[mavam]

🔍 Problem

Messages could only steer an agent's current work or wait for it to finish. Correcting a long-running or stuck agent required stopping it, which also withdrew its job and stopped the agents it spawned.

🛠️ Solution

• agent_send takes mode: "steer" | "followUp" | "interrupt" (default steer) on both surfaces: Pi's tool and agents' own tool. Agents can interrupt whoever they can message; stopped, missing, and not-yet-started recipients are refused before anything is aborted.
• An interrupt aborts the recipient's work under its turn, records the interrupt watermark, then sends the message. The job and spawned agents stay; owed results wait for the next answer, as with Esc. Results report how many queued messages were dropped.
• Replay-safe correction for agents' sends: the message is logged only after the abort and watermark, so a resumed send that finds it logged or submitted never aborts the work the message started; one that crashed earlier aborts again, then logs.
• Cancellation: the abort plus watermark run in the service, not bound to the sender's call, while the sender holds the recipient's turn. A cancelled sender stops waiting at once and logs nothing, so agents interrupting each other can't deadlock, and the watermark still lands, so the recipient's results don't restart it. Callers cancelled while queued for a turn give it up.
• Compatibility: followUp: true still works when mode is absent, including in the call-line view.

💬 Review

• Focus: turn holding in MessagingHub.turn and the ordering abort → watermark → log → submit.
• Not crash-atomic: a crash between abort and Pi's request record leaves the agent interrupted without the correction; a cancelled agent sender leaves the recipient interrupted (results deferred) without the message.
• Not guaranteed first: other agents' messages can't land in between, but user prompts and plain parent sends bypass the turn.
• Aborts wait for the recipient's tools to stop; a tool ignoring cancellation holds up the send and that recipient's turn.
• Tests: parent interrupt (dropped count, job and child kept), agent interrupt answering the job, crash replay without re-abort, mutual interrupts, a sender cancelled mid-interrupt keeping results deferred, legacy followUp.