promote-digest

Promote digest

Purpose

The whole model runs on one invariant: the agent proposes, an authorized human decides, and a deterministic controller applies that exact revision. Everything you mine lands in staged/; nothing becomes accepted until a human says so. That gate is what keeps an autonomous agent from quietly rewriting the company's model of itself. Approval must work directly from chat; requiring the reviewer to open the repository and merge a second pull request makes the queue rot without adding semantic authority.

This skill makes the decision gate cheap enough to use from chat. It turns the staged queue into a digest where every item contains the diff, source, impact, and validation result. The human still decides; the deterministic controller handles application.

It is also a trust-boundary checkpoint. Staged cards may carry facts mined from untrusted inputs — a chat export, a spreadsheet, a connector. The digest is where provenance becomes visible before anything is committed: every item names its source and its source's trust level, so a human can see that a "fact" is really a chat opinion before it gets stamped accepted. You surface that; you never launder it by promoting on your own.

This is a self-initiated skill — you reach for it when the staged queue is worth a human's attention, not because someone asked you to file a report.

When to use

Reach for this skill when:

• A mining pass or capture-loop session has left two or more cards in staged/ and you want them committed.
• A drift-sweep produced status changes, gaps, or new open questions that need a human decision.
• The staged queue has grown to the point where pointing the human at the folder is more friction than a digest would be.
• A human asks "what's waiting?", "anything to approve?", "show me what you staged", or similar.
• A conflict was detected (two sources disagree, a proposed change contradicts an accepted card) and the human needs to choose — the digest is how you put the choice in front of them.