A model per subject, and the fallback

Both are off until set, and with neither set every subject answers on ROOM_MODEL.

Per subject

A subject may carry a model in subjects.json, which replaces ROOM_MODEL for that subject alone: {"id": "gem", "label": "Violet", "system_file": "gem.md", "model": "evaluation/mistral-medium-3.5"}. The room never shows it. Snapshots and the event stream drop model the way they drop system. A subject may also carry model_label, the provider's name for the model such as "GLM-4.7-FlashX", the only model text a snapshot shows. It is a plain name of at most 60 characters with no slash, so a route id is refused at load.

The fallback

ROOM_FALLBACK_MODEL (or --fallback-model) names one route for the whole room. When every answer to a prompt is terminal and any ended failed after its retry, the room logs a fallback event and asks every subject again on that route, so a round is never split across models. It happens once per prompt, and not after the presenter has picked it, since the answers would change under the grading room.

There is no default, so with none set a failed answer stays failed. Attendees and the presenter stream see that a switch happened and never the model. The event log and the server log carry the model name.

Route limits and retries

A provider may cap how many calls it takes at once. ROOM_ROUTE_LIMITS (--route-limits) is model=count,..., the most answers in flight on that route, and a route not listed has no cap. ROOM_RETRIES (--retries, default 1) is how many extra asks a 429 or 5xx gets, waiting 2s, 4s, ... inside the answer deadline. Each subject call sends x-agent-session-id: case:<case id>:<sha256 of the user message>, and the prompt carries that hash as input_sha256, shown to a viewer only with its text. The proxy drops the OpenAI user field and keeps that header.

Restart

Every event is appended before it is applied, and a start replays the log, dropping a torn last line. Answers a restart cut off are asked again. An answer that strips to nothing, a subject that tried to run a command, is asked once more, and each empty reply's raw text (2048 characters at most) is kept in a raw_empty log line that takes no rev and reaches no page. The switch is one log event, and replaying it re-queues every answer of that prompt. A restart mid-round neither loses the switch nor repeats it. Answers the restart cut off resume on the model the event recorded, whatever the restart's own setting says.

No Anthropic models

The room refuses to load a subject model or a fallback whose name contains claude or anthropic, in any case. That is a name heuristic only. What the proxy routes a name to is the real control, so an Anthropic model behind a neutral alias is not caught here.