feat(translator): support Claude thinking type adaptive

fix(amp): rewrite response.model in Responses API SSE events

README.md

@@ -27,10 +27,6 @@ Get 10% OFF GLM CODING PLAN：https://z.ai/subscribe?ic=8JVLJQFSKB
 <td>Thanks to PackyCode for sponsoring this project! PackyCode is a reliable and efficient API relay service provider, offering relay services for Claude Code, Codex, Gemini, and more. PackyCode provides special discounts for our software users: register using <a href="https://www.packyapi.com/register?aff=cliproxyapi">this link</a> and enter the "cliproxyapi" promo code during recharge to get 10% off.</td>
 </tr>
 <tr>
-<td width="180"><a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa"><img src="./assets/cubence.png" alt="Cubence" width="150"></a></td>
-<td>Thanks to Cubence for sponsoring this project! Cubence is a reliable and efficient API relay service provider, offering relay services for Claude Code, Codex, Gemini, and more. Cubence provides special discounts for our software users: register using <a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa">this link</a> and enter the "CLIPROXYAPI" promo code during recharge to get 10% off.</td>
-</tr>
-<tr>
 <td width="180"><a href="https://www.aicodemirror.com/register?invitecode=TJNAIF"><img src="./assets/aicodemirror.png" alt="AICodeMirror" width="150"></a></td>
 <td>Thanks to AICodeMirror for sponsoring this project! AICodeMirror provides official high-stability relay services for Claude Code / Codex / Gemini CLI, with enterprise-grade concurrency, fast invoicing, and 24/7 dedicated technical support. Claude Code / Codex / Gemini official channels at 38% / 2% / 9% of original price, with extra discounts on top-ups! AICodeMirror offers special benefits for CLIProxyAPI users: register via <a href="https://www.aicodemirror.com/register?invitecode=TJNAIF">this link</a> to enjoy 20% off your first top-up, and enterprise customers can get up to 25% off!</td>
 </tr>

README_CN.md

@@ -27,10 +27,6 @@ GLM CODING PLAN 是专为AI编码打造的订阅套餐，每月最低仅需20元
 <td>感谢 PackyCode 对本项目的赞助！PackyCode 是一家可靠高效的 API 中转服务商，提供 Claude Code、Codex、Gemini 等多种服务的中转。PackyCode 为本软件用户提供了特别优惠：使用<a href="https://www.packyapi.com/register?aff=cliproxyapi">此链接</a>注册，并在充值时输入 "cliproxyapi" 优惠码即可享受九折优惠。</td>
 </tr>
 <tr>
-<td width="180"><a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa"><img src="./assets/cubence.png" alt="Cubence" width="150"></a></td>
-<td>感谢 Cubence 对本项目的赞助！Cubence 是一家可靠高效的 API 中转服务商，提供 Claude Code、Codex、Gemini 等多种服务的中转。Cubence 为本软件用户提供了特别优惠：使用<a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa">此链接</a>注册，并在充值时输入 "CLIPROXYAPI" 优惠码即可享受九折优惠。</td>
-</tr>
-<tr>
 <td width="180"><a href="https://www.aicodemirror.com/register?invitecode=TJNAIF"><img src="./assets/aicodemirror.png" alt="AICodeMirror" width="150"></a></td>
 <td>感谢 AICodeMirror 赞助了本项目！AICodeMirror 提供 Claude Code / Codex / Gemini CLI 官方高稳定中转服务，支持企业级高并发、极速开票、7×24 专属技术支持。 Claude Code / Codex / Gemini 官方渠道低至 3.8 / 0.2 / 0.9 折，充值更有折上折！AICodeMirror 为 CLIProxyAPI 的用户提供了特别福利，通过<a href="https://www.aicodemirror.com/register?invitecode=TJNAIF">此链接</a>注册的用户，可享受首充8折，企业客户最高可享 7.5 折！</td>
 </tr>

