Why every AI action in Loop carries a receipt
Auto-Pilot proposes board changes from commits and chat. It never applies them silently, and each proposal shows the trigger that produced it.
The reason boards go stale is not laziness. It is that updating a card is a second, manual copy of information that already exists in a commit, a pull request or a chat message. Auto-Pilot closes that gap without taking the keyboard away from you.
Rules first, model second
A pull request on a branch called PAY-12 is not an ambiguous signal — a regular expression handles it, deterministically, at 94% confidence. The model is only invoked for the genuinely ambiguous case: somebody says "I am stuck on the payment API" without naming a task. It is then given the fifteen most recently active open tasks and asked to pick one, with a confidence score and permission to answer null.
Proposals, not writes
Every outcome becomes a card in the suggestions inbox showing the trigger, the quoted evidence, the confidence and the exact field change. Accept writes to the task and to the audit log. Reject closes it. A project owner can switch on auto-apply, and even then only rule-based suggestions above 90% qualify — model-matched ones are capped below that line deliberately.
What this buys you on stage
A wrong suggestion is not a failure, it is a demonstration of the accept step. That is a much better property than a system that quietly moved the wrong card and nobody noticed until Friday.