██╗     ██╗     ███╗   ███╗    ██╗    ██╗██╗██╗  ██╗██╗
██║     ██║     ████╗ ████║    ██║    ██║██║██║ ██╔╝██║
██║     ██║     ██╔████╔██║    ██║ █╗ ██║██║█████╔╝ ██║
██║     ██║     ██║╚██╔╝██║    ██║███╗██║██║██╔═██╗ ██║
███████╗███████╗██║ ╚═╝ ██║    ╚███╔███╔╝██║██║  ██╗██║
╚══════╝╚══════╝╚═╝     ╚═╝     ╚══╝╚══╝ ╚═╝╚═╝  ╚═╝╚═╝

github.com/nvk/llm-wiki

LLM-compiled knowledge bases for any AI agent. Parallel multi-agent research, thesis-driven investigation, source ingestion, wiki compilation, truth-seeking audits, querying, and artifact generation. Ships as a Claude Code plugin, an OpenAI Codex plugin, an OpenCode instruction file, or a portable AGENTS.md for any other LLM agent. Obsidian-compatible.

Install · Quick Start · Commands · How It Works · Research Modes · Thesis Research · Query Depths · Linking · Obsidian · Architecture · Nono Sandbox · Upgrade · Changelog · Credits

v0.7.0 — PDF, message archive, and Wayback adapters. PDF ingest now prefers real markdown extraction over metadata stubs, with pdftotext plus Python-library fallback guidance. Collection ingestion now covers CSV/TSV/JSON/JSONL message archives as per-message markdown sources and Internet Archive CDX inventories as readability-to-markdown Wayback snapshot imports.

v0.6.0 — Collection ingestion for external wikis and spec repos. Added first-class /wiki:ingest-collection for bounded upstream corpora such as Git document repositories, BIP-style proposal sets, MediaWiki XML dumps, and MediaWiki API sites. Collection imports now create a raw/repos/ manifest plus immutable child sources with upstream revision metadata, while compilation stays synthesized instead of copying another wiki wholesale. The router now detects bulk import intent, and lint/schema docs recognize collection provenance fields and manifest coverage exemptions.

v0.5.0–v0.5.1 — Truth-seeking umbrella audit & durable session provenance. /wiki:audit combines the wiki-only librarian pass with output drift checks, provenance review, and fresh research when local evidence is not enough, while the follow-up provenance work makes multi-round research replayable: research keeps .session-events.jsonl and .session-checkpoint.json, audit records its own provenance milestones, and query --resume can fall back to the checkpoint trail when no active session is running. /wiki:librarian stays available as the focused tool for keeping the wiki/ layer in check.

v0.4.4 — Faster Codex first @wiki load. The generated Codex skill now ships as a smaller router prompt instead of loading the full workflow prose up front. Shared Claude/OpenCode behavior stays unchanged, but Codex first-hit prompt expansion is materially smaller and faster.

v0.4.3 — Codex first-class install fixes. Codex packaging, bootstrap, and verification now line up with the current @wiki plugin layout. The generated Codex mirror lives under plugins/llm-wiki/skills/wiki/, references are validated as copied files for marketplace installs, and the local bootstrap flow now uses the real llm-wiki marketplace name instead of the stale llm-wiki-local alias.

v0.4.2 — Config-first hub resolution & Codex marketplace install. Hub resolution now checks ~/.config/llm-wiki/config.json first, falling back to ~/wiki only when no config exists. Fixes sandbox permission errors in nono where ~/wiki isn't an allowed path. Codex plugin installable directly from GitHub via codex plugin marketplace add nvk/llm-wiki. References changed from symlink to real copy so Codex marketplace caching works. Nono docs updated with per-runtime profiles and $HOME/.codex r+w requirement for Codex plugin install.

