Problem

The nightly reflector task was failing with JSON parse errors. Investigation showed that message content containing <thinking> tags (from extended reasoning models) was being passed directly to the LLM. The LLM's output would then contain confused/mixed XML-like structures, breaking the JSON parser.

Solution

Strip <thinking>, <notepad>, and <thought> tags from message content before including it in the LLM prompt — matching what the web UI already does for normal responses.

Changes

• src/scheduler.rs: Apply strip_reflector_thinking_tags() to message content in reflect_for_chat()

Testing

• Reflector task fires on schedule ✓
• JSON parse errors in logs cease ✓