1984-agent

1984-agent

Agent skills for team workflows. Currently: mr-review: a single, team-wide standard for reviewing and testing GitLab merge requests and GitHub pull requests.

mr-review

/mr-review [MR] [standard|hard|ultra] [--focus <area>]

No more re-explaining the context for every review. The command picks the MR, the depth and the review areas, runs the project's existing build and tests, and prints one report in a fixed format. It posts comments to the MR only after you confirm.

Every finding states why it is a problem, what it rests on (a rule, an analogous implementation, a spec, a test or a docs link), why it has its severity (🔴 blocker · 🟡 should fix · ⚪ nit), and 1–3 ways to fix it with tradeoffs.

Impact sets the colour, proof can only lower it: 🔴 means security, data loss, a broken spec rule, a broken core flow or a broken build/deploy — and shown by a reproduction, a failing check or the spec quote next to the breaking line. A reproducible cosmetic glitch stays ⚪.

Before reporting, reviewers check each finding against the code around the diff and the project's own docs (AGENTS.md, CLAUDE.md, CONTRIBUTING.md, MR templates, REVIEW.md → Context), always read from the target branch. Findings about a library or runtime API cite the docs of the version the project pins.

The intent role checks that the MR does what its task or spec asks — each requirement quoted next to the code — and nothing it did not ask for.

Small diffs on risky paths (CI, Dockerfile and compose, deploy scripts, migrations, auth/, review rules, REVIEW.md → Risky paths) get at least hard. An MR that touches two or more key user journeys or domains (bounded contexts), listed by the team in REVIEW.md → Journeys / Domains, gets ultra proposed, even when it is small and touches no risky path.

Guarantees

• Your working tree is never touched. The MR code runs in a temporary git worktree, which is removed afterwards.
• Only failures caused by the MR are blockers. A test that also fails on the target branch is reported as pre-existing. A failure from a missing DB, secret or network is reported as "not verified".
• MRs from forks: the commands are shown before any of the fork's code runs, and nothing runs until you confirm.
• Nothing is posted without an explicit answer. You choose all / finding numbers / no. Re-runs do not duplicate earlier comments.

Requirements

• git, plus gh (GitHub) and/or glab (GitLab), logged in.
• hard and ultra need an agent with subagents: Claude Code, Codex, Cursor, … standard works everywhere.

Install

Claude Code (plugin):

/plugin marketplace add tru3ty/1984-agent
/plugin install 1984-agent@1984-agent

Any agent supporting Agent Skills: copy skills/mr-review into the agent's skills folder, e.g. ~/.claude/skills/mr-review (Claude Code) or ~/.agents/skills/mr-review (Codex and other agents).

Team rules: REVIEW.md

Put a REVIEW.md in the repository root to give every review the same context: stack, conventions, extra blocker rules, docs to check findings against (Context), paths that always need a deeper review (Risky paths), key user journeys and domains that raise the depth (Journeys, Domains), team and library docs (Docs), build/test commands, ignored paths and the language of MR comments. Start from the template.

Development

bun install
bun run test                                  # script unit tests + skill structure checks
sh evals/setup-scenario.sh E2 evals/.work/E2   # behavioural eval fixture, see evals/scenarios.md

License

MIT