fix(cursor): send Grok 4.5 Fast parameters

docs-site/src/content/docs/ja/reference/adapters.md

@@ -104,6 +104,8 @@ filtered incomplete になります。実際のツール呼び出しを伴わな
 
 - 通常の fetch/parse 経路の代わりに `runTurn` を使います。リクエスト、サーバーイベント、ツール引数、使用量 checkpoint、クライアントレスポンスは `cursor/gen/agent_pb.ts` の `@bufbuild/protobuf` スキーマでエンコードしたのち Connect メッセージとして framing します。
 - content-addressed blob で対話状態を再生し、サーバーツール呼び出しを Codex に再マッピングします。protobuf の `GetUsableModels` RPC でリアルタイム Cursor モデルを探し、run リクエストが wire に commit される前だけリトライします。
+- `cursor/grok-4.5-fast` は選択可能なモデルとして維持しつつ、Cursor には正規の `grok-4.5`
+  モデルを送信し、個別の `effort` および `fast=true` 値は `requested_model.parameters` に格納します。
 - Cursor ネイティブのローカルファイルシステム/shell/network 実行はデフォルトで拒否します。明示的な `mcpServers` と `desktopExecutor` 統合はそれぞれ別の opt-in です。`unsafeAllowNativeLocalExec` はより広い組み込み executor を有効にし、Codex の承認/サンドボックスルールを迂回します。
 
 ## `azure-openai`（別名: `azure`）

docs-site/src/content/docs/ko/reference/adapters.md

@@ -123,6 +123,8 @@ commentary로 유지하고 비공개 완료 툴을 한 번 검증합니다.
 - content-addressed blob으로 대화 상태를 재생하고 서버 툴 호출을 Codex에 다시 매핑합니다. protobuf
   `GetUsableModels` RPC로 실시간 Cursor 모델을 찾으며, run 요청이 wire에 commit되기 전까지만
   재시도합니다.
+- `cursor/grok-4.5-fast`는 선택 가능한 모델로 유지하되, Cursor에는 정식 `grok-4.5` 모델을 보내고
+  별도의 `effort`, `fast=true` 값은 `requested_model.parameters`에 담습니다.
 - Cursor 네이티브 로컬 파일시스템/shell/network 실행은 기본적으로 거부합니다. 명시적인
   `mcpServers`와 `desktopExecutor` 통합은 각각 별도 opt-in입니다. `unsafeAllowNativeLocalExec`은
   더 넓은 내장 executor를 켜며 Codex 승인/샌드박스 규칙을 우회합니다.

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

@@ -168,6 +168,8 @@ advertised effort control on those models as proof of upstream-native reasoning
 - Exposes Cursor Router as `cursor/auto` plus explicit `cursor/auto-cost`,
   `cursor/auto-balance`, and `cursor/auto-intelligence` entries. Explicit levels are encoded in
   `requested_model.parameters` while the legacy `cursor/auto` entry retains the account/team default.
+- Keeps `cursor/grok-4.5-fast` as a selectable model while sending Cursor's canonical `grok-4.5`
+  model with separate `effort` and `fast=true` parameters.
 - Cursor-native local filesystem/shell/network execution is denied by default. Explicit `mcpServers`
   and `desktopExecutor` integrations have separate opt-ins; `nativeLocalExec: "on"` enables the
   broader built-in executor and bypasses Codex approval/sandbox semantics, and legacy

docs-site/src/content/docs/ru/reference/adapters.md

@@ -134,6 +134,8 @@ authorization.
 - Воспроизводит состояние диалога через content-addressed blob'ы, отображает серверные вызовы
   инструментов обратно в Codex, обнаруживает актуальные модели Cursor через protobuf RPC
   `GetUsableModels` и повторяет попытки только до того, как run-запрос зафиксирован на wire.
+- Сохраняет `cursor/grok-4.5-fast` доступной для выбора, но отправляет Cursor каноническую модель
+  `grok-4.5`, помещая отдельные значения `effort` и `fast=true` в `requested_model.parameters`.
 - Нативное для Cursor локальное выполнение операций с файловой системой/shell/сетью по умолчанию
   запрещено. Явные интеграции `mcpServers` и `desktopExecutor` включаются отдельно;
   `unsafeAllowNativeLocalExec` включает более широкий встроенный executor и обходит семантику

