convez.ai
SkillAutomaçãoAvançado
michael-denyer

pstack para Claude Code

michael-denyer/pstack-claude

Fluxos de trabalho paralelos para código de qualidade com menos escrita

Ver no GitHub
InicialPluginScore · de 0 a 100

Agent que porta o plugin pstack do Cursor para Claude Code, trazendo fluxos de trabalho rigorosos com 39 skills para escrita de código de alta qualidade. Implementa padrões de paralelização e revisão profunda antes da execução rápida, com slash commands para arquitetura, investigação e refinamento iterativo.

Instalação

$npx convez-ai michael-denyer/pstack-claude

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
4,5 mil
Forks
408
Acompanhando
9
Saúde
42%
Último commit
há 1 dia
Versão
v0.9.138
Issues abertas
12
Contribuidores
1

Casos de uso

  • Revisar e refinar código antes de PR
  • Debugar CI/CD e resolver merge conflicts automaticamente
  • Arquitetar sistemas e validar tipos antes da implementação
  • Executar múltiplas tentativas paralelas e selecionar melhores soluções

Veredito

Ferramenta sólida para equipes que valorizam qualidade sobre velocidade bruta. Bem documentada, com auditoria completa de mudanças e licença MIT — mas exige familiaridade com Claude Code e CLI do GitHub para extrair seu máximo potencial.

README

A tradução para o português deste README ainda está na fila.

pstack for Claude Code

Claude Code port of poteto's pstack plugin. Original by Lauren Tan; ships MIT. Imports seven skills from cursor-team-kit (also MIT): deslop, thermo-nuclear-code-quality-review, make-pr-easy-to-review, fix-ci, fix-merge-conflicts, get-pr-comments, what-did-i-get-done.

if you want to go fast, go deep first. pstack helps you write less, but higher quality code. rigorous agent workflows you can parallelize with confidence.

This is not a verbatim copy. Skill bodies have been edited so every Cursor-specific primitive resolves to its Claude Code equivalent — see Differences from upstream for the full list. The exhaustive per-skill audit lives in CHANGES.md; license attribution lives in NOTICE.md; the upstream README is preserved verbatim at README-UPSTREAM.md.

Install

This repo ships as a Claude Code marketplace containing one plugin (pstack).

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

Layout

.
├── .claude-plugin/marketplace.json   # marketplace manifest (repo root)
├── plugins/pstack/                   # the plugin itself
│   ├── .claude-plugin/plugin.json
│   ├── skills/                       # 39 skills
│   ├── commands/                     # 20 slash command stubs
│   └── agents/poteto-agent.md
├── LICENSE                           # pstack upstream MIT
├── LICENSE-cursor-team-kit           # cursor-team-kit upstream MIT
├── NOTICE.md                         # attribution table
├── CHANGES.md                        # per-skill substitution audit
└── README.md                         # this file

Plugin-internal path references in the docs below (skills/<name>/, commands/<name>.md) are relative to plugins/pstack/.

Dependencies

Declared in plugin.json and auto-resolved on install:

  • plugin-dev (from the claude-plugins-official marketplace) — required. The rewiring routes skill-authoring tasks (in automate-me, reflect, poteto-mode) to the plugin-dev:skill-development skill. If you haven't added the official marketplace yet, /plugin install pstack@pstack-claude will pull it in automatically, provided claude-plugins-official is already added; otherwise:

    /plugin marketplace add anthropics/claude-plugins-official
    

Not declared as deps, but referenced in skill bodies:

  • run, verify, loop — Claude Code CLI built-ins (ship with the binary, always available).
  • gh CLI — system-level requirement of the babysit skill. Install via brew install gh and authenticate with gh auth login.

No third-party plugins. The harsher-critique escape hatch lives in the bundled thermo-nuclear-code-quality-review skill (imported from cursor-team-kit), not in an external plugin.

Slash commands

commanduse it when
/poteto-modedefault entry point for any non-trivial task
/howwalk through how a subsystem works
/whyinvestigate why something was built this way (parallel multi-MCP evidence)
/architectsettle types and module shape before writing code that crosses a function boundary
/arenarun N parallel attempts at the same task and pick the best parts
/interrogatehave four different models try to break a diff
/automate-medraft your own personal -mode skill from recent transcripts
/reflectcapture a long task's lessons as a skill edit
/tddfix a bug by writing the failing test first, then the fix
/typescript-best-practicesground type-system discipline in TypeScript syntax
/figure-it-outdesign a rigorous, auditable playbook for a task no bundled playbook fits
/show-me-your-worklog decisions to a reviewable tsv decision trail
/unslopclean up writing by removing AI tells
/deslopdeslop a diff before commit
/babysitmonitor an open PR, fix CI/comments, keep it merge-ready
/thermo-nuclear-code-quality-reviewextremely strict maintainability audit
/make-pr-easy-to-reviewclean noisy history and improve PR description before review
/fix-cifind failing PR checks, inspect logs, apply focused fixes
/fix-merge-conflictsnon-interactively resolve merge conflicts, validate, finalize
/get-pr-commentsfetch and summarize review comments from the active PR
/what-did-i-get-donesummarize authored commits over a user-chosen period

Subagent

poteto-agent ships unchanged. Spawn from a parent with subagent_type: "poteto-agent".

