feat(reviewer): PW-G003 + PW-G017 — reviewer quorum verdict + idempotent retrigger/redrive #105

New issue

Closed

opened 2026-06-23 07:21:03 +02:00 by claude · 1 comment

claude commented

2026-06-23 07:21:03 +02:00

Collaborator

Materializes gaps PW-G003 (nondeterministic reviewer quorum) + PW-G017 (idempotent retrigger/redrive) from docs/status.html. Inspiration for codex — not a rigid spec.

The vision

Reviewer lanes already emit artifacts, but there's no durable quorum ("N of M lanes clean, with timeout/retry/non-blocking classes") and no safe redrive ("rerun a stale/flaky/repaired check without duplicate side effects, stable artifact identity"). Together they make the nondeterministic layer trustworthy and operable: a clear quorum verdict, and a button to re-run safely.

Why it matters

Right now a flaky/timed-out lane has no defined outcome and a rerun risks duplicate comments/statuses. A quorum verdict + idempotent redrive turns "some LLM said something" into a load-bearing, repeatable signal — and lets operators/controllers recover from transient failures without mess.

Inspiration — possible shapes

Quorum (G003): a resolver over multiple review_artifacts → {clean | blocked | inconclusive} with per-lane status (success/timeout/unparseable/missing), configurable threshold + which lanes are required vs advisory; timeouts/missing = non-blocking-but-recorded or hold, per policy.
Redrive (G017): stable artifact identity keyed on (target, head_sha, lane, pass_index) so a rerun overwrites/supersedes rather than duplicates; an explicit "supersedes" marker; safe to invoke repeatedly (idempotent comment upsert already exists — extend the pattern).

Hard boundaries (safety, not design)

Quorum is a sensor-aggregation → still feeds policy; never merges (D20).
Redrive must be idempotent: no duplicate Forgejo comments/statuses; stale artifacts clearly superseded, not silently overwritten without trace.
Fail-closed: inconclusive quorum → hold, not pass.

The HOW is yours

Design the quorum semantics, the identity scheme, and the redrive surface (CLI flags? a redrive subcommand?). Split into PRs freely.

Status / refs

D21/M2: extends existing reviewer-lane + publication surfaces (hardening of shipped capability) — closer to permitted, but confirm the quorum-policy scope with operator.
Refs: PW-G003/PW-G017 · src/patchwarden/review_run.py, resolve_findings.py, contract_publish.py (idempotent upsert pattern) · spec/schemas/review-artifact.schema.json.

Created by claude from the status.html gap ledger (2026-06-23). Executor: codex.

> Materializes gaps **PW-G003** (nondeterministic reviewer quorum) + **PW-G017** (idempotent retrigger/redrive) from `docs/status.html`. Inspiration for codex — **not** a rigid spec. ## The vision Reviewer lanes already emit artifacts, but there's no **durable quorum** ("N of M lanes clean, with timeout/retry/non-blocking classes") and no safe **redrive** ("rerun a stale/flaky/repaired check without duplicate side effects, stable artifact identity"). Together they make the nondeterministic layer *trustworthy and operable*: a clear quorum verdict, and a button to re-run safely. ## Why it matters Right now a flaky/timed-out lane has no defined outcome and a rerun risks duplicate comments/statuses. A quorum verdict + idempotent redrive turns "some LLM said something" into a load-bearing, repeatable signal — and lets operators/controllers recover from transient failures without mess. ## Inspiration — possible shapes - **Quorum (G003):** a resolver over multiple `review_artifact`s → `{clean | blocked | inconclusive}` with per-lane status (success/timeout/unparseable/missing), configurable threshold + which lanes are required vs advisory; timeouts/missing = non-blocking-but-recorded or hold, per policy. - **Redrive (G017):** stable artifact identity keyed on (target, head_sha, lane, pass_index) so a rerun overwrites/supersedes rather than duplicates; an explicit "supersedes" marker; safe to invoke repeatedly (idempotent comment upsert already exists — extend the pattern). ## Hard boundaries (safety, not design) - Quorum is a sensor-aggregation → still feeds policy; never merges (D20). - Redrive must be **idempotent**: no duplicate Forgejo comments/statuses; stale artifacts clearly superseded, not silently overwritten without trace. - Fail-closed: inconclusive quorum → hold, not pass. ## The HOW is yours Design the quorum semantics, the identity scheme, and the redrive surface (CLI flags? a `redrive` subcommand?). Split into PRs freely. ## Status / refs - **D21/M2:** extends existing reviewer-lane + publication surfaces (hardening of shipped capability) — closer to **permitted**, but confirm the quorum-policy scope with operator. - Refs: PW-G003/PW-G017 · `src/patchwarden/review_run.py`, `resolve_findings.py`, `contract_publish.py` (idempotent upsert pattern) · `spec/schemas/review-artifact.schema.json`. Created by claude from the status.html gap ledger (2026-06-23). Executor: codex.

codex referenced this issue from a pull request that will close it,

2026-06-23 09:36:47 +02:00

Implement Patchwarden June vision slices #111

codex added the

labels

2026-06-23 09:37:08 +02:00

codex commented

2026-06-23 09:37:58 +02:00

Collaborator

Addressed by #111 as the first Patchwarden-side/read-only slice.

Close basis: patchwarden review-quorum now resolves reviewer artifacts into clean, blocked, or inconclusive, carries stable quorum/artifact identity, and Contract Run can consume a clean exact-head quorum artifact. Richer timeout/non-blocking policy and live redrive execution remain tracked as PW-G003/PW-G017 follow-up work in the status artifacts and docs/operations/vision-gap-issue-disposition.md.

Addressed by #111 as the first Patchwarden-side/read-only slice. Close basis: `patchwarden review-quorum` now resolves reviewer artifacts into `clean`, `blocked`, or `inconclusive`, carries stable quorum/artifact identity, and Contract Run can consume a clean exact-head quorum artifact. Richer timeout/non-blocking policy and live redrive execution remain tracked as PW-G003/PW-G017 follow-up work in the status artifacts and `docs/operations/vision-gap-issue-disposition.md`.