revert .gitignore

fix: remove SOFT_LIMIT_REACHED logic, tool compression, and fix bugs

README.md

@@ -10,7 +10,7 @@ The Plus release stays in lockstep with the mainline features.
 
 ## Differences from the Mainline
 
-[![z.ai](https://assets.router-for.me/english-5.png)](https://z.ai/subscribe?ic=8JVLJQFSKB)
+[![z.ai](https://assets.router-for.me/english-5-0.jpg)](https://z.ai/subscribe?ic=8JVLJQFSKB)
 
 ## New Features (Plus Enhanced)
 

README_CN.md

@@ -10,7 +10,7 @@
 
 ## 与主线版本版本差异
 
-[![bigmodel.cn](https://assets.router-for.me/chinese-5.png)](https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII)
+[![bigmodel.cn](https://assets.router-for.me/chinese-5-0.jpg)](https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII)
 
 ## 新增功能 (Plus 增强版)
 

config.example.yaml

@@ -233,6 +233,9 @@ nonstream-keepalive-interval: 0
 #         alias: "vertex-flash"                   # client-visible alias
 #       - name: "gemini-2.5-pro"
 #         alias: "vertex-pro"
+#     excluded-models:                            # optional: models to exclude from listing
+#       - "imagen-3.0-generate-002"
+#       - "imagen-*"
 
 # Amp Integration
 # ampcode:

internal/api/handlers/management/auth_files.go

@@ -16,6 +16,7 @@ import (
 	"net/url"
 	"os"
 	"path/filepath"
+	"runtime"
 	"sort"
 	"strconv"
 	"strings"
@@ -192,17 +193,6 @@ func startCallbackForwarder(port int, provider, targetBase string) (*callbackFor
 	return forwarder, nil
 }
 
-func stopCallbackForwarder(port int) {
-	callbackForwardersMu.Lock()
-	forwarder := callbackForwarders[port]
-	if forwarder != nil {
-		delete(callbackForwarders, port)
-	}
-	callbackForwardersMu.Unlock()
-
-	stopForwarderInstance(port, forwarder)
-}
-
 func stopCallbackForwarderInstance(port int, forwarder *callbackForwarder) {
 	if forwarder == nil {
 		return
@@ -644,44 +634,85 @@ func (h *Handler) DeleteAuthFile(c *gin.Context) {
 		c.JSON(400, gin.H{"error": "invalid name"})
 		return
 	}
-	full := filepath.Join(h.cfg.AuthDir, filepath.Base(name))
-	if !filepath.IsAbs(full) {
-		if abs, errAbs := filepath.Abs(full); errAbs == nil {
-			full = abs
+
+	targetPath := filepath.Join(h.cfg.AuthDir, filepath.Base(name))
+	targetID := ""
+	if targetAuth := h.findAuthForDelete(name); targetAuth != nil {
+		targetID = strings.TrimSpace(targetAuth.ID)
+		if path := strings.TrimSpace(authAttribute(targetAuth, "path")); path != "" {
+			targetPath = path
 		}
 	}
-	if err := os.Remove(full); err != nil {
-		if os.IsNotExist(err) {
+	if !filepath.IsAbs(targetPath) {
+		if abs, errAbs := filepath.Abs(targetPath); errAbs == nil {
+			targetPath = abs
+		}
+	}
+	if errRemove := os.Remove(targetPath); errRemove != nil {
+		if os.IsNotExist(errRemove) {
 			c.JSON(404, gin.H{"error": "file not found"})
 		} else {
-			c.JSON(500, gin.H{"error": fmt.Sprintf("failed to remove file: %v", err)})
+			c.JSON(500, gin.H{"error": fmt.Sprintf("failed to remove file: %v", errRemove)})
 		}
 		return
 	}
-	if err := h.deleteTokenRecord(ctx, full); err != nil {
-		c.JSON(500, gin.H{"error": err.Error()})
+	if errDeleteRecord := h.deleteTokenRecord(ctx, targetPath); errDeleteRecord != nil {
+		c.JSON(500, gin.H{"error": errDeleteRecord.Error()})
 		return
 	}
-	h.disableAuth(ctx, full)
+	if targetID != "" {
+		h.disableAuth(ctx, targetID)
+	} else {
+		h.disableAuth(ctx, targetPath)
+	}
 	c.JSON(200, gin.H{"status": "ok"})
 }
 
+func (h *Handler) findAuthForDelete(name string) *coreauth.Auth {
+	if h == nil || h.authManager == nil {
+		return nil
+	}
+	name = strings.TrimSpace(name)
+	if name == "" {
+		return nil
+	}
+	if auth, ok := h.authManager.GetByID(name); ok {
+		return auth
+	}
+	auths := h.authManager.List()
+	for _, auth := range auths {
+		if auth == nil {
+			continue
+		}
+		if strings.TrimSpace(auth.FileName) == name {
+			return auth
+		}
+		if filepath.Base(strings.TrimSpace(authAttribute(auth, "path"))) == name {
+			return auth
+		}
+	}
+	return nil
+}
+
 func (h *Handler) authIDForPath(path string) string {
 	path = strings.TrimSpace(path)
 	if path == "" {
 		return ""
 	}
-	if h == nil || h.cfg == nil {
-		return path
+	id := path
+	if h != nil && h.cfg != nil {
+		authDir := strings.TrimSpace(h.cfg.AuthDir)
+		if authDir != "" {
+			if rel, errRel := filepath.Rel(authDir, path); errRel == nil && rel != "" {
+				id = rel
+			}
+		}
 	}
-	authDir := strings.TrimSpace(h.cfg.AuthDir)
-	if authDir == "" {
-		return path
+	// On Windows, normalize ID casing to avoid duplicate auth entries caused by case-insensitive paths.
+	if runtime.GOOS == "windows" {
+		id = strings.ToLower(id)
 	}
-	if rel, err := filepath.Rel(authDir, path); err == nil && rel != "" {
-		return rel
-	}
-	return path
+	return id
 }
 
 func (h *Handler) registerAuthFromFile(ctx context.Context, path string, data []byte) error {
@@ -899,10 +930,19 @@ func (h *Handler) disableAuth(ctx context.Context, id string) {
 	if h == nil || h.authManager == nil {
 		return
 	}
-	authID := h.authIDForPath(id)
-	if authID == "" {
-		authID = strings.TrimSpace(id)
+	id = strings.TrimSpace(id)
+	if id == "" {
+		return
 	}
+	if auth, ok := h.authManager.GetByID(id); ok {
+		auth.Disabled = true
+		auth.Status = coreauth.StatusDisabled
+		auth.StatusMessage = "removed via management API"
+		auth.UpdatedAt = time.Now()
+		_, _ = h.authManager.Update(ctx, auth)
+		return
+	}
+	authID := h.authIDForPath(id)
 	if authID == "" {
 		return
 	}
@@ -2549,6 +2589,7 @@ func PopulateAuthContext(ctx context.Context, c *gin.Context) context.Context {
 	}
 	return coreauth.WithRequestInfo(ctx, info)
 }
+
 const kiroCallbackPort = 9876
 
 func (h *Handler) RequestKiroToken(c *gin.Context) {
@@ -2685,6 +2726,7 @@ func (h *Handler) RequestKiroToken(c *gin.Context) {
 		}
 
 		isWebUI := isWebUIRequest(c)
+		var forwarder *callbackForwarder
 		if isWebUI {
 			targetURL, errTarget := h.managementCallbackURL("/kiro/callback")
 			if errTarget != nil {
@@ -2692,7 +2734,8 @@ func (h *Handler) RequestKiroToken(c *gin.Context) {
 				c.JSON(http.StatusInternalServerError, gin.H{"error": "callback server unavailable"})
 				return
 			}
-			if _, errStart := startCallbackForwarder(kiroCallbackPort, "kiro", targetURL); errStart != nil {
+			var errStart error
+			if forwarder, errStart = startCallbackForwarder(kiroCallbackPort, "kiro", targetURL); errStart != nil {
 				log.WithError(errStart).Error("failed to start kiro callback forwarder")
 				c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to start callback server"})
 				return
@@ -2701,7 +2744,7 @@ func (h *Handler) RequestKiroToken(c *gin.Context) {
 
 		go func() {
 			if isWebUI {
-				defer stopCallbackForwarder(kiroCallbackPort)
+				defer stopCallbackForwarderInstance(kiroCallbackPort, forwarder)
 			}
 
 			socialClient := kiroauth.NewSocialAuthClient(h.cfg)
@@ -2904,7 +2947,7 @@ func (h *Handler) RequestKiloToken(c *gin.Context) {
 			Metadata: map[string]any{
 				"email":           status.UserEmail,
 				"organization_id": orgID,
-				"model":          defaults.Model,
+				"model":           defaults.Model,
 			},
 		}
 

