Merge branch 'main' into plus

This commit is contained in:
Luis Pater
2026-01-20 13:50:41 +08:00
committed by GitHub
82 changed files with 16125 additions and 374 deletions

View File

@@ -50,6 +50,10 @@ type ToolCallAccumulator struct {
// Returns:
// - []string: A slice of strings, each containing an OpenAI-compatible JSON response
func ConvertClaudeResponseToOpenAI(_ context.Context, modelName string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, param *any) []string {
var localParam any
if param == nil {
param = &localParam
}
if *param == nil {
*param = &ConvertAnthropicResponseToOpenAIParams{
CreatedAt: 0,

View File

@@ -33,4 +33,7 @@ import (
_ "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/antigravity/gemini"
_ "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/antigravity/openai/chat-completions"
_ "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/antigravity/openai/responses"
_ "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/claude"
_ "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/openai"
)

View File

@@ -0,0 +1,20 @@
// Package claude provides translation between Kiro and Claude formats.
package claude
import (
. "github.com/router-for-me/CLIProxyAPI/v6/internal/constant"
"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/translator"
)
func init() {
translator.Register(
Claude,
Kiro,
ConvertClaudeRequestToKiro,
interfaces.TranslateResponse{
Stream: ConvertKiroStreamToClaude,
NonStream: ConvertKiroNonStreamToClaude,
},
)
}

View File

@@ -0,0 +1,21 @@
// Package claude provides translation between Kiro and Claude formats.
// Since Kiro executor generates Claude-compatible SSE format internally (with event: prefix),
// translations are pass-through for streaming, but responses need proper formatting.
package claude
import (
"context"
)
// ConvertKiroStreamToClaude converts Kiro streaming response to Claude format.
// Kiro executor already generates complete SSE format with "event:" prefix,
// so this is a simple pass-through.
func ConvertKiroStreamToClaude(ctx context.Context, model string, originalRequest, request, rawResponse []byte, param *any) []string {
return []string{string(rawResponse)}
}
// ConvertKiroNonStreamToClaude converts Kiro non-streaming response to Claude format.
// The response is already in Claude format, so this is a pass-through.
func ConvertKiroNonStreamToClaude(ctx context.Context, model string, originalRequest, request, rawResponse []byte, param *any) string {
return string(rawResponse)
}

View File

@@ -0,0 +1,817 @@
// Package claude provides request translation functionality for Claude API to Kiro format.
// It handles parsing and transforming Claude API requests into the Kiro/Amazon Q API format,
// extracting model information, system instructions, message contents, and tool declarations.
package claude
import (
"encoding/json"
"fmt"
"net/http"
"strings"
"time"
"unicode/utf8"
"github.com/google/uuid"
kirocommon "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/common"
log "github.com/sirupsen/logrus"
"github.com/tidwall/gjson"
)
// Kiro API request structs - field order determines JSON key order
// KiroPayload is the top-level request structure for Kiro API
type KiroPayload struct {
ConversationState KiroConversationState `json:"conversationState"`
ProfileArn string `json:"profileArn,omitempty"`
InferenceConfig *KiroInferenceConfig `json:"inferenceConfig,omitempty"`
}
// KiroInferenceConfig contains inference parameters for the Kiro API.
type KiroInferenceConfig struct {
MaxTokens int `json:"maxTokens,omitempty"`
Temperature float64 `json:"temperature,omitempty"`
TopP float64 `json:"topP,omitempty"`
}
// KiroConversationState holds the conversation context
type KiroConversationState struct {
ChatTriggerType string `json:"chatTriggerType"` // Required: "MANUAL" - must be first field
ConversationID string `json:"conversationId"`
CurrentMessage KiroCurrentMessage `json:"currentMessage"`
History []KiroHistoryMessage `json:"history,omitempty"`
}
// KiroCurrentMessage wraps the current user message
type KiroCurrentMessage struct {
UserInputMessage KiroUserInputMessage `json:"userInputMessage"`
}
// KiroHistoryMessage represents a message in the conversation history
type KiroHistoryMessage struct {
UserInputMessage *KiroUserInputMessage `json:"userInputMessage,omitempty"`
AssistantResponseMessage *KiroAssistantResponseMessage `json:"assistantResponseMessage,omitempty"`
}
// KiroImage represents an image in Kiro API format
type KiroImage struct {
Format string `json:"format"`
Source KiroImageSource `json:"source"`
}
// KiroImageSource contains the image data
type KiroImageSource struct {
Bytes string `json:"bytes"` // base64 encoded image data
}
// KiroUserInputMessage represents a user message
type KiroUserInputMessage struct {
Content string `json:"content"`
ModelID string `json:"modelId"`
Origin string `json:"origin"`
Images []KiroImage `json:"images,omitempty"`
UserInputMessageContext *KiroUserInputMessageContext `json:"userInputMessageContext,omitempty"`
}
// KiroUserInputMessageContext contains tool-related context
type KiroUserInputMessageContext struct {
ToolResults []KiroToolResult `json:"toolResults,omitempty"`
Tools []KiroToolWrapper `json:"tools,omitempty"`
}
// KiroToolResult represents a tool execution result
type KiroToolResult struct {
Content []KiroTextContent `json:"content"`
Status string `json:"status"`
ToolUseID string `json:"toolUseId"`
}
// KiroTextContent represents text content
type KiroTextContent struct {
Text string `json:"text"`
}
// KiroToolWrapper wraps a tool specification
type KiroToolWrapper struct {
ToolSpecification KiroToolSpecification `json:"toolSpecification"`
}
// KiroToolSpecification defines a tool's schema
type KiroToolSpecification struct {
Name string `json:"name"`
Description string `json:"description"`
InputSchema KiroInputSchema `json:"inputSchema"`
}
// KiroInputSchema wraps the JSON schema for tool input
type KiroInputSchema struct {
JSON interface{} `json:"json"`
}
// KiroAssistantResponseMessage represents an assistant message
type KiroAssistantResponseMessage struct {
Content string `json:"content"`
ToolUses []KiroToolUse `json:"toolUses,omitempty"`
}
// KiroToolUse represents a tool invocation by the assistant
type KiroToolUse struct {
ToolUseID string `json:"toolUseId"`
Name string `json:"name"`
Input map[string]interface{} `json:"input"`
}
// ConvertClaudeRequestToKiro converts a Claude API request to Kiro format.
// This is the main entry point for request translation.
func ConvertClaudeRequestToKiro(modelName string, inputRawJSON []byte, stream bool) []byte {
// For Kiro, we pass through the Claude format since buildKiroPayload
// expects Claude format and does the conversion internally.
// The actual conversion happens in the executor when building the HTTP request.
return inputRawJSON
}
// BuildKiroPayload constructs the Kiro API request payload from Claude format.
// Supports tool calling - tools are passed via userInputMessageContext.
// origin parameter determines which quota to use: "CLI" for Amazon Q, "AI_EDITOR" for Kiro IDE.
// isAgentic parameter enables chunked write optimization prompt for -agentic model variants.
// isChatOnly parameter disables tool calling for -chat model variants (pure conversation mode).
// headers parameter allows checking Anthropic-Beta header for thinking mode detection.
// metadata parameter is kept for API compatibility but no longer used for thinking configuration.
// Supports thinking mode - when enabled, injects thinking tags into system prompt.
// Returns the payload and a boolean indicating whether thinking mode was injected.
func BuildKiroPayload(claudeBody []byte, modelID, profileArn, origin string, isAgentic, isChatOnly bool, headers http.Header, metadata map[string]any) ([]byte, bool) {
// Extract max_tokens for potential use in inferenceConfig
// Handle -1 as "use maximum" (Kiro max output is ~32000 tokens)
const kiroMaxOutputTokens = 32000
var maxTokens int64
if mt := gjson.GetBytes(claudeBody, "max_tokens"); mt.Exists() {
maxTokens = mt.Int()
if maxTokens == -1 {
maxTokens = kiroMaxOutputTokens
log.Debugf("kiro: max_tokens=-1 converted to %d", kiroMaxOutputTokens)
}
}
// Extract temperature if specified
var temperature float64
var hasTemperature bool
if temp := gjson.GetBytes(claudeBody, "temperature"); temp.Exists() {
temperature = temp.Float()
hasTemperature = true
}
// Extract top_p if specified
var topP float64
var hasTopP bool
if tp := gjson.GetBytes(claudeBody, "top_p"); tp.Exists() {
topP = tp.Float()
hasTopP = true
log.Debugf("kiro: extracted top_p: %.2f", topP)
}
// Normalize origin value for Kiro API compatibility
origin = normalizeOrigin(origin)
log.Debugf("kiro: normalized origin value: %s", origin)
messages := gjson.GetBytes(claudeBody, "messages")
// For chat-only mode, don't include tools
var tools gjson.Result
if !isChatOnly {
tools = gjson.GetBytes(claudeBody, "tools")
}
// Extract system prompt
systemPrompt := extractSystemPrompt(claudeBody)
// Check for thinking mode using the comprehensive IsThinkingEnabledWithHeaders function
// This supports Claude API format, OpenAI reasoning_effort, AMP/Cursor format, and Anthropic-Beta header
thinkingEnabled := IsThinkingEnabledWithHeaders(claudeBody, headers)
// Inject timestamp context
timestamp := time.Now().Format("2006-01-02 15:04:05 MST")
timestampContext := fmt.Sprintf("[Context: Current time is %s]", timestamp)
if systemPrompt != "" {
systemPrompt = timestampContext + "\n\n" + systemPrompt
} else {
systemPrompt = timestampContext
}
log.Debugf("kiro: injected timestamp context: %s", timestamp)
// Inject agentic optimization prompt for -agentic model variants
if isAgentic {
if systemPrompt != "" {
systemPrompt += "\n"
}
systemPrompt += kirocommon.KiroAgenticSystemPrompt
}
// Handle tool_choice parameter - Kiro doesn't support it natively, so we inject system prompt hints
// Claude tool_choice values: {"type": "auto/any/tool", "name": "..."}
toolChoiceHint := extractClaudeToolChoiceHint(claudeBody)
if toolChoiceHint != "" {
if systemPrompt != "" {
systemPrompt += "\n"
}
systemPrompt += toolChoiceHint
log.Debugf("kiro: injected tool_choice hint into system prompt")
}
// Convert Claude tools to Kiro format
kiroTools := convertClaudeToolsToKiro(tools)
// Thinking mode implementation:
// Kiro API supports official thinking/reasoning mode via <thinking_mode> tag.
// When set to "enabled", Kiro returns reasoning content as official reasoningContentEvent
// rather than inline <thinking> tags in assistantResponseEvent.
// We use a high max_thinking_length to allow extensive reasoning.
if thinkingEnabled {
thinkingHint := `<thinking_mode>enabled</thinking_mode>
<max_thinking_length>200000</max_thinking_length>`
if systemPrompt != "" {
systemPrompt = thinkingHint + "\n\n" + systemPrompt
} else {
systemPrompt = thinkingHint
}
log.Infof("kiro: injected thinking prompt (official mode), has_tools: %v", len(kiroTools) > 0)
}
// Process messages and build history
history, currentUserMsg, currentToolResults := processMessages(messages, modelID, origin)
// Build content with system prompt (only on first turn to avoid re-injection)
if currentUserMsg != nil {
effectiveSystemPrompt := systemPrompt
if len(history) > 0 {
effectiveSystemPrompt = "" // Don't re-inject on subsequent turns
}
currentUserMsg.Content = buildFinalContent(currentUserMsg.Content, effectiveSystemPrompt, currentToolResults)
// Deduplicate currentToolResults
currentToolResults = deduplicateToolResults(currentToolResults)
// Build userInputMessageContext with tools and tool results
if len(kiroTools) > 0 || len(currentToolResults) > 0 {
currentUserMsg.UserInputMessageContext = &KiroUserInputMessageContext{
Tools: kiroTools,
ToolResults: currentToolResults,
}
}
}
// Build payload
var currentMessage KiroCurrentMessage
if currentUserMsg != nil {
currentMessage = KiroCurrentMessage{UserInputMessage: *currentUserMsg}
} else {
fallbackContent := ""
if systemPrompt != "" {
fallbackContent = "--- SYSTEM PROMPT ---\n" + systemPrompt + "\n--- END SYSTEM PROMPT ---\n"
}
currentMessage = KiroCurrentMessage{UserInputMessage: KiroUserInputMessage{
Content: fallbackContent,
ModelID: modelID,
Origin: origin,
}}
}
// Build inferenceConfig if we have any inference parameters
// Note: Kiro API doesn't actually use max_tokens for thinking budget
var inferenceConfig *KiroInferenceConfig
if maxTokens > 0 || hasTemperature || hasTopP {
inferenceConfig = &KiroInferenceConfig{}
if maxTokens > 0 {
inferenceConfig.MaxTokens = int(maxTokens)
}
if hasTemperature {
inferenceConfig.Temperature = temperature
}
if hasTopP {
inferenceConfig.TopP = topP
}
}
payload := KiroPayload{
ConversationState: KiroConversationState{
ChatTriggerType: "MANUAL",
ConversationID: uuid.New().String(),
CurrentMessage: currentMessage,
History: history,
},
ProfileArn: profileArn,
InferenceConfig: inferenceConfig,
}
result, err := json.Marshal(payload)
if err != nil {
log.Debugf("kiro: failed to marshal payload: %v", err)
return nil, false
}
return result, thinkingEnabled
}
// normalizeOrigin normalizes origin value for Kiro API compatibility
func normalizeOrigin(origin string) string {
switch origin {
case "KIRO_CLI":
return "CLI"
case "KIRO_AI_EDITOR":
return "AI_EDITOR"
case "AMAZON_Q":
return "CLI"
case "KIRO_IDE":
return "AI_EDITOR"
default:
return origin
}
}
// extractSystemPrompt extracts system prompt from Claude request
func extractSystemPrompt(claudeBody []byte) string {
systemField := gjson.GetBytes(claudeBody, "system")
if systemField.IsArray() {
var sb strings.Builder
for _, block := range systemField.Array() {
if block.Get("type").String() == "text" {
sb.WriteString(block.Get("text").String())
} else if block.Type == gjson.String {
sb.WriteString(block.String())
}
}
return sb.String()
}
return systemField.String()
}
// checkThinkingMode checks if thinking mode is enabled in the Claude request
func checkThinkingMode(claudeBody []byte) (bool, int64) {
thinkingEnabled := false
var budgetTokens int64 = 24000
thinkingField := gjson.GetBytes(claudeBody, "thinking")
if thinkingField.Exists() {
thinkingType := thinkingField.Get("type").String()
if thinkingType == "enabled" {
thinkingEnabled = true
if bt := thinkingField.Get("budget_tokens"); bt.Exists() {
budgetTokens = bt.Int()
if budgetTokens <= 0 {
thinkingEnabled = false
log.Debugf("kiro: thinking mode disabled via budget_tokens <= 0")
}
}
if thinkingEnabled {
log.Debugf("kiro: thinking mode enabled via Claude API parameter, budget_tokens: %d", budgetTokens)
}
}
}
return thinkingEnabled, budgetTokens
}
// hasThinkingTagInBody checks if the request body already contains thinking configuration tags.
// This is used to prevent duplicate injection when client (e.g., AMP/Cursor) already includes thinking config.
func hasThinkingTagInBody(body []byte) bool {
bodyStr := string(body)
return strings.Contains(bodyStr, "<thinking_mode>") || strings.Contains(bodyStr, "<max_thinking_length>")
}
// IsThinkingEnabledFromHeader checks if thinking mode is enabled via Anthropic-Beta header.
// Claude CLI uses "Anthropic-Beta: interleaved-thinking-2025-05-14" to enable thinking.
func IsThinkingEnabledFromHeader(headers http.Header) bool {
if headers == nil {
return false
}
betaHeader := headers.Get("Anthropic-Beta")
if betaHeader == "" {
return false
}
// Check for interleaved-thinking beta feature
if strings.Contains(betaHeader, "interleaved-thinking") {
log.Debugf("kiro: thinking mode enabled via Anthropic-Beta header: %s", betaHeader)
return true
}
return false
}
// IsThinkingEnabled is a public wrapper to check if thinking mode is enabled.
// This is used by the executor to determine whether to parse <thinking> tags in responses.
// When thinking is NOT enabled in the request, <thinking> tags in responses should be
// treated as regular text content, not as thinking blocks.
//
// Supports multiple formats:
// - Claude API format: thinking.type = "enabled"
// - OpenAI format: reasoning_effort parameter
// - AMP/Cursor format: <thinking_mode>interleaved</thinking_mode> in system prompt
func IsThinkingEnabled(body []byte) bool {
return IsThinkingEnabledWithHeaders(body, nil)
}
// IsThinkingEnabledWithHeaders checks if thinking mode is enabled from body or headers.
// This is the comprehensive check that supports all thinking detection methods:
// - Claude API format: thinking.type = "enabled"
// - OpenAI format: reasoning_effort parameter
// - AMP/Cursor format: <thinking_mode>interleaved</thinking_mode> in system prompt
// - Anthropic-Beta header: interleaved-thinking-2025-05-14
func IsThinkingEnabledWithHeaders(body []byte, headers http.Header) bool {
// Check Anthropic-Beta header first (Claude Code uses this)
if IsThinkingEnabledFromHeader(headers) {
return true
}
// Check Claude API format first (thinking.type = "enabled")
enabled, _ := checkThinkingMode(body)
if enabled {
log.Debugf("kiro: IsThinkingEnabled returning true (Claude API format)")
return true
}
// Check OpenAI format: reasoning_effort parameter
// Valid values: "low", "medium", "high", "auto" (not "none")
reasoningEffort := gjson.GetBytes(body, "reasoning_effort")
if reasoningEffort.Exists() {
effort := reasoningEffort.String()
if effort != "" && effort != "none" {
log.Debugf("kiro: thinking mode enabled via OpenAI reasoning_effort: %s", effort)
return true
}
}
// Check AMP/Cursor format: <thinking_mode>interleaved</thinking_mode> in system prompt
// This is how AMP client passes thinking configuration
bodyStr := string(body)
if strings.Contains(bodyStr, "<thinking_mode>") && strings.Contains(bodyStr, "</thinking_mode>") {
// Extract thinking mode value
startTag := "<thinking_mode>"
endTag := "</thinking_mode>"
startIdx := strings.Index(bodyStr, startTag)
if startIdx >= 0 {
startIdx += len(startTag)
endIdx := strings.Index(bodyStr[startIdx:], endTag)
if endIdx >= 0 {
thinkingMode := bodyStr[startIdx : startIdx+endIdx]
if thinkingMode == "interleaved" || thinkingMode == "enabled" {
log.Debugf("kiro: thinking mode enabled via AMP/Cursor format: %s", thinkingMode)
return true
}
}
}
}
// Check OpenAI format: max_completion_tokens with reasoning (o1-style)
// Some clients use this to indicate reasoning mode
if gjson.GetBytes(body, "max_completion_tokens").Exists() {
// If max_completion_tokens is set, check if model name suggests reasoning
model := gjson.GetBytes(body, "model").String()
if strings.Contains(strings.ToLower(model), "thinking") ||
strings.Contains(strings.ToLower(model), "reason") {
log.Debugf("kiro: thinking mode enabled via model name hint: %s", model)
return true
}
}
log.Debugf("kiro: IsThinkingEnabled returning false (no thinking mode detected)")
return false
}
// shortenToolNameIfNeeded shortens tool names that exceed 64 characters.
// MCP tools often have long names like "mcp__server-name__tool-name".
// This preserves the "mcp__" prefix and last segment when possible.
func shortenToolNameIfNeeded(name string) string {
const limit = 64
if len(name) <= limit {
return name
}
// For MCP tools, try to preserve prefix and last segment
if strings.HasPrefix(name, "mcp__") {
idx := strings.LastIndex(name, "__")
if idx > 0 {
cand := "mcp__" + name[idx+2:]
if len(cand) > limit {
return cand[:limit]
}
return cand
}
}
return name[:limit]
}
// convertClaudeToolsToKiro converts Claude tools to Kiro format
func convertClaudeToolsToKiro(tools gjson.Result) []KiroToolWrapper {
var kiroTools []KiroToolWrapper
if !tools.IsArray() {
return kiroTools
}
for _, tool := range tools.Array() {
name := tool.Get("name").String()
description := tool.Get("description").String()
inputSchema := tool.Get("input_schema").Value()
// Shorten tool name if it exceeds 64 characters (common with MCP tools)
originalName := name
name = shortenToolNameIfNeeded(name)
if name != originalName {
log.Debugf("kiro: shortened tool name from '%s' to '%s'", originalName, name)
}
// CRITICAL FIX: Kiro API requires non-empty description
if strings.TrimSpace(description) == "" {
description = fmt.Sprintf("Tool: %s", name)
log.Debugf("kiro: tool '%s' has empty description, using default: %s", name, description)
}
// Truncate long descriptions (individual tool limit)
if len(description) > kirocommon.KiroMaxToolDescLen {
truncLen := kirocommon.KiroMaxToolDescLen - 30
for truncLen > 0 && !utf8.RuneStart(description[truncLen]) {
truncLen--
}
description = description[:truncLen] + "... (description truncated)"
}
kiroTools = append(kiroTools, KiroToolWrapper{
ToolSpecification: KiroToolSpecification{
Name: name,
Description: description,
InputSchema: KiroInputSchema{JSON: inputSchema},
},
})
}
// Apply dynamic compression if total tools size exceeds threshold
// This prevents 500 errors when Claude Code sends too many tools
kiroTools = compressToolsIfNeeded(kiroTools)
return kiroTools
}
// processMessages processes Claude messages and builds Kiro history
func processMessages(messages gjson.Result, modelID, origin string) ([]KiroHistoryMessage, *KiroUserInputMessage, []KiroToolResult) {
var history []KiroHistoryMessage
var currentUserMsg *KiroUserInputMessage
var currentToolResults []KiroToolResult
// Merge adjacent messages with the same role
messagesArray := kirocommon.MergeAdjacentMessages(messages.Array())
for i, msg := range messagesArray {
role := msg.Get("role").String()
isLastMessage := i == len(messagesArray)-1
if role == "user" {
userMsg, toolResults := BuildUserMessageStruct(msg, modelID, origin)
if isLastMessage {
currentUserMsg = &userMsg
currentToolResults = toolResults
} else {
// CRITICAL: Kiro API requires content to be non-empty for history messages too
if strings.TrimSpace(userMsg.Content) == "" {
if len(toolResults) > 0 {
userMsg.Content = "Tool results provided."
} else {
userMsg.Content = "Continue"
}
}
// For history messages, embed tool results in context
if len(toolResults) > 0 {
userMsg.UserInputMessageContext = &KiroUserInputMessageContext{
ToolResults: toolResults,
}
}
history = append(history, KiroHistoryMessage{
UserInputMessage: &userMsg,
})
}
} else if role == "assistant" {
assistantMsg := BuildAssistantMessageStruct(msg)
if isLastMessage {
history = append(history, KiroHistoryMessage{
AssistantResponseMessage: &assistantMsg,
})
// Create a "Continue" user message as currentMessage
currentUserMsg = &KiroUserInputMessage{
Content: "Continue",
ModelID: modelID,
Origin: origin,
}
} else {
history = append(history, KiroHistoryMessage{
AssistantResponseMessage: &assistantMsg,
})
}
}
}
return history, currentUserMsg, currentToolResults
}
// buildFinalContent builds the final content with system prompt
func buildFinalContent(content, systemPrompt string, toolResults []KiroToolResult) string {
var contentBuilder strings.Builder
if systemPrompt != "" {
contentBuilder.WriteString("--- SYSTEM PROMPT ---\n")
contentBuilder.WriteString(systemPrompt)
contentBuilder.WriteString("\n--- END SYSTEM PROMPT ---\n\n")
}
contentBuilder.WriteString(content)
finalContent := contentBuilder.String()
// CRITICAL: Kiro API requires content to be non-empty
if strings.TrimSpace(finalContent) == "" {
if len(toolResults) > 0 {
finalContent = "Tool results provided."
} else {
finalContent = "Continue"
}
log.Debugf("kiro: content was empty, using default: %s", finalContent)
}
return finalContent
}
// deduplicateToolResults removes duplicate tool results
func deduplicateToolResults(toolResults []KiroToolResult) []KiroToolResult {
if len(toolResults) == 0 {
return toolResults
}
seenIDs := make(map[string]bool)
unique := make([]KiroToolResult, 0, len(toolResults))
for _, tr := range toolResults {
if !seenIDs[tr.ToolUseID] {
seenIDs[tr.ToolUseID] = true
unique = append(unique, tr)
} else {
log.Debugf("kiro: skipping duplicate toolResult in currentMessage: %s", tr.ToolUseID)
}
}
return unique
}
// extractClaudeToolChoiceHint extracts tool_choice from Claude request and returns a system prompt hint.
// Claude tool_choice values:
// - {"type": "auto"}: Model decides (default, no hint needed)
// - {"type": "any"}: Must use at least one tool
// - {"type": "tool", "name": "..."}: Must use specific tool
func extractClaudeToolChoiceHint(claudeBody []byte) string {
toolChoice := gjson.GetBytes(claudeBody, "tool_choice")
if !toolChoice.Exists() {
return ""
}
toolChoiceType := toolChoice.Get("type").String()
switch toolChoiceType {
case "any":
return "[INSTRUCTION: You MUST use at least one of the available tools to respond. Do not respond with text only - always make a tool call.]"
case "tool":
toolName := toolChoice.Get("name").String()
if toolName != "" {
return fmt.Sprintf("[INSTRUCTION: You MUST use the tool named '%s' to respond. Do not use any other tool or respond with text only.]", toolName)
}
case "auto":
// Default behavior, no hint needed
return ""
}
return ""
}
// BuildUserMessageStruct builds a user message and extracts tool results
func BuildUserMessageStruct(msg gjson.Result, modelID, origin string) (KiroUserInputMessage, []KiroToolResult) {
content := msg.Get("content")
var contentBuilder strings.Builder
var toolResults []KiroToolResult
var images []KiroImage
// Track seen toolUseIds to deduplicate
seenToolUseIDs := make(map[string]bool)
if content.IsArray() {
for _, part := range content.Array() {
partType := part.Get("type").String()
switch partType {
case "text":
contentBuilder.WriteString(part.Get("text").String())
case "image":
mediaType := part.Get("source.media_type").String()
data := part.Get("source.data").String()
format := ""
if idx := strings.LastIndex(mediaType, "/"); idx != -1 {
format = mediaType[idx+1:]
}
if format != "" && data != "" {
images = append(images, KiroImage{
Format: format,
Source: KiroImageSource{
Bytes: data,
},
})
}
case "tool_result":
toolUseID := part.Get("tool_use_id").String()
// Skip duplicate toolUseIds
if seenToolUseIDs[toolUseID] {
log.Debugf("kiro: skipping duplicate tool_result with toolUseId: %s", toolUseID)
continue
}
seenToolUseIDs[toolUseID] = true
isError := part.Get("is_error").Bool()
resultContent := part.Get("content")
var textContents []KiroTextContent
if resultContent.IsArray() {
for _, item := range resultContent.Array() {
if item.Get("type").String() == "text" {
textContents = append(textContents, KiroTextContent{Text: item.Get("text").String()})
} else if item.Type == gjson.String {
textContents = append(textContents, KiroTextContent{Text: item.String()})
}
}
} else if resultContent.Type == gjson.String {
textContents = append(textContents, KiroTextContent{Text: resultContent.String()})
}
if len(textContents) == 0 {
textContents = append(textContents, KiroTextContent{Text: "Tool use was cancelled by the user"})
}
status := "success"
if isError {
status = "error"
}
toolResults = append(toolResults, KiroToolResult{
ToolUseID: toolUseID,
Content: textContents,
Status: status,
})
}
}
} else {
contentBuilder.WriteString(content.String())
}
userMsg := KiroUserInputMessage{
Content: contentBuilder.String(),
ModelID: modelID,
Origin: origin,
}
if len(images) > 0 {
userMsg.Images = images
}
return userMsg, toolResults
}
// BuildAssistantMessageStruct builds an assistant message with tool uses
func BuildAssistantMessageStruct(msg gjson.Result) KiroAssistantResponseMessage {
content := msg.Get("content")
var contentBuilder strings.Builder
var toolUses []KiroToolUse
if content.IsArray() {
for _, part := range content.Array() {
partType := part.Get("type").String()
switch partType {
case "text":
contentBuilder.WriteString(part.Get("text").String())
case "tool_use":
toolUseID := part.Get("id").String()
toolName := part.Get("name").String()
toolInput := part.Get("input")
var inputMap map[string]interface{}
if toolInput.IsObject() {
inputMap = make(map[string]interface{})
toolInput.ForEach(func(key, value gjson.Result) bool {
inputMap[key.String()] = value.Value()
return true
})
}
toolUses = append(toolUses, KiroToolUse{
ToolUseID: toolUseID,
Name: toolName,
Input: inputMap,
})
}
}
} else {
contentBuilder.WriteString(content.String())
}
return KiroAssistantResponseMessage{
Content: contentBuilder.String(),
ToolUses: toolUses,
}
}

