@roll-agent/browser-use-agent 0.26.0 → 0.27.0

package/dist/native-visual-activity-session.d.ts

@@ -1,16 +1,53 @@
 import type { VisualActivityTone } from "./visual-activity.ts";
 import type { NativeMouseClickPreview, NativeMouseMotionObserver, NativeMouseMotionPreview } from "./native-mouse-motion.ts";
 type NativeVisualTarget = {
-    evaluateJson<T = unknown>(expression: string): Promise<T>;
+    evaluateJson<T = unknown>(expression: string, options?: {
+        timeoutMs?: number;
+    }): Promise<T>;
 };
 type NativeVisualHighlightOptions = {
     readonly label?: string;
     readonly padding?: number;
     readonly tone?: VisualActivityTone;
 };
+export type NativeExecutionCard = {
+    readonly ownerId: string;
+    readonly epoch: number;
+    readonly revision: number;
+    readonly actionRevision: number;
+    readonly title: string;
+    readonly stage: string;
+    readonly action?: string;
+    readonly target?: string;
+    readonly recent: readonly string[];
+    readonly tone?: VisualActivityTone;
+    readonly lingerMs?: number;
+};
+export type NativeExecutionRect = {
+    readonly ownerId: string;
+    readonly epoch: number;
+    readonly actionRevision: number;
+    readonly x: number;
+    readonly y: number;
+    readonly width: number;
+    readonly height: number;
+};
+export type NativeExecutionPointer = {
+    readonly ownerId: string;
+    readonly epoch: number;
+    readonly actionRevision: number;
+    readonly type: "mouseMoved" | "mousePressed";
+    readonly x: number;
+    readonly y: number;
+};
 export declare class NativeVisualActivitySession implements NativeMouseMotionObserver {
     private readonly target;
+    private executionLifecycleStarted;
     constructor(target: NativeVisualTarget);
+    showExecutionCard(card: NativeExecutionCard, mode?: "begin" | "update" | "complete"): Promise<boolean>;
+    clearExecutionCard(card: Pick<NativeExecutionCard, "ownerId" | "epoch" | "revision" | "actionRevision">): Promise<boolean>;
+    highlightExecutionRect(rect: NativeExecutionRect): Promise<boolean>;
+    previewExecutionPointer(pointer: NativeExecutionPointer): Promise<boolean>;
     begin(label: string, tone?: VisualActivityTone): Promise<boolean>;
     highlightSelector(selector: string, options?: NativeVisualHighlightOptions): Promise<boolean>;
     previewMouseMotion(preview: NativeMouseMotionPreview): Promise<void>;