internal/api/handlers/management/auth_files_delete_test.go

@@ -0,0 +1,129 @@
+package management
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+)
+
+func TestDeleteAuthFile_UsesAuthPathFromManager(t *testing.T) {
+	t.Setenv("MANAGEMENT_PASSWORD", "")
+	gin.SetMode(gin.TestMode)
+
+	tempDir := t.TempDir()
+	authDir := filepath.Join(tempDir, "auth")
+	externalDir := filepath.Join(tempDir, "external")
+	if errMkdirAuth := os.MkdirAll(authDir, 0o700); errMkdirAuth != nil {
+		t.Fatalf("failed to create auth dir: %v", errMkdirAuth)
+	}
+	if errMkdirExternal := os.MkdirAll(externalDir, 0o700); errMkdirExternal != nil {
+		t.Fatalf("failed to create external dir: %v", errMkdirExternal)
+	}
+
+	fileName := "codex-user@example.com-plus.json"
+	shadowPath := filepath.Join(authDir, fileName)
+	realPath := filepath.Join(externalDir, fileName)
+	if errWriteShadow := os.WriteFile(shadowPath, []byte(`{"type":"codex","email":"shadow@example.com"}`), 0o600); errWriteShadow != nil {
+		t.Fatalf("failed to write shadow file: %v", errWriteShadow)
+	}
+	if errWriteReal := os.WriteFile(realPath, []byte(`{"type":"codex","email":"real@example.com"}`), 0o600); errWriteReal != nil {
+		t.Fatalf("failed to write real file: %v", errWriteReal)
+	}
+
+	manager := coreauth.NewManager(nil, nil, nil)
+	record := &coreauth.Auth{
+		ID:          "legacy/" + fileName,
+		FileName:    fileName,
+		Provider:    "codex",
+		Status:      coreauth.StatusError,
+		Unavailable: true,
+		Attributes: map[string]string{
+			"path": realPath,
+		},
+		Metadata: map[string]any{
+			"type":  "codex",
+			"email": "real@example.com",
+		},
+	}
+	if _, errRegister := manager.Register(context.Background(), record); errRegister != nil {
+		t.Fatalf("failed to register auth record: %v", errRegister)
+	}
+
+	h := NewHandlerWithoutConfigFilePath(&config.Config{AuthDir: authDir}, manager)
+	h.tokenStore = &memoryAuthStore{}
+
+	deleteRec := httptest.NewRecorder()
+	deleteCtx, _ := gin.CreateTestContext(deleteRec)
+	deleteReq := httptest.NewRequest(http.MethodDelete, "/v0/management/auth-files?name="+url.QueryEscape(fileName), nil)
+	deleteCtx.Request = deleteReq
+	h.DeleteAuthFile(deleteCtx)
+
+	if deleteRec.Code != http.StatusOK {
+		t.Fatalf("expected delete status %d, got %d with body %s", http.StatusOK, deleteRec.Code, deleteRec.Body.String())
+	}
+	if _, errStatReal := os.Stat(realPath); !os.IsNotExist(errStatReal) {
+		t.Fatalf("expected managed auth file to be removed, stat err: %v", errStatReal)
+	}
+	if _, errStatShadow := os.Stat(shadowPath); errStatShadow != nil {
+		t.Fatalf("expected shadow auth file to remain, stat err: %v", errStatShadow)
+	}
+
+	listRec := httptest.NewRecorder()
+	listCtx, _ := gin.CreateTestContext(listRec)
+	listReq := httptest.NewRequest(http.MethodGet, "/v0/management/auth-files", nil)
+	listCtx.Request = listReq
+	h.ListAuthFiles(listCtx)
+
+	if listRec.Code != http.StatusOK {
+		t.Fatalf("expected list status %d, got %d with body %s", http.StatusOK, listRec.Code, listRec.Body.String())
+	}
+	var listPayload map[string]any
+	if errUnmarshal := json.Unmarshal(listRec.Body.Bytes(), &listPayload); errUnmarshal != nil {
+		t.Fatalf("failed to decode list payload: %v", errUnmarshal)
+	}
+	filesRaw, ok := listPayload["files"].([]any)
+	if !ok {
+		t.Fatalf("expected files array, payload: %#v", listPayload)
+	}
+	if len(filesRaw) != 0 {
+		t.Fatalf("expected removed auth to be hidden from list, got %d entries", len(filesRaw))
+	}
+}
+
+func TestDeleteAuthFile_FallbackToAuthDirPath(t *testing.T) {
+	t.Setenv("MANAGEMENT_PASSWORD", "")
+	gin.SetMode(gin.TestMode)
+
+	authDir := t.TempDir()
+	fileName := "fallback-user.json"
+	filePath := filepath.Join(authDir, fileName)
+	if errWrite := os.WriteFile(filePath, []byte(`{"type":"codex"}`), 0o600); errWrite != nil {
+		t.Fatalf("failed to write auth file: %v", errWrite)
+	}
+
+	manager := coreauth.NewManager(nil, nil, nil)
+	h := NewHandlerWithoutConfigFilePath(&config.Config{AuthDir: authDir}, manager)
+	h.tokenStore = &memoryAuthStore{}
+
+	deleteRec := httptest.NewRecorder()
+	deleteCtx, _ := gin.CreateTestContext(deleteRec)
+	deleteReq := httptest.NewRequest(http.MethodDelete, "/v0/management/auth-files?name="+url.QueryEscape(fileName), nil)
+	deleteCtx.Request = deleteReq
+	h.DeleteAuthFile(deleteCtx)
+
+	if deleteRec.Code != http.StatusOK {
+		t.Fatalf("expected delete status %d, got %d with body %s", http.StatusOK, deleteRec.Code, deleteRec.Body.String())
+	}
+	if _, errStat := os.Stat(filePath); !os.IsNotExist(errStat) {
+		t.Fatalf("expected auth file to be removed from auth dir, stat err: %v", errStat)
+	}
+}

internal/api/handlers/management/config_lists.go

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

internal/auth/copilot/copilot_auth.go

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

internal/config/oauth_model_alias_defaults.go

@@ -0,0 +1,61 @@
+package config
+
+import "strings"
+
+// defaultKiroAliases returns default oauth-model-alias entries for Kiro.
+// These aliases expose standard Claude IDs for Kiro-prefixed upstream models.
+func defaultKiroAliases() []OAuthModelAlias {
+	return []OAuthModelAlias{
+		// Sonnet 4.6
+		{Name: "kiro-claude-sonnet-4-6", Alias: "claude-sonnet-4-6", Fork: true},
+		// 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},
+	}
+}
+
+// defaultGitHubCopilotAliases returns default oauth-model-alias entries for
+// GitHub Copilot Claude models. It exposes hyphen-style IDs used by clients.
+func defaultGitHubCopilotAliases() []OAuthModelAlias {
+	return []OAuthModelAlias{
+		{Name: "claude-haiku-4.5", Alias: "claude-haiku-4-5", Fork: true},
+		{Name: "claude-opus-4.1", Alias: "claude-opus-4-1", Fork: true},
+		{Name: "claude-opus-4.5", Alias: "claude-opus-4-5", Fork: true},
+		{Name: "claude-opus-4.6", Alias: "claude-opus-4-6", Fork: true},
+		{Name: "claude-sonnet-4.5", Alias: "claude-sonnet-4-5", Fork: true},
+		{Name: "claude-sonnet-4.6", Alias: "claude-sonnet-4-6", Fork: true},
+	}
+}
+
+// GitHubCopilotAliasesFromModels generates oauth-model-alias entries from a dynamic
+// list of model IDs fetched from the Copilot API. It auto-creates aliases for
+// models whose ID contains a dot (e.g. "claude-opus-4.6" → "claude-opus-4-6"),
+// which is the pattern used by Claude models on Copilot.
+func GitHubCopilotAliasesFromModels(modelIDs []string) []OAuthModelAlias {
+	var aliases []OAuthModelAlias
+	seen := make(map[string]struct{})
+	for _, id := range modelIDs {
+		if !strings.Contains(id, ".") {
+			continue
+		}
+		hyphenID := strings.ReplaceAll(id, ".", "-")
+		key := id + "→" + hyphenID
+		if _, ok := seen[key]; ok {
+			continue
+		}
+		seen[key] = struct{}{}
+		aliases = append(aliases, OAuthModelAlias{Name: id, Alias: hyphenID, Fork: true})
+	}
+	return aliases
+}

internal/config/vertex_compat.go

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

internal/registry/model_definitions.go

@@ -23,7 +23,6 @@ import (
 //   - kiro
 //   - kilo
 //   - github-copilot
-//   - kiro
 //   - amazonq
 //   - antigravity (returns static overrides only)
 func GetStaticModelDefinitionsByChannel(channel string) []*ModelInfo {

internal/registry/model_definitions_static_data.go

@@ -37,7 +37,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4.6 Sonnet",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false, Levels: []string{"low", "medium", "high"}},
 		},
 		{
 			ID:                  "claude-opus-4-6",
@@ -49,7 +49,7 @@ func GetClaudeModels() []*ModelInfo {
 			Description:         "Premium model combining maximum intelligence with practical performance",
 			ContextLength:       1000000,
 			MaxCompletionTokens: 128000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false, Levels: []string{"low", "medium", "high", "max"}},
 		},
 		{
 			ID:                  "claude-sonnet-4-6",

internal/registry/model_registry.go

@@ -47,6 +47,8 @@ type ModelInfo struct {
 	MaxCompletionTokens int `json:"max_completion_tokens,omitempty"`
 	// SupportedParameters lists supported parameters
 	SupportedParameters []string `json:"supported_parameters,omitempty"`
+	// SupportedEndpoints lists supported API endpoints (e.g., "/chat/completions", "/responses").
+	SupportedEndpoints []string `json:"supported_endpoints,omitempty"`
 	// SupportedInputModalities lists supported input modalities (e.g., TEXT, IMAGE, VIDEO, AUDIO)
 	SupportedInputModalities []string `json:"supportedInputModalities,omitempty"`
 	// SupportedOutputModalities lists supported output modalities (e.g., TEXT, IMAGE)

internal/runtime/executor/antigravity_executor_buildrequest_test.go

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

internal/runtime/executor/claude_executor.go

@@ -634,6 +634,12 @@ func disableThinkingIfToolChoiceForced(body []byte) []byte {
 	if toolChoiceType == "any" || toolChoiceType == "tool" {
 		// Remove thinking configuration entirely to avoid API error
 		body, _ = sjson.DeleteBytes(body, "thinking")
+		// Adaptive thinking may also set output_config.effort; remove it to avoid
+		// leaking thinking controls when tool_choice forces tool use.
+		body, _ = sjson.DeleteBytes(body, "output_config.effort")
+		if oc := gjson.GetBytes(body, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			body, _ = sjson.DeleteBytes(body, "output_config")
+		}
 	}
 	return body
 }

internal/runtime/executor/github_copilot_executor.go

@@ -14,6 +14,7 @@ import (
 	"github.com/google/uuid"
 	copilotauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/copilot"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -490,18 +491,46 @@ func (e *GitHubCopilotExecutor) applyHeaders(r *http.Request, apiToken string, b
 	r.Header.Set("X-Request-Id", uuid.NewString())
 
 	initiator := "user"
-	if len(body) > 0 {
-		if messages := gjson.GetBytes(body, "messages"); messages.Exists() && messages.IsArray() {
-			for _, msg := range messages.Array() {
-				role := msg.Get("role").String()
-				if role == "assistant" || role == "tool" {
-					initiator = "agent"
-					break
-				}
+	if role := detectLastConversationRole(body); role == "assistant" || role == "tool" {
+		initiator = "agent"
+	}
+	r.Header.Set("X-Initiator", initiator)
+}
+
+func detectLastConversationRole(body []byte) string {
+	if len(body) == 0 {
+		return ""
+	}
+
+	if messages := gjson.GetBytes(body, "messages"); messages.Exists() && messages.IsArray() {
+		arr := messages.Array()
+		for i := len(arr) - 1; i >= 0; i-- {
+			if role := arr[i].Get("role").String(); role != "" {
+				return role
 			}
 		}
 	}
-	r.Header.Set("X-Initiator", initiator)
+
+	if inputs := gjson.GetBytes(body, "input"); inputs.Exists() && inputs.IsArray() {
+		arr := inputs.Array()
+		for i := len(arr) - 1; i >= 0; i-- {
+			item := arr[i]
+
+			// Most Responses input items carry a top-level role.
+			if role := item.Get("role").String(); role != "" {
+				return role
+			}
+
+			switch item.Get("type").String() {
+			case "function_call", "function_call_arguments":
+				return "assistant"
+			case "function_call_output", "function_call_response", "tool_result":
+				return "tool"
+			}
+		}
+	}
+
+	return ""
 }
 
