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

fix(kiro): Support token extraction from Metadata for file-based authentication

internal/runtime/executor/caching_verify_test.go

@@ -0,0 +1,210 @@
+package executor
+
+import (
+	"fmt"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestEnsureCacheControl(t *testing.T) {
+	// Test case 1: System prompt as string
+	t.Run("String System Prompt", func(t *testing.T) {
+		input := []byte(`{"model": "claude-3-5-sonnet", "system": "This is a long system prompt", "messages": []}`)
+		output := ensureCacheControl(input)
+
+		res := gjson.GetBytes(output, "system.0.cache_control.type")
+		if res.String() != "ephemeral" {
+			t.Errorf("cache_control not found in system string. Output: %s", string(output))
+		}
+	})
+
+	// Test case 2: System prompt as array
+	t.Run("Array System Prompt", func(t *testing.T) {
+		input := []byte(`{"model": "claude-3-5-sonnet", "system": [{"type": "text", "text": "Part 1"}, {"type": "text", "text": "Part 2"}], "messages": []}`)
+		output := ensureCacheControl(input)
+
+		// cache_control should only be on the LAST element
+		res0 := gjson.GetBytes(output, "system.0.cache_control")
+		res1 := gjson.GetBytes(output, "system.1.cache_control.type")
+
+		if res0.Exists() {
+			t.Errorf("cache_control should NOT be on the first element")
+		}
+		if res1.String() != "ephemeral" {
+			t.Errorf("cache_control not found on last system element. Output: %s", string(output))
+		}
+	})
+
+	// Test case 3: Tools are cached
+	t.Run("Tools Caching", func(t *testing.T) {
+		input := []byte(`{
+			"model": "claude-3-5-sonnet",
+			"tools": [
+				{"name": "tool1", "description": "First tool", "input_schema": {"type": "object"}},
+				{"name": "tool2", "description": "Second tool", "input_schema": {"type": "object"}}
+			],
+			"system": "System prompt",
+			"messages": []
+		}`)
+		output := ensureCacheControl(input)
+
+		// cache_control should only be on the LAST tool
+		tool0Cache := gjson.GetBytes(output, "tools.0.cache_control")
+		tool1Cache := gjson.GetBytes(output, "tools.1.cache_control.type")
+
+		if tool0Cache.Exists() {
+			t.Errorf("cache_control should NOT be on the first tool")
+		}
+		if tool1Cache.String() != "ephemeral" {
+			t.Errorf("cache_control not found on last tool. Output: %s", string(output))
+		}
+
+		// System should also have cache_control
+		systemCache := gjson.GetBytes(output, "system.0.cache_control.type")
+		if systemCache.String() != "ephemeral" {
+			t.Errorf("cache_control not found in system. Output: %s", string(output))
+		}
+	})
+
+	// Test case 4: Tools and system are INDEPENDENT breakpoints
+	// Per Anthropic docs: Up to 4 breakpoints allowed, tools and system are cached separately
+	t.Run("Independent Cache Breakpoints", func(t *testing.T) {
+		input := []byte(`{
+			"model": "claude-3-5-sonnet",
+			"tools": [
+				{"name": "tool1", "description": "First tool", "input_schema": {"type": "object"}, "cache_control": {"type": "ephemeral"}}
+			],
+			"system": [{"type": "text", "text": "System"}],
+			"messages": []
+		}`)
+		output := ensureCacheControl(input)
+
+		// Tool already has cache_control - should not be changed
+		tool0Cache := gjson.GetBytes(output, "tools.0.cache_control.type")
+		if tool0Cache.String() != "ephemeral" {
+			t.Errorf("existing cache_control was incorrectly removed")
+		}
+
+		// System SHOULD get cache_control because it is an INDEPENDENT breakpoint
+		// Tools and system are separate cache levels in the hierarchy
+		systemCache := gjson.GetBytes(output, "system.0.cache_control.type")
+		if systemCache.String() != "ephemeral" {
+			t.Errorf("system should have its own cache_control breakpoint (independent of tools)")
+		}
+	})
+
+	// Test case 5: Only tools, no system
+	t.Run("Only Tools No System", func(t *testing.T) {
+		input := []byte(`{
+			"model": "claude-3-5-sonnet",
+			"tools": [
+				{"name": "tool1", "description": "Tool", "input_schema": {"type": "object"}}
+			],
+			"messages": [{"role": "user", "content": "Hi"}]
+		}`)
+		output := ensureCacheControl(input)
+
+		toolCache := gjson.GetBytes(output, "tools.0.cache_control.type")
+		if toolCache.String() != "ephemeral" {
+			t.Errorf("cache_control not found on tool. Output: %s", string(output))
+		}
+	})
+
+	// Test case 6: Many tools (Claude Code scenario)
+	t.Run("Many Tools (Claude Code Scenario)", func(t *testing.T) {
+		// Simulate Claude Code with many tools
+		toolsJSON := `[`
+		for i := 0; i < 50; i++ {
+			if i > 0 {
+				toolsJSON += ","
+			}
+			toolsJSON += fmt.Sprintf(`{"name": "tool%d", "description": "Tool %d", "input_schema": {"type": "object"}}`, i, i)
+		}
+		toolsJSON += `]`
+
+		input := []byte(fmt.Sprintf(`{
+			"model": "claude-3-5-sonnet",
+			"tools": %s,
+			"system": [{"type": "text", "text": "You are Claude Code"}],
+			"messages": [{"role": "user", "content": "Hello"}]
+		}`, toolsJSON))
+
+		output := ensureCacheControl(input)
+
+		// Only the last tool (index 49) should have cache_control
+		for i := 0; i < 49; i++ {
+			path := fmt.Sprintf("tools.%d.cache_control", i)
+			if gjson.GetBytes(output, path).Exists() {
+				t.Errorf("tool %d should NOT have cache_control", i)
+			}
+		}
+
+		lastToolCache := gjson.GetBytes(output, "tools.49.cache_control.type")
+		if lastToolCache.String() != "ephemeral" {
+			t.Errorf("last tool (49) should have cache_control")
+		}
+
+		// System should also have cache_control
+		systemCache := gjson.GetBytes(output, "system.0.cache_control.type")
+		if systemCache.String() != "ephemeral" {
+			t.Errorf("system should have cache_control")
+		}
+
+		t.Log("test passed: 50 tools - cache_control only on last tool")
+	})
+
+	// Test case 7: Empty tools array
+	t.Run("Empty Tools Array", func(t *testing.T) {
+		input := []byte(`{"model": "claude-3-5-sonnet", "tools": [], "system": "Test", "messages": []}`)
+		output := ensureCacheControl(input)
+
+		// System should still get cache_control
+		systemCache := gjson.GetBytes(output, "system.0.cache_control.type")
+		if systemCache.String() != "ephemeral" {
+			t.Errorf("system should have cache_control even with empty tools array")
+		}
+	})
+}
+
+// TestCacheControlOrder verifies the correct order: tools -> system -> messages
+func TestCacheControlOrder(t *testing.T) {
+	input := []byte(`{
+		"model": "claude-sonnet-4",
+		"tools": [
+			{"name": "Read", "description": "Read file", "input_schema": {"type": "object", "properties": {"path": {"type": "string"}}}},
+			{"name": "Write", "description": "Write file", "input_schema": {"type": "object", "properties": {"path": {"type": "string"}, "content": {"type": "string"}}}}
+		],
+		"system": [
+			{"type": "text", "text": "You are Claude Code, Anthropic's official CLI for Claude."},
+			{"type": "text", "text": "Additional instructions here..."}
+		],
+		"messages": [
+			{"role": "user", "content": "Hello"}
+		]
+	}`)
+
+	output := ensureCacheControl(input)
+
+	// 1. Last tool has cache_control
+	if gjson.GetBytes(output, "tools.1.cache_control.type").String() != "ephemeral" {
+		t.Error("last tool should have cache_control")
+	}
+
+	// 2. First tool has NO cache_control
+	if gjson.GetBytes(output, "tools.0.cache_control").Exists() {
+		t.Error("first tool should NOT have cache_control")
+	}
+
+	// 3. Last system element has cache_control
+	if gjson.GetBytes(output, "system.1.cache_control.type").String() != "ephemeral" {
+		t.Error("last system element should have cache_control")
+	}
+
+	// 4. First system element has NO cache_control
+	if gjson.GetBytes(output, "system.0.cache_control").Exists() {
+		t.Error("first system element should NOT have cache_control")
+	}
+
+	t.Log("cache order correct: tools -> system")
+}

