[Bug] "thinking blocks in the latest assistant message cannot be modified": Error when context is over the limit

[jonasstabenow]

Environment

• OS: Linux RHEL9
• Eclipse Version: 4.38
• Plugin Version: 0.17 & 0.18
• Model: Claude Opus 4.6 (likely affects any Anthropic model with extended thinking)

Bug description
When using MCP tools that produce large outputs (pushing context utilization to (way) over 100%), requests to Anthropic models fail with:

Sorry, an error occurred while generating a response. Details: unhandled status from server: 400
{"message":"messages.1.content.2: thinking or redacted_thinking blocks in the latest assistant message cannot be modified. These blocks must remain as t..."}

Request ID: xxx | GitHub Request ID: xxx

To Reproduce

1. Start a Copilot chat session in agent mode
2. Use an Anthropic model (I used Claude Opus 4.6)
3. Invoke MCP tools that return large outputs (We have a (Bad) tool that can fill the context with a single call)
4. Continue the conversation so the server needs to truncate history
5. The 400 error appears and you cannot recover from it, the chat is broken

Expected behavior

1. The session should never exceed context size by using a summarizing or compression feature.
2. Dont cut thinking blocks in half (I suspect thats whats happening and causing the error, but not 100% sure)

Screenshots

Workarounds

• Try avoid filling the context
• Dont use antrophic models

Additional context

• It seems like Anthropic's API requires that thinking blocks in assistant messages are passed back identical or omitted entirely https://platform.claude.com/docs/en/docs/build-with-claude/extended-thinking#preserving-thinking-blocks.
• This requirments for thinking blocks seems to be older: [BUG] Session becomes completely unrecoverable due to thinking block API error anthropics/claude-code#12743
• This issue was not present in plugin version 0.15 and seems to have been introduced somewhere between 0.15 and 0.17. It might also likely depend on the used language server version.
• This compression feature feat: Implement auto compression feature for chat context microsoft/copilot-for-eclipse#250 might fix the issue, by never needing to cut thinking blocks in half.