Validate LLM responses and warn on unexpected formats

[ColonistOne]

Summary

The agent now catches and logs when the LLM ignores the expected response format, instead of silently doing nothing.

What's validated

Vote parsing — warns when the response contains none of UPVOTE, DOWNVOTE, or SKIP:

WARNING LLM response missing vote keyword (UPVOTE/DOWNVOTE/SKIP) for 'Some post': I think this post is really interesting...

Comment extraction — improved parsing:

• Explicit COMMENT: SKIP now correctly returns empty (was falling through)
• SKIP anywhere in response treated as skip intent
• Short ambiguous responses (<20 chars) logged as warning
• Freeform comments (no COMMENT: prefix) logged at debug level

Reply to comments — rejects replies under 10 characters:

WARNING LLM reply to alice too short to post (7 chars): 'Thanks!'

Prevents posting low-effort single-word replies.

Test plan

• Missing vote keyword → warning logged, no vote cast
• Vote-only response → no comment extracted
• SKIP in middle of text → treated as skip
• Explicit COMMENT: SKIP → empty comment
• Short reply rejected with warning
• 163 tests pass, ruff + mypy clean

🤖 Generated with Claude Code