AFK — Pipeline de Codificação Agentica para Claude Code
alexanderop/afk
Pipeline agentico com memória: spec → implementação → QA → ship com julgamento humano nos limites
Ver no GitHubAgent que orquestra um fluxo de desenvolvimento completo: especificação → slices verticais → loops TDD → refatoração → QA agentica → revisão multi-agente. Mantém um vault de memória (brain/) persistente que armazena princípios de engenharia, decisões arquiteturais e aprendizados, injetando-os em cada skill para decisões fundamentadas.
Instalação
npx convez-ai alexanderop/afk1 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
- Orquestração completa de features do planejamento ao merge verificado
- Paralelização inteligente de trabalho independente com /afk:batch
- Captura e reutilização de princípios de projeto através do brain vault
- QA agentica diferenciada por tipo de aplicação (frontend/backend/hybrid)
Veredito

Excelente para times que querem automação de processo sem sacrificar julgamento humano. O vault de memória é inovador — transforma conhecimento tácito em decisões estruturadas. Requer prática para extrair máximo valor das skills individuais versus orquestração full.
README
A tradução para o português deste README ainda está na fila.
afk — a simple coding flow for Claude Code
One help router, one optional end-to-end orchestrator, focused workflow skills,
a packaged implementation agent pair, and a persistent brain/ memory vault the
flow reads before acting and writes back to as it learns. Just the steps that
matter:
/afk:help Inspects the current repo state and recommends the next AFK
skill to run, with a short explanation.
/afk:ship Runs the AFK loop to evidence: plans when needed, implements,
simplifies when useful, QA-checks behavior, and ends with a
ship/no-ship verdict.
/afk:write-good-goal
Turns a vague objective into a concrete /goal condition with
verification evidence, constraints, and optional stop bounds.
/afk:prototype Builds a throwaway logic TUI or UI variant route to answer
design uncertainty before committed planning or implementation.
/afk:grill AI interviews YOU about the plan, one question at a time,
challenging it against your domain glossary and ADRs.
Output: docs/plans/<slug>.md
/afk:implement The lead classifies the work. Simple edits stay local;
complex plans route through an Opus read-only orchestrator
and Sonnet implementation workers, then every diff is
reviewed.
/afk:batch The fan-out alternative to implement: splits an
independently-mergeable plan into many units and runs one
parallel worktree worker per unit, each opening its own PR.
/afk:simplify 4 cleanup agents in parallel review the diff for reuse,
simplification, efficiency, and altitude issues — then the
fixes get applied. Quality only, no bug hunting.
/afk:qa Routes by project shape: dogfood-style browser QA for
frontend apps, contract-level API/service QA for backend
apps, and both for hybrids. Output: evidence-backed report.
Each skill works standalone. Run the focused skills in sequence for a feature,
grab one on its own, or use /afk:ship when you want the loop driven to a
verified verdict.
Memory: the brain vault
AFK keeps a persistent brain/ vault — an Obsidian-compatible markdown store of
your project's engineering principles, codebase gotchas, and decisions. Two
hooks run the plumbing: a SessionStart hook injects brain/index.md so every
session knows what's there, and a PostToolUse hook rebuilds the index when
brain/ files change. The flow is wired to use it: /afk:grill, the implement
orchestrator, and /afk:qa read the brain's principles before acting, and
/afk:ship calls /afk:reflect to persist learnings afterward.
/afk:init-brain Scaffold the brain/ vault in a project (optional — the vault
is also created on demand the first time something writes to it).
/afk:brain Read or write the vault directly.
/afk:reflect Capture this session's learnings into the brain. Most brain
content comes from here.
/afk:ruminate Mine past Claude Code conversations for patterns reflect missed.
/afk:meditate Audit and prune the vault; distill cross-cutting principles.
/afk:plan Break a medium-to-large task into phased, principle-grounded
plans under brain/plans/. Planning only.
/afk:review Principle-grounded review of code or plans, ending in a verdict.
Review only — no changes.
The brain skills and hooks are derived from brainmaxxing by Lauren Tan (MIT) and rewritten to AFK's skill conventions.
AFK also ships two Claude Code subagents for implementation:
afk:implement-orchestrator: read-only Opus planner for complex contracts, slice boundaries, and worker delegation.afk:implementation-worker: Sonnet worker for one bounded TDD slice with edit and verification tools.
Install
Inside any Claude Code session:
/plugin marketplace add alexanderop/afk
/plugin install afk@afk
Verify with /help — the skills appear under the afk: namespace.
Update later with /plugin marketplace update afk.
To try it without installing:
git clone https://github.com/alexanderop/afk
claude --plugin-dir ./afk
/afk:qa needs Vercel's
agent-browser CLI installed and
available on your PATH when it routes to frontend browser QA.
If the skills don't auto-trigger
AFK works by letting Claude pick a skill from its description. If a skill
doesn't fire on its own (some Claude Code versions don't auto-discover plugin
skills, and crowded installs can drop less-used ones from the listing), invoke
it directly: /afk:help, /afk:prototype, /afk:grill, /afk:implement,
/afk:batch, /afk:simplify, /afk:qa, /afk:ship,
/afk:write-good-goal, and the memory skills /afk:brain, /afk:init-brain,
/afk:reflect, /afk:ruminate,
/afk:meditate, /afk:plan, /afk:review. Run /doctor to check plugin
loading, and raise skillListingBudgetFraction in settings to keep more skill
descriptions listed.
What lands in your repo
CONTEXT.md # domain glossary, grown by /afk:grill
docs/adr/NNNN-*.md # decisions worth recording, offered sparingly by grill
docs/plans/<slug>.md # the agreed plan, input to /afk:implement
docs/prototypes/ # prototype verdict notes for throwaway exploration
qa/ # QA reports + screenshots (gitignored)
brain/ # persistent memory vault: principles, gotchas, plans
Testing the plugin
bun run test # zero-token unit + integration checks
bun run test:unit # file-level markdown/manifest checks
bun run test:integration # cross-file plugin structure checks
bun run test:e2e # one cheap headless turn (~$0.01): plugin actually loads
bun run test:evals # model-backed behavioral evals via claude -p
The testing approach is split into unit, integration, and end-to-end checks in
docs/testing-strategy.md. Static lint also validates
behavioral eval specs under tests/e2e/evals/specs/.
bun run test:e2e and bun run test:evals require Claude Code
non-interactive auth, such as ANTHROPIC_API_KEY in CI or a local Claude Code
login that works with claude -p.
Credits
- The grilling interview prompt is based on Matt Pocock's "interview me about the plan" technique, extended with domain-glossary and ADR awareness.
- Skill structure and red-flags tables borrowed from obra/superpowers.
- The
brain/memory vault — its skills (brain,init-brain,reflect,ruminate,meditate,plan,review) and hooks — is derived from brainmaxxing by Lauren Tan (MIT), rewritten to AFK's skill conventions.