CLIProxyAPIPlus/internal/runtime/executor/kiro_executor.go

package executor

import (
	"bufio"
	"bytes"
	"context"
	"encoding/base64"
	"encoding/binary"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"net"
	"net/http"
	"os"
	"path/filepath"
	"strings"
	"sync"
	"syscall"
	"time"

	"github.com/google/uuid"
	kiroauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/kiro"
	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
	kiroclaude "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/claude"
	kirocommon "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/common"
	kiroopenai "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/openai"
	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
	"github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/usage"
	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
	log "github.com/sirupsen/logrus"
)

const (
	// Kiro API common constants
	kiroContentType  = "application/x-amz-json-1.0"
	kiroAcceptStream = "*/*"

	// Event Stream frame size constants for boundary protection
	// AWS Event Stream binary format: prelude (12 bytes) + headers + payload + message_crc (4 bytes)
	// Prelude consists of: total_length (4) + headers_length (4) + prelude_crc (4)
	minEventStreamFrameSize = 16       // Minimum: 4(total_len) + 4(headers_len) + 4(prelude_crc) + 4(message_crc)
	maxEventStreamMsgSize   = 10 << 20 // Maximum message length: 10MB

	// 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 = "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 = "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"

	// Socket retry configuration constants (based on kiro2Api reference implementation)
	// Maximum number of retry attempts for socket/network errors
	kiroSocketMaxRetries = 3
	// Base delay between retry attempts (uses exponential backoff: delay * 2^attempt)
	kiroSocketBaseRetryDelay = 1 * time.Second
	// Maximum delay between retry attempts (cap for exponential backoff)
	kiroSocketMaxRetryDelay = 30 * time.Second
	// First token timeout for streaming responses (how long to wait for first response)
	kiroFirstTokenTimeout = 15 * time.Second
	// Streaming read timeout (how long to wait between chunks)
	kiroStreamingReadTimeout = 300 * time.Second
)

// retryableHTTPStatusCodes defines HTTP status codes that are considered retryable.
// Based on kiro2Api reference: 502 (Bad Gateway), 503 (Service Unavailable), 504 (Gateway Timeout)
var retryableHTTPStatusCodes = map[int]bool{
	502: true, // Bad Gateway - upstream server error
	503: true, // Service Unavailable - server temporarily overloaded
	504: true, // Gateway Timeout - upstream server timeout
}

// Real-time usage estimation configuration
// These control how often usage updates are sent during streaming
var (
	usageUpdateCharThreshold = 5000             // Send usage update every 5000 characters
	usageUpdateTimeInterval  = 15 * time.Second // Or every 15 seconds, whichever comes first
)

// Global FingerprintManager for dynamic User-Agent generation per token
// Each token gets a unique fingerprint on first use, which is cached for subsequent requests
var (
	globalFingerprintManager     *kiroauth.FingerprintManager
	globalFingerprintManagerOnce sync.Once
)

// getGlobalFingerprintManager returns the global FingerprintManager instance
func getGlobalFingerprintManager() *kiroauth.FingerprintManager {
	globalFingerprintManagerOnce.Do(func() {
		globalFingerprintManager = kiroauth.NewFingerprintManager()
		log.Infof("kiro: initialized global FingerprintManager for dynamic UA generation")
	})
	return globalFingerprintManager
}

// retryConfig holds configuration for socket retry logic.
// Based on kiro2Api Python implementation patterns.
type retryConfig struct {
	MaxRetries       int           // Maximum number of retry attempts
	BaseDelay        time.Duration // Base delay between retries (exponential backoff)
	MaxDelay         time.Duration // Maximum delay cap
	RetryableErrors  []string      // List of retryable error patterns
	RetryableStatus  map[int]bool  // HTTP status codes to retry
	FirstTokenTmout  time.Duration // Timeout for first token in streaming
	StreamReadTmout  time.Duration // Timeout between stream chunks
}

// defaultRetryConfig returns the default retry configuration for Kiro socket operations.
func defaultRetryConfig() retryConfig {
	return retryConfig{
		MaxRetries:      kiroSocketMaxRetries,
		BaseDelay:       kiroSocketBaseRetryDelay,
		MaxDelay:        kiroSocketMaxRetryDelay,
		RetryableStatus: retryableHTTPStatusCodes,
		RetryableErrors: []string{
			"connection reset",
			"connection refused",
			"broken pipe",
			"EOF",
			"timeout",
			"temporary failure",
			"no such host",
			"network is unreachable",
			"i/o timeout",
		},
		FirstTokenTmout: kiroFirstTokenTimeout,
		StreamReadTmout: kiroStreamingReadTimeout,
	}
}

// isRetryableError checks if an error is retryable based on error type and message.
// Returns true for network timeouts, connection resets, and temporary failures.
// Based on kiro2Api's retry logic patterns.
func isRetryableError(err error) bool {
	if err == nil {
		return false
	}

	// Check for context cancellation - not retryable
	if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
		return false
	}

	// Check for net.Error (timeout, temporary)
	var netErr net.Error
	if errors.As(err, &netErr) {
		if netErr.Timeout() {
			log.Debugf("kiro: isRetryableError: network timeout detected")
			return true
		}
		// Note: Temporary() is deprecated but still useful for some error types
	}

	// Check for specific syscall errors (connection reset, broken pipe, etc.)
	var syscallErr syscall.Errno
	if errors.As(err, &syscallErr) {
		switch syscallErr {
		case syscall.ECONNRESET: // Connection reset by peer
			log.Debugf("kiro: isRetryableError: ECONNRESET detected")
			return true
		case syscall.ECONNREFUSED: // Connection refused
			log.Debugf("kiro: isRetryableError: ECONNREFUSED detected")
			return true
		case syscall.EPIPE: // Broken pipe
			log.Debugf("kiro: isRetryableError: EPIPE (broken pipe) detected")
			return true
		case syscall.ETIMEDOUT: // Connection timed out
			log.Debugf("kiro: isRetryableError: ETIMEDOUT detected")
			return true
		case syscall.ENETUNREACH: // Network is unreachable
			log.Debugf("kiro: isRetryableError: ENETUNREACH detected")
			return true
		case syscall.EHOSTUNREACH: // No route to host
			log.Debugf("kiro: isRetryableError: EHOSTUNREACH detected")
			return true
		}
	}

	// Check for net.OpError wrapping other errors
	var opErr *net.OpError
	if errors.As(err, &opErr) {
		log.Debugf("kiro: isRetryableError: net.OpError detected, op=%s", opErr.Op)
		// Recursively check the wrapped error
		if opErr.Err != nil {
			return isRetryableError(opErr.Err)
		}
		return true
	}

	// Check error message for retryable patterns
	errMsg := strings.ToLower(err.Error())
	cfg := defaultRetryConfig()
	for _, pattern := range cfg.RetryableErrors {
		if strings.Contains(errMsg, pattern) {
			log.Debugf("kiro: isRetryableError: pattern '%s' matched in error: %s", pattern, errMsg)
			return true
		}
	}

	// Check for EOF which may indicate connection was closed
	if errors.Is(err, io.EOF) || errors.Is(err, io.ErrUnexpectedEOF) {
		log.Debugf("kiro: isRetryableError: EOF/UnexpectedEOF detected")
		return true
	}

	return false
}

// isRetryableHTTPStatus checks if an HTTP status code is retryable.
// Based on kiro2Api: 502, 503, 504 are retryable server errors.
func isRetryableHTTPStatus(statusCode int) bool {
	return retryableHTTPStatusCodes[statusCode]
}

// calculateRetryDelay calculates the delay for the next retry attempt using exponential backoff.
// delay = min(baseDelay * 2^attempt, maxDelay)
// Adds ±30% jitter to prevent thundering herd.
func calculateRetryDelay(attempt int, cfg retryConfig) time.Duration {
	return kiroauth.ExponentialBackoffWithJitter(attempt, cfg.BaseDelay, cfg.MaxDelay)
}

// logRetryAttempt logs a retry attempt with relevant context.
func logRetryAttempt(attempt, maxRetries int, reason string, delay time.Duration, endpoint string) {
	log.Warnf("kiro: retry attempt %d/%d for %s, waiting %v before next attempt (endpoint: %s)",
		attempt+1, maxRetries, reason, delay, endpoint)
}

// kiroHTTPClientPool provides a shared HTTP client with connection pooling for Kiro API.
// This reduces connection overhead and improves performance for concurrent requests.
// Based on kiro2Api's connection pooling pattern.
var (
	kiroHTTPClientPool     *http.Client
	kiroHTTPClientPoolOnce sync.Once
)

// getKiroPooledHTTPClient returns a shared HTTP client with optimized connection pooling.
// The client is lazily initialized on first use and reused across requests.
// This is especially beneficial for:
// - Reducing TCP handshake overhead
// - Enabling HTTP/2 multiplexing
// - Better handling of keep-alive connections
func getKiroPooledHTTPClient() *http.Client {
	kiroHTTPClientPoolOnce.Do(func() {
		transport := &http.Transport{
			// Connection pool settings
			MaxIdleConns:        100,              // Max idle connections across all hosts
			MaxIdleConnsPerHost: 20,               // Max idle connections per host
			MaxConnsPerHost:     50,               // Max total connections per host
			IdleConnTimeout:     90 * time.Second, // How long idle connections stay in pool

			// Timeouts for connection establishment
			DialContext: (&net.Dialer{
				Timeout:   30 * time.Second, // TCP connection timeout
				KeepAlive: 30 * time.Second, // TCP keep-alive interval
			}).DialContext,

			// TLS handshake timeout
			TLSHandshakeTimeout: 10 * time.Second,

			// Response header timeout
			ResponseHeaderTimeout: 30 * time.Second,

			// Expect 100-continue timeout
			ExpectContinueTimeout: 1 * time.Second,

			// Enable HTTP/2 when available
			ForceAttemptHTTP2: true,
		}

		kiroHTTPClientPool = &http.Client{
			Transport: transport,
			// No global timeout - let individual requests set their own timeouts via context
		}

		log.Debugf("kiro: initialized pooled HTTP client (MaxIdleConns=%d, MaxIdleConnsPerHost=%d, MaxConnsPerHost=%d)",
			transport.MaxIdleConns, transport.MaxIdleConnsPerHost, transport.MaxConnsPerHost)
	})

	return kiroHTTPClientPool
}

// newKiroHTTPClientWithPooling creates an HTTP client that uses connection pooling when appropriate.
// It respects proxy configuration from auth or config, falling back to the pooled client.
// This provides the best of both worlds: custom proxy support + connection reuse.
func newKiroHTTPClientWithPooling(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth, timeout time.Duration) *http.Client {
	// Check if a proxy is configured - if so, we need a custom client
	var proxyURL string
	if auth != nil {
		proxyURL = strings.TrimSpace(auth.ProxyURL)
	}
	if proxyURL == "" && cfg != nil {
		proxyURL = strings.TrimSpace(cfg.ProxyURL)
	}

	// If proxy is configured, use the existing proxy-aware client (doesn't pool)
	if proxyURL != "" {
		log.Debugf("kiro: using proxy-aware HTTP client (proxy=%s)", proxyURL)
		return newProxyAwareHTTPClient(ctx, cfg, auth, timeout)
	}

	// No proxy - use pooled client for better performance
	pooledClient := getKiroPooledHTTPClient()

	// If timeout is specified, we need to wrap the pooled transport with timeout
	if timeout > 0 {
		return &http.Client{
			Transport: pooledClient.Transport,
			Timeout:   timeout,
		}
	}

	return pooledClient
}

// kiroEndpointConfig bundles endpoint URL with its compatible Origin and AmzTarget values.
// This solves the "triple mismatch" problem where different endpoints require matching
// Origin and X-Amz-Target header values.
//
// Based on reference implementations:
// - amq2api-main: Uses Amazon Q endpoint with CLI origin and AmazonQDeveloperStreamingService target
// - AIClient-2-API: Uses CodeWhisperer endpoint with AI_EDITOR origin and AmazonCodeWhispererStreamingService target
type kiroEndpointConfig struct {
	URL       string // Endpoint URL
	Origin    string // Request Origin: "CLI" for Amazon Q quota, "AI_EDITOR" for Kiro IDE quota
	AmzTarget string // X-Amz-Target header value
	Name      string // Endpoint name for logging
}

// kiroEndpointConfigs defines the available Kiro API endpoints with their compatible configurations.
// The order determines fallback priority: primary endpoint first, then fallbacks.
//
// CRITICAL: Each endpoint MUST use its compatible Origin and AmzTarget values:
// - CodeWhisperer endpoint (codewhisperer.us-east-1.amazonaws.com): Uses AI_EDITOR origin and AmazonCodeWhispererStreamingService target
// - Amazon Q endpoint (q.us-east-1.amazonaws.com): Uses CLI origin and AmazonQDeveloperStreamingService target
//
// 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.
var kiroEndpointConfigs = []kiroEndpointConfig{
	{
		URL:       "https://codewhisperer.us-east-1.amazonaws.com/generateAssistantResponse",
		Origin:    "AI_EDITOR",
		AmzTarget: "AmazonCodeWhispererStreamingService.GenerateAssistantResponse",
		Name:      "CodeWhisperer",
	},
	{
		URL:       "https://q.us-east-1.amazonaws.com/",
		Origin:    "CLI",
		AmzTarget: "AmazonQDeveloperStreamingService.SendMessage",
		Name:      "AmazonQ",
	},
}

// getKiroEndpointConfigs returns the list of Kiro API endpoint configurations to try in order.
// Supports reordering based on "preferred_endpoint" in auth metadata/attributes.
// For IDC auth method, automatically uses CodeWhisperer endpoint with CLI origin.
func getKiroEndpointConfigs(auth *cliproxyauth.Auth) []kiroEndpointConfig {
	if auth == nil {
		return kiroEndpointConfigs
	}

	// 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
	// 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 authMethod == "idc" {
			log.Debugf("kiro: IDC auth, using CodeWhisperer endpoint")
			return kiroEndpointConfigs
		}
	}

	// Check for preference
	var preference string
	if auth.Metadata != nil {
		if p, ok := auth.Metadata["preferred_endpoint"].(string); ok {
			preference = p
		}
	}
	// Check attributes as fallback (e.g. from HTTP headers)
	if preference == "" && auth.Attributes != nil {
		preference = auth.Attributes["preferred_endpoint"]
	}

	if preference == "" {
		return kiroEndpointConfigs
	}

	preference = strings.ToLower(strings.TrimSpace(preference))

	// Create new slice to avoid modifying global state
	var sorted []kiroEndpointConfig
	var remaining []kiroEndpointConfig

	for _, cfg := range kiroEndpointConfigs {
		name := strings.ToLower(cfg.Name)
		// Check for matches
		// CodeWhisperer aliases: codewhisperer, ide
		// AmazonQ aliases: amazonq, q, cli
		isMatch := false
		if (preference == "codewhisperer" || preference == "ide") && name == "codewhisperer" {
			isMatch = true
		} else if (preference == "amazonq" || preference == "q" || preference == "cli") && name == "amazonq" {
			isMatch = true
		}

		if isMatch {
			sorted = append(sorted, cfg)
		} else {
			remaining = append(remaining, cfg)
		}
	}

	// If preference didn't match anything, return default
	if len(sorted) == 0 {
		return kiroEndpointConfigs
	}

	// Combine: preferred first, then others
	return append(sorted, remaining...)
}

// KiroExecutor handles requests to AWS CodeWhisperer (Kiro) API.
type KiroExecutor struct {
	cfg       *config.Config
	refreshMu sync.Mutex // Serializes token refresh operations to prevent race conditions
}

// isIDCAuth checks if the auth uses IDC (Identity Center) authentication method.
func isIDCAuth(auth *cliproxyauth.Auth) bool {
	if auth == nil || auth.Metadata == nil {
		return false
	}
	authMethod, _ := auth.Metadata["auth_method"].(string)
	return authMethod == "idc"
}

// buildKiroPayloadForFormat builds the Kiro API payload based on the source format.
// This is critical because OpenAI and Claude formats have different tool structures:
// - OpenAI: tools[].function.name, tools[].function.description
// - Claude: tools[].name, tools[].description
// headers parameter allows checking Anthropic-Beta header for thinking mode detection.
// Returns the serialized JSON payload and a boolean indicating whether thinking mode was injected.
func buildKiroPayloadForFormat(body []byte, modelID, profileArn, origin string, isAgentic, isChatOnly bool, sourceFormat sdktranslator.Format, headers http.Header) ([]byte, bool) {
	switch sourceFormat.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)
	default:
		// Default to Claude format (also handles "claude", "kiro", etc.)
		log.Debugf("kiro: using Claude payload builder for source format: %s", sourceFormat.String())
		return kiroclaude.BuildKiroPayload(body, modelID, profileArn, origin, isAgentic, isChatOnly, headers, nil)
	}
}

// NewKiroExecutor creates a new Kiro executor instance.
func NewKiroExecutor(cfg *config.Config) *KiroExecutor {
	return &KiroExecutor{cfg: cfg}
}

// Identifier returns the unique identifier for this executor.
func (e *KiroExecutor) Identifier() string { return "kiro" }

// applyDynamicFingerprint applies token-specific fingerprint headers to the request
// For IDC auth, uses dynamic fingerprint-based User-Agent
// For other auth types, uses static Amazon Q CLI style headers
func applyDynamicFingerprint(req *http.Request, auth *cliproxyauth.Auth) {
	if isIDCAuth(auth) {
		// Get token-specific fingerprint for dynamic UA generation
		tokenKey := getTokenKey(auth)
		fp := getGlobalFingerprintManager().GetFingerprint(tokenKey)
		
		// 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)
		
		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)
	} else {
		// Use static Amazon Q CLI style headers for non-IDC auth
		req.Header.Set("User-Agent", kiroUserAgent)
		req.Header.Set("X-Amz-User-Agent", kiroFullUserAgent)
	}
}

// PrepareRequest prepares the HTTP request before execution.
func (e *KiroExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth.Auth) error {
	if req == nil {
		return nil
	}
	accessToken, _ := kiroCredentials(auth)
	if strings.TrimSpace(accessToken) == "" {
		return statusErr{code: http.StatusUnauthorized, msg: "missing access token"}
	}
	
	// Apply dynamic fingerprint-based headers
	applyDynamicFingerprint(req, auth)
	
	req.Header.Set("Amz-Sdk-Request", "attempt=1; max=3")
	req.Header.Set("Amz-Sdk-Invocation-Id", uuid.New().String())
	req.Header.Set("Authorization", "Bearer "+accessToken)
	var attrs map[string]string
	if auth != nil {
		attrs = auth.Attributes
	}
	util.ApplyCustomHeadersFromAttrs(req, attrs)
	return nil
}

// HttpRequest injects Kiro credentials into the request and executes it.
func (e *KiroExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
	if req == nil {
		return nil, fmt.Errorf("kiro executor: request is nil")
	}
	if ctx == nil {
		ctx = req.Context()
	}
	httpReq := req.WithContext(ctx)
	if errPrepare := e.PrepareRequest(httpReq, auth); errPrepare != nil {
		return nil, errPrepare
	}
	httpClient := newKiroHTTPClientWithPooling(ctx, e.cfg, auth, 0)
	return httpClient.Do(httpReq)
}

// getTokenKey returns a unique key for rate limiting based on auth credentials.
// Uses auth ID if available, otherwise falls back to a hash of the access token.
func getTokenKey(auth *cliproxyauth.Auth) string {
	if auth != nil && auth.ID != "" {
		return auth.ID
	}
	accessToken, _ := kiroCredentials(auth)
	if len(accessToken) > 16 {
		return accessToken[:16]
	}
	return accessToken
}

