Merge pull request #237 from router-for-me/plus

v6.8.18

.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: |

.gitignore

@@ -6,7 +6,7 @@ cliproxy
 # Configuration
 config.yaml
 .env
-
+.mcp.json
 # Generated content
 bin/*
 logs/*

Dockerfile

@@ -1,4 +1,4 @@
-FROM golang:1.24-alpine AS builder
+FROM golang:1.26-alpine AS builder
 
 WORKDIR /app
 

cmd/server/main.go

@@ -77,6 +77,7 @@ func main() {
 	var noBrowser bool
 	var oauthCallbackPort int
 	var antigravityLogin bool
+	var kimiLogin bool
 	var kiroLogin bool
 	var kiroGoogleLogin bool
 	var kiroAWSLogin bool
@@ -102,6 +103,7 @@ func main() {
 	flag.BoolVar(&useIncognito, "incognito", false, "Open browser in incognito/private mode for OAuth (useful for multiple accounts)")
 	flag.BoolVar(&noIncognito, "no-incognito", false, "Force disable incognito mode (uses existing browser session)")
 	flag.BoolVar(&antigravityLogin, "antigravity-login", false, "Login to Antigravity using OAuth")
+	flag.BoolVar(&kimiLogin, "kimi-login", false, "Login to Kimi using OAuth")
 	flag.BoolVar(&kiroLogin, "kiro-login", false, "Login to Kiro using Google OAuth")
 	flag.BoolVar(&kiroGoogleLogin, "kiro-google-login", false, "Login to Kiro using Google OAuth (same as --kiro-login)")
 	flag.BoolVar(&kiroAWSLogin, "kiro-aws-login", false, "Login to Kiro using AWS Builder ID (device code flow)")
@@ -473,7 +475,7 @@ func main() {
 	}
 
 	// Register built-in access providers before constructing services.
-	configaccess.Register()
+	configaccess.Register(&cfg.SDKConfig)
 
 	// Handle different command modes based on the provided flags.
 
@@ -501,6 +503,8 @@ func main() {
 		cmd.DoIFlowLogin(cfg, options)
 	} else if iflowCookie {
 		cmd.DoIFlowCookieAuth(cfg, options)
+	} else if kimiLogin {
+		cmd.DoKimiLogin(cfg, options)
 	} else if kiroLogin {
 		// For Kiro auth, default to incognito mode for multi-account support
 		// Users can explicitly override with --no-incognito

config.example.yaml

@@ -236,10 +236,10 @@ nonstream-keepalive-interval: 0
 
 # Global OAuth model name aliases (per channel)
 # These aliases rename model IDs for both model listing and request routing.
-# Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow, kiro, github-copilot.
+# Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow, kiro, github-copilot, kimi.
 # NOTE: Aliases do not apply to gemini-api-key, codex-api-key, claude-api-key, openai-compatibility, vertex-api-key, or ampcode.
 # You can repeat the same name with different aliases to expose multiple client model names.
-#oauth-model-alias:
+# oauth-model-alias:
 #  antigravity:
 #    - name: "rev19-uic3-1p"
 #      alias: "gemini-2.5-computer-use-preview-10-2025"
@@ -265,9 +265,6 @@ nonstream-keepalive-interval: 0
 #   aistudio:
 #     - name: "gemini-2.5-pro"
 #       alias: "g2.5p"
-#   antigravity:
-#     - name: "gemini-3-pro-preview"
-#       alias: "g3p"
 #   claude:
 #     - name: "claude-sonnet-4-5-20250929"
 #       alias: "cs4.5"
@@ -280,6 +277,9 @@ nonstream-keepalive-interval: 0
 #   iflow:
 #     - name: "glm-4.7"
 #       alias: "glm-god"
+#   kimi:
+#     - name: "kimi-k2.5"
+#       alias: "k2.5"
 #   kiro:
 #     - name: "kiro-claude-opus-4-5"
 #       alias: "op45"
@@ -309,6 +309,8 @@ nonstream-keepalive-interval: 0
 #     - "vision-model"
 #   iflow:
 #     - "tstars2.0"
+#   kimi:
+#     - "kimi-k2-thinking"
 #   kiro:
 #     - "kiro-claude-haiku-4-5"
 #   github-copilot:

docs/sdk-access.md

@@ -7,80 +7,71 @@ The `github.com/router-for-me/CLIProxyAPI/v6/sdk/access` package centralizes inb
 ```go
 import (
     sdkaccess "github.com/router-for-me/CLIProxyAPI/v6/sdk/access"
-    "github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 )
 ```
 
 Add the module with `go get github.com/router-for-me/CLIProxyAPI/v6/sdk/access`.
 
+## Provider Registry
+
+Providers are registered globally and then attached to a `Manager` as a snapshot:
+
+- `RegisterProvider(type, provider)` installs a pre-initialized provider instance.
+- Registration order is preserved the first time each `type` is seen.
+- `RegisteredProviders()` returns the providers in that order.
+
 ## Manager Lifecycle
 
 ```go
 manager := sdkaccess.NewManager()
-providers, err := sdkaccess.BuildProviders(cfg)
-if err != nil {
-    return err
-}
-manager.SetProviders(providers)
+manager.SetProviders(sdkaccess.RegisteredProviders())
 ```
 
 * `NewManager` constructs an empty manager.
 * `SetProviders` replaces the provider slice using a defensive copy.
 * `Providers` retrieves a snapshot that can be iterated safely from other goroutines.
-* `BuildProviders` translates `config.Config` access declarations into runnable providers. When the config omits explicit providers but defines inline API keys, the helper auto-installs the built-in `config-api-key` provider.
+
+If the manager itself is `nil` or no providers are configured, the call returns `nil, nil`, allowing callers to treat access control as disabled.
 
 ## Authenticating Requests
 
 ```go
-result, err := manager.Authenticate(ctx, req)
+result, authErr := manager.Authenticate(ctx, req)
 switch {
-case err == nil:
+case authErr == nil:
     // Authentication succeeded; result describes the provider and principal.
-case errors.Is(err, sdkaccess.ErrNoCredentials):
+case sdkaccess.IsAuthErrorCode(authErr, sdkaccess.AuthErrorCodeNoCredentials):
     // No recognizable credentials were supplied.
-case errors.Is(err, sdkaccess.ErrInvalidCredential):
+case sdkaccess.IsAuthErrorCode(authErr, sdkaccess.AuthErrorCodeInvalidCredential):
     // Supplied credentials were present but rejected.
 default:
-    // Transport-level failure was returned by a provider.
+    // Internal/transport failure was returned by a provider.
 }
 ```
 
-`Manager.Authenticate` walks the configured providers in order. It returns on the first success, skips providers that surface `ErrNotHandled`, and tracks whether any provider reported `ErrNoCredentials` or `ErrInvalidCredential` for downstream error reporting.
-
-If the manager itself is `nil` or no providers are registered, the call returns `nil, nil`, allowing callers to treat access control as disabled without branching on errors.
+`Manager.Authenticate` walks the configured providers in order. It returns on the first success, skips providers that return `AuthErrorCodeNotHandled`, and aggregates `AuthErrorCodeNoCredentials` / `AuthErrorCodeInvalidCredential` for a final result.
 
 Each `Result` includes the provider identifier, the resolved principal, and optional metadata (for example, which header carried the credential).
 
-## Configuration Layout
+## Built-in `config-api-key` Provider
 
-The manager expects access providers under the `auth.providers` key inside `config.yaml`:
+The proxy includes one built-in access provider:
+
+- `config-api-key`: Validates API keys declared under top-level `api-keys`.
+  - Credential sources: `Authorization: Bearer`, `X-Goog-Api-Key`, `X-Api-Key`, `?key=`, `?auth_token=`
+  - Metadata: `Result.Metadata["source"]` is set to the matched source label.
+
+In the CLI server and `sdk/cliproxy`, this provider is registered automatically based on the loaded configuration.
 
 ```yaml
-auth:
-  providers:
-    - name: inline-api
-      type: config-api-key
-      api-keys:
-        - sk-test-123
-        - sk-prod-456
+api-keys:
+  - sk-test-123
+  - sk-prod-456
 ```
 
-Fields map directly to `config.AccessProvider`: `name` labels the provider, `type` selects the registered factory, `sdk` can name an external module, `api-keys` seeds inline credentials, and `config` passes provider-specific options.
+## Loading Providers from External Go Modules
 
-### Loading providers from external SDK modules
-
-To consume a provider shipped in another Go module, point the `sdk` field at the module path and import it for its registration side effect:
-
-```yaml
-auth:
-  providers:
-    - name: partner-auth
-      type: partner-token
-      sdk: github.com/acme/xplatform/sdk/access/providers/partner
-      config:
-        region: us-west-2
-        audience: cli-proxy
-```
+To consume a provider shipped in another Go module, import it for its registration side effect:
 
 ```go
 import (
@@ -89,19 +80,11 @@ import (
 )
 ```
 
-The blank identifier import ensures `init` runs so `sdkaccess.RegisterProvider` executes before `BuildProviders` is called.
-
-## Built-in Providers
-
-The SDK ships with one provider out of the box:
-
-- `config-api-key`: Validates API keys declared inline or under top-level `api-keys`. It accepts the key from `Authorization: Bearer`, `X-Goog-Api-Key`, `X-Api-Key`, or the `?key=` query string and reports `ErrInvalidCredential` when no match is found.
-
-Additional providers can be delivered by third-party packages. When a provider package is imported, it registers itself with `sdkaccess.RegisterProvider`.
+The blank identifier import ensures `init` runs so `sdkaccess.RegisterProvider` executes before you call `RegisteredProviders()` (or before `cliproxy.NewBuilder().Build()`).
 
 ### Metadata and auditing
 
-`Result.Metadata` carries provider-specific context. The built-in `config-api-key` provider, for example, stores the credential source (`authorization`, `x-goog-api-key`, `x-api-key`, or `query-key`). Populate this map in custom providers to enrich logs and downstream auditing.
+`Result.Metadata` carries provider-specific context. The built-in `config-api-key` provider, for example, stores the credential source (`authorization`, `x-goog-api-key`, `x-api-key`, `query-key`, `query-auth-token`). Populate this map in custom providers to enrich logs and downstream auditing.
 
 ## Writing Custom Providers
 
@@ -110,13 +93,13 @@ type customProvider struct{}
 
 func (p *customProvider) Identifier() string { return "my-provider" }
 
-func (p *customProvider) Authenticate(ctx context.Context, r *http.Request) (*sdkaccess.Result, error) {
+func (p *customProvider) Authenticate(ctx context.Context, r *http.Request) (*sdkaccess.Result, *sdkaccess.AuthError) {
     token := r.Header.Get("X-Custom")
     if token == "" {
-        return nil, sdkaccess.ErrNoCredentials
+        return nil, sdkaccess.NewNotHandledError()
     }
     if token != "expected" {
-        return nil, sdkaccess.ErrInvalidCredential
+        return nil, sdkaccess.NewInvalidCredentialError()
     }
     return &sdkaccess.Result{
         Provider:  p.Identifier(),
@@ -126,51 +109,46 @@ func (p *customProvider) Authenticate(ctx context.Context, r *http.Request) (*sd
 }
 
 func init() {
-    sdkaccess.RegisterProvider("custom", func(cfg *config.AccessProvider, root *config.Config) (sdkaccess.Provider, error) {
-        return &customProvider{}, nil
-    })
+    sdkaccess.RegisterProvider("custom", &customProvider{})
 }
 ```
 
-A provider must implement `Identifier()` and `Authenticate()`. To expose it to configuration, call `RegisterProvider` inside `init`. Provider factories receive the specific `AccessProvider` block plus the full root configuration for contextual needs.
+A provider must implement `Identifier()` and `Authenticate()`. To make it available to the access manager, call `RegisterProvider` inside `init` with an initialized provider instance.
 
 ## Error Semantics
 
-- `ErrNoCredentials`: no credentials were present or recognized by any provider.
-- `ErrInvalidCredential`: at least one provider processed the credentials but rejected them.
-- `ErrNotHandled`: instructs the manager to fall through to the next provider without affecting aggregate error reporting.
+- `NewNoCredentialsError()` (`AuthErrorCodeNoCredentials`): no credentials were present or recognized. (HTTP 401)
+- `NewInvalidCredentialError()` (`AuthErrorCodeInvalidCredential`): credentials were present but rejected. (HTTP 401)
+- `NewNotHandledError()` (`AuthErrorCodeNotHandled`): fall through to the next provider.
+- `NewInternalAuthError(message, cause)` (`AuthErrorCodeInternal`): transport/system failure. (HTTP 500)
 
-Return custom errors to surface transport failures; they propagate immediately to the caller instead of being masked.
+Errors propagate immediately to the caller unless they are classified as `not_handled` / `no_credentials` / `invalid_credential` and can be aggregated by the manager.
 
 ## Integration with cliproxy Service
 
-`sdk/cliproxy` wires `@sdk/access` automatically when you build a CLI service via `cliproxy.NewBuilder`. Supplying a preconfigured manager allows you to extend or override the default providers:
+`sdk/cliproxy` wires `@sdk/access` automatically when you build a CLI service via `cliproxy.NewBuilder`. Supplying a manager lets you reuse the same instance in your host process:
 
 ```go
 coreCfg, _ := config.LoadConfig("config.yaml")
-providers, _ := sdkaccess.BuildProviders(coreCfg)
-manager := sdkaccess.NewManager()
-manager.SetProviders(providers)
+accessManager := sdkaccess.NewManager()
 
 svc, _ := cliproxy.NewBuilder().
   WithConfig(coreCfg).
-  WithAccessManager(manager).
+  WithConfigPath("config.yaml").
+  WithRequestAccessManager(accessManager).
   Build()
 ```
 
-The service reuses the manager for every inbound request, ensuring consistent authentication across embedded deployments and the canonical CLI binary.
+Register any custom providers (typically via blank imports) before calling `Build()` so they are present in the global registry snapshot.
 
-### Hot reloading providers
+### Hot reloading
 
-When configuration changes, rebuild providers and swap them into the manager:
+When configuration changes, refresh any config-backed providers and then reset the manager's provider chain:
 
 ```go
-providers, err := sdkaccess.BuildProviders(newCfg)
-if err != nil {
-    log.Errorf("reload auth providers failed: %v", err)
-    return
-}
-accessManager.SetProviders(providers)
+// configaccess is github.com/router-for-me/CLIProxyAPI/v6/internal/access/config_access
+configaccess.Register(&newCfg.SDKConfig)
+accessManager.SetProviders(sdkaccess.RegisteredProviders())
 ```
 
-This mirrors the behaviour in `cliproxy.Service.refreshAccessProviders` and `api.Server.applyAccessConfig`, enabling runtime updates without restarting the process.
+This mirrors the behaviour in `internal/access.ApplyAccessProviders`, enabling runtime updates without restarting the process.

docs/sdk-access_CN.md

@@ -7,80 +7,71 @@
 ```go
 import (
     sdkaccess "github.com/router-for-me/CLIProxyAPI/v6/sdk/access"
-    "github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 )
 ```
 
 通过 `go get github.com/router-for-me/CLIProxyAPI/v6/sdk/access` 添加依赖。
 
+## Provider Registry
+
+访问提供者是全局注册，然后以快照形式挂到 `Manager` 上：
+
+- `RegisterProvider(type, provider)` 注册一个已经初始化好的 provider 实例。
+- 每个 `type` 第一次出现时会记录其注册顺序。
+- `RegisteredProviders()` 会按该顺序返回 provider 列表。
+
 ## 管理器生命周期
 
 ```go
 manager := sdkaccess.NewManager()
-providers, err := sdkaccess.BuildProviders(cfg)
-if err != nil {
-    return err
-}
-manager.SetProviders(providers)
+manager.SetProviders(sdkaccess.RegisteredProviders())
 ```
 
 - `NewManager` 创建空管理器。
 - `SetProviders` 替换提供者切片并做防御性拷贝。
 - `Providers` 返回适合并发读取的快照。
-- `BuildProviders` 将 `config.Config` 中的访问配置转换成可运行的提供者。当配置没有显式声明但包含顶层 `api-keys` 时，会自动挂载内建的 `config-api-key` 提供者。
+
+如果管理器本身为 `nil` 或未配置任何 provider，调用会返回 `nil, nil`，可视为关闭访问控制。
 
 ## 认证请求
 
 ```go
-result, err := manager.Authenticate(ctx, req)
+result, authErr := manager.Authenticate(ctx, req)
 switch {
-case err == nil:
+case authErr == nil:
     // Authentication succeeded; result carries provider and principal.
-case errors.Is(err, sdkaccess.ErrNoCredentials):
+case sdkaccess.IsAuthErrorCode(authErr, sdkaccess.AuthErrorCodeNoCredentials):
     // No recognizable credentials were supplied.
-case errors.Is(err, sdkaccess.ErrInvalidCredential):
+case sdkaccess.IsAuthErrorCode(authErr, sdkaccess.AuthErrorCodeInvalidCredential):
     // Credentials were present but rejected.
 default:
     // Provider surfaced a transport-level failure.
 }
 ```
 
-`Manager.Authenticate` 按配置顺序遍历提供者。遇到成功立即返回，`ErrNotHandled` 会继续尝试下一个；若发现 `ErrNoCredentials` 或 `ErrInvalidCredential`，会在遍历结束后汇总给调用方。
-
-若管理器本身为 `nil` 或尚未注册提供者，调用会返回 `nil, nil`，让调用方无需针对错误做额外分支即可关闭访问控制。
+`Manager.Authenticate` 会按顺序遍历 provider：遇到成功立即返回，`AuthErrorCodeNotHandled` 会继续尝试下一个；`AuthErrorCodeNoCredentials` / `AuthErrorCodeInvalidCredential` 会在遍历结束后汇总给调用方。
 
 `Result` 提供认证提供者标识、解析出的主体以及可选元数据（例如凭证来源）。
 
-## 配置结构
+## 内建 `config-api-key` Provider
 
-在 `config.yaml` 的 `auth.providers` 下定义访问提供者：
+代理内置一个访问提供者：
+
+- `config-api-key`：校验 `config.yaml` 顶层的 `api-keys`。
+  - 凭证来源：`Authorization: Bearer`、`X-Goog-Api-Key`、`X-Api-Key`、`?key=`、`?auth_token=`
+  - 元数据：`Result.Metadata["source"]` 会写入匹配到的来源标识
+
+在 CLI 服务端与 `sdk/cliproxy` 中，该 provider 会根据加载到的配置自动注册。
 
 ```yaml
-auth:
-  providers:
-    - name: inline-api
-      type: config-api-key
-      api-keys:
-        - sk-test-123
-        - sk-prod-456
+api-keys:
+  - sk-test-123
+  - sk-prod-456
 ```
 
-条目映射到 `config.AccessProvider`：`name` 指定实例名，`type` 选择注册的工厂，`sdk` 可引用第三方模块，`api-keys` 提供内联凭证，`config` 用于传递特定选项。
+## 引入外部 Go 模块提供者
 
-### 引入外部 SDK 提供者
-
-若要消费其它 Go 模块输出的访问提供者，可在配置里填写 `sdk` 字段并在代码中引入该包，利用其 `init` 注册过程：
-
-```yaml
-auth:
-  providers:
-    - name: partner-auth
-      type: partner-token
-      sdk: github.com/acme/xplatform/sdk/access/providers/partner
-      config:
-        region: us-west-2
-        audience: cli-proxy
-```
+若要消费其它 Go 模块输出的访问提供者，直接用空白标识符导入以触发其 `init` 注册即可：
 
 ```go
 import (
@@ -89,19 +80,11 @@ import (
 )
 ```
 
-通过空白标识符导入即可确保 `init` 调用，先于 `BuildProviders` 完成 `sdkaccess.RegisterProvider`。
-
-## 内建提供者
-
-当前 SDK 默认内置：
-
-- `config-api-key`：校验配置中的 API Key。它从 `Authorization: Bearer`、`X-Goog-Api-Key`、`X-Api-Key` 以及查询参数 `?key=` 提取凭证，不匹配时抛出 `ErrInvalidCredential`。
-
-导入第三方包即可通过 `sdkaccess.RegisterProvider` 注册更多类型。
+空白导入可确保 `init` 先执行，从而在你调用 `RegisteredProviders()`（或 `cliproxy.NewBuilder().Build()`）之前完成 `sdkaccess.RegisterProvider`。
 
 ### 元数据与审计
 
-`Result.Metadata` 用于携带提供者特定的上下文信息。内建的 `config-api-key` 会记录凭证来源（`authorization`、`x-goog-api-key`、`x-api-key` 或 `query-key`）。自定义提供者同样可以填充该 Map，以便丰富日志与审计场景。
+`Result.Metadata` 用于携带提供者特定的上下文信息。内建的 `config-api-key` 会记录凭证来源（`authorization`、`x-goog-api-key`、`x-api-key`、`query-key`、`query-auth-token`）。自定义提供者同样可以填充该 Map，以便丰富日志与审计场景。
 
 ## 编写自定义提供者
 
@@ -110,13 +93,13 @@ type customProvider struct{}
 
 func (p *customProvider) Identifier() string { return "my-provider" }
 
-func (p *customProvider) Authenticate(ctx context.Context, r *http.Request) (*sdkaccess.Result, error) {
+func (p *customProvider) Authenticate(ctx context.Context, r *http.Request) (*sdkaccess.Result, *sdkaccess.AuthError) {
     token := r.Header.Get("X-Custom")
     if token == "" {
-        return nil, sdkaccess.ErrNoCredentials
+        return nil, sdkaccess.NewNotHandledError()
     }
     if token != "expected" {
-        return nil, sdkaccess.ErrInvalidCredential
+        return nil, sdkaccess.NewInvalidCredentialError()
     }
     return &sdkaccess.Result{
         Provider:  p.Identifier(),
@@ -126,51 +109,46 @@ func (p *customProvider) Authenticate(ctx context.Context, r *http.Request) (*sd
 }
 
 func init() {
-    sdkaccess.RegisterProvider("custom", func(cfg *config.AccessProvider, root *config.Config) (sdkaccess.Provider, error) {
-        return &customProvider{}, nil
-    })
+    sdkaccess.RegisterProvider("custom", &customProvider{})
 }
 ```
 
-自定义提供者需要实现 `Identifier()` 与 `Authenticate()`。在 `init` 中调用 `RegisterProvider` 暴露给配置层，工厂函数既能读取当前条目，也能访问完整根配置。
+自定义提供者需要实现 `Identifier()` 与 `Authenticate()`。在 `init` 中用已初始化实例调用 `RegisterProvider` 注册到全局 registry。
 
 ## 错误语义
 
-- `ErrNoCredentials`：任何提供者都未识别到凭证。
-- `ErrInvalidCredential`：至少一个提供者处理了凭证但判定无效。
-- `ErrNotHandled`：告诉管理器跳到下一个提供者，不影响最终错误统计。
+- `NewNoCredentialsError()`（`AuthErrorCodeNoCredentials`）：未提供或未识别到凭证。（HTTP 401）
+- `NewInvalidCredentialError()`（`AuthErrorCodeInvalidCredential`）：凭证存在但校验失败。（HTTP 401）
+- `NewNotHandledError()`（`AuthErrorCodeNotHandled`）：告诉管理器跳到下一个 provider。
+- `NewInternalAuthError(message, cause)`（`AuthErrorCodeInternal`）：网络/系统错误。（HTTP 500）
 
-自定义错误（例如网络异常）会马上冒泡返回。
+除可汇总的 `not_handled` / `no_credentials` / `invalid_credential` 外，其它错误会立即冒泡返回。
 
 ## 与 cliproxy 集成
 
-使用 `sdk/cliproxy` 构建服务时会自动接入 `@sdk/access`。如果需要扩展内置行为，可传入自定义管理器：
+使用 `sdk/cliproxy` 构建服务时会自动接入 `@sdk/access`。如果希望在宿主进程里复用同一个 `Manager` 实例，可传入自定义管理器：
 
 ```go
 coreCfg, _ := config.LoadConfig("config.yaml")
-providers, _ := sdkaccess.BuildProviders(coreCfg)
-manager := sdkaccess.NewManager()
-manager.SetProviders(providers)
+accessManager := sdkaccess.NewManager()
 
 svc, _ := cliproxy.NewBuilder().
   WithConfig(coreCfg).
-  WithAccessManager(manager).
+  WithConfigPath("config.yaml").
+  WithRequestAccessManager(accessManager).
   Build()
 ```
 
-服务会复用该管理器处理每一个入站请求，实现与 CLI 二进制一致的访问控制体验。
+请在调用 `Build()` 之前完成自定义 provider 的注册（通常通过空白导入触发 `init`），以确保它们被包含在全局 registry 的快照中。
 
 ### 动态热更新提供者
 
-当配置发生变化时，可以重新构建提供者并替换当前列表：
+当配置发生变化时，刷新依赖配置的 provider，然后重置 manager 的 provider 链：
 
 ```go
-providers, err := sdkaccess.BuildProviders(newCfg)
-if err != nil {
-    log.Errorf("reload auth providers failed: %v", err)
-    return
-}
-accessManager.SetProviders(providers)
+// configaccess is github.com/router-for-me/CLIProxyAPI/v6/internal/access/config_access
+configaccess.Register(&newCfg.SDKConfig)
+accessManager.SetProviders(sdkaccess.RegisteredProviders())
 ```
 
-这一流程与 `cliproxy.Service.refreshAccessProviders` 和 `api.Server.applyAccessConfig` 保持一致，避免为更新访问策略而重启进程。
+这一流程与 `internal/access.ApplyAccessProviders` 保持一致，避免为更新访问策略而重启进程。

go.mod

@@ -1,6 +1,6 @@
 module github.com/router-for-me/CLIProxyAPI/v6
 
-go 1.24.0
+go 1.26.0
 
 require (
 	github.com/andybalholm/brotli v1.0.6

internal/access/config_access/provider.go

@@ -4,19 +4,28 @@ import (
 	"context"
 	"net/http"
 	"strings"
-	"sync"
 
 	sdkaccess "github.com/router-for-me/CLIProxyAPI/v6/sdk/access"
 	sdkconfig "github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
 )
 
-var registerOnce sync.Once
-
 // Register ensures the config-access provider is available to the access manager.
-func Register() {
-	registerOnce.Do(func() {
-		sdkaccess.RegisterProvider(sdkconfig.AccessProviderTypeConfigAPIKey, newProvider)
-	})
+func Register(cfg *sdkconfig.SDKConfig) {
+	if cfg == nil {
+		sdkaccess.UnregisterProvider(sdkaccess.AccessProviderTypeConfigAPIKey)
+		return
+	}
+
+	keys := normalizeKeys(cfg.APIKeys)
+	if len(keys) == 0 {
+		sdkaccess.UnregisterProvider(sdkaccess.AccessProviderTypeConfigAPIKey)
+		return
+	}
+
+	sdkaccess.RegisterProvider(
+		sdkaccess.AccessProviderTypeConfigAPIKey,
+		newProvider(sdkaccess.DefaultAccessProviderName, keys),
+	)
 }
 
 type provider struct {
@@ -24,34 +33,31 @@ type provider struct {
 	keys map[string]struct{}
 }
 
-func newProvider(cfg *sdkconfig.AccessProvider, _ *sdkconfig.SDKConfig) (sdkaccess.Provider, error) {
-	name := cfg.Name
-	if name == "" {
-		name = sdkconfig.DefaultAccessProviderName
+func newProvider(name string, keys []string) *provider {
+	providerName := strings.TrimSpace(name)
+	if providerName == "" {
+		providerName = sdkaccess.DefaultAccessProviderName
 	}
-	keys := make(map[string]struct{}, len(cfg.APIKeys))
-	for _, key := range cfg.APIKeys {
-		if key == "" {
-			continue
-		}
-		keys[key] = struct{}{}
+	keySet := make(map[string]struct{}, len(keys))
+	for _, key := range keys {
+		keySet[key] = struct{}{}
 	}
-	return &provider{name: name, keys: keys}, nil
+	return &provider{name: providerName, keys: keySet}
 }
 
 func (p *provider) Identifier() string {
 	if p == nil || p.name == "" {
-		return sdkconfig.DefaultAccessProviderName
+		return sdkaccess.DefaultAccessProviderName
 	}
 	return p.name
 }
 
-func (p *provider) Authenticate(_ context.Context, r *http.Request) (*sdkaccess.Result, error) {
+func (p *provider) Authenticate(_ context.Context, r *http.Request) (*sdkaccess.Result, *sdkaccess.AuthError) {
 	if p == nil {
-		return nil, sdkaccess.ErrNotHandled
+		return nil, sdkaccess.NewNotHandledError()
 	}
 	if len(p.keys) == 0 {
-		return nil, sdkaccess.ErrNotHandled
+		return nil, sdkaccess.NewNotHandledError()
 	}
 	authHeader := r.Header.Get("Authorization")
 	authHeaderGoogle := r.Header.Get("X-Goog-Api-Key")
@@ -63,7 +69,7 @@ func (p *provider) Authenticate(_ context.Context, r *http.Request) (*sdkaccess.
 		queryAuthToken = r.URL.Query().Get("auth_token")
 	}
 	if authHeader == "" && authHeaderGoogle == "" && authHeaderAnthropic == "" && queryKey == "" && queryAuthToken == "" {
-		return nil, sdkaccess.ErrNoCredentials
+		return nil, sdkaccess.NewNoCredentialsError()
 	}
 
 	apiKey := extractBearerToken(authHeader)
@@ -94,7 +100,7 @@ func (p *provider) Authenticate(_ context.Context, r *http.Request) (*sdkaccess.
 		}
 	}
 
-	return nil, sdkaccess.ErrInvalidCredential
+	return nil, sdkaccess.NewInvalidCredentialError()
 }
 
 func extractBearerToken(header string) string {
@@ -110,3 +116,26 @@ func extractBearerToken(header string) string {
 	}
 	return strings.TrimSpace(parts[1])
 }
+
+func normalizeKeys(keys []string) []string {
+	if len(keys) == 0 {
+		return nil
+	}
+	normalized := make([]string, 0, len(keys))
+	seen := make(map[string]struct{}, len(keys))
+	for _, key := range keys {
+		trimmedKey := strings.TrimSpace(key)
+		if trimmedKey == "" {
+			continue
+		}
+		if _, exists := seen[trimmedKey]; exists {
+			continue
+		}
+		seen[trimmedKey] = struct{}{}
+		normalized = append(normalized, trimmedKey)
+	}
+	if len(normalized) == 0 {
+		return nil
+	}
+	return normalized
+}

internal/access/reconcile.go

@@ -6,9 +6,9 @@ import (
 	"sort"
 	"strings"
 
+	configaccess "github.com/router-for-me/CLIProxyAPI/v6/internal/access/config_access"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	sdkaccess "github.com/router-for-me/CLIProxyAPI/v6/sdk/access"
-	sdkConfig "github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
 	log "github.com/sirupsen/logrus"
 )
 
@@ -17,26 +17,26 @@ import (
 // ordered provider slice along with the identifiers of providers that were added, updated, or
 // removed compared to the previous configuration.
 func ReconcileProviders(oldCfg, newCfg *config.Config, existing []sdkaccess.Provider) (result []sdkaccess.Provider, added, updated, removed []string, err error) {
+	_ = oldCfg
 	if newCfg == nil {
 		return nil, nil, nil, nil, nil
 	}
 
+	result = sdkaccess.RegisteredProviders()
+
 	existingMap := make(map[string]sdkaccess.Provider, len(existing))
 	for _, provider := range existing {
-		if provider == nil {
+		providerID := identifierFromProvider(provider)
+		if providerID == "" {
 			continue
 		}
-		existingMap[provider.Identifier()] = provider
+		existingMap[providerID] = provider
 	}
 
-	oldCfgMap := accessProviderMap(oldCfg)
-	newEntries := collectProviderEntries(newCfg)
-
-	result = make([]sdkaccess.Provider, 0, len(newEntries))
-	finalIDs := make(map[string]struct{}, len(newEntries))
+	finalIDs := make(map[string]struct{}, len(result))
 
 	isInlineProvider := func(id string) bool {
-		return strings.EqualFold(id, sdkConfig.DefaultAccessProviderName)
+		return strings.EqualFold(id, sdkaccess.DefaultAccessProviderName)
 	}
 	appendChange := func(list *[]string, id string) {
 		if isInlineProvider(id) {
@@ -45,85 +45,28 @@ func ReconcileProviders(oldCfg, newCfg *config.Config, existing []sdkaccess.Prov
 		*list = append(*list, id)
 	}
 
-	for _, providerCfg := range newEntries {
-		key := providerIdentifier(providerCfg)
-		if key == "" {
+	for _, provider := range result {
+		providerID := identifierFromProvider(provider)
+		if providerID == "" {
 			continue
 		}
+		finalIDs[providerID] = struct{}{}
 
-		forceRebuild := strings.EqualFold(strings.TrimSpace(providerCfg.Type), sdkConfig.AccessProviderTypeConfigAPIKey)
-		if oldCfgProvider, ok := oldCfgMap[key]; ok {
-			isAliased := oldCfgProvider == providerCfg
-			if !forceRebuild && !isAliased && providerConfigEqual(oldCfgProvider, providerCfg) {
-				if existingProvider, okExisting := existingMap[key]; okExisting {
-					result = append(result, existingProvider)
-					finalIDs[key] = struct{}{}
-					continue
-				}
-			}
+		existingProvider, exists := existingMap[providerID]
+		if !exists {
+			appendChange(&added, providerID)
+			continue
 		}
-
-		provider, buildErr := sdkaccess.BuildProvider(providerCfg, &newCfg.SDKConfig)
-		if buildErr != nil {
-			return nil, nil, nil, nil, buildErr
-		}
-		if _, ok := oldCfgMap[key]; ok {
-			if _, existed := existingMap[key]; existed {
-				appendChange(&updated, key)
-			} else {
-				appendChange(&added, key)
-			}
-		} else {
-			appendChange(&added, key)
-		}
-		result = append(result, provider)
-		finalIDs[key] = struct{}{}
-	}
-
-	if len(result) == 0 {
-		if inline := sdkConfig.MakeInlineAPIKeyProvider(newCfg.APIKeys); inline != nil {
-			key := providerIdentifier(inline)
-			if key != "" {
-				if oldCfgProvider, ok := oldCfgMap[key]; ok {
-					if providerConfigEqual(oldCfgProvider, inline) {
-						if existingProvider, okExisting := existingMap[key]; okExisting {
-							result = append(result, existingProvider)
-							finalIDs[key] = struct{}{}
-							goto inlineDone
-						}
-					}
-				}
-				provider, buildErr := sdkaccess.BuildProvider(inline, &newCfg.SDKConfig)
-				if buildErr != nil {
-					return nil, nil, nil, nil, buildErr
-				}
-				if _, existed := existingMap[key]; existed {
-					appendChange(&updated, key)
-				} else if _, hadOld := oldCfgMap[key]; hadOld {
-					appendChange(&updated, key)
-				} else {
-					appendChange(&added, key)
-				}
-				result = append(result, provider)
-				finalIDs[key] = struct{}{}
-			}
-		}
-	inlineDone:
-	}
-
-	removedSet := make(map[string]struct{})
-	for id := range existingMap {
-		if _, ok := finalIDs[id]; !ok {
-			if isInlineProvider(id) {
-				continue
-			}
-			removedSet[id] = struct{}{}
+		if !providerInstanceEqual(existingProvider, provider) {
+			appendChange(&updated, providerID)
 		}
 	}
 
