Cairn — Documentation skill for Claude Code

Cairn

Cairn turns a plan document a human wrote into a graph of verified, concurrent work.

Cairn generates workflows for **Dagu**, an external DAG engine; a `dagu` binary on PATH is what runs them. "The engine" throughout these documents means Dagu. Cairn's own code is Python 3 with no third-party imports.

**Type `/cairn`, then say what you want.** Cairn is a skill and it opens only when it is asked for by name: author a workflow from a plan, change one, run it, put it on a schedule, read what a past run did, or ask what a workflow would do, what a verdict means, or why a step was excluded. Six capabilities, all reachable by asking, and none of them needing a second surface — though where the engine's own view is the better answer, the skill links to it and says what it will never tell you.

Nothing else opens it. A sentence that describes the work without naming Cairn — "run my plan", "schedule this nightly" — reaches whatever else is installed, and Cairn stays silent rather than competing for it. That is the price of a tool that spends money and commits to a repository: it is entered on purpose, and naming it is itself the first consent.

A run is never a default. It happens on an unambiguous instruction or an accepted offer, the offer states what the run costs before you answer it, and one acceptance authorises exactly one execution. [SKILL.md](SKILL.md) is where those rules live.

**What is under it:** derive a plan graph from a plan's documents, validate it, and read the parse report; answer the authoring conversation for a step whose plan named no assertion; run a step under the committed-marker protocol so a resumed run skips work that is already done, with the verify gate deciding what gets recorded; converge worktrees, commit, land a wave's branches on the parent branch one at a time, and take and give back the repository's run lock; generate a whole workflow file for a plan, refusing to write one the engine would run into a silent failure; and read ba