fix(providers): default DeepSeek V4 Flash to Responses

Author: estelledc

docs-site/src/content/docs/reference/configuration.md

@@ -325,7 +325,7 @@ or bind the forward explicitly to loopback (`ssh -L 127.0.0.1:20100:localhost:10
 | `modelReasoningEfforts?` | `Record<string,string[]>` | Model-specific reasoning labels. An empty list hides the effort control for that model. |
 | `modelSupportsReasoningSummaries?` | `Record<string,boolean>` | Model-specific reasoning-summary capability. Set a model to `false` to stop advertising summaries and strip summary-delivery fields before an `openai-responses` request. |
 | `modelReasoningSummaryDelivery?` | `Record<string,"sequential" \| "sequential_cutoff" \| "concurrent" \| "concurrent_cutoff">` | Model-specific Responses delivery enum. A configured model stays summary-capable and only an existing `stream_options.reasoning_summary_delivery` is rewritten; do not combine it with a `false` summary capability for the same model. |
-| `modelAdapters?` | `Record<string,string>` | Per-model wire override for a gateway that fronts models speaking different wires. Keys are upstream native model ids; values must be `openai-chat` or `openai-responses`. Useful when one model needs the Responses API for hosted tools such as `web_search` while its siblings are fine on chat completions. Models the upstream pins to a single wire, and the canonical ChatGPT forward provider, reject overrides. |
+| `modelAdapters?` | `Record<string,string>` | Per-model wire override for a gateway that fronts models speaking different wires. Keys are upstream native model ids; values must be `openai-chat` or `openai-responses`. Built-in registry defaults may select a verified mixed-wire route automatically (the DeepSeek preset sends `deepseek-v4-flash` over native Responses); an explicit entry wins and can opt a model back out. Models the upstream pins to a single wire, and the canonical ChatGPT forward provider, reject overrides. |
 | `reasoningEffortMap?` | `Record<string,string>` | Provider-wide wire aliases for reasoning labels. Use only when the upstream expects a different value. |
 | `modelReasoningEffortMap?` | `Record<string,Record<string,string>>` | Model-specific wire aliases for reasoning labels. |
 | `noReasoningModels?` | `string[]` | Models that reject a reasoning/thinking param — the adapter drops `reasoning_effort` for them. |

docs-site/src/content/docs/zh-cn/reference/configuration.md

@@ -172,6 +172,7 @@ x-opencodex-api-key: your-secret-token
 | `modelReasoningEfforts?` | `Record<string,string[]>` | 模型级 reasoning label。空数组会隐藏该模型的 effort 控件。 |
 | `modelSupportsReasoningSummaries?` | `Record<string,boolean>` | 模型级 reasoning summary 能力。设为 `false` 时不再声明 summary 支持，并在 `openai-responses` 请求前移除 summary-delivery 字段。 |
 | `modelReasoningSummaryDelivery?` | `Record<string,"sequential" \| "sequential_cutoff" \| "concurrent" \| "concurrent_cutoff">` | 模型级 Responses delivery enum。已配置模型保持 summary 能力，适配器只改写现有的 `stream_options.reasoning_summary_delivery`；同一模型不能同时将 summary 能力设为 `false`。 |
+| `modelAdapters?` | `Record<string,string>` | 面向同一 gateway 混合使用不同 wire 的模型级覆盖。键是上游原生模型 id，值只能是 `openai-chat` 或 `openai-responses`。已验证的混合 wire 路由会由 registry 自动提供默认值（DeepSeek preset 会让 `deepseek-v4-flash` 使用原生 Responses）；显式配置优先，也可以把模型切回 Chat。上游固定单一 wire 的模型和 canonical ChatGPT forward provider 不接受覆盖。 |
 | `reasoningEffortMap?` | `Record<string,string>` | provider 级 reasoning label wire alias。只在上游需要不同值时使用。 |
 | `modelReasoningEffortMap?` | `Record<string,Record<string,string>>` | 模型级 reasoning label wire alias。 |
 | `noReasoningModels?` | `string[]` | 拒绝 reasoning/thinking 参数的模型；adapter 会为它们移除 `reasoning_effort`。 |

