adlc-prosecute

name	adlc-prosecute
description	Run the multi-lens P5 adversarial prosecution fan-out and record ADLC P5 review evidence and P6 acceptance packet workflows in Codex. Fans out five independent read-only lenses, dedupes findings, verifies each independently, and loops until dry before recording evidence.

ADLC_CODEX_SENTINEL_PROSECUTE_V1

ADLC Prosecute

The multi-lens prosecution fan-out (P5)

Codex has no primitive identical to Claude Code's Task-tool subagent fan-out, so this skill instructs the sequence explicitly rather than claiming automatic parallel execution. Run it as a loop:

1. Fan out five independent lenses. Invoke each of the following read-only agent profiles against the change diff, one at a time, collecting each lens's JSON findings array before moving to the next: adlc-prosecutor-correctness, adlc-prosecutor-security, adlc-prosecutor-contract, adlc-prosecutor-diff, adlc-prosecutor-tests (plugins/adlc-codex/agents/adlc-prosecutor-*.toml). Each lens hunts one specific failure class only — do not let one lens's prompt bleed into another's judgment.
2. Dedupe. Combine all five lenses' findings and dedupe with dedupeFindings from lib/prosecutor.mjs (re-exported, by reference, from @adlc/core), keeping the highest-severity instance of each findingKey-identical finding.
3. Verify independently. For each deduped finding, invoke adlc-prosecutor-verifier once, with a fresh context and no memory of other findings' verdicts. Its job is to try to refute the finding, not confirm it — default to refuted on weak evidence. Collect { real, reason, repro } per finding.
4. Decide survival. A finding survives only if survivesVerification (strict majority of valid real: true votes) says so. A finding with no valid vote survives as an unverified blocker (fail closed) rather than being silently dropped.
5. Loop until dry. If any finding survived this round, fix it, then repeat steps 1-4 against the updated diff. Use shouldContinue (maxDry: 2, the default) to decide when to stop: two consecutive rounds with zero surviving findings ends the loop.

This is the same fan-out → dedupe → independent-verify → loop-until-dry shape Claude Code's prosecutor-{correctness,security,contract,diff,tests,verifier} subagents implement — Codex reuses the identical pure orchestration logic (findingKey/dedupeFindings/survivesVerification/shouldContinue) from @adlc/core via plugins/adlc-codex/lib/prosecutor.mjs, driven by a skill-instructed sequential loop instead of a host-level fan-out primitive.

Record every surviving finding (the P5 → P7 bridge)

As soon as a finding survives verification (step 4) — and before you fix it in step 5 — record it. Once you fix it the finding stops existing, and a finding that was never recorded cannot be clustered by lesson-foundry (P7), so the lifecycle stops compounding. Run this exactly once per surviving finding, at verification time (not deferred to the end):

adlc prosecute --record-finding \
  --file <repo-relative path> \
  --desc "<plain prose: the pattern, not this instance>" \
  --category <correctness|security|contract|diff|tests> \
  --severity <high|medium|low>

--file and --desc are required — the recorder fails closed rather than appending a junk entry. Write --desc as plain prose describing the pattern, with no quoted or backticked literals and no identifiers from this diff: --desc is the clustering key, so a description tied to one instance clusters with nothing. This is distinct from the adlc prosecute --input / adlc run p5 evidence in "Recording evidence" below, which records that a prosecution ran; this records what it found — only the second one compounds.

Automatic Stop-time reminder

A native Stop hook checks the session's changed files against the risk-tier categories (auth/trust-boundary, security-control deny-path, secrets, data-loss/destructive, schema-migration, CI/CD supply-chain) and, if any are touched with no recorded adversarial-review gate-manifest entry for the active ticket, surfaces an advisory nudge to run the loop above and record its evidence. This does not replace running the loop yourself — it is a safety net for the case where it was skipped.

Recording evidence

Once the loop above is dry, capture the transcript and record the reviewer-produced evidence with adlc prosecute. The transcript must name the ticket and reviewed git-worktree:<hash> revision that P5 records. Do not pass --revision in normal git worktrees; auto-resolved revisions keep P6 staleness protection active. The P5 input must also include review_packet with prompt path/hash, reviewed-input path/hash, and clean_worktree equal to the reviewed revision.

For scoped P5 evidence:

adlc prosecute --input .adlc/p5-passes.json --ticket <ticket-id> --dir .adlc --json
adlc run p5 --ticket <ticket-id> --dir .adlc --json

P6 strict mode requires P5 evidence:

adlc behavior-diff capture --config behavior.json --out .adlc/before.json
adlc behavior-diff compare .adlc/before.json .adlc/after.json --json
adlc accept --ticket <ticket-id> --packet .adlc/packet.json --before .adlc/before.json --after .adlc/after.json --dir .adlc --json
adlc run p6 --ticket <ticket-id> --dir .adlc --json

The bundled docs fixture is static. Use it only with its fixture revision:

adlc prosecute --input docs/examples/p5-passes.json --ticket T1 --revision docs-example-revision --dir .adlc --json