A proposal is a validated typed diff — reviewed on the Proposals tab, never deployed by itself
A completed, metric-bearing diagnosis can mint a proposal: a typed diff against an existing draft configuration resource (today, the domain performance policy), validated with that resource's real validation function before it can be proposed at all.
Generating one#
Press Generate proposal on a completed diagnosis row (toast: Proposal generated — see it on the Proposals tab.). The API is strict about prerequisites: Only a diagnosis with at least one ranked cause (status "completed") can produce a proposal., A proposal requires a diagnosis triggered by a specific RUM regression metric (lcp, inp, cls, or ttfb)., and The diagnosis has no ranked causes to propose a configuration change for.
The Proposals tab#
/autopilot/proposals, eyebrow Typed configuration diffs. Its hint states the contract: proposals are validated through the resource's own validation (a failing one is Rejected with the reason captured, never silently applied) and Nothing here is deployed to live traffic: replay/canary experiments and promotion are separate, not-yet-built work. Columns: Domain | Status | Target | Confidence | Estimated impact | Created, statuses Proposed / Rejected / Draft, confidence as n%, impact as LCP ~n% with an estimate badge. Details expands Typed diff, Affected traffic: n%, Risks, Validation plan:, Rollback condition:. Empty: No proposals yet. Generate one from a completed diagnosis on the Diagnoses tab.
From proposal to traffic — honestly#
The experiments API exists server-side (POST /v1/domains/:id/autopilot/experiments with evaluate, promote, rollback, plus auto-rollback; promotion can require manual approval: Promotion requires manual approval unless the automatic-promotion policy is enabled.) but no dashboard tab exposes it yet. The visible, guarded path for advancing a performance-policy change is the separate Performance rollouts screen (plan-gated performance_rollouts.enabled): it advances each domain only when baseline, matched audit, compatibility, and live canary evidence pass, with Refresh evidence, progress steps from 0% to 100%, and Roll back per domain (second screenshot below). Until an experiment UI ships, treat the proposal as reviewed analysis and use rollouts as the place where policy changes earn their promotion.