Release Hardening Lives in the Details
The headline features are complete. This afternoon is about the failures that only appear after an agent system has been running for hours: a long poll treated like a normal API request, a typing timer that survives the final answer, and discovery paths that disagree about what exists.
Cloud Connect Is a Long Poll
The Cloud Server connect path intentionally waits for work. Applying the general HTTP timeout to that request turns healthy idle time into a stream of false failures. v1.7.1 gives connect its own long-poll timeout and retry policy, separate from ordinary API calls.
Cancellation still stops promptly, transient transport failures back off, and the connection loop can distinguish a normal empty poll from a network problem. The result is quieter logs, less reconnect churn, and a Cloud bridge that behaves like the protocol it implements.
Typing Must End Before the Answer
The Microsoft 365 Copilot bridge sends informative updates and keeps a typing indicator alive while a FabrCore agent works. A timer race could previously produce one final typing activity after the completed response or error had already been sent.
The completion paths now stop the timer first, then deliver the final result. It is a small sequencing fix with an outsized effect on trust: the channel no longer looks as though the agent is still working after it has answered.
One Release Baseline
We also aligned the Microsoft Agent Framework packages on 1.16.0 and tightened the shared assembly scope used by services and the registry. These changes reduce dependency drift and remove cases where one component can resolve an agent that another component never registered.
None of these items makes a spectacular screenshot. Together they make the larger features credible. Harness loops, durable WebSockets, and squads all depend on boring lifecycle correctness beneath them.
Documentation and migration guidance are now the remaining gate. On Monday, v1.7.1 ships.