GitHubCopilotExecutor handles requests to the GitHub Copilot API.
@@ -51,8 +52,9 @@ type GitHubCopilotExecutor struct {

 // cachedAPIToken stores a cached Copilot API token with its expiry.
 type cachedAPIToken struct {
-	token     string
-	expiresAt time.Time
+	token       string
+	apiEndpoint string
+	expiresAt   time.Time
 }

 // NewGitHubCopilotExecutor constructs a new executor instance.
@@ -75,7 +77,7 @@ func (e *GitHubCopilotExecutor) PrepareRequest(req *http.Request, auth *cliproxy
 	if ctx == nil {
 		ctx = context.Background()
 	}
-	apiToken, errToken := e.ensureAPIToken(ctx, auth)
+	apiToken, _, errToken := e.ensureAPIToken(ctx, auth)
 	if errToken != nil {
 		return errToken
 	}
@@ -101,7 +103,7 @@ func (e *GitHubCopilotExecutor) HttpRequest(ctx context.Context, auth *cliproxya

 // Execute handles non-streaming requests to GitHub Copilot.
 func (e *GitHubCopilotExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
-	apiToken, errToken := e.ensureAPIToken(ctx, auth)
+	apiToken, baseURL, errToken := e.ensureAPIToken(ctx, auth)
 	if errToken != nil {
 		return resp, errToken
 	}
@@ -110,7 +112,7 @@ func (e *GitHubCopilotExecutor) Execute(ctx context.Context, auth *cliproxyauth.
 	defer reporter.trackFailure(ctx, &err)

 	from := opts.SourceFormat
-	useResponses := useGitHubCopilotResponsesEndpoint(from)
+	useResponses := useGitHubCopilotResponsesEndpoint(from, req.Model)
 	to := sdktranslator.FromString("openai")
 	if useResponses {
 		to = sdktranslator.FromString("openai-response")
@@ -123,6 +125,25 @@ func (e *GitHubCopilotExecutor) Execute(ctx context.Context, auth *cliproxyauth.
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
 	body = e.normalizeModel(req.Model, body)
 	body = flattenAssistantContent(body)
+
+	// Detect vision content before input normalization removes messages
+	hasVision := detectVisionContent(body)
+
+	thinkingProvider := "openai"
+	if useResponses {
+		thinkingProvider = "codex"
+	}
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), thinkingProvider, e.Identifier())
+	if err != nil {
+		return resp, err
+	}
+
+	if useResponses {
+		body = normalizeGitHubCopilotResponsesInput(body)
+		body = normalizeGitHubCopilotResponsesTools(body)
+	} else {
+		body = normalizeGitHubCopilotChatTools(body)
+	}
 	requestedModel := payloadRequestedModel(opts, req.Model)
 	body = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", body, originalTranslated, requestedModel)
 	body, _ = sjson.SetBytes(body, "stream", false)
@@ -131,7 +152,7 @@ func (e *GitHubCopilotExecutor) Execute(ctx context.Context, auth *cliproxyauth.
 	if useResponses {
 		path = githubCopilotResponsesPath
 	}
-	url := githubCopilotBaseURL + path
+	url := baseURL + path
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
 	if err != nil {
 		return resp, err
@@ -139,7 +160,7 @@ func (e *GitHubCopilotExecutor) Execute(ctx context.Context, auth *cliproxyauth.
 	e.applyHeaders(httpReq, apiToken, body)

 	// Add Copilot-Vision-Request header if the request contains vision content
-	if detectVisionContent(body) {
+	if hasVision {
 		httpReq.Header.Set("Copilot-Vision-Request", "true")
 	}

@@ -199,15 +220,20 @@ func (e *GitHubCopilotExecutor) Execute(ctx context.Context, auth *cliproxyauth.
 	}

 	var param any
-	converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
+	converted := ""
+	if useResponses && from.String() == "claude" {
+		converted = translateGitHubCopilotResponsesNonStreamToClaude(data)
+	} else {
+		converted = sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
+	}
 	resp = cliproxyexecutor.Response{Payload: []byte(converted)}
 	reporter.ensurePublished(ctx)
 	return resp, nil
 }

 // ExecuteStream handles streaming requests to GitHub Copilot.
-func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
-	apiToken, errToken := e.ensureAPIToken(ctx, auth)
+func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
+	apiToken, baseURL, errToken := e.ensureAPIToken(ctx, auth)
 	if errToken != nil {
 		return nil, errToken
 	}
@@ -216,7 +242,7 @@ func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliprox
 	defer reporter.trackFailure(ctx, &err)

 	from := opts.SourceFormat
-	useResponses := useGitHubCopilotResponsesEndpoint(from)
+	useResponses := useGitHubCopilotResponsesEndpoint(from, req.Model)
 	to := sdktranslator.FromString("openai")
 	if useResponses {
 		to = sdktranslator.FromString("openai-response")
@@ -229,6 +255,25 @@ func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliprox
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 	body = e.normalizeModel(req.Model, body)
 	body = flattenAssistantContent(body)
+
+	// Detect vision content before input normalization removes messages
+	hasVision := detectVisionContent(body)
+
+	thinkingProvider := "openai"
+	if useResponses {
+		thinkingProvider = "codex"
+	}
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), thinkingProvider, e.Identifier())
+	if err != nil {
+		return nil, err
+	}
+
+	if useResponses {
+		body = normalizeGitHubCopilotResponsesInput(body)
+		body = normalizeGitHubCopilotResponsesTools(body)
+	} else {
+		body = normalizeGitHubCopilotChatTools(body)
+	}
 	requestedModel := payloadRequestedModel(opts, req.Model)
 	body = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", body, originalTranslated, requestedModel)
 	body, _ = sjson.SetBytes(body, "stream", true)
@@ -241,7 +286,7 @@ func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliprox
 	if useResponses {
 		path = githubCopilotResponsesPath
 	}
-	url := githubCopilotBaseURL + path
+	url := baseURL + path
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
 	if err != nil {
 		return nil, err
@@ -249,7 +294,7 @@ func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliprox
 	e.applyHeaders(httpReq, apiToken, body)

 	// Add Copilot-Vision-Request header if the request contains vision content
-	if detectVisionContent(body) {
+	if hasVision {
 		httpReq.Header.Set("Copilot-Vision-Request", "true")
 	}

@@ -296,7 +341,6 @@ func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliprox
 	}

 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out

 	go func() {
 		defer close(out)
@@ -329,7 +373,12 @@ func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliprox
 				}
 			}

-			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(line), &param)
+			var chunks []string
+			if useResponses && from.String() == "claude" {
+				chunks = translateGitHubCopilotResponsesStreamToClaude(bytes.Clone(line), &param)
+			} else {
+				chunks = sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(line), &param)
+			}
 			for i := range chunks {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
 			}
@@ -344,7 +393,10 @@ func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliprox
 		}
 	}()

-	return stream, nil
+	return &cliproxyexecutor.StreamResult{
+		Headers: httpResp.Header.Clone(),
+		Chunks:  out,
+	}, nil
 }

 // CountTokens is not supported for GitHub Copilot.
@@ -376,22 +428,22 @@ func (e *GitHubCopilotExecutor) Refresh(ctx context.Context, auth *cliproxyauth.
 }

 // ensureAPIToken gets or refreshes the Copilot API token.
-func (e *GitHubCopilotExecutor) ensureAPIToken(ctx context.Context, auth *cliproxyauth.Auth) (string, error) {
+func (e *GitHubCopilotExecutor) ensureAPIToken(ctx context.Context, auth *cliproxyauth.Auth) (string, string, error) {
 	if auth == nil {
-		return "", statusErr{code: http.StatusUnauthorized, msg: "missing auth"}
+		return "", "", statusErr{code: http.StatusUnauthorized, msg: "missing auth"}
 	}

 	// Get the GitHub access token
 	accessToken := metaStringValue(auth.Metadata, "access_token")
 	if accessToken == "" {
-		return "", statusErr{code: http.StatusUnauthorized, msg: "missing github access token"}
+		return "", "", statusErr{code: http.StatusUnauthorized, msg: "missing github access token"}
 	}

 	// Check for cached API token using thread-safe access
 	e.mu.RLock()
 	if cached, ok := e.cache[accessToken]; ok && cached.expiresAt.After(time.Now().Add(tokenExpiryBuffer)) {
 		e.mu.RUnlock()
-		return cached.token, nil
+		return cached.token, cached.apiEndpoint, nil
 	}
 	e.mu.RUnlock()

@@ -399,7 +451,13 @@ func (e *GitHubCopilotExecutor) ensureAPIToken(ctx context.Context, auth *clipro
 	copilotAuth := copilotauth.NewCopilotAuth(e.cfg)
 	apiToken, err := copilotAuth.GetCopilotAPIToken(ctx, accessToken)
 	if err != nil {
-		return "", statusErr{code: http.StatusUnauthorized, msg: fmt.Sprintf("failed to get copilot api token: %v", err)}
+		return "", "", statusErr{code: http.StatusUnauthorized, msg: fmt.Sprintf("failed to get copilot api token: %v", err)}
+	}
+
+	// Use endpoint from token response, fall back to default
+	apiEndpoint := githubCopilotBaseURL
+	if apiToken.Endpoints.API != "" {
+		apiEndpoint = strings.TrimRight(apiToken.Endpoints.API, "/")
 	}

 	// Cache the token with thread-safe access
@@ -409,12 +467,13 @@ func (e *GitHubCopilotExecutor) ensureAPIToken(ctx context.Context, auth *clipro
 	}
 	e.mu.Lock()
 	e.cache[accessToken] = &cachedAPIToken{
-		token:     apiToken.Token,
-		expiresAt: expiresAt,
+		token:       apiToken.Token,
+		apiEndpoint: apiEndpoint,
+		expiresAt:   expiresAt,
 	}
 	e.mu.Unlock()

-	return apiToken.Token, nil
+	return apiToken.Token, apiEndpoint, nil
 }

 // applyHeaders sets the required headers for GitHub Copilot API requests.
@@ -427,16 +486,17 @@ func (e *GitHubCopilotExecutor) applyHeaders(r *http.Request, apiToken string, b
 	r.Header.Set("Editor-Plugin-Version", copilotPluginVersion)
 	r.Header.Set("Openai-Intent", copilotOpenAIIntent)
 	r.Header.Set("Copilot-Integration-Id", copilotIntegrationID)
+	r.Header.Set("X-Github-Api-Version", copilotGitHubAPIVer)
 	r.Header.Set("X-Request-Id", uuid.NewString())

 	initiator := "user"
 	if len(body) > 0 {
 		if messages := gjson.GetBytes(body, "messages"); messages.Exists() && messages.IsArray() {
-			arr := messages.Array()
-			if len(arr) > 0 {
-				lastRole := arr[len(arr)-1].Get("role").String()
-				if lastRole != "" && lastRole != "user" {
+			for _, msg := range messages.Array() {
+				role := msg.Get("role").String()
+				if role == "assistant" || role == "tool" {
 					initiator = "agent"
+					break
 				}
 			}
 		}
@@ -483,8 +543,12 @@ func (e *GitHubCopilotExecutor) normalizeModel(model string, body []byte) []byte
 	return body
 }

-func useGitHubCopilotResponsesEndpoint(sourceFormat sdktranslator.Format) bool {
-	return sourceFormat.String() == "openai-response"
+func useGitHubCopilotResponsesEndpoint(sourceFormat sdktranslator.Format, model string) bool {
+	if sourceFormat.String() == "openai-response" {
+		return true
+	}
+	baseModel := strings.ToLower(thinking.ParseSuffix(model).ModelName)
+	return strings.Contains(baseModel, "codex")
 }

 // flattenAssistantContent converts assistant message content from array format
@@ -504,6 +568,17 @@ func flattenAssistantContent(body []byte) []byte {
 		if !content.Exists() || !content.IsArray() {
 			continue
 		}
+		// Skip flattening if the content contains non-text blocks (tool_use, thinking, etc.)
+		hasNonText := false
+		for _, part := range content.Array() {
+			if t := part.Get("type").String(); t != "" && t != "text" {
+				hasNonText = true
+				break
+			}
+		}
+		if hasNonText {
+			continue
+		}
 		var textParts []string
 		for _, part := range content.Array() {
 			if part.Get("type").String() == "text" {
@@ -519,6 +594,644 @@ func flattenAssistantContent(body []byte) []byte {
 	return result
 }

+func normalizeGitHubCopilotChatTools(body []byte) []byte {
+	tools := gjson.GetBytes(body, "tools")
+	if tools.Exists() {
+		filtered := "[]"
+		if tools.IsArray() {
+			for _, tool := range tools.Array() {
+				if tool.Get("type").String() != "function" {
+					continue
+				}
+				filtered, _ = sjson.SetRaw(filtered, "-1", tool.Raw)
+			}
+		}
+		body, _ = sjson.SetRawBytes(body, "tools", []byte(filtered))
+	}
+
+	toolChoice := gjson.GetBytes(body, "tool_choice")
+	if !toolChoice.Exists() {
+		return body
+	}
+	if toolChoice.Type == gjson.String {
+		switch toolChoice.String() {
+		case "auto", "none", "required":
+			return body
+		}
+	}
+	body, _ = sjson.SetBytes(body, "tool_choice", "auto")
+	return body
+}
+
+func normalizeGitHubCopilotResponsesInput(body []byte) []byte {
+	input := gjson.GetBytes(body, "input")
+	if input.Exists() {
+		// If input is already a string or array, keep it as-is.
+		if input.Type == gjson.String || input.IsArray() {
+			return body
+		}
+		// Non-string/non-array input: stringify as fallback.
+		body, _ = sjson.SetBytes(body, "input", input.Raw)
+		return body
+	}
+
+	// Convert Claude messages format to OpenAI Responses API input array.
+	// This preserves the conversation structure (roles, tool calls, tool results)
+	// which is critical for multi-turn tool-use conversations.
+	inputArr := "[]"
+
+	// System messages → developer role
+	if system := gjson.GetBytes(body, "system"); system.Exists() {
+		var systemParts []string
+		if system.IsArray() {
+			for _, part := range system.Array() {
+				if txt := part.Get("text").String(); txt != "" {
+					systemParts = append(systemParts, txt)
+				}
+			}
+		} else if system.Type == gjson.String {
+			systemParts = append(systemParts, system.String())
+		}
+		if len(systemParts) > 0 {
+			msg := `{"type":"message","role":"developer","content":[]}`
+			for _, txt := range systemParts {
+				part := `{"type":"input_text","text":""}`
+				part, _ = sjson.Set(part, "text", txt)
+				msg, _ = sjson.SetRaw(msg, "content.-1", part)
+			}
+			inputArr, _ = sjson.SetRaw(inputArr, "-1", msg)
+		}
+	}
+
+	// Messages → structured input items
+	if messages := gjson.GetBytes(body, "messages"); messages.Exists() && messages.IsArray() {
+		for _, msg := range messages.Array() {
+			role := msg.Get("role").String()
+			content := msg.Get("content")
+
+			if !content.Exists() {
+				continue
+			}
+
+			// Simple string content
+			if content.Type == gjson.String {
+				textType := "input_text"
+				if role == "assistant" {
+					textType = "output_text"
+				}
+				item := `{"type":"message","role":"","content":[]}`
+				item, _ = sjson.Set(item, "role", role)
+				part := fmt.Sprintf(`{"type":"%s","text":""}`, textType)
+				part, _ = sjson.Set(part, "text", content.String())
+				item, _ = sjson.SetRaw(item, "content.-1", part)
+				inputArr, _ = sjson.SetRaw(inputArr, "-1", item)
+				continue
+			}
+
+			if !content.IsArray() {
+				continue
+			}
+
+			// Array content: split into message parts vs tool items
+			var msgParts []string
+			for _, c := range content.Array() {
+				cType := c.Get("type").String()
+				switch cType {
+				case "text":
+					textType := "input_text"
+					if role == "assistant" {
+						textType = "output_text"
+					}
+					part := fmt.Sprintf(`{"type":"%s","text":""}`, textType)
+					part, _ = sjson.Set(part, "text", c.Get("text").String())
+					msgParts = append(msgParts, part)
+				case "image":
+					source := c.Get("source")
+					if source.Exists() {
+						data := source.Get("data").String()
+						if data == "" {
+							data = source.Get("base64").String()
+						}
+						mediaType := source.Get("media_type").String()
+						if mediaType == "" {
+							mediaType = source.Get("mime_type").String()
+						}
+						if mediaType == "" {
+							mediaType = "application/octet-stream"
+						}
+						if data != "" {
+							part := `{"type":"input_image","image_url":""}`
+							part, _ = sjson.Set(part, "image_url", fmt.Sprintf("data:%s;base64,%s", mediaType, data))
+							msgParts = append(msgParts, part)
+						}
+					}
+				case "tool_use":
+					// Flush any accumulated message parts first
+					if len(msgParts) > 0 {
+						item := `{"type":"message","role":"","content":[]}`
+						item, _ = sjson.Set(item, "role", role)
+						for _, p := range msgParts {
+							item, _ = sjson.SetRaw(item, "content.-1", p)
+						}
+						inputArr, _ = sjson.SetRaw(inputArr, "-1", item)
+						msgParts = nil
+					}
+					fc := `{"type":"function_call","call_id":"","name":"","arguments":""}`
+					fc, _ = sjson.Set(fc, "call_id", c.Get("id").String())
+					fc, _ = sjson.Set(fc, "name", c.Get("name").String())
+					if inputRaw := c.Get("input"); inputRaw.Exists() {
+						fc, _ = sjson.Set(fc, "arguments", inputRaw.Raw)
+					}
+					inputArr, _ = sjson.SetRaw(inputArr, "-1", fc)
+				case "tool_result":
+					// Flush any accumulated message parts first
+					if len(msgParts) > 0 {
+						item := `{"type":"message","role":"","content":[]}`
+						item, _ = sjson.Set(item, "role", role)
+						for _, p := range msgParts {
+							item, _ = sjson.SetRaw(item, "content.-1", p)
+						}
+						inputArr, _ = sjson.SetRaw(inputArr, "-1", item)
+						msgParts = nil
+					}
+					fco := `{"type":"function_call_output","call_id":"","output":""}`
+					fco, _ = sjson.Set(fco, "call_id", c.Get("tool_use_id").String())
+					// Extract output text
+					resultContent := c.Get("content")
+					if resultContent.Type == gjson.String {
+						fco, _ = sjson.Set(fco, "output", resultContent.String())
+					} else if resultContent.IsArray() {
+						var resultParts []string
+						for _, rc := range resultContent.Array() {
+							if txt := rc.Get("text").String(); txt != "" {
+								resultParts = append(resultParts, txt)
+							}
+						}
+						fco, _ = sjson.Set(fco, "output", strings.Join(resultParts, "\n"))
+					} else if resultContent.Exists() {
+						fco, _ = sjson.Set(fco, "output", resultContent.String())
+					}
+					inputArr, _ = sjson.SetRaw(inputArr, "-1", fco)
+				case "thinking":
+					// Skip thinking blocks - not part of the API input
+				}
+			}
+
+			// Flush remaining message parts
+			if len(msgParts) > 0 {
+				item := `{"type":"message","role":"","content":[]}`
+				item, _ = sjson.Set(item, "role", role)
+				for _, p := range msgParts {
+					item, _ = sjson.SetRaw(item, "content.-1", p)
+				}
+				inputArr, _ = sjson.SetRaw(inputArr, "-1", item)
+			}
+		}
+	}
+
+	body, _ = sjson.SetRawBytes(body, "input", []byte(inputArr))
+	// Remove messages/system since we've converted them to input
+	body, _ = sjson.DeleteBytes(body, "messages")
+	body, _ = sjson.DeleteBytes(body, "system")
+	return body
+}
+
+func normalizeGitHubCopilotResponsesTools(body []byte) []byte {
+	tools := gjson.GetBytes(body, "tools")
+	if tools.Exists() {
+		filtered := "[]"
+		if tools.IsArray() {
+			for _, tool := range tools.Array() {
+				toolType := tool.Get("type").String()
+				// Accept OpenAI format (type="function") and Claude format
+				// (no type field, but has top-level name + input_schema).
+				if toolType != "" && toolType != "function" {
+					continue
+				}
+				name := tool.Get("name").String()
+				if name == "" {
+					name = tool.Get("function.name").String()
+				}
+				if name == "" {
+					continue
+				}
+				normalized := `{"type":"function","name":""}`
+				normalized, _ = sjson.Set(normalized, "name", name)
+				if desc := tool.Get("description").String(); desc != "" {
+					normalized, _ = sjson.Set(normalized, "description", desc)
+				} else if desc = tool.Get("function.description").String(); desc != "" {
+					normalized, _ = sjson.Set(normalized, "description", desc)
+				}
+				if params := tool.Get("parameters"); params.Exists() {
+					normalized, _ = sjson.SetRaw(normalized, "parameters", params.Raw)
+				} else if params = tool.Get("function.parameters"); params.Exists() {
+					normalized, _ = sjson.SetRaw(normalized, "parameters", params.Raw)
+				} else if params = tool.Get("input_schema"); params.Exists() {
+					normalized, _ = sjson.SetRaw(normalized, "parameters", params.Raw)
+				}
+				filtered, _ = sjson.SetRaw(filtered, "-1", normalized)
+			}
+		}
+		body, _ = sjson.SetRawBytes(body, "tools", []byte(filtered))
+	}
+
+	toolChoice := gjson.GetBytes(body, "tool_choice")
+	if !toolChoice.Exists() {
+		return body
+	}
+	if toolChoice.Type == gjson.String {
+		switch toolChoice.String() {
+		case "auto", "none", "required":
+			return body
+		default:
+			body, _ = sjson.SetBytes(body, "tool_choice", "auto")
+			return body
+		}
+	}
+	if toolChoice.Type == gjson.JSON {
+		choiceType := toolChoice.Get("type").String()
+		if choiceType == "function" {
+			name := toolChoice.Get("name").String()
+			if name == "" {
+				name = toolChoice.Get("function.name").String()
+			}
+			if name != "" {
+				normalized := `{"type":"function","name":""}`
+				normalized, _ = sjson.Set(normalized, "name", name)
+				body, _ = sjson.SetRawBytes(body, "tool_choice", []byte(normalized))
+				return body
+			}
+		}
+	}
+	body, _ = sjson.SetBytes(body, "tool_choice", "auto")
+	return body
+}
+
+func collectTextFromNode(node gjson.Result) string {
+	if !node.Exists() {
+		return ""
+	}
+	if node.Type == gjson.String {
+		return node.String()
+	}
+	if node.IsArray() {
+		var parts []string
+		for _, item := range node.Array() {
+			if item.Type == gjson.String {
+				if text := item.String(); text != "" {
+					parts = append(parts, text)
+				}
+				continue
+			}
+			if text := item.Get("text").String(); text != "" {
+				parts = append(parts, text)
+				continue
+			}
+			if nested := collectTextFromNode(item.Get("content")); nested != "" {
+				parts = append(parts, nested)
+			}
+		}
+		return strings.Join(parts, "\n")
+	}
+	if node.Type == gjson.JSON {
+		if text := node.Get("text").String(); text != "" {
+			return text
+		}
+		if nested := collectTextFromNode(node.Get("content")); nested != "" {
+			return nested
+		}
+		return node.Raw
+	}
+	return node.String()
+}
+
+type githubCopilotResponsesStreamToolState struct {
+	Index int
+	ID    string
+	Name  string
+}
+
+type githubCopilotResponsesStreamState struct {
+	MessageStarted    bool
+	MessageStopSent   bool
+	TextBlockStarted  bool
+	TextBlockIndex    int
+	NextContentIndex  int
+	HasToolUse        bool
+	ReasoningActive   bool
+	ReasoningIndex    int
+	OutputIndexToTool map[int]*githubCopilotResponsesStreamToolState
+	ItemIDToTool      map[string]*githubCopilotResponsesStreamToolState
+}
+
+func translateGitHubCopilotResponsesNonStreamToClaude(data []byte) string {
+	root := gjson.ParseBytes(data)
+	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())
+
+	hasToolUse := false
+	if output := root.Get("output"); output.Exists() && output.IsArray() {
+		for _, item := range output.Array() {
+			switch item.Get("type").String() {
+			case "reasoning":
+				var thinkingText string
+				if summary := item.Get("summary"); summary.Exists() && summary.IsArray() {
+					var parts []string
+					for _, part := range summary.Array() {
+						if txt := part.Get("text").String(); txt != "" {
+							parts = append(parts, txt)
+						}
+					}
+					thinkingText = strings.Join(parts, "")
+				}
+				if thinkingText == "" {
+					if content := item.Get("content"); content.Exists() && content.IsArray() {
+						var parts []string
+						for _, part := range content.Array() {
+							if txt := part.Get("text").String(); txt != "" {
+								parts = append(parts, txt)
+							}
+						}
+						thinkingText = strings.Join(parts, "")
+					}
+				}
+				if thinkingText != "" {
+					block := `{"type":"thinking","thinking":""}`
+					block, _ = sjson.Set(block, "thinking", thinkingText)
+					out, _ = sjson.SetRaw(out, "content.-1", block)
+				}
+			case "message":
+				if content := item.Get("content"); content.Exists() && content.IsArray() {
+					for _, part := range content.Array() {
+						if part.Get("type").String() != "output_text" {
+							continue
+						}
+						text := part.Get("text").String()
+						if text == "" {
+							continue
+						}
+						block := `{"type":"text","text":""}`
+						block, _ = sjson.Set(block, "text", text)
+						out, _ = sjson.SetRaw(out, "content.-1", block)
+					}
+				}
+			case "function_call":
+				hasToolUse = true
+				toolUse := `{"type":"tool_use","id":"","name":"","input":{}}`
+				toolID := item.Get("call_id").String()
+				if toolID == "" {
+					toolID = item.Get("id").String()
+				}
+				toolUse, _ = sjson.Set(toolUse, "id", toolID)
+				toolUse, _ = sjson.Set(toolUse, "name", item.Get("name").String())
+				if args := item.Get("arguments").String(); args != "" && gjson.Valid(args) {
+					argObj := gjson.Parse(args)
+					if argObj.IsObject() {
+						toolUse, _ = sjson.SetRaw(toolUse, "input", argObj.Raw)
+					}
+				}
+				out, _ = sjson.SetRaw(out, "content.-1", toolUse)
+			}
+		}
+	}
+
+	inputTokens := root.Get("usage.input_tokens").Int()
+	outputTokens := root.Get("usage.output_tokens").Int()
+	cachedTokens := root.Get("usage.input_tokens_details.cached_tokens").Int()
+	if cachedTokens > 0 && inputTokens >= cachedTokens {
+		inputTokens -= cachedTokens
+	}
+	out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+	out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
+	if cachedTokens > 0 {
+		out, _ = sjson.Set(out, "usage.cache_read_input_tokens", cachedTokens)
+	}
+	if hasToolUse {
+		out, _ = sjson.Set(out, "stop_reason", "tool_use")
+	} else if sr := root.Get("stop_reason").String(); sr == "max_tokens" || sr == "stop" {
+		out, _ = sjson.Set(out, "stop_reason", sr)
+	} else {
+		out, _ = sjson.Set(out, "stop_reason", "end_turn")
+	}
+	return out
+}
+
+func translateGitHubCopilotResponsesStreamToClaude(line []byte, param *any) []string {
+	if *param == nil {
+		*param = &githubCopilotResponsesStreamState{
+			TextBlockIndex:    -1,
+			OutputIndexToTool: make(map[int]*githubCopilotResponsesStreamToolState),
+			ItemIDToTool:      make(map[string]*githubCopilotResponsesStreamToolState),
+		}
+	}
+	state := (*param).(*githubCopilotResponsesStreamState)
+
+	if !bytes.HasPrefix(line, dataTag) {
+		return nil
+	}
+	payload := bytes.TrimSpace(line[5:])
+	if bytes.Equal(payload, []byte("[DONE]")) {
+		return nil
+	}
+	if !gjson.ValidBytes(payload) {
+		return nil
+	}
+
+	event := gjson.GetBytes(payload, "type").String()
+	results := make([]string, 0, 4)
+	ensureMessageStart := func() {
+		if state.MessageStarted {
+			return
+		}
+		messageStart := `{"type":"message_start","message":{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}}`
+		messageStart, _ = sjson.Set(messageStart, "message.id", gjson.GetBytes(payload, "response.id").String())
+		messageStart, _ = sjson.Set(messageStart, "message.model", gjson.GetBytes(payload, "response.model").String())
+		results = append(results, "event: message_start\ndata: "+messageStart+"\n\n")
+		state.MessageStarted = true
+	}
+	startTextBlockIfNeeded := func() {
+		if state.TextBlockStarted {
+			return
+		}
+		if state.TextBlockIndex < 0 {
+			state.TextBlockIndex = state.NextContentIndex
+			state.NextContentIndex++
+		}
+		contentBlockStart := `{"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}`
+		contentBlockStart, _ = sjson.Set(contentBlockStart, "index", state.TextBlockIndex)
+		results = append(results, "event: content_block_start\ndata: "+contentBlockStart+"\n\n")
+		state.TextBlockStarted = true
+	}
+	stopTextBlockIfNeeded := func() {
+		if !state.TextBlockStarted {
+			return
+		}
+		contentBlockStop := `{"type":"content_block_stop","index":0}`
+		contentBlockStop, _ = sjson.Set(contentBlockStop, "index", state.TextBlockIndex)
+		results = append(results, "event: content_block_stop\ndata: "+contentBlockStop+"\n\n")
+		state.TextBlockStarted = false
+		state.TextBlockIndex = -1
+	}
+	resolveTool := func(itemID string, outputIndex int) *githubCopilotResponsesStreamToolState {
+		if itemID != "" {
+			if tool, ok := state.ItemIDToTool[itemID]; ok {
+				return tool
+			}
+		}
+		if tool, ok := state.OutputIndexToTool[outputIndex]; ok {
+			if itemID != "" {
+				state.ItemIDToTool[itemID] = tool
+			}
+			return tool
+		}
+		return nil
+	}
+
+	switch event {
+	case "response.created":
+		ensureMessageStart()
+	case "response.output_text.delta":
+		ensureMessageStart()
+		startTextBlockIfNeeded()
+		delta := gjson.GetBytes(payload, "delta").String()
+		if delta != "" {
+			contentDelta := `{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":""}}`
+			contentDelta, _ = sjson.Set(contentDelta, "index", state.TextBlockIndex)
+			contentDelta, _ = sjson.Set(contentDelta, "delta.text", delta)
+			results = append(results, "event: content_block_delta\ndata: "+contentDelta+"\n\n")
+		}
+	case "response.reasoning_summary_part.added":
+		ensureMessageStart()
+		state.ReasoningActive = true
+		state.ReasoningIndex = state.NextContentIndex
+		state.NextContentIndex++
+		thinkingStart := `{"type":"content_block_start","index":0,"content_block":{"type":"thinking","thinking":""}}`
+		thinkingStart, _ = sjson.Set(thinkingStart, "index", state.ReasoningIndex)
+		results = append(results, "event: content_block_start\ndata: "+thinkingStart+"\n\n")
+	case "response.reasoning_summary_text.delta":
+		if state.ReasoningActive {
+			delta := gjson.GetBytes(payload, "delta").String()
+			if delta != "" {
+				thinkingDelta := `{"type":"content_block_delta","index":0,"delta":{"type":"thinking_delta","thinking":""}}`
+				thinkingDelta, _ = sjson.Set(thinkingDelta, "index", state.ReasoningIndex)
+				thinkingDelta, _ = sjson.Set(thinkingDelta, "delta.thinking", delta)
+				results = append(results, "event: content_block_delta\ndata: "+thinkingDelta+"\n\n")
+			}
+		}
+	case "response.reasoning_summary_part.done":
+		if state.ReasoningActive {
+			thinkingStop := `{"type":"content_block_stop","index":0}`
+			thinkingStop, _ = sjson.Set(thinkingStop, "index", state.ReasoningIndex)
+			results = append(results, "event: content_block_stop\ndata: "+thinkingStop+"\n\n")
+			state.ReasoningActive = false
+		}
+	case "response.output_item.added":
+		if gjson.GetBytes(payload, "item.type").String() != "function_call" {
+			break
+		}
+		ensureMessageStart()
+		stopTextBlockIfNeeded()
+		state.HasToolUse = true
+		tool := &githubCopilotResponsesStreamToolState{
+			Index: state.NextContentIndex,
+			ID:    gjson.GetBytes(payload, "item.call_id").String(),
+			Name:  gjson.GetBytes(payload, "item.name").String(),
+		}
+		if tool.ID == "" {
+			tool.ID = gjson.GetBytes(payload, "item.id").String()
+		}
+		state.NextContentIndex++
+		outputIndex := int(gjson.GetBytes(payload, "output_index").Int())
+		state.OutputIndexToTool[outputIndex] = tool
+		if itemID := gjson.GetBytes(payload, "item.id").String(); itemID != "" {
+			state.ItemIDToTool[itemID] = tool
+		}
+		contentBlockStart := `{"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`
+		contentBlockStart, _ = sjson.Set(contentBlockStart, "index", tool.Index)
+		contentBlockStart, _ = sjson.Set(contentBlockStart, "content_block.id", tool.ID)
+		contentBlockStart, _ = sjson.Set(contentBlockStart, "content_block.name", tool.Name)
+		results = append(results, "event: content_block_start\ndata: "+contentBlockStart+"\n\n")
+	case "response.output_item.delta":
+		item := gjson.GetBytes(payload, "item")
+		if item.Get("type").String() != "function_call" {
+			break
+		}
+		tool := resolveTool(item.Get("id").String(), int(gjson.GetBytes(payload, "output_index").Int()))
+		if tool == nil {
+			break
+		}
+		partial := gjson.GetBytes(payload, "delta").String()
+		if partial == "" {
+			partial = item.Get("arguments").String()
+		}
+		if partial == "" {
+			break
+		}
+		inputDelta := `{"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""}}`
+		inputDelta, _ = sjson.Set(inputDelta, "index", tool.Index)
+		inputDelta, _ = sjson.Set(inputDelta, "delta.partial_json", partial)
+		results = append(results, "event: content_block_delta\ndata: "+inputDelta+"\n\n")
+	case "response.function_call_arguments.delta":
+		// Copilot sends tool call arguments via this event type (not response.output_item.delta).
+		// Data format: {"delta":"...", "item_id":"...", "output_index":N, ...}
+		itemID := gjson.GetBytes(payload, "item_id").String()
+		outputIndex := int(gjson.GetBytes(payload, "output_index").Int())
+		tool := resolveTool(itemID, outputIndex)
+		if tool == nil {
+			break
+		}
+		partial := gjson.GetBytes(payload, "delta").String()
+		if partial == "" {
+			break
+		}
+		inputDelta := `{"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""}}`
+		inputDelta, _ = sjson.Set(inputDelta, "index", tool.Index)
+		inputDelta, _ = sjson.Set(inputDelta, "delta.partial_json", partial)
+		results = append(results, "event: content_block_delta\ndata: "+inputDelta+"\n\n")
+	case "response.output_item.done":
+		if gjson.GetBytes(payload, "item.type").String() != "function_call" {
+			break
+		}
+		tool := resolveTool(gjson.GetBytes(payload, "item.id").String(), int(gjson.GetBytes(payload, "output_index").Int()))
+		if tool == nil {
+			break
+		}
+		contentBlockStop := `{"type":"content_block_stop","index":0}`
+		contentBlockStop, _ = sjson.Set(contentBlockStop, "index", tool.Index)
+		results = append(results, "event: content_block_stop\ndata: "+contentBlockStop+"\n\n")
+	case "response.completed":
+		ensureMessageStart()
+		stopTextBlockIfNeeded()
+		if !state.MessageStopSent {
+			stopReason := "end_turn"
+			if state.HasToolUse {
+				stopReason = "tool_use"
+			} else if sr := gjson.GetBytes(payload, "response.stop_reason").String(); sr == "max_tokens" || sr == "stop" {
+				stopReason = sr
+			}
+			inputTokens := gjson.GetBytes(payload, "response.usage.input_tokens").Int()
+			outputTokens := gjson.GetBytes(payload, "response.usage.output_tokens").Int()
+			cachedTokens := gjson.GetBytes(payload, "response.usage.input_tokens_details.cached_tokens").Int()
+			if cachedTokens > 0 && inputTokens >= cachedTokens {
+				inputTokens -= cachedTokens
+			}
+			messageDelta := `{"type":"message_delta","delta":{"stop_reason":"","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+			messageDelta, _ = sjson.Set(messageDelta, "delta.stop_reason", stopReason)
+			messageDelta, _ = sjson.Set(messageDelta, "usage.input_tokens", inputTokens)
+			messageDelta, _ = sjson.Set(messageDelta, "usage.output_tokens", outputTokens)
+			if cachedTokens > 0 {
+				messageDelta, _ = sjson.Set(messageDelta, "usage.cache_read_input_tokens", cachedTokens)
+			}
+			results = append(results, "event: message_delta\ndata: "+messageDelta+"\n\n")
+			results = append(results, "event: message_stop\ndata: {\"type\":\"message_stop\"}\n\n")
+			state.MessageStopSent = true
+		}
+	}
+
+	return results
+}
+
 // isHTTPSuccess checks if the status code indicates success (2xx).
 func isHTTPSuccess(statusCode int) bool {
 	return statusCode >= 200 && statusCode < 300
--- a/internal/runtime/executor/github_copilot_executor_test.go
+++ b/internal/runtime/executor/github_copilot_executor_test.go
@@ -1,8 +1,11 @@
 package executor

 import (
+	"net/http"
+	"strings"
 	"testing"

+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
 	"github.com/tidwall/gjson"
 )

@@ -52,3 +55,279 @@ func TestGitHubCopilotNormalizeModel_StripsSuffix(t *testing.T) {
 		})
 	}
 }
+
+func TestUseGitHubCopilotResponsesEndpoint_OpenAIResponseSource(t *testing.T) {
+	t.Parallel()
+	if !useGitHubCopilotResponsesEndpoint(sdktranslator.FromString("openai-response"), "claude-3-5-sonnet") {
+		t.Fatal("expected openai-response source to use /responses")
+	}
+}
+
+func TestUseGitHubCopilotResponsesEndpoint_CodexModel(t *testing.T) {
+	t.Parallel()
+	if !useGitHubCopilotResponsesEndpoint(sdktranslator.FromString("openai"), "gpt-5-codex") {
+		t.Fatal("expected codex model to use /responses")
+	}
+}
+
+func TestUseGitHubCopilotResponsesEndpoint_DefaultChat(t *testing.T) {
+	t.Parallel()
+	if useGitHubCopilotResponsesEndpoint(sdktranslator.FromString("openai"), "claude-3-5-sonnet") {
+		t.Fatal("expected default openai source with non-codex model to use /chat/completions")
+	}
+}
+
+func TestNormalizeGitHubCopilotChatTools_KeepFunctionOnly(t *testing.T) {
+	t.Parallel()
+	body := []byte(`{"tools":[{"type":"function","function":{"name":"ok"}},{"type":"code_interpreter"}],"tool_choice":"auto"}`)
+	got := normalizeGitHubCopilotChatTools(body)
+	tools := gjson.GetBytes(got, "tools").Array()
+	if len(tools) != 1 {
+		t.Fatalf("tools len = %d, want 1", len(tools))
+	}
+	if tools[0].Get("type").String() != "function" {
+		t.Fatalf("tool type = %q, want function", tools[0].Get("type").String())
+	}
+}
+
+func TestNormalizeGitHubCopilotChatTools_InvalidToolChoiceDowngradeToAuto(t *testing.T) {
+	t.Parallel()
+	body := []byte(`{"tools":[],"tool_choice":{"type":"function","function":{"name":"x"}}}`)
+	got := normalizeGitHubCopilotChatTools(body)
+	if gjson.GetBytes(got, "tool_choice").String() != "auto" {
+		t.Fatalf("tool_choice = %s, want auto", gjson.GetBytes(got, "tool_choice").Raw)
+	}
+}
+
+func TestNormalizeGitHubCopilotResponsesInput_MissingInputExtractedFromSystemAndMessages(t *testing.T) {
+	t.Parallel()
+	body := []byte(`{"system":"sys text","messages":[{"role":"user","content":"user text"},{"role":"assistant","content":[{"type":"text","text":"assistant text"}]}]}`)
+	got := normalizeGitHubCopilotResponsesInput(body)
+	in := gjson.GetBytes(got, "input")
+	if !in.IsArray() {
+		t.Fatalf("input type = %v, want array", in.Type)
+	}
+	raw := in.Raw
+	if !strings.Contains(raw, "sys text") || !strings.Contains(raw, "user text") || !strings.Contains(raw, "assistant text") {
+		t.Fatalf("input = %s, want structured array with all texts", raw)
+	}
+	if gjson.GetBytes(got, "messages").Exists() {
+		t.Fatal("messages should be removed after conversion")
+	}
+	if gjson.GetBytes(got, "system").Exists() {
+		t.Fatal("system should be removed after conversion")
+	}
+}
+
+func TestNormalizeGitHubCopilotResponsesInput_NonStringInputStringified(t *testing.T) {
+	t.Parallel()
+	body := []byte(`{"input":{"foo":"bar"}}`)
+	got := normalizeGitHubCopilotResponsesInput(body)
+	in := gjson.GetBytes(got, "input")
+	if in.Type != gjson.String {
+		t.Fatalf("input type = %v, want string", in.Type)
+	}
+	if !strings.Contains(in.String(), "foo") {
+		t.Fatalf("input = %q, want stringified object", in.String())
+	}
+}
+
+func TestNormalizeGitHubCopilotResponsesTools_FlattenFunctionTools(t *testing.T) {
+	t.Parallel()
+	body := []byte(`{"tools":[{"type":"function","function":{"name":"sum","description":"d","parameters":{"type":"object"}}},{"type":"web_search"}]}`)
+	got := normalizeGitHubCopilotResponsesTools(body)
+	tools := gjson.GetBytes(got, "tools").Array()
+	if len(tools) != 1 {
+		t.Fatalf("tools len = %d, want 1", len(tools))
+	}
+	if tools[0].Get("name").String() != "sum" {
+		t.Fatalf("tools[0].name = %q, want sum", tools[0].Get("name").String())
+	}
+	if !tools[0].Get("parameters").Exists() {
+		t.Fatal("expected parameters to be preserved")
+	}
+}
+
+func TestNormalizeGitHubCopilotResponsesTools_ClaudeFormatTools(t *testing.T) {
+	t.Parallel()
+	body := []byte(`{"tools":[{"name":"Bash","description":"Run commands","input_schema":{"type":"object","properties":{"command":{"type":"string"}},"required":["command"]}},{"name":"Read","description":"Read files","input_schema":{"type":"object","properties":{"path":{"type":"string"}}}}]}`)
+	got := normalizeGitHubCopilotResponsesTools(body)
+	tools := gjson.GetBytes(got, "tools").Array()
+	if len(tools) != 2 {
+		t.Fatalf("tools len = %d, want 2", len(tools))
+	}
+	if tools[0].Get("type").String() != "function" {
+		t.Fatalf("tools[0].type = %q, want function", tools[0].Get("type").String())
+	}
+	if tools[0].Get("name").String() != "Bash" {
+		t.Fatalf("tools[0].name = %q, want Bash", tools[0].Get("name").String())
+	}
+	if tools[0].Get("description").String() != "Run commands" {
+		t.Fatalf("tools[0].description = %q, want 'Run commands'", tools[0].Get("description").String())
+	}
+	if !tools[0].Get("parameters").Exists() {
+		t.Fatal("expected parameters to be set from input_schema")
+	}
+	if tools[0].Get("parameters.properties.command").Exists() != true {
+		t.Fatal("expected parameters.properties.command to exist")
+	}
+	if tools[1].Get("name").String() != "Read" {
+		t.Fatalf("tools[1].name = %q, want Read", tools[1].Get("name").String())
+	}
+}
+
+func TestNormalizeGitHubCopilotResponsesTools_FlattenToolChoiceFunctionObject(t *testing.T) {
+	t.Parallel()
+	body := []byte(`{"tool_choice":{"type":"function","function":{"name":"sum"}}}`)
+	got := normalizeGitHubCopilotResponsesTools(body)
+	if gjson.GetBytes(got, "tool_choice.type").String() != "function" {
+		t.Fatalf("tool_choice.type = %q, want function", gjson.GetBytes(got, "tool_choice.type").String())
+	}
+	if gjson.GetBytes(got, "tool_choice.name").String() != "sum" {
+		t.Fatalf("tool_choice.name = %q, want sum", gjson.GetBytes(got, "tool_choice.name").String())
+	}
+}
+
+func TestNormalizeGitHubCopilotResponsesTools_InvalidToolChoiceDowngradeToAuto(t *testing.T) {
+	t.Parallel()
+	body := []byte(`{"tool_choice":{"type":"function"}}`)
+	got := normalizeGitHubCopilotResponsesTools(body)
+	if gjson.GetBytes(got, "tool_choice").String() != "auto" {
+		t.Fatalf("tool_choice = %s, want auto", gjson.GetBytes(got, "tool_choice").Raw)
+	}
+}
+
+func TestTranslateGitHubCopilotResponsesNonStreamToClaude_TextMapping(t *testing.T) {
+	t.Parallel()
+	resp := []byte(`{"id":"resp_1","model":"gpt-5-codex","output":[{"type":"message","content":[{"type":"output_text","text":"hello"}]}],"usage":{"input_tokens":3,"output_tokens":5}}`)
+	out := translateGitHubCopilotResponsesNonStreamToClaude(resp)
+	if gjson.Get(out, "type").String() != "message" {
+		t.Fatalf("type = %q, want message", gjson.Get(out, "type").String())
+	}
+	if gjson.Get(out, "content.0.type").String() != "text" {
+		t.Fatalf("content.0.type = %q, want text", gjson.Get(out, "content.0.type").String())
+	}
+	if gjson.Get(out, "content.0.text").String() != "hello" {
+		t.Fatalf("content.0.text = %q, want hello", gjson.Get(out, "content.0.text").String())
+	}
+}
+
+func TestTranslateGitHubCopilotResponsesNonStreamToClaude_ToolUseMapping(t *testing.T) {
+	t.Parallel()
+	resp := []byte(`{"id":"resp_2","model":"gpt-5-codex","output":[{"type":"function_call","id":"fc_1","call_id":"call_1","name":"sum","arguments":"{\"a\":1}"}],"usage":{"input_tokens":1,"output_tokens":2}}`)
+	out := translateGitHubCopilotResponsesNonStreamToClaude(resp)
+	if gjson.Get(out, "content.0.type").String() != "tool_use" {
+		t.Fatalf("content.0.type = %q, want tool_use", gjson.Get(out, "content.0.type").String())
+	}
+	if gjson.Get(out, "content.0.name").String() != "sum" {
+		t.Fatalf("content.0.name = %q, want sum", gjson.Get(out, "content.0.name").String())
+	}
+	if gjson.Get(out, "stop_reason").String() != "tool_use" {
+		t.Fatalf("stop_reason = %q, want tool_use", gjson.Get(out, "stop_reason").String())
+	}
+}
+
+func TestTranslateGitHubCopilotResponsesStreamToClaude_TextLifecycle(t *testing.T) {
+	t.Parallel()
+	var param any
+
+	created := translateGitHubCopilotResponsesStreamToClaude([]byte(`data: {"type":"response.created","response":{"id":"resp_1","model":"gpt-5-codex"}}`), &param)
+	if len(created) == 0 || !strings.Contains(created[0], "message_start") {
+		t.Fatalf("created events = %#v, want message_start", created)
+	}
+
+	delta := translateGitHubCopilotResponsesStreamToClaude([]byte(`data: {"type":"response.output_text.delta","delta":"he"}`), &param)
+	joinedDelta := strings.Join(delta, "")
+	if !strings.Contains(joinedDelta, "content_block_start") || !strings.Contains(joinedDelta, "text_delta") {
+		t.Fatalf("delta events = %#v, want content_block_start + text_delta", delta)
+	}
+
+	completed := translateGitHubCopilotResponsesStreamToClaude([]byte(`data: {"type":"response.completed","response":{"usage":{"input_tokens":7,"output_tokens":9}}}`), &param)
+	joinedCompleted := strings.Join(completed, "")
+	if !strings.Contains(joinedCompleted, "message_delta") || !strings.Contains(joinedCompleted, "message_stop") {
+		t.Fatalf("completed events = %#v, want message_delta + message_stop", completed)
+	}
+}
+
+// --- Tests for X-Initiator detection logic (Problem L) ---
+
+func TestApplyHeaders_XInitiator_UserOnly(t *testing.T) {
+	t.Parallel()
+	e := &GitHubCopilotExecutor{}
+	req, _ := http.NewRequest(http.MethodPost, "https://example.com", nil)
+	body := []byte(`{"messages":[{"role":"system","content":"sys"},{"role":"user","content":"hello"}]}`)
+	e.applyHeaders(req, "token", body)
+	if got := req.Header.Get("X-Initiator"); got != "user" {
+		t.Fatalf("X-Initiator = %q, want user", got)
+	}
+}
+
+func TestApplyHeaders_XInitiator_AgentWithAssistantAndUserToolResult(t *testing.T) {
+	t.Parallel()
+	e := &GitHubCopilotExecutor{}
+	req, _ := http.NewRequest(http.MethodPost, "https://example.com", nil)
+	// Claude Code typical flow: last message is user (tool result), but has assistant in history
+	body := []byte(`{"messages":[{"role":"user","content":"hello"},{"role":"assistant","content":"I will read the file"},{"role":"user","content":"tool result here"}]}`)
+	e.applyHeaders(req, "token", body)
+	if got := req.Header.Get("X-Initiator"); got != "agent" {
+		t.Fatalf("X-Initiator = %q, want agent (assistant exists in messages)", got)
+	}
+}
+
+func TestApplyHeaders_XInitiator_AgentWithToolRole(t *testing.T) {
+	t.Parallel()
+	e := &GitHubCopilotExecutor{}
+	req, _ := http.NewRequest(http.MethodPost, "https://example.com", nil)
+	body := []byte(`{"messages":[{"role":"user","content":"hello"},{"role":"tool","content":"result"}]}`)
+	e.applyHeaders(req, "token", body)
+	if got := req.Header.Get("X-Initiator"); got != "agent" {
+		t.Fatalf("X-Initiator = %q, want agent (tool role exists)", got)
+	}
+}
+
+// --- Tests for x-github-api-version header (Problem M) ---
+
+func TestApplyHeaders_GitHubAPIVersion(t *testing.T) {
+	t.Parallel()
+	e := &GitHubCopilotExecutor{}
+	req, _ := http.NewRequest(http.MethodPost, "https://example.com", nil)
+	e.applyHeaders(req, "token", nil)
+	if got := req.Header.Get("X-Github-Api-Version"); got != "2025-04-01" {
+		t.Fatalf("X-Github-Api-Version = %q, want 2025-04-01", got)
+	}
+}
+
+// --- Tests for vision detection (Problem P) ---
+
+func TestDetectVisionContent_WithImageURL(t *testing.T) {
+	t.Parallel()
+	body := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"describe"},{"type":"image_url","image_url":{"url":"data:image/png;base64,abc"}}]}]}`)
+	if !detectVisionContent(body) {
+		t.Fatal("expected vision content to be detected")
+	}
+}
+
+func TestDetectVisionContent_WithImageType(t *testing.T) {
+	t.Parallel()
+	body := []byte(`{"messages":[{"role":"user","content":[{"type":"image","source":{"data":"abc","media_type":"image/png"}}]}]}`)
+	if !detectVisionContent(body) {
+		t.Fatal("expected image type to be detected")
+	}
+}
+
+func TestDetectVisionContent_NoVision(t *testing.T) {
+	t.Parallel()
+	body := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hello"}]}]}`)
+	if detectVisionContent(body) {
+		t.Fatal("expected no vision content")
+	}
+}
+
+func TestDetectVisionContent_NoMessages(t *testing.T) {
+	t.Parallel()
+	// After Responses API normalization, messages is removed — detection should return false
+	body := []byte(`{"input":[{"type":"message","role":"user","content":[{"type":"input_text","text":"hello"}]}]}`)
+	if detectVisionContent(body) {
+		t.Fatal("expected no vision content when messages field is absent")
+	}
+}
--- a/internal/runtime/executor/iflow_executor.go
+++ b/internal/runtime/executor/iflow_executor.go
@@ -169,12 +169,12 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
 	// the original model name in the response for client compatibility.
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }

