[BUG] Routines created via natural language cannot be edited: permission mode change does not save

Open 💬 0 comments Opened Jul 10, 2026 by Fediac

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?

Routines created by describing them in natural language in a Code session are broken when edited afterwards. If I open such a routine from the Routines list, change the permission mode to "Bypass permissions", and click Save, the change does not persist — reopening the routine shows the old mode. Routines created normally through the Routines → New routine → Local form do not have this problem; their edits save correctly. The issue is specific to routines created via natural language in a session.

What Should Happen?

A routine created via natural language should be fully editable, just like one created through the New routine form. After selecting "Bypass permissions" and clicking Save, the mode should persist and show as "Bypass permissions" when the routine is reopened.

Error Messages/Logs

No error message. Save appears to succeed, but the change silently reverts.

Steps to Reproduce

  1. In a Code session (New session), create a one-time routine by typing a natural-language instruction, e.g. "create a one-time scheduled task that runs tonight at 02:30, works in folder C:\Users\...\my-project, and does X". Claude creates it and it appears in the Routines list.
  2. Open that routine from the Routines list and click Edit (pencil icon).
  3. Change the permission mode dropdown to "Bypass permissions".
  4. Click Save.
  5. Reopen the routine (or open Edit again) and check the permission mode.

Result: the mode is not "Bypass permissions"; the change did not persist.

Note: a routine created through Routines → New routine → Local does NOT show this bug — its edits save correctly. Only routines created via natural language in a session are affected.

Claude Model

Not sure / Multiple models

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

v2.1.206

Platform

Other

Operating System

Windows

Terminal/Shell

Other

Additional Information

Related observation: even when a routine is meant to run in bypass mode, some Bash operations still trigger an approval prompt regardless of permission mode. For example, a step running echo $((RANDOM % 37)) shows an "Arithmetic expansion references variable or non-literal: RANDOM" prompt, which stalls unattended/scheduled runs. This may be a separate issue but is mentioned for context.

Environment: Claude Code Desktop on Windows, Pro plan. Screenshots attached: (1) the Edit view with "Bypass permissions" selected before saving, (2) the approval prompt during a run.

Btw used claude desktop when i had this bug.

View original on GitHub ↗