internal/runtime/executor/claude_executor.go

@@ -120,6 +120,9 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 
+	// Auto-inject cache_control if missing (optimization for ClawdBot/clients without caching support)
+	body = ensureCacheControl(body)
+
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
@@ -252,6 +255,9 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 
+	// Auto-inject cache_control if missing (optimization for ClawdBot/clients without caching support)
+	body = ensureCacheControl(body)
+
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
@@ -636,7 +642,7 @@ func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string,
 		ginHeaders = ginCtx.Request.Header
 	}
 
-	baseBetas := "claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14"
+	baseBetas := "claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14,prompt-caching-2024-07-31"
 	if val := strings.TrimSpace(ginHeaders.Get("Anthropic-Beta")); val != "" {
 		baseBetas = val
 		if !strings.Contains(val, "oauth") {
@@ -990,3 +996,214 @@ func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.A
 
 	return payload
 }
+
+// ensureCacheControl injects cache_control breakpoints into the payload for optimal prompt caching.
+// According to Anthropic's documentation, cache prefixes are created in order: tools -> system -> messages.
+// This function adds cache_control to:
+// 1. The LAST tool in the tools array (caches all tool definitions)
+// 2. The LAST element in the system array (caches system prompt)
+// 3. The SECOND-TO-LAST user turn (caches conversation history for multi-turn)
+//
+// Up to 4 cache breakpoints are allowed per request. Tools, System, and Messages are INDEPENDENT breakpoints.
+// This enables up to 90% cost reduction on cached tokens (cache read = 0.1x base price).
+// See: https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
+func ensureCacheControl(payload []byte) []byte {
+	// 1. Inject cache_control into the LAST tool (caches all tool definitions)
+	// Tools are cached first in the hierarchy, so this is the most important breakpoint.
+	payload = injectToolsCacheControl(payload)
+
+	// 2. Inject cache_control into the LAST system prompt element
+	// System is the second level in the cache hierarchy.
+	payload = injectSystemCacheControl(payload)
+
+	// 3. Inject cache_control into messages for multi-turn conversation caching
+	// This caches the conversation history up to the second-to-last user turn.
+	payload = injectMessagesCacheControl(payload)
+
+	return payload
+}
+
+// injectMessagesCacheControl adds cache_control to the second-to-last user turn for multi-turn caching.
+// Per Anthropic docs: "Place cache_control on the second-to-last User message to let the model reuse the earlier cache."
+// This enables caching of conversation history, which is especially beneficial for long multi-turn conversations.
+// Only adds cache_control if:
+// - There are at least 2 user turns in the conversation
+// - No message content already has cache_control
+func injectMessagesCacheControl(payload []byte) []byte {
+	messages := gjson.GetBytes(payload, "messages")
+	if !messages.Exists() || !messages.IsArray() {
+		return payload
+	}
+
+	// Check if ANY message content already has cache_control
+	hasCacheControlInMessages := false
+	messages.ForEach(func(_, msg gjson.Result) bool {
+		content := msg.Get("content")
+		if content.IsArray() {
+			content.ForEach(func(_, item gjson.Result) bool {
+				if item.Get("cache_control").Exists() {
+					hasCacheControlInMessages = true
+					return false
+				}
+				return true
+			})
+		}
+		return !hasCacheControlInMessages
+	})
+	if hasCacheControlInMessages {
+		return payload
+	}
+
+	// Find all user message indices
+	var userMsgIndices []int
+	messages.ForEach(func(index gjson.Result, msg gjson.Result) bool {
+		if msg.Get("role").String() == "user" {
+			userMsgIndices = append(userMsgIndices, int(index.Int()))
+		}
+		return true
+	})
+
+	// Need at least 2 user turns to cache the second-to-last
+	if len(userMsgIndices) < 2 {
+		return payload
+	}
+
+	// Get the second-to-last user message index
+	secondToLastUserIdx := userMsgIndices[len(userMsgIndices)-2]
+
+	// Get the content of this message
+	contentPath := fmt.Sprintf("messages.%d.content", secondToLastUserIdx)
+	content := gjson.GetBytes(payload, contentPath)
