Merge branch 'router-for-me:main' into main

- Simplified connection logic by removing `connCreateSent` and related state handling.
- Updated `buildCodexWebsocketRequestBody` to always use `response.create`.
- Added unit tests to validate `response.create` behavior and beta header preservation.
- Dropped unsupported `response.append` and outdated `response.done` event types.

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"
@@ -698,17 +699,20 @@ func (h *Handler) authIDForPath(path string) string {
 	if path == "" {
 		return ""
 	}
-	if h == nil || h.cfg == nil {
-		return path
+	id := path
+	if h != nil && h.cfg != nil {
+		authDir := strings.TrimSpace(h.cfg.AuthDir)
+		if authDir != "" {
+			if rel, errRel := filepath.Rel(authDir, path); errRel == nil && rel != "" {
+				id = rel
+			}
+		}
 	}
-	authDir := strings.TrimSpace(h.cfg.AuthDir)
-	if authDir == "" {
-		return path
+	// On Windows, normalize ID casing to avoid duplicate auth entries caused by case-insensitive paths.
+	if runtime.GOOS == "windows" {
+		id = strings.ToLower(id)
 	}
-	if rel, err := filepath.Rel(authDir, path); err == nil && rel != "" {
-		return rel
-	}
-	return path
+	return id
 }
 
 func (h *Handler) registerAuthFromFile(ctx context.Context, path string, data []byte) error {

internal/api/handlers/management/config_lists.go

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

internal/auth/copilot/copilot_auth.go

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

internal/config/oauth_model_alias_defaults.go

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

internal/config/vertex_compat.go

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

internal/registry/model_definitions.go

@@ -23,7 +23,6 @@ import (
 //   - kiro
 //   - kilo
 //   - github-copilot
-//   - kiro
 //   - amazonq
 //   - antigravity (returns static overrides only)
 func GetStaticModelDefinitionsByChannel(channel string) []*ModelInfo {
@@ -152,6 +151,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			Description:         "OpenAI GPT-4.1 via GitHub Copilot",
 			ContextLength:       128000,
 			MaxCompletionTokens: 16384,
+			SupportedEndpoints:  []string{"/chat/completions", "/responses"},
 		},
 	}
 
@@ -166,6 +166,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			Description:         entry.Description,
 			ContextLength:       128000,
 			MaxCompletionTokens: 16384,
+			SupportedEndpoints:  []string{"/chat/completions", "/responses"},
 		})
 	}
 

internal/registry/model_definitions_static_data.go

@@ -220,12 +220,27 @@ func GetGeminiModels() []*ModelInfo {
 			Name:                       "models/gemini-3-flash-preview",
 			Version:                    "3.0",
 			DisplayName:                "Gemini 3 Flash Preview",
-			Description:                "Gemini 3 Flash Preview",
+			Description:                "Our most intelligent model built for speed, combining frontier intelligence with superior search and grounding.",
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-flash-lite-preview",
+			Object:                     "model",
+			Created:                    1776288000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-flash-lite-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Flash Lite Preview",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}},
+		},
 		{
 			ID:                         "gemini-3-pro-image-preview",
 			Object:                     "model",
@@ -336,6 +351,21 @@ func GetGeminiVertexModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-flash-lite-preview",
+			Object:                     "model",
+			Created:                    1776288000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-flash-lite-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Flash Lite Preview",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}},
+		},
 		{
 			ID:                         "gemini-3-pro-image-preview",
 			Object:                     "model",
@@ -508,6 +538,21 @@ func GetGeminiCLIModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-flash-lite-preview",
+			Object:                     "model",
+			Created:                    1776288000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-flash-lite-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Flash Lite Preview",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}},
+		},
 	}
 }
 
@@ -604,6 +649,21 @@ func GetAIStudioModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
 		},
+		{
+			ID:                         "gemini-3.1-flash-lite-preview",
+			Object:                     "model",
+			Created:                    1776288000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-flash-lite-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Flash Lite Preview",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}},
+		},
 		{
 			ID:                         "gemini-pro-latest",
 			Object:                     "model",
@@ -839,6 +899,20 @@ func GetOpenAIModels() []*ModelInfo {
 			SupportedParameters: []string{"tools"},
 			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
 		},
+		{
+			ID:                  "gpt-5.4",
+			Object:              "model",
+			Created:             1772668800,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5.4",
+			DisplayName:         "GPT 5.4",
+			Description:         "Stable version of GPT 5.4",
+			ContextLength:       1_050_000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
+		},
 	}
 }
 
