fix: use parent_message_id and fresh PoW headers for empty-output retry and continue

Previously retry/continue requests reused the initial PoW header and
lacked parent_message_id, causing them to land as disconnected root
messages in the DeepSeek session instead of proper follow-up turns.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

internal/httpapi/openai/chat/chat_stream_runtime.go

@@ -38,6 +38,7 @@ type chatStreamRuntime struct {
 	thinking              strings.Builder
 	toolDetectionThinking strings.Builder
 	text                  strings.Builder
+	responseMessageID     int
 
 	finalThinking     string
 	finalText         string
@@ -234,6 +235,9 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
+	if parsed.ResponseMessageID > 0 {
+		s.responseMessageID = parsed.ResponseMessageID
+	}
 	if parsed.ContentFilter {
 		if strings.TrimSpace(s.text.String()) == "" {
 			return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReason("content_filter")}

internal/httpapi/openai/chat/empty_retry_runtime.go

@@ -23,6 +23,7 @@ type chatNonStreamResult struct {
 	detectedCalls         int
 	body                  map[string]any
 	finishReason          string
+	responseMessageID     int
 }
 
 func (h *Handler) handleNonStreamWithRetry(w http.ResponseWriter, ctx context.Context, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, completionID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, historySession *chatHistorySession) {
@@ -50,9 +51,14 @@ func (h *Handler) handleNonStreamWithRetry(w http.ResponseWriter, ctx context.Co
 		}
 
 		attempts++
-		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "chat.completions", "stream", false, "retry_attempt", attempts)
-		retryPayload := clonePayloadWithEmptyOutputRetryPrompt(payload)
-		nextResp, err := h.DS.CallCompletion(ctx, a, retryPayload, pow, 3)
+		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "chat.completions", "stream", false, "retry_attempt", attempts, "parent_message_id", result.responseMessageID)
+		retryPow, powErr := h.DS.GetPow(ctx, a, 3)
+		if powErr != nil {
+			config.Logger.Warn("[openai_empty_retry] retry PoW fetch failed, falling back to original PoW", "surface", "chat.completions", "stream", false, "retry_attempt", attempts, "error", powErr)
+			retryPow = pow
+		}
+		retryPayload := clonePayloadForEmptyOutputRetry(payload, result.responseMessageID)
+		nextResp, err := h.DS.CallCompletion(ctx, a, retryPayload, retryPow, 3)
 		if err != nil {
 			if historySession != nil {
 				historySession.error(http.StatusInternalServerError, "Failed to get completion.", "error", result.thinking, result.text)
@@ -91,6 +97,7 @@ func (h *Handler) collectChatNonStreamAttempt(w http.ResponseWriter, resp *http.
 		detectedCalls:         len(detected.Calls),
 		body:                  respBody,
 		finishReason:          chatFinishReason(respBody),
+		responseMessageID:     result.ResponseMessageID,
 	}, true
 }
 
@@ -152,8 +159,13 @@ func (h *Handler) handleStreamWithRetry(w http.ResponseWriter, r *http.Request,
 			return
 		}
 		attempts++
-		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "chat.completions", "stream", true, "retry_attempt", attempts)
-		nextResp, err := h.DS.CallCompletion(r.Context(), a, clonePayloadWithEmptyOutputRetryPrompt(payload), pow, 3)
+		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "chat.completions", "stream", true, "retry_attempt", attempts, "parent_message_id", streamRuntime.responseMessageID)
+		retryPow, powErr := h.DS.GetPow(r.Context(), a, 3)
+		if powErr != nil {
+			config.Logger.Warn("[openai_empty_retry] retry PoW fetch failed, falling back to original PoW", "surface", "chat.completions", "stream", true, "retry_attempt", attempts, "error", powErr)
+			retryPow = pow
+		}
+		nextResp, err := h.DS.CallCompletion(r.Context(), a, clonePayloadForEmptyOutputRetry(payload, streamRuntime.responseMessageID), retryPow, 3)
 		if err != nil {
 			failChatStreamRetry(streamRuntime, historySession, http.StatusInternalServerError, "Failed to get completion.", "error")
 			config.Logger.Warn("[openai_empty_retry] retry request failed", "surface", "chat.completions", "stream", true, "retry_attempt", attempts, "error", err)

internal/httpapi/openai/chat/handler.go

@@ -131,8 +131,8 @@ func emptyOutputRetryMaxAttempts() int {
 	return shared.EmptyOutputRetryMaxAttempts()
 }
 
