What
MCP (Model Context Protocol) lets Claude Code use tools from outside servers, and some of those tools run as background tasks. Claude Code changes what it tells the model about them:
- A new notification says that Claude Code cannot follow, or has stopped checking, a background MCP task and that the outcome is not known. The model is told not to report the task as done or failed and not to call the tool again. If you stop the task yourself, this notification is dropped.
- When a stopped
mcp_taskhas the status killed, the stop result can carry a note. It says cancellations are fire and forget, the MCP server gives no confirmation, it may already have done some of the work, and the model should only check on it if you ask. The note appears only when an internal check,Tge(), passes. - Failed-task messages are now built from "Task failed: " plus a helper that blanks control characters and cuts long error text with "… [truncated]". Before, the error text was inserted in full. Outcome-unknown notifications for MCP tasks now use a shared formatter.
Why
The model is less likely to claim a long-running MCP task finished or failed when nobody knows. Very long or garbled error text from a task no longer floods the conversation.