Stop a run, choose the fleet's model, see the spend
A per-run Stop button with a distinct Stopped status, fleet-wide model and effort selection from the admin console, and admin cost pages.
After this entry you'll know the three new levers operators have over the fleet. The biggest one: you can finally stop a run. Until now a run that was chewing the wrong repo or burning spend could only be watched until it finished or crashed.
What changes for you
- Stop a run. Every running run's page now has a Stop button. One confirm and the run halts with a distinct Stopped status, recorded against your name, so a deliberate stop reads differently from a crash in every list and report. Stopping cascades: any child bees the run dispatched and any background E2E verification stop with it, and a reconciler makes sure no task keeps running unseen after its run has ended.
- Choose the model and effort fleet-wide. The admin console now lets you pick which Claude model and effort level every swarm run uses, from a curated catalog, with the change audited and applied without a deploy. Previously the fleet silently ran whatever the underlying tooling defaulted to, which nobody had chosen.
- See where the money goes. New admin cost pages break spend down by model, repo and run over a chosen period, each run shows its own cost, and the runs table gained a Triggered by column so you can see who or what started every run you are paying for.
Cost detail stays admin-only; see the admin guide for where these live.
E2E verification is now visible from dispatch to verdict
A live lifecycle card in the run view, a verifying chip on Done runs, and failed verifications land in Needs you with a one-click retry.
Secrets are harder to reach and worth less if reached
Secret values are now redacted from transcripts and timelines, the agent's shell sees a minimal environment, and GitHub tokens are scoped to one repo.