Claude Code (native plugin):

claude plugin install wiki@llm-wiki

OpenAI Codex (marketplace plugin):

Install from GitHub:

codex plugin marketplace add nvk/llm-wiki
# Then open /plugins in Codex, enable "LLM Wiki", and use @wiki

Install from a local checkout with the managed bootstrap helper:

./scripts/bootstrap-codex-plugin.sh --scope user --verify

Or register the local checkout manually:

codex plugin marketplace add /absolute/path/to/llm-wiki

Canonical explicit invocation:

@wiki research "hardware wallet threat models"
@wiki ingest https://example.com/article
@wiki audit --project coldcard-threat-model
@wiki ll "codex plugin install gotchas"

Upgrade:

codex plugin marketplace upgrade llm-wiki

Remove:

codex plugin marketplace remove llm-wiki

Troubleshooting:

• After installing the marketplace, open /plugins in Codex and enable "LLM Wiki" — first install requires the interactive enable step.
• @wiki is the canonical explicit entry point in Codex. Natural-language wiki requests can still auto-activate the skill.
• Restart Codex after changing config if an existing session does not pick up the new plugin state.
• If you run Codex under a sandbox wrapper like nono, see Nono Sandbox Permissions — Codex needs r+w to $HOME/.codex for plugin install.

OpenCode (instruction file):

Add to your opencode.json (project-level or ~/.config/opencode/.opencode.json for global):

{
  "instructions": ["https://raw.githubusercontent.com/nvk/llm-wiki/master/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md"],
  "permission": {
    "external_directory": {
      "~/.config/llm-wiki/**": "allow",
      "~/Library/Mobile Documents/com~apple~CloudDocs/wiki/**": "allow"
    }
  }
}

OpenCode fetches the URL fresh on every session start — no manual updates needed. If you prefer a local copy instead:

curl -sL https://raw.githubusercontent.com/nvk/llm-wiki/master/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md > ~/.config/opencode/AGENTS.md

The external_directory permission is required because the wiki hub lives outside the project directory. Set the paths to match your hub location. Alternatively, use --local mode (.wiki/ in the project) to skip permissions entirely.

Web search requires export OPENCODE_ENABLE_EXA=1.

Pi (instruction file — best for local models):

Pi's minimal system prompt (~1K tokens) leaves room for the full wiki skill on 32K context local models.

pi --instructions path/to/llm-wiki/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md

With a local llama-server backend (no cloud API needed):

OPENAI_BASE_URL=http://127.0.0.1:8080/v1 OPENAI_API_KEY=local \
  pi --instructions path/to/llm-wiki/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md

Pi uses the same OpenCode skill file — no separate packaging needed.

Any LLM Agent (idea file):

# Copy AGENTS.md into your agent's context or project root
cp AGENTS.md ~/your-project/AGENTS.md

The AGENTS.md file contains the complete wiki protocol as a single portable document — works with any LLM agent that can read/write files and search the web.

Claude Code is the principal user. Keep one shared behavior layer and thin packaging layers per runtime:

• claude-plugin/ is the primary distribution target and UX surface.
• claude-plugin/skills/wiki-manager/ is the behavioral source of truth.
• plugins/llm-wiki/skills/wiki/ is the generated Codex packaging target behind @wiki.
• plugins/llm-wiki-opencode/ is the OpenCode and Pi packaging target.
• .agents/plugins/marketplace.json makes the Codex plugin installable from this repo.
• AGENTS.md is the portable single-file protocol for any other LLM agent.

Supported clients:

Both runtime mirrors are generated, not hand-maintained. Rebuild from the Claude source of truth:

./scripts/sync-codex-plugin.sh      # regenerates plugins/llm-wiki/
./scripts/sync-opencode-plugin.sh   # regenerates plugins/llm-wiki-opencode/

Each sync script:

• copies claude-plugin/skills/wiki-manager/SKILL.md into the target tree and reapplies a small list of runtime-specific wording patches
• copies references/ from the Claude source — references are runtime-neutral and shared verbatim (previously a symlink, now a real copy so Codex marketplace caching works)
• (Codex only) recreates agents/openai.yaml for Codex UI metadata and syncs the plugin version

Drift is caught by ./tests/test-codex-sync.sh and ./tests/test-opencode-sync.sh, which run the sync scripts and fail (with self-healing fix instructions) if the generated directories differ from HEAD.

Practical rule: design workflows first for Claude commands and behavior, but keep the underlying knowledge model and references runtime-neutral. Runtime wrappers adapt invocation and metadata, not wiki logic.

