From 6551b4f6c35162b16a672d049fdeac1cafb8f4b8 Mon Sep 17 00:00:00 2001
From: OpenClaw Agent <openclaw-agent@local>
Date: Fri, 10 Jul 2026 10:46:04 +0000
Subject: [PATCH] Add JSON fallback for Ollama character analysis

---
 ltx_director.py | 41 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 41 insertions(+)

diff --git a/ltx_director.py b/ltx_director.py
index 83c192b..4581aee 100644
--- a/ltx_director.py
+++ b/ltx_director.py
@@ -564,6 +564,24 @@ def _prepare_analysis_images(cleaned_b64_list: list[str], max_dim: int = 768, qu
         _compress_analysis_image_b64(b64, max_dim=max_dim, quality=quality)
         for b64 in cleaned_b64_list
     ]
+
+
+def _extract_json_description(text: str) -> str:
+    text = (text or "").strip()
+    if not text:
+        return ""
+    try:
+        start = text.find("{")
+        end = text.rfind("}")
+        if start != -1 and end != -1 and end > start:
+            payload = json.loads(text[start:end + 1])
+            if isinstance(payload, dict):
+                description = payload.get("description")
+                if isinstance(description, str):
+                    return description.strip()
+    except Exception:
+        return ""
+    return ""
 
 
 def _resolve_provider(data):
@@ -683,6 +701,29 @@ async def analyze_character_endpoint(request):
                                             return web.json_response({"status": "error", "message": f"Ollama HTTP {retry_response.status}: {err_txt}"})
                                 else:
                                     return web.json_response({"status": "error", "message": f"Ollama HTTP {response.status}: {err_txt}"})
+                    if not _analysis_text_is_usable(generated_text) and analysis_images:
+                        json_prompt = (
+                            'Return JSON only with this exact shape: '
+                            '{"description":"two concise sentences describing only the visible character\'s appearance"}'
+                        )
+                        single_image = [analysis_images[0]]
+                        json_payload = {
+                            "model": model_name,
+                            "prompt": json_prompt,
+                            "images": single_image,
+                            "stream": False,
+                            "format": "json",
+                            "keep_alive": 0,
+                            "options": {
+                                "temperature": 0.1,
+                                "num_predict": 120,
+                            },
+                        }
+                        async with session.post(f"{base_url}/api/generate", json=json_payload, timeout=300) as response:
+                            if response.status == 200:
+                                resp_json = await response.json()
+                                raw_text = _extract_ollama_generated_text(resp_json)
+                                generated_text = _extract_json_description(raw_text) or raw_text
                 else:
                     # OpenAI-compatible vision chat (LM Studio / Custom).
                     content = [{"type": "text", "text": _ANALYZE_PROMPT}]