fix(providers): address Codex review on Copilot Responses wire defaults

Move the per-model wire map onto PROVIDER_REGISTRY. A second provider fact
table next to the Copilot seed it had to stay aligned with is exactly the
drift-prone split src/AGENTS.md forbids. `modelWireDefaults` is registry-only
metadata, deliberately not part of ProviderConfigSeed: seeding it into a saved
config would freeze today's map on disk and make it indistinguishable from an
explicit user `modelAdapters` entry, which must always win. A registry
invariant test now pins every default to a seeded model and an allowed wire.

Keep sampling params on routed Responses gateways. Both compat handlers
stripped max_output_tokens for ANY openai-responses route, though the intent —
stated in claude-messages.ts as "routed providers keep them" — was the native
ChatGPT forward path only. Gate it on `authMode === "forward"`, matching the
adapter's own stripUnsupportedForwardParams condition so the two layers agree.
temperature/top_p/stop/user stay stripped on every Responses route: `stop` is
not a Responses parameter, reasoning models reject temperature/top_p, and
noTemperatureModels is honored only by the openai-chat adapter, so there is no
per-model filter on this wire.

Restore credential recovery for routed Responses gateways. Moving a model onto
this wire moves it from the normal adapter loop into the passthrough branch,
which formats non-2xx immediately — so a Copilot model would surface a 401 a
token refresh fixes, or a 429 a healthy pool key absorbs, while its chat-wire
siblings on the same provider still recover. Add a bounded recovery before the
error formatting: one forced OAuth refresh, then one replay per remaining pool
key. Nothing has streamed and the body is a replayable string, the same
precondition the transient-5xx retry relies on. The canonical ChatGPT forward
provider is excluded so its own pool-retry path is untouched.

Also feed registry `modelWireDefaults` and configured `modelAdapters` keys into
knownModelIdsForProvider, so a provider that names a model only in
`modelAdapters` is selectable as `provider/model` and appears in the bare-id
selector hint.

Docs: fix the heading nesting, and carry the gpt-5.4 text-only caveat and the
/chat/completions failure message into the ko/ja/zh-cn/ru locales.

Full suite: 6069 pass, 6 fail — all six also fail on the unmodified base
(claude-desktop-cli x2, cli-help service usage, provider-workspace-auth GUI
grep, decodeSchtasksOutput) except cli-restore-back, which passes in isolation
and times out only under full-suite load.

Author: mushikingh

docs-site/src/content/docs/guides/providers.md

@@ -90,7 +90,7 @@ ocx logout <provider>
 | `cursor` | `cursor` | `https://api2.cursor.sh` | Experimental PKCE login, live HTTP/2 transport, and account-filtered model discovery. |
 | `github-copilot` | `openai-chat` | `https://api.githubcopilot.com` | Experimental. GitHub device flow + `copilot_internal` exchange (VS Code OAuth client). Requires an active Copilot subscription; not an official third-party API. The catalog is mixed-wire — see [GitHub Copilot model wires](#github-copilot-model-wires). |
 
-#### GitHub Copilot model wires
+### GitHub Copilot model wires
 
 Copilot's catalog is not uniformly served by `/chat/completions`. The provider adapter stays
 `openai-chat`, which is correct for the Claude, Gemini, GPT-3.5/4 and `gpt-5-mini` entries, but the

