A paired remote host that misses check-ins is probed before being written off, with a new 'still connected, reply lost' message

What

When Claude Code sends tool calls to a paired remote host (another machine that runs commands for the session), it watches for regular check-ins. What happens after too many missed check-ins has changed.

• The transport (the connection to the host) gained a probeLive method, and Claude Code now calls it before reporting the host as unresponsive. If a late answer arrives in the meantime, that is reported.
• If the probe gets no answer, the host is marked away with probe_missed.
• The host_unresponsive result now carries a stillConnected field.
• When stillConnected is true, the connection is no longer marked unresponsive. Before, it always was.
• In that case the message Claude receives says the host is still connected but its reply to this command was not received, that the reply may have been too large to deliver, and that it should try a narrower command rather than stop using the host.
• The old message, which said the host stopped answering and told Claude not to call it again this turn, is still used when the host does not answer the probe.

Why

One command with a very large output could previously make Claude treat a working host as gone for the rest of the turn. Now a live connection stays usable and Claude is steered toward asking for less output.