Domain decisions

Architecture: domain. Related: backends, Message.

2026-10-08 — Speaker rendered into the message content

Decision: Message.speaker is an optional non-empty single-line name. It is hashed into the snapshot id and written as "speaker" in snapshot JSON only when set. Backends send "<speaker>: <text>" under the unchanged role (message_chat_text); a message emptied by delete is sent empty. PlayerSpan.speaker carries it to labels and to_dict.

Why: Conversations where several people share the user role need to say who wrote what, and chat templates have no speaker slot.

Consequences: Snapshot ids and outputs without speakers do not change. Rejected: one role per speaker (templates accept only user, assistant and system); the OpenAI message name field (Hugging Face templates and the LM Studio SDK drop it, so backends would see different prompts).