Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions internal/proxy/streaming.go
Original file line number Diff line number Diff line change
Expand Up @@ -130,7 +130,8 @@ func (p *Proxy) handleStreamingResponse(w http.ResponseWriter, resp *http.Respon
if writeErr != nil {
return fmt.Errorf("%w: %v", errClientDisconnected, writeErr)
}
if hasStreamEventType(transformedEvent, "response.completed") {
if hasStreamEventType(transformedEvent, "response.completed") ||
hasStreamEventType(transformedEvent, "response.incomplete") {
responseCompleted = true
}
flusher.Flush()
Expand Down Expand Up @@ -246,7 +247,8 @@ func (p *Proxy) handleStreamingAsNonStreaming(w http.ResponseWriter, resp *http.
if err := json.Unmarshal([]byte(jsonData), &event); err != nil {
continue
}
if eventType, _ := event["type"].(string); eventType != "response.completed" {
// A response cut off by the output token limit ends with response.incomplete.
if eventType, _ := event["type"].(string); eventType != "response.completed" && eventType != "response.incomplete" {
continue
}

Expand Down
6 changes: 6 additions & 0 deletions internal/proxy/streaming_completion_test.go
Original file line number Diff line number Diff line change
Expand Up @@ -47,6 +47,7 @@ func waitForSwitchTestSignal(t *testing.T, signal <-chan struct{}, name string)

func TestHandleStreamingResponseRequiresResponsesCompletion(t *testing.T) {
const completed = `data: {"type":"response.completed","response":{"usage":{"input_tokens":1,"output_tokens":1}}}`
const incomplete = `data: {"type":"response.incomplete","response":{"status":"incomplete","incomplete_details":{"reason":"max_output_tokens"},"usage":{"input_tokens":1,"output_tokens":1}}}`
doneText := "data:" + " [DONE]"
embeddedDone := `data: {"type":"response.output_text.delta","delta":"` + doneText + ` is documentation"}` + "\n\n"

Expand Down Expand Up @@ -86,6 +87,11 @@ func TestHandleStreamingResponseRequiresResponsesCompletion(t *testing.T) {
upstream: completed + "\n\n",
wantBody: completed + "\n\n",
},
{
name: "truncated response ends with response.incomplete",
upstream: incomplete + "\n\n",
wantBody: incomplete + "\n\n",
},
{
name: "completion marker inside JSON delta",
upstream: embeddedDone + completed + "\n\n",
Expand Down
32 changes: 22 additions & 10 deletions internal/transformer/convert/claude_openai2.go
Original file line number Diff line number Diff line change
Expand Up @@ -229,6 +229,7 @@ func ClaudeRespToOpenAI2(claudeResp []byte) ([]byte, error) {
"total_tokens": resp.Usage.InputTokens + resp.Usage.OutputTokens,
},
}
markResponsesTruncated(openai2Resp, resp.StopReason == "max_tokens")

return json.Marshal(openai2Resp)
}
Expand Down Expand Up @@ -268,6 +269,10 @@ func OpenAI2RespToClaude(openai2Resp []byte) ([]byte, error) {
}
}

if responsesTruncated(&resp) {
stopReason = "max_tokens"
}

claudeResp := map[string]interface{}{
"id": resp.ID,
"type": "message",
Expand Down Expand Up @@ -527,6 +532,11 @@ func ClaudeStreamToOpenAI2(event []byte, ctx *transformer.StreamContext) ([]byte
}

case "message_delta":
if delta, ok := data["delta"].(map[string]interface{}); ok {
if reason, ok := delta["stop_reason"].(string); ok && reason != "" {
ctx.StopReason = reason
}
}
if usage, ok := data["usage"].(map[string]interface{}); ok {
if out, ok := usage["output_tokens"].(float64); ok {
ctx.OutputTokens = int(out)
Expand All @@ -537,16 +547,15 @@ func ClaudeStreamToOpenAI2(event []byte, ctx *transformer.StreamContext) ([]byte
flushThinkTaggedStream(ctx, emitText, emitThinking)
finishThinking()
finishText()
writeEvent(map[string]interface{}{
"type": "response.completed",
"response": map[string]interface{}{
"id": ctx.MessageID, "object": "response", "status": "completed",
"usage": map[string]interface{}{
"input_tokens": ctx.InputTokens, "output_tokens": ctx.OutputTokens,
"total_tokens": ctx.InputTokens + ctx.OutputTokens,
},
response := map[string]interface{}{
"id": ctx.MessageID, "object": "response", "status": "completed",
"usage": map[string]interface{}{
"input_tokens": ctx.InputTokens, "output_tokens": ctx.OutputTokens,
"total_tokens": ctx.InputTokens + ctx.OutputTokens,
},
})
}
eventType := markResponsesTruncated(response, ctx.StopReason == "max_tokens")
writeEvent(map[string]interface{}{"type": eventType, "response": response})
result.WriteString("data: [DONE]\n\n")
}

