convez.ai
SkillDesenvolvimento WebIntermediário
youhaowei

WyStack Agent Kit

youhaowei/wystack-agent-kit

Ciclo de desenvolvimento completo (plan → build → review → ship) com humano no controle

Ver no GitHub
InicialPluginScore · de 0 a 100

Kit completo de workflow para agentes de código que gerencia o ciclo de vida inteiro do desenvolvimento — desde brainstorming até deploy — em uma única abordagem integrada. Funciona em Claude Code, Codex, Cursor e Pi, mantendo o agente no trabalho técnico enquanto você toma as decisões críticas.

Instalação

$npx convez-ai youhaowei/wystack-agent-kit

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
0
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 3 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Automação de desenvolvimento full-stack com supervisão humana
  • Revisão de código e verificação de requisitos em paralelo
  • Execução paralela de múltiplas tarefas com rastreabilidade de decisões

Veredito

Ferramenta sólida que trata agentes como coautores reais do código, não assistentes superficiais. O foco em auditoria, rastreamento de decisões e supervisão humana obrigatória a diferencia — especialmente útil para equipes que precisam de responsabilidade clara no ciclo de desenvolvimento.

README

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

WyStack Agent Kit

The agent carries the work. You carry the decisions.

npm License: MIT Harnesses

WyStack Agent Kit takes a coding agent through the whole job — planning a feature, building it, reviewing it, shipping it — as one connected method rather than scattered commands. One rule holds across all of it: the agent does the work, you make the decisions.

It's a synthesis of the best of the agent-skills scene, built on four pillars:

  • Full lifecycle — brainstorm → PRD → spec → breakdown → build → review → verify → ship, as one continuous method. Not isolated tricks.
  • Human in the loop — the agent does the work and surfaces the calls that are yours. You decide scope, trade-offs, and what ships.
  • Auditable by design — decisions, review verdicts, and run evidence become durable project state, not lost chat history. Every claim of success carries proof you can review.
  • Autonomous when you want it — scale to parallel, multi-ticket execution with orchestrate; you still gate what merges. Supervised by default, autonomous by choice.

Runs wherever you do — Claude Code, Codex, Cursor, Pi.

Install — 30 seconds

Claude Code

/plugin marketplace add github.com/youhaowei/wystack-agent-kit
/plugin install wystack-agent-kit@wystack-agent-kit

Reload the session after install. Other harnesses (Codex, Cursor, Pi) → jump to install.

The method, made visible

The kit isn't 35 loose skills — it's one software lifecycle an agent drives end to end. At every stage the split is the same: the agent does the work; you make the decision.

brainstorm → prd → spec → breakdown → start-task → code-review → verify → finish-task
StageThe agent doesYou decide
brainstormInterviews, pressure-tests the idea, proposes approachesWhich direction is right
prd / specDrafts the requirements and the technical designSign-off on scope and shape
breakdownSplits the work into vertical-slice ticketsWhat's in, what's cut
start-taskSets up the worktree, plans the slices, writes the codeApprove the plan; own the scope
code-reviewRuns multi-angle review, flags real findingsWhich findings block the merge
verifyRuns the app, gathers runtime evidenceWhether the behavior is acceptable
finish-taskLands the branch, cleans up, updates the trackerThe go/no-go to ship

That split is enforced, not aspirational. Every skill inherits a constitution: the agent never hands you a checklist of chores, never silently makes a call that's yours, and never claims success it can't show evidence for.

It leaves a trail. Decisions, review verdicts, and run evidence are written as durable run records — typed, queryable project state under .wystack/, not chat history that dies with the session. A later retro or review reads what an earlier run found instead of forgetting it.

It scales when you let it. Running one task is start-task. Running many is orchestrate — a conductor that dispatches parallel agents per ticket and brings the results back for you to gate. Supervised by default; autonomous when you choose.

First run

Lifecycle skills need a project workspace. Run once per repo:

wystack-agent-kit:setup-agent-kit

It creates .wystack/ at the project root:

.wystack/
  workspace.md       # project identity, conventions, domain doc layout
  storage.json       # task/doc provider and status mappings
  tasks/             # local markdown fallback for work items
  docs/              # local markdown fallback for PRDs/specs/notes

Default storage is local markdown — zero external setup. Map the same contract to GitHub Issues, GitLab, Linear, Jira, or Notion by editing .wystack/storage.json. The agent's procedure stays constant; only the underlying tools vary per project.

Start here

Four skills cover the everyday loop. Learn these and you're productive; the rest are there when a task calls for them.

SkillWhen you reach for it
brainstormA rough idea needs shaping into a real design before you build
start-taskTime to do a piece of work — plan, worktree, write the code
code-reviewA diff or branch needs a thorough, multi-angle review
finish-taskWork is done — gate it, land it, clean up, update the tracker

Everything else is depth for when you need it — drafting a PRD, verifying runtime behavior, running parallel work across tickets, bootstrapping a design system, writing product copy. Skills self-describe through their frontmatter description (the trigger an LLM matches against), so the agent reaches for the right one without an index to maintain.

