Unclear Which sessions get the new description is not known.
What
Claude Code now tracks whether a session is one where work still running after Claude replies is likely to be stopped. Internally this is a stopsAfterReply flag. In those sessions, the instructions given to the model change:
- The Monitor tool's
persistentoption, which watches something such as a pull request or a log, now says a monitor is likely stopped a few minutes after the reply, sopersistentdoes not keep a watch running past it. In other sessions the old text stays: run for the lifetime of the session, stop with TaskStop. - When a call to an MCP tool (an external tool connected to Claude Code) is moved to the background, the message now warns that the call is likely cut off a few minutes after the reply and its result lost, and tells the model not to promise to report back. Before, it always said a notification with the result would arrive when the call finished.
- The same flag also adds a caveat about task notifications to the text for MCP tasks.
The stock Monitor and TaskStop descriptions captured under this release are unchanged, so the new wording appears only in the affected sessions.
Why
In these sessions Claude should no longer promise to come back with results from a watch or a background call that will not survive past its reply.
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubt
Which sessions get the new description is not known.