feat(responses): add an opt-in bounded JSON fallback for custom providers

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

@@ -84,6 +84,7 @@ account を削除しても mapping は保持され、同じ id を再追加す
 | `modelSupportsReasoningSummaries?` | `Record<string, boolean>` |モデルを `false` に設定して、概要の広告を停止し、概要配信フィールドを削除します。 |
 | `modelReasoningSummaryDelivery?` | `Record<string, "sequential" \| "sequential_cutoff" \| "concurrent" \| "concurrent_cutoff">` |モデルごとの応答配信列挙型。既存の配信フィールドを書き換えます。 |
 | `modelAdapters?` | `Record<string, string>` | 混合配線ゲートウェイのモデルごとの `openai-chat` または `openai-responses` 配線オーバーライド。明示的なエントリはレジストリのデフォルトを破ります。DeepSeek のプリセットは `deepseek-v4-flash` のネイティブ Responses を選択でき、GitHub Copilot は GPT-5 ファミリー (`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`) を Responses 専用デフォルトとして宣言します。これらのモデルはエージェント トラフィックで `/chat/completions` を拒否するためです。`gpt-5.4-nano` のようなビルトイン デフォルトのないモデルはここでオプトインできます。単線アップストリーム ピンと正規の ChatGPT 転送はオーバーライドを拒否します。 |
+| `modelResponsesUpstreamStreaming?` | `Record<string, boolean>` | forward 以外の `openai-responses` プロバイダー向けモデル別 upstream Responses ポリシーです。`false` は upstream に bounded JSON を要求し、検証済み terminal オブジェクトを streaming client 用 Responses イベントへ再構成します。`true` は registry の `false` 既定値を明示的に上書きします。照合は大文字小文字を区別せず、public virtual id を優先し、最終 wire-model id をフォールバックに使います。この correctness-first fallback では incremental delta がなくなり、bounded JSON のサイズと timeout 制限が適用されます。 |
 | `modelPreferHostedTools?` | `Record<string,string[]>` | hosted tool namespace を予約する非 forward Responses gateway 向けの完全一致モデル opt-in。現在は `["image_generation"]` のみを受け付けます。一致したモデルは `openai-responses` wire を使い、その hosted tool をサポートする必要があります。競合するクライアント `image_gen` 宣言を除去し、呼び出し元の tool choice を維持するため selector も書き換えます。OpenAI API の仮想 `-pro` モデルでは、まず選択した公開 ID に一致させ、解決後のベース wire-model ID をフォールバックとして使用します。`modelAdapters` は公開 ID、次にベース ID の順に解決し、後者の結果が最終 wire を決めます。未設定のモデルは通常の alias 動作を維持します。 |
 | `reasoningEffortMap?` | `Record<string, string>` |ラベルを推論するためのプロバイダー全体のワイヤ エイリアス。 |
 | `modelReasoningEffortMap?` | `Record<string, Record<string, string>>` |推論ラベルのモデルごとのワイヤ エイリアス。 |

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

@@ -84,6 +84,7 @@ managed map을 활성화하면 privacy-safe selector를 만들고, 이후 계정
 | `modelSupportsReasoningSummaries?` | `Record<string, boolean>` | 모델을 `false`로 두면 summary 광고를 멈추고 summary 전달 필드를 제거합니다. |
 | `modelReasoningSummaryDelivery?` | `Record<string, "sequential" \| "sequential_cutoff" \| "concurrent" \| "concurrent_cutoff">` | 모델별 Responses 전달 enum입니다. 기존 delivery 필드를 다시 씁니다. |
 | `modelAdapters?` | `Record<string, string>` | 혼합 와이어 게이트웨이를 위한 모델별 `openai-chat` 또는 `openai-responses` 와이어 재정의입니다. 명시적 항목이 레지스트리 기본값보다 우선합니다. DeepSeek 프리셋은 `deepseek-v4-flash`에 네이티브 Responses를 선택할 수 있고, GitHub Copilot은 GPT-5 계열(`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`)을 Responses 전용 기본값으로 선언합니다. 이 모델들은 에이전트 트래픽에서 `/chat/completions`를 거부하기 때문입니다. `gpt-5.4-nano`처럼 기본값이 없는 모델은 여기서 직접 옵트인할 수 있습니다. 단일 와이어 상위 항목과 정식 ChatGPT forward는 재정의를 거부합니다. |
+| `modelResponsesUpstreamStreaming?` | `Record<string, boolean>` | Forward가 아닌 `openai-responses` provider의 모델별 upstream Responses 정책입니다. `false`는 upstream에 bounded JSON을 요청하고 검증된 terminal 객체를 streaming client용 Responses event로 다시 구성합니다. `true`는 registry의 `false` 기본값을 명시적으로 해제합니다. 대소문자를 구분하지 않으며 public virtual id가 먼저, 최종 wire-model id가 fallback으로 일치합니다. 이 correctness-first fallback에서는 incremental delta가 사라지고 bounded JSON 크기·시간 제한이 적용됩니다. |
 | `modelPreferHostedTools?` | `Record<string,string[]>` | hosted tool namespace를 예약하는 non-forward Responses gateway용 정확한 모델 ID opt-in입니다. 현재 `["image_generation"]`만 허용하며, 일치하는 모델은 `openai-responses` wire를 사용하고 해당 hosted tool을 지원해야 합니다. 충돌하는 클라이언트 `image_gen` 선언을 제거하고 호출자의 tool choice를 유지하도록 selector도 다시 씁니다. OpenAI API 가상 `-pro` 모델은 선택한 공개 ID를 먼저 일치시키고, 해석된 기본 wire-model ID를 대체값으로 사용합니다. `modelAdapters`는 공개 ID를 먼저, 그 다음 기본 ID를 해석하며, 두 번째 결과가 최종 wire를 결정합니다. 설정하지 않은 모델은 일반 alias 동작을 유지합니다. |
 | `reasoningEffortMap?` | `Record<string, string>` | reasoning 레이블의 공급자 전반 와이어 별칭입니다. |
 | `modelReasoningEffortMap?` | `Record<string, Record<string, string>>` | reasoning 레이블의 모델별 와이어 별칭입니다. |

