Sanitize leaked reasoning in character analysis

ltx_director.py

@@ -1,9 +1,10 @@
-import logging
-import asyncio
-import json
-import base64
-import io as _io
-import math
+import logging
+import asyncio
+import json
+import base64
+import io as _io
+import math
+import re
 
 import numpy as np
 import torch
@@ -482,6 +483,65 @@ def _analysis_text_is_usable(text: str) -> bool:
     return True
 
 
+def _sanitize_analysis_text(text: str) -> str:
+    """Strip prompt-echo / reasoning junk and keep the shortest usable visual description."""
+    text = (text or "").strip()
+    if not text:
+        return ""
+
+    if "<think>" in text:
+        text = text.split("</think>")[-1].strip()
+
+    # Drop markdown emphasis and common prompt-echo scaffolding.
+    lines = []
+    for raw_line in text.splitlines():
+        line = raw_line.strip()
+        if not line:
+            continue
+        lower = line.lower()
+        if lower.startswith("the user wants"):
+            continue
+        if lower.startswith("sentence 1 requirements"):
+            continue
+        if lower.startswith("sentence 2 requirements"):
+            continue
+        if lower.startswith("requirements:"):
+            continue
+        if lower.startswith("let's interpret"):
+            continue
+        if lower.startswith("wait, the prompt says"):
+            continue
+        if lower.startswith("do not "):
+            continue
+        if line.startswith("- ") or line.startswith("* "):
+            continue
+        cleaned = line.replace("**", "").strip()
+        if cleaned:
+            lines.append(cleaned)
+
+    text = " ".join(lines).strip()
+    text = re.sub(r"\s+", " ", text)
+    if not text:
+        return ""
+
+    # If the model echoed instructions and then described the subject, keep the last descriptive sentences.
+    sentences = re.split(r"(?<=[.!?])\s+", text)
+    descriptive = []
+    for sentence in sentences:
+        sentence = sentence.strip()
+        if not sentence:
+            continue
+        lower = sentence.lower()
+        if "prompt says" in lower or "requirements" in lower or "the user wants" in lower:
+            continue
+        descriptive.append(sentence)
+
+    if descriptive:
+        text = " ".join(descriptive[-2:]).strip()
+
+    return text
+
+
 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:
@@ -653,8 +713,7 @@ async def analyze_character_endpoint(request):
                 "message": f"Could not connect to {provider} at {base_url}. Make sure the server is running and reachable.",
             })
 
-        if "<think>" in generated_text:
-            generated_text = generated_text.split("</think>")[-1].strip()
+        generated_text = _sanitize_analysis_text(generated_text)
         if not _analysis_text_is_usable(generated_text):
             return web.json_response({
                 "status": "error",