fix: route ai polish through available models

api/main.py

@@ -107,6 +107,11 @@ def gpt_model_env(name: str, default: str | None = None) -> str:
 
 
 REWRITE_MODEL = gpt_model_env("REWRITE_MODEL")
+REWRITE_MODEL_FALLBACKS = [
+    model.strip()
+    for model in os.getenv("REWRITE_MODEL_FALLBACKS", "gpt-4o-mini,gemini-2.5-flash").split(",")
+    if model.strip()
+]
 VISION_MODEL = gpt_model_env("VISION_MODEL")
 IMAGE_BASE_URL = os.getenv("IMAGE_BASE_URL", LLM_BASE_URL).strip()
 IMAGE_API_KEY = os.getenv("IMAGE_API_KEY", LLM_API_KEY).strip()
@@ -5914,6 +5919,7 @@ class CreativeCopyResp(BaseModel):
 class PromptPolishReq(BaseModel):
     text: str
     system_prompt: str = ""
+    model: str = ""
     mode: Literal["image", "video", "general", "chat"] = "image"
     target_language: Literal["en", "zh", "keep"] = "en"
 
@@ -6320,11 +6326,26 @@ def _prompt_polish_fallback(req: PromptPolishReq) -> PromptPolishResp:
     return PromptPolishResp(model="fallback", text=_sanitize_polished_prompt(req, intent, _basic_polished_prompt(req, intent)))
 
 
-def _repair_polished_prompt(req: PromptPolishReq, intent: PromptIntent, output: str, *, allow_llm: bool = False) -> str:
+def _prompt_polish_model_candidates(req: PromptPolishReq) -> list[str]:
+    requested = (req.model or "").strip()
+    candidates = [requested, REWRITE_MODEL, *REWRITE_MODEL_FALLBACKS]
+    seen: set[str] = set()
+    out: list[str] = []
+    for model in candidates:
+        model = model.strip()
+        key = model.lower()
+        if model and key not in seen:
+            out.append(model)
+            seen.add(key)
+    return out
+
+
+def _repair_polished_prompt(req: PromptPolishReq, intent: PromptIntent, output: str, *, allow_llm: bool = False, model: str | None = None) -> str:
     out = _sanitize_polished_prompt(req, intent, output)
     issue = _polished_prompt_issue(intent, out)
     if not issue or not allow_llm or not LLM_API_KEY:
         return out
+    repair_model = (model or REWRITE_MODEL).strip() or REWRITE_MODEL
     repair_prompt = (
         "Repair the rewritten generation prompt so it follows the source input exactly.\n"
         f"Issue to fix: {issue}.\n"
@@ -6340,7 +6361,7 @@ def _repair_polished_prompt(req: PromptPolishReq, intent: PromptIntent, output:
     )
     try:
         resp = llm().chat.completions.create(
-            model=REWRITE_MODEL,
+            model=repair_model,
             messages=[
                 {"role": "system", "content": "You repair generation prompts by removing contradictions and preserving only source intent."},
                 {"role": "user", "content": repair_prompt},
@@ -6410,23 +6431,28 @@ def polish_prompt(req: PromptPolishReq) -> PromptPolishResp:
         prompt += f"\nUser-selected polishing guidance:\n{user_system[:1000]}\n"
     prompt += f"\nSource input:\n{intent.cleaned_text[:2500]}"
 
-    try:
-        resp = llm().chat.completions.create(
-            model=REWRITE_MODEL,
-            messages=[
-                {"role": "system", "content": "You are a neutral professional prompt editor. Preserve source intent exactly and never inject SKG or unrelated brands, products, platforms, people, or marketing context."},
-                {"role": "user", "content": prompt},
-            ],
-            temperature=0.45,
-            max_tokens=900,
-        )
-        out = _clean_prompt_output(resp.choices[0].message.content or "")
-        if not out:
-            out = _prompt_polish_fallback(req).text
-        return PromptPolishResp(model=REWRITE_MODEL, text=_repair_polished_prompt(req, intent, out, allow_llm=True))
-    except Exception as e:
-        print(f"[prompt polish fallback] {e}", flush=True)
-        return _prompt_polish_fallback(req)
+    model_errors: list[str] = []
+    for model in _prompt_polish_model_candidates(req):
+        try:
+            resp = llm().chat.completions.create(
+                model=model,
+                messages=[
+                    {"role": "system", "content": "You are a neutral professional prompt editor. Preserve source intent exactly and never inject SKG or unrelated brands, products, platforms, people, or marketing context."},
+                    {"role": "user", "content": prompt},
+                ],
+                temperature=0.45,
+                max_tokens=900,
+            )
+            out = _clean_prompt_output(resp.choices[0].message.content or "")
+            if not out:
+                raise RuntimeError("empty prompt polish response")
+            return PromptPolishResp(model=model, text=_repair_polished_prompt(req, intent, out, allow_llm=True, model=model))
+        except Exception as e:
+            message = str(e).replace("\n", " ")[:400]
+            model_errors.append(f"{model}: {message}")
+            print(f"[prompt polish model fallback] model={model} error={message}", flush=True)
+    print(f"[prompt polish fallback] {' | '.join(model_errors)}", flush=True)
+    return _prompt_polish_fallback(req)
 
 
 @app.post("/translate")