config.example.yaml

@@ -221,25 +221,10 @@ 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.
+# Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow, 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:
-  antigravity:
-    - name: "rev19-uic3-1p"
-      alias: "gemini-2.5-computer-use-preview-10-2025"
-    - name: "gemini-3-pro-image"
-      alias: "gemini-3-pro-image-preview"
-    - name: "gemini-3-pro-high"
-      alias: "gemini-3-pro-preview"
-    - name: "gemini-3-flash"
-      alias: "gemini-3-flash-preview"
-    - name: "claude-sonnet-4-5"
-      alias: "gemini-claude-sonnet-4-5"
-    - name: "claude-sonnet-4-5-thinking"
-      alias: "gemini-claude-sonnet-4-5-thinking"
-    - name: "claude-opus-4-5-thinking"
-      alias: "gemini-claude-opus-4-5-thinking"
+# oauth-model-alias:
 #   gemini-cli:
 #     - name: "gemini-2.5-pro"          # original model name under this channel
 #       alias: "g2.5p"                  # client-visible alias
@@ -250,6 +235,9 @@ oauth-model-alias:
 #   aistudio:
 #     - name: "gemini-2.5-pro"
 #       alias: "g2.5p"
+#   antigravity:
+#     - name: "gemini-3-pro-high"
+#       alias: "gemini-3-pro-preview"
 #   claude:
 #     - name: "claude-sonnet-4-5-20250929"
 #       alias: "cs4.5"
@@ -262,6 +250,9 @@ oauth-model-alias:
 #   iflow:
 #     - name: "glm-4.7"
 #       alias: "glm-god"
+#   kimi:
+#     - name: "kimi-k2.5"
+#       alias: "k2.5"
 
 # OAuth provider excluded models
 # oauth-excluded-models:
@@ -284,6 +275,8 @@ oauth-model-alias:
 #     - "vision-model"
 #   iflow:
 #     - "tstars2.0"
+#   kimi:
+#     - "kimi-k2-thinking"
 
 # Optional payload configuration
 # payload:

go.mod

@@ -22,6 +22,7 @@ require (
 	golang.org/x/crypto v0.45.0
 	golang.org/x/net v0.47.0
 	golang.org/x/oauth2 v0.30.0
+	golang.org/x/sync v0.18.0
 	gopkg.in/natefinch/lumberjack.v2 v2.2.1
 	gopkg.in/yaml.v3 v3.0.1
 )
@@ -69,7 +70,6 @@ require (
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.12 // indirect
 	golang.org/x/arch v0.8.0 // indirect
-	golang.org/x/sync v0.18.0 // indirect
 	golang.org/x/sys v0.38.0 // indirect
 	golang.org/x/text v0.31.0 // indirect
 	google.golang.org/protobuf v1.34.1 // indirect

internal/api/modules/amp/response_rewriter.go

@@ -66,7 +66,7 @@ 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

@@ -655,14 +655,17 @@ 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)
@@ -952,10 +955,6 @@ 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)

internal/auth/kimi/kimi.go

@@ -30,7 +30,7 @@ const (
 	// 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/v1"
+	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.

internal/config/config.go

@@ -1098,8 +1098,13 @@ 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
 	}
@@ -1113,16 +1118,19 @@ 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)
 		}
 	}
 }
@@ -1130,7 +1138,12 @@ 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
 	}
@@ -1139,7 +1152,7 @@ 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 {
@@ -1162,7 +1175,7 @@ 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])
@@ -1204,6 +1217,94 @@ 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/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/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

@@ -20,11 +20,13 @@ import (
 	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
 }
 
@@ -64,6 +66,12 @@ func (e *KimiExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth,
 
 // 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)
@@ -71,12 +79,12 @@ func (e *KimiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
-	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		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)
 
