adlc-distill

name	adlc-distill
description	Run ADLC P7 distillation and maintenance workflows in Codex, including lesson-foundry, rejection-mining, skill-mining, scheduled maintenance, skill-rot, model-ratchet, review-calibration, and gate-fuzzing.

ADLC_CODEX_SENTINEL_DISTILL_V1

ADLC Distill

P7 converts repeated findings into deterministic defenses and keeps cached guidance fresh.

Distillation

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

lesson-foundry is LLM-backed — use --prompt-only and answer the printed prompt yourself, mining repeated findings into deterministic defenses (lint checks, skills). rejection-mining mines human PR rejections into reusable review lenses (needs the gh CLI). Record a no-op manifest entry when there is nothing to distill so the runner can distinguish "checked and empty" from "skipped."

Maintenance (decay-driven, idle-time work — C10/C12 + calibration)

Assumptions rot over time or after a model/repo change: skill cache metadata goes stale, files churn enough to deserve re-prosecution, and gates that once held may now be defeatable. Run these on a schedule or after a model upgrade, and report a single honest summary including any check that did not apply.

1. Skill rot — stale validation metadata (C10)

adlc skill-rot .agents/skills plugins/adlc-codex/skills --json

• Exit 0: skills are fresh.
• Exit 2: one or more skills have stale validation metadata — list them and recommend re-validating (--write to stamp freshness once re-checked).
• Exit 1 with nothing to verify: the targeted skills carry no validation metadata to check — informational, not a failure.

2. Model ratchet — hot files to re-prosecute (C12)

adlc model-ratchet --dry-run --json

Lists the highest-churn / highest-dependency files, the best candidates to re-prosecute after model or repo drift. This is a plan, not a gate — it never fails. Report the top files and suggest running the multi-lens prosecution fan-out ($adlc-prosecute) against them. With --review-cmd, model-ratchet can run a review over those files and append findings to .adlc/findings.jsonl (which later feeds $adlc-distill's lesson-foundry).

3. Ticket prune — stale ticket hygiene

adlc ticket-prune --json

Dry-run by default (never writes on its own): reports tickets that look already shipped — an explicit status: done-shaped field, or every declared scope glob already resolving to a file tracked on HEAD. A plan, not a gate — exit 0 either way; exit 1 only on an operational error (a missing/invalid ticket store). List stale tickets found and recommend confirming by hand, then adlc ticket-prune --write to archive into the gitignored .adlc/tickets.archive.json (never deletes outright). Treat --write as a human-confirmed action — the ticket store is shared, hand-edited state.

4. Gate fuzzing — can hostile candidates defeat the gates? (calibration)

Only run if a gate suite exists at .adlc/gate-suite.json; without one the tool exits 1 (Gate suite not found) — note that calibration was skipped.

adlc gate-fuzzing --suite .adlc/gate-suite.json --prompt-only

LLM-backed: use --prompt-only and play the adversary yourself against each gate in the suite. A gate you can defeat is a calibration gap — report it; the gate needs strengthening (the gate-fuzzing exit-2 condition when run with a real provider).

Scheduling

The deterministic checks (skill-rot, model-ratchet, ticket-prune) are keyless and run well on a cron — see docs/ci/adlc-maintenance.yml. The LLM-backed gate-fuzzing check runs via a scheduled Codex session invoking this skill, where Codex is the model and no API keys are needed — CI cron is the deterministic fallback for the other three when no scheduled session is configured.

Close the loop — the lesson-foundry gate (P7)

Whether you just distilled findings or ran the maintenance checks above, finish on the ledger's own gate — it must exit 0:

adlc lesson-foundry --gate

A non-zero exit names each cluster in .adlc/findings.jsonl that recurs but still has no banked lesson; each is an undefended cluster. Bank or refine its lesson and re-run until the gate is green. It is deterministic, keyless, and portable: .adlc/findings.jsonl is tracked in git (ADR 0014), so the cluster set travels with the repo and the gate returns the same verdict anywhere it is checked out. It is not yet wired into the maintenance cron.