-	removed = make([]string, 0, len(removedSet))
-	for id := range removedSet {
-		removed = append(removed, id)
+	for providerID := range existingMap {
+		if _, exists := finalIDs[providerID]; exists {
+			continue
+		}
+		appendChange(&removed, providerID)
 	}
 
 	sort.Strings(added)
@@ -142,6 +85,7 @@ func ApplyAccessProviders(manager *sdkaccess.Manager, oldCfg, newCfg *config.Con
 	}
 
 	existing := manager.Providers()
+	configaccess.Register(&newCfg.SDKConfig)
 	providers, added, updated, removed, err := ReconcileProviders(oldCfg, newCfg, existing)
 	if err != nil {
 		log.Errorf("failed to reconcile request auth providers: %v", err)
@@ -160,111 +104,24 @@ func ApplyAccessProviders(manager *sdkaccess.Manager, oldCfg, newCfg *config.Con
 	return false, nil
 }
 
-func accessProviderMap(cfg *config.Config) map[string]*sdkConfig.AccessProvider {
-	result := make(map[string]*sdkConfig.AccessProvider)
-	if cfg == nil {
-		return result
-	}
-	for i := range cfg.Access.Providers {
-		providerCfg := &cfg.Access.Providers[i]
-		if providerCfg.Type == "" {
-			continue
-		}
-		key := providerIdentifier(providerCfg)
-		if key == "" {
-			continue
-		}
-		result[key] = providerCfg
-	}
-	if len(result) == 0 && len(cfg.APIKeys) > 0 {
-		if provider := sdkConfig.MakeInlineAPIKeyProvider(cfg.APIKeys); provider != nil {
-			if key := providerIdentifier(provider); key != "" {
-				result[key] = provider
-			}
-		}
-	}
-	return result
-}
-
-func collectProviderEntries(cfg *config.Config) []*sdkConfig.AccessProvider {
-	entries := make([]*sdkConfig.AccessProvider, 0, len(cfg.Access.Providers))
-	for i := range cfg.Access.Providers {
-		providerCfg := &cfg.Access.Providers[i]
-		if providerCfg.Type == "" {
-			continue
-		}
-		if key := providerIdentifier(providerCfg); key != "" {
-			entries = append(entries, providerCfg)
-		}
-	}
-	if len(entries) == 0 && len(cfg.APIKeys) > 0 {
-		if inline := sdkConfig.MakeInlineAPIKeyProvider(cfg.APIKeys); inline != nil {
-			entries = append(entries, inline)
-		}
-	}
-	return entries
-}
-
-func providerIdentifier(provider *sdkConfig.AccessProvider) string {
+func identifierFromProvider(provider sdkaccess.Provider) string {
 	if provider == nil {
 		return ""
 	}
-	if name := strings.TrimSpace(provider.Name); name != "" {
-		return name
-	}
-	typ := strings.TrimSpace(provider.Type)
-	if typ == "" {
-		return ""
-	}
-	if strings.EqualFold(typ, sdkConfig.AccessProviderTypeConfigAPIKey) {
-		return sdkConfig.DefaultAccessProviderName
-	}
-	return typ
+	return strings.TrimSpace(provider.Identifier())
 }
 
-func providerConfigEqual(a, b *sdkConfig.AccessProvider) bool {
+func providerInstanceEqual(a, b sdkaccess.Provider) bool {
 	if a == nil || b == nil {
 		return a == nil && b == nil
 	}
-	if !strings.EqualFold(strings.TrimSpace(a.Type), strings.TrimSpace(b.Type)) {
+	if reflect.TypeOf(a) != reflect.TypeOf(b) {
 		return false
 	}
-	if strings.TrimSpace(a.SDK) != strings.TrimSpace(b.SDK) {
-		return false
+	valueA := reflect.ValueOf(a)
+	valueB := reflect.ValueOf(b)
+	if valueA.Kind() == reflect.Pointer && valueB.Kind() == reflect.Pointer {
+		return valueA.Pointer() == valueB.Pointer()
 	}
-	if !stringSetEqual(a.APIKeys, b.APIKeys) {
-		return false
-	}
-	if len(a.Config) != len(b.Config) {
-		return false
-	}
-	if len(a.Config) > 0 && !reflect.DeepEqual(a.Config, b.Config) {
-		return false
-	}
-	return true
-}
-
-func stringSetEqual(a, b []string) bool {
-	if len(a) != len(b) {
-		return false
-	}
-	if len(a) == 0 {
-		return true
-	}
-	seen := make(map[string]int, len(a))
-	for _, val := range a {
-		seen[val]++
-	}
-	for _, val := range b {
-		count := seen[val]
-		if count == 0 {
-			return false
-		}
-		if count == 1 {
-			delete(seen, val)
-		} else {
-			seen[val] = count - 1
-		}
-	}
-	return len(seen) == 0
+	return reflect.DeepEqual(a, 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"`

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/kimi"
 	kiroauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/kiro"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/qwen"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
@@ -1192,6 +1193,30 @@ 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)
@@ -1613,6 +1638,82 @@ func (h *Handler) RequestQwenToken(c *gin.Context) {
 	c.JSON(200, gin.H{"status": "ok", "url": authURL, "state": state})
 }
 
+func (h *Handler) RequestKimiToken(c *gin.Context) {
+	ctx := context.Background()
+
+	fmt.Println("Initializing Kimi authentication...")
+
+	state := fmt.Sprintf("kmi-%d", time.Now().UnixNano())
+	// Initialize Kimi auth service
+	kimiAuth := kimi.NewKimiAuth(h.cfg)
+
+	// Generate authorization URL
+	deviceFlow, errStartDeviceFlow := kimiAuth.StartDeviceFlow(ctx)
+	if errStartDeviceFlow != nil {
+		log.Errorf("Failed to generate authorization URL: %v", errStartDeviceFlow)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to generate authorization url"})
+		return
+	}
+	authURL := deviceFlow.VerificationURIComplete
+	if authURL == "" {
+		authURL = deviceFlow.VerificationURI
+	}
+
+	RegisterOAuthSession(state, "kimi")
+
+	go func() {
+		fmt.Println("Waiting for authentication...")
+		authBundle, errWaitForAuthorization := kimiAuth.WaitForAuthorization(ctx, deviceFlow)
+		if errWaitForAuthorization != nil {
+			SetOAuthSessionError(state, "Authentication failed")
+			fmt.Printf("Authentication failed: %v\n", errWaitForAuthorization)
+			return
+		}
+
+		// Create token storage
+		tokenStorage := kimiAuth.CreateTokenStorage(authBundle)
+
+		metadata := map[string]any{
+			"type":          "kimi",
+			"access_token":  authBundle.TokenData.AccessToken,
+			"refresh_token": authBundle.TokenData.RefreshToken,
+			"token_type":    authBundle.TokenData.TokenType,
+			"scope":         authBundle.TokenData.Scope,
+			"timestamp":     time.Now().UnixMilli(),
+		}
+		if authBundle.TokenData.ExpiresAt > 0 {
+			expired := time.Unix(authBundle.TokenData.ExpiresAt, 0).UTC().Format(time.RFC3339)
+			metadata["expired"] = expired
+		}
+		if strings.TrimSpace(authBundle.DeviceID) != "" {
+			metadata["device_id"] = strings.TrimSpace(authBundle.DeviceID)
+		}
+
+		fileName := fmt.Sprintf("kimi-%d.json", time.Now().UnixMilli())
+		record := &coreauth.Auth{
+			ID:       fileName,
+			Provider: "kimi",
+			FileName: fileName,
+			Label:    "Kimi User",
+			Storage:  tokenStorage,
+			Metadata: metadata,
+		}
+		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)
+		fmt.Println("You can now use Kimi services through this CLI")
+		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("kimi")
+	}()
+
+	c.JSON(200, gin.H{"status": "ok", "url": authURL, "state": state})
+}
+
 func (h *Handler) RequestIFlowToken(c *gin.Context) {
 	ctx := context.Background()
 
@@ -2047,7 +2148,48 @@ 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{

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 {

internal/api/handlers/management/config_lists.go

@@ -109,14 +109,13 @@ func (h *Handler) GetAPIKeys(c *gin.Context) { c.JSON(200, gin.H{"api-keys": h.c
 func (h *Handler) PutAPIKeys(c *gin.Context) {
 	h.putStringList(c, func(v []string) {
 		h.cfg.APIKeys = append([]string(nil), v...)
-		h.cfg.Access.Providers = nil
 	}, nil)
 }
 func (h *Handler) PatchAPIKeys(c *gin.Context) {
-	h.patchStringList(c, &h.cfg.APIKeys, func() { h.cfg.Access.Providers = nil })
+	h.patchStringList(c, &h.cfg.APIKeys, func() {})
 }
 func (h *Handler) DeleteAPIKeys(c *gin.Context) {
-	h.deleteFromStringList(c, &h.cfg.APIKeys, func() { h.cfg.Access.Providers = nil })
+	h.deleteFromStringList(c, &h.cfg.APIKeys, func() {})
 }
 
 // gemini-api-key: []GeminiKey
@@ -797,10 +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)
 }
 

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)

internal/api/modules/amp/response_rewriter.go

@@ -122,7 +122,7 @@ func (rw *ResponseRewriter) Flush() {
 }
 
 // modelFieldPaths lists all JSON paths where model name may appear
-var modelFieldPaths = []string{"model", "modelVersion", "response.modelVersion", "message.model"}
+var modelFieldPaths = []string{"message.model", "model", "modelVersion", "response.model", "response.modelVersion"}
 
 // rewriteModelInResponse replaces all occurrences of the mapped model with the original model in JSON
 // It also suppresses "thinking" blocks if "tool_use" is present to ensure Amp client compatibility

internal/api/modules/amp/response_rewriter_test.go

@@ -0,0 +1,110 @@
+package amp
+
+import (
+	"testing"
+)
+
+func TestRewriteModelInResponse_TopLevel(t *testing.T) {
+	rw := &ResponseRewriter{originalModel: "gpt-5.2-codex"}
+
+	input := []byte(`{"id":"resp_1","model":"gpt-5.3-codex","output":[]}`)
+	result := rw.rewriteModelInResponse(input)
+
+	expected := `{"id":"resp_1","model":"gpt-5.2-codex","output":[]}`
+	if string(result) != expected {
+		t.Errorf("expected %s, got %s", expected, string(result))
+	}
+}
+
+func TestRewriteModelInResponse_ResponseModel(t *testing.T) {
+	rw := &ResponseRewriter{originalModel: "gpt-5.2-codex"}
+
+	input := []byte(`{"type":"response.completed","response":{"id":"resp_1","model":"gpt-5.3-codex","status":"completed"}}`)
+	result := rw.rewriteModelInResponse(input)
+
+	expected := `{"type":"response.completed","response":{"id":"resp_1","model":"gpt-5.2-codex","status":"completed"}}`
+	if string(result) != expected {
+		t.Errorf("expected %s, got %s", expected, string(result))
+	}
+}
+
+func TestRewriteModelInResponse_ResponseCreated(t *testing.T) {
+	rw := &ResponseRewriter{originalModel: "gpt-5.2-codex"}
+
+	input := []byte(`{"type":"response.created","response":{"id":"resp_1","model":"gpt-5.3-codex","status":"in_progress"}}`)
+	result := rw.rewriteModelInResponse(input)
+
+	expected := `{"type":"response.created","response":{"id":"resp_1","model":"gpt-5.2-codex","status":"in_progress"}}`
+	if string(result) != expected {
+		t.Errorf("expected %s, got %s", expected, string(result))
+	}
+}
+
+func TestRewriteModelInResponse_NoModelField(t *testing.T) {
+	rw := &ResponseRewriter{originalModel: "gpt-5.2-codex"}
+
+	input := []byte(`{"type":"response.output_item.added","item":{"id":"item_1","type":"message"}}`)
+	result := rw.rewriteModelInResponse(input)
+
+	if string(result) != string(input) {
+		t.Errorf("expected no modification, got %s", string(result))
+	}
+}
+
+func TestRewriteModelInResponse_EmptyOriginalModel(t *testing.T) {
+	rw := &ResponseRewriter{originalModel: ""}
+
+	input := []byte(`{"model":"gpt-5.3-codex"}`)
+	result := rw.rewriteModelInResponse(input)
+
+	if string(result) != string(input) {
+		t.Errorf("expected no modification when originalModel is empty, got %s", string(result))
+	}
+}
+
+func TestRewriteStreamChunk_SSEWithResponseModel(t *testing.T) {
+	rw := &ResponseRewriter{originalModel: "gpt-5.2-codex"}
+
+	chunk := []byte("data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_1\",\"model\":\"gpt-5.3-codex\",\"status\":\"completed\"}}\n\n")
+	result := rw.rewriteStreamChunk(chunk)
+
+	expected := "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_1\",\"model\":\"gpt-5.2-codex\",\"status\":\"completed\"}}\n\n"
+	if string(result) != expected {
+		t.Errorf("expected %s, got %s", expected, string(result))
+	}
+}
+
+func TestRewriteStreamChunk_MultipleEvents(t *testing.T) {
+	rw := &ResponseRewriter{originalModel: "gpt-5.2-codex"}
+
+	chunk := []byte("data: {\"type\":\"response.created\",\"response\":{\"model\":\"gpt-5.3-codex\"}}\n\ndata: {\"type\":\"response.output_item.added\",\"item\":{\"id\":\"item_1\"}}\n\n")
+	result := rw.rewriteStreamChunk(chunk)
+
+	if string(result) == string(chunk) {
+		t.Error("expected response.model to be rewritten in SSE stream")
+	}
+	if !contains(result, []byte(`"model":"gpt-5.2-codex"`)) {
+		t.Errorf("expected rewritten model in output, got %s", string(result))
+	}
+}
+
+func TestRewriteStreamChunk_MessageModel(t *testing.T) {
+	rw := &ResponseRewriter{originalModel: "claude-opus-4.5"}
+
+	chunk := []byte("data: {\"message\":{\"model\":\"claude-sonnet-4\",\"role\":\"assistant\"}}\n\n")
+	result := rw.rewriteStreamChunk(chunk)
+
+	expected := "data: {\"message\":{\"model\":\"claude-opus-4.5\",\"role\":\"assistant\"}}\n\n"
+	if string(result) != expected {
+		t.Errorf("expected %s, got %s", expected, string(result))
+	}
+}
+
+func contains(data, substr []byte) bool {
+	for i := 0; i <= len(data)-len(substr); i++ {
+		if string(data[i:i+len(substr)]) == string(substr) {
+			return true
+		}
+	}
+	return false
+}

internal/api/server.go

@@ -649,6 +649,7 @@ 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("/kimi-auth-url", s.mgmt.RequestKimiToken)
 		mgmt.GET("/iflow-auth-url", s.mgmt.RequestIFlowToken)
 		mgmt.POST("/iflow-auth-url", s.mgmt.RequestIFlowCookieToken)
 		mgmt.GET("/kiro-auth-url", s.mgmt.RequestKiroToken)
@@ -682,14 +683,17 @@ func (s *Server) serveManagementControlPanel(c *gin.Context) {
 
 	if _, err := os.Stat(filePath); err != nil {
 		if os.IsNotExist(err) {
-			go managementasset.EnsureLatestManagementHTML(context.Background(), managementasset.StaticDir(s.configFilePath), cfg.ProxyURL, cfg.RemoteManagement.PanelGitHubRepository)
-			c.AbortWithStatus(http.StatusNotFound)
+			// Synchronously ensure management.html is available with a detached context.
+			// Control panel bootstrap should not be canceled by client disconnects.
+			if !managementasset.EnsureLatestManagementHTML(context.Background(), managementasset.StaticDir(s.configFilePath), cfg.ProxyURL, cfg.RemoteManagement.PanelGitHubRepository) {
+				c.AbortWithStatus(http.StatusNotFound)
+				return
+			}
+		} else {
+			log.WithError(err).Error("failed to stat management control panel asset")
+			c.AbortWithStatus(http.StatusInternalServerError)
 			return
 		}
-
-		log.WithError(err).Error("failed to stat management control panel asset")
-		c.AbortWithStatus(http.StatusInternalServerError)
-		return
 	}
 
 	c.File(filePath)
@@ -979,10 +983,6 @@ func (s *Server) UpdateClients(cfg *config.Config) {
 
 	s.handlers.UpdateClients(&cfg.SDKConfig)
 
-	if !cfg.RemoteManagement.DisableControlPanel {
-		staticDir := managementasset.StaticDir(s.configFilePath)
-		go managementasset.EnsureLatestManagementHTML(context.Background(), staticDir, cfg.ProxyURL, cfg.RemoteManagement.PanelGitHubRepository)
-	}
 	if s.mgmt != nil {
 		s.mgmt.SetConfig(cfg)
 		s.mgmt.SetAuthManager(s.handlers.AuthManager)
@@ -1061,14 +1061,10 @@ func AuthMiddleware(manager *sdkaccess.Manager) gin.HandlerFunc {
 			return
 		}
 
-		switch {
-		case errors.Is(err, sdkaccess.ErrNoCredentials):
-			c.AbortWithStatusJSON(http.StatusUnauthorized, gin.H{"error": "Missing API key"})
-		case errors.Is(err, sdkaccess.ErrInvalidCredential):
-			c.AbortWithStatusJSON(http.StatusUnauthorized, gin.H{"error": "Invalid API key"})
-		default:
+		statusCode := err.HTTPStatusCode()
+		if statusCode >= http.StatusInternalServerError {
 			log.Errorf("authentication middleware error: %v", err)
-			c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": "Authentication service error"})
 		}
+		c.AbortWithStatusJSON(statusCode, gin.H{"error": err.Message})
 	}
 }

internal/auth/kimi/kimi.go

@@ -0,0 +1,396 @@
+// Package kimi provides authentication and token management for Kimi (Moonshot AI) API.
+// It handles the RFC 8628 OAuth2 Device Authorization Grant flow for secure authentication.
+package kimi
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"os"
+	"runtime"
+	"strings"
+	"time"
+
+	"github.com/google/uuid"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	log "github.com/sirupsen/logrus"
+)
+
+const (
+	// kimiClientID is Kimi Code's OAuth client ID.
+	kimiClientID = "17e5f671-d194-4dfb-9706-5516cb48c098"
+	// kimiOAuthHost is the OAuth server endpoint.
+	kimiOAuthHost = "https://auth.kimi.com"
+	// kimiDeviceCodeURL is the endpoint for requesting device codes.
+	kimiDeviceCodeURL = kimiOAuthHost + "/api/oauth/device_authorization"
+	// kimiTokenURL is the endpoint for exchanging device codes for tokens.
+	kimiTokenURL = kimiOAuthHost + "/api/oauth/token"
+	// KimiAPIBaseURL is the base URL for Kimi API requests.
+	KimiAPIBaseURL = "https://api.kimi.com/coding"
+	// defaultPollInterval is the default interval for polling token endpoint.
+	defaultPollInterval = 5 * time.Second
+	// maxPollDuration is the maximum time to wait for user authorization.
+	maxPollDuration = 15 * time.Minute
+	// refreshThresholdSeconds is when to refresh token before expiry (5 minutes).
+	refreshThresholdSeconds = 300
+)
+
+// KimiAuth handles Kimi authentication flow.
+type KimiAuth struct {
+	deviceClient *DeviceFlowClient
+	cfg          *config.Config
+}
+
+// NewKimiAuth creates a new KimiAuth service instance.
+func NewKimiAuth(cfg *config.Config) *KimiAuth {
+	return &KimiAuth{
+		deviceClient: NewDeviceFlowClient(cfg),
+		cfg:          cfg,
+	}
+}
+
+// StartDeviceFlow initiates the device flow authentication.
+func (k *KimiAuth) StartDeviceFlow(ctx context.Context) (*DeviceCodeResponse, error) {
+	return k.deviceClient.RequestDeviceCode(ctx)
+}
+
+// WaitForAuthorization polls for user authorization and returns the auth bundle.
+func (k *KimiAuth) WaitForAuthorization(ctx context.Context, deviceCode *DeviceCodeResponse) (*KimiAuthBundle, error) {
+	tokenData, err := k.deviceClient.PollForToken(ctx, deviceCode)
+	if err != nil {
+		return nil, err
+	}
+
+	return &KimiAuthBundle{
+		TokenData: tokenData,
+		DeviceID:  k.deviceClient.deviceID,
+	}, nil
+}
+
+// CreateTokenStorage creates a new KimiTokenStorage from auth bundle.
+func (k *KimiAuth) CreateTokenStorage(bundle *KimiAuthBundle) *KimiTokenStorage {
+	expired := ""
+	if bundle.TokenData.ExpiresAt > 0 {
+		expired = time.Unix(bundle.TokenData.ExpiresAt, 0).UTC().Format(time.RFC3339)
+	}
+	return &KimiTokenStorage{
+		AccessToken:  bundle.TokenData.AccessToken,
+		RefreshToken: bundle.TokenData.RefreshToken,
+		TokenType:    bundle.TokenData.TokenType,
+		Scope:        bundle.TokenData.Scope,
+		DeviceID:     strings.TrimSpace(bundle.DeviceID),
+		Expired:      expired,
+		Type:         "kimi",
+	}
+}
+
+// DeviceFlowClient handles the OAuth2 device flow for Kimi.
+type DeviceFlowClient struct {
+	httpClient *http.Client
+	cfg        *config.Config
+	deviceID   string
+}
+
+// NewDeviceFlowClient creates a new device flow client.
+func NewDeviceFlowClient(cfg *config.Config) *DeviceFlowClient {
+	return NewDeviceFlowClientWithDeviceID(cfg, "")
+}
+
+// NewDeviceFlowClientWithDeviceID creates a new device flow client with the specified device ID.
+func NewDeviceFlowClientWithDeviceID(cfg *config.Config, deviceID string) *DeviceFlowClient {
+	client := &http.Client{Timeout: 30 * time.Second}
+	if cfg != nil {
+		client = util.SetProxy(&cfg.SDKConfig, client)
+	}
+	resolvedDeviceID := strings.TrimSpace(deviceID)
+	if resolvedDeviceID == "" {
+		resolvedDeviceID = getOrCreateDeviceID()
+	}
+	return &DeviceFlowClient{
+		httpClient: client,
+		cfg:        cfg,
+		deviceID:   resolvedDeviceID,
+	}
+}
+
+// getOrCreateDeviceID returns an in-memory device ID for the current authentication flow.
+func getOrCreateDeviceID() string {
+	return uuid.New().String()
+}
+
+// getDeviceModel returns a device model string.
+func getDeviceModel() string {
+	osName := runtime.GOOS
+	arch := runtime.GOARCH
+
+	switch osName {
+	case "darwin":
+		return fmt.Sprintf("macOS %s", arch)
+	case "windows":
+		return fmt.Sprintf("Windows %s", arch)
+	case "linux":
+		return fmt.Sprintf("Linux %s", arch)
+	default:
+		return fmt.Sprintf("%s %s", osName, arch)
+	}
+}
+
+// getHostname returns the machine hostname.
+func getHostname() string {
+	hostname, err := os.Hostname()
+	if err != nil {
+		return "unknown"
+	}
+	return hostname
+}
+
+// commonHeaders returns headers required for Kimi API requests.
+func (c *DeviceFlowClient) commonHeaders() map[string]string {
+	return map[string]string{
+		"X-Msh-Platform":     "cli-proxy-api",
+		"X-Msh-Version":      "1.0.0",
+		"X-Msh-Device-Name":  getHostname(),
+		"X-Msh-Device-Model": getDeviceModel(),
+		"X-Msh-Device-Id":    c.deviceID,
+	}
+}
+
+// RequestDeviceCode initiates the device flow by requesting a device code from Kimi.
+func (c *DeviceFlowClient) RequestDeviceCode(ctx context.Context) (*DeviceCodeResponse, error) {
+	data := url.Values{}
+	data.Set("client_id", kimiClientID)
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, kimiDeviceCodeURL, strings.NewReader(data.Encode()))
+	if err != nil {
+		return nil, fmt.Errorf("kimi: failed to create device code request: %w", err)
+	}
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	req.Header.Set("Accept", "application/json")
+	for k, v := range c.commonHeaders() {
+		req.Header.Set(k, v)
+	}
+
+	resp, err := c.httpClient.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("kimi: device code request failed: %w", err)
+	}
+	defer func() {
+		if errClose := resp.Body.Close(); errClose != nil {
+			log.Errorf("kimi device code: close body error: %v", errClose)
+		}
+	}()
+
+	bodyBytes, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, fmt.Errorf("kimi: failed to read device code response: %w", err)
+	}
+
+	if resp.StatusCode != http.StatusOK {
+		return nil, fmt.Errorf("kimi: device code request failed with status %d: %s", resp.StatusCode, string(bodyBytes))
+	}
+
+	var deviceCode DeviceCodeResponse
+	if err = json.Unmarshal(bodyBytes, &deviceCode); err != nil {
+		return nil, fmt.Errorf("kimi: failed to parse device code response: %w", err)
+	}
+
+	return &deviceCode, nil
+}
+
+// PollForToken polls the token endpoint until the user authorizes or the device code expires.
+func (c *DeviceFlowClient) PollForToken(ctx context.Context, deviceCode *DeviceCodeResponse) (*KimiTokenData, error) {
+	if deviceCode == nil {
+		return nil, fmt.Errorf("kimi: device code is nil")
+	}
+
+	interval := time.Duration(deviceCode.Interval) * time.Second
+	if interval < defaultPollInterval {
+		interval = defaultPollInterval
+	}
+
+	deadline := time.Now().Add(maxPollDuration)
+	if deviceCode.ExpiresIn > 0 {
+		codeDeadline := time.Now().Add(time.Duration(deviceCode.ExpiresIn) * time.Second)
+		if codeDeadline.Before(deadline) {
+			deadline = codeDeadline
+		}
+	}
+
+	ticker := time.NewTicker(interval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return nil, fmt.Errorf("kimi: context cancelled: %w", ctx.Err())
+		case <-ticker.C:
+			if time.Now().After(deadline) {
+				return nil, fmt.Errorf("kimi: device code expired")
+			}
+
+			token, pollErr, shouldContinue := c.exchangeDeviceCode(ctx, deviceCode.DeviceCode)
+			if token != nil {
+				return token, nil
+			}
+			if !shouldContinue {
+				return nil, pollErr
+			}
+			// Continue polling
+		}
+	}
+}
+
+// exchangeDeviceCode attempts to exchange the device code for an access token.
+// Returns (token, error, shouldContinue).
+func (c *DeviceFlowClient) exchangeDeviceCode(ctx context.Context, deviceCode string) (*KimiTokenData, error, bool) {
+	data := url.Values{}
+	data.Set("client_id", kimiClientID)
+	data.Set("device_code", deviceCode)
+	data.Set("grant_type", "urn:ietf:params:oauth:grant-type:device_code")
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, kimiTokenURL, strings.NewReader(data.Encode()))
+	if err != nil {
+		return nil, fmt.Errorf("kimi: failed to create token request: %w", err), false
+	}
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	req.Header.Set("Accept", "application/json")
+	for k, v := range c.commonHeaders() {
+		req.Header.Set(k, v)
+	}
+
+	resp, err := c.httpClient.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("kimi: token request failed: %w", err), false
+	}
+	defer func() {
+		if errClose := resp.Body.Close(); errClose != nil {
+			log.Errorf("kimi token exchange: close body error: %v", errClose)
+		}
+	}()
+
+	bodyBytes, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, fmt.Errorf("kimi: failed to read token response: %w", err), false
+	}
+
+	// Parse response - Kimi returns 200 for both success and pending states
+	var oauthResp struct {
+		Error            string  `json:"error"`
+		ErrorDescription string  `json:"error_description"`
+		AccessToken      string  `json:"access_token"`
+		RefreshToken     string  `json:"refresh_token"`
+		TokenType        string  `json:"token_type"`
+		ExpiresIn        float64 `json:"expires_in"`
+		Scope            string  `json:"scope"`
+	}
+
+	if err = json.Unmarshal(bodyBytes, &oauthResp); err != nil {
+		return nil, fmt.Errorf("kimi: failed to parse token response: %w", err), false
+	}
+
+	if oauthResp.Error != "" {
+		switch oauthResp.Error {
+		case "authorization_pending":
+			return nil, nil, true // Continue polling
+		case "slow_down":
+			return nil, nil, true // Continue polling (with increased interval handled by caller)
+		case "expired_token":
+			return nil, fmt.Errorf("kimi: device code expired"), false
+		case "access_denied":
+			return nil, fmt.Errorf("kimi: access denied by user"), false
+		default:
+			return nil, fmt.Errorf("kimi: OAuth error: %s - %s", oauthResp.Error, oauthResp.ErrorDescription), false
+		}
+	}
+
+	if oauthResp.AccessToken == "" {
+		return nil, fmt.Errorf("kimi: empty access token in response"), false
+	}
+
+	var expiresAt int64
+	if oauthResp.ExpiresIn > 0 {
+		expiresAt = time.Now().Unix() + int64(oauthResp.ExpiresIn)
+	}
+
+	return &KimiTokenData{
+		AccessToken:  oauthResp.AccessToken,
+		RefreshToken: oauthResp.RefreshToken,
+		TokenType:    oauthResp.TokenType,
+		ExpiresAt:    expiresAt,
+		Scope:        oauthResp.Scope,
+	}, nil, false
+}
+
+// RefreshToken exchanges a refresh token for a new access token.
+func (c *DeviceFlowClient) RefreshToken(ctx context.Context, refreshToken string) (*KimiTokenData, error) {
+	data := url.Values{}
+	data.Set("client_id", kimiClientID)
+	data.Set("grant_type", "refresh_token")
+	data.Set("refresh_token", refreshToken)
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, kimiTokenURL, strings.NewReader(data.Encode()))
+	if err != nil {
+		return nil, fmt.Errorf("kimi: failed to create refresh request: %w", err)
+	}
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	req.Header.Set("Accept", "application/json")
+	for k, v := range c.commonHeaders() {
+		req.Header.Set(k, v)
+	}
+
+	resp, err := c.httpClient.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("kimi: refresh request failed: %w", err)
+	}
+	defer func() {
+		if errClose := resp.Body.Close(); errClose != nil {
+			log.Errorf("kimi refresh token: close body error: %v", errClose)
+		}
+	}()
+
+	bodyBytes, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, fmt.Errorf("kimi: failed to read refresh response: %w", err)
+	}
+
+	if resp.StatusCode == http.StatusUnauthorized || resp.StatusCode == http.StatusForbidden {
+		return nil, fmt.Errorf("kimi: refresh token rejected (status %d)", resp.StatusCode)
+	}
+
+	if resp.StatusCode != http.StatusOK {
+		return nil, fmt.Errorf("kimi: refresh failed with status %d: %s", resp.StatusCode, string(bodyBytes))
+	}
+
+	var tokenResp struct {
+		AccessToken  string  `json:"access_token"`
+		RefreshToken string  `json:"refresh_token"`
+		TokenType    string  `json:"token_type"`
+		ExpiresIn    float64 `json:"expires_in"`
+		Scope        string  `json:"scope"`
+	}
+
+	if err = json.Unmarshal(bodyBytes, &tokenResp); err != nil {
+		return nil, fmt.Errorf("kimi: failed to parse refresh response: %w", err)
+	}
+
+	if tokenResp.AccessToken == "" {
+		return nil, fmt.Errorf("kimi: empty access token in refresh response")
+	}
+
+	var expiresAt int64
+	if tokenResp.ExpiresIn > 0 {
+		expiresAt = time.Now().Unix() + int64(tokenResp.ExpiresIn)
+	}
+
+	return &KimiTokenData{
+		AccessToken:  tokenResp.AccessToken,
+		RefreshToken: tokenResp.RefreshToken,
+		TokenType:    tokenResp.TokenType,
+		ExpiresAt:    expiresAt,
+		Scope:        tokenResp.Scope,
+	}, nil
+}

internal/auth/kimi/token.go

@@ -0,0 +1,116 @@
+// Package kimi provides authentication and token management functionality
+// for Kimi (Moonshot AI) services. It handles OAuth2 device flow token storage,
+// serialization, and retrieval for maintaining authenticated sessions with the Kimi API.
+package kimi
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
+)
+
+// KimiTokenStorage stores OAuth2 token information for Kimi API authentication.
+type KimiTokenStorage struct {
+	// AccessToken is the OAuth2 access token used for authenticating API requests.
+	AccessToken string `json:"access_token"`
+	// RefreshToken is the OAuth2 refresh token used to obtain new access tokens.
+	RefreshToken string `json:"refresh_token"`
+	// TokenType is the type of token, typically "Bearer".
+	TokenType string `json:"token_type"`
+	// Scope is the OAuth2 scope granted to the token.
+	Scope string `json:"scope,omitempty"`
+	// DeviceID is the OAuth device flow identifier used for Kimi requests.
+	DeviceID string `json:"device_id,omitempty"`
+	// Expired is the RFC3339 timestamp when the access token expires.
+	Expired string `json:"expired,omitempty"`
+	// Type indicates the authentication provider type, always "kimi" for this storage.
+	Type string `json:"type"`
+}
+
+// KimiTokenData holds the raw OAuth token response from Kimi.
+type KimiTokenData struct {
+	// AccessToken is the OAuth2 access token.
+	AccessToken string `json:"access_token"`
+	// RefreshToken is the OAuth2 refresh token.
+	RefreshToken string `json:"refresh_token"`
+	// TokenType is the type of token, typically "Bearer".
+	TokenType string `json:"token_type"`
+	// ExpiresAt is the Unix timestamp when the token expires.
+	ExpiresAt int64 `json:"expires_at"`
+	// Scope is the OAuth2 scope granted to the token.
+	Scope string `json:"scope"`
+}
+
+// KimiAuthBundle bundles authentication data for storage.
+type KimiAuthBundle struct {
+	// TokenData contains the OAuth token information.
+	TokenData *KimiTokenData
+	// DeviceID is the device identifier used during OAuth device flow.
+	DeviceID string
+}
+
+// DeviceCodeResponse represents Kimi's device code response.
+type DeviceCodeResponse struct {
+	// DeviceCode is the device verification code.
+	DeviceCode string `json:"device_code"`
+	// UserCode is the code the user must enter at the verification URI.
+	UserCode string `json:"user_code"`
+	// VerificationURI is the URL where the user should enter the code.
+	VerificationURI string `json:"verification_uri,omitempty"`
+	// VerificationURIComplete is the URL with the code pre-filled.
+	VerificationURIComplete string `json:"verification_uri_complete"`
+	// ExpiresIn is the number of seconds until the device code expires.
+	ExpiresIn int `json:"expires_in"`
+	// Interval is the minimum number of seconds to wait between polling requests.
+	Interval int `json:"interval"`
+}
+
+// SaveTokenToFile serializes the Kimi token storage to a JSON file.
+func (ts *KimiTokenStorage) SaveTokenToFile(authFilePath string) error {
+	misc.LogSavingCredentials(authFilePath)
+	ts.Type = "kimi"
+
+	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() {
+		_ = f.Close()
+	}()
+
+	encoder := json.NewEncoder(f)
+	encoder.SetIndent("", "  ")
+	if err = encoder.Encode(ts); err != nil {
+		return fmt.Errorf("failed to write token to file: %w", err)
+	}
+	return nil
+}
+
+// IsExpired checks if the token has expired.
+func (ts *KimiTokenStorage) IsExpired() bool {
+	if ts.Expired == "" {
+		return false // No expiry set, assume valid
+	}
+	t, err := time.Parse(time.RFC3339, ts.Expired)
+	if err != nil {
+		return true // Has expiry string but can't parse
+	}
+	// Consider expired if within refresh threshold
+	return time.Now().Add(time.Duration(refreshThresholdSeconds) * time.Second).After(t)
+}
+
+// NeedsRefresh checks if the token should be refreshed.
+func (ts *KimiTokenStorage) NeedsRefresh() bool {
+	if ts.RefreshToken == "" {
+		return false // Can't refresh without refresh token
+	}
+	return ts.IsExpired()
+}

internal/auth/kiro/aws_auth.go

@@ -238,7 +238,7 @@ func (k *KiroAuth) ListAvailableModels(ctx context.Context, tokenData *KiroToken
 			Description    string  `json:"description"`
 			RateMultiplier float64 `json:"rateMultiplier"`
 			RateUnit       string  `json:"rateUnit"`
-			TokenLimits    struct {
+			TokenLimits    *struct {
 				MaxInputTokens int `json:"maxInputTokens"`
 			} `json:"tokenLimits"`
 		} `json:"models"`
@@ -250,13 +250,17 @@ func (k *KiroAuth) ListAvailableModels(ctx context.Context, tokenData *KiroToken
 
 	models := make([]*KiroModel, 0, len(result.Models))
 	for _, m := range result.Models {
+		maxInputTokens := 0
+		if m.TokenLimits != nil {
+			maxInputTokens = m.TokenLimits.MaxInputTokens
+		}
 		models = append(models, &KiroModel{
 			ModelID:        m.ModelID,
 			ModelName:      m.ModelName,
 			Description:    m.Description,
 			RateMultiplier: m.RateMultiplier,
 			RateUnit:       m.RateUnit,
-			MaxInputTokens: m.TokenLimits.MaxInputTokens,
+			MaxInputTokens: maxInputTokens,
 		})
 	}
 