// Execute sends the request to Kiro API and returns the response.
// Supports automatic token refresh on 401/403 errors.
func (e *KiroExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
	accessToken, profileArn := kiroCredentials(auth)
	if accessToken == "" {
		return resp, fmt.Errorf("kiro: access token not found in auth")
	}

	// Rate limiting: get token key for tracking
	tokenKey := getTokenKey(auth)
	rateLimiter := kiroauth.GetGlobalRateLimiter()
	cooldownMgr := kiroauth.GetGlobalCooldownManager()

	// Check if token is in cooldown period
	if cooldownMgr.IsInCooldown(tokenKey) {
		remaining := cooldownMgr.GetRemainingCooldown(tokenKey)
		reason := cooldownMgr.GetCooldownReason(tokenKey)
		log.Warnf("kiro: token %s is in cooldown (reason: %s), remaining: %v", tokenKey, reason, remaining)
		return resp, fmt.Errorf("kiro: token is in cooldown for %v (reason: %s)", remaining, reason)
	}

	// Wait for rate limiter before proceeding
	log.Debugf("kiro: waiting for rate limiter for token %s", tokenKey)
	rateLimiter.WaitForToken(tokenKey)
	log.Debugf("kiro: rate limiter cleared for token %s", tokenKey)

	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
	defer reporter.trackFailure(ctx, &err)

	// Check if token is expired before making request
	if e.isTokenExpired(accessToken) {
		log.Infof("kiro: access token expired, attempting recovery")

		// 方案 B: 先尝试从文件重新加载 token（后台刷新器可能已更新文件）
		reloadedAuth, reloadErr := e.reloadAuthFromFile(auth)
		if reloadErr == nil && reloadedAuth != nil {
			// 文件中有更新的 token，使用它
			auth = reloadedAuth
			accessToken, profileArn = kiroCredentials(auth)
			log.Infof("kiro: recovered token from file (background refresh), expires_at: %v", auth.Metadata["expires_at"])
		} else {
			// 文件中的 token 也过期了，执行主动刷新
			log.Debugf("kiro: file reload failed (%v), attempting active refresh", reloadErr)
			refreshedAuth, refreshErr := e.Refresh(ctx, auth)
			if refreshErr != nil {
				log.Warnf("kiro: pre-request token refresh failed: %v", refreshErr)
			} else if refreshedAuth != nil {
				auth = refreshedAuth
				// Persist the refreshed auth to file so subsequent requests use it
				if persistErr := e.persistRefreshedAuth(auth); persistErr != nil {
					log.Warnf("kiro: failed to persist refreshed auth: %v", persistErr)
				}
				accessToken, profileArn = kiroCredentials(auth)
				log.Infof("kiro: token refreshed successfully before request")
			}
		}
	}

	from := opts.SourceFormat
	to := sdktranslator.FromString("kiro")
	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)

	kiroModelID := e.mapModelToKiro(req.Model)

	// Determine agentic mode and effective profile ARN using helper functions
	isAgentic, isChatOnly := determineAgenticMode(req.Model)
	effectiveProfileArn := getEffectiveProfileArnWithWarning(auth, profileArn)

	// Execute with retry on 401/403 and 429 (quota exhausted)
	// Note: currentOrigin and kiroPayload are built inside executeWithRetry for each endpoint
	resp, err = e.executeWithRetry(ctx, auth, req, opts, accessToken, effectiveProfileArn, nil, body, from, to, reporter, "", kiroModelID, isAgentic, isChatOnly, tokenKey)
	return resp, err
}

// executeWithRetry performs the actual HTTP request with automatic retry on auth errors.
// Supports automatic fallback between endpoints with different quotas:
// - Amazon Q endpoint (CLI origin) uses Amazon Q Developer quota
// - CodeWhisperer endpoint (AI_EDITOR origin) uses Kiro IDE quota
// Also supports multi-endpoint fallback similar to Antigravity implementation.
// tokenKey is used for rate limiting and cooldown tracking.
func (e *KiroExecutor) executeWithRetry(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, accessToken, profileArn string, kiroPayload, body []byte, from, to sdktranslator.Format, reporter *usageReporter, currentOrigin, kiroModelID string, isAgentic, isChatOnly bool, tokenKey string) (cliproxyexecutor.Response, error) {
	var resp cliproxyexecutor.Response
	maxRetries := 2 // Allow retries for token refresh + endpoint fallback
	rateLimiter := kiroauth.GetGlobalRateLimiter()
	cooldownMgr := kiroauth.GetGlobalCooldownManager()
	endpointConfigs := getKiroEndpointConfigs(auth)
	var last429Err error

	for endpointIdx := 0; endpointIdx < len(endpointConfigs); endpointIdx++ {
		endpointConfig := endpointConfigs[endpointIdx]
		url := endpointConfig.URL
		// Use this endpoint's compatible Origin (critical for avoiding 403 errors)
		currentOrigin = endpointConfig.Origin

		// Rebuild payload with the correct origin for this endpoint
		// Each endpoint requires its matching Origin value in the request body
		kiroPayload, _ = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from, opts.Headers)

		log.Debugf("kiro: trying endpoint %d/%d: %s (Name: %s, Origin: %s)",
			endpointIdx+1, len(endpointConfigs), url, endpointConfig.Name, currentOrigin)

		for attempt := 0; attempt <= maxRetries; attempt++ {
			// Apply human-like delay before first request (not on retries)
			// This mimics natural user behavior patterns
			if attempt == 0 && endpointIdx == 0 {
				kiroauth.ApplyHumanLikeDelay()
			}

			httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(kiroPayload))
			if err != nil {
				return resp, err
			}

			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)

			// Apply dynamic fingerprint-based headers
			applyDynamicFingerprint(httpReq, auth)
			
			httpReq.Header.Set("Amz-Sdk-Request", "attempt=1; max=3")
			httpReq.Header.Set("Amz-Sdk-Invocation-Id", uuid.New().String())

			// Bearer token authentication for all auth types (Builder ID, IDC, social, etc.)
			httpReq.Header.Set("Authorization", "Bearer "+accessToken)

			var attrs map[string]string
			if auth != nil {
				attrs = auth.Attributes
			}
			util.ApplyCustomHeadersFromAttrs(httpReq, attrs)

			var authID, authLabel, authType, authValue string
			if auth != nil {
				authID = auth.ID
				authLabel = auth.Label
				authType, authValue = auth.AccountInfo()
			}
			recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
				URL:       url,
				Method:    http.MethodPost,
				Headers:   httpReq.Header.Clone(),
				Body:      kiroPayload,
				Provider:  e.Identifier(),
				AuthID:    authID,
				AuthLabel: authLabel,
				AuthType:  authType,
				AuthValue: authValue,
			})

			httpClient := newKiroHTTPClientWithPooling(ctx, e.cfg, auth, 120*time.Second)
			httpResp, err := httpClient.Do(httpReq)
			if err != nil {
				// Check for context cancellation first - client disconnected, not a server error
				// Use 499 (Client Closed Request - nginx convention) instead of 500
				if errors.Is(err, context.Canceled) {
					log.Debugf("kiro: request canceled by client (context.Canceled)")
					return resp, statusErr{code: 499, msg: "client canceled request"}
				}

				// Check for context deadline exceeded - request timed out
				// Return 504 Gateway Timeout instead of 500
				if errors.Is(err, context.DeadlineExceeded) {
					log.Debugf("kiro: request timed out (context.DeadlineExceeded)")
					return resp, statusErr{code: http.StatusGatewayTimeout, msg: "upstream request timed out"}
				}

				recordAPIResponseError(ctx, e.cfg, err)

				// Enhanced socket retry: Check if error is retryable (network timeout, connection reset, etc.)
				retryCfg := defaultRetryConfig()
				if isRetryableError(err) && attempt < retryCfg.MaxRetries {
					delay := calculateRetryDelay(attempt, retryCfg)
					logRetryAttempt(attempt, retryCfg.MaxRetries, fmt.Sprintf("socket error: %v", err), delay, endpointConfig.Name)
					time.Sleep(delay)
					continue
				}

				return resp, err
			}
			recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())

			// Handle 429 errors (quota exhausted) - try next endpoint
			// Each endpoint has its own quota pool, so we can try different endpoints
			if httpResp.StatusCode == 429 {
				respBody, _ := io.ReadAll(httpResp.Body)
				_ = httpResp.Body.Close()
				appendAPIResponseChunk(ctx, e.cfg, respBody)

				// Record failure and set cooldown for 429
				rateLimiter.MarkTokenFailed(tokenKey)
				cooldownDuration := kiroauth.CalculateCooldownFor429(attempt)
				cooldownMgr.SetCooldown(tokenKey, cooldownDuration, kiroauth.CooldownReason429)
				log.Warnf("kiro: rate limit hit (429), token %s set to cooldown for %v", tokenKey, cooldownDuration)

				// Preserve last 429 so callers can correctly backoff when all endpoints are exhausted
				last429Err = statusErr{code: httpResp.StatusCode, msg: string(respBody)}

				log.Warnf("kiro: %s endpoint quota exhausted (429), will try next endpoint, body: %s",
					endpointConfig.Name, summarizeErrorBody(httpResp.Header.Get("Content-Type"), respBody))

				// Break inner retry loop to try next endpoint (which has different quota)
				break
			}

			// Handle 5xx server errors with exponential backoff retry
			// Enhanced: Use retryConfig for consistent retry behavior
			if httpResp.StatusCode >= 500 && httpResp.StatusCode < 600 {
				respBody, _ := io.ReadAll(httpResp.Body)
				_ = httpResp.Body.Close()
				appendAPIResponseChunk(ctx, e.cfg, respBody)

				retryCfg := defaultRetryConfig()
				// Check if this specific 5xx code is retryable (502, 503, 504)
				if isRetryableHTTPStatus(httpResp.StatusCode) && attempt < retryCfg.MaxRetries {
					delay := calculateRetryDelay(attempt, retryCfg)
					logRetryAttempt(attempt, retryCfg.MaxRetries, fmt.Sprintf("HTTP %d", httpResp.StatusCode), delay, endpointConfig.Name)
					time.Sleep(delay)
					continue
				} else if attempt < maxRetries {
					// Fallback for other 5xx errors (500, 501, etc.)
					backoff := time.Duration(1<<attempt) * time.Second
					if backoff > 30*time.Second {
						backoff = 30 * time.Second
					}
					log.Warnf("kiro: server error %d, retrying in %v (attempt %d/%d)", httpResp.StatusCode, backoff, attempt+1, maxRetries)
					time.Sleep(backoff)
					continue
				}
				log.Errorf("kiro: server error %d after %d retries", httpResp.StatusCode, maxRetries)
				return resp, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
			}

			// Handle 401 errors with token refresh and retry
			// 401 = Unauthorized (token expired/invalid) - refresh token
			if httpResp.StatusCode == 401 {
				respBody, _ := io.ReadAll(httpResp.Body)
				_ = httpResp.Body.Close()
				appendAPIResponseChunk(ctx, e.cfg, respBody)

				log.Warnf("kiro: received 401 error, attempting token refresh")
				refreshedAuth, refreshErr := e.Refresh(ctx, auth)
				if refreshErr != nil {
					log.Errorf("kiro: token refresh failed: %v", refreshErr)
					return resp, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
				}

				if refreshedAuth != nil {
					auth = refreshedAuth
					// Persist the refreshed auth to file so subsequent requests use it
					if persistErr := e.persistRefreshedAuth(auth); persistErr != nil {
						log.Warnf("kiro: failed to persist refreshed auth: %v", persistErr)
						// Continue anyway - the token is valid for this request
					}
					accessToken, profileArn = kiroCredentials(auth)
					// Rebuild payload with new profile ARN if changed
					kiroPayload, _ = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from, opts.Headers)
					if attempt < maxRetries {
						log.Infof("kiro: token refreshed successfully, retrying request (attempt %d/%d)", attempt+1, maxRetries+1)
						continue
					}
					log.Infof("kiro: token refreshed successfully, no retries remaining")
				}

				log.Warnf("kiro request error, status: 401, body: %s", summarizeErrorBody(httpResp.Header.Get("Content-Type"), respBody))
				return resp, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
			}

			// Handle 402 errors - Monthly Limit Reached
			if httpResp.StatusCode == 402 {
				respBody, _ := io.ReadAll(httpResp.Body)
				_ = httpResp.Body.Close()
				appendAPIResponseChunk(ctx, e.cfg, respBody)

				log.Warnf("kiro: received 402 (monthly limit). Upstream body: %s", string(respBody))

				// Return upstream error body directly
				return resp, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
			}

			// Handle 403 errors - Access Denied / Token Expired
			// Do NOT switch endpoints for 403 errors
			if httpResp.StatusCode == 403 {
				respBody, _ := io.ReadAll(httpResp.Body)
				_ = httpResp.Body.Close()
				appendAPIResponseChunk(ctx, e.cfg, respBody)

				// Log the 403 error details for debugging
				log.Warnf("kiro: received 403 error (attempt %d/%d), body: %s", attempt+1, maxRetries+1, summarizeErrorBody(httpResp.Header.Get("Content-Type"), respBody))

				respBodyStr := string(respBody)

				// Check for SUSPENDED status - return immediately without retry
				if strings.Contains(respBodyStr, "SUSPENDED") || strings.Contains(respBodyStr, "TEMPORARILY_SUSPENDED") {
					// Set long cooldown for suspended accounts
					rateLimiter.CheckAndMarkSuspended(tokenKey, respBodyStr)
					cooldownMgr.SetCooldown(tokenKey, kiroauth.LongCooldown, kiroauth.CooldownReasonSuspended)
					log.Errorf("kiro: account is suspended, token %s set to cooldown for %v", tokenKey, kiroauth.LongCooldown)
					return resp, statusErr{code: httpResp.StatusCode, msg: "account suspended: " + string(respBody)}
				}

				// Check if this looks like a token-related 403 (some APIs return 403 for expired tokens)
				isTokenRelated := strings.Contains(respBodyStr, "token") ||
					strings.Contains(respBodyStr, "expired") ||
					strings.Contains(respBodyStr, "invalid") ||
					strings.Contains(respBodyStr, "unauthorized")

				if isTokenRelated && attempt < maxRetries {
					log.Warnf("kiro: 403 appears token-related, attempting token refresh")
					refreshedAuth, refreshErr := e.Refresh(ctx, auth)
					if refreshErr != nil {
						log.Errorf("kiro: token refresh failed: %v", refreshErr)
						// Token refresh failed - return error immediately
						return resp, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
					}
					if refreshedAuth != nil {
						auth = refreshedAuth
						// Persist the refreshed auth to file so subsequent requests use it
						if persistErr := e.persistRefreshedAuth(auth); persistErr != nil {
							log.Warnf("kiro: failed to persist refreshed auth: %v", persistErr)
							// Continue anyway - the token is valid for this request
						}
						accessToken, profileArn = kiroCredentials(auth)
						kiroPayload, _ = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from, opts.Headers)
						log.Infof("kiro: token refreshed for 403, retrying request")
						continue
					}
				}

				// For non-token 403 or after max retries, return error immediately
				// Do NOT switch endpoints for 403 errors
				log.Warnf("kiro: 403 error, returning immediately (no endpoint switch)")
				return resp, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
			}

			if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
				b, _ := io.ReadAll(httpResp.Body)
				appendAPIResponseChunk(ctx, e.cfg, b)
				log.Debugf("kiro request error, status: %d, body: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
				err = statusErr{code: httpResp.StatusCode, msg: string(b)}
				if errClose := httpResp.Body.Close(); errClose != nil {
					log.Errorf("response body close error: %v", errClose)
				}
				return resp, err
			}

			defer func() {
				if errClose := httpResp.Body.Close(); errClose != nil {
					log.Errorf("response body close error: %v", errClose)
				}
			}()

			content, toolUses, usageInfo, stopReason, err := e.parseEventStream(httpResp.Body)
			if err != nil {
				recordAPIResponseError(ctx, e.cfg, err)
				return resp, err
			}

			// Fallback for usage if missing from upstream
			if usageInfo.TotalTokens == 0 {
				if enc, encErr := getTokenizer(req.Model); encErr == nil {
					if inp, countErr := countOpenAIChatTokens(enc, opts.OriginalRequest); countErr == nil {
						usageInfo.InputTokens = inp
					}
				}
				if len(content) > 0 {
					// Use tiktoken for more accurate output token calculation
					if enc, encErr := getTokenizer(req.Model); encErr == nil {
						if tokenCount, countErr := enc.Count(content); countErr == nil {
							usageInfo.OutputTokens = int64(tokenCount)
						}
					}
					// Fallback to character count estimation if tiktoken fails
					if usageInfo.OutputTokens == 0 {
						usageInfo.OutputTokens = int64(len(content) / 4)
						if usageInfo.OutputTokens == 0 {
							usageInfo.OutputTokens = 1
						}
					}
				}
				usageInfo.TotalTokens = usageInfo.InputTokens + usageInfo.OutputTokens
			}

			appendAPIResponseChunk(ctx, e.cfg, []byte(content))
			reporter.publish(ctx, usageInfo)

			// Record success for rate limiting
			rateLimiter.MarkTokenSuccess(tokenKey)
			log.Debugf("kiro: request successful, token %s marked as success", tokenKey)

			// Build response in Claude format for Kiro translator
			// stopReason is extracted from upstream response by parseEventStream
			kiroResponse := kiroclaude.BuildClaudeResponse(content, toolUses, req.Model, usageInfo, stopReason)
			out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, kiroResponse, nil)
			resp = cliproxyexecutor.Response{Payload: []byte(out)}
			return resp, nil
		}
		// Inner retry loop exhausted for this endpoint, try next endpoint
		// Note: This code is unreachable because all paths in the inner loop
		// either return or continue. Kept as comment for documentation.
	}

	// All endpoints exhausted
	if last429Err != nil {
		return resp, last429Err
	}
	return resp, fmt.Errorf("kiro: all endpoints exhausted")
}

// ExecuteStream handles streaming requests to Kiro API.
// Supports automatic token refresh on 401/403 errors and quota fallback on 429.
func (e *KiroExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
	accessToken, profileArn := kiroCredentials(auth)
	if accessToken == "" {
		return nil, fmt.Errorf("kiro: access token not found in auth")
	}

	// Rate limiting: get token key for tracking
	tokenKey := getTokenKey(auth)
	rateLimiter := kiroauth.GetGlobalRateLimiter()
	cooldownMgr := kiroauth.GetGlobalCooldownManager()

	// Check if token is in cooldown period
	if cooldownMgr.IsInCooldown(tokenKey) {
		remaining := cooldownMgr.GetRemainingCooldown(tokenKey)
		reason := cooldownMgr.GetCooldownReason(tokenKey)
		log.Warnf("kiro: token %s is in cooldown (reason: %s), remaining: %v", tokenKey, reason, remaining)
		return nil, fmt.Errorf("kiro: token is in cooldown for %v (reason: %s)", remaining, reason)
	}

	// Wait for rate limiter before proceeding
	log.Debugf("kiro: stream waiting for rate limiter for token %s", tokenKey)
	rateLimiter.WaitForToken(tokenKey)
	log.Debugf("kiro: stream rate limiter cleared for token %s", tokenKey)

	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
	defer reporter.trackFailure(ctx, &err)

	// Check if token is expired before making request
	if e.isTokenExpired(accessToken) {
		log.Infof("kiro: access token expired, attempting recovery before stream request")