docs-site/src/content/docs/ja/guides/providers.md

@@ -199,8 +199,12 @@ Amazon Bedrock ネイティブ API のような、これらの実装のいずれ
 サポートします(非公式ブリッジ — VS Code 公開クライアント ID でログイン後、短期 Copilot API トークンに
 交換し、有効な Copilot サブスクリプションが必要で GitHub ポリシー変更でブロックされる可能性あり)。Copilot の
 カタログは wire が混在します — `gpt-5.3-codex`、`gpt-5.4`、`gpt-5.4-mini`、`gpt-5.5`、`gpt-5.6-luna`、
-`gpt-5.6-sol`、`gpt-5.6-terra` は Responses API でのみ応答するため、opencodex が自動的に
-`openai-responses` へルーティングします。明示的な `modelAdapters` の指定が常に優先されます。
+`gpt-5.6-sol`、`gpt-5.6-terra` は Chat Completions では
+`model "…" is not accessible via the /chat/completions endpoint` で失敗するため、opencodex が自動的に
+`openai-responses` へルーティングします。`gpt-5.4` はテキストのみの Chat Completions リクエストなら
+成功しますが、Codex が function tool と reasoning effort を伴う実リクエストを送ると失敗するため、
+既定では Responses のままです。その他のモデルは `openai-chat` のままです。明示的な
+`modelAdapters` の指定が常に優先されます。
 `gpt-5.6-sol` のような bare な OpenAI 名は canonical `openai` provider 専用なので、
 `github-copilot/gpt-5.6-sol` のように provider を付けて選択してください。GitLab Duo は Bearer
 **サブスクリプショントークン**(通常の API キーではない)で認証します。**Cloudflare AI

docs-site/src/content/docs/ko/guides/providers.md

@@ -200,8 +200,11 @@ Amazon Bedrock 네이티브 API처럼 이 구현 중 어느 것과도 맞지 않
 지원합니다(비공식 브리지 — VS Code 공개 클라이언트 id로 로그인 후 단기 Copilot API 토큰으로
 교환하며, 활성 Copilot 구독이 필요하고 GitHub 정책 변경으로 막힐 수 있음). Copilot 카탈로그는
 wire가 섞여 있습니다 — `gpt-5.3-codex`, `gpt-5.4`, `gpt-5.4-mini`, `gpt-5.5`, `gpt-5.6-luna`,
-`gpt-5.6-sol`, `gpt-5.6-terra`는 Responses API로만 응답하므로 opencodex가 자동으로
-`openai-responses`로 라우팅합니다. 명시적인 `modelAdapters` 항목이 항상 우선합니다.
+`gpt-5.6-sol`, `gpt-5.6-terra`는 Chat Completions에서
+`model "…" is not accessible via the /chat/completions endpoint`로 실패하므로 opencodex가 자동으로
+`openai-responses`로 라우팅합니다. `gpt-5.4`는 텍스트 전용 Chat Completions 요청은 성공하지만
+Codex가 function tool과 reasoning effort를 함께 보내면 실패하므로 기본값은 Responses 유지입니다.
+나머지 모델은 `openai-chat` 그대로입니다. 명시적인 `modelAdapters` 항목이 항상 우선합니다.
 `gpt-5.6-sol` 같은 bare OpenAI 이름은 canonical `openai` provider 전용이므로
 `github-copilot/gpt-5.6-sol`처럼 provider를 붙여 선택해야 합니다. GitLab Duo는 Bearer
 **구독 토큰**(일반 API 키가 아님)으로 인증합니다. **Cloudflare AI

