| name | self-compiler |
| description | Rewrite or tune the agent CLAUDE.md + rules.md + USER.md based on accumulated LEARNINGS.md and operator feedback. Invoke when the operator says "recompile yourself", "update your rules", or when LEARNINGS.md has grown beyond ~30 entries. |
self-compiler
Read LEARNINGS.md + USER.md + rules.md, merge the highest-frequency lessons
into rules.md, and propose an updated CLAUDE.md. This is a stub -- full
classifier-backed self-tuning ships on Day-N.
When to use
core/LEARNINGS.md has 30+ entries.- Operator asks to "update your rules" or "recompile yourself".
- The agent notices recurring corrections on the same topic.
What to do (v2.2.0 stub)
- Read
core/LEARNINGS.md and group by first word / tag.
- Promote any group that appears 3+ times into a one-line rule in
core/rules.md.
- Show the operator the diff and wait for confirmation before writing.
- Never silently edit
CLAUDE.md -- always propose, never apply.
Not yet implemented (Day-N)
- Scored promotion pipeline (episodes.jsonl + learnings-engine).
- Automatic rule lint / regression testing.
- Cross-session pattern detection via Firebase.
