Build on Aktor
Evals
Expected behaviour, checked by running the workspace for real; fails CI on regressions.
name: urgent leak is triaged
workspace: maintenance-desk
input: "Ticket created: water is pouring through the ceiling."
expect:
status: Completed
summary_contains: [urgent]
records:
Ticket: { at_least: 1, where: { priority: urgent } }
The Evals tab starts a real run for each case, waits, and checks the status, the summary, and
the records the run wrote. aktor-cli test will do the same from CI once the CLI is available. With
the Mock model, evals check the plumbing; with a real model, they check behaviour.