Skip to content

The VS Code extension's Settings view has a Models group, what troupe models --json lists for the folder, each model with its window, its price and where they came from, the three roles marked, and one its provider does not serve said so and given no window; --refresh only from the group's own button.

Decision 794 · accepted · 2026-10-05 · issue #387

  • Troupe's answer, not the extension's reading. The group is the rows models.ts makes of what troupe models --json --workspace <folder> prints, run where the terminal runs as config --explain --json is (765). The extension reads neither the config nor the catalog's cache and asks no provider, so the group cannot disagree with troupe models or a session. The words follow the text report's: a window in whole thousands (131k), millions past a million, prices to the places troupe models prints them ($0.10/$0.50, dollars a million tokens in/out), (models.prices) after a price the config gives.
  • Where. In the Settings view, under the Model group (the model in use, then the models to choose from), before what else is changed. More than 20 models start the group folded, so a gateway's list does not push the files and problems out of sight.
  • A row a model. The id; then its roles, its window, its price and where they came from: from <provider>'s list (source: catalog), from your config (config or yaml) or from opencode; no key when key is false. The default, cheap and expensive models are starred and named. A role left unset resolves to the default's id (783), so the default model then says all three, which is what a session would use. Hover shows the whole of it: the window in tokens, the price in and out, who said so.
  • Not served is said, and has no window. A model with served: false has a warning, not served by <provider>, and on hover what the provider does serve (nearest). Its context is Config.models/1's fallback, not a window anyone said (defects D70), so it is not shown, rather than waiting on troupe_core to send null. A price the config gives it is still shown: someone wrote it down. served: null (the provider has not answered) shows the model as it is, with that said on hover.
  • The providers' lists. One row each, before the models, as the text's catalog: lines come first: how many models and when, or that it did not answer, when and why, and how much of it the cache still has.
  • When it asks, and --refresh. With the settings, on the same occasions (the view shown, the folder worked in changing, a file the answer named saved, Refresh), and without --refresh: troupe models asks the providers itself when its cache is stale (778). --refresh asks every provider now, with the person's keys, up to 30 seconds each, so only an explicit act passes it: the button on the group's own row, troupe.refreshModels, "Ask the Providers for Their Models", also in the palette. The settings are asked first and shown while the models are asked for, one troupe at a time; the models' call may take 120 seconds, the settings' 30.
  • When it cannot. No troupe: the one sentence of 765 is the whole view. troupe models failing (a config that does not load, a troupe before 0.8.3, which has no --json, a timeout): the group is one line, the first line of its standard error, the rest and the version on hover, and the settings stay. When the settings could not be read either, the models are asked all the same, since troupe models says on its standard error why a config does not load.
  • No key, even in an error. The JSON has none (783). A failure's text is shown only after anything a key could be is taken out (the value after api_key, token, secret or password, a bearer token, an sk-…); the settings' own failure line goes through the same, as one helper makes both.

Governs: clients/vscode/src/models.ts, clients/vscode/src/settingsView.ts