fix(stream-translate): surface image_generation_call.partial_image as content chunks

Author: ImogeneOctaviap794

proxy/images_generations.go

@@ -36,7 +36,9 @@ import (
 //   - moderation       (string)  "auto" / "low"
 //   - response_format  (string)  "url" / "b64_json"，默认 "url"（我们的扩展，
 //                                OpenAI gpt-image-1 官方只返回 b64_json）
-//   - partial_images   (int)     0-3（流式，本实现暂返回终稿）
+//   - partial_images   (int)     0-3（流式渐进预览，v1.7.50 起经 ChatCompletions
+//                                 翻译层透出为多张 markdown image content chunk；
+//                                 /v1/images/generations 内部走非流式仍返回终稿）
 //   - user             (string)  终端用户标识（仅透传到日志）
 //
 // 响应格式（与 OpenAI gpt-image-1 对齐）：

proxy/model_overrides_test.go

@@ -216,3 +216,55 @@ func TestTranslateStreamChunk_NonImageOutputItemDoneIgnored(t *testing.T) {
 		t.Fatalf("non-image output_item.done should return (nil, false), got (%v, %v)", chunk, done)
 	}
 }
+
+// 回归 v1.7.49 bug：metadata.image_partial_images 被透传到上游，但翻译层把
+// response.image_generation_call.partial_image 事件直接 drop 掉。
+// 修复后必须 emit 为 markdown image content chunk（和终稿同源路径）。
+func TestTranslateStreamChunk_PartialImageEmitsMarkdown(t *testing.T) {
+	b64 := "iVBORw0KGgoAAAANSUhEUgAAAAEAAAAB"
+	event := []byte(`{"type":"response.image_generation_call.partial_image","item_id":"ig_x","output_index":0,"sequence_number":0,"partial_image_index":0,"partial_image_b64":"` + b64 + `"}`)
+
+	chunk, done := TranslateStreamChunk(event, "gpt-image-2", "chatcmpl-x", 0)
+	if done {
+		t.Fatal("partial_image must not finish the stream")
+	}
+	if chunk == nil {
+		t.Fatal("expected content chunk for partial_image, got nil (regression: drop bug returned)")
+	}
+	s := string(chunk)
+	if !strings.Contains(s, "data:image/png;base64,"+b64) {
+		t.Fatalf("chunk missing markdown image; got %s", s)
+	}
+	if !strings.Contains(s, `"chat.completion.chunk"`) {
+		t.Fatalf("not a chat.completion.chunk; got %s", s)
+	}
+}
+
+func TestStreamTranslator_PartialImageEmitsMarkdown(t *testing.T) {
+	b64 := "iVBORw0KGgoAAAANSUhEUgAAAAEAAAAB"
+	event := []byte(`{"type":"response.image_generation_call.partial_image","partial_image_b64":"` + b64 + `"}`)
+
+	st := NewStreamTranslator("chatcmpl-x", "gpt-image-2", 0)
+	chunk, done := st.Translate(event)
+	if done {
+		t.Fatal("partial_image must not finish the stream")
+	}
+	if chunk == nil {
+		t.Fatal("expected content chunk, got nil (regression: drop bug returned)")
+	}
+	if !strings.Contains(string(chunk), "data:image/png;base64,"+b64) {
+		t.Fatalf("chunk missing markdown image; got %s", string(chunk))
+	}
+	if st.HasToolCalls {
+		t.Fatal("partial_image must not set HasToolCalls")
+	}
+}
+
+// partial_image 无 b64 时（理论上不会，但防御）必须静默忽略。
+func TestTranslateStreamChunk_PartialImageMissingB64Ignored(t *testing.T) {
+	event := []byte(`{"type":"response.image_generation_call.partial_image","partial_image_index":0}`)
+	chunk, done := TranslateStreamChunk(event, "gpt-image-2", "id", 0)
+	if done || chunk != nil {
+		t.Fatalf("missing partial_image_b64 should be silently ignored, got (%v, %v)", chunk, done)
+	}
+}

proxy/translator.go

@@ -1035,14 +1035,25 @@ func TranslateStreamChunk(eventData []byte, model string, chunkID string, create
 		}
 		return newErrorResponse(errMsg), true
 
+	case "response.image_generation_call.partial_image":
+		// 渐进式预览帧：客户端通过 metadata.image_partial_images = N
+		// 让上游在终稿之前发 N 张中间帧。每帧 partial_image_b64 字段都是
+		// 一张完整的 base64 PNG，按图像生成的 SSE 顺序到达。
+		// 这里和终稿走同样的落盘路径，作为 markdown image 注入 content delta；
+		// 客户端可以收到一系列 ![image](url1) ![image](url2) ![image](finalurl)。
+		if b64 := gjson.GetBytes(eventData, "partial_image_b64").String(); b64 != "" {
+			return newContentChunk(chunkID, model, created, imageMarkdownFromBase64(b64)), false
+		}
+		return nil, false
+
 	case "response.content_part.done",
 		"response.created", "response.in_progress",
 		"response.output_item.added", "response.content_part.added",
 		"response.reasoning_summary_text.done",
 		"response.reasoning.encrypted_content.delta", "response.reasoning.encrypted_content.done",
 		"response.reasoning_summary_part.added", "response.reasoning_summary_part.done",
 		"response.image_generation_call.in_progress", "response.image_generation_call.generating",
-		"response.image_generation_call.partial_image", "response.image_generation_call.completed":
+		"response.image_generation_call.completed":
 		return nil, false
 
 	default:
@@ -1147,14 +1158,21 @@ func (st *StreamTranslator) Translate(eventData []byte) ([]byte, bool) {
 		}
 		return newErrorResponse(errMsg), true
 
+	case "response.image_generation_call.partial_image":
+		// 渐进式预览帧：见 stateless Translate 路径的同名 case 说明。
+		if b64 := gjson.GetBytes(eventData, "partial_image_b64").String(); b64 != "" {
+			return newContentChunk(st.ChunkID, st.Model, st.Created, imageMarkdownFromBase64(b64)), false
+		}
+		return nil, false
+
 	case "response.content_part.done",
 		"response.created", "response.in_progress",
 		"response.content_part.added",
 		"response.reasoning_summary_text.done",
 		"response.reasoning.encrypted_content.delta", "response.reasoning.encrypted_content.done",
 		"response.reasoning_summary_part.added", "response.reasoning_summary_part.done",
 		"response.image_generation_call.in_progress", "response.image_generation_call.generating",
-		"response.image_generation_call.partial_image", "response.image_generation_call.completed":
+		"response.image_generation_call.completed":
 		return nil, false
 
 	default: