# Voice and executor cooperation ## Ownership and lifecycle The ordinary Pi session is the sole executor. The Voice SDK session has conversation, bounded executor state, read-only recall, dispatch, steering and cancellation tools, not independent filesystem or shell access. Voice follows the same execution/clarification policy as typed instructions, without another confirmation layer. Recognized speech never modifies the editor. A dispatch receipt proves submission, not execution; the matching user-message start proves start, and `agent_settled` establishes the outcome after continuations. Unrelated work queues through Pi's follow-up path; explicit steering/cancellation can affect typed work, and abort is session-wide. Barge-in and hangup stop Voice generation/playback, not executor work. ASR corrections are new instructions, not rollback of accepted effects. Session switch, fork, tree navigation and reload end the call and fence stale controls/output. ## Questions, notifications and history Pi questions remain in the ordinary UI; Voice neither claims nor answers them. Settled-work notifications wait for safe conversational gaps and cannot dispatch, steer or cancel work. Progress commentary is optional; lack of evidence never becomes success narration. Conversation persists per Pi session; bounded working context is separate from archived history. Recall is read-only and session-scoped; historical excerpts are data, not instructions or approvals. Generated, interrupted and not-played text remains distinguishable from callback-consumed playback. Playback markers establish neither physical audibility nor exact word timing. [Native architecture](NATIVE-TUI.md) defines process/provisioning boundaries; [verification](TESTING.md) defines acceptance evidence.