@@ -966,6 +1040,7 @@ func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 		"gemini-3.1-pro-high":      {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
 		"gemini-3.1-pro-low":       {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
 		"gemini-3.1-flash-image":   {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}}},
+		"gemini-3.1-flash-lite-preview": {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}}},
 		"gemini-3-flash":           {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}},
 		"claude-opus-4-6-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 64000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-sonnet-4-6":        {Thinking: &ThinkingSupport{Min: 1024, Max: 64000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},

internal/runtime/executor/antigravity_executor_buildrequest_test.go

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

internal/runtime/executor/claude_executor.go

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

internal/runtime/executor/claude_executor_test.go

@@ -2,6 +2,7 @@ package executor
 
 import (
 	"bytes"
+	"compress/gzip"
 	"context"
 	"io"
 	"net/http"
@@ -9,6 +10,7 @@ import (
 	"strings"
 	"testing"
 
+	"github.com/klauspost/compress/zstd"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
@@ -583,3 +585,385 @@ func testClaudeExecutorInvalidCompressedErrorBody(
 		t.Fatalf("expected status code 400, got: %v", err)
 	}
 }
+
+// TestClaudeExecutor_ExecuteStream_SetsIdentityAcceptEncoding verifies that streaming
+// requests use Accept-Encoding: identity so the upstream cannot respond with a
+// compressed SSE body that would silently break the line scanner.
+func TestClaudeExecutor_ExecuteStream_SetsIdentityAcceptEncoding(t *testing.T) {
+	var gotEncoding, gotAccept string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotEncoding = r.Header.Get("Accept-Encoding")
+		gotAccept = r.Header.Get("Accept")
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte("data: {\"type\":\"message_stop\"}\n\n"))
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	result, err := executor.ExecuteStream(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err != nil {
+		t.Fatalf("ExecuteStream error: %v", err)
+	}
+	for chunk := range result.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("unexpected chunk error: %v", chunk.Err)
+		}
+	}
+
+	if gotEncoding != "identity" {
+		t.Errorf("Accept-Encoding = %q, want %q", gotEncoding, "identity")
+	}
+	if gotAccept != "text/event-stream" {
+		t.Errorf("Accept = %q, want %q", gotAccept, "text/event-stream")
+	}
+}
+
+// TestClaudeExecutor_Execute_SetsCompressedAcceptEncoding verifies that non-streaming
+// requests keep the full accept-encoding to allow response compression (which
+// decodeResponseBody handles correctly).
+func TestClaudeExecutor_Execute_SetsCompressedAcceptEncoding(t *testing.T) {
+	var gotEncoding, gotAccept string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotEncoding = r.Header.Get("Accept-Encoding")
+		gotAccept = r.Header.Get("Accept")
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"id":"msg_1","type":"message","model":"claude-3-5-sonnet-20241022","role":"assistant","content":[{"type":"text","text":"hi"}],"usage":{"input_tokens":1,"output_tokens":1}}`))
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	_, err := executor.Execute(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err != nil {
+		t.Fatalf("Execute error: %v", err)
+	}
+
+	if gotEncoding != "gzip, deflate, br, zstd" {
+		t.Errorf("Accept-Encoding = %q, want %q", gotEncoding, "gzip, deflate, br, zstd")
+	}
+	if gotAccept != "application/json" {
+		t.Errorf("Accept = %q, want %q", gotAccept, "application/json")
+	}
+}
+
+// TestClaudeExecutor_ExecuteStream_GzipSuccessBodyDecoded verifies that a streaming
+// HTTP 200 response with Content-Encoding: gzip is correctly decompressed before
+// the line scanner runs, so SSE chunks are not silently dropped.
+func TestClaudeExecutor_ExecuteStream_GzipSuccessBodyDecoded(t *testing.T) {
+	var buf bytes.Buffer
+	gz := gzip.NewWriter(&buf)
+	_, _ = gz.Write([]byte("data: {\"type\":\"message_stop\"}\n"))
+	_ = gz.Close()
+	compressedBody := buf.Bytes()
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		w.Header().Set("Content-Encoding", "gzip")
+		_, _ = w.Write(compressedBody)
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	result, err := executor.ExecuteStream(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err != nil {
+		t.Fatalf("ExecuteStream error: %v", err)
+	}
+
+	var combined strings.Builder
+	for chunk := range result.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("chunk error: %v", chunk.Err)
+		}
+		combined.Write(chunk.Payload)
+	}
+
+	if combined.Len() == 0 {
+		t.Fatal("expected at least one chunk from gzip-encoded SSE body, got none (body was not decompressed)")
+	}
+	if !strings.Contains(combined.String(), "message_stop") {
+		t.Errorf("expected SSE content in chunks, got: %q", combined.String())
+	}
+}
+
+// TestDecodeResponseBody_MagicByteGzipNoHeader verifies that decodeResponseBody
+// detects gzip-compressed content via magic bytes even when Content-Encoding is absent.
+func TestDecodeResponseBody_MagicByteGzipNoHeader(t *testing.T) {
+	const plaintext = "data: {\"type\":\"message_stop\"}\n"
+
+	var buf bytes.Buffer
+	gz := gzip.NewWriter(&buf)
+	_, _ = gz.Write([]byte(plaintext))
+	_ = gz.Close()
+
+	rc := io.NopCloser(&buf)
+	decoded, err := decodeResponseBody(rc, "")
+	if err != nil {
+		t.Fatalf("decodeResponseBody error: %v", err)
+	}
+	defer decoded.Close()
+
+	got, err := io.ReadAll(decoded)
+	if err != nil {
+		t.Fatalf("ReadAll error: %v", err)
+	}
+	if string(got) != plaintext {
+		t.Errorf("decoded = %q, want %q", got, plaintext)
+	}
+}
+
+// TestDecodeResponseBody_PlainTextNoHeader verifies that decodeResponseBody returns
+// plain text untouched when Content-Encoding is absent and no magic bytes match.
+func TestDecodeResponseBody_PlainTextNoHeader(t *testing.T) {
+	const plaintext = "data: {\"type\":\"message_stop\"}\n"
+	rc := io.NopCloser(strings.NewReader(plaintext))
+	decoded, err := decodeResponseBody(rc, "")
+	if err != nil {
+		t.Fatalf("decodeResponseBody error: %v", err)