		// 方案 B: 先尝试从文件重新加载 token（后台刷新器可能已更新文件）
		reloadedAuth, reloadErr := e.reloadAuthFromFile(auth)
		if reloadErr == nil && reloadedAuth != nil {
			// 文件中有更新的 token，使用它
			auth = reloadedAuth
			accessToken, profileArn = kiroCredentials(auth)
			log.Infof("kiro: recovered token from file (background refresh) for stream, expires_at: %v", auth.Metadata["expires_at"])
		} else {
			// 文件中的 token 也过期了，执行主动刷新
			log.Debugf("kiro: file reload failed (%v), attempting active refresh for stream", reloadErr)
			refreshedAuth, refreshErr := e.Refresh(ctx, auth)
			if refreshErr != nil {
				log.Warnf("kiro: pre-request token refresh failed: %v", refreshErr)
			} else if refreshedAuth != nil {
				auth = refreshedAuth
				// Persist the refreshed auth to file so subsequent requests use it
				if persistErr := e.persistRefreshedAuth(auth); persistErr != nil {
					log.Warnf("kiro: failed to persist refreshed auth: %v", persistErr)
				}
				accessToken, profileArn = kiroCredentials(auth)
				log.Infof("kiro: token refreshed successfully before stream request")
			}
		}
	}

	from := opts.SourceFormat
	to := sdktranslator.FromString("kiro")
	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)

	kiroModelID := e.mapModelToKiro(req.Model)

	// Determine agentic mode and effective profile ARN using helper functions
	isAgentic, isChatOnly := determineAgenticMode(req.Model)
	effectiveProfileArn := getEffectiveProfileArnWithWarning(auth, profileArn)

	// Execute stream with retry on 401/403 and 429 (quota exhausted)
	// Note: currentOrigin and kiroPayload are built inside executeStreamWithRetry for each endpoint
	return e.executeStreamWithRetry(ctx, auth, req, opts, accessToken, effectiveProfileArn, nil, body, from, reporter, "", kiroModelID, isAgentic, isChatOnly, tokenKey)
}

// executeStreamWithRetry performs the streaming HTTP request with automatic retry on auth errors.
// Supports automatic fallback between endpoints with different quotas:
// - Amazon Q endpoint (CLI origin) uses Amazon Q Developer quota
// - CodeWhisperer endpoint (AI_EDITOR origin) uses Kiro IDE quota
// Also supports multi-endpoint fallback similar to Antigravity implementation.
// tokenKey is used for rate limiting and cooldown tracking.
func (e *KiroExecutor) executeStreamWithRetry(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, accessToken, profileArn string, kiroPayload, body []byte, from sdktranslator.Format, reporter *usageReporter, currentOrigin, kiroModelID string, isAgentic, isChatOnly bool, tokenKey string) (<-chan cliproxyexecutor.StreamChunk, error) {
	maxRetries := 2 // Allow retries for token refresh + endpoint fallback
	rateLimiter := kiroauth.GetGlobalRateLimiter()
	cooldownMgr := kiroauth.GetGlobalCooldownManager()
	endpointConfigs := getKiroEndpointConfigs(auth)
	var last429Err error

	for endpointIdx := 0; endpointIdx < len(endpointConfigs); endpointIdx++ {
		endpointConfig := endpointConfigs[endpointIdx]
		url := endpointConfig.URL
		// Use this endpoint's compatible Origin (critical for avoiding 403 errors)
		currentOrigin = endpointConfig.Origin

		// Rebuild payload with the correct origin for this endpoint
		// Each endpoint requires its matching Origin value in the request body
		kiroPayload, thinkingEnabled := buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from, opts.Headers)

		log.Debugf("kiro: stream trying endpoint %d/%d: %s (Name: %s, Origin: %s)",
			endpointIdx+1, len(endpointConfigs), url, endpointConfig.Name, currentOrigin)

		for attempt := 0; attempt <= maxRetries; attempt++ {
			// Apply human-like delay before first streaming request (not on retries)
			// This mimics natural user behavior patterns
			// Note: Delay is NOT applied during streaming response - only before initial request
			if attempt == 0 && endpointIdx == 0 {
				kiroauth.ApplyHumanLikeDelay()
			}

			httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(kiroPayload))
			if err != nil {
				return nil, err
			}

			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)

			// Apply dynamic fingerprint-based headers
			applyDynamicFingerprint(httpReq, auth)
			
			httpReq.Header.Set("Amz-Sdk-Request", "attempt=1; max=3")
			httpReq.Header.Set("Amz-Sdk-Invocation-Id", uuid.New().String())

			// Bearer token authentication for all auth types (Builder ID, IDC, social, etc.)
			httpReq.Header.Set("Authorization", "Bearer "+accessToken)

			var attrs map[string]string
			if auth != nil {
				attrs = auth.Attributes
			}
			util.ApplyCustomHeadersFromAttrs(httpReq, attrs)

			var authID, authLabel, authType, authValue string
			if auth != nil {
				authID = auth.ID
				authLabel = auth.Label
				authType, authValue = auth.AccountInfo()
			}
			recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
				URL:       url,
				Method:    http.MethodPost,
				Headers:   httpReq.Header.Clone(),
				Body:      kiroPayload,
				Provider:  e.Identifier(),
				AuthID:    authID,
				AuthLabel: authLabel,
				AuthType:  authType,
				AuthValue: authValue,
			})

			httpClient := newKiroHTTPClientWithPooling(ctx, e.cfg, auth, 0)
			httpResp, err := httpClient.Do(httpReq)
			if err != nil {
				recordAPIResponseError(ctx, e.cfg, err)

				// Enhanced socket retry for streaming: Check if error is retryable (network timeout, connection reset, etc.)
				retryCfg := defaultRetryConfig()
				if isRetryableError(err) && attempt < retryCfg.MaxRetries {
					delay := calculateRetryDelay(attempt, retryCfg)
					logRetryAttempt(attempt, retryCfg.MaxRetries, fmt.Sprintf("stream socket error: %v", err), delay, endpointConfig.Name)
					time.Sleep(delay)
					continue
				}

				return nil, err
			}
			recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())

			// Handle 429 errors (quota exhausted) - try next endpoint
			// Each endpoint has its own quota pool, so we can try different endpoints
			if httpResp.StatusCode == 429 {
				respBody, _ := io.ReadAll(httpResp.Body)
				_ = httpResp.Body.Close()
				appendAPIResponseChunk(ctx, e.cfg, respBody)

				// Record failure and set cooldown for 429
				rateLimiter.MarkTokenFailed(tokenKey)
				cooldownDuration := kiroauth.CalculateCooldownFor429(attempt)
				cooldownMgr.SetCooldown(tokenKey, cooldownDuration, kiroauth.CooldownReason429)
				log.Warnf("kiro: stream rate limit hit (429), token %s set to cooldown for %v", tokenKey, cooldownDuration)

				// Preserve last 429 so callers can correctly backoff when all endpoints are exhausted
				last429Err = statusErr{code: httpResp.StatusCode, msg: string(respBody)}

				log.Warnf("kiro: stream %s endpoint quota exhausted (429), will try next endpoint, body: %s",
					endpointConfig.Name, summarizeErrorBody(httpResp.Header.Get("Content-Type"), respBody))

				// Break inner retry loop to try next endpoint (which has different quota)
				break
			}

			// Handle 5xx server errors with exponential backoff retry
			// Enhanced: Use retryConfig for consistent retry behavior
			if httpResp.StatusCode >= 500 && httpResp.StatusCode < 600 {
				respBody, _ := io.ReadAll(httpResp.Body)
				_ = httpResp.Body.Close()
				appendAPIResponseChunk(ctx, e.cfg, respBody)

				retryCfg := defaultRetryConfig()
				// Check if this specific 5xx code is retryable (502, 503, 504)
				if isRetryableHTTPStatus(httpResp.StatusCode) && attempt < retryCfg.MaxRetries {
					delay := calculateRetryDelay(attempt, retryCfg)
					logRetryAttempt(attempt, retryCfg.MaxRetries, fmt.Sprintf("stream HTTP %d", httpResp.StatusCode), delay, endpointConfig.Name)
					time.Sleep(delay)
					continue
				} else if attempt < maxRetries {
					// Fallback for other 5xx errors (500, 501, etc.)
					backoff := time.Duration(1<<attempt) * time.Second
					if backoff > 30*time.Second {
						backoff = 30 * time.Second
					}
					log.Warnf("kiro: stream server error %d, retrying in %v (attempt %d/%d)", httpResp.StatusCode, backoff, attempt+1, maxRetries)
					time.Sleep(backoff)
					continue
				}
				log.Errorf("kiro: stream server error %d after %d retries", httpResp.StatusCode, maxRetries)
				return nil, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
			}

			// Handle 400 errors - Credential/Validation issues
			// Do NOT switch endpoints - return error immediately
			if httpResp.StatusCode == 400 {
				respBody, _ := io.ReadAll(httpResp.Body)
				_ = httpResp.Body.Close()
				appendAPIResponseChunk(ctx, e.cfg, respBody)

				log.Warnf("kiro: received 400 error (attempt %d/%d), body: %s", attempt+1, maxRetries+1, summarizeErrorBody(httpResp.Header.Get("Content-Type"), respBody))

				// 400 errors indicate request validation issues - return immediately without retry
				return nil, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
			}

			// Handle 401 errors with token refresh and retry
			// 401 = Unauthorized (token expired/invalid) - refresh token
			if httpResp.StatusCode == 401 {
				respBody, _ := io.ReadAll(httpResp.Body)
				_ = httpResp.Body.Close()
				appendAPIResponseChunk(ctx, e.cfg, respBody)

				log.Warnf("kiro: stream received 401 error, attempting token refresh")
				refreshedAuth, refreshErr := e.Refresh(ctx, auth)
				if refreshErr != nil {
					log.Errorf("kiro: token refresh failed: %v", refreshErr)
					return nil, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
				}

				if refreshedAuth != nil {
					auth = refreshedAuth
					// Persist the refreshed auth to file so subsequent requests use it
					if persistErr := e.persistRefreshedAuth(auth); persistErr != nil {
						log.Warnf("kiro: failed to persist refreshed auth: %v", persistErr)
						// Continue anyway - the token is valid for this request
					}
					accessToken, profileArn = kiroCredentials(auth)
					// Rebuild payload with new profile ARN if changed
					kiroPayload, _ = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from, opts.Headers)
					if attempt < maxRetries {
						log.Infof("kiro: token refreshed successfully, retrying stream request (attempt %d/%d)", attempt+1, maxRetries+1)
						continue
					}
					log.Infof("kiro: token refreshed successfully, no retries remaining")
				}

				log.Warnf("kiro stream error, status: 401, body: %s", string(respBody))
				return nil, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
			}

			// Handle 402 errors - Monthly Limit Reached
			if httpResp.StatusCode == 402 {
				respBody, _ := io.ReadAll(httpResp.Body)
				_ = httpResp.Body.Close()
				appendAPIResponseChunk(ctx, e.cfg, respBody)

				log.Warnf("kiro: stream received 402 (monthly limit). Upstream body: %s", string(respBody))

				// Return upstream error body directly
				return nil, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
			}

			// Handle 403 errors - Access Denied / Token Expired
			// Do NOT switch endpoints for 403 errors
			if httpResp.StatusCode == 403 {
				respBody, _ := io.ReadAll(httpResp.Body)
				_ = httpResp.Body.Close()
				appendAPIResponseChunk(ctx, e.cfg, respBody)

				// Log the 403 error details for debugging
				log.Warnf("kiro: stream received 403 error (attempt %d/%d), body: %s", attempt+1, maxRetries+1, string(respBody))

				respBodyStr := string(respBody)

				// Check for SUSPENDED status - return immediately without retry
				if strings.Contains(respBodyStr, "SUSPENDED") || strings.Contains(respBodyStr, "TEMPORARILY_SUSPENDED") {
					// Set long cooldown for suspended accounts
					rateLimiter.CheckAndMarkSuspended(tokenKey, respBodyStr)
					cooldownMgr.SetCooldown(tokenKey, kiroauth.LongCooldown, kiroauth.CooldownReasonSuspended)
					log.Errorf("kiro: stream account is suspended, token %s set to cooldown for %v", tokenKey, kiroauth.LongCooldown)
					return nil, statusErr{code: httpResp.StatusCode, msg: "account suspended: " + string(respBody)}
				}

				// Check if this looks like a token-related 403 (some APIs return 403 for expired tokens)
				isTokenRelated := strings.Contains(respBodyStr, "token") ||
					strings.Contains(respBodyStr, "expired") ||
					strings.Contains(respBodyStr, "invalid") ||
					strings.Contains(respBodyStr, "unauthorized")

				if isTokenRelated && attempt < maxRetries {
					log.Warnf("kiro: 403 appears token-related, attempting token refresh")
					refreshedAuth, refreshErr := e.Refresh(ctx, auth)
					if refreshErr != nil {
						log.Errorf("kiro: token refresh failed: %v", refreshErr)
						// Token refresh failed - return error immediately
						return nil, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
					}
					if refreshedAuth != nil {
						auth = refreshedAuth
						// Persist the refreshed auth to file so subsequent requests use it
						if persistErr := e.persistRefreshedAuth(auth); persistErr != nil {
							log.Warnf("kiro: failed to persist refreshed auth: %v", persistErr)
							// Continue anyway - the token is valid for this request
						}
						accessToken, profileArn = kiroCredentials(auth)
						kiroPayload, _ = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from, opts.Headers)
						log.Infof("kiro: token refreshed for 403, retrying stream request")
						continue
					}
				}

				// For non-token 403 or after max retries, return error immediately
				// Do NOT switch endpoints for 403 errors
				log.Warnf("kiro: 403 error, returning immediately (no endpoint switch)")
				return nil, statusErr{code: httpResp.StatusCode, msg: string(respBody)}
			}

			if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
				b, _ := io.ReadAll(httpResp.Body)
				appendAPIResponseChunk(ctx, e.cfg, b)
				log.Debugf("kiro stream error, status: %d, body: %s", httpResp.StatusCode, string(b))
				if errClose := httpResp.Body.Close(); errClose != nil {
					log.Errorf("response body close error: %v", errClose)
				}
				return nil, statusErr{code: httpResp.StatusCode, msg: string(b)}
			}

			out := make(chan cliproxyexecutor.StreamChunk)

			// Record success immediately since connection was established successfully
			// Streaming errors will be handled separately
			rateLimiter.MarkTokenSuccess(tokenKey)
			log.Debugf("kiro: stream request successful, token %s marked as success", tokenKey)

			go func(resp *http.Response, thinkingEnabled bool) {
				defer close(out)
				defer func() {
					if r := recover(); r != nil {
						log.Errorf("kiro: panic in stream handler: %v", r)
						out <- cliproxyexecutor.StreamChunk{Err: fmt.Errorf("internal error: %v", r)}
					}
				}()
				defer func() {
					if errClose := resp.Body.Close(); errClose != nil {
						log.Errorf("response body close error: %v", errClose)
					}
				}()

				// Kiro API always returns <thinking> tags regardless of request parameters
				// So we always enable thinking parsing for Kiro responses
				log.Debugf("kiro: stream thinkingEnabled = %v (always true for Kiro)", thinkingEnabled)

				e.streamToChannel(ctx, resp.Body, out, from, req.Model, opts.OriginalRequest, body, reporter, thinkingEnabled)
			}(httpResp, thinkingEnabled)

			return out, nil
		}
		// Inner retry loop exhausted for this endpoint, try next endpoint
		// Note: This code is unreachable because all paths in the inner loop
		// either return or continue. Kept as comment for documentation.
	}

	// All endpoints exhausted
	if last429Err != nil {
		return nil, last429Err
	}
	return nil, fmt.Errorf("kiro: stream all endpoints exhausted")
}

// kiroCredentials extracts access token and profile ARN from auth.
func kiroCredentials(auth *cliproxyauth.Auth) (accessToken, profileArn string) {
	if auth == nil {
		return "", ""
	}

	// Try Metadata first (wrapper format)
	if auth.Metadata != nil {
		if token, ok := auth.Metadata["access_token"].(string); ok {
			accessToken = token
		}
		if arn, ok := auth.Metadata["profile_arn"].(string); ok {
			profileArn = arn
		}
	}

	// Try Attributes
	if accessToken == "" && auth.Attributes != nil {
		accessToken = auth.Attributes["access_token"]
		profileArn = auth.Attributes["profile_arn"]
	}

	// Try direct fields from flat JSON format (new AWS Builder ID format)
	if accessToken == "" && auth.Metadata != nil {
		if token, ok := auth.Metadata["accessToken"].(string); ok {
			accessToken = token
		}
		if arn, ok := auth.Metadata["profileArn"].(string); ok {
			profileArn = arn
		}
	}

	return accessToken, profileArn
}

// findRealThinkingEndTag finds the real </thinking> end tag, skipping false positives.
// Returns -1 if no real end tag is found.
//
// Real </thinking> tags from Kiro API have specific characteristics:
// - Usually preceded by newline (.\n</thinking>)
// - Usually followed by newline (\n\n)
// - Not inside code blocks or inline code
//
// False positives (discussion text) have characteristics:
// - In the middle of a sentence
// - Preceded by discussion words like "标签", "tag", "returns"
// - Inside code blocks or inline code
//
// Parameters:
// - content: the content to search in
// - alreadyInCodeBlock: whether we're already inside a code block from previous chunks
// - alreadyInInlineCode: whether we're already inside inline code from previous chunks
func findRealThinkingEndTag(content string, alreadyInCodeBlock, alreadyInInlineCode bool) int {
	searchStart := 0
	for {
		endIdx := strings.Index(content[searchStart:], kirocommon.ThinkingEndTag)
		if endIdx < 0 {
			return -1
		}
		endIdx += searchStart // Adjust to absolute position

		textBeforeEnd := content[:endIdx]
		textAfterEnd := content[endIdx+len(kirocommon.ThinkingEndTag):]

		// Check 1: Is it inside inline code?
		// Count backticks in current content and add state from previous chunks
		backtickCount := strings.Count(textBeforeEnd, "`")
		effectiveInInlineCode := alreadyInInlineCode
		if backtickCount%2 == 1 {
			effectiveInInlineCode = !effectiveInInlineCode
		}
		if effectiveInInlineCode {
			log.Debugf("kiro: found </thinking> inside inline code at pos %d, skipping", endIdx)
			searchStart = endIdx + len(kirocommon.ThinkingEndTag)
			continue
		}

		// Check 2: Is it inside a code block?
		// Count fences in current content and add state from previous chunks
		fenceCount := strings.Count(textBeforeEnd, "```")
		altFenceCount := strings.Count(textBeforeEnd, "~~~")
		effectiveInCodeBlock := alreadyInCodeBlock
		if fenceCount%2 == 1 || altFenceCount%2 == 1 {
			effectiveInCodeBlock = !effectiveInCodeBlock
		}
		if effectiveInCodeBlock {
			log.Debugf("kiro: found </thinking> inside code block at pos %d, skipping", endIdx)
			searchStart = endIdx + len(kirocommon.ThinkingEndTag)
			continue
		}

		// Check 3: Real </thinking> tags are usually preceded by newline or at start
		// and followed by newline or at end. Check the format.
		charBeforeTag := byte(0)
		if endIdx > 0 {
			charBeforeTag = content[endIdx-1]
		}
		charAfterTag := byte(0)
		if len(textAfterEnd) > 0 {
			charAfterTag = textAfterEnd[0]
		}

