dynos-work
dynos-fit/dynos-work
IA que constrói, testa e corrige código enquanto aprende com seus erros
Ver no GitHubPlugin para Claude Code e Codex que constrói features de forma autônoma, audita seu próprio código, corrige bugs automaticamente e aprende com os erros a cada tarefa. Funciona em pipeline paralelo com aprovação humana nos pontos críticos (spec, plano, revisão).
Instalação
npx convez-ai dynos-fit/dynos-work1 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
- Implementar features completas com spec e plano aprovados
- Auditar e corrigir automaticamente bugs encontrados durante build
- Melhorar qualidade incremental do código a cada nova tarefa
- Acelerar desenvolvimento backend com paralelização de tarefas independentes
Veredito

Plugin maduro que automatiza o ciclo completo de desenvolvimento (spec → build → audit → repair). Excelente para quem quer delegação inteligente mantendo controle; requer Python 3 local e funciona melhor com permissões ajustadas. Curva de aprendizado moderada pela orquestração multi-agente.
README
A tradução para o português deste README ainda está na fila.
dynos-work
A Claude Code and Codex plugin that builds features, catches its own bugs, and gets better at your codebase over time.
Install
Claude Code
claude plugin marketplace add dynos-fit/dynos-work
claude plugin install dynos-work
Codex
This repo also includes a Codex plugin manifest at .codex-plugin/plugin.json. Install it through your Codex plugin workflow from this repository root.
Use
Claude Code
/dynos-work:start build a user settings page
Codex
Invoke the start skill with your task, for example: "Use dynos-work start to build a user settings page."
That's it. You'll answer a few questions about what you want, approve a spec, approve a plan — then it builds.
What happens
When you start a task, the plugin:
- Asks what you need — a few targeted questions to nail down scope
- Writes a spec — exact behavior, edge cases, what's out of scope. You approve it.
- Writes a plan — which files change, in what order, how it'll be tested. You approve it.
- Builds in parallel — multiple agents work on independent pieces simultaneously
- Audits its own work — independent agents check for bugs, security issues, and spec gaps
- Repairs what it finds — automatically, up to 3 attempts before it asks you
After the task, it learns from what went wrong and builds prevention rules for next time. Your 10th task is better than your 1st.
Commands
/dynos-work:start [task] start a new task
/dynos-work:execute run the approved plan
/dynos-work:audit audit, repair, and close
/dynos-work:status check where your task is
/dynos-work:resume continue after interruption
/dynos-work:investigate [bug] deep bug investigation
/dynos-work:residual inspect or drain non-blocking findings
Requirements
Claude Code or Codex, plus Python 3 for the local dynos-work control plane.
Running with permissions ON
The pipeline works without --dangerously-skip-permissions. Every deterministic step funnels through one binary — <plugin-root>/bin/dynos — so you can allow that single command prefix once (via the permission prompt's "don't ask again" or a Bash(<plugin-root>/bin/dynos ...) allow rule in your settings) and then you only interact with the decisions that are actually yours: discovery questions, spec approval, plan approval, and TDD review. Plugin guardrail denials are not Claude Code permission errors — they say so explicitly and name the sanctioned alternative. Details: docs/write-boundary-spec.md.
Links
Completion is a control decision, not a model opinion.