convez.ai
SkillAutomaçãoAvançado
curtyo18

Agent Config — Configuração de Padrões e Skills para Claude Code

curtyo18/agent-config

Padrões, hooks, skills e commands para agentes Claude — código, segurança e contexto

Ver no GitHub
InicialPluginScore · de 0 a 100

Repositório de configuração centralizada para Claude Code com padrões de código, git hooks, slash commands e skills de agente. Inclui auditorias de segurança, revisão de código em camadas, planejamento executável e ferramentas para compactação de contexto. Clone em ~/.claude para integração automática.

Instalação

$npx convez-ai curtyo18/agent-config

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
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 3 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Definir padrões de qualidade de código automatizados para equipes
  • Executar revisões de código em profundidade com múltiplos agentes
  • Escanear secrets e vulnerabilidades pré-commit
  • Compactar contexto e gerenciar tokens em sessões longas
  • Estruturar planejamento executável (idea → plan → execução)

Veredito

Configuração robusta e bem documentada para equipes que usam Claude Code como agente de desenvolvimento. Valor real em automação de padrões, segurança e gerenciamento de contexto; exige setup inicial em ~/.claude mas depois opera transparente. Melhor aproveitado por times que já usam Claude Code nativamente.

README

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

agent-config

Claude Code configuration: coding standards, git hooks, slash commands, and skills.

What's here

  • standards/ — Code quality rules loaded by /code-review: naming, testing, duplication, module depth, dependency discipline, UI preferences.
  • hooks/ — Pre-commit secret scanning (filename, regex, gitleaks).
  • commands/ — Slash commands for code review (code-review plus the layered multi-agent deep-code-review), release checklists, cloud agent briefs, Chrome extension scaffolding, pre-public-repo audits, and selective context compaction (context-collector).
  • skills/ — Agent skills: idea (vague → grilled spec → plan), run-plan (subagent execution), to-issues, grill-me, debug (root-cause-first investigation).
  • CLAUDE.md — Global conventions cloned in as ~/.claude/CLAUDE.md: scope discipline, two-strike rule, git workflow, permission hygiene.
  • settings.json — Claude Code settings: default model, permission allowlist + default mode, and statusline wiring.
  • statusline.py — Optional Claude Code status line script showing session usage metrics.
  • gitleaks.toml — Secret pattern rules for the pre-commit scan.
  • network-allowlist.conf — Squid ACL entries used by agent-sandbox.

Using this

Clone into ~/.claude:

git clone https://github.com/curtyo18/agent-config.git ~/.claude
# Wire the pre-commit secret scan (the sandbox does this for you; standalone clones must opt in):
git config --global core.hooksPath ~/.claude/hooks

Or let agent-sandbox do it automatically on container start.

Heads-up on defaults. settings.json ships permissive defaults (bypassPermissions, skipDangerousModePermissionPrompt) tuned for the agent-sandbox container, where the container is the sandbox. If you clone this standalone onto a host, review settings.json first — outside the sandbox you'll likely want a stricter permissions.defaultMode.

Status line

statusline.py renders a compact, colour-coded session readout in the Claude Code status bar:

agent-config status line

Left to right:

  • cwd — current directory (dimmed; ~ for home, long paths truncated) for orientation.
  • ctx:N% (Mk) — context window used: percentage of the window plus the absolute token count (in thousands). Its colour tracks the absolute tokens, not the percentage — what degrades a session ("context rot") is how many tokens are in play, not how full the window is, so ctx:45% can already be red if that's ~80k tokens.
  • 5hr:N% (Xh Ym) — Claude.ai Pro/Max 5-hour rate-limit usage and time to reset; falls back to session:$X.XX (session cost) when no rate-limit data is present.
  • weekly:N% — Claude.ai Pro/Max 7-day rate-limit usage.

The context colour escalates with absolute size — red at ~75k, and bold red at ~100k, where context rot starts to bite (your cue to compact, summarise, or start a fresh session):

Context tokensColour
< 25kgreen
25k–50kyellow
50k–75korange
75k–100kred
≥ 100kbold red (rot territory)

(The 5-hour and weekly bars use a percentage scale instead: green < 40%, then yellow, orange, red ≥ 75%, and bold red ≥ 90%.)

Planning & execution skills

Self-contained — no plugin dependency; clone the repo and they work as-is. idea is the front door (vague idea → grilled spec → numbered plan → handoff menu); run-plan executes the plan via per-task subagent dispatch with two-stage review (or inline); to-issues converts it to GitHub issues instead; grill-me (a relentless-interview approach inspired by Matt Pocock) drives the questioning in idea's shaping and grilling phases and also runs standalone to stress-test a plan or design. They compose, but each stands alone. (Reimplemented from what the superpowers plugin used to provide.)

Adapting it

Fork and edit directly. The standards, hooks, and commands are designed to be modified — they represent opinions, not rules.

Ver o original no GitHub

Bom para

Relacionados