docs-site/src/content/docs/ru/guides/providers.md

@@ -91,7 +91,7 @@ ocx logout <provider>
 | `kiro` | `kiro` | `https://runtime.us-east-1.kiro.dev` | Первый вход импортирует существующую сессию после установки Kiro CLI (`curl -fsSL https://cli.kiro.dev/install | bash`) и входа через `kiro-cli login`. **Добавить аккаунт** выполняет выход из `kiro-cli`, запускает новый вход через браузер, переключает аккаунт самого `kiro-cli` и сохраняет метаданные профиля отдельно для каждого аккаунта. Существующие аккаунты OpenCodex сохраняются; при отмене или сбое восстанавливается предыдущая сессия `kiro-cli`. |
 | `google-antigravity` | `google` | `https://daily-cloudcode-pa.googleapis.com` | Google OAuth поверх протокола Cloud Code Assist. |
 | `cursor` | `cursor` | `https://api2.cursor.sh` | Экспериментальный PKCE-вход, живой транспорт HTTP/2 и обнаружение моделей с фильтрацией по аккаунту. |
-| `github-copilot` | `openai-chat` | `https://api.githubcopilot.com` | Экспериментально. Device flow GitHub + обмен `copilot_internal` (OAuth-клиент VS Code). Требуется активная подписка Copilot; это не официальный сторонний API. Каталог смешанный по wire: `gpt-5.3-codex`, `gpt-5.4`, `gpt-5.4-mini`, `gpt-5.5`, `gpt-5.6-luna`, `gpt-5.6-sol`, `gpt-5.6-terra` отвечают только на Responses API, и opencodex маршрутизирует их через `openai-responses` автоматически. Явная запись `modelAdapters` всегда важнее. Голые имена OpenAI зарезервированы за каноническим провайдером `openai`, поэтому выбирайте модель как `github-copilot/gpt-5.6-sol`. |
+| `github-copilot` | `openai-chat` | `https://api.githubcopilot.com` | Экспериментально. Device flow GitHub + обмен `copilot_internal` (OAuth-клиент VS Code). Требуется активная подписка Copilot; это не официальный сторонний API. Каталог смешанный по wire: `gpt-5.3-codex`, `gpt-5.4`, `gpt-5.4-mini`, `gpt-5.5`, `gpt-5.6-luna`, `gpt-5.6-sol`, `gpt-5.6-terra` падают на Chat Completions с `model "…" is not accessible via the /chat/completions endpoint`, поэтому opencodex автоматически маршрутизирует их через `openai-responses`. У `gpt-5.4` текстовый запрос к Chat Completions проходит, но реальный запрос Codex с function tools и reasoning effort — нет, поэтому по умолчанию он остаётся на Responses. Остальные модели каталога остаются на `openai-chat`. Явная запись `modelAdapters` всегда важнее. Голые имена OpenAI зарезервированы за каноническим провайдером `openai`, поэтому выбирайте модель командой `codex -m github-copilot/gpt-5.6-sol`. |
 
 Для канонических пресетов Kimi Coding Plan (вход через аккаунт `kimi` и API-ключ `kimi-code`)
 opencodex передаёт в запрос Chat Completions только стабильный `prompt_cache_key`, предоставленный

docs-site/src/content/docs/zh-cn/guides/providers.md

@@ -186,7 +186,10 @@ GPT-5.6 Sol/Terra/Luna 会预置在提供商的回退列表中，因此即使实
 `ocx login github-copilot` 使用 GitHub 设备流 OAuth 登录（非官方桥接 — 使用 VS Code 公开客户端 id
 登录后换取短期 Copilot API 令牌，需要有效的 Copilot 订阅，GitHub 政策收紧时可能失效）。Copilot 目录混用
 多种 wire —— `gpt-5.3-codex`、`gpt-5.4`、`gpt-5.4-mini`、`gpt-5.5`、`gpt-5.6-luna`、`gpt-5.6-sol`、
-`gpt-5.6-terra` 仅在 Responses API 上响应，opencodex 会自动将它们路由到 `openai-responses`；
+`gpt-5.6-terra` 在 Chat Completions 上会失败并返回
+`model "…" is not accessible via the /chat/completions endpoint`，opencodex 会自动将它们路由到
+`openai-responses`。`gpt-5.4` 的纯文本 Chat Completions 请求可以成功，但 Codex 携带 function tools
+和 reasoning effort 的真实请求会失败，因此默认仍走 Responses。其余模型保持 `openai-chat`。
 显式的 `modelAdapters` 配置始终优先。`gpt-5.6-sol` 这类裸 OpenAI 名称保留给 canonical `openai`
 provider，因此必须使用 `github-copilot/gpt-5.6-sol` 这样带 provider 前缀的选择器。GitLab Duo
 使用 Bearer **订阅令牌**（而非普通 API 密钥）进行认证。