 // detectVisionContent checks if the request body contains vision/image content.
@@ -1236,3 +1265,99 @@ func translateGitHubCopilotResponsesStreamToClaude(line []byte, param *any) []st
 func isHTTPSuccess(statusCode int) bool {
 	return statusCode >= 200 && statusCode < 300
 }
+
+const (
+	// defaultCopilotContextLength is the default context window for unknown Copilot models.
+	defaultCopilotContextLength = 128000
+	// defaultCopilotMaxCompletionTokens is the default max output tokens for unknown Copilot models.
+	defaultCopilotMaxCompletionTokens = 16384
+)
+
+// FetchGitHubCopilotModels dynamically fetches available models from the GitHub Copilot API.
+// It exchanges the GitHub access token stored in auth.Metadata for a Copilot API token,
+// then queries the /models endpoint. Falls back to the static registry on any failure.
+func FetchGitHubCopilotModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *config.Config) []*registry.ModelInfo {
+	if auth == nil {
+		log.Debug("github-copilot: auth is nil, using static models")
+		return registry.GetGitHubCopilotModels()
+	}
+
+	accessToken := metaStringValue(auth.Metadata, "access_token")
+	if accessToken == "" {
+		log.Debug("github-copilot: no access_token in auth metadata, using static models")
+		return registry.GetGitHubCopilotModels()
+	}
+
+	copilotAuth := copilotauth.NewCopilotAuth(cfg)
+
+	entries, err := copilotAuth.ListModelsWithGitHubToken(ctx, accessToken)
+	if err != nil {
+		log.Warnf("github-copilot: failed to fetch dynamic models: %v, using static models", err)
+		return registry.GetGitHubCopilotModels()
+	}
+
+	if len(entries) == 0 {
+		log.Debug("github-copilot: API returned no models, using static models")
+		return registry.GetGitHubCopilotModels()
+	}
+
+	// Build a lookup from the static definitions so we can enrich dynamic entries
+	// with known context lengths, thinking support, etc.
+	staticMap := make(map[string]*registry.ModelInfo)
+	for _, m := range registry.GetGitHubCopilotModels() {
+		staticMap[m.ID] = m
+	}
+
+	now := time.Now().Unix()
+	models := make([]*registry.ModelInfo, 0, len(entries))
+	seen := make(map[string]struct{}, len(entries))
+	for _, entry := range entries {
+		if entry.ID == "" {
+			continue
+		}
+		// Deduplicate model IDs to avoid incorrect reference counting.
+		if _, dup := seen[entry.ID]; dup {
+			continue
+		}
+		seen[entry.ID] = struct{}{}
+
+		m := &registry.ModelInfo{
+			ID:      entry.ID,
+			Object:  "model",
+			Created: now,
+			OwnedBy: "github-copilot",
+			Type:    "github-copilot",
+		}
+
+		if entry.Created > 0 {
+			m.Created = entry.Created
+		}
+		if entry.Name != "" {
+			m.DisplayName = entry.Name
+		} else {
+			m.DisplayName = entry.ID
+		}
+
+		// Merge known metadata from the static fallback list
+		if static, ok := staticMap[entry.ID]; ok {
+			if m.DisplayName == entry.ID && static.DisplayName != "" {
+				m.DisplayName = static.DisplayName
+			}
+			m.Description = static.Description
+			m.ContextLength = static.ContextLength
+			m.MaxCompletionTokens = static.MaxCompletionTokens
+			m.SupportedEndpoints = static.SupportedEndpoints
+			m.Thinking = static.Thinking
+		} else {
+			// Sensible defaults for models not in the static list
+			m.Description = entry.ID + " via GitHub Copilot"
+			m.ContextLength = defaultCopilotContextLength
+			m.MaxCompletionTokens = defaultCopilotMaxCompletionTokens
+		}
+
+		models = append(models, m)
+	}
+
+	log.Infof("github-copilot: fetched %d models from API", len(models))
+	return models
+}

internal/runtime/executor/github_copilot_executor_test.go

@@ -262,15 +262,15 @@ func TestApplyHeaders_XInitiator_UserOnly(t *testing.T) {
 	}
 }
 
-func TestApplyHeaders_XInitiator_AgentWithAssistantAndUserToolResult(t *testing.T) {
+func TestApplyHeaders_XInitiator_UserWhenLastRoleIsUser(t *testing.T) {
 	t.Parallel()
 	e := &GitHubCopilotExecutor{}
 	req, _ := http.NewRequest(http.MethodPost, "https://example.com", nil)
-	// Claude Code typical flow: last message is user (tool result), but has assistant in history
+	// Last role governs the initiator decision.
 	body := []byte(`{"messages":[{"role":"user","content":"hello"},{"role":"assistant","content":"I will read the file"},{"role":"user","content":"tool result here"}]}`)
 	e.applyHeaders(req, "token", body)
-	if got := req.Header.Get("X-Initiator"); got != "agent" {
-		t.Fatalf("X-Initiator = %q, want agent (assistant exists in messages)", got)
+	if got := req.Header.Get("X-Initiator"); got != "user" {
+		t.Fatalf("X-Initiator = %q, want user (last role is user)", got)
 	}
 }
 
@@ -285,6 +285,39 @@ func TestApplyHeaders_XInitiator_AgentWithToolRole(t *testing.T) {
 	}
 }
 
+func TestApplyHeaders_XInitiator_InputArrayLastAssistantMessage(t *testing.T) {
+	t.Parallel()
+	e := &GitHubCopilotExecutor{}
+	req, _ := http.NewRequest(http.MethodPost, "https://example.com", nil)
+	body := []byte(`{"input":[{"type":"message","role":"user","content":[{"type":"input_text","text":"Hi"}]},{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Hello"}]}]}`)
+	e.applyHeaders(req, "token", body)
+	if got := req.Header.Get("X-Initiator"); got != "agent" {
+		t.Fatalf("X-Initiator = %q, want agent (last role is assistant)", got)
+	}
+}
+
+func TestApplyHeaders_XInitiator_InputArrayLastUserMessage(t *testing.T) {
+	t.Parallel()
+	e := &GitHubCopilotExecutor{}
+	req, _ := http.NewRequest(http.MethodPost, "https://example.com", nil)
+	body := []byte(`{"input":[{"type":"message","role":"assistant","content":[{"type":"output_text","text":"I can help"}]},{"type":"message","role":"user","content":[{"type":"input_text","text":"Do X"}]}]}`)
+	e.applyHeaders(req, "token", body)
+	if got := req.Header.Get("X-Initiator"); got != "user" {
+		t.Fatalf("X-Initiator = %q, want user (last role is user)", got)
+	}
+}
+
+func TestApplyHeaders_XInitiator_InputArrayLastFunctionCallOutput(t *testing.T) {
+	t.Parallel()
+	e := &GitHubCopilotExecutor{}
+	req, _ := http.NewRequest(http.MethodPost, "https://example.com", nil)
+	body := []byte(`{"input":[{"type":"message","role":"user","content":[{"type":"input_text","text":"Use tool"}]},{"type":"function_call","call_id":"c1","name":"Read","arguments":"{}"},{"type":"function_call_output","call_id":"c1","output":"ok"}]}`)
+	e.applyHeaders(req, "token", body)
+	if got := req.Header.Get("X-Initiator"); got != "agent" {
+		t.Fatalf("X-Initiator = %q, want agent (last item maps to tool role)", got)
+	}
+}
+
 // --- Tests for x-github-api-version header (Problem M) ---
 
 func TestApplyHeaders_GitHubAPIVersion(t *testing.T) {

internal/runtime/executor/kiro_executor.go

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

internal/thinking/apply.go

@@ -293,7 +293,7 @@ func normalizeUserDefinedConfig(config ThinkingConfig, fromFormat, toFormat stri
 	if config.Mode != ModeLevel {
 		return config
 	}
-	if !isBudgetBasedProvider(toFormat) || !isLevelBasedProvider(fromFormat) {
+	if !isBudgetCapableProvider(toFormat) {
 		return config
 	}
 	budget, ok := ConvertLevelToBudget(string(config.Level))
@@ -353,6 +353,26 @@ func extractClaudeConfig(body []byte) ThinkingConfig {
 	if thinkingType == "disabled" {
 		return ThinkingConfig{Mode: ModeNone, Budget: 0}
 	}
+	if thinkingType == "adaptive" || thinkingType == "auto" {
+		// Claude adaptive thinking uses output_config.effort (low/medium/high/max).
+		// We only treat it as a thinking config when effort is explicitly present;
+		// otherwise we passthrough and let upstream defaults apply.
+		if effort := gjson.GetBytes(body, "output_config.effort"); effort.Exists() && effort.Type == gjson.String {
+			value := strings.ToLower(strings.TrimSpace(effort.String()))
+			if value == "" {
+				return ThinkingConfig{}
+			}
+			switch value {
+			case "none":
+				return ThinkingConfig{Mode: ModeNone, Budget: 0}
+			case "auto":
+				return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+			default:
+				return ThinkingConfig{Mode: ModeLevel, Level: ThinkingLevel(value)}
+			}
+		}
+		return ThinkingConfig{}
+	}
 
 	// Check budget_tokens
 	if budget := gjson.GetBytes(body, "thinking.budget_tokens"); budget.Exists() {

internal/thinking/convert.go

@@ -16,6 +16,9 @@ var levelToBudgetMap = map[string]int{
 	"medium":  8192,
 	"high":    24576,
 	"xhigh":   32768,
+	// "max" is used by Claude adaptive thinking effort. We map it to a large budget
+	// and rely on per-model clamping when converting to budget-only providers.
+	"max": 128000,
 }
 
 // ConvertLevelToBudget converts a thinking level to a budget value.
@@ -31,6 +34,7 @@ var levelToBudgetMap = map[string]int{
 //   - medium  → 8192
 //   - high    → 24576
 //   - xhigh   → 32768
+//   - max     → 128000
 //
 // Returns:
 //   - budget: The converted budget value
@@ -92,6 +96,43 @@ func ConvertBudgetToLevel(budget int) (string, bool) {
 	}
 }
 
+// HasLevel reports whether the given target level exists in the levels slice.
+// Matching is case-insensitive with leading/trailing whitespace trimmed.
+func HasLevel(levels []string, target string) bool {
+	for _, level := range levels {
+		if strings.EqualFold(strings.TrimSpace(level), target) {
+			return true
+		}
+	}
+	return false
+}
+
+// MapToClaudeEffort maps a generic thinking level string to a Claude adaptive
+// thinking effort value (low/medium/high/max).
+//
+// supportsMax indicates whether the target model supports "max" effort.
+// Returns the mapped effort and true if the level is valid, or ("", false) otherwise.
+func MapToClaudeEffort(level string, supportsMax bool) (string, bool) {
+	level = strings.ToLower(strings.TrimSpace(level))
+	switch level {
+	case "":
+		return "", false
+	case "minimal":
+		return "low", true
+	case "low", "medium", "high":
+		return level, true
+	case "xhigh", "max":
+		if supportsMax {
+			return "max", true
+		}
+		return "high", true
+	case "auto":
+		return "high", true
+	default:
+		return "", false
+	}
+}
+
 // ModelCapability describes the thinking format support of a model.
 type ModelCapability int
 

internal/thinking/provider/claude/apply.go

@@ -1,8 +1,10 @@
 // Package claude implements thinking configuration scaffolding for Claude models.
 //
-// Claude models use the thinking.budget_tokens format with values in the range
-// 1024-128000. Some Claude models support ZeroAllowed (sonnet-4-5, opus-4-5),
-// while older models do not.
+// Claude models support two thinking control styles:
+//   - Manual thinking: thinking.type="enabled" with thinking.budget_tokens (token budget)
+//   - Adaptive thinking (Claude 4.6): thinking.type="adaptive" with output_config.effort (low/medium/high/max)
+//
+// Some Claude models support ZeroAllowed (sonnet-4-5, opus-4-5), while older models do not.
 // See: _bmad-output/planning-artifacts/architecture.md#Epic-6
 package claude
 
