[BUG] CCR remote agent commits to local branch without pushing — work lost when sandbox terminates

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?

When a scheduled remote agent (CCR) creates a git commit, it does so on a local branch inside the ephemeral sandbox. The sandbox terminates after the session ends, and the commit is permanently lost — there is no automatic git push before teardown.

What Should Happen?

The CCR runtime should automatically push any committed branches to the remote repository before the sandbox terminates. Alternatively, the documentation and/or the agent's system prompt should make it explicitly clear that the user's prompt must include git push — and ideally the runtime should warn or fail visibly if commits exist but were not pushed at session end.

Error Messages/Logs

- No error, no warning, no indication that the push was missing — the routine status showed ended_reason: "run_once_fired" as if everything succeeded

Steps to Reproduce

1. Create a one-time scheduled routine via the RemoteTrigger API with a git repository source
2. Include instructions for the agent to write a file and commit it
3. Wait for the routine to fire
4. Check the GitHub repository — the branch and commit are absent
5. Check the routine's job_config response — it shows outcomes[].git_repository.git_info.branches: ["claude/quirky-euler"], confirming the branch was created inside the sandbox but never

pushed

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.119 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

Impact

• ~12% of weekly token quota consumed with zero recoverable output
• No error, no warning, no indication that the push was missing — the routine status showed ended_reason: "run_once_fired" as if everything succeeded

Environment

• Claude Code CLI (scheduled routines / CCR)
• Environment: env_01MKmuuDckQzSYZBBa7xYzEG (anthropic_cloud)
• Model: claude-opus-4-7
• Trigger type: run_once_at (one-shot)
• Repository: GitHub (public)

Suggested fix

Either:

1. Auto-push all local commits before sandbox teardown, or
2. Detect unpushed commits at session end and surface a clear warning/error in the routine's outcome metadata, or
3. Add a prominent note in the CCR documentation that prompts must explicitly include git push to persist work