 // ExecuteStream performs a streaming chat completion request.
-func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -262,7 +262,6 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	}

 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -294,7 +293,7 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 		reporter.ensurePublished(ctx)
 	}()

-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }

 func (e *IFlowExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
--- a/internal/runtime/executor/kilo_executor.go
+++ b/internal/runtime/executor/kilo_executor.go
@@ -0,0 +1,460 @@
+package executor
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
+	log "github.com/sirupsen/logrus"
+	"github.com/tidwall/gjson"
+)
+
+// KiloExecutor handles requests to Kilo API.
+type KiloExecutor struct {
+	cfg *config.Config
+}
+
+// NewKiloExecutor creates a new Kilo executor instance.
+func NewKiloExecutor(cfg *config.Config) *KiloExecutor {
+	return &KiloExecutor{cfg: cfg}
+}
+
+// Identifier returns the unique identifier for this executor.
+func (e *KiloExecutor) Identifier() string { return "kilo" }
+
+// PrepareRequest prepares the HTTP request before execution.
+func (e *KiloExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	accessToken, _ := kiloCredentials(auth)
+	if strings.TrimSpace(accessToken) == "" {
+		return fmt.Errorf("kilo: missing access token")
+	}
+
+	req.Header.Set("Authorization", "Bearer "+accessToken)
+	var attrs map[string]string
+	if auth != nil {
+		attrs = auth.Attributes
+	}
+	util.ApplyCustomHeadersFromAttrs(req, attrs)
+	return nil
+}
+
+// HttpRequest executes a raw HTTP request.
+func (e *KiloExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("kilo executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
+
+// Execute performs a non-streaming request.
+func (e *KiloExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
+
+	accessToken, orgID := kiloCredentials(auth)
+	if accessToken == "" {
+		return resp, fmt.Errorf("kilo: missing access token")
+	}
+
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("openai")
+	endpoint := "/api/openrouter/chat/completions"
+
+	originalPayloadSource := req.Payload
+	if len(opts.OriginalRequest) > 0 {
+		originalPayloadSource = opts.OriginalRequest
+	}
+	originalPayload := originalPayloadSource
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, opts.Stream)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, opts.Stream)
+	requestedModel := payloadRequestedModel(opts, req.Model)
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated, requestedModel)
+
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
+	if err != nil {
+		return resp, err
+	}
+
+	url := "https://api.kilo.ai" + endpoint
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(translated))
+	if err != nil {
+		return resp, err
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	httpReq.Header.Set("Authorization", "Bearer "+accessToken)
+	if orgID != "" {
+		httpReq.Header.Set("X-Kilocode-OrganizationID", orgID)
+	}
+	httpReq.Header.Set("User-Agent", "cli-proxy-kilo")
+	var attrs map[string]string
+	if auth != nil {
+		attrs = auth.Attributes
+	}
+	util.ApplyCustomHeadersFromAttrs(httpReq, attrs)
+
+	var authID, authLabel, authType, authValue string
+	if auth != nil {
+		authID = auth.ID
+		authLabel = auth.Label
+		authType, authValue = auth.AccountInfo()
+	}
+	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
+		URL:       url,
+		Method:    http.MethodPost,
+		Headers:   httpReq.Header.Clone(),
+		Body:      translated,
+		Provider:  e.Identifier(),
+		AuthID:    authID,
+		AuthLabel: authLabel,
+		AuthType:  authType,
+		AuthValue: authValue,
+	})
+
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpResp, err := httpClient.Do(httpReq)
+	if err != nil {
+		recordAPIResponseError(ctx, e.cfg, err)
+		return resp, err
+	}
+	defer httpResp.Body.Close()
+
+	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
+		b, _ := io.ReadAll(httpResp.Body)
+		appendAPIResponseChunk(ctx, e.cfg, b)
+		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
+		return resp, err
+	}
+
+	body, err := io.ReadAll(httpResp.Body)
+	if err != nil {
+		recordAPIResponseError(ctx, e.cfg, err)
+		return resp, err
+	}
+	appendAPIResponseChunk(ctx, e.cfg, body)
+	reporter.publish(ctx, parseOpenAIUsage(body))
+	reporter.ensurePublished(ctx)
+
+	var param any
+	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, body, &param)
+	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	return resp, nil
+}
+
+// ExecuteStream performs a streaming request.
+func (e *KiloExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
+
+	accessToken, orgID := kiloCredentials(auth)
+	if accessToken == "" {
+		return nil, fmt.Errorf("kilo: missing access token")
+	}
+
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("openai")
+	endpoint := "/api/openrouter/chat/completions"
+
+	originalPayloadSource := req.Payload
+	if len(opts.OriginalRequest) > 0 {
+		originalPayloadSource = opts.OriginalRequest
+	}
+	originalPayload := originalPayloadSource
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
+	requestedModel := payloadRequestedModel(opts, req.Model)
+	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated, requestedModel)
+
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
+	if err != nil {
+		return nil, err
+	}
+
+	url := "https://api.kilo.ai" + endpoint
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(translated))
+	if err != nil {
+		return nil, err
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	httpReq.Header.Set("Authorization", "Bearer "+accessToken)
+	if orgID != "" {
+		httpReq.Header.Set("X-Kilocode-OrganizationID", orgID)
+	}
+	httpReq.Header.Set("User-Agent", "cli-proxy-kilo")
+	httpReq.Header.Set("Accept", "text/event-stream")
+	httpReq.Header.Set("Cache-Control", "no-cache")
+
+	var attrs map[string]string
+	if auth != nil {
+		attrs = auth.Attributes
+	}
+	util.ApplyCustomHeadersFromAttrs(httpReq, attrs)
+
+	var authID, authLabel, authType, authValue string
+	if auth != nil {
+		authID = auth.ID
+		authLabel = auth.Label
+		authType, authValue = auth.AccountInfo()
+	}
+	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
+		URL:       url,
+		Method:    http.MethodPost,
+		Headers:   httpReq.Header.Clone(),
+		Body:      translated,
+		Provider:  e.Identifier(),
+		AuthID:    authID,
+		AuthLabel: authLabel,
+		AuthType:  authType,
+		AuthValue: authValue,
+	})
+
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpResp, err := httpClient.Do(httpReq)
+	if err != nil {
+		recordAPIResponseError(ctx, e.cfg, err)
+		return nil, err
+	}
+
+	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
+		b, _ := io.ReadAll(httpResp.Body)
+		appendAPIResponseChunk(ctx, e.cfg, b)
+		httpResp.Body.Close()
+		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
+		return nil, err
+	}
+
+	out := make(chan cliproxyexecutor.StreamChunk)
+	go func() {
+		defer close(out)
+		defer httpResp.Body.Close()
+
+		scanner := bufio.NewScanner(httpResp.Body)
+		scanner.Buffer(nil, 52_428_800)
+		var param any
+		for scanner.Scan() {
+			line := scanner.Bytes()
+			appendAPIResponseChunk(ctx, e.cfg, line)
+			if detail, ok := parseOpenAIStreamUsage(line); ok {
+				reporter.publish(ctx, detail)
+			}
+			if len(line) == 0 {
+				continue
+			}
+			if !bytes.HasPrefix(line, []byte("data:")) {
+				continue
+			}
+			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, bytes.Clone(line), &param)
+			for i := range chunks {
+				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
+			}
+		}
+		if errScan := scanner.Err(); errScan != nil {
+			recordAPIResponseError(ctx, e.cfg, errScan)
+			reporter.publishFailure(ctx)
+			out <- cliproxyexecutor.StreamChunk{Err: errScan}
+		}
+		reporter.ensurePublished(ctx)
+	}()
+
+	return &cliproxyexecutor.StreamResult{
+		Headers: httpResp.Header.Clone(),
+		Chunks:  out,
+	}, nil
+}
+
+// Refresh validates the Kilo token.
+func (e *KiloExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
+	if auth == nil {
+		return nil, fmt.Errorf("missing auth")
+	}
+	return auth, nil
+}
+
+// CountTokens returns the token count for the given request.
+func (e *KiloExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, fmt.Errorf("kilo: count tokens not supported")
+}
+
+// kiloCredentials extracts access token and other info from auth.
+func kiloCredentials(auth *cliproxyauth.Auth) (accessToken, orgID string) {
+	if auth == nil {
+		return "", ""
+	}
+
+	// Prefer kilocode specific keys, then fall back to generic keys.
+	// Check metadata first, then attributes.
+	if auth.Metadata != nil {
+		if token, ok := auth.Metadata["kilocodeToken"].(string); ok && token != "" {
+			accessToken = token
+		} else if token, ok := auth.Metadata["access_token"].(string); ok && token != "" {
+			accessToken = token
+		}
+
+		if org, ok := auth.Metadata["kilocodeOrganizationId"].(string); ok && org != "" {
+			orgID = org
+		} else if org, ok := auth.Metadata["organization_id"].(string); ok && org != "" {
+			orgID = org
+		}
+	}
+
+	if accessToken == "" && auth.Attributes != nil {
+		if token := auth.Attributes["kilocodeToken"]; token != "" {
+			accessToken = token
+		} else if token := auth.Attributes["access_token"]; token != "" {
+			accessToken = token
+		}
+	}
+
+	if orgID == "" && auth.Attributes != nil {
+		if org := auth.Attributes["kilocodeOrganizationId"]; org != "" {
+			orgID = org
+		} else if org := auth.Attributes["organization_id"]; org != "" {
+			orgID = org
+		}
+	}
+
+	return accessToken, orgID
+}
+
+// FetchKiloModels fetches models from Kilo API.
+func FetchKiloModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *config.Config) []*registry.ModelInfo {
+	accessToken, orgID := kiloCredentials(auth)
+	if accessToken == "" {
+		log.Infof("kilo: no access token found, skipping dynamic model fetch (using static kilo/auto)")
+		return registry.GetKiloModels()
+	}
+
+	log.Debugf("kilo: fetching dynamic models (orgID: %s)", orgID)
+
+	httpClient := newProxyAwareHTTPClient(ctx, cfg, auth, 0)
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, "https://api.kilo.ai/api/openrouter/models", nil)
+	if err != nil {
+		log.Warnf("kilo: failed to create model fetch request: %v", err)
+		return registry.GetKiloModels()
+	}
+
+	req.Header.Set("Authorization", "Bearer "+accessToken)
+	if orgID != "" {
+		req.Header.Set("X-Kilocode-OrganizationID", orgID)
+	}
+	req.Header.Set("User-Agent", "cli-proxy-kilo")
+
+	resp, err := httpClient.Do(req)
+	if err != nil {
+		if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
+			log.Warnf("kilo: fetch models canceled: %v", err)
+		} else {
+			log.Warnf("kilo: using static models (API fetch failed: %v)", err)
+		}
+		return registry.GetKiloModels()
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		log.Warnf("kilo: failed to read models response: %v", err)
+		return registry.GetKiloModels()
+	}
+
+	if resp.StatusCode != http.StatusOK {
+		log.Warnf("kilo: fetch models failed: status %d, body: %s", resp.StatusCode, string(body))
+		return registry.GetKiloModels()
+	}
+
+	result := gjson.GetBytes(body, "data")
+	if !result.Exists() {
+		// Try root if data field is missing
+		result = gjson.ParseBytes(body)
+		if !result.IsArray() {
+			log.Debugf("kilo: response body: %s", string(body))
+			log.Warn("kilo: invalid API response format (expected array or data field with array)")
+			return registry.GetKiloModels()
+		}
+	}
+
+	var dynamicModels []*registry.ModelInfo
+	now := time.Now().Unix()
+	count := 0
+	totalCount := 0
+
+	result.ForEach(func(key, value gjson.Result) bool {
+		totalCount++
+		id := value.Get("id").String()
+		pIdxResult := value.Get("preferredIndex")
+		preferredIndex := pIdxResult.Int()
+
+		// Filter models where preferredIndex > 0 (Kilo-curated models)
+		if preferredIndex <= 0 {
+			return true
+		}
+
+		// Check if it's free. We look for :free suffix, is_free flag, or zero pricing.
+		isFree := strings.HasSuffix(id, ":free") || id == "giga-potato" || value.Get("is_free").Bool()
+		if !isFree {
+			// Check pricing as fallback
+			promptPricing := value.Get("pricing.prompt").String()
+			if promptPricing == "0" || promptPricing == "0.0" {
+				isFree = true
+			}
+		}
+
+		if !isFree {
+			log.Debugf("kilo: skipping curated paid model: %s", id)
+			return true
+		}
+
+		log.Debugf("kilo: found curated model: %s (preferredIndex: %d)", id, preferredIndex)
+
+		dynamicModels = append(dynamicModels, &registry.ModelInfo{
+			ID:            id,
+			DisplayName:   value.Get("name").String(),
+			ContextLength: int(value.Get("context_length").Int()),
+			OwnedBy:       "kilo",
+			Type:          "kilo",
+			Object:        "model",
+			Created:       now,
+		})
+		count++
+		return true
+	})
+
+	log.Infof("kilo: fetched %d models from API, %d curated free (preferredIndex > 0)", totalCount, count)
+	if count == 0 && totalCount > 0 {
+		log.Warn("kilo: no curated free models found (check API response fields)")
+	}
+
+	staticModels := registry.GetKiloModels()
+	// Always include kilo/auto (first static model)
+	allModels := append(staticModels[:1], dynamicModels...)
+
+	return allModels
+}
--- a/internal/runtime/executor/kimi_executor.go
+++ b/internal/runtime/executor/kimi_executor.go
@@ -161,12 +161,12 @@ func (e *KimiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
 	// the original model name in the response for client compatibility.
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }

 // ExecuteStream performs a streaming chat completion request to Kimi.
-func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	from := opts.SourceFormat
 	if from.String() == "claude" {
 		auth.Attributes["base_url"] = kimiauth.KimiAPIBaseURL
@@ -253,7 +253,6 @@ func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -285,7 +284,7 @@ func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }

 // CountTokens estimates token count for Kimi requests.
--- a/internal/runtime/executor/kiro_executor.go
+++ b/internal/runtime/executor/kiro_executor.go
--- a/internal/runtime/executor/openai_compat_executor.go
+++ b/internal/runtime/executor/openai_compat_executor.go
@@ -172,11 +172,11 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	// Translate response back to source format when needed
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, body, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }

-func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	baseModel := thinking.ParseSuffix(req.Model).ModelName

 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
@@ -258,7 +258,6 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -298,7 +297,7 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 		// Ensure we record the request if no usage chunk was ever seen
 		reporter.ensurePublished(ctx)
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }

 func (e *OpenAICompatExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
--- a/internal/runtime/executor/qwen_executor.go
+++ b/internal/runtime/executor/qwen_executor.go
@@ -22,9 +22,7 @@ import (
 )

 const (
-	qwenUserAgent           = "google-api-nodejs-client/9.15.1"
-	qwenXGoogAPIClient      = "gl-node/22.17.0"
-	qwenClientMetadataValue = "ideType=IDE_UNSPECIFIED,platform=PLATFORM_UNSPECIFIED,pluginType=GEMINI"
+	qwenUserAgent = "QwenCode/0.10.3 (darwin; arm64)"
 )

 // QwenExecutor is a stateless executor for Qwen Code using OpenAI-compatible chat completions.
@@ -152,11 +150,11 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
 	// the original model name in the response for client compatibility.
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }

-func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -238,7 +236,6 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -270,7 +267,7 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }

 func (e *QwenExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
@@ -344,8 +341,18 @@ func applyQwenHeaders(r *http.Request, token string, stream bool) {
 	r.Header.Set("Content-Type", "application/json")
 	r.Header.Set("Authorization", "Bearer "+token)
 	r.Header.Set("User-Agent", qwenUserAgent)
-	r.Header.Set("X-Goog-Api-Client", qwenXGoogAPIClient)
-	r.Header.Set("Client-Metadata", qwenClientMetadataValue)
+	r.Header.Set("X-Dashscope-Useragent", qwenUserAgent)
+	r.Header.Set("X-Stainless-Runtime-Version", "v22.17.0")
+	r.Header.Set("Sec-Fetch-Mode", "cors")
+	r.Header.Set("X-Stainless-Lang", "js")
+	r.Header.Set("X-Stainless-Arch", "arm64")
+	r.Header.Set("X-Stainless-Package-Version", "5.11.0")
+	r.Header.Set("X-Dashscope-Cachecontrol", "enable")
+	r.Header.Set("X-Stainless-Retry-Count", "0")
+	r.Header.Set("X-Stainless-Os", "MacOS")
+	r.Header.Set("X-Dashscope-Authtype", "qwen-oauth")
+	r.Header.Set("X-Stainless-Runtime", "node")
+
 	if stream {
 		r.Header.Set("Accept", "text/event-stream")
 		return
--- a/internal/thinking/provider/openai/apply.go
+++ b/internal/thinking/provider/openai/apply.go
@@ -10,10 +10,53 @@ import (

 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )

+// validReasoningEffortLevels contains the standard values accepted by the
+// OpenAI reasoning_effort field. Provider-specific extensions (xhigh, minimal,
+// auto) are NOT in this set and must be clamped before use.
+var validReasoningEffortLevels = map[string]struct{}{
+	"none":   {},
+	"low":    {},
+	"medium": {},
+	"high":   {},
+}
+
+// clampReasoningEffort maps any thinking level string to a value that is safe
+// to send as OpenAI reasoning_effort. Non-standard CPA-internal values are
+// mapped to the nearest standard equivalent.
+//
+// Mapping rules:
+//   - none / low / medium / high  → returned as-is (already valid)
+//   - xhigh                       → "high" (nearest lower standard level)
+//   - minimal                     → "low" (nearest higher standard level)
+//   - auto                        → "medium" (reasonable default)
+//   - anything else               → "medium" (safe default)
+func clampReasoningEffort(level string) string {
+	if _, ok := validReasoningEffortLevels[level]; ok {
+		return level
+	}
+	var clamped string
+	switch level {
+	case string(thinking.LevelXHigh):
+		clamped = string(thinking.LevelHigh)
+	case string(thinking.LevelMinimal):
+		clamped = string(thinking.LevelLow)
+	case string(thinking.LevelAuto):
+		clamped = string(thinking.LevelMedium)
+	default:
+		clamped = string(thinking.LevelMedium)
+	}
+	log.WithFields(log.Fields{
+		"original": level,
+		"clamped":  clamped,
+	}).Debug("openai: reasoning_effort clamped to nearest valid standard value")
+	return clamped
+}
+
 // Applier implements thinking.ProviderApplier for OpenAI models.
 //
 // OpenAI-specific behavior:
@@ -58,7 +101,7 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 	}

 	if config.Mode == thinking.ModeLevel {
-		result, _ := sjson.SetBytes(body, "reasoning_effort", string(config.Level))
+		result, _ := sjson.SetBytes(body, "reasoning_effort", clampReasoningEffort(string(config.Level)))
 		return result, nil
 	}

@@ -79,7 +122,7 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 		return body, nil
 	}

-	result, _ := sjson.SetBytes(body, "reasoning_effort", effort)
+	result, _ := sjson.SetBytes(body, "reasoning_effort", clampReasoningEffort(effort))
 	return result, nil
 }

@@ -114,7 +157,7 @@ func applyCompatibleOpenAI(body []byte, config thinking.ThinkingConfig) ([]byte,
 		return body, nil
 	}

-	result, _ := sjson.SetBytes(body, "reasoning_effort", effort)
+	result, _ := sjson.SetBytes(body, "reasoning_effort", clampReasoningEffort(effort))
 	return result, nil
 }

--- a/internal/translator/antigravity/claude/antigravity_claude_request.go
+++ b/internal/translator/antigravity/claude/antigravity_claude_request.go
@@ -231,8 +231,12 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _

 							} else if functionResponseResult.IsObject() {
 								functionResponseJSON, _ = sjson.SetRaw(functionResponseJSON, "response.result", functionResponseResult.Raw)
-							} else {
+							} else if functionResponseResult.Raw != "" {
 								functionResponseJSON, _ = sjson.SetRaw(functionResponseJSON, "response.result", functionResponseResult.Raw)
+							} else {
+								// Content field is missing entirely — .Raw is empty which
+								// causes sjson.SetRaw to produce invalid JSON (e.g. "result":}).
+								functionResponseJSON, _ = sjson.Set(functionResponseJSON, "response.result", "")
 							}

 							partJSON := `{}`
--- a/internal/translator/antigravity/claude/antigravity_claude_request_test.go
+++ b/internal/translator/antigravity/claude/antigravity_claude_request_test.go
@@ -661,6 +661,85 @@ func TestConvertClaudeRequestToAntigravity_ThinkingOnly_NoHint(t *testing.T) {
 	}
 }

+func TestConvertClaudeRequestToAntigravity_ToolResultNoContent(t *testing.T) {
+	// Bug repro: tool_result with no content field produces invalid JSON
+	inputJSON := []byte(`{
+		"model": "claude-opus-4-6-thinking",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{
+						"type": "tool_use",
+						"id": "MyTool-123-456",
+						"name": "MyTool",
+						"input": {"key": "value"}
+					}
+				]
+			},
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "tool_result",
+						"tool_use_id": "MyTool-123-456"
+					}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-opus-4-6-thinking", inputJSON, true)
+	outputStr := string(output)
+
+	if !gjson.Valid(outputStr) {
+		t.Errorf("Result is not valid JSON:\n%s", outputStr)
+	}
+
+	// Verify the functionResponse has a valid result value
+	fr := gjson.Get(outputStr, "request.contents.1.parts.0.functionResponse.response.result")
+	if !fr.Exists() {
+		t.Error("functionResponse.response.result should exist")
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ToolResultNullContent(t *testing.T) {
+	// Bug repro: tool_result with null content produces invalid JSON
+	inputJSON := []byte(`{
+		"model": "claude-opus-4-6-thinking",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{
+						"type": "tool_use",
+						"id": "MyTool-123-456",
+						"name": "MyTool",
+						"input": {"key": "value"}
+					}
+				]
+			},
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "tool_result",
+						"tool_use_id": "MyTool-123-456",
+						"content": null
+					}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-opus-4-6-thinking", inputJSON, true)
+	outputStr := string(output)
+
+	if !gjson.Valid(outputStr) {
+		t.Errorf("Result is not valid JSON:\n%s", outputStr)
+	}
+}
+
 func TestConvertClaudeRequestToAntigravity_ToolAndThinking_NoExistingSystem(t *testing.T) {
 	// When tools + thinking but no system instruction, should create one with hint
 	inputJSON := []byte(`{
--- a/internal/translator/codex/openai/chat-completions/codex_openai_response.go
+++ b/internal/translator/codex/openai/chat-completions/codex_openai_response.go
@@ -20,10 +20,12 @@ var (

 // ConvertCliToOpenAIParams holds parameters for response conversion.
 type ConvertCliToOpenAIParams struct {
-	ResponseID        string
-	CreatedAt         int64
-	Model             string
-	FunctionCallIndex int
+	ResponseID                string
+	CreatedAt                 int64
+	Model                     string
+	FunctionCallIndex         int
+	HasReceivedArgumentsDelta bool
+	HasToolCallAnnounced      bool
 }

 // ConvertCodexResponseToOpenAI translates a single chunk of a streaming response from the
@@ -43,10 +45,12 @@ type ConvertCliToOpenAIParams struct {
 func ConvertCodexResponseToOpenAI(_ context.Context, modelName string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, param *any) []string {
 	if *param == nil {
 		*param = &ConvertCliToOpenAIParams{
-			Model:             modelName,
-			CreatedAt:         0,
-			ResponseID:        "",
-			FunctionCallIndex: -1,
+			Model:                     modelName,
+			CreatedAt:                 0,
+			ResponseID:                "",
+			FunctionCallIndex:         -1,
+			HasReceivedArgumentsDelta: false,
+			HasToolCallAnnounced:      false,
 		}
 	}

@@ -90,6 +94,9 @@ func ConvertCodexResponseToOpenAI(_ context.Context, modelName string, originalR
 		if inputTokensResult := usageResult.Get("input_tokens"); inputTokensResult.Exists() {
 			template, _ = sjson.Set(template, "usage.prompt_tokens", inputTokensResult.Int())
 		}
+		if cachedTokensResult := usageResult.Get("input_tokens_details.cached_tokens"); cachedTokensResult.Exists() {
+			template, _ = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cachedTokensResult.Int())
+		}
 		if reasoningTokensResult := usageResult.Get("output_tokens_details.reasoning_tokens"); reasoningTokensResult.Exists() {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", reasoningTokensResult.Int())
 		}
@@ -115,35 +122,93 @@ func ConvertCodexResponseToOpenAI(_ context.Context, modelName string, originalR
 		}
 		template, _ = sjson.Set(template, "choices.0.finish_reason", finishReason)
 		template, _ = sjson.Set(template, "choices.0.native_finish_reason", finishReason)
-	} else if dataType == "response.output_item.done" {
-		functionCallItemTemplate := `{"index":0,"id":"","type":"function","function":{"name":"","arguments":""}}`
+	} else if dataType == "response.output_item.added" {
 		itemResult := rootResult.Get("item")
-		if itemResult.Exists() {
-			if itemResult.Get("type").String() != "function_call" {
-				return []string{}
-			}
-
-			// set the index
-			(*param).(*ConvertCliToOpenAIParams).FunctionCallIndex++
-			functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "index", (*param).(*ConvertCliToOpenAIParams).FunctionCallIndex)
-
-			template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls", `[]`)
-			functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "id", itemResult.Get("call_id").String())
-
-			// Restore original tool name if it was shortened
-			name := itemResult.Get("name").String()
-			// Build reverse map on demand from original request tools
-			rev := buildReverseMapFromOriginalOpenAI(originalRequestRawJSON)
-			if orig, ok := rev[name]; ok {
-				name = orig
-			}
-			functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.name", name)
-
-			functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.arguments", itemResult.Get("arguments").String())
-			template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
-			template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls.-1", functionCallItemTemplate)
+		if !itemResult.Exists() || itemResult.Get("type").String() != "function_call" {
+			return []string{}
 		}

+		// Increment index for this new function call item.
+		(*param).(*ConvertCliToOpenAIParams).FunctionCallIndex++
+		(*param).(*ConvertCliToOpenAIParams).HasReceivedArgumentsDelta = false
+		(*param).(*ConvertCliToOpenAIParams).HasToolCallAnnounced = true
+
+		functionCallItemTemplate := `{"index":0,"id":"","type":"function","function":{"name":"","arguments":""}}`
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "index", (*param).(*ConvertCliToOpenAIParams).FunctionCallIndex)
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "id", itemResult.Get("call_id").String())
+
+		// Restore original tool name if it was shortened.
+		name := itemResult.Get("name").String()
+		rev := buildReverseMapFromOriginalOpenAI(originalRequestRawJSON)
+		if orig, ok := rev[name]; ok {
+			name = orig
+		}
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.name", name)
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.arguments", "")
+
+		template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
+		template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls", `[]`)
+		template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls.-1", functionCallItemTemplate)
+
+	} else if dataType == "response.function_call_arguments.delta" {
+		(*param).(*ConvertCliToOpenAIParams).HasReceivedArgumentsDelta = true
+
+		deltaValue := rootResult.Get("delta").String()
+		functionCallItemTemplate := `{"index":0,"function":{"arguments":""}}`
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "index", (*param).(*ConvertCliToOpenAIParams).FunctionCallIndex)
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.arguments", deltaValue)
+
+		template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls", `[]`)
+		template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls.-1", functionCallItemTemplate)
+
+	} else if dataType == "response.function_call_arguments.done" {
+		if (*param).(*ConvertCliToOpenAIParams).HasReceivedArgumentsDelta {
+			// Arguments were already streamed via delta events; nothing to emit.
+			return []string{}
+		}
+
+		// Fallback: no delta events were received, emit the full arguments as a single chunk.
+		fullArgs := rootResult.Get("arguments").String()
+		functionCallItemTemplate := `{"index":0,"function":{"arguments":""}}`
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "index", (*param).(*ConvertCliToOpenAIParams).FunctionCallIndex)
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.arguments", fullArgs)
+
+		template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls", `[]`)
+		template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls.-1", functionCallItemTemplate)
+
+	} else if dataType == "response.output_item.done" {
+		itemResult := rootResult.Get("item")
+		if !itemResult.Exists() || itemResult.Get("type").String() != "function_call" {
+			return []string{}
+		}
+
+		if (*param).(*ConvertCliToOpenAIParams).HasToolCallAnnounced {
+			// Tool call was already announced via output_item.added; skip emission.
+			(*param).(*ConvertCliToOpenAIParams).HasToolCallAnnounced = false
+			return []string{}
+		}
+
+		// Fallback path: model skipped output_item.added, so emit complete tool call now.
+		(*param).(*ConvertCliToOpenAIParams).FunctionCallIndex++
+
+		functionCallItemTemplate := `{"index":0,"id":"","type":"function","function":{"name":"","arguments":""}}`
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "index", (*param).(*ConvertCliToOpenAIParams).FunctionCallIndex)
+
+		template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls", `[]`)
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "id", itemResult.Get("call_id").String())
+
+		// Restore original tool name if it was shortened.
+		name := itemResult.Get("name").String()
+		rev := buildReverseMapFromOriginalOpenAI(originalRequestRawJSON)
+		if orig, ok := rev[name]; ok {
+			name = orig
+		}
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.name", name)
+
+		functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.arguments", itemResult.Get("arguments").String())
+		template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
+		template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls.-1", functionCallItemTemplate)
+
 	} else {
 		return []string{}
 	}
@@ -205,6 +270,9 @@ func ConvertCodexResponseToOpenAINonStream(_ context.Context, _ string, original
 		if inputTokensResult := usageResult.Get("input_tokens"); inputTokensResult.Exists() {
 			template, _ = sjson.Set(template, "usage.prompt_tokens", inputTokensResult.Int())
 		}
+		if cachedTokensResult := usageResult.Get("input_tokens_details.cached_tokens"); cachedTokensResult.Exists() {
+			template, _ = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cachedTokensResult.Int())
+		}
 		if reasoningTokensResult := usageResult.Get("output_tokens_details.reasoning_tokens"); reasoningTokensResult.Exists() {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", reasoningTokensResult.Int())
 		}
--- a/internal/translator/kiro/claude/kiro_claude_request.go
+++ b/internal/translator/kiro/claude/kiro_claude_request.go
@@ -586,6 +586,17 @@ func processMessages(messages gjson.Result, modelID, origin string) ([]KiroHisto

 	// Merge adjacent messages with the same role
 	messagesArray := kirocommon.MergeAdjacentMessages(messages.Array())
+
+	// FIX: Kiro API requires history to start with a user message.
+	// Some clients (e.g., OpenClaw) send conversations starting with an assistant message,
+	// which is valid for the Claude API but causes "Improperly formed request" on Kiro.
+	// Prepend a placeholder user message so the history alternation is correct.
+	if len(messagesArray) > 0 && messagesArray[0].Get("role").String() == "assistant" {
+		placeholder := `{"role":"user","content":"."}`
+		messagesArray = append([]gjson.Result{gjson.Parse(placeholder)}, messagesArray...)
+		log.Infof("kiro: messages started with assistant role, prepended placeholder user message for Kiro API compatibility")
+	}
+
 	for i, msg := range messagesArray {
 		role := msg.Get("role").String()
 		isLastMessage := i == len(messagesArray)-1
--- a/internal/translator/kiro/claude/kiro_claude_stream.go
+++ b/internal/translator/kiro/claude/kiro_claude_stream.go
@@ -183,4 +183,124 @@ func PendingTagSuffix(buffer, tag string) int {
 		}
 	}
 	return 0
-}
+}
+
+// GenerateSearchIndicatorEvents generates ONLY the search indicator SSE events
+// (server_tool_use + web_search_tool_result) without text summary or message termination.
+// These events trigger Claude Code's search indicator UI.
+// The caller is responsible for sending message_start before and message_delta/stop after.
+func GenerateSearchIndicatorEvents(
+	query string,
+	toolUseID string,
+	searchResults *WebSearchResults,
+	startIndex int,
+) [][]byte {
+	events := make([][]byte, 0, 5)
+
+	// 1. content_block_start (server_tool_use)
+	event1 := map[string]interface{}{
+		"type":  "content_block_start",
+		"index": startIndex,
+		"content_block": map[string]interface{}{
+			"id":    toolUseID,
+			"type":  "server_tool_use",
+			"name":  "web_search",
+			"input": map[string]interface{}{},
+		},
+	}
+	data1, _ := json.Marshal(event1)
+	events = append(events, []byte("event: content_block_start\ndata: "+string(data1)+"\n\n"))
+
+	// 2. content_block_delta (input_json_delta)
+	inputJSON, _ := json.Marshal(map[string]string{"query": query})
+	event2 := map[string]interface{}{
+		"type":  "content_block_delta",
+		"index": startIndex,
+		"delta": map[string]interface{}{
+			"type":         "input_json_delta",
+			"partial_json": string(inputJSON),
+		},
+	}
+	data2, _ := json.Marshal(event2)
+	events = append(events, []byte("event: content_block_delta\ndata: "+string(data2)+"\n\n"))
+
+	// 3. content_block_stop (server_tool_use)
+	event3 := map[string]interface{}{
+		"type":  "content_block_stop",
+		"index": startIndex,
+	}
+	data3, _ := json.Marshal(event3)
+	events = append(events, []byte("event: content_block_stop\ndata: "+string(data3)+"\n\n"))
+
+	// 4. content_block_start (web_search_tool_result)
+	searchContent := make([]map[string]interface{}, 0)
+	if searchResults != nil {
+		for _, r := range searchResults.Results {
+			snippet := ""
+			if r.Snippet != nil {
+				snippet = *r.Snippet
+			}
+			searchContent = append(searchContent, map[string]interface{}{
+				"type":              "web_search_result",
+				"title":             r.Title,
+				"url":               r.URL,
+				"encrypted_content": snippet,
+				"page_age":          nil,
+			})
+		}
+	}
+	event4 := map[string]interface{}{
+		"type":  "content_block_start",
+		"index": startIndex + 1,
+		"content_block": map[string]interface{}{
+			"type":        "web_search_tool_result",
+			"tool_use_id": toolUseID,
+			"content":     searchContent,
+		},
+	}
+	data4, _ := json.Marshal(event4)
+	events = append(events, []byte("event: content_block_start\ndata: "+string(data4)+"\n\n"))
+
+	// 5. content_block_stop (web_search_tool_result)
+	event5 := map[string]interface{}{
+		"type":  "content_block_stop",
+		"index": startIndex + 1,
+	}
+	data5, _ := json.Marshal(event5)
+	events = append(events, []byte("event: content_block_stop\ndata: "+string(data5)+"\n\n"))
+
+	return events
+}
+
+// BuildFallbackTextEvents generates SSE events for a fallback text response
+// when the Kiro API fails during the search loop. Uses BuildClaude*Event()
+// functions to align with streamToChannel patterns.
+// Returns raw SSE byte slices ready to be sent to the client channel.
+func BuildFallbackTextEvents(contentBlockIndex int, query string, results *WebSearchResults) [][]byte {
+	summary := FormatSearchContextPrompt(query, results)
+	outputTokens := len(summary) / 4
+	if len(summary) > 0 && outputTokens == 0 {
+		outputTokens = 1
+	}
+
+	var events [][]byte
+
+	// content_block_start (text)
+	events = append(events, BuildClaudeContentBlockStartEvent(contentBlockIndex, "text", "", ""))
+
+	// content_block_delta (text_delta)
+	events = append(events, BuildClaudeStreamEvent(summary, contentBlockIndex))
+
+	// content_block_stop
+	events = append(events, BuildClaudeContentBlockStopEvent(contentBlockIndex))
+
+	// message_delta with end_turn
+	events = append(events, BuildClaudeMessageDeltaEvent("end_turn", usage.Detail{
+		OutputTokens: int64(outputTokens),
+	}))
+
+	// message_stop
+	events = append(events, BuildClaudeMessageStopOnlyEvent())
+
+	return events
+}
--- a/internal/translator/kiro/claude/kiro_claude_stream_parser.go
+++ b/internal/translator/kiro/claude/kiro_claude_stream_parser.go
@@ -0,0 +1,350 @@
+package claude
+
+import (
+	"encoding/json"
+	"strings"
+
+	log "github.com/sirupsen/logrus"
+)
+
+// sseEvent represents a Server-Sent Event
+type sseEvent struct {
+	Event string
+	Data  interface{}
+}
+
+// ToSSEString converts the event to SSE wire format
+func (e *sseEvent) ToSSEString() string {
+	dataBytes, _ := json.Marshal(e.Data)
+	return "event: " + e.Event + "\ndata: " + string(dataBytes) + "\n\n"
+}
+
+// AdjustStreamIndices adjusts content block indices in SSE event data by adding an offset.
+// It also suppresses duplicate message_start events (returns shouldForward=false).
+// This is used to combine search indicator events (indices 0,1) with Kiro model response events.
+//
+// The data parameter is a single SSE "data:" line payload (JSON).
+// Returns: adjusted data, shouldForward (false = skip this event).
+func AdjustStreamIndices(data []byte, offset int) ([]byte, bool) {
+	if len(data) == 0 {
+		return data, true
+	}
+
+	// Quick check: parse the JSON
+	var event map[string]interface{}
+	if err := json.Unmarshal(data, &event); err != nil {
+		// Not valid JSON, pass through
+		return data, true
+	}
+
+	eventType, _ := event["type"].(string)
+
+	// Suppress duplicate message_start events
+	if eventType == "message_start" {
+		return data, false
+	}
+
+	// Adjust index for content_block events
+	switch eventType {
+	case "content_block_start", "content_block_delta", "content_block_stop":
+		if idx, ok := event["index"].(float64); ok {
+			event["index"] = int(idx) + offset
+			adjusted, err := json.Marshal(event)
+			if err != nil {
+				return data, true
+			}
+			return adjusted, true
+		}
+	}
+
+	// Pass through all other events unchanged (message_delta, message_stop, ping, etc.)
+	return data, true
+}
+
+// AdjustSSEChunk processes a raw SSE chunk (potentially containing multiple "event:/data:" pairs)
+// and adjusts content block indices. Suppresses duplicate message_start events.
+// Returns the adjusted chunk and whether it should be forwarded.
+func AdjustSSEChunk(chunk []byte, offset int) ([]byte, bool) {
+	chunkStr := string(chunk)
+
+	// Fast path: if no "data:" prefix, pass through
+	if !strings.Contains(chunkStr, "data: ") {
+		return chunk, true
+	}
+
+	var result strings.Builder
+	hasContent := false
+
+	lines := strings.Split(chunkStr, "\n")
+	for i := 0; i < len(lines); i++ {
+		line := lines[i]
+
+		if strings.HasPrefix(line, "data: ") {
+			dataPayload := strings.TrimPrefix(line, "data: ")
+			dataPayload = strings.TrimSpace(dataPayload)
+
+			if dataPayload == "[DONE]" {
+				result.WriteString(line + "\n")
+				hasContent = true
+				continue
+			}
+
+			adjusted, shouldForward := AdjustStreamIndices([]byte(dataPayload), offset)
+			if !shouldForward {
+				// Skip this event and its preceding "event:" line
+				// Also skip the trailing empty line
+				continue
+			}
+
+			result.WriteString("data: " + string(adjusted) + "\n")
+			hasContent = true
+		} else if strings.HasPrefix(line, "event: ") {
+			// Check if the next data line will be suppressed
+			if i+1 < len(lines) && strings.HasPrefix(lines[i+1], "data: ") {
+				dataPayload := strings.TrimPrefix(lines[i+1], "data: ")
+				dataPayload = strings.TrimSpace(dataPayload)
+
+				var event map[string]interface{}
+				if err := json.Unmarshal([]byte(dataPayload), &event); err == nil {
+					if eventType, ok := event["type"].(string); ok && eventType == "message_start" {
+						// Skip both the event: and data: lines
+						i++ // skip the data: line too
+						continue
+					}
+				}
+			}
+			result.WriteString(line + "\n")
+			hasContent = true
+		} else {
+			result.WriteString(line + "\n")
+			if strings.TrimSpace(line) != "" {
+				hasContent = true
+			}
+		}
+	}
+
+	if !hasContent {
+		return nil, false
+	}
+
+	return []byte(result.String()), true
+}
+
+// BufferedStreamResult contains the analysis of buffered SSE chunks from a Kiro API response.
+type BufferedStreamResult struct {
+	// StopReason is the detected stop_reason from the stream (e.g., "end_turn", "tool_use")
+	StopReason string
+	// WebSearchQuery is the extracted query if the model requested another web_search
+	WebSearchQuery string
+	// WebSearchToolUseId is the tool_use ID from the model's response (needed for toolResults)
+	WebSearchToolUseId string
+	// HasWebSearchToolUse indicates whether the model requested web_search
+	HasWebSearchToolUse bool
+	// WebSearchToolUseIndex is the content_block index of the web_search tool_use
+	WebSearchToolUseIndex int
+}
+
+// AnalyzeBufferedStream scans buffered SSE chunks to detect stop_reason and web_search tool_use.
+// This is used in the search loop to determine if the model wants another search round.
+func AnalyzeBufferedStream(chunks [][]byte) BufferedStreamResult {
+	result := BufferedStreamResult{WebSearchToolUseIndex: -1}
+
+	// Track tool use state across chunks
+	var currentToolName string
+	var currentToolIndex int = -1
+	var toolInputBuilder strings.Builder
+
+	for _, chunk := range chunks {
+		chunkStr := string(chunk)
+		lines := strings.Split(chunkStr, "\n")
+		for _, line := range lines {
+			if !strings.HasPrefix(line, "data: ") {
+				continue
+			}
+			dataPayload := strings.TrimPrefix(line, "data: ")
+			dataPayload = strings.TrimSpace(dataPayload)
+			if dataPayload == "[DONE]" || dataPayload == "" {
+				continue
+			}
+
+			var event map[string]interface{}
+			if err := json.Unmarshal([]byte(dataPayload), &event); err != nil {
+				continue
+			}
+
+			eventType, _ := event["type"].(string)
+
+			switch eventType {
+			case "message_delta":
+				// Extract stop_reason from message_delta
+				if delta, ok := event["delta"].(map[string]interface{}); ok {
+					if sr, ok := delta["stop_reason"].(string); ok && sr != "" {
+						result.StopReason = sr
+					}
+				}
+
+			case "content_block_start":
+				// Detect tool_use content blocks
+				if cb, ok := event["content_block"].(map[string]interface{}); ok {
+					if cbType, ok := cb["type"].(string); ok && cbType == "tool_use" {
+						if name, ok := cb["name"].(string); ok {
+							currentToolName = strings.ToLower(name)
+							if idx, ok := event["index"].(float64); ok {
+								currentToolIndex = int(idx)
+							}
+							// Capture tool use ID for toolResults handshake
+							if id, ok := cb["id"].(string); ok {
+								result.WebSearchToolUseId = id
+							}
+							toolInputBuilder.Reset()
+						}
+					}
+				}
+
+			case "content_block_delta":
+				// Accumulate tool input JSON
+				if currentToolName != "" {
+					if delta, ok := event["delta"].(map[string]interface{}); ok {
+						if deltaType, ok := delta["type"].(string); ok && deltaType == "input_json_delta" {
+							if partial, ok := delta["partial_json"].(string); ok {
+								toolInputBuilder.WriteString(partial)
+							}
+						}
+					}
+				}
+
+			case "content_block_stop":
+				// Finalize tool use detection
+				if currentToolName == "web_search" || currentToolName == "websearch" || currentToolName == "remote_web_search" {
+					result.HasWebSearchToolUse = true
+					result.WebSearchToolUseIndex = currentToolIndex
+					// Extract query from accumulated input JSON
+					inputJSON := toolInputBuilder.String()
+					var input map[string]string
+					if err := json.Unmarshal([]byte(inputJSON), &input); err == nil {
+						if q, ok := input["query"]; ok {
+							result.WebSearchQuery = q
+						}
+					}
+					log.Debugf("kiro/websearch: detected web_search tool_use")
+				}
+				currentToolName = ""
+				currentToolIndex = -1
+				toolInputBuilder.Reset()
+			}
+		}
+	}
+
+	return result
+}
+
+// FilterChunksForClient processes buffered SSE chunks and removes web_search tool_use
+// content blocks. This prevents the client from seeing "Tool use" prompts for web_search
+// when the proxy is handling the search loop internally.
+// Also suppresses message_start and message_delta/message_stop events since those
+// are managed by the outer handleWebSearchStream.
+func FilterChunksForClient(chunks [][]byte, wsToolIndex int, indexOffset int) [][]byte {
+	var filtered [][]byte
+
+	for _, chunk := range chunks {
+		chunkStr := string(chunk)
+		lines := strings.Split(chunkStr, "\n")
+
+		var resultBuilder strings.Builder
+		hasContent := false
+
+		for i := 0; i < len(lines); i++ {
+			line := lines[i]
+
+			if strings.HasPrefix(line, "data: ") {
+				dataPayload := strings.TrimPrefix(line, "data: ")
+				dataPayload = strings.TrimSpace(dataPayload)
+
+				if dataPayload == "[DONE]" {
+					// Skip [DONE] — the outer loop manages stream termination
+					continue
+				}
+
+				var event map[string]interface{}
+				if err := json.Unmarshal([]byte(dataPayload), &event); err != nil {
+					resultBuilder.WriteString(line + "\n")
+					hasContent = true
+					continue
+				}
+
+				eventType, _ := event["type"].(string)
+
+				// Skip message_start (outer loop sends its own)
+				if eventType == "message_start" {
+					continue
+				}
+
+				// Skip message_delta and message_stop (outer loop manages these)
+				if eventType == "message_delta" || eventType == "message_stop" {
+					continue
+				}
+
+				// Check if this event belongs to the web_search tool_use block
+				if wsToolIndex >= 0 {
+					if idx, ok := event["index"].(float64); ok && int(idx) == wsToolIndex {
+						// Skip events for the web_search tool_use block
+						continue
+					}
+				}
+
+				// Apply index offset for remaining events
+				if indexOffset > 0 {
+					switch eventType {
+					case "content_block_start", "content_block_delta", "content_block_stop":
+						if idx, ok := event["index"].(float64); ok {
+							event["index"] = int(idx) + indexOffset
+							adjusted, err := json.Marshal(event)
+							if err == nil {
+								resultBuilder.WriteString("data: " + string(adjusted) + "\n")
+								hasContent = true
+								continue
+							}
+						}
+					}
+				}
+
+				resultBuilder.WriteString(line + "\n")
+				hasContent = true
+			} else if strings.HasPrefix(line, "event: ") {
+				// Check if the next data line will be suppressed
+				if i+1 < len(lines) && strings.HasPrefix(lines[i+1], "data: ") {
+					nextData := strings.TrimPrefix(lines[i+1], "data: ")
+					nextData = strings.TrimSpace(nextData)
+
+					var nextEvent map[string]interface{}
+					if err := json.Unmarshal([]byte(nextData), &nextEvent); err == nil {
+						nextType, _ := nextEvent["type"].(string)
+						if nextType == "message_start" || nextType == "message_delta" || nextType == "message_stop" {
+							i++ // skip the data line
+							continue
+						}
+						if wsToolIndex >= 0 {
+							if idx, ok := nextEvent["index"].(float64); ok && int(idx) == wsToolIndex {
+								i++ // skip the data line
+								continue
+							}
+						}
+					}
+				}
+				resultBuilder.WriteString(line + "\n")
+				hasContent = true
+			} else {
+				resultBuilder.WriteString(line + "\n")
+				if strings.TrimSpace(line) != "" {
+					hasContent = true
+				}
+			}
+		}
+
+		if hasContent {
+			filtered = append(filtered, []byte(resultBuilder.String()))
+		}
+	}
+
+	return filtered
+}
--- a/internal/translator/kiro/claude/kiro_websearch.go
+++ b/internal/translator/kiro/claude/kiro_websearch.go
@@ -1,11 +1,14 @@
 // Package claude provides web search functionality for Kiro translator.
-// This file implements detection and MCP request/response types for web search.
+// This file implements detection, MCP request/response types, and pure data
+// transformation utilities for web search. SSE event generation, stream analysis,
+// and HTTP I/O logic reside in the executor package (kiro_executor.go).
 package claude

