Skip to content

Reconnect Cancellation Validation

Matthew Feroz edited this page Oct 14, 2026 · 1 revision

Reconnect and cancellation validation

Closeout audit: October 14, 2026. Scope: Sprint 1 issue #10. Public release #14 transferred the recovery implementation and integration checks. Public-main CI passed at dd69e2b.

Verified outcomes

Scenario Observed outcome and evidence
Browser closes/reopens The saved experiment is reopened rather than submitted again; browser smoke covers single creation, and a local-backend reopen check was recorded.
Daemon restarts during Oracle work Claim, supervisor and Harbor process identity remain unchanged; one child run exists after recovery.
Telemetry delivery fails Heartbeats stay online while monitoring becomes delayed. UI provisional timers freeze at the last sample; cancellation and final delivery still work.
Queued cancellation Terminal cancellation creates no local run/task.
Active cancellation while worker offline Status stays cancelling until the worker returns. Recovery stops the original process, verifies Docker cleanup and durably acknowledges cancellation.
Terminal partial results A completed verified pass is saved; the unfinished trial is excluded. Child and combined reports arrive without duplicate executions or reports.

The completed isolated October 3 integration run rejected 19 telemetry requests and verified two Oracle passes on the recovered run, followed by a cancellation case retaining one completed pass. No model calls or provider credentials were involved. The retained regression harness is scripts/connected-evaluation-ci.ts; its assertions and CI logs retain scenario/run evidence. UI smoke covers delayed/offline states and provisional timers.

Reproduction and limits

On an isolated Linux checkout with the pinned runtime and Docker prerequisites, run bun run test:evaluations:e2e and bun run test:evaluations:ui. Use the connected-runner guide for recovery behavior.

These checks exercise local Convex and signed disposable identities, not production Clerk credential entry. A daemon restart is not a host reboot; lost supervisor/state requires inspection and run-specific cleanup, never an automatic paid rerun. Monitoring is a bounded snapshot. A partial report's denominator includes completed verifier outcomes only. The first-run and full hosted report-role acceptance remain separately tracked in #9 and #11.

Clone this wiki locally