LLM/CLIProxyAPIPlus
1
0
Fork 0
mirror of https://github.com/router-for-me/CLIProxyAPIPlus.git synced 2026-04-17 20:03:42 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #25 from Ravens2121/master

Browse Source 
fix(kiro): Always parse thinking tags from Kiro API responses
This commit is contained in:
Luis Pater
2025-12-14 20:29:14 +08:00
committed by GitHub
parent 6a4cff6699 de0ea3ac49
commit 8fb1f114bc
3 changed files with 24 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
internal/runtime/executor/kiro_executor.go
View File

@@ -166,7 +166,8 @@ type KiroExecutor struct {
// This is critical because OpenAI and Claude formats have different tool structures:
// - OpenAI: tools[].function.name, tools[].function.description
// - Claude: tools[].name, tools[].description
func buildKiroPayloadForFormat(body []byte, modelID, profileArn, origin string, isAgentic, isChatOnly bool, sourceFormat sdktranslator.Format) []byte {
// Returns the serialized JSON payload and a boolean indicating whether thinking mode was injected.
func buildKiroPayloadForFormat(body []byte, modelID, profileArn, origin string, isAgentic, isChatOnly bool, sourceFormat sdktranslator.Format) ([]byte, bool) {
switch sourceFormat.String() {
case "openai":
log.Debugf("kiro: using OpenAI payload builder for source format: %s", sourceFormat.String())
@@ -248,7 +249,7 @@ func (e *KiroExecutor) executeWithRetry(ctx context.Context, auth *cliproxyauth.
// Rebuild payload with the correct origin for this endpoint
// Each endpoint requires its matching Origin value in the request body
kiroPayload = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
kiroPayload, _ = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
log.Debugf("kiro: trying endpoint %d/%d: %s (Name: %s, Origin: %s)",
endpointIdx+1, len(endpointConfigs), url, endpointConfig.Name, currentOrigin)
@@ -358,7 +359,7 @@ func (e *KiroExecutor) executeWithRetry(ctx context.Context, auth *cliproxyauth.
auth = refreshedAuth
accessToken, profileArn = kiroCredentials(auth)
// Rebuild payload with new profile ARN if changed
kiroPayload = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
kiroPayload, _ = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
log.Infof("kiro: token refreshed successfully, retrying request")
continue
}
@@ -415,7 +416,7 @@ func (e *KiroExecutor) executeWithRetry(ctx context.Context, auth *cliproxyauth.
if refreshedAuth != nil {
auth = refreshedAuth
accessToken, profileArn = kiroCredentials(auth)
kiroPayload = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
kiroPayload, _ = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
log.Infof("kiro: token refreshed for 403, retrying request")
continue
}
@@ -554,7 +555,10 @@ func (e *KiroExecutor) executeStreamWithRetry(ctx context.Context, auth *cliprox
// Rebuild payload with the correct origin for this endpoint
// Each endpoint requires its matching Origin value in the request body
kiroPayload = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
kiroPayload, _ = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
// Kiro API always returns <thinking> tags regardless of whether thinking mode was requested
// So we always enable thinking parsing for Kiro responses
thinkingEnabled := true
log.Debugf("kiro: stream trying endpoint %d/%d: %s (Name: %s, Origin: %s)",
endpointIdx+1, len(endpointConfigs), url, endpointConfig.Name, currentOrigin)
@@ -677,7 +681,7 @@ func (e *KiroExecutor) executeStreamWithRetry(ctx context.Context, auth *cliprox
auth = refreshedAuth
accessToken, profileArn = kiroCredentials(auth)
// Rebuild payload with new profile ARN if changed
kiroPayload = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
kiroPayload, _ = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
log.Infof("kiro: token refreshed successfully, retrying stream request")
continue
}
@@ -734,7 +738,7 @@ func (e *KiroExecutor) executeStreamWithRetry(ctx context.Context, auth *cliprox
if refreshedAuth != nil {
auth = refreshedAuth
accessToken, profileArn = kiroCredentials(auth)
kiroPayload = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
kiroPayload, _ = buildKiroPayloadForFormat(body, kiroModelID, profileArn, currentOrigin, isAgentic, isChatOnly, from)
log.Infof("kiro: token refreshed for 403, retrying stream request")
continue
}
@@ -758,7 +762,7 @@ func (e *KiroExecutor) executeStreamWithRetry(ctx context.Context, auth *cliprox
out := make(chan cliproxyexecutor.StreamChunk)
go func(resp *http.Response) {
go func(resp *http.Response, thinkingEnabled bool) {
defer close(out)
defer func() {
if r := recover(); r != nil {
@@ -772,21 +776,12 @@ func (e *KiroExecutor) executeStreamWithRetry(ctx context.Context, auth *cliprox
}
}()
// Check if thinking mode was enabled in the original request
// Only parse <thinking> tags when thinking was explicitly requested
// Check multiple sources: original request, pre-translation payload, and translated body
// This handles different client formats (Claude API, OpenAI, AMP/Cursor)
thinkingEnabled := kiroclaude.IsThinkingEnabled(opts.OriginalRequest)
if !thinkingEnabled {
thinkingEnabled = kiroclaude.IsThinkingEnabled(req.Payload)
}
if !thinkingEnabled {
thinkingEnabled = kiroclaude.IsThinkingEnabled(body)
}
log.Debugf("kiro: stream thinkingEnabled = %v", thinkingEnabled)
// Kiro API always returns <thinking> tags regardless of request parameters
// So we always enable thinking parsing for Kiro responses
log.Debugf("kiro: stream thinkingEnabled = %v (always true for Kiro)", thinkingEnabled)
e.streamToChannel(ctx, resp.Body, out, from, req.Model, opts.OriginalRequest, body, reporter, thinkingEnabled)
}(httpResp)
}(httpResp, thinkingEnabled)
return out, nil
}