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)

internal/cmd/auth_manager.go

@@ -19,6 +19,7 @@ func newAuthManager() *sdkAuth.Manager {
 		sdkAuth.NewQwenAuthenticator(),
 		sdkAuth.NewIFlowAuthenticator(),
 		sdkAuth.NewAntigravityAuthenticator(),
+		sdkAuth.NewKimiAuthenticator(),
 		sdkAuth.NewKiroAuthenticator(),
 		sdkAuth.NewGitHubCopilotAuthenticator(),
 	)

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
 		}

internal/cmd/kimi_login.go

@@ -0,0 +1,44 @@
+package cmd
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	sdkAuth "github.com/router-for-me/CLIProxyAPI/v6/sdk/auth"
+	log "github.com/sirupsen/logrus"
+)
+
+// DoKimiLogin triggers the OAuth device flow for Kimi (Moonshot AI) and saves tokens.
+// It initiates the device flow authentication, displays the verification URL for the user,
+// and waits for authorization before saving the tokens.
+//
+// Parameters:
+//   - cfg: The application configuration containing proxy and auth directory settings
+//   - options: Login options including browser behavior settings
+func DoKimiLogin(cfg *config.Config, options *LoginOptions) {
+	if options == nil {
+		options = &LoginOptions{}
+	}
+
+	manager := newAuthManager()
+	authOpts := &sdkAuth.LoginOptions{
+		NoBrowser: options.NoBrowser,
+		Metadata:  map[string]string{},
+		Prompt:    options.Prompt,
+	}
+
+	record, savedPath, err := manager.Login(context.Background(), "kimi", cfg, authOpts)
+	if err != nil {
+		log.Errorf("Kimi authentication failed: %v", err)
+		return
+	}
+
+	if savedPath != "" {
+		fmt.Printf("Authentication saved to %s\n", savedPath)
+	}
+	if record != nil && record.Label != "" {
+		fmt.Printf("Authenticated as %s\n", record.Label)
+	}
+	fmt.Println("Kimi authentication successful!")
+}

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)

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)

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
 		}

internal/config/config.go

@@ -535,14 +535,15 @@ func LoadConfig(configFile string) (*Config, error) {
 // If optional is true and the file is missing, it returns an empty Config.
 // If optional is true and the file is empty or invalid, it returns an empty Config.
 func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
-	// Perform oauth-model-alias migration before loading config.
-	// This migrates oauth-model-mappings to oauth-model-alias if needed.
-	if migrated, err := MigrateOAuthModelAlias(configFile); err != nil {
-		// Log warning but don't fail - config loading should still work
-		fmt.Printf("Warning: oauth-model-alias migration failed: %v\n", err)
-	} else if migrated {
-		fmt.Println("Migrated oauth-model-mappings to oauth-model-alias")
-	}
+	// NOTE: Startup oauth-model-alias migration is intentionally disabled.
+	// Reason: avoid mutating config.yaml during server startup.
+	// Re-enable the block below if automatic startup migration is needed again.
+	// if migrated, err := MigrateOAuthModelAlias(configFile); err != nil {
+	// 	// Log warning but don't fail - config loading should still work
+	// 	fmt.Printf("Warning: oauth-model-alias migration failed: %v\n", err)
+	// } else if migrated {
+	// 	fmt.Println("Migrated oauth-model-mappings to oauth-model-alias")
+	// }
 
 	// Read the entire configuration file into memory.
 	data, err := os.ReadFile(configFile)
@@ -583,18 +584,21 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 		return nil, fmt.Errorf("failed to parse config file: %w", err)
 	}
 
-	var legacy legacyConfigData
-	if errLegacy := yaml.Unmarshal(data, &legacy); errLegacy == nil {
-		if cfg.migrateLegacyGeminiKeys(legacy.LegacyGeminiKeys) {
-			cfg.legacyMigrationPending = true
-		}
-		if cfg.migrateLegacyOpenAICompatibilityKeys(legacy.OpenAICompat) {
-			cfg.legacyMigrationPending = true
-		}
-		if cfg.migrateLegacyAmpConfig(&legacy) {
-			cfg.legacyMigrationPending = true
-		}
-	}
+	// NOTE: Startup legacy key migration is intentionally disabled.
+	// Reason: avoid mutating config.yaml during server startup.
+	// Re-enable the block below if automatic startup migration is needed again.
+	// var legacy legacyConfigData
+	// if errLegacy := yaml.Unmarshal(data, &legacy); errLegacy == nil {
+	// 	if cfg.migrateLegacyGeminiKeys(legacy.LegacyGeminiKeys) {
+	// 		cfg.legacyMigrationPending = true
+	// 	}
+	// 	if cfg.migrateLegacyOpenAICompatibilityKeys(legacy.OpenAICompat) {
+	// 		cfg.legacyMigrationPending = true
+	// 	}
+	// 	if cfg.migrateLegacyAmpConfig(&legacy) {
+	// 		cfg.legacyMigrationPending = true
+	// 	}
+	// }
 
 	// Hash remote management key if plaintext is detected (nested)
 	// We consider a value to be already hashed if it looks like a bcrypt hash ($2a$, $2b$, or $2y$ prefix).
@@ -628,9 +632,6 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 		cfg.ErrorLogsMaxFiles = 10
 	}
 
-	// Sync request authentication providers with inline API keys for backwards compatibility.
-	syncInlineAccessProvider(&cfg)
-
 	// Sanitize Gemini API key configuration and migrate legacy entries.
 	cfg.SanitizeGeminiKeys()
 
@@ -658,17 +659,20 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 	// Validate raw payload rules and drop invalid entries.
 	cfg.SanitizePayloadRules()
 
-	if cfg.legacyMigrationPending {
-		fmt.Println("Detected legacy configuration keys, attempting to persist the normalized config...")
-		if !optional && configFile != "" {
-			if err := SaveConfigPreserveComments(configFile, &cfg); err != nil {
-				return nil, fmt.Errorf("failed to persist migrated legacy config: %w", err)
-			}
-			fmt.Println("Legacy configuration normalized and persisted.")
-		} else {
-			fmt.Println("Legacy configuration normalized in memory; persistence skipped.")
-		}
-	}
+	// NOTE: Legacy migration persistence is intentionally disabled together with
+	// startup legacy migration to keep startup read-only for config.yaml.
+	// Re-enable the block below if automatic startup migration is needed again.
+	// if cfg.legacyMigrationPending {
+	// 	fmt.Println("Detected legacy configuration keys, attempting to persist the normalized config...")
+	// 	if !optional && configFile != "" {
+	// 		if err := SaveConfigPreserveComments(configFile, &cfg); err != nil {
+	// 			return nil, fmt.Errorf("failed to persist migrated legacy config: %w", err)
+	// 		}
+	// 		fmt.Println("Legacy configuration normalized and persisted.")
+	// 	} else {
+	// 		fmt.Println("Legacy configuration normalized in memory; persistence skipped.")
+	// 	}
+	// }
 
 	// Return the populated configuration struct.
 	return &cfg, nil
@@ -732,14 +736,44 @@ func payloadRawString(value any) ([]byte, bool) {
 // SanitizeOAuthModelAlias normalizes and deduplicates global OAuth model name aliases.
 // It trims whitespace, normalizes channel keys to lower-case, drops empty entries,
 // allows multiple aliases per upstream name, and ensures aliases are unique within each channel.
+// It also injects default aliases for channels that have built-in defaults (e.g., kiro)
+// when no user-configured aliases exist for those channels.
 func (cfg *Config) SanitizeOAuthModelAlias() {
-	if cfg == nil || len(cfg.OAuthModelAlias) == 0 {
+	if cfg == nil {
+		return
+	}
+
+	// Inject default Kiro aliases if no user-configured kiro aliases exist
+	if cfg.OAuthModelAlias == nil {
+		cfg.OAuthModelAlias = make(map[string][]OAuthModelAlias)
+	}
+	if _, hasKiro := cfg.OAuthModelAlias["kiro"]; !hasKiro {
+		// Check case-insensitive too
+		found := false
+		for k := range cfg.OAuthModelAlias {
+			if strings.EqualFold(strings.TrimSpace(k), "kiro") {
+				found = true
+				break
+			}
+		}
+		if !found {
+			cfg.OAuthModelAlias["kiro"] = defaultKiroAliases()
+		}
+	}
+
+	if len(cfg.OAuthModelAlias) == 0 {
 		return
 	}
 	out := make(map[string][]OAuthModelAlias, len(cfg.OAuthModelAlias))
 	for rawChannel, aliases := range cfg.OAuthModelAlias {
 		channel := strings.ToLower(strings.TrimSpace(rawChannel))
-		if channel == "" || len(aliases) == 0 {
+		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))
@@ -881,18 +915,6 @@ func normalizeModelPrefix(prefix string) string {
 	return trimmed
 }
 
-func syncInlineAccessProvider(cfg *Config) {
-	if cfg == nil {
-		return
-	}
-	if len(cfg.APIKeys) == 0 {
-		if provider := cfg.ConfigAPIKeyProvider(); provider != nil && len(provider.APIKeys) > 0 {
-			cfg.APIKeys = append([]string(nil), provider.APIKeys...)
-		}
-	}
-	cfg.Access.Providers = nil
-}
-
 // looksLikeBcrypt returns true if the provided string appears to be a bcrypt hash.
 func looksLikeBcrypt(s string) bool {
 	return len(s) > 4 && (s[:4] == "$2a$" || s[:4] == "$2b$" || s[:4] == "$2y$")
@@ -980,7 +1002,7 @@ func hashSecret(secret string) (string, error) {
 // SaveConfigPreserveComments writes the config back to YAML while preserving existing comments
 // and key ordering by loading the original file into a yaml.Node tree and updating values in-place.
 func SaveConfigPreserveComments(configFile string, cfg *Config) error {
-	persistCfg := sanitizeConfigForPersist(cfg)
+	persistCfg := cfg
 	// Load original YAML as a node tree to preserve comments and ordering.
 	data, err := os.ReadFile(configFile)
 	if err != nil {
@@ -1048,16 +1070,6 @@ func SaveConfigPreserveComments(configFile string, cfg *Config) error {
 	return err
 }
 
-func sanitizeConfigForPersist(cfg *Config) *Config {
-	if cfg == nil {
-		return nil
-	}
-	clone := *cfg
-	clone.SDKConfig = cfg.SDKConfig
-	clone.SDKConfig.Access = AccessConfig{}
-	return &clone
-}
-
 // SaveConfigPreserveCommentsUpdateNestedScalar updates a nested scalar key path like ["a","b"]
 // while preserving comments and positions.
 func SaveConfigPreserveCommentsUpdateNestedScalar(configFile string, path []string, value string) error {
@@ -1154,8 +1166,13 @@ func getOrCreateMapValue(mapNode *yaml.Node, key string) *yaml.Node {
 
 // mergeMappingPreserve merges keys from src into dst mapping node while preserving
 // key order and comments of existing keys in dst. New keys are only added if their
-// value is non-zero to avoid polluting the config with defaults.
-func mergeMappingPreserve(dst, src *yaml.Node) {
+// value is non-zero and not a known default to avoid polluting the config with defaults.
+func mergeMappingPreserve(dst, src *yaml.Node, path ...[]string) {
+	var currentPath []string
+	if len(path) > 0 {
+		currentPath = path[0]
+	}
+
 	if dst == nil || src == nil {
 		return
 	}
@@ -1169,16 +1186,19 @@ func mergeMappingPreserve(dst, src *yaml.Node) {
 		sk := src.Content[i]
 		sv := src.Content[i+1]
 		idx := findMapKeyIndex(dst, sk.Value)
+		childPath := appendPath(currentPath, sk.Value)
 		if idx >= 0 {
 			// Merge into existing value node (always update, even to zero values)
 			dv := dst.Content[idx+1]
-			mergeNodePreserve(dv, sv)
+			mergeNodePreserve(dv, sv, childPath)
 		} else {
-			// New key: only add if value is non-zero to avoid polluting config with defaults
-			if isZeroValueNode(sv) {
+			// New key: only add if value is non-zero and not a known default
+			candidate := deepCopyNode(sv)
+			pruneKnownDefaultsInNewNode(childPath, candidate)
+			if isKnownDefaultValue(childPath, candidate) {
 				continue
 			}
-			dst.Content = append(dst.Content, deepCopyNode(sk), deepCopyNode(sv))
+			dst.Content = append(dst.Content, deepCopyNode(sk), candidate)
 		}
 	}
 }
@@ -1186,7 +1206,12 @@ func mergeMappingPreserve(dst, src *yaml.Node) {
 // mergeNodePreserve merges src into dst for scalars, mappings and sequences while
 // reusing destination nodes to keep comments and anchors. For sequences, it updates
 // in-place by index.
-func mergeNodePreserve(dst, src *yaml.Node) {
+func mergeNodePreserve(dst, src *yaml.Node, path ...[]string) {
+	var currentPath []string
+	if len(path) > 0 {
+		currentPath = path[0]
+	}
+
 	if dst == nil || src == nil {
 		return
 	}
@@ -1195,7 +1220,7 @@ func mergeNodePreserve(dst, src *yaml.Node) {
 		if dst.Kind != yaml.MappingNode {
 			copyNodeShallow(dst, src)
 		}
-		mergeMappingPreserve(dst, src)
+		mergeMappingPreserve(dst, src, currentPath)
 	case yaml.SequenceNode:
 		// Preserve explicit null style if dst was null and src is empty sequence
 		if dst.Kind == yaml.ScalarNode && dst.Tag == "!!null" && len(src.Content) == 0 {
@@ -1218,7 +1243,7 @@ func mergeNodePreserve(dst, src *yaml.Node) {
 				dst.Content[i] = deepCopyNode(src.Content[i])
 				continue
 			}
-			mergeNodePreserve(dst.Content[i], src.Content[i])
+			mergeNodePreserve(dst.Content[i], src.Content[i], currentPath)
 			if dst.Content[i] != nil && src.Content[i] != nil &&
 				dst.Content[i].Kind == yaml.MappingNode && src.Content[i].Kind == yaml.MappingNode {
 				pruneMissingMapKeys(dst.Content[i], src.Content[i])
@@ -1260,6 +1285,94 @@ func findMapKeyIndex(mapNode *yaml.Node, key string) int {
 	return -1
 }
 
+// appendPath appends a key to the path, returning a new slice to avoid modifying the original.
+func appendPath(path []string, key string) []string {
+	if len(path) == 0 {
+		return []string{key}
+	}
+	newPath := make([]string, len(path)+1)
+	copy(newPath, path)
+	newPath[len(path)] = key
+	return newPath
+}
+
+// isKnownDefaultValue returns true if the given node at the specified path
+// represents a known default value that should not be written to the config file.
+// This prevents non-zero defaults from polluting the config.
+func isKnownDefaultValue(path []string, node *yaml.Node) bool {
+	// First check if it's a zero value
+	if isZeroValueNode(node) {
+		return true
+	}
+
+	// Match known non-zero defaults by exact dotted path.
+	if len(path) == 0 {
+		return false
+	}
+
+	fullPath := strings.Join(path, ".")
+
+	// Check string defaults
+	if node.Kind == yaml.ScalarNode && node.Tag == "!!str" {
+		switch fullPath {
+		case "pprof.addr":
+			return node.Value == DefaultPprofAddr
+		case "remote-management.panel-github-repository":
+			return node.Value == DefaultPanelGitHubRepository
+		case "routing.strategy":
+			return node.Value == "round-robin"
+		}
+	}
+
+	// Check integer defaults
+	if node.Kind == yaml.ScalarNode && node.Tag == "!!int" {
+		switch fullPath {
+		case "error-logs-max-files":
+			return node.Value == "10"
+		}
+	}
+
+	return false
+}
+
+// pruneKnownDefaultsInNewNode removes default-valued descendants from a new node
+// before it is appended into the destination YAML tree.
+func pruneKnownDefaultsInNewNode(path []string, node *yaml.Node) {
+	if node == nil {
+		return
+	}
+
+	switch node.Kind {
+	case yaml.MappingNode:
+		filtered := make([]*yaml.Node, 0, len(node.Content))
+		for i := 0; i+1 < len(node.Content); i += 2 {
+			keyNode := node.Content[i]
+			valueNode := node.Content[i+1]
+			if keyNode == nil || valueNode == nil {
+				continue
+			}
+
+			childPath := appendPath(path, keyNode.Value)
+			if isKnownDefaultValue(childPath, valueNode) {
+				continue
+			}
+
+			pruneKnownDefaultsInNewNode(childPath, valueNode)
+			if (valueNode.Kind == yaml.MappingNode || valueNode.Kind == yaml.SequenceNode) &&
+				len(valueNode.Content) == 0 {
+				continue
+			}
+
+			filtered = append(filtered, keyNode, valueNode)
+		}
+		node.Content = filtered
+	case yaml.SequenceNode:
+		for _, child := range node.Content {
+			pruneKnownDefaultsInNewNode(path, child)
+		}
+	}
+}
+
 // isZeroValueNode returns true if the YAML node represents a zero/default value
 // that should not be written as a new key to preserve config cleanliness.
 // For mappings and sequences, recursively checks if all children are zero values.

internal/config/oauth_model_alias_migration.go

@@ -20,6 +20,28 @@ var antigravityModelConversionTable = map[string]string{
 	"gemini-claude-opus-4-6-thinking":         "claude-opus-4-6-thinking",
 }
 
+// defaultKiroAliases returns the default oauth-model-alias configuration
+// for the kiro channel. Maps kiro-prefixed model names to standard Claude model
+// names so that clients like Claude Code can use standard names directly.
+func defaultKiroAliases() []OAuthModelAlias {
+	return []OAuthModelAlias{
+		// Sonnet 4.5
+		{Name: "kiro-claude-sonnet-4-5", Alias: "claude-sonnet-4-5-20250929", Fork: true},
+		{Name: "kiro-claude-sonnet-4-5", Alias: "claude-sonnet-4-5", Fork: true},
+		// Sonnet 4
+		{Name: "kiro-claude-sonnet-4", Alias: "claude-sonnet-4-20250514", Fork: true},
+		{Name: "kiro-claude-sonnet-4", Alias: "claude-sonnet-4", Fork: true},
+		// Opus 4.6
+		{Name: "kiro-claude-opus-4-6", Alias: "claude-opus-4-6", Fork: true},
+		// Opus 4.5
+		{Name: "kiro-claude-opus-4-5", Alias: "claude-opus-4-5-20251101", Fork: true},
+		{Name: "kiro-claude-opus-4-5", Alias: "claude-opus-4-5", Fork: true},
+		// Haiku 4.5
+		{Name: "kiro-claude-haiku-4-5", Alias: "claude-haiku-4-5-20251001", Fork: true},
+		{Name: "kiro-claude-haiku-4-5", Alias: "claude-haiku-4-5", Fork: true},
+	}
+}
+
 // defaultAntigravityAliases returns the default oauth-model-alias configuration
 // for the antigravity channel when neither field exists.
 func defaultAntigravityAliases() []OAuthModelAlias {

internal/config/oauth_model_alias_test.go

@@ -54,3 +54,132 @@ func TestSanitizeOAuthModelAlias_AllowsMultipleAliasesForSameName(t *testing.T)
 		}
 	}
 }
+
+func TestSanitizeOAuthModelAlias_InjectsDefaultKiroAliases(t *testing.T) {
+	// When no kiro aliases are configured, defaults should be injected
+	cfg := &Config{
+		OAuthModelAlias: map[string][]OAuthModelAlias{
+			"codex": {
+				{Name: "gpt-5", Alias: "g5"},
+			},
+		},
+	}
+
+	cfg.SanitizeOAuthModelAlias()
+
+	kiroAliases := cfg.OAuthModelAlias["kiro"]
+	if len(kiroAliases) == 0 {
+		t.Fatal("expected default kiro aliases to be injected")
+	}
+
+	// Check that standard Claude model names are present
+	aliasSet := make(map[string]bool)
+	for _, a := range kiroAliases {
+		aliasSet[a.Alias] = true
+	}
+	expectedAliases := []string{
+		"claude-sonnet-4-5-20250929",
+		"claude-sonnet-4-5",
+		"claude-sonnet-4-20250514",
+		"claude-sonnet-4",
+		"claude-opus-4-6",
+		"claude-opus-4-5-20251101",
+		"claude-opus-4-5",
+		"claude-haiku-4-5-20251001",
+		"claude-haiku-4-5",
+	}
+	for _, expected := range expectedAliases {
+		if !aliasSet[expected] {
+			t.Fatalf("expected default kiro alias %q to be present", expected)
+		}
+	}
+
+	// All should have fork=true
+	for _, a := range kiroAliases {
+		if !a.Fork {
+			t.Fatalf("expected all default kiro aliases to have fork=true, got fork=false for %q", a.Alias)
+		}
+	}
+
+	// Codex aliases should still be preserved
+	if len(cfg.OAuthModelAlias["codex"]) != 1 {
+		t.Fatal("expected codex aliases to be preserved")
+	}
+}
+
+func TestSanitizeOAuthModelAlias_DoesNotOverrideUserKiroAliases(t *testing.T) {
+	// When user has configured kiro aliases, defaults should NOT be injected
+	cfg := &Config{
+		OAuthModelAlias: map[string][]OAuthModelAlias{
+			"kiro": {
+				{Name: "kiro-claude-sonnet-4", Alias: "my-custom-sonnet", Fork: true},
+			},
+		},
+	}
+
+	cfg.SanitizeOAuthModelAlias()
+
+	kiroAliases := cfg.OAuthModelAlias["kiro"]
+	if len(kiroAliases) != 1 {
+		t.Fatalf("expected 1 user-configured kiro alias, got %d", len(kiroAliases))
+	}
+	if kiroAliases[0].Alias != "my-custom-sonnet" {
+		t.Fatalf("expected user alias to be preserved, got %q", kiroAliases[0].Alias)
+	}
+}
+
+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{}
+
+	cfg.SanitizeOAuthModelAlias()
+
+	kiroAliases := cfg.OAuthModelAlias["kiro"]
+	if len(kiroAliases) == 0 {
+		t.Fatal("expected default kiro aliases to be injected when OAuthModelAlias is nil")
+	}
+}

internal/config/sdk_config.go

@@ -20,9 +20,6 @@ type SDKConfig struct {
 	// APIKeys is a list of keys for authenticating clients to this proxy server.
 	APIKeys []string `yaml:"api-keys" json:"api-keys"`
 
-	// Access holds request authentication provider configuration.
-	Access AccessConfig `yaml:"auth,omitempty" json:"auth,omitempty"`
-
 	// Streaming configures server-side streaming behavior (keep-alives and safe bootstrap retries).
 	Streaming StreamingConfig `yaml:"streaming" json:"streaming"`
 
@@ -42,65 +39,3 @@ type StreamingConfig struct {
 	// <= 0 disables bootstrap retries. Default is 0.
 	BootstrapRetries int `yaml:"bootstrap-retries,omitempty" json:"bootstrap-retries,omitempty"`
 }
-
-// AccessConfig groups request authentication providers.
-type AccessConfig struct {
-	// Providers lists configured authentication providers.
-	Providers []AccessProvider `yaml:"providers,omitempty" json:"providers,omitempty"`
-}
-
-// AccessProvider describes a request authentication provider entry.
-type AccessProvider struct {
-	// Name is the instance identifier for the provider.
-	Name string `yaml:"name" json:"name"`
-
-	// Type selects the provider implementation registered via the SDK.
-	Type string `yaml:"type" json:"type"`
-
-	// SDK optionally names a third-party SDK module providing this provider.
-	SDK string `yaml:"sdk,omitempty" json:"sdk,omitempty"`
-
-	// APIKeys lists inline keys for providers that require them.
-	APIKeys []string `yaml:"api-keys,omitempty" json:"api-keys,omitempty"`
-
-	// Config passes provider-specific options to the implementation.
-	Config map[string]any `yaml:"config,omitempty" json:"config,omitempty"`
-}
-
-const (
-	// AccessProviderTypeConfigAPIKey is the built-in provider validating inline API keys.
-	AccessProviderTypeConfigAPIKey = "config-api-key"
-
-	// DefaultAccessProviderName is applied when no provider name is supplied.
-	DefaultAccessProviderName = "config-inline"
-)
-
-// ConfigAPIKeyProvider returns the first inline API key provider if present.
-func (c *SDKConfig) ConfigAPIKeyProvider() *AccessProvider {
-	if c == nil {
-		return nil
-	}
-	for i := range c.Access.Providers {
-		if c.Access.Providers[i].Type == AccessProviderTypeConfigAPIKey {
-			if c.Access.Providers[i].Name == "" {
-				c.Access.Providers[i].Name = DefaultAccessProviderName
-			}
-			return &c.Access.Providers[i]
-		}
-	}
-	return nil
-}
-
-// MakeInlineAPIKeyProvider constructs an inline API key provider configuration.
-// It returns nil when no keys are supplied.
-func MakeInlineAPIKeyProvider(keys []string) *AccessProvider {
-	if len(keys) == 0 {
-		return nil
-	}
-	provider := &AccessProvider{
-		Name:    DefaultAccessProviderName,
-		Type:    AccessProviderTypeConfigAPIKey,
-		APIKeys: append([]string(nil), keys...),
-	}
-	return provider
-}

internal/managementasset/updater.go

@@ -21,6 +21,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	sdkconfig "github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
 	log "github.com/sirupsen/logrus"
+	"golang.org/x/sync/singleflight"
 )
 
 const (
@@ -28,6 +29,7 @@ const (
 	defaultManagementFallbackURL = "https://cpamc.router-for.me/"
 	managementAssetName          = "management.html"
 	httpUserAgent                = "CLIProxyAPI-management-updater"
+	managementSyncMinInterval    = 30 * time.Second
 	updateCheckInterval          = 3 * time.Hour
 )
 
@@ -37,11 +39,10 @@ const ManagementFileName = managementAssetName
 var (
 	lastUpdateCheckMu   sync.Mutex
 	lastUpdateCheckTime time.Time
-
 	currentConfigPtr    atomic.Pointer[config.Config]
-	disableControlPanel atomic.Bool
 	schedulerOnce       sync.Once
 	schedulerConfigPath atomic.Value
+	sfGroup             singleflight.Group
 )
 
 // SetCurrentConfig stores the latest configuration snapshot for management asset decisions.
@@ -50,16 +51,7 @@ func SetCurrentConfig(cfg *config.Config) {
 		currentConfigPtr.Store(nil)
 		return
 	}
-
-	prevDisabled := disableControlPanel.Load()
 	currentConfigPtr.Store(cfg)
-	disableControlPanel.Store(cfg.RemoteManagement.DisableControlPanel)
-
-	if prevDisabled && !cfg.RemoteManagement.DisableControlPanel {
-		lastUpdateCheckMu.Lock()
-		lastUpdateCheckTime = time.Time{}
-		lastUpdateCheckMu.Unlock()
-	}
 }
 
 // StartAutoUpdater launches a background goroutine that periodically ensures the management asset is up to date.
@@ -92,7 +84,7 @@ func runAutoUpdater(ctx context.Context) {
 			log.Debug("management asset auto-updater skipped: config not yet available")
 			return
 		}
-		if disableControlPanel.Load() {
+		if cfg.RemoteManagement.DisableControlPanel {
 			log.Debug("management asset auto-updater skipped: control panel disabled")
 			return
 		}
@@ -181,103 +173,106 @@ func FilePath(configFilePath string) string {
 }
 
 // EnsureLatestManagementHTML checks the latest management.html asset and updates the local copy when needed.
-// The function is designed to run in a background goroutine and will never panic.
-// It enforces a 3-hour rate limit to avoid frequent checks on config/auth file changes.
-func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL string, panelRepository string) {
+// It coalesces concurrent sync attempts and returns whether the asset exists after the sync attempt.
+func EnsureLatestManagementHTML(ctx context.Context, staticDir string, proxyURL string, panelRepository string) bool {
 	if ctx == nil {
 		ctx = context.Background()
 	}
 
-	if disableControlPanel.Load() {
-		log.Debug("management asset sync skipped: control panel disabled by configuration")
-		return
-	}
-
 	staticDir = strings.TrimSpace(staticDir)
 	if staticDir == "" {
 		log.Debug("management asset sync skipped: empty static directory")
-		return
+		return false
 	}
-
 	localPath := filepath.Join(staticDir, managementAssetName)
-	localFileMissing := false
-	if _, errStat := os.Stat(localPath); errStat != nil {
-		if errors.Is(errStat, os.ErrNotExist) {
-			localFileMissing = true
-		} else {
-			log.WithError(errStat).Debug("failed to stat local management asset")
-		}
-	}
 
-	// Rate limiting: check only once every 3 hours
-	lastUpdateCheckMu.Lock()
-	now := time.Now()
-	timeSinceLastCheck := now.Sub(lastUpdateCheckTime)
-	if timeSinceLastCheck < updateCheckInterval {
+	_, _, _ = sfGroup.Do(localPath, func() (interface{}, error) {
+		lastUpdateCheckMu.Lock()
+		now := time.Now()
+		timeSinceLastAttempt := now.Sub(lastUpdateCheckTime)
+		if !lastUpdateCheckTime.IsZero() && timeSinceLastAttempt < managementSyncMinInterval {
+			lastUpdateCheckMu.Unlock()
+			log.Debugf(
+				"management asset sync skipped by throttle: last attempt %v ago (interval %v)",
+				timeSinceLastAttempt.Round(time.Second),
+				managementSyncMinInterval,
+			)
+			return nil, nil
+		}
+		lastUpdateCheckTime = now
 		lastUpdateCheckMu.Unlock()
-		log.Debugf("management asset update check skipped: last check was %v ago (interval: %v)", timeSinceLastCheck.Round(time.Second), updateCheckInterval)
-		return
-	}
-	lastUpdateCheckTime = now
-	lastUpdateCheckMu.Unlock()
 
-	if errMkdirAll := os.MkdirAll(staticDir, 0o755); errMkdirAll != nil {
-		log.WithError(errMkdirAll).Warn("failed to prepare static directory for management asset")
-		return
-	}
-
-	releaseURL := resolveReleaseURL(panelRepository)
-	client := newHTTPClient(proxyURL)
-
-	localHash, err := fileSHA256(localPath)
-	if err != nil {
-		if !errors.Is(err, os.ErrNotExist) {
-			log.WithError(err).Debug("failed to read local management asset hash")
-		}
-		localHash = ""
-	}
-
-	asset, remoteHash, err := fetchLatestAsset(ctx, client, releaseURL)
-	if err != nil {
-		if localFileMissing {
-			log.WithError(err).Warn("failed to fetch latest management release information, trying fallback page")
-			if ensureFallbackManagementHTML(ctx, client, localPath) {
-				return
+		localFileMissing := false
+		if _, errStat := os.Stat(localPath); errStat != nil {
+			if errors.Is(errStat, os.ErrNotExist) {
+				localFileMissing = true
+			} else {
+				log.WithError(errStat).Debug("failed to stat local management asset")
 			}
-			return
 		}
-		log.WithError(err).Warn("failed to fetch latest management release information")
-		return
-	}
 
-	if remoteHash != "" && localHash != "" && strings.EqualFold(remoteHash, localHash) {
-		log.Debug("management asset is already up to date")
-		return
-	}
+		if errMkdirAll := os.MkdirAll(staticDir, 0o755); errMkdirAll != nil {
+			log.WithError(errMkdirAll).Warn("failed to prepare static directory for management asset")
+			return nil, nil
+		}
 
-	data, downloadedHash, err := downloadAsset(ctx, client, asset.BrowserDownloadURL)
-	if err != nil {
-		if localFileMissing {
-			log.WithError(err).Warn("failed to download management asset, trying fallback page")
-			if ensureFallbackManagementHTML(ctx, client, localPath) {
-				return
+		releaseURL := resolveReleaseURL(panelRepository)
+		client := newHTTPClient(proxyURL)
+
+		localHash, err := fileSHA256(localPath)
+		if err != nil {
+			if !errors.Is(err, os.ErrNotExist) {
+				log.WithError(err).Debug("failed to read local management asset hash")
 			}
-			return
+			localHash = ""
 		}
-		log.WithError(err).Warn("failed to download management asset")
-		return
-	}
 
-	if remoteHash != "" && !strings.EqualFold(remoteHash, downloadedHash) {
-		log.Warnf("remote digest mismatch for management asset: expected %s got %s", remoteHash, downloadedHash)
-	}
+		asset, remoteHash, err := fetchLatestAsset(ctx, client, releaseURL)
+		if err != nil {
+			if localFileMissing {
+				log.WithError(err).Warn("failed to fetch latest management release information, trying fallback page")
+				if ensureFallbackManagementHTML(ctx, client, localPath) {
+					return nil, nil
+				}
+				return nil, nil
+			}
+			log.WithError(err).Warn("failed to fetch latest management release information")
+			return nil, nil
+		}
 
-	if err = atomicWriteFile(localPath, data); err != nil {
-		log.WithError(err).Warn("failed to update management asset on disk")
-		return
-	}
+		if remoteHash != "" && localHash != "" && strings.EqualFold(remoteHash, localHash) {
+			log.Debug("management asset is already up to date")
+			return nil, nil
+		}
 