 import (
 	"encoding/json"
 	"fmt"
 	"strings"
+	"sync/atomic"
 	"time"

 	"github.com/google/uuid"
@@ -14,6 +17,26 @@ import (
 	"github.com/tidwall/sjson"
 )

+// cachedToolDescription stores the dynamically-fetched web_search tool description.
+// Written by the executor via SetWebSearchDescription, read by the translator
+// when building the remote_web_search tool for Kiro API requests.
+var cachedToolDescription atomic.Value // stores string
+
+// GetWebSearchDescription returns the cached web_search tool description,
+// or empty string if not yet fetched. Lock-free via atomic.Value.
+func GetWebSearchDescription() string {
+	if v := cachedToolDescription.Load(); v != nil {
+		return v.(string)
+	}
+	return ""
+}
+
+// SetWebSearchDescription stores the dynamically-fetched web_search tool description.
+// Called by the executor after fetching from MCP tools/list.
+func SetWebSearchDescription(desc string) {
+	cachedToolDescription.Store(desc)
+}
+
 // McpRequest represents a JSON-RPC 2.0 request to Kiro MCP API
 type McpRequest struct {
 	ID      string    `json:"id"`
@@ -191,36 +214,11 @@ func CreateMcpRequest(query string) (string, *McpRequest) {
 	return toolUseID, request
 }

-// GenerateMessageID generates a Claude-style message ID
-func GenerateMessageID() string {
-	return "msg_" + strings.ReplaceAll(uuid.New().String(), "-", "")[:24]
-}
-
 // GenerateToolUseID generates a Kiro-style tool use ID (base62-like UUID)
 func GenerateToolUseID() string {
 	return strings.ReplaceAll(uuid.New().String(), "-", "")[:22]
 }

-// ContainsWebSearchTool checks if the request contains a web_search tool (among any tools).
-// Unlike HasWebSearchTool, this detects web_search even in mixed-tool arrays.
-func ContainsWebSearchTool(body []byte) bool {
-	tools := gjson.GetBytes(body, "tools")
-	if !tools.IsArray() {
-		return false
-	}
-
-	for _, tool := range tools.Array() {
-		name := strings.ToLower(tool.Get("name").String())
-		toolType := strings.ToLower(tool.Get("type").String())
-
-		if isWebSearchTool(name, toolType) {
-			return true
-		}
-	}
-
-	return false
-}
-
 // ReplaceWebSearchToolDescription replaces the web_search tool description with
 // a minimal version that allows re-search without the restrictive "do not search
 // non-coding topics" instruction from the original Kiro tools/list response.
@@ -275,48 +273,6 @@ func ReplaceWebSearchToolDescription(body []byte) ([]byte, error) {
 	return result, nil
 }

-// StripWebSearchTool removes web_search tool entries from the request's tools array.
-// If the tools array becomes empty after removal, it is removed entirely.
-func StripWebSearchTool(body []byte) ([]byte, error) {
-	tools := gjson.GetBytes(body, "tools")
-	if !tools.IsArray() {
-		return body, nil
-	}
-
-	var filtered []json.RawMessage
-	for _, tool := range tools.Array() {
-		name := strings.ToLower(tool.Get("name").String())
-		toolType := strings.ToLower(tool.Get("type").String())
-
-		if !isWebSearchTool(name, toolType) {
-			filtered = append(filtered, json.RawMessage(tool.Raw))
-		}
-	}
-
-	var result []byte
-	var err error
-
-	if len(filtered) == 0 {
-		// Remove tools array entirely
-		result, err = sjson.DeleteBytes(body, "tools")
-		if err != nil {
-			return body, fmt.Errorf("failed to delete tools: %w", err)
-		}
-	} else {
-		// Replace with filtered array
-		filteredJSON, marshalErr := json.Marshal(filtered)
-		if marshalErr != nil {
-			return body, fmt.Errorf("failed to marshal filtered tools: %w", marshalErr)
-		}
-		result, err = sjson.SetRawBytes(body, "tools", filteredJSON)
-		if err != nil {
-			return body, fmt.Errorf("failed to set filtered tools: %w", err)
-		}
-	}
-
-	return result, nil
-}
-
 // FormatSearchContextPrompt formats search results as a structured text block
 // for injection into the system prompt.
 func FormatSearchContextPrompt(query string, results *WebSearchResults) string {
@@ -365,7 +321,7 @@ func FormatToolResultText(results *WebSearchResults) string {
 //
 // This produces the exact same GAR request format as the Kiro IDE (HAR captures).
 // IMPORTANT: The web_search tool must remain in the "tools" array for this to work.
-// Use ReplaceWebSearchToolDescription (not StripWebSearchTool) to keep the tool available.
+// Use ReplaceWebSearchToolDescription to keep the tool available with a minimal description.
 func InjectToolResultsClaude(claudePayload []byte, toolUseId, query string, results *WebSearchResults) ([]byte, error) {
 	var payload map[string]interface{}
 	if err := json.Unmarshal(claudePayload, &payload); err != nil {
@@ -432,8 +388,8 @@ Do NOT apologize for bad results without first attempting a re-search.
 		return claudePayload, fmt.Errorf("failed to marshal updated payload: %w", err)
 	}

-	log.Infof("kiro/websearch: injected tool_use+tool_result (toolUseId=%s, query=%s, messages=%d)",
-		toolUseId, query, len(messages))
+	log.Infof("kiro/websearch: injected tool_use+tool_result (toolUseId=%s, messages=%d)",
+		toolUseId, len(messages))

 	return result, nil
 }
@@ -512,658 +468,28 @@ type SearchIndicator struct {
 	Results   *WebSearchResults
 }

-// ══════════════════════════════════════════════════════════════════════════════
-// SSE Event Generation
-// ══════════════════════════════════════════════════════════════════════════════
-
-// SseEvent represents a Server-Sent Event
-type SseEvent struct {
-	Event string
-	Data  interface{}
+// BuildMcpEndpoint constructs the MCP endpoint URL for the given AWS region.
+// Centralizes the URL pattern used by both handleWebSearch and handleWebSearchStream.
+func BuildMcpEndpoint(region string) string {
+	return fmt.Sprintf("https://q.%s.amazonaws.com/mcp", region)
 }

-// ToSSEString converts the event to SSE wire format
-func (e *SseEvent) ToSSEString() string {
-	dataBytes, _ := json.Marshal(e.Data)
-	return fmt.Sprintf("event: %s\ndata: %s\n\n", e.Event, string(dataBytes))
-}
-
-// GenerateWebSearchEvents generates the 11-event SSE sequence for web search.
-// Events: message_start, content_block_start(server_tool_use), content_block_delta(input_json),
-// content_block_stop, content_block_start(web_search_tool_result), content_block_stop,
-// content_block_start(text), content_block_delta(text), content_block_stop, message_delta, message_stop
-func GenerateWebSearchEvents(
-	model string,
-	query string,
-	toolUseID string,
-	searchResults *WebSearchResults,
-	inputTokens int,
-) []SseEvent {
-	events := make([]SseEvent, 0, 15)
-	messageID := GenerateMessageID()
-
-	// 1. message_start
-	events = append(events, SseEvent{
-		Event: "message_start",
-		Data: map[string]interface{}{
-			"type": "message_start",
-			"message": map[string]interface{}{
-				"id":            messageID,
-				"type":          "message",
-				"role":          "assistant",
-				"model":         model,
-				"content":       []interface{}{},
-				"stop_reason":   nil,
-				"stop_sequence": nil,
-				"usage": map[string]interface{}{
-					"input_tokens":                inputTokens,
-					"output_tokens":               0,
-					"cache_creation_input_tokens": 0,
-					"cache_read_input_tokens":     0,
-				},
-			},
-		},
-	})
-
-	// 2. content_block_start (server_tool_use)
-	events = append(events, SseEvent{
-		Event: "content_block_start",
-		Data: map[string]interface{}{
-			"type":  "content_block_start",
-			"index": 0,
-			"content_block": map[string]interface{}{
-				"id":    toolUseID,
-				"type":  "server_tool_use",
-				"name":  "web_search",
-				"input": map[string]interface{}{},
-			},
-		},
-	})
-
-	// 3. content_block_delta (input_json_delta)
-	inputJSON, _ := json.Marshal(map[string]string{"query": query})
-	events = append(events, SseEvent{
-		Event: "content_block_delta",
-		Data: map[string]interface{}{
-			"type":  "content_block_delta",
-			"index": 0,
-			"delta": map[string]interface{}{
-				"type":         "input_json_delta",
-				"partial_json": string(inputJSON),
-			},
-		},
-	})
-
-	// 4. content_block_stop (server_tool_use)
-	events = append(events, SseEvent{
-		Event: "content_block_stop",
-		Data: map[string]interface{}{
-			"type":  "content_block_stop",
-			"index": 0,
-		},
-	})
-
-	// 5. content_block_start (web_search_tool_result)
-	searchContent := make([]map[string]interface{}, 0)
-	if searchResults != nil {
-		for _, r := range searchResults.Results {
-			snippet := ""
-			if r.Snippet != nil {
-				snippet = *r.Snippet
-			}
-			searchContent = append(searchContent, map[string]interface{}{
-				"type":              "web_search_result",
-				"title":             r.Title,
-				"url":               r.URL,
-				"encrypted_content": snippet,
-				"page_age":          nil,
-			})
-		}
-	}
-	events = append(events, SseEvent{
-		Event: "content_block_start",
-		Data: map[string]interface{}{
-			"type":  "content_block_start",
-			"index": 1,
-			"content_block": map[string]interface{}{
-				"type":        "web_search_tool_result",
-				"tool_use_id": toolUseID,
-				"content":     searchContent,
-			},
-		},
-	})
-
-	// 6. content_block_stop (web_search_tool_result)
-	events = append(events, SseEvent{
-		Event: "content_block_stop",
-		Data: map[string]interface{}{
-			"type":  "content_block_stop",
-			"index": 1,
-		},
-	})
-
-	// 7. content_block_start (text)
-	events = append(events, SseEvent{
-		Event: "content_block_start",
-		Data: map[string]interface{}{
-			"type":  "content_block_start",
-			"index": 2,
-			"content_block": map[string]interface{}{
-				"type": "text",
-				"text": "",
-			},
-		},
-	})
-
-	// 8. content_block_delta (text_delta) - generate search summary
-	summary := generateSearchSummary(query, searchResults)
-
-	// Split text into chunks for streaming effect
-	chunkSize := 100
-	runes := []rune(summary)
-	for i := 0; i < len(runes); i += chunkSize {
-		end := i + chunkSize
-		if end > len(runes) {
-			end = len(runes)
-		}
-		chunk := string(runes[i:end])
-		events = append(events, SseEvent{
-			Event: "content_block_delta",
-			Data: map[string]interface{}{
-				"type":  "content_block_delta",
-				"index": 2,
-				"delta": map[string]interface{}{
-					"type": "text_delta",
-					"text": chunk,
-				},
-			},
-		})
-	}
-
-	// 9. content_block_stop (text)
-	events = append(events, SseEvent{
-		Event: "content_block_stop",
-		Data: map[string]interface{}{
-			"type":  "content_block_stop",
-			"index": 2,
-		},
-	})
-
-	// 10. message_delta
-	outputTokens := (len(summary) + 3) / 4 // Simple estimation
-	events = append(events, SseEvent{
-		Event: "message_delta",
-		Data: map[string]interface{}{
-			"type": "message_delta",
-			"delta": map[string]interface{}{
-				"stop_reason":   "end_turn",
-				"stop_sequence": nil,
-			},
-			"usage": map[string]interface{}{
-				"output_tokens": outputTokens,
-			},
-		},
-	})
-
-	// 11. message_stop
-	events = append(events, SseEvent{
-		Event: "message_stop",
-		Data: map[string]interface{}{
-			"type": "message_stop",
-		},
-	})
-
-	return events
-}
-
-// generateSearchSummary generates a text summary of search results
-func generateSearchSummary(query string, results *WebSearchResults) string {
-	var sb strings.Builder
-	sb.WriteString(fmt.Sprintf("Here are the search results for \"%s\":\n\n", query))
-
-	if results != nil && len(results.Results) > 0 {
-		for i, r := range results.Results {
-			sb.WriteString(fmt.Sprintf("%d. **%s**\n", i+1, r.Title))
-			if r.Snippet != nil {
-				snippet := *r.Snippet
-				if len(snippet) > 200 {
-					snippet = snippet[:200] + "..."
-				}
-				sb.WriteString(fmt.Sprintf("   %s\n", snippet))
-			}
-			sb.WriteString(fmt.Sprintf("   Source: %s\n\n", r.URL))
-		}
-	} else {
-		sb.WriteString("No results found.\n")
-	}
-
-	sb.WriteString("\nPlease note that these are web search results and may not be fully accurate or up-to-date.")
-
-	return sb.String()
-}
-
-// GenerateSearchIndicatorEvents generates ONLY the search indicator SSE events
-// (server_tool_use + web_search_tool_result) without text summary or message termination.
-// These events trigger Claude Code's search indicator UI.
-// The caller is responsible for sending message_start before and message_delta/stop after.
-func GenerateSearchIndicatorEvents(
-	query string,
-	toolUseID string,
-	searchResults *WebSearchResults,
-	startIndex int,
-) []SseEvent {
-	events := make([]SseEvent, 0, 4)
-
-	// 1. content_block_start (server_tool_use)
-	events = append(events, SseEvent{
-		Event: "content_block_start",
-		Data: map[string]interface{}{
-			"type":  "content_block_start",
-			"index": startIndex,
-			"content_block": map[string]interface{}{
-				"id":    toolUseID,
-				"type":  "server_tool_use",
-				"name":  "web_search",
-				"input": map[string]interface{}{},
-			},
-		},
-	})
-
-	// 2. content_block_delta (input_json_delta)
-	inputJSON, _ := json.Marshal(map[string]string{"query": query})
-	events = append(events, SseEvent{
-		Event: "content_block_delta",
-		Data: map[string]interface{}{
-			"type":  "content_block_delta",
-			"index": startIndex,
-			"delta": map[string]interface{}{
-				"type":         "input_json_delta",
-				"partial_json": string(inputJSON),
-			},
-		},
-	})
-
-	// 3. content_block_stop (server_tool_use)
-	events = append(events, SseEvent{
-		Event: "content_block_stop",
-		Data: map[string]interface{}{
-			"type":  "content_block_stop",
-			"index": startIndex,
-		},
-	})
-
-	// 4. content_block_start (web_search_tool_result)
-	searchContent := make([]map[string]interface{}, 0)
-	if searchResults != nil {
-		for _, r := range searchResults.Results {
-			snippet := ""
-			if r.Snippet != nil {
-				snippet = *r.Snippet
-			}
-			searchContent = append(searchContent, map[string]interface{}{
-				"type":              "web_search_result",
-				"title":             r.Title,
-				"url":               r.URL,
-				"encrypted_content": snippet,
-				"page_age":          nil,
-			})
-		}
-	}
-	events = append(events, SseEvent{
-		Event: "content_block_start",
-		Data: map[string]interface{}{
-			"type":  "content_block_start",
-			"index": startIndex + 1,
-			"content_block": map[string]interface{}{
-				"type":        "web_search_tool_result",
-				"tool_use_id": toolUseID,
-				"content":     searchContent,
-			},
-		},
-	})
-
-	// 5. content_block_stop (web_search_tool_result)
-	events = append(events, SseEvent{
-		Event: "content_block_stop",
-		Data: map[string]interface{}{
-			"type":  "content_block_stop",
-			"index": startIndex + 1,
-		},
-	})
-
-	return events
-}
-
-// ══════════════════════════════════════════════════════════════════════════════
-// Stream Analysis & Manipulation
-// ══════════════════════════════════════════════════════════════════════════════
-
-// AdjustStreamIndices adjusts content block indices in SSE event data by adding an offset.
-// It also suppresses duplicate message_start events (returns shouldForward=false).
-// This is used to combine search indicator events (indices 0,1) with Kiro model response events.
-//
-// The data parameter is a single SSE "data:" line payload (JSON).
-// Returns: adjusted data, shouldForward (false = skip this event).
-func AdjustStreamIndices(data []byte, offset int) ([]byte, bool) {
-	if len(data) == 0 {
-		return data, true
-	}
-
-	// Quick check: parse the JSON
-	var event map[string]interface{}
-	if err := json.Unmarshal(data, &event); err != nil {
-		// Not valid JSON, pass through
-		return data, true
-	}
-
-	eventType, _ := event["type"].(string)
-
-	// Suppress duplicate message_start events
-	if eventType == "message_start" {
-		return data, false
-	}
-
-	// Adjust index for content_block events
-	switch eventType {
-	case "content_block_start", "content_block_delta", "content_block_stop":
-		if idx, ok := event["index"].(float64); ok {
-			event["index"] = int(idx) + offset
-			adjusted, err := json.Marshal(event)
-			if err != nil {
-				return data, true
-			}
-			return adjusted, true
-		}
-	}
-
-	// Pass through all other events unchanged (message_delta, message_stop, ping, etc.)
-	return data, true
-}
-
-// AdjustSSEChunk processes a raw SSE chunk (potentially containing multiple "event:/data:" pairs)
-// and adjusts content block indices. Suppresses duplicate message_start events.
-// Returns the adjusted chunk and whether it should be forwarded.
-func AdjustSSEChunk(chunk []byte, offset int) ([]byte, bool) {
-	chunkStr := string(chunk)
-
-	// Fast path: if no "data:" prefix, pass through
-	if !strings.Contains(chunkStr, "data: ") {
-		return chunk, true
-	}
-
-	var result strings.Builder
-	hasContent := false
-
-	lines := strings.Split(chunkStr, "\n")
-	for i := 0; i < len(lines); i++ {
-		line := lines[i]
-
-		if strings.HasPrefix(line, "data: ") {
-			dataPayload := strings.TrimPrefix(line, "data: ")
-			dataPayload = strings.TrimSpace(dataPayload)
-
-			if dataPayload == "[DONE]" {
-				result.WriteString(line + "\n")
-				hasContent = true
-				continue
-			}
-
-			adjusted, shouldForward := AdjustStreamIndices([]byte(dataPayload), offset)
-			if !shouldForward {
-				// Skip this event and its preceding "event:" line
-				// Also skip the trailing empty line
-				continue
-			}
-
-			result.WriteString("data: " + string(adjusted) + "\n")
-			hasContent = true
-		} else if strings.HasPrefix(line, "event: ") {
-			// Check if the next data line will be suppressed
-			if i+1 < len(lines) && strings.HasPrefix(lines[i+1], "data: ") {
-				dataPayload := strings.TrimPrefix(lines[i+1], "data: ")
-				dataPayload = strings.TrimSpace(dataPayload)
-
-				var event map[string]interface{}
-				if err := json.Unmarshal([]byte(dataPayload), &event); err == nil {
-					if eventType, ok := event["type"].(string); ok && eventType == "message_start" {
-						// Skip both the event: and data: lines
-						i++ // skip the data: line too
-						continue
-					}
-				}
-			}
-			result.WriteString(line + "\n")
-			hasContent = true
-		} else {
-			result.WriteString(line + "\n")
-			if strings.TrimSpace(line) != "" {
-				hasContent = true
-			}
-		}
-	}
-
-	if !hasContent {
-		return nil, false
-	}
-
-	return []byte(result.String()), true
-}
-
-// BufferedStreamResult contains the analysis of buffered SSE chunks from a Kiro API response.
-type BufferedStreamResult struct {
-	// StopReason is the detected stop_reason from the stream (e.g., "end_turn", "tool_use")
-	StopReason string
-	// WebSearchQuery is the extracted query if the model requested another web_search
-	WebSearchQuery string
-	// WebSearchToolUseId is the tool_use ID from the model's response (needed for toolResults)
-	WebSearchToolUseId string
-	// HasWebSearchToolUse indicates whether the model requested web_search
-	HasWebSearchToolUse bool
-	// WebSearchToolUseIndex is the content_block index of the web_search tool_use
-	WebSearchToolUseIndex int
-}
-
-// AnalyzeBufferedStream scans buffered SSE chunks to detect stop_reason and web_search tool_use.
-// This is used in the search loop to determine if the model wants another search round.
-func AnalyzeBufferedStream(chunks [][]byte) BufferedStreamResult {
-	result := BufferedStreamResult{WebSearchToolUseIndex: -1}
-
-	// Track tool use state across chunks
-	var currentToolName string
-	var currentToolIndex int = -1
-	var toolInputBuilder strings.Builder
-
-	for _, chunk := range chunks {
-		chunkStr := string(chunk)
-		lines := strings.Split(chunkStr, "\n")
-		for _, line := range lines {
-			if !strings.HasPrefix(line, "data: ") {
-				continue
-			}
-			dataPayload := strings.TrimPrefix(line, "data: ")
-			dataPayload = strings.TrimSpace(dataPayload)
-			if dataPayload == "[DONE]" || dataPayload == "" {
-				continue
-			}
-
-			var event map[string]interface{}
-			if err := json.Unmarshal([]byte(dataPayload), &event); err != nil {
-				continue
-			}
-
-			eventType, _ := event["type"].(string)
-
-			switch eventType {
-			case "message_delta":
-				// Extract stop_reason from message_delta
-				if delta, ok := event["delta"].(map[string]interface{}); ok {
-					if sr, ok := delta["stop_reason"].(string); ok && sr != "" {
-						result.StopReason = sr
-					}
-				}
-
-			case "content_block_start":
-				// Detect tool_use content blocks
-				if cb, ok := event["content_block"].(map[string]interface{}); ok {
-					if cbType, ok := cb["type"].(string); ok && cbType == "tool_use" {
-						if name, ok := cb["name"].(string); ok {
-							currentToolName = strings.ToLower(name)
-							if idx, ok := event["index"].(float64); ok {
-								currentToolIndex = int(idx)
-							}
-							// Capture tool use ID for toolResults handshake
-							if id, ok := cb["id"].(string); ok {
-								result.WebSearchToolUseId = id
-							}
-							toolInputBuilder.Reset()
-						}
-					}
-				}
-
-			case "content_block_delta":
-				// Accumulate tool input JSON
-				if currentToolName != "" {
-					if delta, ok := event["delta"].(map[string]interface{}); ok {
-						if deltaType, ok := delta["type"].(string); ok && deltaType == "input_json_delta" {
-							if partial, ok := delta["partial_json"].(string); ok {
-								toolInputBuilder.WriteString(partial)
-							}
-						}
-					}
-				}
-
-			case "content_block_stop":
-				// Finalize tool use detection
-				if currentToolName == "web_search" || currentToolName == "websearch" || currentToolName == "remote_web_search" {
-					result.HasWebSearchToolUse = true
-					result.WebSearchToolUseIndex = currentToolIndex
-					// Extract query from accumulated input JSON
-					inputJSON := toolInputBuilder.String()
-					var input map[string]string
-					if err := json.Unmarshal([]byte(inputJSON), &input); err == nil {
-						if q, ok := input["query"]; ok {
-							result.WebSearchQuery = q
-						}
-					}
-					log.Debugf("kiro/websearch: detected web_search tool_use, query: %s", result.WebSearchQuery)
-				}
-				currentToolName = ""
-				currentToolIndex = -1
-				toolInputBuilder.Reset()
-			}
-		}
-	}
-
-	return result
-}
-
-// FilterChunksForClient processes buffered SSE chunks and removes web_search tool_use
-// content blocks. This prevents the client from seeing "Tool use" prompts for web_search
-// when the proxy is handling the search loop internally.
-// Also suppresses message_start and message_delta/message_stop events since those
-// are managed by the outer handleWebSearchStream.
-func FilterChunksForClient(chunks [][]byte, wsToolIndex int, indexOffset int) [][]byte {
-	var filtered [][]byte
-
-	for _, chunk := range chunks {
-		chunkStr := string(chunk)
-		lines := strings.Split(chunkStr, "\n")
-
-		var resultBuilder strings.Builder
-		hasContent := false
-
-		for i := 0; i < len(lines); i++ {
-			line := lines[i]
-
-			if strings.HasPrefix(line, "data: ") {
-				dataPayload := strings.TrimPrefix(line, "data: ")
-				dataPayload = strings.TrimSpace(dataPayload)
-
-				if dataPayload == "[DONE]" {
-					// Skip [DONE] — the outer loop manages stream termination
-					continue
-				}
-
-				var event map[string]interface{}
-				if err := json.Unmarshal([]byte(dataPayload), &event); err != nil {
-					resultBuilder.WriteString(line + "\n")
-					hasContent = true
-					continue
-				}
-
-				eventType, _ := event["type"].(string)
-
-				// Skip message_start (outer loop sends its own)
-				if eventType == "message_start" {
-					continue
-				}
-
-				// Skip message_delta and message_stop (outer loop manages these)
-				if eventType == "message_delta" || eventType == "message_stop" {
-					continue
-				}
-
-				// Check if this event belongs to the web_search tool_use block
-				if wsToolIndex >= 0 {
-					if idx, ok := event["index"].(float64); ok && int(idx) == wsToolIndex {
-						// Skip events for the web_search tool_use block
-						continue
-					}
-				}
-
-				// Apply index offset for remaining events
-				if indexOffset > 0 {
-					switch eventType {
-					case "content_block_start", "content_block_delta", "content_block_stop":
-						if idx, ok := event["index"].(float64); ok {
-							event["index"] = int(idx) + indexOffset
-							adjusted, err := json.Marshal(event)
-							if err == nil {
-								resultBuilder.WriteString("data: " + string(adjusted) + "\n")
-								hasContent = true
-								continue
-							}
-						}
-					}
-				}
-
-				resultBuilder.WriteString(line + "\n")
-				hasContent = true
-			} else if strings.HasPrefix(line, "event: ") {
-				// Check if the next data line will be suppressed
-				if i+1 < len(lines) && strings.HasPrefix(lines[i+1], "data: ") {
-					nextData := strings.TrimPrefix(lines[i+1], "data: ")
-					nextData = strings.TrimSpace(nextData)
-
-					var nextEvent map[string]interface{}
-					if err := json.Unmarshal([]byte(nextData), &nextEvent); err == nil {
-						nextType, _ := nextEvent["type"].(string)
-						if nextType == "message_start" || nextType == "message_delta" || nextType == "message_stop" {
-							i++ // skip the data line
-							continue
-						}
-						if wsToolIndex >= 0 {
-							if idx, ok := nextEvent["index"].(float64); ok && int(idx) == wsToolIndex {
-								i++ // skip the data line
-								continue
-							}
-						}
-					}
-				}
-				resultBuilder.WriteString(line + "\n")
-				hasContent = true
-			} else {
-				resultBuilder.WriteString(line + "\n")
-				if strings.TrimSpace(line) != "" {
-					hasContent = true
-				}
-			}
-		}
-
-		if hasContent {
-			filtered = append(filtered, []byte(resultBuilder.String()))
-		}
-	}
-
-	return filtered
+// ParseSearchResults extracts WebSearchResults from MCP response
+func ParseSearchResults(response *McpResponse) *WebSearchResults {
+	if response == nil || response.Result == nil || len(response.Result.Content) == 0 {
+		return nil
+	}
+
+	content := response.Result.Content[0]
+	if content.ContentType != "text" {
+		return nil
+	}
+
+	var results WebSearchResults
+	if err := json.Unmarshal([]byte(content.Text), &results); err != nil {
+		log.Warnf("kiro/websearch: failed to parse search results: %v", err)
+		return nil
+	}
+
+	return &results
 }
--- a/internal/translator/kiro/claude/kiro_websearch_handler.go
+++ b/internal/translator/kiro/claude/kiro_websearch_handler.go
@@ -1,270 +0,0 @@
-// Package claude provides web search handler for Kiro translator.
-// This file implements the MCP API call and response handling.
-package claude
-
-import (
-	"bytes"
-	"encoding/json"
-	"fmt"
-	"io"
-	"net/http"
-	"sync"
-	"sync/atomic"
-	"time"
-
-	"github.com/google/uuid"
-	kiroauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/kiro"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
-	log "github.com/sirupsen/logrus"
-)
-
-// Cached web_search tool description fetched from MCP tools/list.
-// Uses atomic.Pointer[sync.Once] for lock-free reads with retry-on-failure:
-// - sync.Once prevents race conditions and deduplicates concurrent calls
-// - On failure, a fresh sync.Once is swapped in to allow retry on next call
-// - On success, sync.Once stays "done" forever — zero overhead for subsequent calls
-var (
-	cachedToolDescription atomic.Value // stores string
-	toolDescOnce          atomic.Pointer[sync.Once]
-	fallbackFpOnce        sync.Once
-	fallbackFp            *kiroauth.Fingerprint
-)
-
-func init() {
-	toolDescOnce.Store(&sync.Once{})
-}
-
-// FetchToolDescription calls MCP tools/list to get the web_search tool description
-// and caches it. Safe to call concurrently — only one goroutine fetches at a time.
-// If the fetch fails, subsequent calls will retry. On success, no further fetches occur.
-// The httpClient parameter allows reusing a shared pooled HTTP client.
-func FetchToolDescription(mcpEndpoint, authToken string, httpClient *http.Client, fp *kiroauth.Fingerprint, authAttrs map[string]string) {
-	toolDescOnce.Load().Do(func() {
-		handler := NewWebSearchHandler(mcpEndpoint, authToken, httpClient, fp, authAttrs)
-		reqBody := []byte(`{"id":"tools_list","jsonrpc":"2.0","method":"tools/list"}`)
-		log.Debugf("kiro/websearch MCP tools/list request: %d bytes", len(reqBody))
-
-		req, err := http.NewRequest("POST", mcpEndpoint, bytes.NewReader(reqBody))
-		if err != nil {
-			log.Warnf("kiro/websearch: failed to create tools/list request: %v", err)
-			toolDescOnce.Store(&sync.Once{}) // allow retry
-			return
-		}
-
-		// Reuse same headers as CallMcpAPI
-		handler.setMcpHeaders(req)
-
-		resp, err := handler.HTTPClient.Do(req)
-		if err != nil {
-			log.Warnf("kiro/websearch: tools/list request failed: %v", err)
-			toolDescOnce.Store(&sync.Once{}) // allow retry
-			return
-		}
-		defer resp.Body.Close()
-
-		body, err := io.ReadAll(resp.Body)
-		if err != nil || resp.StatusCode != http.StatusOK {
-			log.Warnf("kiro/websearch: tools/list returned status %d", resp.StatusCode)
-			toolDescOnce.Store(&sync.Once{}) // allow retry
-			return
-		}
-		log.Debugf("kiro/websearch MCP tools/list response: [%d] %d bytes", resp.StatusCode, len(body))
-
-		// Parse: {"result":{"tools":[{"name":"web_search","description":"..."}]}}
-		var result struct {
-			Result *struct {
-				Tools []struct {
-					Name        string `json:"name"`
-					Description string `json:"description"`
-				} `json:"tools"`
-			} `json:"result"`
-		}
-		if err := json.Unmarshal(body, &result); err != nil || result.Result == nil {
-			log.Warnf("kiro/websearch: failed to parse tools/list response")
-			toolDescOnce.Store(&sync.Once{}) // allow retry
-			return
-		}
-
-		for _, tool := range result.Result.Tools {
-			if tool.Name == "web_search" && tool.Description != "" {
-				cachedToolDescription.Store(tool.Description)
-				log.Infof("kiro/websearch: cached web_search description from tools/list (%d bytes)", len(tool.Description))
-				return // success — sync.Once stays "done", no more fetches
-			}
-		}
-
-		// web_search tool not found in response
-		toolDescOnce.Store(&sync.Once{}) // allow retry
-	})
-}
-
-// GetWebSearchDescription returns the cached web_search tool description,
-// or empty string if not yet fetched. Lock-free via atomic.Value.
-func GetWebSearchDescription() string {
-	if v := cachedToolDescription.Load(); v != nil {
-		return v.(string)
-	}
-	return ""
-}
-
-// WebSearchHandler handles web search requests via Kiro MCP API
-type WebSearchHandler struct {
-	McpEndpoint string
-	HTTPClient  *http.Client
-	AuthToken   string
-	Fingerprint *kiroauth.Fingerprint // optional, for dynamic headers
-	AuthAttrs   map[string]string     // optional, for custom headers from auth.Attributes
-}
-
-// NewWebSearchHandler creates a new WebSearchHandler.
-// If httpClient is nil, a default client with 30s timeout is used.
-// If fingerprint is nil, a random one-off fingerprint is generated.
-// Pass a shared pooled client (e.g. from getKiroPooledHTTPClient) for connection reuse.
-func NewWebSearchHandler(mcpEndpoint, authToken string, httpClient *http.Client, fp *kiroauth.Fingerprint, authAttrs map[string]string) *WebSearchHandler {
-	if httpClient == nil {
-		httpClient = &http.Client{
-			Timeout: 30 * time.Second,
-		}
-	}
-	if fp == nil {
-		// Use a shared fallback fingerprint for callers without token context
-		fallbackFpOnce.Do(func() {
-			mgr := kiroauth.NewFingerprintManager()
-			fallbackFp = mgr.GetFingerprint("mcp-fallback")
-		})
-		fp = fallbackFp
-	}
-	return &WebSearchHandler{
-		McpEndpoint: mcpEndpoint,
-		HTTPClient:  httpClient,
-		AuthToken:   authToken,
-		Fingerprint: fp,
-		AuthAttrs:   authAttrs,
-	}
-}
-
-// setMcpHeaders sets standard MCP API headers on the request,
-// aligned with the GAR request pattern in kiro_executor.go.
-func (h *WebSearchHandler) setMcpHeaders(req *http.Request) {
-	fp := h.Fingerprint
-
-	// 1. Content-Type & Accept (aligned with GAR)
-	req.Header.Set("Content-Type", "application/json")
-	req.Header.Set("Accept", "*/*")
-
-	// 2. Kiro-specific headers (aligned with GAR)
-	req.Header.Set("x-amzn-kiro-agent-mode", "vibe")
-	req.Header.Set("x-amzn-codewhisperer-optout", "true")
-
-	// 3. Dynamic fingerprint headers
-	req.Header.Set("User-Agent", fp.BuildUserAgent())
-	req.Header.Set("X-Amz-User-Agent", fp.BuildAmzUserAgent())
-
-	// 4. AWS SDK identifiers (casing aligned with GAR)
-	req.Header.Set("Amz-Sdk-Request", "attempt=1; max=3")
-	req.Header.Set("Amz-Sdk-Invocation-Id", uuid.New().String())
-
-	// 5. Authentication
-	req.Header.Set("Authorization", "Bearer "+h.AuthToken)
-
-	// 6. Custom headers from auth attributes
-	util.ApplyCustomHeadersFromAttrs(req, h.AuthAttrs)
-}
-
-// mcpMaxRetries is the maximum number of retries for MCP API calls.
-const mcpMaxRetries = 2
-
-// CallMcpAPI calls the Kiro MCP API with the given request.
-// Includes retry logic with exponential backoff for retryable errors,
-// aligned with the GAR request retry pattern.
-func (h *WebSearchHandler) CallMcpAPI(request *McpRequest) (*McpResponse, error) {
-	requestBody, err := json.Marshal(request)
-	if err != nil {
-		return nil, fmt.Errorf("failed to marshal MCP request: %w", err)
-	}
-	log.Debugf("kiro/websearch MCP request → %s (%d bytes)", h.McpEndpoint, len(requestBody))
-
-	var lastErr error
-	for attempt := 0; attempt <= mcpMaxRetries; attempt++ {
-		if attempt > 0 {
-			backoff := time.Duration(1<<attempt) * time.Second
-			if backoff > 10*time.Second {
-				backoff = 10 * time.Second
-			}
-			log.Warnf("kiro/websearch: MCP retry %d/%d after %v (last error: %v)", attempt, mcpMaxRetries, backoff, lastErr)
-			time.Sleep(backoff)
-		}
-
-		req, err := http.NewRequest("POST", h.McpEndpoint, bytes.NewReader(requestBody))
-		if err != nil {
-			return nil, fmt.Errorf("failed to create HTTP request: %w", err)
-		}
-
-		h.setMcpHeaders(req)
-
-		resp, err := h.HTTPClient.Do(req)
-		if err != nil {
-			lastErr = fmt.Errorf("MCP API request failed: %w", err)
-			continue // network error → retry
-		}
-
-		body, err := io.ReadAll(resp.Body)
-		resp.Body.Close()
-		if err != nil {
-			lastErr = fmt.Errorf("failed to read MCP response: %w", err)
-			continue // read error → retry
-		}
-		log.Debugf("kiro/websearch MCP response ← [%d] (%d bytes)", resp.StatusCode, len(body))
-
-		// Retryable HTTP status codes (aligned with GAR: 502, 503, 504)
-		if resp.StatusCode >= 502 && resp.StatusCode <= 504 {
-			lastErr = fmt.Errorf("MCP API returned retryable status %d: %s", resp.StatusCode, string(body))
-			continue
-		}
-
-		if resp.StatusCode != http.StatusOK {
-			return nil, fmt.Errorf("MCP API returned status %d: %s", resp.StatusCode, string(body))
-		}
-
-		var mcpResponse McpResponse
-		if err := json.Unmarshal(body, &mcpResponse); err != nil {
-			return nil, fmt.Errorf("failed to parse MCP response: %w", err)
-		}
-
-		if mcpResponse.Error != nil {
-			code := -1
-			if mcpResponse.Error.Code != nil {
-				code = *mcpResponse.Error.Code
-			}
-			msg := "Unknown error"
-			if mcpResponse.Error.Message != nil {
-				msg = *mcpResponse.Error.Message
-			}
-			return nil, fmt.Errorf("MCP error %d: %s", code, msg)
-		}
-
-		return &mcpResponse, nil
-	}
-
-	return nil, lastErr
-}
-
-// ParseSearchResults extracts WebSearchResults from MCP response
-func ParseSearchResults(response *McpResponse) *WebSearchResults {
-	if response == nil || response.Result == nil || len(response.Result.Content) == 0 {
-		return nil
-	}
-
-	content := response.Result.Content[0]
-	if content.ContentType != "text" {
-		return nil
-	}
-
-	var results WebSearchResults
-	if err := json.Unmarshal([]byte(content.Text), &results); err != nil {
-		log.Warnf("kiro/websearch: failed to parse search results: %v", err)
-		return nil
-	}
-
-	return &results
-}
--- a/internal/translator/kiro/common/message_merge.go
+++ b/internal/translator/kiro/common/message_merge.go
@@ -36,8 +36,14 @@ func MergeAdjacentMessages(messages []gjson.Result) []gjson.Result {
 		if currentRole == lastRole {
 			// Merge content from current message into last message
 			mergedContent := mergeMessageContent(lastMsg, msg)
-			// Create a new merged message JSON
-			mergedMsg := createMergedMessage(lastRole, mergedContent)
+			var mergedToolCalls []interface{}
+			if currentRole == "assistant" {
+				// Preserve assistant tool_calls when adjacent assistant messages are merged.
+				mergedToolCalls = mergeToolCalls(lastMsg.Get("tool_calls"), msg.Get("tool_calls"))
+			}
+
+			// Create a new merged message JSON.
+			mergedMsg := createMergedMessage(lastRole, mergedContent, mergedToolCalls)
 			merged[len(merged)-1] = gjson.Parse(mergedMsg)
 		} else {
 			merged = append(merged, msg)
@@ -121,12 +127,34 @@ func blockToMap(block gjson.Result) map[string]interface{} {
 	return result
 }

