swarm-pr-review · diff
git:20260814.66999cb to git:20260823.b2733f0
13 added, 4 removed. Audit A to A.
---
name: swarm-pr-review
audience: swarm-plugin
description: >
Use when asked to review a pull request, PR URL, or PR #N with a broad,
read-only review and low false-positive tolerance. This Claude Code adapter's
canonical protocol lives in .opencode and owns comment ingestion,
CI/conflict/staleness intake, parallel explorer lanes, independent reviewer
validation, critic challenge, and the explicit handoff into
swarm-pr-feedback for approved fix work.
---
# Swarm PR Review
Read and follow `../../../.opencode/skills/swarm-pr-review/SKILL.md` as the canonical workflow.
## Claude Code Execution Notes
- A plain Claude Code session has none of the swarm plugin's controller tools:
that is canonical Profile B, not an error. Never report BLOCKED merely
because controller tools are absent — dispatch the review through the
`Agent`/`Task` subagent tool, following the canonical phases, role
boundaries, row contracts, and join barriers.
- `PR_REVIEW` is read-only with respect to the PR branch: fetch refs, inspect metadata,
check out its head after verifying a clean tree; do not fix, commit, push, rebase, or reset.
- Before dispatching explorer lanes, fetch the PR head and verify it with
`git rev-parse --verify <full_pr_head_sha>^0` and `git cat-file -t <full_pr_head_sha>` (must print `commit`);
run `git switch --detach <full_pr_head_sha>`, bind that exact HEAD, and do not use `--track FETCH_HEAD`.
- Ingest every review signal before explorer lanes: PR comments, review
summaries, requested changes, bot findings, CI/check failures,
mergeability/conflicts, stale branch/base drift, PR body claims, linked
issues, and commit messages — each a claim until reviewer validation proves
or disproves it with file:line evidence.
- If the repository defines a PR publication contract in local docs, templates, skills,
or CI, ingest it as an obligation source; do not assume this repo's title/body sections.
- Prefer GitHub connector/MCP tools when available, or `gh`, to inspect PR
metadata, comments, review threads, checks, conflicts, and head SHA.
- Profile B accounting: classify the depth tier (S/M/L); cover all six base
dimensions and evaluate all 11 risk families with lane counts scaled per the
canonical depth-tier table; stamp every subagent prompt and ledger row with
its workflow-lane id and exact-head provenance (the bound `pr_head_sha`);
persist findings and trigger ledgers as files in the session task workspace,
never under `.swarm/`.
- Clean lanes still require a fully populated row: base lanes use
`[CLEAN] | lane | coverage_scope | evidence`; micro and council lanes use
`[CLEAN] | micro_lane | coverage_scope | evidence`. A missing attestation is
a coverage gap — retry it or surface BLOCKED, never a degraded review.
- Use fresh reviewer subagents for candidate validation and fresh critic
subagents after review; settlement composes per item, so a reviewer retry
invalidates only the critic claim for an item whose reviewer row changed.
The Pre-Synthesis Gate checklist is the completion gate.
- Only if this session actually exposes the swarm controller tools
(`dispatch_lanes_async`, `collect_lane_results`, `retrieve_lane_output`),
run Profile A instead: structured modes, incremental polling, full-text
retrieval for every `output_ref`, `complete_pr_workflow` before the final
- response, and an initial base wave of one exact-six batch at tier L or a
- smaller `owned_workflow_lanes`-partitioned batch at tiers S/M (controller
- computes the tier from the bound diff). While active, blocking dispatch
- and direct-Task dispatch are not equivalent — never bypass it.
+ response, and controller-computed depth tiers from the bound diff. With the
+ default `pr_review_resilience` policy enabled, Profile A must stage
+ depth-tier M/L base attempts as a singleton canary batch followed by a
+ fanout batch (`pr_review_wave_stage` / `pr_review_wave_attempt`), carrying
+ forward only unresolved obligations; typed `retry_exhausted` and
+ `circuit_open` outcomes are hard blockers, not prompts to degrade the
+ review. Tier S keeps the legacy single consolidated
+ `owned_workflow_lanes`-partitioned batch because staged resilience does not
+ apply there. If that policy is disabled, Profile A falls back to the legacy
+ non-staged base wave: tier L may dispatch one exact-six batch, while tiers
+ S/M use a smaller `owned_workflow_lanes`-partitioned batch that still covers
+ all six dimensions. While active, blocking dispatch and direct-Task dispatch
+ are not equivalent — never bypass it.
- If actionable findings remain, write the canonical handoff artifact and ask
whether to continue with `swarm-pr-feedback`; carry validated findings
forward with their original IDs and provenance.