-	log.Infof("management asset updated successfully (hash=%s)", downloadedHash)
+		data, downloadedHash, err := downloadAsset(ctx, client, asset.BrowserDownloadURL)
+		if err != nil {
+			if localFileMissing {
+				log.WithError(err).Warn("failed to download management asset, trying fallback page")
+				if ensureFallbackManagementHTML(ctx, client, localPath) {
+					return nil, nil
+				}
+				return nil, nil
+			}
+			log.WithError(err).Warn("failed to download management asset")
+			return nil, nil
+		}
+
+		if remoteHash != "" && !strings.EqualFold(remoteHash, downloadedHash) {
+			log.Warnf("remote digest mismatch for management asset: expected %s got %s", remoteHash, downloadedHash)
+		}
+
+		if err = atomicWriteFile(localPath, data); err != nil {
+			log.WithError(err).Warn("failed to update management asset on disk")
+			return nil, nil
+		}
+
+		log.Infof("management asset updated successfully (hash=%s)", downloadedHash)
+		return nil, nil
+	})
+
+	_, err := os.Stat(localPath)
+	return err == nil
 }
 
 func ensureFallbackManagementHTML(ctx context.Context, client *http.Client, localPath string) bool {

internal/registry/model_definitions.go

@@ -19,6 +19,7 @@ import (
 //   - codex
 //   - qwen
 //   - iflow
+//   - kimi
 //   - kiro
 //   - github-copilot
 //   - kiro
@@ -43,6 +44,8 @@ func GetStaticModelDefinitionsByChannel(channel string) []*ModelInfo {
 		return GetQwenModels()
 	case "iflow":
 		return GetIFlowModels()
+	case "kimi":
+		return GetKimiModels()
 	case "github-copilot":
 		return GetGitHubCopilotModels()
 	case "kiro":
@@ -93,6 +96,7 @@ func LookupStaticModelInfo(modelID string) *ModelInfo {
 		GetOpenAIModels(),
 		GetQwenModels(),
 		GetIFlowModels(),
+		GetKimiModels(),
 		GetGitHubCopilotModels(),
 		GetKiroModels(),
 		GetAmazonQModels(),
@@ -144,6 +148,7 @@ 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 +161,7 @@ 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 +174,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       200000,
 			MaxCompletionTokens: 32768,
 			SupportedEndpoints:  []string{"/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high"}},
 		},
 		{
 			ID:                  "gpt-5.1",
@@ -180,6 +187,7 @@ 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 +200,7 @@ 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 +213,7 @@ 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 +226,7 @@ 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 +239,7 @@ 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 +252,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       200000,
 			MaxCompletionTokens: 32768,
 			SupportedEndpoints:  []string{"/responses"},
+			Thinking:            &ThinkingSupport{Levels: []string{"none", "low", "medium", "high", "xhigh"}},
 		},
 		{
 			ID:                  "claude-haiku-4.5",
@@ -277,6 +290,18 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			MaxCompletionTokens: 64000,
 			SupportedEndpoints:  []string{"/chat/completions"},
 		},
+		{
+			ID:                  "claude-opus-4.6",
+			Object:              "model",
+			Created:             now,
+			OwnedBy:             "github-copilot",
+			Type:                "github-copilot",
+			DisplayName:         "Claude Opus 4.6",
+			Description:         "Anthropic Claude Opus 4.6 via GitHub Copilot",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			SupportedEndpoints:  []string{"/chat/completions"},
+		},
 		{
 			ID:                  "claude-sonnet-4",
 			Object:              "model",
@@ -436,6 +461,87 @@ 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",

internal/registry/model_definitions_static_data.go

@@ -742,6 +742,20 @@ 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 +788,19 @@ 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",
@@ -814,6 +841,7 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "kimi-k2-0905", DisplayName: "Kimi-K2-Instruct-0905", Description: "Moonshot Kimi K2 instruct 0905", Created: 1757030400},
 		{ID: "glm-4.6", DisplayName: "GLM-4.6", Description: "Zhipu GLM 4.6 general model", Created: 1759190400, Thinking: iFlowThinkingSupport},
 		{ID: "glm-4.7", DisplayName: "GLM-4.7", Description: "Zhipu GLM 4.7 general model", Created: 1766448000, Thinking: iFlowThinkingSupport},
+		{ID: "glm-5", DisplayName: "GLM-5", Description: "Zhipu GLM 5 general model", Created: 1770768000, Thinking: iFlowThinkingSupport},
 		{ID: "kimi-k2", DisplayName: "Kimi-K2", Description: "Moonshot Kimi K2 general model", Created: 1752192000},
 		{ID: "kimi-k2-thinking", DisplayName: "Kimi-K2-Thinking", Description: "Moonshot Kimi K2 thinking model", Created: 1762387200},
 		{ID: "deepseek-v3.2-chat", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2 Chat", Created: 1764576000},
@@ -828,6 +856,7 @@ 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},
 	}
@@ -866,9 +895,50 @@ func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 		"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: 128000},
+		"claude-opus-4-6-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-sonnet-4-5":          {MaxCompletionTokens: 64000},
 		"gpt-oss-120b-medium":        {},
 		"tab_flash_lite_preview":     {},
 	}
 }
+
+// GetKimiModels returns the standard Kimi (Moonshot AI) model definitions
+func GetKimiModels() []*ModelInfo {
+	return []*ModelInfo{
+		{
+			ID:                  "kimi-k2",
+			Object:              "model",
+			Created:             1752192000, // 2025-07-11
+			OwnedBy:             "moonshot",
+			Type:                "kimi",
+			DisplayName:         "Kimi K2",
+			Description:         "Kimi K2 - Moonshot AI's flagship coding model",
+			ContextLength:       131072,
+			MaxCompletionTokens: 32768,
+		},
+		{
+			ID:                  "kimi-k2-thinking",
+			Object:              "model",
+			Created:             1762387200, // 2025-11-06
+			OwnedBy:             "moonshot",
+			Type:                "kimi",
+			DisplayName:         "Kimi K2 Thinking",
+			Description:         "Kimi K2 Thinking - Extended reasoning model",
+			ContextLength:       131072,
+			MaxCompletionTokens: 32768,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
+		},
+		{
+			ID:                  "kimi-k2.5",
+			Object:              "model",
+			Created:             1769472000, // 2026-01-26
+			OwnedBy:             "moonshot",
+			Type:                "kimi",
+			DisplayName:         "Kimi K2.5",
+			Description:         "Kimi K2.5 - Latest Moonshot AI coding model with improved capabilities",
+			ContextLength:       131072,
+			MaxCompletionTokens: 32768,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
+		},
+	}
+}

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)
 	}
 }

internal/runtime/executor/antigravity_executor.go

@@ -1007,7 +1007,12 @@ 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 +1026,7 @@ 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 +1039,14 @@ 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 +1059,7 @@ 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 +1067,13 @@ 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
 		}
 

internal/runtime/executor/antigravity_executor_buildrequest_test.go

@@ -0,0 +1,159 @@
+package executor
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"testing"
+
+	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+)
+
+func TestAntigravityBuildRequest_SanitizesGeminiToolSchema(t *testing.T) {
+	body := buildRequestBodyFromPayload(t, "gemini-2.5-pro")
+
+	decl := extractFirstFunctionDeclaration(t, body)
+	if _, ok := decl["parametersJsonSchema"]; ok {
+		t.Fatalf("parametersJsonSchema should be renamed to parameters")
+	}
+
+	params, ok := decl["parameters"].(map[string]any)
+	if !ok {
+		t.Fatalf("parameters missing or invalid type")
+	}
+	assertSchemaSanitizedAndPropertyPreserved(t, params)
+}
+
+func TestAntigravityBuildRequest_SanitizesAntigravityToolSchema(t *testing.T) {
+	body := buildRequestBodyFromPayload(t, "claude-opus-4-6")
+
+	decl := extractFirstFunctionDeclaration(t, body)
+	params, ok := decl["parameters"].(map[string]any)
+	if !ok {
+		t.Fatalf("parameters missing or invalid type")
+	}
+	assertSchemaSanitizedAndPropertyPreserved(t, params)
+}
+
+func buildRequestBodyFromPayload(t *testing.T, modelName string) map[string]any {
+	t.Helper()
+
+	executor := &AntigravityExecutor{}
+	auth := &cliproxyauth.Auth{}
+	payload := []byte(`{
+		"request": {
+			"tools": [
+				{
+					"function_declarations": [
+						{
+							"name": "tool_1",
+							"parametersJsonSchema": {
+								"$schema": "http://json-schema.org/draft-07/schema#",
+								"$id": "root-schema",
+								"type": "object",
+								"properties": {
+									"$id": {"type": "string"},
+									"arg": {
+										"type": "object",
+										"prefill": "hello",
+										"properties": {
+											"mode": {
+												"type": "string",
+												"enum": ["a", "b"],
+												"enumTitles": ["A", "B"]
+											}
+										}
+									}
+								},
+								"patternProperties": {
+									"^x-": {"type": "string"}
+								}
+							}
+						}
+					]
+				}
+			]
+		}
+	}`)
+
+	req, err := executor.buildRequest(context.Background(), auth, "token", modelName, payload, false, "", "https://example.com")
+	if err != nil {
+		t.Fatalf("buildRequest error: %v", err)
+	}
+
+	raw, err := io.ReadAll(req.Body)
+	if err != nil {
+		t.Fatalf("read request body error: %v", err)
+	}
+
+	var body map[string]any
+	if err := json.Unmarshal(raw, &body); err != nil {
+		t.Fatalf("unmarshal request body error: %v, body=%s", err, string(raw))
+	}
+	return body
+}
+
+func extractFirstFunctionDeclaration(t *testing.T, body map[string]any) map[string]any {
+	t.Helper()
+
+	request, ok := body["request"].(map[string]any)
+	if !ok {
+		t.Fatalf("request missing or invalid type")
+	}
+	tools, ok := request["tools"].([]any)
+	if !ok || len(tools) == 0 {
+		t.Fatalf("tools missing or empty")
+	}
+	tool, ok := tools[0].(map[string]any)
+	if !ok {
+		t.Fatalf("first tool invalid type")
+	}
+	decls, ok := tool["function_declarations"].([]any)
+	if !ok || len(decls) == 0 {
+		t.Fatalf("function_declarations missing or empty")
+	}
+	decl, ok := decls[0].(map[string]any)
+	if !ok {
+		t.Fatalf("first function declaration invalid type")
+	}
+	return decl
+}
+
+func assertSchemaSanitizedAndPropertyPreserved(t *testing.T, params map[string]any) {
+	t.Helper()
+
+	if _, ok := params["$id"]; ok {
+		t.Fatalf("root $id should be removed from schema")
+	}
+	if _, ok := params["patternProperties"]; ok {
+		t.Fatalf("patternProperties should be removed from schema")
+	}
+
+	props, ok := params["properties"].(map[string]any)
+	if !ok {
+		t.Fatalf("properties missing or invalid type")
+	}
+	if _, ok := props["$id"]; !ok {
+		t.Fatalf("property named $id should be preserved")
+	}
+
+	arg, ok := props["arg"].(map[string]any)
+	if !ok {
+		t.Fatalf("arg property missing or invalid type")
+	}
+	if _, ok := arg["prefill"]; ok {
+		t.Fatalf("prefill should be removed from nested schema")
+	}
+
+	argProps, ok := arg["properties"].(map[string]any)
+	if !ok {
+		t.Fatalf("arg.properties missing or invalid type")
+	}
+	mode, ok := argProps["mode"].(map[string]any)
+	if !ok {
+		t.Fatalf("mode property missing or invalid type")
+	}
+	if _, ok := mode["enumTitles"]; ok {
+		t.Fatalf("enumTitles should be removed from nested schema")
+	}
+}

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:")

internal/runtime/executor/gemini_cli_executor.go

@@ -899,8 +899,7 @@ 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
 			}
 		}
 	}

internal/runtime/executor/github_copilot_executor.go

@@ -7,12 +7,14 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"strings"
 	"sync"
 	"time"
 
 	"github.com/google/uuid"
 	copilotauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/copilot"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
@@ -33,11 +35,11 @@ const (
 	maxScannerBufferSize = 20_971_520
 
 	// Copilot API header values.
-	copilotUserAgent     = "GithubCopilot/1.0"
-	copilotEditorVersion = "vscode/1.100.0"
-	copilotPluginVersion = "copilot/1.300.0"
+	copilotUserAgent     = "GitHubCopilotChat/0.35.0"
+	copilotEditorVersion = "vscode/1.107.0"
+	copilotPluginVersion = "copilot-chat/0.35.0"
 	copilotIntegrationID = "vscode-chat"
-	copilotOpenAIIntent  = "conversation-panel"
+	copilotOpenAIIntent  = "conversation-edits"
 )
 
 // GitHubCopilotExecutor handles requests to the GitHub Copilot API.
@@ -77,7 +79,7 @@ func (e *GitHubCopilotExecutor) PrepareRequest(req *http.Request, auth *cliproxy
 	if errToken != nil {
 		return errToken
 	}
-	e.applyHeaders(req, apiToken)
+	e.applyHeaders(req, apiToken, nil)
 	return nil
 }
 
@@ -108,7 +110,7 @@ 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")
@@ -120,6 +122,23 @@ func (e *GitHubCopilotExecutor) Execute(ctx context.Context, auth *cliproxyauth.
 	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, false)
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
 	body = e.normalizeModel(req.Model, body)
+	body = flattenAssistantContent(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)
@@ -133,7 +152,7 @@ func (e *GitHubCopilotExecutor) Execute(ctx context.Context, auth *cliproxyauth.
 	if err != nil {
 		return resp, err
 	}
-	e.applyHeaders(httpReq, apiToken)
+	e.applyHeaders(httpReq, apiToken, body)
 
 	// Add Copilot-Vision-Request header if the request contains vision content
 	if detectVisionContent(body) {
@@ -196,7 +215,12 @@ 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
@@ -213,7 +237,7 @@ 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")
@@ -225,6 +249,23 @@ func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliprox
 	originalTranslated := sdktranslator.TranslateRequest(from, to, req.Model, originalPayload, false)
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 	body = e.normalizeModel(req.Model, body)
+	body = flattenAssistantContent(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)
@@ -242,7 +283,7 @@ func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliprox
 	if err != nil {
 		return nil, err
 	}
-	e.applyHeaders(httpReq, apiToken)
+	e.applyHeaders(httpReq, apiToken, body)
 
 	// Add Copilot-Vision-Request header if the request contains vision content
 	if detectVisionContent(body) {
@@ -325,7 +366,12 @@ 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])}
 			}
@@ -414,7 +460,7 @@ func (e *GitHubCopilotExecutor) ensureAPIToken(ctx context.Context, auth *clipro
 }
 
 // applyHeaders sets the required headers for GitHub Copilot API requests.
-func (e *GitHubCopilotExecutor) applyHeaders(r *http.Request, apiToken string) {
+func (e *GitHubCopilotExecutor) applyHeaders(r *http.Request, apiToken string, body []byte) {
 	r.Header.Set("Content-Type", "application/json")
 	r.Header.Set("Authorization", "Bearer "+apiToken)
 	r.Header.Set("Accept", "application/json")
@@ -424,6 +470,20 @@ func (e *GitHubCopilotExecutor) applyHeaders(r *http.Request, apiToken string) {
 	r.Header.Set("Openai-Intent", copilotOpenAIIntent)
 	r.Header.Set("Copilot-Integration-Id", copilotIntegrationID)
 	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" {
+					initiator = "agent"
+				}
+			}
+		}
+	}
+	r.Header.Set("X-Initiator", initiator)
 }
 
 // detectVisionContent checks if the request body contains vision/image content.
@@ -454,14 +514,460 @@ func detectVisionContent(body []byte) bool {
 	return false
 }
 
-// normalizeModel is a no-op as GitHub Copilot accepts model names directly.
-// Model mapping should be done at the registry level if needed.
-func (e *GitHubCopilotExecutor) normalizeModel(_ string, body []byte) []byte {
+// normalizeModel strips the suffix (e.g. "(medium)") from the model name
+// before sending to GitHub Copilot, as the upstream API does not accept
+// suffixed model identifiers.
+func (e *GitHubCopilotExecutor) normalizeModel(model string, body []byte) []byte {
+	baseModel := thinking.ParseSuffix(model).ModelName
+	if baseModel != model {
+		body, _ = sjson.SetBytes(body, "model", baseModel)
+	}
 	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
+// to a joined string. GitHub Copilot requires assistant content as a string;
+// sending it as an array causes Claude models to re-answer all previous prompts.
+func flattenAssistantContent(body []byte) []byte {
+	messages := gjson.GetBytes(body, "messages")
+	if !messages.Exists() || !messages.IsArray() {
+		return body
+	}
+	result := body
+	for i, msg := range messages.Array() {
+		if msg.Get("role").String() != "assistant" {
+			continue
+		}
+		content := msg.Get("content")
+		if !content.Exists() || !content.IsArray() {
+			continue
+		}
+		var textParts []string
+		for _, part := range content.Array() {
+			if part.Get("type").String() == "text" {
+				if t := part.Get("text").String(); t != "" {
+					textParts = append(textParts, t)
+				}
+			}
+		}
+		joined := strings.Join(textParts, "")
+		path := fmt.Sprintf("messages.%d.content", i)
+		result, _ = sjson.SetBytes(result, path, joined)
+	}
+	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.Type == gjson.String {
+			return body
+		}
+		inputString := input.Raw
+		if input.Type != gjson.JSON {
+			inputString = input.String()
+		}
+		body, _ = sjson.SetBytes(body, "input", inputString)
+		return body
+	}
+
+	var parts []string
+	if system := gjson.GetBytes(body, "system"); system.Exists() {
+		if text := strings.TrimSpace(collectTextFromNode(system)); text != "" {
+			parts = append(parts, text)
+		}
+	}
+	if messages := gjson.GetBytes(body, "messages"); messages.Exists() && messages.IsArray() {
+		for _, msg := range messages.Array() {
+			if text := strings.TrimSpace(collectTextFromNode(msg.Get("content"))); text != "" {
+				parts = append(parts, text)
+			}
+		}
+	}
+	body, _ = sjson.SetBytes(body, "input", strings.Join(parts, "\n"))
+	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
+	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 "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()
+	out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+	out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
+	if hasToolUse {
+		out, _ = sjson.Set(out, "stop_reason", "tool_use")
+	} 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.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.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"
+			}
+			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", gjson.GetBytes(payload, "response.usage.input_tokens").Int())
+			messageDelta, _ = sjson.Set(messageDelta, "usage.output_tokens", gjson.GetBytes(payload, "response.usage.output_tokens").Int())
+			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).

internal/runtime/executor/github_copilot_executor_test.go

@@ -0,0 +1,242 @@
+package executor
+
+import (
+	"strings"
+	"testing"
+
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
+	"github.com/tidwall/gjson"
+)
+
+func TestGitHubCopilotNormalizeModel_StripsSuffix(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name      string
+		model     string
+		wantModel string
+	}{
+		{
+			name:      "suffix stripped",
+			model:     "claude-opus-4.6(medium)",
+			wantModel: "claude-opus-4.6",
+		},
+		{
+			name:      "no suffix unchanged",
+			model:     "claude-opus-4.6",
+			wantModel: "claude-opus-4.6",
+		},
+		{
+			name:      "different suffix stripped",
+			model:     "gpt-4o(high)",
+			wantModel: "gpt-4o",
+		},
+		{
+			name:      "numeric suffix stripped",
+			model:     "gemini-2.5-pro(8192)",
+			wantModel: "gemini-2.5-pro",
+		},
+	}
+
+	e := &GitHubCopilotExecutor{}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			body := []byte(`{"model":"` + tt.model + `","messages":[]}`)
+			got := e.normalizeModel(tt.model, body)
+
+			gotModel := gjson.GetBytes(got, "model").String()
+			if gotModel != tt.wantModel {
+				t.Fatalf("normalizeModel() model = %q, want %q", gotModel, tt.wantModel)
+			}
+		})
+	}
+}
+
+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.Type != gjson.String {
+		t.Fatalf("input type = %v, want string", in.Type)
+	}
+	if !strings.Contains(in.String(), "sys text") || !strings.Contains(in.String(), "user text") || !strings.Contains(in.String(), "assistant text") {
+		t.Fatalf("input = %q, want merged text", in.String())
+	}
+}
+
+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)
+	}
+}

internal/runtime/executor/iflow_executor.go

@@ -4,12 +4,16 @@ import (
 	"bufio"
 	"bytes"
 	"context"
+	"crypto/hmac"
+	"crypto/sha256"
+	"encoding/hex"
 	"fmt"
 	"io"
 	"net/http"
 	"strings"
 	"time"
 
+	"github.com/google/uuid"
 	iflowauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/iflow"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
@@ -453,6 +457,20 @@ func applyIFlowHeaders(r *http.Request, apiKey string, stream bool) {
 	r.Header.Set("Content-Type", "application/json")
 	r.Header.Set("Authorization", "Bearer "+apiKey)
 	r.Header.Set("User-Agent", iflowUserAgent)
+
+	// Generate session-id
+	sessionID := "session-" + generateUUID()
+	r.Header.Set("session-id", sessionID)
+
+	// Generate timestamp and signature
+	timestamp := time.Now().UnixMilli()
+	r.Header.Set("x-iflow-timestamp", fmt.Sprintf("%d", timestamp))
+
+	signature := createIFlowSignature(iflowUserAgent, sessionID, timestamp, apiKey)
+	if signature != "" {
+		r.Header.Set("x-iflow-signature", signature)
+	}
+
 	if stream {
 		r.Header.Set("Accept", "text/event-stream")
 	} else {
@@ -460,6 +478,23 @@ func applyIFlowHeaders(r *http.Request, apiKey string, stream bool) {
 	}
 }
 
+// createIFlowSignature generates HMAC-SHA256 signature for iFlow API requests.
+// The signature payload format is: userAgent:sessionId:timestamp
+func createIFlowSignature(userAgent, sessionID string, timestamp int64, apiKey string) string {
+	if apiKey == "" {
+		return ""
+	}
+	payload := fmt.Sprintf("%s:%s:%d", userAgent, sessionID, timestamp)
+	h := hmac.New(sha256.New, []byte(apiKey))
+	h.Write([]byte(payload))
+	return hex.EncodeToString(h.Sum(nil))
+}
+
+// generateUUID generates a random UUID v4 string.
+func generateUUID() string {
+	return uuid.New().String()
+}
+
 func iflowCreds(a *cliproxyauth.Auth) (apiKey, baseURL string) {
 	if a == nil {
 		return "", ""

internal/runtime/executor/kimi_executor.go

@@ -0,0 +1,618 @@
+package executor
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"time"
+
+	kimiauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/kimi"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
+	log "github.com/sirupsen/logrus"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// KimiExecutor is a stateless executor for Kimi API using OpenAI-compatible chat completions.
+type KimiExecutor struct {
+	ClaudeExecutor
+	cfg *config.Config
+}
+
+// NewKimiExecutor creates a new Kimi executor.
+func NewKimiExecutor(cfg *config.Config) *KimiExecutor { return &KimiExecutor{cfg: cfg} }
+
+// Identifier returns the executor identifier.
+func (e *KimiExecutor) Identifier() string { return "kimi" }
+
+// PrepareRequest injects Kimi credentials into the outgoing HTTP request.
+func (e *KimiExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
+	if req == nil {
+		return nil
+	}
+	token := kimiCreds(auth)
+	if strings.TrimSpace(token) != "" {
+		req.Header.Set("Authorization", "Bearer "+token)
+	}
+	return nil
+}
+
+// HttpRequest injects Kimi credentials into the request and executes it.
+func (e *KimiExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
+	if req == nil {
+		return nil, fmt.Errorf("kimi executor: request is nil")
+	}
+	if ctx == nil {
+		ctx = req.Context()
+	}
+	httpReq := req.WithContext(ctx)
+	if err := e.PrepareRequest(httpReq, auth); err != nil {
+		return nil, err
+	}
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	return httpClient.Do(httpReq)
+}
+
+// Execute performs a non-streaming chat completion request to Kimi.
+func (e *KimiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
+	from := opts.SourceFormat
+	if from.String() == "claude" {
+		auth.Attributes["base_url"] = kimiauth.KimiAPIBaseURL
+		return e.ClaudeExecutor.Execute(ctx, auth, req, opts)
+	}
+
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+
+	token := kimiCreds(auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
+
+	to := sdktranslator.FromString("openai")
+	originalPayloadSource := req.Payload
+	if len(opts.OriginalRequest) > 0 {
+		originalPayloadSource = opts.OriginalRequest
+	}
+	originalPayload := bytes.Clone(originalPayloadSource)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+	// Strip kimi- prefix for upstream API
+	upstreamModel := stripKimiPrefix(baseModel)
+	body, err = sjson.SetBytes(body, "model", upstreamModel)
+	if err != nil {
+		return resp, fmt.Errorf("kimi executor: failed to set model in payload: %w", err)
+	}
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "kimi", e.Identifier())
+	if err != nil {
+		return resp, err
+	}
+
+	requestedModel := payloadRequestedModel(opts, req.Model)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated, requestedModel)
+	body, err = normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		return resp, err
+	}
+
+	url := kimiauth.KimiAPIBaseURL + "/v1/chat/completions"
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	if err != nil {
+		return resp, err
+	}
+	applyKimiHeadersWithAuth(httpReq, token, false, auth)
+	var authID, authLabel, authType, authValue string
+	if auth != nil {
+		authID = auth.ID
+		authLabel = auth.Label
+		authType, authValue = auth.AccountInfo()
+	}
+	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
+		URL:       url,
+		Method:    http.MethodPost,
+		Headers:   httpReq.Header.Clone(),
+		Body:      body,
+		Provider:  e.Identifier(),
+		AuthID:    authID,
+		AuthLabel: authLabel,
+		AuthType:  authType,
+		AuthValue: authValue,
+	})
+
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpResp, err := httpClient.Do(httpReq)
+	if err != nil {
+		recordAPIResponseError(ctx, e.cfg, err)
+		return resp, err
+	}
+	defer func() {
+		if errClose := httpResp.Body.Close(); errClose != nil {
+			log.Errorf("kimi executor: close response body error: %v", errClose)
+		}
+	}()
+	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
+		b, _ := io.ReadAll(httpResp.Body)
+		appendAPIResponseChunk(ctx, e.cfg, b)
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
+		return resp, err
+	}
+	data, err := io.ReadAll(httpResp.Body)
+	if err != nil {
+		recordAPIResponseError(ctx, e.cfg, err)
+		return resp, err
+	}
+	appendAPIResponseChunk(ctx, e.cfg, data)
+	reporter.publish(ctx, parseOpenAIUsage(data))
+	var param any
+	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
+	// the original model name in the response for client compatibility.
+	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
+	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	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) {
+	from := opts.SourceFormat
+	if from.String() == "claude" {
+		auth.Attributes["base_url"] = kimiauth.KimiAPIBaseURL
+		return e.ClaudeExecutor.ExecuteStream(ctx, auth, req, opts)
+	}
+
+	baseModel := thinking.ParseSuffix(req.Model).ModelName
+	token := kimiCreds(auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
+	defer reporter.trackFailure(ctx, &err)
+
+	to := sdktranslator.FromString("openai")
+	originalPayloadSource := req.Payload
+	if len(opts.OriginalRequest) > 0 {
+		originalPayloadSource = opts.OriginalRequest
+	}
+	originalPayload := bytes.Clone(originalPayloadSource)
+	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+
+	// Strip kimi- prefix for upstream API
+	upstreamModel := stripKimiPrefix(baseModel)
+	body, err = sjson.SetBytes(body, "model", upstreamModel)
+	if err != nil {
+		return nil, fmt.Errorf("kimi executor: failed to set model in payload: %w", err)
+	}
+
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "kimi", e.Identifier())
+	if err != nil {
+		return nil, err
+	}
+
+	body, err = sjson.SetBytes(body, "stream_options.include_usage", true)
+	if err != nil {
+		return nil, fmt.Errorf("kimi executor: failed to set stream_options in payload: %w", err)
+	}
+	requestedModel := payloadRequestedModel(opts, req.Model)
+	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated, requestedModel)
+	body, err = normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		return nil, err
+	}
+
+	url := kimiauth.KimiAPIBaseURL + "/v1/chat/completions"
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	if err != nil {
+		return nil, err
+	}
+	applyKimiHeadersWithAuth(httpReq, token, true, auth)
+	var authID, authLabel, authType, authValue string
+	if auth != nil {
+		authID = auth.ID
+		authLabel = auth.Label
+		authType, authValue = auth.AccountInfo()
+	}
+	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
+		URL:       url,
+		Method:    http.MethodPost,
+		Headers:   httpReq.Header.Clone(),
+		Body:      body,
+		Provider:  e.Identifier(),
+		AuthID:    authID,
+		AuthLabel: authLabel,
+		AuthType:  authType,
+		AuthValue: authValue,
+	})
+
+	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpResp, err := httpClient.Do(httpReq)
+	if err != nil {
+		recordAPIResponseError(ctx, e.cfg, err)
+		return 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)
+		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		if errClose := httpResp.Body.Close(); errClose != nil {
+			log.Errorf("kimi executor: close response body error: %v", errClose)
+		}
+		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
+		return nil, err
+	}
+	out := make(chan cliproxyexecutor.StreamChunk)
+	stream = out
+	go func() {
+		defer close(out)
+		defer func() {
+			if errClose := httpResp.Body.Close(); errClose != nil {
+				log.Errorf("kimi executor: close response body error: %v", errClose)
+			}
+		}()
+		scanner := bufio.NewScanner(httpResp.Body)
+		scanner.Buffer(nil, 1_048_576) // 1MB
+		var param any
+		for scanner.Scan() {
+			line := scanner.Bytes()
+			appendAPIResponseChunk(ctx, e.cfg, line)
+			if detail, ok := parseOpenAIStreamUsage(line); ok {
+				reporter.publish(ctx, detail)
+			}
+			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, body, bytes.Clone(line), &param)
+			for i := range chunks {
+				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
+			}
+		}
+		doneChunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, body, []byte("[DONE]"), &param)
+		for i := range doneChunks {
+			out <- cliproxyexecutor.StreamChunk{Payload: []byte(doneChunks[i])}
+		}
+		if errScan := scanner.Err(); errScan != nil {
+			recordAPIResponseError(ctx, e.cfg, errScan)
+			reporter.publishFailure(ctx)
+			out <- cliproxyexecutor.StreamChunk{Err: errScan}
+		}
+	}()
+	return stream, nil
+}
+
+// CountTokens estimates token count for Kimi requests.
+func (e *KimiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	auth.Attributes["base_url"] = kimiauth.KimiAPIBaseURL
+	return e.ClaudeExecutor.CountTokens(ctx, auth, req, opts)
+}
+
+func normalizeKimiToolMessageLinks(body []byte) ([]byte, error) {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		return body, nil
+	}
+
+	messages := gjson.GetBytes(body, "messages")
+	if !messages.Exists() || !messages.IsArray() {
+		return body, nil
+	}
+
+	out := body
+	pending := make([]string, 0)
+	patched := 0
+	patchedReasoning := 0
+	ambiguous := 0
+	latestReasoning := ""
+	hasLatestReasoning := false
+
+	removePending := func(id string) {
+		for idx := range pending {
+			if pending[idx] != id {
+				continue
+			}
+			pending = append(pending[:idx], pending[idx+1:]...)
+			return
+		}
+	}
+
+	msgs := messages.Array()
+	for msgIdx := range msgs {
+		msg := msgs[msgIdx]
+		role := strings.TrimSpace(msg.Get("role").String())
+		switch role {
+		case "assistant":
+			reasoning := msg.Get("reasoning_content")
+			if reasoning.Exists() {
+				reasoningText := reasoning.String()
+				if strings.TrimSpace(reasoningText) != "" {
+					latestReasoning = reasoningText
+					hasLatestReasoning = true
+				}
+			}
+
+			toolCalls := msg.Get("tool_calls")
+			if !toolCalls.Exists() || !toolCalls.IsArray() || len(toolCalls.Array()) == 0 {
+				continue
+			}
+
+			if !reasoning.Exists() || strings.TrimSpace(reasoning.String()) == "" {
+				reasoningText := fallbackAssistantReasoning(msg, hasLatestReasoning, latestReasoning)
+				path := fmt.Sprintf("messages.%d.reasoning_content", msgIdx)
+				next, err := sjson.SetBytes(out, path, reasoningText)
+				if err != nil {
+					return body, fmt.Errorf("kimi executor: failed to set assistant reasoning_content: %w", err)
+				}
+				out = next
+				patchedReasoning++
+			}
+
+			for _, tc := range toolCalls.Array() {
+				id := strings.TrimSpace(tc.Get("id").String())
+				if id == "" {
+					continue
+				}
+				pending = append(pending, id)
+			}
+		case "tool":
+			toolCallID := strings.TrimSpace(msg.Get("tool_call_id").String())
+			if toolCallID == "" {
+				toolCallID = strings.TrimSpace(msg.Get("call_id").String())
+				if toolCallID != "" {
+					path := fmt.Sprintf("messages.%d.tool_call_id", msgIdx)
+					next, err := sjson.SetBytes(out, path, toolCallID)
+					if err != nil {
+						return body, fmt.Errorf("kimi executor: failed to set tool_call_id from call_id: %w", err)
+					}
+					out = next
+					patched++
+				}
+			}
+			if toolCallID == "" {
+				if len(pending) == 1 {
+					toolCallID = pending[0]
+					path := fmt.Sprintf("messages.%d.tool_call_id", msgIdx)
+					next, err := sjson.SetBytes(out, path, toolCallID)
+					if err != nil {
+						return body, fmt.Errorf("kimi executor: failed to infer tool_call_id: %w", err)
+					}
+					out = next
+					patched++
+				} else if len(pending) > 1 {
+					ambiguous++
+				}
+			}
+			if toolCallID != "" {
+				removePending(toolCallID)
+			}
+		}
+	}
+
+	if patched > 0 || patchedReasoning > 0 {
+		log.WithFields(log.Fields{
+			"patched_tool_messages":      patched,
+			"patched_reasoning_messages": patchedReasoning,
+		}).Debug("kimi executor: normalized tool message fields")
+	}
+	if ambiguous > 0 {
+		log.WithFields(log.Fields{
+			"ambiguous_tool_messages": ambiguous,
+			"pending_tool_calls":      len(pending),
+		}).Warn("kimi executor: tool messages missing tool_call_id with ambiguous candidates")
+	}
+
+	return out, nil
+}
+
+func fallbackAssistantReasoning(msg gjson.Result, hasLatest bool, latest string) string {
+	if hasLatest && strings.TrimSpace(latest) != "" {
+		return latest
+	}
+
+	content := msg.Get("content")
+	if content.Type == gjson.String {
+		if text := strings.TrimSpace(content.String()); text != "" {
+			return text
+		}
+	}
+	if content.IsArray() {
+		parts := make([]string, 0, len(content.Array()))
+		for _, item := range content.Array() {
+			text := strings.TrimSpace(item.Get("text").String())
+			if text == "" {
+				continue
+			}
+			parts = append(parts, text)
+		}
+		if len(parts) > 0 {
+			return strings.Join(parts, "\n")
+		}
+	}
+
+	return "[reasoning unavailable]"
+}
+
+// Refresh refreshes the Kimi token using the refresh token.
+func (e *KimiExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
+	log.Debugf("kimi executor: refresh called")
+	if auth == nil {
+		return nil, fmt.Errorf("kimi executor: auth is nil")
+	}
+	// Expect refresh_token in metadata for OAuth-based accounts
+	var refreshToken string
+	if auth.Metadata != nil {
+		if v, ok := auth.Metadata["refresh_token"].(string); ok && strings.TrimSpace(v) != "" {
+			refreshToken = v
+		}
+	}
+	if strings.TrimSpace(refreshToken) == "" {
+		// Nothing to refresh
+		return auth, nil
+	}
+
+	client := kimiauth.NewDeviceFlowClientWithDeviceID(e.cfg, resolveKimiDeviceID(auth))
+	td, err := client.RefreshToken(ctx, refreshToken)
+	if err != nil {
+		return nil, err
+	}
+	if auth.Metadata == nil {
+		auth.Metadata = make(map[string]any)
+	}
+	auth.Metadata["access_token"] = td.AccessToken
+	if td.RefreshToken != "" {
+		auth.Metadata["refresh_token"] = td.RefreshToken
+	}
+	if td.ExpiresAt > 0 {
+		exp := time.Unix(td.ExpiresAt, 0).UTC().Format(time.RFC3339)
+		auth.Metadata["expired"] = exp
+	}
+	auth.Metadata["type"] = "kimi"
+	now := time.Now().Format(time.RFC3339)
+	auth.Metadata["last_refresh"] = now
+	return auth, nil
+}
+
+// applyKimiHeaders sets required headers for Kimi API requests.
+// Headers match kimi-cli client for compatibility.
+func applyKimiHeaders(r *http.Request, token string, stream bool) {
+	r.Header.Set("Content-Type", "application/json")
+	r.Header.Set("Authorization", "Bearer "+token)
+	// Match kimi-cli headers exactly
+	r.Header.Set("User-Agent", "KimiCLI/1.10.6")
+	r.Header.Set("X-Msh-Platform", "kimi_cli")
+	r.Header.Set("X-Msh-Version", "1.10.6")
+	r.Header.Set("X-Msh-Device-Name", getKimiHostname())
+	r.Header.Set("X-Msh-Device-Model", getKimiDeviceModel())
+	r.Header.Set("X-Msh-Device-Id", getKimiDeviceID())
+	if stream {
+		r.Header.Set("Accept", "text/event-stream")
+		return
+	}
+	r.Header.Set("Accept", "application/json")
+}
+
+func resolveKimiDeviceIDFromAuth(auth *cliproxyauth.Auth) string {
+	if auth == nil || auth.Metadata == nil {
+		return ""
+	}
+
+	deviceIDRaw, ok := auth.Metadata["device_id"]
+	if !ok {
+		return ""
+	}
+
+	deviceID, ok := deviceIDRaw.(string)
+	if !ok {
+		return ""
+	}
+
+	return strings.TrimSpace(deviceID)
+}
+
+func resolveKimiDeviceIDFromStorage(auth *cliproxyauth.Auth) string {
+	if auth == nil {
+		return ""
+	}
+
+	storage, ok := auth.Storage.(*kimiauth.KimiTokenStorage)
+	if !ok || storage == nil {
+		return ""
+	}
+
+	return strings.TrimSpace(storage.DeviceID)
+}
+
+func resolveKimiDeviceID(auth *cliproxyauth.Auth) string {
+	deviceID := resolveKimiDeviceIDFromAuth(auth)
+	if deviceID != "" {
+		return deviceID
+	}
+	return resolveKimiDeviceIDFromStorage(auth)
+}
+
+func applyKimiHeadersWithAuth(r *http.Request, token string, stream bool, auth *cliproxyauth.Auth) {
+	applyKimiHeaders(r, token, stream)
+
+	if deviceID := resolveKimiDeviceID(auth); deviceID != "" {
+		r.Header.Set("X-Msh-Device-Id", deviceID)
+	}
+}
+
+// getKimiHostname returns the machine hostname.
+func getKimiHostname() string {
+	hostname, err := os.Hostname()
+	if err != nil {
+		return "unknown"
+	}
+	return hostname
+}
+
+// getKimiDeviceModel returns a device model string matching kimi-cli format.
+func getKimiDeviceModel() string {
+	return fmt.Sprintf("%s %s", runtime.GOOS, runtime.GOARCH)
+}
+
+// getKimiDeviceID returns a stable device ID, matching kimi-cli storage location.
+func getKimiDeviceID() string {
+	homeDir, err := os.UserHomeDir()
+	if err != nil {
+		return "cli-proxy-api-device"
+	}
+	// Check kimi-cli's device_id location first (platform-specific)
+	var kimiShareDir string
+	switch runtime.GOOS {
+	case "darwin":
+		kimiShareDir = filepath.Join(homeDir, "Library", "Application Support", "kimi")
+	case "windows":
+		appData := os.Getenv("APPDATA")
+		if appData == "" {
+			appData = filepath.Join(homeDir, "AppData", "Roaming")
+		}
+		kimiShareDir = filepath.Join(appData, "kimi")
+	default: // linux and other unix-like
+		kimiShareDir = filepath.Join(homeDir, ".local", "share", "kimi")
+	}
+	deviceIDPath := filepath.Join(kimiShareDir, "device_id")
+	if data, err := os.ReadFile(deviceIDPath); err == nil {
+		return strings.TrimSpace(string(data))
+	}
+	return "cli-proxy-api-device"
+}
+
+// kimiCreds extracts the access token from auth.
+func kimiCreds(a *cliproxyauth.Auth) (token string) {
+	if a == nil {
+		return ""
+	}
+	// Check metadata first (OAuth flow stores tokens here)
+	if a.Metadata != nil {
+		if v, ok := a.Metadata["access_token"].(string); ok && strings.TrimSpace(v) != "" {
+			return v
+		}
+	}
+	// Fallback to attributes (API key style)
+	if a.Attributes != nil {
+		if v := a.Attributes["access_token"]; v != "" {
+			return v
+		}
+		if v := a.Attributes["api_key"]; v != "" {
+			return v
+		}
+	}
+	return ""
+}
+
+// stripKimiPrefix removes the "kimi-" prefix from model names for the upstream API.
+func stripKimiPrefix(model string) string {
+	model = strings.TrimSpace(model)
+	if strings.HasPrefix(strings.ToLower(model), "kimi-") {
+		return model[5:]
+	}
+	return model
+}