		// Real end tag format: preceded by newline OR end of sentence (. ! ?)
		// and followed by newline OR end of content
		isPrecededByNewlineOrSentenceEnd := charBeforeTag == '\n' || charBeforeTag == '.' ||
			charBeforeTag == '!' || charBeforeTag == '?' || charBeforeTag == 0
		isFollowedByNewlineOrEnd := charAfterTag == '\n' || charAfterTag == 0

		// If the tag has proper formatting (newline before/after), it's likely real
		if isPrecededByNewlineOrSentenceEnd && isFollowedByNewlineOrEnd {
			log.Debugf("kiro: found properly formatted </thinking> at pos %d", endIdx)
			return endIdx
		}

		// Check 4: Is the tag preceded by discussion keywords on the same line?
		lastNewlineIdx := strings.LastIndex(textBeforeEnd, "\n")
		lineBeforeTag := textBeforeEnd
		if lastNewlineIdx >= 0 {
			lineBeforeTag = textBeforeEnd[lastNewlineIdx+1:]
		}
		lineBeforeTagLower := strings.ToLower(lineBeforeTag)

		// Discussion patterns - if found, this is likely discussion text
		discussionPatterns := []string{
			"标签", "返回", "输出", "包含", "使用", "解析", "转换", "生成", // Chinese
			"tag", "return", "output", "contain", "use", "parse", "emit", "convert", "generate", // English
			"<thinking>",    // discussing both tags together
			"`</thinking>`", // explicitly in inline code
		}
		isDiscussion := false
		for _, pattern := range discussionPatterns {
			if strings.Contains(lineBeforeTagLower, pattern) {
				isDiscussion = true
				break
			}
		}
		if isDiscussion {
			log.Debugf("kiro: found </thinking> after discussion text at pos %d, skipping", endIdx)
			searchStart = endIdx + len(kirocommon.ThinkingEndTag)
			continue
		}

		// Check 5: Is there text immediately after on the same line?
		// Real end tags don't have text immediately after on the same line
		if len(textAfterEnd) > 0 && charAfterTag != '\n' && charAfterTag != 0 {
			// Find the next newline
			nextNewline := strings.Index(textAfterEnd, "\n")
			var textOnSameLine string
			if nextNewline >= 0 {
				textOnSameLine = textAfterEnd[:nextNewline]
			} else {
				textOnSameLine = textAfterEnd
			}
			// If there's non-whitespace text on the same line after the tag, it's discussion
			if strings.TrimSpace(textOnSameLine) != "" {
				log.Debugf("kiro: found </thinking> with text after on same line at pos %d, skipping", endIdx)
				searchStart = endIdx + len(kirocommon.ThinkingEndTag)
				continue
			}
		}

		// Check 6: Is there another <thinking> tag after this </thinking>?
		if strings.Contains(textAfterEnd, kirocommon.ThinkingStartTag) {
			nextStartIdx := strings.Index(textAfterEnd, kirocommon.ThinkingStartTag)
			textBeforeNextStart := textAfterEnd[:nextStartIdx]
			nextBacktickCount := strings.Count(textBeforeNextStart, "`")
			nextFenceCount := strings.Count(textBeforeNextStart, "```")
			nextAltFenceCount := strings.Count(textBeforeNextStart, "~~~")

			// If the next <thinking> is NOT in code, then this </thinking> is discussion text
			if nextBacktickCount%2 == 0 && nextFenceCount%2 == 0 && nextAltFenceCount%2 == 0 {
				log.Debugf("kiro: found </thinking> followed by <thinking> at pos %d, likely discussion text, skipping", endIdx)
				searchStart = endIdx + len(kirocommon.ThinkingEndTag)
				continue
			}
		}

		// This looks like a real end tag
		return endIdx
	}
}

// determineAgenticMode determines if the model is an agentic or chat-only variant.
// Returns (isAgentic, isChatOnly) based on model name suffixes.
func determineAgenticMode(model string) (isAgentic, isChatOnly bool) {
	isAgentic = strings.HasSuffix(model, "-agentic")
	isChatOnly = strings.HasSuffix(model, "-chat")
	return isAgentic, isChatOnly
}

// getEffectiveProfileArn determines if profileArn should be included based on auth method.
// profileArn is only needed for social auth (Google OAuth), not for builder-id (AWS SSO).
func getEffectiveProfileArn(auth *cliproxyauth.Auth, profileArn string) string {
	if auth != nil && auth.Metadata != nil {
		if authMethod, ok := auth.Metadata["auth_method"].(string); ok && authMethod == "builder-id" {
			return "" // Don't include profileArn for builder-id auth
		}
	}
	return profileArn
}

// getEffectiveProfileArnWithWarning determines if profileArn should be included based on auth method,
// and logs a warning if profileArn is missing for non-builder-id auth.
// This consolidates the auth_method check that was previously done separately.
func getEffectiveProfileArnWithWarning(auth *cliproxyauth.Auth, profileArn string) string {
	if auth != nil && auth.Metadata != nil {
		if authMethod, ok := auth.Metadata["auth_method"].(string); ok && (authMethod == "builder-id" || authMethod == "idc") {
			// builder-id and idc auth don't need profileArn
			return ""
		}
	}
	// For non-builder-id/idc auth (social auth), profileArn is required
	if profileArn == "" {
		log.Warnf("kiro: profile ARN not found in auth, API calls may fail")
	}
	return profileArn
}

// mapModelToKiro maps external model names to Kiro model IDs.
// Supports both Kiro and Amazon Q prefixes since they use the same API.
// Agentic variants (-agentic suffix) map to the same backend model IDs.
func (e *KiroExecutor) mapModelToKiro(model string) string {
	modelMap := map[string]string{
		// Amazon Q format (amazonq- prefix) - same API as Kiro
		"amazonq-auto":                       "auto",
		"amazonq-claude-opus-4-5":            "claude-opus-4.5",
		"amazonq-claude-sonnet-4-5":          "claude-sonnet-4.5",
		"amazonq-claude-sonnet-4-5-20250929": "claude-sonnet-4.5",
		"amazonq-claude-sonnet-4":            "claude-sonnet-4",
		"amazonq-claude-sonnet-4-20250514":   "claude-sonnet-4",
		"amazonq-claude-haiku-4-5":           "claude-haiku-4.5",
		// Kiro format (kiro- prefix) - valid model names that should be preserved
		"kiro-claude-opus-4-5":            "claude-opus-4.5",
		"kiro-claude-sonnet-4-5":          "claude-sonnet-4.5",
		"kiro-claude-sonnet-4-5-20250929": "claude-sonnet-4.5",
		"kiro-claude-sonnet-4":            "claude-sonnet-4",
		"kiro-claude-sonnet-4-20250514":   "claude-sonnet-4",
		"kiro-claude-haiku-4-5":           "claude-haiku-4.5",
		"kiro-auto":                       "auto",
		// Native format (no prefix) - used by Kiro IDE directly
		"claude-opus-4-5":            "claude-opus-4.5",
		"claude-opus-4.5":            "claude-opus-4.5",
		"claude-haiku-4-5":           "claude-haiku-4.5",
		"claude-haiku-4.5":           "claude-haiku-4.5",
		"claude-sonnet-4-5":          "claude-sonnet-4.5",
		"claude-sonnet-4-5-20250929": "claude-sonnet-4.5",
		"claude-sonnet-4.5":          "claude-sonnet-4.5",
		"claude-sonnet-4":            "claude-sonnet-4",
		"claude-sonnet-4-20250514":   "claude-sonnet-4",
		"auto":                       "auto",
		// Agentic variants (same backend model IDs, but with special system prompt)
		"claude-opus-4.5-agentic":        "claude-opus-4.5",
		"claude-sonnet-4.5-agentic":      "claude-sonnet-4.5",
		"claude-sonnet-4-agentic":        "claude-sonnet-4",
		"claude-haiku-4.5-agentic":       "claude-haiku-4.5",
		"kiro-claude-opus-4-5-agentic":   "claude-opus-4.5",
		"kiro-claude-sonnet-4-5-agentic": "claude-sonnet-4.5",
		"kiro-claude-sonnet-4-agentic":   "claude-sonnet-4",
		"kiro-claude-haiku-4-5-agentic":  "claude-haiku-4.5",
	}
	if kiroID, ok := modelMap[model]; ok {
		return kiroID
	}

	// Smart fallback: try to infer model type from name patterns
	modelLower := strings.ToLower(model)

	// Check for Haiku variants
	if strings.Contains(modelLower, "haiku") {
		log.Debugf("kiro: unknown Haiku model '%s', mapping to claude-haiku-4.5", model)
		return "claude-haiku-4.5"
	}

	// Check for Sonnet variants
	if strings.Contains(modelLower, "sonnet") {
		// Check for specific version patterns
		if strings.Contains(modelLower, "3-7") || strings.Contains(modelLower, "3.7") {
			log.Debugf("kiro: unknown Sonnet 3.7 model '%s', mapping to claude-3-7-sonnet-20250219", model)
			return "claude-3-7-sonnet-20250219"
		}
		if strings.Contains(modelLower, "4-5") || strings.Contains(modelLower, "4.5") {
			log.Debugf("kiro: unknown Sonnet 4.5 model '%s', mapping to claude-sonnet-4.5", model)
			return "claude-sonnet-4.5"
		}
		// Default to Sonnet 4
		log.Debugf("kiro: unknown Sonnet model '%s', mapping to claude-sonnet-4", model)
		return "claude-sonnet-4"
	}

	// Check for Opus variants
	if strings.Contains(modelLower, "opus") {
		log.Debugf("kiro: unknown Opus model '%s', mapping to claude-opus-4.5", model)
		return "claude-opus-4.5"
	}

	// Final fallback to Sonnet 4.5 (most commonly used model)
	log.Warnf("kiro: unknown model '%s', falling back to claude-sonnet-4.5", model)
	return "claude-sonnet-4.5"
}

// EventStreamError represents an Event Stream processing error
type EventStreamError struct {
	Type    string // "fatal", "malformed"
	Message string
	Cause   error
}

func (e *EventStreamError) Error() string {
	if e.Cause != nil {
		return fmt.Sprintf("event stream %s: %s: %v", e.Type, e.Message, e.Cause)
	}
	return fmt.Sprintf("event stream %s: %s", e.Type, e.Message)
}

// eventStreamMessage represents a parsed AWS Event Stream message
type eventStreamMessage struct {
	EventType string // Event type from headers (e.g., "assistantResponseEvent")
	Payload   []byte // JSON payload of the message
}

// NOTE: Request building functions moved to internal/translator/kiro/claude/kiro_claude_request.go
// The executor now uses kiroclaude.BuildKiroPayload() instead

// parseEventStream parses AWS Event Stream binary format.
// Extracts text content, tool uses, and stop_reason from the response.
// Supports embedded [Called ...] tool calls and input buffering for toolUseEvent.
// Returns: content, toolUses, usageInfo, stopReason, error
func (e *KiroExecutor) parseEventStream(body io.Reader) (string, []kiroclaude.KiroToolUse, usage.Detail, string, error) {
	var content strings.Builder
	var toolUses []kiroclaude.KiroToolUse
	var usageInfo usage.Detail
	var stopReason string // Extracted from upstream response
	reader := bufio.NewReader(body)

	// Tool use state tracking for input buffering and deduplication
	processedIDs := make(map[string]bool)
	var currentToolUse *kiroclaude.ToolUseState

	// Upstream usage tracking - Kiro API returns credit usage and context percentage
	var upstreamContextPercentage float64 // Context usage percentage from upstream (e.g., 78.56)

	for {
		msg, eventErr := e.readEventStreamMessage(reader)
		if eventErr != nil {
			log.Errorf("kiro: parseEventStream error: %v", eventErr)
			return content.String(), toolUses, usageInfo, stopReason, eventErr
		}
		if msg == nil {
			// Normal end of stream (EOF)
			break
		}

		eventType := msg.EventType
		payload := msg.Payload
		if len(payload) == 0 {
			continue
		}

		var event map[string]interface{}
		if err := json.Unmarshal(payload, &event); err != nil {
			log.Debugf("kiro: skipping malformed event: %v", err)
			continue
		}

		// Check for error/exception events in the payload (Kiro API may return errors with HTTP 200)
		// These can appear as top-level fields or nested within the event
		if errType, hasErrType := event["_type"].(string); hasErrType {
			// AWS-style error: {"_type": "com.amazon.aws.codewhisperer#ValidationException", "message": "..."}
			errMsg := ""
			if msg, ok := event["message"].(string); ok {
				errMsg = msg
			}
			log.Errorf("kiro: received AWS error in event stream: type=%s, message=%s", errType, errMsg)
			return "", nil, usageInfo, stopReason, fmt.Errorf("kiro API error: %s - %s", errType, errMsg)
		}
		if errType, hasErrType := event["type"].(string); hasErrType && (errType == "error" || errType == "exception") {
			// Generic error event
			errMsg := ""
			if msg, ok := event["message"].(string); ok {
				errMsg = msg
			} else if errObj, ok := event["error"].(map[string]interface{}); ok {
				if msg, ok := errObj["message"].(string); ok {
					errMsg = msg
				}
			}
			log.Errorf("kiro: received error event in stream: type=%s, message=%s", errType, errMsg)
			return "", nil, usageInfo, stopReason, fmt.Errorf("kiro API error: %s", errMsg)
		}

		// Extract stop_reason from various event formats
		// Kiro/Amazon Q API may include stop_reason in different locations
		if sr := kirocommon.GetString(event, "stop_reason"); sr != "" {
			stopReason = sr
			log.Debugf("kiro: parseEventStream found stop_reason (top-level): %s", stopReason)
		}
		if sr := kirocommon.GetString(event, "stopReason"); sr != "" {
			stopReason = sr
			log.Debugf("kiro: parseEventStream found stopReason (top-level): %s", stopReason)
		}

		// Handle different event types
		switch eventType {
		case "followupPromptEvent":
			// Filter out followupPrompt events - these are UI suggestions, not content
			log.Debugf("kiro: parseEventStream ignoring followupPrompt event")
			continue

		case "assistantResponseEvent":
			if assistantResp, ok := event["assistantResponseEvent"].(map[string]interface{}); ok {
				if contentText, ok := assistantResp["content"].(string); ok {
					content.WriteString(contentText)
				}
				// Extract stop_reason from assistantResponseEvent
				if sr := kirocommon.GetString(assistantResp, "stop_reason"); sr != "" {
					stopReason = sr
					log.Debugf("kiro: parseEventStream found stop_reason in assistantResponseEvent: %s", stopReason)
				}
				if sr := kirocommon.GetString(assistantResp, "stopReason"); sr != "" {
					stopReason = sr
					log.Debugf("kiro: parseEventStream found stopReason in assistantResponseEvent: %s", stopReason)
				}
				// Extract tool uses from response
				if toolUsesRaw, ok := assistantResp["toolUses"].([]interface{}); ok {
					for _, tuRaw := range toolUsesRaw {
						if tu, ok := tuRaw.(map[string]interface{}); ok {
							toolUseID := kirocommon.GetStringValue(tu, "toolUseId")
							// Check for duplicate
							if processedIDs[toolUseID] {
								log.Debugf("kiro: skipping duplicate tool use from assistantResponse: %s", toolUseID)
								continue
							}
							processedIDs[toolUseID] = true

							toolUse := kiroclaude.KiroToolUse{
								ToolUseID: toolUseID,
								Name:      kirocommon.GetStringValue(tu, "name"),
							}
							if input, ok := tu["input"].(map[string]interface{}); ok {
								toolUse.Input = input
							}
							toolUses = append(toolUses, toolUse)
						}
					}
				}
			}
			// Also try direct format
			if contentText, ok := event["content"].(string); ok {
				content.WriteString(contentText)
			}
			// Direct tool uses
			if toolUsesRaw, ok := event["toolUses"].([]interface{}); ok {
				for _, tuRaw := range toolUsesRaw {
					if tu, ok := tuRaw.(map[string]interface{}); ok {
						toolUseID := kirocommon.GetStringValue(tu, "toolUseId")
						// Check for duplicate
						if processedIDs[toolUseID] {
							log.Debugf("kiro: skipping duplicate direct tool use: %s", toolUseID)
							continue
						}
						processedIDs[toolUseID] = true

						toolUse := kiroclaude.KiroToolUse{
							ToolUseID: toolUseID,
							Name:      kirocommon.GetStringValue(tu, "name"),
						}
						if input, ok := tu["input"].(map[string]interface{}); ok {
							toolUse.Input = input
						}
						toolUses = append(toolUses, toolUse)
					}
				}
			}

		case "toolUseEvent":
			// Handle dedicated tool use events with input buffering
			completedToolUses, newState := kiroclaude.ProcessToolUseEvent(event, currentToolUse, processedIDs)
			currentToolUse = newState
			toolUses = append(toolUses, completedToolUses...)

		case "supplementaryWebLinksEvent":
			if inputTokens, ok := event["inputTokens"].(float64); ok {
				usageInfo.InputTokens = int64(inputTokens)
			}
			if outputTokens, ok := event["outputTokens"].(float64); ok {
				usageInfo.OutputTokens = int64(outputTokens)
			}

		case "messageStopEvent", "message_stop":
			// Handle message stop events which may contain stop_reason
			if sr := kirocommon.GetString(event, "stop_reason"); sr != "" {
				stopReason = sr
				log.Debugf("kiro: parseEventStream found stop_reason in messageStopEvent: %s", stopReason)
			}
			if sr := kirocommon.GetString(event, "stopReason"); sr != "" {
				stopReason = sr
				log.Debugf("kiro: parseEventStream found stopReason in messageStopEvent: %s", stopReason)
			}

		case "messageMetadataEvent", "metadataEvent":
			// Handle message metadata events which contain token counts
			// Official format: { tokenUsage: { outputTokens, totalTokens, uncachedInputTokens, cacheReadInputTokens, cacheWriteInputTokens, contextUsagePercentage } }
			var metadata map[string]interface{}
			if m, ok := event["messageMetadataEvent"].(map[string]interface{}); ok {
				metadata = m
			} else if m, ok := event["metadataEvent"].(map[string]interface{}); ok {
				metadata = m
			} else {
				metadata = event // event itself might be the metadata
			}

			// Check for nested tokenUsage object (official format)
			if tokenUsage, ok := metadata["tokenUsage"].(map[string]interface{}); ok {
				// outputTokens - precise output token count
				if outputTokens, ok := tokenUsage["outputTokens"].(float64); ok {
					usageInfo.OutputTokens = int64(outputTokens)
					log.Infof("kiro: parseEventStream found precise outputTokens in tokenUsage: %d", usageInfo.OutputTokens)
				}
				// totalTokens - precise total token count
				if totalTokens, ok := tokenUsage["totalTokens"].(float64); ok {
					usageInfo.TotalTokens = int64(totalTokens)
					log.Infof("kiro: parseEventStream found precise totalTokens in tokenUsage: %d", usageInfo.TotalTokens)
				}
				// uncachedInputTokens - input tokens not from cache
				if uncachedInputTokens, ok := tokenUsage["uncachedInputTokens"].(float64); ok {
					usageInfo.InputTokens = int64(uncachedInputTokens)
					log.Infof("kiro: parseEventStream found uncachedInputTokens in tokenUsage: %d", usageInfo.InputTokens)
				}
				// cacheReadInputTokens - tokens read from cache
				if cacheReadTokens, ok := tokenUsage["cacheReadInputTokens"].(float64); ok {
					// Add to input tokens if we have uncached tokens, otherwise use as input
					if usageInfo.InputTokens > 0 {
						usageInfo.InputTokens += int64(cacheReadTokens)
					} else {
						usageInfo.InputTokens = int64(cacheReadTokens)
					}
					log.Debugf("kiro: parseEventStream found cacheReadInputTokens in tokenUsage: %d", int64(cacheReadTokens))
				}
				// contextUsagePercentage - can be used as fallback for input token estimation
				if ctxPct, ok := tokenUsage["contextUsagePercentage"].(float64); ok {
					upstreamContextPercentage = ctxPct
					log.Debugf("kiro: parseEventStream found contextUsagePercentage in tokenUsage: %.2f%%", ctxPct)
				}
			}