View File

@@ -0,0 +1,204 @@
// Package claude provides response translation functionality for Kiro API to Claude format.
// This package handles the conversion of Kiro API responses into Claude-compatible format,
// including support for thinking blocks and tool use.
package claude
import (
"crypto/sha256"
"encoding/base64"
"encoding/json"
"strings"
"github.com/google/uuid"
"github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/usage"
log "github.com/sirupsen/logrus"
kirocommon "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/common"
)
// generateThinkingSignature generates a signature for thinking content.
// This is required by Claude API for thinking blocks in non-streaming responses.
// The signature is a base64-encoded hash of the thinking content.
func generateThinkingSignature(thinkingContent string) string {
if thinkingContent == "" {
return ""
}
// Generate a deterministic signature based on content hash
hash := sha256.Sum256([]byte(thinkingContent))
return base64.StdEncoding.EncodeToString(hash[:])
}
// Local references to kirocommon constants for thinking block parsing
var (
thinkingStartTag = kirocommon.ThinkingStartTag
thinkingEndTag = kirocommon.ThinkingEndTag
)
// BuildClaudeResponse constructs a Claude-compatible response.
// Supports tool_use blocks when tools are present in the response.
// Supports thinking blocks - parses <thinking> tags and converts to Claude thinking content blocks.
// stopReason is passed from upstream; fallback logic applied if empty.
func BuildClaudeResponse(content string, toolUses []KiroToolUse, model string, usageInfo usage.Detail, stopReason string) []byte {
var contentBlocks []map[string]interface{}
// Extract thinking blocks and text from content
if content != "" {
blocks := ExtractThinkingFromContent(content)
contentBlocks = append(contentBlocks, blocks...)
// Log if thinking blocks were extracted
for _, block := range blocks {
if block["type"] == "thinking" {
thinkingContent := block["thinking"].(string)
log.Infof("kiro: buildClaudeResponse extracted thinking block (len: %d)", len(thinkingContent))
}
}
}
// Add tool_use blocks
for _, toolUse := range toolUses {
contentBlocks = append(contentBlocks, map[string]interface{}{
"type": "tool_use",
"id": toolUse.ToolUseID,
"name": toolUse.Name,
"input": toolUse.Input,
})
}
// Ensure at least one content block (Claude API requires non-empty content)
if len(contentBlocks) == 0 {
contentBlocks = append(contentBlocks, map[string]interface{}{
"type": "text",
"text": "",
})
}
// Use upstream stopReason; apply fallback logic if not provided
if stopReason == "" {
stopReason = "end_turn"
if len(toolUses) > 0 {
stopReason = "tool_use"
}
log.Debugf("kiro: buildClaudeResponse using fallback stop_reason: %s", stopReason)
}
// Log warning if response was truncated due to max_tokens
if stopReason == "max_tokens" {
log.Warnf("kiro: response truncated due to max_tokens limit (buildClaudeResponse)")
}
response := map[string]interface{}{
"id": "msg_" + uuid.New().String()[:24],
"type": "message",
"role": "assistant",
"model": model,
"content": contentBlocks,
"stop_reason": stopReason,
"usage": map[string]interface{}{
"input_tokens": usageInfo.InputTokens,
"output_tokens": usageInfo.OutputTokens,
},
}
result, _ := json.Marshal(response)
return result
}
// ExtractThinkingFromContent parses content to extract thinking blocks and text.
// Returns a list of content blocks in the order they appear in the content.
// Handles interleaved thinking and text blocks correctly.
func ExtractThinkingFromContent(content string) []map[string]interface{} {
var blocks []map[string]interface{}
if content == "" {
return blocks
}
// Check if content contains thinking tags at all
if !strings.Contains(content, thinkingStartTag) {
// No thinking tags, return as plain text
return []map[string]interface{}{
{
"type": "text",
"text": content,
},
}
}
log.Debugf("kiro: extractThinkingFromContent - found thinking tags in content (len: %d)", len(content))
remaining := content
for len(remaining) > 0 {
// Look for <thinking> tag
startIdx := strings.Index(remaining, thinkingStartTag)
if startIdx == -1 {
// No more thinking tags, add remaining as text
if strings.TrimSpace(remaining) != "" {
blocks = append(blocks, map[string]interface{}{
"type": "text",
"text": remaining,
})
}
break
}
// Add text before thinking tag (if any meaningful content)
if startIdx > 0 {
textBefore := remaining[:startIdx]
if strings.TrimSpace(textBefore) != "" {
blocks = append(blocks, map[string]interface{}{
"type": "text",
"text": textBefore,
})
}
}
// Move past the opening tag
remaining = remaining[startIdx+len(thinkingStartTag):]
// Find closing tag
endIdx := strings.Index(remaining, thinkingEndTag)
if endIdx == -1 {
// No closing tag found, treat rest as thinking content (incomplete response)
if strings.TrimSpace(remaining) != "" {
// Generate signature for thinking content (required by Claude API)
signature := generateThinkingSignature(remaining)
blocks = append(blocks, map[string]interface{}{
"type": "thinking",
"thinking": remaining,
"signature": signature,
})
log.Warnf("kiro: extractThinkingFromContent - missing closing </thinking> tag")
}
break
}
// Extract thinking content between tags
thinkContent := remaining[:endIdx]
if strings.TrimSpace(thinkContent) != "" {
// Generate signature for thinking content (required by Claude API)
signature := generateThinkingSignature(thinkContent)
blocks = append(blocks, map[string]interface{}{
"type": "thinking",
"thinking": thinkContent,
"signature": signature,
})
log.Debugf("kiro: extractThinkingFromContent - extracted thinking block (len: %d)", len(thinkContent))
}
// Move past the closing tag
remaining = remaining[endIdx+len(thinkingEndTag):]
}
// If no blocks were created (all whitespace), return empty text block
if len(blocks) == 0 {
blocks = append(blocks, map[string]interface{}{
"type": "text",
"text": "",
})
}
return blocks
}

