Skip to content

[Bug]: Native Claude subagents stay Working for hours after their account hit its usage limit, and cannot be stopped #39604

Closed as not planned

Description

@MikaMcFlurry

Before submitting

Area

apps/server (orchestration-v2, ProviderSessionManager / Claude adapter), Agents panel

Summary

Provider-native Claude subagents whose account hit its usage limit stay Working in the Agents panel for hours with a running timer, although their child threads have no run and no activity. t3_thread_interrupt cannot clear them ("no active run"). The parent shows "Waiting on N subagents" the whole time.

What I observed (2026-11-20, read from the live server state)

  • Parent Claude thread (account A) started 3 native subagents (thread:subagent:…, Claude Opus) at 01:41 UTC.
  • All three child threads' last item is at 02:00:04–02:00:25 UTC; the parent's run failed with a usage-limit error at about 02:00:40, and the thread then continued on another provider/account (limit-switch continuation, feat: continue on another account with usage left when one runs out #28595).
  • At 08:19 UTC (6h 38m later) the three children were still listed as Working 6h 39m / 6h 38m in the Agents panel. Server state for each child at that moment: status: idle, latestRunId: null, activeRunId: null, runCount: 0, pendingRequestCount: 0. So nothing on the child can be stopped.
  • A stop sent to each child (t3_thread_interrupt) answered "no active run" every time. After the children were then marked settled (08:19), the Agents panel still listed them as Working with the timer running, and the parent still showed "Waiting on 2 subagents". I did not check what the panel's own stop buttons do for these rows.

Likely cause (code reading of current main, not verified at runtime)

  1. hasPendingBackgroundWork in ClaudeAdapterV2.ts returns true while any entry of sessionSubagentsByTaskId has status === "running". A subagent that died on the account's usage limit never reports completion, so that entry stays running.
  2. releaseIfStillIdle in ProviderSessionManager.ts then pins the session (DEFAULT_MAX_IDLE_PIN_MS = 4h), and any provider activity on the same session resets that budget (pinnedSinceMs: null). The pin therefore never expires while the thread keeps being used.
  3. The orphan cleanup that would end the records as interrupted (writeReleasedSubagentEvents → endOrphanedNativeSubagent) only runs when the session is released, so it never fires.
  4. A native subagent child thread has no runs, so stop paths that look for an active run have nothing to interrupt and nothing settles the subagent record.

Expected behavior

  • When a run ends with a usage-limit failure, or the thread moves to another account/provider, the native subagents of the abandoned session end as interrupted (or the stale "running" entries stop pinning the session).
  • Stopping a subagent whose child thread has no run settles its subagent record instead of answering "no active run".
  • The Agents panel stops the timer once the child thread is settled.

Steps to reproduce

Depends on the account hitting its limit while background subagents run:

  1. Run a Claude thread on an account close to its usage limit and have it start several native subagents.
  2. Let the account hit the limit while they run (or continue the thread on another account right after).
  3. Open the Agents panel after a few hours and try to stop the rows (or call t3_thread_interrupt on a child).

I have not reproduced this deterministically, and I have not checked whether the same happens without the account-switch continuation; a manual switch to another account after the failed run should leave the same session state.

Environment

Linux server (remote environment), mobile client, Claude Code provider with several accounts, build based on main plus #28595.

Activity

  1. MikaMcFlurry commented on Nov 20, 2026

    @MikaMcFlurry
    ContributorAuthor

    Follow-up from the same thread, measured on the server afterwards:

    • Besides the three children above, two more native subagents of the same parent (started 01:19 and 01:23 UTC, last item 01:34 and 01:40) were still listed as unsettled with status: idle, no run, and kept the parent in waiting ("Waiting on 2 subagents").
    • The user could not close the parent chat or the children from the app while this lasted. Settling the five children (t3_thread_organize settle) and then the parent worked, but t3_thread_list still reports the parent as waiting afterwards, while t3_thread_read reports completed. So the "waiting" state seems to come from the leftover subagent records, which settling does not end.

    So besides ending these records when the session dies, a settle/close of the parent should not be blocked by, or leave behind, records for runless native subagents.

  2. juliusmarminge commented on Nov 20, 2026

    @juliusmarminge
    Member

    I tried to reproduce this on the current build and could not.

    Commit d151236 (origin/main). Could not run the reporter's scenario live: it needs a first-party Claude account hitting its usage limit, which no VM has. Reproduced the core defect as an adapter-level SDK replay in apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts (the test from open PR #24347, adapted to provide NodeCrypto.layer). It starts parent turn 1, which starts two background subagents (task_started) and completes. Parent turn 2 then fails with an assistant rate_limit error and an api_error result. After that, failed system/task_notification frames arrive for both subagents. On main, after both failed notifications, hasPendingBackgroundWork is still true and the subagent records never end: vp test run src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts -t "usage limit" fails with 'expected true to equal false'. With #24347's one-line ClaudeAdapterV2.ts change applied, it passes. While that entry stays running, ProviderSessionManager.releaseIfStillIdle keeps the session pinned, so the release-time orphan cleanup (writeReleasedSubagentEvents -> endOrphanedNativeSubagent) never runs. Not verified: the 'no active run' answer from t3_thread_interrupt on a runless subagent child thread, and the Agents panel timer.

    Closing as not planned. If it still happens for you, reply with the exact steps and the version you are on and it will be looked at again.

  3. MikaMcFlurry commented on Nov 20, 2026

    @MikaMcFlurry
    ContributorAuthor

    Thanks for the replay. For completeness, the version: my observation was on a fork build based on upstream c28c6acaf8 (2026-11-12) plus my own branches, which include changes to ClaudeAdapterV2.ts and the account switch from #28595. So I can't give you clean upstream steps for the live case.

    Your SDK replay on d15123668e already shows the core on main, though: after the failed task_notification frames, hasPendingBackgroundWork stays true and the session stays pinned, so the orphan cleanup never runs. The two parts you couldn't verify (t3_thread_interrupt answering "no active run" on a runless child, and the Agents panel timer) are what that leftover record looks like from the outside.

    Since #24347 turns that test green, would it make sense to track this there instead? Next time one of my accounts hits its limit with native subagents running, I'll check whether the records end on a build with #24347 applied and report back on that PR.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions