Crew Plugin — Orquestrador de Agentes Especializados
jircdev/crew-plugin
Equipe de agentes especializados com governança de documentação e estimação
Ver no GitHubSistema de 20 papéis de agentes conversacionais reutilizáveis para Claude Code, com foco em entrega orientada por especificação, taxonomia de documentação e disciplina de estimação. Inclui comandos slash, templates de projeto, hooks de sessão e guardrails para imutabilidade de artefatos.
Instalação
npx convez-ai jircdev/crew-plugin1 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
Casos de uso
- Orquestrar múltiplos agentes para arquitetura, estratégia e revisão de código
- Padronizar documentação e rastreabilidade em projetos multi-agente
- Implementar circuito spec-driven com estimação obrigatória e desvios rastreados
Veredito

Ferramenta robusta para times que precisam de governança forte em projetos multi-agente. A estrutura de hooks, documentação obrigatória e estimação diferencia-se de crews genéricos, mas exige disciplina organizacional — não é plug-and-play.
README
A tradução para o português deste README ainda está na fila.
crew plugin
Stack-agnostic catalog of specialized agent roles — your on-demand crew of architects, strategists, and reviewers. Plus reusable code-quality rules and documentation skeleton for new projects.
Designed to grow beyond roles: future versions may bundle MCP-backed lookups, project-memory helpers, and other on-demand specialists.
Bundles:
- 20 subagents (
agents/) — one per role, with frontmatter for direct activation as Claude Code subagents. - 20 slash commands (
commands/) —/crew:prod,/crew:sys,/crew:ux,/crew:da, etc., each spawning the matching subagent. Conversational by default; structured deliverables only on explicit request. - Templates (
templates/) —AGENTS.md(canonical agent context, AGENTS.md open standard; includes rule precedence, role ownership map, interop contract),CLAUDE.md(thin@AGENTS.mdpointer for Claude Code),.cursor/rules/, and the fulldocs/taxonomy:stories/(functional work items),requirements/(technical work items),decisions/(ADRs, state-in-file),proposals/,guides/delivery-circuit.md(the spec-driven circuit + chaining policy),work/(history — evidence, never truth), andDEVIATIONS.md(accepted-deviations registry, binding for agents). Work items embed a mandatory estimation table (milestones, estimated vs. actual hours) to measure each agentic iteration. - Hooks (
hooks/) — two automatic behaviors, loaded with the plugin:SessionStart→ injectsstandards/session-context.md(~25 lines: rule precedence, delivery circuit, estimation discipline, code-quality core, conversation style, consult-don't-defer) into every session. One local file read, no network.PreToolUse(onEdit|Write) → denies edits to existingdocs/work/entries and to stories/requirements withStatus: Closed— they are immutable by standard. The denial message explains the rule; the guard fails open on any internal error so it never blocks legitimate work.
- Session baseline (
standards/session-context.md) — the always-on context the SessionStart hook injects. Suggestive defaults: the project's own rules always win (see Rule precedence intemplates/AGENTS.md). - Bootstrap script (
bin/init-project.sh) — scaffolds the templates into a new project.
Install in a project
# 1) Add this plugin's directory as a marketplace (one-time; replace the path
# with your local clone, or use the repo URL for team distribution)
/plugin marketplace add C:/w/crew-plugin
# 2) Install
/plugin install crew
# 3) Verify
/help # should list /crew:sys, /crew:ux, /crew:da, /crew:fe, /crew:qa, /crew:sec, /crew:infra, etc.
On first session after install, Claude Code will ask you to approve the plugin's two hooks (see Bundles above) — one-time approval.
For team distribution: push this folder to a Git repo and use /plugin marketplace add <repo-url>.
Bootstrap a new project
From the root of the new (empty) project:
bash C:/w/crew-plugin/bin/init-project.sh
This drops in AGENTS.md, CLAUDE.md, the .cursor/rules/ baseline, and the full docs/ taxonomy (stories, requirements, decisions, proposals, guides, work, DEVIATIONS). Existing files are preserved (skipped, not overwritten).
Onboard an EXISTING project
Do not scaffold over a project that already has docs and conventions. The entry point is the audit:
/crew:doc audit this project's docs against the crew standard
The documentation-steward inventories the project against the plugin taxonomy, reports aligned/deviated/missing findings, and you decide per finding: converge (becomes a story/requirement) or keep the deviation. Kept deviations are recorded in docs/DEVIATIONS.md and the precedence resolution is written into the project's root AGENTS.md — binding for all agents, never re-litigated per session. The plugin baseline is suggestive; the project's own rules always win.
