Message queue for non-blocking input during streaming

[dnouri]

Problem

Pi.el blocks input while the agent is streaming. Users have to wait for the response to complete before they can send follow-up messages or course-correct the agent.

Backend Support

The pi backend already supports message queuing during streaming with two distinct behaviors:

Steering interrupts the current response - useful for "stop, do this instead" corrections. Follow-up waits until the agent finishes - useful for queuing the next question while reading the response.

The TUI uses Enter for steering and Alt+Enter for follow-up. During streaming, it displays queued messages below the editor (e.g., Steering: stop and fix the tests first). On abort, queued messages are restored to the editor so nothing is lost.

RPC Commands

The backend exposes steer and follow_up commands, plus prompt accepts a streamingBehavior option. State includes pendingMessageCount for header display. Mode settings control whether queued messages are processed all-at-once or one-at-a-time.

One gap: clearQueue exists in agent-session.ts but isn't exposed via RPC. This would be needed for proper abort handling (restoring queued messages to the input buffer).

Implementation Considerations

A minimal version could just accept input during streaming and queue via steer, showing the count in the header line. Full parity with the TUI would require the missing clear_queue RPC command for abort restoration.

Reference

• Queue implementation: packages/coding-agent/src/core/agent-session.ts (lines 150-160, 708-730, 831-857)
• RPC commands: packages/coding-agent/src/modes/rpc/rpc-mode.ts (lines 327-335, 417-425)
• TUI display: packages/coding-agent/src/modes/interactive/interactive-mode.ts (updatePendingMessagesDisplay, line 1907)
• TUI abort restore: same file, setupKeyHandlers (line 993)