| name | source-command-pipeline-retro |
| description | L6 — learning loop. Turn recurring drift/debt/patterns from recent work into concrete governance (rules + ADRs). |
source-command-pipeline-retro
Use this skill when the user asks to run the migrated source command retro.
Command Template
🔁 Retro (learning loop)
Look back at recent work and convert repeated friction into durable improvements
to how this project is built. Output proposals — apply only with the user's OK.
-
Gather signal:
node contextkit/tools/scripts/stats.mjs --json (drift rate, cadence).node contextkit/tools/scripts/tech-debt-scan.mjs --json (recurring smells).- The last ~10 sessions as a compact digest [ADR-0027]:
node contextkit/tools/scripts/session-digest.mjs --last 10 (corrections the
user repeated, conventions that emerged, decisions made informally) — open a
full log only when a digest flags something to inspect.
git log since the last few sessions (what actually changed).- Advisor outcomes [ADR-0032]:
node contextkit/tools/scripts/advise-review.mjs
— per-lane hit-rate of /advise findings (acted-on vs still-open). A lane with
many open + low hit-rate is a pattern: either the owner is noisy (tune it down)
or the lane is under-served (schedule the work).
-
Find patterns, not one-offs: the same correction 3×, a debt category that
keeps growing, a rule that's implied but unwritten, a decision never recorded.
-
Propose concrete governance:
- AGENTS.md edits — a new immutable rule or convention (keep it lean).
This overlaps with
/distill-sessions; reuse it for the AGENTS.md diff.
Deprecate by strikethrough, never delete [ADR-0065]: when a rule is
superseded, mark the old line ~~old rule~~ — deprecated: <reason>. Use <new>.
instead of removing it, so the why of the change stays in history. Standards
are living, not a graveyard — only prune a struck-through rule much later,
once it's clearly fossilized.
- New ADRs — for decisions that were made but never written down.
- Config tweaks — e.g. add a path to
l5.highRiskPaths or
qa.criticalPaths that keeps breaking.
- Habit nudges — e.g. "drift rate 40% → register sessions".
-
Present the proposals ranked by impact. On approval, apply via /distill-apply
(AGENTS.md + ADR in one commit) and /context-config (config), then /log-session.
The point: the platform should get smarter about this project over time, not
just enforce static rules.
