Agent skill · gstack
/review Catch production bugs in a diff before it merges
This pre-landing review compares the branch with its base, checks for scope drift and missing requirements, then reads the full diff against a checklist. It lists every edit to tests, CI or snapshots for scrutiny, dispatches specialist and adversarial reviewers, and classifies each finding as auto-fix or ask. It never commits or opens PRs; that belongs to ship.
Works with
- Claude Code
- Codex, Cursor, OpenCode, Gemini CLI and other agents
Used forCode review
When to use it
- A feature branch is ready and you want a careful review before opening a PR.
- An agent changed tests or CI settings and you want those edits looked at by name.
- You want review bot comments triaged alongside your own findings.
How it works
- 1Check branch and scope
The agent stops if there is no diff, then compares commit messages, TODOs and the PR body with the changed files.
- 2Read the full diff
It reads the checklist and the complete diff, including untracked source, before commenting on anything.
- 3Scan gate edits
Every change to tests, CI, lint config or snapshots is listed and reviewed, and an unavailable scan counts as missing coverage.
- 4Critical pass and specialists
Only issues that would make tests fail, break production or lose data block the pass, and each one is searched for at sibling call sites.
- 5Fix first
Findings become auto-fix or ask items; auto-fixes are applied and ask items wait for your approval.
What it steers the agent away from
- Comments on issues the diff already fixes.
- Quietly weakened tests, timeouts or suppressions.
- Committing, pushing or opening a PR from the review.
Install this skill
It installs with the gstack pack.
Claude CodeThe README wraps this command in a prompt you paste into Claude Code, which also adds a short gstack section to CLAUDE.md.
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setupClaude Code (team mode for a shared repo)Run inside your repo; it commits a bootstrap so teammates get gstack with auto-updates, and needs Node.js for the required hook.
(cd ~/.claude/skills/gstack && ./setup --team) && ~/.claude/skills/gstack/bin/gstack-team-init required && git add .claude/ CLAUDE.md && git commit -m "require gstack for AI-assisted work"Codex, Cursor, OpenCode, Gemini CLI and other agentsUse --host auto for every detected agent or name one host; non-Claude hosts are listed as experimental.
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/gstack
cd ~/gstack && ./setup --host auto # or: ./setup --host codex
./setup --status # one row per install: host, tier, scope, version, pathCommands come from the author’s README; check it for changes before you run them.
Questions
Will it change my code?
Yes for obvious, mechanical fixes. Critical or judgement calls are asked first, and gate edits such as test changes always need approval.
Does it repeat findings I skipped before?
It suppresses a finding only if you skipped the same one earlier and its file has not changed since.
More guides from gstack
/csoRun a security audit that keeps evidence separate from claimsAudits code, infrastructure, supply chain or skills for exploitable defects, challenges each finding and can reproduce fixes in Docker.Read the guide →/office-hoursPressure-test an idea before your agent writes any codeQuestions your idea in startup or builder mode, challenges premises, compares two or three approaches and writes a design doc, with no code.Read the guide →/plan-ceo-reviewReview a plan for scope and ambition like a founderReviews a plan as a founder would, in expansion, selective, hold or reduction mode, and asks before adding or cutting any scope.Read the guide →/plan-eng-reviewLock architecture, edge cases and tests before you buildLocks a plan's architecture, data flow, edge cases and test coverage, with ASCII diagrams and a review report appended to the plan.Read the guide →/qaFind, reproduce and fix bugs with regression testsTests a site, API, CLI, job or webhook, reproduces each bug, writes a regression test first and commits each verified fix.Read the guide →/shipTake a branch from finished code to an open PRMerges the base branch, runs tests and reviews, bumps VERSION and CHANGELOG, verifies the frozen content, pushes and opens a PR.Read the guide →
Also in this pack
/autoplanRuns the CEO, design, developer experience and engineering reviews in sequence and brings only taste decisions to you.↗/benchmarkMeasures page load timings, Core Web Vitals and resource sizes in a real browser and compares them with a saved baseline.↗/browseDrives a real browser, preferring your signed-in Aside sessions, to open pages, click through flows, read the console and take screenshots.↗/canaryWatches a live URL after a deploy for console errors, failed pages and visual changes against baseline screenshots.↗/carefulWarns before destructive commands such as recursive deletes, dropped tables or force pushes, and hard-blocks the riskiest ones.↗/design-consultationResearches your product space and proposes a full design system, written to a DESIGN.md file with font and colour previews.↗/design-htmlTurns an approved mockup into production HTML and CSS with computed text layout, detecting React, Svelte or Vue.↗/design-reviewAudits a live site for spacing, hierarchy and generated-looking patterns, then fixes them with atomic commits and before and after screenshots.↗/design-shotgunProduces several AI design mockups side by side in a browser board, collects your feedback and iterates until you pick a direction.↗/devex-reviewTests a developer product live by following its docs and getting-started flow, timing each step and comparing against the plan's scores.↗/diagramTurns an English description into a mermaid source, an editable Excalidraw file and rendered SVG and PNG, offline.↗/document-releaseChecks project docs against what changed in a release and updates the facts, asking about subjective edits.↗
Summary by trendhub. The skill itself lives in garrytan/gstack/review by Garry Tan (MIT); read the full, current text there.