Atlas

atlas

Harness engineering. Harness your repo before it harnesses you.

Claude Code plugin that keeps docs/ as a 1-to-1 mirror of your codebase. Two slash commands, one skill, one subagent, one hook. Minimal footprint.

Install

In any Claude Code session:

/plugin marketplace add lzvsk/atlas-pack
/plugin install atlas@atlas-pack
/reload-plugins

Verify with /plugin — you should see atlas as enabled.

Naming:

• atlas-pack — the marketplace (this repo).
• atlas — the plugin inside, used in slash commands (/atlas:init, /atlas:sync).

Alternative — include via another marketplace

If you maintain a collection marketplace, add this entry to its marketplace.json:

{
  "name": "atlas",
  "source": {
    "source": "github",
    "repo": "lzvsk/atlas-pack",
    "ref": "main"
  }
}

Usage

/atlas:init              interactive bootstrap: AGENTS.md + docs/ + full mirror of the code tree
/atlas:sync              drift audit + fix (one confirmation for the whole edit list)
/atlas:sync --check      report only, no edits

After install, it works mechanically:

• The atlas skill auto-activates in any repo with AGENTS.md + docs/ at the root.
• A PostToolUse hook on Edit/Write/MultiEdit resolves the edited path via mirrors:-frontmatter and tells the agent which doc to update.
• The atlas-auditor subagent runs heavy drift scans in a fresh context.

Principle

Optimize the repository for agent reading, not human reading. If the agent can't see it, it doesn't exist. Docs are the versioned source of truth; code is the implementation. Disagreement = a bug in docs.

Methodology adapted from OpenAI harness engineering (Ryan Lopopolo, 2026-02-11).

Documentation

• Agent entry point — AGENTS.md (navigation into docs/ lives there)
• Plugin architecture (stack, components, flows, boundaries) — docs/ARCHITECTURE.md
• Top-level docs index — docs/ABOUT.md
• Code mirrors — docs/{commands,skills,agents,hooks}/ (each with ABOUT.md + per-file descriptions)

License

WTFPL — see license.txt.