Interrupt the LLM, keep the chat: turn-scoped cancellation (v1.22.0)

Loom SDK v1.22.0 introduces turn-scoped cancellation, allowing users to stop an in-flight model generation and its tool calls while preserving the session, transcript, and context. The update adds new SSE frames, gRPC and HTTP endpoints, and front-end support via the Esc key, alongside model pattern auto-tracking and bundled chat agents with HTTP tools.

RFC BH splits stop-the-turn from stop-the-run. POST /v1/runs/{run id}/cancel {reason?} stops the current turn of an interactive run the in-flight model generation plus the tool calls it started and parks it at awaiting input with the session, transcript, and context intact — instead of throwing the session away like whole-run cancel did. The loop keeps the partial output and synthesizes cancelled-by-operator tool results for any in-flight tool use so the next turn's history stays valid. Run-cancel vs. turn-cancel is double-guarded on ctx.Err plus the ErrTurnCancelled cause. A new turn cancelled SSE frame distinct from the run-level cancelled. Interruption decline: a pending Question can be resolved with disposition:"declined" so the agent proceeds without an answer. Cross-replica routing by runs.replica id new internal/turncancel registry + coord.TurnCancelCoordinator . Every transport: gRPC CancelTurn + ResolveInterrupt RPCs, @loomcycle/client cancelTurn / cancelInterrupt, Python cancel turn / resolve interrupt / cancel interrupt; HTTP + gRPC share extracted connector cores. LoomBoard's Esc key was the front-end driver. Also folded in: RFC BG P1 model pattern for auto-tracking a model family claude-haiku- , and bundled chat agents gaining the HTTP tool.