@@ -94,8 +102,12 @@ func (e *KimiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 
 	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 + "/chat/completions"
+	url := kimiauth.KimiAPIBaseURL + "/v1/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
 	if err != nil {
 		return resp, err
@@ -148,26 +160,31 @@ func (e *KimiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	var param any
 	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
 	// the original model name in the response for client compatibility.
-	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
+	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) {
-	baseModel := thinking.ParseSuffix(req.Model).ModelName
+	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)
 
-	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		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)
 
@@ -189,8 +206,12 @@ func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	}
 	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 + "/chat/completions"
+	url := kimiauth.KimiAPIBaseURL + "/v1/chat/completions"
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
 	if err != nil {
 		return nil, err
@@ -249,12 +270,12 @@ func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 			if detail, ok := parseOpenAIStreamUsage(line); ok {
 				reporter.publish(ctx, detail)
 			}
-			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(line), &param)
+			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, 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, bytes.Clone(opts.OriginalRequest), body, bytes.Clone([]byte("[DONE]")), &param)
+		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])}
 		}
@@ -269,26 +290,152 @@ func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 
 // 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) {
-	baseModel := thinking.ParseSuffix(req.Model).ModelName
+	auth.Attributes["base_url"] = kimiauth.KimiAPIBaseURL
+	return e.ClaudeExecutor.CountTokens(ctx, auth, req, opts)
+}
 
-	from := opts.SourceFormat
-	to := sdktranslator.FromString("openai")
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
-
-	// Use a generic tokenizer for estimation
-	enc, err := tokenizerForModel("gpt-4")
-	if err != nil {
-		return cliproxyexecutor.Response{}, fmt.Errorf("kimi executor: tokenizer init failed: %w", err)
+func normalizeKimiToolMessageLinks(body []byte) ([]byte, error) {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		return body, nil
 	}
 
-	count, err := countOpenAIChatTokens(enc, body)
-	if err != nil {
-		return cliproxyexecutor.Response{}, fmt.Errorf("kimi executor: token counting failed: %w", err)
+	messages := gjson.GetBytes(body, "messages")
+	if !messages.Exists() || !messages.IsArray() {
+		return body, nil
 	}
 
-	usageJSON := buildOpenAIUsageJSON(count)
-	translated := sdktranslator.TranslateTokenCount(ctx, to, from, count, usageJSON)
-	return cliproxyexecutor.Response{Payload: []byte(translated)}, 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.

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/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/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/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/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
 					})

sdk/cliproxy/auth/conductor.go

@@ -607,6 +607,9 @@ func (m *Manager) executeMixedOnce(ctx context.Context, providers []string, req
 				result.RetryAfter = ra
 			}
 			m.MarkResult(execCtx, result)
+			if isRequestInvalidError(errExec) {
+				return cliproxyexecutor.Response{}, errExec
+			}
 			lastErr = errExec
 			continue
 		}
@@ -660,6 +663,9 @@ func (m *Manager) executeCountMixedOnce(ctx context.Context, providers []string,
 				result.RetryAfter = ra
 			}
 			m.MarkResult(execCtx, result)
+			if isRequestInvalidError(errExec) {
+				return cliproxyexecutor.Response{}, errExec
+			}
 			lastErr = errExec
 			continue
 		}
@@ -711,6 +717,9 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 			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
 		}
@@ -1110,6 +1119,9 @@ 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 +1311,7 @@ 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) {
@@ -1430,6 +1442,21 @@ 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.
+// Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow, 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":
+	case "gemini-cli", "aistudio", "antigravity", "qwen", "iflow", "kimi":
 		return provider
 	default:
 		return ""

sdk/cliproxy/auth/oauth_model_alias_test.go

@@ -70,6 +70,15 @@ func TestResolveOAuthUpstreamModel_SuffixPreservation(t *testing.T) {
 			input:   "gemini-2.5-pro(none)",
 			want:    "gemini-2.5-pro-exp-03-25(none)",
 		},
+		{
+			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{
@@ -152,11 +161,21 @@ func createAuthForChannel(channel string) *Auth {
 		return &Auth{Provider: "qwen"}
 	case "iflow":
 		return &Auth{Provider: "iflow"}
+	case "kimi":
+		return &Auth{Provider: "kimi"}
 	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 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")
+	}
+}

