agentharnesses.dev

A field guide to agentic coding harnesses, for teams that have been told to adopt one.


Harnesses > Workflow methodologies

gstack

Draft. This entry has not been checked against its sources yet.

Garry Tan's set of 55 skills, written for Claude Code, that runs a sprint through role-based steps, from product questioning and plan reviews to specialist code review, QA in a real browser, and a /ship command that ends with a pull request.

MaintainerGarry Tan (garrytan)
CategoryWorkflow methodologies
HostsClaude Code, Codex, Cursor, Factory Droid, Kiro, OpenCode
Components55 skills, 11 hooks
LicenseMIT
Version1.91.2.0
Source codehttps://github.com/garrytan/gstack
Websitenone
Docshttps://github.com/garrytan/gstack/blob/main/docs/skills.md
DeepWikihttps://deepwiki.com/garrytan/gstack
Last researched2026-09-28, at commit 01593aa

Features

Definitions are on the compare page.

FeatureSupportedNotes
Stages of work it covers
Brainstormingyes/office-hours asks one question at a time, challenges your premises, and makes you pick one of two or three approaches before it writes a design doc. A hard gate forbids code.
Written planspartialYou draft the plan in the host's plan mode; /plan-ceo-review, /plan-eng-review, and /autoplan review it. Design docs, review reports, and test plans go to ~/.gstack/projects/, not the repository.
Test firstnoTests follow the code: /ship writes missing coverage tests, and /qa and /investigate add a regression test after each fix. The README's 'test-first execution' is not in the ship skill.
Debugging processyes/investigate: 'NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST', hypotheses tested before any fix, a stop after three failed hypotheses, and edits locked to the affected directory.
Code reviewyes/review sends specialist reviewer subagents, chosen by what the diff touches, plus an adversarial pass that goes to the Codex CLI when installed. /ship runs the same review before it pushes.
Verificationyes/ship: 'NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE', checked against a ledger of test runs. /qa re-tests each fix in a browser. An opt-in Stop hook blocks the turn until a declared command passes.
Learning captureyes48 of 55 skills end by logging learnings to ~/.gstack/projects/<slug>/learnings.jsonl, and 13 search it. Kept on your machine unless you turn on sync to a private git repository.
Branch to PRyes/ship: merge the base branch, test, review, bump VERSION, write CHANGELOG, split commits, push, and open the PR (GitHub or GitLab). /land-and-deploy merges and checks production (GitHub only). It does not create branches.
How it runs
Single entry pointyesThe gstack skill sends a request to the right skill. A 'Skill routing' section it offers to add to CLAUDE.md applies the same rules in every session.
Automatic activationyesThe proactive setting is on by default and skills say 'Proactively invoke this skill'. On Claude Code descriptions are cut to one line, so matching leans on the CLAUDE.md routing section.
Session-start hooknoThe SessionStart hook, registered in team mode only, updates gstack in the background and injects nothing. Standing instructions go into CLAUDE.md.
Blocking hooksyes/careful blocks destructive commands, /freeze blocks edits outside one directory, /autoplan blocks skipping a phase, and an opt-in Stop hook holds the turn until a declared check passes.
Subagentsyes
Parallel agentsyesSpecialist reviewers and design variants are launched in one message. Running many sessions at once ('10-15 parallel sprints') relies on the separate Conductor app.
Git worktreespartialOnly /spec --execute creates a worktree, for a claude -p agent it spawns. The README leaves parallel sessions to Conductor workspaces.
Adopting it
Multiple hostsyes./setup installs for six hosts. Claude Code is the source; the others get generated copies without hooks. OpenClaw, Hermes, and Slate get lighter integrations.
Per-project setupyesTeam mode commits a CLAUDE.md section and, in required mode, a project hook that blocks skills until each developer installs gstack. Installs stay per user and pull the latest version; no version pin.
Team extensionspartialConfig keys, custom /careful warning patterns, /plan-tune question preferences, and /skillify browser scripts. No documented way to add workflow skills or change one without forking.
No extra servicesyesNeeds Bun and builds a local Chromium. Reviews use the Codex CLI by default when it is installed and signed in; design mockups need an OpenAI API key. Telemetry is off unless you opt in.

