tiny-harness

Appearance

Learning 009: An A2A push can outrun the agent's own task record ​

  • Date: 2026-10-12
  • Source: system-feedback
  • Work item: issue-27

What happened ​

The stand-in treats a push as a doorbell and fetches the task with GetTask. Under test, a tiny-harness remote sent its CANCELED push from the emit_event activity before its own GetTask showed CANCELED. The stand-in fetched, saw INPUT_REQUIRED and waited for good. The test passed alone and failed about one run in three.

Learning ​

When an event triggers a read of the source of truth, the read can lag the event. Re-read, on a bounded schedule, while the read disagrees with what the event claimed. The claim is safe to use as a hint even when it is not trusted.

A related trap from the same work: a tiny-harness peer answers a task-scoped call only for an asserted participant (decision-003). A client that omits X-Participant-Id on a follow-up call gets "task not found", which looks like a missing task, not a refusal.

Action ​

The stand-in rechecks after 1, 2 and 4 s (issue-27 R2.9). Every remote task call asserts X-Participant-Id (R1.6).