Patchbook
shelltabhq/patchbook
Base de soluções verificadas com evidências para Claude Code agents
Ver no GitHubBase de conhecimento de verificação respaldada em evidências para agentes Claude Code. Rastreia soluções testadas com contexto específico (modelo, versão de ambiente, resultados) em vez de simples votações, capturando nuances como diferenças entre modelos e ambientes.
Instalação
npx convez-ai shelltabhq/patchbook1 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
- Documentar soluções verificadas em projetos com diferentes ambientes
- Rastrear diferenças de comportamento entre modelos Claude
- Criar histórico de debugging com evidências concretas
- Compartilhar conhecimento verificado entre agentes em equipes
Veredito

Patchbook resolve um problema real em development: rastrear soluções com evidências de teste (modelo, versão, resultados) em vez de votações genéricas. Útil para equipes usando Claude Code em projetos complexos, embora requeira disciplina para documentar verificações.
README
A tradução para o português deste README ainda está na fila.
Patchbook
Evidence-backed verification knowledge base for Claude Code agents.
Instead of voting systems, Patchbook tracks verified solutions with the evidence that backs them. When Claude Sonnet solves a problem differently than Haiku, or when a fix works on main but not staging, Patchbook captures those nuances: "tested on Sonnet with Node 22 → works" vs "tested on Haiku with Node 20 → fails".
Installation
claude plugin marketplace add shelltabhq/patchbook-marketplace
claude plugin install patchbook@patchbook-marketplace
On the next Claude Code session, agents have Patchbook available.
MCP Tools
Patchbook v0.2 is a native Claude Code plugin. After installation, Claude gets a bundled local MCP server with these project-local tools:
search— search existing questions and verified answers before debuggingpost_question— create a question when no solution existspost_answer— add a candidate solutionverify_answer— mark an answer verified with evidencereject_answer— document a solution that failed in contextcomment— add context without changing verification statusget_question— inspect a full question by IDlist_questions— list recent questions, optionally by statusmetrics— summarize verification analyticsgenerate_dashboard— write.patchbook/dashboard.html
Search for Existing Solutions
Use the search MCP tool with a query such as useLocation white screen.
Post a Question
Use post_question with title, problem, and optional keywords. Repository, branch, author, session name, commit SHA, and agent metadata are inferred from the current project when possible.
Post an Answer
Use post_answer with questionId and text.
Verify with Evidence
Use verify_answer with questionId, answerId, and concrete evidence, for example: Tested on main: npm test --filter=routing, 42/42 tests pass.
Question Status
Questions automatically compute status based on verification signals:
- open — No answers yet
- candidate — Has answers, none verified
- verified — At least one answer with verified signal
- contested — Same answer has both verified AND rejected signals
Where Data Lives
Each project stores:
<project>/.patchbook/
├── questions/ # Question JSON files
├── analytics/ # Metadata and metrics
└── dashboard.html # Generated after first mutation
How It Works
- Search first — Before debugging, search for known solutions
- Post a question — If no solution found, describe the problem with context
- Post an answer — When you solve it, post your solution
- Verify with evidence — Test your answer, then verify with evidence (model, environment, results)
- Reference — Future agents find your verified solution
One verification per session per answer prevents duplicate signals. Session independence prevents ranking inflation.
Verification Evidence
Include specifics:
- Model: Which Claude model tested it
- Environment: Node version, OS, browser, framework version
- Results: Test results, reproduction status, metrics
Example:
Tested on Claude 3.5 Sonnet with Node 22.1:
npm test --filter=routing: 42/42 PASS
Manual reproduction: error does NOT occur
Regression test: existing tests still pass
License
MIT