Restore MSR character analyze config and parsing

ltx_director.py

@@ -445,6 +445,27 @@ _ANALYZE_PROMPT = (
 )
 
 
+def _extract_ollama_generated_text(resp_json: dict) -> str:
+    if not isinstance(resp_json, dict):
+        return ""
+
+    candidates = [resp_json.get("response"), resp_json.get("content"), resp_json.get("thinking")]
+    message = resp_json.get("message")
+    if isinstance(message, dict):
+        candidates.extend([
+            message.get("content"),
+            message.get("reasoning_content"),
+            message.get("thinking"),
+        ])
+
+    for candidate in candidates:
+        if isinstance(candidate, str):
+            text = candidate.strip()
+            if text:
+                return text
+    return ""
+
+
 def _compress_analysis_image_b64(b64_payload: str, max_dim: int = 768, quality: int = 82) -> str:
     """Shrink analysis images so multimodal providers do not burn their full context on pixels."""
     try:
@@ -531,12 +552,12 @@ async def analyze_character_endpoint(request):
                                         retry_err = await retry_response.text()
                                         return web.json_response({"status": "error", "message": f"Ollama HTTP {retry_response.status}: {retry_err}"})
                                     resp_json = await retry_response.json()
-                                    generated_text = (resp_json.get("response") or "").strip()
+                                    generated_text = _extract_ollama_generated_text(resp_json)
                             else:
                                 return web.json_response({"status": "error", "message": f"Ollama HTTP {response.status}: {err_txt}"})
                         else:
                             resp_json = await response.json()
-                            generated_text = (resp_json.get("response") or "").strip()
+                            generated_text = _extract_ollama_generated_text(resp_json)
                 else:
                     # OpenAI-compatible vision chat (LM Studio / Custom).
                     content = [{"type": "text", "text": _ANALYZE_PROMPT}]