internal/runtime/executor/kimi_executor_test.go

@@ -0,0 +1,205 @@
+package executor
+
+import (
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestNormalizeKimiToolMessageLinks_UsesCallIDFallback(t *testing.T) {
+	body := []byte(`{
+		"messages":[
+			{"role":"assistant","tool_calls":[{"id":"list_directory:1","type":"function","function":{"name":"list_directory","arguments":"{}"}}]},
+			{"role":"tool","call_id":"list_directory:1","content":"[]"}
+		]
+	}`)
+
+	out, err := normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		t.Fatalf("normalizeKimiToolMessageLinks() error = %v", err)
+	}
+
+	got := gjson.GetBytes(out, "messages.1.tool_call_id").String()
+	if got != "list_directory:1" {
+		t.Fatalf("messages.1.tool_call_id = %q, want %q", got, "list_directory:1")
+	}
+}
+
+func TestNormalizeKimiToolMessageLinks_InferSinglePendingID(t *testing.T) {
+	body := []byte(`{
+		"messages":[
+			{"role":"assistant","tool_calls":[{"id":"call_123","type":"function","function":{"name":"read_file","arguments":"{}"}}]},
+			{"role":"tool","content":"file-content"}
+		]
+	}`)
+
+	out, err := normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		t.Fatalf("normalizeKimiToolMessageLinks() error = %v", err)
+	}
+
+	got := gjson.GetBytes(out, "messages.1.tool_call_id").String()
+	if got != "call_123" {
+		t.Fatalf("messages.1.tool_call_id = %q, want %q", got, "call_123")
+	}
+}
+
+func TestNormalizeKimiToolMessageLinks_AmbiguousMissingIDIsNotInferred(t *testing.T) {
+	body := []byte(`{
+		"messages":[
+			{"role":"assistant","tool_calls":[
+				{"id":"call_1","type":"function","function":{"name":"list_directory","arguments":"{}"}},
+				{"id":"call_2","type":"function","function":{"name":"read_file","arguments":"{}"}}
+			]},
+			{"role":"tool","content":"result-without-id"}
+		]
+	}`)
+
+	out, err := normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		t.Fatalf("normalizeKimiToolMessageLinks() error = %v", err)
+	}
+
+	if gjson.GetBytes(out, "messages.1.tool_call_id").Exists() {
+		t.Fatalf("messages.1.tool_call_id should be absent for ambiguous case, got %q", gjson.GetBytes(out, "messages.1.tool_call_id").String())
+	}
+}
+
+func TestNormalizeKimiToolMessageLinks_PreservesExistingToolCallID(t *testing.T) {
+	body := []byte(`{
+		"messages":[
+			{"role":"assistant","tool_calls":[{"id":"call_1","type":"function","function":{"name":"list_directory","arguments":"{}"}}]},
+			{"role":"tool","tool_call_id":"call_1","call_id":"different-id","content":"result"}
+		]
+	}`)
+
+	out, err := normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		t.Fatalf("normalizeKimiToolMessageLinks() error = %v", err)
+	}
+
+	got := gjson.GetBytes(out, "messages.1.tool_call_id").String()
+	if got != "call_1" {
+		t.Fatalf("messages.1.tool_call_id = %q, want %q", got, "call_1")
+	}
+}
+
+func TestNormalizeKimiToolMessageLinks_InheritsPreviousReasoningForAssistantToolCalls(t *testing.T) {
+	body := []byte(`{
+		"messages":[
+			{"role":"assistant","content":"plan","reasoning_content":"previous reasoning"},
+			{"role":"assistant","tool_calls":[{"id":"call_1","type":"function","function":{"name":"list_directory","arguments":"{}"}}]}
+		]
+	}`)
+
+	out, err := normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		t.Fatalf("normalizeKimiToolMessageLinks() error = %v", err)
+	}
+
+	got := gjson.GetBytes(out, "messages.1.reasoning_content").String()
+	if got != "previous reasoning" {
+		t.Fatalf("messages.1.reasoning_content = %q, want %q", got, "previous reasoning")
+	}
+}
+
+func TestNormalizeKimiToolMessageLinks_InsertsFallbackReasoningWhenMissing(t *testing.T) {
+	body := []byte(`{
+		"messages":[
+			{"role":"assistant","tool_calls":[{"id":"call_1","type":"function","function":{"name":"list_directory","arguments":"{}"}}]}
+		]
+	}`)
+
+	out, err := normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		t.Fatalf("normalizeKimiToolMessageLinks() error = %v", err)
+	}
+
+	reasoning := gjson.GetBytes(out, "messages.0.reasoning_content")
+	if !reasoning.Exists() {
+		t.Fatalf("messages.0.reasoning_content should exist")
+	}
+	if reasoning.String() != "[reasoning unavailable]" {
+		t.Fatalf("messages.0.reasoning_content = %q, want %q", reasoning.String(), "[reasoning unavailable]")
+	}
+}
+
+func TestNormalizeKimiToolMessageLinks_UsesContentAsReasoningFallback(t *testing.T) {
+	body := []byte(`{
+		"messages":[
+			{"role":"assistant","content":[{"type":"text","text":"first line"},{"type":"text","text":"second line"}],"tool_calls":[{"id":"call_1","type":"function","function":{"name":"list_directory","arguments":"{}"}}]}
+		]
+	}`)
+
+	out, err := normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		t.Fatalf("normalizeKimiToolMessageLinks() error = %v", err)
+	}
+
+	got := gjson.GetBytes(out, "messages.0.reasoning_content").String()
+	if got != "first line\nsecond line" {
+		t.Fatalf("messages.0.reasoning_content = %q, want %q", got, "first line\nsecond line")
+	}
+}
+
+func TestNormalizeKimiToolMessageLinks_ReplacesEmptyReasoningContent(t *testing.T) {
+	body := []byte(`{
+		"messages":[
+			{"role":"assistant","content":"assistant summary","tool_calls":[{"id":"call_1","type":"function","function":{"name":"list_directory","arguments":"{}"}}],"reasoning_content":""}
+		]
+	}`)
+
+	out, err := normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		t.Fatalf("normalizeKimiToolMessageLinks() error = %v", err)
+	}
+
+	got := gjson.GetBytes(out, "messages.0.reasoning_content").String()
+	if got != "assistant summary" {
+		t.Fatalf("messages.0.reasoning_content = %q, want %q", got, "assistant summary")
+	}
+}
+
+func TestNormalizeKimiToolMessageLinks_PreservesExistingAssistantReasoning(t *testing.T) {
+	body := []byte(`{
+		"messages":[
+			{"role":"assistant","tool_calls":[{"id":"call_1","type":"function","function":{"name":"list_directory","arguments":"{}"}}],"reasoning_content":"keep me"}
+		]
+	}`)
+
+	out, err := normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		t.Fatalf("normalizeKimiToolMessageLinks() error = %v", err)
+	}
+
+	got := gjson.GetBytes(out, "messages.0.reasoning_content").String()
+	if got != "keep me" {
+		t.Fatalf("messages.0.reasoning_content = %q, want %q", got, "keep me")
+	}
+}
+
+func TestNormalizeKimiToolMessageLinks_RepairsIDsAndReasoningTogether(t *testing.T) {
+	body := []byte(`{
+		"messages":[
+			{"role":"assistant","tool_calls":[{"id":"call_1","type":"function","function":{"name":"list_directory","arguments":"{}"}}],"reasoning_content":"r1"},
+			{"role":"tool","call_id":"call_1","content":"[]"},
+			{"role":"assistant","tool_calls":[{"id":"call_2","type":"function","function":{"name":"read_file","arguments":"{}"}}]},
+			{"role":"tool","call_id":"call_2","content":"file"}
+		]
+	}`)
+
+	out, err := normalizeKimiToolMessageLinks(body)
+	if err != nil {
+		t.Fatalf("normalizeKimiToolMessageLinks() error = %v", err)
+	}
+
+	if got := gjson.GetBytes(out, "messages.1.tool_call_id").String(); got != "call_1" {
+		t.Fatalf("messages.1.tool_call_id = %q, want %q", got, "call_1")
+	}
+	if got := gjson.GetBytes(out, "messages.3.tool_call_id").String(); got != "call_2" {
+		t.Fatalf("messages.3.tool_call_id = %q, want %q", got, "call_2")
+	}
+	if got := gjson.GetBytes(out, "messages.2.reasoning_content").String(); got != "r1" {
+		t.Fatalf("messages.2.reasoning_content = %q, want %q", got, "r1")
+	}
+}

internal/runtime/executor/kiro_executor.go

@@ -16,6 +16,7 @@ import (
 	"path/filepath"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"syscall"
 	"time"
 
@@ -385,6 +386,35 @@ func buildKiroEndpointConfigs(region string) []kiroEndpointConfig {
 	}
 }
 
+// resolveKiroAPIRegion determines the AWS region for Kiro API calls.
+// Region priority:
+// 1. auth.Metadata["api_region"] - explicit API region override
+// 2. ProfileARN region - extracted from arn:aws:service:REGION:account:resource
+// 3. kiroDefaultRegion (us-east-1) - fallback
+// Note: OIDC "region" is NOT used - it's for token refresh, not API calls
+func resolveKiroAPIRegion(auth *cliproxyauth.Auth) string {
+	if auth == nil || auth.Metadata == nil {
+		return kiroDefaultRegion
+	}
+	// Priority 1: Explicit api_region override
+	if r, ok := auth.Metadata["api_region"].(string); ok && r != "" {
+		log.Debugf("kiro: using region %s (source: api_region)", r)
+		return r
+	}
+	// Priority 2: Extract from ProfileARN
+	if profileArn, ok := auth.Metadata["profile_arn"].(string); ok && profileArn != "" {
+		if arnRegion := extractRegionFromProfileARN(profileArn); arnRegion != "" {
+			log.Debugf("kiro: using region %s (source: profile_arn)", arnRegion)
+			return arnRegion
+		}
+	}
+	// Note: OIDC "region" field is NOT used for API endpoint
+	// Kiro API only exists in us-east-1, while OIDC region can vary (e.g., ap-northeast-2)
+	// Using OIDC region for API calls causes DNS failures
+	log.Debugf("kiro: using region %s (source: default)", kiroDefaultRegion)
+	return kiroDefaultRegion
+}
+
 // kiroEndpointConfigs is kept for backward compatibility with default us-east-1 region.
 // Prefer using buildKiroEndpointConfigs(region) for dynamic region support.
 var kiroEndpointConfigs = buildKiroEndpointConfigs(kiroDefaultRegion)
@@ -403,30 +433,8 @@ func getKiroEndpointConfigs(auth *cliproxyauth.Auth) []kiroEndpointConfig {
 		return kiroEndpointConfigs
 	}
 
-	// Determine API region with priority: api_region > profile_arn > region > default
-	region := kiroDefaultRegion
-	regionSource := "default"
-
-	if auth.Metadata != nil {
-		// Priority 1: Explicit api_region override
-		if r, ok := auth.Metadata["api_region"].(string); ok && r != "" {
-			region = r
-			regionSource = "api_region"
-		} else {
-			// Priority 2: Extract from ProfileARN
-			if profileArn, ok := auth.Metadata["profile_arn"].(string); ok && profileArn != "" {
-				if arnRegion := extractRegionFromProfileARN(profileArn); arnRegion != "" {
-					region = arnRegion
-					regionSource = "profile_arn"
-				}
-			}
-			// Note: OIDC "region" field is NOT used for API endpoint
-			// Kiro API only exists in us-east-1, while OIDC region can vary (e.g., ap-northeast-2)
-			// Using OIDC region for API calls causes DNS failures
-		}
-	}
-
-	log.Debugf("kiro: using region %s (source: %s)", region, regionSource)
+	// Determine API region using shared resolution logic
+	region := resolveKiroAPIRegion(auth)
 
 	// Build endpoint configs for the specified region
 	endpointConfigs := buildKiroEndpointConfigs(region)
@@ -519,8 +527,12 @@ func buildKiroPayloadForFormat(body []byte, modelID, profileArn, origin string,
 	case "openai":
 		log.Debugf("kiro: using OpenAI payload builder for source format: %s", sourceFormat.String())
 		return kiroopenai.BuildKiroPayloadFromOpenAI(body, modelID, profileArn, origin, isAgentic, isChatOnly, headers, nil)
+	case "kiro":
+		// Body is already in Kiro format — pass through directly
+		log.Debugf("kiro: body already in Kiro format, passing through directly")
+		return body, false
 	default:
-		// Default to Claude format (also handles "claude", "kiro", etc.)
+		// Default to Claude format
 		log.Debugf("kiro: using Claude payload builder for source format: %s", sourceFormat.String())
 		return kiroclaude.BuildKiroPayload(body, modelID, profileArn, origin, isAgentic, isChatOnly, headers, nil)
 	}
@@ -636,10 +648,7 @@ func (e *KiroExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	rateLimiter.WaitForToken(tokenKey)
 	log.Debugf("kiro: rate limiter cleared for token %s", tokenKey)
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
-
-	// Check if token is expired before making request
+	// Check if token is expired before making request (covers both normal and web_search paths)
 	if e.isTokenExpired(accessToken) {
 		log.Infof("kiro: access token expired, attempting recovery")
 
@@ -668,6 +677,16 @@ func (e *KiroExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 		}
 	}
 
+	// Check for pure web_search request
+	// Route to MCP endpoint instead of normal Kiro API
+	if kiroclaude.HasWebSearchTool(req.Payload) {
+		log.Infof("kiro: detected pure web_search request (non-stream), routing to MCP endpoint")
+		return e.handleWebSearch(ctx, auth, req, opts, accessToken, profileArn)
+	}
+
+	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	defer reporter.trackFailure(ctx, &err)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("kiro")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
@@ -1014,8 +1033,9 @@ func (e *KiroExecutor) executeWithRetry(ctx context.Context, auth *cliproxyauth.
 
 			// Build response in Claude format for Kiro translator
 			// stopReason is extracted from upstream response by parseEventStream
-			kiroResponse := kiroclaude.BuildClaudeResponse(content, toolUses, req.Model, usageInfo, stopReason)
-			out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, kiroResponse, nil)
+			requestedModel := payloadRequestedModel(opts, req.Model)
+			kiroResponse := kiroclaude.BuildClaudeResponse(content, toolUses, requestedModel, usageInfo, stopReason)
+			out := sdktranslator.TranslateNonStream(ctx, to, from, requestedModel, bytes.Clone(opts.OriginalRequest), body, kiroResponse, nil)
 			resp = cliproxyexecutor.Response{Payload: []byte(out)}
 			return resp, nil
 		}
@@ -1057,10 +1077,7 @@ func (e *KiroExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	rateLimiter.WaitForToken(tokenKey)
 	log.Debugf("kiro: stream rate limiter cleared for token %s", tokenKey)
 
-	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
-	defer reporter.trackFailure(ctx, &err)
-
-	// Check if token is expired before making request
+	// Check if token is expired before making request (covers both normal and web_search paths)
 	if e.isTokenExpired(accessToken) {
 		log.Infof("kiro: access token expired, attempting recovery before stream request")
 
@@ -1089,6 +1106,16 @@ func (e *KiroExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		}
 	}
 
+	// Check for pure web_search request
+	// Route to MCP endpoint instead of normal Kiro API
+	if kiroclaude.HasWebSearchTool(req.Payload) {
+		log.Infof("kiro: detected pure web_search request, routing to MCP endpoint")
+		return e.handleWebSearchStream(ctx, auth, req, opts, accessToken, profileArn)
+	}
+
+	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	defer reporter.trackFailure(ctx, &err)
+
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("kiro")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
@@ -1405,7 +1432,7 @@ func (e *KiroExecutor) executeStreamWithRetry(ctx context.Context, auth *cliprox
 				// So we always enable thinking parsing for Kiro responses
 				log.Debugf("kiro: stream thinkingEnabled = %v (always true for Kiro)", thinkingEnabled)
 
-				e.streamToChannel(ctx, resp.Body, out, from, req.Model, opts.OriginalRequest, body, reporter, thinkingEnabled)
+				e.streamToChannel(ctx, resp.Body, out, from, payloadRequestedModel(opts, req.Model), opts.OriginalRequest, body, reporter, thinkingEnabled)
 			}(httpResp, thinkingEnabled)
 
 			return out, nil
@@ -2102,6 +2129,22 @@ func (e *KiroExecutor) parseEventStream(body io.Reader) (string, []kiroclaude.Ki
 				}
 			}
 
+		case "contextUsageEvent":
+			// Handle context usage events from Kiro API
+			// Format: {"contextUsageEvent": {"contextUsagePercentage": 0.53}}
+			if ctxUsage, ok := event["contextUsageEvent"].(map[string]interface{}); ok {
+				if ctxPct, ok := ctxUsage["contextUsagePercentage"].(float64); ok {
+					upstreamContextPercentage = ctxPct
+					log.Debugf("kiro: parseEventStream received contextUsageEvent: %.2f%%", ctxPct*100)
+				}
+			} else {
+				// Try direct field (fallback)
+				if ctxPct, ok := event["contextUsagePercentage"].(float64); ok {
+					upstreamContextPercentage = ctxPct
+					log.Debugf("kiro: parseEventStream received contextUsagePercentage (direct): %.2f%%", ctxPct*100)
+				}
+			}
+
 		case "error", "exception", "internalServerException", "invalidStateEvent":
 			// Handle error events from Kiro API stream
 			errMsg := ""
@@ -2705,6 +2748,22 @@ func (e *KiroExecutor) streamToChannel(ctx context.Context, body io.Reader, out
 				}
 			}
 
+		case "contextUsageEvent":
+			// Handle context usage events from Kiro API
+			// Format: {"contextUsageEvent": {"contextUsagePercentage": 0.53}}
+			if ctxUsage, ok := event["contextUsageEvent"].(map[string]interface{}); ok {
+				if ctxPct, ok := ctxUsage["contextUsagePercentage"].(float64); ok {
+					upstreamContextPercentage = ctxPct
+					log.Debugf("kiro: streamToChannel received contextUsageEvent: %.2f%%", ctxPct*100)
+				}
+			} else {
+				// Try direct field (fallback)
+				if ctxPct, ok := event["contextUsagePercentage"].(float64); ok {
+					upstreamContextPercentage = ctxPct
+					log.Debugf("kiro: streamToChannel received contextUsagePercentage (direct): %.2f%%", ctxPct*100)
+				}
+			}
+
 		case "error", "exception", "internalServerException":
 			// Handle error events from Kiro API stream
 			errMsg := ""
@@ -4064,6 +4123,659 @@ func (e *KiroExecutor) isTokenExpired(accessToken string) bool {
 	return isExpired
 }
 
-// NOTE: Message merging functions moved to internal/translator/kiro/common/message_merge.go
-// NOTE: Tool calling support functions moved to internal/translator/kiro/claude/kiro_claude_tools.go
-// The executor now uses kiroclaude.* and kirocommon.* functions instead
+// ══════════════════════════════════════════════════════════════════════════════
+// Web Search Handler (MCP API)
+// ══════════════════════════════════════════════════════════════════════════════
+
+// fetchToolDescription caching:
+// Uses a mutex + fetched flag to ensure only one goroutine fetches at a time,
+// with automatic retry on failure:
+// - On failure, fetched stays false so subsequent calls will retry
+// - On success, fetched is set to true — subsequent calls skip immediately (mutex-free fast path)
+// The cached description is stored in the translator package via kiroclaude.SetWebSearchDescription(),
+// enabling the translator's convertClaudeToolsToKiro to read it when building Kiro requests.
+var (
+	toolDescMu      sync.Mutex
+	toolDescFetched atomic.Bool
+)
+
+// 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(ctx context.Context, mcpEndpoint, authToken string, httpClient *http.Client, auth *cliproxyauth.Auth, authAttrs map[string]string) {
+	// Fast path: already fetched successfully, no lock needed
+	if toolDescFetched.Load() {
+		return
+	}
+
+	toolDescMu.Lock()
+	defer toolDescMu.Unlock()
+
+	// Double-check after acquiring lock
+	if toolDescFetched.Load() {
+		return
+	}
+
+	handler := newWebSearchHandler(ctx, mcpEndpoint, authToken, httpClient, auth, 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.NewRequestWithContext(ctx, "POST", mcpEndpoint, bytes.NewReader(reqBody))
+	if err != nil {
+		log.Warnf("kiro/websearch: failed to create tools/list request: %v", err)
+		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)
+		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)
+		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")
+		return
+	}
+
+	for _, tool := range result.Result.Tools {
+		if tool.Name == "web_search" && tool.Description != "" {
+			kiroclaude.SetWebSearchDescription(tool.Description)
+			toolDescFetched.Store(true) // success — no more fetches
+			log.Infof("kiro/websearch: cached web_search description from tools/list (%d bytes)", len(tool.Description))
+			return
+		}
+	}
+
+	// web_search tool not found in response
+	log.Warnf("kiro/websearch: web_search tool not found in tools/list response")
+}
+
+// webSearchHandler handles web search requests via Kiro MCP API
+type webSearchHandler struct {
+	ctx         context.Context
+	mcpEndpoint string
+	httpClient  *http.Client
+	authToken   string
+	auth        *cliproxyauth.Auth // for applyDynamicFingerprint
+	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.
+// Pass a shared pooled client (e.g. from getKiroPooledHTTPClient) for connection reuse.
+func newWebSearchHandler(ctx context.Context, mcpEndpoint, authToken string, httpClient *http.Client, auth *cliproxyauth.Auth, authAttrs map[string]string) *webSearchHandler {
+	if httpClient == nil {
+		httpClient = &http.Client{
+			Timeout: 30 * time.Second,
+		}
+	}
+	return &webSearchHandler{
+		ctx:         ctx,
+		mcpEndpoint: mcpEndpoint,
+		httpClient:  httpClient,
+		authToken:   authToken,
+		auth:        auth,
+		authAttrs:   authAttrs,
+	}
+}
+
+// setMcpHeaders sets standard MCP API headers on the request,
+// aligned with the GAR request pattern.
+func (h *webSearchHandler) setMcpHeaders(req *http.Request) {
+	// 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. User-Agent: Reuse applyDynamicFingerprint for consistency
+	applyDynamicFingerprint(req, h.auth)
+
+	// 4. AWS SDK identifiers
+	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.
+func (h *webSearchHandler) callMcpAPI(request *kiroclaude.McpRequest) (*kiroclaude.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)
+			select {
+			case <-h.ctx.Done():
+				return nil, h.ctx.Err()
+			case <-time.After(backoff):
+			}
+		}
+
+		req, err := http.NewRequestWithContext(h.ctx, "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 kiroclaude.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
+}
+
+// webSearchAuthAttrs extracts auth attributes for MCP calls.
+// Used by handleWebSearch and handleWebSearchStream to pass custom headers.
+func webSearchAuthAttrs(auth *cliproxyauth.Auth) map[string]string {
+	if auth != nil {
+		return auth.Attributes
+	}
+	return nil
+}
+
+const maxWebSearchIterations = 5
+
+// handleWebSearchStream handles web_search requests:
+// Step 1: tools/list (sync) → fetch/cache tool description
+// Step 2+: MCP search → InjectToolResultsClaude → callKiroAndBuffer loop
+// Note: We skip the "model decides to search" step because Claude Code already
+// decided to use web_search. The Kiro tool description restricts non-coding
+// topics, so asking the model again would cause it to refuse valid searches.
+func (e *KiroExecutor) handleWebSearchStream(
+	ctx context.Context,
+	auth *cliproxyauth.Auth,
+	req cliproxyexecutor.Request,
+	opts cliproxyexecutor.Options,
+	accessToken, profileArn string,
+) (<-chan cliproxyexecutor.StreamChunk, error) {
+	// Extract search query from Claude Code's web_search tool_use
+	query := kiroclaude.ExtractSearchQuery(req.Payload)
+	if query == "" {
+		log.Warnf("kiro/websearch: failed to extract search query, falling back to normal flow")
+		return e.callKiroDirectStream(ctx, auth, req, opts, accessToken, profileArn)
+	}
+
+	// Build MCP endpoint using shared region resolution (supports api_region + ProfileARN fallback)
+	region := resolveKiroAPIRegion(auth)
+	mcpEndpoint := kiroclaude.BuildMcpEndpoint(region)
+
+	// ── Step 1: tools/list (SYNC) — cache tool description ──
+	{
+		authAttrs := webSearchAuthAttrs(auth)
+		fetchToolDescription(ctx, mcpEndpoint, accessToken, newKiroHTTPClientWithPooling(ctx, e.cfg, auth, 30*time.Second), auth, authAttrs)
+	}
+
+	// Create output channel
+	out := make(chan cliproxyexecutor.StreamChunk)
+
+	// Usage reporting: track web search requests like normal streaming requests
+	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+
+	go func() {
+		var wsErr error
+		defer reporter.trackFailure(ctx, &wsErr)
+		defer close(out)
+
+		// Estimate input tokens using tokenizer (matching streamToChannel pattern)
+		var totalUsage usage.Detail
+		if enc, tokErr := getTokenizer(req.Model); tokErr == nil {
+			if inp, e := countClaudeChatTokens(enc, req.Payload); e == nil && inp > 0 {
+				totalUsage.InputTokens = inp
+			} else {
+				totalUsage.InputTokens = int64(len(req.Payload) / 4)
+			}
+		} else {
+			totalUsage.InputTokens = int64(len(req.Payload) / 4)
+		}
+		if totalUsage.InputTokens == 0 && len(req.Payload) > 0 {
+			totalUsage.InputTokens = 1
+		}
+		var accumulatedOutputLen int
+		defer func() {
+			if wsErr != nil {
+				return // let trackFailure handle failure reporting
+			}
+			totalUsage.OutputTokens = int64(accumulatedOutputLen / 4)
+			if accumulatedOutputLen > 0 && totalUsage.OutputTokens == 0 {
+				totalUsage.OutputTokens = 1
+			}
+			reporter.publish(ctx, totalUsage)
+		}()
+
+		// Send message_start event to client (aligned with streamToChannel pattern)
+		// Use payloadRequestedModel to return user's original model alias
+		msgStart := kiroclaude.BuildClaudeMessageStartEvent(
+			payloadRequestedModel(opts, req.Model),
+			totalUsage.InputTokens,
+		)
+		select {
+		case <-ctx.Done():
+			return
+		case out <- cliproxyexecutor.StreamChunk{Payload: append(msgStart, '\n', '\n')}:
+		}
+
+		// ── Step 2+: MCP search → InjectToolResultsClaude → callKiroAndBuffer loop ──
+		contentBlockIndex := 0
+		currentQuery := query
+
+		// Replace web_search tool description with a minimal one that allows re-search.
+		// The original tools/list description from Kiro restricts non-coding topics,
+		// but we've already decided to search. We keep the tool so the model can
+		// request additional searches when results are insufficient.
+		simplifiedPayload, simplifyErr := kiroclaude.ReplaceWebSearchToolDescription(bytes.Clone(req.Payload))
+		if simplifyErr != nil {
+			log.Warnf("kiro/websearch: failed to simplify web_search tool: %v, using original payload", simplifyErr)
+			simplifiedPayload = bytes.Clone(req.Payload)
+		}
+
+		currentClaudePayload := simplifiedPayload
+		totalSearches := 0
+
+		// Generate toolUseId for the first iteration (Claude Code already decided to search)
+		currentToolUseId := fmt.Sprintf("srvtoolu_%s", kiroclaude.GenerateToolUseID())
+
+		for iteration := 0; iteration < maxWebSearchIterations; iteration++ {
+			log.Infof("kiro/websearch: search iteration %d/%d",
+				iteration+1, maxWebSearchIterations)
+
+			// MCP search
+			_, mcpRequest := kiroclaude.CreateMcpRequest(currentQuery)
+
+			authAttrs := webSearchAuthAttrs(auth)
+			handler := newWebSearchHandler(ctx, mcpEndpoint, accessToken, newKiroHTTPClientWithPooling(ctx, e.cfg, auth, 30*time.Second), auth, authAttrs)
+			mcpResponse, mcpErr := handler.callMcpAPI(mcpRequest)
+
+			var searchResults *kiroclaude.WebSearchResults
+			if mcpErr != nil {
+				log.Warnf("kiro/websearch: MCP API call failed: %v, continuing with empty results", mcpErr)
+			} else {
+				searchResults = kiroclaude.ParseSearchResults(mcpResponse)
+			}
+
+			resultCount := 0
+			if searchResults != nil {
+				resultCount = len(searchResults.Results)
+			}
+			totalSearches++
+			log.Infof("kiro/websearch: iteration %d — got %d search results", iteration+1, resultCount)
+
+			// Send search indicator events to client
+			searchEvents := kiroclaude.GenerateSearchIndicatorEvents(currentQuery, currentToolUseId, searchResults, contentBlockIndex)
+			for _, event := range searchEvents {
+				select {
+				case <-ctx.Done():
+					return
+				case out <- cliproxyexecutor.StreamChunk{Payload: event}:
+				}
+			}
+			contentBlockIndex += 2
+
+			// Inject tool_use + tool_result into Claude payload, then call GAR
+			var err error
+			currentClaudePayload, err = kiroclaude.InjectToolResultsClaude(currentClaudePayload, currentToolUseId, currentQuery, searchResults)
+			if err != nil {
+				log.Warnf("kiro/websearch: failed to inject tool results: %v", err)
+				wsErr = fmt.Errorf("failed to inject tool results: %w", err)
+				e.sendFallbackText(ctx, out, contentBlockIndex, currentQuery, searchResults)
+				return
+			}
+
+			// Call GAR with modified Claude payload (full translation pipeline)
+			modifiedReq := req
+			modifiedReq.Payload = currentClaudePayload
+			kiroChunks, kiroErr := e.callKiroAndBuffer(ctx, auth, modifiedReq, opts, accessToken, profileArn)
+			if kiroErr != nil {
+				log.Warnf("kiro/websearch: Kiro API failed at iteration %d: %v", iteration+1, kiroErr)
+				wsErr = fmt.Errorf("Kiro API failed at iteration %d: %w", iteration+1, kiroErr)
+				e.sendFallbackText(ctx, out, contentBlockIndex, currentQuery, searchResults)
+				return
+			}
+
+			// Analyze response
+			analysis := kiroclaude.AnalyzeBufferedStream(kiroChunks)
+			log.Infof("kiro/websearch: iteration %d — stop_reason: %s, has_tool_use: %v",
+				iteration+1, analysis.StopReason, analysis.HasWebSearchToolUse)
+
+			if analysis.HasWebSearchToolUse && analysis.WebSearchQuery != "" && iteration+1 < maxWebSearchIterations {
+				// Model wants another search
+				filteredChunks := kiroclaude.FilterChunksForClient(kiroChunks, analysis.WebSearchToolUseIndex, contentBlockIndex)
+				for _, chunk := range filteredChunks {
+					select {
+					case <-ctx.Done():
+						return
+					case out <- cliproxyexecutor.StreamChunk{Payload: chunk}:
+					}
+				}
+
+				currentQuery = analysis.WebSearchQuery
+				currentToolUseId = analysis.WebSearchToolUseId
+				continue
+			}
+
+			// Model returned final response — stream to client
+			for _, chunk := range kiroChunks {
+				if contentBlockIndex > 0 && len(chunk) > 0 {
+					adjusted, shouldForward := kiroclaude.AdjustSSEChunk(chunk, contentBlockIndex)
+					if !shouldForward {
+						continue
+					}
+					accumulatedOutputLen += len(adjusted)
+					select {
+					case <-ctx.Done():
+						return
+					case out <- cliproxyexecutor.StreamChunk{Payload: adjusted}:
+					}
+				} else {
+					accumulatedOutputLen += len(chunk)
+					select {
+					case <-ctx.Done():
+						return
+					case out <- cliproxyexecutor.StreamChunk{Payload: chunk}:
+					}
+				}
+			}
+			log.Infof("kiro/websearch: completed after %d search iteration(s), total searches: %d", iteration+1, totalSearches)
+			return
+		}
+
+		log.Warnf("kiro/websearch: reached max iterations (%d), stopping search loop", maxWebSearchIterations)
+	}()
+
+	return out, nil
+}
+
+// handleWebSearch handles web_search requests for non-streaming Execute path.
+// Performs MCP search synchronously, injects results into the request payload,
+// then calls the normal non-streaming Kiro API path which returns a proper
+// Claude JSON response (not SSE chunks).
+func (e *KiroExecutor) handleWebSearch(
+	ctx context.Context,
+	auth *cliproxyauth.Auth,
+	req cliproxyexecutor.Request,
+	opts cliproxyexecutor.Options,
+	accessToken, profileArn string,
+) (cliproxyexecutor.Response, error) {
+	// Extract search query from Claude Code's web_search tool_use
+	query := kiroclaude.ExtractSearchQuery(req.Payload)
+	if query == "" {
+		log.Warnf("kiro/websearch: non-stream: failed to extract search query, falling back to normal Execute")
+		// Fall through to normal non-streaming path
+		return e.executeNonStreamFallback(ctx, auth, req, opts, accessToken, profileArn)
+	}
+
+	// Build MCP endpoint using shared region resolution (supports api_region + ProfileARN fallback)
+	region := resolveKiroAPIRegion(auth)
+	mcpEndpoint := kiroclaude.BuildMcpEndpoint(region)
+
+	// Step 1: Fetch/cache tool description (sync)
+	{
+		authAttrs := webSearchAuthAttrs(auth)
+		fetchToolDescription(ctx, mcpEndpoint, accessToken, newKiroHTTPClientWithPooling(ctx, e.cfg, auth, 30*time.Second), auth, authAttrs)
+	}
+
+	// Step 2: Perform MCP search
+	_, mcpRequest := kiroclaude.CreateMcpRequest(query)
+
+	authAttrs := webSearchAuthAttrs(auth)
+	handler := newWebSearchHandler(ctx, mcpEndpoint, accessToken, newKiroHTTPClientWithPooling(ctx, e.cfg, auth, 30*time.Second), auth, authAttrs)
+	mcpResponse, mcpErr := handler.callMcpAPI(mcpRequest)
+
+	var searchResults *kiroclaude.WebSearchResults
+	if mcpErr != nil {
+		log.Warnf("kiro/websearch: non-stream: MCP API call failed: %v, continuing with empty results", mcpErr)
+	} else {
+		searchResults = kiroclaude.ParseSearchResults(mcpResponse)
+	}
+
+	resultCount := 0
+	if searchResults != nil {
+		resultCount = len(searchResults.Results)
+	}
+	log.Infof("kiro/websearch: non-stream: got %d search results", resultCount)
+
+	// Step 3: Replace restrictive web_search tool description (align with streaming path)
+	simplifiedPayload, simplifyErr := kiroclaude.ReplaceWebSearchToolDescription(bytes.Clone(req.Payload))
+	if simplifyErr != nil {
+		log.Warnf("kiro/websearch: non-stream: failed to simplify web_search tool: %v, using original payload", simplifyErr)
+		simplifiedPayload = bytes.Clone(req.Payload)
+	}
+
+	// Step 4: Inject search tool_use + tool_result into Claude payload
+	currentToolUseId := fmt.Sprintf("srvtoolu_%s", kiroclaude.GenerateToolUseID())
+	modifiedPayload, err := kiroclaude.InjectToolResultsClaude(simplifiedPayload, currentToolUseId, query, searchResults)
+	if err != nil {
+		log.Warnf("kiro/websearch: non-stream: failed to inject tool results: %v, falling back", err)
+		return e.executeNonStreamFallback(ctx, auth, req, opts, accessToken, profileArn)
+	}
+
+	// Step 5: Call Kiro API via the normal non-streaming path (executeWithRetry)
+	// This path uses parseEventStream → BuildClaudeResponse → TranslateNonStream
+	// to produce a proper Claude JSON response
+	modifiedReq := req
+	modifiedReq.Payload = modifiedPayload
+
+	resp, err := e.executeNonStreamFallback(ctx, auth, modifiedReq, opts, accessToken, profileArn)
+	if err != nil {
+		return resp, err
+	}
+
+	// Step 6: Inject server_tool_use + web_search_tool_result into response
+	// so Claude Code can display "Did X searches in Ys"
+	indicators := []kiroclaude.SearchIndicator{
+		{
+			ToolUseID: currentToolUseId,
+			Query:     query,
+			Results:   searchResults,
+		},
+	}
+	injectedPayload, injErr := kiroclaude.InjectSearchIndicatorsInResponse(resp.Payload, indicators)
+	if injErr != nil {
+		log.Warnf("kiro/websearch: non-stream: failed to inject search indicators: %v", injErr)
+	} else {
+		resp.Payload = injectedPayload
+	}
+
+	return resp, nil
+}
+
+// callKiroAndBuffer calls the Kiro API and buffers all response chunks.
+// Returns the buffered chunks for analysis before forwarding to client.
+// Usage reporting is NOT done here — the caller (handleWebSearchStream) manages its own reporter.
+func (e *KiroExecutor) callKiroAndBuffer(
+	ctx context.Context,
+	auth *cliproxyauth.Auth,
+	req cliproxyexecutor.Request,
+	opts cliproxyexecutor.Options,
+	accessToken, profileArn string,
+) ([][]byte, error) {
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("kiro")
+	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+	log.Debugf("kiro/websearch GAR request: %d bytes", len(body))
+
+	kiroModelID := e.mapModelToKiro(req.Model)
+	isAgentic, isChatOnly := determineAgenticMode(req.Model)
+	effectiveProfileArn := getEffectiveProfileArnWithWarning(auth, profileArn)
+
+	tokenKey := getTokenKey(auth)
+
+	kiroStream, err := e.executeStreamWithRetry(
+		ctx, auth, req, opts, accessToken, effectiveProfileArn,
+		nil, body, from, nil, "", kiroModelID, isAgentic, isChatOnly, tokenKey,
+	)
+	if err != nil {
+		return nil, err
+	}
+
+	// Buffer all chunks
+	var chunks [][]byte
+	for chunk := range kiroStream {
+		if chunk.Err != nil {
+			return chunks, chunk.Err
+		}
+		if len(chunk.Payload) > 0 {
+			chunks = append(chunks, bytes.Clone(chunk.Payload))
+		}
+	}
+
+	log.Debugf("kiro/websearch GAR response: %d chunks buffered", len(chunks))
+
+	return chunks, nil
+}
+
+// callKiroDirectStream creates a direct streaming channel to Kiro API without search.
+func (e *KiroExecutor) callKiroDirectStream(
+	ctx context.Context,
+	auth *cliproxyauth.Auth,
+	req cliproxyexecutor.Request,
+	opts cliproxyexecutor.Options,
+	accessToken, profileArn string,
+) (<-chan cliproxyexecutor.StreamChunk, error) {
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("kiro")
+	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+
+	kiroModelID := e.mapModelToKiro(req.Model)
+	isAgentic, isChatOnly := determineAgenticMode(req.Model)
+	effectiveProfileArn := getEffectiveProfileArnWithWarning(auth, profileArn)
+
+	tokenKey := getTokenKey(auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	var streamErr error
+	defer reporter.trackFailure(ctx, &streamErr)
+
+	stream, streamErr := e.executeStreamWithRetry(
+		ctx, auth, req, opts, accessToken, effectiveProfileArn,
+		nil, body, from, reporter, "", kiroModelID, isAgentic, isChatOnly, tokenKey,
+	)
+	return stream, streamErr
+}
+
+// sendFallbackText sends a simple text response when the Kiro API fails during the search loop.
+// Delegates SSE event construction to kiroclaude.BuildFallbackTextEvents() for alignment
+// with how streamToChannel() uses BuildClaude*Event() functions.
+func (e *KiroExecutor) sendFallbackText(
+	ctx context.Context,
+	out chan<- cliproxyexecutor.StreamChunk,
+	contentBlockIndex int,
+	query string,
+	searchResults *kiroclaude.WebSearchResults,
+) {
+	events := kiroclaude.BuildFallbackTextEvents(contentBlockIndex, query, searchResults)
+	for _, event := range events {
+		select {
+		case <-ctx.Done():
+			return
+		case out <- cliproxyexecutor.StreamChunk{Payload: append(event, '\n', '\n')}:
+		}
+	}
+}
+
+// executeNonStreamFallback runs the standard non-streaming Execute path for a request.
+// Used by handleWebSearch after injecting search results, or as a fallback.
+func (e *KiroExecutor) executeNonStreamFallback(
+	ctx context.Context,
+	auth *cliproxyauth.Auth,
+	req cliproxyexecutor.Request,
+	opts cliproxyexecutor.Options,
+	accessToken, profileArn string,
+) (cliproxyexecutor.Response, error) {
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("kiro")
+	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+
+	kiroModelID := e.mapModelToKiro(req.Model)
+	isAgentic, isChatOnly := determineAgenticMode(req.Model)
+	effectiveProfileArn := getEffectiveProfileArnWithWarning(auth, profileArn)
+	tokenKey := getTokenKey(auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	var err error
+	defer reporter.trackFailure(ctx, &err)
+
+	resp, err := e.executeWithRetry(ctx, auth, req, opts, accessToken, effectiveProfileArn, nil, body, from, to, reporter, "", kiroModelID, isAgentic, isChatOnly, tokenKey)
+	return resp, err
+}

