adlc-distill

name	adlc-distill
description	Run ADLC P7 distillation in the GitHub Copilot CLI — lesson-foundry, rejection-mining, and review-calibration — turning repeated review findings and PR rejections into permanent, deterministic defenses. Decay-driven maintenance lives in the adlc-maintain skill.

ADLC Distill

P7 is where the lifecycle compounds: repeated findings become deterministic defenses (lint checks, skills, spec-gap templates) so the same mistake cannot recur. This is idle-time work — run it after a batch of reviews, or on a schedule.

Both gates here are LLM-backed, and inside the GitHub Copilot CLI you are the model — use --prompt-only, answer the printed prompt yourself, and apply the result. No API keys. Prerequisite: adlc --version works (else npm i -g @adlc/cli).

Distillation

adlc lesson-foundry --prompt-only
adlc rejection-mining --prompt-only
adlc review-calibration --review-cmd "npx adversarial-review --base {base}" --json

• lesson-foundry mines repeated findings in .adlc/findings.jsonl into deterministic defenses. If it prints (no clusters to refine), there are not yet enough repeated findings to distill — report that and stop. Otherwise answer each cluster's prompt yourself, deciding the cheapest deterministic defense (a lint rule, a skill, or a spec-gap template) that would have caught the whole cluster. After the user approves, scaffold with adlc lesson-foundry --write --out-dir .adlc/lessons (the writer is dry-run by default), then edit the scaffolded files to match the wording you decided — --write alone does not apply your prompt-only refinement.
• For any defense that is a skill (a SKILL.md), validate it before PR: hand the scaffolded stub to skill-mining (npx skills add voodootikigod/skill-mining) for dedup against the public ecosystem and a Gate B red-team. Only PR a SHIP-verdict skill; hold un-validated stubs for human review rather than landing them.
• rejection-mining reads recent PR review rejections via the gh CLI and turns each repeated human objection into a reusable review lens. If it errors with a gh/auth/repo message, the repo is not GitHub-linked or gh is not authenticated — note that this gate was skipped and why, then continue. Materialize with --write only after approval.
• Record a no-op manifest entry when there is nothing to distill so the runner can distinguish "checked and empty" from "skipped."

Close the loop — the lesson-foundry gate must be green

Distillation is done only when no recurring finding cluster is left undefended. After writing and refining the lessons above (only materialized lessons count), run the gate as the closing check:

adlc lesson-foundry --gate

It must exit 0. A non-zero exit names each cluster in .adlc/findings.jsonl that recurs but still has no banked lesson — write or refine its lesson and re-run until the gate is green. In an advisory or headless run where nothing was materialized, expect the gate to stay red: report the named unbanked clusters as the outstanding P7 work rather than treating red as complete.

Summarize

Report how many finding clusters and rejection lenses were found, the concrete defenses proposed, which were written (if any), and which gates were skipped (e.g. rejection-mining when gh is unavailable) so the coverage stays honest. For any skill defense, report its skill-mining verdict or flag it as held for human review. Point the user at /adlc-maintain for the decay-driven maintenance checks (skill-rot, model-ratchet, ticket-prune, gate-fuzzing).

Scheduling

This skill is idle-time metabolism. Headless runs are advisory by default: the write steps above require human approval, so an unattended scheduled Copilot session will propose defenses in its summary without materializing them — auto-writing lint rules or skills from clustered findings unattended is risky. The skill-mining handoff is likewise interactive only; a headless run must never npx skills add or auto-validate skills.