+	}
+	defer decoded.Close()
+
+	got, err := io.ReadAll(decoded)
+	if err != nil {
+		t.Fatalf("ReadAll error: %v", err)
+	}
+	if string(got) != plaintext {
+		t.Errorf("decoded = %q, want %q", got, plaintext)
+	}
+}
+
+// TestClaudeExecutor_ExecuteStream_GzipNoContentEncodingHeader verifies the full
+// pipeline: when the upstream returns a gzip-compressed SSE body WITHOUT setting
+// Content-Encoding (a misbehaving upstream), the magic-byte sniff in
+// decodeResponseBody still decompresses it, so chunks reach the caller.
+func TestClaudeExecutor_ExecuteStream_GzipNoContentEncodingHeader(t *testing.T) {
+	var buf bytes.Buffer
+	gz := gzip.NewWriter(&buf)
+	_, _ = gz.Write([]byte("data: {\"type\":\"message_stop\"}\n"))
+	_ = gz.Close()
+	compressedBody := buf.Bytes()
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		// Intentionally omit Content-Encoding to simulate misbehaving upstream.
+		_, _ = w.Write(compressedBody)
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	result, err := executor.ExecuteStream(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err != nil {
+		t.Fatalf("ExecuteStream error: %v", err)
+	}
+
+	var combined strings.Builder
+	for chunk := range result.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("chunk error: %v", chunk.Err)
+		}
+		combined.Write(chunk.Payload)
+	}
+
+	if combined.Len() == 0 {
+		t.Fatal("expected chunks from gzip body without Content-Encoding header, got none (magic-byte sniff failed)")
+	}
+	if !strings.Contains(combined.String(), "message_stop") {
+		t.Errorf("unexpected chunk content: %q", combined.String())
+	}
+}
+
+// TestClaudeExecutor_ExecuteStream_AcceptEncodingOverrideCannotBypassIdentity verifies
+// that injecting Accept-Encoding via auth.Attributes cannot override the stream
+// path's enforced identity encoding.
+func TestClaudeExecutor_ExecuteStream_AcceptEncodingOverrideCannotBypassIdentity(t *testing.T) {
+	var gotEncoding string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotEncoding = r.Header.Get("Accept-Encoding")
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte("data: {\"type\":\"message_stop\"}\n\n"))
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	// Inject Accept-Encoding via the custom header attribute mechanism.
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":             "key-123",
+		"base_url":            server.URL,
+		"header:Accept-Encoding": "gzip, deflate, br, zstd",
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	result, err := executor.ExecuteStream(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err != nil {
+		t.Fatalf("ExecuteStream error: %v", err)
+	}
+	for chunk := range result.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("unexpected chunk error: %v", chunk.Err)
+		}
+	}
+
+	if gotEncoding != "identity" {
+		t.Errorf("Accept-Encoding = %q; stream path must enforce identity regardless of auth.Attributes override", gotEncoding)
+	}
+}
+
+// TestDecodeResponseBody_MagicByteZstdNoHeader verifies that decodeResponseBody
+// detects zstd-compressed content via magic bytes (28 b5 2f fd) even when
+// Content-Encoding is absent.
+func TestDecodeResponseBody_MagicByteZstdNoHeader(t *testing.T) {
+	const plaintext = "data: {\"type\":\"message_stop\"}\n"
+
+	var buf bytes.Buffer
+	enc, err := zstd.NewWriter(&buf)
+	if err != nil {
+		t.Fatalf("zstd.NewWriter: %v", err)
+	}
+	_, _ = enc.Write([]byte(plaintext))
+	_ = enc.Close()
+
+	rc := io.NopCloser(&buf)
+	decoded, err := decodeResponseBody(rc, "")
+	if err != nil {
+		t.Fatalf("decodeResponseBody error: %v", err)
+	}
+	defer decoded.Close()
+
+	got, err := io.ReadAll(decoded)
+	if err != nil {
+		t.Fatalf("ReadAll error: %v", err)
+	}
+	if string(got) != plaintext {
+		t.Errorf("decoded = %q, want %q", got, plaintext)
+	}
+}
+
+// TestClaudeExecutor_Execute_GzipErrorBodyNoContentEncodingHeader verifies that the
+// error path (4xx) correctly decompresses a gzip body even when the upstream omits
+// the Content-Encoding header.  This closes the gap left by PR #1771, which only
+// fixed header-declared compression on the error path.
+func TestClaudeExecutor_Execute_GzipErrorBodyNoContentEncodingHeader(t *testing.T) {
+	const errJSON = `{"type":"error","error":{"type":"invalid_request_error","message":"test error"}}`
+
+	var buf bytes.Buffer
+	gz := gzip.NewWriter(&buf)
+	_, _ = gz.Write([]byte(errJSON))
+	_ = gz.Close()
+	compressedBody := buf.Bytes()
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		// Intentionally omit Content-Encoding to simulate misbehaving upstream.
+		w.WriteHeader(http.StatusBadRequest)
+		_, _ = w.Write(compressedBody)
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	_, err := executor.Execute(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err == nil {
+		t.Fatal("expected an error for 400 response, got nil")
+	}
+	if !strings.Contains(err.Error(), "test error") {
+		t.Errorf("error message should contain decompressed JSON, got: %q", err.Error())
+	}
+}
+
+// TestClaudeExecutor_ExecuteStream_GzipErrorBodyNoContentEncodingHeader verifies
+// the same for the streaming executor: 4xx gzip body without Content-Encoding is
+// decoded and the error message is readable.
+func TestClaudeExecutor_ExecuteStream_GzipErrorBodyNoContentEncodingHeader(t *testing.T) {
+	const errJSON = `{"type":"error","error":{"type":"invalid_request_error","message":"stream test error"}}`
+
+	var buf bytes.Buffer
+	gz := gzip.NewWriter(&buf)
+	_, _ = gz.Write([]byte(errJSON))
+	_ = gz.Close()
+	compressedBody := buf.Bytes()
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		// Intentionally omit Content-Encoding to simulate misbehaving upstream.
+		w.WriteHeader(http.StatusBadRequest)
+		_, _ = w.Write(compressedBody)
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	_, err := executor.ExecuteStream(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-sonnet-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FromString("claude"),
+	})
+	if err == nil {
+		t.Fatal("expected an error for 400 response, got nil")
+	}
+	if !strings.Contains(err.Error(), "stream test error") {
+		t.Errorf("error message should contain decompressed JSON, got: %q", err.Error())
+	}
+}

