nodwyn.
Menu

ADR 0040: A conversation's model is a choice

  • Status: accepted; implemented 2026-10-03 (slice 1: the log, the session and the wire)
  • Date: 2026-10-03

Context

Configuring models already separates the key from the model: a server ([servers.NAME]) holds one address and one key, [models.NAME] names a model on it (with its context length, reply cap, sampling and prices), and [control] is the model a session talks to unless a strategy asks another. A provider such as OpenRouter offers hundreds of models at one address under one key, and models.offered already lists them to the console without sending the key.

What was fixed was the conversation: Options.Config.Control was read when the session was built, and only Settings › Model (one form, one [control] table) or a hand-edited config.toml could change it — for the next session. A strategy could hand one turn to another model ((model-turn "NAME"), ADR 27), but a person could not choose the model a conversation talks to, and could not move one to another model as it went on. So "one model per key" was true of the app, not of the storage.

Decision

  • A conversation's model is a choice. It is config.toml's [control] model, a [models.NAME] entry, or a model on a configured [servers.NAME] — the last needs no [models.NAME] block, so the models a server lists can be chosen without writing one line each. The chosen model answers (model-turn) from the next turn on; (model-turn "NAME") still asks a named model for one turn, unchanged.
  • The choice is the conversation's, so it is in its log. session.model records the model id, where it is reached, the server and the registry name, who chose it, and the window's recipe from then on. It is written when a model is chosen for a conversation as it opens, and at each switch; session.start and session.resume record the model the session talks to when it opens, which is the chosen one. Resuming restores the last choice, so a conversation picks up where it was left; a choice config.toml no longer offers falls back to [control] and the log says so, rather than making the task unopenable.
  • Who writes it. Only the daemon, or another nodwyn process holding the log's lock, writes a session log: it lives in nodwyn's data directory, never in a repository, so no stranger's clone can point a conversation at another model. The restore reads only what such a writer put there, and resolves it against config.toml again, so a recorded choice can reach nothing the person has not configured.
  • One key per server, unchanged. A switch brings the model, never a key and never a binding: a model on another server is reached with that server's key, and every request checks its binding (egress.check) as before. No choice carries an address or a credential class, and a model whose key is in a keychain this task does not have is refused with what to do about it.
  • Between turns, never mid-request. A switch is refused while a turn or an attachment is running: the turn under way has already asked its model, and an attachment was confirmed for the model it named.
  • Images are asked of the model that answers. An attached image goes only to the host it was confirmed for (ADR 22 §3), and only to a model that takes images. Whether one does is asked of the model, not of the name control: a switch forgets what the previous model said, so on a server that offers both kinds (OpenRouter's shape, one host) an image is withheld from a text-only model, with the reason recorded.
  • The window is sized for the model that answers. A switch sizes the water marks, the result cap and the ceiling as opening the conversation on that model would (config.Window.Sized): from its num_ctx, or from what the server's model list says of its context when config.toml does not set one, and the log says how in a note. session.model carries the recipe the requests after it are built by, as session.resume does, and nodwyn replay --window takes it from there, so every request still replays to its recorded hash. The tool catalog and the prompts stay the session's (ADR 9).
  • Refused when the target cannot hold the conversation. The window built with the marks sized for the target must fit its context less the room its reply cap needs; user messages are never trimmed, so a conversation can outgrow any window. Otherwise the switch says so and says to compact first. A model whose context neither config.toml nor its server says is left to its server, and nothing is judged.
  • It is this computer's act. A paired device opens a task on the [control] model; choosing another, at session.open or with session.model.set, is refused to it: another device answers and steers a task, but where its content goes stays with the computer nodwyn runs on (ADR 30). The daemon refuses it by the device's via, as it refuses an unwatched run, and the bridge does not offer session.model.set to a device at all.
  • Methods. session.model.set {id, name} or {id, server, model}, answering with the model now; session.open takes the same fields, so a conversation can start on a chosen model; session details carry model (name, server, model), never a key.

Consequences

  • The console and the TUI can offer a picker (the model list already arrives through models.offered, models.status and now session.info). Slice 1 builds neither, nor a Settings editor for [models.NAME]: the wire, the log and the session are what exist.
  • A conversation's model is no longer one value per machine, so a run's model is the conversation's: a run started on a switched session asks the switched model.
  • A conversation moved to a model with a smaller context is trimmed by the window sized for it from the next turn, as one opened on that model would be. The request prefix changes at the switch, so the server's prompt cache starts again; so it would anyway, since the cache is the model's.
  • The context gauges of the console and the TUI read the high water from session.model too, so they measure against the model that answers.
  • Two models on one server are two choices over one key, so choosing between them changes nothing about consent, bindings or the keychain: what changed is which one answers.

Amendment 2026-10-04: the picker, a queued choice, no fallback (ADR 41)

Slice 2 builds what slice 1 left: the console's picker and the Settings editor for [models.NAME]; the TUI's command is still not built. Two clauses of the Decision change:

  • "Between turns, never mid-request" no longer refuses a choice made while a turn runs. It is queued, said so in the state and the log, and applied when the turn ends, before the next can start; one that cannot apply says why and leaves the conversation on the model it had. An attachment in progress and an active run still refuse.
  • A recorded choice config.toml no longer offers no longer "falls back to [control]". The conversation opens, says what it was left on, and asks no turn until a model is chosen for it, so it is never answered by a model, or sent to a provider, nobody chose for it.

A choice also carries the effort its requests ask for (session.model effort), and session.model.set with the model the conversation is on changes the effort alone, without sizing or fitting anything again.

Amendment 2026-10-08: the TUI chooses too

The TUI's command, "still not built" above, is built: :model and :model NAME [EFFORT] (ADR 41, amended 2026-10-08).

Amendment 2026-10-17: the person's device chooses (ADR 30)

A device signed in with the person's passkey chooses a conversation's model, as this computer does (ADR 30, amended 2026-10-17); a device paired by a plain invitation still only sees it.