@@ -34,7 +36,11 @@ func init() {
 //   - Budget clamping to model range
 //   - ZeroAllowed constraint enforcement
 //
-// Apply only processes ModeBudget and ModeNone; other modes are passed through unchanged.
+// Apply processes:
+//   - ModeBudget: manual thinking budget_tokens
+//   - ModeLevel: adaptive thinking effort (Claude 4.6)
+//   - ModeAuto: provider default adaptive/manual behavior
+//   - ModeNone: disabled
 //
 // Expected output format when enabled:
 //
@@ -45,6 +51,17 @@ func init() {
 //	  }
 //	}
 //
+// Expected output format for adaptive:
+//
+//	{
+//	  "thinking": {
+//	    "type": "adaptive"
+//	  },
+//	  "output_config": {
+//	    "effort": "high"
+//	  }
+//	}
+//
 // Expected output format when disabled:
 //
 //	{
@@ -60,30 +77,91 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 		return body, nil
 	}
 
-	// Only process ModeBudget and ModeNone; other modes pass through
-	// (caller should use ValidateConfig first to normalize modes)
-	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone {
-		return body, nil
-	}
-
 	if len(body) == 0 || !gjson.ValidBytes(body) {
 		body = []byte(`{}`)
 	}
 
-	// Budget is expected to be pre-validated by ValidateConfig (clamped, ZeroAllowed enforced)
-	// Decide enabled/disabled based on budget value
-	if config.Budget == 0 {
+	supportsAdaptive := modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) > 0
+
+	switch config.Mode {
+	case thinking.ModeNone:
 		result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
 		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
 		return result, nil
+
+	case thinking.ModeLevel:
+		// Adaptive thinking effort is only valid when the model advertises discrete levels.
+		// (Claude 4.6 uses output_config.effort.)
+		if supportsAdaptive && config.Level != "" {
+			result, _ := sjson.SetBytes(body, "thinking.type", "adaptive")
+			result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+			result, _ = sjson.SetBytes(result, "output_config.effort", string(config.Level))
+			return result, nil
+		}
+
+		// Fallback for non-adaptive Claude models: convert level to budget_tokens.
+		if budget, ok := thinking.ConvertLevelToBudget(string(config.Level)); ok {
+			config.Mode = thinking.ModeBudget
+			config.Budget = budget
+			config.Level = ""
+		} else {
+			return body, nil
+		}
+		fallthrough
+
+	case thinking.ModeBudget:
+		// Budget is expected to be pre-validated by ValidateConfig (clamped, ZeroAllowed enforced).
+		// Decide enabled/disabled based on budget value.
+		if config.Budget == 0 {
+			result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
+			result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+			result, _ = sjson.DeleteBytes(result, "output_config.effort")
+			if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+				result, _ = sjson.DeleteBytes(result, "output_config")
+			}
+			return result, nil
+		}
+
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
+
+		// Ensure max_tokens > thinking.budget_tokens (Anthropic API constraint).
+		result = a.normalizeClaudeBudget(result, config.Budget, modelInfo)
+		return result, nil
+
+	case thinking.ModeAuto:
+		// For Claude 4.6 models, auto maps to adaptive thinking with upstream defaults.
+		if supportsAdaptive {
+			result, _ := sjson.SetBytes(body, "thinking.type", "adaptive")
+			result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+			// Explicit effort is optional for adaptive thinking; omit it to allow upstream default.
+			result, _ = sjson.DeleteBytes(result, "output_config.effort")
+			if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+				result, _ = sjson.DeleteBytes(result, "output_config")
+			}
+			return result, nil
+		}
+
+		// Legacy fallback: enable thinking without specifying budget_tokens.
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
+		return result, nil
+
+	default:
+		return body, nil
 	}
-
-	result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
-	result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
-
-	// Ensure max_tokens > thinking.budget_tokens (Anthropic API constraint)
-	result = a.normalizeClaudeBudget(result, config.Budget, modelInfo)
-	return result, nil
 }
 
 // normalizeClaudeBudget applies Claude-specific constraints to ensure max_tokens > budget_tokens.
@@ -141,7 +219,7 @@ func (a *Applier) effectiveMaxTokens(body []byte, modelInfo *registry.ModelInfo)
 }
 
 func applyCompatibleClaude(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
-	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto && config.Mode != thinking.ModeLevel {
 		return body, nil
 	}
 
@@ -153,14 +231,36 @@ func applyCompatibleClaude(body []byte, config thinking.ThinkingConfig) ([]byte,
 	case thinking.ModeNone:
 		result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
 		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
 		return result, nil
 	case thinking.ModeAuto:
 		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
 		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
+		return result, nil
+	case thinking.ModeLevel:
+		// For user-defined models, interpret ModeLevel as Claude adaptive thinking effort.
+		// Upstream is responsible for validating whether the target model supports it.
+		if config.Level == "" {
+			return body, nil
+		}
+		result, _ := sjson.SetBytes(body, "thinking.type", "adaptive")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		result, _ = sjson.SetBytes(result, "output_config.effort", string(config.Level))
 		return result, nil
 	default:
 		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
 		result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
 		return result, nil
 	}
 }

internal/thinking/provider/codex/apply.go

@@ -7,8 +7,6 @@
 package codex
 
 import (
-	"strings"
-
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
@@ -68,7 +66,7 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 	effort := ""
 	support := modelInfo.Thinking
 	if config.Budget == 0 {
-		if support.ZeroAllowed || hasLevel(support.Levels, string(thinking.LevelNone)) {
+		if support.ZeroAllowed || thinking.HasLevel(support.Levels, string(thinking.LevelNone)) {
 			effort = string(thinking.LevelNone)
 		}
 	}
@@ -120,12 +118,3 @@ func applyCompatibleCodex(body []byte, config thinking.ThinkingConfig) ([]byte,
 	result, _ := sjson.SetBytes(body, "reasoning.effort", effort)
 	return result, nil
 }
-
-func hasLevel(levels []string, target string) bool {
-	for _, level := range levels {
-		if strings.EqualFold(strings.TrimSpace(level), target) {
-			return true
-		}
-	}
-	return false
-}

internal/thinking/provider/openai/apply.go

@@ -6,8 +6,6 @@
 package openai
 
 import (
-	"strings"
-
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
@@ -65,7 +63,7 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 	effort := ""
 	support := modelInfo.Thinking
 	if config.Budget == 0 {
-		if support.ZeroAllowed || hasLevel(support.Levels, string(thinking.LevelNone)) {
+		if support.ZeroAllowed || thinking.HasLevel(support.Levels, string(thinking.LevelNone)) {
 			effort = string(thinking.LevelNone)
 		}
 	}
@@ -117,12 +115,3 @@ func applyCompatibleOpenAI(body []byte, config thinking.ThinkingConfig) ([]byte,
 	result, _ := sjson.SetBytes(body, "reasoning_effort", effort)
 	return result, nil
 }
-
-func hasLevel(levels []string, target string) bool {
-	for _, level := range levels {
-		if strings.EqualFold(strings.TrimSpace(level), target) {
-			return true
-		}
-	}
-	return false
-}

internal/thinking/strip.go

@@ -30,7 +30,7 @@ func StripThinkingConfig(body []byte, provider string) []byte {
 	var paths []string
 	switch provider {
 	case "claude":
-		paths = []string{"thinking"}
+		paths = []string{"thinking", "output_config.effort"}
 	case "gemini":
 		paths = []string{"generationConfig.thinkingConfig"}
 	case "gemini-cli", "antigravity":
@@ -59,5 +59,12 @@ func StripThinkingConfig(body []byte, provider string) []byte {
 	for _, path := range paths {
 		result, _ = sjson.DeleteBytes(result, path)
 	}
+
+	// Avoid leaving an empty output_config object for Claude when effort was the only field.
+	if provider == "claude" {
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
+	}
 	return result
 }

internal/thinking/suffix.go

@@ -109,7 +109,7 @@ func ParseSpecialSuffix(rawSuffix string) (mode ThinkingMode, ok bool) {
 // ParseLevelSuffix attempts to parse a raw suffix as a discrete thinking level.
 //
 // This function parses the raw suffix content (from ParseSuffix.RawSuffix) as a level.
-// Only discrete effort levels are valid: minimal, low, medium, high, xhigh.
+// Only discrete effort levels are valid: minimal, low, medium, high, xhigh, max.
 // Level matching is case-insensitive.
 //
 // Special values (none, auto) are NOT handled by this function; use ParseSpecialSuffix
@@ -140,6 +140,8 @@ func ParseLevelSuffix(rawSuffix string) (level ThinkingLevel, ok bool) {
 		return LevelHigh, true
 	case "xhigh":
 		return LevelXHigh, true
+	case "max":
+		return LevelMax, true
 	default:
 		return "", false
 	}

internal/thinking/types.go

@@ -54,6 +54,9 @@ const (
 	LevelHigh ThinkingLevel = "high"
 	// LevelXHigh sets extra-high thinking effort
 	LevelXHigh ThinkingLevel = "xhigh"
+	// LevelMax sets maximum thinking effort.
+	// This is currently used by Claude 4.6 adaptive thinking (opus supports "max").
+	LevelMax ThinkingLevel = "max"
 )
 
 // ThinkingConfig represents a unified thinking configuration.

internal/thinking/validate.go

@@ -53,7 +53,17 @@ func ValidateConfig(config ThinkingConfig, modelInfo *registry.ModelInfo, fromFo
 		return &config, nil
 	}
 
-	allowClampUnsupported := isBudgetBasedProvider(fromFormat) && isLevelBasedProvider(toFormat)
+	// allowClampUnsupported determines whether to clamp unsupported levels instead of returning an error.
+	// This applies when crossing provider families (e.g., openai→gemini, claude→gemini) and the target
+	// model supports discrete levels. Same-family conversions require strict validation.
+	toCapability := detectModelCapability(modelInfo)
+	toHasLevelSupport := toCapability == CapabilityLevelOnly || toCapability == CapabilityHybrid
+	allowClampUnsupported := toHasLevelSupport && !isSameProviderFamily(fromFormat, toFormat)
+
+	// strictBudget determines whether to enforce strict budget range validation.
+	// This applies when: (1) config comes from request body (not suffix), (2) source format is known,
+	// and (3) source and target are in the same provider family. Cross-family or suffix-based configs
+	// are clamped instead of rejected to improve interoperability.
 	strictBudget := !fromSuffix && fromFormat != "" && isSameProviderFamily(fromFormat, toFormat)
 	budgetDerivedFromLevel := false
 
@@ -201,7 +211,7 @@ func convertAutoToMidRange(config ThinkingConfig, support *registry.ThinkingSupp
 }
 
 // standardLevelOrder defines the canonical ordering of thinking levels from lowest to highest.
-var standardLevelOrder = []ThinkingLevel{LevelMinimal, LevelLow, LevelMedium, LevelHigh, LevelXHigh}
+var standardLevelOrder = []ThinkingLevel{LevelMinimal, LevelLow, LevelMedium, LevelHigh, LevelXHigh, LevelMax}
 
 // clampLevel clamps the given level to the nearest supported level.
 // On tie, prefers the lower level.
@@ -325,7 +335,9 @@ func normalizeLevels(levels []string) []string {
 	return out
 }
 
-func isBudgetBasedProvider(provider string) bool {
+// isBudgetCapableProvider returns true if the provider supports budget-based thinking.
+// These providers may also support level-based thinking (hybrid models).
+func isBudgetCapableProvider(provider string) bool {
 	switch provider {
 	case "gemini", "gemini-cli", "antigravity", "claude":
 		return true
@@ -334,15 +346,6 @@ func isBudgetBasedProvider(provider string) bool {
 	}
 }
 