-func clonePayloadWithEmptyOutputRetryPrompt(payload map[string]any) map[string]any {
-	return shared.ClonePayloadWithEmptyOutputRetryPrompt(payload)
+func clonePayloadForEmptyOutputRetry(payload map[string]any, parentMessageID int) map[string]any {
+	return shared.ClonePayloadForEmptyOutputRetry(payload, parentMessageID)
 }
 
 func usagePromptWithEmptyOutputRetry(originalPrompt string, retryAttempts int) string {

internal/httpapi/openai/chat/handler_chat.go

@@ -22,6 +22,10 @@ func (h *Handler) ChatCompletions(w http.ResponseWriter, r *http.Request) {
 		h.handleVercelStreamRelease(w, r)
 		return
 	}
+	if isVercelStreamPowRequest(r) {
+		h.handleVercelStreamPow(w, r)
+		return
+	}
 	if isVercelStreamPrepareRequest(r) {
 		h.handleVercelStreamPrepare(w, r)
 		return

internal/httpapi/openai/chat/vercel_stream.go

@@ -150,6 +150,44 @@ func (h *Handler) handleVercelStreamRelease(w http.ResponseWriter, r *http.Reque
 	writeJSON(w, http.StatusOK, map[string]any{"success": true})
 }
 
+func (h *Handler) handleVercelStreamPow(w http.ResponseWriter, r *http.Request) {
+	if !config.IsVercel() {
+		http.NotFound(w, r)
+		return
+	}
+	internalSecret := vercelInternalSecret()
+	internalToken := strings.TrimSpace(r.Header.Get("X-Ds2-Internal-Token"))
+	if internalSecret == "" || subtle.ConstantTimeCompare([]byte(internalToken), []byte(internalSecret)) != 1 {
+		writeOpenAIError(w, http.StatusUnauthorized, "unauthorized internal request")
+		return
+	}
+
+	var req map[string]any
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		writeOpenAIError(w, http.StatusBadRequest, "invalid json")
+		return
+	}
+	leaseID, _ := req["lease_id"].(string)
+	leaseID = strings.TrimSpace(leaseID)
+	if leaseID == "" {
+		writeOpenAIError(w, http.StatusBadRequest, "lease_id is required")
+		return
+	}
+	leaseAuth := h.lookupStreamLeaseAuth(leaseID)
+	if leaseAuth == nil {
+		writeOpenAIError(w, http.StatusNotFound, "stream lease not found or expired")
+		return
+	}
+	powHeader, err := h.DS.GetPow(r.Context(), leaseAuth, 3)
+	if err != nil {
+		writeOpenAIError(w, http.StatusInternalServerError, "Failed to get PoW.")
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]any{
+		"pow_header": powHeader,
+	})
+}
+
 func isVercelStreamPrepareRequest(r *http.Request) bool {
 	if r == nil {
 		return false
@@ -164,6 +202,13 @@ func isVercelStreamReleaseRequest(r *http.Request) bool {
 	return strings.TrimSpace(r.URL.Query().Get("__stream_release")) == "1"
 }
 
+func isVercelStreamPowRequest(r *http.Request) bool {
+	if r == nil {
+		return false
+	}
+	return strings.TrimSpace(r.URL.Query().Get("__stream_pow")) == "1"
+}
+
 func vercelInternalSecret() string {
 	if v := strings.TrimSpace(os.Getenv("DS2API_VERCEL_INTERNAL_SECRET")); v != "" {
 		return v
@@ -199,6 +244,20 @@ func (h *Handler) holdStreamLease(a *auth.RequestAuth) string {
 	return leaseID
 }
 
+func (h *Handler) lookupStreamLeaseAuth(leaseID string) *auth.RequestAuth {
+	leaseID = strings.TrimSpace(leaseID)
+	if leaseID == "" {
+		return nil
+	}
+	h.leaseMu.Lock()
+	lease, ok := h.streamLeases[leaseID]
+	h.leaseMu.Unlock()
+	if !ok || time.Now().After(lease.ExpiresAt) {
+		return nil
+	}
+	return lease.Auth
+}
+
 func (h *Handler) releaseStreamLease(leaseID string) bool {
 	leaseID = strings.TrimSpace(leaseID)
 	if leaseID == "" {

internal/httpapi/openai/responses/empty_retry_runtime.go

@@ -24,6 +24,7 @@ type responsesNonStreamResult struct {
 	contentFilter         bool
 	parsed                toolcall.ToolCallParseResult
 	body                  map[string]any
+	responseMessageID     int
 }
 
 func (h *Handler) handleResponsesNonStreamWithRetry(w http.ResponseWriter, ctx context.Context, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, owner, responseID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolChoice promptcompat.ToolChoicePolicy, traceID string) {
@@ -50,8 +51,13 @@ func (h *Handler) handleResponsesNonStreamWithRetry(w http.ResponseWriter, ctx c
 		}
 