internal/runtime/executor/codex_executor.go

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

internal/runtime/executor/codex_executor_cache_test.go

@@ -0,0 +1,64 @@
+package executor
+
+import (
+	"context"
+	"io"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/uuid"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
+	"github.com/tidwall/gjson"
+)
+
+func TestCodexExecutorCacheHelper_OpenAIChatCompletions_StablePromptCacheKeyFromAPIKey(t *testing.T) {
+	recorder := httptest.NewRecorder()
+	ginCtx, _ := gin.CreateTestContext(recorder)
+	ginCtx.Set("apiKey", "test-api-key")
+
+	ctx := context.WithValue(context.Background(), "gin", ginCtx)
+	executor := &CodexExecutor{}
+	rawJSON := []byte(`{"model":"gpt-5.3-codex","stream":true}`)
+	req := cliproxyexecutor.Request{
+		Model:   "gpt-5.3-codex",
+		Payload: []byte(`{"model":"gpt-5.3-codex"}`),
+	}
+	url := "https://example.com/responses"
+
+	httpReq, err := executor.cacheHelper(ctx, sdktranslator.FromString("openai"), url, req, rawJSON)
+	if err != nil {
+		t.Fatalf("cacheHelper error: %v", err)
+	}
+
+	body, errRead := io.ReadAll(httpReq.Body)
+	if errRead != nil {
+		t.Fatalf("read request body: %v", errRead)
+	}
+
+	expectedKey := uuid.NewSHA1(uuid.NameSpaceOID, []byte("cli-proxy-api:codex:prompt-cache:test-api-key")).String()
+	gotKey := gjson.GetBytes(body, "prompt_cache_key").String()
+	if gotKey != expectedKey {
+		t.Fatalf("prompt_cache_key = %q, want %q", gotKey, expectedKey)
+	}
+	if gotConversation := httpReq.Header.Get("Conversation_id"); gotConversation != expectedKey {
+		t.Fatalf("Conversation_id = %q, want %q", gotConversation, expectedKey)
+	}
+	if gotSession := httpReq.Header.Get("Session_id"); gotSession != expectedKey {
+		t.Fatalf("Session_id = %q, want %q", gotSession, expectedKey)
+	}
+
+	httpReq2, err := executor.cacheHelper(ctx, sdktranslator.FromString("openai"), url, req, rawJSON)
+	if err != nil {
+		t.Fatalf("cacheHelper error (second call): %v", err)
+	}
+	body2, errRead2 := io.ReadAll(httpReq2.Body)
+	if errRead2 != nil {
+		t.Fatalf("read request body (second call): %v", errRead2)
+	}
+	gotKey2 := gjson.GetBytes(body2, "prompt_cache_key").String()
+	if gotKey2 != expectedKey {
+		t.Fatalf("prompt_cache_key (second call) = %q, want %q", gotKey2, expectedKey)
+	}
+}

