convez.ai
SkillDesenvolvimento WebAvançado
jamieafk

Fable Check

jamieafk/fable-check

Revisão de código estruturada e adversarial com Fable 5

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill de revisão de código extensiva alimentada por Claude Fable 5, executável no Claude Code e Codex. Oferece análises estruturadas (aprovação/atenção necessária), modo adversarial para desafiar decisões de design, e reviews profundas em paralelo (correção, segurança, arquitetura). Funciona read-only com progresso em tempo real via stderr.

Instalação

$npx convez-ai jamieafk/fable-check

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
42%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Análise profunda de pull requests antes de merge
  • Desafiar designs de retry e tratamento de concorrência
  • Validação de segurança e integridade de dados em mudanças
  • Aconselhamento arquitetural em decisões de tradeoff
  • Verificação contínua em CI/CD via terminal

Veredito

Ferramenta robusta e bem arquitetada para times que querem code review assistido por IA com rigor técnico real. A abordagem adversarial e os múltiplos modos (simples, deep, ask) cobrem casos de uso sofisticados; a implementação read-only + streaming de progresso mostra engenharia madura. Exige Node.js 18+ e Claude Code ativo, mas o ROI é alto em projetos que usam esse stack.

README

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

fable-check

Extensive code review and advisory powered by Claude Fable 5 — usable as a skill from both Claude Code and Codex. The inverse of openai/codex-plugin-cc: instead of Claude Code calling Codex for reviews, any agent (or your terminal) calls Fable 5 for reviews.

What you get

  • review — a thorough, structured review of your uncommitted changes or your branch vs a base. Verdict (approve / needs-attention), severity-sorted findings with file:line locations, confidence scores, and next steps.
  • review --adversarial [focus text] — a skeptical review that actively tries to block the change. Steerable: "challenge the retry design", "look for race conditions", etc.
  • review --deep — three parallel review passes (correctness, security & data safety, design & failure modes) merged and de-duplicated into one report. The most extensive mode.
  • ask "<question>" — advisory mode: Fable explores your repo read-only and answers an architecture/design/tradeoff question in prose, with one clear recommendation.
  • status / result / cancel + --background — run long jobs in the background with live progress (phase, elapsed time, tool activity).
  • setup — checks that everything is installed and logged in.

The reviewer runs read-only: it can read files, grep, and run read-only git commands to verify findings against the real code, but it can never modify anything.

Runs are never silent: every tool call the reviewer makes streams to stderr as a progress line, with heartbeats every ~20s during thinking stretches — so a watching agent (or human) always knows it's alive, what phase it's in, and what it's doing.

Requirements

  • Claude Code installed and logged in (reviews use your existing Claude subscription — no API key)
  • Node.js 18+
  • git

Install

git clone https://github.com/jamieafk/fable-check.git
cd fable-check
./install.sh

This symlinks the skill into ~/.claude/skills/fable-check and ~/.codex/skills/fable-check and runs a setup check. After that:

  • In Claude Code or Codex: say "run a fable check" / "do a fable review of my changes".
  • From a terminal, directly:
node skill/scripts/fable-check.mjs review
node skill/scripts/fable-check.mjs review --base main
node skill/scripts/fable-check.mjs review --adversarial question whether this caching design is safe
node skill/scripts/fable-check.mjs review --deep --background
node skill/scripts/fable-check.mjs ask "is the detached-worker design for background jobs sound, or should this use a queue?"
node skill/scripts/fable-check.mjs status   # live progress of a background job
node skill/scripts/fable-check.mjs result

How it works

The script collects your git context (working-tree diff or branch diff against a base; auto-detected like the original plugin). Small diffs are inlined into the prompt; large ones switch the reviewer into self-collect mode, where it explores the repo itself with read-only tools. It then runs the local claude CLI headlessly:

claude -p --model claude-fable-5 --effort xhigh --output-format stream-json --verbose \
  --json-schema <review schema> --allowedTools <read-only set>

The streamed events power live progress reporting (per-tool-call lines + heartbeats on stderr; live status for background jobs); the final event carries the structured review, which is validated against the schema. Reports are printed, saved under ~/.fable-check/jobs/<repo>/, and include a claude -r <session-id> command to reopen the review session interactively.

Differences from codex-plugin-cc

codex-plugin-ccfable-check
Reviewer modelGPT-5.x via CodexClaude Fable 5
HostClaude Code (plugin)Claude Code and Codex (skill), or plain terminal
Reviewer capabilitiesinline diff or self-collectsame, plus agentic read-only repo exploration in every mode
Deep mode—--deep: 3 parallel lenses + merge pass
Task delegation (rescue)yesnot included (out of scope: this is a review tool)
Stop-hook review gateoptionalnot included (plugin-only machinery; drains usage)

License & attribution

Apache-2.0 — see LICENSE.

This project is inspired by, and adapts portions of, openai/codex-plugin-cc (Apache-2.0): the review/adversarial prompt structure, the structured findings schema, and the review-target selection design. See NOTICE for details. All other code is original.

fable-check is an independent project, not affiliated with or endorsed by OpenAI or Anthropic. "Claude" and "Codex" refer to the third-party tools it integrates with.

Ver o original no GitHub

Bom para

Relacionados