fix2

backend/internal/model/ai/provider/openai_compatible.go

@@ -84,7 +84,11 @@ func (p *OpenAICompatible) GenerateText(ctx context.Context, req domai.GenerateR
 		return nil, app.For(code.AI).InputMissingRequired("missing AI provider token")
 	}
 
-	body := p.buildChatCompletionRequest(req, false)
+	// Stream even for one-shot generation: long outputs (e.g. knowledge graph /
+	// research map) can exceed the provider's Cloudflare ~100s origin timeout in
+	// non-stream mode and surface as HTTP 524. Streaming keeps bytes flowing and
+	// we accumulate them back into a single result.
+	body := p.buildChatCompletionRequest(req, true)
 	payload, err := json.Marshal(body)
 	if err != nil {
 		return nil, err
@@ -96,7 +100,7 @@ func (p *OpenAICompatible) GenerateText(ctx context.Context, req domai.GenerateR
 	}
 	httpReq.Header.Set("Authorization", "Bearer "+req.Credential.APIKey)
 	httpReq.Header.Set("Content-Type", "application/json")
-	httpReq.Header.Set("Accept", "application/json")
+	httpReq.Header.Set("Accept", "text/event-stream")
 
 	resp, err := p.client.Do(httpReq)
 	if err != nil {
@@ -104,26 +108,70 @@ func (p *OpenAICompatible) GenerateText(ctx context.Context, req domai.GenerateR
 	}
 	defer resp.Body.Close()
 
-	data, err := io.ReadAll(io.LimitReader(resp.Body, 1<<20))
-	if err != nil {
-		return nil, err
-	}
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		_, _ = io.Copy(io.Discard, io.LimitReader(resp.Body, 4096))
 		return nil, providerHTTPError("AI provider request failed", resp.StatusCode, resp.Status)
 	}
 
-	var chatResp openAIChatResponse
+	return accumulateChatStream(resp.Body)
-	if err := json.Unmarshal(data, &chatResp); err != nil {
+}
-		return nil, app.For(code.AI).SvcThirdParty("failed to parse AI provider chat response")
+
+// accumulateChatStream consumes an OpenAI-compatible SSE stream and folds the
+// chunks back into a single GenerateResult. Final answer content and reasoning
+// are kept apart so reasoning never gets prepended to (and corrupts) the parsed
+// answer; reasoning is only used as a fallback when no content was emitted.
+func accumulateChatStream(body io.Reader) (*domai.GenerateResult, error) {
+	scanner := bufio.NewScanner(body)
+	scanner.Buffer(make([]byte, 0, 64*1024), 1024*1024)
+
+	var content strings.Builder
+	var reasoning strings.Builder
+	finishReason := ""
+
+	for scanner.Scan() {
+		line := strings.TrimSpace(scanner.Text())
+		if line == "" || !strings.HasPrefix(line, "data:") {
+			continue
+		}
+		data := strings.TrimSpace(strings.TrimPrefix(line, "data:"))
+		if data == "[DONE]" {
+			break
+		}
+		var chunk openAIStreamChunk
+		if err := json.Unmarshal([]byte(data), &chunk); err != nil {
+			return nil, app.For(code.AI).SvcThirdParty("failed to parse AI provider chat response")
+		}
+		for _, choice := range chunk.Choices {
+			delta := choice.Delta
+			if delta.Content != "" {
+				content.WriteString(delta.Content)
+			} else if delta.Text != "" {
+				content.WriteString(delta.Text)
+			}
+			if delta.ReasoningContent != "" {
+				reasoning.WriteString(delta.ReasoningContent)
+			} else if delta.Reasoning != "" {
+				reasoning.WriteString(delta.Reasoning)
+			}
+			if choice.FinishReason != "" {
+				finishReason = choice.FinishReason
+			}
+		}
 	}
-	if len(chatResp.Choices) == 0 {
+	if err := scanner.Err(); err != nil {
-		return nil, app.For(code.AI).SvcThirdParty("AI provider returned no choices")
+		return nil, app.For(code.AI).SvcThirdParty("AI provider stream read failed: " + err.Error())
 	}
 
-	choice := chatResp.Choices[0]
+	text := content.String()
+	if strings.TrimSpace(text) == "" {
+		text = reasoning.String()
+	}
+	if strings.TrimSpace(text) == "" {
+		return nil, app.For(code.AI).SvcThirdParty("AI provider returned no choices")
+	}
 	return &domai.GenerateResult{
-		Text:         extractAssistantText(choice.Message),
+		Text:         text,
-		FinishReason: choice.FinishReason,
+		FinishReason: finishReason,
 	}, nil
 }
 

backend/internal/model/ai/provider/openai_compatible_test.go

@@ -60,8 +60,8 @@ func TestOpenAICompatible_GenerateText_UsesContent(t *testing.T) {
 		if err := json.Unmarshal(body, &req); err != nil {
 			t.Fatalf("unmarshal request: %v", err)
 		}
-		if req.Stream {
+		if !req.Stream {
-			t.Fatal("expected non-stream request")
+			t.Fatal("expected stream request to avoid upstream 524 on long generations")
 		}
 		if req.Thinking == nil || req.Thinking.Type != "disabled" {
 			t.Fatalf("thinking = %+v, want disabled for deepseek on opencode", req.Thinking)
@@ -69,7 +69,10 @@ func TestOpenAICompatible_GenerateText_UsesContent(t *testing.T) {
 		if req.MaxTokens == nil || *req.MaxTokens < 2048 {
 			t.Fatalf("max_tokens = %+v, want >= 2048 for deepseek on opencode", req.MaxTokens)
 		}
-		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"hello"},"finish_reason":"stop"}]}`))
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte("data: {\"choices\":[{\"delta\":{\"content\":\"hel\"}}]}\n\n"))
+		_, _ = w.Write([]byte("data: {\"choices\":[{\"delta\":{\"content\":\"lo\"},\"finish_reason\":\"stop\"}]}\n\n"))
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
 	}))
 	defer server.Close()
 
@@ -92,7 +95,9 @@ func TestOpenAICompatible_GenerateText_UsesContent(t *testing.T) {
 
 func TestOpenAICompatible_GenerateText_FallsBackToReasoningContent(t *testing.T) {
 	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"","reasoning_content":"thinking only"},"finish_reason":"length"}]}`))
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte("data: {\"choices\":[{\"delta\":{\"reasoning_content\":\"thinking only\"},\"finish_reason\":\"length\"}]}\n\n"))
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
 	}))
 	defer server.Close()