Differences from upstream

The port is editorial, not mechanical. Anywhere upstream pstack assumed Cursor-specific primitives, this port substitutes the Claude Code equivalent so refs actually resolve. Two prior ports (v1truv1us/ai-eng-system, Evan-Kim2028/agent-fleet) stop at namespacing — they vendor pstack under pstack/ and leave the Cursor refs intact. This port does the content surgery.

What's added

  • skills/babysit/ — Claude Code analog of Cursor's closed-source /babysit built-in. Wraps gh pr view / gh pr checks / gh run view --log-failed plus the loop skill for pacing. Independently authored; workflow informed by Cursor's public /babysit behavior — not a copy of Cursor's implementation.
  • skills/deslop/ — imported verbatim from cursor-team-kit. Cleans AI tells out of diffs before commit.
  • skills/thermo-nuclear-code-quality-review/ — imported verbatim from cursor-team-kit. Used as the harsher-critique escape hatch in arena, interrogate, architect, and how (replaces the Cursor-original cross-vendor bridge).
  • skills/make-pr-easy-to-review/ — imported verbatim from cursor-team-kit. Composes with opening-a-pr and babysit.
  • skills/fix-ci/ — imported verbatim from cursor-team-kit. Narrower CI-fix primitive that babysit can route to.
  • skills/fix-merge-conflicts/ — imported verbatim from cursor-team-kit. Pairs with babysit step 5.
  • skills/get-pr-comments/ — imported verbatim from cursor-team-kit. Primitive for babysit step 4 and reflect.
  • skills/what-did-i-get-done/ — imported verbatim from cursor-team-kit. Commit summary over a chosen period.

What's substituted in skill bodies

Upstream (Cursor)This port (Claude Code)
Task tool, subagent_type: generalPurpose, readonly: false/trueAgent tool, subagent_type: "general-purpose", no readonly flag (subagent_type controls MCP access)
AskQuestion toolAskUserQuestion tool
Cursor's built-in /loopClaude Code's built-in loop skill
Cursor's built-in /babysitbabysit skill bundled in this plugin
Cursor's built-in /create-skillplugin-dev:skill-development skill
cursor-team-kit control-cli (CLI/TUI driver)Claude Code's run skill
cursor-team-kit control-ui (browser/Electron driver)Claude Code's verify skill
Transcripts at ~/.cursor/projects/*/ or agent-transcripts/~/.claude/projects/<encoded-cwd>/*.jsonl (where <encoded-cwd> is the workspace cwd with / → -)
Skill paths .cursor/skills/, ~/.cursor/plugins/.claude/skills/, ~/.claude/plugins/
MCP discovery via Cursor's mcps/ directoryTool list at top of system prompt (mcp__<server>__<name> entries), or .mcp.json, or claude mcp list
Model composer-2.5-fast (Cursor)claude-sonnet-4-6
Model claude-opus-4-X-thinking-xhigh (Cursor UI variant)claude-opus-4-8 (extended thinking configured separately)
Models gpt-5.3-codex-high-fast, gpt-5.5-high-fast (via Cursor)claude-sonnet-4-6, claude-haiku-4-5 (Claude family)

What's lost in translation

Cross-vendor model diversity. arena, interrogate, architect, and how all rely on stress-testing a design against four different model families. Claude Code is single-vendor, so the four-way split collapses to four Claude variants by tier and thinking budget. Instead of bridging to an external CLI for that diversity, the rewiring routes the "harsher pass" to the bundled thermo-nuclear-code-quality-review skill — different style of pressure (strict maintainability rubric), not vendor diversity, but it lives in-plugin with no extra installs.

What's deliberately kept

  • The poteto-agent subagent ID and all references to it.
  • run_in_background: true on Agent calls (Claude Code supports it).
  • /loop, /deslop, /babysit slash references in skill bodies — they all resolve in Claude Code now.
  • The principle/playbook structure and every word of the principles themselves.

Forking note

Editing skill bodies forks this from upstream. Re-syncing to a future pstack release means re-applying the substitution table. The full re-port recipe is in CHANGES.md.

License

MIT. Two upstream LICENSE files are preserved:

Ver o original no GitHub

Bom para

Relacionados