[BUG] Cowork (Windows): no scheduled-task setting can be saved — "Sign in again on this computer, then save your changes again" — while that same computer still signs successfully from a cloud session

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

In the Claude desktop app (Cowork) on Windows, every attempt to save any scheduled-task setting fails with the toast:

Sign in again on this computer, then save your changes again.

This happens for all three of my scheduled tasks — both the device-bound ones and the cloud-only one — and for every field I tried (folder list, model, prompt). Signing out and back in, restarting the app, and rebooting the machine all leave the message unchanged. It has persisted across two days.

The important part: the computer's signing path is not actually broken. At the same time the app refuses to save, the exact same machine:

• is linked to a cloud Cowork session and serves every remote-devices tool normally,
• runs its scheduled tasks successfully (all three report SUCCEEDED), and
• successfully re-signs a scheduled task's prompt when the same kind of edit is made through the Claude Code Remote MCP instead of the app.

So this is not a stale session taking the device offline, as in #81512 / #77596 / #95963. The device is online and demonstrably able to sign. Only the desktop app's settings-save path demands a re-login, and that re-login does not help.

Evidence that the device signature is healthy

All of the following happened while the app was still refusing to save, from a cloud Cowork session linked to this same computer:

1. device_request_folder_access for a local folder → {"granted":["<local folder>"]}
2. device_bash mounted that folder read/write and wrote a 31 KB file into it successfully.
3. list_triggers shows all three scheduled tasks with last_run.status = ROUTINE_RUN_STATUS_SUCCEEDED.
4. update_trigger with a new prompt on a device-bound task succeeded, and the server appended prompt re-signed by <this computer> — this scheduled task still requires that computer. The task kept folders_state: FOLDERS_STATE_PRESENT and its device binding.

Point 4 is the crux: the device produced a valid signed approval on demand, seconds after the app had refused the same kind of edit.

Why this is worse than it looks: the model becomes unreachable

update_trigger deliberately refuses model changes on a device-bound task:

failed to update trigger: model cannot be patched directly on a device-bound routine;
set the model inside the signed session_request edit

That is reasonable on its own — the model is meant to be changed through the app's signed edit. But the app's signed edit is exactly what is broken here. The combination means:

• The model of a device-bound scheduled task currently cannot be changed by any available path.
• The same applies to the folder list, which has no MCP equivalent at all.

Delete-and-recreate is not a workaround: per #92268 an agent-created task cannot be re-bound to the computer afterwards, so recreating would permanently lose the device binding and the run history.

What Should Happen?

1. If the app needs a fresh device signature, signing in again should actually produce one. Right now re-login, app restart and a full reboot all leave the same latched state.
2. The app should not demand a re-login when the device is demonstrably able to sign — the MCP path proves it is, in the same minute.
3. The error should name which credential is missing or stale rather than a generic "sign in again", and ideally perform the same re-sign the MCP path performs.
4. Failing all of that: allow the model of a device-bound task to be changed through a signed MCP edit, so that a broken app save path cannot make a setting permanently unreachable.

Error Messages/Logs

Toast shown by the desktop app on every save attempt:

    Sign in again on this computer, then save your changes again.

For contrast, from a cloud Cowork session linked to the same computer, in the same minute:

    update_trigger (prompt change, device-bound task) -> success
    prompt re-signed by <this computer> — this scheduled task still requires that computer

    update_trigger (model change, device-bound task) -> error
    failed to update trigger: model cannot be patched directly on a device-bound routine; set the model inside the signed session_request edit

I have not attached claude.ai-web.log excerpts yet — tell me which lines you want and I will add them.

Steps to Reproduce

1. In the Claude desktop app (Cowork) on Windows, open any scheduled task and go to its settings.
2. Change anything — attach a folder, change the model, or edit the prompt.
3. Click Save.
4. A toast appears: "Sign in again on this computer, then save your changes again." Nothing is saved.

This reproduces every time, on all three of my scheduled tasks, device-bound and cloud-only alike, and for every field.

Already tried, none of which changes anything:

• signed out of the app and signed back in
• killed the Claude process and restarted the app
• rebooted the machine

To see the contrast, from a cloud Cowork session linked to the same computer, call update_trigger with a new prompt on one of the device-bound tasks: it succeeds and the server reports the prompt was re-signed by that computer.

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude desktop app 2.9939.2 (Electron 44.4.3, Node 24.21.0) — desktop Cowork, not a CLI install

Platform

Other

Operating System

Windows

Terminal/Shell

Other

Additional Information

Related issues, and how this one differs

• #81512 — Cowork device session force-logged-out every 24–36h (elevated_auth / session_stale_relogin), with continuous remote_cowork.sign_for_session_header_failed. Closest match, since signing is what fails. Difference: there, re-login fixes it immediately and the device goes offline. Here re-login does nothing and the device stays fully online.
• #77596 — device_key_missing, device key not persisting on Windows 11. Difference: there, re-login fixes it at least temporarily and scheduled tasks silently fail. Here the tasks all run fine.
• #95963 — daily forced re-login (session_stale_relogin) plus frozen scheduled-task runs. Difference: my runs are not frozen; only the settings editor is affected.
• #92268 — agent-created scheduled tasks can never be device-bound. Relevant because it removes delete-and-recreate as a workaround.

Setup: individual claude.ai account (no Team/Enterprise). Three scheduled tasks, two device-bound with connected folders, one cloud-only. All were created through the Claude Code Remote MCP rather than in the app.

Redaction: task IDs, folder paths and the device name are replaced with placeholders above. I can supply them privately, along with claude.ai-web.log excerpts — just say which log lines you need.