-// createMergedMessage creates a JSON string for a merged message
-func createMergedMessage(role string, content string) string {
+// createMergedMessage creates a JSON string for a merged message.
+// toolCalls is optional and only emitted for assistant role.
+func createMergedMessage(role string, content string, toolCalls []interface{}) string {
 	msg := map[string]interface{}{
 		"role":    role,
 		"content": json.RawMessage(content),
 	}
+	if role == "assistant" && len(toolCalls) > 0 {
+		msg["tool_calls"] = toolCalls
+	}
 	result, _ := json.Marshal(msg)
 	return string(result)
-}
+}
+
+// mergeToolCalls combines tool_calls from two assistant messages while preserving order.
+func mergeToolCalls(tc1, tc2 gjson.Result) []interface{} {
+	var merged []interface{}
+
+	if tc1.IsArray() {
+		for _, tc := range tc1.Array() {
+			merged = append(merged, tc.Value())
+		}
+	}
+	if tc2.IsArray() {
+		for _, tc := range tc2.Array() {
+			merged = append(merged, tc.Value())
+		}
+	}
+
+	return merged
+}
--- a/internal/translator/kiro/common/message_merge_test.go
+++ b/internal/translator/kiro/common/message_merge_test.go
@@ -0,0 +1,106 @@
+package common
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func parseMessages(t *testing.T, raw string) []gjson.Result {
+	t.Helper()
+	parsed := gjson.Parse(raw)
+	if !parsed.IsArray() {
+		t.Fatalf("expected JSON array, got: %s", raw)
+	}
+	return parsed.Array()
+}
+
+func TestMergeAdjacentMessages_AssistantMergePreservesToolCalls(t *testing.T) {
+	messages := parseMessages(t, `[
+		{"role":"assistant","content":"part1"},
+		{
+			"role":"assistant",
+			"content":"part2",
+			"tool_calls":[
+				{
+					"id":"call_1",
+					"type":"function",
+					"function":{"name":"Read","arguments":"{}"}
+				}
+			]
+		},
+		{"role":"tool","tool_call_id":"call_1","content":"ok"}
+	]`)
+
+	merged := MergeAdjacentMessages(messages)
+	if len(merged) != 2 {
+		t.Fatalf("expected 2 messages after merge, got %d", len(merged))
+	}
+
+	assistant := merged[0]
+	if assistant.Get("role").String() != "assistant" {
+		t.Fatalf("expected first message role assistant, got %q", assistant.Get("role").String())
+	}
+
+	toolCalls := assistant.Get("tool_calls")
+	if !toolCalls.IsArray() || len(toolCalls.Array()) != 1 {
+		t.Fatalf("expected assistant.tool_calls length 1, got: %s", toolCalls.Raw)
+	}
+	if toolCalls.Array()[0].Get("id").String() != "call_1" {
+		t.Fatalf("expected tool call id call_1, got %q", toolCalls.Array()[0].Get("id").String())
+	}
+
+	contentRaw := assistant.Get("content").Raw
+	if !strings.Contains(contentRaw, "part1") || !strings.Contains(contentRaw, "part2") {
+		t.Fatalf("expected merged content to contain both parts, got: %s", contentRaw)
+	}
+
+	if merged[1].Get("role").String() != "tool" {
+		t.Fatalf("expected second message role tool, got %q", merged[1].Get("role").String())
+	}
+}
+
+func TestMergeAdjacentMessages_AssistantMergeCombinesMultipleToolCalls(t *testing.T) {
+	messages := parseMessages(t, `[
+		{
+			"role":"assistant",
+			"content":"first",
+			"tool_calls":[
+				{"id":"call_1","type":"function","function":{"name":"Read","arguments":"{}"}}
+			]
+		},
+		{
+			"role":"assistant",
+			"content":"second",
+			"tool_calls":[
+				{"id":"call_2","type":"function","function":{"name":"Write","arguments":"{}"}}
+			]
+		}
+	]`)
+
+	merged := MergeAdjacentMessages(messages)
+	if len(merged) != 1 {
+		t.Fatalf("expected 1 message after merge, got %d", len(merged))
+	}
+
+	toolCalls := merged[0].Get("tool_calls").Array()
+	if len(toolCalls) != 2 {
+		t.Fatalf("expected 2 merged tool calls, got %d", len(toolCalls))
+	}
+	if toolCalls[0].Get("id").String() != "call_1" || toolCalls[1].Get("id").String() != "call_2" {
+		t.Fatalf("unexpected merged tool call ids: %q, %q", toolCalls[0].Get("id").String(), toolCalls[1].Get("id").String())
+	}
+}
+
+func TestMergeAdjacentMessages_ToolMessagesRemainUnmerged(t *testing.T) {
+	messages := parseMessages(t, `[
+		{"role":"tool","tool_call_id":"call_1","content":"r1"},
+		{"role":"tool","tool_call_id":"call_2","content":"r2"}
+	]`)
+
+	merged := MergeAdjacentMessages(messages)
+	if len(merged) != 2 {
+		t.Fatalf("expected tool messages to remain separate, got %d", len(merged))
+	}
+}
--- a/internal/tui/app.go
+++ b/internal/tui/app.go
@@ -0,0 +1,542 @@
+package tui
+
+import (
+	"fmt"
+	"io"
+	"os"
+	"strings"
+
+	"github.com/charmbracelet/bubbles/textinput"
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+)
+
+// Tab identifiers
+const (
+	tabDashboard = iota
+	tabConfig
+	tabAuthFiles
+	tabAPIKeys
+	tabOAuth
+	tabUsage
+	tabLogs
+)
+
+// App is the root bubbletea model that contains all tab sub-models.
+type App struct {
+	activeTab int
+	tabs      []string
+
+	standalone  bool
+	logsEnabled bool
+
+	authenticated  bool
+	authInput      textinput.Model
+	authError      string
+	authConnecting bool
+
+	dashboard dashboardModel
+	config    configTabModel
+	auth      authTabModel
+	keys      keysTabModel
+	oauth     oauthTabModel
+	usage     usageTabModel
+	logs      logsTabModel
+
+	client *Client
+
+	width  int
+	height int
+	ready  bool
+
+	// Track which tabs have been initialized (fetched data)
+	initialized [7]bool
+}
+
+type authConnectMsg struct {
+	cfg map[string]any
+	err error
+}
+
+// NewApp creates the root TUI application model.
+func NewApp(port int, secretKey string, hook *LogHook) App {
+	standalone := hook != nil
+	authRequired := !standalone
+	ti := textinput.New()
+	ti.CharLimit = 512
+	ti.EchoMode = textinput.EchoPassword
+	ti.EchoCharacter = '*'
+	ti.SetValue(strings.TrimSpace(secretKey))
+	ti.Focus()
+
+	client := NewClient(port, secretKey)
+	app := App{
+		activeTab:     tabDashboard,
+		standalone:    standalone,
+		logsEnabled:   true,
+		authenticated: !authRequired,
+		authInput:     ti,
+		dashboard:     newDashboardModel(client),
+		config:        newConfigTabModel(client),
+		auth:          newAuthTabModel(client),
+		keys:          newKeysTabModel(client),
+		oauth:         newOAuthTabModel(client),
+		usage:         newUsageTabModel(client),
+		logs:          newLogsTabModel(client, hook),
+		client:        client,
+		initialized: [7]bool{
+			tabDashboard: true,
+			tabLogs:      true,
+		},
+	}
+
+	app.refreshTabs()
+	if authRequired {
+		app.initialized = [7]bool{}
+	}
+	app.setAuthInputPrompt()
+	return app
+}
+
+func (a App) Init() tea.Cmd {
+	if !a.authenticated {
+		return textinput.Blink
+	}
+	cmds := []tea.Cmd{a.dashboard.Init()}
+	if a.logsEnabled {
+		cmds = append(cmds, a.logs.Init())
+	}
+	return tea.Batch(cmds...)
+}
+
+func (a App) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
+	switch msg := msg.(type) {
+	case tea.WindowSizeMsg:
+		a.width = msg.Width
+		a.height = msg.Height
+		a.ready = true
+		if a.width > 0 {
+			a.authInput.Width = a.width - 6
+		}
+		contentH := a.height - 4 // tab bar + status bar
+		if contentH < 1 {
+			contentH = 1
+		}
+		contentW := a.width
+		a.dashboard.SetSize(contentW, contentH)
+		a.config.SetSize(contentW, contentH)
+		a.auth.SetSize(contentW, contentH)
+		a.keys.SetSize(contentW, contentH)
+		a.oauth.SetSize(contentW, contentH)
+		a.usage.SetSize(contentW, contentH)
+		a.logs.SetSize(contentW, contentH)
+		return a, nil
+
+	case authConnectMsg:
+		a.authConnecting = false
+		if msg.err != nil {
+			a.authError = fmt.Sprintf(T("auth_gate_connect_fail"), msg.err.Error())
+			return a, nil
+		}
+		a.authError = ""
+		a.authenticated = true
+		a.logsEnabled = a.standalone || isLogsEnabledFromConfig(msg.cfg)
+		a.refreshTabs()
+		a.initialized = [7]bool{}
+		a.initialized[tabDashboard] = true
+		cmds := []tea.Cmd{a.dashboard.Init()}
+		if a.logsEnabled {
+			a.initialized[tabLogs] = true
+			cmds = append(cmds, a.logs.Init())
+		}
+		return a, tea.Batch(cmds...)
+
+	case configUpdateMsg:
+		var cmdLogs tea.Cmd
+		if !a.standalone && msg.err == nil && msg.path == "logging-to-file" {
+			logsEnabledConfig, okConfig := msg.value.(bool)
+			if okConfig {
+				logsEnabledBefore := a.logsEnabled
+				a.logsEnabled = logsEnabledConfig
+				if logsEnabledBefore != a.logsEnabled {
+					a.refreshTabs()
+				}
+				if !a.logsEnabled {
+					a.initialized[tabLogs] = false
+				}
+				if !logsEnabledBefore && a.logsEnabled {
+					a.initialized[tabLogs] = true
+					cmdLogs = a.logs.Init()
+				}
+			}
+		}
+
+		var cmdConfig tea.Cmd
+		a.config, cmdConfig = a.config.Update(msg)
+		if cmdConfig != nil && cmdLogs != nil {
+			return a, tea.Batch(cmdConfig, cmdLogs)
+		}
+		if cmdConfig != nil {
+			return a, cmdConfig
+		}
+		return a, cmdLogs
+
+	case tea.KeyMsg:
+		if !a.authenticated {
+			switch msg.String() {
+			case "ctrl+c", "q":
+				return a, tea.Quit
+			case "L":
+				ToggleLocale()
+				a.refreshTabs()
+				a.setAuthInputPrompt()
+				return a, nil
+			case "enter":
+				if a.authConnecting {
+					return a, nil
+				}
+				password := strings.TrimSpace(a.authInput.Value())
+				if password == "" {
+					a.authError = T("auth_gate_password_required")
+					return a, nil
+				}
+				a.authError = ""
+				a.authConnecting = true
+				return a, a.connectWithPassword(password)
+			default:
+				var cmd tea.Cmd
+				a.authInput, cmd = a.authInput.Update(msg)
+				return a, cmd
+			}
+		}
+
+		switch msg.String() {
+		case "ctrl+c":
+			return a, tea.Quit
+		case "q":
+			// Only quit if not in logs tab (where 'q' might be useful)
+			if !a.logsEnabled || a.activeTab != tabLogs {
+				return a, tea.Quit
+			}
+		case "L":
+			ToggleLocale()
+			a.refreshTabs()
+			return a.broadcastToAllTabs(localeChangedMsg{})
+		case "tab":
+			if len(a.tabs) == 0 {
+				return a, nil
+			}
+			prevTab := a.activeTab
+			a.activeTab = (a.activeTab + 1) % len(a.tabs)
+			return a, a.initTabIfNeeded(prevTab)
+		case "shift+tab":
+			if len(a.tabs) == 0 {
+				return a, nil
+			}
+			prevTab := a.activeTab
+			a.activeTab = (a.activeTab - 1 + len(a.tabs)) % len(a.tabs)
+			return a, a.initTabIfNeeded(prevTab)
+		}
+	}
+
+	if !a.authenticated {
+		var cmd tea.Cmd
+		a.authInput, cmd = a.authInput.Update(msg)
+		return a, cmd
+	}
+
+	// Route msg to active tab
+	var cmd tea.Cmd
+	switch a.activeTab {
+	case tabDashboard:
+		a.dashboard, cmd = a.dashboard.Update(msg)
+	case tabConfig:
+		a.config, cmd = a.config.Update(msg)
+	case tabAuthFiles:
+		a.auth, cmd = a.auth.Update(msg)
+	case tabAPIKeys:
+		a.keys, cmd = a.keys.Update(msg)
+	case tabOAuth:
+		a.oauth, cmd = a.oauth.Update(msg)
+	case tabUsage:
+		a.usage, cmd = a.usage.Update(msg)
+	case tabLogs:
+		a.logs, cmd = a.logs.Update(msg)
+	}
+
+	// Keep logs polling alive even when logs tab is not active.
+	if a.logsEnabled && a.activeTab != tabLogs {
+		switch msg.(type) {
+		case logsPollMsg, logsTickMsg, logLineMsg:
+			var logCmd tea.Cmd
+			a.logs, logCmd = a.logs.Update(msg)
+			if logCmd != nil {
+				cmd = logCmd
+			}
+		}
+	}
+
+	return a, cmd
+}
+
+// localeChangedMsg is broadcast to all tabs when the user toggles locale.
+type localeChangedMsg struct{}
+
+func (a *App) refreshTabs() {
+	names := TabNames()
+	if a.logsEnabled {
+		a.tabs = names
+	} else {
+		filtered := make([]string, 0, len(names)-1)
+		for idx, name := range names {
+			if idx == tabLogs {
+				continue
+			}
+			filtered = append(filtered, name)
+		}
+		a.tabs = filtered
+	}
+
+	if len(a.tabs) == 0 {
+		a.activeTab = tabDashboard
+		return
+	}
+	if a.activeTab >= len(a.tabs) {
+		a.activeTab = len(a.tabs) - 1
+	}
+}
+
+func (a *App) initTabIfNeeded(_ int) tea.Cmd {
+	if a.initialized[a.activeTab] {
+		return nil
+	}
+	a.initialized[a.activeTab] = true
+	switch a.activeTab {
+	case tabDashboard:
+		return a.dashboard.Init()
+	case tabConfig:
+		return a.config.Init()
+	case tabAuthFiles:
+		return a.auth.Init()
+	case tabAPIKeys:
+		return a.keys.Init()
+	case tabOAuth:
+		return a.oauth.Init()
+	case tabUsage:
+		return a.usage.Init()
+	case tabLogs:
+		if !a.logsEnabled {
+			return nil
+		}
+		return a.logs.Init()
+	}
+	return nil
+}
+
+func (a App) View() string {
+	if !a.authenticated {
+		return a.renderAuthView()
+	}
+
+	if !a.ready {
+		return T("initializing_tui")
+	}
+
+	var sb strings.Builder
+
+	// Tab bar
+	sb.WriteString(a.renderTabBar())
+	sb.WriteString("\n")
+
+	// Content
+	switch a.activeTab {
+	case tabDashboard:
+		sb.WriteString(a.dashboard.View())
+	case tabConfig:
+		sb.WriteString(a.config.View())
+	case tabAuthFiles:
+		sb.WriteString(a.auth.View())
+	case tabAPIKeys:
+		sb.WriteString(a.keys.View())
+	case tabOAuth:
+		sb.WriteString(a.oauth.View())
+	case tabUsage:
+		sb.WriteString(a.usage.View())
+	case tabLogs:
+		if a.logsEnabled {
+			sb.WriteString(a.logs.View())
+		}
+	}
+
+	// Status bar
+	sb.WriteString("\n")
+	sb.WriteString(a.renderStatusBar())
+
+	return sb.String()
+}
+
+func (a App) renderAuthView() string {
+	var sb strings.Builder
+
+	sb.WriteString(titleStyle.Render(T("auth_gate_title")))
+	sb.WriteString("\n")
+	sb.WriteString(helpStyle.Render(T("auth_gate_help")))
+	sb.WriteString("\n\n")
+	if a.authConnecting {
+		sb.WriteString(warningStyle.Render(T("auth_gate_connecting")))
+		sb.WriteString("\n\n")
+	}
+	if strings.TrimSpace(a.authError) != "" {
+		sb.WriteString(errorStyle.Render(a.authError))
+		sb.WriteString("\n\n")
+	}
+	sb.WriteString(a.authInput.View())
+	sb.WriteString("\n")
+	sb.WriteString(helpStyle.Render(T("auth_gate_enter")))
+	return sb.String()
+}
+
+func (a App) renderTabBar() string {
+	var tabs []string
+	for i, name := range a.tabs {
+		if i == a.activeTab {
+			tabs = append(tabs, tabActiveStyle.Render(name))
+		} else {
+			tabs = append(tabs, tabInactiveStyle.Render(name))
+		}
+	}
+	tabBar := lipgloss.JoinHorizontal(lipgloss.Top, tabs...)
+	return tabBarStyle.Width(a.width).Render(tabBar)
+}
+
+func (a App) renderStatusBar() string {
+	left := strings.TrimRight(T("status_left"), " ")
+	right := strings.TrimRight(T("status_right"), " ")
+
+	width := a.width
+	if width < 1 {
+		width = 1
+	}
+
+	// statusBarStyle has left/right padding(1), so content area is width-2.
+	contentWidth := width - 2
+	if contentWidth < 0 {
+		contentWidth = 0
+	}
+
+	if lipgloss.Width(left) > contentWidth {
+		left = fitStringWidth(left, contentWidth)
+		right = ""
+	}
+
+	remaining := contentWidth - lipgloss.Width(left)
+	if remaining < 0 {
+		remaining = 0
+	}
+	if lipgloss.Width(right) > remaining {
+		right = fitStringWidth(right, remaining)
+	}
+
+	gap := contentWidth - lipgloss.Width(left) - lipgloss.Width(right)
+	if gap < 0 {
+		gap = 0
+	}
+	return statusBarStyle.Width(width).Render(left + strings.Repeat(" ", gap) + right)
+}
+
+func fitStringWidth(text string, maxWidth int) string {
+	if maxWidth <= 0 {
+		return ""
+	}
+	if lipgloss.Width(text) <= maxWidth {
+		return text
+	}
+
+	out := ""
+	for _, r := range text {
+		next := out + string(r)
+		if lipgloss.Width(next) > maxWidth {
+			break
+		}
+		out = next
+	}
+	return out
+}
+
+func isLogsEnabledFromConfig(cfg map[string]any) bool {
+	if cfg == nil {
+		return true
+	}
+	value, ok := cfg["logging-to-file"]
+	if !ok {
+		return true
+	}
+	enabled, ok := value.(bool)
+	if !ok {
+		return true
+	}
+	return enabled
+}
+
+func (a *App) setAuthInputPrompt() {
+	if a == nil {
+		return
+	}
+	a.authInput.Prompt = fmt.Sprintf("  %s: ", T("auth_gate_password"))
+}
+
+func (a App) connectWithPassword(password string) tea.Cmd {
+	return func() tea.Msg {
+		a.client.SetSecretKey(password)
+		cfg, errGetConfig := a.client.GetConfig()
+		return authConnectMsg{cfg: cfg, err: errGetConfig}
+	}
+}
+
+// Run starts the TUI application.
+// output specifies where bubbletea renders. If nil, defaults to os.Stdout.
+func Run(port int, secretKey string, hook *LogHook, output io.Writer) error {
+	if output == nil {
+		output = os.Stdout
+	}
+	app := NewApp(port, secretKey, hook)
+	p := tea.NewProgram(app, tea.WithAltScreen(), tea.WithOutput(output))
+	_, err := p.Run()
+	return err
+}
+
+func (a App) broadcastToAllTabs(msg tea.Msg) (tea.Model, tea.Cmd) {
+	var cmds []tea.Cmd
+	var cmd tea.Cmd
+
+	a.dashboard, cmd = a.dashboard.Update(msg)
+	if cmd != nil {
+		cmds = append(cmds, cmd)
+	}
+	a.config, cmd = a.config.Update(msg)
+	if cmd != nil {
+		cmds = append(cmds, cmd)
+	}
+	a.auth, cmd = a.auth.Update(msg)
+	if cmd != nil {
+		cmds = append(cmds, cmd)
+	}
+	a.keys, cmd = a.keys.Update(msg)
+	if cmd != nil {
+		cmds = append(cmds, cmd)
+	}
+	a.oauth, cmd = a.oauth.Update(msg)
+	if cmd != nil {
+		cmds = append(cmds, cmd)
+	}
+	a.usage, cmd = a.usage.Update(msg)
+	if cmd != nil {
+		cmds = append(cmds, cmd)
+	}
+	a.logs, cmd = a.logs.Update(msg)
+	if cmd != nil {
+		cmds = append(cmds, cmd)
+	}
+
+	return a, tea.Batch(cmds...)
+}
--- a/internal/tui/auth_tab.go
+++ b/internal/tui/auth_tab.go
@@ -0,0 +1,456 @@
+package tui
+
+import (
+	"fmt"
+	"strconv"
+	"strings"
+
+	"github.com/charmbracelet/bubbles/textinput"
+	"github.com/charmbracelet/bubbles/viewport"
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+)
+
+// editableField represents an editable field on an auth file.
+type editableField struct {
+	label string
+	key   string // API field key: "prefix", "proxy_url", "priority"
+}
+
+var authEditableFields = []editableField{
+	{label: "Prefix", key: "prefix"},
+	{label: "Proxy URL", key: "proxy_url"},
+	{label: "Priority", key: "priority"},
+}
+
+// authTabModel displays auth credential files with interactive management.
+type authTabModel struct {
+	client   *Client
+	viewport viewport.Model
+	files    []map[string]any
+	err      error
+	width    int
+	height   int
+	ready    bool
+	cursor   int
+	expanded int // -1 = none expanded, >=0 = expanded index
+	confirm  int // -1 = no confirmation, >=0 = confirm delete for index
+	status   string
+
+	// Editing state
+	editing      bool            // true when editing a field
+	editField    int             // index into authEditableFields
+	editInput    textinput.Model // text input for editing
+	editFileName string          // name of file being edited
+}
+
+type authFilesMsg struct {
+	files []map[string]any
+	err   error
+}
+
+type authActionMsg struct {
+	action string // "deleted", "toggled", "updated"
+	err    error
+}
+
+func newAuthTabModel(client *Client) authTabModel {
+	ti := textinput.New()
+	ti.CharLimit = 256
+	return authTabModel{
+		client:    client,
+		expanded:  -1,
+		confirm:   -1,
+		editInput: ti,
+	}
+}
+
+func (m authTabModel) Init() tea.Cmd {
+	return m.fetchFiles
+}
+
+func (m authTabModel) fetchFiles() tea.Msg {
+	files, err := m.client.GetAuthFiles()
+	return authFilesMsg{files: files, err: err}
+}
+
+func (m authTabModel) Update(msg tea.Msg) (authTabModel, tea.Cmd) {
+	switch msg := msg.(type) {
+	case localeChangedMsg:
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+	case authFilesMsg:
+		if msg.err != nil {
+			m.err = msg.err
+		} else {
+			m.err = nil
+			m.files = msg.files
+			if m.cursor >= len(m.files) {
+				m.cursor = max(0, len(m.files)-1)
+			}
+			m.status = ""
+		}
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+
+	case authActionMsg:
+		if msg.err != nil {
+			m.status = errorStyle.Render("✗ " + msg.err.Error())
+		} else {
+			m.status = successStyle.Render("✓ " + msg.action)
+		}
+		m.confirm = -1
+		m.viewport.SetContent(m.renderContent())
+		return m, m.fetchFiles
+
+	case tea.KeyMsg:
+		// ---- Editing mode ----
+		if m.editing {
+			return m.handleEditInput(msg)
+		}
+
+		// ---- Delete confirmation mode ----
+		if m.confirm >= 0 {
+			return m.handleConfirmInput(msg)
+		}
+
+		// ---- Normal mode ----
+		return m.handleNormalInput(msg)
+	}
+
+	var cmd tea.Cmd
+	m.viewport, cmd = m.viewport.Update(msg)
+	return m, cmd
+}
+
+// startEdit activates inline editing for a field on the currently selected auth file.
+func (m *authTabModel) startEdit(fieldIdx int) tea.Cmd {
+	if m.cursor >= len(m.files) {
+		return nil
+	}
+	f := m.files[m.cursor]
+	m.editFileName = getString(f, "name")
+	m.editField = fieldIdx
+	m.editing = true
+
+	// Pre-populate with current value
+	key := authEditableFields[fieldIdx].key
+	currentVal := getAnyString(f, key)
+	m.editInput.SetValue(currentVal)
+	m.editInput.Focus()
+	m.editInput.Prompt = fmt.Sprintf("  %s: ", authEditableFields[fieldIdx].label)
+	m.viewport.SetContent(m.renderContent())
+	return textinput.Blink
+}
+
+func (m *authTabModel) SetSize(w, h int) {
+	m.width = w
+	m.height = h
+	m.editInput.Width = w - 20
+	if !m.ready {
+		m.viewport = viewport.New(w, h)
+		m.viewport.SetContent(m.renderContent())
+		m.ready = true
+	} else {
+		m.viewport.Width = w
+		m.viewport.Height = h
+	}
+}
+
+func (m authTabModel) View() string {
+	if !m.ready {
+		return T("loading")
+	}
+	return m.viewport.View()
+}
+
+func (m authTabModel) renderContent() string {
+	var sb strings.Builder
+
+	sb.WriteString(titleStyle.Render(T("auth_title")))
+	sb.WriteString("\n")
+	sb.WriteString(helpStyle.Render(T("auth_help1")))
+	sb.WriteString("\n")
+	sb.WriteString(helpStyle.Render(T("auth_help2")))
+	sb.WriteString("\n")
+	sb.WriteString(strings.Repeat("─", m.width))
+	sb.WriteString("\n")
+
+	if m.err != nil {
+		sb.WriteString(errorStyle.Render("⚠ Error: " + m.err.Error()))
+		sb.WriteString("\n")
+		return sb.String()
+	}
+
+	if len(m.files) == 0 {
+		sb.WriteString(subtitleStyle.Render(T("no_auth_files")))
+		sb.WriteString("\n")
+		return sb.String()
+	}
+
+	for i, f := range m.files {
+		name := getString(f, "name")
+		channel := getString(f, "channel")
+		email := getString(f, "email")
+		disabled := getBool(f, "disabled")
+
+		statusIcon := successStyle.Render("●")
+		statusText := T("status_active")
+		if disabled {
+			statusIcon = lipgloss.NewStyle().Foreground(colorMuted).Render("○")
+			statusText = T("status_disabled")
+		}
+
+		cursor := "  "
+		rowStyle := lipgloss.NewStyle()
+		if i == m.cursor {
+			cursor = "▸ "
+			rowStyle = lipgloss.NewStyle().Bold(true)
+		}
+
+		displayName := name
+		if len(displayName) > 24 {
+			displayName = displayName[:21] + "..."
+		}
+		displayEmail := email
+		if len(displayEmail) > 28 {
+			displayEmail = displayEmail[:25] + "..."
+		}
+
+		row := fmt.Sprintf("%s%s %-24s %-12s %-28s %s",
+			cursor, statusIcon, displayName, channel, displayEmail, statusText)
+		sb.WriteString(rowStyle.Render(row))
+		sb.WriteString("\n")
+
+		// Delete confirmation
+		if m.confirm == i {
+			sb.WriteString(warningStyle.Render(fmt.Sprintf("    "+T("confirm_delete"), name)))
+			sb.WriteString("\n")
+		}
+
+		// Inline edit input
+		if m.editing && i == m.cursor {
+			sb.WriteString(m.editInput.View())
+			sb.WriteString("\n")
+			sb.WriteString(helpStyle.Render("    " + T("enter_save") + " • " + T("esc_cancel")))
+			sb.WriteString("\n")
+		}
+
+		// Expanded detail view
+		if m.expanded == i {
+			sb.WriteString(m.renderDetail(f))
+		}
+	}
+
+	if m.status != "" {
+		sb.WriteString("\n")
+		sb.WriteString(m.status)
+		sb.WriteString("\n")
+	}
+
+	return sb.String()
+}
+
+func (m authTabModel) renderDetail(f map[string]any) string {
+	var sb strings.Builder
+
+	labelStyle := lipgloss.NewStyle().
+		Foreground(lipgloss.Color("111")).
+		Bold(true)
+	valueStyle := lipgloss.NewStyle().
+		Foreground(lipgloss.Color("252"))
+	editableMarker := lipgloss.NewStyle().
+		Foreground(lipgloss.Color("214")).
+		Render(" ✎")
+
+	sb.WriteString("    ┌─────────────────────────────────────────────\n")
+
+	fields := []struct {
+		label    string
+		key      string
+		editable bool
+	}{
+		{"Name", "name", false},
+		{"Channel", "channel", false},
+		{"Email", "email", false},
+		{"Status", "status", false},
+		{"Status Msg", "status_message", false},
+		{"File Name", "file_name", false},
+		{"Auth Type", "auth_type", false},
+		{"Prefix", "prefix", true},
+		{"Proxy URL", "proxy_url", true},
+		{"Priority", "priority", true},
+		{"Project ID", "project_id", false},
+		{"Disabled", "disabled", false},
+		{"Created", "created_at", false},
+		{"Updated", "updated_at", false},
+	}
+
+	for _, field := range fields {
+		val := getAnyString(f, field.key)
+		if val == "" || val == "<nil>" {
+			if field.editable {
+				val = T("not_set")
+			} else {
+				continue
+			}
+		}
+		editMark := ""
+		if field.editable {
+			editMark = editableMarker
+		}
+		line := fmt.Sprintf("    │ %s %s%s",
+			labelStyle.Render(fmt.Sprintf("%-12s:", field.label)),
+			valueStyle.Render(val),
+			editMark)
+		sb.WriteString(line)
+		sb.WriteString("\n")
+	}
+
+	sb.WriteString("    └─────────────────────────────────────────────\n")
+	return sb.String()
+}
+
+// getAnyString converts any value to its string representation.
+func getAnyString(m map[string]any, key string) string {
+	v, ok := m[key]
+	if !ok || v == nil {
+		return ""
+	}
+	return fmt.Sprintf("%v", v)
+}
+
+func max(a, b int) int {
+	if a > b {
+		return a
+	}
+	return b
+}
+
+func (m authTabModel) handleEditInput(msg tea.KeyMsg) (authTabModel, tea.Cmd) {
+	switch msg.String() {
+	case "enter":
+		value := m.editInput.Value()
+		fieldKey := authEditableFields[m.editField].key
+		fileName := m.editFileName
+		m.editing = false
+		m.editInput.Blur()
+		fields := map[string]any{}
+		if fieldKey == "priority" {
+			p, err := strconv.Atoi(value)
+			if err != nil {
+				return m, func() tea.Msg {
+					return authActionMsg{err: fmt.Errorf("%s: %s", T("invalid_int"), value)}
+				}
+			}
+			fields[fieldKey] = p
+		} else {
+			fields[fieldKey] = value
+		}
+		return m, func() tea.Msg {
+			err := m.client.PatchAuthFileFields(fileName, fields)
+			if err != nil {
+				return authActionMsg{err: err}
+			}
+			return authActionMsg{action: fmt.Sprintf(T("updated_field"), fieldKey, fileName)}
+		}
+	case "esc":
+		m.editing = false
+		m.editInput.Blur()
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+	default:
+		var cmd tea.Cmd
+		m.editInput, cmd = m.editInput.Update(msg)
+		m.viewport.SetContent(m.renderContent())
+		return m, cmd
+	}
+}
+
+func (m authTabModel) handleConfirmInput(msg tea.KeyMsg) (authTabModel, tea.Cmd) {
+	switch msg.String() {
+	case "y", "Y":
+		idx := m.confirm
+		m.confirm = -1
+		if idx < len(m.files) {
+			name := getString(m.files[idx], "name")
+			return m, func() tea.Msg {
+				err := m.client.DeleteAuthFile(name)
+				if err != nil {
+					return authActionMsg{err: err}
+				}
+				return authActionMsg{action: fmt.Sprintf(T("deleted"), name)}
+			}
+		}
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+	case "n", "N", "esc":
+		m.confirm = -1
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+	}
+	return m, nil
+}
+
+func (m authTabModel) handleNormalInput(msg tea.KeyMsg) (authTabModel, tea.Cmd) {
+	switch msg.String() {
+	case "j", "down":
+		if len(m.files) > 0 {
+			m.cursor = (m.cursor + 1) % len(m.files)
+			m.viewport.SetContent(m.renderContent())
+		}
+		return m, nil
+	case "k", "up":
+		if len(m.files) > 0 {
+			m.cursor = (m.cursor - 1 + len(m.files)) % len(m.files)
+			m.viewport.SetContent(m.renderContent())
+		}
+		return m, nil
+	case "enter", " ":
+		if m.expanded == m.cursor {
+			m.expanded = -1
+		} else {
+			m.expanded = m.cursor
+		}
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+	case "d", "D":
+		if m.cursor < len(m.files) {
+			m.confirm = m.cursor
+			m.viewport.SetContent(m.renderContent())
+		}
+		return m, nil
+	case "e", "E":
+		if m.cursor < len(m.files) {
+			f := m.files[m.cursor]
+			name := getString(f, "name")
+			disabled := getBool(f, "disabled")
+			newDisabled := !disabled
+			return m, func() tea.Msg {
+				err := m.client.ToggleAuthFile(name, newDisabled)
+				if err != nil {
+					return authActionMsg{err: err}
+				}
+				action := T("enabled")
+				if newDisabled {
+					action = T("disabled")
+				}
+				return authActionMsg{action: fmt.Sprintf("%s %s", action, name)}
+			}
+		}
+		return m, nil
+	case "1":
+		return m, m.startEdit(0) // prefix
+	case "2":
+		return m, m.startEdit(1) // proxy_url
+	case "3":
+		return m, m.startEdit(2) // priority
+	case "r":
+		m.status = ""
+		return m, m.fetchFiles
+	default:
+		var cmd tea.Cmd
+		m.viewport, cmd = m.viewport.Update(msg)
+		return m, cmd
+	}
+}
--- a/internal/tui/browser.go
+++ b/internal/tui/browser.go
@@ -0,0 +1,20 @@
+package tui
+
+import (
+	"os/exec"
+	"runtime"
+)
+
+// openBrowser opens the specified URL in the user's default browser.
+func openBrowser(url string) error {
+	switch runtime.GOOS {
+	case "darwin":
+		return exec.Command("open", url).Start()
+	case "linux":
+		return exec.Command("xdg-open", url).Start()
+	case "windows":
+		return exec.Command("rundll32", "url.dll,FileProtocolHandler", url).Start()
+	default:
+		return exec.Command("xdg-open", url).Start()
+	}
+}
--- a/internal/tui/client.go
+++ b/internal/tui/client.go
@@ -0,0 +1,400 @@
+package tui
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// Client wraps HTTP calls to the management API.
+type Client struct {
+	baseURL   string
+	secretKey string
+	http      *http.Client
+}
+
+// NewClient creates a new management API client.
+func NewClient(port int, secretKey string) *Client {
+	return &Client{
+		baseURL:   fmt.Sprintf("http://127.0.0.1:%d", port),
+		secretKey: strings.TrimSpace(secretKey),
+		http: &http.Client{
+			Timeout: 10 * time.Second,
+		},
+	}
+}
+
+// SetSecretKey updates management API bearer token used by this client.
+func (c *Client) SetSecretKey(secretKey string) {
+	c.secretKey = strings.TrimSpace(secretKey)
+}
+
+func (c *Client) doRequest(method, path string, body io.Reader) ([]byte, int, error) {
+	url := c.baseURL + path
+	req, err := http.NewRequest(method, url, body)
+	if err != nil {
+		return nil, 0, err
+	}
+	if c.secretKey != "" {
+		req.Header.Set("Authorization", "Bearer "+c.secretKey)
+	}
+	if body != nil {
+		req.Header.Set("Content-Type", "application/json")
+	}
+	resp, err := c.http.Do(req)
+	if err != nil {
+		return nil, 0, err
+	}
+	defer resp.Body.Close()
+	data, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, resp.StatusCode, err
+	}
+	return data, resp.StatusCode, nil
+}
+
+func (c *Client) get(path string) ([]byte, error) {
+	data, code, err := c.doRequest("GET", path, nil)
+	if err != nil {
+		return nil, err
+	}
+	if code >= 400 {
+		return nil, fmt.Errorf("HTTP %d: %s", code, strings.TrimSpace(string(data)))
+	}
+	return data, nil
+}
+
+func (c *Client) put(path string, body io.Reader) ([]byte, error) {
+	data, code, err := c.doRequest("PUT", path, body)
+	if err != nil {
+		return nil, err
+	}
+	if code >= 400 {
+		return nil, fmt.Errorf("HTTP %d: %s", code, strings.TrimSpace(string(data)))
+	}
+	return data, nil
+}
+
+func (c *Client) patch(path string, body io.Reader) ([]byte, error) {
+	data, code, err := c.doRequest("PATCH", path, body)
+	if err != nil {
+		return nil, err
+	}
+	if code >= 400 {
+		return nil, fmt.Errorf("HTTP %d: %s", code, strings.TrimSpace(string(data)))
+	}
+	return data, nil
+}
+
+// getJSON fetches a path and unmarshals JSON into a generic map.
+func (c *Client) getJSON(path string) (map[string]any, error) {
+	data, err := c.get(path)
+	if err != nil {
+		return nil, err
+	}
+	var result map[string]any
+	if err := json.Unmarshal(data, &result); err != nil {
+		return nil, err
+	}
+	return result, nil
+}
+
+// postJSON sends a JSON body via POST and checks for errors.
+func (c *Client) postJSON(path string, body any) error {
+	jsonBody, err := json.Marshal(body)
+	if err != nil {
+		return err
+	}
+	_, code, err := c.doRequest("POST", path, strings.NewReader(string(jsonBody)))
+	if err != nil {
+		return err
+	}
+	if code >= 400 {
+		return fmt.Errorf("HTTP %d", code)
+	}
+	return nil
+}
+
+// GetConfig fetches the parsed config.
+func (c *Client) GetConfig() (map[string]any, error) {
+	return c.getJSON("/v0/management/config")
+}
+
+// GetConfigYAML fetches the raw config.yaml content.
+func (c *Client) GetConfigYAML() (string, error) {
+	data, err := c.get("/v0/management/config.yaml")
+	if err != nil {
+		return "", err
+	}
+	return string(data), nil
+}
+
+// PutConfigYAML uploads new config.yaml content.
+func (c *Client) PutConfigYAML(yamlContent string) error {
+	_, err := c.put("/v0/management/config.yaml", strings.NewReader(yamlContent))
+	return err
+}
+
+// GetUsage fetches usage statistics.
+func (c *Client) GetUsage() (map[string]any, error) {
+	return c.getJSON("/v0/management/usage")
+}
+
+// GetAuthFiles lists auth credential files.
+// API returns {"files": [...]}.
+func (c *Client) GetAuthFiles() ([]map[string]any, error) {
+	wrapper, err := c.getJSON("/v0/management/auth-files")
+	if err != nil {
+		return nil, err
+	}
+	return extractList(wrapper, "files")
+}
+
+// DeleteAuthFile deletes a single auth file by name.
+func (c *Client) DeleteAuthFile(name string) error {
+	query := url.Values{}
+	query.Set("name", name)
+	path := "/v0/management/auth-files?" + query.Encode()
+	_, code, err := c.doRequest("DELETE", path, nil)
+	if err != nil {
+		return err
+	}
+	if code >= 400 {
+		return fmt.Errorf("delete failed (HTTP %d)", code)
+	}
+	return nil
+}
+
+// ToggleAuthFile enables or disables an auth file.
+func (c *Client) ToggleAuthFile(name string, disabled bool) error {
+	body, _ := json.Marshal(map[string]any{"name": name, "disabled": disabled})
+	_, err := c.patch("/v0/management/auth-files/status", strings.NewReader(string(body)))
+	return err
+}
+
+// PatchAuthFileFields updates editable fields on an auth file.
+func (c *Client) PatchAuthFileFields(name string, fields map[string]any) error {
+	fields["name"] = name
+	body, _ := json.Marshal(fields)
+	_, err := c.patch("/v0/management/auth-files/fields", strings.NewReader(string(body)))
+	return err
+}
+
+// GetLogs fetches log lines from the server.
+func (c *Client) GetLogs(after int64, limit int) ([]string, int64, error) {
+	query := url.Values{}
+	if limit > 0 {
+		query.Set("limit", strconv.Itoa(limit))
+	}
+	if after > 0 {
+		query.Set("after", strconv.FormatInt(after, 10))
+	}
+
+	path := "/v0/management/logs"
+	encodedQuery := query.Encode()
+	if encodedQuery != "" {
+		path += "?" + encodedQuery
+	}
+
+	wrapper, err := c.getJSON(path)
+	if err != nil {
+		return nil, after, err
+	}
+
+	lines := []string{}
+	if rawLines, ok := wrapper["lines"]; ok && rawLines != nil {
+		rawJSON, errMarshal := json.Marshal(rawLines)
+		if errMarshal != nil {
+			return nil, after, errMarshal
+		}
+		if errUnmarshal := json.Unmarshal(rawJSON, &lines); errUnmarshal != nil {
+			return nil, after, errUnmarshal
+		}
+	}
+
+	latest := after
+	if rawLatest, ok := wrapper["latest-timestamp"]; ok {
+		switch value := rawLatest.(type) {
+		case float64:
+			latest = int64(value)
+		case json.Number:
+			if parsed, errParse := value.Int64(); errParse == nil {
+				latest = parsed
+			}
+		case int64:
+			latest = value
+		case int:
+			latest = int64(value)
+		}
+	}
+	if latest < after {
+		latest = after
+	}
+
+	return lines, latest, nil
+}
+
+// GetAPIKeys fetches the list of API keys.
+// API returns {"api-keys": [...]}.
+func (c *Client) GetAPIKeys() ([]string, error) {
+	wrapper, err := c.getJSON("/v0/management/api-keys")
+	if err != nil {
+		return nil, err
+	}
+	arr, ok := wrapper["api-keys"]
+	if !ok {
+		return nil, nil
+	}
+	raw, err := json.Marshal(arr)
+	if err != nil {
+		return nil, err
+	}
+	var result []string
+	if err := json.Unmarshal(raw, &result); err != nil {
+		return nil, err
+	}
+	return result, nil
+}
+
+// AddAPIKey adds a new API key by sending old=nil, new=key which appends.
+func (c *Client) AddAPIKey(key string) error {
+	body := map[string]any{"old": nil, "new": key}
+	jsonBody, _ := json.Marshal(body)
+	_, err := c.patch("/v0/management/api-keys", strings.NewReader(string(jsonBody)))
+	return err
+}
+
+// EditAPIKey replaces an API key at the given index.
+func (c *Client) EditAPIKey(index int, newValue string) error {
+	body := map[string]any{"index": index, "value": newValue}
+	jsonBody, _ := json.Marshal(body)
+	_, err := c.patch("/v0/management/api-keys", strings.NewReader(string(jsonBody)))
+	return err
+}
+
+// DeleteAPIKey deletes an API key by index.
+func (c *Client) DeleteAPIKey(index int) error {
+	_, code, err := c.doRequest("DELETE", fmt.Sprintf("/v0/management/api-keys?index=%d", index), nil)
+	if err != nil {
+		return err
+	}
+	if code >= 400 {
+		return fmt.Errorf("delete failed (HTTP %d)", code)
+	}
+	return nil
+}
+
+// GetGeminiKeys fetches Gemini API keys.
+// API returns {"gemini-api-key": [...]}.
+func (c *Client) GetGeminiKeys() ([]map[string]any, error) {
+	return c.getWrappedKeyList("/v0/management/gemini-api-key", "gemini-api-key")
+}
+
+// GetClaudeKeys fetches Claude API keys.
+func (c *Client) GetClaudeKeys() ([]map[string]any, error) {
+	return c.getWrappedKeyList("/v0/management/claude-api-key", "claude-api-key")
+}
+
+// GetCodexKeys fetches Codex API keys.
+func (c *Client) GetCodexKeys() ([]map[string]any, error) {
+	return c.getWrappedKeyList("/v0/management/codex-api-key", "codex-api-key")
+}
+
+// GetVertexKeys fetches Vertex API keys.
+func (c *Client) GetVertexKeys() ([]map[string]any, error) {
+	return c.getWrappedKeyList("/v0/management/vertex-api-key", "vertex-api-key")
+}
+
+// GetOpenAICompat fetches OpenAI compatibility entries.
+func (c *Client) GetOpenAICompat() ([]map[string]any, error) {
+	return c.getWrappedKeyList("/v0/management/openai-compatibility", "openai-compatibility")
+}
+
+// getWrappedKeyList fetches a wrapped list from the API.
+func (c *Client) getWrappedKeyList(path, key string) ([]map[string]any, error) {
+	wrapper, err := c.getJSON(path)
+	if err != nil {
+		return nil, err
+	}
+	return extractList(wrapper, key)
+}
+
+// extractList pulls an array of maps from a wrapper object by key.
+func extractList(wrapper map[string]any, key string) ([]map[string]any, error) {
+	arr, ok := wrapper[key]
+	if !ok || arr == nil {
+		return nil, nil
+	}
+	raw, err := json.Marshal(arr)
+	if err != nil {
+		return nil, err
+	}
+	var result []map[string]any
+	if err := json.Unmarshal(raw, &result); err != nil {
+		return nil, err
+	}
+	return result, nil
+}
+
+// GetDebug fetches the current debug setting.
+func (c *Client) GetDebug() (bool, error) {
+	wrapper, err := c.getJSON("/v0/management/debug")
+	if err != nil {
+		return false, err
+	}
+	if v, ok := wrapper["debug"]; ok {
+		if b, ok := v.(bool); ok {
+			return b, nil
+		}
+	}
+	return false, nil
+}
+
+// GetAuthStatus polls the OAuth session status.
+// Returns status ("wait", "ok", "error") and optional error message.
+func (c *Client) GetAuthStatus(state string) (string, string, error) {
+	query := url.Values{}
+	query.Set("state", state)
+	path := "/v0/management/get-auth-status?" + query.Encode()
+	wrapper, err := c.getJSON(path)
+	if err != nil {
+		return "", "", err
+	}
+	status := getString(wrapper, "status")
+	errMsg := getString(wrapper, "error")
+	return status, errMsg, nil
+}
+
+// ----- Config field update methods -----
+
+// PutBoolField updates a boolean config field.
+func (c *Client) PutBoolField(path string, value bool) error {
+	body, _ := json.Marshal(map[string]any{"value": value})
+	_, err := c.put("/v0/management/"+path, strings.NewReader(string(body)))
+	return err
+}
+
+// PutIntField updates an integer config field.
+func (c *Client) PutIntField(path string, value int) error {
+	body, _ := json.Marshal(map[string]any{"value": value})
+	_, err := c.put("/v0/management/"+path, strings.NewReader(string(body)))
+	return err
+}
+
+// PutStringField updates a string config field.
+func (c *Client) PutStringField(path string, value string) error {
+	body, _ := json.Marshal(map[string]any{"value": value})
+	_, err := c.put("/v0/management/"+path, strings.NewReader(string(body)))
+	return err
+}
+
+// DeleteField sends a DELETE request for a config field.
+func (c *Client) DeleteField(path string) error {
+	_, _, err := c.doRequest("DELETE", "/v0/management/"+path, nil)
+	return err
+}
--- a/internal/tui/config_tab.go
+++ b/internal/tui/config_tab.go
@@ -0,0 +1,413 @@
+package tui
+
+import (
+	"fmt"
+	"strconv"
+	"strings"
+
+	"github.com/charmbracelet/bubbles/textinput"
+	"github.com/charmbracelet/bubbles/viewport"
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+)
+
+// configField represents a single editable config field.
+type configField struct {
+	label    string
+	apiPath  string // management API path (e.g. "debug", "proxy-url")
