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

2026-03-09 15:25:17 +00:00 · 2026-02-11 16:20:20 +00:00
parent 544238772a 2334a2b174
commit ab0bf1b517
15 changed files with 2476 additions and 48 deletions
--- a/internal/runtime/executor/github_copilot_executor.go
+++ b/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/thinking"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
@@ -471,9 +472,14 @@ func detectVisionContent(body []byte) bool {
 	return false
 }

-// normalizeModel is a no-op as GitHub Copilot accepts model names directly.
-// Model mapping should be done at the registry level if needed.
-func (e *GitHubCopilotExecutor) normalizeModel(_ string, body []byte) []byte {
+// normalizeModel strips the suffix (e.g. "(medium)") from the model name
+// before sending to GitHub Copilot, as the upstream API does not accept
+// suffixed model identifiers.
+func (e *GitHubCopilotExecutor) normalizeModel(model string, body []byte) []byte {
+	baseModel := thinking.ParseSuffix(model).ModelName
+	if baseModel != model {
+		body, _ = sjson.SetBytes(body, "model", baseModel)
+	}
 	return body
 }

--- a/internal/runtime/executor/github_copilot_executor_test.go
+++ b/internal/runtime/executor/github_copilot_executor_test.go
@@ -0,0 +1,54 @@
+package executor
+
+import (
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestGitHubCopilotNormalizeModel_StripsSuffix(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name      string
+		model     string
+		wantModel string
+	}{
+		{
+			name:      "suffix stripped",
+			model:     "claude-opus-4.6(medium)",
+			wantModel: "claude-opus-4.6",
+		},
+		{
+			name:      "no suffix unchanged",
+			model:     "claude-opus-4.6",
+			wantModel: "claude-opus-4.6",
+		},
+		{
+			name:      "different suffix stripped",
+			model:     "gpt-4o(high)",
+			wantModel: "gpt-4o",
+		},
+		{
+			name:      "numeric suffix stripped",
+			model:     "gemini-2.5-pro(8192)",
+			wantModel: "gemini-2.5-pro",
+		},
+	}
+
+	e := &GitHubCopilotExecutor{}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			body := []byte(`{"model":"` + tt.model + `","messages":[]}`)
+			got := e.normalizeModel(tt.model, body)
+
+			gotModel := gjson.GetBytes(got, "model").String()
+			if gotModel != tt.wantModel {
+				t.Fatalf("normalizeModel() model = %q, want %q", gotModel, tt.wantModel)
+			}
+		})
+	}
+}
--- a/internal/runtime/executor/kiro_executor.go
+++ b/internal/runtime/executor/kiro_executor.go
@@ -519,8 +519,12 @@ func buildKiroPayloadForFormat(body []byte, modelID, profileArn, origin string,
 	case "openai":
 		log.Debugf("kiro: using OpenAI payload builder for source format: %s", sourceFormat.String())
 		return kiroopenai.BuildKiroPayloadFromOpenAI(body, modelID, profileArn, origin, isAgentic, isChatOnly, headers, nil)
+	case "kiro":
+		// Body is already in Kiro format — pass through directly (used by callKiroRawAndBuffer)
+		log.Debugf("kiro: body already in Kiro format, passing through directly")
+		return body, false
 	default:
-		// Default to Claude format (also handles "claude", "kiro", etc.)
+		// Default to Claude format
 		log.Debugf("kiro: using Claude payload builder for source format: %s", sourceFormat.String())
 		return kiroclaude.BuildKiroPayload(body, modelID, profileArn, origin, isAgentic, isChatOnly, headers, nil)
 	}
@@ -636,6 +640,13 @@ func (e *KiroExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	rateLimiter.WaitForToken(tokenKey)
 	log.Debugf("kiro: rate limiter cleared for token %s", tokenKey)

+	// Check for pure web_search request
+	// Route to MCP endpoint instead of normal Kiro API
+	if kiroclaude.HasWebSearchTool(req.Payload) {
+		log.Infof("kiro: detected pure web_search request (non-stream), routing to MCP endpoint")
+		return e.handleWebSearch(ctx, auth, req, opts, accessToken, profileArn)
+	}
+
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)

@@ -1057,6 +1068,13 @@ func (e *KiroExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	rateLimiter.WaitForToken(tokenKey)
 	log.Debugf("kiro: stream rate limiter cleared for token %s", tokenKey)

+	// Check for pure web_search request
+	// Route to MCP endpoint instead of normal Kiro API
+	if kiroclaude.HasWebSearchTool(req.Payload) {
+		log.Infof("kiro: detected pure web_search request, routing to MCP endpoint")
+		return e.handleWebSearchStream(ctx, auth, req, opts, accessToken, profileArn)
+	}
+
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
 	defer reporter.trackFailure(ctx, &err)

@@ -4096,6 +4114,539 @@ func (e *KiroExecutor) isTokenExpired(accessToken string) bool {
 	return isExpired
 }