			// Fallback: check for direct fields in metadata (legacy format)
			if usageInfo.InputTokens == 0 {
				if inputTokens, ok := metadata["inputTokens"].(float64); ok {
					usageInfo.InputTokens = int64(inputTokens)
					log.Debugf("kiro: parseEventStream found inputTokens in messageMetadataEvent: %d", usageInfo.InputTokens)
				}
			}
			if usageInfo.OutputTokens == 0 {
				if outputTokens, ok := metadata["outputTokens"].(float64); ok {
					usageInfo.OutputTokens = int64(outputTokens)
					log.Debugf("kiro: parseEventStream found outputTokens in messageMetadataEvent: %d", usageInfo.OutputTokens)
				}
			}
			if usageInfo.TotalTokens == 0 {
				if totalTokens, ok := metadata["totalTokens"].(float64); ok {
					usageInfo.TotalTokens = int64(totalTokens)
					log.Debugf("kiro: parseEventStream found totalTokens in messageMetadataEvent: %d", usageInfo.TotalTokens)
				}
			}

		case "usageEvent", "usage":
			// Handle dedicated usage events
			if inputTokens, ok := event["inputTokens"].(float64); ok {
				usageInfo.InputTokens = int64(inputTokens)
				log.Debugf("kiro: parseEventStream found inputTokens in usageEvent: %d", usageInfo.InputTokens)
			}
			if outputTokens, ok := event["outputTokens"].(float64); ok {
				usageInfo.OutputTokens = int64(outputTokens)
				log.Debugf("kiro: parseEventStream found outputTokens in usageEvent: %d", usageInfo.OutputTokens)
			}
			if totalTokens, ok := event["totalTokens"].(float64); ok {
				usageInfo.TotalTokens = int64(totalTokens)
				log.Debugf("kiro: parseEventStream found totalTokens in usageEvent: %d", usageInfo.TotalTokens)
			}
			// Also check nested usage object
			if usageObj, ok := event["usage"].(map[string]interface{}); ok {
				if inputTokens, ok := usageObj["input_tokens"].(float64); ok {
					usageInfo.InputTokens = int64(inputTokens)
				} else if inputTokens, ok := usageObj["prompt_tokens"].(float64); ok {
					usageInfo.InputTokens = int64(inputTokens)
				}
				if outputTokens, ok := usageObj["output_tokens"].(float64); ok {
					usageInfo.OutputTokens = int64(outputTokens)
				} else if outputTokens, ok := usageObj["completion_tokens"].(float64); ok {
					usageInfo.OutputTokens = int64(outputTokens)
				}
				if totalTokens, ok := usageObj["total_tokens"].(float64); ok {
					usageInfo.TotalTokens = int64(totalTokens)
				}
				log.Debugf("kiro: parseEventStream found usage object: input=%d, output=%d, total=%d",
					usageInfo.InputTokens, usageInfo.OutputTokens, usageInfo.TotalTokens)
			}

		case "metricsEvent":
			// Handle metrics events which may contain usage data
			if metrics, ok := event["metricsEvent"].(map[string]interface{}); ok {
				if inputTokens, ok := metrics["inputTokens"].(float64); ok {
					usageInfo.InputTokens = int64(inputTokens)
				}
				if outputTokens, ok := metrics["outputTokens"].(float64); ok {
					usageInfo.OutputTokens = int64(outputTokens)
				}
				log.Debugf("kiro: parseEventStream found metricsEvent: input=%d, output=%d",
					usageInfo.InputTokens, usageInfo.OutputTokens)
			}

		case "meteringEvent":
			// Handle metering events from Kiro API (usage billing information)
			// Official format: { unit: string, unitPlural: string, usage: number }
			if metering, ok := event["meteringEvent"].(map[string]interface{}); ok {
				unit := ""
				if u, ok := metering["unit"].(string); ok {
					unit = u
				}
				usageVal := 0.0
				if u, ok := metering["usage"].(float64); ok {
					usageVal = u
				}
				log.Infof("kiro: parseEventStream received meteringEvent: usage=%.2f %s", usageVal, unit)
				// Store metering info for potential billing/statistics purposes
				// Note: This is separate from token counts - it's AWS billing units
			} else {
				// Try direct fields
				unit := ""
				if u, ok := event["unit"].(string); ok {
					unit = u
				}
				usageVal := 0.0
				if u, ok := event["usage"].(float64); ok {
					usageVal = u
				}
				if unit != "" || usageVal > 0 {
					log.Infof("kiro: parseEventStream received meteringEvent (direct): usage=%.2f %s", usageVal, unit)
				}
			}

		case "error", "exception", "internalServerException", "invalidStateEvent":
			// Handle error events from Kiro API stream
			errMsg := ""
			errType := eventType

			// Try to extract error message from various formats
			if msg, ok := event["message"].(string); ok {
				errMsg = msg
			} else if errObj, ok := event[eventType].(map[string]interface{}); ok {
				if msg, ok := errObj["message"].(string); ok {
					errMsg = msg
				}
				if t, ok := errObj["type"].(string); ok {
					errType = t
				}
			} else if errObj, ok := event["error"].(map[string]interface{}); ok {
				if msg, ok := errObj["message"].(string); ok {
					errMsg = msg
				}
				if t, ok := errObj["type"].(string); ok {
					errType = t
				}
			}

			// Check for specific error reasons
			if reason, ok := event["reason"].(string); ok {
				errMsg = fmt.Sprintf("%s (reason: %s)", errMsg, reason)
			}

			log.Errorf("kiro: parseEventStream received error event: type=%s, message=%s", errType, errMsg)

			// For invalidStateEvent, we may want to continue processing other events
			if eventType == "invalidStateEvent" {
				log.Warnf("kiro: invalidStateEvent received, continuing stream processing")
				continue
			}

			// For other errors, return the error
			if errMsg != "" {
				return "", nil, usageInfo, stopReason, fmt.Errorf("kiro API error (%s): %s", errType, errMsg)
			}

		default:
			// Check for contextUsagePercentage in any event
			if ctxPct, ok := event["contextUsagePercentage"].(float64); ok {
				upstreamContextPercentage = ctxPct
				log.Debugf("kiro: parseEventStream received context usage: %.2f%%", upstreamContextPercentage)
			}
			// Log unknown event types for debugging (to discover new event formats)
			log.Debugf("kiro: parseEventStream unknown event type: %s, payload: %s", eventType, string(payload))
		}

		// Check for direct token fields in any event (fallback)
		if usageInfo.InputTokens == 0 {
			if inputTokens, ok := event["inputTokens"].(float64); ok {
				usageInfo.InputTokens = int64(inputTokens)
				log.Debugf("kiro: parseEventStream found direct inputTokens: %d", usageInfo.InputTokens)
			}
		}
		if usageInfo.OutputTokens == 0 {
			if outputTokens, ok := event["outputTokens"].(float64); ok {
				usageInfo.OutputTokens = int64(outputTokens)
				log.Debugf("kiro: parseEventStream found direct outputTokens: %d", usageInfo.OutputTokens)
			}
		}

		// Check for usage object in any event (OpenAI format)
		if usageInfo.InputTokens == 0 || usageInfo.OutputTokens == 0 {
			if usageObj, ok := event["usage"].(map[string]interface{}); ok {
				if usageInfo.InputTokens == 0 {
					if inputTokens, ok := usageObj["input_tokens"].(float64); ok {
						usageInfo.InputTokens = int64(inputTokens)
					} else if inputTokens, ok := usageObj["prompt_tokens"].(float64); ok {
						usageInfo.InputTokens = int64(inputTokens)
					}
				}
				if usageInfo.OutputTokens == 0 {
					if outputTokens, ok := usageObj["output_tokens"].(float64); ok {
						usageInfo.OutputTokens = int64(outputTokens)
					} else if outputTokens, ok := usageObj["completion_tokens"].(float64); ok {
						usageInfo.OutputTokens = int64(outputTokens)
					}
				}
				if usageInfo.TotalTokens == 0 {
					if totalTokens, ok := usageObj["total_tokens"].(float64); ok {
						usageInfo.TotalTokens = int64(totalTokens)
					}
				}
				log.Debugf("kiro: parseEventStream found usage object (fallback): input=%d, output=%d, total=%d",
					usageInfo.InputTokens, usageInfo.OutputTokens, usageInfo.TotalTokens)
			}
		}

		// Also check nested supplementaryWebLinksEvent
		if usageEvent, ok := event["supplementaryWebLinksEvent"].(map[string]interface{}); ok {
			if inputTokens, ok := usageEvent["inputTokens"].(float64); ok {
				usageInfo.InputTokens = int64(inputTokens)
			}
			if outputTokens, ok := usageEvent["outputTokens"].(float64); ok {
				usageInfo.OutputTokens = int64(outputTokens)
			}
		}
	}

	// Parse embedded tool calls from content (e.g., [Called tool_name with args: {...}])
	contentStr := content.String()
	cleanedContent, embeddedToolUses := kiroclaude.ParseEmbeddedToolCalls(contentStr, processedIDs)
	toolUses = append(toolUses, embeddedToolUses...)

	// Deduplicate all tool uses
	toolUses = kiroclaude.DeduplicateToolUses(toolUses)

	// Apply fallback logic for stop_reason if not provided by upstream
	// Priority: upstream stopReason > tool_use detection > end_turn default
	if stopReason == "" {
		if len(toolUses) > 0 {
			stopReason = "tool_use"
			log.Debugf("kiro: parseEventStream using fallback stop_reason: tool_use (detected %d tool uses)", len(toolUses))
		} else {
			stopReason = "end_turn"
			log.Debugf("kiro: parseEventStream using fallback stop_reason: end_turn")
		}
	}

	// Log warning if response was truncated due to max_tokens
	if stopReason == "max_tokens" {
		log.Warnf("kiro: response truncated due to max_tokens limit")
	}

	// Use contextUsagePercentage to calculate more accurate input tokens
	// Kiro model has 200k max context, contextUsagePercentage represents the percentage used
	// Formula: input_tokens = contextUsagePercentage * 200000 / 100
	if upstreamContextPercentage > 0 {
		calculatedInputTokens := int64(upstreamContextPercentage * 200000 / 100)
		if calculatedInputTokens > 0 {
			localEstimate := usageInfo.InputTokens
			usageInfo.InputTokens = calculatedInputTokens
			usageInfo.TotalTokens = usageInfo.InputTokens + usageInfo.OutputTokens
			log.Infof("kiro: parseEventStream using contextUsagePercentage (%.2f%%) to calculate input tokens: %d (local estimate was: %d)",
				upstreamContextPercentage, calculatedInputTokens, localEstimate)
		}
	}

	return cleanedContent, toolUses, usageInfo, stopReason, nil
}

// readEventStreamMessage reads and validates a single AWS Event Stream message.
// Returns the parsed message or a structured error for different failure modes.
// This function implements boundary protection and detailed error classification.
//
// AWS Event Stream binary format:
// - Prelude (12 bytes): total_length (4) + headers_length (4) + prelude_crc (4)
// - Headers (variable): header entries
// - Payload (variable): JSON data
// - Message CRC (4 bytes): CRC32C of entire message (not validated, just skipped)
func (e *KiroExecutor) readEventStreamMessage(reader *bufio.Reader) (*eventStreamMessage, *EventStreamError) {
	// Read prelude (first 12 bytes: total_len + headers_len + prelude_crc)
	prelude := make([]byte, 12)
	_, err := io.ReadFull(reader, prelude)
	if err == io.EOF {
		return nil, nil // Normal end of stream
	}
	if err != nil {
		return nil, &EventStreamError{
			Type:    ErrStreamFatal,
			Message: "failed to read prelude",
			Cause:   err,
		}
	}

	totalLength := binary.BigEndian.Uint32(prelude[0:4])
	headersLength := binary.BigEndian.Uint32(prelude[4:8])
	// Note: prelude[8:12] is prelude_crc - we read it but don't validate (no CRC check per requirements)

	// Boundary check: minimum frame size
	if totalLength < minEventStreamFrameSize {
		return nil, &EventStreamError{
			Type:    ErrStreamMalformed,
			Message: fmt.Sprintf("invalid message length: %d (minimum is %d)", totalLength, minEventStreamFrameSize),
		}
	}

	// Boundary check: maximum message size
	if totalLength > maxEventStreamMsgSize {
		return nil, &EventStreamError{
			Type:    ErrStreamMalformed,
			Message: fmt.Sprintf("message too large: %d bytes (maximum is %d)", totalLength, maxEventStreamMsgSize),
		}
	}

	// Boundary check: headers length within message bounds
	// Message structure: prelude(12) + headers(headersLength) + payload + message_crc(4)
	// So: headersLength must be <= totalLength - 16 (12 for prelude + 4 for message_crc)
	if headersLength > totalLength-16 {
		return nil, &EventStreamError{
			Type:    ErrStreamMalformed,
			Message: fmt.Sprintf("headers length %d exceeds message bounds (total: %d)", headersLength, totalLength),
		}
	}

	// Read the rest of the message (total - 12 bytes already read)
	remaining := make([]byte, totalLength-12)
	_, err = io.ReadFull(reader, remaining)
	if err != nil {
		return nil, &EventStreamError{
			Type:    ErrStreamFatal,
			Message: "failed to read message body",
			Cause:   err,
		}
	}

	// Extract event type from headers
	// Headers start at beginning of 'remaining', length is headersLength
	var eventType string
	if headersLength > 0 && headersLength <= uint32(len(remaining)) {
		eventType = e.extractEventTypeFromBytes(remaining[:headersLength])
	}

	// Calculate payload boundaries
	// Payload starts after headers, ends before message_crc (last 4 bytes)
	payloadStart := headersLength
	payloadEnd := uint32(len(remaining)) - 4 // Skip message_crc at end

	// Validate payload boundaries
	if payloadStart >= payloadEnd {
		// No payload, return empty message
		return &eventStreamMessage{
			EventType: eventType,
			Payload:   nil,
		}, nil
	}

	payload := remaining[payloadStart:payloadEnd]

	return &eventStreamMessage{
		EventType: eventType,
		Payload:   payload,
	}, nil
}

func skipEventStreamHeaderValue(headers []byte, offset int, valueType byte) (int, bool) {
	switch valueType {
	case 0, 1: // bool true / bool false
		return offset, true
	case 2: // byte
		if offset+1 > len(headers) {
			return offset, false
		}
		return offset + 1, true
	case 3: // short
		if offset+2 > len(headers) {
			return offset, false
		}
		return offset + 2, true
	case 4: // int
		if offset+4 > len(headers) {
			return offset, false
		}
		return offset + 4, true
	case 5: // long
		if offset+8 > len(headers) {
			return offset, false
		}
		return offset + 8, true
	case 6: // byte array (2-byte length + data)
		if offset+2 > len(headers) {
			return offset, false
		}
		valueLen := int(binary.BigEndian.Uint16(headers[offset : offset+2]))
		offset += 2
		if offset+valueLen > len(headers) {
			return offset, false
		}
		return offset + valueLen, true
	case 8: // timestamp
		if offset+8 > len(headers) {
			return offset, false
		}
		return offset + 8, true
	case 9: // uuid
		if offset+16 > len(headers) {
			return offset, false
		}
		return offset + 16, true
	default:
		return offset, false
	}
}

// extractEventTypeFromBytes extracts the event type from raw header bytes (without prelude CRC prefix)
func (e *KiroExecutor) extractEventTypeFromBytes(headers []byte) string {
	offset := 0
	for offset < len(headers) {
		nameLen := int(headers[offset])
		offset++
		if offset+nameLen > len(headers) {
			break
		}
		name := string(headers[offset : offset+nameLen])
		offset += nameLen

		if offset >= len(headers) {
			break
		}
		valueType := headers[offset]
		offset++

		if valueType == 7 { // String type
			if offset+2 > len(headers) {
				break
			}
			valueLen := int(binary.BigEndian.Uint16(headers[offset : offset+2]))
			offset += 2
			if offset+valueLen > len(headers) {
				break
			}
			value := string(headers[offset : offset+valueLen])
			offset += valueLen

			if name == ":event-type" {
				return value
			}
			continue
		}

		nextOffset, ok := skipEventStreamHeaderValue(headers, offset, valueType)
		if !ok {
			break
		}
		offset = nextOffset
	}
	return ""
}

// NOTE: Response building functions moved to internal/translator/kiro/claude/kiro_claude_response.go
// The executor now uses kiroclaude.BuildClaudeResponse() and kiroclaude.ExtractThinkingFromContent() instead

// streamToChannel converts AWS Event Stream to channel-based streaming.
// Supports tool calling - emits tool_use content blocks when tools are used.
// Includes embedded [Called ...] tool call parsing and input buffering for toolUseEvent.
// Implements duplicate content filtering using lastContentEvent detection (based on AIClient-2-API).
// Extracts stop_reason from upstream events when available.
// thinkingEnabled controls whether <thinking> tags are parsed - only parse when request enabled thinking.
func (e *KiroExecutor) streamToChannel(ctx context.Context, body io.Reader, out chan<- cliproxyexecutor.StreamChunk, targetFormat sdktranslator.Format, model string, originalReq, claudeBody []byte, reporter *usageReporter, thinkingEnabled bool) {
	reader := bufio.NewReaderSize(body, 20*1024*1024) // 20MB buffer to match other providers
	var totalUsage usage.Detail
	var hasToolUses bool          // Track if any tool uses were emitted
	var upstreamStopReason string // Track stop_reason from upstream events

	// Tool use state tracking for input buffering and deduplication
	processedIDs := make(map[string]bool)
	var currentToolUse *kiroclaude.ToolUseState

	// NOTE: Duplicate content filtering removed - it was causing legitimate repeated
	// content (like consecutive newlines) to be incorrectly filtered out.
	// The previous implementation compared lastContentEvent == contentDelta which
	// is too aggressive for streaming scenarios.

