Active Runs Hardening Signoff Inventory

Date: 2026-03-13

This inventory is the shared coverage list for provider-backed browser signoff of the active runs board.

Runtime Assumptions

  • Docker stack is running locally.
  • Backend and worker use the standard LiteLLM client path.
  • A real provider gateway is reachable from containers at http://host.docker.internal:1234.
  • Baseline model: qwen3.5-0.8b.

User-Visible Claims To Sign Off

  • streaming success is visible on the current item card
  • waiting, streaming, quiet, and buffered wait remain visually distinct
  • canceling an in-flight run first shows cancellation requested
  • worker interruption resumes as recovery, not fake user cancel
  • the active runs board, archive/history list, and inspector converge on the same truth
  • desktop and mobile layouts remain readable without horizontal dependency

Coverage Matrix

Claim / ControlStates to inspectFunctional checkExpected evidence
Launch runpending -> runninglaunch a provider-backed run from Runsdesktop active-runs screenshot
Current item activitywaiting, streaming, quietobserve timer and activity badge transitionscurrent-item screenshot
Cancel actioncancellation requested -> canceledcancel while request is in flightcancel-requested screenshot
Worker interruption / resumeretry scheduled -> resumedrestart worker and confirm recoveryresumed screenshot
Archive/history convergencerunning / terminalcompare board, archive, and inspectorconvergence screenshot
Mobile layoutnarrow viewportrevisit board on phone-sized viewportmobile screenshot

Evidence To Capture

  • Desktop Runs active board while streaming
  • Desktop active board with cancellation requested
  • Desktop active board after worker resume / recovery
  • Desktop dashboard or archive showing converged terminal state
  • Mobile active board proving no horizontal dependency

Exit Criteria

  • every claim above has matching functional evidence
  • no clipping, contradictory status text, broken assets, or SSE hydration gaps are observed
  • if buffered fallback does not naturally reproduce, automated regression remains the hard gate