internal/runtime/executor/codex_websockets_executor.go

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

internal/runtime/executor/codex_websockets_executor_test.go

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

internal/runtime/executor/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/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/codex/claude/codex_claude_request.go

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

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

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

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

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

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

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

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

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

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

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

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

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

internal/translator/gemini/claude/gemini_claude_request.go

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

internal/translator/gemini/claude/gemini_claude_request_test.go

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

internal/translator/gemini/claude/gemini_claude_response.go

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

internal/translator/gemini/openai/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_response.go

@@ -22,9 +22,11 @@ var (
 
 // ConvertOpenAIResponseToAnthropicParams holds parameters for response conversion
 type ConvertOpenAIResponseToAnthropicParams struct {
-	MessageID string
-	Model     string
-	CreatedAt int64
+	MessageID   string
+	Model       string
+	CreatedAt   int64
+	ToolNameMap map[string]string
+	SawToolCall bool
 	// Content accumulator for streaming
 	ContentAccumulator strings.Builder
 	// Tool calls accumulator for streaming
@@ -78,6 +80,8 @@ func ConvertOpenAIResponseToClaude(_ context.Context, _ string, originalRequestR
 			MessageID:                   "",
 			Model:                       "",
 			CreatedAt:                   0,
+			ToolNameMap:                 nil,
+			SawToolCall:                 false,
 			ContentAccumulator:          strings.Builder{},
 			ToolCallsAccumulator:        nil,
 			TextContentBlockStarted:     false,
@@ -97,6 +101,10 @@ func ConvertOpenAIResponseToClaude(_ context.Context, _ string, originalRequestR
 	}
 	rawJSON = bytes.TrimSpace(rawJSON[5:])
 
+	if (*param).(*ConvertOpenAIResponseToAnthropicParams).ToolNameMap == nil {
+		(*param).(*ConvertOpenAIResponseToAnthropicParams).ToolNameMap = util.ToolNameMapFromClaudeRequest(originalRequestRawJSON)
+	}
+
 	// Check if this is the [DONE] marker
 	rawStr := strings.TrimSpace(string(rawJSON))
 	if rawStr == "[DONE]" {
@@ -111,6 +119,16 @@ func ConvertOpenAIResponseToClaude(_ context.Context, _ string, originalRequestR
 	}
 }
 
+func effectiveOpenAIFinishReason(param *ConvertOpenAIResponseToAnthropicParams) string {
+	if param == nil {
+		return ""
+	}
+	if param.SawToolCall {
+		return "tool_calls"
+	}
+	return param.FinishReason
+}
+
 // convertOpenAIStreamingChunkToAnthropic converts OpenAI streaming chunk to Anthropic streaming events
 func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAIResponseToAnthropicParams) []string {
 	root := gjson.ParseBytes(rawJSON)
@@ -197,6 +215,7 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 			}
 
 			toolCalls.ForEach(func(_, toolCall gjson.Result) bool {
+				param.SawToolCall = true
 				index := int(toolCall.Get("index").Int())
 				blockIndex := param.toolContentBlockIndex(index)
 
@@ -215,7 +234,7 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 				// Handle function name
 				if function := toolCall.Get("function"); function.Exists() {
 					if name := function.Get("name"); name.Exists() {
-						accumulator.Name = name.String()
+						accumulator.Name = util.MapToolName(param.ToolNameMap, name.String())
 
 						stopThinkingContentBlock(param, &results)
 
@@ -246,7 +265,11 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 	// Handle finish_reason (but don't send message_delta/message_stop yet)
 	if finishReason := root.Get("choices.0.finish_reason"); finishReason.Exists() && finishReason.String() != "" {
 		reason := finishReason.String()
-		param.FinishReason = reason
+		if param.SawToolCall {
+			param.FinishReason = "tool_calls"
+		} else {
+			param.FinishReason = reason
+		}
 