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

@@ -94,6 +94,7 @@ differing backup and rewrites known legacy namespaced selected ids to bare ids.
 | `modelSupportsReasoningSummaries?` | `Record<string, boolean>` | Set a model to `false` to stop advertising summaries and strip summary-delivery fields. |
 | `modelReasoningSummaryDelivery?` | `Record<string, "sequential" \| "sequential_cutoff" \| "concurrent" \| "concurrent_cutoff">` | Per-model Responses delivery enum; rewrites an existing delivery field. |
 | `modelAdapters?` | `Record<string, string>` | Per-model `openai-chat` or `openai-responses` wire override for mixed-wire gateways. Explicit entries beat registry defaults; DeepSeek's preset can select native Responses for `deepseek-v4-flash`, and GitHub Copilot declares Responses-only defaults for its GPT-5 family (`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`) because those models reject `/chat/completions` for agent traffic. Models without a built-in default (for example `gpt-5.4-nano`) can be opted in here. Single-wire upstream pins and canonical ChatGPT forward reject overrides. |
+| `modelResponsesUpstreamStreaming?` | `Record<string, boolean>` | Per-model Responses upstream policy for non-forward `openai-responses` providers. `false` asks upstream for bounded JSON and reframes the validated terminal object as Responses events for a streaming client; `true` explicitly overrides a registry `false` default. Matching is case-insensitive and supports public virtual ids with their final wire-model id as fallback. This correctness-first fallback removes incremental deltas and remains subject to the bounded JSON size and timeout limits. |
 | `modelPreferHostedTools?` | `Record<string,string[]>` | Exact-model opt-in for non-forward Responses gateways that reserve a hosted-tool namespace. Currently accepts only `["image_generation"]`; a matching model must use the `openai-responses` wire and support that hosted tool. It removes colliding client `image_gen` declarations and rewrites their selectors to preserve caller tool choice. For OpenAI API virtual `-pro` models, the selected public ID is matched first and the resolved base wire-model ID is a fallback. `modelAdapters` resolves the public ID first, then the base ID; the second resolution determines the final wire. Other models retain normal alias behavior. |
 | `reasoningEffortMap?` | `Record<string, string>` | Provider-wide wire aliases for reasoning labels. |
 | `modelReasoningEffortMap?` | `Record<string, Record<string, string>>` | Per-model wire aliases for reasoning labels. |