View File

@@ -0,0 +1,186 @@
// Package claude provides streaming SSE event building for Claude format.
// This package handles the construction of Claude-compatible Server-Sent Events (SSE)
// for streaming responses from Kiro API.
package claude
import (
"encoding/json"
"github.com/google/uuid"
"github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/usage"
)
// BuildClaudeMessageStartEvent creates the message_start SSE event
func BuildClaudeMessageStartEvent(model string, inputTokens int64) []byte {
event := map[string]interface{}{
"type": "message_start",
"message": map[string]interface{}{
"id": "msg_" + uuid.New().String()[:24],
"type": "message",
"role": "assistant",
"content": []interface{}{},
"model": model,
"stop_reason": nil,
"stop_sequence": nil,
"usage": map[string]interface{}{"input_tokens": inputTokens, "output_tokens": 0},
},
}
result, _ := json.Marshal(event)
return []byte("event: message_start\ndata: " + string(result))
}
// BuildClaudeContentBlockStartEvent creates a content_block_start SSE event
func BuildClaudeContentBlockStartEvent(index int, blockType, toolUseID, toolName string) []byte {
var contentBlock map[string]interface{}
switch blockType {
case "tool_use":
contentBlock = map[string]interface{}{
"type": "tool_use",
"id": toolUseID,
"name": toolName,
"input": map[string]interface{}{},
}
case "thinking":
contentBlock = map[string]interface{}{
"type": "thinking",
"thinking": "",
}
default:
contentBlock = map[string]interface{}{
"type": "text",
"text": "",
}
}
event := map[string]interface{}{
"type": "content_block_start",
"index": index,
"content_block": contentBlock,
}
result, _ := json.Marshal(event)
return []byte("event: content_block_start\ndata: " + string(result))
}
// BuildClaudeStreamEvent creates a text_delta content_block_delta SSE event
func BuildClaudeStreamEvent(contentDelta string, index int) []byte {
event := map[string]interface{}{
"type": "content_block_delta",
"index": index,
"delta": map[string]interface{}{
"type": "text_delta",
"text": contentDelta,
},
}
result, _ := json.Marshal(event)
return []byte("event: content_block_delta\ndata: " + string(result))
}
// BuildClaudeInputJsonDeltaEvent creates an input_json_delta event for tool use streaming
func BuildClaudeInputJsonDeltaEvent(partialJSON string, index int) []byte {
event := map[string]interface{}{
"type": "content_block_delta",
"index": index,
"delta": map[string]interface{}{
"type": "input_json_delta",
"partial_json": partialJSON,
},
}
result, _ := json.Marshal(event)
return []byte("event: content_block_delta\ndata: " + string(result))
}
// BuildClaudeContentBlockStopEvent creates a content_block_stop SSE event
func BuildClaudeContentBlockStopEvent(index int) []byte {
event := map[string]interface{}{
"type": "content_block_stop",
"index": index,
}
result, _ := json.Marshal(event)
return []byte("event: content_block_stop\ndata: " + string(result))
}
// BuildClaudeThinkingBlockStopEvent creates a content_block_stop SSE event for thinking blocks.
func BuildClaudeThinkingBlockStopEvent(index int) []byte {
event := map[string]interface{}{
"type": "content_block_stop",
"index": index,
}
result, _ := json.Marshal(event)
return []byte("event: content_block_stop\ndata: " + string(result))
}
// BuildClaudeMessageDeltaEvent creates the message_delta event with stop_reason and usage
func BuildClaudeMessageDeltaEvent(stopReason string, usageInfo usage.Detail) []byte {
deltaEvent := map[string]interface{}{
"type": "message_delta",
"delta": map[string]interface{}{
"stop_reason": stopReason,
"stop_sequence": nil,
},
"usage": map[string]interface{}{
"input_tokens": usageInfo.InputTokens,
"output_tokens": usageInfo.OutputTokens,
},
}
deltaResult, _ := json.Marshal(deltaEvent)
return []byte("event: message_delta\ndata: " + string(deltaResult))
}
// BuildClaudeMessageStopOnlyEvent creates only the message_stop event
func BuildClaudeMessageStopOnlyEvent() []byte {
stopEvent := map[string]interface{}{
"type": "message_stop",
}
stopResult, _ := json.Marshal(stopEvent)
return []byte("event: message_stop\ndata: " + string(stopResult))
}
// BuildClaudePingEventWithUsage creates a ping event with embedded usage information.
// This is used for real-time usage estimation during streaming.
func BuildClaudePingEventWithUsage(inputTokens, outputTokens int64) []byte {
event := map[string]interface{}{
"type": "ping",
"usage": map[string]interface{}{
"input_tokens": inputTokens,
"output_tokens": outputTokens,
"total_tokens": inputTokens + outputTokens,
"estimated": true,
},
}
result, _ := json.Marshal(event)
return []byte("event: ping\ndata: " + string(result))
}
// BuildClaudeThinkingDeltaEvent creates a thinking_delta event for Claude API compatibility.
// This is used when streaming thinking content wrapped in <thinking> tags.
func BuildClaudeThinkingDeltaEvent(thinkingDelta string, index int) []byte {
event := map[string]interface{}{
"type": "content_block_delta",
"index": index,
"delta": map[string]interface{}{
"type": "thinking_delta",
"thinking": thinkingDelta,
},
}
result, _ := json.Marshal(event)
return []byte("event: content_block_delta\ndata: " + string(result))
}
// PendingTagSuffix detects if the buffer ends with a partial prefix of the given tag.
// Returns the length of the partial match (0 if no match).
// Based on amq2api implementation for handling cross-chunk tag boundaries.
func PendingTagSuffix(buffer, tag string) int {
if buffer == "" || tag == "" {
return 0
}
maxLen := len(buffer)
if maxLen > len(tag)-1 {
maxLen = len(tag) - 1
}
for length := maxLen; length > 0; length-- {
if len(buffer) >= length && buffer[len(buffer)-length:] == tag[:length] {
return length
}
}
return 0
}

View File

@@ -0,0 +1,522 @@
// Package claude provides tool calling support for Kiro to Claude translation.
// This package handles parsing embedded tool calls, JSON repair, and deduplication.
package claude
import (
"encoding/json"
"regexp"
"strings"
"github.com/google/uuid"
kirocommon "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/common"
log "github.com/sirupsen/logrus"
)
// ToolUseState tracks the state of an in-progress tool use during streaming.
type ToolUseState struct {
ToolUseID string
Name string
InputBuffer strings.Builder
IsComplete bool
}
// Pre-compiled regex patterns for performance
var (
// embeddedToolCallPattern matches [Called tool_name with args: {...}] format
embeddedToolCallPattern = regexp.MustCompile(`\[Called\s+([A-Za-z0-9_.-]+)\s+with\s+args:\s*`)
// trailingCommaPattern matches trailing commas before closing braces/brackets
trailingCommaPattern = regexp.MustCompile(`,\s*([}\]])`)
)
// ParseEmbeddedToolCalls extracts [Called tool_name with args: {...}] format from text.
// Kiro sometimes embeds tool calls in text content instead of using toolUseEvent.
// Returns the cleaned text (with tool calls removed) and extracted tool uses.
func ParseEmbeddedToolCalls(text string, processedIDs map[string]bool) (string, []KiroToolUse) {
if !strings.Contains(text, "[Called") {
return text, nil
}
var toolUses []KiroToolUse
cleanText := text
// Find all [Called markers
matches := embeddedToolCallPattern.FindAllStringSubmatchIndex(text, -1)
if len(matches) == 0 {
return text, nil
}
// Process matches in reverse order to maintain correct indices
for i := len(matches) - 1; i >= 0; i-- {
matchStart := matches[i][0]
toolNameStart := matches[i][2]
toolNameEnd := matches[i][3]
if toolNameStart < 0 || toolNameEnd < 0 {
continue
}
toolName := text[toolNameStart:toolNameEnd]
// Find the JSON object start (after "with args:")
jsonStart := matches[i][1]
if jsonStart >= len(text) {
continue
}
// Skip whitespace to find the opening brace
for jsonStart < len(text) && (text[jsonStart] == ' ' || text[jsonStart] == '\t') {
jsonStart++
}
if jsonStart >= len(text) || text[jsonStart] != '{' {
continue
}
// Find matching closing bracket
jsonEnd := findMatchingBracket(text, jsonStart)
if jsonEnd < 0 {
continue
}
// Extract JSON and find the closing bracket of [Called ...]
jsonStr := text[jsonStart : jsonEnd+1]
// Find the closing ] after the JSON
closingBracket := jsonEnd + 1
for closingBracket < len(text) && text[closingBracket] != ']' {
closingBracket++
}
if closingBracket >= len(text) {
continue
}
// End index of the full tool call (closing ']' inclusive)
matchEnd := closingBracket + 1
// Repair and parse JSON
repairedJSON := RepairJSON(jsonStr)
var inputMap map[string]interface{}
if err := json.Unmarshal([]byte(repairedJSON), &inputMap); err != nil {
log.Debugf("kiro: failed to parse embedded tool call JSON: %v, raw: %s", err, jsonStr)
continue
}
// Generate unique tool ID
toolUseID := "toolu_" + uuid.New().String()[:12]
// Check for duplicates using name+input as key
dedupeKey := toolName + ":" + repairedJSON
if processedIDs != nil {
if processedIDs[dedupeKey] {
log.Debugf("kiro: skipping duplicate embedded tool call: %s", toolName)
// Still remove from text even if duplicate
if matchStart >= 0 && matchEnd <= len(cleanText) && matchStart <= matchEnd {
cleanText = cleanText[:matchStart] + cleanText[matchEnd:]
}
continue
}
processedIDs[dedupeKey] = true
}
toolUses = append(toolUses, KiroToolUse{
ToolUseID: toolUseID,
Name: toolName,
Input: inputMap,
})
log.Infof("kiro: extracted embedded tool call: %s (ID: %s)", toolName, toolUseID)
// Remove from clean text (index-based removal to avoid deleting the wrong occurrence)
if matchStart >= 0 && matchEnd <= len(cleanText) && matchStart <= matchEnd {
cleanText = cleanText[:matchStart] + cleanText[matchEnd:]
}
}
return cleanText, toolUses
}
// findMatchingBracket finds the index of the closing brace/bracket that matches
// the opening one at startPos. Handles nested objects and strings correctly.
func findMatchingBracket(text string, startPos int) int {
if startPos >= len(text) {
return -1
}
openChar := text[startPos]
var closeChar byte
switch openChar {
case '{':
closeChar = '}'
case '[':
closeChar = ']'
default:
return -1
}
depth := 1
inString := false
escapeNext := false
for i := startPos + 1; i < len(text); i++ {
char := text[i]
if escapeNext {
escapeNext = false
continue
}
if char == '\\' && inString {
escapeNext = true
continue
}
if char == '"' {
inString = !inString
continue
}
if !inString {
if char == openChar {
depth++
} else if char == closeChar {
depth--
if depth == 0 {
return i
}
}
}
}
return -1
}
// RepairJSON attempts to fix common JSON issues that may occur in tool call arguments.
// Conservative repair strategy:
// 1. First try to parse JSON directly - if valid, return as-is
// 2. Only attempt repair if parsing fails
// 3. After repair, validate the result - if still invalid, return original
func RepairJSON(jsonString string) string {
// Handle empty or invalid input
if jsonString == "" {
return "{}"
}
str := strings.TrimSpace(jsonString)
if str == "" {
return "{}"
}
// CONSERVATIVE STRATEGY: First try to parse directly
var testParse interface{}
if err := json.Unmarshal([]byte(str), &testParse); err == nil {
log.Debugf("kiro: repairJSON - JSON is already valid, returning unchanged")
return str
}
log.Debugf("kiro: repairJSON - JSON parse failed, attempting repair")
originalStr := str
// First, escape unescaped newlines/tabs within JSON string values
str = escapeNewlinesInStrings(str)
// Remove trailing commas before closing braces/brackets
str = trailingCommaPattern.ReplaceAllString(str, "$1")
// Calculate bracket balance
braceCount := 0
bracketCount := 0
inString := false
escape := false
lastValidIndex := -1
for i := 0; i < len(str); i++ {
char := str[i]
if escape {
escape = false
continue
}
if char == '\\' {
escape = true
continue
}
if char == '"' {
inString = !inString
continue
}
if inString {
continue
}
switch char {
case '{':
braceCount++
case '}':
braceCount--
case '[':
bracketCount++
case ']':
bracketCount--
}
if braceCount >= 0 && bracketCount >= 0 {
lastValidIndex = i
}
}
// If brackets are unbalanced, try to repair
if braceCount > 0 || bracketCount > 0 {
if lastValidIndex > 0 && lastValidIndex < len(str)-1 {
truncated := str[:lastValidIndex+1]
// Recount brackets after truncation
braceCount = 0
bracketCount = 0
inString = false
escape = false
for i := 0; i < len(truncated); i++ {
char := truncated[i]
if escape {
escape = false
continue
}
if char == '\\' {
escape = true
continue
}
if char == '"' {
inString = !inString
continue
}
if inString {
continue
}
switch char {
case '{':
braceCount++
case '}':
braceCount--
case '[':
bracketCount++
case ']':
bracketCount--
}
}
str = truncated
}
// Add missing closing brackets
for braceCount > 0 {
str += "}"
braceCount--
}
for bracketCount > 0 {
str += "]"
bracketCount--
}
}
// Validate repaired JSON
if err := json.Unmarshal([]byte(str), &testParse); err != nil {
log.Warnf("kiro: repairJSON - repair failed to produce valid JSON, returning original")
return originalStr
}
log.Debugf("kiro: repairJSON - successfully repaired JSON")
return str
}
// escapeNewlinesInStrings escapes literal newlines, tabs, and other control characters
// that appear inside JSON string values.
func escapeNewlinesInStrings(raw string) string {
var result strings.Builder
result.Grow(len(raw) + 100)
inString := false
escaped := false
for i := 0; i < len(raw); i++ {
c := raw[i]
if escaped {
result.WriteByte(c)
escaped = false
continue
}
if c == '\\' && inString {
result.WriteByte(c)
escaped = true
continue
}
if c == '"' {
inString = !inString
result.WriteByte(c)
continue
}
if inString {
switch c {
case '\n':
result.WriteString("\\n")
case '\r':
result.WriteString("\\r")
case '\t':
result.WriteString("\\t")
default:
result.WriteByte(c)
}
} else {
result.WriteByte(c)
}
}
return result.String()
}
// ProcessToolUseEvent handles a toolUseEvent from the Kiro stream.
// It accumulates input fragments and emits tool_use blocks when complete.
// Returns events to emit and updated state.
func ProcessToolUseEvent(event map[string]interface{}, currentToolUse *ToolUseState, processedIDs map[string]bool) ([]KiroToolUse, *ToolUseState) {
var toolUses []KiroToolUse
// Extract from nested toolUseEvent or direct format
tu := event
if nested, ok := event["toolUseEvent"].(map[string]interface{}); ok {
tu = nested
}
toolUseID := kirocommon.GetString(tu, "toolUseId")
toolName := kirocommon.GetString(tu, "name")
isStop := false
if stop, ok := tu["stop"].(bool); ok {
isStop = stop
}
// Get input - can be string (fragment) or object (complete)
var inputFragment string
var inputMap map[string]interface{}
if inputRaw, ok := tu["input"]; ok {
switch v := inputRaw.(type) {
case string:
inputFragment = v
case map[string]interface{}:
inputMap = v
}
}
// New tool use starting
if toolUseID != "" && toolName != "" {
if currentToolUse != nil && currentToolUse.ToolUseID != toolUseID {
log.Warnf("kiro: interleaved tool use detected - new ID %s arrived while %s in progress, completing previous",
toolUseID, currentToolUse.ToolUseID)
if !processedIDs[currentToolUse.ToolUseID] {
incomplete := KiroToolUse{
ToolUseID: currentToolUse.ToolUseID,
Name: currentToolUse.Name,
}
if currentToolUse.InputBuffer.Len() > 0 {
raw := currentToolUse.InputBuffer.String()
repaired := RepairJSON(raw)
var input map[string]interface{}
if err := json.Unmarshal([]byte(repaired), &input); err != nil {
log.Warnf("kiro: failed to parse interleaved tool input: %v, raw: %s", err, raw)
input = make(map[string]interface{})
}
incomplete.Input = input
}
toolUses = append(toolUses, incomplete)
processedIDs[currentToolUse.ToolUseID] = true
}
currentToolUse = nil
}
if currentToolUse == nil {
if processedIDs != nil && processedIDs[toolUseID] {
log.Debugf("kiro: skipping duplicate toolUseEvent: %s", toolUseID)
return nil, nil
}
currentToolUse = &ToolUseState{
ToolUseID: toolUseID,
Name: toolName,
}
log.Infof("kiro: starting new tool use: %s (ID: %s)", toolName, toolUseID)
}
}
// Accumulate input fragments
if currentToolUse != nil && inputFragment != "" {
currentToolUse.InputBuffer.WriteString(inputFragment)
log.Debugf("kiro: accumulated input fragment, total length: %d", currentToolUse.InputBuffer.Len())
}
// If complete input object provided directly
if currentToolUse != nil && inputMap != nil {
inputBytes, _ := json.Marshal(inputMap)
currentToolUse.InputBuffer.Reset()
currentToolUse.InputBuffer.Write(inputBytes)
}
// Tool use complete
if isStop && currentToolUse != nil {
fullInput := currentToolUse.InputBuffer.String()
// Repair and parse the accumulated JSON
repairedJSON := RepairJSON(fullInput)
var finalInput map[string]interface{}
if err := json.Unmarshal([]byte(repairedJSON), &finalInput); err != nil {
log.Warnf("kiro: failed to parse accumulated tool input: %v, raw: %s", err, fullInput)
finalInput = make(map[string]interface{})
}
toolUse := KiroToolUse{
ToolUseID: currentToolUse.ToolUseID,
Name: currentToolUse.Name,
Input: finalInput,
}
toolUses = append(toolUses, toolUse)
if processedIDs != nil {
processedIDs[currentToolUse.ToolUseID] = true
}
log.Infof("kiro: completed tool use: %s (ID: %s)", currentToolUse.Name, currentToolUse.ToolUseID)
return toolUses, nil
}
return toolUses, currentToolUse
}
// DeduplicateToolUses removes duplicate tool uses based on toolUseId and content.
func DeduplicateToolUses(toolUses []KiroToolUse) []KiroToolUse {
seenIDs := make(map[string]bool)
seenContent := make(map[string]bool)
var unique []KiroToolUse
for _, tu := range toolUses {
if seenIDs[tu.ToolUseID] {
log.Debugf("kiro: removing ID-duplicate tool use: %s (name: %s)", tu.ToolUseID, tu.Name)
continue
}
inputJSON, _ := json.Marshal(tu.Input)
contentKey := tu.Name + ":" + string(inputJSON)
if seenContent[contentKey] {
log.Debugf("kiro: removing content-duplicate tool use: %s (id: %s)", tu.Name, tu.ToolUseID)
continue
}
seenIDs[tu.ToolUseID] = true
seenContent[contentKey] = true
unique = append(unique, tu)
}
return unique
}

