Help us improve
Share bugs, ideas, or general feedback.
Share bugs, ideas, or general feedback.
Share bugs, ideas, or general feedback.
By Sungmin-Cho
LLM-managed markdown wiki for persistent knowledge accumulation — based on Karpathy's LLM Wiki philosophy. 5 skill-based entry surfaces (cross-platform: Claude Code slash + Codex/Copilot CLI/Gemini CLI/SDK via Skill()).
npx claudepluginhub sungmin-cho/claude-deep-suite --plugin deep-wikiBody-generation worker for A5 page-level fanout. Receives one page_plan entry with pre-captured existing_page_body bytes, emits one page_content draft. Invoked by /wiki-ingest in single-source A5 fanout mode (page_plan ≥ a5_fanout_threshold).
Stage 1 of single-source A5 page-fanout (v1.4.0+). Reads source + candidate pages, decides create-vs-update per topic, emits a `page_plan` (with `inline_bodies` for sub-threshold cases). DOES NOT write or version pages — main owns Stage 3 atomic-write under lock. Invoked by /wiki-ingest Step 7.5 single-source branch.
Dormant in v1.4.x+; preserved for future restoration of v1.3.0 single-source byte-identical inline path. Reads source + candidate pages, decides create-vs-update, version-backs-up under .wiki-meta/.versions/, atomic-writes pages under <wiki_root>/pages/, returns manifest. NO active caller in v1.4.x+ (skills/wiki-ingest/SKILL.md Step 7.5 routes to wiki-synthesizer-analysis instead).
Multi-source A4 fanout worker (v1.3.0+). Reads one source-shard + cross-page candidates, decides create/update/skip per topic, emits structured `drafts[]` for main session to aggregate + atomic-write under the global lock during Phase 3. Also handles second-pass cross-worker collision merge via `colliding_drafts` input. Invoked by /wiki-ingest Step 7.5.M-A (per-shard) and Step 7.5.M-B Case B2 (collision second-pass).
Use when the user wants to inspect the deep-wiki for health issues — orphan pages, broken inter-page links, schema violations, contradictions, stale content, missing source provenance YAML, `pages_created` exactly-once invariant breaks, and `.last-scan` monotonicity violations. Triggers on `/wiki-lint`, "lint wiki", "check wiki health", "wiki status", "wiki audit", "wiki diagnose", "wiki dashboard", "위키 린트", "위키 점검", "위키 상태", "위키 헬스체크", "위키 무결성". Accepts an optional `--fix` flag that auto-repairs the auto-fixable subset (broken links, orphan removal, sources/ slug normalization) while leaving audit-only items in the dashboard.
Use when the user wants to ask a question against the deep-wiki and get an answer grounded in the accumulated knowledge — searching pages, reading their bodies, and (when the answer is novel cross-page synthesis) auto-filing the answer back into the wiki as a new page via the `query-filed` lifecycle action. Triggers on `/wiki-query`, "ask the wiki", "wiki query", "search the wiki", "ask wiki", "위키 질의", "위키 검색", "위키에서 찾기", "위키에 물어보기", "위키 답변". The question string is passed as the sole argument.
Use when the user wants to regenerate derived wiki artifacts (M3-envelope-wrapped `.wiki-meta/index.json`) from the source-of-truth page frontmatter — used when the index is suspected stale, corrupted, or drifted relative to actual page files. Triggers on `/wiki-rebuild`, "rebuild wiki index", "regenerate index", "rebuild wiki", "wiki reindex", "wiki rebuild", "위키 인덱스 재생성", "위키 재빌드", "위키 인덱스 복구", "wiki 재구성". Takes no arguments; always acquires the wiki lock and emits a `rebuild` lifecycle event to `log.jsonl`.
Defines schema, invariants, and lifecycle rules for any deep-wiki knowledge base operation. Activate whenever wiki pages or metadata are read, created, modified, or validated — specifically during /wiki-ingest, /wiki-query (including auto-filing of cross-page synthesis), /wiki-lint, /wiki-rebuild, /wiki-setup, page frontmatter validation, source provenance tracking, or any direct manipulation of pages/, index.json, log.jsonl, sources/, or .versions/ under a wiki root. Covers required frontmatter fields, kebab-case naming, markdown link conventions, source provenance YAML, M3-envelope-wrapped index.json catalog, append-only log.jsonl event log with 10 lifecycle actions (ingest, ingest-skip, ingest-repair, ingest-fail, update, lint, rebuild, delete, query-filed, setup), mkdir-based concurrency locking, page versioning, and the 4 critical invariants (pages_created exactly-once, .last-scan monotonicity, lock atomicity, source provenance correspondence).
Use when the user wants to initialize the deep-wiki knowledge base for first use — creating the wiki root directory, the config file at `~/.claude/deep-wiki-config.yaml`, the initial directory structure (`pages/`, `.wiki-meta/`), and a seed `welcome.md` page. Triggers on `/wiki-setup`, "init wiki", "set up wiki", "scaffold wiki", "create wiki", "wiki bootstrap", "위키 초기화", "위키 셋업", "위키 설정", "위키 만들기", "위키 시작". Accepts an optional `<wiki_root_path>` argument; otherwise prompts the user via AskUserQuestion (option A — inside an Obsidian vault, option B — standalone directory). Windows users must supply a POSIX form path (`/c/...` or `/mnt/c/...`).
Uses power tools
Uses Bash, Write, or Edit tools
Share bugs, ideas, or general feedback.