If you run any AI coding agent inside a nono sandbox, the wiki needs filesystem access beyond the default profile.

{
  "extends": "claude-code",
  "policy": {
    "add_allow_read": [
      "$HOME/.config/llm-wiki"
    ],
    "add_allow_readwrite": [
      "$HOME/Library/Mobile Documents/com~apple~CloudDocs/wiki"
    ]
  }
}

Replace "extends": "claude-code" with "opencode" for OpenCode.

Codex needs r+w to its own $HOME/.codex directory for plugin install, marketplace cache, state, and skill registration:

{
  "extends": "codex",
  "policy": {
    "add_allow_read": [
      "$HOME/.config/llm-wiki"
    ],
    "add_allow_readwrite": [
      "$HOME/.codex",
      "$HOME/Library/Mobile Documents/com~apple~CloudDocs/wiki"
    ]
  }
}

As of v0.4.2, hub resolution checks ~/.config/llm-wiki/config.json first and only falls back to ~/wiki when no config exists. If your wiki lives on iCloud or any non-default path, set the config and you don't need $HOME/wiki in the sandbox at all:

# Set once — agents will resolve from config, never touch ~/wiki
/wiki config hub-path "~/Library/Mobile Documents/com~apple~CloudDocs/wiki"

If you prefer ~/wiki as a symlink to iCloud, nono's Seatbelt follows symlinks — the target path must be allowed, not the symlink itself.

Without the right permissions, Seatbelt silently blocks file access — reads return empty, writes disappear, and the plugin looks broken with no error messages. Use nono why to diagnose:

nono why --path ~/.config/llm-wiki --op read
nono why --path ~/Library/Mobile\ Documents/com~apple~CloudDocs/wiki --op readwrite

OpenCode also needs the external_directory permission in opencode.json (see Install) — nono and OpenCode have independent sandboxes that both need the same paths allowed.

Agents and sandboxed sessions should use GitHub CLI web login with HTTPS git transport, not SSH. This avoids SSH host-key prompts and known_hosts writes inside nono:

gh auth login --web --git-protocol https
gh auth setup-git

Claude Code — if claude plugin update pulls the latest correctly:

git -C ~/.claude/plugins/marketplaces/llm-wiki remote set-url origin https://github.com/nvk/llm-wiki.git
claude plugin update wiki@llm-wiki
# Restart Claude Code to apply

If the update command doesn't pick up the new version (stale marketplace cache), sync manually from the repo:

# Clone or pull the latest
git clone https://github.com/nvk/llm-wiki.git  # or: git -C ~/llm-wiki pull

# Sync plugin files to Claude Code's plugin cache
REPO=~/llm-wiki/claude-plugin
DEST=~/.claude/plugins/cache/llm-wiki/wiki
VERSION=$(grep '"version"' "$REPO/.claude-plugin/plugin.json" | grep -o '[0-9.]*')
rm -rf "$DEST"/*
mkdir -p "$DEST/$VERSION"
cp -R "$REPO/.claude-plugin" "$REPO/commands" "$REPO/skills" "$DEST/$VERSION/"

# Restart Claude Code to apply

Codex — upgrade from the marketplace:

codex plugin marketplace upgrade llm-wiki

OpenCode — if using the GitHub URL in instructions, updates are automatic (fetched every session). If using a local copy:

curl -sL https://raw.githubusercontent.com/nvk/llm-wiki/master/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md > ~/.config/opencode/AGENTS.md

AGENTS.md — just pull the latest and replace:

curl -sL https://raw.githubusercontent.com/nvk/llm-wiki/master/AGENTS.md > ~/your-project/AGENTS.md

Check your installed version:

• Claude Code: look for the version in /wiki status output or check ~/.claude/plugins/installed_plugins.json
• Codex: run ./scripts/verify-codex-plugin.sh --scope project (or --scope user) and confirm the resolved skill path points at this repo
• If the verify script reports PENDING, finish the first-time enable in /plugins and rerun it

New to a topic? One command, from anywhere:

/wiki:research "gut microbiome" --new-topic --min-time 1h

Creates a topic wiki, launches parallel agents, and keeps researching for an hour — drilling into subtopics each round finds. Come back to a fully compiled wiki.

