[BUG] Context summary from earlier in session injected mid-conversation, causing data inconsistencies

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?

During a long active session, a summary of earlier conversation
parts was automatically injected mid-session when context limit
was approached. This summary contained inaccurate data
(e.g. wrong plant count: "~40" instead of actual 56/57).
The injected summary then influenced subsequent responses
with incorrect information, without user awareness.
Expected: User is notified when context is being compressed/summarized.
Actual: Silent injection of inaccurate summary mid-session.
OS: Windows 11

What Should Happen?

User should be notified when context is being compressed or summarized,
so they are aware that subsequent responses may be based on an
inaccurate or incomplete summary rather than the full conversation.

Error Messages/Logs

Steps to Reproduce

1. Start a long Claude Code session with many file edits

(multiple large files, repeated tool calls)

1. Continue session until context limit is approached

(noticeable when responses slow down or tool calls hang)

1. Continue conversation after that point
2. Observe that a summary of earlier conversation parts

has been silently injected into the context

1. Ask Claude about a specific fact from earlier in the

session — response may contain inaccurate data from
the injected summary

Claude Model

Sonnet (default)

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

4.6

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

_No response_