Lmg

Use lmg for codebase research

Use lmg when answering a focused repository question would otherwise require several searches or produce noisy output. It runs its own search loop and returns only a concise JSON codemap with file and line evidence.

Run it from the repository root:

lmg "Where is session expiry enforced?"

Or identify the repository explicitly:

lmg -C /path/to/repo "Trace how requests become authenticated users"

Ask one concrete investigative question at a time. Include the behavior, symbol, configuration, or execution path you need to understand. Prefer direct shell inspection when the exact file and location are already known.

On success, parse the JSON codemap on stdout:

{
  "summary": "direct answer",
  "nodes": [
    {
      "id": "unique-id",
      "location": "relative/path:line-line",
      "symbol": "optional symbol",
      "description": "fact established by this location"
    }
  ],
  "edges": [
    {"from": "node-id", "to": "node-id", "relation": "calls or data flow"}
  ],
  "usage": {"input": 100, "output": 20, "cache_read": 80, "cache_create": 0}
}

Every edge references node IDs. symbol is optional, and edges is empty when the answer does not need a relationship graph. usage is added by lmg and aggregates provider token counts across every agent round.

lmg automatically supplies repository-root instructions from AGENTS.md, falling back to CLAUDE.md only when AGENTS.md is absent. Callers do not need to include those instructions in the question.

Treat the answer as research evidence, not authorization to modify anything. Before editing or making a high-impact decision, inspect the cited locations directly and reconcile any ambiguity or missing context.

The default sandbox is read-only, has no network access, and hides the user's home and parent environment. Keep it enabled for ordinary research. Use --yolo only when the task is authorized to grant the model the invoking process's filesystem and network access; environment scrubbing and other limits still remain active.

Use --verbose to put tool activity and per-round input, output, cache-read, and cache-create token counts on stderr. Do not substitute --yolo for missing endpoint, API-key, model, or sandbox setup.

Network failures and retryable HTTP responses are retried indefinitely with bounded exponential backoff, so an invocation may keep waiting during a provider outage until it succeeds or the caller terminates it.

Parallel tool calling is enabled by default, allowing the model to request multiple repository inspections in one turn.

Run lmg --skill to print this skill from the installed binary.