[Tong-bit-art]

Problem and change

assignTaskToDomain() can leave a task with two AgentState owners. #4245 deliberately keeps the previous generic owner busy when a task is retargeted to another domain because the retarget does not prove that worker stopped executing; its regression test asserts both agents are busy with currentTask pointing at the same task.

The terminal handlers only cleared one side:

• handleTaskComplete() cleared and released the reporting agent;
• handleTaskFail()'s terminal branch cleared and released the reporting agent;
• cancelTask() cleared and released task.assignedTo.

The other owner kept currentTask === taskId forever. It stayed busy, its domain-pool slot stayed reserved, and the shared reservation could never be resolved even though the task was already completed/failed/cancelled.

Change: add a private releaseStaleTaskOwners(taskId, exceptAgentId?) that runs after a task reaches a terminal state and clears + releases every other agent whose currentTask still points at that task. The reporting (or assignedTo) agent remains handled by the existing path; retries are deliberately untouched because the task is still live and another owner may still be executing it.

This completes the multi-owner model introduced by #4245 without changing retargeting semantics.

Validation

• 3 new regressions in v3/@claude-flow/swarm/__tests__/terminal-task-stale-owners.test.ts:
  • completion from the second owner releases the first;
  • cancellation releases the non-assigned owner;
  • terminal failure releases the non-reporting owner.
• Against main b93144803: 3/3 fail (the non-reporting owner stays busy with currentTask === taskId).
• Candidate: 3/3 pass; full @claude-flow/swarm suite 335/335, 28/28 files; tsc --noEmit 0 errors; git diff --check clean.
• Root v3 vitest config also collects and passes the new file (3/3).

Scope / known limitations

• Retargeting itself is unchanged: both owners still exist until a terminal report. This PR only resolves the stale reservation once the task is actually terminal.
• The retry branch is intentionally not cleared; the task is still live, and fix(swarm): honor shared worker reservations in agent pools #4245's reservation model allows another owner to continue.
• This does not prevent two workers from executing the same task after a retarget; fix(swarm): honor shared worker reservations in agent pools #4245 accepts that trade-off and this change only stops the stale owner from leaking after the task ends.
• No public API or return-shape changes.

Compatibility

• The helper is private and only touches AgentState/existing domain release paths.
• Non-domain agents are cleared and become generically available; domain agents are released through releaseDomainAgent() and their pool state self-heals through reconcileReservations().
• No dependency, lockfile or persistence-format changes.