View File

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

View File

@@ -0,0 +1,83 @@
// Package common provides shared constants and utilities for Kiro translator.
package common
const (
// KiroMaxToolDescLen is the maximum description length for Kiro API tools.
// Kiro API limit is 10240 bytes, leave room for "..."
KiroMaxToolDescLen = 10237
// ToolCompressionTargetSize is the target total size for compressed tools (20KB).
// If tools exceed this size, compression will be applied.
ToolCompressionTargetSize = 20 * 1024 // 20KB
// MinToolDescriptionLength is the minimum description length after compression.
// Descriptions will not be shortened below this length.
MinToolDescriptionLength = 50
// ThinkingStartTag is the start tag for thinking blocks in responses.
ThinkingStartTag = "<thinking>"
// ThinkingEndTag is the end tag for thinking blocks in responses.
ThinkingEndTag = "</thinking>"
// CodeFenceMarker is the markdown code fence marker.
CodeFenceMarker = "```"
// AltCodeFenceMarker is the alternative markdown code fence marker.
AltCodeFenceMarker = "~~~"
// InlineCodeMarker is the markdown inline code marker (backtick).
InlineCodeMarker = "`"
// KiroAgenticSystemPrompt is injected only for -agentic models to prevent timeouts on large writes.
// AWS Kiro API has a 2-3 minute timeout for large file write operations.
KiroAgenticSystemPrompt = `
# CRITICAL: CHUNKED WRITE PROTOCOL (MANDATORY)
You MUST follow these rules for ALL file operations. Violation causes server timeouts and task failure.
## ABSOLUTE LIMITS
- **MAXIMUM 350 LINES** per single write/edit operation - NO EXCEPTIONS
- **RECOMMENDED 300 LINES** or less for optimal performance
- **NEVER** write entire files in one operation if >300 lines
## MANDATORY CHUNKED WRITE STRATEGY
### For NEW FILES (>300 lines total):
1. FIRST: Write initial chunk (first 250-300 lines) using write_to_file/fsWrite
2. THEN: Append remaining content in 250-300 line chunks using file append operations
3. REPEAT: Continue appending until complete
### For EDITING EXISTING FILES:
1. Use surgical edits (apply_diff/targeted edits) - change ONLY what's needed
2. NEVER rewrite entire files - use incremental modifications
3. Split large refactors into multiple small, focused edits
### For LARGE CODE GENERATION:
1. Generate in logical sections (imports, types, functions separately)
2. Write each section as a separate operation
3. Use append operations for subsequent sections
## EXAMPLES OF CORRECT BEHAVIOR
✅ CORRECT: Writing a 600-line file
- Operation 1: Write lines 1-300 (initial file creation)
- Operation 2: Append lines 301-600
✅ CORRECT: Editing multiple functions
- Operation 1: Edit function A
- Operation 2: Edit function B
- Operation 3: Edit function C
❌ WRONG: Writing 500 lines in single operation → TIMEOUT
❌ WRONG: Rewriting entire file to change 5 lines → TIMEOUT
❌ WRONG: Generating massive code blocks without chunking → TIMEOUT
## WHY THIS MATTERS
- Server has 2-3 minute timeout for operations
- Large writes exceed timeout and FAIL completely
- Chunked writes are FASTER and more RELIABLE
- Failed writes waste time and require retry
REMEMBER: When in doubt, write LESS per operation. Multiple small operations > one large operation.`
)

View File

@@ -0,0 +1,132 @@
// Package common provides shared utilities for Kiro translators.
package common
import (
"encoding/json"
"github.com/tidwall/gjson"
)
// MergeAdjacentMessages merges adjacent messages with the same role.
// This reduces API call complexity and improves compatibility.
// Based on AIClient-2-API implementation.
// NOTE: Tool messages are NOT merged because each has a unique tool_call_id that must be preserved.
func MergeAdjacentMessages(messages []gjson.Result) []gjson.Result {
if len(messages) <= 1 {
return messages
}
var merged []gjson.Result
for _, msg := range messages {
if len(merged) == 0 {
merged = append(merged, msg)
continue
}
lastMsg := merged[len(merged)-1]
currentRole := msg.Get("role").String()
lastRole := lastMsg.Get("role").String()
// Don't merge tool messages - each has a unique tool_call_id
if currentRole == "tool" || lastRole == "tool" {
merged = append(merged, msg)
continue
}
if currentRole == lastRole {
// Merge content from current message into last message
mergedContent := mergeMessageContent(lastMsg, msg)
// Create a new merged message JSON
mergedMsg := createMergedMessage(lastRole, mergedContent)
merged[len(merged)-1] = gjson.Parse(mergedMsg)
} else {
merged = append(merged, msg)
}
}
return merged
}
// mergeMessageContent merges the content of two messages with the same role.
// Handles both string content and array content (with text, tool_use, tool_result blocks).
func mergeMessageContent(msg1, msg2 gjson.Result) string {
content1 := msg1.Get("content")
content2 := msg2.Get("content")
// Extract content blocks from both messages
var blocks1, blocks2 []map[string]interface{}
if content1.IsArray() {
for _, block := range content1.Array() {
blocks1 = append(blocks1, blockToMap(block))
}
} else if content1.Type == gjson.String {
blocks1 = append(blocks1, map[string]interface{}{
"type": "text",
"text": content1.String(),
})
}
if content2.IsArray() {
for _, block := range content2.Array() {
blocks2 = append(blocks2, blockToMap(block))
}
} else if content2.Type == gjson.String {
blocks2 = append(blocks2, map[string]interface{}{
"type": "text",
"text": content2.String(),
})
}
// Merge text blocks if both end/start with text
if len(blocks1) > 0 && len(blocks2) > 0 {
if blocks1[len(blocks1)-1]["type"] == "text" && blocks2[0]["type"] == "text" {
// Merge the last text block of msg1 with the first text block of msg2
text1 := blocks1[len(blocks1)-1]["text"].(string)
text2 := blocks2[0]["text"].(string)
blocks1[len(blocks1)-1]["text"] = text1 + "\n" + text2
blocks2 = blocks2[1:] // Remove the merged block from blocks2
}
}
// Combine all blocks
allBlocks := append(blocks1, blocks2...)
// Convert to JSON
result, _ := json.Marshal(allBlocks)
return string(result)
}
// blockToMap converts a gjson.Result block to a map[string]interface{}
func blockToMap(block gjson.Result) map[string]interface{} {
result := make(map[string]interface{})
block.ForEach(func(key, value gjson.Result) bool {
if value.IsObject() {
result[key.String()] = blockToMap(value)
} else if value.IsArray() {
var arr []interface{}
for _, item := range value.Array() {
if item.IsObject() {
arr = append(arr, blockToMap(item))
} else {
arr = append(arr, item.Value())
}
}
result[key.String()] = arr
} else {
result[key.String()] = value.Value()
}
return true
})
return result
}
// createMergedMessage creates a JSON string for a merged message
func createMergedMessage(role string, content string) string {
msg := map[string]interface{}{
"role": role,
"content": json.RawMessage(content),
}
result, _ := json.Marshal(msg)
return string(result)
}

View File

@@ -0,0 +1,16 @@
// Package common provides shared constants and utilities for Kiro translator.
package common
// GetString safely extracts a string from a map.
// Returns empty string if the key doesn't exist or the value is not a string.
func GetString(m map[string]interface{}, key string) string {
if v, ok := m[key].(string); ok {
return v
}
return ""
}
// GetStringValue is an alias for GetString for backward compatibility.
func GetStringValue(m map[string]interface{}, key string) string {
return GetString(m, key)
}

View File

@@ -0,0 +1,20 @@
// Package openai provides translation between OpenAI Chat Completions and Kiro formats.
package openai
import (
. "github.com/router-for-me/CLIProxyAPI/v6/internal/constant"
"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/translator"
)
func init() {
translator.Register(
OpenAI, // source format
Kiro, // target format
ConvertOpenAIRequestToKiro,
interfaces.TranslateResponse{
Stream: ConvertKiroStreamToOpenAI,
NonStream: ConvertKiroNonStreamToOpenAI,
},
)
}

