| name | swarm-pr-review |
| audience | swarm-plugin |
| description | Codex adapter for deep PR review in opencode-swarm. Use when the user wants a broad, read-only PR review with low false-positive tolerance. The 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.
Codex Execution Notes
PR_REVIEW is read-only with respect to the PR branch. You may fetch refs,
inspect metadata, and check out the PR head after verifying a clean working
tree, but do not fix code, resolve conflicts, commit, push, rebase, or reset
from this mode.- 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.
- Treat every ingested signal as a claim until reviewer validation proves or
disproves it with file:line evidence or explicit counter-evidence.
- Prefer GitHub connector tools when available, or
gh, to inspect PR metadata,
comments, review threads, checks, conflicts, and head SHA.
- Use the canonical deterministic lane flow:
dispatch_lanes_async plus
incremental collect_lane_results polling (without wait) to process
settled lanes while continuing independent work; fall back to wait: true
only when no independent work remains, and to blocking dispatch_lanes
only when async collection is unavailable. All lanes must be settled
before synthesis or phase transitions.
- If lane tools cannot close required coverage after retry/re-collection,
Task-tool dispatch is the final fallback, but only as a verified equivalent:
same agent type, same prompt, same scope, same isolation. If equivalence cannot
be proven, stop and surface the lane failure as BLOCKED; do not produce a
degraded review or partial verdict.
- When lane results include
output_ref, call retrieve_lane_output for
full text, then parse_lane_candidates to extract structured candidates
for reviewer dispatch; degraded or incomplete outputs are coverage gaps.
- If actionable findings remain, write the handoff artifact described by the
canonical skill and ask the user whether to continue with
swarm-pr-feedback.
Do not improvise a fix path from review mode. If the user approves follow-up
work, switch to swarm-pr-feedback and carry validated findings forward with
their original IDs and provenance.