	// Streaming token calculation - accumulate content for real-time token counting
	// Based on AIClient-2-API implementation
	var accumulatedContent strings.Builder
	accumulatedContent.Grow(4096) // Pre-allocate 4KB capacity to reduce reallocations

	// Real-time usage estimation state
	// These track when to send periodic usage updates during streaming
	var lastUsageUpdateLen int           // Last accumulated content length when usage was sent
	var lastUsageUpdateTime = time.Now() // Last time usage update was sent
	var lastReportedOutputTokens int64   // Last reported output token count

	// Upstream usage tracking - Kiro API returns credit usage and context percentage
	var upstreamCreditUsage float64       // Credit usage from upstream (e.g., 1.458)
	var upstreamContextPercentage float64 // Context usage percentage from upstream (e.g., 78.56)
	var hasUpstreamUsage bool             // Whether we received usage from upstream

	// Translator param for maintaining tool call state across streaming events
	// IMPORTANT: This must persist across all TranslateStream calls
	var translatorParam any

	// Thinking mode state tracking - tag-based parsing for <thinking> tags in content
	inThinkBlock := false                          // Whether we're currently inside a <thinking> block
	isThinkingBlockOpen := false                   // Track if thinking content block SSE event is open
	thinkingBlockIndex := -1                       // Index of the thinking content block
	var accumulatedThinkingContent strings.Builder // Accumulate thinking content for token counting

	// Buffer for handling partial tag matches at chunk boundaries
	var pendingContent strings.Builder // Buffer content that might be part of a tag

	// Pre-calculate input tokens from request if possible
	// Kiro uses Claude format, so try Claude format first, then OpenAI format, then fallback
	if enc, err := getTokenizer(model); err == nil {
		var inputTokens int64
		var countMethod string

		// Try Claude format first (Kiro uses Claude API format)
		if inp, err := countClaudeChatTokens(enc, claudeBody); err == nil && inp > 0 {
			inputTokens = inp
			countMethod = "claude"
		} else if inp, err := countOpenAIChatTokens(enc, originalReq); err == nil && inp > 0 {
			// Fallback to OpenAI format (for OpenAI-compatible requests)
			inputTokens = inp
			countMethod = "openai"
		} else {
			// Final fallback: estimate from raw request size (roughly 4 chars per token)
			inputTokens = int64(len(claudeBody) / 4)
			if inputTokens == 0 && len(claudeBody) > 0 {
				inputTokens = 1
			}
			countMethod = "estimate"
		}

		totalUsage.InputTokens = inputTokens
		log.Debugf("kiro: streamToChannel pre-calculated input tokens: %d (method: %s, claude body: %d bytes, original req: %d bytes)",
			totalUsage.InputTokens, countMethod, len(claudeBody), len(originalReq))
	}

	contentBlockIndex := -1
	messageStartSent := false
	isTextBlockOpen := false
	var outputLen int

	// Ensure usage is published even on early return
	defer func() {
		reporter.publish(ctx, totalUsage)
	}()

	for {
		select {
		case <-ctx.Done():
			return
		default:
		}

		msg, eventErr := e.readEventStreamMessage(reader)
		if eventErr != nil {
			// Log the error
			log.Errorf("kiro: streamToChannel error: %v", eventErr)

			// Send error to channel for client notification
			out <- cliproxyexecutor.StreamChunk{Err: eventErr}
			return
		}
		if msg == nil {
			// Normal end of stream (EOF)
			// Flush any incomplete tool use before ending stream
			if currentToolUse != nil && !processedIDs[currentToolUse.ToolUseID] {
				log.Warnf("kiro: flushing incomplete tool use at EOF: %s (ID: %s)", currentToolUse.Name, currentToolUse.ToolUseID)
				fullInput := currentToolUse.InputBuffer.String()
				repairedJSON := kiroclaude.RepairJSON(fullInput)
				var finalInput map[string]interface{}
				if err := json.Unmarshal([]byte(repairedJSON), &finalInput); err != nil {
					log.Warnf("kiro: failed to parse incomplete tool input at EOF: %v", err)
					finalInput = make(map[string]interface{})
				}

				processedIDs[currentToolUse.ToolUseID] = true
				contentBlockIndex++

				// Send tool_use content block
				blockStart := kiroclaude.BuildClaudeContentBlockStartEvent(contentBlockIndex, "tool_use", currentToolUse.ToolUseID, currentToolUse.Name)
				sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStart, &translatorParam)
				for _, chunk := range sseData {
					if chunk != "" {
						out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
					}
				}

				// Send tool input as delta
				inputBytes, _ := json.Marshal(finalInput)
				inputDelta := kiroclaude.BuildClaudeInputJsonDeltaEvent(string(inputBytes), contentBlockIndex)
				sseData = sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, inputDelta, &translatorParam)
				for _, chunk := range sseData {
					if chunk != "" {
						out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
					}
				}

				// Close block
				blockStop := kiroclaude.BuildClaudeContentBlockStopEvent(contentBlockIndex)
				sseData = sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
				for _, chunk := range sseData {
					if chunk != "" {
						out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
					}
				}

				hasToolUses = true
				currentToolUse = nil
			}

			// DISABLED: Tag-based pending character flushing
			// This code block was used for tag-based thinking detection which has been
			// replaced by reasoningContentEvent handling. No pending tag chars to flush.
			// Original code preserved in git history.
			break
		}

		eventType := msg.EventType
		payload := msg.Payload
		if len(payload) == 0 {
			continue
		}
		appendAPIResponseChunk(ctx, e.cfg, payload)

		var event map[string]interface{}
		if err := json.Unmarshal(payload, &event); err != nil {
			log.Warnf("kiro: failed to unmarshal event payload: %v, raw: %s", err, string(payload))
			continue
		}

		// Check for error/exception events in the payload (Kiro API may return errors with HTTP 200)
		// These can appear as top-level fields or nested within the event
		if errType, hasErrType := event["_type"].(string); hasErrType {
			// AWS-style error: {"_type": "com.amazon.aws.codewhisperer#ValidationException", "message": "..."}
			errMsg := ""
			if msg, ok := event["message"].(string); ok {
				errMsg = msg
			}
			log.Errorf("kiro: received AWS error in stream: type=%s, message=%s", errType, errMsg)
			out <- cliproxyexecutor.StreamChunk{Err: fmt.Errorf("kiro API error: %s - %s", errType, errMsg)}
			return
		}
		if errType, hasErrType := event["type"].(string); hasErrType && (errType == "error" || errType == "exception") {
			// Generic error event
			errMsg := ""
			if msg, ok := event["message"].(string); ok {
				errMsg = msg
			} else if errObj, ok := event["error"].(map[string]interface{}); ok {
				if msg, ok := errObj["message"].(string); ok {
					errMsg = msg
				}
			}
			log.Errorf("kiro: received error event in stream: type=%s, message=%s", errType, errMsg)
			out <- cliproxyexecutor.StreamChunk{Err: fmt.Errorf("kiro API error: %s", errMsg)}
			return
		}

		// Extract stop_reason from various event formats (streaming)
		// Kiro/Amazon Q API may include stop_reason in different locations
		if sr := kirocommon.GetString(event, "stop_reason"); sr != "" {
			upstreamStopReason = sr
			log.Debugf("kiro: streamToChannel found stop_reason (top-level): %s", upstreamStopReason)
		}
		if sr := kirocommon.GetString(event, "stopReason"); sr != "" {
			upstreamStopReason = sr
			log.Debugf("kiro: streamToChannel found stopReason (top-level): %s", upstreamStopReason)
		}

		// Send message_start on first event
		if !messageStartSent {
			msgStart := kiroclaude.BuildClaudeMessageStartEvent(model, totalUsage.InputTokens)
			sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, msgStart, &translatorParam)
			for _, chunk := range sseData {
				if chunk != "" {
					out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
				}
			}
			messageStartSent = true
		}

		switch eventType {
		case "followupPromptEvent":
			// Filter out followupPrompt events - these are UI suggestions, not content
			log.Debugf("kiro: streamToChannel ignoring followupPrompt event")
			continue

		case "messageStopEvent", "message_stop":
			// Handle message stop events which may contain stop_reason
			if sr := kirocommon.GetString(event, "stop_reason"); sr != "" {
				upstreamStopReason = sr
				log.Debugf("kiro: streamToChannel found stop_reason in messageStopEvent: %s", upstreamStopReason)
			}
			if sr := kirocommon.GetString(event, "stopReason"); sr != "" {
				upstreamStopReason = sr
				log.Debugf("kiro: streamToChannel found stopReason in messageStopEvent: %s", upstreamStopReason)
			}

		case "meteringEvent":
			// Handle metering events from Kiro API (usage billing information)
			// Official format: { unit: string, unitPlural: string, usage: number }
			if metering, ok := event["meteringEvent"].(map[string]interface{}); ok {
				unit := ""
				if u, ok := metering["unit"].(string); ok {
					unit = u
				}
				usageVal := 0.0
				if u, ok := metering["usage"].(float64); ok {
					usageVal = u
				}
				upstreamCreditUsage = usageVal
				hasUpstreamUsage = true
				log.Infof("kiro: streamToChannel received meteringEvent: usage=%.4f %s", usageVal, unit)
			} else {
				// Try direct fields (event is meteringEvent itself)
				if unit, ok := event["unit"].(string); ok {
					if usage, ok := event["usage"].(float64); ok {
						upstreamCreditUsage = usage
						hasUpstreamUsage = true
						log.Infof("kiro: streamToChannel received meteringEvent (direct): usage=%.4f %s", usage, unit)
					}
				}
			}

		case "error", "exception", "internalServerException":
			// Handle error events from Kiro API stream
			errMsg := ""
			errType := eventType

			// Try to extract error message from various formats
			if msg, ok := event["message"].(string); ok {
				errMsg = msg
			} else if errObj, ok := event[eventType].(map[string]interface{}); ok {
				if msg, ok := errObj["message"].(string); ok {
					errMsg = msg
				}
				if t, ok := errObj["type"].(string); ok {
					errType = t
				}
			} else if errObj, ok := event["error"].(map[string]interface{}); ok {
				if msg, ok := errObj["message"].(string); ok {
					errMsg = msg
				}
			}

			log.Errorf("kiro: streamToChannel received error event: type=%s, message=%s", errType, errMsg)

			// Send error to the stream and exit
			if errMsg != "" {
				out <- cliproxyexecutor.StreamChunk{
					Err: fmt.Errorf("kiro API error (%s): %s", errType, errMsg),
				}
				return
			}

		case "invalidStateEvent":
			// Handle invalid state events - log and continue (non-fatal)
			errMsg := ""
			if msg, ok := event["message"].(string); ok {
				errMsg = msg
			} else if stateEvent, ok := event["invalidStateEvent"].(map[string]interface{}); ok {
				if msg, ok := stateEvent["message"].(string); ok {
					errMsg = msg
				}
			}
			log.Warnf("kiro: streamToChannel received invalidStateEvent: %s, continuing", errMsg)
			continue

		default:
			// Check for upstream usage events from Kiro API
			// Format: {"unit":"credit","unitPlural":"credits","usage":1.458}
			if unit, ok := event["unit"].(string); ok && unit == "credit" {
				if usage, ok := event["usage"].(float64); ok {
					upstreamCreditUsage = usage
					hasUpstreamUsage = true
					log.Debugf("kiro: received upstream credit usage: %.4f", upstreamCreditUsage)
				}
			}
			// Format: {"contextUsagePercentage":78.56}
			if ctxPct, ok := event["contextUsagePercentage"].(float64); ok {
				upstreamContextPercentage = ctxPct
				log.Debugf("kiro: received upstream context usage: %.2f%%", upstreamContextPercentage)
			}

			// Check for token counts in unknown events
			if inputTokens, ok := event["inputTokens"].(float64); ok {
				totalUsage.InputTokens = int64(inputTokens)
				hasUpstreamUsage = true
				log.Debugf("kiro: streamToChannel found inputTokens in event %s: %d", eventType, totalUsage.InputTokens)
			}
			if outputTokens, ok := event["outputTokens"].(float64); ok {
				totalUsage.OutputTokens = int64(outputTokens)
				hasUpstreamUsage = true
				log.Debugf("kiro: streamToChannel found outputTokens in event %s: %d", eventType, totalUsage.OutputTokens)
			}
			if totalTokens, ok := event["totalTokens"].(float64); ok {
				totalUsage.TotalTokens = int64(totalTokens)
				log.Debugf("kiro: streamToChannel found totalTokens in event %s: %d", eventType, totalUsage.TotalTokens)
			}

			// Check for usage object in unknown events (OpenAI/Claude format)
			if usageObj, ok := event["usage"].(map[string]interface{}); ok {
				if inputTokens, ok := usageObj["input_tokens"].(float64); ok {
					totalUsage.InputTokens = int64(inputTokens)
					hasUpstreamUsage = true
				} else if inputTokens, ok := usageObj["prompt_tokens"].(float64); ok {
					totalUsage.InputTokens = int64(inputTokens)
					hasUpstreamUsage = true
				}
				if outputTokens, ok := usageObj["output_tokens"].(float64); ok {
					totalUsage.OutputTokens = int64(outputTokens)
					hasUpstreamUsage = true
				} else if outputTokens, ok := usageObj["completion_tokens"].(float64); ok {
					totalUsage.OutputTokens = int64(outputTokens)
					hasUpstreamUsage = true
				}
				if totalTokens, ok := usageObj["total_tokens"].(float64); ok {
					totalUsage.TotalTokens = int64(totalTokens)
				}
				log.Debugf("kiro: streamToChannel found usage object in event %s: input=%d, output=%d, total=%d",
					eventType, totalUsage.InputTokens, totalUsage.OutputTokens, totalUsage.TotalTokens)
			}

			// Log unknown event types for debugging (to discover new event formats)
			if eventType != "" {
				log.Debugf("kiro: streamToChannel unknown event type: %s, payload: %s", eventType, string(payload))
			}

		case "assistantResponseEvent":
			var contentDelta string
			var toolUses []map[string]interface{}

			if assistantResp, ok := event["assistantResponseEvent"].(map[string]interface{}); ok {
				if c, ok := assistantResp["content"].(string); ok {
					contentDelta = c
				}
				// Extract stop_reason from assistantResponseEvent
				if sr := kirocommon.GetString(assistantResp, "stop_reason"); sr != "" {
					upstreamStopReason = sr
					log.Debugf("kiro: streamToChannel found stop_reason in assistantResponseEvent: %s", upstreamStopReason)
				}
				if sr := kirocommon.GetString(assistantResp, "stopReason"); sr != "" {
					upstreamStopReason = sr
					log.Debugf("kiro: streamToChannel found stopReason in assistantResponseEvent: %s", upstreamStopReason)
				}
				// Extract tool uses from response
				if tus, ok := assistantResp["toolUses"].([]interface{}); ok {
					for _, tuRaw := range tus {
						if tu, ok := tuRaw.(map[string]interface{}); ok {
							toolUses = append(toolUses, tu)
						}
					}
				}
			}
			if contentDelta == "" {
				if c, ok := event["content"].(string); ok {
					contentDelta = c
				}
			}
			// Direct tool uses
			if tus, ok := event["toolUses"].([]interface{}); ok {
				for _, tuRaw := range tus {
					if tu, ok := tuRaw.(map[string]interface{}); ok {
						toolUses = append(toolUses, tu)
					}
				}
			}

			// Handle text content with thinking mode support
			if contentDelta != "" {
				// NOTE: Duplicate content filtering was removed because it incorrectly
				// filtered out legitimate repeated content (like consecutive newlines "\n\n").
				// Streaming naturally can have identical chunks that are valid content.

				outputLen += len(contentDelta)
				// Accumulate content for streaming token calculation
				accumulatedContent.WriteString(contentDelta)

				// Real-time usage estimation: Check if we should send a usage update
				// This helps clients track context usage during long thinking sessions
				shouldSendUsageUpdate := false
				if accumulatedContent.Len()-lastUsageUpdateLen >= usageUpdateCharThreshold {
					shouldSendUsageUpdate = true
				} else if time.Since(lastUsageUpdateTime) >= usageUpdateTimeInterval && accumulatedContent.Len() > lastUsageUpdateLen {
					shouldSendUsageUpdate = true
				}

				if shouldSendUsageUpdate {
					// Calculate current output tokens using tiktoken
					var currentOutputTokens int64
					if enc, encErr := getTokenizer(model); encErr == nil {
						if tokenCount, countErr := enc.Count(accumulatedContent.String()); countErr == nil {
							currentOutputTokens = int64(tokenCount)
						}
					}
					// Fallback to character estimation if tiktoken fails
					if currentOutputTokens == 0 {
						currentOutputTokens = int64(accumulatedContent.Len() / 4)
						if currentOutputTokens == 0 {
							currentOutputTokens = 1
						}
					}

					// Only send update if token count has changed significantly (at least 10 tokens)
					if currentOutputTokens > lastReportedOutputTokens+10 {
						// Send ping event with usage information
						// This is a non-blocking update that clients can optionally process
						pingEvent := kiroclaude.BuildClaudePingEventWithUsage(totalUsage.InputTokens, currentOutputTokens)
						sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, pingEvent, &translatorParam)
						for _, chunk := range sseData {
							if chunk != "" {
								out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
							}
						}

						lastReportedOutputTokens = currentOutputTokens
						log.Debugf("kiro: sent real-time usage update - input: %d, output: %d (accumulated: %d chars)",
							totalUsage.InputTokens, currentOutputTokens, accumulatedContent.Len())
					}

					lastUsageUpdateLen = accumulatedContent.Len()
					lastUsageUpdateTime = time.Now()
				}

				// TAG-BASED THINKING PARSING: Parse <thinking> tags from content
				// Combine pending content with new content for processing
				pendingContent.WriteString(contentDelta)
				processContent := pendingContent.String()
				pendingContent.Reset()

