pstack

What it is

pstack is a set of skills, agents and hooks that gives a coding agent a method for its work. Lauren Tan wrote it for Cursor, and Michael Denyer’s repository ports it to Claude Code, Codex, Pi, GitHub Copilot, Muse and other agents. You give its entry skill, poteto-mode, a goal, and it picks the workflow that fits. The port tracks upstream and also carries named policy forks.

1. You give poteto-mode a goal
2. It chooses a playbook for the task
3. architect, arena or swarm plan and delegate the work
4. interrogate, tests and measurements review and verify it
5. You get the result and the evidence it ran

Who it suits

Developers who let a coding agent change real code and want it to plan, test and report what it checked, not hand back one unchecked answer. People who already use Lauren Tan’s Cursor skills can keep the same skills in another agent. Skip it for small edits, which the routing hook leaves to the agent on its own.

Where to use it

The README has an install block for each of these agents, and a shared route for the rest.

• Claude Code: a marketplace plugin, with skills called as /pstack:poteto-mode and a hook that routes tasks
• Codex: a native plugin from the repository’s Codex marketplace, with a hook you trust through /hooks
• GitHub Copilot: a plugin for the CLI and the app, with a first-run model setup
• Pi: a package whose extension adds subagent, question and wake-up tools
• Muse: a native plugin bundle that you approve with muse plugins approve pstack
• OpenCode, Gemini CLI and Prime Agent: skills linked into ~/.agents/skills/

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with each agent’s steps from the maker’s README and reference page. For Claude Code, run two lines inside a session.

1. Add the marketplace and install the plugin with the two lines below.
2. Run /pstack:setup-pstack to choose models and a reasoning effort, and to turn routing on or off.
3. Give your agent a task and name poteto-mode in it.

/plugin marketplace add michael-denyer/pstack-claude
/plugin install pstack@pstack-claude

Claude Code leaves auto-update off for third-party marketplaces. To update once by hand, run this in a terminal:

claude plugin update pstack@pstack-claude

The other agents have their own lines in the README:

codex plugin marketplace add michael-denyer/pstack-claude
codex plugin add pstack@pstack-claude

copilot plugin marketplace add michael-denyer/pstack-claude
copilot plugin install pstack@pstack-claude

pi install git:github.com/michael-denyer/pstack-claude

git clone https://github.com/michael-denyer/pstack-claude
cd pstack-claude
muse plugins install ./plugins/pstack
muse plugins approve pstack

How to use it

Tell poteto-mode your goal in plain words. A hook points the agent at it for a task that touches more than one file or changes a signature other files call. It does the same for a design choice, a bug with an unknown cause or a performance issue. Smaller tasks go straight to the agent.

Use poteto-mode to fix the search filter resetting when I change pages.

Claude Code calls each public skill /pstack:<name>, Pi calls it /skill:<name>, and Muse lists it as /<name>. In Codex you ask for a skill by name, or install the optional shortcuts for the /name form. These are some of its slash commands, from the reference page.

To turn routing off, run setup-pstack and answer off for the session hook. You can also write this line in the model sheet:

session hook: off

Example

You work in a web app where the search filter resets whenever you change pages, and pstack is installed in Claude Code. You type:

Use poteto-mode to fix the search filter resetting when I change pages.

Its first reply restates your goal and the smallest change it expects. For a bug, the README says it reproduces the failure, uses how and why to investigate, delegates the fix, then reruns the failing case. If the fix crosses a function boundary, it brings in architect first. You get the fix and the failing and passing evidence.

For newcomers and experts

If you are new. A skill is a Markdown file of instructions that your agent reads. pstack bundles a large set of them, plus agents and scripts, so you do not learn each one. Install it, name poteto-mode in a real task, and read what it reports back. Run setup-pstack once to pick models.

For experts. Each role has a default model, and a line in the pstack-models.md sheet overrides it. A value can carry an effort, as in opus @xhigh. Panels such as arena runners and interrogate reviewers take a list, one subagent per entry, and the reference page advises distinct models. Tools and model names differ by agent, so Codex, Pi, Copilot and Muse each have a mapping file.

A required independent review does not fall back to your own check. If no reviewer can run, poteto-mode keeps the review gate closed and records it. A repository can add playbooks under .agents/playbooks/, and a script flags any that extends a missing playbook or quotes step text the new version dropped.

For a property tests cannot reach, create-formal-verification-skill picks the riskiest one and leaves a project skill. The README points to a separate plugin, agent-formal-verify, for guided TLA+ model checking, Lean proofs and Kani for Rust.

Good to know

Most of pstack is text your agent reads, and a few parts run on your machine.

• Data: the README says pstack has no server or telemetry. Anything its skills ask your agent to read, including session transcripts, goes to your model provider
• Scripts: they run locally, and the pull request tools use your GitHub CLI login
• Autonomy: poteto-mode lets reversible work and external actions such as team chat and ticket updates proceed without asking. It pauses for force-pushes to shared branches, deploys, data deletion and customer messages
• Hooks: a SessionStart hook runs a script at session start. Codex and Muse run it only after you trust or approve it. On Windows, Codex and Muse run a PowerShell adapter with -ExecutionPolicy Bypass, which a machine or user execution policy set by Group Policy overrides. Then the hook fails and no instruction loads
• Parallel models: arena and interrogate run several models at once. Pi warns that Anthropic bills Claude used through Pi per token, as extra usage, even on a Claude subscription, and every subagent adds to that bill
• Untested paths: Gemini CLI and Prime Agent were not tried in a live session
• Copilot: the README names the CLI versions it was tested on. If a later version keeps only one plugin’s context, setup-pstack offers a standing instruction instead
• Removal: the README and the reference page give no uninstall steps for Claude Code, Codex, Copilot or Pi. Muse has muse plugins remove pstack, and linked skills are removed by deleting their links
• License: MIT. The LICENSE names Lauren Tan and Michael Denyer, and the imported cursor-team-kit and thermos skills are Cursor’s