convez.ai
SkillDesenvolvimento WebAvançado
steloit

Steloit Squad Skills

steloit/squad-skills

Pipeline de AI-team colaborativo com kanban e code-review inteligente

Ver no GitHub
InicialPluginScore · de 0 a 100

Conjunto de skills para orquestrar um pipeline de AI-team (Planner → Critic → Builder → Shield → Inspector → Ranger) em kanban compartilhado, com review inteligente de PRs. Portável entre Claude Code, Cursor, Codex e 50+ agents via padrão aberto Agent Skills.

Instalação

$npx convez-ai steloit/squad-skills

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
37%
Último commit
há 4 meses
Versão
—
Issues abertas
1
Contribuidores
2

Casos de uso

  • Orquestração de tarefas de desenvolvimento em equipe de IA
  • Code review assistido por múltiplos agentes especializados
  • Planejamento e execução de sprints com IA
  • Documentação automática de projetos (wiki gerado)

Veredito

Solução robusta e bem-arquitetada para equipes que querem delegar orchestração de tarefas a múltiplos agentes IA em paralelo. Requer configuração inicial (token + agent detection), mas compensa com evals rigorosos e suporte multiplataforma via Agent Skills.

README

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

Squad runs your tasks through an AI-team pipeline (Planner → Critic → Builder → Shield → Inspector → Ranger) on a shared kanban board, plus intent-aware PR review skills. Portable Agent Skills (SKILL.md) — works with Claude Code, Codex, Cursor, Gemini CLI, and 50+ other agents.

Install

One command for any agent — the skills CLI auto-detects your installed agents (Claude Code, Codex, Cursor, …) and installs to each:

npx skills add steloit/squad-skills          # the 5 core skills, to detected agents
npx skills add steloit/squad-skills -a codex -g   # one agent, global (user-wide)
npx skills update                            # update installed skills later

You get the 5 core skills: squad (board), squad-init (register a project), squad-run (run the pipeline), squad-refine (refine requirements), squad-explore (scope work).

Maintainers — internal skills

The repo also ships specialized skills (squad-batch-run, squad-gen-wiki, squad-heartbeat, squad-kickstart) flagged metadata.internal: true, so they stay hidden from the install above. To pull them too:

INSTALL_INTERNAL_SKILLS=1 npx skills add steloit/squad-skills

Configure

The skills default to the deployed board, so you only supply the shared token — read tool-agnostically, so set it whichever way suits you:

# Option A — environment variable (works for every agent: Claude Code, Codex, Cursor)
export SQUAD_AUTH_TOKEN='<your-shared-token>'        # add to ~/.zshrc or ~/.profile to persist

# Option B — credential file (mode 600; keeps the secret out of your shell profile)
mkdir -p ~/.squad && printf 'SQUAD_AUTH_TOKEN=%s\n' '<your-shared-token>' > ~/.squad/auth && chmod 600 ~/.squad/auth

SQUAD_BASE_URL is optional — it defaults to the deployed board; set it (env, or ~/.squad/config) only to point at a self-hosted board. Then register a project from its directory:

/squad-init

Commands

CommandWhat it does
/squadView / manage the board
/squad-initRegister the current project
/squad-refineTurn a rough task into a concrete spec
/squad-runRun the AI-team pipeline on a task
/squad-batch-runRolling-wave batch execution across tasks
/squad-exploreCodebase exploration → phased task plan
/squad-gen-wikiSynthesize a project wiki
/squad-heartbeatDetect stagnant tasks

Testing & evals

Two tiers — full detail in EVALS.md:

  • tests/ — deterministic (pytest): the non-LLM logic — selector parsing, grouping/parallel rules, auth resolution, prompt rendering. No keys, runs on every PR.
    pip install -r tests/requirements.txt && python -m pytest tests/ -q
    
  • evals/ — behavioral (DeepEval): runs the real agent headless against a squad-eval board project (N trials/scenario), scored on board-state ground truth + a GEval rubric. Records to a git-versioned history.jsonl and flags regressions vs a trailing baseline (Welch's t-test) + writes an HTML trend dashboard. Uses your Claude Code login — no API key needed locally (+ a board token); on-demand/nightly (costs tokens/credits).
    pip install -r evals/requirements.txt && python evals/run_evals.py --trials 3
    

Known gap

ToolCorrectnessMetric is intentionally not wired in yet. While the skills talk to the board over Bash(curl), every tool call is just "Bash" — there's nothing meaningful to score. It becomes valuable after an MCP migration exposes typed squad_* tools; the harness is structured so it drops in then without rework.

License

MIT — see LICENSE. Derived from prior MIT-licensed work; see NOTICE for attribution.

Ver o original no GitHub

Bom para

Relacionados