Quickstart: auto-remediation

1. Give it something to work on. A tracked repo is enough to start: the loop runs on repo findings alone. Connecting Sentry adds the production lane, and until you do, the queue says the production lane is empty rather than implying production is clean.
2. Set a floor and a cap. Sensing, triage and scoping always run at cents. Build spend is bounded by a severity floor and a daily dollar cap, so decide what is worth building before anything builds.
3. Work the Today queue. It holds what is built and waiting on you, what is building now with its live spend, and what your policy is holding back.

Approve is a person, every time, and there is no setting that changes that. Approving re-runs the bar server-side before merging: the fix builds, tests pass, a new regression test fails on the commit that caused the defect and passes on the fix, the guard blocks that commit on replay, and CI is green. A miss blocks the merge rather than shipping. Reject takes a reason and discards the diff, the guard and the lesson together.

Nothing merges on its own. A fix that fails twice stops and lands on Today as a finding with both attempts attached, rather than retrying quietly. A guard is derived from the fix, never invented: when no rule, diff pattern or regression test can be derived, the guard check fails with that reason and the record carries no guard source.

The loop gets better the more it can attribute, so pair it with attribution, and let the readiness score feed it findings worth fixing.

Auto-remediation covers all nine stages, the brief that bounds each build, and the five-check bar in full. Combined remediation covers taking several cases as one ordered plan.