# llm-wiki-kit

Keep project decisions and debugging knowledge available across coding sessions. `llm-wiki-kit` connects Codex, Claude Code, and opencode to a project-local Markdown wiki: hooks bring relevant context into your work, and your agent reviews reusable findings and maintains the pages.

Routine turns stay transient. Durable knowledge remains in files you can read, edit, and version. The kit does not provide a model or make model calls; your coding agent keeps using its configured provider.

## Quick Start

Requires Node.js 20 or later.

```bash
npm install -g llm-wiki-kit@latest
llm-wiki install --workspace /path/to/project --profile standard
llm-wiki doctor --workspace /path/to/project
```

Restart active Codex or Claude Code sessions after installation. For opencode, start a new session to load its plugin bridge.

```powershell
npm install -g llm-wiki-kit@latest
llm-wiki install --workspace C:\path\to\project --profile standard
llm-wiki doctor --workspace C:\path\to\project
```

Native Windows uses the npm-generated `llm-wiki.cmd` shim. Use WSL2 when your repository and tools already run in Linux.

## What it does

- **Brings useful context into the work.** Hooks search the project's Markdown knowledge and provide compact context to supported agents.
- **Keeps durable knowledge reviewable.** The agent decides which conclusions are reusable and writes them to Markdown; routine answers are not automatically promoted.
- **Finds pages with evidence.** Search returns section references and line ranges. Lint and retrieval eval help maintain the wiki.
- **Keeps project data local.** Wiki files live under `llm-wiki/` in the project. The kit itself does not need a model API key; install and update checks may contact npm, and your agent may contact its configured model provider.

You can ask questions naturally, for example: “What should we check before a release?” or “Document the cause of this error and how we verified the fix.” The agent can search existing pages first, then update a curated page when the finding is worth keeping.

<p align="center">
  <img src="https://cdn.jsdelivr.net/npm/llm-wiki-kit@0.26.1/docs/assets/product-overview.svg" alt="Coding signals move through agent review into a local Markdown wiki, which supplies relevant context to later sessions" width="960" />
</p>

## Supported agents and languages

| Agent | Integration |
|---|---|
| Codex | Native hooks and bundled skills |
| Claude Code | Native hooks and bundled skills |
| opencode | Managed plugin bridge; some hook events are experimental. [Details](docs/integrations/opencode.md) |

Korean and English prompts are supported. Hook guidance follows the current prompt language where it is clear; commands, paths, identifiers, and original error text stay unchanged. The [manual](docs/manual.md#language-behavior) describes fallback behavior and unsupported languages.

## Install choices

`install` sets up the project wiki and the selected agent integrations. By default, Codex and Claude Code hooks and skills and the opencode bridge are managed. Use `--providers codex,claude` to choose integrations explicitly, `--providers none` for the core wiki only, or `--no-skills` to skip bundled skills. Normal updates preserve edited integrations; `llm-wiki update --replace-hooks` explicitly refreshes only content with proven kit ownership. See the [manual](docs/manual.md#install) for provider selection, setup, and update behavior.

## Documentation

- [Full manual](docs/manual.md) — commands, search, page metadata, maintenance, and release verification.
- [Security](docs/security.md) — redaction, local data, exports, and filesystem boundaries.
- [Troubleshooting](docs/troubleshooting.md) — installation, provider, search, and release issues.
- [Operations](docs/operations.md) — operator and maintainer procedures.
- [Concepts](docs/concepts.md) — the wiki lifecycle and retrieval model.

## Development

```bash
git clone <your-git-url>/llm-wiki-kit.git /path/to/llm-wiki-kit
cd /path/to/llm-wiki-kit
npm install
node --test
```

Before publishing a release, verify the same candidate tarball in isolated Linux and native Windows environments. Do not install a prepublish candidate globally or point it at a real workspace. The [release section of the manual](docs/manual.md#release-verification) describes the manifest and verification steps.
