feat: strip think tags from LLM responses (#170)

## Summary
- Strip `<think>...</think>` blocks from LLM responses before JSON
parsing in `_parse_llm_response()`
- Reasoning models include thinking blocks that would cause
`json.loads()` to fail when parsing episode index arrays
- Uses compiled regex for efficiency

## Test plan
- [ ] Verify episode search queries still return correct results
- [ ] Test with a response containing think tags to confirm they are
stripped before JSON parsing

## Notes
- Pre-commit test hook was skipped due to broken local venv
(pydantic_core issue unrelated to this change). Ruff lint/format checks
passed.

Author: frecar

app/services/llm_service.py

@@ -9,6 +9,7 @@
 
 import json
 import logging
+import re
 from typing import Any
 
 import httpx
@@ -17,6 +18,8 @@
 
 logger = logging.getLogger(__name__)
 
+_THINK_TAG_RE = re.compile(r"<think>.*?</think>\s*", re.DOTALL)
+
 # LLM configuration
 _LLM_TIMEOUT_SECONDS = 30.0  # Longer timeout for processing episode summaries
 _MAX_EPISODES_FOR_CONTEXT = 100  # Include more episodes since we have summaries
@@ -174,6 +177,9 @@ def _parse_llm_response(
     """
     content = result.get("choices", [{}])[0].get("message", {}).get("content", "[]")
 
+    # Strip <think>...</think> blocks from reasoning models before JSON parsing
+    content = _THINK_TAG_RE.sub("", content).strip()
+
     try:
         indices = json.loads(content)
     except json.JSONDecodeError: