convez.ai
AgenteAutomaçãoAvançado
cuibuaa

FlowCrew

cuibuaa/flow-crew

Orquestração de agentes com planejamento DAG, gates de QA e memória de campanha

Ver no GitHub
BomPluginScore · de 0 a 100

Agent orquestrador que transforma tarefas em fluxos supervisionados de planejamento, codificação paralela e gates de QA. Integra Claude Code para planejamento conversacional e Codex para execução, mantendo memória de conhecimento através de tentativas e campanhas.

Instalação

$npx convez-ai cuibuaa/flow-crew

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
9
Forks
1
Acompanhando
1
Saúde
57%
Último commit
há 1 dia
Versão
v0.11.0
Issues abertas
0
Contribuidores
2

Casos de uso

  • Automação de fluxos de trabalho complexos que exigem múltiplas tentativas e verificações
  • Execução supervisionada de tarefas de engenharia de software com retry loops
  • Construção de inteligência de campanha a partir de históricos de falhas e insights
  • Prototipagem de sistemas multi-agente com observabilidade e determinismo

Veredito

FlowCrew é uma solução robusta para orquestração de agentes complexos, trazendo abstrações sólidas como DAG, gates de QA e memória de conhecimento. Exige domínio técnico significativo, mas oferece controle fino sobre fluxos de trabalho de longa duração com supervisão determinística.

README

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

FlowCrew

flowcrew quick "fix the checkout race condition and prove it with a regression test"

FlowCrew turns a task brief into a supervised team of agents: planner, coder, researcher, reviewer, QA, and supervisor. It is built for work that needs retries, evidence, long-running execution, and memory across attempts.

It is not just a prompt runner. FlowCrew gives your AI workflow a run state, a dashboard, deterministic reality checks, and a knowledge graph of decisions, findings, results, and dead ends.

Highlights

HighlightWhat It Gives You
Claude Code planning -> Codex executionDiscuss the plan conversationally, then use /ship to dispatch the confirmed work to FlowCrew's Codex-default backend.
Planner-generated DAGThe planner turns a brief into explicit stages with dependencies, gates, and retries.
QA gate retry loopFailed gates trigger targeted fix stages and re-checks instead of asking you to babysit the run.
Backend-driven supervisorA configurable observer reads stage output and run state, then guides, aborts, replans, or detects completion.
Run MemoryGoals, approaches, findings, insights, results, sources, and dead ends stay attached to each run, and roll up into a ranked per-campaign knowledge digest.
Campaign intelligenceRelated runs share metrics and failure history so future attempts can pivot instead of repeat.
Reality-GateDeterministic checks block fabricated or unsupported terminal success.
Dashboard + CLI + skillsUse the interface that fits the moment: /ship, terminal, or browser dashboard.

The Recommended Loop

FlowCrew works best as the execution layer behind your interactive coding agent:

1. Discuss scope, constraints, and acceptance criteria in Claude Code.
2. Type /ship.
3. FlowCrew creates a task brief and dispatches it to Codex by default.
4. Agents plan, execute, verify, retry, and summarize.
5. You inspect the dashboard, run summary, artifacts, and knowledge graph.

Why this pairing works: Claude Code is strong for collaborative plan shaping, while Codex is the default FlowCrew execution backend and is often the better fit for long multi-agent sub-runs when your Codex subscription has more generous execution budget.

Install the skills once:

./skills/install.sh

Then ship from the conversation:

> Split auth into token validation and session management.
> Keep the public API compatible and add focused regression tests.
> /ship

From Brief To Verified Run

The important boundary: the supervisor steers, but it does not edit files or run commands directly. Work happens in worker stages; evidence is checked by gates and Reality-Gate.

