2.7 KiB
2.7 KiB
id, title, status, wave, depends_on, plan, spec
| id | title | status | wave | depends_on | plan | spec | ||||
|---|---|---|---|---|---|---|---|---|---|---|
| 05-integration-regression | Integrated stop regression | done | 5 |
|
plan.md | ../../specs/tasks/requirements/parent-child-task-stop.md |
Task 05: Integrated stop regression
Acceptance
- Real orchestrator, dispatcher, and MCP backend-handler wiring lets a direct parent stop a long-running child without waiting for turn completion or dispatching a replacement prompt.
- Action response returns promptly; child session becomes
CANCELLED, eligible child task becomesREVIEW, and the simulated lifecycle reports the execution is no longer running. The test does not require simulated-manager map removal. - A repeat call returns
not_runningand changes no additional state. - The test does not claim to exercise the local task-mode MCP transport adapter; its exact action/payload forwarding is covered in Task 04.
Verification
cd apps/backend && go test ./internal/integration -run TestMCPStopTask
Files Likely Touched
apps/backend/internal/integration/mcp_stop_task_test.go
Dependencies
- Tasks 01-04 complete the lifecycle, coordinator operation, backend action, and tool adapter.
Inputs
- Spec scenarios for direct-parent success, no replacement prompt, and idempotent retry.
- Existing orchestrator integration harness and mock-agent long-turn patterns.
Output Contract
Update this task to done, update plan.md, and report observed timing/state,
files changed, tests run, blockers, and any unverified process-cleanup risks.
Completion
- Added a real orchestrator, MCP-handler, and shared-dispatcher regression for a long-running direct child.
- Deliberately blocked simulated runtime teardown and proved the response lands
after session
CANCELLEDand taskREVIEW, but before process stop finishes. - Verified no replacement prompt or queued message is created, runtime stop
completes after release, and a repeat returns
not_runningwithout mutation. - Passed the focused test, 20 repeated runs, and the focused race-detector run.
- After CI exposed a
RUNNING-versus-stream ordering race, added a deterministicmessage.addedbarrier for the persisted initial tool call before capturing the pre-stop message baseline. The strict no-new-message assertion remains; no sleep, retry, or relaxed expectation was added. - Passed 300 race-detector repetitions across
-cpu=1,2,4, the complete integration package, and the full 133-package shard-2 race test. - Remaining lifecycle boundary: the test proves the simulated runtime reaches stopped status, not operating-system cleanup for every production provider.