View File

@@ -0,0 +1,371 @@
// Package openai provides translation between OpenAI Chat Completions and Kiro formats.
// This package enables direct OpenAI → Kiro translation, bypassing the Claude intermediate layer.
//
// The Kiro executor generates Claude-compatible SSE format internally, so the streaming response
// translation converts from Claude SSE format to OpenAI SSE format.
package openai
import (
"bytes"
"context"
"encoding/json"
"strings"
kirocommon "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/common"
"github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/usage"
log "github.com/sirupsen/logrus"
"github.com/tidwall/gjson"
)
// ConvertKiroStreamToOpenAI converts Kiro streaming response to OpenAI format.
// The Kiro executor emits Claude-compatible SSE events, so this function translates
// from Claude SSE format to OpenAI SSE format.
//
// Claude SSE format:
// - event: message_start\ndata: {...}
// - event: content_block_start\ndata: {...}
// - event: content_block_delta\ndata: {...}
// - event: content_block_stop\ndata: {...}
// - event: message_delta\ndata: {...}
// - event: message_stop\ndata: {...}
//
// OpenAI SSE format:
// - data: {"id":"...","object":"chat.completion.chunk",...}
// - data: [DONE]
func ConvertKiroStreamToOpenAI(ctx context.Context, model string, originalRequest, request, rawResponse []byte, param *any) []string {
// Initialize state if needed
if *param == nil {
*param = NewOpenAIStreamState(model)
}
state := (*param).(*OpenAIStreamState)
// Parse the Claude SSE event
responseStr := string(rawResponse)
// Handle raw event format (event: xxx\ndata: {...})
var eventType string
var eventData string
if strings.HasPrefix(responseStr, "event:") {
// Parse event type and data
lines := strings.SplitN(responseStr, "\n", 2)
if len(lines) >= 1 {
eventType = strings.TrimSpace(strings.TrimPrefix(lines[0], "event:"))
}
if len(lines) >= 2 && strings.HasPrefix(lines[1], "data:") {
eventData = strings.TrimSpace(strings.TrimPrefix(lines[1], "data:"))
}
} else if strings.HasPrefix(responseStr, "data:") {
// Just data line
eventData = strings.TrimSpace(strings.TrimPrefix(responseStr, "data:"))
} else {
// Try to parse as raw JSON
eventData = strings.TrimSpace(responseStr)
}
if eventData == "" {
return []string{}
}
// Parse the event data as JSON
eventJSON := gjson.Parse(eventData)
if !eventJSON.Exists() {
return []string{}
}
// Determine event type from JSON if not already set
if eventType == "" {
eventType = eventJSON.Get("type").String()
}
var results []string
switch eventType {
case "message_start":
// Send first chunk with role
firstChunk := BuildOpenAISSEFirstChunk(state)
results = append(results, firstChunk)
case "content_block_start":
// Check block type
blockType := eventJSON.Get("content_block.type").String()
switch blockType {
case "text":
// Text block starting - nothing to emit yet
case "thinking":
// Thinking block starting - nothing to emit yet for OpenAI
case "tool_use":
// Tool use block starting
toolUseID := eventJSON.Get("content_block.id").String()
toolName := eventJSON.Get("content_block.name").String()
chunk := BuildOpenAISSEToolCallStart(state, toolUseID, toolName)
results = append(results, chunk)
state.ToolCallIndex++
}
case "content_block_delta":
deltaType := eventJSON.Get("delta.type").String()
switch deltaType {
case "text_delta":
textDelta := eventJSON.Get("delta.text").String()
if textDelta != "" {
chunk := BuildOpenAISSETextDelta(state, textDelta)
results = append(results, chunk)
}
case "thinking_delta":
// Convert thinking to reasoning_content for o1-style compatibility
thinkingDelta := eventJSON.Get("delta.thinking").String()
if thinkingDelta != "" {
chunk := BuildOpenAISSEReasoningDelta(state, thinkingDelta)
results = append(results, chunk)
}
case "input_json_delta":
// Tool call arguments delta
partialJSON := eventJSON.Get("delta.partial_json").String()
if partialJSON != "" {
// Get the tool index from content block index
blockIndex := int(eventJSON.Get("index").Int())
chunk := BuildOpenAISSEToolCallArgumentsDelta(state, partialJSON, blockIndex-1) // Adjust for 0-based tool index
results = append(results, chunk)
}
}
case "content_block_stop":
// Content block ended - nothing to emit for OpenAI
case "message_delta":
// Message delta with stop_reason
stopReason := eventJSON.Get("delta.stop_reason").String()
finishReason := mapKiroStopReasonToOpenAI(stopReason)
if finishReason != "" {
chunk := BuildOpenAISSEFinish(state, finishReason)
results = append(results, chunk)
}
// Extract usage if present
if eventJSON.Get("usage").Exists() {
inputTokens := eventJSON.Get("usage.input_tokens").Int()
outputTokens := eventJSON.Get("usage.output_tokens").Int()
usageInfo := usage.Detail{
InputTokens: inputTokens,
OutputTokens: outputTokens,
TotalTokens: inputTokens + outputTokens,
}
chunk := BuildOpenAISSEUsage(state, usageInfo)
results = append(results, chunk)
}
case "message_stop":
// Final event - do NOT emit [DONE] here
// The handler layer (openai_handlers.go) will send [DONE] when the stream closes
// Emitting [DONE] here would cause duplicate [DONE] markers
case "ping":
// Ping event with usage - optionally emit usage chunk
if eventJSON.Get("usage").Exists() {
inputTokens := eventJSON.Get("usage.input_tokens").Int()
outputTokens := eventJSON.Get("usage.output_tokens").Int()
usageInfo := usage.Detail{
InputTokens: inputTokens,
OutputTokens: outputTokens,
TotalTokens: inputTokens + outputTokens,
}
chunk := BuildOpenAISSEUsage(state, usageInfo)
results = append(results, chunk)
}
}
return results
}
// ConvertKiroNonStreamToOpenAI converts Kiro non-streaming response to OpenAI format.
// The Kiro executor returns Claude-compatible JSON responses, so this function translates
// from Claude format to OpenAI format.
func ConvertKiroNonStreamToOpenAI(ctx context.Context, model string, originalRequest, request, rawResponse []byte, param *any) string {
// Parse the Claude-format response
response := gjson.ParseBytes(rawResponse)
// Extract content
var content string
var reasoningContent string
var toolUses []KiroToolUse
var stopReason string
// Get stop_reason
stopReason = response.Get("stop_reason").String()
// Process content blocks
contentBlocks := response.Get("content")
if contentBlocks.IsArray() {
for _, block := range contentBlocks.Array() {
blockType := block.Get("type").String()
switch blockType {
case "text":
content += block.Get("text").String()
case "thinking":
// Convert thinking blocks to reasoning_content for OpenAI format
reasoningContent += block.Get("thinking").String()
case "tool_use":
toolUseID := block.Get("id").String()
toolName := block.Get("name").String()
toolInput := block.Get("input")
var inputMap map[string]interface{}
if toolInput.IsObject() {
inputMap = make(map[string]interface{})
toolInput.ForEach(func(key, value gjson.Result) bool {
inputMap[key.String()] = value.Value()
return true
})
}
toolUses = append(toolUses, KiroToolUse{
ToolUseID: toolUseID,
Name: toolName,
Input: inputMap,
})
}
}
}
// Extract usage
usageInfo := usage.Detail{
InputTokens: response.Get("usage.input_tokens").Int(),
OutputTokens: response.Get("usage.output_tokens").Int(),
}
usageInfo.TotalTokens = usageInfo.InputTokens + usageInfo.OutputTokens
// Build OpenAI response with reasoning_content support
openaiResponse := BuildOpenAIResponseWithReasoning(content, reasoningContent, toolUses, model, usageInfo, stopReason)
return string(openaiResponse)
}
// ParseClaudeEvent parses a Claude SSE event and returns the event type and data
func ParseClaudeEvent(rawEvent []byte) (eventType string, eventData []byte) {
lines := bytes.Split(rawEvent, []byte("\n"))
for _, line := range lines {
line = bytes.TrimSpace(line)
if bytes.HasPrefix(line, []byte("event:")) {
eventType = string(bytes.TrimSpace(bytes.TrimPrefix(line, []byte("event:"))))
} else if bytes.HasPrefix(line, []byte("data:")) {
eventData = bytes.TrimSpace(bytes.TrimPrefix(line, []byte("data:")))
}
}
return eventType, eventData
}
// ExtractThinkingFromContent parses content to extract thinking blocks.
// Returns cleaned content (without thinking tags) and whether thinking was found.
func ExtractThinkingFromContent(content string) (string, string, bool) {
if !strings.Contains(content, kirocommon.ThinkingStartTag) {
return content, "", false
}
var cleanedContent strings.Builder
var thinkingContent strings.Builder
hasThinking := false
remaining := content
for len(remaining) > 0 {
startIdx := strings.Index(remaining, kirocommon.ThinkingStartTag)
if startIdx == -1 {
cleanedContent.WriteString(remaining)
break
}
// Add content before thinking tag
cleanedContent.WriteString(remaining[:startIdx])
// Move past opening tag
remaining = remaining[startIdx+len(kirocommon.ThinkingStartTag):]
// Find closing tag
endIdx := strings.Index(remaining, kirocommon.ThinkingEndTag)
if endIdx == -1 {
// No closing tag - treat rest as thinking
thinkingContent.WriteString(remaining)
hasThinking = true
break
}
// Extract thinking content
thinkingContent.WriteString(remaining[:endIdx])
hasThinking = true
remaining = remaining[endIdx+len(kirocommon.ThinkingEndTag):]
}
return strings.TrimSpace(cleanedContent.String()), strings.TrimSpace(thinkingContent.String()), hasThinking
}
// ConvertOpenAIToolsToKiroFormat is a helper that converts OpenAI tools format to Kiro format
func ConvertOpenAIToolsToKiroFormat(tools []map[string]interface{}) []KiroToolWrapper {
var kiroTools []KiroToolWrapper
for _, tool := range tools {
toolType, _ := tool["type"].(string)
if toolType != "function" {
continue
}
fn, ok := tool["function"].(map[string]interface{})
if !ok {
continue
}
name := kirocommon.GetString(fn, "name")
description := kirocommon.GetString(fn, "description")
parameters := fn["parameters"]
if name == "" {
continue
}
if description == "" {
description = "Tool: " + name
}
kiroTools = append(kiroTools, KiroToolWrapper{
ToolSpecification: KiroToolSpecification{
Name: name,
Description: description,
InputSchema: KiroInputSchema{JSON: parameters},
},
})
}
return kiroTools
}
// OpenAIStreamParams holds parameters for OpenAI streaming conversion
type OpenAIStreamParams struct {
State *OpenAIStreamState
ThinkingState *ThinkingTagState
ToolCallsEmitted map[string]bool
}
// NewOpenAIStreamParams creates new streaming parameters
func NewOpenAIStreamParams(model string) *OpenAIStreamParams {
return &OpenAIStreamParams{
State: NewOpenAIStreamState(model),
ThinkingState: NewThinkingTagState(),
ToolCallsEmitted: make(map[string]bool),
}
}
// ConvertClaudeToolUseToOpenAI converts a Claude tool_use block to OpenAI tool_calls format
func ConvertClaudeToolUseToOpenAI(toolUseID, toolName string, input map[string]interface{}) map[string]interface{} {
inputJSON, _ := json.Marshal(input)
return map[string]interface{}{
"id": toolUseID,
"type": "function",
"function": map[string]interface{}{
"name": toolName,
"arguments": string(inputJSON),
},
}
}
// LogStreamEvent logs a streaming event for debugging
func LogStreamEvent(eventType, data string) {
log.Debugf("kiro-openai: stream event type=%s, data_len=%d", eventType, len(data))
}