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

@@ -97,6 +97,7 @@ cross-route credential fallback не существует. Строки API GPT-
 | `modelSupportsReasoningSummaries?` | `Record<string, boolean>` | Установите `false` для модели, чтобы перестать рекламировать summary и вырезать поля доставки summary. |
 | `modelReasoningSummaryDelivery?` | `Record<string, "sequential" \| "sequential_cutoff" \| "concurrent" \| "concurrent_cutoff">` | Responses delivery enum по моделям; переписывает уже существующее поле delivery. |
 | `modelAdapters?` | `Record<string, string>` | Wire-override по модели для `openai-chat` или `openai-responses` в gateway с несколькими wire-форматами. Явные записи имеют приоритет над default'ами registry; preset DeepSeek может выбирать native Responses для `deepseek-v4-flash`, а GitHub Copilot объявляет Responses-only default'ы для семейства GPT-5 (`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`), потому что эти модели отклоняют `/chat/completions` для агентного трафика. Модели без встроенного default'а (например, `gpt-5.4-nano`) можно включить здесь. Single-wire upstream pin'ы и canonical ChatGPT forward override не принимают. |
+| `modelResponsesUpstreamStreaming?` | `Record<string, boolean>` | Модельная политика upstream Responses для non-forward провайдеров `openai-responses`. `false` запрашивает bounded JSON и преобразует проверенный terminal-объект в события Responses для streaming client; `true` явно перекрывает registry default `false`. Сопоставление не зависит от регистра: сначала используется public virtual id, затем итоговый wire-model id. Этот correctness-first fallback убирает incremental delta и подчиняется лимитам размера и timeout bounded JSON. |
 | `modelPreferHostedTools?` | `Record<string,string[]>` | Opt-in для точного model ID в non-forward Responses gateway, который резервирует namespace hosted tool. Сейчас допускается только `["image_generation"]`; совпавшая модель должна использовать wire `openai-responses` и поддерживать этот hosted tool. Прокси удаляет конфликтующие клиентские объявления `image_gen` и переписывает их selectors, сохраняя caller tool choice. Для виртуальных моделей OpenAI API `-pro` сначала сопоставляется выбранный публичный ID, а затем в качестве fallback используется ID базовой wire-модели. `modelAdapters` сначала разрешается по публичному ID, затем по базовому ID; второй результат определяет итоговый wire. Остальные модели сохраняют обычное alias-поведение. |
 | `reasoningEffortMap?` | `Record<string, string>` | Provider-wide wire-alias'ы для reasoning-label'ов. |
 | `modelReasoningEffortMap?` | `Record<string, Record<string, string>>` | Wire-alias'ы для reasoning-label'ов по отдельным моделям. |

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

@@ -84,6 +84,7 @@ selector，而不是分配一个新名称。
 | `modelSupportsReasoningSummaries?` | `Record<string, boolean>` | 将某个模型设为 `false`，即可停止暴露摘要并移除摘要交付字段。 |
 | `modelReasoningSummaryDelivery?` | `Record<string, "sequential" \| "sequential_cutoff" \| "concurrent" \| "concurrent_cutoff">` | 按模型设置的 Responses 交付枚举；会重写现有的 delivery 字段。 |
 | `modelAdapters?` | `Record<string, string>` | 按模型设置的 `openai-chat` 或 `openai-responses` 线协议覆盖项，用于混合线协议网关。显式条目优先于注册表默认值；DeepSeek 预设可以为 `deepseek-v4-flash` 选择原生 Responses，GitHub Copilot 则为 GPT-5 系列（`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`）声明了 Responses 专用默认值，因为这些模型在代理流量下会拒绝 `/chat/completions`。没有内置默认值的模型（例如 `gpt-5.4-nano`）可以在此手动启用。单一线协议上游固定项和规范 ChatGPT forward 会拒绝覆盖。 |
+| `modelResponsesUpstreamStreaming?` | `Record<string, boolean>` | 非 forward `openai-responses` provider 的逐模型 upstream Responses 策略。`false` 请求 bounded JSON，并将验证后的 terminal 对象重构为 streaming client 所需的 Responses 事件；`true` 显式覆盖 registry 的 `false` 默认值。匹配不区分大小写，优先使用 public virtual id，并以最终 wire-model id 作为回退。此 correctness-first fallback 不提供 incremental delta，并受 bounded JSON 大小与 timeout 限制。 |
 | `modelPreferHostedTools?` | `Record<string,string[]>` | 非 forward Responses gateway 的精确模型 ID opt-in，用于上游预留 hosted tool namespace 的情况。目前只支持 `["image_generation"]`；匹配模型必须使用 `openai-responses` wire 且支持该 hosted 工具。它会移除冲突的客户端 `image_gen` 声明，并改写其 selector 以保持调用方的 tool choice。对于 OpenAI API 的虚拟 `-pro` 模型，先匹配所选公开 ID，未命中时才使用解析出的基础 wire-model ID 作为回退。`modelAdapters` 会先按公开 ID、再按基础 ID 解析；后一次结果决定最终 wire。未配置模型保持普通 alias 行为。 |
 | `reasoningEffortMap?` | `Record<string, string>` | 提供者级、用于推理标签的线协议别名。 |
 | `modelReasoningEffortMap?` | `Record<string, Record<string, string>>` | 按模型设置的推理标签线协议别名。 |

src/config.ts

@@ -616,6 +616,7 @@ const providerConfigSchema = z.object({
   requiresAdjacentResponsesToolResults: z.boolean().optional(),
   supportsServiceTier: z.boolean().optional(),
   preserveResponsesReasoningContent: z.boolean().optional(),
+  modelResponsesUpstreamStreaming: z.record(z.string(), z.boolean()).optional(),
   allowPrivateNetwork: z.boolean().optional(),
   retryOn429: retryOn429PolicySchema.optional(),
   codexAccountMode: z.enum(["pool", "direct"]).optional(),
@@ -908,6 +909,83 @@ export function modelAdapterRecordConfigError(
   return null;
 }
 