-func isLevelBasedProvider(provider string) bool {
-	switch provider {
-	case "openai", "openai-response", "codex":
-		return true
-	default:
-		return false
-	}
-}
-
 func isGeminiFamily(provider string) bool {
 	switch provider {
 	case "gemini", "gemini-cli", "antigravity":
@@ -352,11 +355,21 @@ func isGeminiFamily(provider string) bool {
 	}
 }
 
+func isOpenAIFamily(provider string) bool {
+	switch provider {
+	case "openai", "openai-response", "codex":
+		return true
+	default:
+		return false
+	}
+}
+
 func isSameProviderFamily(from, to string) bool {
 	if from == to {
 		return true
 	}
-	return isGeminiFamily(from) && isGeminiFamily(to)
+	return (isGeminiFamily(from) && isGeminiFamily(to)) ||
+		(isOpenAIFamily(from) && isOpenAIFamily(to))
 }
 
 func abs(x int) int {

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -431,6 +431,33 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 		out, _ = sjson.SetRaw(out, "request.tools", toolsJSON)
 	}
 
+	// tool_choice
+	toolChoiceResult := gjson.GetBytes(rawJSON, "tool_choice")
+	if toolChoiceResult.Exists() {
+		toolChoiceType := ""
+		toolChoiceName := ""
+		if toolChoiceResult.IsObject() {
+			toolChoiceType = toolChoiceResult.Get("type").String()
+			toolChoiceName = toolChoiceResult.Get("name").String()
+		} else if toolChoiceResult.Type == gjson.String {
+			toolChoiceType = toolChoiceResult.String()
+		}
+
+		switch toolChoiceType {
+		case "auto":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "AUTO")
+		case "none":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "NONE")
+		case "any":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "ANY")
+		case "tool":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "ANY")
+			if toolChoiceName != "" {
+				out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.allowedFunctionNames", []string{toolChoiceName})
+			}
+		}
+	}
+
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
 	if t := gjson.GetBytes(rawJSON, "thinking"); enableThoughtTranslate && t.Exists() && t.IsObject() {
 		switch t.Get("type").String() {
@@ -441,9 +468,22 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		case "adaptive", "auto":
-			// Keep adaptive/auto as a high level sentinel; ApplyThinking resolves it
-			// to model-specific max capability.
-			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			// For adaptive thinking:
+			// - If output_config.effort is explicitly present, pass through as thinkingLevel.
+			// - Otherwise, treat it as "enabled with target-model maximum" and emit high.
+			// ApplyThinking handles clamping to target model's supported levels.
+			effort := ""
+			if v := gjson.GetBytes(rawJSON, "output_config.effort"); v.Exists() && v.Type == gjson.String {
+				effort = strings.ToLower(strings.TrimSpace(v.String()))
+			}
+			if effort != "" {
+				if effort == "max" {
+					effort = "high"
+				}
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", effort)
+			} else {
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			}
 			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}

internal/translator/antigravity/claude/antigravity_claude_request_test.go

@@ -193,6 +193,42 @@ func TestConvertClaudeRequestToAntigravity_ToolDeclarations(t *testing.T) {
 	}
 }
 
+func TestConvertClaudeRequestToAntigravity_ToolChoice_SpecificTool(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "gemini-3-flash-preview",
+		"messages": [
+			{
+				"role": "user",
+				"content": [
+					{"type": "text", "text": "hi"}
+				]
+			}
+		],
+		"tools": [
+			{
+				"name": "json",
+				"description": "A JSON tool",
+				"input_schema": {
+					"type": "object",
+					"properties": {}
+				}
+			}
+		],
+		"tool_choice": {"type": "tool", "name": "json"}
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("gemini-3-flash-preview", inputJSON, false)
+	outputStr := string(output)
+
+	if got := gjson.Get(outputStr, "request.toolConfig.functionCallingConfig.mode").String(); got != "ANY" {
+		t.Fatalf("Expected toolConfig.functionCallingConfig.mode 'ANY', got '%s'", got)
+	}
+	allowed := gjson.Get(outputStr, "request.toolConfig.functionCallingConfig.allowedFunctionNames").Array()
+	if len(allowed) != 1 || allowed[0].String() != "json" {
+		t.Fatalf("Expected allowedFunctionNames ['json'], got %s", gjson.Get(outputStr, "request.toolConfig.functionCallingConfig.allowedFunctionNames").Raw)
+	}
+}
+
 func TestConvertClaudeRequestToAntigravity_ToolUse(t *testing.T) {
 	inputJSON := []byte(`{
 		"model": "claude-3-5-sonnet-20240620",
@@ -1199,3 +1235,64 @@ func TestConvertClaudeRequestToAntigravity_ToolAndThinking_NoExistingSystem(t *t
 		t.Errorf("Interleaved thinking hint should be in created systemInstruction, got: %v", sysInstruction.Raw)
 	}
 }
+
+func TestConvertClaudeRequestToAntigravity_AdaptiveThinking_EffortLevels(t *testing.T) {
+	tests := []struct {
+		name     string
+		effort   string
+		expected string
+	}{
+		{"low", "low", "low"},
+		{"medium", "medium", "medium"},
+		{"high", "high", "high"},
+		{"max", "max", "high"},
+	}
+
+	for _, tt := range tests {
+		tt := tt
+		t.Run(tt.name, func(t *testing.T) {
+			inputJSON := []byte(`{
+				"model": "claude-opus-4-6-thinking",
+				"messages": [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}],
+				"thinking": {"type": "adaptive"},
+				"output_config": {"effort": "` + tt.effort + `"}
+			}`)
+
+			output := ConvertClaudeRequestToAntigravity("claude-opus-4-6-thinking", inputJSON, false)
+			outputStr := string(output)
+
+			thinkingConfig := gjson.Get(outputStr, "request.generationConfig.thinkingConfig")
+			if !thinkingConfig.Exists() {
+				t.Fatal("thinkingConfig should exist for adaptive thinking")
+			}
+			if thinkingConfig.Get("thinkingLevel").String() != tt.expected {
+				t.Errorf("Expected thinkingLevel %q, got %q", tt.expected, thinkingConfig.Get("thinkingLevel").String())
+			}
+			if !thinkingConfig.Get("includeThoughts").Bool() {
+				t.Error("includeThoughts should be true")
+			}
+		})
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_AdaptiveThinking_NoEffort(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "claude-opus-4-6-thinking",
+		"messages": [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}],
+		"thinking": {"type": "adaptive"}
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-opus-4-6-thinking", inputJSON, false)
+	outputStr := string(output)
+
+	thinkingConfig := gjson.Get(outputStr, "request.generationConfig.thinkingConfig")
+	if !thinkingConfig.Exists() {
+		t.Fatal("thinkingConfig should exist for adaptive thinking without effort")
+	}
+	if thinkingConfig.Get("thinkingLevel").String() != "high" {
+		t.Errorf("Expected default thinkingLevel \"high\", got %q", thinkingConfig.Get("thinkingLevel").String())
+	}
+	if !thinkingConfig.Get("includeThoughts").Bool() {
+		t.Error("includeThoughts should be true")
+	}
+}

internal/translator/claude/gemini/claude_gemini_request.go

@@ -14,6 +14,7 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
@@ -115,24 +116,47 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 		// Include thoughts configuration for reasoning process visibility
 		// Translator only does format conversion, ApplyThinking handles model capability validation.
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
+			mi := registry.LookupModelInfo(modelName, "claude")
+			supportsAdaptive := mi != nil && mi.Thinking != nil && len(mi.Thinking.Levels) > 0
+			supportsMax := supportsAdaptive && thinking.HasLevel(mi.Thinking.Levels, string(thinking.LevelMax))
+
+			// MapToClaudeEffort normalizes levels (e.g. minimal→low, xhigh→high) to avoid
+			// validation errors since validate treats same-provider unsupported levels as errors.
 			thinkingLevel := thinkingConfig.Get("thinkingLevel")
 			if !thinkingLevel.Exists() {
 				thinkingLevel = thinkingConfig.Get("thinking_level")
 			}
 			if thinkingLevel.Exists() {
 				level := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
-				switch level {
-				case "":
-				case "none":
-					out, _ = sjson.Set(out, "thinking.type", "disabled")
-					out, _ = sjson.Delete(out, "thinking.budget_tokens")
-				case "auto":
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
-					out, _ = sjson.Delete(out, "thinking.budget_tokens")
-				default:
-					if budget, ok := thinking.ConvertLevelToBudget(level); ok {
+				if supportsAdaptive {
+					switch level {
+					case "":
+					case "none":
+						out, _ = sjson.Set(out, "thinking.type", "disabled")
+						out, _ = sjson.Delete(out, "thinking.budget_tokens")
+						out, _ = sjson.Delete(out, "output_config.effort")
+					default:
+						if mapped, ok := thinking.MapToClaudeEffort(level, supportsMax); ok {
+							level = mapped
+						}
+						out, _ = sjson.Set(out, "thinking.type", "adaptive")
+						out, _ = sjson.Delete(out, "thinking.budget_tokens")
+						out, _ = sjson.Set(out, "output_config.effort", level)
+					}
+				} else {
+					switch level {
+					case "":
+					case "none":
+						out, _ = sjson.Set(out, "thinking.type", "disabled")
+						out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					case "auto":
 						out, _ = sjson.Set(out, "thinking.type", "enabled")
-						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+						out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					default:
+						if budget, ok := thinking.ConvertLevelToBudget(level); ok {
+							out, _ = sjson.Set(out, "thinking.type", "enabled")
+							out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+						}
 					}
 				}
 			} else {
@@ -142,16 +166,35 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 				}
 				if thinkingBudget.Exists() {
 					budget := int(thinkingBudget.Int())
-					switch budget {
-					case 0:
-						out, _ = sjson.Set(out, "thinking.type", "disabled")
-						out, _ = sjson.Delete(out, "thinking.budget_tokens")
-					case -1:
-						out, _ = sjson.Set(out, "thinking.type", "enabled")
-						out, _ = sjson.Delete(out, "thinking.budget_tokens")
-					default:
-						out, _ = sjson.Set(out, "thinking.type", "enabled")
-						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					if supportsAdaptive {
+						switch budget {
+						case 0:
+							out, _ = sjson.Set(out, "thinking.type", "disabled")
+							out, _ = sjson.Delete(out, "thinking.budget_tokens")
+							out, _ = sjson.Delete(out, "output_config.effort")
+						default:
+							level, ok := thinking.ConvertBudgetToLevel(budget)
+							if ok {
+								if mapped, okM := thinking.MapToClaudeEffort(level, supportsMax); okM {
+									level = mapped
+								}
+								out, _ = sjson.Set(out, "thinking.type", "adaptive")
+								out, _ = sjson.Delete(out, "thinking.budget_tokens")
+								out, _ = sjson.Set(out, "output_config.effort", level)
+							}
+						}
+					} else {
+						switch budget {
+						case 0:
+							out, _ = sjson.Set(out, "thinking.type", "disabled")
+							out, _ = sjson.Delete(out, "thinking.budget_tokens")
+						case -1:
+							out, _ = sjson.Set(out, "thinking.type", "enabled")
+							out, _ = sjson.Delete(out, "thinking.budget_tokens")
+						default:
+							out, _ = sjson.Set(out, "thinking.type", "enabled")
+							out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+						}
 					}
 				} else if includeThoughts := thinkingConfig.Get("includeThoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
 					out, _ = sjson.Set(out, "thinking.type", "enabled")

