pstack-claude: Poteto-Style Agent Workflows for Claude Code, Codex and Copilot

pstack-claude is a GitHub repository by Michael Denyer that takes a Cursor skill stack, Lauren Tan's pstack, and translates it into plugins for Claude Code, Codex, GitHub Copilot, Pi, OpenCode, Gemini CLI and Prime Agent. It was trending on GitHub in week 42 of 2026 with about 393 new stars, and the repository page showed roughly 1.9k stars, 204 forks and 244 commits when we read it. The pitch in the repo description: "Rigorous agent workflows with Cursor primitives translated for other harnesses."

This post covers what it contains, how to install it on each harness, what it can and cannot do, and the trust questions that apply to any skill pack you let run inside your editor.

TL;DR: questions people ask first

What is in the stack

The center of the repo is poteto-mode, a router. You tell the agent your goal in plain language and the router matches it to a playbook. According to the README, the playbooks cover planning, building features, refactoring, performance work, investigations, prototypes, pull-request maintenance, shipping, and longer projects.

Around it sit supporting skills with short names: how, why, architect, arena, swarm, interrogate and unslop. From their names you can guess the intent: explain how something works, trace why it is that way, design before coding, pit approaches against each other, fan out parallel agents, question the plan, and clean up machine-sounding output. We have not run each skill, so read the individual skill files before relying on a specific behavior.

A setup skill, setup-pstack, lets you choose models and reasoning effort per role and turn automatic routing off. That last option matters: a router that silently changes how your agent behaves is useful only if you can switch it off.

There is also a companion plugin, agent-formal-verify, which adds TLA+ model checking and Lean proofs for concurrency bugs that ordinary tests cannot reach. That one is aimed at a narrow audience, but it shows the project's bias toward rigor over speed.

How it differs from a plain CLAUDE.md or skills folder

If you already steer your agent with a CLAUDE.md, skills, hooks and subagents, pstack-claude is the same ingredients packaged as a single installable unit with a router on top. The advantages are consistency across tools and a ready-made set of procedures. The cost is that you inherit someone else's opinions about how a refactor or an investigation should proceed.

Compare with approaches we have covered before:

• Hand-written rules files give you full control and no surprises, but you carry the maintenance.
• A registry pack such as the ones described in our guide to installing skills with npx gives you single skills without a router.
• Automated skill improvement, like Microsoft's SkillOpt, tries to tune skills from feedback rather than from a fixed stack.
• Whole-harness comparisons, such as our agent harness roundup, show where plugins like this fit against the harness itself.

How to install it

The README gives one path per harness. The commands below are copied from it and may change.

# Claude Code
/plugin marketplace add michael-denyer/pstack-claude
/plugin install pstack@pstack-claude

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

For Codex you add the marketplace and install the plugin from the terminal, then approve the routing hook through /hooks before it will run. For GitHub Copilot you add the marketplace from the Copilot CLI and start a new session afterward. OpenCode, Gemini CLI, Prime Agent and skills-only setups follow a shared installation document in the repo.

After install on a harness with no default models, such as Copilot, the first skill that needs a model triggers setup-pstack. Plan a few minutes to choose models, and decide up front whether you want automatic routing on.

Limitations the README states

The author is candid about several constraints, which we repeat because they decide whether this fits your setup:

1. Copilot version range. The Copilot build was tested only on Copilot CLI 1.0.87 through 1.0.92. Later versions may drop the routing context, and on 1.0.92 non-interactive copilot -p sessions may not list plugin skills.
2. Manual trust step on Codex. The routing hook needs your explicit approval. That is the right design; do not approve hooks you have not read.
3. No default models on Copilot. You configure them yourself.
4. Data handling. No server and no telemetry, but session transcripts and anything the skills read go to your model provider. Pull-request tools use your GitHub CLI login, so they act with your permissions.

Security and trust: what to check before you install

A skill pack is instructions plus sometimes scripts and hooks that your agent executes with your credentials. Treat installing one like adding a dependency.

• Read the hook definitions and any shell scripts before approving them.
• Check which tools the skills request. A planning skill should not need your GitHub token; PR-maintenance skills legitimately do.
• Pin to a commit if you need reproducibility, since the repo tracks an upstream and declares policy forks in tools/forks.json.
• Keep permission prompts on for destructive commands. Our guide to Claude Code permission modes and hooks explains the controls.
• The extra license files for imported Cursor and thermos material mean redistribution inside a company needs a quick license check, even though the main license is MIT.

If you run agents with real credentials, a layer that blocks dangerous actions before they execute is worth having; AgentBeam is the agent security platform from the explainx.ai team built for that.

A worked trial you can run in 20 minutes

Plugins are easy to install and hard to judge, so give yourself a controlled test. The plugin mechanism itself is documented in Anthropic's Claude Code plugin docs; use it to see where the files land and how to remove them.

1. Create a scratch branch of a repository you know well, with a failing test or a small tangled function.
2. Run setup-pstack and pick a cheaper model for the exploratory roles and a stronger one for planning. Write down the choices.
3. State a goal such as "refactor this module to remove the duplicated parsing, keep tests green" and let poteto-mode route it. Note which playbook it chose and whether that was the one you would have picked.
4. Reset the branch, turn routing off, and give the same goal to the bare agent.
5. Compare four numbers: wall-clock time, tool calls, tokens, and whether the final diff passes your tests without edits.

If the routed run is clearly better on a task you do weekly, keep it. If it only looks more thorough, but costs more and produces the same diff, install just the supporting skills you actually call. Remember that one session proves little; repeat it on three different task types before you decide, and remove the plugin if you cannot explain what it changed.

Is it worth trying?

It depends on how you work.

Try it if you switch between harnesses and want the same procedures in each, if you like structured playbooks for refactors and investigations, or if you already use Cursor's pstack and want the same habits in Claude Code or Codex.

Skip it if your harness version is outside the tested range, if you prefer a short hand-written rules file, or if you do not want a router deciding which workflow runs. The router can be turned off, so you can also install the supporting skills and call them by name.

A reasonable trial: install on a scratch repository, ask for a small refactor with routing on, then repeat with routing off, and compare the plan quality, the number of tool calls, and the tokens used. Longer, more rigorous procedures usually cost more tokens. Our notes on loop engineering for coding agents cover how to measure that trade-off.

What the trend tells us

Stars alone do not prove quality, and we have not benchmarked pstack-claude against a bare agent. What the week's growth does show is demand for portable agent workflows. The same skill content now runs in Claude Code's plugin system, Codex's marketplace, Copilot CLI and Pi, which suggests plugin formats are converging enough that a single repo can serve several harnesses. For teams, that lowers lock-in: your procedures live in a repository rather than in one vendor's settings. The Pi harness is one example of a smaller tool gaining first-class support in packs like this.

Related reading

• Steering Claude Code with CLAUDE.md, skills, hooks and subagents
• Top 10 closed and open-source agent harnesses
• Claude Code hooks: automate actions on tool calls
• Claude Code permission modes explained
• Install Claude Code skills with npx
• Loop engineering for coding agents

Repository details are as of October 13, 2026 and change as the project updates.