feat(kennel): strip personality emote lines before storing drawers

[SrividyaGanapathi]

Summary

• Adds _strip_noise() to recorder.py that removes Code Puppy personality emote lines (e.g. *wags tail excitedly*, *zooms around happily*) before a response is stored as a kennel drawer
• Adds PUPPY_KENNEL_STRIP_NOISE env var to config.py (default on) so users can opt out without code changes
• Adds tests/plugins/test_puppy_kennel_recorder.py with 13 correctness tests and 3 quantitative evals

Problem

Agent responses were stored verbatim, including personality emotes. These got packed back into the system prompt each session, consuming token budget with zero memory signal.

How it works

The sanitizer processes responses line-by-line before add_drawer is called:

• Strips lines matching ^\s*\*(?!\*)[^*\n]+\*\s*$ (the RPG emote pattern)
• Skips lines inside code fences — content between ``` is never touched
• Leaves **bold**, * bullets, and inline *italic* mid-sentence untouched
• Collapses blank lines left behind by stripped emotes

Evals (measured on a representative mixed response)

Metric	Result
Char reduction per drawer	23.5%
Tokens saved per drawer	~42 tokens
Packer: drawers in 1500-token budget	8 → 11 (+37%)
Latency at 500 chars	0.014 ms
Latency at 5,000 chars	0.118 ms
Latency at 32,000 chars (max)	0.737 ms

Test plan

• _strip_noise() strips emote lines
• Bold, bullets, inline italic, code fences all preserved
• All-emote response → blank → record_run_end no-ops (no drawer written)
• PUPPY_KENNEL_STRIP_NOISE=0 disables filtering end-to-end
• Token savings eval asserts ≥15% reduction
• Packer eval asserts signal drawers surface in recall block
• Latency eval asserts <5ms at all sizes (actual: <1ms)

uv run pytest tests/plugins/test_puppy_kennel_recorder.py -v
# 18 passed

🤖 Generated with Claude Code