/wiki:research "nutrition" --new-topic            # Create wiki + research in one shot
/wiki:research "gut-brain axis" --wiki nutrition   # Add more research to existing wiki
/wiki:research "fasting" --deep --min-time 2h     # 8 agents, keep going for 2 hours
/wiki:research "keto" --retardmax                 # 10 agents, max speed, ingest everything
/wiki:research "What makes long form articles go viral?" --new-topic  # Question → decompose → playbook
/wiki:thesis "fiber reduces neuroinflammation via SCFAs"  # Thesis-driven: evidence for + against → verdict
/wiki:thesis "cold exposure upregulates BDNF" --min-time 1h  # Deep thesis investigation
/wiki:query "How does fiber affect mood?"         # Ask the wiki
/wiki:query "compare keto and mediterranean" --deep  # Deep cross-referenced answer
/wiki:query --resume                              # Where did I leave off?
/wiki add https://example.com/article             # Fuzzy router detects URL → ingest
/wiki what do we know about CRISPR?               # Fuzzy router detects question → query
/wiki:ingest https://example.com/article          # Manually ingest a source
/wiki:ingest --inbox                              # Process files dropped in inbox/
/wiki:ingest-collection https://github.com/bitcoin/bips --wiki bitcoin  # Bulk import spec repos
/wiki:ingest-collection https://dump.bitcoin.it/dump_20260429_en.xml.bz2 --wiki bitcoin  # Import MediaWiki dumps
/wiki:ingest-collection messages.csv --adapter csv-messages --wiki bitcoin  # Split message archives
/wiki:ingest-collection "https://example.com/*" --adapter wayback-cdx --from 20100101 --to 20200101  # Import archived snapshots
/wiki:compile                                     # Compile any unprocessed sources
/wiki:audit --project gut-brain-playbook          # Truth-seeking audit across outputs + wiki + fresh research
/wiki:output report --topic gut-brain             # Generate a report
/wiki:output slides --retardmax                   # Ship a rough slide deck NOW
/wiki:assess /path/to/my-app --wiki nutrition     # Gap analysis: repo vs wiki vs market
/wiki:lint --fix                                  # Clean up inconsistencies

All commands accept --wiki <name> to target a specific topic wiki and --local to target the project wiki. Commands that generate content (query, output, plan) also accept --with <wiki> to load supplementary wikis as cross-wiki context — e.g., --with article-writing applies writing craft knowledge when generating output from a domain wiki.

/wiki:librarian is the focused wiki-maintenance tool. /wiki:audit is broader and may perform fresh research to decide whether the current knowledge or generated outputs are still trustworthy.

~/wiki/                                 # Hub — lightweight, no content
├── wikis.json                          # Registry of all topic wikis
├── _index.md                           # Lists topic wikis with stats
├── log.md                              # Global activity log
└── topics/                             # Each topic is an isolated wiki
    ├── nutrition/                      # Example topic wiki
    │   ├── .obsidian/                  # Obsidian vault config
    │   ├── inbox/                      # Drop zone for this topic
    │   ├── raw/                        # Immutable sources
    │   ├── wiki/                       # Compiled articles
    │   │   ├── concepts/
    │   │   ├── topics/
    │   │   └── references/
    │   ├── output/                     # Generated artifacts
    │   ├── _index.md
    │   ├── config.md
    │   └── log.md
    ├── woodworking/                    # Another topic wiki
    └── ...

The hub is just a registry — no content directories, no .obsidian/. All content lives in topic sub-wikis with isolated indexes and articles. Queries stay focused. The multi-wiki peek finds overlap across topics when relevant.

