fix: 优化架构

engine/main.py

@@ -12,13 +12,14 @@ from typing import Any
 from moviepy import ImageClip
 from PIL import Image, ImageDraw, ImageFont
 
-from engine.audio_gen import synthesize_scenes
+from engine.model_factory import get_model
+from engine.prompt_injector import inject_prompt
+from engine.adapters.image.mock_adapter import MockImageGen
 from engine.assembler import assemble_clips
 from engine.comfy_client import ComfyClient
 from engine.config import AppConfig
 from engine.director import scenes_to_shots
 from engine.shot_executor import render_shot
-from engine.script_gen import generate_scenes, refine_scene
 from engine.task_store import create_task, update_shot_status, update_task_status
 from engine.types import Scene
 from engine.video_editor import Segment, render_final
@@ -28,13 +29,15 @@ def _emit(line: str) -> None:
     print(line, flush=True)
 
 
-def _emit_scene(scene_idx: int, scene: Scene) -> None:
+def _emit_scene(scene_idx: int, scene: Scene, extra: dict[str, Any] | None = None) -> None:
     payload = {
         "index": scene_idx,
         "image_prompt": scene.image_prompt,
         "video_motion": scene.video_motion,
         "narration": scene.narration,
     }
+    if extra:
+        payload.update(extra)
     _emit("SCENE_JSON " + json.dumps(payload, ensure_ascii=False))
 
 
@@ -136,9 +139,50 @@ def _fallback_scenes(prompt: str) -> list[Scene]:
     ]
 
 
+def _generate_scene_preview(
+    *,
+    cfg: AppConfig,
+    out_dir: Path,
+    image_prompt: str,
+    style: str | None,
+    character: str | None,
+) -> str | None:
+    try:
+        image_gen = get_model("image", cfg)
+    except Exception:
+        image_gen = get_model("image_fallback", cfg)
+
+    global_cfg = dict(cfg.get("global", {}) or {})
+    if style:
+        global_cfg["style"] = style
+    if character:
+        global_cfg["character"] = character
+
+    prompt_obj = inject_prompt(global_cfg, {"prompt": image_prompt})
+    try:
+        image_path = image_gen.generate(prompt_obj, out_dir)
+    except Exception:
+        try:
+            image_path = get_model("image_fallback", cfg).generate(prompt_obj, out_dir)
+        except Exception:
+            # Last-resort hard fallback: never block script stage due to preview failures.
+            image_path = MockImageGen().generate(prompt_obj, out_dir)
+
+    p = Path(str(image_path))
+    if not p.exists():
+        return None
+    return f"/api/static/{out_dir.name}/{p.name}"
+
+
 def _has_llm_key(cfg: AppConfig) -> bool:
-    api_key_env = str(cfg.get("openai.api_key_env", "OPENAI_API_KEY"))
-    return bool(os.environ.get(api_key_env))
+    api_key_env = str(cfg.get("openai.api_key_env", "OPENAI_API_KEY") or "OPENAI_API_KEY").strip()
+    # Env var name case.
+    if os.environ.get(api_key_env):
+        return True
+    # Literal key case (DashScope / OpenAI-compatible).
+    if api_key_env.startswith("sk-"):
+        return True
+    return False
 
 
 def _parse_scenes_from_obj(obj: Any) -> list[Scene]:
@@ -239,7 +283,8 @@ def step_script(prompt: str, cfg: AppConfig, mock: bool, *, style: str | None, c
         # fallback scenes still should include global injection
         scenes = _fallback_scenes(prompt)
     else:
-        scenes = generate_scenes(prompt2, cfg)
+        llm = get_model("llm", cfg)
+        scenes = llm.generate_script(prompt2, context=None)
 
     out_dir.mkdir(parents=True, exist_ok=True)
     _emit("SCRIPT_BEGIN")
@@ -249,7 +294,14 @@ def step_script(prompt: str, cfg: AppConfig, mock: bool, *, style: str | None, c
             video_motion=s.video_motion,
             narration=s.narration,
         )
-        _emit_scene(idx, s2)
+        preview_url = _generate_scene_preview(
+            cfg=cfg,
+            out_dir=out_dir,
+            image_prompt=s2.image_prompt,
+            style=style,
+            character=character,
+        )
+        _emit_scene(idx, s2, extra={"preview_url": preview_url or ""})
     _emit("SCRIPT_END")
     (out_dir / "scenes.json").write_text(
         json.dumps(
@@ -292,8 +344,9 @@ def step_refine(
             narration=(s.narration + "（更凝练）")[:30],
         )
     else:
-        # Ensure globals are visible to LLM, and inject to output image prompt.
-        refined0 = refine_scene(prompt=prompt2, scenes=scenes, target_index=target_index, cfg=cfg)
+        llm = get_model("llm", cfg)
+        # Context carries prompt + scenes for consistent refinement.
+        refined0 = llm.refine_scene(scenes[target_index - 1], context={"prompt": prompt2, "scenes": scenes, "target_index": target_index})
         refined = Scene(
             image_prompt=_decorate_image_prompt(refined0.image_prompt, style=style, character=character),
             video_motion=refined0.video_motion,
@@ -301,7 +354,14 @@ def step_refine(
         )
 
     # Keep the original index for frontend replacement.
-    _emit_scene(scene_index, refined)
+    preview_url = _generate_scene_preview(
+        cfg=cfg,
+        out_dir=out_dir,
+        image_prompt=refined.image_prompt,
+        style=style,
+        character=character,
+    )
+    _emit_scene(scene_index, refined, extra={"preview_url": preview_url or ""})
     out_dir.mkdir(parents=True, exist_ok=True)
     (out_dir / f"refine_scene_{scene_index}.json").write_text(
         json.dumps(