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.
@@ -344,8 +342,18 @@ 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

internal/runtime/executor/thinking_providers.go

@@ -7,5 +7,6 @@ import (
 	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/gemini"
 	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/geminicli"
 	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/iflow"
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/kimi"
 	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/openai"
 )

internal/store/gitstore.go

@@ -21,6 +21,9 @@ import (
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 )
 
+// gcInterval defines minimum time between garbage collection runs.
+const gcInterval = 5 * time.Minute
+
 // GitTokenStore persists token records and auth metadata using git as the backing storage.
 type GitTokenStore struct {
 	mu        sync.Mutex
@@ -31,6 +34,7 @@ type GitTokenStore struct {
 	remote    string
 	username  string
 	password  string
+	lastGC    time.Time
 }
 
 // NewGitTokenStore creates a token store that saves credentials to disk through the
@@ -613,6 +617,7 @@ func (s *GitTokenStore) commitAndPushLocked(message string, relPaths ...string)
 	} else if errRewrite := s.rewriteHeadAsSingleCommit(repo, headRef.Name(), commitHash, message, signature); errRewrite != nil {
 		return errRewrite
 	}
+	s.maybeRunGC(repo)
 	if err = repo.Push(&git.PushOptions{Auth: s.gitAuth(), Force: true}); err != nil {
 		if errors.Is(err, git.NoErrAlreadyUpToDate) {
 			return nil
@@ -652,6 +657,23 @@ func (s *GitTokenStore) rewriteHeadAsSingleCommit(repo *git.Repository, branch p
 	return nil
 }
 
+func (s *GitTokenStore) maybeRunGC(repo *git.Repository) {
+	now := time.Now()
+	if now.Sub(s.lastGC) < gcInterval {
+		return
+	}
+	s.lastGC = now
+
+	pruneOpts := git.PruneOptions{
+		OnlyObjectsOlderThan: now,
+		Handler:              repo.DeleteObject,
+	}
+	if err := repo.Prune(pruneOpts); err != nil && !errors.Is(err, git.ErrLooseObjectsNotSupported) {
+		return
+	}
+	_ = repo.RepackObjects(&git.RepackConfig{})
+}
+
 // PersistConfig commits and pushes configuration changes to git.
 func (s *GitTokenStore) PersistConfig(_ context.Context) error {
 	if err := s.EnsureRepository(); err != nil {

internal/thinking/apply.go

@@ -18,6 +18,7 @@ var providerAppliers = map[string]ProviderApplier{
 	"codex":       nil,
 	"iflow":       nil,
 	"antigravity": nil,
+	"kimi":        nil,
 }
 
 // GetProviderApplier returns the ProviderApplier for the given provider name.
@@ -326,6 +327,9 @@ func extractThinkingConfig(body []byte, provider string) ThinkingConfig {
 			return config
 		}
 		return extractOpenAIConfig(body)
+	case "kimi":
+		// Kimi uses OpenAI-compatible reasoning_effort format
+		return extractOpenAIConfig(body)
 	default:
 		return ThinkingConfig{}
 	}

internal/thinking/provider/kimi/apply.go

@@ -0,0 +1,126 @@
+// Package kimi implements thinking configuration for Kimi (Moonshot AI) models.
+//
+// Kimi models use the OpenAI-compatible reasoning_effort format with discrete levels
+// (low/medium/high). The provider strips any existing thinking config and applies
+// the unified ThinkingConfig in OpenAI format.
+package kimi
+
+import (
+	"fmt"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier implements thinking.ProviderApplier for Kimi models.
+//
+// Kimi-specific behavior:
+//   - Output format: reasoning_effort (string: low/medium/high)
+//   - Uses OpenAI-compatible format
+//   - Supports budget-to-level conversion
+type Applier struct{}
+
+var _ thinking.ProviderApplier = (*Applier)(nil)
+
+// NewApplier creates a new Kimi thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("kimi", NewApplier())
+}
+
+// Apply applies thinking configuration to Kimi request body.
+//
+// Expected output format:
+//
+//	{
+//	  "reasoning_effort": "high"
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return applyCompatibleKimi(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	var effort string
+	switch config.Mode {
+	case thinking.ModeLevel:
+		if config.Level == "" {
+			return body, nil
+		}
+		effort = string(config.Level)
+	case thinking.ModeNone:
+		// Kimi uses "none" to disable thinking
+		effort = string(thinking.LevelNone)
+	case thinking.ModeBudget:
+		// Convert budget to level using threshold mapping
+		level, ok := thinking.ConvertBudgetToLevel(config.Budget)
+		if !ok {
+			return body, nil
+		}
+		effort = level
+	case thinking.ModeAuto:
+		// Auto mode maps to "auto" effort
+		effort = string(thinking.LevelAuto)
+	default:
+		return body, nil
+	}
+
+	if effort == "" {
+		return body, nil
+	}
+
+	result, err := sjson.SetBytes(body, "reasoning_effort", effort)
+	if err != nil {
+		return body, fmt.Errorf("kimi thinking: failed to set reasoning_effort: %w", err)
+	}
+	return result, nil
+}
+
+// applyCompatibleKimi applies thinking config for user-defined Kimi models.
+func applyCompatibleKimi(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	var effort string
+	switch config.Mode {
+	case thinking.ModeLevel:
+		if config.Level == "" {
+			return body, nil
+		}
+		effort = string(config.Level)
+	case thinking.ModeNone:
+		effort = string(thinking.LevelNone)
+		if config.Level != "" {
+			effort = string(config.Level)
+		}
+	case thinking.ModeAuto:
+		effort = string(thinking.LevelAuto)
+	case thinking.ModeBudget:
+		// Convert budget to level
+		level, ok := thinking.ConvertBudgetToLevel(config.Budget)
+		if !ok {
+			return body, nil
+		}
+		effort = level
+	default:
+		return body, nil
+	}
+
+	result, err := sjson.SetBytes(body, "reasoning_effort", effort)
+	if err != nil {
+		return body, fmt.Errorf("kimi thinking: failed to set reasoning_effort: %w", err)
+	}
+	return result, nil
+}

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -344,7 +344,8 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 	// Inject interleaved thinking hint when both tools and thinking are active
 	hasTools := toolDeclCount > 0
 	thinkingResult := gjson.GetBytes(rawJSON, "thinking")
-	hasThinking := thinkingResult.Exists() && thinkingResult.IsObject() && thinkingResult.Get("type").String() == "enabled"
+	thinkingType := thinkingResult.Get("type").String()
+	hasThinking := thinkingResult.Exists() && thinkingResult.IsObject() && (thinkingType == "enabled" || thinkingType == "adaptive")
 	isClaudeThinking := util.IsClaudeThinkingModel(modelName)
 
 	if hasTools && hasThinking && isClaudeThinking {
@@ -377,12 +378,18 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
 	if t := gjson.GetBytes(rawJSON, "thinking"); enableThoughtTranslate && t.Exists() && t.IsObject() {
-		if t.Get("type").String() == "enabled" {
+		switch t.Get("type").String() {
+		case "enabled":
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
+		case "adaptive":
+			// Keep adaptive as a high level sentinel; ApplyThinking resolves it
+			// to model-specific max capability.
+			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}
 	if v := gjson.GetBytes(rawJSON, "temperature"); v.Exists() && v.Type == gjson.Number {

internal/translator/codex/claude/codex_claude_request.go

@@ -222,6 +222,10 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 					reasoningEffort = effort
 				}
 			}
+		case "adaptive":
+			// Claude adaptive means "enable with max capacity"; keep it as highest level
+			// and let ApplyThinking normalize per target model capability.
+			reasoningEffort = string(thinking.LevelXHigh)
 		case "disabled":
 			if effort, ok := thinking.ConvertBudgetToLevel(0); ok && effort != "" {
 				reasoningEffort = effort

internal/translator/codex/claude/codex_claude_response.go

@@ -113,10 +113,10 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 		template = `{"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
 		p := (*param).(*ConvertCodexResponseToClaudeParams).HasToolCall
 		stopReason := rootResult.Get("response.stop_reason").String()
-		if stopReason != "" {
-			template, _ = sjson.Set(template, "delta.stop_reason", stopReason)
-		} else if p {
+		if p {
 			template, _ = sjson.Set(template, "delta.stop_reason", "tool_use")
+		} else if stopReason == "max_tokens" || stopReason == "stop" {
+			template, _ = sjson.Set(template, "delta.stop_reason", stopReason)
 		} else {
 			template, _ = sjson.Set(template, "delta.stop_reason", "end_turn")
 		}

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

@@ -90,6 +90,9 @@ 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())
 		}
@@ -205,6 +208,9 @@ 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())
 		}

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

@@ -27,6 +27,9 @@ func ConvertOpenAIResponsesRequestToCodex(modelName string, inputRawJSON []byte,
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "top_p")
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "service_tier")
 
+	// Delete the user field as it is not supported by the Codex upstream.
+	rawJSON, _ = sjson.DeleteBytes(rawJSON, "user")
+
 	// Convert role "system" to "developer" in input array to comply with Codex API requirements.
 	rawJSON = convertSystemRoleToDeveloper(rawJSON)
 

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

@@ -263,3 +263,20 @@ func TestConvertSystemRoleToDeveloper_AssistantRole(t *testing.T) {
 		t.Errorf("Expected third role 'assistant', got '%s'", thirdRole.String())
 	}
 }
+
+func TestUserFieldDeletion(t *testing.T) {  
+	inputJSON := []byte(`{  
+		"model": "gpt-5.2",  
+		"user": "test-user",  
+		"input": [{"role": "user", "content": "Hello"}]  
+	}`)  
+	  
+	output := ConvertOpenAIResponsesRequestToCodex("gpt-5.2", inputJSON, false)  
+	outputStr := string(output)  
+	  
+	// Verify user field is deleted  
+	userField := gjson.Get(outputStr, "user")  
+	if userField.Exists() {
+		t.Errorf("user field should be deleted, but it was found with value: %s", userField.Raw)
+	}
+}

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

@@ -173,12 +173,18 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
-		if t.Get("type").String() == "enabled" {
+		switch t.Get("type").String() {
+		case "enabled":
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
+		case "adaptive":
+			// Keep adaptive as a high level sentinel; ApplyThinking resolves it
+			// to model-specific max capability.
+			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}
 	if v := gjson.GetBytes(rawJSON, "temperature"); v.Exists() && v.Type == gjson.Number {

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

@@ -14,6 +14,7 @@ import (
 	"time"
 
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/openai/chat-completions"
+	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -77,14 +78,20 @@ func ConvertCliResponseToOpenAI(_ context.Context, _ string, originalRequestRawJ
 		template, _ = sjson.Set(template, "id", responseIDResult.String())
 	}
 
-	// Extract and set the finish reason.
-	if finishReasonResult := gjson.GetBytes(rawJSON, "response.candidates.0.finishReason"); finishReasonResult.Exists() {
-		template, _ = sjson.Set(template, "choices.0.finish_reason", strings.ToLower(finishReasonResult.String()))
-		template, _ = sjson.Set(template, "choices.0.native_finish_reason", strings.ToLower(finishReasonResult.String()))
+	finishReason := ""
+	if stopReasonResult := gjson.GetBytes(rawJSON, "response.stop_reason"); stopReasonResult.Exists() {
+		finishReason = stopReasonResult.String()
 	}
+	if finishReason == "" {
+		if finishReasonResult := gjson.GetBytes(rawJSON, "response.candidates.0.finishReason"); finishReasonResult.Exists() {
+			finishReason = finishReasonResult.String()
+		}
+	}
+	finishReason = strings.ToLower(finishReason)
 
 	// Extract and set usage metadata (token counts).
 	if usageResult := gjson.GetBytes(rawJSON, "response.usageMetadata"); usageResult.Exists() {
+		cachedTokenCount := usageResult.Get("cachedContentTokenCount").Int()
 		if candidatesTokenCountResult := usageResult.Get("candidatesTokenCount"); candidatesTokenCountResult.Exists() {
 			template, _ = sjson.Set(template, "usage.completion_tokens", candidatesTokenCountResult.Int())
 		}
@@ -97,6 +104,14 @@ func ConvertCliResponseToOpenAI(_ context.Context, _ string, originalRequestRawJ
 		if thoughtsTokenCount > 0 {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", thoughtsTokenCount)
 		}
+		// Include cached token count if present (indicates prompt caching is working)
+		if cachedTokenCount > 0 {
+			var err error
+			template, err = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cachedTokenCount)
+			if err != nil {
+				log.Warnf("gemini-cli openai response: failed to set cached_tokens: %v", err)
+			}
+		}
 	}
 
 	// Process the main content part of the response.
@@ -187,6 +202,12 @@ func ConvertCliResponseToOpenAI(_ context.Context, _ string, originalRequestRawJ
 	if hasFunctionCall {
 		template, _ = sjson.Set(template, "choices.0.finish_reason", "tool_calls")
 		template, _ = sjson.Set(template, "choices.0.native_finish_reason", "tool_calls")
+	} else if finishReason != "" && (*param).(*convertCliResponseToOpenAIChatParams).FunctionIndex == 0 {
+		// Only pass through specific finish reasons
+		if finishReason == "max_tokens" || finishReason == "stop" {
+			template, _ = sjson.Set(template, "choices.0.finish_reason", finishReason)
+			template, _ = sjson.Set(template, "choices.0.native_finish_reason", finishReason)
+		}
 	}
 
 	return []string{template}

internal/translator/gemini/claude/gemini_claude_request.go

@@ -154,12 +154,18 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when enabled
 	// Translator only does format conversion, ApplyThinking handles model capability validation.
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
-		if t.Get("type").String() == "enabled" {
+		switch t.Get("type").String() {
+		case "enabled":
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
 				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
 				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.includeThoughts", true)
 			}
+		case "adaptive":
+			// Keep adaptive as a high level sentinel; ApplyThinking resolves it
+			// to model-specific max capability.
+			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingLevel", "high")
+			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}
 	if v := gjson.GetBytes(rawJSON, "temperature"); v.Exists() && v.Type == gjson.Number {

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

@@ -129,11 +129,16 @@ func ConvertGeminiResponseToOpenAI(_ context.Context, _ string, originalRequestR
 			candidateIndex := int(candidate.Get("index").Int())
 			template, _ = sjson.Set(template, "choices.0.index", candidateIndex)
 
-			// Extract and set the finish reason.
-			if finishReasonResult := candidate.Get("finishReason"); finishReasonResult.Exists() {
-				template, _ = sjson.Set(template, "choices.0.finish_reason", strings.ToLower(finishReasonResult.String()))
-				template, _ = sjson.Set(template, "choices.0.native_finish_reason", strings.ToLower(finishReasonResult.String()))
+			finishReason := ""
+			if stopReasonResult := gjson.GetBytes(rawJSON, "stop_reason"); stopReasonResult.Exists() {
+				finishReason = stopReasonResult.String()
 			}
+			if finishReason == "" {
+				if finishReasonResult := gjson.GetBytes(rawJSON, "candidates.0.finishReason"); finishReasonResult.Exists() {
+					finishReason = finishReasonResult.String()
+				}
+			}
+			finishReason = strings.ToLower(finishReason)
 
 			partsResult := candidate.Get("content.parts")
 			hasFunctionCall := false
@@ -225,6 +230,12 @@ func ConvertGeminiResponseToOpenAI(_ context.Context, _ string, originalRequestR
 			if hasFunctionCall {
 				template, _ = sjson.Set(template, "choices.0.finish_reason", "tool_calls")
 				template, _ = sjson.Set(template, "choices.0.native_finish_reason", "tool_calls")
+			} else if finishReason != "" {
+				// Only pass through specific finish reasons
+				if finishReason == "max_tokens" || finishReason == "stop" {
+					template, _ = sjson.Set(template, "choices.0.finish_reason", finishReason)
+					template, _ = sjson.Set(template, "choices.0.native_finish_reason", finishReason)
+				}
 			}
 
 			responseStrings = append(responseStrings, template)

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

@@ -117,19 +117,29 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 			switch itemType {
 			case "message":
 				if strings.EqualFold(itemRole, "system") {
-					if contentArray := item.Get("content"); contentArray.Exists() && contentArray.IsArray() {
-						var builder strings.Builder
-						contentArray.ForEach(func(_, contentItem gjson.Result) bool {
-							text := contentItem.Get("text").String()
-							if builder.Len() > 0 && text != "" {
-								builder.WriteByte('\n')
-							}
-							builder.WriteString(text)
-							return true
-						})
-						if !gjson.Get(out, "system_instruction").Exists() {
-							systemInstr := `{"parts":[{"text":""}]}`
-							systemInstr, _ = sjson.Set(systemInstr, "parts.0.text", builder.String())
+					if contentArray := item.Get("content"); contentArray.Exists() {
+						systemInstr := ""
+						if systemInstructionResult := gjson.Get(out, "system_instruction"); systemInstructionResult.Exists() {
+							systemInstr = systemInstructionResult.Raw
+						} else {
+							systemInstr = `{"parts":[]}`
+						}
+
+						if contentArray.IsArray() {
+							contentArray.ForEach(func(_, contentItem gjson.Result) bool {
+								part := `{"text":""}`
+								text := contentItem.Get("text").String()
+								part, _ = sjson.Set(part, "text", text)
+								systemInstr, _ = sjson.SetRaw(systemInstr, "parts.-1", part)
+								return true
+							})
+						} else if contentArray.Type == gjson.String {
+							part := `{"text":""}`
+							part, _ = sjson.Set(part, "text", contentArray.String())
+							systemInstr, _ = sjson.SetRaw(systemInstr, "parts.-1", part)
+						}
+
+						if systemInstr != `{"parts":[]}` {
 							out, _ = sjson.SetRaw(out, "system_instruction", systemInstr)
 						}
 					}
@@ -236,8 +246,22 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 					})
 
 					flush()
-				}
+				} else if contentArray.Type == gjson.String {
+					effRole := "user"
+					if itemRole != "" {
+						switch strings.ToLower(itemRole) {
+						case "assistant", "model":
+							effRole = "model"
+						default:
+							effRole = strings.ToLower(itemRole)
+						}
+					}
 
+					one := `{"role":"","parts":[{"text":""}]}`
+					one, _ = sjson.Set(one, "role", effRole)
+					one, _ = sjson.Set(one, "parts.0.text", contentArray.String())
+					out, _ = sjson.SetRaw(out, "contents.-1", one)
+				}
 			case "function_call":
 				// Handle function calls - convert to model message with functionCall
 				name := item.Get("name").String()

internal/translator/kiro/claude/kiro_claude_request.go

@@ -17,6 +17,9 @@ import (
 	"github.com/tidwall/gjson"
 )
 
+// remoteWebSearchDescription is a minimal fallback for when dynamic fetch from MCP tools/list hasn't completed yet.
+const remoteWebSearchDescription = "WebSearch looks up information outside the model's training data. Supports multiple queries to gather comprehensive information."
+
 // Kiro API request structs - field order determines JSON key order
 
 // KiroPayload is the top-level request structure for Kiro API
@@ -219,26 +222,7 @@ func BuildKiroPayload(claudeBody []byte, modelID, profileArn, origin string, isA
 	}
 
 	// Convert Claude tools to Kiro format
-	kiroTools, hasWebSearch := convertClaudeToolsToKiro(tools)
-
-	// If web_search was requested but filtered, inject alternative hint
-	if hasWebSearch {
-		webSearchHint := `[CRITICAL WEB ACCESS INSTRUCTION]
-You have the Fetch/read_url_content tool available. When the user asks about current events, weather, news, or any information that requires web access:
-- DO NOT say you cannot search the web
-- DO NOT refuse to help with web-related queries
-- IMMEDIATELY use the Fetch tool to access relevant URLs
-- Use well-known official websites, documentation sites, or API endpoints
-- Construct appropriate URLs based on the query context
-
-IMPORTANT: Always attempt to fetch information FIRST before declining. You CAN access the web via Fetch.`
-		if systemPrompt != "" {
-			systemPrompt += "\n" + webSearchHint
-		} else {
-			systemPrompt = webSearchHint
-		}
-		log.Infof("kiro: injected web_search alternative hint (tool was filtered)")
-	}
+	kiroTools := convertClaudeToolsToKiro(tools)
 
 	// Thinking mode implementation:
 	// Kiro API supports official thinking/reasoning mode via <thinking_mode> tag.
@@ -527,27 +511,15 @@ func ensureKiroInputSchema(parameters interface{}) interface{} {
 	}
 }
 
-// convertClaudeToolsToKiro converts Claude tools to Kiro format.
-// Returns the converted tools and a boolean indicating if web_search was filtered.
-func convertClaudeToolsToKiro(tools gjson.Result) ([]KiroToolWrapper, bool) {
+// convertClaudeToolsToKiro converts Claude tools to Kiro format
+func convertClaudeToolsToKiro(tools gjson.Result) []KiroToolWrapper {
 	var kiroTools []KiroToolWrapper
-	hasWebSearch := false
 	if !tools.IsArray() {
-		return kiroTools, hasWebSearch
+		return kiroTools
 	}
 
 	for _, tool := range tools.Array() {
 		name := tool.Get("name").String()
-
-		// Filter out web_search/websearch tools (Kiro API doesn't support them)
-		// This matches the behavior in AIClient-2-API/claude-kiro.js
-		nameLower := strings.ToLower(name)
-		if nameLower == "web_search" || nameLower == "websearch" {
-			log.Debugf("kiro: skipping unsupported tool: %s", name)
-			hasWebSearch = true
-			continue
-		}
-
 		description := tool.Get("description").String()
 		inputSchemaResult := tool.Get("input_schema")
 		var inputSchema interface{}
@@ -569,6 +541,18 @@ func convertClaudeToolsToKiro(tools gjson.Result) ([]KiroToolWrapper, bool) {
 			log.Debugf("kiro: tool '%s' has empty description, using default: %s", name, description)
 		}
 
+		// Rename web_search → remote_web_search for Kiro API compatibility
+		if name == "web_search" {
+			name = "remote_web_search"
+			// Prefer dynamically fetched description, fall back to hardcoded constant
+			if cached := GetWebSearchDescription(); cached != "" {
+				description = cached
+			} else {
+				description = remoteWebSearchDescription
+			}
+			log.Debugf("kiro: renamed tool web_search → remote_web_search")
+		}
+
 		// Truncate long descriptions (individual tool limit)
 		if len(description) > kirocommon.KiroMaxToolDescLen {
 			truncLen := kirocommon.KiroMaxToolDescLen - 30
@@ -591,7 +575,7 @@ func convertClaudeToolsToKiro(tools gjson.Result) ([]KiroToolWrapper, bool) {
 	// This prevents 500 errors when Claude Code sends too many tools
 	kiroTools = compressToolsIfNeeded(kiroTools)
 
-	return kiroTools, hasWebSearch
+	return kiroTools
 }
 
 // processMessages processes Claude messages and builds Kiro history
@@ -602,24 +586,39 @@ 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
 
 		if role == "user" {
 			userMsg, toolResults := BuildUserMessageStruct(msg, modelID, origin)
+			// CRITICAL: Kiro API requires content to be non-empty for ALL user messages
+			// This includes both history messages and the current message.
+			// When user message contains only tool_result (no text), content will be empty.
+			// This commonly happens in compaction requests from OpenCode.
+			if strings.TrimSpace(userMsg.Content) == "" {
+				if len(toolResults) > 0 {
+					userMsg.Content = kirocommon.DefaultUserContentWithToolResults
+				} else {
+					userMsg.Content = kirocommon.DefaultUserContent
+				}
+				log.Debugf("kiro: user content was empty, using default: %s", userMsg.Content)
+			}
 			if isLastMessage {
 				currentUserMsg = &userMsg
 				currentToolResults = toolResults
 			} else {
-				// CRITICAL: Kiro API requires content to be non-empty for history messages too
-				if strings.TrimSpace(userMsg.Content) == "" {
-					if len(toolResults) > 0 {
-						userMsg.Content = "Tool results provided."
-					} else {
-						userMsg.Content = "Continue"
-					}
-				}
 				// For history messages, embed tool results in context
 				if len(toolResults) > 0 {
 					userMsg.UserInputMessageContext = &KiroUserInputMessageContext{
@@ -650,6 +649,57 @@ func processMessages(messages gjson.Result, modelID, origin string) ([]KiroHisto
 		}
 	}
 
+	// POST-PROCESSING: Remove orphaned tool_results that have no matching tool_use
+	// in any assistant message. This happens when Claude Code compaction truncates
+	// the conversation and removes the assistant message containing the tool_use,
+	// but keeps the user message with the corresponding tool_result.
+	// Without this fix, Kiro API returns "Improperly formed request".
+	validToolUseIDs := make(map[string]bool)
+	for _, h := range history {
+		if h.AssistantResponseMessage != nil {
+			for _, tu := range h.AssistantResponseMessage.ToolUses {
+				validToolUseIDs[tu.ToolUseID] = true
+			}
+		}
+	}
+
+	// Filter orphaned tool results from history user messages
+	for i, h := range history {
+		if h.UserInputMessage != nil && h.UserInputMessage.UserInputMessageContext != nil {
+			ctx := h.UserInputMessage.UserInputMessageContext
+			if len(ctx.ToolResults) > 0 {
+				filtered := make([]KiroToolResult, 0, len(ctx.ToolResults))
+				for _, tr := range ctx.ToolResults {
+					if validToolUseIDs[tr.ToolUseID] {
+						filtered = append(filtered, tr)
+					} else {
+						log.Debugf("kiro: dropping orphaned tool_result in history[%d]: toolUseId=%s (no matching tool_use)", i, tr.ToolUseID)
+					}
+				}
+				ctx.ToolResults = filtered
+				if len(ctx.ToolResults) == 0 && len(ctx.Tools) == 0 {
+					h.UserInputMessage.UserInputMessageContext = nil
+				}
+			}
+		}
+	}
+
+	// Filter orphaned tool results from current message
+	if len(currentToolResults) > 0 {
+		filtered := make([]KiroToolResult, 0, len(currentToolResults))
+		for _, tr := range currentToolResults {
+			if validToolUseIDs[tr.ToolUseID] {
+				filtered = append(filtered, tr)
+			} else {
+				log.Debugf("kiro: dropping orphaned tool_result in currentMessage: toolUseId=%s (no matching tool_use)", tr.ToolUseID)
+			}
+		}
+		if len(filtered) != len(currentToolResults) {
+			log.Infof("kiro: dropped %d orphaned tool_result(s) from currentMessage (compaction artifact)", len(currentToolResults)-len(filtered))
+		}
+		currentToolResults = filtered
+	}
+
 	return history, currentUserMsg, currentToolResults
 }
 
