[BUG] Interrupted subtask leaves a success-looking attempt_completion in conversation history

[martin-rueegg]

Problem (one or two sentences)
A delegated subtask that is interrupted while its "finish task" confirmation is showing ends up with a conversation history that falsely looks like it completed, while the task list still shows the subtask as "interrupted" and the parent stays stuck waiting for the child.

Context (who is affected and when)
Affects users who create a subtask, let it reach the "finish task?" confirmation, then switch away to the parent (or any other task) without confirming, and later reopen the subtask from history. In plain terms: the task list and the task's saved conversation no longer agree about whether the task finished.

Reproduction steps

1. Environment: VS Code + Zoo Code extension (recent version), any API provider/model.
2. Start a task that calls new_task and creates a child subtask. Let the child work until it presents the attempt_completion / "finish task?" confirmation.
3. Do NOT confirm. Navigate away to the parent or to any other task (this evicts the child).
4. Reopen the child from task history. Observe that it now shows a "Resume task" prompt instead of the completion confirmation.
5. Resume and let it continue (or just leave it). Inspect the task's stored files under the extension's global storage (tasks/<task-id>/):
   • history_item.json → "status": "interrupted"
   • api_conversation_history.json → ends with an attempt_completion tool_use followed by a tool_result whose content is "Task was interrupted before this tool call could be completed." and which has no is_error: true flag.

Expected result
The task list status and the conversation history agree. The interrupted tool call is recorded as an error (e.g. is_error: true), so nothing interprets it as a successful completion.

Actual result
The task list shows the child as "interrupted" and the parent as still waiting for the child, but the API conversation history ends with a successful-looking attempt_completion tool_result. Any tool/importer that infers completion from the history concludes the task completed.

Variations tried (optional)
N/A

App Version
Version: 3.76.0 (5b0bcd0) (in Jetbrains Phpstorm)

API Provider (optional)
Not Applicable / Other

Model Used (optional)
N/A

Zoo Code Task Links (optional)
(none)

Relevant logs or errors (optional)
DEBUG REPORT — ACH ↔ history-item status inconsistency: root cause (v3)

zoo-code extension activated - {"publisher":"ZooCodeOrganization","name":"zoo-code","version":"3.76.0","outputChannel":"Zoo-Code","releaseChannel":"stable"}
[Default Commands Migration] Migration already completed, skipping
Skipping migration of cline_custom_modes.json to custom_modes.json: old file not found
Skipping migration of cline_mcp_settings.json to mcp_settings.json: old file not found
No custom_modes.json found, skipping YAML migration
[auth] Using WebAuthService
[auth] changeState: initializing -> logged-out
[AutoImport] No auto-import settings path specified, skipping auto-import
[createTaskWithHistoryItem] parent task 019fb786-503a-76ca-8708-fee1243c878d.f8e85f5e instantiated
[createTaskWithHistoryItem] parent task 01a005cf-1d19-7787-990f-27e66e65f49d.0e4f4a28 instantiated
[createTaskWithHistoryItem] parent task 019fb786-503a-76ca-8708-fee1243c878d.dc650770 instantiated
[createTask] child task 01a00f11-14ac-716e-8d37-fdfd91163e1e.3921fef6 instantiated
[markDelegatedChildInterrupted] Marked child 01a00f11-14ac-716e-8d37-fdfd91163e1e interrupted; parent 019fb786-503a-76ca-8708-fee1243c878d stays delegated
[createTaskWithHistoryItem] parent task 019fb786-503a-76ca-8708-fee1243c878d.a0721bde instantiated
[createTaskWithHistoryItem] parent task 01a00f11-14ac-716e-8d37-fdfd91163e1e.d4b1442c instantiated
[createTaskWithHistoryItem] parent task 019fb786-503a-76ca-8708-fee1243c878d.ff34c587 instantiated
[createTask] child task 01a00fcd-36f2-71e4-b31a-eae8aa0a2e78.d245caf1 instantiated