-// NOTE: Message merging functions moved to internal/translator/kiro/common/message_merge.go
-// NOTE: Tool calling support functions moved to internal/translator/kiro/claude/kiro_claude_tools.go
-// The executor now uses kiroclaude.* and kirocommon.* functions instead
+const maxWebSearchIterations = 5
+
+// handleWebSearchStream handles web_search requests:
+// Step 1: tools/list (sync) → fetch/cache tool description
+// Step 2+: MCP search → InjectToolResultsClaude → callKiroAndBuffer loop
+// Note: We skip the "model decides to search" step because Claude Code already
+// decided to use web_search. The Kiro tool description restricts non-coding
+// topics, so asking the model again would cause it to refuse valid searches.
+func (e *KiroExecutor) handleWebSearchStream(
+	ctx context.Context,
+	auth *cliproxyauth.Auth,
+	req cliproxyexecutor.Request,
+	opts cliproxyexecutor.Options,
+	accessToken, profileArn string,
+) (<-chan cliproxyexecutor.StreamChunk, error) {
+	// Extract search query from Claude Code's web_search tool_use
+	query := kiroclaude.ExtractSearchQuery(req.Payload)
+	if query == "" {
+		log.Warnf("kiro/websearch: failed to extract search query, falling back to normal flow")
+		return e.callKiroDirectStream(ctx, auth, req, opts, accessToken, profileArn)
+	}
+
+	// Build MCP endpoint based on region
+	region := kiroDefaultRegion
+	if auth != nil && auth.Metadata != nil {
+		if r, ok := auth.Metadata["api_region"].(string); ok && r != "" {
+			region = r
+		}
+	}
+	mcpEndpoint := fmt.Sprintf("https://q.%s.amazonaws.com/mcp", region)
+
+	// ── Step 1: tools/list (SYNC) — cache tool description ──
+	{
+		tokenKey := getTokenKey(auth)
+		fp := getGlobalFingerprintManager().GetFingerprint(tokenKey)
+		var authAttrs map[string]string
+		if auth != nil {
+			authAttrs = auth.Attributes
+		}
+		kiroclaude.FetchToolDescription(mcpEndpoint, accessToken, newKiroHTTPClientWithPooling(ctx, e.cfg, auth, 30*time.Second), fp, authAttrs)
+	}
+
+	// Create output channel
+	out := make(chan cliproxyexecutor.StreamChunk)
+
+	go func() {
+		defer close(out)
+
+		// Send message_start event to client
+		messageStartEvent := kiroclaude.SseEvent{
+			Event: "message_start",
+			Data: map[string]interface{}{
+				"type": "message_start",
+				"message": map[string]interface{}{
+					"id":            kiroclaude.GenerateMessageID(),
+					"type":          "message",
+					"role":          "assistant",
+					"model":         req.Model,
+					"content":       []interface{}{},
+					"stop_reason":   nil,
+					"stop_sequence": nil,
+					"usage": map[string]interface{}{
+						"input_tokens":                len(req.Payload) / 4,
+						"output_tokens":               0,
+						"cache_creation_input_tokens": 0,
+						"cache_read_input_tokens":     0,
+					},
+				},
+			},
+		}
+		select {
+		case <-ctx.Done():
+			return
+		case out <- cliproxyexecutor.StreamChunk{Payload: []byte(messageStartEvent.ToSSEString())}:
+		}
+
+		// ── Step 2+: MCP search → InjectToolResultsClaude → callKiroAndBuffer loop ──
+		contentBlockIndex := 0
+		currentQuery := query
+
+		// Replace web_search tool description with a minimal one that allows re-search.
+		// The original tools/list description from Kiro restricts non-coding topics,
+		// but we've already decided to search. We keep the tool so the model can
+		// request additional searches when results are insufficient.
+		simplifiedPayload, simplifyErr := kiroclaude.ReplaceWebSearchToolDescription(bytes.Clone(req.Payload))
+		if simplifyErr != nil {
+			log.Warnf("kiro/websearch: failed to simplify web_search tool: %v, using original payload", simplifyErr)
+			simplifiedPayload = bytes.Clone(req.Payload)
+		}
+
+		currentClaudePayload := simplifiedPayload
+		totalSearches := 0
+
+		// Generate toolUseId for the first iteration (Claude Code already decided to search)
+		currentToolUseId := fmt.Sprintf("srvtoolu_%s", kiroclaude.GenerateToolUseID())
+
+		for iteration := 0; iteration < maxWebSearchIterations; iteration++ {
+			log.Infof("kiro/websearch: search iteration %d/%d — query: %s",
+				iteration+1, maxWebSearchIterations, currentQuery)
+
+			// MCP search
+			_, mcpRequest := kiroclaude.CreateMcpRequest(currentQuery)
+			tokenKey := getTokenKey(auth)
+			fp := getGlobalFingerprintManager().GetFingerprint(tokenKey)
+			var authAttrs map[string]string
+			if auth != nil {
+				authAttrs = auth.Attributes
+			}
+			handler := kiroclaude.NewWebSearchHandler(mcpEndpoint, accessToken, newKiroHTTPClientWithPooling(ctx, e.cfg, auth, 30*time.Second), fp, authAttrs)
+			mcpResponse, mcpErr := handler.CallMcpAPI(mcpRequest)
+
+			var searchResults *kiroclaude.WebSearchResults
+			if mcpErr != nil {
+				log.Warnf("kiro/websearch: MCP API call failed: %v, continuing with empty results", mcpErr)
+			} else {
+				searchResults = kiroclaude.ParseSearchResults(mcpResponse)
+			}
+
+			resultCount := 0
+			if searchResults != nil {
+				resultCount = len(searchResults.Results)
+			}
+			totalSearches++
+			log.Infof("kiro/websearch: iteration %d — got %d search results", iteration+1, resultCount)
+
+			// Send search indicator events to client
+			searchEvents := kiroclaude.GenerateSearchIndicatorEvents(currentQuery, currentToolUseId, searchResults, contentBlockIndex)
+			for _, event := range searchEvents {
+				select {
+				case <-ctx.Done():
+					return
+				case out <- cliproxyexecutor.StreamChunk{Payload: []byte(event.ToSSEString())}:
+				}
+			}
+			contentBlockIndex += 2
+
+			// Inject tool_use + tool_result into Claude payload, then call GAR
+			var err error
+			currentClaudePayload, err = kiroclaude.InjectToolResultsClaude(currentClaudePayload, currentToolUseId, currentQuery, searchResults)
+			if err != nil {
+				log.Warnf("kiro/websearch: failed to inject tool results: %v", err)
+				e.sendFallbackText(ctx, out, contentBlockIndex, currentQuery, searchResults)
+				break
+			}
+
+			// Call GAR with modified Claude payload (full translation pipeline)
+			modifiedReq := req
+			modifiedReq.Payload = currentClaudePayload
+			kiroChunks, kiroErr := e.callKiroAndBuffer(ctx, auth, modifiedReq, opts, accessToken, profileArn)
+			if kiroErr != nil {
+				log.Warnf("kiro/websearch: Kiro API failed at iteration %d: %v", iteration+1, kiroErr)
+				e.sendFallbackText(ctx, out, contentBlockIndex, currentQuery, searchResults)
+				break
+			}
+
+			// Analyze response
+			analysis := kiroclaude.AnalyzeBufferedStream(kiroChunks)
+			log.Infof("kiro/websearch: iteration %d — stop_reason: %s, has_tool_use: %v, query: %s, toolUseId: %s",
+				iteration+1, analysis.StopReason, analysis.HasWebSearchToolUse, analysis.WebSearchQuery, analysis.WebSearchToolUseId)
+
+			if analysis.HasWebSearchToolUse && analysis.WebSearchQuery != "" && iteration+1 < maxWebSearchIterations {
+				// Model wants another search
+				filteredChunks := kiroclaude.FilterChunksForClient(kiroChunks, analysis.WebSearchToolUseIndex, contentBlockIndex)
+				for _, chunk := range filteredChunks {
+					select {
+					case <-ctx.Done():
+						return
+					case out <- cliproxyexecutor.StreamChunk{Payload: chunk}:
+					}
+				}
+
+				currentQuery = analysis.WebSearchQuery
+				currentToolUseId = analysis.WebSearchToolUseId
+				continue
+			}
+
+			// Model returned final response — stream to client
+			for _, chunk := range kiroChunks {
+				if contentBlockIndex > 0 && len(chunk) > 0 {
+					adjusted, shouldForward := kiroclaude.AdjustSSEChunk(chunk, contentBlockIndex)
+					if !shouldForward {
+						continue
+					}
+					select {
+					case <-ctx.Done():
+						return
+					case out <- cliproxyexecutor.StreamChunk{Payload: adjusted}:
+					}
+				} else {
+					select {
+					case <-ctx.Done():
+						return
+					case out <- cliproxyexecutor.StreamChunk{Payload: chunk}:
+					}
+				}
+			}
+			log.Infof("kiro/websearch: completed after %d search iteration(s), total searches: %d", iteration+1, totalSearches)
+			return
+		}
+
+		log.Warnf("kiro/websearch: reached max iterations (%d), stopping search loop", maxWebSearchIterations)
+	}()
+
+	return out, nil
+}
+
+// callKiroAndBuffer calls the Kiro API and buffers all response chunks.
+// Returns the buffered chunks for analysis before forwarding to client.
+func (e *KiroExecutor) callKiroAndBuffer(
+	ctx context.Context,
+	auth *cliproxyauth.Auth,
+	req cliproxyexecutor.Request,
+	opts cliproxyexecutor.Options,
+	accessToken, profileArn string,
+) ([][]byte, error) {
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("kiro")
+	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+	log.Debugf("kiro/websearch GAR request: %d bytes", len(body))
+
+	kiroModelID := e.mapModelToKiro(req.Model)
+	isAgentic, isChatOnly := determineAgenticMode(req.Model)
+	effectiveProfileArn := getEffectiveProfileArnWithWarning(auth, profileArn)
+
+	tokenKey := ""
+	if auth != nil {
+		tokenKey = auth.ID
+	}
+
+	kiroStream, err := e.executeStreamWithRetry(
+		ctx, auth, req, opts, accessToken, effectiveProfileArn,
+		nil, body, from, nil, "", kiroModelID, isAgentic, isChatOnly, tokenKey,
+	)
+	if err != nil {
+		return nil, err
+	}
+
+	// Buffer all chunks
+	var chunks [][]byte
+	for chunk := range kiroStream {
+		if chunk.Err != nil {
+			return chunks, chunk.Err
+		}
+		if len(chunk.Payload) > 0 {
+			chunks = append(chunks, bytes.Clone(chunk.Payload))
+		}
+	}
+
+	log.Debugf("kiro/websearch GAR response: %d chunks buffered", len(chunks))
+
+	return chunks, nil
+}
+
+// callKiroRawAndBuffer calls the Kiro API with a pre-built Kiro payload (no translation).