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

@@ -14,6 +14,7 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
+	"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"
@@ -68,17 +69,45 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 	if v := root.Get("reasoning_effort"); v.Exists() {
 		effort := strings.ToLower(strings.TrimSpace(v.String()))
 		if effort != "" {
-			budget, ok := thinking.ConvertLevelToBudget(effort)
-			if ok {
-				switch budget {
-				case 0:
+			mi := registry.LookupModelInfo(modelName, "claude")
+			supportsAdaptive := mi != nil && mi.Thinking != nil && len(mi.Thinking.Levels) > 0
+			supportsMax := supportsAdaptive && thinking.HasLevel(mi.Thinking.Levels, string(thinking.LevelMax))
+
+			// Claude 4.6 supports adaptive thinking with output_config.effort.
+			// MapToClaudeEffort normalizes levels (e.g. minimal→low, xhigh→high) to avoid
+			// validation errors since validate treats same-provider unsupported levels as errors.
+			if supportsAdaptive {
+				switch effort {
+				case "none":
 					out, _ = sjson.Set(out, "thinking.type", "disabled")
-				case -1:
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Delete(out, "output_config.effort")
+				case "auto":
+					out, _ = sjson.Set(out, "thinking.type", "adaptive")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Delete(out, "output_config.effort")
 				default:
-					if budget > 0 {
+					if mapped, ok := thinking.MapToClaudeEffort(effort, supportsMax); ok {
+						effort = mapped
+					}
+					out, _ = sjson.Set(out, "thinking.type", "adaptive")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Set(out, "output_config.effort", effort)
+				}
+			} else {
+				// Legacy/manual thinking (budget_tokens).
+				budget, ok := thinking.ConvertLevelToBudget(effort)
+				if ok {
+					switch budget {
+					case 0:
+						out, _ = sjson.Set(out, "thinking.type", "disabled")
+					case -1:
 						out, _ = sjson.Set(out, "thinking.type", "enabled")
-						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					default:
+						if budget > 0 {
+							out, _ = sjson.Set(out, "thinking.type", "enabled")
+							out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+						}
 					}
 				}
 			}

internal/translator/claude/openai/responses/claude_openai-responses_request.go

@@ -9,6 +9,7 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
+	"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"
@@ -56,17 +57,45 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 	if v := root.Get("reasoning.effort"); v.Exists() {
 		effort := strings.ToLower(strings.TrimSpace(v.String()))
 		if effort != "" {
-			budget, ok := thinking.ConvertLevelToBudget(effort)
-			if ok {
-				switch budget {
-				case 0:
+			mi := registry.LookupModelInfo(modelName, "claude")
+			supportsAdaptive := mi != nil && mi.Thinking != nil && len(mi.Thinking.Levels) > 0
+			supportsMax := supportsAdaptive && thinking.HasLevel(mi.Thinking.Levels, string(thinking.LevelMax))
+
+			// Claude 4.6 supports adaptive thinking with output_config.effort.
+			// MapToClaudeEffort normalizes levels (e.g. minimal→low, xhigh→high) to avoid
+			// validation errors since validate treats same-provider unsupported levels as errors.
+			if supportsAdaptive {
+				switch effort {
+				case "none":
 					out, _ = sjson.Set(out, "thinking.type", "disabled")
-				case -1:
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Delete(out, "output_config.effort")
+				case "auto":
+					out, _ = sjson.Set(out, "thinking.type", "adaptive")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Delete(out, "output_config.effort")
 				default:
-					if budget > 0 {
+					if mapped, ok := thinking.MapToClaudeEffort(effort, supportsMax); ok {
+						effort = mapped
+					}
+					out, _ = sjson.Set(out, "thinking.type", "adaptive")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Set(out, "output_config.effort", effort)
+				}
+			} else {
+				// Legacy/manual thinking (budget_tokens).
+				budget, ok := thinking.ConvertLevelToBudget(effort)
+				if ok {
+					switch budget {
+					case 0:
+						out, _ = sjson.Set(out, "thinking.type", "disabled")
+					case -1:
 						out, _ = sjson.Set(out, "thinking.type", "enabled")
-						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					default:
+						if budget > 0 {
+							out, _ = sjson.Set(out, "thinking.type", "enabled")
+							out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+						}
 					}
 				}
 			}

internal/translator/codex/claude/codex_claude_request.go

@@ -160,7 +160,51 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 						flushMessage()
 						functionCallOutputMessage := `{"type":"function_call_output"}`
 						functionCallOutputMessage, _ = sjson.Set(functionCallOutputMessage, "call_id", messageContentResult.Get("tool_use_id").String())
-						functionCallOutputMessage, _ = sjson.Set(functionCallOutputMessage, "output", messageContentResult.Get("content").String())
+
+						contentResult := messageContentResult.Get("content")
+						if contentResult.IsArray() {
+							toolResultContentIndex := 0
+							toolResultContent := `[]`
+							contentResults := contentResult.Array()
+							for k := 0; k < len(contentResults); k++ {
+								toolResultContentType := contentResults[k].Get("type").String()
+								if toolResultContentType == "image" {
+									sourceResult := contentResults[k].Get("source")
+									if sourceResult.Exists() {
+										data := sourceResult.Get("data").String()
+										if data == "" {
+											data = sourceResult.Get("base64").String()
+										}
+										if data != "" {
+											mediaType := sourceResult.Get("media_type").String()
+											if mediaType == "" {
+												mediaType = sourceResult.Get("mime_type").String()
+											}
+											if mediaType == "" {
+												mediaType = "application/octet-stream"
+											}
+											dataURL := fmt.Sprintf("data:%s;base64,%s", mediaType, data)
+
+											toolResultContent, _ = sjson.Set(toolResultContent, fmt.Sprintf("%d.type", toolResultContentIndex), "input_image")
+											toolResultContent, _ = sjson.Set(toolResultContent, fmt.Sprintf("%d.image_url", toolResultContentIndex), dataURL)
+											toolResultContentIndex++
+										}
+									}
+								} else if toolResultContentType == "text" {
+									toolResultContent, _ = sjson.Set(toolResultContent, fmt.Sprintf("%d.type", toolResultContentIndex), "input_text")
+									toolResultContent, _ = sjson.Set(toolResultContent, fmt.Sprintf("%d.text", toolResultContentIndex), contentResults[k].Get("text").String())
+									toolResultContentIndex++
+								}
+							}
+							if toolResultContent != `[]` {
+								functionCallOutputMessage, _ = sjson.SetRaw(functionCallOutputMessage, "output", toolResultContent)
+							} else {
+								functionCallOutputMessage, _ = sjson.Set(functionCallOutputMessage, "output", messageContentResult.Get("content").String())
+							}
+						} else {
+							functionCallOutputMessage, _ = sjson.Set(functionCallOutputMessage, "output", messageContentResult.Get("content").String())
+						}
+
 						template, _ = sjson.SetRaw(template, "input.-1", functionCallOutputMessage)
 					}
 				}
@@ -211,6 +255,8 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 			tool, _ = sjson.SetRaw(tool, "parameters", normalizeToolParameters(toolResult.Get("input_schema").Raw))
 			tool, _ = sjson.Delete(tool, "input_schema")
 			tool, _ = sjson.Delete(tool, "parameters.$schema")
+			tool, _ = sjson.Delete(tool, "cache_control")
+			tool, _ = sjson.Delete(tool, "defer_loading")
 			tool, _ = sjson.Set(tool, "strict", false)
 			template, _ = sjson.SetRaw(template, "tools.-1", tool)
 		}
@@ -231,9 +277,17 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 				}
 			}
 		case "adaptive", "auto":
-			// Claude adaptive/auto means "enable with max capacity"; keep it as highest level
-			// and let ApplyThinking normalize per target model capability.
-			reasoningEffort = string(thinking.LevelXHigh)
+			// Adaptive thinking can carry an explicit effort in output_config.effort (Claude 4.6).
+			// Pass through directly; ApplyThinking handles clamping to target model's levels.
+			effort := ""
+			if v := rootResult.Get("output_config.effort"); v.Exists() && v.Type == gjson.String {
+				effort = strings.ToLower(strings.TrimSpace(v.String()))
+			}
+			if effort != "" {
+				reasoningEffort = effort
+			} else {
+				reasoningEffort = string(thinking.LevelXHigh)
+			}
 		case "disabled":
 			if effort, ok := thinking.ConvertBudgetToLevel(0); ok && effort != "" {
 				reasoningEffort = effort

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

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

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

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

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

@@ -156,6 +156,7 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 				tool, _ = sjson.Delete(tool, "input_examples")
 				tool, _ = sjson.Delete(tool, "type")
 				tool, _ = sjson.Delete(tool, "cache_control")
+				tool, _ = sjson.Delete(tool, "defer_loading")
 				if gjson.Valid(tool) && gjson.Parse(tool).IsObject() {
 					if !hasTools {
 						out, _ = sjson.SetRaw(out, "request.tools", `[{"functionDeclarations":[]}]`)
@@ -171,7 +172,35 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 		}
 	}
 
-	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
+	// tool_choice
+	toolChoiceResult := gjson.GetBytes(rawJSON, "tool_choice")
+	if toolChoiceResult.Exists() {
+		toolChoiceType := ""
+		toolChoiceName := ""
+		if toolChoiceResult.IsObject() {
+			toolChoiceType = toolChoiceResult.Get("type").String()
+			toolChoiceName = toolChoiceResult.Get("name").String()
+		} else if toolChoiceResult.Type == gjson.String {
+			toolChoiceType = toolChoiceResult.String()
+		}
+
+		switch toolChoiceType {
+		case "auto":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "AUTO")
+		case "none":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "NONE")
+		case "any":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "ANY")
+		case "tool":
+			out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.mode", "ANY")
+			if toolChoiceName != "" {
+				out, _ = sjson.Set(out, "request.toolConfig.functionCallingConfig.allowedFunctionNames", []string{toolChoiceName})
+			}
+		}
+	}
+
+	// Map Anthropic thinking -> Gemini CLI thinkingConfig when enabled
+	// Translator only does format conversion, ApplyThinking handles model capability validation.
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
 		switch t.Get("type").String() {
 		case "enabled":
@@ -181,9 +210,19 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		case "adaptive", "auto":
-			// Keep adaptive/auto as a high level sentinel; ApplyThinking resolves it
-			// to model-specific max capability.
-			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			// For adaptive thinking:
+			// - If output_config.effort is explicitly present, pass through as thinkingLevel.
+			// - Otherwise, treat it as "enabled with target-model maximum" and emit high.
+			// ApplyThinking handles clamping to target model's supported levels.
+			effort := ""
+			if v := gjson.GetBytes(rawJSON, "output_config.effort"); v.Exists() && v.Type == gjson.String {
+				effort = strings.ToLower(strings.TrimSpace(v.String()))
+			}
+			if effort != "" {
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", effort)
+			} else {
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			}
 			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}

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

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

internal/translator/gemini/claude/gemini_claude_request.go

@@ -9,6 +9,7 @@ import (
 	"bytes"
 	"strings"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -136,6 +137,7 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 				tool, _ = sjson.Delete(tool, "input_examples")
 				tool, _ = sjson.Delete(tool, "type")
 				tool, _ = sjson.Delete(tool, "cache_control")
+				tool, _ = sjson.Delete(tool, "defer_loading")
 				if gjson.Valid(tool) && gjson.Parse(tool).IsObject() {
 					if !hasTools {
 						out, _ = sjson.SetRaw(out, "tools", `[{"functionDeclarations":[]}]`)
@@ -151,7 +153,34 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 		}
 	}
 