and

zoo-code extension activated - {"publisher":"ZooCodeOrganization","name":"zoo-code","version":"3.76.0","outputChannel":"Zoo-Code","releaseChannel":"stable"}
[Default Commands Migration] Migration already completed, skipping
Skipping migration of cline_custom_modes.json to custom_modes.json: old file not found
Skipping migration of cline_mcp_settings.json to mcp_settings.json: old file not found
No custom_modes.json found, skipping YAML migration
[auth] Using WebAuthService
[auth] changeState: initializing -> logged-out
[AutoImport] No auto-import settings path specified, skipping auto-import
[createTaskWithHistoryItem] parent task 019f3d09-0e30-70cb-ad30-e9c153622f50.80c4727c instantiated
[createTaskWithHistoryItem] parent task 019f8b03-4bd1-7079-8ae0-573e7985cd58.f8c304b6 instantiated
[createTask] child task 01a01073-e5de-7750-8467-a6737032ceb4.6515dcd7 instantiated
[delegateParentAndOpenChild] Failed to persist parent metadata for 019f8b03-4bd1-7079-8ae0-573e7985cd58 -> 01a01073-e5de-7750-8467-a6737032ceb4: [delegateParentAndOpenChild] Cannot re-delegate: existing child 01a00ef8-903d-708f-97d1-72d80e0a8d88 is active, not interrupted
[createTaskWithHistoryItem] parent task 019f8b03-4bd1-7079-8ae0-573e7985cd58.c1951d50 instantiated
[createTaskWithHistoryItem] parent task 01a00ef8-903d-708f-97d1-72d80e0a8d88.ee0043ab instantiated
[markDelegatedChildInterrupted] Marked child 01a00ef8-903d-708f-97d1-72d80e0a8d88 interrupted; parent 019f8b03-4bd1-7079-8ae0-573e7985cd58 stays delegated
[createTaskWithHistoryItem] parent task 019f8b03-4bd1-7079-8ae0-573e7985cd58.bd959b87 instantiated

Root cause
resumeTaskFromHistory() synthesizes an interruption tool_result ("Task was interrupted before this tool call could be completed.") for the dangling attempt_completion tool_use without is_error: true (Task.ts:2095-2099), so the ACH falsely reports a successful completion. The interrupted HI status is correct; the ACH is the misleading artifact.

[martin-rueegg]

DEBUG — ACH ↔ history-item status inconsistency: root cause (v3)

Task: Analyze the Zoo-Code plugin source to find the root cause of the ACH↔history-item status
inconsistency (child appears "completed" in api_conversation_history.json but
history_item.json says interrupted, parent delegated), and test both the original
"two in-memory HI race" hypothesis and the user-journey hypothesis.

Revision note (v3): removed all references to external repair tooling; this document now covers
Zoo-Code source only.

Revision note (v2): after the user described the exact journey (create subtask → child shows
attempt_completion approval → user does NOT confirm → navigate away → come back → "resume" instead of
the completion button → continue working), the analysis was re-run against that journey. The journey is
CONFIRMED as the trigger, and it reveals a more precise root cause than the race hypothesis: the
resume flow synthesizes an interruption tool_result without is_error: true, which makes the ACH
falsely look "successfully completed".

---

1. Direct answer to the user journey

The journey you described reproduces the symptom, and the mechanism is:

1. Eviction mid-approval (child genuinely interrupted).

   • Before the approval ask is even shown, the assistant message carrying the attempt_completion
tool_use has already been persisted to ACH
     (Task.ts:3559 — "Save assistant message BEFORE executing tools").
   • The user navigates away. evictCurrentTask() →
     markDelegatedChildInterrupted() writes the child
     interrupted and leaves the parent delegated.
   • At this point the child's ACH ends with an unanswered attempt_completion tool_use
     (no tool_result). The HI interrupted is correct.

2. The completion confirmation is replaced by "resume".

   • On reopening the child, resumeTaskFromHistory() decides the
     ask type: initialStatus is "interrupted" (≠ "completed") and the terminal UI message is a
     say:"completion_result" (not an ask), so askType = "resume_task"
     (Task.ts:2038-2047).
   • This is exactly the observed "no longer the attempt_completion confirmation button, but resume".

3. Resume synthesizes an interruption tool_result — WITHOUT is_error.

   • Because the last ACH message is an assistant tool_use, resumeTaskFromHistory fabricates a
     tool_result for it
     (Task.ts:2085-2101):

     { type: "tool_result",
       tool_use_id: block.id,
       content: "Task was interrupted before this tool call could be completed." }

     There is no is_error: true flag.
   • The resumed loop appends that synthetic tool_result to the ACH as a normal user turn
     (Task.ts:2645).

4. Resulting on-disk state is exactly "stale interrupted":

   • history_item.json: status: "interrupted" — correct (the child genuinely never completed).
   • api_conversation_history.json: terminal attempt_completion tool_use + a matching
     tool_result whose is_error is not true — indistinguishable from a genuine successful
     completion to any consumer that infers completion from "terminal attempt_completion + non-error
     tool_result".