7
internal/translator/kiro/claude/kiro_claude_request.go
View File

@@ -135,7 +135,8 @@ func ConvertClaudeRequestToKiro(modelName string, inputRawJSON []byte, stream bo
// isAgentic parameter enables chunked write optimization prompt for -agentic model variants.
// isChatOnly parameter disables tool calling for -chat model variants (pure conversation mode).
// Supports thinking mode - when Claude API thinking parameter is present, injects thinkingHint.
func BuildKiroPayload(claudeBody []byte, modelID, profileArn, origin string, isAgentic, isChatOnly bool) []byte {
// Returns the payload and a boolean indicating whether thinking mode was injected.
func BuildKiroPayload(claudeBody []byte, modelID, profileArn, origin string, isAgentic, isChatOnly bool) ([]byte, bool) {
// Extract max_tokens for potential use in inferenceConfig
// Handle -1 as "use maximum" (Kiro max output is ~32000 tokens)
const kiroMaxOutputTokens = 32000
@@ -307,10 +308,10 @@ func BuildKiroPayload(claudeBody []byte, modelID, profileArn, origin string, isA
result, err := json.Marshal(payload)
if err != nil {
log.Debugf("kiro: failed to marshal payload: %v", err)
return nil
return nil, false
}
return result
return result, thinkingEnabled
}
// normalizeOrigin normalizes origin value for Kiro API compatibility

7
internal/translator/kiro/openai/kiro_openai_request.go
View File

@@ -133,7 +133,8 @@ func ConvertOpenAIRequestToKiro(modelName string, inputRawJSON []byte, stream bo
// 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).
func BuildKiroPayloadFromOpenAI(openaiBody []byte, modelID, profileArn, origin string, isAgentic, isChatOnly bool) []byte {
// Returns the payload and a boolean indicating whether thinking mode was injected.
func BuildKiroPayloadFromOpenAI(openaiBody []byte, modelID, profileArn, origin string, isAgentic, isChatOnly bool) ([]byte, bool) {
// Extract max_tokens for potential use in inferenceConfig
// Handle -1 as "use maximum" (Kiro max output is ~32000 tokens)
const kiroMaxOutputTokens = 32000
@@ -311,10 +312,10 @@ func BuildKiroPayloadFromOpenAI(openaiBody []byte, modelID, profileArn, origin s
result, err := json.Marshal(payload)
if err != nil {
log.Debugf("kiro-openai: failed to marshal payload: %v", err)
return nil
return nil, false
}
return result
return result, thinkingEnabled
}
// normalizeOrigin normalizes origin value for Kiro API compatibility