+// Used in the web search loop where the payload is modified directly in Kiro format.
+func (e *KiroExecutor) callKiroRawAndBuffer(
+	ctx context.Context,
+	auth *cliproxyauth.Auth,
+	req cliproxyexecutor.Request,
+	opts cliproxyexecutor.Options,
+	accessToken, profileArn string,
+	kiroBody []byte,
+) ([][]byte, error) {
+	kiroModelID := e.mapModelToKiro(req.Model)
+	isAgentic, isChatOnly := determineAgenticMode(req.Model)
+	effectiveProfileArn := getEffectiveProfileArnWithWarning(auth, profileArn)
+
+	tokenKey := ""
+	if auth != nil {
+		tokenKey = auth.ID
+	}
+	log.Debugf("kiro/websearch GAR raw request: %d bytes", len(kiroBody))
+
+	kiroFormat := sdktranslator.FromString("kiro")
+	kiroStream, err := e.executeStreamWithRetry(
+		ctx, auth, req, opts, accessToken, effectiveProfileArn,
+		nil, kiroBody, kiroFormat, nil, "", kiroModelID, isAgentic, isChatOnly, tokenKey,
+	)
+	if err != nil {
+		return nil, err
+	}
+
+	// Buffer all chunks
+	var chunks [][]byte
+	for chunk := range kiroStream {
+		if chunk.Err != nil {
+			return chunks, chunk.Err
+		}
+		if len(chunk.Payload) > 0 {
+			chunks = append(chunks, bytes.Clone(chunk.Payload))
+		}
+	}
+
+	log.Debugf("kiro/websearch GAR raw response: %d chunks buffered", len(chunks))
+
+	return chunks, nil
+}
+
+// callKiroDirectStream creates a direct streaming channel to Kiro API without search.
+func (e *KiroExecutor) callKiroDirectStream(
+	ctx context.Context,
+	auth *cliproxyauth.Auth,
+	req cliproxyexecutor.Request,
+	opts cliproxyexecutor.Options,
+	accessToken, profileArn string,
+) (<-chan cliproxyexecutor.StreamChunk, error) {
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("kiro")
+	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+
+	kiroModelID := e.mapModelToKiro(req.Model)
+	isAgentic, isChatOnly := determineAgenticMode(req.Model)
+	effectiveProfileArn := getEffectiveProfileArnWithWarning(auth, profileArn)
+
+	tokenKey := ""
+	if auth != nil {
+		tokenKey = auth.ID
+	}
+
+	return e.executeStreamWithRetry(
+		ctx, auth, req, opts, accessToken, effectiveProfileArn,
+		nil, body, from, nil, "", kiroModelID, isAgentic, isChatOnly, tokenKey,
+	)
+}
+
+// sendFallbackText sends a simple text response when the Kiro API fails during the search loop.
+func (e *KiroExecutor) sendFallbackText(
+	ctx context.Context,
+	out chan<- cliproxyexecutor.StreamChunk,
+	contentBlockIndex int,
+	query string,
+	searchResults *kiroclaude.WebSearchResults,
+) {
+	// Generate a simple text summary from search results
+	summary := kiroclaude.FormatSearchContextPrompt(query, searchResults)
+
+	events := []kiroclaude.SseEvent{
+		{
+			Event: "content_block_start",
+			Data: map[string]interface{}{
+				"type":  "content_block_start",
+				"index": contentBlockIndex,
+				"content_block": map[string]interface{}{
+					"type": "text",
+					"text": "",
+				},
+			},
+		},
+		{
+			Event: "content_block_delta",
+			Data: map[string]interface{}{
+				"type":  "content_block_delta",
+				"index": contentBlockIndex,
+				"delta": map[string]interface{}{
+					"type": "text_delta",
+					"text": summary,
+				},
+			},
+		},
+		{
+			Event: "content_block_stop",
+			Data: map[string]interface{}{
+				"type":  "content_block_stop",
+				"index": contentBlockIndex,
+			},
+		},
+	}
+
+	for _, event := range events {
+		select {
+		case <-ctx.Done():
+			return
+		case out <- cliproxyexecutor.StreamChunk{Payload: []byte(event.ToSSEString())}:
+		}
+	}
+
+	// Send message_delta with end_turn and message_stop
+	msgDelta := kiroclaude.SseEvent{
+		Event: "message_delta",
+		Data: map[string]interface{}{
+			"type": "message_delta",
+			"delta": map[string]interface{}{
+				"stop_reason":   "end_turn",
+				"stop_sequence": nil,
+			},
+			"usage": map[string]interface{}{
+				"output_tokens": len(summary) / 4,
+			},
+		},
+	}
+	select {
+	case <-ctx.Done():
+		return
+	case out <- cliproxyexecutor.StreamChunk{Payload: []byte(msgDelta.ToSSEString())}:
+	}
+
+	msgStop := kiroclaude.SseEvent{
+		Event: "message_stop",
+		Data: map[string]interface{}{
+			"type": "message_stop",
+		},
+	}
+	select {
+	case <-ctx.Done():
+		return
+	case out <- cliproxyexecutor.StreamChunk{Payload: []byte(msgStop.ToSSEString())}:
+	}
+
+}
+
+// handleWebSearch handles web_search requests for non-streaming Execute path.
+// Performs MCP search synchronously, injects results into the request payload,
+// then calls the normal non-streaming Kiro API path which returns a proper
+// Claude JSON response (not SSE chunks).
+func (e *KiroExecutor) handleWebSearch(
+	ctx context.Context,
+	auth *cliproxyauth.Auth,
+	req cliproxyexecutor.Request,
+	opts cliproxyexecutor.Options,
+	accessToken, profileArn string,
+) (cliproxyexecutor.Response, error) {
+	// Extract search query from Claude Code's web_search tool_use
+	query := kiroclaude.ExtractSearchQuery(req.Payload)
+	if query == "" {
+		log.Warnf("kiro/websearch: non-stream: failed to extract search query, falling back to normal Execute")
+		// Fall through to normal non-streaming path
+		return e.executeNonStreamFallback(ctx, auth, req, opts, accessToken, profileArn)
+	}
+
+	// Build MCP endpoint based on region
+	region := kiroDefaultRegion
+	if auth != nil && auth.Metadata != nil {
+		if r, ok := auth.Metadata["api_region"].(string); ok && r != "" {
+			region = r
+		}
+	}
+	mcpEndpoint := fmt.Sprintf("https://q.%s.amazonaws.com/mcp", region)
+
+	// Step 1: Fetch/cache tool description (sync)
+	{
+		tokenKey := getTokenKey(auth)
+		fp := getGlobalFingerprintManager().GetFingerprint(tokenKey)
+		var authAttrs map[string]string
+		if auth != nil {
+			authAttrs = auth.Attributes
+		}
+		kiroclaude.FetchToolDescription(mcpEndpoint, accessToken, newKiroHTTPClientWithPooling(ctx, e.cfg, auth, 30*time.Second), fp, authAttrs)
+	}
+
+	// Step 2: Perform MCP search
+	_, mcpRequest := kiroclaude.CreateMcpRequest(query)
+	tokenKey := getTokenKey(auth)
+	fp := getGlobalFingerprintManager().GetFingerprint(tokenKey)
+	var authAttrs map[string]string
+	if auth != nil {
+		authAttrs = auth.Attributes
+	}
+	handler := kiroclaude.NewWebSearchHandler(mcpEndpoint, accessToken, newKiroHTTPClientWithPooling(ctx, e.cfg, auth, 30*time.Second), fp, authAttrs)
+	mcpResponse, mcpErr := handler.CallMcpAPI(mcpRequest)
+
+	var searchResults *kiroclaude.WebSearchResults
+	if mcpErr != nil {
+		log.Warnf("kiro/websearch: non-stream: MCP API call failed: %v, continuing with empty results", mcpErr)
+	} else {
+		searchResults = kiroclaude.ParseSearchResults(mcpResponse)
+	}
+
+	resultCount := 0
+	if searchResults != nil {
+		resultCount = len(searchResults.Results)
+	}
+	log.Infof("kiro/websearch: non-stream: got %d search results for query: %s", resultCount, query)
+
+	// Step 3: Inject search tool_use + tool_result into Claude payload
+	currentToolUseId := fmt.Sprintf("srvtoolu_%s", kiroclaude.GenerateToolUseID())
+	modifiedPayload, err := kiroclaude.InjectToolResultsClaude(bytes.Clone(req.Payload), currentToolUseId, query, searchResults)
+	if err != nil {
+		log.Warnf("kiro/websearch: non-stream: failed to inject tool results: %v, falling back", err)
+		return e.executeNonStreamFallback(ctx, auth, req, opts, accessToken, profileArn)
+	}
+
+	// Step 4: Call Kiro API via the normal non-streaming path (executeWithRetry)
+	// This path uses parseEventStream → BuildClaudeResponse → TranslateNonStream
+	// to produce a proper Claude JSON response
+	modifiedReq := req
+	modifiedReq.Payload = modifiedPayload
+
+	resp, err := e.executeNonStreamFallback(ctx, auth, modifiedReq, opts, accessToken, profileArn)
+	if err != nil {
+		return resp, err
+	}
+
+	// Step 5: Inject server_tool_use + web_search_tool_result into response
+	// so Claude Code can display "Did X searches in Ys"
+	indicators := []kiroclaude.SearchIndicator{
+		{
+			ToolUseID: currentToolUseId,
+			Query:     query,
+			Results:   searchResults,
+		},
+	}
+	injectedPayload, injErr := kiroclaude.InjectSearchIndicatorsInResponse(resp.Payload, indicators)
+	if injErr != nil {
+		log.Warnf("kiro/websearch: non-stream: failed to inject search indicators: %v", injErr)
+	} else {
+		resp.Payload = injectedPayload
+	}
+
+	return resp, nil
+}
+
+// executeNonStreamFallback runs the standard non-streaming Execute path for a request.
+// Used by handleWebSearch after injecting search results, or as a fallback.
+func (e *KiroExecutor) executeNonStreamFallback(
+	ctx context.Context,
+	auth *cliproxyauth.Auth,
+	req cliproxyexecutor.Request,
+	opts cliproxyexecutor.Options,
+	accessToken, profileArn string,
+) (cliproxyexecutor.Response, error) {
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("kiro")
+	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
+
+	kiroModelID := e.mapModelToKiro(req.Model)
+	isAgentic, isChatOnly := determineAgenticMode(req.Model)
+	effectiveProfileArn := getEffectiveProfileArnWithWarning(auth, profileArn)
+	tokenKey := getTokenKey(auth)
+
+	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
+	var err error
+	defer reporter.trackFailure(ctx, &err)
+
+	resp, err := e.executeWithRetry(ctx, auth, req, opts, accessToken, effectiveProfileArn, nil, body, from, to, reporter, "", kiroModelID, isAgentic, isChatOnly, tokenKey)
+	return resp, err
+}
--- a/internal/translator/codex/openai/responses/codex_openai-responses_request.go
+++ b/internal/translator/codex/openai/responses/codex_openai-responses_request.go
@@ -27,6 +27,9 @@ func ConvertOpenAIResponsesRequestToCodex(modelName string, inputRawJSON []byte,
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "top_p")
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "service_tier")

