Versions, evals, staged rollout, and rollback

Edit the Agent Team, open Trigger Hub, and choose Compare update on an existing automation. Codelit reruns hosted readiness checks, snapshots the executable permission manifest, creates deterministic golden cases from the run brief, tools, approval gates, evaluations, and evidence-capable writes, then stores a new immutable candidate receipt.

Projects shows active and candidate run pass rate, average cost, average latency, proven action count, golden pass rate, and added or removed permissions. Deterministic checks use no model tokens. Optional judge evaluates at most five declared cases in one 256-token managed request and is charged against the existing managed-AI allowance.

Staging is always an operator decision. Candidate assignment is stable for a delivery identity and capped at half of traffic until promotion.

1. 1Review every added read or write permission.
2. 2Choose 5%, 10%, 25%, or 50%. Delivery identities are assigned deterministically, so retries stay on the same version.
3. 3Inspect real candidate runs, then choose Promote. Codelit never promotes or expands permissions on its own.

Roll back restores the previous immutable workflow, pauses the automation, and halts only unfinished runs pinned to the replaced version. Completed runs and action checkpoints remain evidence; Codelit does not replay their external writes. Review recovery, then resume as a separate manual action.