Auto-Routines: Automação Evolutiva de Repositórios
paipeline/auto-routines
Repositório que se desenvolve e mantém a si mesmo, automaticamente.
Ver no GitHubUma skill Claude Code que instala um harness de automação permanente em seu repositório, executando rotinas que se adaptam lendo commits, PRs e logs de CI. Suporta até 100 agentes trabalhando 24/7 em modo totalmente automático ou orientado por objetivos, com cada mudança rastreável via git commits.
Instalação
npx convez-ai paipeline/auto-routines1 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
- Manter testes rodando automaticamente em cada commit
- Gerar relatórios diários e digests de status do projeto
- Atualizar documentação (README, docs) quando código diverge
- Fechar PRs técnicas de manutenção sem intervenção manual
- Executar múltiplas rotinas de qualidade e linting em paralelo
- Remover código e automações obsoletas após inatividade
Veredito

Projeto ambicioso e bem-executado para quem quer delegar manutenção contínua de repos a agentes. Exige confiança em automação e Git sólido; melhor para equipes que mantêm máquinas ligadas 24/7. Ideia de 'harness via git hooks + MCP + branches isoláveis' é genuinamente criativa.
README
A tradução para o português deste README ainda está na fila.
_ _ _
__ _ _ _| |_ ___ _ __ ___ _ _| |_(_)_ __ ___ ___
/ _` | | | | __/ _ \ _____| '__/ _ \| | | | __| | '_ \ / _ \/ __|
| (_| | |_| | || (_) |_____| | | (_) | |_| | |_| | | | | __/\__ \
\__,_|\__,_|\__\___/ |_| \___/ \__,_|\__|_|_| |_|\___||___/
always-on evolving harness on your repo.
auto-routines
Automation is the best harness. A discipline agents don't maintain is a harness agents don't have. Let your repo wear the harness for you.
1–100 agents, working on your repo 24/7. Seconds to put to work.
1. cd into any repo
2. answer the setup questions
3. keep the laptop / server running
4. go get some fresh air for a few days
5. come back to a repo that has been developing and maintaining itself
auto-routines is not a planner — it actually installs the harness on disk: writes .git/hooks/post-commit shell scripts, creates scheduled tasks via the scheduled-tasks MCP, generates per-routine SKILL.md files under .claude/skills/, and registers Stop hooks in .claude/settings.json. After install it verifies every artifact landed; if anything is missing it aborts with a written failure report. From there it runs and evolves itself. Each routine writes code on a routines/<id> branch and opens a PR — none of them just print findings. Pick one of two modes:
fully-auto— the meta-agent picks direction from signals alone (CI flake rate, PR queue depth, doc drift, commit cadence). The repo keeps itself healthy.goal-driven— you set an iteration goal. The meta-agent picks routines that close the gap to it.
Every change the meta-agent makes is a git commit. Revert any of them with one command. You stay in the loop via a plain-text status block (am/pm times, FSM state per routine, no cron, no mermaid) refreshed after every run.
Before vs. after
| Before auto-routines | After auto-routines | |
|---|---|---|
| Who pushes the PRD forward | You, when you have time | A 4-hour routine that reads .iteration/goal.md, picks the next slice, writes code + tests, opens a PR |
| Who runs the tests | You, when you remember | A post-commit routine, every commit |
| Who reads the CI log | You, after a Slack ping | A 15-min watcher, comments the failing line on the PR |
| Who writes the daily digest | Nobody — you "will later" | An 18:00 routine, drops it in .iteration/digests/ |
| Who keeps the README in sync | Nobody — it rots | A weekday drift fixer, opens a PR when it diverges |
| Who removes stale automation | Nobody — it accumulates | The daily meta-agent, neutralizes routines that go quiet |
| Agents working on your repo | 0 | up to 100, on cron / hooks / loops, 24/7 |
| Your job | Hold the discipline yourself | Read the diff in the morning |
Quick start
git clone https://github.com/paipeline/auto-routines ~/.claude/skills/auto-routines
cd /your/project
claude --dangerously-skip-permissions # required: routines fire while you're away
> /auto-routines
Requires gh CLI, Python 3.9+ with pyyaml, and the scheduled-tasks MCP. The meta-routine and the scheduled tasks all invoke Claude — they only run unattended if Claude is in auto mode (no per-tool prompts). Without that, the harness sits there waiting for your approval on every fire.
