Skip to content

[Bug]: Subagent updates reset the main thread working timer during ongoing work #22931

Open

Description

@Kamkmgamer

What happened

While a main thread was working with delegated subagents, messages labeled "Sent by another agent" reset its elapsed working timer. The user reports that a thread working for about an hour starts counting again when a subagent sends an update, even though the same task is still underway and the user has sent no new request.

These updates appear in user-style message bubbles, sometimes marked "Steer", making the reset look as though the user started a new request.

What I expected

A subagent reporting progress during ongoing work should preserve the elapsed time of that work. A steer should not reset the timer. If coordination starts a separate continuation turn, the duration of the ongoing work should remain visible rather than appear to restart from zero.

Steps to reproduce

  1. Start a long-running task in a main thread and have it delegate work to subagents.
  2. Let the main thread accumulate a noticeable working duration.
  3. Have a subagent send an update to the main thread with t3_thread_send while the main task remains underway.
  4. Compare the elapsed working timer immediately before and after the update. Repeat with another update.

Additional context

  • Reported on Windows desktop. The main thread's composer shows Codex, GPT 6.1 Sol. Exact installed version was not captured.
  • This report records the user's observation; the timer reset was not independently reproduced in a running client.
  • The shared timer logic explicitly says timers count from the work start, never last activity, and that a wake keeps the start of the work it continues. Existing tests also expect repeated steers to preserve the original duration. Source inspected at commit 9a43c64ec0ef69d57efe96e3635527f085f80b1e.
  • Related but different reports: [Bug]: A message queued behind a running turn pushes the "Working for…" header below it #19736 concerns timer/header placement after queued input; Agent steers on Claude threads cancel tools and look like user refusals #21895 concerns agent steers cancelling Claude tools. Neither covers the elapsed timer resetting.
  • Impact: misleading duration during long-running delegated work. No workaround confirmed.

Before submitting

  • I searched existing open issues and did not find a duplicate covering the timer reset.
  • I included steps and enough context to investigate; independent reproduction is still needed.

Activity

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