+	// Delete the user field as it is not supported by the Codex upstream.
+	rawJSON, _ = sjson.DeleteBytes(rawJSON, "user")
+
 	// Convert role "system" to "developer" in input array to comply with Codex API requirements.
 	rawJSON = convertSystemRoleToDeveloper(rawJSON)

--- a/internal/translator/codex/openai/responses/codex_openai-responses_request_test.go
+++ b/internal/translator/codex/openai/responses/codex_openai-responses_request_test.go
@@ -263,3 +263,20 @@ func TestConvertSystemRoleToDeveloper_AssistantRole(t *testing.T) {
 		t.Errorf("Expected third role 'assistant', got '%s'", thirdRole.String())
 	}
 }
+
+func TestUserFieldDeletion(t *testing.T) {  
+	inputJSON := []byte(`{  
+		"model": "gpt-5.2",  
+		"user": "test-user",  
+		"input": [{"role": "user", "content": "Hello"}]  
+	}`)  
+	  
+	output := ConvertOpenAIResponsesRequestToCodex("gpt-5.2", inputJSON, false)  
+	outputStr := string(output)  
+	  
+	// Verify user field is deleted  
+	userField := gjson.Get(outputStr, "user")  
+	if userField.Exists() {
+		t.Errorf("user field should be deleted, but it was found with value: %s", userField.Raw)
+	}
+}
--- a/internal/translator/kiro/claude/kiro_claude_request.go
+++ b/internal/translator/kiro/claude/kiro_claude_request.go
@@ -17,6 +17,9 @@ import (
 	"github.com/tidwall/gjson"
 )

+// remoteWebSearchDescription is a minimal fallback for when dynamic fetch from MCP tools/list hasn't completed yet.
+const remoteWebSearchDescription = "WebSearch looks up information outside the model's training data. Supports multiple queries to gather comprehensive information."
+
 // Kiro API request structs - field order determines JSON key order

 // KiroPayload is the top-level request structure for Kiro API
@@ -219,26 +222,7 @@ func BuildKiroPayload(claudeBody []byte, modelID, profileArn, origin string, isA
 	}

 	// Convert Claude tools to Kiro format
-	kiroTools, hasWebSearch := convertClaudeToolsToKiro(tools)
-
-	// If web_search was requested but filtered, inject alternative hint
-	if hasWebSearch {
-		webSearchHint := `[CRITICAL WEB ACCESS INSTRUCTION]
-You have the Fetch/read_url_content tool available. When the user asks about current events, weather, news, or any information that requires web access:
- DO NOT say you cannot search the web
- DO NOT refuse to help with web-related queries
- IMMEDIATELY use the Fetch tool to access relevant URLs
- Use well-known official websites, documentation sites, or API endpoints
- Construct appropriate URLs based on the query context
-
-IMPORTANT: Always attempt to fetch information FIRST before declining. You CAN access the web via Fetch.`
-		if systemPrompt != "" {
-			systemPrompt += "\n" + webSearchHint
-		} else {
-			systemPrompt = webSearchHint
-		}
-		log.Infof("kiro: injected web_search alternative hint (tool was filtered)")
-	}
+	kiroTools := convertClaudeToolsToKiro(tools)

 	// Thinking mode implementation:
 	// Kiro API supports official thinking/reasoning mode via <thinking_mode> tag.
@@ -527,27 +511,15 @@ func ensureKiroInputSchema(parameters interface{}) interface{} {
 	}
 }

-// convertClaudeToolsToKiro converts Claude tools to Kiro format.
-// Returns the converted tools and a boolean indicating if web_search was filtered.
-func convertClaudeToolsToKiro(tools gjson.Result) ([]KiroToolWrapper, bool) {
+// convertClaudeToolsToKiro converts Claude tools to Kiro format
+func convertClaudeToolsToKiro(tools gjson.Result) []KiroToolWrapper {
 	var kiroTools []KiroToolWrapper
-	hasWebSearch := false
 	if !tools.IsArray() {
-		return kiroTools, hasWebSearch
+		return kiroTools
 	}

 	for _, tool := range tools.Array() {
 		name := tool.Get("name").String()
-
-		// Filter out web_search/websearch tools (Kiro API doesn't support them)
-		// This matches the behavior in AIClient-2-API/claude-kiro.js
-		nameLower := strings.ToLower(name)
-		if nameLower == "web_search" || nameLower == "websearch" {
-			log.Debugf("kiro: skipping unsupported tool: %s", name)
-			hasWebSearch = true
-			continue
-		}
-
 		description := tool.Get("description").String()
 		inputSchemaResult := tool.Get("input_schema")
 		var inputSchema interface{}
@@ -569,6 +541,18 @@ func convertClaudeToolsToKiro(tools gjson.Result) ([]KiroToolWrapper, bool) {
 			log.Debugf("kiro: tool '%s' has empty description, using default: %s", name, description)
 		}