Then:
- Edit
AGENTS.md— fill in{PROJECT_NAME}, stack table, folder layout, and commands. - Write
docs/spec.mdwith the project's technical spec. - Pick stack-specific rules from
templates/.cursor/rules-stack-examples/(Tauri+React, Rust, Node sidecar) and copy the relevant ones into.cursor/rules/. - Install the plugin (above) so
/crew:sys,/crew:ux, etc. are available.
Usage
Slash command
/crew:sys add an IPC command for spawning a process with custom env vars
Spawns the system-architect subagent. It reads its role doc, restricts itself to the role's authority, and returns the canonical deliverable (in this case: an architectural spec).
Alias prefix (interpreted by AGENTS.md)
SYS: add an IPC command for spawning a process with custom env vars
UX: redesign the process panel for the "All" view
LEA: where is the workspace detection implemented?
The project's AGENTS.md instructs the main agent to honor the prefix and either spawn the subagent or read the role doc and adopt its constraints.
Composition rules
From agents/ (each role doc):
- One owner per decision.
- Specs before code. Roles can search and edit code, but implement only after the direction converges or the user explicitly asks — never as the first response.
researcheris the exception: strictly read-only. security-compliancemay interrupt any role.researcherreturns findings, never recommendations.- Roles know the full catalog. The "Role relationships" section in each role doc lists typical handoffs and consults; any role may invoke any other when the situation warrants it.
- Consult, don't defer. When a complete answer needs another role's judgment, the agent obtains it now (spawn or lens-adoption) and answers in the same turn — never closes with "review this with ROLE".
- Chaining policy. After a deliverable, the next role chains in-session only if its human owner (per the project's
AGENTS.md§ Role ownership map) is the session user; otherwise the work stops at the artifact and awaits that human's approval. - Estimation discipline. Every story/requirement embeds a milestone estimation table (estimated vs. actual hours), filled before implementation — this is how teams measure the cost of each agentic iteration.
Folder structure
crew-plugin/
├── .claude-plugin/
│ ├── plugin.json
│ └── marketplace.json
├── agents/
│ ├── product-strategist.md
│ ├── functional-analyst.md
│ ├── system-architect.md
│ ├── ... (20 total)
├── commands/
│ ├── prod.md
│ ├── fa.md
│ ├── sys.md
│ ├── ... (20 total)
├── hooks/
│ ├── hooks.json # registers the two hooks below
│ ├── session-start.js # SessionStart: inject standards/session-context.md
│ └── guard-immutable.js # PreToolUse: deny edits to immutable artifacts
├── standards/
│ └── session-context.md # always-on session baseline (suggestive defaults)
├── templates/
│ ├── AGENTS.md # canonical agent context (precedence, ownership map, interop)
│ ├── CLAUDE.md # thin @AGENTS.md pointer
│ ├── .cursor/rules/
│ │ ├── general.mdc
│ │ ├── communication.mdc
│ │ └── code-quality.mdc # universal core (suggestive; project rules win)
│ ├── .cursor/rules-stack-examples/
│ │ ├── code-quality-tauri-react.mdc
│ │ ├── frontend-tauri-react.mdc
│ │ ├── rust-core.mdc
│ │ └── node-sidecar.mdc
│ └── docs/
│ ├── INDEX.md
│ ├── AGENTS.md
│ ├── MAINTAINING.md
│ ├── DEVIATIONS.md
│ ├── stories/README.md
│ ├── requirements/README.md
│ ├── proposals/README.md
│ ├── guides/delivery-circuit.md
│ ├── decisions/
│ │ ├── README.md
│ │ └── 0000-template.md
│ └── work/README.md
├── bin/
│ └── init-project.sh
├── LICENSE
└── README.md
Updating the plugin
Roles and templates evolve. To propagate changes to consumers:
- Edit the relevant file in
agents/,commands/, ortemplates/. - Bump
versionin.claude-plugin/plugin.json. - Commit and push.
- Consumers run
/plugin update crew.
For template changes, existing projects must re-run bin/init-project.sh (which skips existing files) or merge the new template manually.
Maintenance
- Adding a new role: drop a new
agents/<name>.md(with frontmatter), a newcommands/<alias>.md, and add a row totemplates/AGENTS.mdalias table. - Renaming a role: don't. Aliases are a shared vocabulary; renaming breaks all downstream projects.
- Stack-specific rule: add to
templates/.cursor/rules-stack-examples/, never to the universal.cursor/rules/set.
License
MIT.