Remote routines on private GitHub repos auto-disabled with auto_disabled_repo_access after first run

Describe the bug
Scheduled remote routines (/code/routines) that target a private GitHub repository are repeatedly auto-disabled with ended_reason: "auto_disabled_repo_access" immediately on first fire, even when the routine ran successfully on a prior session.

To reproduce

1. Create a remote routine targeting a private GitHub repo
2. Routine fires, runs, and completes successfully
3. Routine is subsequently auto-disabled (enabled: false, ended_reason: "auto_disabled_repo_access")
4. Re-enabling via API restores it, but it dies again on the next fire

Observed pattern

• Affects 3 of 4 routines targeting the same private repo
• The 4th routine has been healthy for 5+ days — it was re-enabled earlier and appears to have cached credentials
• The 3 affected routines die within minutes of firing — before any git push, confirming the failure is at repo checkout (read), not push (write)
• One routine explicitly does no git push, yet still dies — rules out write access as the cause
• No new commits appear in the repo from failed runs, confirming nothing executes

Expected behavior
Routines should maintain repo access across fires without requiring manual re-enable.

Environment

• Claude Code remote routines (CCR)
• Environment: env_014srqQBonPPddvGcLwVryun (Anthropic Cloud, Default)
• Repo visibility: private