Experimental: support user steering hints during in-progress agent turns

[LyalinDotCom]

Problem

When an agent turn is already in progress, users need a low-friction way to steer execution without waiting for the full turn to finish. Today this is inconsistent and not clearly acknowledged in the UX.

Proposed Direction

Introduce experimental in-progress steering hints:

• Keep normal input available while the agent is running.
• Treat non-slash input during active execution as a steering hint.
• Inject a hidden steering instruction into continuation/follow-up turns with minimal scope changes.
• Preserve unaffected tasks unless explicitly canceled.
• Show a short neutral acknowledgement message in history (not warning/error styling).

Why This Matters

• Reduces user frustration during long-running turns.
• Improves controllability and course-correction latency.
• Makes hint handling explicit and auditable in the session history.

Experimental Scope

This should ship as an experimental behavior first so we can gather feedback and adjust prompt/system behavior and UX details before hardening.

Initial Acceptance Criteria

• Input stays active during Responding state.
• Non-slash input during active turns is captured as steering.
• If immediate continuation injection is not possible, hints are sent after turn completion.
• Steering acknowledgement is visible and non-warning styled.
• Hidden steering instruction avoids over-canceling unrelated tasks.