learning-looplisted
Install: claude install-skill sumitake/agent-collab
# Learning loop — capture, consult, recur, promote
Maintain a `.learnings/` ledger inside the user's project so lessons survive
session boundaries and tool restarts. The loop has four verbs: **capture** a
lesson when you learn it, **consult** the ledger before re-deriving a known
failure, **recur** when a known pattern strikes again on a new task, and
**promote** proven patterns into the project's standing docs through your own
review process. The ledger is file-based, offline, and deterministic — no
network, no daemon, no automatic context injection.
## The bundled CLI
Resolve the **plugin root** from this loaded file: `SKILL.md` is at
`<plugin-root>/skills/learning-loop/SKILL.md`. The deterministic tool is
`<plugin-root>/learning_ledger.py` (stdlib-only, Python 3.10+):
```text
python3 "<plugin-root>/learning_ledger.py" add --type learning --area <topic> \
--priority medium --pattern-key some.lower.snake.key \
[--class mechanical|judgmental] [--model M --effort E] [--task <ref>]
python3 "<plugin-root>/learning_ledger.py" suggest --error "<message>"
python3 "<plugin-root>/learning_ledger.py" recur <pattern_key> --task <ref>
python3 "<plugin-root>/learning_ledger.py" index
python3 "<plugin-root>/learning_ledger.py" check
python3 "<plugin-root>/learning_ledger.py" lint
```
All commands accept `--root <dir>` (default `./.learnings`). `add` and
`recur` write one new file each; `index` regenerates `INDEX.md`
deterministically; `lint`/`check`/`suggest` are read-only.