llmwiki

---
name: llmwiki
description: |
license: Apache-2.0
---

# llmwiki — your LLM's persistent memory

The commands below are CLI invocations. Execute them with your command-execution primitive, not the one that read this file.

This is the entrypoint. The CLI is the source of truth.

**Always start with `llmwiki --help`** — it shows the active wiki
alias, workspace, and which resolution step matched, so you know exactly
which wiki you're operating on. The same info is also printed when you
run `llmwiki` with no subcommand at all (try it: `llmwiki`).

**Install shell completion once per host** (zero-config TAB-completion
for every command, subcommand, and flag):

```bash
# bash
llmwiki completion bash > ~/.local/share/bash-completion/completions/llmwiki
# zsh
llmwiki completion zsh > "${fpath[1]}/_llmwiki"
# fish
llmwiki completion fish > ~/.config/fish/completions/llmwiki.fish
```

```bash
llmwiki --help                # active wiki + full command reference
llmwiki doctor                # diagnose install + config + NIM
llmwiki config current        # alias + workspace + resolution source
llmwiki skill list            # enumerate every sub-skill (3 categories)
llmwiki skill get <topic>     # load one (version-matched with your binary)
llmwiki <command> --help      # full flag reference
```

If the binary isn't installed: `git clone https://github.com/Git-Fg/llmwiki.git && cd llmwiki && cargo build --release`, then add `./target/release` to your `PATH`. There is no prebuilt artifact or package-registry install.

Sub-skills are NOT installed to disk — they live inside the CLI binary,
served on demand. **Always prefer `skill get` over guessing commands.**

Multi-wiki quick reference:

- `llmwiki config list` — every registered alias
- `llmwiki config prune-stale [--dry-run]` — drop aliases whose path no longer resolves (auto-runs inside `llmwiki doctor`)
- `llmwiki status --all` — one-line health summary per wiki
- `llmwiki use <alias>` — pin this workspace to a specific wiki
  (writes `<workspace>/.llmwiki-cli/state/active-wiki`, gitignored by
  `llmwiki init`)

When in doubt, run `llmwiki doctor` first — catches missing API keys,
NIM connectivity, broken config, orphans in one pass. If you don't know
which wiki the CLI is operating on, run `llmwiki config current`
(pinned to the active alias + resolution source + registry file path).

For the full Config schema: run `llmwiki config show-schema`.

## Sub-Skills (3 Categories)

`llmwiki skill list` shows exactly 3 categories. Load the one that
matches your task:

| Category | Topic | Use when... |
|----------|-------|-------------|
| **Core wiki workflow** | `core` | Search, ingest, config, lint, models, sync (90% of tasks) |
| **OKF bundles** | `okf` | Produce, validate, consume, visualize, status, maintain `.okf/` bundles |
| **System admin** | `admin` | Install skill, shell completion, registry management |

**Quick routing:**

- "Find X in my wiki" / "Search the wiki" → `llmwiki skill get core`
- "Answer a question against my notes" → `llmwiki skill get core`
- "Add a source file to the wiki" → `llmwiki skill get core`
- "Change the embedding model" → `llmwiki skill get core`
- "Check wiki health / broken links" → `llmwiki skill get core`
- "Create / validate an OKF bundle" → `llmwiki skill get okf`
- "Visualize the OKF graph" → `llmwiki skill get okf`
- "Install the skill for my AI agent" → `llmwiki skill get admin`
- "Set up shell completion" → `llmwiki skill get admin`

## OKF (Open Knowledge Format) — v0.4.0+

The `okf` subcommand group manages a `.okf/` bundle of typed concepts
sitting next to the wiki. OKF is opt-in: a workspace without
`.okf/index.md` is unchanged. Concepts and wiki pages share the same
embedder, so `search --type` and `search --source` filter across both.

The §9 permissive-consumption contract guarantees that `consume`,
`visualize`, `embed`, `search`, `doctor`, and `status` never fail on a
malformed bundle — only `okf validate` does, and only under `--strict`.

## When to skip this skill

- **You already know the exact command.** Run `llmwiki <command> --help` directly. The hub is a router; if the routing decision is already made, do not load this skill.
- **You need a one-shot quick lookup.** Run `llmwiki skill list` (and `skill get <topic>` for one sub-skill) directly without loading the hub first.
- **The user asked about a non-wiki memory tool.** This skill is for `llmwiki` and Karpathy-style markdown wikis. For other note systems (Notion, vanilla Obsidian, Apple Notes, etc.) this is the wrong skill.
- **The user wants a generic RAG framework.** llmwiki is a specific tool, not a framework. The OKF subcommand is a typed-bundle extension, not a general RAG library.
- **The user asked about the source code of llmwiki itself.** Use code-reading tools and the GitHub repo. This skill is the user-facing guide, not the developer guide.

## Tips

- **Not sure which skill to load?** Run `llmwiki skill list` to enumerate all 3 categories. Then `llmwiki skill get <topic>` to load one — the content is version-matched with your binary, never write commands from memory.
- **First time?** Start with `llmwiki skill get core` to verify the install, then `llmwiki doctor` for a health check.
- **Stuck on an error?** `llmwiki skill get core` covers the 9 most common symptoms with first-thing-to-try for each (see the Diagnostics section in the core skill).
- **Ingested something and want to find it?** The search workflow is in the core skill. `llmwiki skill get core` to enumerate.
- **Want OKF concepts, not just wiki pages?** `llmwiki skill get okf` shows how to import markdown into a typed `.okf/` bundle.
- **All error messages include "Next:" and "Tip:" lines** with the exact fix and the relevant sub-skill to load. Read them before retrying.