fix(claude): preserve hosted web search choices

docs-site/src/content/docs/guides/claude-code.md

@@ -371,7 +371,7 @@ Claude Code's `/effort` setting is preserved across the adapter:
 | --- | --- |
 | `thinking.type: "adaptive"` + `output_config.effort` | Effort passed directly (`minimal`\|`low`\|`medium`\|`high`\|`xhigh`\|`max`\|`ultra`) |
 | `thinking.type: "enabled"` + `budget_tokens` | ≤4096→`low`, ≤16384→`medium`, above→`high` |
-| `thinking.type: "disabled"` | Reasoning parameters omitted entirely |
+| `thinking.type: "disabled"` | `reasoning: { effort: "none" }`; summary omitted |
 
 The resolved value appears in the request log's **Reasoning effort** column.
 
@@ -389,7 +389,7 @@ The proxy translates every Anthropic Messages API request into the Codex Respons
 | User `tool_result` | `function_call_output` (`is_error` → `[tool error]` prefix) |
 | `thinking` / `redacted_thinking` replay | Dropped |
 | Function tools | `{type: "function"}` (`web_search*` → `{type: "web_search"}`) |
-| `tool_choice` | `auto`→`auto`, `none`→`none`, `any`→`required`, named→`{type:"function",name}` |
+| `tool_choice` | `auto`→`auto`, `none`→`none`, `any`→`required`, named function→`{type:"function",name}`, hosted WebSearch/web_search→`{type:"web_search"}` |
 | `max_tokens` | `max_output_tokens` |
 | `stop_sequences` | `stop` |
 

docs-site/src/content/docs/ja/guides/claude-code.md

@@ -250,7 +250,7 @@ Claude Code の `/effort` 設定はアダプターでも維持されます。
  --- | --- |
 | `thinking.type: "adaptive"` + `output_config.effort` | Effort をそのまま渡します(`minimal`\|`low`\|`medium`\|`high`\|`xhigh`\|`max`\|`ultra`) |
 | `thinking.type: "enabled"` + `budget_tokens` | ≤4096→`low`、≤16384→`medium`、それより大→`high` |
-| `thinking.type: "disabled"` | 推論パラメータをすべて省略します |
+| `thinking.type: "disabled"` | `reasoning: { effort: "none" }` を明示し、`summary` は省略します |
 
 解釈された値はリクエストログの **Reasoning effort** 列に表示されます。
 
@@ -268,7 +268,7 @@ Claude Code の `/effort` 設定はアダプターでも維持されます。
 | ユーザー `tool_result` | `function_call_output`(`is_error` → `[tool error]` 接頭辞) |
 | `thinking` / `redacted_thinking` 再生 | 破棄 |
 | Function ツール | `{type: "function"}`(`web_search*` → `{type: "web_search"}`) |
-| `tool_choice` | `auto`→`auto`、`none`→`none`、`any`→`required`、名前指定→`{type:"function",name}` |
+| `tool_choice` | `auto`→`auto`、`none`→`none`、`any`→`required`、名前指定関数→`{type:"function",name}`、ホスト型 WebSearch/web_search→`{type:"web_search"}` |
 | `max_tokens` | `max_output_tokens` |
 | `stop_sequences` | `stop` |
 

docs-site/src/content/docs/ko/guides/claude-code.md

@@ -287,7 +287,7 @@ Claude Code의 `/effort` 설정은 어댑터에서도 유지돼요.
 | --- | --- |
 | `thinking.type: "adaptive"` + `output_config.effort` | Effort를 그대로 전달해요(`minimal`\|`low`\|`medium`\|`high`\|`xhigh`\|`max`\|`ultra`) |
 | `thinking.type: "enabled"` + `budget_tokens` | ≤4096→`low`, ≤16384→`medium`, 그보다 크면→`high` |
-| `thinking.type: "disabled"` | 추론 매개변수를 모두 생략해요 |
+| `thinking.type: "disabled"` | `reasoning: { effort: "none" }`을 명시하고 `summary`는 생략해요 |
 
 해석된 값은 요청 로그의 **Reasoning effort** 열에 표시돼요.
 