@@ -872,6 +922,11 @@ func BuildAssistantMessageStruct(msg gjson.Result) KiroAssistantResponseMessage
 					})
 				}
 
+				// Rename web_search → remote_web_search to match convertClaudeToolsToKiro
+				if toolName == "web_search" {
+					toolName = "remote_web_search"
+				}
+
 				toolUses = append(toolUses, KiroToolUse{
 					ToolUseID: toolUseID,
 					Name:      toolName,

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
+}

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
+}

internal/translator/kiro/claude/kiro_websearch.go

@@ -0,0 +1,495 @@
+// Package claude provides web search functionality for Kiro translator.
+// 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"
+	log "github.com/sirupsen/logrus"
+	"github.com/tidwall/gjson"
+	"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"`
+	JSONRPC string    `json:"jsonrpc"`
+	Method  string    `json:"method"`
+	Params  McpParams `json:"params"`
+}
+
+// McpParams represents MCP request parameters
+type McpParams struct {
+	Name      string       `json:"name"`
+	Arguments McpArguments `json:"arguments"`
+}
+
+// McpArgumentsMeta represents the _meta field in MCP arguments
+type McpArgumentsMeta struct {
+	IsValid        bool       `json:"_isValid"`
+	ActivePath     []string   `json:"_activePath"`
+	CompletedPaths [][]string `json:"_completedPaths"`
+}
+
+// McpArguments represents MCP request arguments
+type McpArguments struct {
+	Query string            `json:"query"`
+	Meta  *McpArgumentsMeta `json:"_meta,omitempty"`
+}
+
+// McpResponse represents a JSON-RPC 2.0 response from Kiro MCP API
+type McpResponse struct {
+	Error   *McpError  `json:"error,omitempty"`
+	ID      string     `json:"id"`
+	JSONRPC string     `json:"jsonrpc"`
+	Result  *McpResult `json:"result,omitempty"`
+}
+
+// McpError represents an MCP error
+type McpError struct {
+	Code    *int    `json:"code,omitempty"`
+	Message *string `json:"message,omitempty"`
+}
+
+// McpResult represents MCP result
+type McpResult struct {
+	Content []McpContent `json:"content"`
+	IsError bool         `json:"isError"`
+}
+
+// McpContent represents MCP content item
+type McpContent struct {
+	ContentType string `json:"type"`
+	Text        string `json:"text"`
+}
+
+// WebSearchResults represents parsed search results
+type WebSearchResults struct {
+	Results      []WebSearchResult `json:"results"`
+	TotalResults *int              `json:"totalResults,omitempty"`
+	Query        *string           `json:"query,omitempty"`
+	Error        *string           `json:"error,omitempty"`
+}
+
+// WebSearchResult represents a single search result
+type WebSearchResult struct {
+	Title                string  `json:"title"`
+	URL                  string  `json:"url"`
+	Snippet              *string `json:"snippet,omitempty"`
+	PublishedDate        *int64  `json:"publishedDate,omitempty"`
+	ID                   *string `json:"id,omitempty"`
+	Domain               *string `json:"domain,omitempty"`
+	MaxVerbatimWordLimit *int    `json:"maxVerbatimWordLimit,omitempty"`
+	PublicDomain         *bool   `json:"publicDomain,omitempty"`
+}
+
+// isWebSearchTool checks if a tool name or type indicates a web_search tool.
+func isWebSearchTool(name, toolType string) bool {
+	return name == "web_search" ||
+		strings.HasPrefix(toolType, "web_search") ||
+		toolType == "web_search_20250305"
+}
+
+// HasWebSearchTool checks if the request contains ONLY a web_search tool.
+// Returns true only if tools array has exactly one tool named "web_search".
+// Only intercept pure web_search requests (single-tool array).
+func HasWebSearchTool(body []byte) bool {
+	tools := gjson.GetBytes(body, "tools")
+	if !tools.IsArray() {
+		return false
+	}
+
+	toolsArray := tools.Array()
+	if len(toolsArray) != 1 {
+		return false
+	}
+
+	// Check if the single tool is web_search
+	tool := toolsArray[0]
+
+	// Check both name and type fields for web_search detection
+	name := strings.ToLower(tool.Get("name").String())
+	toolType := strings.ToLower(tool.Get("type").String())
+
+	return isWebSearchTool(name, toolType)
+}
+
+// ExtractSearchQuery extracts the search query from the request.
+// Reads messages[0].content and removes "Perform a web search for the query: " prefix.
+func ExtractSearchQuery(body []byte) string {
+	messages := gjson.GetBytes(body, "messages")
+	if !messages.IsArray() || len(messages.Array()) == 0 {
+		return ""
+	}
+
+	firstMsg := messages.Array()[0]
+	content := firstMsg.Get("content")
+
+	var text string
+	if content.IsArray() {
+		// Array format: [{"type": "text", "text": "..."}]
+		for _, block := range content.Array() {
+			if block.Get("type").String() == "text" {
+				text = block.Get("text").String()
+				break
+			}
+		}
+	} else {
+		// String format
+		text = content.String()
+	}
+
+	// Remove prefix "Perform a web search for the query: "
+	const prefix = "Perform a web search for the query: "
+	if strings.HasPrefix(text, prefix) {
+		text = text[len(prefix):]
+	}
+
+	return strings.TrimSpace(text)
+}
+
+// generateRandomID8 generates an 8-character random lowercase alphanumeric string
+func generateRandomID8() string {
+	u := uuid.New()
+	return strings.ToLower(strings.ReplaceAll(u.String(), "-", "")[:8])
+}
+
+// CreateMcpRequest creates an MCP request for web search.
+// Returns (toolUseID, McpRequest)
+// ID format: web_search_tooluse_{22 random}_{timestamp_millis}_{8 random}
+func CreateMcpRequest(query string) (string, *McpRequest) {
+	random22 := GenerateToolUseID()
+	timestamp := time.Now().UnixMilli()
+	random8 := generateRandomID8()
+
+	requestID := fmt.Sprintf("web_search_tooluse_%s_%d_%s", random22, timestamp, random8)
+
+	// tool_use_id format: srvtoolu_{32 hex chars}
+	toolUseID := "srvtoolu_" + strings.ReplaceAll(uuid.New().String(), "-", "")[:32]
+
+	request := &McpRequest{
+		ID:      requestID,
+		JSONRPC: "2.0",
+		Method:  "tools/call",
+		Params: McpParams{
+			Name: "web_search",
+			Arguments: McpArguments{
+				Query: query,
+				Meta: &McpArgumentsMeta{
+					IsValid:        true,
+					ActivePath:     []string{"query"},
+					CompletedPaths: [][]string{{"query"}},
+				},
+			},
+		},
+	}
+
+	return toolUseID, request
+}
+
+// GenerateToolUseID generates a Kiro-style tool use ID (base62-like UUID)
+func GenerateToolUseID() string {
+	return strings.ReplaceAll(uuid.New().String(), "-", "")[:22]
+}
+
+// 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.
+// This keeps the tool available so the model can request additional searches.
+func ReplaceWebSearchToolDescription(body []byte) ([]byte, error) {
+	tools := gjson.GetBytes(body, "tools")
+	if !tools.IsArray() {
+		return body, nil
+	}
+
+	var updated []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) {
+			// Replace with a minimal web_search tool definition
+			minimalTool := map[string]interface{}{
+				"name":        "web_search",
+				"description": "Search the web for information. Use this when the previous search results are insufficient or when you need additional information on a different aspect of the query. Provide a refined or different search query.",
+				"input_schema": map[string]interface{}{
+					"type": "object",
+					"properties": map[string]interface{}{
+						"query": map[string]interface{}{
+							"type":        "string",
+							"description": "The search query to execute",
+						},
+					},
+					"required":             []string{"query"},
+					"additionalProperties": false,
+				},
+			}
+			minimalJSON, err := json.Marshal(minimalTool)
+			if err != nil {
+				return body, fmt.Errorf("failed to marshal minimal tool: %w", err)
+			}
+			updated = append(updated, json.RawMessage(minimalJSON))
+		} else {
+			updated = append(updated, json.RawMessage(tool.Raw))
+		}
+	}
+
+	updatedJSON, err := json.Marshal(updated)
+	if err != nil {
+		return body, fmt.Errorf("failed to marshal updated tools: %w", err)
+	}
+	result, err := sjson.SetRawBytes(body, "tools", updatedJSON)
+	if err != nil {
+		return body, fmt.Errorf("failed to set updated 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 {
+	var sb strings.Builder
+	sb.WriteString(fmt.Sprintf("[Web Search Results for \"%s\"]\n", query))
+
+	if results != nil && len(results.Results) > 0 {
+		for i, r := range results.Results {
+			sb.WriteString(fmt.Sprintf("%d. %s - %s\n", i+1, r.Title, r.URL))
+			if r.Snippet != nil && *r.Snippet != "" {
+				snippet := *r.Snippet
+				if len(snippet) > 500 {
+					snippet = snippet[:500] + "..."
+				}
+				sb.WriteString(fmt.Sprintf("   %s\n", snippet))
+			}
+		}
+	} else {
+		sb.WriteString("No results found.\n")
+	}
+
+	sb.WriteString("[End Web Search Results]")
+	return sb.String()
+}
+
+// FormatToolResultText formats search results as JSON text for the toolResults content field.
+// This matches the format observed in Kiro IDE HAR captures.
+func FormatToolResultText(results *WebSearchResults) string {
+	if results == nil || len(results.Results) == 0 {
+		return "No search results found."
+	}
+
+	text := fmt.Sprintf("Found %d search result(s):\n\n", len(results.Results))
+	resultJSON, err := json.MarshalIndent(results.Results, "", "  ")
+	if err != nil {
+		return text + "Error formatting results."
+	}
+	return text + string(resultJSON)
+}
+
+// InjectToolResultsClaude modifies a Claude-format JSON payload to append
+// tool_use (assistant) and tool_result (user) messages to the messages array.
+// BuildKiroPayload correctly translates:
+//   - assistant tool_use → KiroAssistantResponseMessage.toolUses
+//   - user tool_result   → KiroUserInputMessageContext.toolResults
+//
+// 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 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 {
+		return claudePayload, fmt.Errorf("failed to parse claude payload: %w", err)
+	}
+
+	messages, _ := payload["messages"].([]interface{})
+
+	// 1. Append assistant message with tool_use (matches HAR: assistantResponseMessage.toolUses)
+	assistantMsg := map[string]interface{}{
+		"role": "assistant",
+		"content": []interface{}{
+			map[string]interface{}{
+				"type":  "tool_use",
+				"id":    toolUseId,
+				"name":  "web_search",
+				"input": map[string]interface{}{"query": query},
+			},
+		},
+	}
+	messages = append(messages, assistantMsg)
+
+	// 2. Append user message with tool_result + search behavior instructions.
+	// NOTE: We embed search instructions HERE (not in system prompt) because
+	// BuildKiroPayload clears the system prompt when len(history) > 0,
+	// which is always true after injecting assistant + user messages.
+	now := time.Now()
+	searchGuidance := fmt.Sprintf(`<search_guidance>
+Current date: %s (%s)
+
+IMPORTANT: Evaluate the search results above carefully. If the results are:
+- Mostly spam, SEO junk, or unrelated websites
+- Missing actual information about the query topic
+- Outdated or not matching the requested time frame
+
+Then you MUST use the web_search tool again with a refined query. Try:
+- Rephrasing in English for better coverage
+- Using more specific keywords
+- Adding date context
+
+Do NOT apologize for bad results without first attempting a re-search.
+</search_guidance>`, now.Format("January 2, 2006"), now.Format("Monday"))
+
+	userMsg := map[string]interface{}{
+		"role": "user",
+		"content": []interface{}{
+			map[string]interface{}{
+				"type":        "tool_result",
+				"tool_use_id": toolUseId,
+				"content":     FormatToolResultText(results),
+			},
+			map[string]interface{}{
+				"type": "text",
+				"text": searchGuidance,
+			},
+		},
+	}
+	messages = append(messages, userMsg)
+
+	payload["messages"] = messages
+
+	result, err := json.Marshal(payload)
+	if err != nil {
+		return claudePayload, fmt.Errorf("failed to marshal updated payload: %w", err)
+	}
+
+	log.Infof("kiro/websearch: injected tool_use+tool_result (toolUseId=%s, messages=%d)",
+		toolUseId, len(messages))
+
+	return result, nil
+}
+
+// InjectSearchIndicatorsInResponse prepends server_tool_use + web_search_tool_result
+// content blocks into a non-streaming Claude JSON response. Claude Code counts
+// server_tool_use blocks to display "Did X searches in Ys".
+//
+// Input response:  {"content": [{"type":"text","text":"..."}], ...}
+// Output response: {"content": [{"type":"server_tool_use",...}, {"type":"web_search_tool_result",...}, {"type":"text","text":"..."}], ...}
+func InjectSearchIndicatorsInResponse(responsePayload []byte, searches []SearchIndicator) ([]byte, error) {
+	if len(searches) == 0 {
+		return responsePayload, nil
+	}
+
+	var resp map[string]interface{}
+	if err := json.Unmarshal(responsePayload, &resp); err != nil {
+		return responsePayload, fmt.Errorf("failed to parse response: %w", err)
+	}
+
+	existingContent, _ := resp["content"].([]interface{})
+
+	// Build new content: search indicators first, then existing content
+	newContent := make([]interface{}, 0, len(searches)*2+len(existingContent))
+
+	for _, s := range searches {
+		// server_tool_use block
+		newContent = append(newContent, map[string]interface{}{
+			"type":  "server_tool_use",
+			"id":    s.ToolUseID,
+			"name":  "web_search",
+			"input": map[string]interface{}{"query": s.Query},
+		})
+
+		// web_search_tool_result block
+		searchContent := make([]map[string]interface{}, 0)
+		if s.Results != nil {
+			for _, r := range s.Results.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,
+				})
+			}
+		}
+		newContent = append(newContent, map[string]interface{}{
+			"type":        "web_search_tool_result",
+			"tool_use_id": s.ToolUseID,
+			"content":     searchContent,
+		})
+	}
+
+	// Append existing content blocks
+	newContent = append(newContent, existingContent...)
+	resp["content"] = newContent
+
+	result, err := json.Marshal(resp)
+	if err != nil {
+		return responsePayload, fmt.Errorf("failed to marshal response: %w", err)
+	}
+
+	log.Infof("kiro/websearch: injected %d search indicator(s) into non-stream response", len(searches))
+	return result, nil
+}
+
+// SearchIndicator holds the data for one search operation to inject into a response.
+type SearchIndicator struct {
+	ToolUseID string
+	Query     string
+	Results   *WebSearchResults
+}
+
+// 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)
+}
+
+// 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
+}

internal/translator/kiro/common/constants.go

@@ -31,11 +31,23 @@ const (
 
 	// DefaultAssistantContentWithTools is the fallback content for assistant messages
 	// that have tool_use but no text content. Kiro API requires non-empty content.
-	DefaultAssistantContentWithTools = "I'll help you with that."
+	// IMPORTANT: Use a minimal neutral string that the model won't mimic in responses.
+	// Previously "I'll help you with that." which caused the model to parrot it back.
+	DefaultAssistantContentWithTools = "."
 
 	// DefaultAssistantContent is the fallback content for assistant messages
 	// that have no content at all. Kiro API requires non-empty content.
-	DefaultAssistantContent = "I understand."
+	// IMPORTANT: Use a minimal neutral string that the model won't mimic in responses.
+	// Previously "I understand." which could leak into model behavior.
+	DefaultAssistantContent = "."
+
+	// DefaultUserContentWithToolResults is the fallback content for user messages
+	// that have only tool_result (no text). Kiro API requires non-empty content.
+	DefaultUserContentWithToolResults = "Tool results provided."
+
+	// DefaultUserContent is the fallback content for user messages
+	// that have no content at all. Kiro API requires non-empty content.
+	DefaultUserContent = "Continue"
 
 	// KiroAgenticSystemPrompt is injected only for -agentic models to prevent timeouts on large writes.
 	// AWS Kiro API has a 2-3 minute timeout for large file write operations.

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
+}

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))
+	}
+}

internal/translator/openai/claude/openai_claude_request.go

@@ -75,6 +75,10 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 						out, _ = sjson.Set(out, "reasoning_effort", effort)
 					}
 				}
+			case "adaptive":
+				// Claude adaptive means "enable with max capacity"; keep it as highest level
+				// and let ApplyThinking normalize per target model capability.
+				out, _ = sjson.Set(out, "reasoning_effort", string(thinking.LevelXHigh))
 			case "disabled":
 				if effort, ok := thinking.ConvertBudgetToLevel(0); ok && effort != "" {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)

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

@@ -70,7 +70,7 @@ func ConvertOpenAIResponsesRequestToOpenAIChatCompletions(modelName string, inpu
 				if role == "developer" {
 					role = "user"
 				}
-				message := `{"role":"","content":""}`
+				message := `{"role":"","content":[]}`
 				message, _ = sjson.Set(message, "role", role)
 
 				if content := item.Get("content"); content.Exists() && content.IsArray() {
@@ -84,20 +84,16 @@ func ConvertOpenAIResponsesRequestToOpenAIChatCompletions(modelName string, inpu
 						}
 
 						switch contentType {
-						case "input_text":
+						case "input_text", "output_text":
 							text := contentItem.Get("text").String()
-							if messageContent != "" {
-								messageContent += "\n" + text
-							} else {
-								messageContent = text
-							}
-						case "output_text":
-							text := contentItem.Get("text").String()
-							if messageContent != "" {
-								messageContent += "\n" + text
-							} else {
-								messageContent = text
-							}
+							contentPart := `{"type":"text","text":""}`
+							contentPart, _ = sjson.Set(contentPart, "text", text)
+							message, _ = sjson.SetRaw(message, "content.-1", contentPart)
+						case "input_image":
+							imageURL := contentItem.Get("image_url").String()
+							contentPart := `{"type":"image_url","image_url":{"url":""}}`
+							contentPart, _ = sjson.Set(contentPart, "image_url.url", imageURL)
+							message, _ = sjson.SetRaw(message, "content.-1", contentPart)
 						}
 						return true
 					})

internal/util/gemini_schema.go

@@ -428,8 +428,9 @@ func flattenTypeArrays(jsonStr string) string {
 
 func removeUnsupportedKeywords(jsonStr string) string {
 	keywords := append(unsupportedConstraints,
-		"$schema", "$defs", "definitions", "const", "$ref", "additionalProperties",
-		"propertyNames", // Gemini doesn't support property name validation
+		"$schema", "$defs", "definitions", "const", "$ref", "$id", "additionalProperties",
+		"propertyNames", "patternProperties", // Gemini doesn't support these schema keywords
+		"enumTitles", "prefill", // Claude/OpenCode schema metadata fields unsupported by Gemini
 	)
 
 	deletePaths := make([]string, 0)

internal/util/gemini_schema_test.go

@@ -870,6 +870,57 @@ func TestCleanJSONSchemaForAntigravity_BooleanEnumToString(t *testing.T) {
 	}
 }
 
+func TestCleanJSONSchemaForGemini_RemovesGeminiUnsupportedMetadataFields(t *testing.T) {
+	input := `{
+		"$schema": "http://json-schema.org/draft-07/schema#",
+		"$id": "root-schema",
+		"type": "object",
+		"properties": {
+			"payload": {
+				"type": "object",
+				"prefill": "hello",
+				"properties": {
+					"mode": {
+						"type": "string",
+						"enum": ["a", "b"],
+						"enumTitles": ["A", "B"]
+					}
+				},
+				"patternProperties": {
+					"^x-": {"type": "string"}
+				}
+			},
+			"$id": {
+				"type": "string",
+				"description": "property name should not be removed"
+			}
+		}
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"payload": {
+				"type": "object",
+				"properties": {
+					"mode": {
+						"type": "string",
+						"enum": ["a", "b"],
+						"description": "Allowed: a, b"
+					}
+				}
+			},
+			"$id": {
+				"type": "string",
+				"description": "property name should not be removed"
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForGemini(input)
+	compareJSON(t, expected, result)
+}
+
 func TestRemoveExtensionFields(t *testing.T) {
 	tests := []struct {
 		name     string

internal/watcher/synthesizer/file.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"strconv"
 	"strings"
 	"time"
 
@@ -92,6 +93,9 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 			status = coreauth.StatusDisabled
 		}
 
+		// Read per-account excluded models from the OAuth JSON file
+		perAccountExcluded := extractExcludedModelsFromMetadata(metadata)
+
 		a := &coreauth.Auth{
 			ID:       id,
 			Provider: provider,
@@ -108,11 +112,23 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 			CreatedAt: now,
 			UpdatedAt: now,
 		}
-		ApplyAuthExcludedModelsMeta(a, cfg, nil, "oauth")
+		// Read priority from auth file
+		if rawPriority, ok := metadata["priority"]; ok {
+			switch v := rawPriority.(type) {
+			case float64:
+				a.Attributes["priority"] = strconv.Itoa(int(v))
+			case string:
+				priority := strings.TrimSpace(v)
+				if _, errAtoi := strconv.Atoi(priority); errAtoi == nil {
+					a.Attributes["priority"] = priority
+				}
+			}
+		}
+		ApplyAuthExcludedModelsMeta(a, cfg, perAccountExcluded, "oauth")
 		if provider == "gemini-cli" {
 			if virtuals := SynthesizeGeminiVirtualAuths(a, metadata, now); len(virtuals) > 0 {
 				for _, v := range virtuals {
-					ApplyAuthExcludedModelsMeta(v, cfg, nil, "oauth")
+					ApplyAuthExcludedModelsMeta(v, cfg, perAccountExcluded, "oauth")
 				}
 				out = append(out, a)
 				out = append(out, virtuals...)
@@ -167,6 +183,10 @@ func SynthesizeGeminiVirtualAuths(primary *coreauth.Auth, metadata map[string]an
 		if authPath != "" {
 			attrs["path"] = authPath
 		}
+		// Propagate priority from primary auth to virtual auths
+		if priorityVal, hasPriority := primary.Attributes["priority"]; hasPriority && priorityVal != "" {
+			attrs["priority"] = priorityVal
+		}
 		metadataCopy := map[string]any{
 			"email":             email,
 			"project_id":        projectID,
@@ -239,3 +259,40 @@ func buildGeminiVirtualID(baseID, projectID string) string {
 	replacer := strings.NewReplacer("/", "_", "\\", "_", " ", "_")
 	return fmt.Sprintf("%s::%s", baseID, replacer.Replace(project))
 }
+
+// extractExcludedModelsFromMetadata reads per-account excluded models from the OAuth JSON metadata.
+// Supports both "excluded_models" and "excluded-models" keys, and accepts both []string and []interface{}.
+func extractExcludedModelsFromMetadata(metadata map[string]any) []string {
+	if metadata == nil {
+		return nil
+	}
+	// Try both key formats
+	raw, ok := metadata["excluded_models"]
+	if !ok {
+		raw, ok = metadata["excluded-models"]
+	}
+	if !ok || raw == nil {
+		return nil
+	}
+	var stringSlice []string
+	switch v := raw.(type) {
+	case []string:
+		stringSlice = v
+	case []interface{}:
+		stringSlice = make([]string, 0, len(v))
+		for _, item := range v {
+			if s, ok := item.(string); ok {
+				stringSlice = append(stringSlice, s)
+			}
+		}
+	default:
+		return nil
+	}
+	result := make([]string, 0, len(stringSlice))
+	for _, s := range stringSlice {
+		if trimmed := strings.TrimSpace(s); trimmed != "" {
+			result = append(result, trimmed)
+		}
+	}
+	return result
+}

internal/watcher/synthesizer/file_test.go

@@ -297,6 +297,117 @@ func TestFileSynthesizer_Synthesize_PrefixValidation(t *testing.T) {
 	}
 }
 
+func TestFileSynthesizer_Synthesize_PriorityParsing(t *testing.T) {
+	tests := []struct {
+		name     string
+		priority any
+		want     string
+		hasValue bool
+	}{
+		{
+			name:     "string with spaces",
+			priority: " 10 ",
+			want:     "10",
+			hasValue: true,
+		},
+		{
+			name:     "number",
+			priority: 8,
+			want:     "8",
+			hasValue: true,
+		},
+		{
+			name:     "invalid string",
+			priority: "1x",
+			hasValue: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			tempDir := t.TempDir()
+			authData := map[string]any{
+				"type":     "claude",
+				"priority": tt.priority,
+			}
+			data, _ := json.Marshal(authData)
+			errWriteFile := os.WriteFile(filepath.Join(tempDir, "auth.json"), data, 0644)
+			if errWriteFile != nil {
+				t.Fatalf("failed to write auth file: %v", errWriteFile)
+			}
+
+			synth := NewFileSynthesizer()
+			ctx := &SynthesisContext{
+				Config:      &config.Config{},
+				AuthDir:     tempDir,
+				Now:         time.Now(),
+				IDGenerator: NewStableIDGenerator(),
+			}
+
+			auths, errSynthesize := synth.Synthesize(ctx)
+			if errSynthesize != nil {
+				t.Fatalf("unexpected error: %v", errSynthesize)
+			}
+			if len(auths) != 1 {
+				t.Fatalf("expected 1 auth, got %d", len(auths))
+			}
+
+			value, ok := auths[0].Attributes["priority"]
+			if tt.hasValue {
+				if !ok {
+					t.Fatal("expected priority attribute to be set")
+				}
+				if value != tt.want {
+					t.Fatalf("expected priority %q, got %q", tt.want, value)
+				}
+				return
+			}
+			if ok {
+				t.Fatalf("expected priority attribute to be absent, got %q", value)
+			}
+		})
+	}
+}
+
+func TestFileSynthesizer_Synthesize_OAuthExcludedModelsMerged(t *testing.T) {
+	tempDir := t.TempDir()
+	authData := map[string]any{
+		"type":            "claude",
+		"excluded_models": []string{"custom-model", "MODEL-B"},
+	}
+	data, _ := json.Marshal(authData)
+	errWriteFile := os.WriteFile(filepath.Join(tempDir, "auth.json"), data, 0644)
+	if errWriteFile != nil {
+		t.Fatalf("failed to write auth file: %v", errWriteFile)
+	}
+
+	synth := NewFileSynthesizer()
+	ctx := &SynthesisContext{
+		Config: &config.Config{
+			OAuthExcludedModels: map[string][]string{
+				"claude": {"shared", "model-b"},
+			},
+		},
+		AuthDir:     tempDir,
+		Now:         time.Now(),
+		IDGenerator: NewStableIDGenerator(),
+	}
+
+	auths, errSynthesize := synth.Synthesize(ctx)
+	if errSynthesize != nil {
+		t.Fatalf("unexpected error: %v", errSynthesize)
+	}
+	if len(auths) != 1 {
+		t.Fatalf("expected 1 auth, got %d", len(auths))
+	}
+
+	got := auths[0].Attributes["excluded_models"]
+	want := "custom-model,model-b,shared"
+	if got != want {
+		t.Fatalf("expected excluded_models %q, got %q", want, got)
+	}
+}
+
 func TestSynthesizeGeminiVirtualAuths_NilInputs(t *testing.T) {
 	now := time.Now()
 
@@ -533,6 +644,7 @@ func TestFileSynthesizer_Synthesize_MultiProjectGemini(t *testing.T) {
 		"type":       "gemini",
 		"email":      "multi@example.com",
 		"project_id": "project-a, project-b, project-c",
+		"priority":   " 10 ",
 	}
 	data, _ := json.Marshal(authData)
 	err := os.WriteFile(filepath.Join(tempDir, "gemini-multi.json"), data, 0644)
@@ -565,6 +677,9 @@ func TestFileSynthesizer_Synthesize_MultiProjectGemini(t *testing.T) {
 	if primary.Status != coreauth.StatusDisabled {
 		t.Errorf("expected primary status disabled, got %s", primary.Status)
 	}
+	if gotPriority := primary.Attributes["priority"]; gotPriority != "10" {
+		t.Errorf("expected primary priority 10, got %q", gotPriority)
+	}
 
 	// Remaining auths should be virtuals
 	for i := 1; i < 4; i++ {
@@ -575,6 +690,9 @@ func TestFileSynthesizer_Synthesize_MultiProjectGemini(t *testing.T) {
 		if v.Attributes["gemini_virtual_parent"] != primary.ID {
 			t.Errorf("expected virtual %d parent to be %s, got %s", i, primary.ID, v.Attributes["gemini_virtual_parent"])
 		}
+		if gotPriority := v.Attributes["priority"]; gotPriority != "10" {
+			t.Errorf("expected virtual %d priority 10, got %q", i, gotPriority)
+		}
 	}
 }
 

internal/watcher/synthesizer/helpers.go

@@ -53,6 +53,8 @@ func (g *StableIDGenerator) Next(kind string, parts ...string) (string, string)
 
 // ApplyAuthExcludedModelsMeta applies excluded models metadata to an auth entry.
 // It computes a hash of excluded models and sets the auth_kind attribute.
+// For OAuth entries, perKey (from the JSON file's excluded-models field) is merged
+// with the global oauth-excluded-models config for the provider.
 func ApplyAuthExcludedModelsMeta(auth *coreauth.Auth, cfg *config.Config, perKey []string, authKind string) {
 	if auth == nil || cfg == nil {
 		return
@@ -72,9 +74,13 @@ func ApplyAuthExcludedModelsMeta(auth *coreauth.Auth, cfg *config.Config, perKey
 	}
 	if authKindKey == "apikey" {
 		add(perKey)
-	} else if cfg.OAuthExcludedModels != nil {
-		providerKey := strings.ToLower(strings.TrimSpace(auth.Provider))
-		add(cfg.OAuthExcludedModels[providerKey])
+	} else {
+		// For OAuth: merge per-account excluded models with global provider-level exclusions
+		add(perKey)
+		if cfg.OAuthExcludedModels != nil {
+			providerKey := strings.ToLower(strings.TrimSpace(auth.Provider))
+			add(cfg.OAuthExcludedModels[providerKey])
+		}
 	}
 	combined := make([]string, 0, len(seen))
 	for k := range seen {
@@ -88,6 +94,10 @@ func ApplyAuthExcludedModelsMeta(auth *coreauth.Auth, cfg *config.Config, perKey
 	if hash != "" {
 		auth.Attributes["excluded_models_hash"] = hash
 	}
+	// Store the combined excluded models list so that routing can read it at runtime
+	if len(combined) > 0 {
+		auth.Attributes["excluded_models"] = strings.Join(combined, ",")
+	}
 	if authKind != "" {
 		auth.Attributes["auth_kind"] = authKind
 	}

internal/watcher/synthesizer/helpers_test.go

@@ -6,6 +6,7 @@ import (
 	"testing"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/watcher/diff"
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 )
 
@@ -200,6 +201,30 @@ func TestApplyAuthExcludedModelsMeta(t *testing.T) {
 	}
 }
 
+func TestApplyAuthExcludedModelsMeta_OAuthMergeWritesCombinedModels(t *testing.T) {
+	auth := &coreauth.Auth{
+		Provider:   "claude",
+		Attributes: make(map[string]string),
+	}
+	cfg := &config.Config{
+		OAuthExcludedModels: map[string][]string{
+			"claude": {"global-a", "shared"},
+		},
+	}
+
+	ApplyAuthExcludedModelsMeta(auth, cfg, []string{"per", "SHARED"}, "oauth")
+
+	const wantCombined = "global-a,per,shared"
+	if gotCombined := auth.Attributes["excluded_models"]; gotCombined != wantCombined {
+		t.Fatalf("expected excluded_models=%q, got %q", wantCombined, gotCombined)
+	}
+
+	expectedHash := diff.ComputeExcludedModelsHash([]string{"global-a", "per", "shared"})
+	if gotHash := auth.Attributes["excluded_models_hash"]; gotHash != expectedHash {
+		t.Fatalf("expected excluded_models_hash=%q, got %q", expectedHash, gotHash)
+	}
+}
+
 func TestAddConfigHeadersToAttrs(t *testing.T) {
 	tests := []struct {
 		name    string

sdk/access/errors.go

@@ -1,12 +1,90 @@
 package access
 
-import "errors"
-
-var (
-	// ErrNoCredentials indicates no recognizable credentials were supplied.
-	ErrNoCredentials = errors.New("access: no credentials provided")
-	// ErrInvalidCredential signals that supplied credentials were rejected by a provider.
-	ErrInvalidCredential = errors.New("access: invalid credential")
-	// ErrNotHandled tells the manager to continue trying other providers.
-	ErrNotHandled = errors.New("access: not handled")
+import (
+	"fmt"
+	"net/http"
+	"strings"
 )
+
+// AuthErrorCode classifies authentication failures.
+type AuthErrorCode string
+
+const (
+	AuthErrorCodeNoCredentials     AuthErrorCode = "no_credentials"
+	AuthErrorCodeInvalidCredential AuthErrorCode = "invalid_credential"
+	AuthErrorCodeNotHandled        AuthErrorCode = "not_handled"
+	AuthErrorCodeInternal          AuthErrorCode = "internal_error"
+)
+
+// AuthError carries authentication failure details and HTTP status.
+type AuthError struct {
+	Code       AuthErrorCode
+	Message    string
+	StatusCode int
+	Cause      error
+}
+
+func (e *AuthError) Error() string {
+	if e == nil {
+		return ""
+	}
+	message := strings.TrimSpace(e.Message)
+	if message == "" {
+		message = "authentication error"
+	}
+	if e.Cause != nil {
+		return fmt.Sprintf("%s: %v", message, e.Cause)
+	}
+	return message
+}
+
+func (e *AuthError) Unwrap() error {
+	if e == nil {
+		return nil
+	}
+	return e.Cause
+}
+
+// HTTPStatusCode returns a safe fallback for missing status codes.
+func (e *AuthError) HTTPStatusCode() int {
+	if e == nil || e.StatusCode <= 0 {
+		return http.StatusInternalServerError
+	}
+	return e.StatusCode
+}
+
+func newAuthError(code AuthErrorCode, message string, statusCode int, cause error) *AuthError {
+	return &AuthError{
+		Code:       code,
+		Message:    message,
+		StatusCode: statusCode,
+		Cause:      cause,
+	}
+}
+
+func NewNoCredentialsError() *AuthError {
+	return newAuthError(AuthErrorCodeNoCredentials, "Missing API key", http.StatusUnauthorized, nil)
+}
+
+func NewInvalidCredentialError() *AuthError {
+	return newAuthError(AuthErrorCodeInvalidCredential, "Invalid API key", http.StatusUnauthorized, nil)
+}
+
+func NewNotHandledError() *AuthError {
+	return newAuthError(AuthErrorCodeNotHandled, "authentication provider did not handle request", 0, nil)
+}
+
+func NewInternalAuthError(message string, cause error) *AuthError {
+	normalizedMessage := strings.TrimSpace(message)
+	if normalizedMessage == "" {
+		normalizedMessage = "Authentication service error"
+	}
+	return newAuthError(AuthErrorCodeInternal, normalizedMessage, http.StatusInternalServerError, cause)
+}
+
+func IsAuthErrorCode(authErr *AuthError, code AuthErrorCode) bool {
+	if authErr == nil {
+		return false
+	}
+	return authErr.Code == code
+}

sdk/access/manager.go

@@ -2,7 +2,6 @@ package access
 
 import (
 	"context"
-	"errors"
 	"net/http"
 	"sync"
 )
@@ -43,7 +42,7 @@ func (m *Manager) Providers() []Provider {
 }
 