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge.
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge.
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
LLM-maintained personal wiki skills for Claude Code. Implements Karpathy's LLM Wiki pattern — persistent, compounding knowledge base for research, codebase documentation, or any long-term knowledge accumulation.
Persistent, compounding knowledge base maintained by LLMs in Obsidian — agent-first edition. Four task-oriented agents (Researcher / Advisor / Curator / Scribe) with citations, confidence, supersession, and rolling session cache. Inspired by Karpathy's LLM Wiki pattern.
A persistent LLM-maintained wiki for your knowledge base. Ingest sources, query your knowledge, and lint for consistency.
Karpathy-style local-first LLM Wiki / Knowledge Compiler for Claude Code. Ingest sources, build a structured markdown knowledge base, and maintain it with AI-powered workflows.
Build and maintain an LLM-curated personal knowledge base in your project — Andrej Karpathy's LLM Wiki pattern, designed to scale to thousands of pages without becoming a context bottleneck. Now with an optional compiled graph layer for typed, provenance-backed relationships.
LLM-maintained knowledge base skill — structured wiki with Obsidian, milestone-based source clustering, proactive write-back, and autonomous lint
Evidence-Driven Development Protocol — 24 command-equivalent skill surfaces plus a skill-native deep-work entry alias (cross-platform: Claude Code skills + Codex / Copilot CLI / Gemini CLI / SDK), 3-layer architecture with Skill-based phase dispatch, computational enforcement (worktree guard + phase transition injection), TDD enforcement, and receipt validation
Autonomous Experimentation Protocol — goal-driven experiment loops that systematically improve any project through measured code modifications
Independent Evaluator for AI coding agents — cross-model code review with Codex integration
Document gardening agent — validates freshness and auto-repairs CLAUDE.md, AGENTS.md, and project docs
English | 한국어
An LLM-managed markdown wiki for persistent knowledge accumulation — a plugin implementation of Karpathy's LLM Wiki philosophy for Claude Code and Codex.
"Most people's experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There's no accumulation." — Andrej Karpathy
Instead of re-discovering knowledge each time (RAG), deep-wiki incrementally builds and maintains a persistent wiki — a structured, interlinked collection of markdown files. When you add a new source, the LLM reads it, extracts key information, and integrates it into the existing wiki. The cross-references are already there; the contradictions have already been flagged; the synthesis already reflects everything you've read. The knowledge is compiled once and kept current, not re-derived on every query.
deep-wiki is the persistent knowledge layer of the deep-suite. In the Harness Engineering 2×2 matrix it acts as an Inferential Guide — accumulated project knowledge that shapes the agent's understanding, replacing repeated RAG queries with a compounding knowledge base. The 5 /wiki-* entry points are skills, so they run natively from Claude Code (slash commands) and from Codex / Copilot CLI / Gemini CLI / the Agent SDK via Skill({ skill: "deep-wiki:wiki-<verb>" }).
Based on Karpathy's three-layer model:
Raw Sources → Wiki (markdown pages) → Schema (management rules)
↑ ↑ ↑
wiki-ingest pages/ wiki-schema skill
| Layer | Description | Owner |
|---|---|---|
| Raw Sources | Immutable inputs — files, URLs, text, reports | You curate |
| Wiki | LLM-generated markdown pages with cross-references | LLM writes, you read |
| Schema | Rules governing how the wiki is structured and maintained | Co-evolved |
# Claude Code
/plugin marketplace add Sungmin-Cho/claude-deep-suite
/plugin install deep-wiki@Sungmin-Cho-claude-deep-suite
# Codex
codex plugin install deep-wiki
/plugin marketplace add Sungmin-Cho/claude-deep-wiki
/plugin install deep-wiki@Sungmin-Cho-claude-deep-wiki
Prerequisite: the Claude Code CLI (or Codex) installed and configured.
# 1. Initialize the wiki
/wiki-setup ~/Obsidian/MyVault/wiki
# 2. Ingest sources into the wiki
/wiki-ingest https://example.com/article
/wiki-ingest ./document.pdf
/wiki-ingest # paste text directly
# 3. Query the wiki
/wiki-query What are the rules of React hooks?
# 4. Health check
/wiki-lint
| Command | Description |
|---|---|
/wiki-setup | Initialize the wiki and create the directory structure |
/wiki-ingest | Read a source (URL, file, text, deep-work report) and create/update wiki pages |
/wiki-query | Search the wiki and generate grounded answers; auto-files cross-page syntheses back into the wiki |
/wiki-lint | Health check — schema violations, orphan pages, broken links, contradictions (also runs automatically after ingest/rebuild) |
/wiki-rebuild | Regenerate the machine-readable index from page frontmatter |
Ingest — Drop a new source and the LLM reads it, writes summary pages, updates the index, updates relevant pages across the wiki, and appends to the log. A single source might touch multiple pages. New information is merged with existing content — pages grow richer with each ingest. Auto-lint runs after every ingest.
Query — Ask questions against the wiki. The LLM searches for relevant pages using a three-layer strategy (index scan → content search → candidate reading) and synthesizes a grounded answer with citations. When a query synthesizes insights across 2+ pages, the result is automatically filed back into the wiki — the knowledge compounds.
Lint — Health-check the wiki: schema violations, contradictions, orphan pages, broken links, stale versions, and index drift. --fix auto-repairs structural issues. Runs automatically after ingest and rebuild — invoke it manually for deep inspections.