+		// Rename web_search → remote_web_search for Kiro API compatibility
+		if name == "web_search" {
+			name = "remote_web_search"
+			// Prefer dynamically fetched description, fall back to hardcoded constant
+			if cached := GetWebSearchDescription(); cached != "" {
+				description = cached
+			} else {
+				description = remoteWebSearchDescription
+			}
+			log.Debugf("kiro: renamed tool web_search → remote_web_search")
+		}
+
 		// Truncate long descriptions (individual tool limit)
 		if len(description) > kirocommon.KiroMaxToolDescLen {
 			truncLen := kirocommon.KiroMaxToolDescLen - 30
@@ -591,7 +575,7 @@ func convertClaudeToolsToKiro(tools gjson.Result) ([]KiroToolWrapper, bool) {
 	// This prevents 500 errors when Claude Code sends too many tools
 	kiroTools = compressToolsIfNeeded(kiroTools)

-	return kiroTools, hasWebSearch
+	return kiroTools
 }

 // processMessages processes Claude messages and builds Kiro history
@@ -654,6 +638,57 @@ func processMessages(messages gjson.Result, modelID, origin string) ([]KiroHisto
 		}
 	}

+	// POST-PROCESSING: Remove orphaned tool_results that have no matching tool_use
+	// in any assistant message. This happens when Claude Code compaction truncates
+	// the conversation and removes the assistant message containing the tool_use,
+	// but keeps the user message with the corresponding tool_result.
+	// Without this fix, Kiro API returns "Improperly formed request".
+	validToolUseIDs := make(map[string]bool)
+	for _, h := range history {
+		if h.AssistantResponseMessage != nil {
+			for _, tu := range h.AssistantResponseMessage.ToolUses {
+				validToolUseIDs[tu.ToolUseID] = true
+			}
+		}
+	}
+
+	// Filter orphaned tool results from history user messages
+	for i, h := range history {
+		if h.UserInputMessage != nil && h.UserInputMessage.UserInputMessageContext != nil {
+			ctx := h.UserInputMessage.UserInputMessageContext
+			if len(ctx.ToolResults) > 0 {
+				filtered := make([]KiroToolResult, 0, len(ctx.ToolResults))
+				for _, tr := range ctx.ToolResults {
+					if validToolUseIDs[tr.ToolUseID] {
+						filtered = append(filtered, tr)
+					} else {
+						log.Debugf("kiro: dropping orphaned tool_result in history[%d]: toolUseId=%s (no matching tool_use)", i, tr.ToolUseID)
+					}
+				}
+				ctx.ToolResults = filtered
+				if len(ctx.ToolResults) == 0 && len(ctx.Tools) == 0 {
+					h.UserInputMessage.UserInputMessageContext = nil
+				}
+			}
+		}
+	}
+
+	// Filter orphaned tool results from current message
+	if len(currentToolResults) > 0 {
+		filtered := make([]KiroToolResult, 0, len(currentToolResults))
+		for _, tr := range currentToolResults {
+			if validToolUseIDs[tr.ToolUseID] {
+				filtered = append(filtered, tr)
+			} else {
+				log.Debugf("kiro: dropping orphaned tool_result in currentMessage: toolUseId=%s (no matching tool_use)", tr.ToolUseID)
+			}
+		}
+		if len(filtered) != len(currentToolResults) {
+			log.Infof("kiro: dropped %d orphaned tool_result(s) from currentMessage (compaction artifact)", len(currentToolResults)-len(filtered))
+		}
+		currentToolResults = filtered
+	}
+
 	return history, currentUserMsg, currentToolResults
 }

@@ -876,6 +911,11 @@ func BuildAssistantMessageStruct(msg gjson.Result) KiroAssistantResponseMessage
 					})
 				}