Why It Is Different

One-shot agentFlowCrew
Best effort answerAuditable run with state, artifacts, verdicts, and summary
One context windowPersistent run memory and campaign history
Manual retry after failureGate -> fix -> re-gate loop
"Looks done"Deterministic evidence checks before terminal success
Lost rationaleKnowledge graph of decisions and evidence
Single backend assumptionCodex default with Claude/Codex per-role overrides

Architecture: Atoms

FlowCrew is built on self-describing atomic semantics. Every composable primitive — roles, skills, deterministic checks, research policies, terminal/verdict vocabularies — describes itself, is collected into a registry, and is injected into the planner at runtime. The planner composes a run from these atoms, and each atom maps to the roles that execute it.

The invariant that keeps this maintainable: semantics live at the atom's own source, injected at runtime — the planner prompt is a stable composition engine, never a semantics dictionary. Domain-specific semantics live in the brief / project contract, never in the engine. Adding a role, check, or skill needs zero planner-prompt edits, and a project's hard constraints (declared in <project>/.flowcrew/contract.yaml) are wired by the planner into deterministic Reality-Gate checks.

See design/atom-architecture.md for the full rationale, the drift problem it solves, and the roadmap.

Quick Start

npm install
npx flowcrew init
npx flowcrew doctor

Ship directly:

flowcrew quick "fix the failing checkout flow and add a regression test"

Ship from a file or stdin:

flowcrew quick --task "$(cat task.md)" --max-iterations 3 --timeout 600000
echo "audit docs for stale API examples" | flowcrew quick -

Start the dashboard:

npx flowcrew start

Open http://localhost:3000 to inspect live stages, QA verdicts, artifacts, campaign scores, summaries, and run memory.

What You Can Run

Unknown bug hunt

flowcrew quick --campaign checkout-bug "Find the root cause of an intermittent checkout failure.
Acceptance:
1. Add a reproducer test that fails before the fix.
2. Document the root cause in docs/root_cause_checkout.md.
3. Fix the bug and make the reproducer pass 50 consecutive times.
4. Do not repeat any hypothesis already marked dead_end in the campaign."

Research loop

flowcrew quick --campaign model-eval "Improve src/model.py on data/validation.jsonl.
Baseline: accuracy 0.72. Target: accuracy >= 0.85.
Each iteration tries one new approach, records the metric, and stops after target hit or two non-improving rounds."

Review-gated writing

flowcrew quick "Polish docs/design.md until reviewer score is >= 8/10 on clarity, evidence, and reproducibility.
Never invent citations, never change reported numbers, and only rewrite passages the review gate flags."

Run Memory

FlowCrew records why a run made decisions, not just what files changed. Every run captures its goals, approaches, findings, insights, results, cited sources, and dead ends as a knowledge graph.

Across a campaign these roll up into a knowledge digest: the best direction and its result, plus ranked, deduped findings, insights, dead ends, and sources — each linking back to the run that produced it — so the signal is readable at a glance instead of buried in a dense node graph.

The relational graph stays available per run for the full goal -> approach -> finding -> result structure:

Common node types:

TypeMeaning
goalThe objective being pursued
approachStrategy selected by the planner
findingEvidence discovered during work
insightReusable lesson from a stage or iteration
resultMeasured outcome
sourceExternal reference (paper/URL) cited during research
dead_endFailed direction that future runs should avoid
user_hintHuman guidance preserved for future stages

Configuration

FlowCrew reads config/defaults.yaml. The current default backend is Codex:

default_timeout_ms: 3600000
default_max_iterations: 5
default_gate_retry_loops: 3

adapter: codex
model: default
reasoning_effort: default

supervisor:
  stuck_threshold_ms: 600000
  # Optional: use Claude for higher-level supervision while stages use Codex.
  # adapter: claude
  # model: claude-opus-4-7
  # reasoning_effort: high

Override a single role when useful:

# config/agents/qa.yaml
adapter: claude
model: claude-opus-4-7
reasoning_effort: xhigh

CLI

flowcrew init
flowcrew quick "task"
flowcrew quick "task" --background
flowcrew status
flowcrew list
flowcrew guide "message"
flowcrew start
flowcrew doctor
flowcrew audit-reality

Documentation

License

MIT

Author

FlowCrew Captain LinkedIn: Profile

Ver o original no GitHub

Bom para

Relacionados