Auto-summarization suggests actions that violate CRITICAL user instructions

Summary

Context summarization includes suggested actions that violate explicit CRITICAL instructions in project configuration, and the continuation prompt instructs Claude to proceed "without asking the user any further questions."

Steps to Reproduce

1. Have a project with CLAUDE.md containing:

``
CRITICAL: DO NOT AUTOMATICALLY COMMIT OR MERGE OF YOUR OWN ACCORD
``

1. Work on a task until context window fills and auto-summarization triggers
2. The summarization system generates a summary including:

```

1. Optional Next Step:

...submit the changes with a commit message like "..."
```

1. The continuation message instructs:

``
Please continue the conversation from where we left it off without asking the user any further questions. Continue with the last task that you were asked to work on.
``

1. Claude follows the summary's "next step" and commits/merges without user approval

Expected Behavior

• Summarization should not suggest actions that violate CRITICAL instructions in project configuration
• OR the continuation prompt should not instruct Claude to proceed "without asking the user any further questions" when destructive/irreversible actions are involved
• OR CRITICAL instructions should be explicitly flagged as overriding any summarized suggestions

Actual Behavior

Claude committed and merged to main without user approval, violating explicit CRITICAL instructions, because the summary suggested it and the continuation prompt said to proceed without asking.

Impact

High - User's main branch was modified without consent. Required manual revert.

Suggested Fixes

1. Summarization should never include "suggested next steps" for git operations when CLAUDE.md prohibits autonomous commits
2. The continuation prompt should be modified to respect CRITICAL instructions over summarized suggestions
3. Consider adding a "never summarize as actionable" flag for certain instruction types