Claude Wiki Pages Plugin
odere-pro/claude-wiki-pages-plugin
Wiki mantida por IA em Obsidian com schema enforced e provenance-tracking
Ver no GitHubPlugin Claude Code para Obsidian que transforma um vault em base de conhecimento estruturada seguindo o padrão LLM Wiki de Andrej Karpathy. Mantém proveniência de fontes com um sistema de hooks em quatro camadas que reforça o schema em cada limite de tool-call, combinando 26 skills, 8 agentes e orquestração multi-agent.
Instalação
npx convez-ai odere-pro/claude-wiki-pages-plugin1 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
- Manter base de conhecimento estruturada sobre ML/LLM com proveniência de fontes
- Orquestrar múltiplos agentes para ingestão, análise e síntese de conteúdo técnico
- Lint e fix automático de documentação seguindo schema definido
- Preencher lacunas em wikis existentes com análise orientada por IA
Veredito

Ferramenta sofisticada e opinada para quem trabalha com conhecimento técnico em Obsidian — excelente arquitetura com hooks enforced e multi-agent orchestration, mas requer compreensão profunda da convenção Karpathy e investimento em onboarding. Ideal para equipes curando wikis estruturadas.
README
A tradução para o português deste README ainda está na fila.
claude-wiki-pages
Karpathy's LLM Wiki, shipped as a Claude Code plugin — four layers, hook-enforced.
A Claude Code plugin that turns an Obsidian vault into a maintained, provenance-tracked knowledge base following Andrej Karpathy's LLM Wiki pattern. The human curates sources; the plugin maintains the wiki; hooks enforce the schema at every tool-call boundary.
The system is convention-driven: the schema lives in skills/init/template/CLAUDE.md, the canonical terms in docs/GLOSSARY.md, the architecture in docs/architecture.md. Every skill, agent, and hook binds to them.
What's inside
| Layer | Surface | Count |
|---|---|---|
| Data | skills/init/template/ — immutable raw/, LLM-maintained wiki/, schema in vault/CLAUDE.md | 1 |
| Skills | 14 short verbs (init, ingest, query, lint, fix, status, synthesize, index, markdown, search, review, draft, sync, fill-gaps) + onboarding + 5 agent-teaching (engine-api, maintain-contract, analyst-modes, curator-fixes, ingest-pipeline) + voice + obsidian-graph-colors + obsidian-vault + 3 third-party obsidian-* (MIT, kepano) | 26 |
| Agents | Orchestrator (entry) + onboarding, ingest, curator, analyst, polish, maintenance + extract-worker (ingest fan-out) — see docs/operations.md | 8 |
| Commands | /claude-wiki-pages:wiki, /claude-wiki-pages:onboarding, /claude-wiki-pages:doctor, /claude-wiki-pages:fill-gaps | 4 |
| Hooks | SessionStart + UserPromptSubmit + 7 PreToolUse + 2 PostToolUse + 3 SubagentStop + Stop + SessionEnd (session-memory persistence) | 16 |
| Rules | Path-scoped guidance under rules/ | 4 |
| Tests | Five tiers — Tier 0 static, Tier 1 Bats unit, Tier 2 smoke, Tier 3 release, Tier 4 adversarial | 5 |
Long-form architecture: docs/architecture.md. Feature list and competitor comparison: docs/features.md.
How it works
You type one verb; the orchestrator routes it through the right skill, the Bun engine does the exact work (verify / fix), and a git checkpoint makes every auto-heal reversible. Hooks enforce the schema at each write.
flowchart TD
U([You]) -->|/claude-wiki-pages:wiki| O[orchestrator agent]
O -->|fresh vault| INIT[init]
O -->|raw/ pending| ING[ingest agent]
O -->|question| ANA[analyst agent]
O -->|drift| CUR[curator agent]
subgraph L2["Layer 2 — Skills (thin orchestrators)"]
INIT & ING & ANA & CUR
end
ING & CUR -->|call| ENG[["Bun engine (13 verbs)<br/>verify · fix · heal · search · doctor"]]
ENG -->|checkpoint then auto-heal| HEAL{{"verify → fix → re-verify<br/>(git-checkpointed)"}}
HEAL -->|clean| VAULT[("Vault<br/>raw/ → wiki/")]
HEAL -.->|rollback| GIT[(git revert)]
PreHook[/"hooks: frontmatter · wikilinks · raw-immutable"/] -.->|gate every write| VAULT
classDef eng fill:#1c2c3c,stroke:#4aa3ff,color:#eaf2ff;
classDef vault fill:#10261a,stroke:#3ddc84,color:#eafff2;
class ENG,HEAL eng;
class VAULT vault;
Prerequisites
| Tool | Purpose | Install |
|---|---|---|
| Claude Code | >= 2.0 | docs.claude.com/code |
bash, git, find | Hook scripts and file walking | Pre-installed on macOS / Linux |
jq | JSON parsing in hooks and resolvers | brew install jq / apt-get install jq |
Bun >= 1.2 | Recommended — runs the deterministic engine (verify/fix/heal/doctor/config) and git-checkpointed self-heal. Without it the plugin still works, but those commands are disabled. | curl -fsSL https://bun.sh/install | bash |
git repo for the vault | So self-heal can checkpoint and git revert | git init in the vault (or /claude-wiki-pages:doctor --fix) |
| Obsidian | Optional — for graph view, Dataview, Web Clipper | obsidian.md |
/claude-wiki-pages:doctor checks all of these and tells you exactly what to install. The SessionStart hook prints a one-line notice if Bun is missing.