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

@@ -117,6 +117,8 @@ Kiro 的 assistant 文本本身没有可靠的回合结束标记，但终止的
   Connect message。
 - 经 content-addressed blob 重放对话状态，把 server tool call 映射回 Codex，用 protobuf
   `GetUsableModels` RPC 发现实时 Cursor 模型，并且只在 run request 尚未 commit 到 wire 前重试。
+- 保留 `cursor/grok-4.5-fast` 作为可选模型，但向 Cursor 发送规范的 `grok-4.5` 模型，并将独立的
+  `effort` 和 `fast=true` 值放入 `requested_model.parameters`。
 - Cursor 原生本地 filesystem/shell/network 执行默认被拒绝。显式 `mcpServers` 与
   `desktopExecutor` 集成分别需要 opt-in；`unsafeAllowNativeLocalExec` 会启用更广泛的内置
   executor，并绕过 Codex 审批和 sandbox 语义。

src/adapters/cursor/effort-map.ts

@@ -116,8 +116,9 @@ export function cursorModelHasEffortTiers(baseModelId: string): boolean {
 }
 
 /**
- * Compose a Cursor wire id from a Codex-facing base id and effort tier.
- * Fast variants put the mode after the effort; other models use the ordinary `{base}-{effort}` form.
+ * Compose Cursor's flattened model id from a Codex-facing base id and effort tier. Discovery uses
+ * this for the ids returned by GetUsableModels. Parameterized Grok Fast requests bypass the flat id
+ * and send the base model plus requested_model parameters instead.
  */
 export function cursorWireModelIdWithEffort(baseModelId: string, effortSuffix: string): string {
   if (baseModelId.endsWith("-fast")) {

src/adapters/cursor/protobuf-request.ts

@@ -621,6 +621,11 @@ function buildPreparedCursorRunRequest(
     tools: request.tools?.length ?? 0,
   });
 
+  const requestedModelParameters = [
+    ...(request.requestedModelParameters ?? []),
+    ...(request.routingLevel ? [{ id: CURSOR_ROUTING_LEVEL_PARAMETER_ID, value: request.routingLevel }] : []),
+  ];
+  const hasExplicitModelParameters = (request.requestedModelParameters?.length ?? 0) > 0;
   const runRequest = create(AgentRunRequestSchema, {
     conversationId: request.conversationId,
     conversationState: create(ConversationStateStructureSchema, {
@@ -637,24 +642,24 @@ function buildPreparedCursorRunRequest(
       readPaths: [],
     }),
     action,
-    modelDetails: create(ModelDetailsSchema, {
-      modelId: request.modelId,
-      displayModelId: request.modelId,
-      displayName: request.modelId,
-      displayNameShort: request.modelId,
-      aliases: [],
-    }),
-    // requested_model is currently a Cursor Router-only surface. External model clients still
-    // send model_details alone; sending both makes external workers reach stepCompleted and then
-    // reject the turn with invalid_argument.
-    ...(request.routingLevel ? {
+    // Explicit model-picker parameters follow current Cursor clients and use requested_model alone.
+    // Keep legacy model_details for flat model ids and the already-live Router path; sending both for
+    // a parameterized external model can resolve conflicting selections and end in invalid_argument.
+    ...(!hasExplicitModelParameters ? {
+      modelDetails: create(ModelDetailsSchema, {
+        modelId: request.modelId,
+        displayModelId: request.modelId,
+        displayName: request.modelId,
+        displayNameShort: request.modelId,
+        aliases: [],
+      }),
+    } : {}),
+    ...(requestedModelParameters.length > 0 ? {
       requestedModel: create(RequestedModelSchema, {
         modelId: request.modelId,
         maxMode: false,
-        parameters: [create(RequestedModel_ModelParameterbytesSchema, {
-          id: CURSOR_ROUTING_LEVEL_PARAMETER_ID,
-          value: request.routingLevel,
-        })],
+        parameters: requestedModelParameters.map(parameter =>
+          create(RequestedModel_ModelParameterbytesSchema, parameter)),
       }),
     } : {}),
     // Mirror the client (Responses) tool definitions into the top-level AgentRunRequest.mcp_tools

src/adapters/cursor/request-builder.ts

@@ -8,7 +8,7 @@ import type {
   OcxToolResultMessage,
 } from "../../types";
 import { isAllowedToolChoice, namespacedToolName, toolChoiceAliases, type OcxTool, type OcxToolChoice } from "../../types";
-import type { CursorRequestMessage, CursorRunRequest } from "./types";
+import type { CursorRequestMessage, CursorRequestedModelParameter, CursorRunRequest } from "./types";
 import { cursorWireModelSelection, type CursorRoutingLevel } from "./discovery";
 import { cursorEffortSuffix, cursorWireModelIdWithEffort } from "./effort-map";
 import {
@@ -117,16 +117,29 @@ function catalogLimitNote(kept: readonly OcxTool[], omitted: readonly OcxTool[])
 }
 