test/config_migration_test.go

@@ -1,195 +0,0 @@
-package test
-
-import (
-	"os"
-	"path/filepath"
-	"strings"
-	"testing"
-
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
-)
-
-func TestLegacyConfigMigration(t *testing.T) {
-	t.Run("onlyLegacyFields", func(t *testing.T) {
-		path := writeConfig(t, `
-port: 8080
-generative-language-api-key:
-  - "legacy-gemini-1"
-openai-compatibility:
-  - name: "legacy-provider"
-    base-url: "https://example.com"
-    api-keys:
-      - "legacy-openai-1"
-amp-upstream-url: "https://amp.example.com"
-amp-upstream-api-key: "amp-legacy-key"
-amp-restrict-management-to-localhost: false
-amp-model-mappings:
-  - from: "old-model"
-    to: "new-model"
-`)
-		cfg, err := config.LoadConfig(path)
-		if err != nil {
-			t.Fatalf("load legacy config: %v", err)
-		}
-		if got := len(cfg.GeminiKey); got != 1 || cfg.GeminiKey[0].APIKey != "legacy-gemini-1" {
-			t.Fatalf("gemini migration mismatch: %+v", cfg.GeminiKey)
-		}
-		if got := len(cfg.OpenAICompatibility); got != 1 {
-			t.Fatalf("expected 1 openai-compat provider, got %d", got)
-		}
-		if entries := cfg.OpenAICompatibility[0].APIKeyEntries; len(entries) != 1 || entries[0].APIKey != "legacy-openai-1" {
-			t.Fatalf("openai-compat migration mismatch: %+v", entries)
-		}
-		if cfg.AmpCode.UpstreamURL != "https://amp.example.com" || cfg.AmpCode.UpstreamAPIKey != "amp-legacy-key" {
-			t.Fatalf("amp migration failed: %+v", cfg.AmpCode)
-		}
-		if cfg.AmpCode.RestrictManagementToLocalhost {
-			t.Fatalf("expected amp restriction to be false after migration")
-		}
-		if got := len(cfg.AmpCode.ModelMappings); got != 1 || cfg.AmpCode.ModelMappings[0].From != "old-model" {
-			t.Fatalf("amp mappings migration mismatch: %+v", cfg.AmpCode.ModelMappings)
-		}
-		updated := readFile(t, path)
-		if strings.Contains(updated, "generative-language-api-key") {
-			t.Fatalf("legacy gemini key still present:\n%s", updated)
-		}
-		if strings.Contains(updated, "amp-upstream-url") || strings.Contains(updated, "amp-restrict-management-to-localhost") {
-			t.Fatalf("legacy amp keys still present:\n%s", updated)
-		}
-		if strings.Contains(updated, "\n    api-keys:") {
-			t.Fatalf("legacy openai compat keys still present:\n%s", updated)
-		}
-	})
-
-	t.Run("mixedLegacyAndNewFields", func(t *testing.T) {
-		path := writeConfig(t, `
-gemini-api-key:
-  - api-key: "new-gemini"
-generative-language-api-key:
-  - "new-gemini"
-  - "legacy-gemini-only"
-openai-compatibility:
-  - name: "mixed-provider"
-    base-url: "https://mixed.example.com"
-    api-key-entries:
-      - api-key: "new-entry"
-    api-keys:
-      - "legacy-entry"
-      - "new-entry"
-`)
-		cfg, err := config.LoadConfig(path)
-		if err != nil {
-			t.Fatalf("load mixed config: %v", err)
-		}
-		if got := len(cfg.GeminiKey); got != 2 {
-			t.Fatalf("expected 2 gemini entries, got %d: %+v", got, cfg.GeminiKey)
-		}
-		seen := make(map[string]struct{}, len(cfg.GeminiKey))
-		for _, entry := range cfg.GeminiKey {
-			if _, exists := seen[entry.APIKey]; exists {
-				t.Fatalf("duplicate gemini key %q after migration", entry.APIKey)
-			}
-			seen[entry.APIKey] = struct{}{}
