Repository navigation
Resumed steps that stopped arriving
After a paused workflow was resumed from a model turn, the next step could stop arriving, and the run waited forever with the status "A step is pending delivery. It starts a new model turn after this turn ends." Since Pi 0.87.0, a turn requested from an agent_settled handler starts only after every settled handler has returned, and Pi reports itself idle until then. The branch report sent right after the step therefore named the new resumed step without an entry, and the server read that as a lost delivery of the same request and cancelled the step.
The server now draws conclusions only about the message a branch report names. A named step that Pi has not added yet starts no recovery, and recovery reopens a cancelled re-issue target, so a request always keeps one step that Pi can receive. The extension reports the entry of a delivered step whose turn it owns and lets the server's turn receipt settle that turn, so one cancelled step can no longer hold back later steps. A new E2E test replays the incident on a real Pi and hangs without the fix.
Pi 1.1.0
Development and CI use Pi 1.1.0, so CI runs the deferred agent_settled timing. The Pi peer ranges move to >=0.99.0. Pi 0.99.0 changed preflightResult to report started, handled, or queued for an accepted prompt and nothing for a rejected one, and Pi agent groups now require started, so a handled or queued prompt no longer passes as accepted. Agent groups also pick the chat entry of a model ID that has image or classifier entries, and the installed live E2E accepts the extensions that Pi 1.0 and later ship built in.
Running Pi sessions pick up the release after a reload.