GroupSkills
Planningbrainstorm, prd, spec, breakdown, groom, estimation, reprioritize, next-task, new-task
Executionstart-task, worktree, orchestrate, finish-task, cleanup, handoff
Review + verificationcode-review, full-review, critique, verify, perspective, retro, improve-codebase
Designestablish, frontend, polish
Writing + distributionux-writing, copywriting, discoverability, competitor-analysis
Workspacesetup-agent-kit, workspace, identify-specialists, engineering-context

Plus 10 universal role agents — pm, principal, qa, devops, task-manager, wiki-librarian, designer, ux-writer, copywriter, marketing-specialist — with project-configured domain specialists on top.

Inspired by the best of the scene

The synthesis draws from three sources, all MIT: the methodology discipline of obra/superpowers, the lifecycle drive of garrytan/gstack, and the everyday-real-engineering ethos of mattpocock/skills. The contribution here is unifying them under one charter — the agent carries the work, the human carries the decisions — with auditability and optional autonomy built in.

See UPSTREAM.md for fork lineage and the upstream-tracking process.

Other harnesses

Codex

codex plugin marketplace add github.com/youhaowei/wystack-agent-kit \
  --manifest .agents/plugins/marketplace.json
codex plugin install wystack-agent-kit

Reload Codex to pick up plugin metadata.

Cursor / any local-path host

git clone https://github.com/youhaowei/wystack-agent-kit.git
# point your host's plugin config at the inner plugin directory:
# { "type": "local", "path": "/abs/path/to/wystack-agent-kit/wystack-agent-kit" }

Pi

git clone https://github.com/youhaowei/wystack-agent-kit.git
cd wystack-agent-kit
pi install .

Loads plugin skills, role agents, the wystack_agent subagent tool, and an agent_browser tool. If Pi is already running, /reload. For MCP-based external tools: pi install npm:pi-mcp-adapter.

How it's built

One skill body runs on every harness because behavior and adapter are separated. Skill and agent bodies are portable principles — generic, project-agnostic. Frontmatter carries the per-harness adapter: name, description, top-level model, and a nested delegation.<harness>.* block. Claude Code reads top-level; Codex reads the nested block. Porting to a new harness is that harness reading the same frontmatter — never rewriting the method.

Every skill inherits the shared methodology docs in wystack-agent-kit/docs/:

  • constitution.md — the behavioral charter: the agent carries work; the human owns decisions.
  • communication-contract.md — output shape and interactive checkpoints.
  • testing-philosophy.md, review-loop.md, run-record.md — quality and evidence.
  • storage-contract.md, extension-contract.md, workspace-model.md, doc-model.md — the project-instance and extension model.

Installing this as an agent

If you are an LLM agent reading this README to install the kit into the current project, follow the block matching your host. Each block is self-contained and does not ask the user to make sub-decisions.

Host: Claude Code

/plugin marketplace add github.com/youhaowei/wystack-agent-kit
/plugin install wystack-agent-kit@wystack-agent-kit

Verify: /plugin should list wystack-agent-kit as installed. Tell the user to reload the session. Then, if the project will use lifecycle skills, invoke /wystack-agent-kit:setup-agent-kit. Skip if the user only wants design / writing / review skills.

Host: Codex

codex plugin marketplace add github.com/youhaowei/wystack-agent-kit \
  --manifest .agents/plugins/marketplace.json
codex plugin install wystack-agent-kit

Verify: codex plugin list should include wystack-agent-kit. Tell the user to reload Codex.

Host: Cursor / generic local-path host

  1. Clone outside the project: git clone https://github.com/youhaowei/wystack-agent-kit.git ~/code/wystack-agent-kit
  2. Register the plugin source in the host's plugin config:
    { "type": "local", "path": "~/code/wystack-agent-kit/wystack-agent-kit" }
    
  3. Reload the host.

After install — workspace bootstrap

Run once per project repo (not per agent session):

wystack-agent-kit:setup-agent-kit

This is interactive — surface its prompts to the user, don't auto-answer. Outcome: a .wystack/ directory with workspace.md and storage.json, read on every subsequent skill invocation.

Verifying the install worked

  • wystack-agent-kit:workspace resolves the project workspace without error.
  • Listing skills shows entries prefixed wystack-agent-kit: (e.g. wystack-agent-kit:prd).
  • Spawning an agent (e.g. principal) runs on the intended model — sonnet for routine roles, opus for principal, pm, designer. If it inherits the parent's model, the install is using an older cached version; reload or reinstall.

Contributing

Build, deploy, and Codex-migration workflows live in CONTRIBUTING.md.

License

MIT — see LICENSE. Built on the MIT-licensed work credited above.

Ver o original no GitHub

Bom para

Relacionados