src/providers/model-wire-defaults.ts

@@ -123,6 +123,16 @@ export interface ProviderRegistryEntry {
   defaultModel?: string;
   models?: string[];
   liveModels?: boolean;
+  /**
+   * Per-model wire defaults for a preset whose upstream catalog is mixed-wire, keyed by
+   * native model id. Values must be OpenAI-shaped wires (MODEL_ADAPTER_OVERRIDE_ALLOWED).
+   *
+   * Registry-only metadata, deliberately not part of `ProviderConfigSeed`: seeding it into
+   * a saved config would freeze today's map on disk and make it indistinguishable from an
+   * explicit user `modelAdapters` entry, which must always win over the default.
+   * Read through `providerModelWireDefault` at resolve time instead.
+   */
+  modelWireDefaults?: Record<string, string>;
   modelDiscovery?: ProviderModelDiscoverySpec;
   contextWindow?: number;
   modelContextWindows?: Record<string, number>;
@@ -1190,8 +1200,6 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     liveModels: true,
     // Seed refreshed 2026-07-30 from a live 37-entry Copilot catalog; embeddings and the
     // internal `trajectory-compaction` entry are omitted because they are not chat models.
-    // The GPT-5.3+ entries are Responses-only on this upstream — see
-    // `builtinModelWireDefault` in providers/model-wire-defaults.ts.
     models: [
       "claude-haiku-4.5", "claude-opus-4.5", "claude-opus-4.6", "claude-opus-4.7",
       "claude-opus-4.8", "claude-sonnet-4.5", "claude-sonnet-4.6", "claude-sonnet-5",
@@ -1200,6 +1208,22 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
       "gpt-5.3-codex", "gpt-5.4", "gpt-5.4-mini", "gpt-5.5",
       "gpt-5.6-luna", "gpt-5.6-sol", "gpt-5.6-terra",
     ],
+    // Copilot's catalog is mixed-wire: `openai-chat` above is right for the Claude, Gemini,
+    // GPT-3.5/4 and gpt-5-mini entries, but these answer only on the Responses API and fail
+    // chat completions with `model "…" is not accessible via the /chat/completions endpoint`.
+    // gpt-5.4 hides it behind a passing smoke test — a text-only chat request succeeds and
+    // only a real Codex request fails, because Codex supplies function tools plus a reasoning
+    // effort. Verified 2026-07-30 with real `codex exec` runs, not minimal text probes:
+    // presence in GET /models proves neither. Recheck when Copilot changes models.
+    modelWireDefaults: {
+      "gpt-5.3-codex": "openai-responses",
+      "gpt-5.4": "openai-responses",
+      "gpt-5.4-mini": "openai-responses",
+      "gpt-5.5": "openai-responses",
+      "gpt-5.6-luna": "openai-responses",
+      "gpt-5.6-sol": "openai-responses",
+      "gpt-5.6-terra": "openai-responses",
+    },
     defaultModel: "gpt-4o",
     note: "Experimental unofficial Copilot bridge. Logs in via GitHub device flow using the public VS Code OAuth client id, then exchanges for a short-lived Copilot API token (copilot_internal). Requires an active Copilot subscription. GitHub may tighten or revoke this path; do not send confidential material you would not paste into Copilot Chat.",
   },
@@ -1246,6 +1270,35 @@ export function providerMatchesRegistryTransport(
   return normalizedProviderEndpoint(provider.baseUrl) === normalizedProviderEndpoint(entry.baseUrl);
 }
 
