fix: add per-model structured output opt-out

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

@@ -93,6 +93,7 @@ account を削除しても mapping は保持され、同じ id を再追加す
 | `noTemperatureModels?` | `string[]` |発信者指定の`temperature`を拒否するモデル。 |
 | `noTopPModels?` | `string[]` |発信者指定の`top_p`を拒否するモデル。 |
 | `noPenaltyModels?` | `string[]` |存在/周波数ペナルティを拒否するモデル。 |
+| `noStructuredOutputModels?` | `string[]` | `openai-chat` エンドポイントが `response_format` を拒否する正確なモデル ID。要求モデルが項目と完全一致する場合だけフィールドを省略し、その他の `openai-chat` モデルでは structured-output 変換を維持します。 |
 | `parallelToolCalls?` | `boolean` |並列ツール呼び出しを切り替えます。 OpenAI Chat はデフォルトでオンになっています。非チャット アダプターは明示的な `true` でのみアドバタイズします。 |
 | `responsesItemIdRepair?` | `{ message?: string[]; reasoning?: string[]; repairMissingTerminalIds?: boolean; repairInvalidIds?: boolean }` |正確なプレースホルダー ID、欠落している端末 ID、および（`repairInvalidIds` で）正規の `msg_`/`rs_` 接頭辞を欠く message/reasoning ID に対するダウンストリーム SSE 修復はデフォルトで無効になっています。関数呼び出し ID は決して書き換えられません。組み込み DeepSeek は最後の 2 つをデフォルトで有効にします。 |
 | `responsesSnapshotRepair?` | `boolean` | デフォルトで無効のクライアント向け修復です。SSE と JSON の Responses ライフサイクルで欠落した status、output、ツールメタデータを補完し、raw 検査と永続化は変更しません。 |

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

@@ -93,6 +93,7 @@ managed map을 활성화하면 privacy-safe selector를 만들고, 이후 계정
 | `noTemperatureModels?` | `string[]` | 호출자가 지정한 `temperature`를 거부하는 모델입니다. |
 | `noTopPModels?` | `string[]` | 호출자가 지정한 `top_p`를 거부하는 모델입니다. |
 | `noPenaltyModels?` | `string[]` | presence/frequency penalty를 허용하지 않는 모델입니다. |
+| `noStructuredOutputModels?` | `string[]` | `openai-chat` 엔드포인트가 `response_format`을 거부하는 정확한 모델 ID입니다. 요청 모델이 항목과 정확히 일치할 때만 필드를 생략하며, 그 외 `openai-chat` 모델에서는 structured-output 변환을 유지합니다. |
 | `parallelToolCalls?` | `boolean` | 병렬 도구 호출을 켜거나 끕니다. OpenAI Chat은 기본으로 켜져 있고, 비-chat 어댑터는 명시적으로 `true`일 때만 이를 노출합니다. |
 | `responsesItemIdRepair?` | `{ message?: string[]; reasoning?: string[]; repairMissingTerminalIds?: boolean; repairInvalidIds?: boolean }` | 기본값이 꺼진 downstream SSE 복구입니다. 정확한 자리표시자 id, 누락된 종료 id, 그리고(`repairInvalidIds`) 정규 `msg_`/`rs_` 접두사가 없는 message/reasoning id를 복구합니다. function-call id는 다시 쓰지 않습니다. 내장 DeepSeek은 마지막 두 가지를 기본으로 켭니다. |
 | `responsesSnapshotRepair?` | `boolean` | 기본값이 꺼진 클라이언트용 복구입니다. SSE와 JSON의 Responses 수명 주기에서 누락된 status, output, 도구 메타데이터를 채우며 raw 검사와 영속화는 변경하지 않습니다. |

