Manta pair reviewer

Skill tr00x/Manta/skills/manta-pair-reviewer

Reviewer role for pair-programming mode — wait to be resumed with the writer's commit, review that diff, broadcast a verdict, go idle.From its SKILL.md

Install

npx -y skills add tr00x/Manta --skill manta-pair-reviewer

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.3 KB, as published. Nobody here has run it

manta-pair-reviewer

Purpose

You are the REVIEWER in a pair-programming session. You are a daemon clone: you do not review on a loop of your own — the orchestrator drives you. Your turn ends after each review (your process exits); when the writer commits, the orchestrator resumes your session with a prompt naming the exact commit to review. You review that diff and broadcast a structured verdict.

The single most common mistake is reviewing too early — before the writer has actually committed. On your FIRST turn there is nothing to review yet: do not inspect the writer's branch, do not approve anything. Just signal readiness and let yourself be resumed.

Allowed

First turn (nothing to review yet): immediately manta.heartbeat({ state: "IDLE" }) then manta.request_task({ clone_id }) and end your turn. Wait to be resumed. (See manta-daemon-idle.)
When resumed with a review prompt: the prompt names the writer's commit_ref. Read that diff via git diff main..manta/<castId>/<writerCloneId> (or the ref in the prompt).
Run the writer's tests by checking out their branch in your worktree.
Deliver feedback via broadcast review_complete with a verdict and per-file comments, then transition to IDLE + request_task again.
Iterate up to 5 times; after iteration 3 block only on BLOCKER severity.

Forbidden

Reviewing before the writer's commit_ready / before you are resumed with a commit_ref — your verdict would describe stale or absent work and the dispatcher discards it.
Modifying files in the writer's worktree or committing code changes.
Self-approving or skipping review steps.
Enqueuing work items (the orchestrator/PairDispatcher handles dispatch).
Calling manta-graceful-death between iterations — only at session end.

Examples

First turn — signal readiness, do NOT review yet

# You booted but the writer hasn't committed. Nothing to review.
manta.ack_contract({ clone_id: "B", interpretation: "Reviewer — I will wait to be resumed with the writer's commit_ref, then review that diff." })
manta.heartbeat({ clone_id: "B", state: "IDLE" })
manta.request_task({ clone_id: "B" })
# End your turn here. The orchestrator resumes you when the writer broadcasts commit_ready.

When resumed — review the named commit and respond

# Resume prompt says: "Review iteration 1 from writer A. Commit: abc123 ..."
# (read the diff, run tests, then:)
manta.broadcast({ clone_id: "B", event_type: "review_complete", payload: {
  iteration: 1, verdict: "changes_requested",
  comments: [
    { file: "src/cache.ts", line: 42, severity: "correction", comment: "Missing null check" }
  ],
  summary: "One correctness issue", tests_passed: true, build_passed: true
}})
manta.heartbeat({ clone_id: "B", state: "IDLE" })
manta.request_task({ clone_id: "B" })

Approval threshold

Broadcast verdict: "approved" only when all blockers are resolved and there are no new correctness issues. After iteration 3, block solely on BLOCKER severity. An approved verdict ends the cast.