michael-denyer/pstack-claude

Claude Code, Codex, OpenCode, Gemini, and Prime Agent versions of Poteto's pstack. Rigorous agent workflows with Cursor primitives translated to other agents.

pstack‑claude – a multi‑runtime agent‑skill plugin

What it is

  • A port of the pstack plugin (originally written for Cursor’s Claude Code) that bundles 54 reusable Agent Skills.
  • The skills are written for Claude Code but are also consumable by the Codex, Prime Agent, opencode, and Gemini‑CLI runtimes via a shared ~/.agents/skills/ directory.
  • The repository contains the plugin manifest, the full skill tree, generated command stubs for Codex, and the tooling needed to keep the skill set in sync with the upstream Cursor version.

Why it matters

  • Provides a ready‑made “poteto‑mode” workflow that injects a short mandate at session start, then offers a suite of higher‑level engineering helpers (TDD, code‑review, PR‑automation, CI fixing, etc.).
  • By exposing the same skill definitions to several AI‑coding agents, a user can switch between Claude Code, Codex, Prime Agent, opencode or Gemini‑CLI without rewriting prompts.
  • All code is MIT‑licensed and includes a thorough audit (CHANGES.md, NOTICE.md) of the adaptations made from the original Cursor plugin.

Installation

  • Claude Code – add the marketplace entry and install the plugin:
    /plugin marketplace add michael-denyer/pstack-claude
    /plugin install pstack@pstack-claude
    
    The plugin auto‑fires on session start (injecting the poteto‑mode mandate). The hook can be disabled by removing hooks/hooks.json.
  • Shared skill install – clone the repo and symlink each skill into ~/.agents/skills/:
    git clone https://github.com/michael-denyer/pstack-claude
    cd pstack-claude
    mkdir -p ~/.agents/skills
    for s in plugins/pstack/skills/*/; do ln -s "$PWD/$s" ~/.agents/skills/$(basename "$s"); done
    
    This single tree is then discovered by Codex, Prime Agent, opencode and Gemini‑CLI.
  • Codex‑only – after the shared install, link the generated slash‑command stubs into ~/.codex/prompts/ and enable multi‑agent support in ~/.codex/config.toml.
  • Gemini‑CLI / opencode / Prime Agent – the same symlinked skill directory works; you only need to reload the runtime’s skill cache.

Key capabilities (selected skills)

Skill Purpose
poteto-mode Sets a short “mandate” that routes any non‑trivial engineering request through a structured workflow before the first model response.
tdd Test‑driven‑development helper that scaffolds tests, runs them, and suggests fixes.
fix-ci Analyzes CI failures and proposes patches.
fix-merge-conflicts Detects and resolves merge conflicts automatically.
make-pr-easy-to-review Generates concise PR summaries and checklists.
get-pr-comments Pulls existing PR discussion into the session for context.
what-did-i-get-done Summarizes recent work for a quick status update.
deslop Performs code‑style linting and cleanup.
thermo-nuclear-code-quality-review Deep static‑analysis pass that surfaces hidden bugs.

Runtime‑specific notes

  • Claude Code – the plugin ships a SessionStart hook that auto‑fires the poteto‑mode mandate. The hook lives in plugins/pstack/hooks/ and can be removed if you prefer manual activation.
  • Codex – no hook system, so you must invoke pstack:poteto-mode explicitly or add a permanent instruction in ~/.codex/AGENTS.md. Slash‑command shortcuts (/tdd, /fix-ci, …) are provided via generated markdown prompts.
  • Prime Agent – discovers skills from the same directory; it ignores the user-invocable: false flag, so internal “principle‑*” leaves stay hidden in the UI but are still readable by poteto-mode.
  • opencode – treats all 54 skills as public; the principle‑* leaves appear in the picker but are harmless.
  • Gemini‑CLI – loads the skills via ~/.gemini/skills/ (or the compatibility alias). Use the activate_skill tool to run poteto-mode or any other skill.

Development & CI

  • CI pipelines (ci.yml, security.yml) regenerate generated files, lint code (Bun, shell, Markdown), run skill‑collision tests, and scan dependencies with osv‑scanner and zizmor.
  • The repo includes helper scripts (tools/generate.mjs, tools/validate-skills.mjs, tools/sync.mjs) that keep the skill tree in lock‑step with the upstream Cursor version (e8d856f).
  • A VERSION file is the single source of truth for the plugin version; manifests are stamped from it during generation.

Dependencies

  • No runtime‑level dependencies are declared in plugin.json. The only optional companion plugin is plugin‑dev (from the official Claude‑plugins marketplace) which improves skill‑authoring workflows.
  • System‑level tools required by some playbooks: gh (GitHub CLI), bun (for bundled scripts), jq and rg (for audit scripts). They are optional unless you run the corresponding playbooks.

License

  • The core plugin and all bundled skills are released under the MIT license (see LICENSE and LICENSE‑cursor‑team‑kit). Attribution details are in NOTICE.md.

Bottom line pstack‑claude is a well‑documented, CI‑tested, multi‑runtime plugin that brings a rich set of engineering‑automation skills to Claude Code and its compatible agents. If you use any of the supported AI coding assistants, installing this repo gives you a ready‑made toolbox for code quality, PR handling, CI debugging, and more, all without having to write the underlying prompts yourself.

Related

  • Project
  • Project
  • Project
  • Project