+/** Copilot publishes dated snapshots of a family (`gpt-4o-2024-05-13`), so match those too. */
+const DATED_SNAPSHOT_SUFFIX = /-\d{4}-\d{2}-\d{2}$/;
+
+/**
+ * The wire a model should use when the user has configured no `modelAdapters` entry for it,
+ * or undefined when the provider-wide adapter already applies.
+ *
+ * Distinct from the hard pins in `types.ts`: a pin means the upstream speaks exactly one wire
+ * and a user override must not apply. This is a default — an explicit `modelAdapters` entry
+ * always wins, including one naming the provider-wide adapter to opt a model back out.
+ *
+ * `currentAdapter` is the provider's configured adapter: a default only nudges a provider
+ * between OpenAI-shaped wires, so a provider a user moved to some other adapter is left alone.
+ * A provider re-pointed away from the preset transport keeps nothing from the preset either.
+ */
+export function providerModelWireDefault(
+  id: string,
+  provider: Pick<OcxProviderConfig, "baseUrl" | "adapter"> & Partial<Pick<OcxProviderConfig, "authMode">>,
+  modelId: string,
+  allowedWires: ReadonlySet<string>,
+): string | undefined {
+  if (!allowedWires.has(provider.adapter)) return undefined;
+  const defaults = getProviderRegistryEntry(id)?.modelWireDefaults;
+  if (!defaults || !providerMatchesRegistryTransport(id, provider)) return undefined;
+  const model = modelId.trim().toLowerCase();
+  const wire = defaults[model] ?? defaults[model.replace(DATED_SNAPSHOT_SUFFIX, "")];
+  return wire !== undefined && allowedWires.has(wire) ? wire : undefined;
+}
+
 /**
  * Effective Codex account mode for a provider. For canonical `openai`, a valid persisted
  * `codexAccountMode` on the provider config wins and a missing/invalid value defaults to

src/providers/registry.ts

@@ -53,6 +53,10 @@ export function knownModelIdsForProvider(provName: string, prov: OcxProviderConf
     registry?.modelDefaultReasoningEfforts,
     registry?.modelReasoningEffortMap,
     registry?.modelMaxOutputTokens,
+    registry?.modelWireDefaults,
+    // A user who names a model in `modelAdapters` has asserted it exists on this provider,
+    // so it is selectable as `provider/model` even when no seed or cache lists it.
+    prov.modelAdapters,
   ]) {
     for (const id of Object.keys(map ?? {})) ids.add(id);
   }

src/router.ts

@@ -9,7 +9,7 @@ import { createResponsesPassthroughAdapter } from "../adapters/openai-responses"
 import type { OcxProviderConfig } from "../types";
 import { isWirePinnedModel, MODEL_ADAPTER_OVERRIDE_ALLOWED, pinnedWireAdapter } from "../types";
 import { isCanonicalOpenAiForwardProvider } from "../providers/openai-tiers";
-import { builtinModelWireDefault } from "../providers/model-wire-defaults";
+import { providerModelWireDefault } from "../providers/registry";
 
 /**
  * Resolve the wire a single model should use: a hard pin first, then a configured
@@ -31,7 +31,7 @@ export function resolveWireProtocolOverride(providerName: string, modelId: strin
   // provider-wide adapter, which is how a user opts a model out of a built-in default.
   const requested = configured && MODEL_ADAPTER_OVERRIDE_ALLOWED.has(configured)
     ? configured
-    : builtinModelWireDefault(providerName, modelId, providerConfig.adapter);
+    : providerModelWireDefault(providerName, providerConfig, modelId, MODEL_ADAPTER_OVERRIDE_ALLOWED);
   if (requested
     && MODEL_ADAPTER_OVERRIDE_ALLOWED.has(requested)
     && requested !== providerConfig.adapter

src/server/adapter-resolve.ts

@@ -88,7 +88,15 @@ export async function handleChatCompletions(
       directRoute = route.codexAccountMode === "direct";
       // ChatGPT backend rejects store:true and unsupported sampling knobs.
       internalBody.store = false;
-      delete internalBody.max_output_tokens;
+      // Forward mode is the only route that 400s on max_output_tokens, and the adapter's
+      // own stripUnsupportedForwardParams already drops it there on the same condition.
+      // A routed Responses gateway (Copilot, api.openai.com, ...) honors the field, and
+      // dropping it there silently ignores the caller's max_tokens.
+      if (route.provider.authMode === "forward") delete internalBody.max_output_tokens;
+      // temperature/top_p/stop/user stay stripped on every Responses route: `stop` is not
+      // a Responses parameter at all, reasoning models reject temperature/top_p, and the
+      // passthrough adapter relays the body verbatim — `noTemperatureModels` and friends
+      // are honored only by the openai-chat adapter, so there is no per-model filter here.
       delete internalBody.temperature;
       delete internalBody.top_p;
       delete internalBody.stop;