# Codex model protocol fixtures These use synthetic models/providers, never guessed production identifiers. Wire shapes were checked against the official app-server documentation and the generated protocol source matching the installed package manifest (Codex 0.163.0): - https://developers.openai.com/codex/app-server - https://github.com/openai/codex/blob/rust-v0.163.0/codex-rs/app-server-protocol/schema/typescript/v2/ModelListResponse.ts - Same directory: `ModelListParams.ts`, `Model.ts`, `Thread.ts`, `ThreadStartResponse.ts`, `TurnStartParams.ts`. `Model.model` is the turn override, not the catalog row's opaque `id`. `Thread.model` is current configured/latest persisted metadata, not per-turn execution telemetry. `thread/start` also returns model/provider at the top level. The fixtures project fields consumed by this adapter and retain real casing and pagination shapes. Tests use an isolated fake WebSocket JSON-RPC server. Local `codex --version` timed out before output, so schema generation was not reached; the installed `codex-package.json` identified version 0.163.0, so version-tagged generated source was read instead. No live app-server socket was contacted. Model selection validates via `thread/read` + `model/list`, persists a Pipelines next-turn choice, and supplies `turn/start.model` only when the user sends a prompt. It never calls `thread/resume`, which can awaken a queued goal, nor any configuration-writing API. The acknowledgement means "saved for the next turn," not "remote settings changed." The override is supplied again after a Pipelines restart; an already-running turn is not reconfigured.