+/** Validate the opt-in bounded-JSON policy against the model's effective Responses wire. */
+export function modelResponsesUpstreamStreamingConfigError(
+  value: unknown,
+  field: string,
+  providerName: string,
+  provider: { adapter?: unknown; authMode?: unknown; baseUrl?: unknown; modelAdapters?: unknown },
+): string | null {
+  const shapeError = booleanRecordConfigError(value, field);
+  if (shapeError) return shapeError;
+  const entries = Object.entries((value ?? {}) as Record<string, boolean>);
+  if (entries.length === 0) return null;
+
+  const registry = getProviderRegistryEntry(providerName);
+  const registryTransportMatches = typeof provider.baseUrl === "string"
+    && providerMatchesRegistryTransport(providerName, {
+      baseUrl: provider.baseUrl,
+      adapter: provider.adapter as OcxProviderConfig["adapter"],
+      ...(typeof provider.authMode === "string"
+        ? { authMode: provider.authMode as OcxProviderConfig["authMode"] }
+        : {}),
+    });
+  const effectiveForwardAuth = registryTransportMatches
+    ? registry?.authKind === "forward"
+    : provider.authMode === "forward";
+  if (effectiveForwardAuth) {
+    return `${field} is not supported on forward-auth Responses providers`;
+  }
+
+  const resolveEffectiveWire = (modelId: string, currentWire: unknown): unknown => {
+    const pinned = pinnedWireAdapter(providerName, modelId);
+    if (pinned) return pinned;
+    const configured = provider.modelAdapters && typeof provider.modelAdapters === "object"
+      && !Array.isArray(provider.modelAdapters)
+      ? (provider.modelAdapters as Record<string, unknown>)[modelId]
+      : undefined;
+    if (typeof configured === "string" && MODEL_ADAPTER_OVERRIDE_ALLOWED.has(configured)) {
+      return configured;
+    }
+    const registryDefault = typeof currentWire === "string" && typeof provider.baseUrl === "string"
+      ? providerModelWireDefault(
+        providerName,
+        {
+          baseUrl: provider.baseUrl,
+          adapter: currentWire,
+          ...(typeof provider.authMode === "string"
+            ? { authMode: provider.authMode as OcxProviderConfig["authMode"] }
+            : {}),
+        },
+        modelId,
+        MODEL_ADAPTER_OVERRIDE_ALLOWED,
+        "responses",
+      )
+      : undefined;
+    return registryDefault ?? currentWire;
+  };
+
+  for (const [modelId] of entries) {
+    const baseWire = registryTransportMatches ? registry?.adapter ?? provider.adapter : provider.adapter;
+    const virtualSelectedModelId = Object.keys(registry?.virtualModels ?? {}).find(
+      candidate => candidate.toLowerCase() === modelId.trim().toLowerCase(),
+    );
+    const effectiveSelectedModelId = virtualSelectedModelId ?? modelId;
+    let effectiveWire = resolveEffectiveWire(effectiveSelectedModelId, baseWire);
+    const virtualWireModel = resolveOpenAiVirtualModel(
+      providerName,
+      effectiveSelectedModelId,
+    )?.wireModelId;
+    if (virtualWireModel && virtualWireModel !== effectiveSelectedModelId) {
+      effectiveWire = resolveEffectiveWire(virtualWireModel, effectiveWire);
+    }
+    if (effectiveWire !== "openai-responses") {
+      return `${field}.${modelId} requires the openai-responses wire`;
+    }
+  }
+  return null;
+}
+
 const CODEX_ACCOUNT_NAMESPACES_RECORD_ERROR =
   "codexAccountNamespaces must be a plain object mapping account selectors to Codex account ids";
 const CODEX_ACCOUNT_NAMESPACE_KEY_ERROR =
@@ -1249,6 +1327,19 @@ const configSchema = z.object({
         message: modelAdaptersError,
       });
     }
+    const responsesUpstreamStreamingError = modelResponsesUpstreamStreamingConfigError(
+      (provider as { modelResponsesUpstreamStreaming?: unknown }).modelResponsesUpstreamStreaming,
+      "modelResponsesUpstreamStreaming",
+      name,
+      provider,
+    );
+    if (responsesUpstreamStreamingError) {
+      ctx.addIssue({
+        code: "custom",
+        path: ["providers", name, "modelResponsesUpstreamStreaming"],
+        message: responsesUpstreamStreamingError,
+      });
+    }
     const preferHostedToolsError = modelPreferHostedToolsConfigError(
       (provider as { modelPreferHostedTools?: unknown }).modelPreferHostedTools,
       "modelPreferHostedTools",