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

@@ -103,6 +103,7 @@ differing backup and rewrites known legacy namespaced selected ids to bare ids.
 | `noTemperatureModels?` | `string[]` | Models that reject caller-specified `temperature`. |
 | `noTopPModels?` | `string[]` | Models that reject caller-specified `top_p`. |
 | `noPenaltyModels?` | `string[]` | Models that reject presence/frequency penalties. |
+| `noStructuredOutputModels?` | `string[]` | Exact model IDs whose `openai-chat` endpoint rejects `response_format`. Only an exact requested-model match omits the field; structured-output translation stays enabled for every other `openai-chat` model. |
 | `parallelToolCalls?` | `boolean` | Toggle parallel tool calls. OpenAI Chat defaults on; non-chat adapters advertise only on explicit `true`. |
 | `responsesItemIdRepair?` | `{ message?: string[]; reasoning?: string[]; repairMissingTerminalIds?: boolean; repairInvalidIds?: boolean }` | Disabled-by-default downstream SSE repair for exact placeholder ids, missing terminal ids, and (with `repairInvalidIds`) message/reasoning ids missing the canonical `msg_`/`rs_` prefix. Function-call ids are never rewritten. Built-in DeepSeek enables the last two by default. |
 | `responsesSnapshotRepair?` | `boolean` | Disabled-by-default client-facing repair for sparse Responses lifecycle snapshots in SSE and JSON. Fills missing canonical status, output, and tool metadata while raw inspection and persistence remain unchanged. |

docs-site/src/content/docs/reference/proxy-formats.md

@@ -149,8 +149,9 @@ Structured output is part of that translation: `response_format` with `json_obje
 `json_schema` is forwarded to routed `openai-chat` models. On `POST /v1/responses` the
 equivalent request field is `text.format`: native Responses routes preserve it in the raw
 Responses body, and it is translated to `response_format` when the model routes to an
-`openai-chat` provider. A backend without structured-output support returns its own error
-instead of the proxy rejecting the request locally.
+`openai-chat` provider. A model listed in the provider's `noStructuredOutputModels` omits
+`response_format` on that chat wire; sibling models keep the translation. Unclassified backends
+receive the field and return their own error instead of the proxy guessing their capability.
 
 Non-streaming output has `object: "chat.completion"`. Streaming output uses SSE objects with
 `object: "chat.completion.chunk"`, choice deltas, a terminal choice with `finish_reason`, and
@@ -287,4 +288,4 @@ structurally valid Fernet runs unchanged. If an `agent_message` loses all encryp
 that repair, it becomes a normal user message. If a current v2 task remains genuinely encrypted
 but the selected routed target cannot read native ChatGPT ciphertext, opencodex fails with
 `unreadable_encrypted_agent_task` instead of sending unreadable bytes to that provider. See
-[Sub-agent Surface](/guides/sub-agent-surface/) for the client behavior around worker tasks.
+[Sub-agent Surface](/guides/sub-agent-surface/) for the client behavior around worker tasks.

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

@@ -106,6 +106,7 @@ cross-route credential fallback не существует. Строки API GPT-
 | `noTemperatureModels?` | `string[]` | Модели, отвергающие переданный вызывающей стороной `temperature`. |
 | `noTopPModels?` | `string[]` | Модели, отвергающие переданный вызывающей стороной `top_p`. |
 | `noPenaltyModels?` | `string[]` | Модели, отвергающие penalty presence/frequency. |
+| `noStructuredOutputModels?` | `string[]` | Точные идентификаторы моделей, чей endpoint `openai-chat` отклоняет `response_format`. Поле опускается только при точном совпадении запрошенной модели; для остальных моделей `openai-chat` преобразование structured output остаётся включённым. |
 | `parallelToolCalls?` | `boolean` | Переключатель parallel tool call'ов. Для OpenAI Chat по умолчанию включено; не-chat adapter'ы рекламируют это только при явном `true`. |
 | `responsesItemIdRepair?` | `{ message?: string[]; reasoning?: string[]; repairMissingTerminalIds?: boolean; repairInvalidIds?: boolean }` | По умолчанию выключенная downstream SSE-repair для exact placeholder-id, отсутствующих terminal-id и (с `repairInvalidIds`) message/reasoning id без канонического префикса `msg_`/`rs_`. Function-call id никогда не переписываются. Встроенный DeepSeek включает последние два по умолчанию. |
 | `responsesSnapshotRepair?` | `boolean` | По умолчанию выключенная клиентская repair для неполных lifecycle snapshot'ов Responses в SSE и JSON. Добавляет отсутствующие status, output и tool metadata, не меняя raw inspection и persistence. |

