Skip to content

commands.md

Latest commit

 

History

History
129 lines (103 loc) · 5.94 KB

commands.md

File metadata and controls

129 lines (103 loc) · 5.94 KB
Raw

Commands & Wiki Structure

Full command reference, plus where files land and how freshness tracking works.

← Back to README

Commands

Command Description
init [path] Create llmwiki.yaml and optionally wire up graymatter hooks
init [path] --hooks Also install a Git pre-commit hook that checks for stale docs
ingest <path> Scan a project and generate/update wiki entries
check [path] Report which wiki entries are stale relative to source code
check --json / --exit-code / --files Machine output / CI exit code / restrict to given files
ingest <path> --no-memory Ingest without memory recall/storage
absorb <path> Extract session facts into memory (near-zero token cost)
absorb <path> --note "..." Absorb with an explicit session note
absorb <path> --note-stdin Absorb note piped from stdin (used by the Claude Code hook)
absorb-drain Drain queued absorb sessions (created when the memory DB was busy)
materialize <project> Rebuild wiki from accumulated memory facts (~10× cheaper than ingest)
list List all tracked projects
context <project> Print wiki context (pipe into CLAUDE.md)
query "<question>" Ask a question across all wiki entries (LLM-backed)
mcp Run an MCP server exposing the wiki to agents over stdio (see Integrations)
docs <path> Generate/update project documentation from wiki + memory
docs <path> --write Write the updated doc to the project directory
docs <path> --target FILE Update a specific file (default: README.md)
index [customer] Generate client and project index files
link Add cross-reference links between wiki files
remember --project <name> "<fact>" Store a fact in memory
recall "<query>" Recall facts from memory
recall --project <name> "<query>" Recall facts for a specific project
hook install Install llmwiki as a Claude Code plugin
hook uninstall Remove the Claude Code plugin
hook status Check if the Claude Code plugin is installed

Wiki Structure

~/llmwiki/wiki/
├── _index.md                              # global project listing
├── clients/
│   ├── acme/
│   │   ├── _index.md                      # client executive summary + C4 diagram
│   │   ├── billing-api.md                 # single-service project
│   │   ├── notification-service.md
│   │   └── ecommerce/
│   │       ├── _index.md                  # project overview + service table
│   │       ├── cart-service.md            # per-service wiki
│   │       ├── payment-service.md
│   │       └── ...
│   └── globex/
│       ├── _index.md                      # client executive summary
│       └── platform/
│           ├── _index.md
│           ├── auth-service.md
│           └── ...
├── personal/
│   └── my-tool.md
└── opensource/
    └── some-lib.md

Plain markdown with YAML front matter. No proprietary format. Works with git, grep, and any text editor.

AI coding integration

Inject wiki context directly into CLAUDE.md (or any file) with marker-based replacement:

<!-- llmwiki:start -->
<!-- llmwiki:end -->
llmwiki context my-project --inject CLAUDE.md

Your AI assistant starts every session with Domain, Architecture, Services, and Flows already in context. No more "can you look at the codebase and figure out what this does."

Change tracking & freshness

Documentation drifts the moment code changes. llmwiki tracks which source files each wiki entry describes, and tells you when those files have changed but the docs haven't.

At ingest time, each wiki entry's front matter gains an llmwiki_tracking block:

llmwiki_tracking:
  area: internal/auth
  files:
    - internal/auth/handler.go
    - internal/auth/middleware.go
  hash: a3f9bc1d2e4f6a8b
  cluster_method: git-cochange
  updated_at: "2026-05-25"

llmwiki figures out areas from git co-change history: files that keep landing in the same commits get grouped together (union-find clustering, 30% co-occurrence threshold). On projects with fewer than 20 commits it falls back to top-level directory heuristics. The hash is a SHA256 over git ls-tree HEAD output for each tracked file, so it changes when file contents change. Timestamps don't enter into it.

Run the check anytime:

llmwiki check                       # report fresh/stale entries for the current project
llmwiki check --json                # machine-readable output
llmwiki check --exit-code           # exit 1 if anything is stale (for CI)
llmwiki check --files a.go,b.go     # restrict to areas containing these files
✓ clients/acme/billing-api.md   fresh   area: internal/auth   updated: 2026-05-25
✗ clients/acme/billing-api.md   STALE   area: internal/billing

Staleness shows up through three paths:

  • Manual / agent — run llmwiki check yourself, or register it as a slash command so an AI agent runs it before handing work back.
  • Git pre-commit hook — llmwiki init --hooks installs a .git/hooks/pre-commit that blocks the commit (--exit-code) when staged files belong to a stale area. If the doc update is deliberately deferred, git commit --no-verify gets you through. An existing pre-commit hook is left alone.
  • AI session Stop hook — the graymatter Stop hook (see below) runs a non-blocking llmwiki check on the files touched during the session and records the result in memory.

Client & project indexes

For consultants and agencies managing multiple clients:

llmwiki index acme    # executive summary across all acme projects

Generates a client-level _index.md with executive summary, C4 diagram, architecture overview, and a projects table — useful for onboarding, handoffs, and architecture reviews.