+	kind     string // "bool", "int", "string", "readonly"
+	value    string // current display value
+	rawValue any    // raw value from API
+}
+
+// configTabModel displays parsed config with interactive editing.
+type configTabModel struct {
+	client    *Client
+	viewport  viewport.Model
+	fields    []configField
+	cursor    int
+	editing   bool
+	textInput textinput.Model
+	err       error
+	message   string // status message (success/error)
+	width     int
+	height    int
+	ready     bool
+}
+
+type configDataMsg struct {
+	config map[string]any
+	err    error
+}
+
+type configUpdateMsg struct {
+	path  string
+	value any
+	err   error
+}
+
+func newConfigTabModel(client *Client) configTabModel {
+	ti := textinput.New()
+	ti.CharLimit = 256
+	return configTabModel{
+		client:    client,
+		textInput: ti,
+	}
+}
+
+func (m configTabModel) Init() tea.Cmd {
+	return m.fetchConfig
+}
+
+func (m configTabModel) fetchConfig() tea.Msg {
+	cfg, err := m.client.GetConfig()
+	return configDataMsg{config: cfg, err: err}
+}
+
+func (m configTabModel) Update(msg tea.Msg) (configTabModel, tea.Cmd) {
+	switch msg := msg.(type) {
+	case localeChangedMsg:
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+	case configDataMsg:
+		if msg.err != nil {
+			m.err = msg.err
+			m.fields = nil
+		} else {
+			m.err = nil
+			m.fields = m.parseConfig(msg.config)
+		}
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+
+	case configUpdateMsg:
+		if msg.err != nil {
+			m.message = errorStyle.Render("✗ " + msg.err.Error())
+		} else {
+			m.message = successStyle.Render(T("updated_ok"))
+		}
+		m.viewport.SetContent(m.renderContent())
+		// Refresh config from server
+		return m, m.fetchConfig
+
+	case tea.KeyMsg:
+		if m.editing {
+			return m.handleEditingKey(msg)
+		}
+		return m.handleNormalKey(msg)
+	}
+
+	var cmd tea.Cmd
+	m.viewport, cmd = m.viewport.Update(msg)
+	return m, cmd
+}
+
+func (m configTabModel) handleNormalKey(msg tea.KeyMsg) (configTabModel, tea.Cmd) {
+	switch msg.String() {
+	case "r":
+		m.message = ""
+		return m, m.fetchConfig
+	case "up", "k":
+		if m.cursor > 0 {
+			m.cursor--
+			m.viewport.SetContent(m.renderContent())
+			// Ensure cursor is visible
+			m.ensureCursorVisible()
+		}
+		return m, nil
+	case "down", "j":
+		if m.cursor < len(m.fields)-1 {
+			m.cursor++
+			m.viewport.SetContent(m.renderContent())
+			m.ensureCursorVisible()
+		}
+		return m, nil
+	case "enter", " ":
+		if m.cursor >= 0 && m.cursor < len(m.fields) {
+			f := m.fields[m.cursor]
+			if f.kind == "readonly" {
+				return m, nil
+			}
+			if f.kind == "bool" {
+				// Toggle directly
+				return m, m.toggleBool(m.cursor)
+			}
+			// Start editing for int/string
+			m.editing = true
+			m.textInput.SetValue(configFieldEditValue(f))
+			m.textInput.Focus()
+			m.viewport.SetContent(m.renderContent())
+			return m, textinput.Blink
+		}
+		return m, nil
+	}
+
+	var cmd tea.Cmd
+	m.viewport, cmd = m.viewport.Update(msg)
+	return m, cmd
+}
+
+func (m configTabModel) handleEditingKey(msg tea.KeyMsg) (configTabModel, tea.Cmd) {
+	switch msg.String() {
+	case "enter":
+		m.editing = false
+		m.textInput.Blur()
+		return m, m.submitEdit(m.cursor, m.textInput.Value())
+	case "esc":
+		m.editing = false
+		m.textInput.Blur()
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+	default:
+		var cmd tea.Cmd
+		m.textInput, cmd = m.textInput.Update(msg)
+		m.viewport.SetContent(m.renderContent())
+		return m, cmd
+	}
+}
+
+func (m configTabModel) toggleBool(idx int) tea.Cmd {
+	return func() tea.Msg {
+		f := m.fields[idx]
+		current := f.value == "true"
+		newValue := !current
+		errPutBool := m.client.PutBoolField(f.apiPath, newValue)
+		return configUpdateMsg{
+			path:  f.apiPath,
+			value: newValue,
+			err:   errPutBool,
+		}
+	}
+}
+
+func (m configTabModel) submitEdit(idx int, newValue string) tea.Cmd {
+	return func() tea.Msg {
+		f := m.fields[idx]
+		var err error
+		var value any
+		switch f.kind {
+		case "int":
+			valueInt, errAtoi := strconv.Atoi(newValue)
+			if errAtoi != nil {
+				return configUpdateMsg{
+					path: f.apiPath,
+					err:  fmt.Errorf("%s: %s", T("invalid_int"), newValue),
+				}
+			}
+			value = valueInt
+			err = m.client.PutIntField(f.apiPath, valueInt)
+		case "string":
+			value = newValue
+			err = m.client.PutStringField(f.apiPath, newValue)
+		}
+		return configUpdateMsg{
+			path:  f.apiPath,
+			value: value,
+			err:   err,
+		}
+	}
+}
+
+func configFieldEditValue(f configField) string {
+	if rawString, ok := f.rawValue.(string); ok {
+		return rawString
+	}
+	return f.value
+}
+
+func (m *configTabModel) SetSize(w, h int) {
+	m.width = w
+	m.height = h
+	if !m.ready {
+		m.viewport = viewport.New(w, h)
+		m.viewport.SetContent(m.renderContent())
+		m.ready = true
+	} else {
+		m.viewport.Width = w
+		m.viewport.Height = h
+	}
+}
+
+func (m *configTabModel) ensureCursorVisible() {
+	// Each field takes ~1 line, header takes ~4 lines
+	targetLine := m.cursor + 5
+	if targetLine < m.viewport.YOffset {
+		m.viewport.SetYOffset(targetLine)
+	}
+	if targetLine >= m.viewport.YOffset+m.viewport.Height {
+		m.viewport.SetYOffset(targetLine - m.viewport.Height + 1)
+	}
+}
+
+func (m configTabModel) View() string {
+	if !m.ready {
+		return T("loading")
+	}
+	return m.viewport.View()
+}
+
+func (m configTabModel) renderContent() string {
+	var sb strings.Builder
+
+	sb.WriteString(titleStyle.Render(T("config_title")))
+	sb.WriteString("\n")
+
+	if m.message != "" {
+		sb.WriteString("  " + m.message)
+		sb.WriteString("\n")
+	}
+
+	sb.WriteString(helpStyle.Render(T("config_help1")))
+	sb.WriteString("\n")
+	sb.WriteString(helpStyle.Render(T("config_help2")))
+	sb.WriteString("\n\n")
+
+	if m.err != nil {
+		sb.WriteString(errorStyle.Render("  ⚠ Error: " + m.err.Error()))
+		return sb.String()
+	}
+
+	if len(m.fields) == 0 {
+		sb.WriteString(subtitleStyle.Render(T("no_config")))
+		return sb.String()
+	}
+
+	currentSection := ""
+	for i, f := range m.fields {
+		// Section headers
+		section := fieldSection(f.apiPath)
+		if section != currentSection {
+			currentSection = section
+			sb.WriteString("\n")
+			sb.WriteString(lipgloss.NewStyle().Bold(true).Foreground(colorHighlight).Render("  ── " + section + " "))
+			sb.WriteString("\n")
+		}
+
+		isSelected := i == m.cursor
+		prefix := "  "
+		if isSelected {
+			prefix = "▸ "
+		}
+
+		labelStr := lipgloss.NewStyle().
+			Foreground(colorInfo).
+			Bold(isSelected).
+			Width(32).
+			Render(f.label)
+
+		var valueStr string
+		if m.editing && isSelected {
+			valueStr = m.textInput.View()
+		} else {
+			switch f.kind {
+			case "bool":
+				if f.value == "true" {
+					valueStr = successStyle.Render("● ON")
+				} else {
+					valueStr = lipgloss.NewStyle().Foreground(colorMuted).Render("○ OFF")
+				}
+			case "readonly":
+				valueStr = lipgloss.NewStyle().Foreground(colorSubtext).Render(f.value)
+			default:
+				valueStr = valueStyle.Render(f.value)
+			}
+		}
+
+		line := prefix + labelStr + "  " + valueStr
+		if isSelected && !m.editing {
+			line = lipgloss.NewStyle().Background(colorSurface).Render(line)
+		}
+		sb.WriteString(line + "\n")
+	}
+
+	return sb.String()
+}
+
+func (m configTabModel) parseConfig(cfg map[string]any) []configField {
+	var fields []configField
+
+	// Server settings
+	fields = append(fields, configField{"Port", "port", "readonly", fmt.Sprintf("%.0f", getFloat(cfg, "port")), nil})
+	fields = append(fields, configField{"Host", "host", "readonly", getString(cfg, "host"), nil})
+	fields = append(fields, configField{"Debug", "debug", "bool", fmt.Sprintf("%v", getBool(cfg, "debug")), nil})
+	fields = append(fields, configField{"Proxy URL", "proxy-url", "string", getString(cfg, "proxy-url"), nil})
+	fields = append(fields, configField{"Request Retry", "request-retry", "int", fmt.Sprintf("%.0f", getFloat(cfg, "request-retry")), nil})
+	fields = append(fields, configField{"Max Retry Interval (s)", "max-retry-interval", "int", fmt.Sprintf("%.0f", getFloat(cfg, "max-retry-interval")), nil})
+	fields = append(fields, configField{"Force Model Prefix", "force-model-prefix", "string", getString(cfg, "force-model-prefix"), nil})
+
+	// Logging
+	fields = append(fields, configField{"Logging to File", "logging-to-file", "bool", fmt.Sprintf("%v", getBool(cfg, "logging-to-file")), nil})
+	fields = append(fields, configField{"Logs Max Total Size (MB)", "logs-max-total-size-mb", "int", fmt.Sprintf("%.0f", getFloat(cfg, "logs-max-total-size-mb")), nil})
+	fields = append(fields, configField{"Error Logs Max Files", "error-logs-max-files", "int", fmt.Sprintf("%.0f", getFloat(cfg, "error-logs-max-files")), nil})
+	fields = append(fields, configField{"Usage Stats Enabled", "usage-statistics-enabled", "bool", fmt.Sprintf("%v", getBool(cfg, "usage-statistics-enabled")), nil})
+	fields = append(fields, configField{"Request Log", "request-log", "bool", fmt.Sprintf("%v", getBool(cfg, "request-log")), nil})
+
+	// Quota exceeded
+	fields = append(fields, configField{"Switch Project on Quota", "quota-exceeded/switch-project", "bool", fmt.Sprintf("%v", getBoolNested(cfg, "quota-exceeded", "switch-project")), nil})
+	fields = append(fields, configField{"Switch Preview Model", "quota-exceeded/switch-preview-model", "bool", fmt.Sprintf("%v", getBoolNested(cfg, "quota-exceeded", "switch-preview-model")), nil})
+
+	// Routing
+	if routing, ok := cfg["routing"].(map[string]any); ok {
+		fields = append(fields, configField{"Routing Strategy", "routing/strategy", "string", getString(routing, "strategy"), nil})
+	} else {
+		fields = append(fields, configField{"Routing Strategy", "routing/strategy", "string", "", nil})
+	}
+
+	// WebSocket auth
+	fields = append(fields, configField{"WebSocket Auth", "ws-auth", "bool", fmt.Sprintf("%v", getBool(cfg, "ws-auth")), nil})
+
+	// AMP settings
+	if amp, ok := cfg["ampcode"].(map[string]any); ok {
+		upstreamURL := getString(amp, "upstream-url")
+		upstreamAPIKey := getString(amp, "upstream-api-key")
+		fields = append(fields, configField{"AMP Upstream URL", "ampcode/upstream-url", "string", upstreamURL, upstreamURL})
+		fields = append(fields, configField{"AMP Upstream API Key", "ampcode/upstream-api-key", "string", maskIfNotEmpty(upstreamAPIKey), upstreamAPIKey})
+		fields = append(fields, configField{"AMP Restrict Mgmt Localhost", "ampcode/restrict-management-to-localhost", "bool", fmt.Sprintf("%v", getBool(amp, "restrict-management-to-localhost")), nil})
+	}
+
+	return fields
+}
+
+func fieldSection(apiPath string) string {
+	if strings.HasPrefix(apiPath, "ampcode/") {
+		return T("section_ampcode")
+	}
+	if strings.HasPrefix(apiPath, "quota-exceeded/") {
+		return T("section_quota")
+	}
+	if strings.HasPrefix(apiPath, "routing/") {
+		return T("section_routing")
+	}
+	switch apiPath {
+	case "port", "host", "debug", "proxy-url", "request-retry", "max-retry-interval", "force-model-prefix":
+		return T("section_server")
+	case "logging-to-file", "logs-max-total-size-mb", "error-logs-max-files", "usage-statistics-enabled", "request-log":
+		return T("section_logging")
+	case "ws-auth":
+		return T("section_websocket")
+	default:
+		return T("section_other")
+	}
+}
+
+func getBoolNested(m map[string]any, keys ...string) bool {
+	current := m
+	for i, key := range keys {
+		if i == len(keys)-1 {
+			return getBool(current, key)
+		}
+		if nested, ok := current[key].(map[string]any); ok {
+			current = nested
+		} else {
+			return false
+		}
+	}
+	return false
+}
+
+func maskIfNotEmpty(s string) string {
+	if s == "" {
+		return T("not_set")
+	}
+	return maskKey(s)
+}
--- a/internal/tui/dashboard.go
+++ b/internal/tui/dashboard.go
@@ -0,0 +1,360 @@
+package tui
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"github.com/charmbracelet/bubbles/viewport"
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+)
+
+// dashboardModel displays server info, stats cards, and config overview.
+type dashboardModel struct {
+	client   *Client
+	viewport viewport.Model
+	content  string
+	err      error
+	width    int
+	height   int
+	ready    bool
+
+	// Cached data for re-rendering on locale change
+	lastConfig    map[string]any
+	lastUsage     map[string]any
+	lastAuthFiles []map[string]any
+	lastAPIKeys   []string
+}
+
+type dashboardDataMsg struct {
+	config    map[string]any
+	usage     map[string]any
+	authFiles []map[string]any
+	apiKeys   []string
+	err       error
+}
+
+func newDashboardModel(client *Client) dashboardModel {
+	return dashboardModel{
+		client: client,
+	}
+}
+
+func (m dashboardModel) Init() tea.Cmd {
+	return m.fetchData
+}
+
+func (m dashboardModel) fetchData() tea.Msg {
+	cfg, cfgErr := m.client.GetConfig()
+	usage, usageErr := m.client.GetUsage()
+	authFiles, authErr := m.client.GetAuthFiles()
+	apiKeys, keysErr := m.client.GetAPIKeys()
+
+	var err error
+	for _, e := range []error{cfgErr, usageErr, authErr, keysErr} {
+		if e != nil {
+			err = e
+			break
+		}
+	}
+	return dashboardDataMsg{config: cfg, usage: usage, authFiles: authFiles, apiKeys: apiKeys, err: err}
+}
+
+func (m dashboardModel) Update(msg tea.Msg) (dashboardModel, tea.Cmd) {
+	switch msg := msg.(type) {
+	case localeChangedMsg:
+		// Re-render immediately with cached data using new locale
+		m.content = m.renderDashboard(m.lastConfig, m.lastUsage, m.lastAuthFiles, m.lastAPIKeys)
+		m.viewport.SetContent(m.content)
+		// Also fetch fresh data in background
+		return m, m.fetchData
+
+	case dashboardDataMsg:
+		if msg.err != nil {
+			m.err = msg.err
+			m.content = errorStyle.Render("⚠ Error: " + msg.err.Error())
+		} else {
+			m.err = nil
+			// Cache data for locale switching
+			m.lastConfig = msg.config
+			m.lastUsage = msg.usage
+			m.lastAuthFiles = msg.authFiles
+			m.lastAPIKeys = msg.apiKeys
+
+			m.content = m.renderDashboard(msg.config, msg.usage, msg.authFiles, msg.apiKeys)
+		}
+		m.viewport.SetContent(m.content)
+		return m, nil
+
+	case tea.KeyMsg:
+		if msg.String() == "r" {
+			return m, m.fetchData
+		}
+		var cmd tea.Cmd
+		m.viewport, cmd = m.viewport.Update(msg)
+		return m, cmd
+	}
+
+	var cmd tea.Cmd
+	m.viewport, cmd = m.viewport.Update(msg)
+	return m, cmd
+}
+
+func (m *dashboardModel) SetSize(w, h int) {
+	m.width = w
+	m.height = h
+	if !m.ready {
+		m.viewport = viewport.New(w, h)
+		m.viewport.SetContent(m.content)
+		m.ready = true
+	} else {
+		m.viewport.Width = w
+		m.viewport.Height = h
+	}
+}
+
+func (m dashboardModel) View() string {
+	if !m.ready {
+		return T("loading")
+	}
+	return m.viewport.View()
+}
+
+func (m dashboardModel) renderDashboard(cfg, usage map[string]any, authFiles []map[string]any, apiKeys []string) string {
+	var sb strings.Builder
+
+	sb.WriteString(titleStyle.Render(T("dashboard_title")))
+	sb.WriteString("\n")
+	sb.WriteString(helpStyle.Render(T("dashboard_help")))
+	sb.WriteString("\n\n")
+
+	// ━━━ Connection Status ━━━
+	connStyle := lipgloss.NewStyle().Bold(true).Foreground(colorSuccess)
+	sb.WriteString(connStyle.Render(T("connected")))
+	sb.WriteString(fmt.Sprintf("  %s", m.client.baseURL))
+	sb.WriteString("\n\n")
+
+	// ━━━ Stats Cards ━━━
+	cardWidth := 25
+	if m.width > 0 {
+		cardWidth = (m.width - 6) / 4
+		if cardWidth < 18 {
+			cardWidth = 18
+		}
+	}
+
+	cardStyle := lipgloss.NewStyle().
+		Border(lipgloss.RoundedBorder()).
+		BorderForeground(lipgloss.Color("240")).
+		Padding(0, 1).
+		Width(cardWidth).
+		Height(2)
+
+	// Card 1: API Keys
+	keyCount := len(apiKeys)
+	card1 := cardStyle.Render(fmt.Sprintf(
+		"%s\n%s",
+		lipgloss.NewStyle().Bold(true).Foreground(lipgloss.Color("111")).Render(fmt.Sprintf("🔑 %d", keyCount)),
+		lipgloss.NewStyle().Foreground(colorMuted).Render(T("mgmt_keys")),
+	))
+
+	// Card 2: Auth Files
+	authCount := len(authFiles)
+	activeAuth := 0
+	for _, f := range authFiles {
+		if !getBool(f, "disabled") {
+			activeAuth++
+		}
+	}
+	card2 := cardStyle.Render(fmt.Sprintf(
+		"%s\n%s",
+		lipgloss.NewStyle().Bold(true).Foreground(lipgloss.Color("76")).Render(fmt.Sprintf("📄 %d", authCount)),
+		lipgloss.NewStyle().Foreground(colorMuted).Render(fmt.Sprintf("%s (%d %s)", T("auth_files_label"), activeAuth, T("active_suffix"))),
+	))
+
+	// Card 3: Total Requests
+	totalReqs := int64(0)
+	successReqs := int64(0)
+	failedReqs := int64(0)
+	totalTokens := int64(0)
+	if usage != nil {
+		if usageMap, ok := usage["usage"].(map[string]any); ok {
+			totalReqs = int64(getFloat(usageMap, "total_requests"))
+			successReqs = int64(getFloat(usageMap, "success_count"))
+			failedReqs = int64(getFloat(usageMap, "failure_count"))
+			totalTokens = int64(getFloat(usageMap, "total_tokens"))
+		}
+	}
+	card3 := cardStyle.Render(fmt.Sprintf(
+		"%s\n%s",
+		lipgloss.NewStyle().Bold(true).Foreground(lipgloss.Color("214")).Render(fmt.Sprintf("📈 %d", totalReqs)),
+		lipgloss.NewStyle().Foreground(colorMuted).Render(fmt.Sprintf("%s (✓%d ✗%d)", T("total_requests"), successReqs, failedReqs)),
+	))
+
+	// Card 4: Total Tokens
+	tokenStr := formatLargeNumber(totalTokens)
+	card4 := cardStyle.Render(fmt.Sprintf(
+		"%s\n%s",
+		lipgloss.NewStyle().Bold(true).Foreground(lipgloss.Color("170")).Render(fmt.Sprintf("🔤 %s", tokenStr)),
+		lipgloss.NewStyle().Foreground(colorMuted).Render(T("total_tokens")),
+	))
+
+	sb.WriteString(lipgloss.JoinHorizontal(lipgloss.Top, card1, " ", card2, " ", card3, " ", card4))
+	sb.WriteString("\n\n")
+
+	// ━━━ Current Config ━━━
+	sb.WriteString(lipgloss.NewStyle().Bold(true).Foreground(colorHighlight).Render(T("current_config")))
+	sb.WriteString("\n")
+	sb.WriteString(strings.Repeat("─", minInt(m.width, 60)))
+	sb.WriteString("\n")
+
+	if cfg != nil {
+		debug := getBool(cfg, "debug")
+		retry := getFloat(cfg, "request-retry")
+		proxyURL := getString(cfg, "proxy-url")
+		loggingToFile := getBool(cfg, "logging-to-file")
+		usageEnabled := true
+		if v, ok := cfg["usage-statistics-enabled"]; ok {
+			if b, ok2 := v.(bool); ok2 {
+				usageEnabled = b
+			}
+		}
+
+		configItems := []struct {
+			label string
+			value string
+		}{
+			{T("debug_mode"), boolEmoji(debug)},
+			{T("usage_stats"), boolEmoji(usageEnabled)},
+			{T("log_to_file"), boolEmoji(loggingToFile)},
+			{T("retry_count"), fmt.Sprintf("%.0f", retry)},
+		}
+		if proxyURL != "" {
+			configItems = append(configItems, struct {
+				label string
+				value string
+			}{T("proxy_url"), proxyURL})
+		}
+
+		// Render config items as a compact row
+		for _, item := range configItems {
+			sb.WriteString(fmt.Sprintf("  %s %s\n",
+				labelStyle.Render(item.label+":"),
+				valueStyle.Render(item.value)))
+		}
+
+		// Routing strategy
+		strategy := "round-robin"
+		if routing, ok := cfg["routing"].(map[string]any); ok {
+			if s := getString(routing, "strategy"); s != "" {
+				strategy = s
+			}
+		}
+		sb.WriteString(fmt.Sprintf("  %s %s\n",
+			labelStyle.Render(T("routing_strategy")+":"),
+			valueStyle.Render(strategy)))
+	}
+
+	sb.WriteString("\n")
+
+	// ━━━ Per-Model Usage ━━━
+	if usage != nil {
+		if usageMap, ok := usage["usage"].(map[string]any); ok {
+			if apis, ok := usageMap["apis"].(map[string]any); ok && len(apis) > 0 {
+				sb.WriteString(lipgloss.NewStyle().Bold(true).Foreground(colorHighlight).Render(T("model_stats")))
+				sb.WriteString("\n")
+				sb.WriteString(strings.Repeat("─", minInt(m.width, 60)))
+				sb.WriteString("\n")
+
+				header := fmt.Sprintf("  %-40s %10s %12s", T("model"), T("requests"), T("tokens"))
+				sb.WriteString(tableHeaderStyle.Render(header))
+				sb.WriteString("\n")
+
+				for _, apiSnap := range apis {
+					if apiMap, ok := apiSnap.(map[string]any); ok {
+						if models, ok := apiMap["models"].(map[string]any); ok {
+							for model, v := range models {
+								if stats, ok := v.(map[string]any); ok {
+									reqs := int64(getFloat(stats, "total_requests"))
+									toks := int64(getFloat(stats, "total_tokens"))
+									row := fmt.Sprintf("  %-40s %10d %12s", truncate(model, 40), reqs, formatLargeNumber(toks))
+									sb.WriteString(tableCellStyle.Render(row))
+									sb.WriteString("\n")
+								}
+							}
+						}
+					}
+				}
+			}
+		}
+	}
+
+	return sb.String()
+}
+
+func formatKV(key, value string) string {
+	return fmt.Sprintf("  %s %s\n", labelStyle.Render(key+":"), valueStyle.Render(value))
+}
+
+func getString(m map[string]any, key string) string {
+	if v, ok := m[key]; ok {
+		if s, ok := v.(string); ok {
+			return s
+		}
+	}
+	return ""
+}
+
+func getFloat(m map[string]any, key string) float64 {
+	if v, ok := m[key]; ok {
+		switch n := v.(type) {
+		case float64:
+			return n
+		case json.Number:
+			f, _ := n.Float64()
+			return f
+		}
+	}
+	return 0
+}
+
+func getBool(m map[string]any, key string) bool {
+	if v, ok := m[key]; ok {
+		if b, ok := v.(bool); ok {
+			return b
+		}
+	}
+	return false
+}
+
+func boolEmoji(b bool) string {
+	if b {
+		return T("bool_yes")
+	}
+	return T("bool_no")
+}
+
+func formatLargeNumber(n int64) string {
+	if n >= 1_000_000 {
+		return fmt.Sprintf("%.1fM", float64(n)/1_000_000)
+	}
+	if n >= 1_000 {
+		return fmt.Sprintf("%.1fK", float64(n)/1_000)
+	}
+	return fmt.Sprintf("%d", n)
+}
+
+func truncate(s string, maxLen int) string {
+	if len(s) > maxLen {
+		return s[:maxLen-3] + "..."
+	}
+	return s
+}
+
+func minInt(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
--- a/internal/tui/i18n.go
+++ b/internal/tui/i18n.go
@@ -0,0 +1,364 @@
+package tui
+
+// i18n provides a simple internationalization system for the TUI.
+// Supported locales: "zh" (Chinese, default), "en" (English).
+
+var currentLocale = "en"
+
+// SetLocale changes the active locale.
+func SetLocale(locale string) {
+	if _, ok := locales[locale]; ok {
+		currentLocale = locale
+	}
+}
+
+// CurrentLocale returns the active locale code.
+func CurrentLocale() string {
+	return currentLocale
+}
+
+// ToggleLocale switches between zh and en.
+func ToggleLocale() {
+	if currentLocale == "zh" {
+		currentLocale = "en"
+	} else {
+		currentLocale = "zh"
+	}
+}
+
+// T returns the translated string for the given key.
+func T(key string) string {
+	if m, ok := locales[currentLocale]; ok {
+		if v, ok := m[key]; ok {
+			return v
+		}
+	}
+	// Fallback to English
+	if m, ok := locales["en"]; ok {
+		if v, ok := m[key]; ok {
+			return v
+		}
+	}
+	return key
+}
+
+var locales = map[string]map[string]string{
+	"zh": zhStrings,
+	"en": enStrings,
+}
+
+// ──────────────────────────────────────────
+// Tab names
+// ──────────────────────────────────────────
+var zhTabNames = []string{"仪表盘", "配置", "认证文件", "API 密钥", "OAuth", "使用统计", "日志"}
+var enTabNames = []string{"Dashboard", "Config", "Auth Files", "API Keys", "OAuth", "Usage", "Logs"}
+
+// TabNames returns tab names in the current locale.
+func TabNames() []string {
+	if currentLocale == "zh" {
+		return zhTabNames
+	}
+	return enTabNames
+}
+
+var zhStrings = map[string]string{
+	// ── Common ──
+	"loading":      "加载中...",
+	"refresh":      "刷新",
+	"save":         "保存",
+	"cancel":       "取消",
+	"confirm":      "确认",
+	"yes":          "是",
+	"no":           "否",
+	"error":        "错误",
+	"success":      "成功",
+	"navigate":     "导航",
+	"scroll":       "滚动",
+	"enter_save":   "Enter: 保存",
+	"esc_cancel":   "Esc: 取消",
+	"enter_submit": "Enter: 提交",
+	"press_r":      "[r] 刷新",
+	"press_scroll": "[↑↓] 滚动",
+	"not_set":      "(未设置)",
+	"error_prefix": "⚠ 错误: ",
+
+	// ── Status bar ──
+	"status_left":                 " CLIProxyAPI 管理终端",
+	"status_right":                "Tab/Shift+Tab: 切换 • L: 语言 • q/Ctrl+C: 退出 ",
+	"initializing_tui":            "正在初始化...",
+	"auth_gate_title":             "🔐 连接管理 API",
+	"auth_gate_help":              " 请输入管理密码并按 Enter 连接",
+	"auth_gate_password":          "密码",
+	"auth_gate_enter":             " Enter: 连接 • q/Ctrl+C: 退出 • L: 语言",
+	"auth_gate_connecting":        "正在连接...",
+	"auth_gate_connect_fail":      "连接失败：%s",
+	"auth_gate_password_required": "请输入密码",
+
+	// ── Dashboard ──
+	"dashboard_title":  "📊 仪表盘",
+	"dashboard_help":   " [r] 刷新 • [↑↓] 滚动",
+	"connected":        "● 已连接",
+	"mgmt_keys":        "管理密钥",
+	"auth_files_label": "认证文件",
+	"active_suffix":    "活跃",
+	"total_requests":   "请求",
+	"success_label":    "成功",
+	"failure_label":    "失败",
+	"total_tokens":     "总 Tokens",
+	"current_config":   "当前配置",
+	"debug_mode":       "启用调试模式",
+	"usage_stats":      "启用使用统计",
+	"log_to_file":      "启用日志记录到文件",
+	"retry_count":      "重试次数",
+	"proxy_url":        "代理 URL",
+	"routing_strategy": "路由策略",
+	"model_stats":      "模型统计",
+	"model":            "模型",
+	"requests":         "请求数",
+	"tokens":           "Tokens",
+	"bool_yes":         "是 ✓",
+	"bool_no":          "否",
+
+	// ── Config ──
+	"config_title":      "⚙ 配置",
+	"config_help1":      "  [↑↓/jk] 导航 • [Enter/Space] 编辑 • [r] 刷新",
+	"config_help2":      "  布尔: Enter 切换 • 文本/数字: Enter 输入, Enter 确认, Esc 取消",
+	"updated_ok":        "✓ 更新成功",
+	"no_config":         "  未加载配置",
+	"invalid_int":       "无效整数",
+	"section_server":    "服务器",
+	"section_logging":   "日志与统计",
+	"section_quota":     "配额超限处理",
+	"section_routing":   "路由",
+	"section_websocket": "WebSocket",
+	"section_ampcode":   "AMP Code",
+	"section_other":     "其他",
+
+	// ── Auth Files ──
+	"auth_title":      "🔑 认证文件",
+	"auth_help1":      " [↑↓/jk] 导航 • [Enter] 展开 • [e] 启用/停用 • [d] 删除 • [r] 刷新",
+	"auth_help2":      " [1] 编辑 prefix • [2] 编辑 proxy_url • [3] 编辑 priority",
+	"no_auth_files":   "  无认证文件",
+	"confirm_delete":  "⚠ 删除 %s? [y/n]",
+	"deleted":         "已删除 %s",
+	"enabled":         "已启用",
+	"disabled":        "已停用",
+	"updated_field":   "已更新 %s 的 %s",
+	"status_active":   "活跃",
+	"status_disabled": "已停用",
+
+	// ── API Keys ──
+	"keys_title":         "🔐 API 密钥",
+	"keys_help":          " [↑↓/jk] 导航 • [a] 添加 • [e] 编辑 • [d] 删除 • [c] 复制 • [r] 刷新",
+	"no_keys":            "  无 API Key，按 [a] 添加",
+	"access_keys":        "Access API Keys",
+	"confirm_delete_key": "⚠ 确认删除 %s? [y/n]",
+	"key_added":          "已添加 API Key",
+	"key_updated":        "已更新 API Key",
+	"key_deleted":        "已删除 API Key",
+	"copied":             "✓ 已复制到剪贴板",
+	"copy_failed":        "✗ 复制失败",
+	"new_key_prompt":     "  New Key: ",
+	"edit_key_prompt":    "  Edit Key: ",
+	"enter_add":          "    Enter: 添加 • Esc: 取消",
+	"enter_save_esc":     "    Enter: 保存 • Esc: 取消",
+
+	// ── OAuth ──
+	"oauth_title":        "🔐 OAuth 登录",
+	"oauth_select":       "  选择提供商并按 [Enter] 开始 OAuth 登录:",
+	"oauth_help":         "  [↑↓/jk] 导航 • [Enter] 登录 • [Esc] 清除状态",
+	"oauth_initiating":   "⏳ 正在初始化 %s 登录...",
+	"oauth_success":      "认证成功! 请刷新 Auth Files 标签查看新凭证。",
+	"oauth_completed":    "认证流程已完成。",
+	"oauth_failed":       "认证失败",
+	"oauth_timeout":      "OAuth 流程超时 (5 分钟)",
+	"oauth_press_esc":    "  按 [Esc] 取消",
+	"oauth_auth_url":     "  授权链接:",
+	"oauth_remote_hint":  "  远程浏览器模式：在浏览器中打开上述链接完成授权后，将回调 URL 粘贴到下方。",
+	"oauth_callback_url": "  回调 URL:",
+	"oauth_press_c":      "  按 [c] 输入回调 URL • [Esc] 返回",
+	"oauth_submitting":   "⏳ 提交回调中...",
+	"oauth_submit_ok":    "✓ 回调已提交，等待处理...",
+	"oauth_submit_fail":  "✗ 提交回调失败",
+	"oauth_waiting":      "  等待认证中...",
+
+	// ── Usage ──
+	"usage_title":         "📈 使用统计",
+	"usage_help":          " [r] 刷新 • [↑↓] 滚动",
+	"usage_no_data":       "  使用数据不可用",
+	"usage_total_reqs":    "总请求数",
+	"usage_total_tokens":  "总 Token 数",
+	"usage_success":       "成功",
+	"usage_failure":       "失败",
+	"usage_total_token_l": "总Token",
+	"usage_rpm":           "RPM",
+	"usage_tpm":           "TPM",
+	"usage_req_by_hour":   "请求趋势 (按小时)",
+	"usage_tok_by_hour":   "Token 使用趋势 (按小时)",
+	"usage_req_by_day":    "请求趋势 (按天)",
+	"usage_api_detail":    "API 详细统计",
+	"usage_input":         "输入",
+	"usage_output":        "输出",
+	"usage_cached":        "缓存",
+	"usage_reasoning":     "思考",
+
+	// ── Logs ──
+	"logs_title":       "📋 日志",
+	"logs_auto_scroll": "● 自动滚动",
+	"logs_paused":      "○ 已暂停",
+	"logs_filter":      "过滤",
+	"logs_lines":       "行数",
+	"logs_help":        " [a] 自动滚动 • [c] 清除 • [1] 全部 [2] info+ [3] warn+ [4] error • [↑↓] 滚动",
+	"logs_waiting":     "  等待日志输出...",
+}
+
+var enStrings = map[string]string{
+	// ── Common ──
+	"loading":      "Loading...",
+	"refresh":      "Refresh",
+	"save":         "Save",
+	"cancel":       "Cancel",
+	"confirm":      "Confirm",
+	"yes":          "Yes",
+	"no":           "No",
+	"error":        "Error",
+	"success":      "Success",
+	"navigate":     "Navigate",
+	"scroll":       "Scroll",
+	"enter_save":   "Enter: Save",
+	"esc_cancel":   "Esc: Cancel",
+	"enter_submit": "Enter: Submit",
+	"press_r":      "[r] Refresh",
+	"press_scroll": "[↑↓] Scroll",
+	"not_set":      "(not set)",
+	"error_prefix": "⚠ Error: ",
+
+	// ── Status bar ──
+	"status_left":                 " CLIProxyAPI Management TUI",
+	"status_right":                "Tab/Shift+Tab: switch • L: lang • q/Ctrl+C: quit ",
+	"initializing_tui":            "Initializing...",
+	"auth_gate_title":             "🔐 Connect Management API",
+	"auth_gate_help":              " Enter management password and press Enter to connect",
+	"auth_gate_password":          "Password",
+	"auth_gate_enter":             " Enter: connect • q/Ctrl+C: quit • L: lang",
+	"auth_gate_connecting":        "Connecting...",
+	"auth_gate_connect_fail":      "Connection failed: %s",
+	"auth_gate_password_required": "password is required",
+
+	// ── Dashboard ──
+	"dashboard_title":  "📊 Dashboard",
+	"dashboard_help":   " [r] Refresh • [↑↓] Scroll",
+	"connected":        "● Connected",
+	"mgmt_keys":        "Mgmt Keys",
+	"auth_files_label": "Auth Files",
+	"active_suffix":    "active",
+	"total_requests":   "Requests",
+	"success_label":    "Success",
+	"failure_label":    "Failed",
+	"total_tokens":     "Total Tokens",
+	"current_config":   "Current Config",
+	"debug_mode":       "Debug Mode",
+	"usage_stats":      "Usage Statistics",
+	"log_to_file":      "Log to File",
+	"retry_count":      "Retry Count",
+	"proxy_url":        "Proxy URL",
+	"routing_strategy": "Routing Strategy",
+	"model_stats":      "Model Stats",
+	"model":            "Model",
+	"requests":         "Requests",
+	"tokens":           "Tokens",
+	"bool_yes":         "Yes ✓",
+	"bool_no":          "No",
+
+	// ── Config ──
+	"config_title":      "⚙ Configuration",
+	"config_help1":      "  [↑↓/jk] Navigate • [Enter/Space] Edit • [r] Refresh",
+	"config_help2":      "  Bool: Enter to toggle • String/Int: Enter to type, Enter to confirm, Esc to cancel",
+	"updated_ok":        "✓ Updated successfully",
+	"no_config":         "  No configuration loaded",
+	"invalid_int":       "invalid integer",
+	"section_server":    "Server",
+	"section_logging":   "Logging & Stats",
+	"section_quota":     "Quota Exceeded Handling",
+	"section_routing":   "Routing",
+	"section_websocket": "WebSocket",
+	"section_ampcode":   "AMP Code",
+	"section_other":     "Other",
+
+	// ── Auth Files ──
+	"auth_title":      "🔑 Auth Files",
+	"auth_help1":      " [↑↓/jk] Navigate • [Enter] Expand • [e] Enable/Disable • [d] Delete • [r] Refresh",
+	"auth_help2":      " [1] Edit prefix • [2] Edit proxy_url • [3] Edit priority",
+	"no_auth_files":   "  No auth files found",
+	"confirm_delete":  "⚠ Delete %s? [y/n]",
+	"deleted":         "Deleted %s",
+	"enabled":         "Enabled",
+	"disabled":        "Disabled",
+	"updated_field":   "Updated %s on %s",
+	"status_active":   "active",
+	"status_disabled": "disabled",
+
+	// ── API Keys ──
+	"keys_title":         "🔐 API Keys",
+	"keys_help":          " [↑↓/jk] Navigate • [a] Add • [e] Edit • [d] Delete • [c] Copy • [r] Refresh",
+	"no_keys":            "  No API Keys. Press [a] to add",
+	"access_keys":        "Access API Keys",
+	"confirm_delete_key": "⚠ Delete %s? [y/n]",
+	"key_added":          "API Key added",
+	"key_updated":        "API Key updated",
+	"key_deleted":        "API Key deleted",
+	"copied":             "✓ Copied to clipboard",
+	"copy_failed":        "✗ Copy failed",
+	"new_key_prompt":     "  New Key: ",
+	"edit_key_prompt":    "  Edit Key: ",
+	"enter_add":          "    Enter: Add • Esc: Cancel",
+	"enter_save_esc":     "    Enter: Save • Esc: Cancel",
+
+	// ── OAuth ──
+	"oauth_title":        "🔐 OAuth Login",
+	"oauth_select":       "  Select a provider and press [Enter] to start OAuth login:",
+	"oauth_help":         "  [↑↓/jk] Navigate • [Enter] Login • [Esc] Clear status",
+	"oauth_initiating":   "⏳ Initiating %s login...",
+	"oauth_success":      "Authentication successful! Refresh Auth Files tab to see the new credential.",
+	"oauth_completed":    "Authentication flow completed.",
+	"oauth_failed":       "Authentication failed",
+	"oauth_timeout":      "OAuth flow timed out (5 minutes)",
+	"oauth_press_esc":    "  Press [Esc] to cancel",
+	"oauth_auth_url":     "  Authorization URL:",
+	"oauth_remote_hint":  "  Remote browser mode: Open the URL above in browser, paste the callback URL below after authorization.",
+	"oauth_callback_url": "  Callback URL:",
+	"oauth_press_c":      "  Press [c] to enter callback URL • [Esc] to go back",
+	"oauth_submitting":   "⏳ Submitting callback...",
+	"oauth_submit_ok":    "✓ Callback submitted, waiting...",
+	"oauth_submit_fail":  "✗ Callback submission failed",
+	"oauth_waiting":      "  Waiting for authentication...",
+
+	// ── Usage ──
+	"usage_title":         "📈 Usage Statistics",
+	"usage_help":          " [r] Refresh • [↑↓] Scroll",
+	"usage_no_data":       "  Usage data not available",
+	"usage_total_reqs":    "Total Requests",
+	"usage_total_tokens":  "Total Tokens",
+	"usage_success":       "Success",
+	"usage_failure":       "Failed",
+	"usage_total_token_l": "Total Tokens",
+	"usage_rpm":           "RPM",
+	"usage_tpm":           "TPM",
+	"usage_req_by_hour":   "Requests by Hour",
+	"usage_tok_by_hour":   "Token Usage by Hour",
+	"usage_req_by_day":    "Requests by Day",
+	"usage_api_detail":    "API Detail Statistics",
+	"usage_input":         "Input",
+	"usage_output":        "Output",
+	"usage_cached":        "Cached",
+	"usage_reasoning":     "Reasoning",
+
+	// ── Logs ──
+	"logs_title":       "📋 Logs",
+	"logs_auto_scroll": "● AUTO-SCROLL",
+	"logs_paused":      "○ PAUSED",
+	"logs_filter":      "Filter",
+	"logs_lines":       "Lines",
+	"logs_help":        " [a] Auto-scroll • [c] Clear • [1] All [2] info+ [3] warn+ [4] error • [↑↓] Scroll",
+	"logs_waiting":     "  Waiting for log output...",
+}
--- a/internal/tui/keys_tab.go
+++ b/internal/tui/keys_tab.go
@@ -0,0 +1,405 @@
+package tui
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/atotto/clipboard"
+	"github.com/charmbracelet/bubbles/textinput"
+	"github.com/charmbracelet/bubbles/viewport"
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+)
+
+// keysTabModel displays and manages API keys.
+type keysTabModel struct {
+	client   *Client
+	viewport viewport.Model
+	keys     []string
+	gemini   []map[string]any
+	claude   []map[string]any
+	codex    []map[string]any
+	vertex   []map[string]any
+	openai   []map[string]any
+	err      error
+	width    int
+	height   int
+	ready    bool
+	cursor   int
+	confirm  int // -1 = no deletion pending
+	status   string
+
+	// Editing / Adding
+	editing   bool
+	adding    bool
+	editIdx   int
+	editInput textinput.Model
+}
+
+type keysDataMsg struct {
+	apiKeys []string
+	gemini  []map[string]any
+	claude  []map[string]any
+	codex   []map[string]any
+	vertex  []map[string]any
+	openai  []map[string]any
+	err     error
+}
+
+type keyActionMsg struct {
+	action string
+	err    error
+}
+
+func newKeysTabModel(client *Client) keysTabModel {
+	ti := textinput.New()
+	ti.CharLimit = 512
+	ti.Prompt = "  Key: "
+	return keysTabModel{
+		client:    client,
+		confirm:   -1,
+		editInput: ti,
+	}
+}
+
+func (m keysTabModel) Init() tea.Cmd {
+	return m.fetchKeys
+}
+
+func (m keysTabModel) fetchKeys() tea.Msg {
+	result := keysDataMsg{}
+	apiKeys, err := m.client.GetAPIKeys()
+	if err != nil {
+		result.err = err
+		return result
+	}
+	result.apiKeys = apiKeys
+	result.gemini, _ = m.client.GetGeminiKeys()
+	result.claude, _ = m.client.GetClaudeKeys()
+	result.codex, _ = m.client.GetCodexKeys()
+	result.vertex, _ = m.client.GetVertexKeys()
+	result.openai, _ = m.client.GetOpenAICompat()
+	return result
+}
+
+func (m keysTabModel) Update(msg tea.Msg) (keysTabModel, tea.Cmd) {
+	switch msg := msg.(type) {
+	case localeChangedMsg:
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+	case keysDataMsg:
+		if msg.err != nil {
+			m.err = msg.err
+		} else {
+			m.err = nil
+			m.keys = msg.apiKeys
+			m.gemini = msg.gemini
+			m.claude = msg.claude
+			m.codex = msg.codex
+			m.vertex = msg.vertex
+			m.openai = msg.openai
+			if m.cursor >= len(m.keys) {
+				m.cursor = max(0, len(m.keys)-1)
+			}
+		}
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+
+	case keyActionMsg:
+		if msg.err != nil {
+			m.status = errorStyle.Render("✗ " + msg.err.Error())
+		} else {
+			m.status = successStyle.Render("✓ " + msg.action)
+		}
+		m.confirm = -1
+		m.viewport.SetContent(m.renderContent())
+		return m, m.fetchKeys
+
+	case tea.KeyMsg:
+		// ---- Editing / Adding mode ----
+		if m.editing || m.adding {
+			switch msg.String() {
+			case "enter":
+				value := strings.TrimSpace(m.editInput.Value())
+				if value == "" {
+					m.editing = false
+					m.adding = false
+					m.editInput.Blur()
+					m.viewport.SetContent(m.renderContent())
+					return m, nil
+				}
+				isAdding := m.adding
+				editIdx := m.editIdx
+				m.editing = false
+				m.adding = false
+				m.editInput.Blur()
+				if isAdding {
+					return m, func() tea.Msg {
+						err := m.client.AddAPIKey(value)
+						if err != nil {
+							return keyActionMsg{err: err}
+						}
+						return keyActionMsg{action: T("key_added")}
+					}
+				}
+				return m, func() tea.Msg {
+					err := m.client.EditAPIKey(editIdx, value)
+					if err != nil {
+						return keyActionMsg{err: err}
+					}
+					return keyActionMsg{action: T("key_updated")}
+				}
+			case "esc":
+				m.editing = false
+				m.adding = false
+				m.editInput.Blur()
+				m.viewport.SetContent(m.renderContent())
+				return m, nil
+			default:
+				var cmd tea.Cmd
+				m.editInput, cmd = m.editInput.Update(msg)
+				m.viewport.SetContent(m.renderContent())
+				return m, cmd
+			}
+		}
+
+		// ---- Delete confirmation ----
+		if m.confirm >= 0 {
+			switch msg.String() {
+			case "y", "Y":
+				idx := m.confirm
+				m.confirm = -1
+				return m, func() tea.Msg {
+					err := m.client.DeleteAPIKey(idx)
+					if err != nil {
+						return keyActionMsg{err: err}
+					}
+					return keyActionMsg{action: T("key_deleted")}
+				}
+			case "n", "N", "esc":
+				m.confirm = -1
+				m.viewport.SetContent(m.renderContent())
+				return m, nil
+			}
+			return m, nil
+		}
+
+		// ---- Normal mode ----
+		switch msg.String() {
+		case "j", "down":
+			if len(m.keys) > 0 {
+				m.cursor = (m.cursor + 1) % len(m.keys)
+				m.viewport.SetContent(m.renderContent())
+			}
+			return m, nil
+		case "k", "up":
+			if len(m.keys) > 0 {
+				m.cursor = (m.cursor - 1 + len(m.keys)) % len(m.keys)
+				m.viewport.SetContent(m.renderContent())
+			}
+			return m, nil
+		case "a":
+			// Add new key
+			m.adding = true
+			m.editing = false
+			m.editInput.SetValue("")
+			m.editInput.Prompt = T("new_key_prompt")
+			m.editInput.Focus()
+			m.viewport.SetContent(m.renderContent())
+			return m, textinput.Blink
+		case "e":
+			// Edit selected key
+			if m.cursor < len(m.keys) {
+				m.editing = true
+				m.adding = false
+				m.editIdx = m.cursor
+				m.editInput.SetValue(m.keys[m.cursor])
+				m.editInput.Prompt = T("edit_key_prompt")
+				m.editInput.Focus()
+				m.viewport.SetContent(m.renderContent())
+				return m, textinput.Blink
+			}
+			return m, nil
+		case "d":
+			// Delete selected key
+			if m.cursor < len(m.keys) {
+				m.confirm = m.cursor
+				m.viewport.SetContent(m.renderContent())
+			}
+			return m, nil
+		case "c":
+			// Copy selected key to clipboard
+			if m.cursor < len(m.keys) {
+				key := m.keys[m.cursor]
+				if err := clipboard.WriteAll(key); err != nil {
+					m.status = errorStyle.Render(T("copy_failed") + ": " + err.Error())
+				} else {
+					m.status = successStyle.Render(T("copied"))
+				}
+				m.viewport.SetContent(m.renderContent())
+			}
+			return m, nil
+		case "r":
+			m.status = ""
+			return m, m.fetchKeys
+		default:
+			var cmd tea.Cmd