-	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when enabled
+	// tool_choice
+	toolChoiceResult := gjson.GetBytes(rawJSON, "tool_choice")
+	if toolChoiceResult.Exists() {
+		toolChoiceType := ""
+		toolChoiceName := ""
+		if toolChoiceResult.IsObject() {
+			toolChoiceType = toolChoiceResult.Get("type").String()
+			toolChoiceName = toolChoiceResult.Get("name").String()
+		} else if toolChoiceResult.Type == gjson.String {
+			toolChoiceType = toolChoiceResult.String()
+		}
+
+		switch toolChoiceType {
+		case "auto":
+			out, _ = sjson.Set(out, "toolConfig.functionCallingConfig.mode", "AUTO")
+		case "none":
+			out, _ = sjson.Set(out, "toolConfig.functionCallingConfig.mode", "NONE")
+		case "any":
+			out, _ = sjson.Set(out, "toolConfig.functionCallingConfig.mode", "ANY")
+		case "tool":
+			out, _ = sjson.Set(out, "toolConfig.functionCallingConfig.mode", "ANY")
+			if toolChoiceName != "" {
+				out, _ = sjson.Set(out, "toolConfig.functionCallingConfig.allowedFunctionNames", []string{toolChoiceName})
+			}
+		}
+	}
+
+	// Map Anthropic thinking -> Gemini thinking config when enabled
 	// Translator only does format conversion, ApplyThinking handles model capability validation.
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
 		switch t.Get("type").String() {
@@ -162,9 +191,27 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		case "adaptive", "auto":
-			// Keep adaptive/auto as a high level sentinel; ApplyThinking resolves it
-			// to model-specific max capability.
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingLevel", "high")
+			// For adaptive thinking:
+			// - If output_config.effort is explicitly present, pass through as thinkingLevel.
+			// - Otherwise, treat it as "enabled with target-model maximum" and emit thinkingBudget=max.
+			// ApplyThinking handles clamping to target model's supported levels.
+			effort := ""
+			if v := gjson.GetBytes(rawJSON, "output_config.effort"); v.Exists() && v.Type == gjson.String {
+				effort = strings.ToLower(strings.TrimSpace(v.String()))
+			}
+			if effort != "" {
+				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingLevel", effort)
+			} else {
+				maxBudget := 0
+				if mi := registry.LookupModelInfo(modelName, "gemini"); mi != nil && mi.Thinking != nil {
+					maxBudget = mi.Thinking.Max
+				}
+				if maxBudget > 0 {
+					out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", maxBudget)
+				} else {
+					out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingLevel", "high")
+				}
+			}
 			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}

internal/translator/gemini/claude/gemini_claude_request_test.go

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

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

@@ -354,22 +354,7 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 					funcDecl, _ = sjson.Set(funcDecl, "description", desc.String())
 				}
 				if params := tool.Get("parameters"); params.Exists() {
-					// Convert parameter types from OpenAI format to Gemini format
-					cleaned := params.Raw
-					// Convert type values to uppercase for Gemini
-					paramsResult := gjson.Parse(cleaned)
-					if properties := paramsResult.Get("properties"); properties.Exists() {
-						properties.ForEach(func(key, value gjson.Result) bool {
-							if propType := value.Get("type"); propType.Exists() {
-								upperType := strings.ToUpper(propType.String())
-								cleaned, _ = sjson.Set(cleaned, "properties."+key.String()+".type", upperType)
-							}
-							return true
-						})
-					}
-					// Set the overall type to OBJECT
-					cleaned, _ = sjson.Set(cleaned, "type", "OBJECT")
-					funcDecl, _ = sjson.SetRaw(funcDecl, "parametersJsonSchema", cleaned)
+					funcDecl, _ = sjson.SetRaw(funcDecl, "parametersJsonSchema", params.Raw)
 				}
 
 				geminiTools, _ = sjson.SetRaw(geminiTools, "0.functionDeclarations.-1", funcDecl)

internal/translator/kiro/claude/kiro_claude_request.go

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

internal/translator/kiro/claude/kiro_claude_response.go

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

internal/translator/kiro/claude/kiro_claude_stream_parser.go

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

internal/translator/kiro/claude/tool_compression.go

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

internal/translator/kiro/claude/truncation_detector.go

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

internal/translator/kiro/common/constants.go

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

internal/translator/openai/claude/openai_claude_request.go

@@ -76,9 +76,17 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 					}
 				}
 			case "adaptive", "auto":
-				// Claude adaptive/auto 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))
+				// Adaptive thinking can carry an explicit effort in output_config.effort (Claude 4.6).
+				// Pass through directly; ApplyThinking handles clamping to target model's levels.
+				effort := ""
+				if v := root.Get("output_config.effort"); v.Exists() && v.Type == gjson.String {
+					effort = strings.ToLower(strings.TrimSpace(v.String()))
+				}
+				if effort != "" {
+					out, _ = sjson.Set(out, "reasoning_effort", effort)
+				} else {
+					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/util/gemini_schema.go

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

internal/watcher/diff/config_diff.go

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

internal/watcher/synthesizer/config.go

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

internal/watcher/synthesizer/file.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"runtime"
 	"strconv"
 	"strings"
 	"time"
@@ -72,6 +73,10 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 		if rel, errRel := filepath.Rel(ctx.AuthDir, full); errRel == nil && rel != "" {
 			id = rel
 		}
+		// On Windows, normalize ID casing to avoid duplicate auth entries caused by case-insensitive paths.
+		if runtime.GOOS == "windows" {
+			id = strings.ToLower(id)
+		}
 
 		proxyURL := ""
 		if p, ok := metadata["proxy_url"].(string); ok {

sdk/auth/filestore.go

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

sdk/cliproxy/auth/conductor.go

@@ -463,9 +463,14 @@ func (m *Manager) Update(ctx context.Context, auth *Auth) (*Auth, error) {
 		return nil, nil
 	}
 	m.mu.Lock()
-	if existing, ok := m.auths[auth.ID]; ok && existing != nil && !auth.indexAssigned && auth.Index == "" {
-		auth.Index = existing.Index
-		auth.indexAssigned = existing.indexAssigned
+	if existing, ok := m.auths[auth.ID]; ok && existing != nil {
+		if !auth.indexAssigned && auth.Index == "" {
+			auth.Index = existing.Index
+			auth.indexAssigned = existing.indexAssigned
+		}
+		if len(auth.ModelStates) == 0 && len(existing.ModelStates) > 0 {
+			auth.ModelStates = existing.ModelStates
+		}
 	}
 	auth.EnsureIndex()
 	m.auths[auth.ID] = auth.Clone()

sdk/cliproxy/auth/conductor_overrides_test.go

@@ -7,6 +7,8 @@ import (
 	"testing"
 	"time"
 
+	"github.com/google/uuid"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 )
 
@@ -115,8 +117,19 @@ func newCredentialRetryLimitTestManager(t *testing.T, maxRetryCredentials int) (
 	executor := &credentialRetryLimitExecutor{id: "claude"}
 	m.RegisterExecutor(executor)
 
-	auth1 := &Auth{ID: "auth-1", Provider: "claude"}
-	auth2 := &Auth{ID: "auth-2", Provider: "claude"}
+	baseID := uuid.NewString()
+	auth1 := &Auth{ID: baseID + "-auth-1", Provider: "claude"}
+	auth2 := &Auth{ID: baseID + "-auth-2", Provider: "claude"}
+
+	// Auth selection requires that the global model registry knows each credential supports the model.
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient(auth1.ID, "claude", []*registry.ModelInfo{{ID: "test-model"}})
+	reg.RegisterClient(auth2.ID, "claude", []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		reg.UnregisterClient(auth1.ID)
+		reg.UnregisterClient(auth2.ID)
+	})
+
 	if _, errRegister := m.Register(context.Background(), auth1); errRegister != nil {
 		t.Fatalf("register auth1: %v", errRegister)
 	}

sdk/cliproxy/auth/conductor_update_test.go

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

sdk/cliproxy/service.go

@@ -301,6 +301,9 @@ func (s *Service) applyCoreAuthAddOrUpdate(ctx context.Context, auth *coreauth.A
 		auth.CreatedAt = existing.CreatedAt
 		auth.LastRefreshedAt = existing.LastRefreshedAt
 		auth.NextRefreshAfter = existing.NextRefreshAfter
+		if len(auth.ModelStates) == 0 && len(existing.ModelStates) > 0 {
+			auth.ModelStates = existing.ModelStates
+		}
 		op = "update"
 		_, err = s.coreManager.Update(ctx, auth)
 	} else {
@@ -817,10 +820,13 @@ func (s *Service) registerModelsForAuth(a *coreauth.Auth) {
 	case "vertex":
 		// Vertex AI Gemini supports the same model identifiers as Gemini.
 		models = registry.GetGeminiVertexModels()
-		if authKind == "apikey" {
-			if entry := s.resolveConfigVertexCompatKey(a); entry != nil && len(entry.Models) > 0 {
+		if entry := s.resolveConfigVertexCompatKey(a); entry != nil {
+			if len(entry.Models) > 0 {
 				models = buildVertexCompatConfigModels(entry)
 			}
+			if authKind == "apikey" {
+				excluded = entry.ExcludedModels
+			}
 		}
 		models = applyExcludedModels(models, excluded)
 	case "gemini-cli":
@@ -866,7 +872,9 @@ func (s *Service) registerModelsForAuth(a *coreauth.Auth) {
 		models = registry.GetKimiModels()
     models = applyExcludedModels(models, excluded)
 	case "github-copilot":
-		models = registry.GetGitHubCopilotModels()
+		ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second)
+		defer cancel()
+		models = executor.FetchGitHubCopilotModels(ctx, a, s.cfg)
 		models = applyExcludedModels(models, excluded)
 	case "kiro":
 		models = s.fetchKiroModels(a)