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

@@ -93,6 +93,7 @@ selector，而不是分配一个新名称。
 | `noTemperatureModels?` | `string[]` | 会拒绝调用方指定 `temperature` 的模型。 |
 | `noTopPModels?` | `string[]` | 会拒绝调用方指定 `top_p` 的模型。 |
 | `noPenaltyModels?` | `string[]` | 会拒绝 presence/frequency penalty 的模型。 |
+| `noStructuredOutputModels?` | `string[]` | `openai-chat` 端点拒绝 `response_format` 的精确模型 ID。仅当请求模型与条目完全匹配时才省略该字段；其他 `openai-chat` 模型仍启用 structured-output 转换。 |
 | `parallelToolCalls?` | `boolean` | 切换并行工具调用。OpenAI Chat 默认开启；非 chat 适配器只有显式 `true` 时才会声明支持。 |
 | `responsesItemIdRepair?` | `{ message?: string[]; reasoning?: string[]; repairMissingTerminalIds?: boolean; repairInvalidIds?: boolean }` | 默认关闭的下游 SSE 修复，用于精确占位 id、缺失的终止 id，以及（`repairInvalidIds`）缺少规范 `msg_`/`rs_` 前缀的 message/reasoning id。function-call id 永远不会被重写。内置 DeepSeek 默认启用后两项。 |
 | `responsesSnapshotRepair?` | `boolean` | 默认关闭的客户端修复，用于补全 SSE 与 JSON 中稀疏 Responses 生命周期快照缺失的 status、output 和工具元数据；原始检查与持久化保持不变。 |

src/adapters/openai-chat.ts

@@ -813,19 +813,25 @@ export function createOpenAIChatAdapter(provider: OcxProviderConfig): ProviderAd
       if (provider.promptCacheKey && parsed.options.promptCacheKey !== undefined) {
         body.prompt_cache_key = parsed.options.promptCacheKey;
       }
-      const textFormat = parsed.options.textFormat;
-      if (textFormat?.type === "json_object") {
-        body.response_format = { type: "json_object" };
-      } else if (textFormat?.type === "json_schema") {
-        body.response_format = {
-          type: "json_schema",
-          json_schema: {
-            name: textFormat.name ?? "response",
-            ...(textFormat.description !== undefined ? { description: textFormat.description } : {}),
-            ...(textFormat.schema !== undefined ? { schema: textFormat.schema } : {}),
-            ...(textFormat.strict !== undefined ? { strict: textFormat.strict } : {}),
-          },
-        };
+      // Structured-output support varies by the physical upstream model even when one
+      // gateway exposes a uniform OpenAI-compatible endpoint. Keep the #1137 translation
+      // as the default, but let an exact model opt out instead of forcing a provider-wide
+      // rollback that would silently return prose for siblings that support JSON Schema.
+      if (!provider.noStructuredOutputModels?.includes(parsed.modelId)) {
+        const textFormat = parsed.options.textFormat;
+        if (textFormat?.type === "json_object") {
+          body.response_format = { type: "json_object" };
+        } else if (textFormat?.type === "json_schema") {
+          body.response_format = {
+            type: "json_schema",
+            json_schema: {
+              name: textFormat.name ?? "response",
+              ...(textFormat.description !== undefined ? { description: textFormat.description } : {}),
+              ...(textFormat.schema !== undefined ? { schema: textFormat.schema } : {}),
+              ...(textFormat.strict !== undefined ? { strict: textFormat.strict } : {}),
+            },
+          };
+        }
       }
 
       if (tools) {

src/config.ts

@@ -629,6 +629,9 @@ const providerConfigSchema = z.object({
   supportsServiceTier: z.boolean().optional(),
   preserveResponsesReasoningContent: z.boolean().optional(),
   allowPrivateNetwork: z.boolean().optional(),
+  noStructuredOutputModels: z.array(z.string().min(1))
+    .transform(normalizeNonBlankStringArray)
+    .optional(),
   retryOn429: retryOn429PolicySchema.optional(),
   codexAccountMode: z.enum(["pool", "direct"]).optional(),
   responsesItemIdRepair: z.object({
@@ -820,6 +823,26 @@ export function positiveIntegerConfigError(value: unknown, field: string): strin
   return null;
 }
 