+				// Rename web_search → remote_web_search to match convertClaudeToolsToKiro
+				if toolName == "web_search" {
+					toolName = "remote_web_search"
+				}
+
 				toolUses = append(toolUses, KiroToolUse{
 					ToolUseID: toolUseID,
 					Name:      toolName,
--- a/internal/translator/kiro/claude/kiro_websearch.go
+++ b/internal/translator/kiro/claude/kiro_websearch.go
--- a/internal/translator/kiro/claude/kiro_websearch_handler.go
+++ b/internal/translator/kiro/claude/kiro_websearch_handler.go
@@ -0,0 +1,270 @@
+// Package claude provides web search handler for Kiro translator.
+// This file implements the MCP API call and response handling.
+package claude
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/google/uuid"
+	kiroauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/kiro"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	log "github.com/sirupsen/logrus"
+)
+
+// Cached web_search tool description fetched from MCP tools/list.
+// Uses atomic.Pointer[sync.Once] for lock-free reads with retry-on-failure:
+// - sync.Once prevents race conditions and deduplicates concurrent calls
+// - On failure, a fresh sync.Once is swapped in to allow retry on next call
+// - On success, sync.Once stays "done" forever — zero overhead for subsequent calls
+var (
+	cachedToolDescription atomic.Value // stores string
+	toolDescOnce          atomic.Pointer[sync.Once]
+	fallbackFpOnce        sync.Once
+	fallbackFp            *kiroauth.Fingerprint
+)
+
+func init() {
+	toolDescOnce.Store(&sync.Once{})
+}
+
+// FetchToolDescription calls MCP tools/list to get the web_search tool description
+// and caches it. Safe to call concurrently — only one goroutine fetches at a time.
+// If the fetch fails, subsequent calls will retry. On success, no further fetches occur.
+// The httpClient parameter allows reusing a shared pooled HTTP client.
+func FetchToolDescription(mcpEndpoint, authToken string, httpClient *http.Client, fp *kiroauth.Fingerprint, authAttrs map[string]string) {
+	toolDescOnce.Load().Do(func() {
+		handler := NewWebSearchHandler(mcpEndpoint, authToken, httpClient, fp, authAttrs)
+		reqBody := []byte(`{"id":"tools_list","jsonrpc":"2.0","method":"tools/list"}`)
+		log.Debugf("kiro/websearch MCP tools/list request: %d bytes", len(reqBody))
+
+		req, err := http.NewRequest("POST", mcpEndpoint, bytes.NewReader(reqBody))
+		if err != nil {
+			log.Warnf("kiro/websearch: failed to create tools/list request: %v", err)
+			toolDescOnce.Store(&sync.Once{}) // allow retry
+			return
+		}
+
+		// Reuse same headers as CallMcpAPI
+		handler.setMcpHeaders(req)
+
+		resp, err := handler.HTTPClient.Do(req)
+		if err != nil {
+			log.Warnf("kiro/websearch: tools/list request failed: %v", err)
+			toolDescOnce.Store(&sync.Once{}) // allow retry
+			return
+		}
+		defer resp.Body.Close()
+
+		body, err := io.ReadAll(resp.Body)
+		if err != nil || resp.StatusCode != http.StatusOK {
+			log.Warnf("kiro/websearch: tools/list returned status %d", resp.StatusCode)
+			toolDescOnce.Store(&sync.Once{}) // allow retry
+			return
+		}
+		log.Debugf("kiro/websearch MCP tools/list response: [%d] %d bytes", resp.StatusCode, len(body))
+
+		// Parse: {"result":{"tools":[{"name":"web_search","description":"..."}]}}
+		var result struct {
+			Result *struct {
+				Tools []struct {
+					Name        string `json:"name"`
+					Description string `json:"description"`
+				} `json:"tools"`
+			} `json:"result"`
+		}
+		if err := json.Unmarshal(body, &result); err != nil || result.Result == nil {
+			log.Warnf("kiro/websearch: failed to parse tools/list response")
+			toolDescOnce.Store(&sync.Once{}) // allow retry
+			return
+		}
+
+		for _, tool := range result.Result.Tools {
+			if tool.Name == "web_search" && tool.Description != "" {
+				cachedToolDescription.Store(tool.Description)
+				log.Infof("kiro/websearch: cached web_search description from tools/list (%d bytes)", len(tool.Description))
+				return // success — sync.Once stays "done", no more fetches
+			}
+		}
+
+		// web_search tool not found in response
+		toolDescOnce.Store(&sync.Once{}) // allow retry
+	})
+}
+
+// GetWebSearchDescription returns the cached web_search tool description,
+// or empty string if not yet fetched. Lock-free via atomic.Value.
+func GetWebSearchDescription() string {
+	if v := cachedToolDescription.Load(); v != nil {
+		return v.(string)
+	}
+	return ""
+}
+
+// WebSearchHandler handles web search requests via Kiro MCP API
+type WebSearchHandler struct {
+	McpEndpoint string
+	HTTPClient  *http.Client
+	AuthToken   string
+	Fingerprint *kiroauth.Fingerprint // optional, for dynamic headers
+	AuthAttrs   map[string]string     // optional, for custom headers from auth.Attributes
+}
+
+// NewWebSearchHandler creates a new WebSearchHandler.
+// If httpClient is nil, a default client with 30s timeout is used.
+// If fingerprint is nil, a random one-off fingerprint is generated.
+// Pass a shared pooled client (e.g. from getKiroPooledHTTPClient) for connection reuse.
+func NewWebSearchHandler(mcpEndpoint, authToken string, httpClient *http.Client, fp *kiroauth.Fingerprint, authAttrs map[string]string) *WebSearchHandler {
+	if httpClient == nil {
+		httpClient = &http.Client{
+			Timeout: 30 * time.Second,
+		}
+	}
+	if fp == nil {
+		// Use a shared fallback fingerprint for callers without token context
+		fallbackFpOnce.Do(func() {
+			mgr := kiroauth.NewFingerprintManager()
+			fallbackFp = mgr.GetFingerprint("mcp-fallback")
+		})
+		fp = fallbackFp
+	}
+	return &WebSearchHandler{
+		McpEndpoint: mcpEndpoint,
+		HTTPClient:  httpClient,
+		AuthToken:   authToken,
+		Fingerprint: fp,
+		AuthAttrs:   authAttrs,
+	}
+}
+
+// setMcpHeaders sets standard MCP API headers on the request,
+// aligned with the GAR request pattern in kiro_executor.go.
+func (h *WebSearchHandler) setMcpHeaders(req *http.Request) {
+	fp := h.Fingerprint
+
+	// 1. Content-Type & Accept (aligned with GAR)
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Accept", "*/*")
+
+	// 2. Kiro-specific headers (aligned with GAR)
+	req.Header.Set("x-amzn-kiro-agent-mode", "vibe")
+	req.Header.Set("x-amzn-codewhisperer-optout", "true")
+
+	// 3. Dynamic fingerprint headers
+	req.Header.Set("User-Agent", fp.BuildUserAgent())
+	req.Header.Set("X-Amz-User-Agent", fp.BuildAmzUserAgent())
+
+	// 4. AWS SDK identifiers (casing aligned with GAR)
+	req.Header.Set("Amz-Sdk-Request", "attempt=1; max=3")
+	req.Header.Set("Amz-Sdk-Invocation-Id", uuid.New().String())
+
+	// 5. Authentication
+	req.Header.Set("Authorization", "Bearer "+h.AuthToken)
+
+	// 6. Custom headers from auth attributes
+	util.ApplyCustomHeadersFromAttrs(req, h.AuthAttrs)
+}
+
+// mcpMaxRetries is the maximum number of retries for MCP API calls.
+const mcpMaxRetries = 2
+
+// CallMcpAPI calls the Kiro MCP API with the given request.
+// Includes retry logic with exponential backoff for retryable errors,
+// aligned with the GAR request retry pattern.
+func (h *WebSearchHandler) CallMcpAPI(request *McpRequest) (*McpResponse, error) {
+	requestBody, err := json.Marshal(request)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal MCP request: %w", err)
+	}
+	log.Debugf("kiro/websearch MCP request → %s (%d bytes)", h.McpEndpoint, len(requestBody))
+
+	var lastErr error
+	for attempt := 0; attempt <= mcpMaxRetries; attempt++ {
+		if attempt > 0 {
+			backoff := time.Duration(1<<attempt) * time.Second
+			if backoff > 10*time.Second {
+				backoff = 10 * time.Second
+			}
+			log.Warnf("kiro/websearch: MCP retry %d/%d after %v (last error: %v)", attempt, mcpMaxRetries, backoff, lastErr)
+			time.Sleep(backoff)
+		}
+
+		req, err := http.NewRequest("POST", h.McpEndpoint, bytes.NewReader(requestBody))
+		if err != nil {
+			return nil, fmt.Errorf("failed to create HTTP request: %w", err)
+		}
+
+		h.setMcpHeaders(req)
+
+		resp, err := h.HTTPClient.Do(req)
+		if err != nil {
+			lastErr = fmt.Errorf("MCP API request failed: %w", err)
+			continue // network error → retry
+		}
+
+		body, err := io.ReadAll(resp.Body)
+		resp.Body.Close()
+		if err != nil {
+			lastErr = fmt.Errorf("failed to read MCP response: %w", err)
+			continue // read error → retry
+		}
+		log.Debugf("kiro/websearch MCP response ← [%d] (%d bytes)", resp.StatusCode, len(body))
+
+		// Retryable HTTP status codes (aligned with GAR: 502, 503, 504)
+		if resp.StatusCode >= 502 && resp.StatusCode <= 504 {
+			lastErr = fmt.Errorf("MCP API returned retryable status %d: %s", resp.StatusCode, string(body))
+			continue
+		}
+
+		if resp.StatusCode != http.StatusOK {
+			return nil, fmt.Errorf("MCP API returned status %d: %s", resp.StatusCode, string(body))
+		}
+
+		var mcpResponse McpResponse
+		if err := json.Unmarshal(body, &mcpResponse); err != nil {
+			return nil, fmt.Errorf("failed to parse MCP response: %w", err)
+		}
+
+		if mcpResponse.Error != nil {
+			code := -1
+			if mcpResponse.Error.Code != nil {
+				code = *mcpResponse.Error.Code
+			}
+			msg := "Unknown error"
+			if mcpResponse.Error.Message != nil {
+				msg = *mcpResponse.Error.Message
+			}
+			return nil, fmt.Errorf("MCP error %d: %s", code, msg)
+		}
+
+		return &mcpResponse, nil
+	}
+
+	return nil, lastErr
+}
+
+// ParseSearchResults extracts WebSearchResults from MCP response
+func ParseSearchResults(response *McpResponse) *WebSearchResults {
+	if response == nil || response.Result == nil || len(response.Result.Content) == 0 {
+		return nil
+	}
+
+	content := response.Result.Content[0]
+	if content.ContentType != "text" {
+		return nil
+	}
+
+	var results WebSearchResults
+	if err := json.Unmarshal([]byte(content.Text), &results); err != nil {
+		log.Warnf("kiro/websearch: failed to parse search results: %v", err)
+		return nil
+	}
+
+	return &results
+}
--- a/internal/watcher/synthesizer/file.go
+++ b/internal/watcher/synthesizer/file.go
@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"strconv"
 	"strings"
 	"time"

@@ -92,6 +93,9 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 			status = coreauth.StatusDisabled
 		}