+
+	if content.IsArray() {
+		// Add cache_control to the last content block of this message
+		contentCount := int(content.Get("#").Int())
+		if contentCount > 0 {
+			cacheControlPath := fmt.Sprintf("messages.%d.content.%d.cache_control", secondToLastUserIdx, contentCount-1)
+			result, err := sjson.SetBytes(payload, cacheControlPath, map[string]string{"type": "ephemeral"})
+			if err != nil {
+				log.Warnf("failed to inject cache_control into messages: %v", err)
+				return payload
+			}
+			payload = result
+		}
+	} else if content.Type == gjson.String {
+		// Convert string content to array with cache_control
+		text := content.String()
+		newContent := []map[string]interface{}{
+			{
+				"type": "text",
+				"text": text,
+				"cache_control": map[string]string{
+					"type": "ephemeral",
+				},
+			},
+		}
+		result, err := sjson.SetBytes(payload, contentPath, newContent)
+		if err != nil {
+			log.Warnf("failed to inject cache_control into message string content: %v", err)
+			return payload
+		}
+		payload = result
+	}
+
+	return payload
+}
+
+// injectToolsCacheControl adds cache_control to the last tool in the tools array.
+// Per Anthropic docs: "The cache_control parameter on the last tool definition caches all tool definitions."
+// This only adds cache_control if NO tool in the array already has it.
+func injectToolsCacheControl(payload []byte) []byte {
+	tools := gjson.GetBytes(payload, "tools")
+	if !tools.Exists() || !tools.IsArray() {
+		return payload
+	}
+
+	toolCount := int(tools.Get("#").Int())
+	if toolCount == 0 {
+		return payload
+	}
+
+	// Check if ANY tool already has cache_control - if so, don't modify tools
+	hasCacheControlInTools := false
+	tools.ForEach(func(_, tool gjson.Result) bool {
+		if tool.Get("cache_control").Exists() {
+			hasCacheControlInTools = true
+			return false
+		}
+		return true
+	})
+	if hasCacheControlInTools {
+		return payload
+	}
+
+	// Add cache_control to the last tool
+	lastToolPath := fmt.Sprintf("tools.%d.cache_control", toolCount-1)
+	result, err := sjson.SetBytes(payload, lastToolPath, map[string]string{"type": "ephemeral"})
+	if err != nil {
+		log.Warnf("failed to inject cache_control into tools array: %v", err)
+		return payload
+	}
+
+	return result
+}
+
+// injectSystemCacheControl adds cache_control to the last element in the system prompt.
+// Converts string system prompts to array format if needed.
+// This only adds cache_control if NO system element already has it.
+func injectSystemCacheControl(payload []byte) []byte {
+	system := gjson.GetBytes(payload, "system")
+	if !system.Exists() {
+		return payload
+	}
+
+	if system.IsArray() {
+		count := int(system.Get("#").Int())
+		if count == 0 {
+			return payload
+		}
+
+		// Check if ANY system element already has cache_control
+		hasCacheControlInSystem := false
+		system.ForEach(func(_, item gjson.Result) bool {
+			if item.Get("cache_control").Exists() {
+				hasCacheControlInSystem = true
+				return false
+			}
+			return true
+		})
+		if hasCacheControlInSystem {
+			return payload
+		}
+
+		// Add cache_control to the last system element
+		lastSystemPath := fmt.Sprintf("system.%d.cache_control", count-1)
+		result, err := sjson.SetBytes(payload, lastSystemPath, map[string]string{"type": "ephemeral"})
+		if err != nil {
+			log.Warnf("failed to inject cache_control into system array: %v", err)
+			return payload
+		}
+		payload = result
+	} else if system.Type == gjson.String {
+		// Convert string system prompt to array with cache_control
+		// "system": "text" -> "system": [{"type": "text", "text": "text", "cache_control": {"type": "ephemeral"}}]
+		text := system.String()
+		newSystem := []map[string]interface{}{
+			{
+				"type": "text",
+				"text": text,
+				"cache_control": map[string]string{
+					"type": "ephemeral",
+				},
+			},
+		}
+		result, err := sjson.SetBytes(payload, "system", newSystem)
+		if err != nil {
+			log.Warnf("failed to inject cache_control into system string: %v", err)
+			return payload
+		}
+		payload = result
+	}
+
+	return payload
+}

internal/runtime/executor/kiro_executor.go

@@ -35,7 +35,7 @@ import (
 
 const (
 	// Kiro API common constants
-	kiroContentType  = "application/x-amz-json-1.0"
+	kiroContentType  = "application/json"
 	kiroAcceptStream = "*/*"
 