+export function nonBlankStringArrayConfigError(value: unknown, field: string): string | null {
+  if (value === undefined) return null;
+  if (!Array.isArray(value)) return `${field} must be an array`;
+  for (const [index, entry] of value.entries()) {
+    if (typeof entry !== "string" || !entry.trim()) {
+      return `${field}.${index} must be a nonblank model id`;
+    }
+  }
+  return null;
+}
+
+/**
+ * Keep hand-edited config and management writes on one canonical model-id list.
+ * Validation happens separately so an all-whitespace value is rejected rather than
+ * normalized into a model id that can never match at runtime.
+ */
+export function normalizeNonBlankStringArray(value: readonly string[]): string[] {
+  return [...new Set(value.map(entry => entry.trim()))];
+}
+
 export function booleanRecordConfigError(value: unknown, field: string): string | null {
   if (value === undefined) return null;
   if (!value || typeof value !== "object" || Array.isArray(value)) return `${field} must be a plain object`;
@@ -1408,6 +1431,17 @@ const configSchema = z.object({
         message: maxOutputError,
       });
     }
+    const structuredOutputOptOutError = nonBlankStringArrayConfigError(
+      (provider as { noStructuredOutputModels?: unknown }).noStructuredOutputModels,
+      "noStructuredOutputModels",
+    );
+    if (structuredOutputOptOutError) {
+      ctx.addIssue({
+        code: "custom",
+        path: ["providers", redactSecretString(name), "noStructuredOutputModels"],
+        message: structuredOutputOptOutError,
+      });
+    }
     if (Object.hasOwn(provider, "codexAccountMode") && provider.codexAccountMode !== undefined) {
       // Persisted account mode is valid ONLY on the canonical built-in `openai` forward provider.
       // Old openai-multi rows stay parseable (they never carry a mode) so startup can migrate them.

src/server/auth-cors.ts

@@ -6,6 +6,7 @@ import {
   modelAdapterRecordConfigError,
   modelPreferHostedToolsConfigError,
   codexAutoStartEnabled,
+  nonBlankStringArrayConfigError,
   positiveIntegerConfigError,
   positiveIntegerRecordConfigError,
   providerBaseUrlConfigError,
@@ -509,6 +510,11 @@ export function providerManagementConfigError(name: unknown, provider: unknown):
   if (defaultMaxOutputError) return `provider ${name} ${defaultMaxOutputError}`;
   const maxOutputError = positiveIntegerRecordConfigError(raw.modelMaxOutputTokens, "modelMaxOutputTokens");
   if (maxOutputError) return `provider ${name} ${maxOutputError}`;
+  const structuredOutputOptOutError = nonBlankStringArrayConfigError(
+    raw.noStructuredOutputModels,
+    "noStructuredOutputModels",
+  );
+  if (structuredOutputOptOutError) return `provider ${name} ${structuredOutputOptOutError}`;
   const openRouterError = openRouterRoutingConfigError(typed);
   if (openRouterError) return `provider ${name} ${openRouterError}`;
   if (typed.authMode === "local") {
@@ -589,6 +595,7 @@ export function safeConfigDTO(config: OcxConfig): unknown {
       "noTemperatureModels",
       "noTopPModels",
       "noPenaltyModels",
+      "noStructuredOutputModels",
       "autoToolChoiceOnlyModels",
       "preserveReasoningContentModels",
       "requiresReasoningPlaceholderModels",