 		attempts++
-		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "responses", "stream", false, "retry_attempt", attempts)
-		nextResp, err := h.DS.CallCompletion(ctx, a, clonePayloadWithEmptyOutputRetryPrompt(payload), pow, 3)
+		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "responses", "stream", false, "retry_attempt", attempts, "parent_message_id", result.responseMessageID)
+		retryPow, powErr := h.DS.GetPow(ctx, a, 3)
+		if powErr != nil {
+			config.Logger.Warn("[openai_empty_retry] retry PoW fetch failed, falling back to original PoW", "surface", "responses", "stream", false, "retry_attempt", attempts, "error", powErr)
+			retryPow = pow
+		}
+		nextResp, err := h.DS.CallCompletion(ctx, a, clonePayloadForEmptyOutputRetry(payload, result.responseMessageID), retryPow, 3)
 		if err != nil {
 			writeOpenAIError(w, http.StatusInternalServerError, "Failed to get completion.")
 			config.Logger.Warn("[openai_empty_retry] retry request failed", "surface", "responses", "stream", false, "retry_attempt", attempts, "error", err)
@@ -86,6 +92,7 @@ func (h *Handler) collectResponsesNonStreamAttempt(w http.ResponseWriter, resp *
 		contentFilter:         result.ContentFilter,
 		parsed:                textParsed,
 		body:                  responseObj,
+		responseMessageID:     result.ResponseMessageID,
 	}, true
 }
 
@@ -135,8 +142,13 @@ func (h *Handler) handleResponsesStreamWithRetry(w http.ResponseWriter, r *http.
 			return
 		}
 		attempts++
-		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "responses", "stream", true, "retry_attempt", attempts)
-		nextResp, err := h.DS.CallCompletion(r.Context(), a, clonePayloadWithEmptyOutputRetryPrompt(payload), pow, 3)
+		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "responses", "stream", true, "retry_attempt", attempts, "parent_message_id", streamRuntime.responseMessageID)
+		retryPow, powErr := h.DS.GetPow(r.Context(), a, 3)
+		if powErr != nil {
+			config.Logger.Warn("[openai_empty_retry] retry PoW fetch failed, falling back to original PoW", "surface", "responses", "stream", true, "retry_attempt", attempts, "error", powErr)
+			retryPow = pow
+		}
+		nextResp, err := h.DS.CallCompletion(r.Context(), a, clonePayloadForEmptyOutputRetry(payload, streamRuntime.responseMessageID), retryPow, 3)
 		if err != nil {
 			streamRuntime.failResponse(http.StatusInternalServerError, "Failed to get completion.", "error")
 			config.Logger.Warn("[openai_empty_retry] retry request failed", "surface", "responses", "stream", true, "retry_attempt", attempts, "error", err)

internal/httpapi/openai/responses/handler.go

@@ -121,8 +121,8 @@ func emptyOutputRetryMaxAttempts() int {
 	return shared.EmptyOutputRetryMaxAttempts()
 }
 
-func clonePayloadWithEmptyOutputRetryPrompt(payload map[string]any) map[string]any {
-	return shared.ClonePayloadWithEmptyOutputRetryPrompt(payload)
+func clonePayloadForEmptyOutputRetry(payload map[string]any, parentMessageID int) map[string]any {
+	return shared.ClonePayloadForEmptyOutputRetry(payload, parentMessageID)
 }
 
 func usagePromptWithEmptyOutputRetry(originalPrompt string, retryAttempts int) string {

internal/httpapi/openai/responses/responses_stream_runtime_core.go

@@ -39,6 +39,7 @@ type responsesStreamRuntime struct {
 	toolDetectionThinking strings.Builder
 	text                  strings.Builder
 	visibleText           strings.Builder
+	responseMessageID     int
 	streamToolCallIDs     map[int]string
 	functionItemIDs       map[int]string
 	functionOutputIDs     map[int]int
@@ -205,6 +206,9 @@ func (s *responsesStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Pa
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
+	if parsed.ResponseMessageID > 0 {
+		s.responseMessageID = parsed.ResponseMessageID
+	}
 	if parsed.ContentFilter || parsed.ErrorMessage != "" {
 		return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReason("content_filter")}
 	}

internal/httpapi/openai/shared/empty_retry.go

@@ -13,12 +13,23 @@ func EmptyOutputRetryMaxAttempts() int {
 }
 
 func ClonePayloadWithEmptyOutputRetryPrompt(payload map[string]any) map[string]any {
+	return ClonePayloadForEmptyOutputRetry(payload, 0)
+}
+
+// ClonePayloadForEmptyOutputRetry creates a retry payload with the suffix
+// appended and, if parentMessageID > 0, sets parent_message_id so the
+// retry is submitted as a proper follow-up turn in the same DeepSeek
+// session rather than a disconnected root message.
+func ClonePayloadForEmptyOutputRetry(payload map[string]any, parentMessageID int) map[string]any {
 	clone := make(map[string]any, len(payload))
 	for k, v := range payload {
 		clone[k] = v
 	}
 	original, _ := payload["prompt"].(string)
 	clone["prompt"] = AppendEmptyOutputRetrySuffix(original)
+	if parentMessageID > 0 {
+		clone["parent_message_id"] = parentMessageID
+	}
 	return clone
 }
 

