Learning 014: A parent hears its sub-task late, and a rollover must not drop a pending cancel

• Date: 2026-10-13
• Source: user-feedback (the owner on PR #37: "the integration tests in main branch are also failing … pls fix those as well in this PR")
• Work item: issue-36

What happened

Three integration tests on main failed sometimes in CI. All three came from one change in issue-34: a /sub-task child became a real A2A task. Its parent now hears about it later, through a stand-in workflow that reads the child's state when the child's push arrives.

• Tests that read the parent's view too early. A TUI test waited on the sub-task's own row and then completed the parent before the parent had heard. A durable test expected a WORKING relay that fetch-on-push may never see.
• A production bug that the delay exposed. A stand-in's row update woke a parent as its cancel arrived. The readiness check then ran continue_as_new, whose snapshot carries no pending cancel, so the cancel was lost and the task never ended.

Learning

• Wait for the observer you assert on. When a parent learns state through a stand-in, a test of the parent waits for the parent's own record, not for the child's.
• A continue-as-new must carry, or wait for, every pending signal. Before a workflow rolls over, check for signals its snapshot does not hold.