convez.ai
SkillAutomaçãoAvançado
Stealthy-McStealth

self-evolve

Stealthy-McStealth/self-evolve

Agentes Claude Code que aprendem com seus erros e mantêm consistência

Ver no GitHub
BomPluginScore · de 0 a 100

Plugin que transforma um agente Claude Code em um sistema autoaprendizado persistente, similar ao Hermes. Extrai habilidades duráveis de execuções anteriores, as armazena e impõe conformidade em trajetórias longas para evitar que o agente desvie de suas instruções ou repita erros.

Instalação

$npx convez-ai Stealthy-McStealth/self-evolve

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
6
Forks
0
Acompanhando
0
Saúde
62%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Desenvolvimento iterativo com aprendizado multi-sessão
  • Automação de tarefas longas mantendo foco e instruções
  • Construção de bibliotecas de skills reutilizáveis no Code agent

Veredito

Implementação rigorosa (com fundamentação em pesquisa peer-reviewed) que resolve um problema real: Claude Code sem persistência de aprendizado. A abordagem de extração de skills e enforcement de aderência é sofisticada, embora exija que o usuário domine o sistema de plugins do Claude para extrair o máximo valor.

README

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

self-evolve

Make your Claude Code agent work like a real harness — the one Hermes uses.

an agent that can't learn from its mistakes isn't an agent. it's a very expensive grep.

A Claude Code plugin that gives your agent self-evolving harness capabilities: it learns from mistakes, encodes durable skills, and actually follows them across long trajectories. The same architecture that makes Hermes effective — applied to your local Claude Code setup.

Backed by peer-reviewed research. Not vibes.

License: MIT


why this exists

Hermes-style agents don't just run tools. They build up a persistent skill library from execution traces and stay on-task across long trajectories. That's what makes them effective over time — they compound.

Out of the box, Claude Code doesn't do this. Every session starts fresh. Skills drift mid-conversation. You repeat yourself.

This plugin fixes both problems with two mechanisms:

  1. Skill evolution — extract learnings from conversations and persist them as reusable skills (like Hermes's harness updating)
  2. Adherence enforcement — prevent mid-trajectory drift where the agent loads instructions but quietly stops following them

the research

This plugin implements findings from:

"Harness Updating is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents" Lin et al., arXiv:2605.30621, May 2026

Key findings:

FindingImplication
Any model can write good skills (spread: 3.1pp max)Skill-writing is pattern extraction, not deep reasoning
Performance is bottlenecked by the agent, not the evolverInvest in following skills, not writing them
Weak agents fail to activate skills (25% vs 96%)Trigger phrases and explicit invocation matter
Adherence decays 4x faster in weak models over long trajectoriesAlways-on drift prevention is essential

what's included

self-evolve/
├── CLAUDE.md                          # always-on adherence protocol
├── skills/
│   ├── improve-urself/SKILL.md        # /improve-urself — encode learnings as skills
│   └── reload-context/                # /reload-context — fix mid-trajectory drift
│       ├── SKILL.md
│       └── references/
│           └── skill-standards.md     # template for writing followable skills
└── .claude-plugin/plugin.json

skills

/improve-urself [optional: topics]

Scans the current conversation for mistakes, corrections, and patterns. Creates or updates a persistent skill so the learning survives beyond this session.

/improve-urself
/improve-urself bash error handling
/improve-urself json formatting tool-use

The key insight: you don't need a frontier model to write the skill. The quality is bounded by what the execution trace reveals — not by how smart the writer is. Even small learnings compound across sessions.

/reload-context

Re-reads every loaded skill and active instruction. Explicitly re-commits to fallback steps. Makes the active harness visible and auditable.

Use it when:

  • Claude is improvising around instructions it should be following
  • A task is getting long and you want to reset attention
  • Output looks right-shaped but something feels off

CLAUDE.md — always-on adherence protocol

Runs every session without explicit invocation:

  • State relevant loaded skills before multi-step tasks
  • Re-check alignment at sub-task transitions
  • Acknowledge fallbacks upfront, not after step 1 fails
  • If improvising around a skill → stop and re-read

install

Claude Code CLI

claude plugin add Stealthy-McStealth/self-evolve

Manual

Clone this repo into your Claude Code plugins directory:

git clone https://github.com/Stealthy-McStealth/self-evolve.git ~/.claude/plugins/self-evolve

limitations

  • Non-monotonic benefit: mid-tier models benefit most from evolved skills. Very weak models struggle even with perfect skills because they can't reliably invoke or follow them.
  • Not fully automatic: you call /improve-urself yourself. Fully automatic skill evolution from conversation traces requires more infrastructure — that's future work.

contributing

PRs welcome. If you've found a pattern that makes skills more followable or adherence more durable, open an issue or submit a skill.


related


license

MIT

Ver o original no GitHub

Bom para

Relacionados