+			m.viewport, cmd = m.viewport.Update(msg)
+			return m, cmd
+		}
+	}
+
+	var cmd tea.Cmd
+	m.viewport, cmd = m.viewport.Update(msg)
+	return m, cmd
+}
+
+func (m *keysTabModel) SetSize(w, h int) {
+	m.width = w
+	m.height = h
+	m.editInput.Width = w - 16
+	if !m.ready {
+		m.viewport = viewport.New(w, h)
+		m.viewport.SetContent(m.renderContent())
+		m.ready = true
+	} else {
+		m.viewport.Width = w
+		m.viewport.Height = h
+	}
+}
+
+func (m keysTabModel) View() string {
+	if !m.ready {
+		return T("loading")
+	}
+	return m.viewport.View()
+}
+
+func (m keysTabModel) renderContent() string {
+	var sb strings.Builder
+
+	sb.WriteString(titleStyle.Render(T("keys_title")))
+	sb.WriteString("\n")
+	sb.WriteString(helpStyle.Render(T("keys_help")))
+	sb.WriteString("\n")
+	sb.WriteString(strings.Repeat("─", m.width))
+	sb.WriteString("\n")
+
+	if m.err != nil {
+		sb.WriteString(errorStyle.Render(T("error_prefix") + m.err.Error()))
+		sb.WriteString("\n")
+		return sb.String()
+	}
+
+	// ━━━ Access API Keys (interactive) ━━━
+	sb.WriteString(tableHeaderStyle.Render(fmt.Sprintf("  %s (%d)", T("access_keys"), len(m.keys))))
+	sb.WriteString("\n")
+
+	if len(m.keys) == 0 {
+		sb.WriteString(subtitleStyle.Render(T("no_keys")))
+		sb.WriteString("\n")
+	}
+
+	for i, key := range m.keys {
+		cursor := "  "
+		rowStyle := lipgloss.NewStyle()
+		if i == m.cursor {
+			cursor = "▸ "
+			rowStyle = lipgloss.NewStyle().Bold(true)
+		}
+
+		row := fmt.Sprintf("%s%d. %s", cursor, i+1, maskKey(key))
+		sb.WriteString(rowStyle.Render(row))
+		sb.WriteString("\n")
+
+		// Delete confirmation
+		if m.confirm == i {
+			sb.WriteString(warningStyle.Render(fmt.Sprintf("    "+T("confirm_delete_key"), maskKey(key))))
+			sb.WriteString("\n")
+		}
+
+		// Edit input
+		if m.editing && m.editIdx == i {
+			sb.WriteString(m.editInput.View())
+			sb.WriteString("\n")
+			sb.WriteString(helpStyle.Render(T("enter_save_esc")))
+			sb.WriteString("\n")
+		}
+	}
+
+	// Add input
+	if m.adding {
+		sb.WriteString("\n")
+		sb.WriteString(m.editInput.View())
+		sb.WriteString("\n")
+		sb.WriteString(helpStyle.Render(T("enter_add")))
+		sb.WriteString("\n")
+	}
+
+	sb.WriteString("\n")
+
+	// ━━━ Provider Keys (read-only display) ━━━
+	renderProviderKeys(&sb, "Gemini API Keys", m.gemini)
+	renderProviderKeys(&sb, "Claude API Keys", m.claude)
+	renderProviderKeys(&sb, "Codex API Keys", m.codex)
+	renderProviderKeys(&sb, "Vertex API Keys", m.vertex)
+
+	if len(m.openai) > 0 {
+		renderSection(&sb, "OpenAI Compatibility", len(m.openai))
+		for i, entry := range m.openai {
+			name := getString(entry, "name")
+			baseURL := getString(entry, "base-url")
+			prefix := getString(entry, "prefix")
+			info := name
+			if prefix != "" {
+				info += " (prefix: " + prefix + ")"
+			}
+			if baseURL != "" {
+				info += " → " + baseURL
+			}
+			sb.WriteString(fmt.Sprintf("  %d. %s\n", i+1, info))
+		}
+		sb.WriteString("\n")
+	}
+
+	if m.status != "" {
+		sb.WriteString(m.status)
+		sb.WriteString("\n")
+	}
+
+	return sb.String()
+}
+
+func renderSection(sb *strings.Builder, title string, count int) {
+	header := fmt.Sprintf("%s (%d)", title, count)
+	sb.WriteString(tableHeaderStyle.Render("  " + header))
+	sb.WriteString("\n")
+}
+
+func renderProviderKeys(sb *strings.Builder, title string, keys []map[string]any) {
+	if len(keys) == 0 {
+		return
+	}
+	renderSection(sb, title, len(keys))
+	for i, key := range keys {
+		apiKey := getString(key, "api-key")
+		prefix := getString(key, "prefix")
+		baseURL := getString(key, "base-url")
+		info := maskKey(apiKey)
+		if prefix != "" {
+			info += " (prefix: " + prefix + ")"
+		}
+		if baseURL != "" {
+			info += " → " + baseURL
+		}
+		sb.WriteString(fmt.Sprintf("  %d. %s\n", i+1, info))
+	}
+	sb.WriteString("\n")
+}
+
+func maskKey(key string) string {
+	if len(key) <= 8 {
+		return strings.Repeat("*", len(key))
+	}
+	return key[:4] + strings.Repeat("*", len(key)-8) + key[len(key)-4:]
+}
--- a/internal/tui/loghook.go
+++ b/internal/tui/loghook.go
@@ -0,0 +1,78 @@
+package tui
+
+import (
+	"fmt"
+	"strings"
+	"sync"
+
+	log "github.com/sirupsen/logrus"
+)
+
+// LogHook is a logrus hook that captures log entries and sends them to a channel.
+type LogHook struct {
+	ch        chan string
+	formatter log.Formatter
+	mu        sync.Mutex
+	levels    []log.Level
+}
+
+// NewLogHook creates a new LogHook with a buffered channel of the given size.
+func NewLogHook(bufSize int) *LogHook {
+	return &LogHook{
+		ch:        make(chan string, bufSize),
+		formatter: &log.TextFormatter{DisableColors: true, FullTimestamp: true},
+		levels:    log.AllLevels,
+	}
+}
+
+// SetFormatter sets a custom formatter for the hook.
+func (h *LogHook) SetFormatter(f log.Formatter) {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	h.formatter = f
+}
+
+// Levels returns the log levels this hook should fire on.
+func (h *LogHook) Levels() []log.Level {
+	return h.levels
+}
+
+// Fire is called by logrus when a log entry is fired.
+func (h *LogHook) Fire(entry *log.Entry) error {
+	h.mu.Lock()
+	f := h.formatter
+	h.mu.Unlock()
+
+	var line string
+	if f != nil {
+		b, err := f.Format(entry)
+		if err == nil {
+			line = strings.TrimRight(string(b), "\n\r")
+		} else {
+			line = fmt.Sprintf("[%s] %s", entry.Level, entry.Message)
+		}
+	} else {
+		line = fmt.Sprintf("[%s] %s", entry.Level, entry.Message)
+	}
+
+	// Non-blocking send
+	select {
+	case h.ch <- line:
+	default:
+		// Drop oldest if full
+		select {
+		case <-h.ch:
+		default:
+		}
+		select {
+		case h.ch <- line:
+		default:
+		}
+	}
+	return nil
+}
+
+// Chan returns the channel to read log lines from.
+func (h *LogHook) Chan() <-chan string {
+	return h.ch
+}
--- a/internal/tui/logs_tab.go
+++ b/internal/tui/logs_tab.go
@@ -0,0 +1,261 @@
+package tui
+
+import (
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/charmbracelet/bubbles/viewport"
+	tea "github.com/charmbracelet/bubbletea"
+)
+
+// logsTabModel displays real-time log lines from hook/API source.
+type logsTabModel struct {
+	client     *Client
+	hook       *LogHook
+	viewport   viewport.Model
+	lines      []string
+	maxLines   int
+	autoScroll bool
+	width      int
+	height     int
+	ready      bool
+	filter     string // "", "debug", "info", "warn", "error"
+	after      int64
+	lastErr    error
+}
+
+type logsPollMsg struct {
+	lines  []string
+	latest int64
+	err    error
+}
+
+type logsTickMsg struct{}
+type logLineMsg string
+
+func newLogsTabModel(client *Client, hook *LogHook) logsTabModel {
+	return logsTabModel{
+		client:     client,
+		hook:       hook,
+		maxLines:   5000,
+		autoScroll: true,
+	}
+}
+
+func (m logsTabModel) Init() tea.Cmd {
+	if m.hook != nil {
+		return m.waitForLog
+	}
+	return m.fetchLogs
+}
+
+func (m logsTabModel) fetchLogs() tea.Msg {
+	lines, latest, err := m.client.GetLogs(m.after, 200)
+	return logsPollMsg{
+		lines:  lines,
+		latest: latest,
+		err:    err,
+	}
+}
+
+func (m logsTabModel) waitForNextPoll() tea.Cmd {
+	return tea.Tick(2*time.Second, func(_ time.Time) tea.Msg {
+		return logsTickMsg{}
+	})
+}
+
+func (m logsTabModel) waitForLog() tea.Msg {
+	if m.hook == nil {
+		return nil
+	}
+	line, ok := <-m.hook.Chan()
+	if !ok {
+		return nil
+	}
+	return logLineMsg(line)
+}
+
+func (m logsTabModel) Update(msg tea.Msg) (logsTabModel, tea.Cmd) {
+	switch msg := msg.(type) {
+	case localeChangedMsg:
+		m.viewport.SetContent(m.renderLogs())
+		return m, nil
+	case logsTickMsg:
+		if m.hook != nil {
+			return m, nil
+		}
+		return m, m.fetchLogs
+	case logsPollMsg:
+		if m.hook != nil {
+			return m, nil
+		}
+		if msg.err != nil {
+			m.lastErr = msg.err
+		} else {
+			m.lastErr = nil
+			m.after = msg.latest
+			if len(msg.lines) > 0 {
+				m.lines = append(m.lines, msg.lines...)
+				if len(m.lines) > m.maxLines {
+					m.lines = m.lines[len(m.lines)-m.maxLines:]
+				}
+			}
+		}
+		m.viewport.SetContent(m.renderLogs())
+		if m.autoScroll {
+			m.viewport.GotoBottom()
+		}
+		return m, m.waitForNextPoll()
+	case logLineMsg:
+		m.lines = append(m.lines, string(msg))
+		if len(m.lines) > m.maxLines {
+			m.lines = m.lines[len(m.lines)-m.maxLines:]
+		}
+		m.viewport.SetContent(m.renderLogs())
+		if m.autoScroll {
+			m.viewport.GotoBottom()
+		}
+		return m, m.waitForLog
+
+	case tea.KeyMsg:
+		switch msg.String() {
+		case "a":
+			m.autoScroll = !m.autoScroll
+			if m.autoScroll {
+				m.viewport.GotoBottom()
+			}
+			return m, nil
+		case "c":
+			m.lines = nil
+			m.lastErr = nil
+			m.viewport.SetContent(m.renderLogs())
+			return m, nil
+		case "1":
+			m.filter = ""
+			m.viewport.SetContent(m.renderLogs())
+			return m, nil
+		case "2":
+			m.filter = "info"
+			m.viewport.SetContent(m.renderLogs())
+			return m, nil
+		case "3":
+			m.filter = "warn"
+			m.viewport.SetContent(m.renderLogs())
+			return m, nil
+		case "4":
+			m.filter = "error"
+			m.viewport.SetContent(m.renderLogs())
+			return m, nil
+		default:
+			wasAtBottom := m.viewport.AtBottom()
+			var cmd tea.Cmd
+			m.viewport, cmd = m.viewport.Update(msg)
+			// If user scrolls up, disable auto-scroll
+			if !m.viewport.AtBottom() && wasAtBottom {
+				m.autoScroll = false
+			}
+			// If user scrolls to bottom, re-enable auto-scroll
+			if m.viewport.AtBottom() {
+				m.autoScroll = true
+			}
+			return m, cmd
+		}
+	}
+
+	var cmd tea.Cmd
+	m.viewport, cmd = m.viewport.Update(msg)
+	return m, cmd
+}
+
+func (m *logsTabModel) SetSize(w, h int) {
+	m.width = w
+	m.height = h
+	if !m.ready {
+		m.viewport = viewport.New(w, h)
+		m.viewport.SetContent(m.renderLogs())
+		m.ready = true
+	} else {
+		m.viewport.Width = w
+		m.viewport.Height = h
+	}
+}
+
+func (m logsTabModel) View() string {
+	if !m.ready {
+		return T("loading")
+	}
+	return m.viewport.View()
+}
+
+func (m logsTabModel) renderLogs() string {
+	var sb strings.Builder
+
+	scrollStatus := successStyle.Render(T("logs_auto_scroll"))
+	if !m.autoScroll {
+		scrollStatus = warningStyle.Render(T("logs_paused"))
+	}
+	filterLabel := "ALL"
+	if m.filter != "" {
+		filterLabel = strings.ToUpper(m.filter) + "+"
+	}
+
+	header := fmt.Sprintf(" %s  %s  %s: %s  %s: %d",
+		T("logs_title"), scrollStatus, T("logs_filter"), filterLabel, T("logs_lines"), len(m.lines))
+	sb.WriteString(titleStyle.Render(header))
+	sb.WriteString("\n")
+	sb.WriteString(helpStyle.Render(T("logs_help")))
+	sb.WriteString("\n")
+	sb.WriteString(strings.Repeat("─", m.width))
+	sb.WriteString("\n")
+
+	if m.lastErr != nil {
+		sb.WriteString(errorStyle.Render("⚠ Error: " + m.lastErr.Error()))
+		sb.WriteString("\n")
+	}
+
+	if len(m.lines) == 0 {
+		sb.WriteString(subtitleStyle.Render(T("logs_waiting")))
+		return sb.String()
+	}
+
+	for _, line := range m.lines {
+		if m.filter != "" && !m.matchLevel(line) {
+			continue
+		}
+		styled := m.styleLine(line)
+		sb.WriteString(styled)
+		sb.WriteString("\n")
+	}
+
+	return sb.String()
+}
+
+func (m logsTabModel) matchLevel(line string) bool {
+	switch m.filter {
+	case "error":
+		return strings.Contains(line, "[error]") || strings.Contains(line, "[fatal]") || strings.Contains(line, "[panic]")
+	case "warn":
+		return strings.Contains(line, "[warn") || strings.Contains(line, "[error]") || strings.Contains(line, "[fatal]")
+	case "info":
+		return !strings.Contains(line, "[debug]")
+	default:
+		return true
+	}
+}
+
+func (m logsTabModel) styleLine(line string) string {
+	if strings.Contains(line, "[error]") || strings.Contains(line, "[fatal]") {
+		return logErrorStyle.Render(line)
+	}
+	if strings.Contains(line, "[warn") {
+		return logWarnStyle.Render(line)
+	}
+	if strings.Contains(line, "[info") {
+		return logInfoStyle.Render(line)
+	}
+	if strings.Contains(line, "[debug]") {
+		return logDebugStyle.Render(line)
+	}
+	return line
+}
--- a/internal/tui/oauth_tab.go
+++ b/internal/tui/oauth_tab.go
@@ -0,0 +1,473 @@
+package tui
+
+import (
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/charmbracelet/bubbles/textinput"
+	"github.com/charmbracelet/bubbles/viewport"
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+)
+
+// oauthProvider represents an OAuth provider option.
+type oauthProvider struct {
+	name    string
+	apiPath string // management API path
+	emoji   string
+}
+
+var oauthProviders = []oauthProvider{
+	{"Gemini CLI", "gemini-cli-auth-url", "🟦"},
+	{"Claude (Anthropic)", "anthropic-auth-url", "🟧"},
+	{"Codex (OpenAI)", "codex-auth-url", "🟩"},
+	{"Antigravity", "antigravity-auth-url", "🟪"},
+	{"Qwen", "qwen-auth-url", "🟨"},
+	{"Kimi", "kimi-auth-url", "🟫"},
+	{"IFlow", "iflow-auth-url", "⬜"},
+}
+
+// oauthTabModel handles OAuth login flows.
+type oauthTabModel struct {
+	client   *Client
+	viewport viewport.Model
+	cursor   int
+	state    oauthState
+	message  string
+	err      error
+	width    int
+	height   int
+	ready    bool
+
+	// Remote browser mode
+	authURL       string // auth URL to display
+	authState     string // OAuth state parameter
+	providerName  string // current provider name
+	callbackInput textinput.Model
+	inputActive   bool // true when user is typing callback URL
+}
+
+type oauthState int
+
+const (
+	oauthIdle oauthState = iota
+	oauthPending
+	oauthRemote // remote browser mode: waiting for manual callback
+	oauthSuccess
+	oauthError
+)
+
+// Messages
+type oauthStartMsg struct {
+	url          string
+	state        string
+	providerName string
+	err          error
+}
+
+type oauthPollMsg struct {
+	done    bool
+	message string
+	err     error
+}
+
+type oauthCallbackSubmitMsg struct {
+	err error
+}
+
+func newOAuthTabModel(client *Client) oauthTabModel {
+	ti := textinput.New()
+	ti.Placeholder = "http://localhost:.../auth/callback?code=...&state=..."
+	ti.CharLimit = 2048
+	ti.Prompt = "  回调 URL: "
+	return oauthTabModel{
+		client:        client,
+		callbackInput: ti,
+	}
+}
+
+func (m oauthTabModel) Init() tea.Cmd {
+	return nil
+}
+
+func (m oauthTabModel) Update(msg tea.Msg) (oauthTabModel, tea.Cmd) {
+	switch msg := msg.(type) {
+	case localeChangedMsg:
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+	case oauthStartMsg:
+		if msg.err != nil {
+			m.state = oauthError
+			m.err = msg.err
+			m.message = errorStyle.Render("✗ " + msg.err.Error())
+			m.viewport.SetContent(m.renderContent())
+			return m, nil
+		}
+		m.authURL = msg.url
+		m.authState = msg.state
+		m.providerName = msg.providerName
+		m.state = oauthRemote
+		m.callbackInput.SetValue("")
+		m.callbackInput.Focus()
+		m.inputActive = true
+		m.message = ""
+		m.viewport.SetContent(m.renderContent())
+		// Also start polling in the background
+		return m, tea.Batch(textinput.Blink, m.pollOAuthStatus(msg.state))
+
+	case oauthPollMsg:
+		if msg.err != nil {
+			m.state = oauthError
+			m.err = msg.err
+			m.message = errorStyle.Render("✗ " + msg.err.Error())
+			m.inputActive = false
+			m.callbackInput.Blur()
+		} else if msg.done {
+			m.state = oauthSuccess
+			m.message = successStyle.Render("✓ " + msg.message)
+			m.inputActive = false
+			m.callbackInput.Blur()
+		} else {
+			m.message = warningStyle.Render("⏳ " + msg.message)
+		}
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+
+	case oauthCallbackSubmitMsg:
+		if msg.err != nil {
+			m.message = errorStyle.Render(T("oauth_submit_fail") + ": " + msg.err.Error())
+		} else {
+			m.message = successStyle.Render(T("oauth_submit_ok"))
+		}
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+
+	case tea.KeyMsg:
+		// ---- Input active: typing callback URL ----
+		if m.inputActive {
+			switch msg.String() {
+			case "enter":
+				callbackURL := m.callbackInput.Value()
+				if callbackURL == "" {
+					return m, nil
+				}
+				m.inputActive = false
+				m.callbackInput.Blur()
+				m.message = warningStyle.Render(T("oauth_submitting"))
+				m.viewport.SetContent(m.renderContent())
+				return m, m.submitCallback(callbackURL)
+			case "esc":
+				m.inputActive = false
+				m.callbackInput.Blur()
+				m.viewport.SetContent(m.renderContent())
+				return m, nil
+			default:
+				var cmd tea.Cmd
+				m.callbackInput, cmd = m.callbackInput.Update(msg)
+				m.viewport.SetContent(m.renderContent())
+				return m, cmd
+			}
+		}
+
+		// ---- Remote mode but not typing ----
+		if m.state == oauthRemote {
+			switch msg.String() {
+			case "c", "C":
+				// Re-activate input
+				m.inputActive = true
+				m.callbackInput.Focus()
+				m.viewport.SetContent(m.renderContent())
+				return m, textinput.Blink
+			case "esc":
+				m.state = oauthIdle
+				m.message = ""
+				m.authURL = ""
+				m.authState = ""
+				m.viewport.SetContent(m.renderContent())
+				return m, nil
+			}
+			var cmd tea.Cmd
+			m.viewport, cmd = m.viewport.Update(msg)
+			return m, cmd
+		}
+
+		// ---- Pending (auto polling) ----
+		if m.state == oauthPending {
+			if msg.String() == "esc" {
+				m.state = oauthIdle
+				m.message = ""
+				m.viewport.SetContent(m.renderContent())
+			}
+			return m, nil
+		}
+
+		// ---- Idle ----
+		switch msg.String() {
+		case "up", "k":
+			if m.cursor > 0 {
+				m.cursor--
+				m.viewport.SetContent(m.renderContent())
+			}
+			return m, nil
+		case "down", "j":
+			if m.cursor < len(oauthProviders)-1 {
+				m.cursor++
+				m.viewport.SetContent(m.renderContent())
+			}
+			return m, nil
+		case "enter":
+			if m.cursor >= 0 && m.cursor < len(oauthProviders) {
+				provider := oauthProviders[m.cursor]
+				m.state = oauthPending
+				m.message = warningStyle.Render(fmt.Sprintf(T("oauth_initiating"), provider.name))
+				m.viewport.SetContent(m.renderContent())
+				return m, m.startOAuth(provider)
+			}
+			return m, nil
+		case "esc":
+			m.state = oauthIdle
+			m.message = ""
+			m.err = nil
+			m.viewport.SetContent(m.renderContent())
+			return m, nil
+		}
+
+		var cmd tea.Cmd
+		m.viewport, cmd = m.viewport.Update(msg)
+		return m, cmd
+	}
+
+	var cmd tea.Cmd
+	m.viewport, cmd = m.viewport.Update(msg)
+	return m, cmd
+}
+
+func (m oauthTabModel) startOAuth(provider oauthProvider) tea.Cmd {
+	return func() tea.Msg {
+		// Call the auth URL endpoint with is_webui=true
+		data, err := m.client.getJSON("/v0/management/" + provider.apiPath + "?is_webui=true")
+		if err != nil {
+			return oauthStartMsg{err: fmt.Errorf("failed to start %s login: %w", provider.name, err)}
+		}
+
+		authURL := getString(data, "url")
+		state := getString(data, "state")
+		if authURL == "" {
+			return oauthStartMsg{err: fmt.Errorf("no auth URL returned for %s", provider.name)}
+		}
+
+		// Try to open browser (best effort)
+		_ = openBrowser(authURL)
+
+		return oauthStartMsg{url: authURL, state: state, providerName: provider.name}
+	}
+}
+
+func (m oauthTabModel) submitCallback(callbackURL string) tea.Cmd {
+	return func() tea.Msg {
+		// Determine provider from current context
+		providerKey := ""
+		for _, p := range oauthProviders {
+			if p.name == m.providerName {
+				// Map provider name to the canonical key the API expects
+				switch p.apiPath {
+				case "gemini-cli-auth-url":
+					providerKey = "gemini"
+				case "anthropic-auth-url":
+					providerKey = "anthropic"
+				case "codex-auth-url":
+					providerKey = "codex"
+				case "antigravity-auth-url":
+					providerKey = "antigravity"
+				case "qwen-auth-url":
+					providerKey = "qwen"
+				case "kimi-auth-url":
+					providerKey = "kimi"
+				case "iflow-auth-url":
+					providerKey = "iflow"
+				}
+				break
+			}
+		}
+
+		body := map[string]string{
+			"provider":     providerKey,
+			"redirect_url": callbackURL,
+			"state":        m.authState,
+		}
+		err := m.client.postJSON("/v0/management/oauth-callback", body)
+		if err != nil {
+			return oauthCallbackSubmitMsg{err: err}
+		}
+		return oauthCallbackSubmitMsg{}
+	}
+}
+
+func (m oauthTabModel) pollOAuthStatus(state string) tea.Cmd {
+	return func() tea.Msg {
+		// Poll session status for up to 5 minutes
+		deadline := time.Now().Add(5 * time.Minute)
+		for {
+			if time.Now().After(deadline) {
+				return oauthPollMsg{done: false, err: fmt.Errorf("%s", T("oauth_timeout"))}
+			}
+
+			time.Sleep(2 * time.Second)
+
+			status, errMsg, err := m.client.GetAuthStatus(state)
+			if err != nil {
+				continue // Ignore transient errors
+			}
+
+			switch status {
+			case "ok":
+				return oauthPollMsg{
+					done:    true,
+					message: T("oauth_success"),
+				}
+			case "error":
+				return oauthPollMsg{
+					done: false,
+					err:  fmt.Errorf("%s: %s", T("oauth_failed"), errMsg),
+				}
+			case "wait":
+				continue
+			default:
+				return oauthPollMsg{
+					done:    true,
+					message: T("oauth_completed"),
+				}
+			}
+		}
+	}
+}
+
+func (m *oauthTabModel) SetSize(w, h int) {
+	m.width = w
+	m.height = h
+	m.callbackInput.Width = w - 16
+	if !m.ready {
+		m.viewport = viewport.New(w, h)
+		m.viewport.SetContent(m.renderContent())
+		m.ready = true
+	} else {
+		m.viewport.Width = w
+		m.viewport.Height = h
+	}
+}
+
+func (m oauthTabModel) View() string {
+	if !m.ready {
+		return T("loading")
+	}
+	return m.viewport.View()
+}
+
+func (m oauthTabModel) renderContent() string {
+	var sb strings.Builder
+
+	sb.WriteString(titleStyle.Render(T("oauth_title")))
+	sb.WriteString("\n\n")
+
+	if m.message != "" {
+		sb.WriteString("  " + m.message)
+		sb.WriteString("\n\n")
+	}
+
+	// ---- Remote browser mode ----
+	if m.state == oauthRemote {
+		sb.WriteString(m.renderRemoteMode())
+		return sb.String()
+	}
+
+	if m.state == oauthPending {
+		sb.WriteString(helpStyle.Render(T("oauth_press_esc")))
+		return sb.String()
+	}
+
+	sb.WriteString(helpStyle.Render(T("oauth_select")))
+	sb.WriteString("\n\n")
+
+	for i, p := range oauthProviders {
+		isSelected := i == m.cursor
+		prefix := "  "
+		if isSelected {
+			prefix = "▸ "
+		}
+
+		label := fmt.Sprintf("%s %s", p.emoji, p.name)
+		if isSelected {
+			label = lipgloss.NewStyle().Bold(true).Foreground(lipgloss.Color("#FFFFFF")).Background(colorPrimary).Padding(0, 1).Render(label)
+		} else {
+			label = lipgloss.NewStyle().Foreground(colorText).Padding(0, 1).Render(label)
+		}
+
+		sb.WriteString(prefix + label + "\n")
+	}
+
+	sb.WriteString("\n")
+	sb.WriteString(helpStyle.Render(T("oauth_help")))
+
+	return sb.String()
+}
+
+func (m oauthTabModel) renderRemoteMode() string {
+	var sb strings.Builder
+
+	providerStyle := lipgloss.NewStyle().Bold(true).Foreground(colorHighlight)
+	sb.WriteString(providerStyle.Render(fmt.Sprintf("  ✦ %s OAuth", m.providerName)))
+	sb.WriteString("\n\n")
+
+	// Auth URL section
+	sb.WriteString(lipgloss.NewStyle().Bold(true).Foreground(colorInfo).Render(T("oauth_auth_url")))
+	sb.WriteString("\n")
+
+	// Wrap URL to fit terminal width
+	urlStyle := lipgloss.NewStyle().Foreground(lipgloss.Color("252"))
+	maxURLWidth := m.width - 6
+	if maxURLWidth < 40 {
+		maxURLWidth = 40
+	}
+	wrappedURL := wrapText(m.authURL, maxURLWidth)
+	for _, line := range wrappedURL {
+		sb.WriteString("  " + urlStyle.Render(line) + "\n")
+	}
+	sb.WriteString("\n")
+
+	sb.WriteString(helpStyle.Render(T("oauth_remote_hint")))
+	sb.WriteString("\n\n")
+
+	// Callback URL input
+	sb.WriteString(lipgloss.NewStyle().Bold(true).Foreground(colorInfo).Render(T("oauth_callback_url")))
+	sb.WriteString("\n")
+
+	if m.inputActive {
+		sb.WriteString(m.callbackInput.View())
+		sb.WriteString("\n")
+		sb.WriteString(helpStyle.Render("  " + T("enter_submit") + " • " + T("esc_cancel")))
+	} else {
+		sb.WriteString(helpStyle.Render(T("oauth_press_c")))
+	}
+
+	sb.WriteString("\n\n")
+	sb.WriteString(warningStyle.Render(T("oauth_waiting")))
+
+	return sb.String()
+}
+
+// wrapText splits a long string into lines of at most maxWidth characters.
+func wrapText(s string, maxWidth int) []string {
+	if maxWidth <= 0 {
+		return []string{s}
+	}
+	var lines []string
+	for len(s) > maxWidth {
+		lines = append(lines, s[:maxWidth])
+		s = s[maxWidth:]
+	}
+	if len(s) > 0 {
+		lines = append(lines, s)
+	}
+	return lines
+}
--- a/internal/tui/styles.go
+++ b/internal/tui/styles.go
@@ -0,0 +1,126 @@
+// Package tui provides a terminal-based management interface for CLIProxyAPI.
+package tui
+
+import "github.com/charmbracelet/lipgloss"
+
+// Color palette
+var (
+	colorPrimary   = lipgloss.Color("#7C3AED") // violet
+	colorSecondary = lipgloss.Color("#6366F1") // indigo
+	colorSuccess   = lipgloss.Color("#22C55E") // green
+	colorWarning   = lipgloss.Color("#EAB308") // yellow
+	colorError     = lipgloss.Color("#EF4444") // red
+	colorInfo      = lipgloss.Color("#3B82F6") // blue
+	colorMuted     = lipgloss.Color("#6B7280") // gray
+	colorBg        = lipgloss.Color("#1E1E2E") // dark bg
+	colorSurface   = lipgloss.Color("#313244") // slightly lighter
+	colorText      = lipgloss.Color("#CDD6F4") // light text
+	colorSubtext   = lipgloss.Color("#A6ADC8") // dimmer text
+	colorBorder    = lipgloss.Color("#45475A") // border
+	colorHighlight = lipgloss.Color("#F5C2E7") // pink highlight
+)
+
+// Tab bar styles
+var (
+	tabActiveStyle = lipgloss.NewStyle().
+			Bold(true).
+			Foreground(lipgloss.Color("#FFFFFF")).
+			Background(colorPrimary).
+			Padding(0, 2)
+
+	tabInactiveStyle = lipgloss.NewStyle().
+				Foreground(colorSubtext).
+				Background(colorSurface).
+				Padding(0, 2)
+
+	tabBarStyle = lipgloss.NewStyle().
+			Background(colorSurface).
+			PaddingLeft(1).
+			PaddingBottom(0)
+)
+
+// Content styles
+var (
+	titleStyle = lipgloss.NewStyle().
+			Bold(true).
+			Foreground(colorHighlight).
+			MarginBottom(1)
+
+	subtitleStyle = lipgloss.NewStyle().
+			Foreground(colorSubtext).
+			Italic(true)
+
+	labelStyle = lipgloss.NewStyle().
+			Foreground(colorInfo).
+			Bold(true).
+			Width(24)
+
+	valueStyle = lipgloss.NewStyle().
+			Foreground(colorText)
+
+	sectionStyle = lipgloss.NewStyle().
+			Border(lipgloss.RoundedBorder()).
+			BorderForeground(colorBorder).
+			Padding(1, 2)
+
+	errorStyle = lipgloss.NewStyle().
+			Foreground(colorError).
+			Bold(true)
+
+	successStyle = lipgloss.NewStyle().
+			Foreground(colorSuccess)
+
+	warningStyle = lipgloss.NewStyle().
+			Foreground(colorWarning)
+
+	statusBarStyle = lipgloss.NewStyle().
+			Foreground(colorSubtext).
+			Background(colorSurface).
+			PaddingLeft(1).
+			PaddingRight(1)
+
+	helpStyle = lipgloss.NewStyle().
+			Foreground(colorMuted)
+)
+
+// Log level styles
+var (
+	logDebugStyle = lipgloss.NewStyle().Foreground(colorMuted)
+	logInfoStyle  = lipgloss.NewStyle().Foreground(colorInfo)
+	logWarnStyle  = lipgloss.NewStyle().Foreground(colorWarning)
+	logErrorStyle = lipgloss.NewStyle().Foreground(colorError)
+)
+
+// Table styles
+var (
+	tableHeaderStyle = lipgloss.NewStyle().
+				Bold(true).
+				Foreground(colorHighlight).
+				BorderBottom(true).
+				BorderStyle(lipgloss.NormalBorder()).
+				BorderForeground(colorBorder)
+
+	tableCellStyle = lipgloss.NewStyle().
+			Foreground(colorText).
+			PaddingRight(2)
+
+	tableSelectedStyle = lipgloss.NewStyle().
+				Foreground(lipgloss.Color("#FFFFFF")).
+				Background(colorPrimary).
+				Bold(true)
+)
+
+func logLevelStyle(level string) lipgloss.Style {
+	switch level {
+	case "debug":
+		return logDebugStyle
+	case "info":
+		return logInfoStyle
+	case "warn", "warning":
+		return logWarnStyle
+	case "error", "fatal", "panic":
+		return logErrorStyle
+	default:
+		return logInfoStyle
+	}
+}
--- a/internal/tui/usage_tab.go
+++ b/internal/tui/usage_tab.go
@@ -0,0 +1,364 @@
+package tui
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+
+	"github.com/charmbracelet/bubbles/viewport"
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+)
+
+// usageTabModel displays usage statistics with charts and breakdowns.
+type usageTabModel struct {
+	client   *Client
+	viewport viewport.Model
+	usage    map[string]any
+	err      error
+	width    int
+	height   int
+	ready    bool
+}
+
+type usageDataMsg struct {
+	usage map[string]any
+	err   error
+}
+
+func newUsageTabModel(client *Client) usageTabModel {
+	return usageTabModel{
+		client: client,
+	}
+}
+
+func (m usageTabModel) Init() tea.Cmd {
+	return m.fetchData
+}
+
+func (m usageTabModel) fetchData() tea.Msg {
+	usage, err := m.client.GetUsage()
+	return usageDataMsg{usage: usage, err: err}
+}
+
+func (m usageTabModel) Update(msg tea.Msg) (usageTabModel, tea.Cmd) {
+	switch msg := msg.(type) {
+	case localeChangedMsg:
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+	case usageDataMsg:
+		if msg.err != nil {
+			m.err = msg.err
+		} else {
+			m.err = nil
+			m.usage = msg.usage
+		}
+		m.viewport.SetContent(m.renderContent())
+		return m, nil
+
+	case tea.KeyMsg:
+		if msg.String() == "r" {
+			return m, m.fetchData
+		}
+		var cmd tea.Cmd
+		m.viewport, cmd = m.viewport.Update(msg)
+		return m, cmd
+	}
+
+	var cmd tea.Cmd
+	m.viewport, cmd = m.viewport.Update(msg)
+	return m, cmd
+}
+
+func (m *usageTabModel) SetSize(w, h int) {
+	m.width = w
+	m.height = h
+	if !m.ready {
+		m.viewport = viewport.New(w, h)
+		m.viewport.SetContent(m.renderContent())
+		m.ready = true
+	} else {
+		m.viewport.Width = w
+		m.viewport.Height = h
+	}
+}
+
+func (m usageTabModel) View() string {
+	if !m.ready {
+		return T("loading")
+	}
+	return m.viewport.View()
+}
+
+func (m usageTabModel) renderContent() string {
+	var sb strings.Builder
+
+	sb.WriteString(titleStyle.Render(T("usage_title")))
+	sb.WriteString("\n")
+	sb.WriteString(helpStyle.Render(T("usage_help")))
+	sb.WriteString("\n\n")
+
+	if m.err != nil {
+		sb.WriteString(errorStyle.Render("⚠ Error: " + m.err.Error()))
+		sb.WriteString("\n")
+		return sb.String()
+	}
+
+	if m.usage == nil {
+		sb.WriteString(subtitleStyle.Render(T("usage_no_data")))
+		sb.WriteString("\n")
+		return sb.String()
+	}
+
+	usageMap, _ := m.usage["usage"].(map[string]any)
+	if usageMap == nil {
+		sb.WriteString(subtitleStyle.Render(T("usage_no_data")))
+		sb.WriteString("\n")
+		return sb.String()
+	}
+
+	totalReqs := int64(getFloat(usageMap, "total_requests"))
+	successCnt := int64(getFloat(usageMap, "success_count"))
+	failureCnt := int64(getFloat(usageMap, "failure_count"))
+	totalTokens := int64(getFloat(usageMap, "total_tokens"))
+
+	// ━━━ Overview Cards ━━━
+	cardWidth := 20
+	if m.width > 0 {
+		cardWidth = (m.width - 6) / 4
+		if cardWidth < 16 {
+			cardWidth = 16
+		}
+	}
+	cardStyle := lipgloss.NewStyle().
+		Border(lipgloss.RoundedBorder()).
+		BorderForeground(lipgloss.Color("240")).
+		Padding(0, 1).
+		Width(cardWidth).
+		Height(3)
+
+	// Total Requests
+	card1 := cardStyle.Copy().BorderForeground(lipgloss.Color("111")).Render(fmt.Sprintf(
+		"%s\n%s\n%s",
+		lipgloss.NewStyle().Foreground(colorMuted).Render(T("usage_total_reqs")),
+		lipgloss.NewStyle().Bold(true).Foreground(lipgloss.Color("111")).Render(fmt.Sprintf("%d", totalReqs)),
+		lipgloss.NewStyle().Foreground(colorMuted).Render(fmt.Sprintf("● %s: %d  ● %s: %d", T("usage_success"), successCnt, T("usage_failure"), failureCnt)),
+	))
+
+	// Total Tokens
+	card2 := cardStyle.Copy().BorderForeground(lipgloss.Color("214")).Render(fmt.Sprintf(
+		"%s\n%s\n%s",
+		lipgloss.NewStyle().Foreground(colorMuted).Render(T("usage_total_tokens")),
+		lipgloss.NewStyle().Bold(true).Foreground(lipgloss.Color("214")).Render(formatLargeNumber(totalTokens)),
+		lipgloss.NewStyle().Foreground(colorMuted).Render(fmt.Sprintf("%s: %s", T("usage_total_token_l"), formatLargeNumber(totalTokens))),
+	))
+
+	// RPM
+	rpm := float64(0)
+	if totalReqs > 0 {
+		if rByH, ok := usageMap["requests_by_hour"].(map[string]any); ok && len(rByH) > 0 {
+			rpm = float64(totalReqs) / float64(len(rByH)) / 60.0
+		}
+	}
+	card3 := cardStyle.Copy().BorderForeground(lipgloss.Color("76")).Render(fmt.Sprintf(
+		"%s\n%s\n%s",
+		lipgloss.NewStyle().Foreground(colorMuted).Render(T("usage_rpm")),
+		lipgloss.NewStyle().Bold(true).Foreground(lipgloss.Color("76")).Render(fmt.Sprintf("%.2f", rpm)),
+		lipgloss.NewStyle().Foreground(colorMuted).Render(fmt.Sprintf("%s: %d", T("usage_total_reqs"), totalReqs)),
+	))
+
+	// TPM
+	tpm := float64(0)
+	if totalTokens > 0 {
+		if tByH, ok := usageMap["tokens_by_hour"].(map[string]any); ok && len(tByH) > 0 {
+			tpm = float64(totalTokens) / float64(len(tByH)) / 60.0
+		}
+	}
+	card4 := cardStyle.Copy().BorderForeground(lipgloss.Color("170")).Render(fmt.Sprintf(
+		"%s\n%s\n%s",
+		lipgloss.NewStyle().Foreground(colorMuted).Render(T("usage_tpm")),
+		lipgloss.NewStyle().Bold(true).Foreground(lipgloss.Color("170")).Render(fmt.Sprintf("%.2f", tpm)),
+		lipgloss.NewStyle().Foreground(colorMuted).Render(fmt.Sprintf("%s: %s", T("usage_total_tokens"), formatLargeNumber(totalTokens))),
+	))
+
+	sb.WriteString(lipgloss.JoinHorizontal(lipgloss.Top, card1, " ", card2, " ", card3, " ", card4))
+	sb.WriteString("\n\n")
+
+	// ━━━ Requests by Hour (ASCII bar chart) ━━━
+	if rByH, ok := usageMap["requests_by_hour"].(map[string]any); ok && len(rByH) > 0 {
+		sb.WriteString(lipgloss.NewStyle().Bold(true).Foreground(colorHighlight).Render(T("usage_req_by_hour")))
+		sb.WriteString("\n")
+		sb.WriteString(strings.Repeat("─", minInt(m.width, 60)))
+		sb.WriteString("\n")
+		sb.WriteString(renderBarChart(rByH, m.width-6, lipgloss.Color("111")))
+		sb.WriteString("\n")
+	}
+
+	// ━━━ Tokens by Hour ━━━
+	if tByH, ok := usageMap["tokens_by_hour"].(map[string]any); ok && len(tByH) > 0 {
+		sb.WriteString(lipgloss.NewStyle().Bold(true).Foreground(colorHighlight).Render(T("usage_tok_by_hour")))
+		sb.WriteString("\n")
+		sb.WriteString(strings.Repeat("─", minInt(m.width, 60)))
+		sb.WriteString("\n")
+		sb.WriteString(renderBarChart(tByH, m.width-6, lipgloss.Color("214")))
+		sb.WriteString("\n")
+	}
+
+	// ━━━ Requests by Day ━━━
+	if rByD, ok := usageMap["requests_by_day"].(map[string]any); ok && len(rByD) > 0 {
+		sb.WriteString(lipgloss.NewStyle().Bold(true).Foreground(colorHighlight).Render(T("usage_req_by_day")))
+		sb.WriteString("\n")
+		sb.WriteString(strings.Repeat("─", minInt(m.width, 60)))
+		sb.WriteString("\n")
+		sb.WriteString(renderBarChart(rByD, m.width-6, lipgloss.Color("76")))
+		sb.WriteString("\n")
+	}
+
+	// ━━━ API Detail Stats ━━━
+	if apis, ok := usageMap["apis"].(map[string]any); ok && len(apis) > 0 {
+		sb.WriteString(lipgloss.NewStyle().Bold(true).Foreground(colorHighlight).Render(T("usage_api_detail")))
+		sb.WriteString("\n")
+		sb.WriteString(strings.Repeat("─", minInt(m.width, 80)))
+		sb.WriteString("\n")
+
+		header := fmt.Sprintf("  %-30s %10s %12s", "API", T("requests"), T("tokens"))
+		sb.WriteString(tableHeaderStyle.Render(header))
+		sb.WriteString("\n")
+
+		for apiName, apiSnap := range apis {
+			if apiMap, ok := apiSnap.(map[string]any); ok {
+				apiReqs := int64(getFloat(apiMap, "total_requests"))
+				apiToks := int64(getFloat(apiMap, "total_tokens"))
+
+				row := fmt.Sprintf("  %-30s %10d %12s",
+					truncate(maskKey(apiName), 30), apiReqs, formatLargeNumber(apiToks))
+				sb.WriteString(lipgloss.NewStyle().Bold(true).Render(row))
+				sb.WriteString("\n")
+
+				// Per-model breakdown
+				if models, ok := apiMap["models"].(map[string]any); ok {
+					for model, v := range models {
+						if stats, ok := v.(map[string]any); ok {
+							mReqs := int64(getFloat(stats, "total_requests"))
+							mToks := int64(getFloat(stats, "total_tokens"))
+							mRow := fmt.Sprintf("    ├─ %-28s %10d %12s",
+								truncate(model, 28), mReqs, formatLargeNumber(mToks))
+							sb.WriteString(tableCellStyle.Render(mRow))
+							sb.WriteString("\n")
+
+							// Token type breakdown from details
+							sb.WriteString(m.renderTokenBreakdown(stats))
+						}
+					}
+				}
+			}
+		}
+	}
+
+	sb.WriteString("\n")
+	return sb.String()
+}
+
+// renderTokenBreakdown aggregates input/output/cached/reasoning tokens from model details.
+func (m usageTabModel) renderTokenBreakdown(modelStats map[string]any) string {
+	details, ok := modelStats["details"]
+	if !ok {
+		return ""
+	}
+	detailList, ok := details.([]any)
+	if !ok || len(detailList) == 0 {
+		return ""
+	}
+
+	var inputTotal, outputTotal, cachedTotal, reasoningTotal int64
+	for _, d := range detailList {
+		dm, ok := d.(map[string]any)
+		if !ok {
+			continue
+		}
+		tokens, ok := dm["tokens"].(map[string]any)
+		if !ok {
+			continue
+		}
+		inputTotal += int64(getFloat(tokens, "input_tokens"))
+		outputTotal += int64(getFloat(tokens, "output_tokens"))
+		cachedTotal += int64(getFloat(tokens, "cached_tokens"))
+		reasoningTotal += int64(getFloat(tokens, "reasoning_tokens"))
+	}
+
+	if inputTotal == 0 && outputTotal == 0 && cachedTotal == 0 && reasoningTotal == 0 {
+		return ""
+	}
+
+	parts := []string{}
+	if inputTotal > 0 {
+		parts = append(parts, fmt.Sprintf("%s:%s", T("usage_input"), formatLargeNumber(inputTotal)))
+	}
+	if outputTotal > 0 {
+		parts = append(parts, fmt.Sprintf("%s:%s", T("usage_output"), formatLargeNumber(outputTotal)))
+	}
+	if cachedTotal > 0 {
+		parts = append(parts, fmt.Sprintf("%s:%s", T("usage_cached"), formatLargeNumber(cachedTotal)))
+	}
+	if reasoningTotal > 0 {
+		parts = append(parts, fmt.Sprintf("%s:%s", T("usage_reasoning"), formatLargeNumber(reasoningTotal)))
+	}
+
+	return fmt.Sprintf("    │  %s\n",
+		lipgloss.NewStyle().Foreground(colorMuted).Render(strings.Join(parts, "  ")))
+}
+
+// renderBarChart renders a simple ASCII horizontal bar chart.
+func renderBarChart(data map[string]any, maxBarWidth int, barColor lipgloss.Color) string {
+	if maxBarWidth < 10 {
+		maxBarWidth = 10
+	}
+
+	// Sort keys
+	keys := make([]string, 0, len(data))
+	for k := range data {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+
+	// Find max value
+	maxVal := float64(0)
+	for _, k := range keys {
+		v := getFloat(data, k)
+		if v > maxVal {
+			maxVal = v
+		}
+	}
+	if maxVal == 0 {
+		return ""
+	}
+
+	barStyle := lipgloss.NewStyle().Foreground(barColor)
+	var sb strings.Builder
+
+	labelWidth := 12
+	barAvail := maxBarWidth - labelWidth - 12
+	if barAvail < 5 {
+		barAvail = 5
+	}
+
+	for _, k := range keys {
+		v := getFloat(data, k)
+		barLen := int(v / maxVal * float64(barAvail))
+		if barLen < 1 && v > 0 {
+			barLen = 1
+		}
+		bar := strings.Repeat("█", barLen)
+		label := k
+		if len(label) > labelWidth {
+			label = label[:labelWidth]
+		}
+		sb.WriteString(fmt.Sprintf("  %-*s %s %s\n",
+			labelWidth, label,
+			barStyle.Render(bar),
+			lipgloss.NewStyle().Foreground(colorMuted).Render(fmt.Sprintf("%.0f", v)),
+		))
+	}
+
+	return sb.String()
+}
--- a/internal/watcher/diff/config_diff.go
+++ b/internal/watcher/diff/config_diff.go
@@ -184,6 +184,9 @@ func BuildConfigChangeDetails(oldCfg, newCfg *config.Config) []string {
 			if strings.TrimSpace(o.Prefix) != strings.TrimSpace(n.Prefix) {
 				changes = append(changes, fmt.Sprintf("codex[%d].prefix: %s -> %s", i, strings.TrimSpace(o.Prefix), strings.TrimSpace(n.Prefix)))
 			}
+			if o.Websockets != n.Websockets {
+				changes = append(changes, fmt.Sprintf("codex[%d].websockets: %t -> %t", i, o.Websockets, n.Websockets))
+			}
 			if strings.TrimSpace(o.APIKey) != strings.TrimSpace(n.APIKey) {
 				changes = append(changes, fmt.Sprintf("codex[%d].api-key: updated", i))
 			}
--- a/internal/watcher/synthesizer/config.go
+++ b/internal/watcher/synthesizer/config.go
@@ -164,6 +164,9 @@ func (s *ConfigSynthesizer) synthesizeCodexKeys(ctx *SynthesisContext) []*coreau
 		if ck.BaseURL != "" {
 			attrs["base_url"] = ck.BaseURL
 		}
+		if ck.Websockets {
+			attrs["websockets"] = "true"
+		}
 		if hash := diff.ComputeCodexModelsHash(ck.Models); hash != "" {
 			attrs["models_hash"] = hash
 		}
--- a/internal/watcher/synthesizer/config_test.go
+++ b/internal/watcher/synthesizer/config_test.go
@@ -231,10 +231,11 @@ func TestConfigSynthesizer_CodexKeys(t *testing.T) {
 		Config: &config.Config{
 			CodexKey: []config.CodexKey{
 				{
-					APIKey:   "codex-key-123",
-					Prefix:   "dev",
-					BaseURL:  "https://api.openai.com",
-					ProxyURL: "http://proxy.local",
+					APIKey:     "codex-key-123",
+					Prefix:     "dev",
+					BaseURL:    "https://api.openai.com",
+					ProxyURL:   "http://proxy.local",
+					Websockets: true,
 				},
 			},
 		},
@@ -259,6 +260,9 @@ func TestConfigSynthesizer_CodexKeys(t *testing.T) {
 	if auths[0].ProxyURL != "http://proxy.local" {
 		t.Errorf("expected proxy_url http://proxy.local, got %s", auths[0].ProxyURL)
 	}
+	if auths[0].Attributes["websockets"] != "true" {
+		t.Errorf("expected websockets=true, got %s", auths[0].Attributes["websockets"])
+	}
 }

