[BUG] CancelTask fails after an ask-user answer resumed the task ("deferred continuation cannot be resumed") #3314

Activity

1. github-project-automation

   added this to kagent kanbanon Oct 21, 2026
2. github-project-automation

   moved this to Backlog in kagent kanbanon Oct 21, 2026
3. astrogilda commented on Oct 26, 2026

   

   astrogilda

   on Oct 26, 2026

   More actions

   @yyewolf, I traced this on main at 92ac08a: in go/harness/runtime/a2a/executor.go, Executor.Cancel calls restorePending() before it looks at the task state.

   restorePending() reads the durable record only when e.state is nil. The Claude driver's RestorePending, in go/harness/claude/internal/driver/deferred.go, returns ErrSessionAbandoned for a record in phaseResuming or phaseExecuting. Say Substrate restores the data directory into a fresh harness, so the executor no longer holds the resumed turn as its activeTask. The answered record then reads as abandoned, and Cancel returns ErrSessionAbandoned before it can reach the live turn.

   A regression pair through Executor.Cancel with a restoring runner would pin it down. Cancel while parked should end TASK_STATE_CANCELED, as it does today. Cancel after the answer, with the record in phaseExecuting, should stop the turn or report that the outcome is unknown. I'd assert three things separately: what Cancel returned, whether the harness process exited, and what the target saw afterwards. In your run, that last one is the clone and the git push.
4. arun-elr commented on Nov 11, 2026

   

   arun-elr

   on Nov 11, 2026

   More actions

   @yyewolf @astrogilda reproduced this on main (fbfd58a) on kind with the claude harness: answer an ask_user, then cancel while the resumed turn is in a model call. CancelTask fails after 30s with runtime cancellation cleanup failed: context deadline exceeded, and the session can't be deleted after that. kagent, codex and byo cancel fine.

   cause: on cancel the claude driver doesn't kill a replay process. it marks the journal canceling and waits up to 30s for claude to hit a hook and stop (process.go#L409-L423). once the replayed call has its result, no hook fires until the next tool call, so a long model call or tool (the git clone in this report) outlasts the wait. after the timeout Resume marks the journal abandoned, which is likely where the deferred continuation cannot be resumed error comes from on the next cancel.

   I have a fix: once the replayed call has its tool result, stop the rest of the turn like any other turn and record the journal as canceled, so a restored harness doesn't report it abandoned. a cancel before the result still waits for claude to settle the call. unit test plus an e2e test (ask, answer, hold the next model call, cancel) that fail on main and pass with it.

   @EItanya ok to assign this to me? branch: https://github.com/arun-elr/kagent/tree/fix-cancel-after-ask-user
5. arun-elr commented on Nov 11, 2026

   

   arun-elr

   on Nov 11, 2026

   More actions

   @astrogilda @yyewolf draft PR up at #3647, so nobody doubles up on it.

   @astrogilda your trace of Executor.Cancel calling restorePending() before it looks at the task is the a2a half. The claude harness has its own: consume sets cancelReplay and then waits cancellationCleanupTimeout (30s) for a hook that never fires if Claude is mid model call, so CancelTask returns runtime cancellation cleanup failed: context deadline exceeded. The PR stops the turn once the replayed call's tool result has arrived, and records the journal as canceled so a restore does not read it as an unfinished replay.

   Happy to fold in the a2a side here if you would rather it were one change, or leave it to you and keep this one to the harness.
6. EItanya

   assigned

   arun-elr

   on Nov 11, 2026
7. EItanya commented on Nov 11, 2026

   

   EItanya

   on Nov 11, 2026

   Contributor

   More actions

   Assigned! Sorry been super busy at KubeCon