internal/httpapi/openai/stream_status_test.go

@@ -285,7 +285,7 @@ func TestChatCompletionsStreamEmitsFailureFrameWhenUpstreamOutputEmpty(t *testin
 
 func TestChatCompletionsStreamRetriesEmptyOutputOnSameSession(t *testing.T) {
 	ds := &streamStatusDSSeqStub{resps: []*http.Response{
-		makeOpenAISSEHTTPResponse(`data: {"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
+		makeOpenAISSEHTTPResponse(`data: {"response_message_id":42,"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
@@ -313,6 +313,10 @@ func TestChatCompletionsStreamRetriesEmptyOutputOnSameSession(t *testing.T) {
 	if !strings.Contains(retryPrompt, "Previous reply had no visible output. Please regenerate the visible final answer or tool call now.") {
 		t.Fatalf("expected retry suffix in prompt, got %q", retryPrompt)
 	}
+	// Verify multi-turn chaining: retry must set parent_message_id from first call's response_message_id.
+	if parentID, ok := ds.payloads[1]["parent_message_id"].(int); !ok || parentID != 42 {
+		t.Fatalf("expected retry parent_message_id=42, got %#v", ds.payloads[1]["parent_message_id"])
+	}
 
 	frames, done := parseSSEDataFrames(t, rec.Body.String())
 	if !done {
@@ -341,7 +345,7 @@ func TestChatCompletionsStreamRetriesEmptyOutputOnSameSession(t *testing.T) {
 
 func TestChatCompletionsNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	ds := &streamStatusDSSeqStub{resps: []*http.Response{
-		makeOpenAISSEHTTPResponse(`data: {"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
+		makeOpenAISSEHTTPResponse(`data: {"response_message_id":99,"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
@@ -362,6 +366,10 @@ func TestChatCompletionsNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	if len(ds.payloads) != 2 {
 		t.Fatalf("expected one synthetic retry call, got %d", len(ds.payloads))
 	}
+	// Verify multi-turn chaining.
+	if parentID, ok := ds.payloads[1]["parent_message_id"].(int); !ok || parentID != 99 {
+		t.Fatalf("expected retry parent_message_id=99, got %#v", ds.payloads[1]["parent_message_id"])
+	}
 	var out map[string]any
 	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
 		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())
@@ -452,7 +460,7 @@ func TestResponsesStreamUsageIgnoresBatchAccumulatedTokenUsage(t *testing.T) {
 
 func TestResponsesStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	ds := &streamStatusDSSeqStub{resps: []*http.Response{
-		makeOpenAISSEHTTPResponse(`data: {"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
+		makeOpenAISSEHTTPResponse(`data: {"response_message_id":77,"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
@@ -473,6 +481,10 @@ func TestResponsesStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	if len(ds.payloads) != 2 {
 		t.Fatalf("expected one synthetic retry call, got %d", len(ds.payloads))
 	}
+	// Verify multi-turn chaining.
+	if parentID, ok := ds.payloads[1]["parent_message_id"].(int); !ok || parentID != 77 {
+		t.Fatalf("expected retry parent_message_id=77, got %#v", ds.payloads[1]["parent_message_id"])
+	}
 	body := rec.Body.String()
 	if strings.Contains(body, "response.failed") {
 		t.Fatalf("did not expect premature response.failed, body=%s", body)
@@ -487,7 +499,7 @@ func TestResponsesStreamRetriesThinkingOnlyOutput(t *testing.T) {
 
 func TestResponsesNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	ds := &streamStatusDSSeqStub{resps: []*http.Response{
-		makeOpenAISSEHTTPResponse(`data: {"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
+		makeOpenAISSEHTTPResponse(`data: {"response_message_id":88,"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
@@ -508,6 +520,10 @@ func TestResponsesNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	if len(ds.payloads) != 2 {
 		t.Fatalf("expected one synthetic retry call, got %d", len(ds.payloads))
 	}
+	// Verify multi-turn chaining.
+	if parentID, ok := ds.payloads[1]["parent_message_id"].(int); !ok || parentID != 88 {
+		t.Fatalf("expected retry parent_message_id=88, got %#v", ds.payloads[1]["parent_message_id"])
+	}
 	var out map[string]any
 	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
 		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())