 	// Event Stream frame size constants for boundary protection
@@ -47,17 +47,18 @@ const (
 	// Event Stream error type constants
 	ErrStreamFatal     = "fatal"     // Connection/authentication errors, not recoverable
 	ErrStreamMalformed = "malformed" // Format errors, data cannot be parsed
-	// kiroUserAgent matches amq2api format for User-Agent header (Amazon Q CLI style)
+
+	// kiroUserAgent matches Amazon Q CLI style for User-Agent header
 	kiroUserAgent = "aws-sdk-rust/1.3.9 os/macos lang/rust/1.87.0"
-	// kiroFullUserAgent is the complete x-amz-user-agent header matching amq2api (Amazon Q CLI style)
+	// kiroFullUserAgent is the complete x-amz-user-agent header (Amazon Q CLI style)
 	kiroFullUserAgent = "aws-sdk-rust/1.3.9 ua/2.1 api/ssooidc/1.88.0 os/macos lang/rust/1.87.0 m/E app/AmazonQ-For-CLI"
 
-	// Kiro IDE style headers (from kiro2api - for IDC auth)
-	kiroIDEUserAgent     = "aws-sdk-js/1.0.18 ua/2.1 os/darwin#25.0.0 lang/js md/nodejs#20.16.0 api/codewhispererstreaming#1.0.18 m/E KiroIDE-0.2.13-66c23a8c5d15afabec89ef9954ef52a119f10d369df04d548fc6c1eac694b0d1"
-	kiroIDEAmzUserAgent  = "aws-sdk-js/1.0.18 KiroIDE-0.2.13-66c23a8c5d15afabec89ef9954ef52a119f10d369df04d548fc6c1eac694b0d1"
-	kiroIDEAgentModeSpec = "spec"
+	// Kiro IDE style headers for IDC auth
+	kiroIDEUserAgent     = "aws-sdk-js/1.0.27 ua/2.1 os/win32#10.0.19044 lang/js md/nodejs#22.21.1 api/codewhispererstreaming#1.0.27 m/E"
+	kiroIDEAmzUserAgent  = "aws-sdk-js/1.0.27"
+	kiroIDEAgentModeVibe = "vibe"
 
-	// Socket retry configuration constants (based on kiro2Api reference implementation)
+	// Socket retry configuration constants
 	// Maximum number of retry attempts for socket/network errors
 	kiroSocketMaxRetries = 3
 	// Base delay between retry attempts (uses exponential backoff: delay * 2^attempt)
@@ -355,34 +356,32 @@ func extractRegionFromProfileARN(profileArn string) string {
 // buildKiroEndpointConfigs creates endpoint configurations for the specified region.
 // This enables dynamic region support for Enterprise/IdC users in non-us-east-1 regions.
 //
-// CRITICAL: Each endpoint MUST use its compatible Origin and AmzTarget values:
-// - CodeWhisperer endpoint (codewhisperer.{region}.amazonaws.com): Uses AI_EDITOR origin and AmazonCodeWhispererStreamingService target
-// - Amazon Q endpoint (q.{region}.amazonaws.com): Uses CLI origin and AmazonQDeveloperStreamingService target
+// Uses Q endpoint (q.{region}.amazonaws.com) as primary for ALL auth types:
+// - Works universally across all AWS regions (CodeWhisperer endpoint only exists in us-east-1)
+// - Uses /generateAssistantResponse path with AI_EDITOR origin
+// - Does NOT require X-Amz-Target header
 //
-// Mismatched combinations will result in 403 Forbidden errors.
-//
-// NOTE: CodeWhisperer is set as the default endpoint because:
-// 1. Most tokens come from Kiro IDE / VSCode extensions (AWS Builder ID auth)
-// 2. These tokens use AI_EDITOR origin which is only compatible with CodeWhisperer endpoint
-// 3. Amazon Q endpoint requires CLI origin which is for Amazon Q CLI tokens
-// This matches the AIClient-2-API-main project's configuration.
+// The AmzTarget field is kept for backward compatibility but should be empty
+// to indicate that the header should NOT be set.
 func buildKiroEndpointConfigs(region string) []kiroEndpointConfig {
 	if region == "" {
 		region = kiroDefaultRegion
 	}
 	return []kiroEndpointConfig{
 		{
+			// Primary: Q endpoint - works for all regions and auth types
+			URL:       fmt.Sprintf("https://q.%s.amazonaws.com/generateAssistantResponse", region),
+			Origin:    "AI_EDITOR",
+			AmzTarget: "", // Empty = don't set X-Amz-Target header
+			Name:      "AmazonQ",
+		},
+		{
+			// Fallback: CodeWhisperer endpoint (legacy, only works in us-east-1)
 			URL:       fmt.Sprintf("https://codewhisperer.%s.amazonaws.com/generateAssistantResponse", region),
 			Origin:    "AI_EDITOR",
 			AmzTarget: "AmazonCodeWhispererStreamingService.GenerateAssistantResponse",
 			Name:      "CodeWhisperer",
 		},
-		{
-			URL:       fmt.Sprintf("https://q.%s.amazonaws.com/", region),
-			Origin:    "CLI",
-			AmzTarget: "AmazonQDeveloperStreamingService.SendMessage",
-			Name:      "AmazonQ",
-		},
 	}
 }
 