View File

@@ -0,0 +1,863 @@
// Package openai provides request translation from OpenAI Chat Completions to Kiro format.
// It handles parsing and transforming OpenAI API requests into the Kiro/Amazon Q API format,
// extracting model information, system instructions, message contents, and tool declarations.
package openai
import (
"encoding/json"
"fmt"
"net/http"
"strings"
"time"
"unicode/utf8"
"github.com/google/uuid"
kiroclaude "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/claude"
kirocommon "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/kiro/common"
log "github.com/sirupsen/logrus"
"github.com/tidwall/gjson"
)
// Kiro API request structs - reuse from kiroclaude package structure
// KiroPayload is the top-level request structure for Kiro API
type KiroPayload struct {
ConversationState KiroConversationState `json:"conversationState"`
ProfileArn string `json:"profileArn,omitempty"`
InferenceConfig *KiroInferenceConfig `json:"inferenceConfig,omitempty"`
}
// KiroInferenceConfig contains inference parameters for the Kiro API.
type KiroInferenceConfig struct {
MaxTokens int `json:"maxTokens,omitempty"`
Temperature float64 `json:"temperature,omitempty"`
TopP float64 `json:"topP,omitempty"`
}
// KiroConversationState holds the conversation context
type KiroConversationState struct {
ChatTriggerType string `json:"chatTriggerType"` // Required: "MANUAL"
ConversationID string `json:"conversationId"`
CurrentMessage KiroCurrentMessage `json:"currentMessage"`
History []KiroHistoryMessage `json:"history,omitempty"`
}
// KiroCurrentMessage wraps the current user message
type KiroCurrentMessage struct {
UserInputMessage KiroUserInputMessage `json:"userInputMessage"`
}
// KiroHistoryMessage represents a message in the conversation history
type KiroHistoryMessage struct {
UserInputMessage *KiroUserInputMessage `json:"userInputMessage,omitempty"`
AssistantResponseMessage *KiroAssistantResponseMessage `json:"assistantResponseMessage,omitempty"`
}
// KiroImage represents an image in Kiro API format
type KiroImage struct {
Format string `json:"format"`
Source KiroImageSource `json:"source"`
}
// KiroImageSource contains the image data
type KiroImageSource struct {
Bytes string `json:"bytes"` // base64 encoded image data
}
// KiroUserInputMessage represents a user message
type KiroUserInputMessage struct {
Content string `json:"content"`
ModelID string `json:"modelId"`
Origin string `json:"origin"`
Images []KiroImage `json:"images,omitempty"`
UserInputMessageContext *KiroUserInputMessageContext `json:"userInputMessageContext,omitempty"`
}
// KiroUserInputMessageContext contains tool-related context
type KiroUserInputMessageContext struct {
ToolResults []KiroToolResult `json:"toolResults,omitempty"`
Tools []KiroToolWrapper `json:"tools,omitempty"`
}
// KiroToolResult represents a tool execution result
type KiroToolResult struct {
Content []KiroTextContent `json:"content"`
Status string `json:"status"`
ToolUseID string `json:"toolUseId"`
}
// KiroTextContent represents text content
type KiroTextContent struct {
Text string `json:"text"`
}
// KiroToolWrapper wraps a tool specification
type KiroToolWrapper struct {
ToolSpecification KiroToolSpecification `json:"toolSpecification"`
}
// KiroToolSpecification defines a tool's schema
type KiroToolSpecification struct {
Name string `json:"name"`
Description string `json:"description"`
InputSchema KiroInputSchema `json:"inputSchema"`
}
// KiroInputSchema wraps the JSON schema for tool input
type KiroInputSchema struct {
JSON interface{} `json:"json"`
}
// KiroAssistantResponseMessage represents an assistant message
type KiroAssistantResponseMessage struct {
Content string `json:"content"`
ToolUses []KiroToolUse `json:"toolUses,omitempty"`
}
// KiroToolUse represents a tool invocation by the assistant
type KiroToolUse struct {
ToolUseID string `json:"toolUseId"`
Name string `json:"name"`
Input map[string]interface{} `json:"input"`
}
// ConvertOpenAIRequestToKiro converts an OpenAI Chat Completions request to Kiro format.
// This is the main entry point for request translation.
// Note: The actual payload building happens in the executor, this just passes through
// the OpenAI format which will be converted by BuildKiroPayloadFromOpenAI.
func ConvertOpenAIRequestToKiro(modelName string, inputRawJSON []byte, stream bool) []byte {
// Pass through the OpenAI format - actual conversion happens in BuildKiroPayloadFromOpenAI
return inputRawJSON
}
// BuildKiroPayloadFromOpenAI constructs the Kiro API request payload from OpenAI format.
// Supports tool calling - tools are passed via userInputMessageContext.
// origin parameter determines which quota to use: "CLI" for Amazon Q, "AI_EDITOR" for Kiro IDE.
// isAgentic parameter enables chunked write optimization prompt for -agentic model variants.
// isChatOnly parameter disables tool calling for -chat model variants (pure conversation mode).
// headers parameter allows checking Anthropic-Beta header for thinking mode detection.
// metadata parameter is kept for API compatibility but no longer used for thinking configuration.
// Returns the payload and a boolean indicating whether thinking mode was injected.
func BuildKiroPayloadFromOpenAI(openaiBody []byte, modelID, profileArn, origin string, isAgentic, isChatOnly bool, headers http.Header, metadata map[string]any) ([]byte, bool) {
// Extract max_tokens for potential use in inferenceConfig
// Handle -1 as "use maximum" (Kiro max output is ~32000 tokens)
const kiroMaxOutputTokens = 32000
var maxTokens int64
if mt := gjson.GetBytes(openaiBody, "max_tokens"); mt.Exists() {
maxTokens = mt.Int()
if maxTokens == -1 {
maxTokens = kiroMaxOutputTokens
log.Debugf("kiro-openai: max_tokens=-1 converted to %d", kiroMaxOutputTokens)
}
}
// Extract temperature if specified
var temperature float64
var hasTemperature bool
if temp := gjson.GetBytes(openaiBody, "temperature"); temp.Exists() {
temperature = temp.Float()
hasTemperature = true
}
// Extract top_p if specified
var topP float64
var hasTopP bool
if tp := gjson.GetBytes(openaiBody, "top_p"); tp.Exists() {
topP = tp.Float()
hasTopP = true
log.Debugf("kiro-openai: extracted top_p: %.2f", topP)
}
// Normalize origin value for Kiro API compatibility
origin = normalizeOrigin(origin)
log.Debugf("kiro-openai: normalized origin value: %s", origin)
messages := gjson.GetBytes(openaiBody, "messages")
// For chat-only mode, don't include tools
var tools gjson.Result
if !isChatOnly {
tools = gjson.GetBytes(openaiBody, "tools")
}
// Extract system prompt from messages
systemPrompt := extractSystemPromptFromOpenAI(messages)
// Inject timestamp context
timestamp := time.Now().Format("2006-01-02 15:04:05 MST")
timestampContext := fmt.Sprintf("[Context: Current time is %s]", timestamp)
if systemPrompt != "" {
systemPrompt = timestampContext + "\n\n" + systemPrompt
} else {
systemPrompt = timestampContext
}
log.Debugf("kiro-openai: injected timestamp context: %s", timestamp)
// Inject agentic optimization prompt for -agentic model variants
if isAgentic {
if systemPrompt != "" {
systemPrompt += "\n"
}
systemPrompt += kirocommon.KiroAgenticSystemPrompt
}
// Handle tool_choice parameter - Kiro doesn't support it natively, so we inject system prompt hints
// OpenAI tool_choice values: "none", "auto", "required", or {"type":"function","function":{"name":"..."}}
toolChoiceHint := extractToolChoiceHint(openaiBody)
if toolChoiceHint != "" {
if systemPrompt != "" {
systemPrompt += "\n"
}
systemPrompt += toolChoiceHint
log.Debugf("kiro-openai: injected tool_choice hint into system prompt")
}
// Handle response_format parameter - Kiro doesn't support it natively, so we inject system prompt hints
// OpenAI response_format: {"type": "json_object"} or {"type": "json_schema", "json_schema": {...}}
responseFormatHint := extractResponseFormatHint(openaiBody)
if responseFormatHint != "" {
if systemPrompt != "" {
systemPrompt += "\n"
}
systemPrompt += responseFormatHint
log.Debugf("kiro-openai: injected response_format hint into system prompt")
}
// Check for thinking mode
// Supports OpenAI reasoning_effort parameter, model name hints, and Anthropic-Beta header
thinkingEnabled := checkThinkingModeFromOpenAIWithHeaders(openaiBody, headers)
// Convert OpenAI tools to Kiro format
kiroTools := convertOpenAIToolsToKiro(tools)
// Thinking mode implementation:
// Kiro API supports official thinking/reasoning mode via <thinking_mode> tag.
// When set to "enabled", Kiro returns reasoning content as official reasoningContentEvent
// rather than inline <thinking> tags in assistantResponseEvent.
// We use a high max_thinking_length to allow extensive reasoning.
if thinkingEnabled {
thinkingHint := `<thinking_mode>enabled</thinking_mode>
<max_thinking_length>200000</max_thinking_length>`
if systemPrompt != "" {
systemPrompt = thinkingHint + "\n\n" + systemPrompt
} else {
systemPrompt = thinkingHint
}
log.Debugf("kiro-openai: injected thinking prompt (official mode)")
}
// Process messages and build history
history, currentUserMsg, currentToolResults := processOpenAIMessages(messages, modelID, origin)
// Build content with system prompt
if currentUserMsg != nil {
currentUserMsg.Content = buildFinalContent(currentUserMsg.Content, systemPrompt, currentToolResults)
// Deduplicate currentToolResults
currentToolResults = deduplicateToolResults(currentToolResults)
// Build userInputMessageContext with tools and tool results
if len(kiroTools) > 0 || len(currentToolResults) > 0 {
currentUserMsg.UserInputMessageContext = &KiroUserInputMessageContext{
Tools: kiroTools,
ToolResults: currentToolResults,
}
}
}
// Build payload
var currentMessage KiroCurrentMessage
if currentUserMsg != nil {
currentMessage = KiroCurrentMessage{UserInputMessage: *currentUserMsg}
} else {
fallbackContent := ""
if systemPrompt != "" {
fallbackContent = "--- SYSTEM PROMPT ---\n" + systemPrompt + "\n--- END SYSTEM PROMPT ---\n"
}
currentMessage = KiroCurrentMessage{UserInputMessage: KiroUserInputMessage{
Content: fallbackContent,
ModelID: modelID,
Origin: origin,
}}
}
// Build inferenceConfig if we have any inference parameters
// Note: Kiro API doesn't actually use max_tokens for thinking budget
var inferenceConfig *KiroInferenceConfig
if maxTokens > 0 || hasTemperature || hasTopP {
inferenceConfig = &KiroInferenceConfig{}
if maxTokens > 0 {
inferenceConfig.MaxTokens = int(maxTokens)
}
if hasTemperature {
inferenceConfig.Temperature = temperature
}
if hasTopP {
inferenceConfig.TopP = topP
}
}
payload := KiroPayload{
ConversationState: KiroConversationState{
ChatTriggerType: "MANUAL",
ConversationID: uuid.New().String(),
CurrentMessage: currentMessage,
History: history,
},
ProfileArn: profileArn,
InferenceConfig: inferenceConfig,
}
result, err := json.Marshal(payload)
if err != nil {
log.Debugf("kiro-openai: failed to marshal payload: %v", err)
return nil, false
}
return result, thinkingEnabled
}
// normalizeOrigin normalizes origin value for Kiro API compatibility
func normalizeOrigin(origin string) string {
switch origin {
case "KIRO_CLI":
return "CLI"
case "KIRO_AI_EDITOR":
return "AI_EDITOR"
case "AMAZON_Q":
return "CLI"
case "KIRO_IDE":
return "AI_EDITOR"
default:
return origin
}
}
// extractSystemPromptFromOpenAI extracts system prompt from OpenAI messages
func extractSystemPromptFromOpenAI(messages gjson.Result) string {
if !messages.IsArray() {
return ""
}
var systemParts []string
for _, msg := range messages.Array() {
if msg.Get("role").String() == "system" {
content := msg.Get("content")
if content.Type == gjson.String {
systemParts = append(systemParts, content.String())
} else if content.IsArray() {
// Handle array content format
for _, part := range content.Array() {
if part.Get("type").String() == "text" {
systemParts = append(systemParts, part.Get("text").String())
}
}
}
}
}
return strings.Join(systemParts, "\n")
}
// shortenToolNameIfNeeded shortens tool names that exceed 64 characters.
// MCP tools often have long names like "mcp__server-name__tool-name".
// This preserves the "mcp__" prefix and last segment when possible.
func shortenToolNameIfNeeded(name string) string {
const limit = 64
if len(name) <= limit {
return name
}
// For MCP tools, try to preserve prefix and last segment
if strings.HasPrefix(name, "mcp__") {
idx := strings.LastIndex(name, "__")
if idx > 0 {
cand := "mcp__" + name[idx+2:]
if len(cand) > limit {
return cand[:limit]
}
return cand
}
}
return name[:limit]
}
// convertOpenAIToolsToKiro converts OpenAI tools to Kiro format
func convertOpenAIToolsToKiro(tools gjson.Result) []KiroToolWrapper {
var kiroTools []KiroToolWrapper
if !tools.IsArray() {
return kiroTools
}
for _, tool := range tools.Array() {
// OpenAI tools have type "function" with function definition inside
if tool.Get("type").String() != "function" {
continue
}
fn := tool.Get("function")
if !fn.Exists() {
continue
}
name := fn.Get("name").String()
description := fn.Get("description").String()
parameters := fn.Get("parameters").Value()
// Shorten tool name if it exceeds 64 characters (common with MCP tools)
originalName := name
name = shortenToolNameIfNeeded(name)
if name != originalName {
log.Debugf("kiro-openai: shortened tool name from '%s' to '%s'", originalName, name)
}
// CRITICAL FIX: Kiro API requires non-empty description
if strings.TrimSpace(description) == "" {
description = fmt.Sprintf("Tool: %s", name)
log.Debugf("kiro-openai: tool '%s' has empty description, using default: %s", name, description)
}
// Truncate long descriptions
if len(description) > kirocommon.KiroMaxToolDescLen {
truncLen := kirocommon.KiroMaxToolDescLen - 30
for truncLen > 0 && !utf8.RuneStart(description[truncLen]) {
truncLen--
}
description = description[:truncLen] + "... (description truncated)"
}
kiroTools = append(kiroTools, KiroToolWrapper{
ToolSpecification: KiroToolSpecification{
Name: name,
Description: description,
InputSchema: KiroInputSchema{JSON: parameters},
},
})
}
return kiroTools
}
// processOpenAIMessages processes OpenAI messages and builds Kiro history
func processOpenAIMessages(messages gjson.Result, modelID, origin string) ([]KiroHistoryMessage, *KiroUserInputMessage, []KiroToolResult) {
var history []KiroHistoryMessage
var currentUserMsg *KiroUserInputMessage
var currentToolResults []KiroToolResult
if !messages.IsArray() {
return history, currentUserMsg, currentToolResults
}
// Merge adjacent messages with the same role
messagesArray := kirocommon.MergeAdjacentMessages(messages.Array())
// Track pending tool results that should be attached to the next user message
// This is critical for LiteLLM-translated requests where tool results appear
// as separate "tool" role messages between assistant and user messages
var pendingToolResults []KiroToolResult
for i, msg := range messagesArray {
role := msg.Get("role").String()
isLastMessage := i == len(messagesArray)-1
switch role {
case "system":
// System messages are handled separately via extractSystemPromptFromOpenAI
continue
case "user":
userMsg, toolResults := buildUserMessageFromOpenAI(msg, modelID, origin)
// Merge any pending tool results from preceding "tool" role messages
toolResults = append(pendingToolResults, toolResults...)
pendingToolResults = nil // Reset pending tool results
if isLastMessage {
currentUserMsg = &userMsg
currentToolResults = toolResults
} else {
// CRITICAL: Kiro API requires content to be non-empty for history messages
if strings.TrimSpace(userMsg.Content) == "" {
if len(toolResults) > 0 {
userMsg.Content = "Tool results provided."
} else {
userMsg.Content = "Continue"
}
}
// For history messages, embed tool results in context
if len(toolResults) > 0 {
userMsg.UserInputMessageContext = &KiroUserInputMessageContext{
ToolResults: toolResults,
}
}
history = append(history, KiroHistoryMessage{
UserInputMessage: &userMsg,
})
}
case "assistant":
assistantMsg := buildAssistantMessageFromOpenAI(msg)
// If there are pending tool results, we need to insert a synthetic user message
// before this assistant message to maintain proper conversation structure
if len(pendingToolResults) > 0 {
syntheticUserMsg := KiroUserInputMessage{
Content: "Tool results provided.",
ModelID: modelID,
Origin: origin,
UserInputMessageContext: &KiroUserInputMessageContext{
ToolResults: pendingToolResults,
},
}
history = append(history, KiroHistoryMessage{
UserInputMessage: &syntheticUserMsg,
})
pendingToolResults = nil
}
if isLastMessage {
history = append(history, KiroHistoryMessage{
AssistantResponseMessage: &assistantMsg,
})
// Create a "Continue" user message as currentMessage
currentUserMsg = &KiroUserInputMessage{
Content: "Continue",
ModelID: modelID,
Origin: origin,
}
} else {
history = append(history, KiroHistoryMessage{
AssistantResponseMessage: &assistantMsg,
})
}
case "tool":
// Tool messages in OpenAI format provide results for tool_calls
// These are typically followed by user or assistant messages
// Collect them as pending and attach to the next user message
toolCallID := msg.Get("tool_call_id").String()
content := msg.Get("content").String()
if toolCallID != "" {
toolResult := KiroToolResult{
ToolUseID: toolCallID,
Content: []KiroTextContent{{Text: content}},
Status: "success",
}
// Collect pending tool results to attach to the next user message
pendingToolResults = append(pendingToolResults, toolResult)
}
}
}
// Handle case where tool results are at the end with no following user message
if len(pendingToolResults) > 0 {
currentToolResults = append(currentToolResults, pendingToolResults...)
// If there's no current user message, create a synthetic one for the tool results
if currentUserMsg == nil {
currentUserMsg = &KiroUserInputMessage{
Content: "Tool results provided.",
ModelID: modelID,
Origin: origin,
}
}
}
return history, currentUserMsg, currentToolResults
}
// buildUserMessageFromOpenAI builds a user message from OpenAI format and extracts tool results
func buildUserMessageFromOpenAI(msg gjson.Result, modelID, origin string) (KiroUserInputMessage, []KiroToolResult) {
content := msg.Get("content")
var contentBuilder strings.Builder
var toolResults []KiroToolResult
var images []KiroImage
if content.IsArray() {
for _, part := range content.Array() {
partType := part.Get("type").String()
switch partType {
case "text":
contentBuilder.WriteString(part.Get("text").String())
case "image_url":
imageURL := part.Get("image_url.url").String()
if strings.HasPrefix(imageURL, "data:") {
// Parse data URL: data:image/png;base64,xxxxx
if idx := strings.Index(imageURL, ";base64,"); idx != -1 {
mediaType := imageURL[5:idx] // Skip "data:"
data := imageURL[idx+8:] // Skip ";base64,"
format := ""
if lastSlash := strings.LastIndex(mediaType, "/"); lastSlash != -1 {
format = mediaType[lastSlash+1:]
}
if format != "" && data != "" {
images = append(images, KiroImage{
Format: format,
Source: KiroImageSource{
Bytes: data,
},
})
}
}
}
}
}
} else if content.Type == gjson.String {
contentBuilder.WriteString(content.String())
}
userMsg := KiroUserInputMessage{
Content: contentBuilder.String(),
ModelID: modelID,
Origin: origin,
}
if len(images) > 0 {
userMsg.Images = images
}
return userMsg, toolResults
}
// buildAssistantMessageFromOpenAI builds an assistant message from OpenAI format
func buildAssistantMessageFromOpenAI(msg gjson.Result) KiroAssistantResponseMessage {
content := msg.Get("content")
var contentBuilder strings.Builder
var toolUses []KiroToolUse
// Handle content
if content.Type == gjson.String {
contentBuilder.WriteString(content.String())
} else if content.IsArray() {
for _, part := range content.Array() {
if part.Get("type").String() == "text" {
contentBuilder.WriteString(part.Get("text").String())
}
}
}
// Handle tool_calls
toolCalls := msg.Get("tool_calls")
if toolCalls.IsArray() {
for _, tc := range toolCalls.Array() {
if tc.Get("type").String() != "function" {
continue
}
toolUseID := tc.Get("id").String()
toolName := tc.Get("function.name").String()
toolArgs := tc.Get("function.arguments").String()
var inputMap map[string]interface{}
if err := json.Unmarshal([]byte(toolArgs), &inputMap); err != nil {
log.Debugf("kiro-openai: failed to parse tool arguments: %v", err)
inputMap = make(map[string]interface{})
}
toolUses = append(toolUses, KiroToolUse{
ToolUseID: toolUseID,
Name: toolName,
Input: inputMap,
})
}
}
return KiroAssistantResponseMessage{
Content: contentBuilder.String(),
ToolUses: toolUses,
}
}
// buildFinalContent builds the final content with system prompt
func buildFinalContent(content, systemPrompt string, toolResults []KiroToolResult) string {
var contentBuilder strings.Builder
if systemPrompt != "" {
contentBuilder.WriteString("--- SYSTEM PROMPT ---\n")
contentBuilder.WriteString(systemPrompt)
contentBuilder.WriteString("\n--- END SYSTEM PROMPT ---\n\n")
}
contentBuilder.WriteString(content)
finalContent := contentBuilder.String()
// CRITICAL: Kiro API requires content to be non-empty
if strings.TrimSpace(finalContent) == "" {
if len(toolResults) > 0 {
finalContent = "Tool results provided."
} else {
finalContent = "Continue"
}
log.Debugf("kiro-openai: content was empty, using default: %s", finalContent)
}
return finalContent
}
// checkThinkingModeFromOpenAI checks if thinking mode is enabled in the OpenAI request.
// Returns thinkingEnabled.
// Supports:
// - reasoning_effort parameter (low/medium/high/auto)
// - Model name containing "thinking" or "reason"
// - <thinking_mode> tag in system prompt (AMP/Cursor format)
func checkThinkingModeFromOpenAI(openaiBody []byte) bool {
return checkThinkingModeFromOpenAIWithHeaders(openaiBody, nil)
}
// checkThinkingModeFromOpenAIWithHeaders checks if thinking mode is enabled in the OpenAI request.
// Returns thinkingEnabled.
// Supports:
// - Anthropic-Beta header with interleaved-thinking (Claude CLI)
// - reasoning_effort parameter (low/medium/high/auto)
// - Model name containing "thinking" or "reason"
// - <thinking_mode> tag in system prompt (AMP/Cursor format)
func checkThinkingModeFromOpenAIWithHeaders(openaiBody []byte, headers http.Header) bool {
// Check Anthropic-Beta header first (Claude CLI uses this)
if kiroclaude.IsThinkingEnabledFromHeader(headers) {
log.Debugf("kiro-openai: thinking mode enabled via Anthropic-Beta header")
return true
}
// Check OpenAI format: reasoning_effort parameter
// Valid values: "low", "medium", "high", "auto" (not "none")
reasoningEffort := gjson.GetBytes(openaiBody, "reasoning_effort")
if reasoningEffort.Exists() {
effort := reasoningEffort.String()
if effort != "" && effort != "none" {
log.Debugf("kiro-openai: thinking mode enabled via reasoning_effort: %s", effort)
return true
}
}
// Check AMP/Cursor format: <thinking_mode>interleaved</thinking_mode> in system prompt
bodyStr := string(openaiBody)
if strings.Contains(bodyStr, "<thinking_mode>") && strings.Contains(bodyStr, "</thinking_mode>") {
startTag := "<thinking_mode>"
endTag := "</thinking_mode>"
startIdx := strings.Index(bodyStr, startTag)
if startIdx >= 0 {
startIdx += len(startTag)
endIdx := strings.Index(bodyStr[startIdx:], endTag)
if endIdx >= 0 {
thinkingMode := bodyStr[startIdx : startIdx+endIdx]
if thinkingMode == "interleaved" || thinkingMode == "enabled" {
log.Debugf("kiro-openai: thinking mode enabled via AMP/Cursor format: %s", thinkingMode)
return true
}
}
}
}
// Check model name for thinking hints
model := gjson.GetBytes(openaiBody, "model").String()
modelLower := strings.ToLower(model)
if strings.Contains(modelLower, "thinking") || strings.Contains(modelLower, "-reason") {
log.Debugf("kiro-openai: thinking mode enabled via model name hint: %s", model)
return true
}
log.Debugf("kiro-openai: no thinking mode detected in OpenAI request")
return false
}
// hasThinkingTagInBody checks if the request body already contains thinking configuration tags.
// This is used to prevent duplicate injection when client (e.g., AMP/Cursor) already includes thinking config.
func hasThinkingTagInBody(body []byte) bool {
bodyStr := string(body)
return strings.Contains(bodyStr, "<thinking_mode>") || strings.Contains(bodyStr, "<max_thinking_length>")
}
// extractToolChoiceHint extracts tool_choice from OpenAI request and returns a system prompt hint.
// OpenAI tool_choice values:
// - "none": Don't use any tools
// - "auto": Model decides (default, no hint needed)
// - "required": Must use at least one tool
// - {"type":"function","function":{"name":"..."}} : Must use specific tool
func extractToolChoiceHint(openaiBody []byte) string {
toolChoice := gjson.GetBytes(openaiBody, "tool_choice")
if !toolChoice.Exists() {
return ""
}
// Handle string values
if toolChoice.Type == gjson.String {
switch toolChoice.String() {
case "none":
// Note: When tool_choice is "none", we should ideally not pass tools at all
// But since we can't modify tool passing here, we add a strong hint
return "[INSTRUCTION: Do NOT use any tools. Respond with text only.]"
case "required":
return "[INSTRUCTION: You MUST use at least one of the available tools to respond. Do not respond with text only - always make a tool call.]"
case "auto":
// Default behavior, no hint needed
return ""
}
}
// Handle object value: {"type":"function","function":{"name":"..."}}
if toolChoice.IsObject() {
if toolChoice.Get("type").String() == "function" {
toolName := toolChoice.Get("function.name").String()
if toolName != "" {
return fmt.Sprintf("[INSTRUCTION: You MUST use the tool named '%s' to respond. Do not use any other tool or respond with text only.]", toolName)
}
}
}
return ""
}
// extractResponseFormatHint extracts response_format from OpenAI request and returns a system prompt hint.
// OpenAI response_format values:
// - {"type": "text"}: Default, no hint needed
// - {"type": "json_object"}: Must respond with valid JSON
// - {"type": "json_schema", "json_schema": {...}}: Must respond with JSON matching schema
func extractResponseFormatHint(openaiBody []byte) string {
responseFormat := gjson.GetBytes(openaiBody, "response_format")
if !responseFormat.Exists() {
return ""
}
formatType := responseFormat.Get("type").String()
switch formatType {
case "json_object":
return "[INSTRUCTION: You MUST respond with valid JSON only. Do not include any text before or after the JSON. Do not wrap the JSON in markdown code blocks. Output raw JSON directly.]"
case "json_schema":
// Extract schema if provided
schema := responseFormat.Get("json_schema.schema")
if schema.Exists() {
schemaStr := schema.Raw
// Truncate if too long
if len(schemaStr) > 500 {
schemaStr = schemaStr[:500] + "..."
}
return fmt.Sprintf("[INSTRUCTION: You MUST respond with valid JSON that matches this schema: %s. Do not include any text before or after the JSON. Do not wrap the JSON in markdown code blocks. Output raw JSON directly.]", schemaStr)
}
return "[INSTRUCTION: You MUST respond with valid JSON only. Do not include any text before or after the JSON. Do not wrap the JSON in markdown code blocks. Output raw JSON directly.]"
case "text":
// Default behavior, no hint needed
return ""
}
return ""
}
// deduplicateToolResults removes duplicate tool results
func deduplicateToolResults(toolResults []KiroToolResult) []KiroToolResult {
if len(toolResults) == 0 {
return toolResults
}
seenIDs := make(map[string]bool)
unique := make([]KiroToolResult, 0, len(toolResults))
for _, tr := range toolResults {
if !seenIDs[tr.ToolUseID] {
seenIDs[tr.ToolUseID] = true
unique = append(unique, tr)
} else {
log.Debugf("kiro-openai: skipping duplicate toolResult: %s", tr.ToolUseID)
}
}
return unique
}