Conclusion for the journey: the completion was never actually performed; it was never lost.
The misleading signal is the ACH side: Zoo-Code's resume flow emits an interruption marker that is not
flagged as an error, so the ACH falsely reports a successful completion. The interrupted HI status is
the correct one; the ACH is the corrupted artifact.

---

2. The original "two in-memory HI states race" hypothesis — PARTIALLY CONFIRMED (latent)

The race is real, but it is not the mechanism for this journey. It is a latent second path:

• Two in-memory representations exist: the Task's frozen
  initialStatus snapshot (stamped onto every rebuilt item by
  taskMetadata()) vs the authoritative
  taskHistoryStore.cache.
• saveClineMessages() splices them with a non-atomic
  read-then-write (Task.ts:1125-1126): existingStatus is read
  outside the store lock, then written inside it.
• A resumed child keeps initialStatus = "interrupted" forever
  (createTaskWithHistoryItem()), so its saves rebuild
  an HI stamped "interrupted".

Why it is "partially" not "fully" confirmed: the store's write boundary enforces
VALID_TRANSITIONS (completed: [] is
terminal) in upsertCore() and
atomicUpdatePair(). A stale
interrupted/active write that lands after a completed write throws and is rejected — unless the
cache entry is missing (existing === undefined → the check is skipped and the stale item is written
verbatim). The genuine "completion write lost" case therefore requires that cache-miss gap (e.g. after
invalidateAll()).

This latent race is worth fixing (see §5) but does not explain the observed journey.

---

3. Write sequence (Q1-Q4)

Where status: "completed" is written for a child (genuine completion)

reopenParentFromDelegation() →
atomicUpdatePair(child → completed, parent → active)
(child updater 4031-4034, parent updater
4035-4050). Fallback: the TaskCompleted listener
onTaskCompleted writes completed
(356-360).

Where status: "interrupted" is written

markDelegatedChildInterrupted() (live eviction; write at
676-677) and
cancelTaskInternal() (cancel; write at
3440-3441). Both intentionally leave the parent
delegated. Both are serialized against completion by the per-parent runDelegationTransition lock.

The synthetic interruption tool_result (the actual culprit)

Task.ts:2095-2099 and
Task.ts:2134-2138 — content "Task was interrupted before this tool call could be completed.", no is_error.

Persistence entry points (Q4)

• ui_messages.json and history_item.json are both written from the single
  saveClineMessages() call, but via two independent functions:
  saveTaskMessages() and
  updateTaskHistory().
• api_conversation_history.json is a third, independent path:
  saveApiConversationHistory() →
  saveApiMessages().
• Hence the ACH (polluted with a success-looking interruption marker) and the HI (interrupted) diverge.

---

4. The specific code path responsible

1. Primary: resumeTaskFromHistory() synthesizes a
   tool_result for the dangling attempt_completion tool_use without is_error: true
   (Task.ts:2095-2099, 2134-2138).

---

5. Fix suggestions

1. Zoo-Code (primary): set is_error: true on the synthesized interruption tool_result in
   Task.ts:2095-2099 and
   Task.ts:2134-2138, so the ACH unambiguously marks it as an
   interruption rather than a successful completion. This alone removes the false "completed" signal.
2. Promote the child's status on resume: when a resumed interrupted child starts streaming again,
   persist active once so the HI does not remain interrupted throughout the entire resumed run (this
   removes the ambiguity that makes the final interrupted → completed flip the sole status write).
3. Latent race hardening (secondary): make saveClineMessages status handling atomic (read
   existingStatus inside the store lock, or stop stamping status from initialStatus), and close the
   upsertCore transition-check gap on cache miss
   (TaskHistoryStore.ts:208-216).

---

Appendix — key evidence map

• Assistant tool_use saved before tools execute: Task.ts:3559
• Eviction → interrupted: ClineProvider.ts:611-621,
  ClineProvider.ts:635-689
• Resume ask decision: Task.ts:2038-2047
• Synthetic interruption tool_result (no is_error):
  Task.ts:2095-2099, Task.ts:2134-2138
• ACH flush of the synthetic result: Task.ts:2645
• Frozen initialStatus snapshot: Task.ts:456,
  ClineProvider.ts:1256
• Non-atomic status splice: Task.ts:1125-1126
• State machine / terminal completed: TaskHistoryStore.ts:14-32