OpenCode 2: a queue-delivered prompt detaches a foreground wait:true call the moment it is enqueued #563

Summary

On OpenCode 2, a prompt delivered with delivery: "queue" detaches a foreground wait: true shell call (and a sync bash_watch) the moment it is enqueued, exactly like delivery: "steer". A queued message is not shown to the agent until the current turn ends. So the detach hands control back for a message the agent cannot read yet, and the wait the caller asked for is lost.

Repro (aft-opencode 0.60.0 on OpenCode 2.0.31)

shell({command: "sleep 60", wait: true}). Then, from a plugin, ctx.session.prompt({sessionID, delivery, text}) with text containing no &detach:

Neither call waited for the command to exit.

Where (main, 5ee867d)

packages/opencode-plugin/src/v2-prompt-detach.ts: the prompt hook (interruptBashWaitsForV2Prompt, line 110) interrupts every waiting call for the session through interruptBashWaitsForChatMessage. V2PromptHookEvent (line 33) reads only sessionID, prompt and metadata, so the delivery mode never reaches the decision.

Expected

A queue delivery does not detach a foreground wait, since the agent cannot act on it until the turn ends. A steer delivery, or typed operator text, still detaches. If the hook event does not expose the mode, the hook could defer the detach until the message is actually admitted to the conversation.

The current workaround, bash.detach_on_user_message: false, also stops steer deliveries and operator messages from detaching, which is the case the detach exists for (#366).