src/providers/registry.ts

@@ -123,6 +123,12 @@ export interface ProviderRegistryEntry {
   defaultModel?: string;
   models?: string[];
   liveModels?: boolean;
+  /**
+   * Registry-only per-model wire defaults for mixed OpenAI-compatible gateways.
+   * These are intentionally not seeded into saved config: an explicit `modelAdapters`
+   * entry must remain distinguishable and must always win over a default.
+   */
+  modelWireDefaults?: Record<string, string>;
   modelDiscovery?: ProviderModelDiscoverySpec;
   contextWindow?: number;
   modelContextWindows?: Record<string, number>;
@@ -813,6 +819,9 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     models: ["deepseek-chat", "deepseek-reasoner", ...DEEPSEEK_THINKING_MODELS],
     defaultModel: "deepseek-v4-flash",
     modelContextWindows: { "deepseek-v4-flash": 1_000_000, "deepseek-v4-pro": 1_000_000 },
+    // DeepSeek documents V4-Flash as a native Responses API model adapted for Codex. The
+    // API id is `deepseek-v4-flash`; `DeepSeek-V4-Flash-0731` is a release/version label.
+    modelWireDefaults: { "deepseek-v4-flash": "openai-responses" },
     /* [Decision Log]
     - 목적: DeepSeek V4 thinking mode multi-turn/tool-call requests must replay prior assistant reasoning_content.
     - 대안 분석: Globally preserve reasoning_content for all OpenAI-compatible models; preserve it for legacy deepseek-reasoner too; mark only V4 thinking models in registry metadata.
@@ -1235,6 +1244,25 @@ export function providerMatchesRegistryTransport(
   return normalizedProviderEndpoint(provider.baseUrl) === normalizedProviderEndpoint(entry.baseUrl);
 }
 
+/**
+ * Resolve a registry-only default for a mixed-wire provider. Defaults only move a provider
+ * between the two OpenAI-shaped adapters and never override a provider configured on another
+ * wire. The resolver receives the allow-list so this helper cannot accidentally widen the
+ * adapter-selection boundary when a new registry entry is added.
+ */
+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 entry = getProviderRegistryEntry(id);
+  if (!entry?.modelWireDefaults || !providerMatchesRegistryTransport(id, provider)) return undefined;
+  const wire = entry.modelWireDefaults[modelId.trim().toLowerCase()];
+  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/server/adapter-resolve.ts

@@ -9,10 +9,12 @@ 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 { providerModelWireDefault } from "../providers/registry";
 
 /**
  * Resolve the wire a single model should use: a hard pin first, then a configured
- * per-model override, then the provider's own adapter.
+ * per-model override, then a registry default for a mixed-wire provider, then the provider's
+ * own adapter.
  *
  * Safe to call more than once on its own output — the pin check does not look at the
  * current adapter, so a second pass cannot let an override displace a pin.
@@ -24,7 +26,12 @@ export function resolveWireProtocolOverride(providerName: string, modelId: strin
   }
   // Re-check the allow-list here, not just in the config validator: the file may have
   // been hand-edited, or written by a build that allowed more values.
-  const requested = providerConfig.modelAdapters?.[modelId];
+  const configured = providerConfig.modelAdapters?.[modelId];
+  // An explicit allowed override wins, including one naming the provider-wide adapter (the
+  // opt-out from a registry default). Invalid hand-edited values fall through to the default.
+  const requested = configured && MODEL_ADAPTER_OVERRIDE_ALLOWED.has(configured)
+    ? configured
+    : providerModelWireDefault(providerName, providerConfig, modelId, MODEL_ADAPTER_OVERRIDE_ALLOWED);
   if (requested
     && MODEL_ADAPTER_OVERRIDE_ALLOWED.has(requested)
     && requested !== providerConfig.adapter

tests/adapter-resolve.test.ts

@@ -90,3 +90,42 @@ describe("per-model wire override (#404)", () => {
       .toBe("openai-responses");
   });
 });
+
+describe("registry per-model wire defaults", () => {
+  function deepseek(overrides: Partial<OcxProviderConfig> = {}): OcxProviderConfig {
+    return gateway({
+      baseUrl: "https://api.deepseek.com",
+      authMode: "key",
+      ...overrides,
+    });
+  }
+
+  test("routes only the official Flash API id through Responses", () => {
+    expect(resolveWireProtocolOverride("deepseek", "deepseek-v4-flash", deepseek()).adapter)
+      .toBe("openai-responses");
+    expect(resolveWireProtocolOverride("deepseek", "deepseek-v4-pro", deepseek()).adapter)
+      .toBe("openai-chat");
+    // The dated release label is not the API model id and must not be silently rewritten.
+    expect(resolveWireProtocolOverride("deepseek", "deepseek-v4-flash-0731", deepseek()).adapter)
+      .toBe("openai-chat");
+  });
+
+  test("an explicit Chat override opts Flash back out of the default", () => {
+    const provider = deepseek({ modelAdapters: { "deepseek-v4-flash": "openai-chat" } });
+    expect(resolveWireProtocolOverride("deepseek", "deepseek-v4-flash", provider).adapter)
+      .toBe("openai-chat");
+  });
+
+  test("defaults do not apply when the provider is already on another wire", () => {
+    expect(resolveWireProtocolOverride("deepseek", "deepseek-v4-flash", deepseek({ adapter: "anthropic" })).adapter)
+      .toBe("anthropic");
+  });
+
+  test("keeps provider credentials and destination untouched", () => {
+    const provider = deepseek({ apiKey: "test-key" });
+    const resolved = resolveWireProtocolOverride("deepseek", "deepseek-v4-flash", provider);
+    expect(provider.adapter).toBe("openai-chat");
+    expect(resolved.apiKey).toBe("test-key");
+    expect(resolved.baseUrl).toBe("https://api.deepseek.com");
+  });
+});