test/thinking_conversion_test.go

@@ -34,6 +34,8 @@ type thinkingTestCase struct {
 	inputJSON       string
 	expectField     string
 	expectValue     string
+	expectField2    string
+	expectValue2    string
 	includeThoughts string
 	expectErr       bool
 }
@@ -384,15 +386,17 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			includeThoughts: "true",
 			expectErr:       false,
 		},
-		// Case 30: Effort xhigh → not in low/high → error
+		// Case 30: Effort xhigh → clamped to high
 		{
-			name:        "30",
-			from:        "openai",
-			to:          "gemini",
-			model:       "gemini-mixed-model(xhigh)",
-			inputJSON:   `{"model":"gemini-mixed-model(xhigh)","messages":[{"role":"user","content":"hi"}]}`,
-			expectField: "",
-			expectErr:   true,
+			name:            "30",
+			from:            "openai",
+			to:              "gemini",
+			model:           "gemini-mixed-model(xhigh)",
+			inputJSON:       `{"model":"gemini-mixed-model(xhigh)","messages":[{"role":"user","content":"hi"}]}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
+			expectValue:     "high",
+			includeThoughts: "true",
+			expectErr:       false,
 		},
 		// Case 31: Effort none → clamped to low (min supported) → includeThoughts=false
 		{
@@ -1782,15 +1786,17 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			includeThoughts: "true",
 			expectErr:       false,
 		},
-		// Case 30: reasoning_effort=xhigh → error (not in low/high)
+		// Case 30: reasoning_effort=xhigh → clamped to high
 		{
-			name:        "30",
-			from:        "openai",
-			to:          "gemini",
-			model:       "gemini-mixed-model",
-			inputJSON:   `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"xhigh"}`,
-			expectField: "",
-			expectErr:   true,
+			name:            "30",
+			from:            "openai",
+			to:              "gemini",
+			model:           "gemini-mixed-model",
+			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"xhigh"}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
+			expectValue:     "high",
+			includeThoughts: "true",
+			expectErr:       false,
 		},
 		// Case 31: reasoning_effort=none → clamped to low → includeThoughts=false
 		{
@@ -2822,9 +2828,8 @@ 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.
+// TestThinkingE2EClaudeAdaptive_Body covers Group 3 cases in docs/thinking-e2e-test-cases.md.
+// It focuses on Claude 4.6 adaptive thinking and effort/level cross-protocol semantics (body-only).
 func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 	reg := registry.GetGlobalRegistry()
 	uid := fmt.Sprintf("thinking-e2e-claude-adaptive-%d", time.Now().UnixNano())
@@ -2833,32 +2838,347 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 	defer reg.UnregisterClient(uid)
 
 	cases := []thinkingTestCase{
-		// A1: Claude adaptive to OpenAI level model -> highest supported level
+		// A subgroup: OpenAI -> Claude (reasoning_effort -> output_config.effort)
 		{
 			name:        "A1",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"minimal"}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "A2",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"low"}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "A3",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"medium"}`,
+			expectField: "output_config.effort",
+			expectValue: "medium",
+			expectErr:   false,
+		},
+		{
+			name:        "A4",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"high"}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "A5",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-opus-4-6-model",
+			inputJSON:   `{"model":"claude-opus-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"xhigh"}`,
+			expectField: "output_config.effort",
+			expectValue: "max",
+			expectErr:   false,
+		},
+		{
+			name:        "A6",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"xhigh"}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "A7",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-opus-4-6-model",
+			inputJSON:   `{"model":"claude-opus-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"max"}`,
+			expectField: "output_config.effort",
+			expectValue: "max",
+			expectErr:   false,
+		},
+		{
+			name:        "A8",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"max"}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+
+		// B subgroup: Gemini -> Claude (thinkingLevel/thinkingBudget -> output_config.effort)
+		{
+			name:        "B1",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"minimal"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "B2",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"low"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "B3",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"medium"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "medium",
+			expectErr:   false,
+		},
+		{
+			name:        "B4",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "B5",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-opus-4-6-model",
+			inputJSON:   `{"model":"claude-opus-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"xhigh"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "max",
+			expectErr:   false,
+		},
+		{
+			name:        "B6",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"xhigh"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "B7",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":512}}}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "B8",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":1024}}}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "B9",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":8192}}}`,
+			expectField: "output_config.effort",
+			expectValue: "medium",
+			expectErr:   false,
+		},
+		{
+			name:        "B10",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":24576}}}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "B11",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-opus-4-6-model",
+			inputJSON:   `{"model":"claude-opus-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":32768}}}`,
+			expectField: "output_config.effort",
+			expectValue: "max",
+			expectErr:   false,
+		},
+		{
+			name:        "B12",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":32768}}}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "B13",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":0}}}`,
+			expectField: "thinking.type",
+			expectValue: "disabled",
+			expectErr:   false,
+		},
+		{
+			name:        "B14",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":-1}}}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+
+		// C subgroup: Claude adaptive + effort cross-protocol conversion
+		{
+			name:        "C1",
 			from:        "claude",
 			to:          "openai",
 			model:       "level-model",
-			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"minimal"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "minimal",
+			expectErr:   false,
+		},
+		{
+			name:        "C2",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"low"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "C3",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"medium"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "medium",
+			expectErr:   false,
+		},
+		{
+			name:        "C4",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
 			expectField: "reasoning_effort",
 			expectValue: "high",
 			expectErr:   false,
 		},
-		// A2: Claude adaptive to Gemini level subset model -> highest supported level
 		{
-			name:            "A2",
+			name:        "C5",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"xhigh"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "C6",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"max"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "C7",
+			from:        "claude",
+			to:          "openai",
+			model:       "no-thinking-model",
+			inputJSON:   `{"model":"no-thinking-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
+			expectField: "",
+			expectErr:   false,
+		},
+
+		{
+			name:            "C8",
 			from:            "claude",
 			to:              "gemini",
 			model:           "level-subset-model",
-			inputJSON:       `{"model":"level-subset-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			inputJSON:       `{"model":"level-subset-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "high",
 			includeThoughts: "true",
 			expectErr:       false,
 		},
-		// A3: Claude adaptive to Gemini budget model -> max budget
 		{
-			name:            "A3",
+			name:            "C9",
+			from:            "claude",
+			to:              "gemini",
+			model:           "gemini-budget-model",
+			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"low"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "1024",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		{
+			name:            "C10",
+			from:            "claude",
+			to:              "gemini",
+			model:           "gemini-budget-model",
+			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"medium"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "8192",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		{
+			name:            "C11",
+			from:            "claude",
+			to:              "gemini",
+			model:           "gemini-budget-model",
+			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "20000",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		{
+			name:            "C12",
 			from:            "claude",
 			to:              "gemini",
 			model:           "gemini-budget-model",
@@ -2868,32 +3188,91 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 			includeThoughts: "true",
 			expectErr:       false,
 		},
-		// A4: Claude adaptive to Gemini mixed model -> highest supported level
 		{
-			name:            "A4",
+			name:            "C13",
 			from:            "claude",
 			to:              "gemini",
 			model:           "gemini-mixed-model",
-			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "high",
 			includeThoughts: "true",
 			expectErr:       false,
 		},
-		// A5: Claude adaptive passthrough for same protocol
+
 		{
-			name:        "A5",
+			name:        "C14",
 			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",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"minimal"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "minimal",
 			expectErr:   false,
 		},
-		// A6: Claude adaptive to Antigravity budget model -> max budget
 		{
-			name:            "A6",
+			name:        "C15",
+			from:        "claude",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"low"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "C16",
+			from:        "claude",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "C17",
+			from:        "claude",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"xhigh"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "C18",
+			from:        "claude",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"max"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+
+		{
+			name:        "C19",
+			from:        "claude",
+			to:          "iflow",
+			model:       "glm-test",
+			inputJSON:   `{"model":"glm-test","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"minimal"}}`,
+			expectField: "chat_template_kwargs.enable_thinking",
+			expectValue: "true",
+			expectErr:   false,
+		},
+		{
+			name:        "C20",
+			from:        "claude",
+			to:          "iflow",
+			model:       "minimax-test",
+			inputJSON:   `{"model":"minimax-test","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
+			expectField: "reasoning_split",
+			expectValue: "true",
+			expectErr:   false,
+		},
+		{
+			name:            "C21",
 			from:            "claude",
 			to:              "antigravity",
 			model:           "antigravity-budget-model",
@@ -2903,48 +3282,66 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 			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,
+			name:         "C22",
+			from:         "claude",
+			to:           "claude",
+			model:        "claude-sonnet-4-6-model",
+			inputJSON:    `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"medium"}}`,
+			expectField:  "thinking.type",
+			expectValue:  "adaptive",
+			expectField2: "output_config.effort",
+			expectValue2: "medium",
+			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,
+			name:         "C23",
+			from:         "claude",
+			to:           "claude",
+			model:        "claude-opus-4-6-model",
+			inputJSON:    `{"model":"claude-opus-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"max"}}`,
+			expectField:  "thinking.type",
+			expectValue:  "adaptive",
+			expectField2: "output_config.effort",
+			expectValue2: "max",
+			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,
+			name:      "C24",
+			from:      "claude",
+			to:        "claude",
+			model:     "claude-opus-4-6-model",
+			inputJSON: `{"model":"claude-opus-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"xhigh"}}`,
+			expectErr: true,
 		},
-		// 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,
+			name:         "C25",
+			from:         "claude",
+			to:           "claude",
+			model:        "claude-sonnet-4-6-model",
+			inputJSON:    `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
+			expectField:  "thinking.type",
+			expectValue:  "adaptive",
+			expectField2: "output_config.effort",
+			expectValue2: "high",
+			expectErr:    false,
+		},
+		{
+			name:      "C26",
+			from:      "claude",
+			to:        "claude",
+			model:     "claude-sonnet-4-6-model",
+			inputJSON: `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"max"}}`,
+			expectErr: true,
+		},
+		{
+			name:      "C27",
+			from:      "claude",
+			to:        "claude",
+			model:     "claude-sonnet-4-6-model",
+			inputJSON: `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"xhigh"}}`,
+			expectErr: true,
 		},
 	}
 
@@ -2999,6 +3396,29 @@ func getTestModels() []*registry.ModelInfo {
 			DisplayName: "Claude Budget Model",
 			Thinking:    &registry.ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
 		},
+		{
+			ID:                  "claude-sonnet-4-6-model",
+			Object:              "model",
+			Created:             1771372800, // 2026-02-17
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4.6 Sonnet",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &registry.ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false, Levels: []string{"low", "medium", "high"}},
+		},
+		{
+			ID:                  "claude-opus-4-6-model",
+			Object:              "model",
+			Created:             1770318000, // 2026-02-05
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4.6 Opus",
+			Description:         "Premium model combining maximum intelligence with practical performance",
+			ContextLength:       1000000,
+			MaxCompletionTokens: 128000,
+			Thinking:            &registry.ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false, Levels: []string{"low", "medium", "high", "max"}},
+		},
 		{
 			ID:          "antigravity-budget-model",
 			Object:      "model",
@@ -3165,17 +3585,23 @@ func runThinkingTests(t *testing.T, cases []thinkingTestCase) {
 				return
 			}
 
-			val := gjson.GetBytes(body, tc.expectField)
-			if !val.Exists() {
-				t.Fatalf("expected field %s not found, body=%s", tc.expectField, string(body))
+			assertField := func(fieldPath, expected string) {
+				val := gjson.GetBytes(body, fieldPath)
+				if !val.Exists() {
+					t.Fatalf("expected field %s not found, body=%s", fieldPath, string(body))
+				}
+				actualValue := val.String()
+				if val.Type == gjson.Number {
+					actualValue = fmt.Sprintf("%d", val.Int())
+				}
+				if actualValue != expected {
+					t.Fatalf("field %s: expected %q, got %q, body=%s", fieldPath, expected, actualValue, string(body))
+				}
 			}
 
-			actualValue := val.String()
-			if val.Type == gjson.Number {
-				actualValue = fmt.Sprintf("%d", val.Int())
-			}
-			if actualValue != tc.expectValue {
-				t.Fatalf("field %s: expected %q, got %q, body=%s", tc.expectField, tc.expectValue, actualValue, string(body))
+			assertField(tc.expectField, tc.expectValue)
+			if tc.expectField2 != "" {
+				assertField(tc.expectField2, tc.expectValue2)
 			}
 
 			if tc.includeThoughts != "" && (tc.to == "gemini" || tc.to == "gemini-cli" || tc.to == "antigravity") {