Repository navigation
[1/1 Live model routing] Select backends behind a fixed Codex model - #161
Conversation
…ection Keep provider save feedback separate from Codex connection state. Supply an explicit private catalog without changing the original OpenAI provider identity, and bound repeated native turn requests to prevent duplicate upstream inference.
Use a two-step provider editor with encrypted draft and selection persistence. Load saved fields without treating a displayed key as a replacement, and add explicit reveal, copy, reload and reuse actions. Keep save failures editable and route a requested restart through the existing owner flow after the dialog closes.
Keep one Codex-facing model and persist the actual model choice in Switch. Pin admission, streaming and retry snapshots while retaining legacy routes, history/image guards, original sign-in and reversible connection setup.
Gaebabi
left a comment
Gaebabi
left a comment
There was a problem hiding this comment.
Author self-review of 04ee1ea (not an independent approval). Reviewed fixed-catalog/default bootstrap and reversible restoration, the global actual-model scope, saved selection and discovery-cache validation, admission before upload, model/key/reasoning snapshots through streams and bounded retries, cross-selection native duplicate refusal, existing route/auth/image/history guards, recovery selection identity and both desktop menus. No dependency/workflow or owner-data changes are included.
All six latest protected contexts pass on current main. Local verification passes 292 provider/current-native-CLI contracts, 124 offscreen WPF checks, 350 account/recovery checks and 42 language API checks with nine 535-message catalogs and both desktop builds. The same current native process/chat routes to A/B/OpenAI and a provider added after startup without changing its model list. The attached pictures are fictional production-component renders. Personal Codex picker/external-provider acceptance and a public binary release remain separate. The owner authorized CI-gated merge and local Switch update while preserving live Codex; initial owner connection/restart is deferred.
A running Codex retains its startup model catalog, so changing provider models cannot update that picker in place. This change exposes one stable
codex-switchmodel/default and moves actual OpenAI/custom model selection to a direct dropdown on Switch Home and provider setup.Selection applies globally to new requests using the virtual model. The gateway snapshots the actual model, destination, key and reasoning default before body upload, retains them through streams/retries, and refuses a duplicate native turn request instead of retargeting it after a selection change. Missing/deleted targets fail explicitly. Original OpenAI authentication, native image routes, history/image compatibility guards and older custom aliases are retained. An existing connection upgrades only through explicit Connect; initial activation still needs one idle Codex restart. Later selections, saved reasoning defaults and newly added providers do not rewrite the fixed catalog or restart Codex.
Validation:
0.167.0-alpha.3.2: same app-server and chat route A/B/OpenAI, use a provider added after startup, and change reasoning defaults while its native model list remains fixed.Application-component evidence uses fictional data and is rendered from the production controls offscreen; it is not a native owner-window capture. Before setup / scope and source hashes.
Refs #138; its broader desktop/provider acceptance remains open.