@@ -393,13 +392,12 @@ var kiroEndpointConfigs = buildKiroEndpointConfigs(kiroDefaultRegion)
 // getKiroEndpointConfigs returns the list of Kiro API endpoint configurations to try in order.
 // Supports dynamic region based on auth metadata "api_region", "profile_arn", or "region" field.
 // Supports reordering based on "preferred_endpoint" in auth metadata/attributes.
-// For IDC auth method, automatically uses CodeWhisperer endpoint with CLI origin.
 //
 // Region priority:
 // 1. auth.Metadata["api_region"] - explicit API region override
 // 2. ProfileARN region - extracted from arn:aws:service:REGION:account:resource
-// 3. auth.Metadata["region"] - OIDC/Identity region (may differ from API region)
-// 4. kiroDefaultRegion (us-east-1) - fallback
+// 3. kiroDefaultRegion (us-east-1) - fallback
+// Note: OIDC "region" is NOT used - it's for token refresh, not API calls
 func getKiroEndpointConfigs(auth *cliproxyauth.Auth) []kiroEndpointConfig {
 	if auth == nil {
 		return kiroEndpointConfigs
@@ -422,13 +420,9 @@ func getKiroEndpointConfigs(auth *cliproxyauth.Auth) []kiroEndpointConfig {
 					regionSource = "profile_arn"
 				}
 			}
-			// Priority 3: OIDC region (only if not already set from profile_arn)
-			if regionSource == "default" {
-				if r, ok := auth.Metadata["region"].(string); ok && r != "" {
-					region = r
-					regionSource = "region"
-				}
-			}
+			// Note: OIDC "region" field is NOT used for API endpoint
+			// Kiro API only exists in us-east-1, while OIDC region can vary (e.g., ap-northeast-2)
+			// Using OIDC region for API calls causes DNS failures
 		}
 	}
 
@@ -437,14 +431,14 @@ func getKiroEndpointConfigs(auth *cliproxyauth.Auth) []kiroEndpointConfig {
 	// Build endpoint configs for the specified region
 	endpointConfigs := buildKiroEndpointConfigs(region)
 
-	// For IDC auth, use CodeWhisperer endpoint with AI_EDITOR origin (same as Social auth)
-	// Based on kiro2api analysis: IDC tokens work with CodeWhisperer endpoint using Bearer auth
+	// For IDC auth, use Q endpoint with AI_EDITOR origin
+	// IDC tokens work with Q endpoint using Bearer auth
 	// The difference is only in how tokens are refreshed (OIDC with clientId/clientSecret for IDC)
 	// NOT in how API calls are made - both Social and IDC use the same endpoint/origin
 	if auth.Metadata != nil {
 		authMethod, _ := auth.Metadata["auth_method"].(string)
 		if strings.ToLower(authMethod) == "idc" {
-			log.Debugf("kiro: IDC auth, using CodeWhisperer endpoint (region: %s)", region)
+			log.Debugf("kiro: IDC auth, using Q endpoint (region: %s)", region)
 			return endpointConfigs
 		}
 	}
@@ -552,7 +546,7 @@ func applyDynamicFingerprint(req *http.Request, auth *cliproxyauth.Auth) {
 		// Use fingerprint-generated dynamic User-Agent
 		req.Header.Set("User-Agent", fp.BuildUserAgent())
 		req.Header.Set("X-Amz-User-Agent", fp.BuildAmzUserAgent())
-		req.Header.Set("x-amzn-kiro-agent-mode", kiroIDEAgentModeSpec)
+		req.Header.Set("x-amzn-kiro-agent-mode", kiroIDEAgentModeVibe)
 
 		log.Debugf("kiro: using dynamic fingerprint for token %s (SDK:%s, OS:%s/%s, Kiro:%s)",
 			tokenKey[:8]+"...", fp.SDKVersion, fp.OSType, fp.OSVersion, fp.KiroVersion)
@@ -731,8 +725,13 @@ func (e *KiroExecutor) executeWithRetry(ctx context.Context, auth *cliproxyauth.
 
 			httpReq.Header.Set("Content-Type", kiroContentType)
 			httpReq.Header.Set("Accept", kiroAcceptStream)
-			// Use endpoint-specific X-Amz-Target (critical for avoiding 403 errors)
-			httpReq.Header.Set("X-Amz-Target", endpointConfig.AmzTarget)
+			// Only set X-Amz-Target if specified (Q endpoint doesn't require it)
+			if endpointConfig.AmzTarget != "" {
+				httpReq.Header.Set("X-Amz-Target", endpointConfig.AmzTarget)
+			}
+			// Kiro-specific headers
+			httpReq.Header.Set("x-amzn-kiro-agent-mode", kiroIDEAgentModeVibe)
+			httpReq.Header.Set("x-amzn-codewhisperer-optout", "true")
 