View File

@@ -0,0 +1,386 @@
package openai
import (
"encoding/json"
"testing"
)
// TestToolResultsAttachedToCurrentMessage verifies that tool results from "tool" role messages
// are properly attached to the current user message (the last message in the conversation).
// This is critical for LiteLLM-translated requests where tool results appear as separate messages.
func TestToolResultsAttachedToCurrentMessage(t *testing.T) {
// OpenAI format request simulating LiteLLM's translation from Anthropic format
// Sequence: user -> assistant (with tool_calls) -> tool (result) -> user
// The last user message should have the tool results attached
input := []byte(`{
"model": "kiro-claude-opus-4-5-agentic",
"messages": [
{"role": "user", "content": "Hello, can you read a file for me?"},
{
"role": "assistant",
"content": "I'll read that file for you.",
"tool_calls": [
{
"id": "call_abc123",
"type": "function",
"function": {
"name": "Read",
"arguments": "{\"file_path\": \"/tmp/test.txt\"}"
}
}
]
},
{
"role": "tool",
"tool_call_id": "call_abc123",
"content": "File contents: Hello World!"
},
{"role": "user", "content": "What did the file say?"}
]
}`)
result, _ := BuildKiroPayloadFromOpenAI(input, "kiro-model", "", "CLI", false, false, nil, nil)
var payload KiroPayload
if err := json.Unmarshal(result, &payload); err != nil {
t.Fatalf("Failed to unmarshal result: %v", err)
}
// The last user message becomes currentMessage
// History should have: user (first), assistant (with tool_calls)
t.Logf("History count: %d", len(payload.ConversationState.History))
if len(payload.ConversationState.History) != 2 {
t.Errorf("Expected 2 history entries (user + assistant), got %d", len(payload.ConversationState.History))
}
// Tool results should be attached to currentMessage (the last user message)
ctx := payload.ConversationState.CurrentMessage.UserInputMessage.UserInputMessageContext
if ctx == nil {
t.Fatal("Expected currentMessage to have UserInputMessageContext with tool results")
}
if len(ctx.ToolResults) != 1 {
t.Fatalf("Expected 1 tool result in currentMessage, got %d", len(ctx.ToolResults))
}
tr := ctx.ToolResults[0]
if tr.ToolUseID != "call_abc123" {
t.Errorf("Expected toolUseId 'call_abc123', got '%s'", tr.ToolUseID)
}
if len(tr.Content) == 0 || tr.Content[0].Text != "File contents: Hello World!" {
t.Errorf("Tool result content mismatch, got: %+v", tr.Content)
}
}
// TestToolResultsInHistoryUserMessage verifies that when there are multiple user messages
// after tool results, the tool results are attached to the correct user message in history.
func TestToolResultsInHistoryUserMessage(t *testing.T) {
// Sequence: user -> assistant (with tool_calls) -> tool (result) -> user -> assistant -> user
// The first user after tool should have tool results in history
input := []byte(`{
"model": "kiro-claude-opus-4-5-agentic",
"messages": [
{"role": "user", "content": "Hello"},
{
"role": "assistant",
"content": "I'll read the file.",
"tool_calls": [
{
"id": "call_1",
"type": "function",
"function": {
"name": "Read",
"arguments": "{}"
}
}
]
},
{
"role": "tool",
"tool_call_id": "call_1",
"content": "File result"
},
{"role": "user", "content": "Thanks for the file"},
{"role": "assistant", "content": "You're welcome"},
{"role": "user", "content": "Bye"}
]
}`)
result, _ := BuildKiroPayloadFromOpenAI(input, "kiro-model", "", "CLI", false, false, nil, nil)
var payload KiroPayload
if err := json.Unmarshal(result, &payload); err != nil {
t.Fatalf("Failed to unmarshal result: %v", err)
}
// History should have: user, assistant, user (with tool results), assistant
// CurrentMessage should be: last user "Bye"
t.Logf("History count: %d", len(payload.ConversationState.History))
// Find the user message in history with tool results
foundToolResults := false
for i, h := range payload.ConversationState.History {
if h.UserInputMessage != nil {
t.Logf("History[%d]: user message content=%q", i, h.UserInputMessage.Content)
if h.UserInputMessage.UserInputMessageContext != nil {
if len(h.UserInputMessage.UserInputMessageContext.ToolResults) > 0 {
foundToolResults = true
t.Logf(" Found %d tool results", len(h.UserInputMessage.UserInputMessageContext.ToolResults))
tr := h.UserInputMessage.UserInputMessageContext.ToolResults[0]
if tr.ToolUseID != "call_1" {
t.Errorf("Expected toolUseId 'call_1', got '%s'", tr.ToolUseID)
}
}
}
}
if h.AssistantResponseMessage != nil {
t.Logf("History[%d]: assistant message content=%q", i, h.AssistantResponseMessage.Content)
}
}
if !foundToolResults {
t.Error("Tool results were not attached to any user message in history")
}
}
// TestToolResultsWithMultipleToolCalls verifies handling of multiple tool calls
func TestToolResultsWithMultipleToolCalls(t *testing.T) {
input := []byte(`{
"model": "kiro-claude-opus-4-5-agentic",
"messages": [
{"role": "user", "content": "Read two files for me"},
{
"role": "assistant",
"content": "I'll read both files.",
"tool_calls": [
{
"id": "call_1",
"type": "function",
"function": {
"name": "Read",
"arguments": "{\"file_path\": \"/tmp/file1.txt\"}"
}
},
{
"id": "call_2",
"type": "function",
"function": {
"name": "Read",
"arguments": "{\"file_path\": \"/tmp/file2.txt\"}"
}
}
]
},
{
"role": "tool",
"tool_call_id": "call_1",
"content": "Content of file 1"
},
{
"role": "tool",
"tool_call_id": "call_2",
"content": "Content of file 2"
},
{"role": "user", "content": "What do they say?"}
]
}`)
result, _ := BuildKiroPayloadFromOpenAI(input, "kiro-model", "", "CLI", false, false, nil, nil)
var payload KiroPayload
if err := json.Unmarshal(result, &payload); err != nil {
t.Fatalf("Failed to unmarshal result: %v", err)
}
t.Logf("History count: %d", len(payload.ConversationState.History))
t.Logf("CurrentMessage content: %q", payload.ConversationState.CurrentMessage.UserInputMessage.Content)
// Check if there are any tool results anywhere
var totalToolResults int
for i, h := range payload.ConversationState.History {
if h.UserInputMessage != nil && h.UserInputMessage.UserInputMessageContext != nil {
count := len(h.UserInputMessage.UserInputMessageContext.ToolResults)
t.Logf("History[%d] user message has %d tool results", i, count)
totalToolResults += count
}
}
ctx := payload.ConversationState.CurrentMessage.UserInputMessage.UserInputMessageContext
if ctx != nil {
t.Logf("CurrentMessage has %d tool results", len(ctx.ToolResults))
totalToolResults += len(ctx.ToolResults)
} else {
t.Logf("CurrentMessage has no UserInputMessageContext")
}
if totalToolResults != 2 {
t.Errorf("Expected 2 tool results total, got %d", totalToolResults)
}
}
// TestToolResultsAtEndOfConversation verifies tool results are handled when
// the conversation ends with tool results (no following user message)
func TestToolResultsAtEndOfConversation(t *testing.T) {
input := []byte(`{
"model": "kiro-claude-opus-4-5-agentic",
"messages": [
{"role": "user", "content": "Read a file"},
{
"role": "assistant",
"content": "Reading the file.",
"tool_calls": [
{
"id": "call_end",
"type": "function",
"function": {
"name": "Read",
"arguments": "{\"file_path\": \"/tmp/test.txt\"}"
}
}
]
},
{
"role": "tool",
"tool_call_id": "call_end",
"content": "File contents here"
}
]
}`)
result, _ := BuildKiroPayloadFromOpenAI(input, "kiro-model", "", "CLI", false, false, nil, nil)
var payload KiroPayload
if err := json.Unmarshal(result, &payload); err != nil {
t.Fatalf("Failed to unmarshal result: %v", err)
}
// When the last message is a tool result, a synthetic user message is created
// and tool results should be attached to it
ctx := payload.ConversationState.CurrentMessage.UserInputMessage.UserInputMessageContext
if ctx == nil || len(ctx.ToolResults) == 0 {
t.Error("Expected tool results to be attached to current message when conversation ends with tool result")
} else {
if ctx.ToolResults[0].ToolUseID != "call_end" {
t.Errorf("Expected toolUseId 'call_end', got '%s'", ctx.ToolResults[0].ToolUseID)
}
}
}
// TestToolResultsFollowedByAssistant verifies handling when tool results are followed
// by an assistant message (no intermediate user message).
// This is the pattern from LiteLLM translation of Anthropic format where:
// user message has ONLY tool_result blocks -> LiteLLM creates tool messages
// then the next message is assistant
func TestToolResultsFollowedByAssistant(t *testing.T) {
// Sequence: user -> assistant (with tool_calls) -> tool -> tool -> assistant -> user
// This simulates LiteLLM's translation of:
// user: "Read files"
// assistant: [tool_use, tool_use]
// user: [tool_result, tool_result] <- becomes multiple "tool" role messages
// assistant: "I've read them"
// user: "What did they say?"
input := []byte(`{
"model": "kiro-claude-opus-4-5-agentic",
"messages": [
{"role": "user", "content": "Read two files for me"},
{
"role": "assistant",
"content": "I'll read both files.",
"tool_calls": [
{
"id": "call_1",
"type": "function",
"function": {
"name": "Read",
"arguments": "{\"file_path\": \"/tmp/a.txt\"}"
}
},
{
"id": "call_2",
"type": "function",
"function": {
"name": "Read",
"arguments": "{\"file_path\": \"/tmp/b.txt\"}"
}
}
]
},
{
"role": "tool",
"tool_call_id": "call_1",
"content": "Contents of file A"
},
{
"role": "tool",
"tool_call_id": "call_2",
"content": "Contents of file B"
},
{
"role": "assistant",
"content": "I've read both files."
},
{"role": "user", "content": "What did they say?"}
]
}`)
result, _ := BuildKiroPayloadFromOpenAI(input, "kiro-model", "", "CLI", false, false, nil, nil)
var payload KiroPayload
if err := json.Unmarshal(result, &payload); err != nil {
t.Fatalf("Failed to unmarshal result: %v", err)
}
t.Logf("History count: %d", len(payload.ConversationState.History))
// Tool results should be attached to a synthetic user message or the history should be valid
var totalToolResults int
for i, h := range payload.ConversationState.History {
if h.UserInputMessage != nil {
t.Logf("History[%d]: user message content=%q", i, h.UserInputMessage.Content)
if h.UserInputMessage.UserInputMessageContext != nil {
count := len(h.UserInputMessage.UserInputMessageContext.ToolResults)
t.Logf(" Has %d tool results", count)
totalToolResults += count
}
}
if h.AssistantResponseMessage != nil {
t.Logf("History[%d]: assistant message content=%q", i, h.AssistantResponseMessage.Content)
}
}
ctx := payload.ConversationState.CurrentMessage.UserInputMessage.UserInputMessageContext
if ctx != nil {
t.Logf("CurrentMessage has %d tool results", len(ctx.ToolResults))
totalToolResults += len(ctx.ToolResults)
}
if totalToolResults != 2 {
t.Errorf("Expected 2 tool results total, got %d", totalToolResults)
}
}
// TestAssistantEndsConversation verifies handling when assistant is the last message
func TestAssistantEndsConversation(t *testing.T) {
input := []byte(`{
"model": "kiro-claude-opus-4-5-agentic",
"messages": [
{"role": "user", "content": "Hello"},
{
"role": "assistant",
"content": "Hi there!"
}
]
}`)
result, _ := BuildKiroPayloadFromOpenAI(input, "kiro-model", "", "CLI", false, false, nil, nil)
var payload KiroPayload
if err := json.Unmarshal(result, &payload); err != nil {
t.Fatalf("Failed to unmarshal result: %v", err)
}
// When assistant is last, a "Continue" user message should be created
if payload.ConversationState.CurrentMessage.UserInputMessage.Content == "" {
t.Error("Expected a 'Continue' message to be created when assistant is last")
}
}