 // Authenticate evaluates providers until one succeeds.
-func (m *Manager) Authenticate(ctx context.Context, r *http.Request) (*Result, error) {
+func (m *Manager) Authenticate(ctx context.Context, r *http.Request) (*Result, *AuthError) {
 	if m == nil {
 		return nil, nil
 	}
@@ -61,29 +60,29 @@ func (m *Manager) Authenticate(ctx context.Context, r *http.Request) (*Result, e
 		if provider == nil {
 			continue
 		}
-		res, err := provider.Authenticate(ctx, r)
-		if err == nil {
+		res, authErr := provider.Authenticate(ctx, r)
+		if authErr == nil {
 			return res, nil
 		}
-		if errors.Is(err, ErrNotHandled) {
+		if IsAuthErrorCode(authErr, AuthErrorCodeNotHandled) {
 			continue
 		}
-		if errors.Is(err, ErrNoCredentials) {
+		if IsAuthErrorCode(authErr, AuthErrorCodeNoCredentials) {
 			missing = true
 			continue
 		}
-		if errors.Is(err, ErrInvalidCredential) {
+		if IsAuthErrorCode(authErr, AuthErrorCodeInvalidCredential) {
 			invalid = true
 			continue
 		}
-		return nil, err
+		return nil, authErr
 	}
 
 	if invalid {
-		return nil, ErrInvalidCredential
+		return nil, NewInvalidCredentialError()
 	}
 	if missing {
-		return nil, ErrNoCredentials
+		return nil, NewNoCredentialsError()
 	}
-	return nil, ErrNoCredentials
+	return nil, NewNoCredentialsError()
 }

sdk/access/registry.go

@@ -2,17 +2,15 @@ package access
 
 import (
 	"context"
-	"fmt"
 	"net/http"
+	"strings"
 	"sync"
-
-	"github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
 )
 
 // Provider validates credentials for incoming requests.
 type Provider interface {
 	Identifier() string
-	Authenticate(ctx context.Context, r *http.Request) (*Result, error)
+	Authenticate(ctx context.Context, r *http.Request) (*Result, *AuthError)
 }
 
 // Result conveys authentication outcome.
@@ -22,66 +20,64 @@ type Result struct {
 	Metadata  map[string]string
 }
 
-// ProviderFactory builds a provider from configuration data.
-type ProviderFactory func(cfg *config.AccessProvider, root *config.SDKConfig) (Provider, error)
-
 var (
 	registryMu sync.RWMutex
-	registry   = make(map[string]ProviderFactory)
+	registry   = make(map[string]Provider)
+	order      []string
 )
 
-// RegisterProvider registers a provider factory for a given type identifier.
-func RegisterProvider(typ string, factory ProviderFactory) {
-	if typ == "" || factory == nil {
+// RegisterProvider registers a pre-built provider instance for a given type identifier.
+func RegisterProvider(typ string, provider Provider) {
+	normalizedType := strings.TrimSpace(typ)
+	if normalizedType == "" || provider == nil {
 		return
 	}
+
 	registryMu.Lock()
-	registry[typ] = factory
+	if _, exists := registry[normalizedType]; !exists {
+		order = append(order, normalizedType)
+	}
+	registry[normalizedType] = provider
 	registryMu.Unlock()
 }
 
-func BuildProvider(cfg *config.AccessProvider, root *config.SDKConfig) (Provider, error) {
-	if cfg == nil {
-		return nil, fmt.Errorf("access: nil provider config")
+// UnregisterProvider removes a provider by type identifier.
+func UnregisterProvider(typ string) {
+	normalizedType := strings.TrimSpace(typ)
+	if normalizedType == "" {
+		return
 	}
-	registryMu.RLock()
-	factory, ok := registry[cfg.Type]
-	registryMu.RUnlock()
-	if !ok {
-		return nil, fmt.Errorf("access: provider type %q is not registered", cfg.Type)
+	registryMu.Lock()
+	if _, exists := registry[normalizedType]; !exists {
+		registryMu.Unlock()
+		return
 	}
-	provider, err := factory(cfg, root)
-	if err != nil {
-		return nil, fmt.Errorf("access: failed to build provider %q: %w", cfg.Name, err)
-	}
-	return provider, nil
-}
-
-// BuildProviders constructs providers declared in configuration.
-func BuildProviders(root *config.SDKConfig) ([]Provider, error) {
-	if root == nil {
-		return nil, nil
-	}
-	providers := make([]Provider, 0, len(root.Access.Providers))
-	for i := range root.Access.Providers {
-		providerCfg := &root.Access.Providers[i]
-		if providerCfg.Type == "" {
+	delete(registry, normalizedType)
+	for index := range order {
+		if order[index] != normalizedType {
 			continue
 		}
-		provider, err := BuildProvider(providerCfg, root)
-		if err != nil {
-			return nil, err
+		order = append(order[:index], order[index+1:]...)
+		break
+	}
+	registryMu.Unlock()
+}
+
+// RegisteredProviders returns the global provider instances in registration order.
+func RegisteredProviders() []Provider {
+	registryMu.RLock()
+	if len(order) == 0 {
+		registryMu.RUnlock()
+		return nil
+	}
+	providers := make([]Provider, 0, len(order))
+	for _, providerType := range order {
+		provider, exists := registry[providerType]
+		if !exists || provider == nil {
+			continue
 		}
 		providers = append(providers, provider)
 	}
-	if len(providers) == 0 {
-		if inline := config.MakeInlineAPIKeyProvider(root.APIKeys); inline != nil {
-			provider, err := BuildProvider(inline, root)
-			if err != nil {
-				return nil, err
-			}
-			providers = append(providers, provider)
-		}
-	}
-	return providers, nil
+	registryMu.RUnlock()
+	return providers
 }

sdk/access/types.go

@@ -0,0 +1,47 @@
+package access
+
+// AccessConfig groups request authentication providers.
+type AccessConfig struct {
+	// Providers lists configured authentication providers.
+	Providers []AccessProvider `yaml:"providers,omitempty" json:"providers,omitempty"`
+}
+
+// AccessProvider describes a request authentication provider entry.
+type AccessProvider struct {
+	// Name is the instance identifier for the provider.
+	Name string `yaml:"name" json:"name"`
+
+	// Type selects the provider implementation registered via the SDK.
+	Type string `yaml:"type" json:"type"`
+
+	// SDK optionally names a third-party SDK module providing this provider.
+	SDK string `yaml:"sdk,omitempty" json:"sdk,omitempty"`
+
+	// APIKeys lists inline keys for providers that require them.
+	APIKeys []string `yaml:"api-keys,omitempty" json:"api-keys,omitempty"`
+
+	// Config passes provider-specific options to the implementation.
+	Config map[string]any `yaml:"config,omitempty" json:"config,omitempty"`
+}
+
+const (
+	// AccessProviderTypeConfigAPIKey is the built-in provider validating inline API keys.
+	AccessProviderTypeConfigAPIKey = "config-api-key"
+
+	// DefaultAccessProviderName is applied when no provider name is supplied.
+	DefaultAccessProviderName = "config-inline"
+)
+
+// MakeInlineAPIKeyProvider constructs an inline API key provider configuration.
+// It returns nil when no keys are supplied.
+func MakeInlineAPIKeyProvider(keys []string) *AccessProvider {
+	if len(keys) == 0 {
+		return nil
+	}
+	provider := &AccessProvider{
+		Name:    DefaultAccessProviderName,
+		Type:    AccessProviderTypeConfigAPIKey,
+		APIKeys: append([]string(nil), keys...),
+	}
+	return provider
+}

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

@@ -185,8 +185,7 @@ 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{

sdk/api/handlers/gemini/gemini_handlers.go

@@ -300,8 +300,7 @@ 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{

sdk/api/management.go

@@ -18,6 +18,7 @@ type ManagementTokenRequester interface {
 	RequestCodexToken(*gin.Context)
 	RequestAntigravityToken(*gin.Context)
 	RequestQwenToken(*gin.Context)
+	RequestKimiToken(*gin.Context)
 	RequestIFlowToken(*gin.Context)
 	RequestIFlowCookieToken(*gin.Context)
 	GetAuthStatus(c *gin.Context)
@@ -55,6 +56,10 @@ func (m *managementTokenRequester) RequestQwenToken(c *gin.Context) {
 	m.handler.RequestQwenToken(c)
 }
 
+func (m *managementTokenRequester) RequestKimiToken(c *gin.Context) {
+	m.handler.RequestKimiToken(c)
+}
+
 func (m *managementTokenRequester) RequestIFlowToken(c *gin.Context) {
 	m.handler.RequestIFlowToken(c)
 }

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.

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) {

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) {

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

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)
+			}
+		})
+	}
+}

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.

sdk/auth/kimi.go

@@ -0,0 +1,123 @@
+package auth
+
+import (
+	"context"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/kimi"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/browser"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	log "github.com/sirupsen/logrus"
+)
+
+// kimiRefreshLead is the duration before token expiry when refresh should occur.
+var kimiRefreshLead = 5 * time.Minute
+
+// KimiAuthenticator implements the OAuth device flow login for Kimi (Moonshot AI).
+type KimiAuthenticator struct{}
+
+// NewKimiAuthenticator constructs a new Kimi authenticator.
+func NewKimiAuthenticator() Authenticator {
+	return &KimiAuthenticator{}
+}
+
+// Provider returns the provider key for kimi.
+func (KimiAuthenticator) Provider() string {
+	return "kimi"
+}
+
+// RefreshLead returns the duration before token expiry when refresh should occur.
+// Kimi tokens expire and need to be refreshed before expiry.
+func (KimiAuthenticator) RefreshLead() *time.Duration {
+	return &kimiRefreshLead
+}
+
+// Login initiates the Kimi device flow authentication.
+func (a KimiAuthenticator) Login(ctx context.Context, cfg *config.Config, opts *LoginOptions) (*coreauth.Auth, error) {
+	if cfg == nil {
+		return nil, fmt.Errorf("cliproxy auth: configuration is required")
+	}
+	if opts == nil {
+		opts = &LoginOptions{}
+	}
+
+	authSvc := kimi.NewKimiAuth(cfg)
+
+	// Start the device flow
+	fmt.Println("Starting Kimi authentication...")
+	deviceCode, err := authSvc.StartDeviceFlow(ctx)
+	if err != nil {
+		return nil, fmt.Errorf("kimi: failed to start device flow: %w", err)
+	}
+
+	// Display the verification URL
+	verificationURL := deviceCode.VerificationURIComplete
+	if verificationURL == "" {
+		verificationURL = deviceCode.VerificationURI
+	}
+
+	fmt.Printf("\nTo authenticate, please visit:\n%s\n\n", verificationURL)
+	if deviceCode.UserCode != "" {
+		fmt.Printf("User code: %s\n\n", deviceCode.UserCode)
+	}
+
+	// Try to open the browser automatically
+	if !opts.NoBrowser {
+		if browser.IsAvailable() {
+			if errOpen := browser.OpenURL(verificationURL); errOpen != nil {
+				log.Warnf("Failed to open browser automatically: %v", errOpen)
+			} else {
+				fmt.Println("Browser opened automatically.")
+			}
+		}
+	}
+
+	fmt.Println("Waiting for authorization...")
+	if deviceCode.ExpiresIn > 0 {
+		fmt.Printf("(This will timeout in %d seconds if not authorized)\n", deviceCode.ExpiresIn)
+	}
+
+	// Wait for user authorization
+	authBundle, err := authSvc.WaitForAuthorization(ctx, deviceCode)
+	if err != nil {
+		return nil, fmt.Errorf("kimi: %w", err)
+	}
+
+	// Create the token storage
+	tokenStorage := authSvc.CreateTokenStorage(authBundle)
+
+	// Build metadata with token information
+	metadata := map[string]any{
+		"type":          "kimi",
+		"access_token":  authBundle.TokenData.AccessToken,
+		"refresh_token": authBundle.TokenData.RefreshToken,
+		"token_type":    authBundle.TokenData.TokenType,
+		"scope":         authBundle.TokenData.Scope,
+		"timestamp":     time.Now().UnixMilli(),
+	}
+
+	if authBundle.TokenData.ExpiresAt > 0 {
+		exp := time.Unix(authBundle.TokenData.ExpiresAt, 0).UTC().Format(time.RFC3339)
+		metadata["expired"] = exp
+	}
+	if strings.TrimSpace(authBundle.DeviceID) != "" {
+		metadata["device_id"] = strings.TrimSpace(authBundle.DeviceID)
+	}
+
+	// Generate a unique filename
+	fileName := fmt.Sprintf("kimi-%d.json", time.Now().UnixMilli())
+
+	fmt.Println("\nKimi authentication successful!")
+
+	return &coreauth.Auth{
+		ID:       fileName,
+		Provider: a.Provider(),
+		FileName: fileName,
+		Label:    "Kimi User",
+		Storage:  tokenStorage,
+		Metadata: metadata,
+	}, nil
+}

sdk/auth/qwen.go

@@ -27,8 +27,7 @@ func (a *QwenAuthenticator) Provider() string {
 }
 
 func (a *QwenAuthenticator) RefreshLead() *time.Duration {
-	d := 3 * time.Hour
-	return &d
+	return new(3 * time.Hour)
 }
 
 func (a *QwenAuthenticator) Login(ctx context.Context, cfg *config.Config, opts *LoginOptions) (*coreauth.Auth, error) {

sdk/auth/refresh_registry.go

@@ -14,6 +14,7 @@ func init() {
 	registerRefreshLead("gemini", func() Authenticator { return NewGeminiAuthenticator() })
 	registerRefreshLead("gemini-cli", func() Authenticator { return NewGeminiAuthenticator() })
 	registerRefreshLead("antigravity", func() Authenticator { return NewAntigravityAuthenticator() })
+	registerRefreshLead("kimi", func() Authenticator { return NewKimiAuthenticator() })
 	registerRefreshLead("kiro", func() Authenticator { return NewKiroAuthenticator() })
 	registerRefreshLead("github-copilot", func() Authenticator { return NewGitHubCopilotAuthenticator() })
 }

sdk/cliproxy/auth/conductor.go

@@ -599,14 +599,16 @@ func (m *Manager) executeMixedOnce(ctx context.Context, providers []string, req
 				return cliproxyexecutor.Response{}, errCtx
 			}
 			result.Error = &Error{Message: errExec.Error()}
-			var se cliproxyexecutor.StatusError
-			if errors.As(errExec, &se) && se != nil {
+			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
 				result.Error.HTTPStatus = se.StatusCode()
 			}
 			if ra := retryAfterFromError(errExec); ra != nil {
 				result.RetryAfter = ra
 			}
 			m.MarkResult(execCtx, result)
+			if isRequestInvalidError(errExec) {
+				return cliproxyexecutor.Response{}, errExec
+			}
 			lastErr = errExec
 			continue
 		}
@@ -652,14 +654,16 @@ func (m *Manager) executeCountMixedOnce(ctx context.Context, providers []string,
 				return cliproxyexecutor.Response{}, errCtx
 			}
 			result.Error = &Error{Message: errExec.Error()}
-			var se cliproxyexecutor.StatusError
-			if errors.As(errExec, &se) && se != nil {
+			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
 				result.Error.HTTPStatus = se.StatusCode()
 			}
 			if ra := retryAfterFromError(errExec); ra != nil {
 				result.RetryAfter = ra
 			}
 			m.MarkResult(execCtx, result)
+			if isRequestInvalidError(errExec) {
+				return cliproxyexecutor.Response{}, errExec
+			}
 			lastErr = errExec
 			continue
 		}
@@ -704,13 +708,15 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 				return nil, errCtx
 			}
 			rerr := &Error{Message: errStream.Error()}
-			var se cliproxyexecutor.StatusError
-			if errors.As(errStream, &se) && se != nil {
+			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errStream); ok && se != nil {
 				rerr.HTTPStatus = se.StatusCode()
 			}
 			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr}
 			result.RetryAfter = retryAfterFromError(errStream)
 			m.MarkResult(execCtx, result)
+			if isRequestInvalidError(errStream) {
+				return nil, errStream
+			}
 			lastErr = errStream
 			continue
 		}
@@ -723,8 +729,7 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 				if chunk.Err != nil && !failed {
 					failed = true
 					rerr := &Error{Message: chunk.Err.Error()}
-					var se cliproxyexecutor.StatusError
-					if errors.As(chunk.Err, &se) && se != nil {
+					if se, ok := errors.AsType[cliproxyexecutor.StatusError](chunk.Err); ok && se != nil {
 						rerr.HTTPStatus = se.StatusCode()
 					}
 					m.MarkResult(streamCtx, Result{AuthID: streamAuth.ID, Provider: streamProvider, Model: routeModel, Success: false, Error: rerr})
@@ -1110,6 +1115,9 @@ func (m *Manager) shouldRetryAfterError(err error, attempt int, providers []stri
 	if status := statusCodeFromError(err); status == http.StatusOK {
 		return 0, false
 	}
+	if isRequestInvalidError(err) {
+		return 0, false
+	}
 	wait, found := m.closestCooldownWait(providers, model, attempt)
 	if !found || wait > maxWait {
 		return 0, false
@@ -1299,7 +1307,7 @@ func updateAggregatedAvailability(auth *Auth, now time.Time) {
 			stateUnavailable = true
 		} else if state.Unavailable {
 			if state.NextRetryAfter.IsZero() {
-				stateUnavailable = true
+				stateUnavailable = false
 			} else if state.NextRetryAfter.After(now) {
 				stateUnavailable = true
 				if earliestRetry.IsZero() || state.NextRetryAfter.Before(earliestRetry) {
@@ -1419,8 +1427,7 @@ func retryAfterFromError(err error) *time.Duration {
 	if retryAfter == nil {
 		return nil
 	}
-	val := *retryAfter
-	return &val
+	return new(*retryAfter)
 }
 
 func statusCodeFromResult(err *Error) int {
@@ -1430,6 +1437,21 @@ func statusCodeFromResult(err *Error) int {
 	return err.StatusCode()
 }
 
+// isRequestInvalidError returns true if the error represents a client request
+// error that should not be retried. Specifically, it checks for 400 Bad Request
+// with "invalid_request_error" in the message, indicating the request itself is
+// malformed and switching to a different auth will not help.
+func isRequestInvalidError(err error) bool {
+	if err == nil {
+		return false
+	}
+	status := statusCodeFromError(err)
+	if status != http.StatusBadRequest {
+		return false
+	}
+	return strings.Contains(err.Error(), "invalid_request_error")
+}
+
 func applyAuthFailureState(auth *Auth, resultErr *Error, retryAfter *time.Duration, now time.Time) {
 	if auth == nil {
 		return

sdk/cliproxy/auth/conductor_availability_test.go

@@ -0,0 +1,61 @@
+package auth
+
+import (
+	"testing"
+	"time"
+)
+
+func TestUpdateAggregatedAvailability_UnavailableWithoutNextRetryDoesNotBlockAuth(t *testing.T) {
+	t.Parallel()
+
+	now := time.Now()
+	model := "test-model"
+	auth := &Auth{
+		ID: "a",
+		ModelStates: map[string]*ModelState{
+			model: {
+				Status:      StatusError,
+				Unavailable: true,
+			},
+		},
+	}
+
+	updateAggregatedAvailability(auth, now)
+
+	if auth.Unavailable {
+		t.Fatalf("auth.Unavailable = true, want false")
+	}
+	if !auth.NextRetryAfter.IsZero() {
+		t.Fatalf("auth.NextRetryAfter = %v, want zero", auth.NextRetryAfter)
+	}
+}
+
+func TestUpdateAggregatedAvailability_FutureNextRetryBlocksAuth(t *testing.T) {
+	t.Parallel()
+
+	now := time.Now()
+	model := "test-model"
+	next := now.Add(5 * time.Minute)
+	auth := &Auth{
+		ID: "a",
+		ModelStates: map[string]*ModelState{
+			model: {
+				Status:         StatusError,
+				Unavailable:    true,
+				NextRetryAfter: next,
+			},
+		},
+	}
+
+	updateAggregatedAvailability(auth, now)
+
+	if !auth.Unavailable {
+		t.Fatalf("auth.Unavailable = false, want true")
+	}
+	if auth.NextRetryAfter.IsZero() {
+		t.Fatalf("auth.NextRetryAfter = zero, want %v", next)
+	}
+	if auth.NextRetryAfter.Sub(next) > time.Second || next.Sub(auth.NextRetryAfter) > time.Second {
+		t.Fatalf("auth.NextRetryAfter = %v, want %v", auth.NextRetryAfter, next)
+	}
+}

sdk/cliproxy/auth/oauth_model_alias.go

@@ -221,7 +221,7 @@ func modelAliasChannel(auth *Auth) string {
 // and auth kind. Returns empty string if the provider/authKind combination doesn't support
 // OAuth model alias (e.g., API key authentication).
 //
-// Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow, kiro, github-copilot.
+// Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow, kiro, github-copilot, kimi.
 func OAuthModelAliasChannel(provider, authKind string) string {
 	provider = strings.ToLower(strings.TrimSpace(provider))
 	authKind = strings.ToLower(strings.TrimSpace(authKind))
@@ -245,7 +245,7 @@ func OAuthModelAliasChannel(provider, authKind string) string {
 			return ""
 		}
 		return "codex"
-	case "gemini-cli", "aistudio", "antigravity", "qwen", "iflow", "kiro", "github-copilot":
+	case "gemini-cli", "aistudio", "antigravity", "qwen", "iflow", "kiro", "github-copilot", "kimi":
 		return provider
 	default:
 		return ""

sdk/cliproxy/auth/oauth_model_alias_test.go

@@ -79,6 +79,33 @@ func TestResolveOAuthUpstreamModel_SuffixPreservation(t *testing.T) {
 			input:   "gemini-2.5-pro(none)",
 			want:    "gemini-2.5-pro-exp-03-25(none)",
 		},
+		{
+			name: "github-copilot suffix preserved",
+			aliases: map[string][]internalconfig.OAuthModelAlias{
+				"github-copilot": {{Name: "claude-opus-4.6", Alias: "opus"}},
+			},
+			channel: "github-copilot",
+			input:   "opus(medium)",
+			want:    "claude-opus-4.6(medium)",
+		},
+		{
+			name: "github-copilot no suffix",
+			aliases: map[string][]internalconfig.OAuthModelAlias{
+				"github-copilot": {{Name: "claude-opus-4.6", Alias: "opus"}},
+			},
+			channel: "github-copilot",
+			input:   "opus",
+			want:    "claude-opus-4.6",
+		},
+		{
+			name: "kimi suffix preserved",
+			aliases: map[string][]internalconfig.OAuthModelAlias{
+				"kimi": {{Name: "kimi-k2.5", Alias: "k2.5"}},
+			},
+			channel: "kimi",
+			input:   "k2.5(high)",
+			want:    "kimi-k2.5(high)",
+		},
 		{
 			name: "case insensitive alias lookup with suffix",
 			aliases: map[string][]internalconfig.OAuthModelAlias{
@@ -161,13 +188,41 @@ func createAuthForChannel(channel string) *Auth {
 		return &Auth{Provider: "qwen"}
 	case "iflow":
 		return &Auth{Provider: "iflow"}
+	case "kimi":
+		return &Auth{Provider: "kimi"}
 	case "kiro":
 		return &Auth{Provider: "kiro"}
+	case "github-copilot":
+		return &Auth{Provider: "github-copilot"}
 	default:
 		return &Auth{Provider: channel}
 	}
 }
 
+func TestOAuthModelAliasChannel_Kimi(t *testing.T) {
+	t.Parallel()
+
+	if got := OAuthModelAliasChannel("kimi", "oauth"); got != "kimi" {
+		t.Fatalf("OAuthModelAliasChannel() = %q, want %q", got, "kimi")
+	}
+}
+
+func TestOAuthModelAliasChannel_GitHubCopilot(t *testing.T) {
+	t.Parallel()
+
+	if got := OAuthModelAliasChannel("github-copilot", ""); got != "github-copilot" {
+		t.Fatalf("OAuthModelAliasChannel() = %q, want %q", got, "github-copilot")
+	}
+}
+
+func TestOAuthModelAliasChannel_Kiro(t *testing.T) {
+	t.Parallel()
+
+	if got := OAuthModelAliasChannel("kiro", ""); got != "kiro" {
+		t.Fatalf("OAuthModelAliasChannel() = %q, want %q", got, "kiro")
+	}
+}
+
 func TestApplyOAuthModelAlias_SuffixPreservation(t *testing.T) {
 	t.Parallel()
 

sdk/cliproxy/auth/selector.go

@@ -12,6 +12,7 @@ import (
 	"sync"
 	"time"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 )
 
@@ -19,6 +20,7 @@ import (
 type RoundRobinSelector struct {
 	mu      sync.Mutex
 	cursors map[string]int
+	maxKeys int
 }
 
 // FillFirstSelector selects the first available credential (deterministic ordering).
@@ -119,6 +121,19 @@ func authPriority(auth *Auth) int {
 	return parsed
 }
 
+func canonicalModelKey(model string) string {
+	model = strings.TrimSpace(model)
+	if model == "" {
+		return ""
+	}
+	parsed := thinking.ParseSuffix(model)
+	modelName := strings.TrimSpace(parsed.ModelName)
+	if modelName == "" {
+		return model
+	}
+	return modelName
+}
+
 func collectAvailableByPriority(auths []*Auth, model string, now time.Time) (available map[int][]*Auth, cooldownCount int, earliest time.Time) {
 	available = make(map[int][]*Auth)
 	for i := 0; i < len(auths); i++ {
@@ -185,11 +200,18 @@ func (s *RoundRobinSelector) Pick(ctx context.Context, provider, model string, o
 	if err != nil {
 		return nil, err
 	}
-	key := provider + ":" + model
+	key := provider + ":" + canonicalModelKey(model)
 	s.mu.Lock()
 	if s.cursors == nil {
 		s.cursors = make(map[string]int)
 	}
+	limit := s.maxKeys
+	if limit <= 0 {
+		limit = 4096
+	}
+	if _, ok := s.cursors[key]; !ok && len(s.cursors) >= limit {
+		s.cursors = make(map[string]int)
+	}
 	index := s.cursors[key]
 
 	if index >= 2_147_483_640 {
@@ -223,7 +245,14 @@ func isAuthBlockedForModel(auth *Auth, model string, now time.Time) (bool, block
 	}
 	if model != "" {
 		if len(auth.ModelStates) > 0 {
-			if state, ok := auth.ModelStates[model]; ok && state != nil {
+			state, ok := auth.ModelStates[model]
+			if (!ok || state == nil) && model != "" {
+				baseModel := canonicalModelKey(model)
+				if baseModel != "" && baseModel != model {
+					state, ok = auth.ModelStates[baseModel]
+				}
+			}
+			if ok && state != nil {
 				if state.Status == StatusDisabled {
 					return true, blockReasonDisabled, time.Time{}
 				}

sdk/cliproxy/auth/selector_test.go

@@ -2,7 +2,9 @@ package auth
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
+	"net/http"
 	"sync"
 	"testing"
 	"time"
@@ -175,3 +177,228 @@ func TestRoundRobinSelectorPick_Concurrent(t *testing.T) {
 	default:
 	}
 }
+
+func TestSelectorPick_AllCooldownReturnsModelCooldownError(t *testing.T) {
+	t.Parallel()
+
+	model := "test-model"
+	now := time.Now()
+	next := now.Add(60 * time.Second)
+	auths := []*Auth{
+		{
+			ID: "a",
+			ModelStates: map[string]*ModelState{
+				model: {
+					Status:         StatusActive,
+					Unavailable:    true,
+					NextRetryAfter: next,
+					Quota: QuotaState{
+						Exceeded:      true,
+						NextRecoverAt: next,
+					},
+				},
+			},
+		},
+		{
+			ID: "b",
+			ModelStates: map[string]*ModelState{
+				model: {
+					Status:         StatusActive,
+					Unavailable:    true,
+					NextRetryAfter: next,
+					Quota: QuotaState{
+						Exceeded:      true,
+						NextRecoverAt: next,
+					},
+				},
+			},
+		},
+	}
+
+	t.Run("mixed provider redacts provider field", func(t *testing.T) {
+		t.Parallel()
+
+		selector := &FillFirstSelector{}
+		_, err := selector.Pick(context.Background(), "mixed", model, cliproxyexecutor.Options{}, auths)
+		if err == nil {
+			t.Fatalf("Pick() error = nil")
+		}
+
+		var mce *modelCooldownError
+		if !errors.As(err, &mce) {
+			t.Fatalf("Pick() error = %T, want *modelCooldownError", err)
+		}
+		if mce.StatusCode() != http.StatusTooManyRequests {
+			t.Fatalf("StatusCode() = %d, want %d", mce.StatusCode(), http.StatusTooManyRequests)
+		}
+
+		headers := mce.Headers()
+		if got := headers.Get("Retry-After"); got == "" {
+			t.Fatalf("Headers().Get(Retry-After) = empty")
+		}
+
+		var payload map[string]any
+		if err := json.Unmarshal([]byte(mce.Error()), &payload); err != nil {
+			t.Fatalf("json.Unmarshal(Error()) error = %v", err)
+		}
+		rawErr, ok := payload["error"].(map[string]any)
+		if !ok {
+			t.Fatalf("Error() payload missing error object: %v", payload)
+		}
+		if got, _ := rawErr["code"].(string); got != "model_cooldown" {
+			t.Fatalf("Error().error.code = %q, want %q", got, "model_cooldown")
+		}
+		if _, ok := rawErr["provider"]; ok {
+			t.Fatalf("Error().error.provider exists for mixed provider: %v", rawErr["provider"])
+		}
+	})
+
+	t.Run("non-mixed provider includes provider field", func(t *testing.T) {
+		t.Parallel()
+
+		selector := &FillFirstSelector{}
+		_, err := selector.Pick(context.Background(), "gemini", model, cliproxyexecutor.Options{}, auths)
+		if err == nil {
+			t.Fatalf("Pick() error = nil")
+		}
+
+		var mce *modelCooldownError
+		if !errors.As(err, &mce) {
+			t.Fatalf("Pick() error = %T, want *modelCooldownError", err)
+		}
+
+		var payload map[string]any
+		if err := json.Unmarshal([]byte(mce.Error()), &payload); err != nil {
+			t.Fatalf("json.Unmarshal(Error()) error = %v", err)
+		}
+		rawErr, ok := payload["error"].(map[string]any)
+		if !ok {
+			t.Fatalf("Error() payload missing error object: %v", payload)
+		}
+		if got, _ := rawErr["provider"].(string); got != "gemini" {
+			t.Fatalf("Error().error.provider = %q, want %q", got, "gemini")
+		}
+	})
+}
+
+func TestIsAuthBlockedForModel_UnavailableWithoutNextRetryIsNotBlocked(t *testing.T) {
+	t.Parallel()
+
+	now := time.Now()
+	model := "test-model"
+	auth := &Auth{
+		ID: "a",
+		ModelStates: map[string]*ModelState{
+			model: {
+				Status:      StatusActive,
+				Unavailable: true,
+				Quota: QuotaState{
+					Exceeded: true,
+				},
+			},
+		},
+	}
+
+	blocked, reason, next := isAuthBlockedForModel(auth, model, now)
+	if blocked {
+		t.Fatalf("blocked = true, want false")
+	}
+	if reason != blockReasonNone {
+		t.Fatalf("reason = %v, want %v", reason, blockReasonNone)
+	}
+	if !next.IsZero() {
+		t.Fatalf("next = %v, want zero", next)
+	}
+}
+
+func TestFillFirstSelectorPick_ThinkingSuffixFallsBackToBaseModelState(t *testing.T) {
+	t.Parallel()
+
+	selector := &FillFirstSelector{}
+	now := time.Now()
+
+	baseModel := "test-model"
+	requestedModel := "test-model(high)"
+
+	high := &Auth{
+		ID:         "high",
+		Attributes: map[string]string{"priority": "10"},
+		ModelStates: map[string]*ModelState{
+			baseModel: {
+				Status:         StatusActive,
+				Unavailable:    true,
+				NextRetryAfter: now.Add(30 * time.Minute),
+				Quota: QuotaState{
+					Exceeded: true,
+				},
+			},
+		},
+	}
+	low := &Auth{
+		ID:         "low",
+		Attributes: map[string]string{"priority": "0"},
+	}
+
+	got, err := selector.Pick(context.Background(), "mixed", requestedModel, cliproxyexecutor.Options{}, []*Auth{high, low})
+	if err != nil {
+		t.Fatalf("Pick() error = %v", err)
+	}
+	if got == nil {
+		t.Fatalf("Pick() auth = nil")
+	}
+	if got.ID != "low" {
+		t.Fatalf("Pick() auth.ID = %q, want %q", got.ID, "low")
+	}
+}
+
+func TestRoundRobinSelectorPick_ThinkingSuffixSharesCursor(t *testing.T) {
+	t.Parallel()
+
+	selector := &RoundRobinSelector{}
+	auths := []*Auth{
+		{ID: "b"},
+		{ID: "a"},
+	}
+
+	first, err := selector.Pick(context.Background(), "gemini", "test-model(high)", cliproxyexecutor.Options{}, auths)
+	if err != nil {
+		t.Fatalf("Pick() first error = %v", err)
+	}
+	second, err := selector.Pick(context.Background(), "gemini", "test-model(low)", cliproxyexecutor.Options{}, auths)
+	if err != nil {
+		t.Fatalf("Pick() second error = %v", err)
+	}
+	if first == nil || second == nil {
+		t.Fatalf("Pick() returned nil auth")
+	}
+	if first.ID != "a" {
+		t.Fatalf("Pick() first auth.ID = %q, want %q", first.ID, "a")
+	}
+	if second.ID != "b" {
+		t.Fatalf("Pick() second auth.ID = %q, want %q", second.ID, "b")
+	}
+}
+
+func TestRoundRobinSelectorPick_CursorKeyCap(t *testing.T) {
+	t.Parallel()
+
+	selector := &RoundRobinSelector{maxKeys: 2}
+	auths := []*Auth{{ID: "a"}}
+
+	_, _ = selector.Pick(context.Background(), "gemini", "m1", cliproxyexecutor.Options{}, auths)
+	_, _ = selector.Pick(context.Background(), "gemini", "m2", cliproxyexecutor.Options{}, auths)
+	_, _ = selector.Pick(context.Background(), "gemini", "m3", cliproxyexecutor.Options{}, auths)
+
+	selector.mu.Lock()
+	defer selector.mu.Unlock()
+
+	if selector.cursors == nil {
+		t.Fatalf("selector.cursors = nil")
+	}
+	if len(selector.cursors) != 1 {
+		t.Fatalf("len(selector.cursors) = %d, want %d", len(selector.cursors), 1)
+	}
+	if _, ok := selector.cursors["gemini:m3"]; !ok {
+		t.Fatalf("selector.cursors missing key %q", "gemini:m3")
+	}
+}