-		}
-		provider := cfg.OpenAICompatibility[0]
-		if got := len(provider.APIKeyEntries); got != 2 {
-			t.Fatalf("expected 2 openai entries, got %d: %+v", got, provider.APIKeyEntries)
-		}
-		entrySeen := make(map[string]struct{}, len(provider.APIKeyEntries))
-		for _, entry := range provider.APIKeyEntries {
-			if _, ok := entrySeen[entry.APIKey]; ok {
-				t.Fatalf("duplicate openai key %q after migration", entry.APIKey)
-			}
-			entrySeen[entry.APIKey] = struct{}{}
-		}
-	})
-
-	t.Run("onlyNewFields", func(t *testing.T) {
-		path := writeConfig(t, `
-gemini-api-key:
-  - api-key: "new-only"
-openai-compatibility:
-  - name: "new-only-provider"
-    base-url: "https://new-only.example.com"
-    api-key-entries:
-      - api-key: "new-only-entry"
-ampcode:
-  upstream-url: "https://amp.new"
-  upstream-api-key: "new-amp-key"
-  restrict-management-to-localhost: true
-  model-mappings:
-    - from: "a"
-      to: "b"
-`)
-		cfg, err := config.LoadConfig(path)
-		if err != nil {
-			t.Fatalf("load new config: %v", err)
-		}
-		if len(cfg.GeminiKey) != 1 || cfg.GeminiKey[0].APIKey != "new-only" {
-			t.Fatalf("unexpected gemini entries: %+v", cfg.GeminiKey)
-		}
-		if len(cfg.OpenAICompatibility) != 1 || len(cfg.OpenAICompatibility[0].APIKeyEntries) != 1 {
-			t.Fatalf("unexpected openai compat entries: %+v", cfg.OpenAICompatibility)
-		}
-		if cfg.AmpCode.UpstreamURL != "https://amp.new" || cfg.AmpCode.UpstreamAPIKey != "new-amp-key" {
-			t.Fatalf("unexpected amp config: %+v", cfg.AmpCode)
-		}
-	})
-
-	t.Run("duplicateNamesDifferentBase", func(t *testing.T) {
-		path := writeConfig(t, `
-openai-compatibility:
-  - name: "dup-provider"
-    base-url: "https://provider-a"
-    api-keys:
-      - "key-a"
-  - name: "dup-provider"
-    base-url: "https://provider-b"
-    api-keys:
-      - "key-b"
-`)
-		cfg, err := config.LoadConfig(path)
-		if err != nil {
-			t.Fatalf("load duplicate config: %v", err)
-		}
-		if len(cfg.OpenAICompatibility) != 2 {
-			t.Fatalf("expected 2 providers, got %d", len(cfg.OpenAICompatibility))
-		}
-		for _, entry := range cfg.OpenAICompatibility {
-			if len(entry.APIKeyEntries) != 1 {
-				t.Fatalf("expected 1 key entry per provider: %+v", entry)
-			}
-			switch entry.BaseURL {
-			case "https://provider-a":
-				if entry.APIKeyEntries[0].APIKey != "key-a" {
-					t.Fatalf("provider-a key mismatch: %+v", entry.APIKeyEntries)
-				}
-			case "https://provider-b":
-				if entry.APIKeyEntries[0].APIKey != "key-b" {
-					t.Fatalf("provider-b key mismatch: %+v", entry.APIKeyEntries)
-				}
-			default:
-				t.Fatalf("unexpected provider base url: %s", entry.BaseURL)
-			}
-		}
-	})
-}
-
-func writeConfig(t *testing.T, content string) string {
-	t.Helper()
-	dir := t.TempDir()
-	path := filepath.Join(dir, "config.yaml")
-	if err := os.WriteFile(path, []byte(strings.TrimSpace(content)+"\n"), 0o644); err != nil {
-		t.Fatalf("write temp config: %v", err)
-	}
-	return path
-}
-
-func readFile(t *testing.T, path string) string {
-	t.Helper()
-	data, err := os.ReadFile(path)
-	if err != nil {
-		t.Fatalf("read temp config: %v", err)
-	}
-	return string(data)
-}