 		// Send content_block_stop for thinking content if needed
 		if param.ThinkingContentBlockStarted {
@@ -294,7 +317,7 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 			inputTokens, outputTokens, cachedTokens = extractOpenAIUsage(usage)
 			// Send message_delta with usage
 			messageDeltaJSON := `{"type":"message_delta","delta":{"stop_reason":"","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
-			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(param.FinishReason))
+			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(effectiveOpenAIFinishReason(param)))
 			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.input_tokens", inputTokens)
 			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.output_tokens", outputTokens)
 			if cachedTokens > 0 {
@@ -348,7 +371,7 @@ func convertOpenAIDoneToAnthropic(param *ConvertOpenAIResponseToAnthropicParams)
 	// If we haven't sent message_delta yet (no usage info was received), send it now
 	if param.FinishReason != "" && !param.MessageDeltaSent {
 		messageDeltaJSON := `{"type":"message_delta","delta":{"stop_reason":"","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
-		messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(param.FinishReason))
+		messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(effectiveOpenAIFinishReason(param)))
 		results = append(results, "event: message_delta\ndata: "+messageDeltaJSON+"\n\n")
 		param.MessageDeltaSent = true
 	}
@@ -531,10 +554,10 @@ func stopTextContentBlock(param *ConvertOpenAIResponseToAnthropicParams, results
 // Returns:
 //   - string: An Anthropic-compatible JSON response.
 func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, _ *any) string {
-	_ = originalRequestRawJSON
 	_ = requestRawJSON
 
 	root := gjson.ParseBytes(rawJSON)
+	toolNameMap := util.ToolNameMapFromClaudeRequest(originalRequestRawJSON)
 	out := `{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}`
 	out, _ = sjson.Set(out, "id", root.Get("id").String())
 	out, _ = sjson.Set(out, "model", root.Get("model").String())
@@ -590,7 +613,7 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 									hasToolCall = true
 									toolUse := `{"type":"tool_use","id":"","name":"","input":{}}`
 									toolUse, _ = sjson.Set(toolUse, "id", tc.Get("id").String())
-									toolUse, _ = sjson.Set(toolUse, "name", tc.Get("function.name").String())
+									toolUse, _ = sjson.Set(toolUse, "name", util.MapToolName(toolNameMap, tc.Get("function.name").String()))
 
 									argsStr := util.FixJSON(tc.Get("function.arguments").String())
 									if argsStr != "" && gjson.Valid(argsStr) {
@@ -647,7 +670,7 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 					hasToolCall = true
 					toolUseBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
 					toolUseBlock, _ = sjson.Set(toolUseBlock, "id", toolCall.Get("id").String())
-					toolUseBlock, _ = sjson.Set(toolUseBlock, "name", toolCall.Get("function.name").String())
+					toolUseBlock, _ = sjson.Set(toolUseBlock, "name", util.MapToolName(toolNameMap, toolCall.Get("function.name").String()))
 
 					argsStr := util.FixJSON(toolCall.Get("function.arguments").String())
 					if argsStr != "" && gjson.Valid(argsStr) {

internal/util/gemini_schema.go

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

internal/util/translator.go

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

internal/watcher/clients.go

@@ -183,7 +183,7 @@ func (w *Watcher) addOrUpdateClient(path string) {
 
 	if w.reloadCallback != nil {
 		log.Debugf("triggering server update callback after add/update")
-		w.reloadCallback(cfg)
+		w.triggerServerUpdate(cfg)
 	}
 	w.persistAuthAsync(fmt.Sprintf("Sync auth %s", filepath.Base(path)), path)
 }
@@ -202,7 +202,7 @@ func (w *Watcher) removeClient(path string) {
 
