Background agent lifecycle incident

Completed task output disappears before retrieval

The interface reports a successful finish, then immediately rejects the same task identifier with No task found with ID.

Repro evidence captured

Affected taska1f48b4de9ae5dc32

AgentAudit remediation scope

Reported run45m 43s

Observed behavior

Agent "Audit remediation scope" finished · 45m 43s Task Output a1f48b4de9ae5dc32 └ Error: No task found with ID: a1f48b4de9ae5dc32. Running background agents: a27796702701fc0e5 Design execution phases a17f7eceeff4ca0ec Inspect IGHE implementation plan a1ac5c814dc2f9c63 Inspect IGHE implementation plan a56a0f82f0ceea951 Inspect IGHE plan a9ef392c0167decb8 Inspect IGHE implementation plan ae027938d0de1d40c Inspect IGHE and plan

Expected behavior

After completion is reported, the task identifier should remain valid and its output should be retrievable for the documented retention period. Expiry or cleanup should produce an explicit lifecycle status rather than claiming the task never existed.

Reproduction

Start one or more background agents.
Allow an agent to run for an extended period.
Wait for the UI to report that the agent finished.
Request output using the completed task ID.
Observe Error: No task found with ID.

Diagnostic data to collect

Creation, completion, cleanup, and lookup timestamps for the affected task.
Mapping between agent ID, background task ID, and output artifact ID.
Session and process identifiers used by completion and lookup paths.
Retention or garbage-collection events around completion.
Whether output remains on disk or in persistent storage after registry removal.
Whether duration or concurrent-agent load changes reproduction frequency.

Impact

Completed work is lost or inaccessible.
Long-running results cannot be reviewed.
Invalid IDs, cleanup, expiry, and state desynchronization are indistinguishable.
Concurrent background workflows become unreliable.

Suspected failure area

Registry entry removed before output consumption.
Completion and lookup paths expose different identifiers.
Completed state is isolated in another process or session.
Retention logic disposes of output prematurely.

These are hypotheses only; the captured evidence does not establish root cause.

Acceptance criteria

Completed output remains retrievable for the documented period.
The completion event exposes the identifier accepted by output retrieval.
Cleanup never removes output before it can be consumed.
Missing, expired, and invalid IDs return distinct actionable errors.
Concurrent agents do not affect completed-output retrieval.