Today, all model-override surfaces (/model, session_status(model=...), sessions.patch, the chat-header picker) write persistent session overrides. There is no atomic single-turn override.

For cost-discipline at scale on a multi-tier model routing scheme, agents need to escalate to a higher-tier model for a single turn and auto-revert without explicit cleanup. The sticky-switch workaround works but adds overhead (every phase boundary needs two sessions.patch calls and live discipline to revert) and creates a recurring failure mode: the session drifts up-tier when the revert is forgotten, burning higher-cost capacity on turns that do not need it.

Proposed API additions (one of)

1. sessions.patch with scope: "next-turn" — patch applies to the next single agent turn and auto-reverts to the prior model on completion.
2. Per-message model field in the chat-send payload — overrides the session model for that send only; resolved model is reported in the response envelope.

Either form would replace the sticky-switch protocol with a single atomic operation and eliminate the drift failure mode.

Acceptance criteria

• Per-turn override does not change persistent session state.
• Resolved model for the turn is reported in the response envelope (so agents can populate a visible per-turn signal in their reply).
• Works with a new modelOverrideSource: "agent" source kind so it does not collide with user overrides.
• Logged as an override event in any tooling that already consumes sessions.patch.

Frame

Required infrastructure for cost-discipline at scale on multi-tier model routing. Sticky-switch works but burdens the agent with discipline drift; atomic per-turn override is the structural fix. Pairs naturally with a visible per-turn model signal in agent replies, which is otherwise hard to populate accurately from a sticky-switch implementation.