 	if w.reloadCallback != nil {
 		log.Debugf("triggering server update callback after removal")
-		w.reloadCallback(cfg)
+		w.triggerServerUpdate(cfg)
 	}
 	w.persistAuthAsync(fmt.Sprintf("Remove auth %s", filepath.Base(path)), path)
 }
@@ -303,3 +303,79 @@ func (w *Watcher) persistAuthAsync(message string, paths ...string) {
 		}
 	}()
 }
+
+func (w *Watcher) stopServerUpdateTimer() {
+	w.serverUpdateMu.Lock()
+	defer w.serverUpdateMu.Unlock()
+	if w.serverUpdateTimer != nil {
+		w.serverUpdateTimer.Stop()
+		w.serverUpdateTimer = nil
+	}
+	w.serverUpdatePend = false
+}
+
+func (w *Watcher) triggerServerUpdate(cfg *config.Config) {
+	if w == nil || w.reloadCallback == nil || cfg == nil {
+		return
+	}
+	if w.stopped.Load() {
+		return
+	}
+
+	now := time.Now()
+
+	w.serverUpdateMu.Lock()
+	if w.serverUpdateLast.IsZero() || now.Sub(w.serverUpdateLast) >= serverUpdateDebounce {
+		w.serverUpdateLast = now
+		if w.serverUpdateTimer != nil {
+			w.serverUpdateTimer.Stop()
+			w.serverUpdateTimer = nil
+		}
+		w.serverUpdatePend = false
+		w.serverUpdateMu.Unlock()
+		w.reloadCallback(cfg)
+		return
+	}
+
+	if w.serverUpdatePend {
+		w.serverUpdateMu.Unlock()
+		return
+	}
+
+	delay := serverUpdateDebounce - now.Sub(w.serverUpdateLast)
+	if delay < 10*time.Millisecond {
+		delay = 10 * time.Millisecond
+	}
+	w.serverUpdatePend = true
+	if w.serverUpdateTimer != nil {
+		w.serverUpdateTimer.Stop()
+		w.serverUpdateTimer = nil
+	}
+	var timer *time.Timer
+	timer = time.AfterFunc(delay, func() {
+		if w.stopped.Load() {
+			return
+		}
+		w.clientsMutex.RLock()
+		latestCfg := w.config
+		w.clientsMutex.RUnlock()
+
+		w.serverUpdateMu.Lock()
+		if w.serverUpdateTimer != timer || !w.serverUpdatePend {
+			w.serverUpdateMu.Unlock()
+			return
+		}
+		w.serverUpdateTimer = nil
+		w.serverUpdatePend = false
+		if latestCfg == nil || w.reloadCallback == nil || w.stopped.Load() {
+			w.serverUpdateMu.Unlock()
+			return
+		}
+
+		w.serverUpdateLast = time.Now()
+		w.serverUpdateMu.Unlock()
+		w.reloadCallback(latestCfg)
+	})
+	w.serverUpdateTimer = timer
+	w.serverUpdateMu.Unlock()
+}

internal/watcher/diff/config_diff.go

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

internal/watcher/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 {

internal/watcher/watcher.go

@@ -6,6 +6,7 @@ import (
 	"context"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"github.com/fsnotify/fsnotify"
@@ -35,6 +36,11 @@ type Watcher struct {
 	clientsMutex      sync.RWMutex
 	configReloadMu    sync.Mutex
 	configReloadTimer *time.Timer
+	serverUpdateMu    sync.Mutex
+	serverUpdateTimer *time.Timer
+	serverUpdateLast  time.Time
+	serverUpdatePend  bool
+	stopped           atomic.Bool
 	reloadCallback    func(*config.Config)
 	watcher           *fsnotify.Watcher
 	lastAuthHashes    map[string]string
@@ -76,6 +82,7 @@ const (
 	replaceCheckDelay        = 50 * time.Millisecond
 	configReloadDebounce     = 150 * time.Millisecond
 	authRemoveDebounceWindow = 1 * time.Second
+	serverUpdateDebounce     = 1 * time.Second
 )
 
 // NewWatcher creates a new file watcher instance
@@ -114,8 +121,10 @@ func (w *Watcher) Start(ctx context.Context) error {
 
 // Stop stops the file watcher
 func (w *Watcher) Stop() error {
+	w.stopped.Store(true)
 	w.stopDispatch()
 	w.stopConfigReloadTimer()
+	w.stopServerUpdateTimer()
 	return w.watcher.Close()
 }
 

internal/watcher/watcher_test.go

@@ -441,6 +441,46 @@ func TestRemoveClientRemovesHash(t *testing.T) {
 	}
 }
 
+func TestTriggerServerUpdateCancelsPendingTimerOnImmediate(t *testing.T) {
+	tmpDir := t.TempDir()
+	cfg := &config.Config{AuthDir: tmpDir}
+
+	var reloads int32
+	w := &Watcher{
+		reloadCallback: func(*config.Config) {
+			atomic.AddInt32(&reloads, 1)
+		},
+	}
+	w.SetConfig(cfg)
+
+	w.serverUpdateMu.Lock()
+	w.serverUpdateLast = time.Now().Add(-(serverUpdateDebounce - 100*time.Millisecond))
+	w.serverUpdateMu.Unlock()
+	w.triggerServerUpdate(cfg)
+
+	if got := atomic.LoadInt32(&reloads); got != 0 {
+		t.Fatalf("expected no immediate reload, got %d", got)
+	}
+
+	w.serverUpdateMu.Lock()
+	if !w.serverUpdatePend || w.serverUpdateTimer == nil {
+		w.serverUpdateMu.Unlock()
+		t.Fatal("expected a pending server update timer")
+	}
+	w.serverUpdateLast = time.Now().Add(-(serverUpdateDebounce + 10*time.Millisecond))
+	w.serverUpdateMu.Unlock()
+
+	w.triggerServerUpdate(cfg)
+	if got := atomic.LoadInt32(&reloads); got != 1 {
+		t.Fatalf("expected immediate reload once, got %d", got)
+	}
+
+	time.Sleep(250 * time.Millisecond)
+	if got := atomic.LoadInt32(&reloads); got != 1 {
+		t.Fatalf("expected pending timer to be cancelled, got %d reloads", got)
+	}
+}
+
 func TestShouldDebounceRemove(t *testing.T) {
 	w := &Watcher{}
 	path := filepath.Clean("test.json")