+		// Read per-account excluded models from the OAuth JSON file
+		perAccountExcluded := extractExcludedModelsFromMetadata(metadata)
+
 		a := &coreauth.Auth{
 			ID:       id,
 			Provider: provider,
@@ -108,11 +112,23 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 			CreatedAt: now,
 			UpdatedAt: now,
 		}
-		ApplyAuthExcludedModelsMeta(a, cfg, nil, "oauth")
+		// Read priority from auth file
+		if rawPriority, ok := metadata["priority"]; ok {
+			switch v := rawPriority.(type) {
+			case float64:
+				a.Attributes["priority"] = strconv.Itoa(int(v))
+			case string:
+				priority := strings.TrimSpace(v)
+				if _, errAtoi := strconv.Atoi(priority); errAtoi == nil {
+					a.Attributes["priority"] = priority
+				}
+			}
+		}
+		ApplyAuthExcludedModelsMeta(a, cfg, perAccountExcluded, "oauth")
 		if provider == "gemini-cli" {
 			if virtuals := SynthesizeGeminiVirtualAuths(a, metadata, now); len(virtuals) > 0 {
 				for _, v := range virtuals {
-					ApplyAuthExcludedModelsMeta(v, cfg, nil, "oauth")
+					ApplyAuthExcludedModelsMeta(v, cfg, perAccountExcluded, "oauth")
 				}
 				out = append(out, a)
 				out = append(out, virtuals...)
@@ -167,6 +183,10 @@ func SynthesizeGeminiVirtualAuths(primary *coreauth.Auth, metadata map[string]an
 		if authPath != "" {
 			attrs["path"] = authPath
 		}
+		// Propagate priority from primary auth to virtual auths
+		if priorityVal, hasPriority := primary.Attributes["priority"]; hasPriority && priorityVal != "" {
+			attrs["priority"] = priorityVal
+		}
 		metadataCopy := map[string]any{
 			"email":             email,
 			"project_id":        projectID,
@@ -239,3 +259,40 @@ func buildGeminiVirtualID(baseID, projectID string) string {
 	replacer := strings.NewReplacer("/", "_", "\\", "_", " ", "_")
 	return fmt.Sprintf("%s::%s", baseID, replacer.Replace(project))
 }
+
+// extractExcludedModelsFromMetadata reads per-account excluded models from the OAuth JSON metadata.
+// Supports both "excluded_models" and "excluded-models" keys, and accepts both []string and []interface{}.
+func extractExcludedModelsFromMetadata(metadata map[string]any) []string {
+	if metadata == nil {
+		return nil
+	}
+	// Try both key formats
+	raw, ok := metadata["excluded_models"]
+	if !ok {
+		raw, ok = metadata["excluded-models"]
+	}
+	if !ok || raw == nil {
+		return nil
+	}
+	var stringSlice []string
+	switch v := raw.(type) {
+	case []string:
+		stringSlice = v
+	case []interface{}:
+		stringSlice = make([]string, 0, len(v))
+		for _, item := range v {
+			if s, ok := item.(string); ok {
+				stringSlice = append(stringSlice, s)
+			}
+		}
+	default:
+		return nil
+	}
+	result := make([]string, 0, len(stringSlice))
+	for _, s := range stringSlice {
+		if trimmed := strings.TrimSpace(s); trimmed != "" {
+			result = append(result, trimmed)
+		}
+	}
+	return result
+}
--- a/internal/watcher/synthesizer/file_test.go
+++ b/internal/watcher/synthesizer/file_test.go
@@ -297,6 +297,117 @@ func TestFileSynthesizer_Synthesize_PrefixValidation(t *testing.T) {
 	}
 }

+func TestFileSynthesizer_Synthesize_PriorityParsing(t *testing.T) {
+	tests := []struct {
+		name     string
+		priority any
+		want     string
+		hasValue bool
+	}{
+		{
+			name:     "string with spaces",
+			priority: " 10 ",
+			want:     "10",
+			hasValue: true,
+		},
+		{
+			name:     "number",
+			priority: 8,
+			want:     "8",
+			hasValue: true,
+		},
+		{
+			name:     "invalid string",
+			priority: "1x",
+			hasValue: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			tempDir := t.TempDir()
+			authData := map[string]any{
+				"type":     "claude",
+				"priority": tt.priority,
+			}
+			data, _ := json.Marshal(authData)
+			errWriteFile := os.WriteFile(filepath.Join(tempDir, "auth.json"), data, 0644)
+			if errWriteFile != nil {
+				t.Fatalf("failed to write auth file: %v", errWriteFile)
+			}
+
+			synth := NewFileSynthesizer()
+			ctx := &SynthesisContext{
+				Config:      &config.Config{},
+				AuthDir:     tempDir,
+				Now:         time.Now(),
+				IDGenerator: NewStableIDGenerator(),
+			}
+
+			auths, errSynthesize := synth.Synthesize(ctx)
+			if errSynthesize != nil {
+				t.Fatalf("unexpected error: %v", errSynthesize)
+			}
+			if len(auths) != 1 {
+				t.Fatalf("expected 1 auth, got %d", len(auths))
+			}
+
+			value, ok := auths[0].Attributes["priority"]
+			if tt.hasValue {
+				if !ok {
+					t.Fatal("expected priority attribute to be set")
+				}
+				if value != tt.want {
+					t.Fatalf("expected priority %q, got %q", tt.want, value)
+				}
+				return
+			}
+			if ok {
+				t.Fatalf("expected priority attribute to be absent, got %q", value)
+			}
+		})
+	}
+}
+
+func TestFileSynthesizer_Synthesize_OAuthExcludedModelsMerged(t *testing.T) {
+	tempDir := t.TempDir()
+	authData := map[string]any{
+		"type":            "claude",
+		"excluded_models": []string{"custom-model", "MODEL-B"},
+	}
+	data, _ := json.Marshal(authData)
+	errWriteFile := os.WriteFile(filepath.Join(tempDir, "auth.json"), data, 0644)
+	if errWriteFile != nil {
+		t.Fatalf("failed to write auth file: %v", errWriteFile)
+	}
+
+	synth := NewFileSynthesizer()
+	ctx := &SynthesisContext{
+		Config: &config.Config{
+			OAuthExcludedModels: map[string][]string{
+				"claude": {"shared", "model-b"},
+			},
+		},
+		AuthDir:     tempDir,
+		Now:         time.Now(),
+		IDGenerator: NewStableIDGenerator(),
+	}
+
+	auths, errSynthesize := synth.Synthesize(ctx)
+	if errSynthesize != nil {
+		t.Fatalf("unexpected error: %v", errSynthesize)
+	}
+	if len(auths) != 1 {
+		t.Fatalf("expected 1 auth, got %d", len(auths))
+	}
+
+	got := auths[0].Attributes["excluded_models"]
+	want := "custom-model,model-b,shared"
+	if got != want {
+		t.Fatalf("expected excluded_models %q, got %q", want, got)
+	}
+}
+
 func TestSynthesizeGeminiVirtualAuths_NilInputs(t *testing.T) {
 	now := time.Now()

@@ -533,6 +644,7 @@ func TestFileSynthesizer_Synthesize_MultiProjectGemini(t *testing.T) {
 		"type":       "gemini",
 		"email":      "multi@example.com",
 		"project_id": "project-a, project-b, project-c",
+		"priority":   " 10 ",
 	}
 	data, _ := json.Marshal(authData)
 	err := os.WriteFile(filepath.Join(tempDir, "gemini-multi.json"), data, 0644)
@@ -565,6 +677,9 @@ func TestFileSynthesizer_Synthesize_MultiProjectGemini(t *testing.T) {
 	if primary.Status != coreauth.StatusDisabled {
 		t.Errorf("expected primary status disabled, got %s", primary.Status)
 	}
+	if gotPriority := primary.Attributes["priority"]; gotPriority != "10" {
+		t.Errorf("expected primary priority 10, got %q", gotPriority)
+	}

 	// Remaining auths should be virtuals
 	for i := 1; i < 4; i++ {
@@ -575,6 +690,9 @@ func TestFileSynthesizer_Synthesize_MultiProjectGemini(t *testing.T) {
 		if v.Attributes["gemini_virtual_parent"] != primary.ID {
 			t.Errorf("expected virtual %d parent to be %s, got %s", i, primary.ID, v.Attributes["gemini_virtual_parent"])
 		}
+		if gotPriority := v.Attributes["priority"]; gotPriority != "10" {
+			t.Errorf("expected virtual %d priority 10, got %q", i, gotPriority)
+		}
 	}
 }

--- a/internal/watcher/synthesizer/helpers.go
+++ b/internal/watcher/synthesizer/helpers.go
@@ -53,6 +53,8 @@ func (g *StableIDGenerator) Next(kind string, parts ...string) (string, string)

