[BUG] Subtask completes but doesn’t return to parent after interruptions

[hannesrudolph]

Type

Bug

Problem / Value

In Boomerang (subtask) mode, after an interruption and resume, the subtask can complete but control does not return to the parent task. The parent–child link breaks, stalling longer workflows and forcing manual recovery.

Context

Affects users running subtasks during longer sessions or when the editor/model session is interrupted (window reloads, crashes, hangs, provider stalls). This occurs frequently. You can resume the subtask and let it finish, but the parent does not resume.

Reproduction

1. Start a parent task that spawns a subtask (Boomerang).
2. Cause an interruption (e.g., window reload, crash, hang) and then resume the subtask.
3. Let the subtask finish.
4. Expected: the parent task automatically resumes when the subtask completes (i.e., on subtask completion).
5. Actual: the parent does not resume; the parent–child link appears broken.

Constraints/Preferences (optional)

• Make the parent–subtask relationship durable across interruptions and reloads.
• Provide a clear “Return to main task” action.
• Offer a list of recoverable subtasks with status and their associated parent.

[roomote]

I'm exploring a potential fix for the parent-subtask link persistence issue after interruptions. I'll open a PR to share progress.

[hannesrudolph]

Hey @roomote-agent, I dug into the subtask “boomerang” flow and it looks like the parent–child handoff breaks specifically after an interruption because the parent isn’t rehydrated onto the provider stack when resuming a subtask from history.

What’s happening

• On subtask completion, attempt_completion routes here: attemptCompletionTool(). If the task is a subtask, it calls ClineProvider.finishSubTask().
• finishSubTask pops the child via ClineProvider.removeClineFromStack() and then tries to resume the parent via Task.completeSubtask().
• After reload, a resumed subtask is created from history with ClineProvider.createTaskWithHistoryItem(), but only the subtask instance is on the stack; the parent object isn’t present (only parentTaskId is persisted). After the child is popped, getCurrentTask() may be undefined, so the parent never resumes.

Why it works only during a single uninterrupted session

• In a normal run the stack is intact (parent under child), so popping reveals the parent and Task.completeSubtask() runs.
• After a reload, the stack only has the child; the parent wasn’t reconstructed.

Proposed minimal fix (localized)

• Make ClineProvider.finishSubTask() resilient:
  1. const parentTaskId = this.getCurrentTask()?.parentTaskId
  2. await removeClineFromStack()
  3. let parent = getCurrentTask()
  4. if (!parent && parentTaskId) { const { historyItem } = await getTaskWithId(parentTaskId); parent = await createTaskWithHistoryItem(historyItem) }
  5. await parent?.completeSubtask(lastMessage)

This rehydrates the missing parent only when needed and preserves current stack semantics, including multi-level chains.

Nice-to-haves

• Provide a “Return to main task” action that invokes the same fallback. The current clearTask path only checks a live parent reference: webviewMessageHandler.ts. Add a branch to check parentTaskId and call the finishSubTask fallback.
• Offer a “Recoverable subtasks” list by filtering history for items with parentTaskId and routing clicks to the same rehydrate-and-return logic.

Tests

• Unit: finishSubTask rehydrates absent parent and calls Task.completeSubtask().
• E2E: Parent spawns subtask; simulate reload (resume child from history); complete child; expect parent to resume.

[roomote]

Thanks @hannesrudolph for the detailed analysis! I've reviewed your findings about the parent-child handoff breaking after interruptions. Your proposed solution to make finishSubTask() resilient by rehydrating the parent when needed looks solid. I'll implement this fix along with the nice-to-have features for better recovery options. Starting work on this now.

[roomote]

@hannesrudolph I've implemented your proposed fix and opened PR #8090: #8090

What I implemented:

✅ Minimal fix in ClineProvider.finishSubTask() - Now rehydrates parent from history when not on stack after interruption
✅ Enhanced clearTask handler - Checks for both parentTask and parentTaskId to ensure proper parent resumption
✅ Comprehensive unit tests - 8 test cases covering various scenarios including multi-level subtask chains

What's pending:

• "Recoverable subtasks" list feature (nice-to-have for future PR)
• E2E test for reload scenario (can be added separately)

The CI checks are currently running. All unit tests are passing locally. Thank you for the detailed analysis - it made the implementation straightforward!

Feedback welcome!