 /**
- * Resolve a `cursor/<model>` selection + Codex reasoning effort to the actual Cursor model id. Cursor
-* encodes the effort as a per-model suffix (`claude-4.6-opus-high`); `cursorEffortSuffix` picks the
- * right tier for that specific model (literal pass-through, with rank clamp fallback) or
-* `undefined` for non-reasoning models like `composer-2.5`. A fully-qualified id (one that isn't a
-* known effort base) passes through unchanged.
+ * Resolve a `cursor/<model>` selection + Codex reasoning effort to Cursor's requested model shape.
+ * Most models encode effort in a flat id (`claude-4.6-opus-high`). Grok 4.5 Fast is parameterized
+ * instead: current Cursor clients send the `grok-4.5` base id plus `effort` and `fast` parameters.
+ * A fully-qualified id (one that is not a known effort base) passes through unchanged.
  */
-function normalizeCursorModelId(modelId: string, reasoning?: string): { modelId: string; routingLevel?: CursorRoutingLevel } {
+function normalizeCursorModelId(modelId: string, reasoning?: string): {
+  modelId: string;
+  requestedModelParameters?: readonly CursorRequestedModelParameter[];
+  routingLevel?: CursorRoutingLevel;
+} {
   const selection = cursorWireModelSelection(modelId);
   const id = selection.modelId;
   const suffix = cursorEffortSuffix(id, reasoning);
+  if (id === "grok-4.5-fast" && suffix) {
+    return {
+      ...selection,
+      modelId: "grok-4.5",
+      requestedModelParameters: [
+        { id: "effort", value: suffix },
+        { id: "fast", value: "true" },
+      ],
+    };
+  }
   return { ...selection, modelId: suffix ? cursorWireModelIdWithEffort(id, suffix) : id };
 }
 
@@ -241,6 +254,7 @@ export function createCursorRequest(
   const model = normalizeCursorModelId(parsed.modelId, parsed.options.reasoning);
   return {
     modelId: model.modelId,
+    ...(model.requestedModelParameters ? { requestedModelParameters: model.requestedModelParameters } : {}),
     ...(model.routingLevel ? { routingLevel: model.routingLevel } : {}),
     conversationId: resolveCursorConversationId(parsed, model.modelId, options),
     system: [...(parsed.context.systemPrompt ?? []), ...(limitNote ? [limitNote] : [])],

src/adapters/cursor/types.ts

@@ -2,8 +2,15 @@ import type { OcxUsage } from "../../types";
 import type { OcxMessage, OcxRequestOptions, OcxTool } from "../../types";
 import type { CursorRoutingLevel } from "./discovery";
 
+export interface CursorRequestedModelParameter {
+  id: string;
+  value: string;
+}
+
 export interface CursorRunRequest {
   modelId: string;
+  /** Cursor model-picker parameters encoded through AgentRunRequest.requested_model. */
+  requestedModelParameters?: readonly CursorRequestedModelParameter[];
   /** Cursor Router optimization parameter; valid only while modelId is the `default` wire model. */
   routingLevel?: CursorRoutingLevel;
   conversationId: string;

structure/04_transports-and-sidecars.md

@@ -267,7 +267,7 @@ replays are explicit and receive the same repair.
 These compatibility guards are covered by focused tests and should stay close to the adapters that
 need them.
 
-## Cursor Router optimization levels
+## Cursor parameterized models
 
 Cursor Router's parameterized `default` model is represented in Codex by four catalog rows:
 `cursor/auto` preserves Cursor's team/account default, while `cursor/auto-cost`,
@@ -277,6 +277,12 @@ All four route to the `default` Cursor wire model. Explicit variants additionall
 parameterized-model channel used by current Cursor clients. Router rows are static capabilities and
 must survive a live `GetUsableModels` response that omits `default`.
 
+`cursor/grok-4.5-fast` is also a stable Codex-facing row, but current Cursor clients do not request
+it as a flat model slug. OpenCodex sends `grok-4.5` through `requested_model` with separate `effort`
+and `fast=true` parameters, leaving legacy `model_details` unset for that parameterized external
+selection. Live discovery still recognizes Cursor's flattened `cursor-grok-4.5-{effort}-fast`
+variants, plus the older `grok-4.5-fast-{effort}` ordering, as availability evidence only.
+
 ## Cursor active-context usage
 
 Cursor's `conversationCheckpointUpdate.tokenDetails.usedTokens` is treated as the authoritative

tests/cursor-blob.test.ts

@@ -356,6 +356,31 @@ describe("Cursor blob handshake", () => {
     expect(run?.requestedModel).toBeUndefined();
   });
 
