Harden Ollama MSR analysis parsing

ltx_director.py

@@ -444,6 +444,32 @@ _ANALYZE_PROMPT = (
 )
 
 
+def _extract_ollama_generated_text(resp_json: dict) -> str:
+    """Handle Ollama generate/chat variants and reasoning-capable models."""
+    if not isinstance(resp_json, dict):
+        return ""
+
+    candidates = []
+    candidates.append(resp_json.get("response"))
+    candidates.append(resp_json.get("thinking"))
+    candidates.append(resp_json.get("content"))
+
+    message = resp_json.get("message")
+    if isinstance(message, dict):
+        candidates.append(message.get("content"))
+        candidates.append(message.get("reasoning_content"))
+        candidates.append(message.get("thinking"))
+
+    for candidate in candidates:
+        if isinstance(candidate, str) and candidate.strip():
+            text = candidate.strip()
+            if "<think>" in text:
+                text = text.split("</think>")[-1].strip()
+            if text:
+                return text
+    return ""
+
+
 def _resolve_provider(data):
     provider = (data.get("provider") or "ollama").lower()
     defs = _PROVIDER_DEFAULTS.get(provider, _PROVIDER_DEFAULTS["ollama"])
@@ -492,12 +518,30 @@ async def analyze_character_endpoint(request):
                         "model": model_name, "prompt": _ANALYZE_PROMPT,
                         "images": cleaned_b64_list, "stream": False, "keep_alive": 0,
                     }
-                    async with session.post(f"{base_url}/api/generate", json=payload, timeout=120) as response:
+                    async with session.post(f"{base_url}/api/generate", json=payload, timeout=300) as response:
                         if response.status != 200:
                             err_txt = await response.text()
                             return web.json_response({"status": "error", "message": f"Ollama HTTP {response.status}: {err_txt}"})
                         resp_json = await response.json()
-                        generated_text = (resp_json.get("response") or "").strip()
+                        generated_text = _extract_ollama_generated_text(resp_json)
+
+                    # Some Ollama model variants return an empty `response` from `/api/generate`
+                    # even though the same request succeeds via the chat endpoint.
+                    if not generated_text:
+                        chat_payload = {
+                            "model": model_name,
+                            "messages": [{
+                                "role": "user",
+                                "content": _ANALYZE_PROMPT,
+                                "images": cleaned_b64_list,
+                            }],
+                            "stream": False,
+                            "keep_alive": 0,
+                        }
+                        async with session.post(f"{base_url}/api/chat", json=chat_payload, timeout=300) as response:
+                            if response.status == 200:
+                                resp_json = await response.json()
+                                generated_text = _extract_ollama_generated_text(resp_json)
                 else:
                     # OpenAI-compatible vision chat (LM Studio / Custom).
                     content = [{"type": "text", "text": _ANALYZE_PROMPT}]
@@ -530,6 +574,11 @@ async def analyze_character_endpoint(request):
 
         if "<think>" in generated_text:
             generated_text = generated_text.split("</think>")[-1].strip()
+        if not generated_text:
+            return web.json_response({
+                "status": "error",
+                "message": f"{provider} returned an empty analysis response.",
+            })
 
         log.info("[LTXDirector] Analysis complete: %s", generated_text)
         return web.json_response({"status": "success", "description": generated_text})