 			// Apply dynamic fingerprint-based headers
 			applyDynamicFingerprint(httpReq, auth)
@@ -1146,8 +1145,13 @@ func (e *KiroExecutor) executeStreamWithRetry(ctx context.Context, auth *cliprox
 
 			httpReq.Header.Set("Content-Type", kiroContentType)
 			httpReq.Header.Set("Accept", kiroAcceptStream)
-			// Use endpoint-specific X-Amz-Target (critical for avoiding 403 errors)
-			httpReq.Header.Set("X-Amz-Target", endpointConfig.AmzTarget)
+			// Only set X-Amz-Target if specified (Q endpoint doesn't require it)
+			if endpointConfig.AmzTarget != "" {
+				httpReq.Header.Set("X-Amz-Target", endpointConfig.AmzTarget)
+			}
+			// Kiro-specific headers
+			httpReq.Header.Set("x-amzn-kiro-agent-mode", kiroIDEAgentModeVibe)
+			httpReq.Header.Set("x-amzn-codewhisperer-optout", "true")
 
 			// Apply dynamic fingerprint-based headers
 			applyDynamicFingerprint(httpReq, auth)

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -155,10 +155,13 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 						clientContentJSON, _ = sjson.SetRaw(clientContentJSON, "parts.-1", partJSON)
 					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "text" {
 						prompt := contentResult.Get("text").String()
-						partJSON := `{}`
-						if prompt != "" {
-							partJSON, _ = sjson.Set(partJSON, "text", prompt)
+						// Skip empty text parts to avoid Gemini API error:
+						// "required oneof field 'data' must have one initialized field"
+						if prompt == "" {
+							continue
 						}
+						partJSON := `{}`
+						partJSON, _ = sjson.Set(partJSON, "text", prompt)
 						clientContentJSON, _ = sjson.SetRaw(clientContentJSON, "parts.-1", partJSON)
 					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "tool_use" {
 						// NOTE: Do NOT inject dummy thinking blocks here.
@@ -285,6 +288,13 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 					}
 				}
 
+				// Skip messages with empty parts array to avoid Gemini API error:
+				// "required oneof field 'data' must have one initialized field"
+				partsCheck := gjson.Get(clientContentJSON, "parts")
+				if !partsCheck.IsArray() || len(partsCheck.Array()) == 0 {
+					continue
+				}
+
 				contentsJSON, _ = sjson.SetRaw(contentsJSON, "-1", clientContentJSON)
 				hasContents = true
 			} else if contentsResult.Type == gjson.String {

internal/translator/antigravity/openai/chat-completions/antigravity_openai_request.go

@@ -305,12 +305,14 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 		}
 	}
 
-	// tools -> request.tools[].functionDeclarations + request.tools[].googleSearch passthrough
+	// tools -> request.tools[].functionDeclarations + request.tools[].googleSearch/codeExecution/urlContext passthrough
 	tools := gjson.GetBytes(rawJSON, "tools")
 	if tools.IsArray() && len(tools.Array()) > 0 {
 		functionToolNode := []byte(`{}`)
 		hasFunction := false
 		googleSearchNodes := make([][]byte, 0)
+		codeExecutionNodes := make([][]byte, 0)
+		urlContextNodes := make([][]byte, 0)
 		for _, t := range tools.Array() {
 			if t.Get("type").String() == "function" {
 				fn := t.Get("function")
@@ -370,8 +372,28 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 				}
 				googleSearchNodes = append(googleSearchNodes, googleToolNode)
 			}
+			if ce := t.Get("code_execution"); ce.Exists() {
+				codeToolNode := []byte(`{}`)
+				var errSet error
+				codeToolNode, errSet = sjson.SetRawBytes(codeToolNode, "codeExecution", []byte(ce.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set codeExecution tool: %v", errSet)
+					continue
+				}
+				codeExecutionNodes = append(codeExecutionNodes, codeToolNode)
+			}
+			if uc := t.Get("url_context"); uc.Exists() {
+				urlToolNode := []byte(`{}`)
+				var errSet error
+				urlToolNode, errSet = sjson.SetRawBytes(urlToolNode, "urlContext", []byte(uc.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set urlContext tool: %v", errSet)
+					continue
+				}
+				urlContextNodes = append(urlContextNodes, urlToolNode)
+			}
 		}
-		if hasFunction || len(googleSearchNodes) > 0 {
+		if hasFunction || len(googleSearchNodes) > 0 || len(codeExecutionNodes) > 0 || len(urlContextNodes) > 0 {
 			toolsNode := []byte("[]")
 			if hasFunction {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", functionToolNode)
@@ -379,6 +401,12 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 			for _, googleNode := range googleSearchNodes {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", googleNode)
 			}
+			for _, codeNode := range codeExecutionNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", codeNode)
+			}
+			for _, urlNode := range urlContextNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", urlNode)
+			}
 			out, _ = sjson.SetRawBytes(out, "request.tools", toolsNode)
 		}
 	}

