Model treats context continuation summary as pending task list, acts autonomously

Bug Description

After a session runs out of context and is continued (with a summary of the previous conversation), the model treats items mentioned in the continuation summary as pending/unfinished tasks and autonomously begins working on them — even when the user didn't ask.

Steps to Reproduce

1. Have a long session that triggers context continuation (ran out of context window)
2. The continuation summary mentions various completed and in-progress work (in my case: SVG logo fetching, branding updates, etc.)
3. Send a short message like "this looks fine to me" (referring to the current state of the project)
4. Model responds correctly ("Great — glad it's looking good!")
5. Then it keeps going — unprompted web searches, file reads, and attempted edits related to items from the summary

Observed Behavior

After responding to my message, the model autonomously:

• Opened the browser
• Searched the web for SVG logos
• Read files
• Attempted to update the HTML header with a logo

This happened repeatedly across multiple continued sessions with the same pattern: correct response first, then unsolicited autonomous actions.

Expected Behavior

The model should only act when the user explicitly asks it to do something. The continuation summary should be treated as historical context, not a to-do list.

Environment

• Claude Code v2.1.41
• macOS (Claude Desktop entrypoint)
• Session with context continuation/compaction

Screenshots

User provided screenshots showing the behavior occurring in multiple sessions.