src/server/management/provider-routes.ts

@@ -8,6 +8,8 @@ import {
   hasOwnProvider,
   isValidProviderName,
   multiAgentGuidanceEnabled,
+  nonBlankStringArrayConfigError,
+  normalizeNonBlankStringArray,
   providerBaseUrlConfigError,
   providerHeadersConfigError,
   saveConfigPreservingClaudeCode,
@@ -205,6 +207,19 @@ function applyProviderPatchFields(
     }
     touched = true;
   }
+  if (Object.hasOwn(rawBody, "noStructuredOutputModels")) {
+    const value = rawBody.noStructuredOutputModels;
+    if (value === null) {
+      delete next.noStructuredOutputModels;
+    } else {
+      const error = nonBlankStringArrayConfigError(value, "noStructuredOutputModels");
+      if (error) return { error };
+      const models = normalizeNonBlankStringArray(value as string[]);
+      if (models.length > 0) next.noStructuredOutputModels = models;
+      else delete next.noStructuredOutputModels;
+    }
+    touched = true;
+  }
 
   // headers is the one object-valued field in the mask. PATCH semantics merge it
   // shallowly into the existing block so a single fingerprint header can be added
@@ -291,6 +306,7 @@ export async function handleProviderRoutes(ctx: ManagementContext): Promise<Resp
       models: p.models ?? [],
       contextWindow: p.contextWindow,
       modelContextWindows: p.modelContextWindows,
+      noStructuredOutputModels: p.noStructuredOutputModels,
       authMode: p.authMode,
       apiKeyTransport: p.apiKeyTransport,
       disabled: p.disabled === true,

src/types.ts

@@ -1393,6 +1393,12 @@ export interface OcxProviderConfig {
   noTopPModels?: string[];
   /** Model ids that reject caller-specified presence/frequency penalty values. */
   noPenaltyModels?: string[];
+  /**
+   * Model ids whose Chat Completions endpoint rejects `response_format`.
+   * Structured-output translation remains enabled by default; this is a narrow
+   * per-model compatibility escape hatch for mixed-capability gateways.
+   */
+  noStructuredOutputModels?: string[];
   /**
    * Allow multiple tool calls per completion. DEFAULT-ON for openai-chat providers (the
    * buffered stream parser assembles interleaved/fragmented multi-call turns safely);

structure/04_transports-and-sidecars.md

@@ -528,6 +528,24 @@ adapters advertise the catalog bit only on explicit `true`; cursor keeps its own
 Providers with flaky parallel streaming can be opted out individually. Evidence and provider
 ledger: `devlog/_fin/260709_parallel_tool_calls/`.
 
+## Chat structured-output compatibility
+
+The `openai-chat` adapter translates Responses `text.format` and Chat Completions
+`response_format` through one internal format, then emits `response_format` on the upstream chat
+wire. That remains the default because silently returning prose breaks clients that requested a
+JSON object or schema. A mixed-capability gateway may list exact native model ids in
+`noStructuredOutputModels`; only those models omit the wire field, while siblings keep the normal
+translation. The proxy does not infer this from provider names, localhost destinations, or a model
+family shared by unrelated upstreams.
+
+[Decision Log]
+- 목적과 의도: Recover chat models that reject `response_format` without removing structured output from models that support it.
+- 기존 구현 및 제약 조건: The adapter forwarded the field to every routed chat model after #1137, while the same model id may sit behind gateways with different capabilities.
+- 검토한 주요 대안: Revert translation globally; blacklist a model id globally; detect a proxy by name or URL; add an explicit provider/model opt-out.
+- 선택한 방식: Preserve default translation and omit it only for exact ids in `noStructuredOutputModels`.
+- 다른 대안 대신 이 방식을 선택한 이유: Global or heuristic rules regress supported providers and make custom gateway names part of the wire contract.
+- 장점, 단점 및 영향: Compatible siblings retain schema enforcement and explicitly incompatible models avoid the upstream 400; operators must classify each unsupported model they route.
+
 ## Reasoning display parity (hideThinkingSummary)
 
 `hideThinkingSummary` (request reasoning summary absent/"none" — the routed catalog default) is