 // ApplyAuthExcludedModelsMeta applies excluded models metadata to an auth entry.
 // It computes a hash of excluded models and sets the auth_kind attribute.
+// For OAuth entries, perKey (from the JSON file's excluded-models field) is merged
+// with the global oauth-excluded-models config for the provider.
 func ApplyAuthExcludedModelsMeta(auth *coreauth.Auth, cfg *config.Config, perKey []string, authKind string) {
 	if auth == nil || cfg == nil {
 		return
@@ -72,9 +74,13 @@ func ApplyAuthExcludedModelsMeta(auth *coreauth.Auth, cfg *config.Config, perKey
 	}
 	if authKindKey == "apikey" {
 		add(perKey)
-	} else if cfg.OAuthExcludedModels != nil {
-		providerKey := strings.ToLower(strings.TrimSpace(auth.Provider))
-		add(cfg.OAuthExcludedModels[providerKey])
+	} else {
+		// For OAuth: merge per-account excluded models with global provider-level exclusions
+		add(perKey)
+		if cfg.OAuthExcludedModels != nil {
+			providerKey := strings.ToLower(strings.TrimSpace(auth.Provider))
+			add(cfg.OAuthExcludedModels[providerKey])
+		}
 	}
 	combined := make([]string, 0, len(seen))
 	for k := range seen {
@@ -88,6 +94,10 @@ func ApplyAuthExcludedModelsMeta(auth *coreauth.Auth, cfg *config.Config, perKey
 	if hash != "" {
 		auth.Attributes["excluded_models_hash"] = hash
 	}
+	// Store the combined excluded models list so that routing can read it at runtime
+	if len(combined) > 0 {
+		auth.Attributes["excluded_models"] = strings.Join(combined, ",")
+	}
 	if authKind != "" {
 		auth.Attributes["auth_kind"] = authKind
 	}
--- a/internal/watcher/synthesizer/helpers_test.go
+++ b/internal/watcher/synthesizer/helpers_test.go
@@ -6,6 +6,7 @@ import (
 	"testing"

 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/watcher/diff"
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 )

@@ -200,6 +201,30 @@ func TestApplyAuthExcludedModelsMeta(t *testing.T) {
 	}
 }

+func TestApplyAuthExcludedModelsMeta_OAuthMergeWritesCombinedModels(t *testing.T) {
+	auth := &coreauth.Auth{
+		Provider:   "claude",
+		Attributes: make(map[string]string),
+	}
+	cfg := &config.Config{
+		OAuthExcludedModels: map[string][]string{
+			"claude": {"global-a", "shared"},
+		},
+	}
+
+	ApplyAuthExcludedModelsMeta(auth, cfg, []string{"per", "SHARED"}, "oauth")
+
+	const wantCombined = "global-a,per,shared"
+	if gotCombined := auth.Attributes["excluded_models"]; gotCombined != wantCombined {
+		t.Fatalf("expected excluded_models=%q, got %q", wantCombined, gotCombined)
+	}
+
+	expectedHash := diff.ComputeExcludedModelsHash([]string{"global-a", "per", "shared"})
+	if gotHash := auth.Attributes["excluded_models_hash"]; gotHash != expectedHash {
+		t.Fatalf("expected excluded_models_hash=%q, got %q", expectedHash, gotHash)
+	}
+}
+
 func TestAddConfigHeadersToAttrs(t *testing.T) {
 	tests := []struct {
 		name    string
--- a/sdk/cliproxy/auth/oauth_model_alias_test.go
+++ b/sdk/cliproxy/auth/oauth_model_alias_test.go
@@ -79,6 +79,24 @@ func TestResolveOAuthUpstreamModel_SuffixPreservation(t *testing.T) {
 			input:   "gemini-2.5-pro(none)",
 			want:    "gemini-2.5-pro-exp-03-25(none)",
 		},
+		{
+			name: "github-copilot suffix preserved",
+			aliases: map[string][]internalconfig.OAuthModelAlias{
+				"github-copilot": {{Name: "claude-opus-4.6", Alias: "opus"}},
+			},
+			channel: "github-copilot",
+			input:   "opus(medium)",
+			want:    "claude-opus-4.6(medium)",
+		},
+		{
+			name: "github-copilot no suffix",
+			aliases: map[string][]internalconfig.OAuthModelAlias{
+				"github-copilot": {{Name: "claude-opus-4.6", Alias: "opus"}},
+			},
+			channel: "github-copilot",
+			input:   "opus",
+			want:    "claude-opus-4.6",
+		},
 		{
 			name: "kimi suffix preserved",
 			aliases: map[string][]internalconfig.OAuthModelAlias{
@@ -174,6 +192,8 @@ func createAuthForChannel(channel string) *Auth {
 		return &Auth{Provider: "kimi"}
 	case "kiro":
 		return &Auth{Provider: "kiro"}
+	case "github-copilot":
+		return &Auth{Provider: "github-copilot"}
 	default:
 		return &Auth{Provider: channel}
 	}
@@ -187,6 +207,22 @@ func TestOAuthModelAliasChannel_Kimi(t *testing.T) {
 	}
 }

+func TestOAuthModelAliasChannel_GitHubCopilot(t *testing.T) {
+	t.Parallel()
+
+	if got := OAuthModelAliasChannel("github-copilot", ""); got != "github-copilot" {
+		t.Fatalf("OAuthModelAliasChannel() = %q, want %q", got, "github-copilot")
+	}
+}
+
+func TestOAuthModelAliasChannel_Kiro(t *testing.T) {
+	t.Parallel()
+
+	if got := OAuthModelAliasChannel("kiro", ""); got != "kiro" {
+		t.Fatalf("OAuthModelAliasChannel() = %q, want %q", got, "kiro")
+	}
+}
+
 func TestApplyOAuthModelAlias_SuffixPreservation(t *testing.T) {
 	t.Parallel()

--- a/sdk/cliproxy/service.go
+++ b/sdk/cliproxy/service.go
@@ -767,6 +767,13 @@ func (s *Service) registerModelsForAuth(a *coreauth.Auth) {
 		provider = "openai-compatibility"
 	}
 	excluded := s.oauthExcludedModels(provider, authKind)
+	// The synthesizer pre-merges per-account and global exclusions into the "excluded_models" attribute.
+	// If this attribute is present, it represents the complete list of exclusions and overrides the global config.
+	if a.Attributes != nil {
+		if val, ok := a.Attributes["excluded_models"]; ok && strings.TrimSpace(val) != "" {
+			excluded = strings.Split(val, ",")
+		}
+	}
 	var models []*ModelInfo
 	switch provider {
 	case "gemini":
--- a/sdk/cliproxy/service_excluded_models_test.go
+++ b/sdk/cliproxy/service_excluded_models_test.go
@@ -0,0 +1,65 @@
+package cliproxy
+
+import (
+	"strings"
+	"testing"
+
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
+)
+
+func TestRegisterModelsForAuth_UsesPreMergedExcludedModelsAttribute(t *testing.T) {
+	service := &Service{
+		cfg: &config.Config{
+			OAuthExcludedModels: map[string][]string{
+				"gemini-cli": {"gemini-2.5-pro"},
+			},
+		},
+	}
+	auth := &coreauth.Auth{
+		ID:       "auth-gemini-cli",
+		Provider: "gemini-cli",
+		Status:   coreauth.StatusActive,
+		Attributes: map[string]string{
+			"auth_kind":       "oauth",
+			"excluded_models": "gemini-2.5-flash",
+		},
+	}
+
+	registry := GlobalModelRegistry()
+	registry.UnregisterClient(auth.ID)
+	t.Cleanup(func() {
+		registry.UnregisterClient(auth.ID)
+	})
+
+	service.registerModelsForAuth(auth)
+
+	models := registry.GetAvailableModelsByProvider("gemini-cli")
+	if len(models) == 0 {
+		t.Fatal("expected gemini-cli models to be registered")
+	}
+
+	for _, model := range models {
+		if model == nil {
+			continue
+		}
+		modelID := strings.TrimSpace(model.ID)
+		if strings.EqualFold(modelID, "gemini-2.5-flash") {
+			t.Fatalf("expected model %q to be excluded by auth attribute", modelID)
+		}
+	}
+
+	seenGlobalExcluded := false
+	for _, model := range models {
+		if model == nil {
+			continue
+		}
+		if strings.EqualFold(strings.TrimSpace(model.ID), "gemini-2.5-pro") {
+			seenGlobalExcluded = true
+			break
+		}
+	}
+	if !seenGlobalExcluded {
+		t.Fatal("expected global excluded model to be present when attribute override is set")
+	}
+}