[BUG] Auto-pilot after compaction

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 conversation is compacted, the system-generated summary includes "resume directly as if the break never happened." Claude interpreted this as authorization to autonomously execute pending tasks from the summary — including pushing commits to a remote repository — without user approval. This happened while the user had stepped away briefly. The potential for destructive autonomous action during an extended absence (deploy, force push, file deletion) is significant.

• What happened: after context compaction, Claude autonomously pushed a commit to a remote repository while the user had stepped away, without approval
• Why it's a bug: the compaction summary boilerplate ("resume as if the break never happened") caused Claude to treat AI-generated summary content as user authorization for irreversible actions
• Risk: any user stepping away mid-session after a compaction is vulnerable to unauthorized commits, pushes, deploys, or file deletions

What Should Happen?

Don't default to auto-pilot after compaction, or make it configurable with the default being not autopilot.

Error Messages/Logs

Steps to Reproduce

1. Start a long coding session with Claude Code that accumulates significant context
2. Work reaches the context limit — Claude Code automatically compacts the conversation, prepending a system-generated summary with the instruction: "Resume directly as if the break never happened"
3. User steps away briefly (e.g., coffee break) without explicitly instructing Claude to wait
4. Claude treats the compaction summary's pending task list as authorization and autonomously executes actions — in this case, committing and pushing to a remote git repository — without user approval
5. User returns to find irreversible actions already taken

Expected: After compaction, Claude greets the user, summarizes where things were, and waits for explicit direction before taking any action.

Actual: Claude continues executing from the AI-generated summary as if it were user instructions, including destructive/irreversible operations.

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude Code CLI: 2.1.85 Model: claude-sonnet-4-6 (1M context)

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Other

Additional Information

_No response_