 func TestConfigSynthesizer_CodexKeys_SkipsEmptyAndHeaders(t *testing.T) {
--- a/sdk/api/handlers/claude/code_handlers.go
+++ b/sdk/api/handlers/claude/code_handlers.go
@@ -112,12 +112,13 @@ func (h *ClaudeCodeAPIHandler) ClaudeCountTokens(c *gin.Context) {

 	modelName := gjson.GetBytes(rawJSON, "model").String()

-	resp, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -165,7 +166,7 @@ func (h *ClaudeCodeAPIHandler) handleNonStreamingResponse(c *gin.Context, rawJSO

 	modelName := gjson.GetBytes(rawJSON, "model").String()

-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
@@ -194,6 +195,7 @@ func (h *ClaudeCodeAPIHandler) handleNonStreamingResponse(c *gin.Context, rawJSO
 		}
 	}

+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -225,7 +227,7 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [
 	// This allows proper cleanup and cancellation of ongoing requests
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())

-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
 		c.Header("Cache-Control", "no-cache")
@@ -257,6 +259,7 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [
 			if !ok {
 				// Stream closed without data? Send DONE or just headers.
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				flusher.Flush()
 				cliCancel(nil)
 				return
@@ -264,6 +267,7 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [

 			// Success! Set headers now.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)

 			// Write the first chunk
 			if len(chunk) > 0 {
--- a/sdk/api/handlers/gemini/gemini-cli_handlers.go
+++ b/sdk/api/handlers/gemini/gemini-cli_handlers.go
@@ -159,7 +159,8 @@ func (h *GeminiCLIAPIHandler) handleInternalStreamGenerateContent(c *gin.Context
 	modelName := modelResult.String()

 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	h.forwardCLIStream(c, flusher, "", func(err error) { cliCancel(err) }, dataChan, errChan)
 	return
 }
@@ -172,12 +173,13 @@ func (h *GeminiCLIAPIHandler) handleInternalGenerateContent(c *gin.Context, rawJ
 	modelName := modelResult.String()

 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -185,8 +187,7 @@ func (h *GeminiCLIAPIHandler) handleInternalGenerateContent(c *gin.Context, rawJ
 func (h *GeminiCLIAPIHandler) forwardCLIStream(c *gin.Context, flusher http.Flusher, alt string, cancel func(error), data <-chan []byte, errs <-chan *interfaces.ErrorMessage) {
 	var keepAliveInterval *time.Duration
 	if alt != "" {
-		disabled := time.Duration(0)
-		keepAliveInterval = &disabled
+		keepAliveInterval = new(time.Duration(0))
 	}

 	h.ForwardStream(c, flusher, cancel, data, errs, handlers.StreamForwardOptions{
--- a/sdk/api/handlers/gemini/gemini_handlers.go
+++ b/sdk/api/handlers/gemini/gemini_handlers.go
@@ -188,7 +188,7 @@ func (h *GeminiAPIHandler) handleStreamGenerateContent(c *gin.Context, modelName
 	}

 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)

 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -223,6 +223,7 @@ func (h *GeminiAPIHandler) handleStreamGenerateContent(c *gin.Context, modelName
 				if alt == "" {
 					setSSEHeaders()
 				}
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				flusher.Flush()
 				cliCancel(nil)
 				return
@@ -232,6 +233,7 @@ func (h *GeminiAPIHandler) handleStreamGenerateContent(c *gin.Context, modelName
 			if alt == "" {
 				setSSEHeaders()
 			}
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)

 			// Write first chunk
 			if alt == "" {
@@ -262,12 +264,13 @@ func (h *GeminiAPIHandler) handleCountTokens(c *gin.Context, modelName string, r
 	c.Header("Content-Type", "application/json")
 	alt := h.GetAlt(c)
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -286,13 +289,14 @@ func (h *GeminiAPIHandler) handleGenerateContent(c *gin.Context, modelName strin
 	alt := h.GetAlt(c)
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -300,8 +304,7 @@ func (h *GeminiAPIHandler) handleGenerateContent(c *gin.Context, modelName strin
 func (h *GeminiAPIHandler) forwardGeminiStream(c *gin.Context, flusher http.Flusher, alt string, cancel func(error), data <-chan []byte, errs <-chan *interfaces.ErrorMessage) {
 	var keepAliveInterval *time.Duration
 	if alt != "" {
-		disabled := time.Duration(0)
-		keepAliveInterval = &disabled
+		keepAliveInterval = new(time.Duration(0))
 	}

 	h.ForwardStream(c, flusher, cancel, data, errs, handlers.StreamForwardOptions{
--- a/sdk/api/handlers/handlers.go
+++ b/sdk/api/handlers/handlers.go
@@ -52,6 +52,45 @@ const (
 	defaultStreamingBootstrapRetries = 0
 )

+type pinnedAuthContextKey struct{}
+type selectedAuthCallbackContextKey struct{}
+type executionSessionContextKey struct{}
+
+// WithPinnedAuthID returns a child context that requests execution on a specific auth ID.
+func WithPinnedAuthID(ctx context.Context, authID string) context.Context {
+	authID = strings.TrimSpace(authID)
+	if authID == "" {
+		return ctx
+	}
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	return context.WithValue(ctx, pinnedAuthContextKey{}, authID)
+}
+
+// WithSelectedAuthIDCallback returns a child context that receives the selected auth ID.
+func WithSelectedAuthIDCallback(ctx context.Context, callback func(string)) context.Context {
+	if callback == nil {
+		return ctx
+	}
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	return context.WithValue(ctx, selectedAuthCallbackContextKey{}, callback)
+}
+
+// WithExecutionSessionID returns a child context tagged with a long-lived execution session ID.
+func WithExecutionSessionID(ctx context.Context, sessionID string) context.Context {
+	sessionID = strings.TrimSpace(sessionID)
+	if sessionID == "" {
+		return ctx
+	}
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	return context.WithValue(ctx, executionSessionContextKey{}, sessionID)
+}
+
 // BuildErrorResponseBody builds an OpenAI-compatible JSON error response body.
 // If errText is already valid JSON, it is returned as-is to preserve upstream error payloads.
 func BuildErrorResponseBody(status int, errText string) []byte {
@@ -140,6 +179,12 @@ func StreamingBootstrapRetries(cfg *config.SDKConfig) int {
 	return retries
 }

+// PassthroughHeadersEnabled returns whether upstream response headers should be forwarded to clients.
+// Default is false.
+func PassthroughHeadersEnabled(cfg *config.SDKConfig) bool {
+	return cfg != nil && cfg.PassthroughHeaders
+}
+
 func requestExecutionMetadata(ctx context.Context) map[string]any {
 	// Idempotency-Key is an optional client-supplied header used to correlate retries.
 	// It is forwarded as execution metadata; when absent we generate a UUID.
@@ -152,7 +197,59 @@ func requestExecutionMetadata(ctx context.Context) map[string]any {
 	if key == "" {
 		key = uuid.NewString()
 	}
-	return map[string]any{idempotencyKeyMetadataKey: key}
+
+	meta := map[string]any{idempotencyKeyMetadataKey: key}
+	if pinnedAuthID := pinnedAuthIDFromContext(ctx); pinnedAuthID != "" {
+		meta[coreexecutor.PinnedAuthMetadataKey] = pinnedAuthID
+	}
+	if selectedCallback := selectedAuthIDCallbackFromContext(ctx); selectedCallback != nil {
+		meta[coreexecutor.SelectedAuthCallbackMetadataKey] = selectedCallback
+	}
+	if executionSessionID := executionSessionIDFromContext(ctx); executionSessionID != "" {
+		meta[coreexecutor.ExecutionSessionMetadataKey] = executionSessionID
+	}
+	return meta
+}
+
+func pinnedAuthIDFromContext(ctx context.Context) string {
+	if ctx == nil {
+		return ""
+	}
+	raw := ctx.Value(pinnedAuthContextKey{})
+	switch v := raw.(type) {
+	case string:
+		return strings.TrimSpace(v)
+	case []byte:
+		return strings.TrimSpace(string(v))
+	default:
+		return ""
+	}
+}
+
+func selectedAuthIDCallbackFromContext(ctx context.Context) func(string) {
+	if ctx == nil {
+		return nil
+	}
+	raw := ctx.Value(selectedAuthCallbackContextKey{})
+	if callback, ok := raw.(func(string)); ok && callback != nil {
+		return callback
+	}
+	return nil
+}
+
+func executionSessionIDFromContext(ctx context.Context) string {
+	if ctx == nil {
+		return ""
+	}
+	raw := ctx.Value(executionSessionContextKey{})
+	switch v := raw.(type) {
+	case string:
+		return strings.TrimSpace(v)
+	case []byte:
+		return strings.TrimSpace(string(v))
+	default:
+		return ""
+	}
 }

 // BaseAPIHandler contains the handlers for API endpoints.
@@ -371,10 +468,10 @@ func appendAPIResponse(c *gin.Context, data []byte) {

 // ExecuteWithAuthManager executes a non-streaming request via the core auth manager.
 // This path is the only supported execution route.
-func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, *interfaces.ErrorMessage) {
+func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, http.Header, *interfaces.ErrorMessage) {
 	providers, normalizedModel, errMsg := h.getRequestDetails(modelName)
 	if errMsg != nil {
-		return nil, errMsg
+		return nil, nil, errMsg
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
@@ -407,17 +504,20 @@ func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType
 				addon = hdr.Clone()
 			}
 		}
-		return nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
+		return nil, nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
 	}
-	return resp.Payload, nil
+	if !PassthroughHeadersEnabled(h.Cfg) {
+		return resp.Payload, nil, nil
+	}
+	return resp.Payload, FilterUpstreamHeaders(resp.Headers), nil
 }

 // ExecuteCountWithAuthManager executes a non-streaming request via the core auth manager.
 // This path is the only supported execution route.
-func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, *interfaces.ErrorMessage) {
+func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, http.Header, *interfaces.ErrorMessage) {
 	providers, normalizedModel, errMsg := h.getRequestDetails(modelName)
 	if errMsg != nil {
-		return nil, errMsg
+		return nil, nil, errMsg
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
@@ -450,20 +550,24 @@ func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handle
 				addon = hdr.Clone()
 			}
 		}
-		return nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
+		return nil, nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
 	}
-	return resp.Payload, nil
+	if !PassthroughHeadersEnabled(h.Cfg) {
+		return resp.Payload, nil, nil
+	}
+	return resp.Payload, FilterUpstreamHeaders(resp.Headers), nil
 }

 // ExecuteStreamWithAuthManager executes a streaming request via the core auth manager.
 // This path is the only supported execution route.
-func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) (<-chan []byte, <-chan *interfaces.ErrorMessage) {
+// The returned http.Header carries upstream response headers captured before streaming begins.
+func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) (<-chan []byte, http.Header, <-chan *interfaces.ErrorMessage) {
 	providers, normalizedModel, errMsg := h.getRequestDetails(modelName)
 	if errMsg != nil {
 		errChan := make(chan *interfaces.ErrorMessage, 1)
 		errChan <- errMsg
 		close(errChan)
-		return nil, errChan
+		return nil, nil, errChan
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
@@ -482,7 +586,7 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 		SourceFormat:    sdktranslator.FromString(handlerType),
 	}
 	opts.Metadata = reqMeta
-	chunks, err := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
+	streamResult, err := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
 	if err != nil {
 		errChan := make(chan *interfaces.ErrorMessage, 1)
 		status := http.StatusInternalServerError
@@ -499,8 +603,19 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 		}
 		errChan <- &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
 		close(errChan)
-		return nil, errChan
+		return nil, nil, errChan
 	}
+	passthroughHeadersEnabled := PassthroughHeadersEnabled(h.Cfg)
+	// Capture upstream headers from the initial connection synchronously before the goroutine starts.
+	// Keep a mutable map so bootstrap retries can replace it before first payload is sent.
+	var upstreamHeaders http.Header
+	if passthroughHeadersEnabled {
+		upstreamHeaders = cloneHeader(FilterUpstreamHeaders(streamResult.Headers))
+		if upstreamHeaders == nil {
+			upstreamHeaders = make(http.Header)
+		}
+	}
+	chunks := streamResult.Chunks
 	dataChan := make(chan []byte)
 	errChan := make(chan *interfaces.ErrorMessage, 1)
 	go func() {
@@ -574,9 +689,12 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 					if !sentPayload {
 						if bootstrapRetries < maxBootstrapRetries && bootstrapEligible(streamErr) {
 							bootstrapRetries++
-							retryChunks, retryErr := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
+							retryResult, retryErr := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
 							if retryErr == nil {
-								chunks = retryChunks
+								if passthroughHeadersEnabled {
+									replaceHeader(upstreamHeaders, FilterUpstreamHeaders(retryResult.Headers))
+								}
+								chunks = retryResult.Chunks
 								continue outer
 							}
 							streamErr = retryErr
@@ -607,7 +725,7 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 			}
 		}
 	}()
-	return dataChan, errChan
+	return dataChan, upstreamHeaders, errChan
 }

 func statusFromError(err error) int {
@@ -667,13 +785,33 @@ func cloneBytes(src []byte) []byte {
 	return dst
 }

+func cloneHeader(src http.Header) http.Header {
+	if src == nil {
+		return nil
+	}
+	dst := make(http.Header, len(src))
+	for key, values := range src {
+		dst[key] = append([]string(nil), values...)
+	}
+	return dst
+}
+
+func replaceHeader(dst http.Header, src http.Header) {
+	for key := range dst {
+		delete(dst, key)
+	}
+	for key, values := range src {
+		dst[key] = append([]string(nil), values...)
+	}
+}
+
 // WriteErrorResponse writes an error message to the response writer using the HTTP status embedded in the message.
 func (h *BaseAPIHandler) WriteErrorResponse(c *gin.Context, msg *interfaces.ErrorMessage) {
 	status := http.StatusInternalServerError
 	if msg != nil && msg.StatusCode > 0 {
 		status = msg.StatusCode
 	}
-	if msg != nil && msg.Addon != nil {
+	if msg != nil && msg.Addon != nil && PassthroughHeadersEnabled(h.Cfg) {
 		for key, values := range msg.Addon {
 			if len(values) == 0 {
 				continue
--- a/sdk/api/handlers/handlers_error_response_test.go
+++ b/sdk/api/handlers/handlers_error_response_test.go
@@ -0,0 +1,68 @@
+package handlers
+
+import (
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"reflect"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	sdkconfig "github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
+)
+
+func TestWriteErrorResponse_AddonHeadersDisabledByDefault(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodGet, "/", nil)
+
+	handler := NewBaseAPIHandlers(nil, nil)
+	handler.WriteErrorResponse(c, &interfaces.ErrorMessage{
+		StatusCode: http.StatusTooManyRequests,
+		Error:      errors.New("rate limit"),
+		Addon: http.Header{
+			"Retry-After":  {"30"},
+			"X-Request-Id": {"req-1"},
+		},
+	})
+
+	if recorder.Code != http.StatusTooManyRequests {
+		t.Fatalf("status = %d, want %d", recorder.Code, http.StatusTooManyRequests)
+	}
+	if got := recorder.Header().Get("Retry-After"); got != "" {
+		t.Fatalf("Retry-After should be empty when passthrough is disabled, got %q", got)
+	}
+	if got := recorder.Header().Get("X-Request-Id"); got != "" {
+		t.Fatalf("X-Request-Id should be empty when passthrough is disabled, got %q", got)
+	}
+}
+
+func TestWriteErrorResponse_AddonHeadersEnabled(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodGet, "/", nil)
+	c.Writer.Header().Set("X-Request-Id", "old-value")
+
+	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{PassthroughHeaders: true}, nil)
+	handler.WriteErrorResponse(c, &interfaces.ErrorMessage{
+		StatusCode: http.StatusTooManyRequests,
+		Error:      errors.New("rate limit"),
+		Addon: http.Header{
+			"Retry-After":  {"30"},
+			"X-Request-Id": {"new-1", "new-2"},
+		},
+	})
+
+	if recorder.Code != http.StatusTooManyRequests {
+		t.Fatalf("status = %d, want %d", recorder.Code, http.StatusTooManyRequests)
+	}
+	if got := recorder.Header().Get("Retry-After"); got != "30" {
+		t.Fatalf("Retry-After = %q, want %q", got, "30")
+	}
+	if got := recorder.Header().Values("X-Request-Id"); !reflect.DeepEqual(got, []string{"new-1", "new-2"}) {
+		t.Fatalf("X-Request-Id = %#v, want %#v", got, []string{"new-1", "new-2"})
+	}
+}
--- a/sdk/api/handlers/handlers_stream_bootstrap_test.go
+++ b/sdk/api/handlers/handlers_stream_bootstrap_test.go
@@ -23,7 +23,7 @@ func (e *failOnceStreamExecutor) Execute(context.Context, *coreauth.Auth, coreex
 	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "Execute not implemented"}
 }

-func (e *failOnceStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+func (e *failOnceStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (*coreexecutor.StreamResult, error) {
 	e.mu.Lock()
 	e.calls++
 	call := e.calls
@@ -40,12 +40,18 @@ func (e *failOnceStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth,
 			},
 		}
 		close(ch)
-		return ch, nil
+		return &coreexecutor.StreamResult{
+			Headers: http.Header{"X-Upstream-Attempt": {"1"}},
+			Chunks:  ch,
+		}, nil
 	}

 	ch <- coreexecutor.StreamChunk{Payload: []byte("ok")}
 	close(ch)
-	return ch, nil
+	return &coreexecutor.StreamResult{
+		Headers: http.Header{"X-Upstream-Attempt": {"2"}},
+		Chunks:  ch,
+	}, nil
 }

 func (e *failOnceStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
@@ -81,7 +87,7 @@ func (e *payloadThenErrorStreamExecutor) Execute(context.Context, *coreauth.Auth
 	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "Execute not implemented"}
 }

-func (e *payloadThenErrorStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+func (e *payloadThenErrorStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (*coreexecutor.StreamResult, error) {
 	e.mu.Lock()
 	e.calls++
 	e.mu.Unlock()
@@ -97,7 +103,7 @@ func (e *payloadThenErrorStreamExecutor) ExecuteStream(context.Context, *coreaut
 		},
 	}
 	close(ch)
-	return ch, nil
+	return &coreexecutor.StreamResult{Chunks: ch}, nil
 }

 func (e *payloadThenErrorStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
@@ -122,6 +128,82 @@ func (e *payloadThenErrorStreamExecutor) Calls() int {
 	return e.calls
 }

+type authAwareStreamExecutor struct {
+	mu      sync.Mutex
+	calls   int
+	authIDs []string
+}
+
+func (e *authAwareStreamExecutor) Identifier() string { return "codex" }
+
+func (e *authAwareStreamExecutor) Execute(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
+	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "Execute not implemented"}
+}
+
+func (e *authAwareStreamExecutor) ExecuteStream(ctx context.Context, auth *coreauth.Auth, req coreexecutor.Request, opts coreexecutor.Options) (*coreexecutor.StreamResult, error) {
+	_ = ctx
+	_ = req
+	_ = opts
+	ch := make(chan coreexecutor.StreamChunk, 1)
+
+	authID := ""
+	if auth != nil {
+		authID = auth.ID
+	}
+
+	e.mu.Lock()
+	e.calls++
+	e.authIDs = append(e.authIDs, authID)
+	e.mu.Unlock()
+
+	if authID == "auth1" {
+		ch <- coreexecutor.StreamChunk{
+			Err: &coreauth.Error{
+				Code:       "unauthorized",
+				Message:    "unauthorized",
+				Retryable:  false,
+				HTTPStatus: http.StatusUnauthorized,
+			},
+		}
+		close(ch)
+		return &coreexecutor.StreamResult{Chunks: ch}, nil
+	}
+
+	ch <- coreexecutor.StreamChunk{Payload: []byte("ok")}
+	close(ch)
+	return &coreexecutor.StreamResult{Chunks: ch}, nil
+}
+
+func (e *authAwareStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
+	return auth, nil
+}
+
+func (e *authAwareStreamExecutor) CountTokens(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
+	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "CountTokens not implemented"}
+}
+
+func (e *authAwareStreamExecutor) HttpRequest(ctx context.Context, auth *coreauth.Auth, req *http.Request) (*http.Response, error) {
+	return nil, &coreauth.Error{
+		Code:       "not_implemented",
+		Message:    "HttpRequest not implemented",
+		HTTPStatus: http.StatusNotImplemented,
+	}
+}
+
+func (e *authAwareStreamExecutor) Calls() int {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.calls
+}
+
+func (e *authAwareStreamExecutor) AuthIDs() []string {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	out := make([]string, len(e.authIDs))
+	copy(out, e.authIDs)
+	return out
+}
+
 func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 	executor := &failOnceStreamExecutor{}
 	manager := coreauth.NewManager(nil, nil, nil)
@@ -154,12 +236,78 @@ func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 		registry.GetGlobalRegistry().UnregisterClient(auth2.ID)
 	})

+	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{
+		PassthroughHeaders: true,
+		Streaming: sdkconfig.StreamingConfig{
+			BootstrapRetries: 1,
+		},
+	}, manager)
+	dataChan, upstreamHeaders, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	if dataChan == nil || errChan == nil {
+		t.Fatalf("expected non-nil channels")
+	}
+
+	var got []byte
+	for chunk := range dataChan {
+		got = append(got, chunk...)
+	}
+
+	for msg := range errChan {
+		if msg != nil {
+			t.Fatalf("unexpected error: %+v", msg)
+		}
+	}
+
+	if string(got) != "ok" {
+		t.Fatalf("expected payload ok, got %q", string(got))
+	}
+	if executor.Calls() != 2 {
+		t.Fatalf("expected 2 stream attempts, got %d", executor.Calls())
+	}
+	upstreamAttemptHeader := upstreamHeaders.Get("X-Upstream-Attempt")
+	if upstreamAttemptHeader != "2" {
+		t.Fatalf("expected upstream header from retry attempt, got %q", upstreamAttemptHeader)
+	}
+}
+
+func TestExecuteStreamWithAuthManager_HeaderPassthroughDisabledByDefault(t *testing.T) {
+	executor := &failOnceStreamExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+
+	auth1 := &coreauth.Auth{
+		ID:       "auth1",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test1@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth1); err != nil {
+		t.Fatalf("manager.Register(auth1): %v", err)
+	}
+
+	auth2 := &coreauth.Auth{
+		ID:       "auth2",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test2@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth2); err != nil {
+		t.Fatalf("manager.Register(auth2): %v", err)
+	}
+
+	registry.GetGlobalRegistry().RegisterClient(auth1.ID, auth1.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	registry.GetGlobalRegistry().RegisterClient(auth2.ID, auth2.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth1.ID)
+		registry.GetGlobalRegistry().UnregisterClient(auth2.ID)
+	})
+
 	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{
 		Streaming: sdkconfig.StreamingConfig{
 			BootstrapRetries: 1,
 		},
 	}, manager)
-	dataChan, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	dataChan, upstreamHeaders, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
 	if dataChan == nil || errChan == nil {
 		t.Fatalf("expected non-nil channels")
 	}
@@ -168,7 +316,6 @@ func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 	for chunk := range dataChan {
 		got = append(got, chunk...)
 	}
-
 	for msg := range errChan {
 		if msg != nil {
 			t.Fatalf("unexpected error: %+v", msg)
@@ -178,8 +325,8 @@ func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 	if string(got) != "ok" {
 		t.Fatalf("expected payload ok, got %q", string(got))
 	}
-	if executor.Calls() != 2 {
-		t.Fatalf("expected 2 stream attempts, got %d", executor.Calls())
+	if upstreamHeaders != nil {
+		t.Fatalf("expected nil upstream headers when passthrough is disabled, got %#v", upstreamHeaders)
 	}
 }

@@ -220,7 +367,7 @@ func TestExecuteStreamWithAuthManager_DoesNotRetryAfterFirstByte(t *testing.T) {
 			BootstrapRetries: 1,
 		},
 	}, manager)
-	dataChan, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	dataChan, _, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
 	if dataChan == nil || errChan == nil {
 		t.Fatalf("expected non-nil channels")
 	}
@@ -252,3 +399,128 @@ func TestExecuteStreamWithAuthManager_DoesNotRetryAfterFirstByte(t *testing.T) {
 		t.Fatalf("expected 1 stream attempt, got %d", executor.Calls())
 	}
 }
+
+func TestExecuteStreamWithAuthManager_PinnedAuthKeepsSameUpstream(t *testing.T) {
+	executor := &authAwareStreamExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+
+	auth1 := &coreauth.Auth{
+		ID:       "auth1",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test1@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth1); err != nil {
+		t.Fatalf("manager.Register(auth1): %v", err)
+	}
+
+	auth2 := &coreauth.Auth{
+		ID:       "auth2",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test2@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth2); err != nil {
+		t.Fatalf("manager.Register(auth2): %v", err)
+	}
+
+	registry.GetGlobalRegistry().RegisterClient(auth1.ID, auth1.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	registry.GetGlobalRegistry().RegisterClient(auth2.ID, auth2.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth1.ID)
+		registry.GetGlobalRegistry().UnregisterClient(auth2.ID)
+	})
+
+	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{
+		Streaming: sdkconfig.StreamingConfig{
+			BootstrapRetries: 1,
+		},
+	}, manager)
+	ctx := WithPinnedAuthID(context.Background(), "auth1")
+	dataChan, _, errChan := handler.ExecuteStreamWithAuthManager(ctx, "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	if dataChan == nil || errChan == nil {
+		t.Fatalf("expected non-nil channels")
+	}
+
+	var got []byte
+	for chunk := range dataChan {
+		got = append(got, chunk...)
+	}
+
+	var gotErr error
+	for msg := range errChan {
+		if msg != nil && msg.Error != nil {
+			gotErr = msg.Error
+		}
+	}
+
+	if len(got) != 0 {
+		t.Fatalf("expected empty payload, got %q", string(got))
+	}
+	if gotErr == nil {
+		t.Fatalf("expected terminal error, got nil")
+	}
+	authIDs := executor.AuthIDs()
+	if len(authIDs) == 0 {
+		t.Fatalf("expected at least one upstream attempt")
+	}
+	for _, authID := range authIDs {
+		if authID != "auth1" {
+			t.Fatalf("expected all attempts on auth1, got sequence %v", authIDs)
+		}
+	}
+}
+
+func TestExecuteStreamWithAuthManager_SelectedAuthCallbackReceivesAuthID(t *testing.T) {
+	executor := &authAwareStreamExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+
+	auth2 := &coreauth.Auth{
+		ID:       "auth2",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test2@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth2); err != nil {
+		t.Fatalf("manager.Register(auth2): %v", err)
+	}
+
+	registry.GetGlobalRegistry().RegisterClient(auth2.ID, auth2.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth2.ID)
+	})
+
+	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{
+		Streaming: sdkconfig.StreamingConfig{
+			BootstrapRetries: 0,
+		},
+	}, manager)
+
+	selectedAuthID := ""
+	ctx := WithSelectedAuthIDCallback(context.Background(), func(authID string) {
+		selectedAuthID = authID
+	})
+	dataChan, _, errChan := handler.ExecuteStreamWithAuthManager(ctx, "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	if dataChan == nil || errChan == nil {
+		t.Fatalf("expected non-nil channels")
+	}
+
+	var got []byte
+	for chunk := range dataChan {
+		got = append(got, chunk...)
+	}
+	for msg := range errChan {
+		if msg != nil {
+			t.Fatalf("unexpected error: %+v", msg)
+		}
+	}
+
+	if string(got) != "ok" {
+		t.Fatalf("expected payload ok, got %q", string(got))
+	}
+	if selectedAuthID != "auth2" {
+		t.Fatalf("selectedAuthID = %q, want %q", selectedAuthID, "auth2")
+	}
+}
--- a/sdk/api/handlers/header_filter.go
+++ b/sdk/api/handlers/header_filter.go
@@ -0,0 +1,80 @@
+package handlers
+
+import (
+	"net/http"
+	"strings"
+)
+
+// hopByHopHeaders lists RFC 7230 Section 6.1 hop-by-hop headers that MUST NOT
+// be forwarded by proxies, plus security-sensitive headers that should not leak.
+var hopByHopHeaders = map[string]struct{}{
+	// RFC 7230 hop-by-hop
+	"Connection":          {},
+	"Keep-Alive":          {},
+	"Proxy-Authenticate":  {},
+	"Proxy-Authorization": {},
+	"Te":                  {},
+	"Trailer":             {},
+	"Transfer-Encoding":   {},
+	"Upgrade":             {},
+	// Security-sensitive
+	"Set-Cookie": {},
+	// CPA-managed (set by handlers, not upstream)
+	"Content-Length":   {},
+	"Content-Encoding": {},
+}
+
+// FilterUpstreamHeaders returns a copy of src with hop-by-hop and security-sensitive
+// headers removed. Returns nil if src is nil or empty after filtering.
+func FilterUpstreamHeaders(src http.Header) http.Header {
+	if src == nil {
+		return nil
+	}
+	connectionScoped := connectionScopedHeaders(src)
+	dst := make(http.Header)
+	for key, values := range src {
+		canonicalKey := http.CanonicalHeaderKey(key)
+		if _, blocked := hopByHopHeaders[canonicalKey]; blocked {
+			continue
+		}
+		if _, scoped := connectionScoped[canonicalKey]; scoped {
+			continue
+		}
+		dst[key] = values
+	}
+	if len(dst) == 0 {
+		return nil
+	}
+	return dst
+}
+
+func connectionScopedHeaders(src http.Header) map[string]struct{} {
+	scoped := make(map[string]struct{})
+	for _, rawValue := range src.Values("Connection") {
+		for _, token := range strings.Split(rawValue, ",") {
+			headerName := strings.TrimSpace(token)
+			if headerName == "" {
+				continue
+			}
+			scoped[http.CanonicalHeaderKey(headerName)] = struct{}{}
+		}
+	}
+	return scoped
+}
+
+// WriteUpstreamHeaders writes filtered upstream headers to the gin response writer.
+// Headers already set by CPA (e.g., Content-Type) are NOT overwritten.
+func WriteUpstreamHeaders(dst http.Header, src http.Header) {
+	if src == nil {
+		return
+	}
+	for key, values := range src {
+		// Don't overwrite headers already set by CPA handlers
+		if dst.Get(key) != "" {
+			continue
+		}
+		for _, v := range values {
+			dst.Add(key, v)
+		}
+	}
+}
--- a/sdk/api/handlers/header_filter_test.go
+++ b/sdk/api/handlers/header_filter_test.go
@@ -0,0 +1,55 @@
+package handlers
+
+import (
+	"net/http"
+	"testing"
+)
+
+func TestFilterUpstreamHeaders_RemovesConnectionScopedHeaders(t *testing.T) {
+	src := http.Header{}
+	src.Add("Connection", "keep-alive, x-hop-a, x-hop-b")
+	src.Add("Connection", "x-hop-c")
+	src.Set("Keep-Alive", "timeout=5")
+	src.Set("X-Hop-A", "a")
+	src.Set("X-Hop-B", "b")
+	src.Set("X-Hop-C", "c")
+	src.Set("X-Request-Id", "req-1")
+	src.Set("Set-Cookie", "session=secret")
+
+	filtered := FilterUpstreamHeaders(src)
+	if filtered == nil {
+		t.Fatalf("expected filtered headers, got nil")
+	}
+
+	requestID := filtered.Get("X-Request-Id")
+	if requestID != "req-1" {
+		t.Fatalf("expected X-Request-Id to be preserved, got %q", requestID)
+	}
+
+	blockedHeaderKeys := []string{
+		"Connection",
+		"Keep-Alive",
+		"X-Hop-A",
+		"X-Hop-B",
+		"X-Hop-C",
+		"Set-Cookie",
+	}
+	for _, key := range blockedHeaderKeys {
+		value := filtered.Get(key)
+		if value != "" {
+			t.Fatalf("expected %s to be removed, got %q", key, value)
+		}
+	}
+}
+
+func TestFilterUpstreamHeaders_ReturnsNilWhenAllHeadersBlocked(t *testing.T) {
+	src := http.Header{}
+	src.Add("Connection", "x-hop-a")
+	src.Set("X-Hop-A", "a")
+	src.Set("Set-Cookie", "session=secret")
+
+	filtered := FilterUpstreamHeaders(src)
+	if filtered != nil {
+		t.Fatalf("expected nil when all headers are filtered, got %#v", filtered)
+	}
+}
--- a/sdk/api/handlers/openai/openai_handlers.go
+++ b/sdk/api/handlers/openai/openai_handlers.go
@@ -420,6 +420,7 @@ func convertChatCompletionsStreamChunkToCompletions(chunkData []byte) []byte {

 	// Check if this chunk has any meaningful content
 	hasContent := false
+	hasUsage := root.Get("usage").Exists()
 	if chatChoices := root.Get("choices"); chatChoices.Exists() && chatChoices.IsArray() {
 		chatChoices.ForEach(func(_, choice gjson.Result) bool {
 			// Check if delta has content or finish_reason
@@ -438,8 +439,8 @@ func convertChatCompletionsStreamChunkToCompletions(chunkData []byte) []byte {
 		})
 	}

-	// If no meaningful content, return nil to indicate this chunk should be skipped
-	if !hasContent {
+	// If no meaningful content and no usage, return nil to indicate this chunk should be skipped
+	if !hasContent && !hasUsage {
 		return nil
 	}

@@ -498,6 +499,11 @@ func convertChatCompletionsStreamChunkToCompletions(chunkData []byte) []byte {
 		out, _ = sjson.SetRaw(out, "choices", string(choicesJSON))
 	}

+	// Copy usage if present
+	if usage := root.Get("usage"); usage.Exists() {
+		out, _ = sjson.SetRaw(out, "usage", usage.Raw)
+	}
+
 	return []byte(out)
 }

@@ -513,12 +519,13 @@ func (h *OpenAIAPIHandler) handleNonStreamingResponse(c *gin.Context, rawJSON []

 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -528,12 +535,13 @@ func (h *OpenAIAPIHandler) handleNonStreamingResponseViaResponses(c *gin.Context

 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, OpenaiResponse, modelName, rawJSON, h.GetAlt(c))
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, OpenaiResponse, modelName, rawJSON, h.GetAlt(c))
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	converted := convertResponsesObjectToChatCompletion(cliCtx, modelName, originalChatJSON, rawJSON, resp)
 	if converted == nil {
 		h.WriteErrorResponse(c, &interfaces.ErrorMessage{
@@ -569,7 +577,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON []byt

 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))

 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -602,6 +610,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON []byt
 			if !ok {
 				// Stream closed without data? Send DONE or just headers.
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = fmt.Fprintf(c.Writer, "data: [DONE]\n\n")
 				flusher.Flush()
 				cliCancel(nil)
@@ -610,6 +619,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON []byt

 			// Success! Commit to streaming headers.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)

 			_, _ = fmt.Fprintf(c.Writer, "data: %s\n\n", string(chunk))
 			flusher.Flush()
@@ -635,7 +645,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponseViaResponses(c *gin.Context, r

 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, OpenaiResponse, modelName, rawJSON, h.GetAlt(c))
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, OpenaiResponse, modelName, rawJSON, h.GetAlt(c))
 	var param any

 	setSSEHeaders := func() {
@@ -666,6 +676,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponseViaResponses(c *gin.Context, r
 		case chunk, ok := <-dataChan:
 			if !ok {
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = fmt.Fprintf(c.Writer, "data: [DONE]\n\n")
 				flusher.Flush()
 				cliCancel(nil)
@@ -673,6 +684,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponseViaResponses(c *gin.Context, r
 			}

