fulcra-agent-review

Implements a review handshake with deterministic verdict tallying for multi-agent teams.

10|1|Updated May 21, 2026
One-click install
npx skills add https://github.com/ashfulcra/fulcra-tools --skill fulcra-agent-review-ashfulcra
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: fulcra-agent-review
Source: https://github.com/ashfulcra/fulcra-tools/tree/main/skills/fulcra-agent-review
Command: npx skills add https://github.com/ashfulcra/fulcra-tools --skill fulcra-agent-review-ashfulcra

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Coordinating code or document reviews across multiple AI agents is fragile: review requests sent as chat messages get acked and forgotten, verdicts get tallied by hand and drift, and merges proceed without the required sign-offs. This Skill adds a durable review handshake to a fulcra-agent-teams space where the request document itself is the obligation and the APPROVED/CHANGES/PENDING state is computed deterministically. ## Core Features & Use Cases - Review requests as durable obligations: coord-engine review request writes a review doc that persists in every required reviewer's needs-me until their verdict file exists, surviving sessions and compaction. - Deterministic verdict tally: coord-engine review status folds verdict files into APPROVED, CHANGES, or PENDING with required-reviewer gating, never trusting cached markers and failing closed on unreadable state. - Head-keyed PR rounds: a stable pr-N slug tracks exact commit heads across pushes, advancing rounds idempotently while keeping prior verdicts append-only. - Use Case: An agent opens a pull request, requests review from the reviewer and security roles, and the merge is gated until both file approve verdicts; a changes verdict blocks until that same reviewer re-uploads an approval. ## Quick Start Ask the agent to request a review of your pull request by running coord-engine review request with the team name, a pr-N slug, the PR URL, the exact commit head, and the required reviewer roles.

Frequently Asked Questions about fulcra-agent-review

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I request a review in a fulcra-agent-teams space?

Run coord-engine review request with the team, a slug, the artifact URL via --of, and required reviewer roles via --reviewer. For pull requests, use a stable pr-N slug and pass the exact commit SHA with --head. Never hand-write the request doc or notify reviewers with a bare tell.

How does a reviewer leave a verdict on a review?

Write a verdict file at the exact path echoed by review request, named verdicts/<head>--<required-token>.md, with verdict: approve or changes in the frontmatter. Then verify with review status that you left pending_required, and only then ack the review-request directive from your inbox.

How is the APPROVED, CHANGES, or PENDING review state computed?

coord-engine review status deterministically folds verdict files: CHANGES if any reviewer requested changes, APPROVED if there is at least one approval, no outstanding changes, and all required reviewers approved, and PENDING otherwise. Never tally verdicts by hand.

What happens when a pull request gets a new commit head?

Re-run review request with the same slug, artifact, requester, and required set but the new exact --head. The review doc advances to the next round, prior verdicts remain append-only, and only verdicts matching the active head count in the tally.

Why does review status exit with code 1?

Exit code 1 means the review doc or verdicts could not be read, so the tally is unknown and retryable, or the slug is a tombstone from an archived review, which is terminal. Treat rc 1 as transport down and retry, never as a settled state.

Does a changes verdict clear when a fix is pushed?

No. The handshake is fail-closed: a changes verdict keeps blocking until that same reviewer re-uploads their verdict file as approve. Pushing new commits does not clear it; the requirement must be explicitly re-affirmed.