What it looks like running
$ /auto-routines evolve
sanity check: OK
checkpoint: iter-010 (sha 3a1f9c2) — triggered by: pr-ci-watcher
changes:
+ added release-tag-checker (every git commit) — release commits without version bump
~ retuned pr-ci-watcher every 30 minutes → every 15 minutes — CI flake rate tripled
→ completed weekly-dep-audit — success_criterion met (0 vulns 4 runs running)
→ stagnant doc-drift-fixer — 11 runs, 0 useful, paused (re-openable)
$ /auto-routines status
goal: ship v1.0 with great test coverage mode: fully-auto
meta evolve: 9:00 AM daily ─ last fired 2h ago, next 9:00 AM tomorrow
routine schedule state runs useful noisy notes
───────────────── ────────────────────── ────────── ──── ────── ───── ──────────────────────
pr-ci-watcher every 15 minutes ACTIVE 147 132 15 retuned at iter-010
release-checker on every git commit ACTIVE 7 7 0 added at iter-010
daily-digest 6:00 PM daily ACTIVE 12 12 0
doc-drift-fixer 5:00 PM weekdays STAGNANT 11 0 0 paused at iter-009
weekly-dep-audit 9:00 AM Mondays COMPLETED 4 2 0 goal: 0 vulns reached
evolve requests pending: 0
last iter: iter-010 — 2h ago — triggered by: pr-ci-watcher
No mermaid, no cron syntax, no black box. Times are am/pm. State is one of:
ACTIVE firing on schedule
EVOLVING meta-agent is currently re-evaluating it (transient)
STAGNANT no incremental work for N runs — paused, re-openable
COMPLETED success_criterion met — paused, re-openable
STOPPED user-disabled or meta-removed — terminal (run /auto-routines start to revive)
A use case
Three weeks into a side project, the discipline you started with has rotted. Tests skipped, README stale, CI red for two days.
You run /auto-routines once. The interview asks the goal, the mode, and — for each candidate routine — the frequency in plain English (every 15 minutes, 5:00 PM weekdays, on every git commit), an optional success criterion ("CI green on last 50 PRs"), and whether the routine may evolve itself.
By week four the meta-agent has marked the drift fixer STAGNANT (no signal in 11 runs), retuned the PR watcher to every 15 minutes (CI got flaky), and added a release-tag-checker on its own — because it noticed you keep forgetting to bump versions. The dep-audit routine hit its COMPLETED state (0 vulns four runs running) and stopped firing. None of it is in your way.
Halfway through week three, the PR watcher itself dropped a line into .iteration/evolve_requests.jsonl saying "CI flake rate is now 0%, recommend reducing my frequency." The meta-agent picked it up at 9:00 AM the next morning and retuned. Routines can ask the skill to re-evaluate them.
You never maintained the discipline. The repo did.
Commands
/auto-routines # init if first run, else show status
/auto-routines evolve # run one iteration (meta calls this daily)
/auto-routines evolve --triggered-by <id> --reason <text> # mid-run, called by routines
/auto-routines status # text status block (am/pm, FSM state)
/auto-routines stop <routine_id> # ACTIVE → STOPPED
/auto-routines start <routine_id> # STAGNANT|STOPPED → ACTIVE
/auto-routines revert iter-007 # restore checkpoint
Contributing
PRs and issues are welcome. The project is developed test-first — every guardrail has a failing test before the check that makes it pass. See CONTRIBUTING.md for the dev loop, ground rules, and the TDD workflow.
pip install pyyaml pytest
pytest -q # 100 tests, ~80ms
Good first issues are tagged good first issue. Routine proposals go through the feature-request template.
License
MIT — see LICENSE. If this is useful, star it. PRs welcome.