convez.ai
SkillProdutividadeIntermediário
so0osh

Retro — Retrospectiva de Sessão para Claude Code

so0osh/claude-plugin-retro

Loop de automelhoramento estruturado para agentes de código com memória e auditoria

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que captura padrões comportamentais do seu agente de codificação após cada sessão, registra aprendizados em memória, atualiza regras de conduta e mantém um audit log. Funciona em dois estágios: análise completa sob demanda e captura leve automática antes da compactação de contexto, melhorando continuamente a qualidade e a aderência do agente ao seu fluxo de trabalho.

Instalação

$npx convez-ai so0osh/claude-plugin-retro

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
1
Forks
1
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Rastrear e corrigir padrões comportamentais negativos do agente entre sessões
  • Documentar decisões boas e ruins para reforço contínuo de boas práticas
  • Detectar e bloquear hallucinations, scope creep e uso inadequado de ferramentas
  • Manter auditoria datada de comportamentos e trends de desempenho

Veredito

Ferramenta sofisticada e bem pensada que endereça um problema real: agentes IA que degradam com o tempo sem feedback estruturado. O design em dois estágios (análise pesada + captura leve) é inteligente. Requer PowerShell e Claude Code, limitando o alcance, mas quem estiver nesse ecossistema achará de alto valor.

README

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

retro

retro gives your coding agent a structured self-improvement loop. After each session it scans what happened — hallucinations, scope creep, behavioral drift, tool misuse, good decisions worth repeating — and acts on the findings by writing memories, updating rules, and maintaining a dated audit log. The longer you use it, the better your agent knows how to work with you.

Quickstart

Install for Claude Code.

How it works

At the end of a session, run /retro. The agent reviews the full conversation against 17 behavioral categories, triages findings by severity, and immediately acts on high-confidence ones: writing memory files, updating settings rules, reinforcing positive patterns. Structural changes (new skills, deleted rules) require your approval first. Everything gets logged to ~/.claude/retro-log.md so you can track trends across sessions.

Compaction is handled in two stages so nothing is lost to timing. Before context is compressed, a PreCompact hook snapshots the full transcript to disk. After compaction completes, a SessionStart hook points the agent at that snapshot and the lightweight retro-quick skill captures ephemeral signals — decisions made, corrections given, intent clarifications — into memory without interrupting you.

Over time, retro builds up a picture of how you and your agent work best together. Good behaviors get reinforced. Bad patterns get flagged and blocked. The agent gets better at the things that matter to you specifically.

Installation

Claude Code

Register the marketplace and install:

claude plugin marketplace add so0osh/claude-plugin-retro
claude plugin install retro@retro

Or from within a Claude Code session:

/plugin marketplace add so0osh/claude-plugin-retro
/plugin install retro@retro

Restart Claude Code after installation for the hooks to take effect.

Usage

TriggerSkillWhen
/retroretroEnd of day, after a major delivery, after a long debug session
Automaticretro-quickSnapshots on PreCompact, captured via SessionStart after compaction

What's Inside

Skills

retro — Full on-demand retrospective. Scans the session, presents triaged findings, writes memories and settings updates assertively, gates structural changes behind your approval, appends a dated entry to the audit log.

retro-quick — Lightweight pre-compaction capture. The PreCompact hook snapshots the transcript before compression; after compaction the SessionStart hook hands the snapshot to this skill, which preserves ephemeral signals (corrections, decisions, intent clarifications) into memory without any user interaction.

Analysis Rubric

17 detection categories shared by both skills:

CategoryOutput
Coding qualitymemory, audit
Best practicesmemory, settings
Bugs introduced or caughtmemory, audit
Security issuesmemory, settings
Behavioral driftsmemory, settings
Hallucinationsmemory, audit
Inaccuracies and ambiguitiesmemory, audit
Intent misinterpretationsmemory, audit
Good / bad decisionsmemory, audit
Guesses vs grounded truthsmemory, audit
Scope creepmemory, settings
Tool misusememory, settings
Confirmation biasmemory, audit
Verbosity driftmemory, settings
Memory stalenessmemory, audit
Permission over-reachsettings, audit
Positive patterns to reinforcememory, skill update

Outputs

ArtifactLocationPurpose
Memory files~/.claude/projects/<project>/memory/Recalled in future sessions
Settings rules~/.claude/settings.jsonEnforced rules and hooks
Audit log~/.claude/retro-log.mdCross-session trend record

Extension

Add or refine detection categories by editing shared/analysis-rubric.md. Each entry needs four fields: what to look for, positive form, negative form, and output type. No changes to the skill files required.

Contributing

  1. Fork the repository
  2. Add or improve categories in shared/analysis-rubric.md, or refine skill behavior in skills/retro/SKILL.md or skills/retro-quick/SKILL.md
  3. Test with real session data before submitting
  4. Submit a PR

Updating

claude plugin update retro

License

MIT

Ver o original no GitHub

Bom para

Relacionados