Expand Down Expand Up @@ -681,7 +690,7 @@ func OpenAI2StreamToClaude(event []byte, ctx *transformer.StreamContext) ([]byte
ctx.ContentIndex++
}

case "response.completed":
case "response.completed", "response.incomplete":
if evt.Response != nil {
if evt.Response.Usage.InputTokens > 0 {
ctx.InputTokens = evt.Response.Usage.InputTokens
Expand All @@ -704,6 +713,9 @@ func OpenAI2StreamToClaude(event []byte, ctx *transformer.StreamContext) ([]byte
if ctx.ToolIndex > 0 || ctx.CurrentToolID != "" {
stopReason = "tool_use"
}
if responsesTruncated(evt.Response) {
stopReason = "max_tokens"
}
result = append(result, buildClaudeEvent("message_delta", map[string]interface{}{
"delta": map[string]interface{}{"stop_reason": stopReason, "stop_sequence": nil},
"usage": map[string]interface{}{"output_tokens": ctx.OutputTokens},
Expand Down
18 changes: 18 additions & 0 deletions internal/transformer/convert/common.go
Original file line number Diff line number Diff line change
Expand Up @@ -79,6 +79,24 @@ func buildOpenAIChunkWithUsage(id, model, content string, toolCalls []map[string
return []byte(fmt.Sprintf("data: %s\n\n", data)), nil
}

// markResponsesTruncated flags a Responses API result that was cut off by the output
// token limit: status "incomplete" with incomplete_details.reason "max_output_tokens".
// It returns the matching stream event type.
func markResponsesTruncated(response map[string]interface{}, truncated bool) string {
if !truncated {
return "response.completed"
}
response["status"] = "incomplete"
response["incomplete_details"] = map[string]interface{}{"reason": "max_output_tokens"}
return "response.incomplete"
}

// responsesTruncated reports whether a Responses API result stopped on the output token limit.
func responsesTruncated(resp *transformer.OpenAI2Response) bool {
return resp != nil && resp.Status == "incomplete" &&
resp.IncompleteDetails != nil && resp.IncompleteDetails.Reason == "max_output_tokens"
}

// syncGeminiUsageMetadata stores Gemini usage metadata in stream context for later usage emission.
func syncGeminiUsageMetadata(resp *transformer.GeminiResponse, ctx *transformer.StreamContext) {
if resp == nil || resp.UsageMetadata == nil || ctx == nil {
Expand Down
14 changes: 7 additions & 7 deletions internal/transformer/convert/openai2_gemini.go
Original file line number Diff line number Diff line change
Expand Up @@ -139,6 +139,7 @@ func GeminiRespToOpenAI2(geminiResp []byte) ([]byte, error) {
if usage != nil {
openai2Resp["usage"] = usage
}
markResponsesTruncated(openai2Resp, len(resp.Candidates) > 0 && resp.Candidates[0].FinishReason == "MAX_TOKENS")

return json.Marshal(openai2Resp)
}
Expand Down Expand Up @@ -255,13 +256,12 @@ func GeminiStreamToOpenAI2(event []byte, ctx *transformer.StreamContext) ([]byte
if resp.UsageMetadata != nil && resp.UsageMetadata.TotalTokenCount > 0 {
totalTokens = resp.UsageMetadata.TotalTokenCount
}
writeEvent(map[string]interface{}{
"type": "response.completed",
"response": map[string]interface{}{
"id": ctx.MessageID, "object": "response", "status": "completed",
"usage": map[string]interface{}{"input_tokens": ctx.InputTokens, "output_tokens": ctx.OutputTokens, "total_tokens": totalTokens},
},
})
response := map[string]interface{}{
"id": ctx.MessageID, "object": "response", "status": "completed",
"usage": map[string]interface{}{"input_tokens": ctx.InputTokens, "output_tokens": ctx.OutputTokens, "total_tokens": totalTokens},
}
eventType := markResponsesTruncated(response, candidate.FinishReason == "MAX_TOKENS")
writeEvent(map[string]interface{}{"type": eventType, "response": response})
result.WriteString("data: [DONE]\n\n")
}

Expand Down
84 changes: 84 additions & 0 deletions internal/transformer/convert/responses_truncation_test.go
Original file line number Diff line number Diff line change
@@ -0,0 +1,84 @@
package convert

import (
"testing"