 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 			writeConvertedResponsesChunk(c, cliCtx, modelName, originalChatJSON, rawJSON, chunk, &param)
 			flusher.Flush()

@@ -698,13 +710,14 @@ func (h *OpenAIAPIHandler) handleCompletionsNonStreamingResponse(c *gin.Context,
 	modelName := gjson.GetBytes(chatCompletionsJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	completionsResp := convertChatCompletionsResponseToCompletions(resp)
 	_, _ = c.Writer.Write(completionsResp)
 	cliCancel()
@@ -735,7 +748,7 @@ func (h *OpenAIAPIHandler) handleCompletionsStreamingResponse(c *gin.Context, ra

 	modelName := gjson.GetBytes(chatCompletionsJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")

 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -766,6 +779,7 @@ func (h *OpenAIAPIHandler) handleCompletionsStreamingResponse(c *gin.Context, ra
 		case chunk, ok := <-dataChan:
 			if !ok {
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = fmt.Fprintf(c.Writer, "data: [DONE]\n\n")
 				flusher.Flush()
 				cliCancel(nil)
@@ -774,6 +788,7 @@ func (h *OpenAIAPIHandler) handleCompletionsStreamingResponse(c *gin.Context, ra

 			// Success! Set headers.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)

 			// Write the first chunk
 			converted := convertChatCompletionsStreamChunkToCompletions(chunk)
--- a/sdk/api/handlers/openai/openai_responses_compact_test.go
+++ b/sdk/api/handlers/openai/openai_responses_compact_test.go
@@ -31,7 +31,7 @@ func (e *compactCaptureExecutor) Execute(ctx context.Context, auth *coreauth.Aut
 	return coreexecutor.Response{Payload: []byte(`{"ok":true}`)}, nil
 }

-func (e *compactCaptureExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+func (e *compactCaptureExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (*coreexecutor.StreamResult, error) {
 	return nil, errors.New("not implemented")
 }

--- a/sdk/api/handlers/openai/openai_responses_handlers.go
+++ b/sdk/api/handlers/openai/openai_responses_handlers.go
@@ -139,13 +139,14 @@ func (h *OpenAIResponsesAPIHandler) Compact(c *gin.Context) {
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "responses/compact")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "responses/compact")
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -164,13 +165,14 @@ func (h *OpenAIResponsesAPIHandler) handleNonStreamingResponse(c *gin.Context, r
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)

-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -180,12 +182,13 @@ func (h *OpenAIResponsesAPIHandler) handleNonStreamingResponseViaChat(c *gin.Con

 	modelName := gjson.GetBytes(chatJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, OpenAI, modelName, chatJSON, "")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, OpenAI, modelName, chatJSON, "")
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	var param any
 	converted := responsesconverter.ConvertOpenAIChatCompletionsResponseToOpenAIResponsesNonStream(cliCtx, modelName, originalResponsesJSON, originalResponsesJSON, resp, &param)
 	if converted == "" {
@@ -223,7 +226,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponse(c *gin.Context, rawJ
 	// New core execution path
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")

 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -256,6 +259,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponse(c *gin.Context, rawJ
 			if !ok {
 				// Stream closed without data? Send headers and done.
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = c.Writer.Write([]byte("\n"))
 				flusher.Flush()
 				cliCancel(nil)
@@ -264,6 +268,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponse(c *gin.Context, rawJ

 			// Success! Set headers.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)

 			// Write first chunk logic (matching forwardResponsesStream)
 			if bytes.HasPrefix(chunk, []byte("event:")) {
@@ -294,7 +299,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponseViaChat(c *gin.Contex

 	modelName := gjson.GetBytes(chatJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, OpenAI, modelName, chatJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, OpenAI, modelName, chatJSON, "")
 	var param any

 	setSSEHeaders := func() {
@@ -324,6 +329,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponseViaChat(c *gin.Contex
 		case chunk, ok := <-dataChan:
 			if !ok {
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = c.Writer.Write([]byte("\n"))
 				flusher.Flush()
 				cliCancel(nil)
@@ -331,6 +337,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponseViaChat(c *gin.Contex
 			}

 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 			writeChatAsResponsesChunk(c, cliCtx, modelName, originalResponsesJSON, chunk, &param)
 			flusher.Flush()

--- a/sdk/api/handlers/openai/openai_responses_websocket.go
+++ b/sdk/api/handlers/openai/openai_responses_websocket.go
@@ -0,0 +1,662 @@
+package openai
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/uuid"
+	"github.com/gorilla/websocket"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/api/handlers"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+	log "github.com/sirupsen/logrus"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+const (
+	wsRequestTypeCreate  = "response.create"
+	wsRequestTypeAppend  = "response.append"
+	wsEventTypeError     = "error"
+	wsEventTypeCompleted = "response.completed"
+	wsEventTypeDone      = "response.done"
+	wsDoneMarker         = "[DONE]"
+	wsTurnStateHeader    = "x-codex-turn-state"
+	wsRequestBodyKey     = "REQUEST_BODY_OVERRIDE"
+	wsPayloadLogMaxSize  = 2048
+)
+
+var responsesWebsocketUpgrader = websocket.Upgrader{
+	ReadBufferSize:  4096,
+	WriteBufferSize: 4096,
+	CheckOrigin: func(r *http.Request) bool {
+		return true
+	},
+}
+
+// ResponsesWebsocket handles websocket requests for /v1/responses.
+// It accepts `response.create` and `response.append` requests and streams
+// response events back as JSON websocket text messages.
+func (h *OpenAIResponsesAPIHandler) ResponsesWebsocket(c *gin.Context) {
+	conn, err := responsesWebsocketUpgrader.Upgrade(c.Writer, c.Request, websocketUpgradeHeaders(c.Request))
+	if err != nil {
+		return
+	}
+	passthroughSessionID := uuid.NewString()
+	clientRemoteAddr := ""
+	if c != nil && c.Request != nil {
+		clientRemoteAddr = strings.TrimSpace(c.Request.RemoteAddr)
+	}
+	log.Infof("responses websocket: client connected id=%s remote=%s", passthroughSessionID, clientRemoteAddr)
+	var wsTerminateErr error
+	var wsBodyLog strings.Builder
+	defer func() {
+		if wsTerminateErr != nil {
+			// log.Infof("responses websocket: session closing id=%s reason=%v", passthroughSessionID, wsTerminateErr)
+		} else {
+			log.Infof("responses websocket: session closing id=%s", passthroughSessionID)
+		}
+		if h != nil && h.AuthManager != nil {
+			h.AuthManager.CloseExecutionSession(passthroughSessionID)
+			log.Infof("responses websocket: upstream execution session closed id=%s", passthroughSessionID)
+		}
+		setWebsocketRequestBody(c, wsBodyLog.String())
+		if errClose := conn.Close(); errClose != nil {
+			log.Warnf("responses websocket: close connection error: %v", errClose)
+		}
+	}()
+
+	var lastRequest []byte
+	lastResponseOutput := []byte("[]")
+	pinnedAuthID := ""
+
+	for {
+		msgType, payload, errReadMessage := conn.ReadMessage()
+		if errReadMessage != nil {
+			wsTerminateErr = errReadMessage
+			appendWebsocketEvent(&wsBodyLog, "disconnect", []byte(errReadMessage.Error()))
+			if websocket.IsCloseError(errReadMessage, websocket.CloseNormalClosure, websocket.CloseGoingAway, websocket.CloseNoStatusReceived) {
+				log.Infof("responses websocket: client disconnected id=%s error=%v", passthroughSessionID, errReadMessage)
+			} else {
+				// log.Warnf("responses websocket: read message failed id=%s error=%v", passthroughSessionID, errReadMessage)
+			}
+			return
+		}
+		if msgType != websocket.TextMessage && msgType != websocket.BinaryMessage {
+			continue
+		}
+		// log.Infof(
+		// 	"responses websocket: downstream_in id=%s type=%d event=%s payload=%s",
+		// 	passthroughSessionID,
+		// 	msgType,
+		// 	websocketPayloadEventType(payload),
+		// 	websocketPayloadPreview(payload),
+		// )
+		appendWebsocketEvent(&wsBodyLog, "request", payload)
+
+		allowIncrementalInputWithPreviousResponseID := websocketUpstreamSupportsIncrementalInput(nil, nil)
+		if pinnedAuthID != "" && h != nil && h.AuthManager != nil {
+			if pinnedAuth, ok := h.AuthManager.GetByID(pinnedAuthID); ok && pinnedAuth != nil {
+				allowIncrementalInputWithPreviousResponseID = websocketUpstreamSupportsIncrementalInput(pinnedAuth.Attributes, pinnedAuth.Metadata)
+			}
+		}
+
+		var requestJSON []byte
+		var updatedLastRequest []byte
+		var errMsg *interfaces.ErrorMessage
+		requestJSON, updatedLastRequest, errMsg = normalizeResponsesWebsocketRequestWithMode(
+			payload,
+			lastRequest,
+			lastResponseOutput,
+			allowIncrementalInputWithPreviousResponseID,
+		)
+		if errMsg != nil {
+			h.LoggingAPIResponseError(context.WithValue(context.Background(), "gin", c), errMsg)
+			markAPIResponseTimestamp(c)
+			errorPayload, errWrite := writeResponsesWebsocketError(conn, errMsg)
+			appendWebsocketEvent(&wsBodyLog, "response", errorPayload)
+			log.Infof(
+				"responses websocket: downstream_out id=%s type=%d event=%s payload=%s",
+				passthroughSessionID,
+				websocket.TextMessage,
+				websocketPayloadEventType(errorPayload),
+				websocketPayloadPreview(errorPayload),
+			)
+			if errWrite != nil {
+				log.Warnf(
+					"responses websocket: downstream_out write failed id=%s event=%s error=%v",
+					passthroughSessionID,
+					websocketPayloadEventType(errorPayload),
+					errWrite,
+				)
+				return
+			}
+			continue
+		}
+		lastRequest = updatedLastRequest
+
+		modelName := gjson.GetBytes(requestJSON, "model").String()
+		cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
+		cliCtx = cliproxyexecutor.WithDownstreamWebsocket(cliCtx)
+		cliCtx = handlers.WithExecutionSessionID(cliCtx, passthroughSessionID)
+		if pinnedAuthID != "" {
+			cliCtx = handlers.WithPinnedAuthID(cliCtx, pinnedAuthID)
+		} else {
+			cliCtx = handlers.WithSelectedAuthIDCallback(cliCtx, func(authID string) {
+				pinnedAuthID = strings.TrimSpace(authID)
+			})
+		}
+		dataChan, _, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, requestJSON, "")
+
+		completedOutput, errForward := h.forwardResponsesWebsocket(c, conn, cliCancel, dataChan, errChan, &wsBodyLog, passthroughSessionID)
+		if errForward != nil {
+			wsTerminateErr = errForward
+			appendWebsocketEvent(&wsBodyLog, "disconnect", []byte(errForward.Error()))
+			log.Warnf("responses websocket: forward failed id=%s error=%v", passthroughSessionID, errForward)
+			return
+		}
+		lastResponseOutput = completedOutput
+	}
+}
+
+func websocketUpgradeHeaders(req *http.Request) http.Header {
+	headers := http.Header{}
+	if req == nil {
+		return headers
+	}
+
+	// Keep the same sticky turn-state across reconnects when provided by the client.
+	turnState := strings.TrimSpace(req.Header.Get(wsTurnStateHeader))
+	if turnState != "" {
+		headers.Set(wsTurnStateHeader, turnState)
+	}
+	return headers
+}
+
+func normalizeResponsesWebsocketRequest(rawJSON []byte, lastRequest []byte, lastResponseOutput []byte) ([]byte, []byte, *interfaces.ErrorMessage) {
+	return normalizeResponsesWebsocketRequestWithMode(rawJSON, lastRequest, lastResponseOutput, true)
+}
+
+func normalizeResponsesWebsocketRequestWithMode(rawJSON []byte, lastRequest []byte, lastResponseOutput []byte, allowIncrementalInputWithPreviousResponseID bool) ([]byte, []byte, *interfaces.ErrorMessage) {
+	requestType := strings.TrimSpace(gjson.GetBytes(rawJSON, "type").String())
+	switch requestType {
+	case wsRequestTypeCreate:
+		// log.Infof("responses websocket: response.create request")
+		if len(lastRequest) == 0 {
+			return normalizeResponseCreateRequest(rawJSON)
+		}
+		return normalizeResponseSubsequentRequest(rawJSON, lastRequest, lastResponseOutput, allowIncrementalInputWithPreviousResponseID)
+	case wsRequestTypeAppend:
+		// log.Infof("responses websocket: response.append request")
+		return normalizeResponseSubsequentRequest(rawJSON, lastRequest, lastResponseOutput, allowIncrementalInputWithPreviousResponseID)
+	default:
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("unsupported websocket request type: %s", requestType),
+		}
+	}
+}
+
+func normalizeResponseCreateRequest(rawJSON []byte) ([]byte, []byte, *interfaces.ErrorMessage) {
+	normalized, errDelete := sjson.DeleteBytes(rawJSON, "type")
+	if errDelete != nil {
+		normalized = bytes.Clone(rawJSON)
+	}
+	normalized, _ = sjson.SetBytes(normalized, "stream", true)
+	if !gjson.GetBytes(normalized, "input").Exists() {
+		normalized, _ = sjson.SetRawBytes(normalized, "input", []byte("[]"))
+	}
+
+	modelName := strings.TrimSpace(gjson.GetBytes(normalized, "model").String())
+	if modelName == "" {
+		return nil, nil, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("missing model in response.create request"),
+		}
+	}
+	return normalized, bytes.Clone(normalized), nil
+}
+
+func normalizeResponseSubsequentRequest(rawJSON []byte, lastRequest []byte, lastResponseOutput []byte, allowIncrementalInputWithPreviousResponseID bool) ([]byte, []byte, *interfaces.ErrorMessage) {
+	if len(lastRequest) == 0 {
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("websocket request received before response.create"),
+		}
+	}
+
+	nextInput := gjson.GetBytes(rawJSON, "input")
+	if !nextInput.Exists() || !nextInput.IsArray() {
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("websocket request requires array field: input"),
+		}
+	}
+
+	// Websocket v2 mode uses response.create with previous_response_id + incremental input.
+	// Do not expand it into a full input transcript; upstream expects the incremental payload.
+	if allowIncrementalInputWithPreviousResponseID {
+		if prev := strings.TrimSpace(gjson.GetBytes(rawJSON, "previous_response_id").String()); prev != "" {
+			normalized, errDelete := sjson.DeleteBytes(rawJSON, "type")
+			if errDelete != nil {
+				normalized = bytes.Clone(rawJSON)
+			}
+			if !gjson.GetBytes(normalized, "model").Exists() {
+				modelName := strings.TrimSpace(gjson.GetBytes(lastRequest, "model").String())
+				if modelName != "" {
+					normalized, _ = sjson.SetBytes(normalized, "model", modelName)
+				}
+			}
+			if !gjson.GetBytes(normalized, "instructions").Exists() {
+				instructions := gjson.GetBytes(lastRequest, "instructions")
+				if instructions.Exists() {
+					normalized, _ = sjson.SetRawBytes(normalized, "instructions", []byte(instructions.Raw))
+				}
+			}
+			normalized, _ = sjson.SetBytes(normalized, "stream", true)
+			return normalized, bytes.Clone(normalized), nil
+		}
+	}
+
+	existingInput := gjson.GetBytes(lastRequest, "input")
+	mergedInput, errMerge := mergeJSONArrayRaw(existingInput.Raw, normalizeJSONArrayRaw(lastResponseOutput))
+	if errMerge != nil {
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("invalid previous response output: %w", errMerge),
+		}
+	}
+
+	mergedInput, errMerge = mergeJSONArrayRaw(mergedInput, nextInput.Raw)
+	if errMerge != nil {
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("invalid request input: %w", errMerge),
+		}
+	}
+
+	normalized, errDelete := sjson.DeleteBytes(rawJSON, "type")
+	if errDelete != nil {
+		normalized = bytes.Clone(rawJSON)
+	}
+	normalized, _ = sjson.DeleteBytes(normalized, "previous_response_id")
+	var errSet error
+	normalized, errSet = sjson.SetRawBytes(normalized, "input", []byte(mergedInput))
+	if errSet != nil {
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("failed to merge websocket input: %w", errSet),
+		}
+	}
+	if !gjson.GetBytes(normalized, "model").Exists() {
+		modelName := strings.TrimSpace(gjson.GetBytes(lastRequest, "model").String())
+		if modelName != "" {
+			normalized, _ = sjson.SetBytes(normalized, "model", modelName)
+		}
+	}
+	if !gjson.GetBytes(normalized, "instructions").Exists() {
+		instructions := gjson.GetBytes(lastRequest, "instructions")
+		if instructions.Exists() {
+			normalized, _ = sjson.SetRawBytes(normalized, "instructions", []byte(instructions.Raw))
+		}
+	}
+	normalized, _ = sjson.SetBytes(normalized, "stream", true)
+	return normalized, bytes.Clone(normalized), nil
+}
+
+func websocketUpstreamSupportsIncrementalInput(attributes map[string]string, metadata map[string]any) bool {
+	if len(attributes) > 0 {
+		if raw := strings.TrimSpace(attributes["websockets"]); raw != "" {
+			parsed, errParse := strconv.ParseBool(raw)
+			if errParse == nil {
+				return parsed
+			}
+		}
+	}
+	if len(metadata) == 0 {
+		return false
+	}
+	raw, ok := metadata["websockets"]
+	if !ok || raw == nil {
+		return false
+	}
+	switch value := raw.(type) {
+	case bool:
+		return value
+	case string:
+		parsed, errParse := strconv.ParseBool(strings.TrimSpace(value))
+		if errParse == nil {
+			return parsed
+		}
+	default:
+	}
+	return false
+}
+
+func mergeJSONArrayRaw(existingRaw, appendRaw string) (string, error) {
+	existingRaw = strings.TrimSpace(existingRaw)
+	appendRaw = strings.TrimSpace(appendRaw)
+	if existingRaw == "" {
+		existingRaw = "[]"
+	}
+	if appendRaw == "" {
+		appendRaw = "[]"
+	}
+
+	var existing []json.RawMessage
+	if err := json.Unmarshal([]byte(existingRaw), &existing); err != nil {
+		return "", err
+	}
+	var appendItems []json.RawMessage
+	if err := json.Unmarshal([]byte(appendRaw), &appendItems); err != nil {
+		return "", err
+	}
+
+	merged := append(existing, appendItems...)
+	out, err := json.Marshal(merged)
+	if err != nil {
+		return "", err
+	}
+	return string(out), nil
+}
+
+func normalizeJSONArrayRaw(raw []byte) string {
+	trimmed := strings.TrimSpace(string(raw))
+	if trimmed == "" {
+		return "[]"
+	}
+	result := gjson.Parse(trimmed)
+	if result.Type == gjson.JSON && result.IsArray() {
+		return trimmed
+	}
+	return "[]"
+}
+
+func (h *OpenAIResponsesAPIHandler) forwardResponsesWebsocket(
+	c *gin.Context,
+	conn *websocket.Conn,
+	cancel handlers.APIHandlerCancelFunc,
+	data <-chan []byte,
+	errs <-chan *interfaces.ErrorMessage,
+	wsBodyLog *strings.Builder,
+	sessionID string,
+) ([]byte, error) {
+	completed := false
+	completedOutput := []byte("[]")
+
+	for {
+		select {
+		case <-c.Request.Context().Done():
+			cancel(c.Request.Context().Err())
+			return completedOutput, c.Request.Context().Err()
+		case errMsg, ok := <-errs:
+			if !ok {
+				errs = nil
+				continue
+			}
+			if errMsg != nil {
+				h.LoggingAPIResponseError(context.WithValue(context.Background(), "gin", c), errMsg)
+				markAPIResponseTimestamp(c)
+				errorPayload, errWrite := writeResponsesWebsocketError(conn, errMsg)
+				appendWebsocketEvent(wsBodyLog, "response", errorPayload)
+				log.Infof(
+					"responses websocket: downstream_out id=%s type=%d event=%s payload=%s",
+					sessionID,
+					websocket.TextMessage,
+					websocketPayloadEventType(errorPayload),
+					websocketPayloadPreview(errorPayload),
+				)
+				if errWrite != nil {
+					// log.Warnf(
+					// 	"responses websocket: downstream_out write failed id=%s event=%s error=%v",
+					// 	sessionID,
+					// 	websocketPayloadEventType(errorPayload),
+					// 	errWrite,
+					// )
+					cancel(errMsg.Error)
+					return completedOutput, errWrite
+				}
+			}
+			if errMsg != nil {
+				cancel(errMsg.Error)
+			} else {
+				cancel(nil)
+			}
+			return completedOutput, nil
+		case chunk, ok := <-data:
+			if !ok {
+				if !completed {
+					errMsg := &interfaces.ErrorMessage{
+						StatusCode: http.StatusRequestTimeout,
+						Error:      fmt.Errorf("stream closed before response.completed"),
+					}
+					h.LoggingAPIResponseError(context.WithValue(context.Background(), "gin", c), errMsg)
+					markAPIResponseTimestamp(c)
+					errorPayload, errWrite := writeResponsesWebsocketError(conn, errMsg)
+					appendWebsocketEvent(wsBodyLog, "response", errorPayload)
+					log.Infof(
+						"responses websocket: downstream_out id=%s type=%d event=%s payload=%s",
+						sessionID,
+						websocket.TextMessage,
+						websocketPayloadEventType(errorPayload),
+						websocketPayloadPreview(errorPayload),
+					)
+					if errWrite != nil {
+						log.Warnf(
+							"responses websocket: downstream_out write failed id=%s event=%s error=%v",
+							sessionID,
+							websocketPayloadEventType(errorPayload),
+							errWrite,
+						)
+						cancel(errMsg.Error)
+						return completedOutput, errWrite
+					}
+					cancel(errMsg.Error)
+					return completedOutput, nil
+				}
+				cancel(nil)
+				return completedOutput, nil
+			}
+
+			payloads := websocketJSONPayloadsFromChunk(chunk)
+			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])
+				}
+				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,
+					)
+					cancel(errWrite)
+					return completedOutput, errWrite
+				}
+			}
+		}
+	}
+}
+
+func responseCompletedOutputFromPayload(payload []byte) []byte {
+	output := gjson.GetBytes(payload, "response.output")
+	if output.Exists() && output.IsArray() {
+		return bytes.Clone([]byte(output.Raw))
+	}
+	return []byte("[]")
+}
+
+func websocketJSONPayloadsFromChunk(chunk []byte) [][]byte {
+	payloads := make([][]byte, 0, 2)
+	lines := bytes.Split(chunk, []byte("\n"))
+	for i := range lines {
+		line := bytes.TrimSpace(lines[i])
+		if len(line) == 0 || bytes.HasPrefix(line, []byte("event:")) {
+			continue
+		}
+		if bytes.HasPrefix(line, []byte("data:")) {
+			line = bytes.TrimSpace(line[len("data:"):])
+		}
+		if len(line) == 0 || bytes.Equal(line, []byte(wsDoneMarker)) {
+			continue
+		}
+		if json.Valid(line) {
+			payloads = append(payloads, bytes.Clone(line))
+		}
+	}
+
+	if len(payloads) > 0 {
+		return payloads
+	}
+
+	trimmed := bytes.TrimSpace(chunk)
+	if bytes.HasPrefix(trimmed, []byte("data:")) {
+		trimmed = bytes.TrimSpace(trimmed[len("data:"):])
+	}
+	if len(trimmed) > 0 && !bytes.Equal(trimmed, []byte(wsDoneMarker)) && json.Valid(trimmed) {
+		payloads = append(payloads, bytes.Clone(trimmed))
+	}
+	return payloads
+}
+
+func writeResponsesWebsocketError(conn *websocket.Conn, errMsg *interfaces.ErrorMessage) ([]byte, error) {
+	status := http.StatusInternalServerError
+	errText := http.StatusText(status)
+	if errMsg != nil {
+		if errMsg.StatusCode > 0 {
+			status = errMsg.StatusCode
+			errText = http.StatusText(status)
+		}
+		if errMsg.Error != nil && strings.TrimSpace(errMsg.Error.Error()) != "" {
+			errText = errMsg.Error.Error()
+		}
+	}
+
+	body := handlers.BuildErrorResponseBody(status, errText)
+	payload := map[string]any{
+		"type":   wsEventTypeError,
+		"status": status,
+	}
+
+	if errMsg != nil && errMsg.Addon != nil {
+		headers := map[string]any{}
+		for key, values := range errMsg.Addon {
+			if len(values) == 0 {
+				continue
+			}
+			headers[key] = values[0]
+		}
+		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 _, ok := payload["error"]; !ok {
+		payload["error"] = map[string]any{
+			"type":    "server_error",
+			"message": errText,
+		}
+	}
+
+	data, err := json.Marshal(payload)
+	if err != nil {
+		return nil, err
+	}
+	return data, conn.WriteMessage(websocket.TextMessage, data)
+}
+
+func appendWebsocketEvent(builder *strings.Builder, eventType string, payload []byte) {
+	if builder == nil {
+		return
+	}
+	trimmedPayload := bytes.TrimSpace(payload)
+	if len(trimmedPayload) == 0 {
+		return
+	}
+	if builder.Len() > 0 {
+		builder.WriteString("\n")
+	}
+	builder.WriteString("websocket.")
+	builder.WriteString(eventType)
+	builder.WriteString("\n")
+	builder.Write(trimmedPayload)
+	builder.WriteString("\n")
+}
+
+func websocketPayloadEventType(payload []byte) string {
+	eventType := strings.TrimSpace(gjson.GetBytes(payload, "type").String())
+	if eventType == "" {
+		return "-"
+	}
+	return eventType
+}
+
+func websocketPayloadPreview(payload []byte) string {
+	trimmedPayload := bytes.TrimSpace(payload)
+	if len(trimmedPayload) == 0 {
+		return "<empty>"
+	}
+	preview := trimmedPayload
+	if len(preview) > wsPayloadLogMaxSize {
+		preview = preview[:wsPayloadLogMaxSize]
+	}
+	previewText := strings.ReplaceAll(string(preview), "\n", "\\n")
+	previewText = strings.ReplaceAll(previewText, "\r", "\\r")
+	if len(trimmedPayload) > wsPayloadLogMaxSize {
+		return fmt.Sprintf("%s...(truncated,total=%d)", previewText, len(trimmedPayload))
+	}
+	return previewText
+}
+
+func setWebsocketRequestBody(c *gin.Context, body string) {
+	if c == nil {
+		return
+	}
+	trimmedBody := strings.TrimSpace(body)
+	if trimmedBody == "" {
+		return
+	}
+	c.Set(wsRequestBodyKey, []byte(trimmedBody))
+}
+
+func markAPIResponseTimestamp(c *gin.Context) {
+	if c == nil {
+		return
+	}
+	if _, exists := c.Get("API_RESPONSE_TIMESTAMP"); exists {
+		return
+	}
+	c.Set("API_RESPONSE_TIMESTAMP", time.Now())
+}
--- a/sdk/api/handlers/openai/openai_responses_websocket_test.go
+++ b/sdk/api/handlers/openai/openai_responses_websocket_test.go
@@ -0,0 +1,249 @@
+package openai
+
+import (
+	"bytes"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/tidwall/gjson"
+)
+
+func TestNormalizeResponsesWebsocketRequestCreate(t *testing.T) {
+	raw := []byte(`{"type":"response.create","model":"test-model","stream":false,"input":[{"type":"message","id":"msg-1"}]}`)
+
+	normalized, last, errMsg := normalizeResponsesWebsocketRequest(raw, nil, nil)
+	if errMsg != nil {
+		t.Fatalf("unexpected error: %v", errMsg.Error)
+	}
+	if gjson.GetBytes(normalized, "type").Exists() {
+		t.Fatalf("normalized create request must not include type field")
+	}
+	if !gjson.GetBytes(normalized, "stream").Bool() {
+		t.Fatalf("normalized create request must force stream=true")
+	}
+	if gjson.GetBytes(normalized, "model").String() != "test-model" {
+		t.Fatalf("unexpected model: %s", gjson.GetBytes(normalized, "model").String())
+	}
+	if !bytes.Equal(last, normalized) {
+		t.Fatalf("last request snapshot should match normalized request")
+	}
+}
+
+func TestNormalizeResponsesWebsocketRequestCreateWithHistory(t *testing.T) {
+	lastRequest := []byte(`{"model":"test-model","stream":true,"input":[{"type":"message","id":"msg-1"}]}`)
+	lastResponseOutput := []byte(`[
+		{"type":"function_call","id":"fc-1","call_id":"call-1"},
+		{"type":"message","id":"assistant-1"}
+	]`)
+	raw := []byte(`{"type":"response.create","input":[{"type":"function_call_output","call_id":"call-1","id":"tool-out-1"}]}`)
+
+	normalized, next, errMsg := normalizeResponsesWebsocketRequest(raw, lastRequest, lastResponseOutput)
+	if errMsg != nil {
+		t.Fatalf("unexpected error: %v", errMsg.Error)
+	}
+	if gjson.GetBytes(normalized, "type").Exists() {
+		t.Fatalf("normalized subsequent create request must not include type field")
+	}
+	if gjson.GetBytes(normalized, "model").String() != "test-model" {
+		t.Fatalf("unexpected model: %s", gjson.GetBytes(normalized, "model").String())
+	}
+
+	input := gjson.GetBytes(normalized, "input").Array()
+	if len(input) != 4 {
+		t.Fatalf("merged input len = %d, want 4", len(input))
+	}
+	if input[0].Get("id").String() != "msg-1" ||
+		input[1].Get("id").String() != "fc-1" ||
+		input[2].Get("id").String() != "assistant-1" ||
+		input[3].Get("id").String() != "tool-out-1" {
+		t.Fatalf("unexpected merged input order")
+	}
+	if !bytes.Equal(next, normalized) {
+		t.Fatalf("next request snapshot should match normalized request")
+	}
+}
+
+func TestNormalizeResponsesWebsocketRequestWithPreviousResponseIDIncremental(t *testing.T) {
+	lastRequest := []byte(`{"model":"test-model","stream":true,"instructions":"be helpful","input":[{"type":"message","id":"msg-1"}]}`)
+	lastResponseOutput := []byte(`[
+		{"type":"function_call","id":"fc-1","call_id":"call-1"},
+		{"type":"message","id":"assistant-1"}
+	]`)
+	raw := []byte(`{"type":"response.create","previous_response_id":"resp-1","input":[{"type":"function_call_output","call_id":"call-1","id":"tool-out-1"}]}`)
+
+	normalized, next, errMsg := normalizeResponsesWebsocketRequestWithMode(raw, lastRequest, lastResponseOutput, true)
+	if errMsg != nil {
+		t.Fatalf("unexpected error: %v", errMsg.Error)
+	}
+	if gjson.GetBytes(normalized, "type").Exists() {
+		t.Fatalf("normalized request must not include type field")
+	}
+	if gjson.GetBytes(normalized, "previous_response_id").String() != "resp-1" {
+		t.Fatalf("previous_response_id must be preserved in incremental mode")
+	}
+	input := gjson.GetBytes(normalized, "input").Array()
+	if len(input) != 1 {
+		t.Fatalf("incremental input len = %d, want 1", len(input))
+	}
+	if input[0].Get("id").String() != "tool-out-1" {
+		t.Fatalf("unexpected incremental input item id: %s", input[0].Get("id").String())
+	}
+	if gjson.GetBytes(normalized, "model").String() != "test-model" {
+		t.Fatalf("unexpected model: %s", gjson.GetBytes(normalized, "model").String())
+	}
+	if gjson.GetBytes(normalized, "instructions").String() != "be helpful" {
+		t.Fatalf("unexpected instructions: %s", gjson.GetBytes(normalized, "instructions").String())
+	}
+	if !bytes.Equal(next, normalized) {
+		t.Fatalf("next request snapshot should match normalized request")
+	}
+}
+
+func TestNormalizeResponsesWebsocketRequestWithPreviousResponseIDMergedWhenIncrementalDisabled(t *testing.T) {
+	lastRequest := []byte(`{"model":"test-model","stream":true,"input":[{"type":"message","id":"msg-1"}]}`)
+	lastResponseOutput := []byte(`[
+		{"type":"function_call","id":"fc-1","call_id":"call-1"},
+		{"type":"message","id":"assistant-1"}
+	]`)
+	raw := []byte(`{"type":"response.create","previous_response_id":"resp-1","input":[{"type":"function_call_output","call_id":"call-1","id":"tool-out-1"}]}`)
+
+	normalized, next, errMsg := normalizeResponsesWebsocketRequestWithMode(raw, lastRequest, lastResponseOutput, false)
+	if errMsg != nil {
+		t.Fatalf("unexpected error: %v", errMsg.Error)
+	}
+	if gjson.GetBytes(normalized, "previous_response_id").Exists() {
+		t.Fatalf("previous_response_id must be removed when incremental mode is disabled")
+	}
+	input := gjson.GetBytes(normalized, "input").Array()
+	if len(input) != 4 {
+		t.Fatalf("merged input len = %d, want 4", len(input))
+	}
+	if input[0].Get("id").String() != "msg-1" ||
+		input[1].Get("id").String() != "fc-1" ||
+		input[2].Get("id").String() != "assistant-1" ||
+		input[3].Get("id").String() != "tool-out-1" {
+		t.Fatalf("unexpected merged input order")
+	}
+	if !bytes.Equal(next, normalized) {
+		t.Fatalf("next request snapshot should match normalized request")
+	}
+}
+
+func TestNormalizeResponsesWebsocketRequestAppend(t *testing.T) {
+	lastRequest := []byte(`{"model":"test-model","stream":true,"input":[{"type":"message","id":"msg-1"}]}`)
+	lastResponseOutput := []byte(`[
+		{"type":"message","id":"assistant-1"},
+		{"type":"function_call_output","id":"tool-out-1"}
+	]`)
+	raw := []byte(`{"type":"response.append","input":[{"type":"message","id":"msg-2"},{"type":"message","id":"msg-3"}]}`)
+
+	normalized, next, errMsg := normalizeResponsesWebsocketRequest(raw, lastRequest, lastResponseOutput)
+	if errMsg != nil {
+		t.Fatalf("unexpected error: %v", errMsg.Error)
+	}
+	input := gjson.GetBytes(normalized, "input").Array()
+	if len(input) != 5 {
+		t.Fatalf("merged input len = %d, want 5", len(input))
+	}
+	if input[0].Get("id").String() != "msg-1" ||
+		input[1].Get("id").String() != "assistant-1" ||
+		input[2].Get("id").String() != "tool-out-1" ||
+		input[3].Get("id").String() != "msg-2" ||
+		input[4].Get("id").String() != "msg-3" {
+		t.Fatalf("unexpected merged input order")
+	}
+	if !bytes.Equal(next, normalized) {
+		t.Fatalf("next request snapshot should match normalized append request")
+	}
+}
+
+func TestNormalizeResponsesWebsocketRequestAppendWithoutCreate(t *testing.T) {
+	raw := []byte(`{"type":"response.append","input":[]}`)
+
+	_, _, errMsg := normalizeResponsesWebsocketRequest(raw, nil, nil)
+	if errMsg == nil {
+		t.Fatalf("expected error for append without previous request")
+	}
+	if errMsg.StatusCode != http.StatusBadRequest {
+		t.Fatalf("status = %d, want %d", errMsg.StatusCode, http.StatusBadRequest)
+	}
+}
+
+func TestWebsocketJSONPayloadsFromChunk(t *testing.T) {
+	chunk := []byte("event: response.created\n\ndata: {\"type\":\"response.created\",\"response\":{\"id\":\"resp-1\"}}\n\ndata: [DONE]\n")
+
+	payloads := websocketJSONPayloadsFromChunk(chunk)
+	if len(payloads) != 1 {
+		t.Fatalf("payloads len = %d, want 1", len(payloads))
+	}
+	if gjson.GetBytes(payloads[0], "type").String() != "response.created" {
+		t.Fatalf("unexpected payload type: %s", gjson.GetBytes(payloads[0], "type").String())
+	}
+}
+
+func TestWebsocketJSONPayloadsFromPlainJSONChunk(t *testing.T) {
+	chunk := []byte(`{"type":"response.completed","response":{"id":"resp-1"}}`)
+
+	payloads := websocketJSONPayloadsFromChunk(chunk)
+	if len(payloads) != 1 {
+		t.Fatalf("payloads len = %d, want 1", len(payloads))
+	}
+	if gjson.GetBytes(payloads[0], "type").String() != "response.completed" {
+		t.Fatalf("unexpected payload type: %s", gjson.GetBytes(payloads[0], "type").String())
+	}
+}
+
+func TestResponseCompletedOutputFromPayload(t *testing.T) {
+	payload := []byte(`{"type":"response.completed","response":{"id":"resp-1","output":[{"type":"message","id":"out-1"}]}}`)
+
+	output := responseCompletedOutputFromPayload(payload)
+	items := gjson.ParseBytes(output).Array()
+	if len(items) != 1 {
+		t.Fatalf("output len = %d, want 1", len(items))
+	}
+	if items[0].Get("id").String() != "out-1" {
+		t.Fatalf("unexpected output id: %s", items[0].Get("id").String())
+	}
+}
+
+func TestAppendWebsocketEvent(t *testing.T) {
+	var builder strings.Builder
+
+	appendWebsocketEvent(&builder, "request", []byte("  {\"type\":\"response.create\"}\n"))
+	appendWebsocketEvent(&builder, "response", []byte("{\"type\":\"response.created\"}"))
+
+	got := builder.String()
+	if !strings.Contains(got, "websocket.request\n{\"type\":\"response.create\"}\n") {
+		t.Fatalf("request event not found in body: %s", got)
+	}
+	if !strings.Contains(got, "websocket.response\n{\"type\":\"response.created\"}\n") {
+		t.Fatalf("response event not found in body: %s", got)
+	}
+}
+
+func TestSetWebsocketRequestBody(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+
+	setWebsocketRequestBody(c, " \n ")
+	if _, exists := c.Get(wsRequestBodyKey); exists {
+		t.Fatalf("request body key should not be set for empty body")
+	}
+
+	setWebsocketRequestBody(c, "event body")
+	value, exists := c.Get(wsRequestBodyKey)
+	if !exists {
+		t.Fatalf("request body key not set")
+	}
+	bodyBytes, ok := value.([]byte)
+	if !ok {
+		t.Fatalf("request body key type mismatch")
+	}
+	if string(bodyBytes) != "event body" {
+		t.Fatalf("request body = %q, want %q", string(bodyBytes), "event body")
+	}
+}
--- a/sdk/auth/antigravity.go
+++ b/sdk/auth/antigravity.go
@@ -28,8 +28,7 @@ func (AntigravityAuthenticator) Provider() string { return "antigravity" }

 // RefreshLead instructs the manager to refresh five minutes before expiry.
 func (AntigravityAuthenticator) RefreshLead() *time.Duration {
-	lead := 5 * time.Minute
-	return &lead
+	return new(5 * time.Minute)
 }

 // Login launches a local OAuth flow to obtain antigravity tokens and persists them.
--- a/sdk/auth/claude.go
+++ b/sdk/auth/claude.go
@@ -32,8 +32,7 @@ func (a *ClaudeAuthenticator) Provider() string {
 }

 func (a *ClaudeAuthenticator) RefreshLead() *time.Duration {
-	d := 4 * time.Hour
-	return &d
+	return new(4 * time.Hour)
 }

 func (a *ClaudeAuthenticator) Login(ctx context.Context, cfg *config.Config, opts *LoginOptions) (*coreauth.Auth, error) {
--- a/sdk/auth/codex.go
+++ b/sdk/auth/codex.go
@@ -34,8 +34,7 @@ func (a *CodexAuthenticator) Provider() string {
 }

 func (a *CodexAuthenticator) RefreshLead() *time.Duration {
-	d := 5 * 24 * time.Hour
-	return &d
+	return new(5 * 24 * time.Hour)
 }

 func (a *CodexAuthenticator) Login(ctx context.Context, cfg *config.Config, opts *LoginOptions) (*coreauth.Auth, error) {
--- a/sdk/auth/filestore.go
+++ b/sdk/auth/filestore.go
@@ -4,8 +4,10 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"io"
 	"io/fs"
 	"net/http"
+	"net/url"
 	"os"
 	"path/filepath"
 	"strings"
@@ -186,15 +188,21 @@ func (s *FileTokenStore) readAuthFile(path, baseDir string) (*cliproxyauth.Auth,
 	if provider == "" {
 		provider = "unknown"
 	}
-	if provider == "antigravity" {
+	if provider == "antigravity" || provider == "gemini" {
 		projectID := ""
 		if pid, ok := metadata["project_id"].(string); ok {
 			projectID = strings.TrimSpace(pid)
 		}
 		if projectID == "" {
-			accessToken := ""
-			if token, ok := metadata["access_token"].(string); ok {
-				accessToken = strings.TrimSpace(token)
+			accessToken := extractAccessToken(metadata)
+			// For gemini type, the stored access_token is likely expired (~1h lifetime).
+			// Refresh it using the long-lived refresh_token before querying.
+			if provider == "gemini" {
+				if tokenMap, ok := metadata["token"].(map[string]any); ok {
+					if refreshed, errRefresh := refreshGeminiAccessToken(tokenMap, http.DefaultClient); errRefresh == nil {
+						accessToken = refreshed
+					}
+				}
 			}
 			if accessToken != "" {
 				fetchedProjectID, errFetch := FetchAntigravityProjectID(context.Background(), accessToken, http.DefaultClient)
@@ -313,6 +321,67 @@ func (s *FileTokenStore) baseDirSnapshot() string {
 	return s.baseDir
 }

+func extractAccessToken(metadata map[string]any) string {
+	if at, ok := metadata["access_token"].(string); ok {
+		if v := strings.TrimSpace(at); v != "" {
+			return v
+		}
+	}
+	if tokenMap, ok := metadata["token"].(map[string]any); ok {
+		if at, ok := tokenMap["access_token"].(string); ok {
+			if v := strings.TrimSpace(at); v != "" {
+				return v
+			}
+		}
+	}
+	return ""
+}
+
+func refreshGeminiAccessToken(tokenMap map[string]any, httpClient *http.Client) (string, error) {
+	refreshToken, _ := tokenMap["refresh_token"].(string)
+	clientID, _ := tokenMap["client_id"].(string)
+	clientSecret, _ := tokenMap["client_secret"].(string)
+	tokenURI, _ := tokenMap["token_uri"].(string)
+
+	if refreshToken == "" || clientID == "" || clientSecret == "" {
+		return "", fmt.Errorf("missing refresh credentials")
+	}
+	if tokenURI == "" {
+		tokenURI = "https://oauth2.googleapis.com/token"
+	}
+
+	data := url.Values{
+		"grant_type":    {"refresh_token"},
+		"refresh_token": {refreshToken},
+		"client_id":     {clientID},
+		"client_secret": {clientSecret},
+	}
+
+	resp, err := httpClient.PostForm(tokenURI, data)
+	if err != nil {
+		return "", fmt.Errorf("refresh request: %w", err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+
+	body, _ := io.ReadAll(resp.Body)
+	if resp.StatusCode != http.StatusOK {
+		return "", fmt.Errorf("refresh failed: status %d", resp.StatusCode)
+	}
+
+	var result map[string]any
+	if errUnmarshal := json.Unmarshal(body, &result); errUnmarshal != nil {
+		return "", fmt.Errorf("decode refresh response: %w", errUnmarshal)
+	}
+
+	newAccessToken, _ := result["access_token"].(string)
+	if newAccessToken == "" {
+		return "", fmt.Errorf("no access_token in refresh response")
+	}
+
+	tokenMap["access_token"] = newAccessToken
+	return newAccessToken, nil
+}
+
 // jsonEqual compares two JSON blobs by parsing them into Go objects and deep comparing.
 func jsonEqual(a, b []byte) bool {
 	var objA any
--- a/sdk/auth/filestore_test.go
+++ b/sdk/auth/filestore_test.go
@@ -0,0 +1,80 @@
+package auth
+
+import "testing"
+
+func TestExtractAccessToken(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name     string
+		metadata map[string]any
+		expected string
+	}{
+		{
+			"antigravity top-level access_token",
+			map[string]any{"access_token": "tok-abc"},
+			"tok-abc",
+		},
+		{
+			"gemini nested token.access_token",
+			map[string]any{
+				"token": map[string]any{"access_token": "tok-nested"},
+			},
+			"tok-nested",
+		},
+		{
+			"top-level takes precedence over nested",
+			map[string]any{
+				"access_token": "tok-top",
+				"token":        map[string]any{"access_token": "tok-nested"},
+			},
+			"tok-top",
+		},
+		{
+			"empty metadata",
+			map[string]any{},
+			"",
+		},
+		{
+			"whitespace-only access_token",
+			map[string]any{"access_token": "   "},
+			"",
+		},
+		{
+			"wrong type access_token",
+			map[string]any{"access_token": 12345},
+			"",
+		},
+		{
+			"token is not a map",
+			map[string]any{"token": "not-a-map"},
+			"",
+		},
+		{
+			"nested whitespace-only",
+			map[string]any{
+				"token": map[string]any{"access_token": "  "},
+			},
+			"",
+		},
+		{
+			"fallback to nested when top-level empty",
+			map[string]any{
+				"access_token": "",
+				"token":        map[string]any{"access_token": "tok-fallback"},
+			},
+			"tok-fallback",
+		},
+	}
+
+	for _, tt := range tests {
+		tt := tt
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+			got := extractAccessToken(tt.metadata)
+			if got != tt.expected {
+				t.Errorf("extractAccessToken() = %q, want %q", got, tt.expected)
+			}
+		})
+	}
+}
--- a/sdk/auth/iflow.go
+++ b/sdk/auth/iflow.go
@@ -26,8 +26,7 @@ func (a *IFlowAuthenticator) Provider() string { return "iflow" }

 // RefreshLead indicates how soon before expiry a refresh should be attempted.
 func (a *IFlowAuthenticator) RefreshLead() *time.Duration {
-	d := 24 * time.Hour
-	return &d
+	return new(24 * time.Hour)
 }

 // Login performs the OAuth code flow using a local callback server.
--- a/Show More
+++ b/Show More