1. Research a topic — parallel agents search the web, ingest sources, and compile articles in one command
2. Ingest additional sources — URLs, files, text, tweets (via Grok MCP), or bulk via inbox
3. Compile raw sources into synthesized wiki articles with cross-references and confidence scores
4. Query the wiki — quick (indexes), standard (articles), or deep (everything)
5. Lessons learned — extract knowledge from the current session (errors, fixes, gotchas) into the wiki
6. Assess a repo against the wiki — gap analysis: what aligns, what's missing, what the market offers
7. Lint for consistency — broken links, missing indexes, orphan articles
8. Output artifacts — summaries, reports, slides — filed back into the wiki

• One topic, one wiki — each research area gets its own sub-wiki with isolated indexes. No cross-topic noise.
• Parallel research agents — 5 standard, 8 deep, 10 retardmax. Each agent searches from a different angle.
• _index.md navigation — every directory has an index. Claude reads indexes first, never scans blindly.
• Articles are synthesized, not copied — they explain, contextualize, cross-reference.
• Raw is immutable — once ingested, sources are never modified.
• Multi-wiki aware — queries peek at sibling wiki indexes for overlap.
• Dual-linking — both [[wikilinks]] (Obsidian) and standard markdown links on every cross-reference. Works everywhere.
• Confidence scoring — articles rated high/medium/low based on source quality and corroboration.
• Structural guardian — auto-checks wiki integrity after operations, fixes trivial issues silently.
• Activity log — log.md tracks every operation, append-only, grep-friendly.
• Zero dependencies — runs entirely on built-in tools (Claude Code, OpenCode, or Codex).

Smart input detection — /wiki:research auto-detects whether you're passing a topic or a question:

Question mode produces a playbook (actionable output artifact) and suggests testable theses derived from the findings.

Modifiers (combine with any mode):

# The full combo — new topic, 2 hours of deep research, from anywhere
/wiki:research "CRISPR gene therapy" --new-topic --deep --min-time 2h

Retardmax mode is inspired by Elisha Long's retardmaxxing philosophy — act first, think later. The antidote to analysis paralysis. Works for both /wiki:research and /wiki:output.

Unlike open-ended research, /wiki:thesis starts with a specific claim and evaluates it:

/wiki:thesis "intermittent fasting reduces neuroinflammation via glymphatic upregulation"

How it works:

1. Decomposes the thesis into key variables, testable predictions, and falsification criteria
2. Launches parallel agents — but each agent has the thesis as a FILTER. Irrelevant sources get skipped (this prevents bloat)
3. Agents are split: supporting, opposing, mechanistic, meta/review, adjacent — balanced by design
4. Compiles evidence into wiki articles + a thesis file with evidence tables
5. Delivers a verdict: supported / partially supported / contradicted / insufficient evidence / mixed

Anti-confirmation-bias: When using --min-time, Round 2 automatically focuses harder on the WEAKER side of the evidence. If Round 1 found mostly supporting evidence, Round 2 hunts for counter-evidence.

The thesis is the bloat filter. Sources that don't relate to the claim's variables don't get ingested. Higher skip rate = tighter focus.

Every cross-reference in the wiki uses dual-link format:

[[gut-brain-axis|Gut-Brain Axis]] ([Gut-Brain Axis](../concepts/gut-brain-axis.md))

The wiki is not locked into any tool:

• Obsidian reads the [[wikilink]] — graph view, backlinks panel, quick-open
• Claude Code follows the standard (relative/path.md) link
• GitHub/any markdown viewer renders the standard link as clickable
• No viewer at all — plain markdown, readable in any text editor

Each topic wiki has its own .obsidian/ config and can be opened as an independent vault:

open ~/wiki/topics/nutrition/     # Open in Obsidian — focused graph for one topic

The hub (~/wiki/) has no .obsidian/ to avoid nested vault confusion. If you want a cross-topic view, open ~/wiki/ manually and let Obsidian create its own config.

What works out of the box:

• .obsidian/ config created on init with sane defaults
• [[wikilinks]] power the graph view
• aliases in frontmatter enable search by alternate names
• tags in frontmatter are natively read
• inbox/ works as a drop zone in both Obsidian and the CLI

Claude Code is the compiler. Obsidian is an optional viewer.

• Andrej Karpathy — the LLM wiki concept and idea file
• Elisha Long — retardmaxxing philosophy (act first, think later)
• tobi/qmd — recommended local search engine for scaling beyond ~100 articles
• rvk7895/llm-knowledge-bases — prior art in Claude Code wiki plugins

MIT License. Copyright (c) 2026 nvk.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files, to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, subject to the following conditions: the above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.