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

@@ -283,12 +283,14 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 		}
 	}
 
-	// tools -> request.tools[].functionDeclarations + request.tools[].googleSearch passthrough
+	// tools -> request.tools[].functionDeclarations + request.tools[].googleSearch/codeExecution/urlContext passthrough
 	tools := gjson.GetBytes(rawJSON, "tools")
 	if tools.IsArray() && len(tools.Array()) > 0 {
 		functionToolNode := []byte(`{}`)
 		hasFunction := false
 		googleSearchNodes := make([][]byte, 0)
+		codeExecutionNodes := make([][]byte, 0)
+		urlContextNodes := make([][]byte, 0)
 		for _, t := range tools.Array() {
 			if t.Get("type").String() == "function" {
 				fn := t.Get("function")
@@ -348,8 +350,28 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 				}
 				googleSearchNodes = append(googleSearchNodes, googleToolNode)
 			}
+			if ce := t.Get("code_execution"); ce.Exists() {
+				codeToolNode := []byte(`{}`)
+				var errSet error
+				codeToolNode, errSet = sjson.SetRawBytes(codeToolNode, "codeExecution", []byte(ce.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set codeExecution tool: %v", errSet)
+					continue
+				}
+				codeExecutionNodes = append(codeExecutionNodes, codeToolNode)
+			}
+			if uc := t.Get("url_context"); uc.Exists() {
+				urlToolNode := []byte(`{}`)
+				var errSet error
+				urlToolNode, errSet = sjson.SetRawBytes(urlToolNode, "urlContext", []byte(uc.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set urlContext tool: %v", errSet)
+					continue
+				}
+				urlContextNodes = append(urlContextNodes, urlToolNode)
+			}
 		}
-		if hasFunction || len(googleSearchNodes) > 0 {
+		if hasFunction || len(googleSearchNodes) > 0 || len(codeExecutionNodes) > 0 || len(urlContextNodes) > 0 {
 			toolsNode := []byte("[]")
 			if hasFunction {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", functionToolNode)
@@ -357,6 +379,12 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 			for _, googleNode := range googleSearchNodes {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", googleNode)
 			}
+			for _, codeNode := range codeExecutionNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", codeNode)
+			}
+			for _, urlNode := range urlContextNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", urlNode)
+			}
 			out, _ = sjson.SetRawBytes(out, "request.tools", toolsNode)
 		}
 	}

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

@@ -289,12 +289,14 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 		}
 	}
 
-	// tools -> tools[].functionDeclarations + tools[].googleSearch passthrough
+	// tools -> tools[].functionDeclarations + tools[].googleSearch/codeExecution/urlContext passthrough
 	tools := gjson.GetBytes(rawJSON, "tools")
 	if tools.IsArray() && len(tools.Array()) > 0 {
 		functionToolNode := []byte(`{}`)
 		hasFunction := false
 		googleSearchNodes := make([][]byte, 0)
+		codeExecutionNodes := make([][]byte, 0)
+		urlContextNodes := make([][]byte, 0)
 		for _, t := range tools.Array() {
 			if t.Get("type").String() == "function" {
 				fn := t.Get("function")
@@ -354,8 +356,28 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 				}
 				googleSearchNodes = append(googleSearchNodes, googleToolNode)
 			}
+			if ce := t.Get("code_execution"); ce.Exists() {
+				codeToolNode := []byte(`{}`)
+				var errSet error
+				codeToolNode, errSet = sjson.SetRawBytes(codeToolNode, "codeExecution", []byte(ce.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set codeExecution tool: %v", errSet)
+					continue
+				}
+				codeExecutionNodes = append(codeExecutionNodes, codeToolNode)
+			}
+			if uc := t.Get("url_context"); uc.Exists() {
+				urlToolNode := []byte(`{}`)
+				var errSet error
+				urlToolNode, errSet = sjson.SetRawBytes(urlToolNode, "urlContext", []byte(uc.Raw))
+				if errSet != nil {
+					log.Warnf("Failed to set urlContext tool: %v", errSet)
+					continue
+				}
+				urlContextNodes = append(urlContextNodes, urlToolNode)
+			}
 		}
-		if hasFunction || len(googleSearchNodes) > 0 {
+		if hasFunction || len(googleSearchNodes) > 0 || len(codeExecutionNodes) > 0 || len(urlContextNodes) > 0 {
 			toolsNode := []byte("[]")
 			if hasFunction {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", functionToolNode)
@@ -363,6 +385,12 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 			for _, googleNode := range googleSearchNodes {
 				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", googleNode)
 			}
+			for _, codeNode := range codeExecutionNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", codeNode)
+			}
+			for _, urlNode := range urlContextNodes {
+				toolsNode, _ = sjson.SetRawBytes(toolsNode, "-1", urlNode)
+			}
 			out, _ = sjson.SetRawBytes(out, "tools", toolsNode)
 		}
 	}

internal/util/gemini_schema.go

@@ -4,6 +4,7 @@ package util
 import (
 	"fmt"
 	"sort"
+	"strconv"
 	"strings"
 
 	"github.com/tidwall/gjson"
@@ -431,9 +432,54 @@ func removeUnsupportedKeywords(jsonStr string) string {
 			jsonStr, _ = sjson.Delete(jsonStr, p)
 		}
 	}
+	// Remove x-* extension fields (e.g., x-google-enum-descriptions) that are not supported by Gemini API
+	jsonStr = removeExtensionFields(jsonStr)
 	return jsonStr
 }
 
+// removeExtensionFields removes all x-* extension fields from the JSON schema.
+// These are OpenAPI/JSON Schema extension fields that Google APIs don't recognize.
+func removeExtensionFields(jsonStr string) string {
+	var paths []string
+	walkForExtensions(gjson.Parse(jsonStr), "", &paths)
+	// walkForExtensions returns paths in a way that deeper paths are added before their ancestors
+	// when they are not deleted wholesale, but since we skip children of deleted x-* nodes,
+	// any collected path is safe to delete. We still use DeleteBytes for efficiency.
+
+	b := []byte(jsonStr)
+	for _, p := range paths {
+		b, _ = sjson.DeleteBytes(b, p)
+	}
+	return string(b)
+}
+
+func walkForExtensions(value gjson.Result, path string, paths *[]string) {
+	if value.IsArray() {
+		arr := value.Array()
+		for i := len(arr) - 1; i >= 0; i-- {
+			walkForExtensions(arr[i], joinPath(path, strconv.Itoa(i)), paths)
+		}
+		return
+	}
+
+	if value.IsObject() {
+		value.ForEach(func(key, val gjson.Result) bool {
+			keyStr := key.String()
+			safeKey := escapeGJSONPathKey(keyStr)
+			childPath := joinPath(path, safeKey)
+
+			// If it's an extension field, we delete it and don't need to look at its children.
+			if strings.HasPrefix(keyStr, "x-") && !isPropertyDefinition(path) {
+				*paths = append(*paths, childPath)
+				return true
+			}
+
+			walkForExtensions(val, childPath, paths)
+			return true
+		})
+	}
+}
+
 func cleanupRequiredFields(jsonStr string) string {
 	for _, p := range findPaths(jsonStr, "required") {
 		parentPath := trimSuffix(p, ".required")

