rstack

rstack is @michaelshimeles's agent development stack: seven agent skills and an AGENTS.md workflow that tells a coding agent how to take a change from a clean branch to a reviewed pull request. It is written for Claude Code, Cursor, Codex, and other agents that load SKILL.md folders.

Use it if you want an agent to work in its own worktree, follow your project's architecture, prove each change with evidence, and stop only when Greptile rates the PR 5/5. Copy AGENTS.md into another repo and swap in that repo's checks.

The workflow

AGENTS.md picks the steps that fit each task. Read-only reviews and investigations skip all of them. Small documentation edits get an isolated branch and content checks, with no runtime evidence or bot review.

1. Isolate. /new-feature puts the edit in its own Git worktree and branch from origin/main, after checking open PRs for overlap. Nothing gets built on main.
2. Build. /code-structure applies when extracting shared operations or choosing boundaries. The project's existing architecture wins. The service layer pattern is the default only when there is none.
3. Prove. Run the relevant checks. /evidence-driven-testing records the failure before the fix and the success after it, for any change in runtime behavior.
4. Ship. Open the PR with evidence that fits the change. /before-and-after adds a screenshot table when visuals help. /greploop runs Greptile review cycles for code and executable workflow changes until the current commit scores 5/5 with no unresolved actionable findings, or until it hits the iteration cap, a timeout, or a missing integration. It then reports the real score.

/unslop runs over anything a person will read along the way: commit messages, the PR title and body, docs, and code comments.

AGENTS.md also sets multi-agent rules: one worktree and branch per task and agent, no commits to main, --force-with-lease only on your own branch, and no merge unless someone asks for it.

Skills

Skill notes

• evidence-driven-testing. The recorder, scripts/evidence.py, needs Python 3 and ffmpeg/ffprobe with libx264 and the ass filter. It runs on Linux, macOS, and Windows with doctor, start, annotate, and stop commands, and writes evidence.mp4, report.md, and manifest.json. Run python3 scripts/evidence.py doctor from the skill folder to check your machine. The agent drives the app with computer use, or cua-driver when it has no computer-use tools. The headless helper, scripts/run-playwright.sh, needs Node 20+ and npm and installs a pinned Playwright in a temp directory, outside your project.
• before-and-after. Drives the @vercel/before-and-after CLI. Install it with npm i -g @vercel/before-and-after agent-browser. The default upload host, 0x0.st, is public. Pass --upload-url for anything sensitive.
• greploop. Needs Greptile installed on the repo and an authenticated gh, glab, or p4. The GitHub helper, scripts/github_review.py, needs Python 3.10+. It records each trigger and rejects stale checks and summaries. The cap is --max-iterations, default 10.
• greploop-apps. Has no implementation of its own. Install it next to greploop.
• unslop. The body matches upstream. The frontmatter drops disable-model-invocation: true and scopes the description to text the agent writes or edits for people, so agents apply it without a typed /unslop. Put the flag back if you want slash-command-only behavior.

Install

With npx skills

The skills CLI installs to Claude Code, Cursor, Codex, and most other agents:

npx skills add michaelshimeles/rstack

Useful flags: --list shows the skills without installing, -s <name> picks skills, -a <agent> picks agents (claude-code, cursor, codex), and -g installs at user level instead of project level.

By hand

Clone the repo and copy skill folders into your agent's skills directory. For Claude Code:

git clone https://github.com/michaelshimeles/rstack.git
cd rstack
mkdir -p ~/.claude/skills/
cp -r new-feature code-structure evidence-driven-testing before-and-after greploop greploop-apps unslop ~/.claude/skills/

greploop-apps only works with greploop installed beside it.

The workflow file

The skills work alone, but the workflow lives in AGENTS.md. Copy it into the root of the repo you want the agent to work in, and replace the "Checks in this repo" section with that repo's checks. Cursor and Codex read AGENTS.md from the repo root. Claude Code reads CLAUDE.md, so add a CLAUDE.md that contains @AGENTS.md.

Once installed, agents pick skills by their descriptions. In Claude Code you can also call one directly, for example /new-feature or /greploop.

Repo layout

AGENTS.md                    workflow and multi-agent rules
new-feature/                 SKILL.md
code-structure/              SKILL.md
evidence-driven-testing/     SKILL.md, scripts/evidence.py, scripts/run-playwright.sh, scripts/playwright/
before-and-after/            SKILL.md, scripts/upload-and-copy.sh, scripts/adapters/, LICENSE
greploop/                    SKILL.md, scripts/github_review.py, references/, LICENSE
greploop-apps/               SKILL.md, LICENSE
unslop/                      SKILL.md, LICENSE
tests/                       pytest suite for the recorder, review helper, and Playwright runner

Checks

Run python3 -m pytest tests/ -q and git diff --check. The tests need Python 3.10+, pytest, and ffmpeg/ffprobe with libx264 and the ass filter. The Playwright runner tests also need Node and Bash, and they stub npm downloads. If you change that helper, run a real Chromium capture with the documented command too. The tests use synthetic recorder input and local API fixtures, so they don't cover live uploads, every native capture backend, or a live Greptile review.

Adding a skill

1. Create a kebab-case folder named after the skill.
2. Add a SKILL.md with name and description frontmatter. Agents use the description to decide when the skill applies, so lead with the trigger ("Use when...").
3. Keep the instructions short. Move long material into reference files in the same folder.

Sources and licenses

new-feature, code-structure, and evidence-driven-testing were written here. The rest are vendored, each with its license in its folder:

The repo has no top-level license file. Only the vendored skills ship a license.