OS: macOS or Linux verified. Windows/WSL unverified for hook scripts; markdown-only paths should work.
Install
/plugin marketplace add odere-pro/claude-software-3-0-marketplace
/plugin install claude-wiki-pages
Local-clone install, update, and uninstall: see docs/install.md.
Quickstart
/claude-wiki-pages:wiki
That is the one verb. The orchestrator probes vault state and dispatches:
- No vault yet → runs the
initwizard. Scaffoldsdocs/vault/with the schema and a bundled sample source ready to ingest — no files needed from you to see a real result. - New files in
raw/→ runsclaude-wiki-pages-ingest-agent. Produces typed wiki pages with citations and awiki/log.mdentry, then runsclaude-wiki-pages-polish-agentto refresh graph colors and indexes. - Pending lint after an ingest → runs
claude-wiki-pages-curator-agentto audit and repair. - Analytical prompt (
what,why,compare,summarize, …) → runsclaude-wiki-pages-analyst-agent. Every answer cites[[wikilinks]]back to source.
Full operations reference: docs/operations.md.
Guided first-run wizard — walks scaffold → ingest → first answer, one step at a time:
/claude-wiki-pages:onboarding
Environment health check — run after install and whenever something feels off:
/claude-wiki-pages:doctor
doctor reports all green when prerequisites are met. If it flags anything, fix the prerequisite it names and run /claude-wiki-pages:wiki.
The basic operations — ingestion and querying — also run with a local model, fully offline, with zero Claude. Both are gated: only a model that passed the per-tier quality gate runs (qwen3-coder:30b today), and a query answer is shown only after every citation verifies verbatim against the wiki — otherwise the plugin warns and denies the operation. doctor is deterministic and needs no model at all.
// .claude/claude-wiki-pages.json
{
"localModel": {
"enabled": true,
"model": "qwen3-coder:30b",
"tier": "query",
"offlinePolicy": "prefer-local",
},
}
# With Claude Code stopped and `ollama serve` running:
bash scripts/offline-draft.sh # ingest raw/ → _proposed/ drafts (tier: ingest-extract)
bash scripts/offline-query.sh --question "<question>" # verified, cited answer (tier: query)
Drafts are promoted later via /claude-wiki-pages:review. Full story, tested models, and how tiers unlock: docs/local-models.md.
Documentation
| Topic | Guide |
|---|---|
| Install / update / uninstall | docs/install.md |
| Day-to-day operations | docs/operations.md |
| Multiple vaults | docs/operations.md — Multi-vault registry |
| Local models (Ollama) | docs/local-models.md |
| Features and comparison | docs/features.md |
| Architecture (four layers) | docs/architecture.md |
| Glossary | docs/GLOSSARY.md |
| Security and threat model | SECURITY.md |
| Step-by-step user guides | docs/llm-wiki/ |
| ADRs | docs/adr/ |
| Agent teams (dev) | docs/teams.md |
| Test harness | tests/README.md |
| Contributing | CONTRIBUTING.md |
| Release log | CHANGELOG.md |
| Vulnerability disclosure | SECURITY.md, SUPPORT.md |
Privacy
No telemetry. The plugin never phones home. Your vault, your hooks, your shell. Settings are local at .claude/claude-wiki-pages/settings.json.
License and non-affiliation
Licensed under Apache 2.0. See NOTICE for bundled third-party code and THIRD_PARTY_LICENSES.md for full license text.
Not affiliated with Anthropic, Obsidian, or Andrej Karpathy — this is an independent implementation built on their public work. Credits:
- Andrej Karpathy's LLM Wiki gist — the pattern this implements.
- kepano/obsidian-skills — MIT —
obsidian-markdown,obsidian-bases,obsidian-cliskills, included unmodified. - Anthropic — Claude Code and the plugin format.
- Obsidian — the vault format this plugin maintains.
Author: odere-pro on GitHub.