+  test("encodes Grok Fast through requested_model parameters without legacy model_details", () => {
+    const bytes = encodeCursorRunRequest({
+      modelId: "grok-4.5",
+      requestedModelParameters: [
+        { id: "effort", value: "high" },
+        { id: "fast", value: "true" },
+      ],
+      conversationId: "c1",
+      system: [],
+      messages: [{ role: "user", content: "hi" }],
+    });
+    const msg = fromBinary(AgentClientMessageSchema, bytes);
+    const run = msg.message.case === "runRequest" ? msg.message.value : undefined;
+
+    expect(run?.modelDetails).toBeUndefined();
+    expect(run?.requestedModel).toMatchObject({
+      modelId: "grok-4.5",
+      maxMode: false,
+      parameters: [
+        { id: "effort", value: "high" },
+        { id: "fast", value: "true" },
+      ],
+    });
+  });
+
   test("adds Cursor exact-tool guidance to system prompt blobs when tools are advertised", () => {
     const bytes = encodeCursorRunRequest({
       modelId: "claude-4.6-sonnet",

tests/cursor-effort-suffix.test.ts

@@ -13,6 +13,17 @@ function modelIdFor(modelId: string, reasoning?: string): string {
   return createCursorRequest(parsed).modelId;
 }
 
+function selectionFor(modelId: string, reasoning?: string) {
+  const parsed: OcxParsedRequest = {
+    modelId,
+    context: { messages: [{ role: "user", content: "hi", timestamp: 1 }] },
+    stream: false,
+    options: reasoning ? { reasoning } : {},
+  };
+  const request = createCursorRequest(parsed);
+  return { modelId: request.modelId, parameters: request.requestedModelParameters };
+}
+
 describe("Cursor per-model reasoning-effort suffix", () => {
   test("literal requested efforts pass through when the model supports that tier", () => {
     expect(modelIdFor("cursor/claude-4.6-opus", "high")).toBe("claude-4.6-opus-high");
@@ -60,18 +71,37 @@ describe("Cursor per-model reasoning-effort suffix", () => {
     expect(modelIdFor("cursor/glm-5.2", "max")).toBe("glm-5.2-max");
   });
 
-  test("grok-4.5 uses current low/medium/high tiers and trailing Fast wire ids", () => {
+  test("grok-4.5 uses current tiers and sends Fast as a separate model parameter", () => {
     expect(modelIdFor("cursor/grok-4.5", "low")).toBe("grok-4.5-low");
     expect(modelIdFor("cursor/grok-4.5", "medium")).toBe("grok-4.5-medium");
     expect(modelIdFor("cursor/grok-4.5", "high")).toBe("grok-4.5-high");
     expect(modelIdFor("cursor/grok-4.5", "xhigh")).toBe("grok-4.5-high");
     expect(modelIdFor("cursor/grok-4.5")).toBe("grok-4.5-high");
-    expect(modelIdFor("cursor/grok-4.5-fast", "low")).toBe("grok-4.5-low-fast");
-    expect(modelIdFor("cursor/grok-4.5-fast", "medium")).toBe("grok-4.5-medium-fast");
-    expect(modelIdFor("cursor/grok-4.5-fast", "high")).toBe("grok-4.5-high-fast");
+    expect(selectionFor("cursor/grok-4.5", "high")).toEqual({
+      modelId: "grok-4.5-high",
+      parameters: undefined,
+    });
+    expect(selectionFor("cursor/grok-4.5-fast", "low")).toEqual({
+      modelId: "grok-4.5",
+      parameters: [{ id: "effort", value: "low" }, { id: "fast", value: "true" }],
+    });
+    expect(selectionFor("cursor/grok-4.5-fast", "medium")).toEqual({
+      modelId: "grok-4.5",
+      parameters: [{ id: "effort", value: "medium" }, { id: "fast", value: "true" }],
+    });
+    expect(selectionFor("cursor/grok-4.5-fast", "high")).toEqual({
+      modelId: "grok-4.5",
+      parameters: [{ id: "effort", value: "high" }, { id: "fast", value: "true" }],
+    });
     // Codex-only upper tiers and an omitted effort clamp to Cursor's current top tier.
-    expect(modelIdFor("cursor/grok-4.5-fast", "xhigh")).toBe("grok-4.5-high-fast");
-    expect(modelIdFor("cursor/grok-4.5-fast")).toBe("grok-4.5-high-fast");
+    expect(selectionFor("cursor/grok-4.5-fast", "xhigh")).toEqual({
+      modelId: "grok-4.5",
+      parameters: [{ id: "effort", value: "high" }, { id: "fast", value: "true" }],
+    });
+    expect(selectionFor("cursor/grok-4.5-fast")).toEqual({
+      modelId: "grok-4.5",
+      parameters: [{ id: "effort", value: "high" }, { id: "fast", value: "true" }],
+    });
     expect(cursorModelEffortLadder("grok-4.5")).toEqual(["low", "medium", "high"]);
     expect(cursorModelEffortLadder("grok-4.5-fast")).toEqual(["low", "medium", "high"]);
   });