View File

@@ -0,0 +1,277 @@
// Package openai provides response translation from Kiro to OpenAI format.
// This package handles the conversion of Kiro API responses into OpenAI Chat Completions-compatible
// JSON format, transforming streaming events and non-streaming responses.
package openai
import (
"encoding/json"
"fmt"
"sync/atomic"
"time"
"github.com/google/uuid"
"github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/usage"
log "github.com/sirupsen/logrus"
)
// functionCallIDCounter provides a process-wide unique counter for function call identifiers.
var functionCallIDCounter uint64
// BuildOpenAIResponse constructs an OpenAI Chat Completions-compatible response.
// Supports tool_calls when tools are present in the response.
// stopReason is passed from upstream; fallback logic applied if empty.
func BuildOpenAIResponse(content string, toolUses []KiroToolUse, model string, usageInfo usage.Detail, stopReason string) []byte {
return BuildOpenAIResponseWithReasoning(content, "", toolUses, model, usageInfo, stopReason)
}
// BuildOpenAIResponseWithReasoning constructs an OpenAI Chat Completions-compatible response with reasoning_content support.
// Supports tool_calls when tools are present in the response.
// reasoningContent is included as reasoning_content field in the message when present.
// stopReason is passed from upstream; fallback logic applied if empty.
func BuildOpenAIResponseWithReasoning(content, reasoningContent string, toolUses []KiroToolUse, model string, usageInfo usage.Detail, stopReason string) []byte {
// Build the message object
message := map[string]interface{}{
"role": "assistant",
"content": content,
}
// Add reasoning_content if present (for thinking/reasoning models)
if reasoningContent != "" {
message["reasoning_content"] = reasoningContent
}
// Add tool_calls if present
if len(toolUses) > 0 {
var toolCalls []map[string]interface{}
for i, tu := range toolUses {
inputJSON, _ := json.Marshal(tu.Input)
toolCalls = append(toolCalls, map[string]interface{}{
"id": tu.ToolUseID,
"type": "function",
"index": i,
"function": map[string]interface{}{
"name": tu.Name,
"arguments": string(inputJSON),
},
})
}
message["tool_calls"] = toolCalls
// When tool_calls are present, content should be null according to OpenAI spec
if content == "" {
message["content"] = nil
}
}
// Use upstream stopReason; apply fallback logic if not provided
finishReason := mapKiroStopReasonToOpenAI(stopReason)
if finishReason == "" {
finishReason = "stop"
if len(toolUses) > 0 {
finishReason = "tool_calls"
}
log.Debugf("kiro-openai: buildOpenAIResponse using fallback finish_reason: %s", finishReason)
}
response := map[string]interface{}{
"id": "chatcmpl-" + uuid.New().String()[:24],
"object": "chat.completion",
"created": time.Now().Unix(),
"model": model,
"choices": []map[string]interface{}{
{
"index": 0,
"message": message,
"finish_reason": finishReason,
},
},
"usage": map[string]interface{}{
"prompt_tokens": usageInfo.InputTokens,
"completion_tokens": usageInfo.OutputTokens,
"total_tokens": usageInfo.InputTokens + usageInfo.OutputTokens,
},
}
result, _ := json.Marshal(response)
return result
}
// mapKiroStopReasonToOpenAI converts Kiro/Claude stop_reason to OpenAI finish_reason
func mapKiroStopReasonToOpenAI(stopReason string) string {
switch stopReason {
case "end_turn":
return "stop"
case "stop_sequence":
return "stop"
case "tool_use":
return "tool_calls"
case "max_tokens":
return "length"
case "content_filtered":
return "content_filter"
default:
return stopReason
}
}
// BuildOpenAIStreamChunk constructs an OpenAI Chat Completions streaming chunk.
// This is the delta format used in streaming responses.
func BuildOpenAIStreamChunk(model string, deltaContent string, deltaToolCalls []map[string]interface{}, finishReason string, index int) []byte {
delta := map[string]interface{}{}
// First chunk should include role
if index == 0 && deltaContent == "" && len(deltaToolCalls) == 0 {
delta["role"] = "assistant"
delta["content"] = ""
} else if deltaContent != "" {
delta["content"] = deltaContent
}
// Add tool_calls delta if present
if len(deltaToolCalls) > 0 {
delta["tool_calls"] = deltaToolCalls
}
choice := map[string]interface{}{
"index": 0,
"delta": delta,
}
if finishReason != "" {
choice["finish_reason"] = finishReason
} else {
choice["finish_reason"] = nil
}
chunk := map[string]interface{}{
"id": "chatcmpl-" + uuid.New().String()[:12],
"object": "chat.completion.chunk",
"created": time.Now().Unix(),
"model": model,
"choices": []map[string]interface{}{choice},
}
result, _ := json.Marshal(chunk)
return result
}
// BuildOpenAIStreamChunkWithToolCallStart creates a stream chunk for tool call start
func BuildOpenAIStreamChunkWithToolCallStart(model string, toolUseID, toolName string, toolIndex int) []byte {
toolCall := map[string]interface{}{
"index": toolIndex,
"id": toolUseID,
"type": "function",
"function": map[string]interface{}{
"name": toolName,
"arguments": "",
},
}
delta := map[string]interface{}{
"tool_calls": []map[string]interface{}{toolCall},
}
choice := map[string]interface{}{
"index": 0,
"delta": delta,
"finish_reason": nil,
}
chunk := map[string]interface{}{
"id": "chatcmpl-" + uuid.New().String()[:12],
"object": "chat.completion.chunk",
"created": time.Now().Unix(),
"model": model,
"choices": []map[string]interface{}{choice},
}
result, _ := json.Marshal(chunk)
return result
}
// BuildOpenAIStreamChunkWithToolCallDelta creates a stream chunk for tool call arguments delta
func BuildOpenAIStreamChunkWithToolCallDelta(model string, argumentsDelta string, toolIndex int) []byte {
toolCall := map[string]interface{}{
"index": toolIndex,
"function": map[string]interface{}{
"arguments": argumentsDelta,
},
}
delta := map[string]interface{}{
"tool_calls": []map[string]interface{}{toolCall},
}
choice := map[string]interface{}{
"index": 0,
"delta": delta,
"finish_reason": nil,
}
chunk := map[string]interface{}{
"id": "chatcmpl-" + uuid.New().String()[:12],
"object": "chat.completion.chunk",
"created": time.Now().Unix(),
"model": model,
"choices": []map[string]interface{}{choice},
}
result, _ := json.Marshal(chunk)
return result
}
// BuildOpenAIStreamDoneChunk creates the final [DONE] stream event
func BuildOpenAIStreamDoneChunk() []byte {
return []byte("data: [DONE]")
}
// BuildOpenAIStreamFinishChunk creates the final chunk with finish_reason
func BuildOpenAIStreamFinishChunk(model string, finishReason string) []byte {
choice := map[string]interface{}{
"index": 0,
"delta": map[string]interface{}{},
"finish_reason": finishReason,
}
chunk := map[string]interface{}{
"id": "chatcmpl-" + uuid.New().String()[:12],
"object": "chat.completion.chunk",
"created": time.Now().Unix(),
"model": model,
"choices": []map[string]interface{}{choice},
}
result, _ := json.Marshal(chunk)
return result
}
// BuildOpenAIStreamUsageChunk creates a chunk with usage information (optional, for stream_options.include_usage)
func BuildOpenAIStreamUsageChunk(model string, usageInfo usage.Detail) []byte {
chunk := map[string]interface{}{
"id": "chatcmpl-" + uuid.New().String()[:12],
"object": "chat.completion.chunk",
"created": time.Now().Unix(),
"model": model,
"choices": []map[string]interface{}{},
"usage": map[string]interface{}{
"prompt_tokens": usageInfo.InputTokens,
"completion_tokens": usageInfo.OutputTokens,
"total_tokens": usageInfo.InputTokens + usageInfo.OutputTokens,
},
}
result, _ := json.Marshal(chunk)
return result
}
// GenerateToolCallID generates a unique tool call ID in OpenAI format
func GenerateToolCallID(toolName string) string {
return fmt.Sprintf("call_%s_%d_%d", toolName[:min(8, len(toolName))], time.Now().UnixNano(), atomic.AddUint64(&functionCallIDCounter, 1))
}
// min returns the minimum of two integers
func min(a, b int) int {
if a < b {
return a
}
return b
}

View File

@@ -0,0 +1,212 @@
// Package openai provides streaming SSE event building for OpenAI format.
// This package handles the construction of OpenAI-compatible Server-Sent Events (SSE)
// for streaming responses from Kiro API.
package openai
import (
"encoding/json"
"time"
"github.com/google/uuid"
"github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/usage"
)
// OpenAIStreamState tracks the state of streaming response conversion
type OpenAIStreamState struct {
ChunkIndex int
ToolCallIndex int
HasSentFirstChunk bool
Model string
ResponseID string
Created int64
}
// NewOpenAIStreamState creates a new stream state for tracking
func NewOpenAIStreamState(model string) *OpenAIStreamState {
return &OpenAIStreamState{
ChunkIndex: 0,
ToolCallIndex: 0,
HasSentFirstChunk: false,
Model: model,
ResponseID: "chatcmpl-" + uuid.New().String()[:24],
Created: time.Now().Unix(),
}
}
// FormatSSEEvent formats a JSON payload for SSE streaming.
// Note: This returns raw JSON data without "data:" prefix.
// The SSE "data:" prefix is added by the Handler layer (e.g., openai_handlers.go)
// to maintain architectural consistency and avoid double-prefix issues.
func FormatSSEEvent(data []byte) string {
return string(data)
}
// BuildOpenAISSETextDelta creates an SSE event for text content delta
func BuildOpenAISSETextDelta(state *OpenAIStreamState, textDelta string) string {
delta := map[string]interface{}{
"content": textDelta,
}
// Include role in first chunk
if !state.HasSentFirstChunk {
delta["role"] = "assistant"
state.HasSentFirstChunk = true
}
chunk := buildBaseChunk(state, delta, nil)
result, _ := json.Marshal(chunk)
state.ChunkIndex++
return FormatSSEEvent(result)
}
// BuildOpenAISSEToolCallStart creates an SSE event for tool call start
func BuildOpenAISSEToolCallStart(state *OpenAIStreamState, toolUseID, toolName string) string {
toolCall := map[string]interface{}{
"index": state.ToolCallIndex,
"id": toolUseID,
"type": "function",
"function": map[string]interface{}{
"name": toolName,
"arguments": "",
},
}
delta := map[string]interface{}{
"tool_calls": []map[string]interface{}{toolCall},
}
// Include role in first chunk if not sent yet
if !state.HasSentFirstChunk {
delta["role"] = "assistant"
state.HasSentFirstChunk = true
}
chunk := buildBaseChunk(state, delta, nil)
result, _ := json.Marshal(chunk)
state.ChunkIndex++
return FormatSSEEvent(result)
}
// BuildOpenAISSEToolCallArgumentsDelta creates an SSE event for tool call arguments delta
func BuildOpenAISSEToolCallArgumentsDelta(state *OpenAIStreamState, argumentsDelta string, toolIndex int) string {
toolCall := map[string]interface{}{
"index": toolIndex,
"function": map[string]interface{}{
"arguments": argumentsDelta,
},
}
delta := map[string]interface{}{
"tool_calls": []map[string]interface{}{toolCall},
}
chunk := buildBaseChunk(state, delta, nil)
result, _ := json.Marshal(chunk)
state.ChunkIndex++
return FormatSSEEvent(result)
}
// BuildOpenAISSEFinish creates an SSE event with finish_reason
func BuildOpenAISSEFinish(state *OpenAIStreamState, finishReason string) string {
chunk := buildBaseChunk(state, map[string]interface{}{}, &finishReason)
result, _ := json.Marshal(chunk)
state.ChunkIndex++
return FormatSSEEvent(result)
}
// BuildOpenAISSEUsage creates an SSE event with usage information
func BuildOpenAISSEUsage(state *OpenAIStreamState, usageInfo usage.Detail) string {
chunk := map[string]interface{}{
"id": state.ResponseID,
"object": "chat.completion.chunk",
"created": state.Created,
"model": state.Model,
"choices": []map[string]interface{}{},
"usage": map[string]interface{}{
"prompt_tokens": usageInfo.InputTokens,
"completion_tokens": usageInfo.OutputTokens,
"total_tokens": usageInfo.InputTokens + usageInfo.OutputTokens,
},
}
result, _ := json.Marshal(chunk)
return FormatSSEEvent(result)
}
// BuildOpenAISSEDone creates the final [DONE] SSE event.
// Note: This returns raw "[DONE]" without "data:" prefix.
// The SSE "data:" prefix is added by the Handler layer (e.g., openai_handlers.go)
// to maintain architectural consistency and avoid double-prefix issues.
func BuildOpenAISSEDone() string {
return "[DONE]"
}
// buildBaseChunk creates a base chunk structure for streaming
func buildBaseChunk(state *OpenAIStreamState, delta map[string]interface{}, finishReason *string) map[string]interface{} {
choice := map[string]interface{}{
"index": 0,
"delta": delta,
}
if finishReason != nil {
choice["finish_reason"] = *finishReason
} else {
choice["finish_reason"] = nil
}
return map[string]interface{}{
"id": state.ResponseID,
"object": "chat.completion.chunk",
"created": state.Created,
"model": state.Model,
"choices": []map[string]interface{}{choice},
}
}
// BuildOpenAISSEReasoningDelta creates an SSE event for reasoning content delta
// This is used for o1/o3 style models that expose reasoning tokens
func BuildOpenAISSEReasoningDelta(state *OpenAIStreamState, reasoningDelta string) string {
delta := map[string]interface{}{
"reasoning_content": reasoningDelta,
}
// Include role in first chunk
if !state.HasSentFirstChunk {
delta["role"] = "assistant"
state.HasSentFirstChunk = true
}
chunk := buildBaseChunk(state, delta, nil)
result, _ := json.Marshal(chunk)
state.ChunkIndex++
return FormatSSEEvent(result)
}
// BuildOpenAISSEFirstChunk creates the first chunk with role only
func BuildOpenAISSEFirstChunk(state *OpenAIStreamState) string {
delta := map[string]interface{}{
"role": "assistant",
"content": "",
}
state.HasSentFirstChunk = true
chunk := buildBaseChunk(state, delta, nil)
result, _ := json.Marshal(chunk)
state.ChunkIndex++
return FormatSSEEvent(result)
}
// ThinkingTagState tracks state for thinking tag detection in streaming
type ThinkingTagState struct {
InThinkingBlock bool
PendingStartChars int
PendingEndChars int
}
// NewThinkingTagState creates a new thinking tag state
func NewThinkingTagState() *ThinkingTagState {
return &ThinkingTagState{
InThinkingBlock: false,
PendingStartChars: 0,
PendingEndChars: 0,
}
}