claude agent throws execution errors when Anthropic > Context Editing: Mode is enabled

[PeterPonyu]

Issue Summary

When I enable GitHub › Copilot › Chat › Anthropic › Context Editing: Mode for Claude models, the Claude agent frequently throws an execution error.

This setting is supposed to automatically manage conversation context as it grows. However, in practice, Claude becomes unstable when it is enabled.

I also noticed that with certain prompts, Claude appears to perform a kind of conversation compaction during the response (similar to /compact behavior), but the process is unreliable and often ends with an execution error.

When the setting is turned off, the issue does not appear to happen in the same way.

Steps to Reproduce

1. Open VS Code with GitHub Copilot Chat installed.
2. Go to GitHub › Copilot › Chat › Anthropic › Context Editing: Mode.
3. Set the mode to any enabled option such as:
   • clear-thinking
   • clear-tooluse
   • clear-both
4. Start a Claude agent session.
5. Use a prompt or workflow that grows the conversation context.
6. Observe that the agent eventually throws an execution error.

Expected Behavior

Claude agent should continue working normally when context editing is enabled, without throwing execution errors.

Actual Behavior

Claude agent becomes unstable and raises execution errors after context editing is triggered or while the conversation is being compacted/rewritten.

Additional Notes

• This seems related specifically to the experimental Context Editing feature for Anthropic models.
• The issue appears intermittent, but it happens often enough to affect usability.
• Claude sometimes seems to partially handle context compaction correctly before failing, which suggests the feature may be triggering but not completing reliably.

[PeterPonyu]

It also seems that some features, such as switching into and exiting Plan mode, can be performed automatically by the Claude agent even without enabling github.copilot.chat.switchAgent.enabled. In this respect, the Claude agent appears to be less constrained by VS Code itself and more likely than the Copilot agent to reflect these new features.

[PeterPonyu]

One more question: can the Claude agent actually use Copilot’s local tools? If commands like /compact can be executed by Claude, why isn’t Copilot’s local toolset shown for Claude?