				// Process content looking for thinking tags
				for len(processContent) > 0 {
					if inThinkBlock {
						// We're inside a thinking block, look for </thinking>
						endIdx := strings.Index(processContent, kirocommon.ThinkingEndTag)
						if endIdx >= 0 {
							// Found end tag - emit thinking content before the tag
							thinkingText := processContent[:endIdx]
							if thinkingText != "" {
								// Ensure thinking block is open
								if !isThinkingBlockOpen {
									contentBlockIndex++
									thinkingBlockIndex = contentBlockIndex
									isThinkingBlockOpen = true
									blockStart := kiroclaude.BuildClaudeContentBlockStartEvent(thinkingBlockIndex, "thinking", "", "")
									sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStart, &translatorParam)
									for _, chunk := range sseData {
										if chunk != "" {
											out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
										}
									}
								}
								// Send thinking delta
								thinkingEvent := kiroclaude.BuildClaudeThinkingDeltaEvent(thinkingText, thinkingBlockIndex)
								sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, thinkingEvent, &translatorParam)
								for _, chunk := range sseData {
									if chunk != "" {
										out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
									}
								}
								accumulatedThinkingContent.WriteString(thinkingText)
							}
							// Close thinking block
							if isThinkingBlockOpen {
								blockStop := kiroclaude.BuildClaudeThinkingBlockStopEvent(thinkingBlockIndex)
								sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
								for _, chunk := range sseData {
									if chunk != "" {
										out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
									}
								}
								isThinkingBlockOpen = false
							}
							inThinkBlock = false
							processContent = processContent[endIdx+len(kirocommon.ThinkingEndTag):]
							log.Debugf("kiro: closed thinking block, remaining content: %d chars", len(processContent))
						} else {
							// No end tag found - check for partial match at end
							partialMatch := false
							for i := 1; i < len(kirocommon.ThinkingEndTag) && i <= len(processContent); i++ {
								if strings.HasSuffix(processContent, kirocommon.ThinkingEndTag[:i]) {
									// Possible partial tag at end, buffer it
									pendingContent.WriteString(processContent[len(processContent)-i:])
									processContent = processContent[:len(processContent)-i]
									partialMatch = true
									break
								}
							}
							if !partialMatch || len(processContent) > 0 {
								// Emit all as thinking content
								if processContent != "" {
									if !isThinkingBlockOpen {
										contentBlockIndex++
										thinkingBlockIndex = contentBlockIndex
										isThinkingBlockOpen = true
										blockStart := kiroclaude.BuildClaudeContentBlockStartEvent(thinkingBlockIndex, "thinking", "", "")
										sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStart, &translatorParam)
										for _, chunk := range sseData {
											if chunk != "" {
												out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
											}
										}
									}
									thinkingEvent := kiroclaude.BuildClaudeThinkingDeltaEvent(processContent, thinkingBlockIndex)
									sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, thinkingEvent, &translatorParam)
									for _, chunk := range sseData {
										if chunk != "" {
											out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
										}
									}
									accumulatedThinkingContent.WriteString(processContent)
								}
							}
							processContent = ""
						}
					} else {
						// Not in thinking block, look for <thinking>
						startIdx := strings.Index(processContent, kirocommon.ThinkingStartTag)
						if startIdx >= 0 {
							// Found start tag - emit text content before the tag
							textBefore := processContent[:startIdx]
							if textBefore != "" {
								// Close thinking block if open
								if isThinkingBlockOpen {
									blockStop := kiroclaude.BuildClaudeThinkingBlockStopEvent(thinkingBlockIndex)
									sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
									for _, chunk := range sseData {
										if chunk != "" {
											out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
										}
									}
									isThinkingBlockOpen = false
								}
								// Ensure text block is open
								if !isTextBlockOpen {
									contentBlockIndex++
									isTextBlockOpen = true
									blockStart := kiroclaude.BuildClaudeContentBlockStartEvent(contentBlockIndex, "text", "", "")
									sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStart, &translatorParam)
									for _, chunk := range sseData {
										if chunk != "" {
											out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
										}
									}
								}
								// Send text delta
								claudeEvent := kiroclaude.BuildClaudeStreamEvent(textBefore, contentBlockIndex)
								sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, claudeEvent, &translatorParam)
								for _, chunk := range sseData {
									if chunk != "" {
										out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
									}
								}
							}
							// Close text block before entering thinking
							if isTextBlockOpen {
								blockStop := kiroclaude.BuildClaudeContentBlockStopEvent(contentBlockIndex)
								sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
								for _, chunk := range sseData {
									if chunk != "" {
										out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
									}
								}
								isTextBlockOpen = false
							}
							inThinkBlock = true
							processContent = processContent[startIdx+len(kirocommon.ThinkingStartTag):]
							log.Debugf("kiro: entered thinking block")
						} else {
							// No start tag found - check for partial match at end
							partialMatch := false
							for i := 1; i < len(kirocommon.ThinkingStartTag) && i <= len(processContent); i++ {
								if strings.HasSuffix(processContent, kirocommon.ThinkingStartTag[:i]) {
									// Possible partial tag at end, buffer it
									pendingContent.WriteString(processContent[len(processContent)-i:])
									processContent = processContent[:len(processContent)-i]
									partialMatch = true
									break
								}
							}
							if !partialMatch || len(processContent) > 0 {
								// Emit all as text content
								if processContent != "" {
									if !isTextBlockOpen {
										contentBlockIndex++
										isTextBlockOpen = true
										blockStart := kiroclaude.BuildClaudeContentBlockStartEvent(contentBlockIndex, "text", "", "")
										sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStart, &translatorParam)
										for _, chunk := range sseData {
											if chunk != "" {
												out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
											}
										}
									}
									claudeEvent := kiroclaude.BuildClaudeStreamEvent(processContent, contentBlockIndex)
									sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, claudeEvent, &translatorParam)
									for _, chunk := range sseData {
										if chunk != "" {
											out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
										}
									}
								}
							}
							processContent = ""
						}
					}
				}
			}

			// Handle tool uses in response (with deduplication)
			for _, tu := range toolUses {
				toolUseID := kirocommon.GetString(tu, "toolUseId")
				toolName := kirocommon.GetString(tu, "name")

				// Check for duplicate
				if processedIDs[toolUseID] {
					log.Debugf("kiro: skipping duplicate tool use in stream: %s", toolUseID)
					continue
				}
				processedIDs[toolUseID] = true

				hasToolUses = true
				// Close text block if open before starting tool_use block
				if isTextBlockOpen && contentBlockIndex >= 0 {
					blockStop := kiroclaude.BuildClaudeContentBlockStopEvent(contentBlockIndex)
					sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
					for _, chunk := range sseData {
						if chunk != "" {
							out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
						}
					}
					isTextBlockOpen = false
				}

				// Emit tool_use content block
				contentBlockIndex++

				blockStart := kiroclaude.BuildClaudeContentBlockStartEvent(contentBlockIndex, "tool_use", toolUseID, toolName)
				sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStart, &translatorParam)
				for _, chunk := range sseData {
					if chunk != "" {
						out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
					}
				}

				// Send input_json_delta with the tool input
				if input, ok := tu["input"].(map[string]interface{}); ok {
					inputJSON, err := json.Marshal(input)
					if err != nil {
						log.Debugf("kiro: failed to marshal tool input: %v", err)
						// Don't continue - still need to close the block
					} else {
						inputDelta := kiroclaude.BuildClaudeInputJsonDeltaEvent(string(inputJSON), contentBlockIndex)
						sseData = sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, inputDelta, &translatorParam)
						for _, chunk := range sseData {
							if chunk != "" {
								out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
							}
						}
					}
				}

				// Close tool_use block (always close even if input marshal failed)
				blockStop := kiroclaude.BuildClaudeContentBlockStopEvent(contentBlockIndex)
				sseData = sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
				for _, chunk := range sseData {
					if chunk != "" {
						out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
					}
				}
			}

		case "reasoningContentEvent":
			// Handle official reasoningContentEvent from Kiro API
			// This replaces tag-based thinking detection with the proper event type
			// Official format: { text: string, signature?: string, redactedContent?: base64 }
			var thinkingText string
			var signature string

			if re, ok := event["reasoningContentEvent"].(map[string]interface{}); ok {
				if text, ok := re["text"].(string); ok {
					thinkingText = text
				}
				if sig, ok := re["signature"].(string); ok {
					signature = sig
					if len(sig) > 20 {
						log.Debugf("kiro: reasoningContentEvent has signature: %s...", sig[:20])
					} else {
						log.Debugf("kiro: reasoningContentEvent has signature: %s", sig)
					}
				}
			} else {
				// Try direct fields
				if text, ok := event["text"].(string); ok {
					thinkingText = text
				}
				if sig, ok := event["signature"].(string); ok {
					signature = sig
				}
			}

			if thinkingText != "" {
				// Close text block if open before starting thinking block
				if isTextBlockOpen && contentBlockIndex >= 0 {
					blockStop := kiroclaude.BuildClaudeContentBlockStopEvent(contentBlockIndex)
					sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
					for _, chunk := range sseData {
						if chunk != "" {
							out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
						}
					}
					isTextBlockOpen = false
				}

				// Start thinking block if not already open
				if !isThinkingBlockOpen {
					contentBlockIndex++
					thinkingBlockIndex = contentBlockIndex
					isThinkingBlockOpen = true
					blockStart := kiroclaude.BuildClaudeContentBlockStartEvent(thinkingBlockIndex, "thinking", "", "")
					sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStart, &translatorParam)
					for _, chunk := range sseData {
						if chunk != "" {
							out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
						}
					}
				}

				// Send thinking content
				thinkingEvent := kiroclaude.BuildClaudeThinkingDeltaEvent(thinkingText, thinkingBlockIndex)
				sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, thinkingEvent, &translatorParam)
				for _, chunk := range sseData {
					if chunk != "" {
						out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
					}
				}

				// Accumulate for token counting
				accumulatedThinkingContent.WriteString(thinkingText)
				log.Debugf("kiro: received reasoningContentEvent, text length: %d, has signature: %v", len(thinkingText), signature != "")
			}

			// Note: We don't close the thinking block here - it will be closed when we see
			// the next assistantResponseEvent or at the end of the stream
			_ = signature // Signature can be used for verification if needed

		case "toolUseEvent":
			// Handle dedicated tool use events with input buffering
			completedToolUses, newState := kiroclaude.ProcessToolUseEvent(event, currentToolUse, processedIDs)
			currentToolUse = newState

			// Emit completed tool uses
			for _, tu := range completedToolUses {
				hasToolUses = true

				// Close text block if open
				if isTextBlockOpen && contentBlockIndex >= 0 {
					blockStop := kiroclaude.BuildClaudeContentBlockStopEvent(contentBlockIndex)
					sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
					for _, chunk := range sseData {
						if chunk != "" {
							out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
						}
					}
					isTextBlockOpen = false
				}

				contentBlockIndex++

				blockStart := kiroclaude.BuildClaudeContentBlockStartEvent(contentBlockIndex, "tool_use", tu.ToolUseID, tu.Name)
				sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStart, &translatorParam)
				for _, chunk := range sseData {
					if chunk != "" {
						out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
					}
				}

				if tu.Input != nil {
					inputJSON, err := json.Marshal(tu.Input)
					if err != nil {
						log.Debugf("kiro: failed to marshal tool input in toolUseEvent: %v", err)
					} else {
						inputDelta := kiroclaude.BuildClaudeInputJsonDeltaEvent(string(inputJSON), contentBlockIndex)
						sseData = sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, inputDelta, &translatorParam)
						for _, chunk := range sseData {
							if chunk != "" {
								out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
							}
						}
					}
				}

				blockStop := kiroclaude.BuildClaudeContentBlockStopEvent(contentBlockIndex)
				sseData = sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
				for _, chunk := range sseData {
					if chunk != "" {
						out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
					}
				}
			}

		case "supplementaryWebLinksEvent":
			if inputTokens, ok := event["inputTokens"].(float64); ok {
				totalUsage.InputTokens = int64(inputTokens)
			}
			if outputTokens, ok := event["outputTokens"].(float64); ok {
				totalUsage.OutputTokens = int64(outputTokens)
			}

		case "messageMetadataEvent", "metadataEvent":
			// Handle message metadata events which contain token counts
			// Official format: { tokenUsage: { outputTokens, totalTokens, uncachedInputTokens, cacheReadInputTokens, cacheWriteInputTokens, contextUsagePercentage } }
			var metadata map[string]interface{}
			if m, ok := event["messageMetadataEvent"].(map[string]interface{}); ok {
				metadata = m
			} else if m, ok := event["metadataEvent"].(map[string]interface{}); ok {
				metadata = m
			} else {
				metadata = event // event itself might be the metadata
			}

			// Check for nested tokenUsage object (official format)
			if tokenUsage, ok := metadata["tokenUsage"].(map[string]interface{}); ok {
				// outputTokens - precise output token count
				if outputTokens, ok := tokenUsage["outputTokens"].(float64); ok {
					totalUsage.OutputTokens = int64(outputTokens)
					hasUpstreamUsage = true
					log.Infof("kiro: streamToChannel found precise outputTokens in tokenUsage: %d", totalUsage.OutputTokens)
				}
				// totalTokens - precise total token count
				if totalTokens, ok := tokenUsage["totalTokens"].(float64); ok {
					totalUsage.TotalTokens = int64(totalTokens)
					log.Infof("kiro: streamToChannel found precise totalTokens in tokenUsage: %d", totalUsage.TotalTokens)
				}
				// uncachedInputTokens - input tokens not from cache
				if uncachedInputTokens, ok := tokenUsage["uncachedInputTokens"].(float64); ok {
					totalUsage.InputTokens = int64(uncachedInputTokens)
					hasUpstreamUsage = true
					log.Infof("kiro: streamToChannel found uncachedInputTokens in tokenUsage: %d", totalUsage.InputTokens)
				}
				// cacheReadInputTokens - tokens read from cache
				if cacheReadTokens, ok := tokenUsage["cacheReadInputTokens"].(float64); ok {
					// Add to input tokens if we have uncached tokens, otherwise use as input
					if totalUsage.InputTokens > 0 {
						totalUsage.InputTokens += int64(cacheReadTokens)
					} else {
						totalUsage.InputTokens = int64(cacheReadTokens)
					}
					hasUpstreamUsage = true
					log.Debugf("kiro: streamToChannel found cacheReadInputTokens in tokenUsage: %d", int64(cacheReadTokens))
				}
				// contextUsagePercentage - can be used as fallback for input token estimation
				if ctxPct, ok := tokenUsage["contextUsagePercentage"].(float64); ok {
					upstreamContextPercentage = ctxPct
					log.Debugf("kiro: streamToChannel found contextUsagePercentage in tokenUsage: %.2f%%", ctxPct)
				}
			}

			// Fallback: check for direct fields in metadata (legacy format)
			if totalUsage.InputTokens == 0 {
				if inputTokens, ok := metadata["inputTokens"].(float64); ok {
					totalUsage.InputTokens = int64(inputTokens)
					hasUpstreamUsage = true
					log.Debugf("kiro: streamToChannel found inputTokens in messageMetadataEvent: %d", totalUsage.InputTokens)
				}
			}
			if totalUsage.OutputTokens == 0 {
				if outputTokens, ok := metadata["outputTokens"].(float64); ok {
					totalUsage.OutputTokens = int64(outputTokens)
					hasUpstreamUsage = true
					log.Debugf("kiro: streamToChannel found outputTokens in messageMetadataEvent: %d", totalUsage.OutputTokens)
				}
			}
			if totalUsage.TotalTokens == 0 {
				if totalTokens, ok := metadata["totalTokens"].(float64); ok {
					totalUsage.TotalTokens = int64(totalTokens)
					log.Debugf("kiro: streamToChannel found totalTokens in messageMetadataEvent: %d", totalUsage.TotalTokens)
				}
			}

		case "usageEvent", "usage":
			// Handle dedicated usage events
			if inputTokens, ok := event["inputTokens"].(float64); ok {
				totalUsage.InputTokens = int64(inputTokens)
				log.Debugf("kiro: streamToChannel found inputTokens in usageEvent: %d", totalUsage.InputTokens)
			}
			if outputTokens, ok := event["outputTokens"].(float64); ok {
				totalUsage.OutputTokens = int64(outputTokens)
				log.Debugf("kiro: streamToChannel found outputTokens in usageEvent: %d", totalUsage.OutputTokens)
			}
			if totalTokens, ok := event["totalTokens"].(float64); ok {
				totalUsage.TotalTokens = int64(totalTokens)
				log.Debugf("kiro: streamToChannel found totalTokens in usageEvent: %d", totalUsage.TotalTokens)
			}
			// Also check nested usage object
			if usageObj, ok := event["usage"].(map[string]interface{}); ok {
				if inputTokens, ok := usageObj["input_tokens"].(float64); ok {
					totalUsage.InputTokens = int64(inputTokens)
				} else if inputTokens, ok := usageObj["prompt_tokens"].(float64); ok {
					totalUsage.InputTokens = int64(inputTokens)
				}
				if outputTokens, ok := usageObj["output_tokens"].(float64); ok {
					totalUsage.OutputTokens = int64(outputTokens)
				} else if outputTokens, ok := usageObj["completion_tokens"].(float64); ok {
					totalUsage.OutputTokens = int64(outputTokens)
				}
				if totalTokens, ok := usageObj["total_tokens"].(float64); ok {
					totalUsage.TotalTokens = int64(totalTokens)
				}
				log.Debugf("kiro: streamToChannel found usage object: input=%d, output=%d, total=%d",
					totalUsage.InputTokens, totalUsage.OutputTokens, totalUsage.TotalTokens)
			}

		case "metricsEvent":
			// Handle metrics events which may contain usage data
			if metrics, ok := event["metricsEvent"].(map[string]interface{}); ok {
				if inputTokens, ok := metrics["inputTokens"].(float64); ok {
					totalUsage.InputTokens = int64(inputTokens)
				}
				if outputTokens, ok := metrics["outputTokens"].(float64); ok {
					totalUsage.OutputTokens = int64(outputTokens)
				}
				log.Debugf("kiro: streamToChannel found metricsEvent: input=%d, output=%d",
					totalUsage.InputTokens, totalUsage.OutputTokens)
			}
		}

		// Check nested usage event
		if usageEvent, ok := event["supplementaryWebLinksEvent"].(map[string]interface{}); ok {
			if inputTokens, ok := usageEvent["inputTokens"].(float64); ok {
				totalUsage.InputTokens = int64(inputTokens)
			}
			if outputTokens, ok := usageEvent["outputTokens"].(float64); ok {
				totalUsage.OutputTokens = int64(outputTokens)
			}
		}

		// Check for direct token fields in any event (fallback)
		if totalUsage.InputTokens == 0 {
			if inputTokens, ok := event["inputTokens"].(float64); ok {
				totalUsage.InputTokens = int64(inputTokens)
				log.Debugf("kiro: streamToChannel found direct inputTokens: %d", totalUsage.InputTokens)
			}
		}
		if totalUsage.OutputTokens == 0 {
			if outputTokens, ok := event["outputTokens"].(float64); ok {
				totalUsage.OutputTokens = int64(outputTokens)
				log.Debugf("kiro: streamToChannel found direct outputTokens: %d", totalUsage.OutputTokens)
			}
		}

		// Check for usage object in any event (OpenAI format)
		if totalUsage.InputTokens == 0 || totalUsage.OutputTokens == 0 {
			if usageObj, ok := event["usage"].(map[string]interface{}); ok {
				if totalUsage.InputTokens == 0 {
					if inputTokens, ok := usageObj["input_tokens"].(float64); ok {
						totalUsage.InputTokens = int64(inputTokens)
					} else if inputTokens, ok := usageObj["prompt_tokens"].(float64); ok {
						totalUsage.InputTokens = int64(inputTokens)
					}
				}
				if totalUsage.OutputTokens == 0 {
					if outputTokens, ok := usageObj["output_tokens"].(float64); ok {
						totalUsage.OutputTokens = int64(outputTokens)
					} else if outputTokens, ok := usageObj["completion_tokens"].(float64); ok {
						totalUsage.OutputTokens = int64(outputTokens)
					}
				}
				if totalUsage.TotalTokens == 0 {
					if totalTokens, ok := usageObj["total_tokens"].(float64); ok {
						totalUsage.TotalTokens = int64(totalTokens)
					}
				}
				log.Debugf("kiro: streamToChannel found usage object (fallback): input=%d, output=%d, total=%d",
					totalUsage.InputTokens, totalUsage.OutputTokens, totalUsage.TotalTokens)
			}
		}
	}

	// Close content block if open
	if isTextBlockOpen && contentBlockIndex >= 0 {
		blockStop := kiroclaude.BuildClaudeContentBlockStopEvent(contentBlockIndex)
		sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStop, &translatorParam)
		for _, chunk := range sseData {
			if chunk != "" {
				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
			}
		}
	}

	// Streaming token calculation - calculate output tokens from accumulated content
	// Only use local estimation if server didn't provide usage (server-side usage takes priority)
	if totalUsage.OutputTokens == 0 && accumulatedContent.Len() > 0 {
		// Try to use tiktoken for accurate counting
		if enc, err := getTokenizer(model); err == nil {
			if tokenCount, countErr := enc.Count(accumulatedContent.String()); countErr == nil {
				totalUsage.OutputTokens = int64(tokenCount)
				log.Debugf("kiro: streamToChannel calculated output tokens using tiktoken: %d", totalUsage.OutputTokens)
			} else {
				// Fallback on count error: estimate from character count
				totalUsage.OutputTokens = int64(accumulatedContent.Len() / 4)
				if totalUsage.OutputTokens == 0 {
					totalUsage.OutputTokens = 1
				}
				log.Debugf("kiro: streamToChannel tiktoken count failed, estimated from chars: %d", totalUsage.OutputTokens)
			}
		} else {
			// Fallback: estimate from character count (roughly 4 chars per token)
			totalUsage.OutputTokens = int64(accumulatedContent.Len() / 4)
			if totalUsage.OutputTokens == 0 {
				totalUsage.OutputTokens = 1
			}
			log.Debugf("kiro: streamToChannel estimated output tokens from chars: %d (content len: %d)", totalUsage.OutputTokens, accumulatedContent.Len())
		}
	} else if totalUsage.OutputTokens == 0 && outputLen > 0 {
		// Legacy fallback using outputLen
		totalUsage.OutputTokens = int64(outputLen / 4)
		if totalUsage.OutputTokens == 0 {
			totalUsage.OutputTokens = 1
		}
	}