@@ -305,7 +305,7 @@ Claude Code의 `/effort` 설정은 어댑터에서도 유지돼요.
 | 사용자 `tool_result` | `function_call_output`(`is_error` → `[tool error]` 접두사) |
 | `thinking` / `redacted_thinking` 재생 | 버려요 |
 | Function 도구 | `{type: "function"}`(`web_search*` → `{type: "web_search"}`) |
-| `tool_choice` | `auto`→`auto`, `none`→`none`, `any`→`required`, 이름 지정→`{type:"function",name}` |
+| `tool_choice` | `auto`→`auto`, `none`→`none`, `any`→`required`, 이름 지정 함수→`{type:"function",name}`, 호스팅 WebSearch/web_search→`{type:"web_search"}` |
 | `max_tokens` | `max_output_tokens` |
 | `stop_sequences` | `stop` |
 

docs-site/src/content/docs/ru/guides/claude-code.md

@@ -266,7 +266,7 @@ Claude Code — это лишь учётные данные для доступ
 | --- | --- |
 | `thinking.type: "adaptive"` + `output_config.effort` | Уровень передаётся напрямую (`minimal`\|`low`\|`medium`\|`high`\|`xhigh`\|`max`\|`ultra`) |
 | `thinking.type: "enabled"` + `budget_tokens` | ≤4096→`low`, ≤16384→`medium`, выше→`high` |
-| `thinking.type: "disabled"` | Параметры рассуждений полностью опускаются |
+| `thinking.type: "disabled"` | Явно передаётся `reasoning: { effort: "none" }`, а `summary` опускается |
 
 Итоговое значение отображается в столбце **Reasoning effort** журнала запросов.
 
@@ -284,7 +284,7 @@ Claude Code — это лишь учётные данные для доступ
 | `tool_result` пользователя | `function_call_output` (`is_error` → префикс `[tool error]`) |
 | Повтор `thinking` / `redacted_thinking` | Отбрасывается |
 | Function-инструменты | `{type: "function"}` (`web_search*` → `{type: "web_search"}`) |
-| `tool_choice` | `auto`→`auto`, `none`→`none`, `any`→`required`, именованный→`{type:"function",name}` |
+| `tool_choice` | `auto`→`auto`, `none`→`none`, `any`→`required`, именованная функция→`{type:"function",name}`, размещённый WebSearch/web_search→`{type:"web_search"}` |
 | `max_tokens` | `max_output_tokens` |
 | `stop_sequences` | `stop` |
 

docs-site/src/content/docs/zh-cn/guides/claude-code.md

@@ -245,7 +245,7 @@ Claude Code 的 `/effort` 设置会完整保留并传递给适配器：
 | --- | --- |
 | `thinking.type: "adaptive"` + `output_config.effort` | 直接传递强度（`minimal`\|`low`\|`medium`\|`high`\|`xhigh`\|`max`\|`ultra`） |
 | `thinking.type: "enabled"` + `budget_tokens` | ≤4096→`low`，≤16384→`medium`，更高→`high` |
-| `thinking.type: "disabled"` | 完全省略推理参数 |
+| `thinking.type: "disabled"` | 显式发送 `reasoning: { effort: "none" }`，并省略 `summary` |
 
 解析后的值会显示在请求日志的 **Reasoning effort** 列中。
 
@@ -263,7 +263,7 @@ Claude Code 的 `/effort` 设置会完整保留并传递给适配器：
 | 用户 `tool_result` | `function_call_output`（`is_error` → `[tool error]` 前缀） |
 | 重放 `thinking` / `redacted_thinking` | 丢弃 |
 | Function 工具 | `{type: "function"}`（`web_search*` → `{type: "web_search"}`） |
-| `tool_choice` | `auto`→`auto`，`none`→`none`，`any`→`required`，指定名称→`{type:"function",name}` |
+| `tool_choice` | `auto`→`auto`，`none`→`none`，`any`→`required`，指定函数→`{type:"function",name}`，托管 WebSearch/web_search→`{type:"web_search"}` |
 | `max_tokens` | `max_output_tokens` |
 | `stop_sequences` | `stop` |
 

src/claude/inbound.ts

@@ -13,6 +13,7 @@ import type { OcxClaudeCodeConfig } from "../types";
 import { resolveAlias } from "./alias";
 import { stripOneMillionMarker } from "./context-windows";
 import { resolveDesktop3pAlias } from "./desktop-3p";
+import { isClaudeWebSearchToolName } from "./outbound";
 import { createHash } from "node:crypto";
 