How it works

The workflow

The README calls gstack "a process, not a collection of tools" and orders the skills as a sprint: Think, Plan, Build, Review, Test, Ship, Reflect. Each step saves something the next one reads: "/office-hours writes a design doc that /plan-ceo-review reads. /plan-eng-review writes a test plan that /qa picks up."

  1. Think (/office-hours). It asks what you are building for, then runs one of two modes. Startup mode asks six "forcing questions" about demand, the status quo, and the narrowest first version; builder mode brainstorms. It states premises for you to accept or reject, requires two or three approaches (one minimal, one ideal), and stops until you pick one. Then it writes a design doc to ~/.gstack/projects/<slug>/. A hard gate forbids any code. /spec is the alternative when the result should be a GitHub issue.
  2. Plan, in the host's plan mode. /plan-ceo-review challenges scope in one of four modes (expansion, selective expansion, hold scope, reduction). /plan-eng-review settles architecture, data flow, edge cases, and a test plan. /plan-design-review and /plan-devex-review cover interfaces and developer-facing work. /autoplan runs them in sequence (design and developer-experience only when they apply, eng always last), answers routine questions with six written decision principles, and saves the judgment calls for one approval at the end. Each review appends a ## GSTACK REVIEW REPORT to the plan file.
  3. Build. gstack has no implementation skill. You approve the plan, leave plan mode, and the host writes the code.
  4. Review (/review). It picks specialist reviewers from the diff: testing and maintainability always (for 50 or more changed lines), and security, performance, data migration, API contract, design, and simplification when the diff touches those areas, plus an adversarial reviewer. It fixes mechanical findings itself, asks about the rest, and repeats, up to three rounds, until a pass makes no edits. /investigate handles bugs, /cso security audits, and /design-review visual problems.
  5. Test (/qa <url>). It drives a browser through the app, fixes each bug in its own commit, re-tests it, and adds a regression test. /qa-only reports without fixing.
  6. Ship (/ship). From a feature branch it merges the base branch, runs the tests, audits coverage and writes missing tests, checks the plan's items were done, runs the pre-landing review, bumps VERSION, writes CHANGELOG.md, updates TODOS.md, splits the work into commits, checks the test evidence, pushes, updates docs through a /document-release subagent, and opens or updates the pull request. /land-and-deploy merges, waits for CI and the deploy, and checks production. /canary watches it afterwards.
  7. Reflect (/retro). A weekly retrospective built from git history and gstack's logs. /learn reviews and prunes saved learnings.

48 of the 55 skills share a generated preamble. It checks for updates and sets the question format and writing rules; in 39 skills it also lists recent artifacts and decisions for the current project and branch. These 48 skills end with a status (DONE, DONE_WITH_CONCERNS, BLOCKED, or NEEDS_CONTEXT) and a step that logs learnings.

Components

How it steers the agent

Files it writes

Install

gstack is not a Claude Code plugin. The README says to paste this into Claude Code:

git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup

and to have Claude add a "gstack" section listing the skills to CLAUDE.md. ./setup builds the browser, links each skill into ~/.claude/skills/, and registers the hooks.

For a team, run ./setup --team, then ~/.claude/skills/gstack/bin/gstack-team-init required (or optional) inside the repository, and commit .claude/ and CLAUDE.md. Each developer still installs gstack; team mode then pulls the latest version at the start of each session, at most once an hour.

Other hosts: ./setup --host codex, opencode, cursor, factory (Factory Droid), or kiro, or let ./setup detect what is installed. These hosts get generated copies of the skills without hooks. OpenClaw and Hermes get prompt templates and generated skill files rather than an install; OpenClaw's main path is to start Claude Code sessions that have gstack, and four skills are published for it on ClawHub. Slate reads the Claude Code install. Any agent that reads a rules file can use the 2 KB summary in agents-digest/gstack-AGENTS.md.

Needs Git and Bun, plus Node.js on Windows. The Aside browser is recommended on macOS. /cso needs a native build toolchain. Remove it with bin/gstack-uninstall.

When to choose it

When not to

The first /office-hours session ends with a note from the author inviting you to apply to Y Combinator; later sessions skip it.

Sources