	// Use contextUsagePercentage to calculate more accurate input tokens
	// Kiro model has 200k max context, contextUsagePercentage represents the percentage used
	// Formula: input_tokens = contextUsagePercentage * 200000 / 100
	// Note: The effective input context is ~170k (200k - 30k reserved for output)
	if upstreamContextPercentage > 0 {
		// Calculate input tokens from context percentage
		// Using 200k as the base since that's what Kiro reports against
		calculatedInputTokens := int64(upstreamContextPercentage * 200000 / 100)

		// Only use calculated value if it's significantly different from local estimate
		// This provides more accurate token counts based on upstream data
		if calculatedInputTokens > 0 {
			localEstimate := totalUsage.InputTokens
			totalUsage.InputTokens = calculatedInputTokens
			log.Debugf("kiro: using contextUsagePercentage (%.2f%%) to calculate input tokens: %d (local estimate was: %d)",
				upstreamContextPercentage, calculatedInputTokens, localEstimate)
		}
	}

	totalUsage.TotalTokens = totalUsage.InputTokens + totalUsage.OutputTokens

	// Log upstream usage information if received
	if hasUpstreamUsage {
		log.Debugf("kiro: upstream usage - credits: %.4f, context: %.2f%%, final tokens - input: %d, output: %d, total: %d",
			upstreamCreditUsage, upstreamContextPercentage,
			totalUsage.InputTokens, totalUsage.OutputTokens, totalUsage.TotalTokens)
	}

	// Determine stop reason: prefer upstream, then detect tool_use, default to end_turn
	stopReason := upstreamStopReason
	if stopReason == "" {
		if hasToolUses {
			stopReason = "tool_use"
			log.Debugf("kiro: streamToChannel using fallback stop_reason: tool_use")
		} else {
			stopReason = "end_turn"
			log.Debugf("kiro: streamToChannel using fallback stop_reason: end_turn")
		}
	}

	// Log warning if response was truncated due to max_tokens
	if stopReason == "max_tokens" {
		log.Warnf("kiro: response truncated due to max_tokens limit (streamToChannel)")
	}

	// Send message_delta event
	msgDelta := kiroclaude.BuildClaudeMessageDeltaEvent(stopReason, totalUsage)
	sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, msgDelta, &translatorParam)
	for _, chunk := range sseData {
		if chunk != "" {
			out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
		}
	}

	// Send message_stop event separately
	msgStop := kiroclaude.BuildClaudeMessageStopOnlyEvent()
	sseData = sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, msgStop, &translatorParam)
	for _, chunk := range sseData {
		if chunk != "" {
			out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
		}
	}
	// reporter.publish is called via defer
}

// NOTE: Claude SSE event builders moved to internal/translator/kiro/claude/kiro_claude_stream.go
// The executor now uses kiroclaude.BuildClaude*Event() functions instead

// CountTokens counts tokens locally using tiktoken since Kiro API doesn't expose a token counting endpoint.
// This provides approximate token counts for client requests.
func (e *KiroExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
	// Use tiktoken for local token counting
	enc, err := getTokenizer(req.Model)
	if err != nil {
		log.Warnf("kiro: CountTokens failed to get tokenizer: %v, falling back to estimate", err)
		// Fallback: estimate from payload size (roughly 4 chars per token)
		estimatedTokens := len(req.Payload) / 4
		if estimatedTokens == 0 && len(req.Payload) > 0 {
			estimatedTokens = 1
		}
		return cliproxyexecutor.Response{
			Payload: []byte(fmt.Sprintf(`{"count":%d}`, estimatedTokens)),
		}, nil
	}

	// Try to count tokens from the request payload
	var totalTokens int64

	// Try OpenAI chat format first
	if tokens, countErr := countOpenAIChatTokens(enc, req.Payload); countErr == nil && tokens > 0 {
		totalTokens = tokens
		log.Debugf("kiro: CountTokens counted %d tokens using OpenAI chat format", totalTokens)
	} else {
		// Fallback: count raw payload tokens
		if tokenCount, countErr := enc.Count(string(req.Payload)); countErr == nil {
			totalTokens = int64(tokenCount)
			log.Debugf("kiro: CountTokens counted %d tokens from raw payload", totalTokens)
		} else {
			// Final fallback: estimate from payload size
			totalTokens = int64(len(req.Payload) / 4)
			if totalTokens == 0 && len(req.Payload) > 0 {
				totalTokens = 1
			}
			log.Debugf("kiro: CountTokens estimated %d tokens from payload size", totalTokens)
		}
	}

	return cliproxyexecutor.Response{
		Payload: []byte(fmt.Sprintf(`{"count":%d}`, totalTokens)),
	}, nil
}

// Refresh refreshes the Kiro OAuth token.
// Supports both AWS Builder ID (SSO OIDC) and Google OAuth (social login).
// Uses mutex to prevent race conditions when multiple concurrent requests try to refresh.
func (e *KiroExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
	// Serialize token refresh operations to prevent race conditions
	e.refreshMu.Lock()
	defer e.refreshMu.Unlock()

	var authID string
	if auth != nil {
		authID = auth.ID
	} else {
		authID = "<nil>"
	}
	log.Debugf("kiro executor: refresh called for auth %s", authID)
	if auth == nil {
		return nil, fmt.Errorf("kiro executor: auth is nil")
	}

	// Double-check: After acquiring lock, verify token still needs refresh
	// Another goroutine may have already refreshed while we were waiting
	// NOTE: This check has a design limitation - it reads from the auth object passed in,
	// not from persistent storage. If another goroutine returns a new Auth object (via Clone),
	// this check won't see those updates. The mutex still prevents truly concurrent refreshes,
	// but queued goroutines may still attempt redundant refreshes. This is acceptable as
	// the refresh operation is idempotent and the extra API calls are infrequent.
	if auth.Metadata != nil {
		if lastRefresh, ok := auth.Metadata["last_refresh"].(string); ok {
			if refreshTime, err := time.Parse(time.RFC3339, lastRefresh); err == nil {
				// If token was refreshed within the last 30 seconds, skip refresh
				if time.Since(refreshTime) < 30*time.Second {
					log.Debugf("kiro executor: token was recently refreshed by another goroutine, skipping")
					return auth, nil
				}
			}
		}
		// Also check if expires_at is now in the future with sufficient buffer
		if expiresAt, ok := auth.Metadata["expires_at"].(string); ok {
			if expTime, err := time.Parse(time.RFC3339, expiresAt); err == nil {
				// If token expires more than 20 minutes from now, it's still valid
				if time.Until(expTime) > 20*time.Minute {
					log.Debugf("kiro executor: token is still valid (expires in %v), skipping refresh", time.Until(expTime))
					// CRITICAL FIX: Set NextRefreshAfter to prevent frequent refresh checks
					// Without this, shouldRefresh() will return true again in 30 seconds
					updated := auth.Clone()
					// Set next refresh to 20 minutes before expiry, or at least 30 seconds from now
					nextRefresh := expTime.Add(-20 * time.Minute)
					minNextRefresh := time.Now().Add(30 * time.Second)
					if nextRefresh.Before(minNextRefresh) {
						nextRefresh = minNextRefresh
					}
					updated.NextRefreshAfter = nextRefresh
					log.Debugf("kiro executor: setting NextRefreshAfter to %v (in %v)", nextRefresh.Format(time.RFC3339), time.Until(nextRefresh))
					return updated, nil
				}
			}
		}
	}

	var refreshToken string
	var clientID, clientSecret string
	var authMethod string
	var region, startURL string

	if auth.Metadata != nil {
		if rt, ok := auth.Metadata["refresh_token"].(string); ok {
			refreshToken = rt
		}
		if cid, ok := auth.Metadata["client_id"].(string); ok {
			clientID = cid
		}
		if cs, ok := auth.Metadata["client_secret"].(string); ok {
			clientSecret = cs
		}
		if am, ok := auth.Metadata["auth_method"].(string); ok {
			authMethod = am
		}
		if r, ok := auth.Metadata["region"].(string); ok {
			region = r
		}
		if su, ok := auth.Metadata["start_url"].(string); ok {
			startURL = su
		}
	}

	if refreshToken == "" {
		return nil, fmt.Errorf("kiro executor: refresh token not found")
	}

	var tokenData *kiroauth.KiroTokenData
	var err error

	ssoClient := kiroauth.NewSSOOIDCClient(e.cfg)

	// Use SSO OIDC refresh for AWS Builder ID or IDC, otherwise use Kiro's OAuth refresh endpoint
	switch {
	case clientID != "" && clientSecret != "" && authMethod == "idc" && region != "":
		// IDC refresh with region-specific endpoint
		log.Debugf("kiro executor: using SSO OIDC refresh for IDC (region=%s)", region)
		tokenData, err = ssoClient.RefreshTokenWithRegion(ctx, clientID, clientSecret, refreshToken, region, startURL)
	case clientID != "" && clientSecret != "" && authMethod == "builder-id":
		// Builder ID refresh with default endpoint
		log.Debugf("kiro executor: using SSO OIDC refresh for AWS Builder ID")
		tokenData, err = ssoClient.RefreshToken(ctx, clientID, clientSecret, refreshToken)
	default:
		// Fallback to Kiro's OAuth refresh endpoint (for social auth: Google/GitHub)
		log.Debugf("kiro executor: using Kiro OAuth refresh endpoint")
		oauth := kiroauth.NewKiroOAuth(e.cfg)
		tokenData, err = oauth.RefreshToken(ctx, refreshToken)
	}

	if err != nil {
		return nil, fmt.Errorf("kiro executor: token refresh failed: %w", err)
	}

	updated := auth.Clone()
	now := time.Now()
	updated.UpdatedAt = now
	updated.LastRefreshedAt = now

	if updated.Metadata == nil {
		updated.Metadata = make(map[string]any)
	}
	updated.Metadata["access_token"] = tokenData.AccessToken
	updated.Metadata["refresh_token"] = tokenData.RefreshToken
	updated.Metadata["expires_at"] = tokenData.ExpiresAt
	updated.Metadata["last_refresh"] = now.Format(time.RFC3339)
	if tokenData.ProfileArn != "" {
		updated.Metadata["profile_arn"] = tokenData.ProfileArn
	}
	if tokenData.AuthMethod != "" {
		updated.Metadata["auth_method"] = tokenData.AuthMethod
	}
	if tokenData.Provider != "" {
		updated.Metadata["provider"] = tokenData.Provider
	}
	// Preserve client credentials for future refreshes (AWS Builder ID)
	if tokenData.ClientID != "" {
		updated.Metadata["client_id"] = tokenData.ClientID
	}
	if tokenData.ClientSecret != "" {
		updated.Metadata["client_secret"] = tokenData.ClientSecret
	}
	// Preserve region and start_url for IDC token refresh
	if tokenData.Region != "" {
		updated.Metadata["region"] = tokenData.Region
	}
	if tokenData.StartURL != "" {
		updated.Metadata["start_url"] = tokenData.StartURL
	}

	if updated.Attributes == nil {
		updated.Attributes = make(map[string]string)
	}
	updated.Attributes["access_token"] = tokenData.AccessToken
	if tokenData.ProfileArn != "" {
		updated.Attributes["profile_arn"] = tokenData.ProfileArn
	}

	// NextRefreshAfter is aligned with RefreshLead (20min)
	if expiresAt, parseErr := time.Parse(time.RFC3339, tokenData.ExpiresAt); parseErr == nil {
		updated.NextRefreshAfter = expiresAt.Add(-20 * time.Minute)
	}

	log.Infof("kiro executor: token refreshed successfully, expires at %s", tokenData.ExpiresAt)
	return updated, nil
}

// persistRefreshedAuth persists a refreshed auth record to disk.
// This ensures token refreshes from inline retry are saved to the auth file.
func (e *KiroExecutor) persistRefreshedAuth(auth *cliproxyauth.Auth) error {
	if auth == nil || auth.Metadata == nil {
		return fmt.Errorf("kiro executor: cannot persist nil auth or metadata")
	}

	// Determine the file path from auth attributes or filename
	var authPath string
	if auth.Attributes != nil {
		if p := strings.TrimSpace(auth.Attributes["path"]); p != "" {
			authPath = p
		}
	}
	if authPath == "" {
		fileName := strings.TrimSpace(auth.FileName)
		if fileName == "" {
			return fmt.Errorf("kiro executor: auth has no file path or filename")
		}
		if filepath.IsAbs(fileName) {
			authPath = fileName
		} else if e.cfg != nil && e.cfg.AuthDir != "" {
			authPath = filepath.Join(e.cfg.AuthDir, fileName)
		} else {
			return fmt.Errorf("kiro executor: cannot determine auth file path")
		}
	}

	// Marshal metadata to JSON
	raw, err := json.Marshal(auth.Metadata)
	if err != nil {
		return fmt.Errorf("kiro executor: marshal metadata failed: %w", err)
	}

	// Write to temp file first, then rename (atomic write)
	tmp := authPath + ".tmp"
	if err := os.WriteFile(tmp, raw, 0o600); err != nil {
		return fmt.Errorf("kiro executor: write temp auth file failed: %w", err)
	}
	if err := os.Rename(tmp, authPath); err != nil {
		return fmt.Errorf("kiro executor: rename auth file failed: %w", err)
	}

	log.Debugf("kiro executor: persisted refreshed auth to %s", authPath)
	return nil
}

// reloadAuthFromFile 从文件重新加载 auth 数据（方案 B: Fallback 机制）
// 当内存中的 token 已过期时，尝试从文件读取最新的 token
// 这解决了后台刷新器已更新文件但内存中 Auth 对象尚未同步的时间差问题
func (e *KiroExecutor) reloadAuthFromFile(auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
	if auth == nil {
		return nil, fmt.Errorf("kiro executor: cannot reload nil auth")
	}

	// 确定文件路径
	var authPath string
	if auth.Attributes != nil {
		if p := strings.TrimSpace(auth.Attributes["path"]); p != "" {
			authPath = p
		}
	}
	if authPath == "" {
		fileName := strings.TrimSpace(auth.FileName)
		if fileName == "" {
			return nil, fmt.Errorf("kiro executor: auth has no file path or filename for reload")
		}
		if filepath.IsAbs(fileName) {
			authPath = fileName
		} else if e.cfg != nil && e.cfg.AuthDir != "" {
			authPath = filepath.Join(e.cfg.AuthDir, fileName)
		} else {
			return nil, fmt.Errorf("kiro executor: cannot determine auth file path for reload")
		}
	}

	// 读取文件
	raw, err := os.ReadFile(authPath)
	if err != nil {
		return nil, fmt.Errorf("kiro executor: failed to read auth file %s: %w", authPath, err)
	}

	// 解析 JSON
	var metadata map[string]any
	if err := json.Unmarshal(raw, &metadata); err != nil {
		return nil, fmt.Errorf("kiro executor: failed to parse auth file %s: %w", authPath, err)
	}

	// 检查文件中的 token 是否比内存中的更新
	fileExpiresAt, _ := metadata["expires_at"].(string)
	fileAccessToken, _ := metadata["access_token"].(string)
	memExpiresAt, _ := auth.Metadata["expires_at"].(string)
	memAccessToken, _ := auth.Metadata["access_token"].(string)

	// 文件中必须有有效的 access_token
	if fileAccessToken == "" {
		return nil, fmt.Errorf("kiro executor: auth file has no access_token field")
	}

	// 如果有 expires_at，检查是否过期
	if fileExpiresAt != "" {
		fileExpTime, parseErr := time.Parse(time.RFC3339, fileExpiresAt)
		if parseErr == nil {
			// 如果文件中的 token 也已过期，不使用它
			if time.Now().After(fileExpTime) {
				log.Debugf("kiro executor: file token also expired at %s, not using", fileExpiresAt)
				return nil, fmt.Errorf("kiro executor: file token also expired")
			}
		}
	}

	// 判断文件中的 token 是否比内存中的更新
	// 条件1: access_token 不同（说明已刷新）
	// 条件2: expires_at 更新（说明已刷新）
	isNewer := false

	// 优先检查 access_token 是否变化
	if fileAccessToken != memAccessToken {
		isNewer = true
		log.Debugf("kiro executor: file access_token differs from memory, using file token")
	}

	// 如果 access_token 相同，检查 expires_at
	if !isNewer && fileExpiresAt != "" && memExpiresAt != "" {
		fileExpTime, fileParseErr := time.Parse(time.RFC3339, fileExpiresAt)
		memExpTime, memParseErr := time.Parse(time.RFC3339, memExpiresAt)
		if fileParseErr == nil && memParseErr == nil && fileExpTime.After(memExpTime) {
			isNewer = true
			log.Debugf("kiro executor: file expires_at (%s) is newer than memory (%s)", fileExpiresAt, memExpiresAt)
		}
	}

	// 如果文件中没有 expires_at 但 access_token 相同，无法判断是否更新
	if !isNewer && fileExpiresAt == "" && fileAccessToken == memAccessToken {
		return nil, fmt.Errorf("kiro executor: cannot determine if file token is newer (no expires_at, same access_token)")
	}

	if !isNewer {
		log.Debugf("kiro executor: file token not newer than memory token")
		return nil, fmt.Errorf("kiro executor: file token not newer")
	}

	// 创建更新后的 auth 对象
	updated := auth.Clone()
	updated.Metadata = metadata
	updated.UpdatedAt = time.Now()

	// 同步更新 Attributes
	if updated.Attributes == nil {
		updated.Attributes = make(map[string]string)
	}
	if accessToken, ok := metadata["access_token"].(string); ok {
		updated.Attributes["access_token"] = accessToken
	}
	if profileArn, ok := metadata["profile_arn"].(string); ok {
		updated.Attributes["profile_arn"] = profileArn
	}

	log.Infof("kiro executor: reloaded auth from file %s, new expires_at: %s", authPath, fileExpiresAt)
	return updated, nil
}

// isTokenExpired checks if a JWT access token has expired.
// Returns true if the token is expired or cannot be parsed.
func (e *KiroExecutor) isTokenExpired(accessToken string) bool {
	if accessToken == "" {
		return true
	}

	// JWT tokens have 3 parts separated by dots
	parts := strings.Split(accessToken, ".")
	if len(parts) != 3 {
		// Not a JWT token, assume not expired
		return false
	}

	// Decode the payload (second part)
	// JWT uses base64url encoding without padding (RawURLEncoding)
	payload := parts[1]
	decoded, err := base64.RawURLEncoding.DecodeString(payload)
	if err != nil {
		// Try with padding added as fallback
		switch len(payload) % 4 {
		case 2:
			payload += "=="
		case 3:
			payload += "="
		}
		decoded, err = base64.URLEncoding.DecodeString(payload)
		if err != nil {
			log.Debugf("kiro: failed to decode JWT payload: %v", err)
			return false
		}
	}

	var claims struct {
		Exp int64 `json:"exp"`
	}
	if err := json.Unmarshal(decoded, &claims); err != nil {
		log.Debugf("kiro: failed to parse JWT claims: %v", err)
		return false
	}

	if claims.Exp == 0 {
		// No expiration claim, assume not expired
		return false
	}

	expTime := time.Unix(claims.Exp, 0)
	now := time.Now()

	// Consider token expired if it expires within 1 minute (buffer for clock skew)
	isExpired := now.After(expTime) || expTime.Sub(now) < time.Minute
	if isExpired {
		log.Debugf("kiro: token expired at %s (now: %s)", expTime.Format(time.RFC3339), now.Format(time.RFC3339))
	}

	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