"github.com/lich0821/ccNexus/internal/transformer"
)

const claudeTruncatedResp = `{"id":"msg_1","type":"message","role":"assistant","model":"claude-sonnet-4-5","content":[{"type":"text","text":"Once upon a"}],"stop_reason":"max_tokens","usage":{"input_tokens":5,"output_tokens":3}}`

const responsesTruncatedResp = `{"id":"resp_1","object":"response","status":"incomplete","incomplete_details":{"reason":"max_output_tokens"},` +
`"output":[{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Once upon a"}]}],"usage":{"input_tokens":5,"output_tokens":3,"total_tokens":8}}`

func TestClaudeRespToOpenAI2MaxTokensReportsIncomplete(t *testing.T) {
out, err := ClaudeRespToOpenAI2([]byte(claudeTruncatedResp))
if err != nil {
t.Fatalf("ClaudeRespToOpenAI2 failed: %v", err)
}

assertContains(t, string(out), `"status":"incomplete"`, "Expected status incomplete for a Claude max_tokens stop")
assertContains(t, string(out), `"reason":"max_output_tokens"`, "Expected incomplete_details.reason max_output_tokens")
}

func TestClaudeStreamToOpenAI2MaxTokensReportsIncomplete(t *testing.T) {
ctx := transformer.NewStreamContext()
ctx.MessageID = "msg_1"

delta := "event: message_delta\ndata: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"max_tokens\"},\"usage\":{\"output_tokens\":3}}\n\n"
if _, err := ClaudeStreamToOpenAI2([]byte(delta), ctx); err != nil {
t.Fatalf("ClaudeStreamToOpenAI2 message_delta failed: %v", err)
}
out, err := ClaudeStreamToOpenAI2([]byte("event: message_stop\ndata: {\"type\":\"message_stop\"}\n\n"), ctx)
if err != nil {
t.Fatalf("ClaudeStreamToOpenAI2 message_stop failed: %v", err)
}

assertContains(t, string(out), `"type":"response.incomplete"`, "Expected a response.incomplete event for a Claude max_tokens stop")
assertContains(t, string(out), `"reason":"max_output_tokens"`, "Expected incomplete_details.reason max_output_tokens")
}

func TestGeminiRespToOpenAI2MaxTokensReportsIncomplete(t *testing.T) {
out, err := GeminiRespToOpenAI2([]byte(geminiTruncatedResp))
if err != nil {
t.Fatalf("GeminiRespToOpenAI2 failed: %v", err)
}

assertContains(t, string(out), `"status":"incomplete"`, "Expected status incomplete for a Gemini MAX_TOKENS finish")
assertContains(t, string(out), `"reason":"max_output_tokens"`, "Expected incomplete_details.reason max_output_tokens")
}

func TestGeminiStreamToOpenAI2MaxTokensReportsIncomplete(t *testing.T) {
ctx := transformer.NewStreamContext()
ctx.ModelName = "gemini-3.8-flash"

out, err := GeminiStreamToOpenAI2([]byte(geminiTruncatedChunk), ctx)
if err != nil {
t.Fatalf("GeminiStreamToOpenAI2 failed: %v", err)
}

assertContains(t, string(out), `"type":"response.incomplete"`, "Expected a response.incomplete event for a Gemini MAX_TOKENS finish")
assertContains(t, string(out), `"reason":"max_output_tokens"`, "Expected incomplete_details.reason max_output_tokens")
}

func TestOpenAI2RespToClaudeIncompleteReportsMaxTokens(t *testing.T) {
out, err := OpenAI2RespToClaude([]byte(responsesTruncatedResp))
if err != nil {
t.Fatalf("OpenAI2RespToClaude failed: %v", err)
}

assertContains(t, string(out), `"stop_reason":"max_tokens"`, "Expected stop_reason max_tokens for an incomplete Responses result")
}

func TestOpenAI2StreamToClaudeIncompleteReportsMaxTokens(t *testing.T) {
ctx := transformer.NewStreamContext()
ctx.MessageID = "resp_1"

event := "data: {\"type\":\"response.incomplete\",\"response\":" + responsesTruncatedResp + "}\n\n"
out, err := OpenAI2StreamToClaude([]byte(event), ctx)
if err != nil {
t.Fatalf("OpenAI2StreamToClaude failed: %v", err)
}

assertContains(t, string(out), `"stop_reason":"max_tokens"`, "Expected stop_reason max_tokens for a response.incomplete event")
}
11 changes: 8 additions & 3 deletions internal/transformer/types.go
Original file line number Diff line number Diff line change
Expand Up @@ -206,6 +206,7 @@ type StreamContext struct {
// Claude -> Responses streaming fields
NextOutputIndex int
OutputText string
StopReason string // Claude stop_reason from message_delta
}

// NewStreamContext creates a new stream context with default values
Expand Down Expand Up @@ -385,9 +386,13 @@ type OpenAI2OutputItem struct {

// OpenAI2Response represents an OpenAI Responses API response
type OpenAI2Response struct {
ID string `json:"id"`
Object string `json:"object"` // "response"
Status string `json:"status"` // "completed", "failed", etc.
ID string `json:"id"`
Object string `json:"object"` // "response"
Status string `json:"status"` // "completed", "failed", etc.
// IncompleteDetails explains an "incomplete" status, e.g. reason "max_output_tokens".
IncompleteDetails *struct {
Reason string `json:"reason"`
} `json:"incomplete_details,omitempty"`
Output []OpenAI2OutputItem `json:"output"`
Usage struct {
InputTokens int `json:"input_tokens"`
Expand Down
Loading