 export class AnthropicRequestError extends Error {}
@@ -375,7 +376,12 @@ function toolChoiceToResponses(choice: unknown, body: Rec): void {
       if (typeof choice.name !== "string" || choice.name.length === 0) {
         throw new AnthropicRequestError("tool_choice.tool requires a name");
       }
-      body.tool_choice = { type: "function", name: choice.name };
+      // Anthropic represents hosted WebSearch as a named tool choice, while
+      // Responses requires the choice type to match the hosted declaration.
+      // Preserve forced-tool intent rather than weakening it to `auto`.
+      body.tool_choice = isClaudeWebSearchToolName(choice.name)
+        ? { type: "web_search" }
+        : { type: "function", name: choice.name };
       break;
     default: break;
   }
@@ -502,8 +508,9 @@ export function anthropicToResponsesTranslation(raw: unknown, cc?: OcxClaudeCode
     // An explicit "disabled" is an instruction, not an absence. Dropping it made this
     // indistinguishable from a request that never mentioned thinking — and for models that
     // think by default, omission means thinking is ON, sharing the caller's max_tokens (#545).
-    // "none" is the parser's disable sentinel (parser.ts REASONING_EFFORTS).
-    body.reasoning = { effort: "none", summary: "none" };
+    // `none` is the effort disable sentinel. It is not a valid OpenAI summary
+    // value, so do not attach the similarly named internal catalog sentinel.
+    body.reasoning = { effort: "none" };
   } else if (isRec(thinking) || outputConfigEffort !== undefined) {
     const reasoning: Rec = { summary: "auto" };
     if (outputConfigEffort !== undefined) {

src/responses/schema.ts

@@ -107,7 +107,7 @@ export const toolSchema = z.object({
 const builtinToolSchema = z.object({ type: z.string() }).loose();
 
 const hostedToolType = z.enum([
-  "web_search_preview", "file_search", "computer_use_preview",
+  "web_search", "web_search_preview", "file_search", "computer_use_preview",
   "code_interpreter", "image_generation", "mcp",
 ]);
 

tests/claude-inbound.test.ts

@@ -96,7 +96,7 @@ describe("claude inbound translation", () => {
     expect((anthropicToResponsesBody({ ...base, thinking: { type: "adaptive" } }) as any).reasoning).toEqual({ summary: "auto" });
     // "disabled" and omitted must NOT collapse to the same state: for a model that thinks by
     // default, omission means thinking is ON and shares the caller's max_tokens (#545).
-    expect((anthropicToResponsesBody({ ...base, thinking: { type: "disabled" } }) as any).reasoning).toEqual({ effort: "none", summary: "none" }); // justified: sibling assertions in this test use the same cast
+    expect((anthropicToResponsesBody({ ...base, thinking: { type: "disabled" } }) as any).reasoning).toEqual({ effort: "none" }); // justified: sibling assertions in this test use the same cast
     expect((anthropicToResponsesBody(base) as any).reasoning).toBeUndefined();
     expect(effortForThinkingBudget(1024)).toBe("low");
     expect(effortForThinkingBudget(8192)).toBe("medium");
@@ -134,7 +134,7 @@ describe("claude inbound translation", () => {
     // rather than left to think anyway (#545).
     expect(reasoningOf(anthropicToResponsesBody({
       ...base, thinking: { type: "disabled" }, output_config: { effort: "high" },
-    }))).toEqual({ effort: "none", summary: "none" });
+    }))).toEqual({ effort: "none" });
     // unknown effort strings are dropped so downstream defaults win
     expect(reasoningOf(anthropicToResponsesBody({
       ...base, thinking: { type: "adaptive" }, output_config: { effort: "turbo" },
@@ -149,6 +149,23 @@ describe("claude inbound translation", () => {
       .toEqual({ type: "function", name: "Read" });
   });
 
+  test("forced Claude WebSearch stays a hosted Responses tool choice", () => {
+    const body = anthropicToResponsesBody({
+      model: "gpt-5.6-luna",
+      max_tokens: 10,
+      messages: [{ role: "user", content: "search" }],
+      tools: [{ type: "web_search_20250305", name: "web_search" }],
+      tool_choice: { type: "tool", name: "web_search" },
+      thinking: { type: "disabled" },
+    }) as Record<string, unknown>;
+
+    expect(body.tools).toEqual([{ type: "web_search" }]);
+    expect(body.tool_choice).toEqual({ type: "web_search" });
+    expect(body.reasoning).toEqual({ effort: "none" });
+    expect(() => responsesRequestSchema.parse(body)).not.toThrow();
+    expect(() => parseRequest(body)).not.toThrow();
+  });
+
   test("system role messages fold into instructions (real Claude Code sends them; native backend rejects system items)", () => {
     const body = anthropicToResponsesBody({
       model: "m", max_tokens: 10,