sdk/api/handlers/openai/openai_responses_websocket.go

@@ -26,7 +26,6 @@ const (
 	wsRequestTypeAppend  = "response.append"
 	wsEventTypeError     = "error"
 	wsEventTypeCompleted = "response.completed"
-	wsEventTypeDone      = "response.done"
 	wsDoneMarker         = "[DONE]"
 	wsTurnStateHeader    = "x-codex-turn-state"
 	wsRequestBodyKey     = "REQUEST_BODY_OVERRIDE"
@@ -469,9 +468,6 @@ func (h *OpenAIResponsesAPIHandler) forwardResponsesWebsocket(
 			for i := range payloads {
 				eventType := gjson.GetBytes(payloads[i], "type").String()
 				if eventType == wsEventTypeCompleted {
-					// log.Infof("replace %s with %s", wsEventTypeCompleted, wsEventTypeDone)
-					payloads[i], _ = sjson.SetBytes(payloads[i], "type", wsEventTypeDone)
-
 					completed = true
 					completedOutput = responseCompletedOutputFromPayload(payloads[i])
 				}

sdk/api/handlers/openai/openai_responses_websocket_test.go

@@ -2,12 +2,15 @@ package openai
 
 import (
 	"bytes"
+	"errors"
 	"net/http"
 	"net/http/httptest"
 	"strings"
 	"testing"
 
 	"github.com/gin-gonic/gin"
+	"github.com/gorilla/websocket"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
 	"github.com/tidwall/gjson"
 )
 
@@ -247,3 +250,79 @@ func TestSetWebsocketRequestBody(t *testing.T) {
 		t.Fatalf("request body = %q, want %q", string(bodyBytes), "event body")
 	}
 }
+
+func TestForwardResponsesWebsocketPreservesCompletedEvent(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	serverErrCh := make(chan error, 1)
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		conn, err := responsesWebsocketUpgrader.Upgrade(w, r, nil)
+		if err != nil {
+			serverErrCh <- err
+			return
+		}
+		defer func() {
+			errClose := conn.Close()
+			if errClose != nil {
+				serverErrCh <- errClose
+			}
+		}()
+
+		ctx, _ := gin.CreateTestContext(httptest.NewRecorder())
+		ctx.Request = r
+
+		data := make(chan []byte, 1)
+		errCh := make(chan *interfaces.ErrorMessage)
+		data <- []byte("data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp-1\",\"output\":[{\"type\":\"message\",\"id\":\"out-1\"}]}}\n\n")
+		close(data)
+		close(errCh)
+
+		var bodyLog strings.Builder
+		completedOutput, err := (*OpenAIResponsesAPIHandler)(nil).forwardResponsesWebsocket(
+			ctx,
+			conn,
+			func(...interface{}) {},
+			data,
+			errCh,
+			&bodyLog,
+			"session-1",
+		)
+		if err != nil {
+			serverErrCh <- err
+			return
+		}
+		if gjson.GetBytes(completedOutput, "0.id").String() != "out-1" {
+			serverErrCh <- errors.New("completed output not captured")
+			return
+		}
+		serverErrCh <- nil
+	}))
+	defer server.Close()
+
+	wsURL := "ws" + strings.TrimPrefix(server.URL, "http")
+	conn, _, err := websocket.DefaultDialer.Dial(wsURL, nil)
+	if err != nil {
+		t.Fatalf("dial websocket: %v", err)
+	}
+	defer func() {
+		errClose := conn.Close()
+		if errClose != nil {
+			t.Fatalf("close websocket: %v", errClose)
+		}
+	}()
+
+	_, payload, errReadMessage := conn.ReadMessage()
+	if errReadMessage != nil {
+		t.Fatalf("read websocket message: %v", errReadMessage)
+	}
+	if gjson.GetBytes(payload, "type").String() != wsEventTypeCompleted {
+		t.Fatalf("payload type = %s, want %s", gjson.GetBytes(payload, "type").String(), wsEventTypeCompleted)
+	}
+	if strings.Contains(string(payload), "response.done") {
+		t.Fatalf("payload unexpectedly rewrote completed event: %s", payload)
+	}
+
+	if errServer := <-serverErrCh; errServer != nil {
+		t.Fatalf("server error: %v", errServer)
+	}
+}

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_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)