| name | lessons-loop |
| description | Capture a durable lesson every time the human corrects the agent — or a task escalates models — so the same mistake doesn't recur. Use immediately after any user correction and after any model escalation. Also use at session start to load relevant past lessons. |
lessons-loop
Vendored from aesop (github:agentmc15/aesop) registry/skills/lessons-loop at commit 5506617,
with a Copilot-harness routing category added. A prompted-Reflexion pattern: the model forgets
between runs, so lessons live on disk. After a correction, write down the pattern and a rule
that prevents it; review lessons at session start. Iterate until the mistake rate drops.
Steps
- On correction: append to
tasks/lessons.md an entry with: the failure pattern, the
lesson (a rule for next time), and the contexts it applies to.
- On session start: read
tasks/lessons.md; surface the lessons relevant to the current
task.
- Hygiene: deduplicate, keep entries concise, add a date, and prune stale/contradicted
ones so bad lessons don't pollute future behavior.
Routing lessons (Copilot-harness category)
A misroute is a correction too. Record an entry with "applies_to": ["routing"] whenever:
- a task escalated — its pinned model failed the verify command and a higher tier had to
finish it (the execute driver marks these as
lesson-candidate (routing): lines in the
kit's NOTES.md); or
- a tier was grossly overprovisioned (frontier spent on work a mid model does routinely).
State the rule in routing terms — task shape → tier — so the
route agent can apply it at
session start. Tiers and model ids come from data/pricing.copilot.json at run time; never
bake a price or a model ranking into a lesson.
Example entry
{"date": "2026-07-01", "failure_pattern": "pinned a cheap-tier model for a multi-file refactor; verify failed twice and the task escalated to strong", "lesson": "multi-file refactors start at the strong tier", "applies_to": ["routing"]}
Notes
- Keep lessons project-scoped to avoid leaking one project's quirks into another.
- This is the same trick the eval harness uses: every production failure becomes a durable
test/rule.