internal/util/gemini_schema_test.go

@@ -869,3 +869,129 @@ func TestCleanJSONSchemaForAntigravity_BooleanEnumToString(t *testing.T) {
 		t.Errorf("Boolean enum values should be converted to string format, got: %s", result)
 	}
 }
+
+func TestRemoveExtensionFields(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    string
+		expected string
+	}{
+		{
+			name: "removes x- fields at root",
+			input: `{
+				"type": "object",
+				"x-custom-meta": "value",
+				"properties": {
+					"foo": { "type": "string" }
+				}
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"foo": { "type": "string" }
+				}
+			}`,
+		},
+		{
+			name: "removes x- fields in nested properties",
+			input: `{
+				"type": "object",
+				"properties": {
+					"foo": {
+						"type": "string",
+						"x-internal-id": 123
+					}
+				}
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"foo": {
+						"type": "string"
+					}
+				}
+			}`,
+		},
+		{
+			name: "does NOT remove properties named x-",
+			input: `{
+				"type": "object",
+				"properties": {
+					"x-data": { "type": "string" },
+					"normal": { "type": "number", "x-meta": "remove" }
+				},
+				"required": ["x-data"]
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"x-data": { "type": "string" },
+					"normal": { "type": "number" }
+				},
+				"required": ["x-data"]
+			}`,
+		},
+		{
+			name: "does NOT remove $schema and other meta fields (as requested)",
+			input: `{
+				"$schema": "http://json-schema.org/draft-07/schema#",
+				"$id": "test",
+				"type": "object",
+				"properties": {
+					"foo": { "type": "string" }
+				}
+			}`,
+			expected: `{
+				"$schema": "http://json-schema.org/draft-07/schema#",
+				"$id": "test",
+				"type": "object",
+				"properties": {
+					"foo": { "type": "string" }
+				}
+			}`,
+		},
+		{
+			name: "handles properties named $schema",
+			input: `{
+				"type": "object",
+				"properties": {
+					"$schema": { "type": "string" }
+				}
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"$schema": { "type": "string" }
+				}
+			}`,
+		},
+		{
+			name: "handles escaping in paths",
+			input: `{
+				"type": "object",
+				"properties": {
+					"foo.bar": {
+						"type": "string",
+						"x-meta": "remove"
+					}
+				},
+				"x-root.meta": "remove"
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"foo.bar": {
+						"type": "string"
+					}
+				}
+			}`,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			actual := removeExtensionFields(tt.input)
+			compareJSON(t, tt.expected, actual)
+		})
+	}
+}

sdk/cliproxy/service.go

@@ -1416,29 +1416,44 @@ func (s *Service) fetchKiroModels(a *coreauth.Auth) []*ModelInfo {
 }
 
 // extractKiroTokenData extracts KiroTokenData from auth attributes and metadata.
+// It supports both config-based tokens (stored in Attributes) and file-based tokens (stored in Metadata).
 func (s *Service) extractKiroTokenData(a *coreauth.Auth) *kiroauth.KiroTokenData {
-	if a == nil || a.Attributes == nil {
+	if a == nil {
 		return nil
 	}
 
-	accessToken := strings.TrimSpace(a.Attributes["access_token"])
+	var accessToken, profileArn, refreshToken string
+
+	// Priority 1: Try to get from Attributes (config.yaml source)
+	if a.Attributes != nil {
+		accessToken = strings.TrimSpace(a.Attributes["access_token"])
+		profileArn = strings.TrimSpace(a.Attributes["profile_arn"])
+		refreshToken = strings.TrimSpace(a.Attributes["refresh_token"])
+	}
+
+	// Priority 2: If not found in Attributes, try Metadata (JSON file source)
+	if accessToken == "" && a.Metadata != nil {
+		if at, ok := a.Metadata["access_token"].(string); ok {
+			accessToken = strings.TrimSpace(at)
+		}
+		if pa, ok := a.Metadata["profile_arn"].(string); ok {
+			profileArn = strings.TrimSpace(pa)
+		}
+		if rt, ok := a.Metadata["refresh_token"].(string); ok {
+			refreshToken = strings.TrimSpace(rt)
+		}
+	}
+
+	// access_token is required
 	if accessToken == "" {
 		return nil
 	}
 
-	tokenData := &kiroauth.KiroTokenData{
-		AccessToken: accessToken,
-		ProfileArn:  strings.TrimSpace(a.Attributes["profile_arn"]),
+	return &kiroauth.KiroTokenData{
+		AccessToken:  accessToken,
+		ProfileArn:   profileArn,
+		RefreshToken: refreshToken,
 	}
-
-	// Also try to get refresh token from metadata
-	if a.Metadata != nil {
-		if rt, ok := a.Metadata["refresh_token"].(string); ok {
-			tokenData.RefreshToken = rt
-		}
-	}
-
-	return tokenData
 }
 
 // convertKiroAPIModels converts Kiro API models to ModelInfo slice.