test/thinking_conversion_test.go

@@ -2590,6 +2590,135 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 	runThinkingTests(t, cases)
 }
 
+// TestThinkingE2EClaudeAdaptive_Body tests Claude thinking.type=adaptive extended body-only cases.
+// These cases validate that adaptive means "thinking enabled without explicit budget", and
+// cross-protocol conversion should resolve to target-model maximum thinking capability.
+func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	uid := fmt.Sprintf("thinking-e2e-claude-adaptive-%d", time.Now().UnixNano())
+
+	reg.RegisterClient(uid, "test", getTestModels())
+	defer reg.UnregisterClient(uid)
+
+	cases := []thinkingTestCase{
+		// A1: Claude adaptive to OpenAI level model -> highest supported level
+		{
+			name:        "A1",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		// A2: Claude adaptive to Gemini level subset model -> highest supported level
+		{
+			name:            "A2",
+			from:            "claude",
+			to:              "gemini",
+			model:           "level-subset-model",
+			inputJSON:       `{"model":"level-subset-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
+			expectValue:     "high",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		// A3: Claude adaptive to Gemini budget model -> max budget
+		{
+			name:            "A3",
+			from:            "claude",
+			to:              "gemini",
+			model:           "gemini-budget-model",
+			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "20000",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		// A4: Claude adaptive to Gemini mixed model -> highest supported level
+		{
+			name:            "A4",
+			from:            "claude",
+			to:              "gemini",
+			model:           "gemini-mixed-model",
+			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
+			expectValue:     "high",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		// A5: Claude adaptive passthrough for same protocol
+		{
+			name:        "A5",
+			from:        "claude",
+			to:          "claude",
+			model:       "claude-budget-model",
+			inputJSON:   `{"model":"claude-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "thinking.type",
+			expectValue: "adaptive",
+			expectErr:   false,
+		},
+		// A6: Claude adaptive to Antigravity budget model -> max budget
+		{
+			name:            "A6",
+			from:            "claude",
+			to:              "antigravity",
+			model:           "antigravity-budget-model",
+			inputJSON:       `{"model":"antigravity-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "20000",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		// A7: Claude adaptive to iFlow GLM -> enabled boolean
+		{
+			name:        "A7",
+			from:        "claude",
+			to:          "iflow",
+			model:       "glm-test",
+			inputJSON:   `{"model":"glm-test","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "chat_template_kwargs.enable_thinking",
+			expectValue: "true",
+			expectErr:   false,
+		},
+		// A8: Claude adaptive to iFlow MiniMax -> enabled boolean
+		{
+			name:        "A8",
+			from:        "claude",
+			to:          "iflow",
+			model:       "minimax-test",
+			inputJSON:   `{"model":"minimax-test","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "reasoning_split",
+			expectValue: "true",
+			expectErr:   false,
+		},
+		// A9: Claude adaptive to Codex level model -> highest supported level
+		{
+			name:        "A9",
+			from:        "claude",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		// A10: Claude adaptive on non-thinking model should still be stripped
+		{
+			name:        "A10",
+			from:        "claude",
+			to:          "openai",
+			model:       "no-thinking-model",
+			inputJSON:   `{"model":"no-thinking-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "",
+			expectErr:   false,
+		},
+	}
+
+	runThinkingTests(t, cases)
+}
+
 // getTestModels returns the shared model definitions for E2E tests.
 func getTestModels() []*registry.ModelInfo {
 	return []*registry.ModelInfo{