package/dist/tools/browser-operate.d.ts

@@ -0,0 +1,252 @@
+export declare const browserOperate: import("@roll-agent/sdk").ToolDefinition<{
+    pageId: string;
+    allowedOrigins: string[];
+    goal: string;
+    values?: {
+        name: string;
+        text: string;
+    }[] | undefined;
+    timeoutMs?: number | undefined;
+    model?: string | undefined;
+    browserActionApproval?: {
+        id: string;
+    } | undefined;
+    formTask?: {
+        mode: "create" | "edit";
+        fields: ({
+            name: string;
+            intent: "set";
+            valueName: string;
+            comparison?: "auto" | "literal" | "semantic" | undefined;
+        } | {
+            name: string;
+            intent: "preserve";
+        })[];
+        stopAt?: "applied" | "current-view" | undefined;
+    } | undefined;
+    readTask?: {
+        target: string;
+        captureView: string;
+        outputs: string[];
+        terminal: {
+            view: string;
+            selectedTab?: string | undefined;
+        };
+    } | undefined;
+    delegatedDecisions?: string[] | undefined;
+    maxTextCalls?: number | undefined;
+    maxRecoveryDecisions?: number | undefined;
+    blockedNames?: string[] | undefined;
+    engine?: "jev" | "sampling" | undefined;
+    strategy?: "task" | "fields" | undefined;
+    maxSteps?: number | undefined;
+}, {
+    status: "failed" | "cancelled" | "needs_input" | "model_done" | "interaction_done" | "needs_reasoning" | "blocked" | "step_limit";
+    elapsedMs: number;
+    verified: false;
+    steps: {
+        executed: boolean;
+        step: number;
+        provider: string;
+        requestedModel: string;
+        resolvedModel: string;
+        operation: string;
+        observationMs: number;
+        decisionMs: number;
+        actionMs: number;
+        target?: string | undefined;
+        error?: string | undefined;
+        usage?: {
+            cost?: number | undefined;
+            inputTokens?: number | undefined;
+            outputTokens?: number | undefined;
+        } | undefined;
+        distributions?: Record<string, {
+            probabilities: Record<string, number>;
+            confidence?: number | undefined;
+        }> | undefined;
+        valueName?: string | undefined;
+        decisionAttempts?: number | undefined;
+        textMs?: number | undefined;
+        recovery?: boolean | undefined;
+        requirement?: string | undefined;
+    }[];
+    error?: string | undefined;
+    progress?: {
+        missing: string[];
+        taskRunId: string;
+        contractDigest: string;
+        phase: "collect" | "return" | "done";
+        evidence: {
+            id: string;
+            url: string;
+            frameId: string;
+            documentId: string;
+            text: string;
+            field: string;
+            end: number;
+            observedAt: string;
+            observationId: string;
+            regionId: string;
+            regionName: string;
+            start: number;
+            attribution: "model-selected-observed-text";
+        }[];
+        invalidatedEvidence: {
+            id: string;
+            url: string;
+            frameId: string;
+            documentId: string;
+            text: string;
+            field: string;
+            end: number;
+            observedAt: string;
+            observationId: string;
+            regionId: string;
+            regionName: string;
+            start: number;
+            attribution: "model-selected-observed-text";
+        }[];
+        conflicts: number;
+        capturedRegion?: {
+            id: string;
+            url: string;
+            name: string;
+            frameId: string;
+            documentId: string;
+            digest: string;
+        } | undefined;
+    } | undefined;
+    execution?: {
+        changes: {
+            scope: "delegated" | "preserved" | "unassigned";
+            kind: "changed" | "became_unknown" | "appeared";
+            field: string;
+            observationId: string;
+            identity: string;
+            before?: string | undefined;
+            after?: string | undefined;
+            afterAction?: string | undefined;
+        }[];
+        revision: number;
+        invocationId: string;
+        documentChanged: boolean;
+        form?: {
+            mode: "create" | "edit";
+            stopAt: "applied" | "current-view";
+            fields: {
+                status: "unknown" | "satisfied" | "unsatisfied";
+                id: string;
+                name: string;
+                reason: string;
+                intent: "set" | "preserve";
+                comparison: "auto" | "literal" | "semantic";
+                expected?: string | undefined;
+                target?: {
+                    label: string;
+                    key: string;
+                    optionsKey: string;
+                } | undefined;
+                fingerprint?: string | undefined;
+                binding?: {
+                    role: string;
+                    name: string;
+                    identity: string;
+                    key?: string | undefined;
+                    fieldLabel?: string | undefined;
+                    readOnlyDisplay?: true | undefined;
+                } | undefined;
+                entry?: {
+                    role: string;
+                    name: string;
+                    identity: string;
+                    key?: string | undefined;
+                    fieldLabel?: string | undefined;
+                    readOnlyDisplay?: true | undefined;
+                } | undefined;
+                entryProof?: {
+                    status: "unknown" | "satisfied" | "unsatisfied";
+                    fingerprint: string;
+                } | undefined;
+                initial?: string | undefined;
+                current?: string | undefined;
+                semanticProof?: {
+                    status: "satisfied" | "unsatisfied";
+                    fingerprint: string;
+                } | undefined;
+            }[];
+            panel?: {
+                status: "pending" | "rejected" | "accepted";
+                key: string;
+                queryIdentity?: string | undefined;
+            } | undefined;
+            focus?: string | undefined;
+        } | undefined;
+        focus?: string | undefined;
+        latestObservation?: string | undefined;
+        lastAction?: {
+            executed: boolean;
+            operation: string;
+            target?: string | undefined;
+            error?: string | undefined;
+        } | undefined;
+    } | undefined;
+    resolvedValues?: {
+        source: "selected" | "supplied" | "derived" | "generated";
+        text: string;
+        field: string;
+        evidence: string;
+        sourceIds: string[];
+        targetKey?: string | undefined;
+    }[] | undefined;
+    question?: string | undefined;
+    handoff?: {
+        fields: {
+            status: "unknown" | "satisfied" | "unsatisfied";
+            id: string;
+            name: string;
+            reason: string;
+            intent: "set" | "preserve";
+            valueTruncated: boolean;
+            expected?: string | undefined;
+            current?: string | undefined;
+        }[];
+        observationFresh: boolean;
+        diagnosticsOmitted: true;
+        totalSteps: number;
+        omittedSteps: number;
+        next: string;
+    } | undefined;
+    recoveryDecisions?: number | undefined;
+    fieldEvidence?: {
+        status: "pending" | "preserved" | "applied";
+        field: string;
+        instruction: string;
+        targetKey: string;
+        expected?: string | undefined;
+        source?: string | undefined;
+        current?: string | undefined;
+    }[] | undefined;
+    pendingRequirements?: string[] | undefined;
+    finalObservation?: unknown;
+    textCalls?: {
+        elapsedMs: number;
+        purpose: "preflight" | "resolve_value" | "resolve_batch" | "validate_value" | "validate_batch" | "review_field" | "review" | "clarify";
+        cached: boolean;
+        startedAtMs?: number | undefined;
+        endedAtMs?: number | undefined;
+        modelCalls?: {
+            elapsedMs: number;
+            provider: string;
+            requestedModel: string;
+            resolvedModel: string;
+            error?: string | undefined;
+            usage?: {
+                cost?: number | undefined;
+                inputTokens?: number | undefined;
+                outputTokens?: number | undefined;
+            } | undefined;
+            fallback?: boolean | undefined;
+        }[] | undefined;
+    }[] | undefined;
+}>;

package/dist/tools/zhipin-diagnose-browser-state.d.ts

@@ -14,8 +14,8 @@ declare const StorageEntrySummarySchema: z.ZodObject<{
 }, "strip", z.ZodTypeAny, {
     key: string;
     area: "localStorage" | "sessionStorage";
-    valueLength: number;
     valueKind: "string" | "json" | "empty";
+    valueLength: number;
     jsonKind?: "string" | "number" | "boolean" | "object" | "null" | "array" | undefined;
     jsonTopLevelKeys?: string[] | undefined;
     jsonArrayLength?: number | undefined;
@@ -25,8 +25,8 @@ declare const StorageEntrySummarySchema: z.ZodObject<{
 }, {
     key: string;
     area: "localStorage" | "sessionStorage";
-    valueLength: number;
     valueKind: "string" | "json" | "empty";
+    valueLength: number;
     jsonKind?: "string" | "number" | "boolean" | "object" | "null" | "array" | undefined;
     jsonTopLevelKeys?: string[] | undefined;
     jsonArrayLength?: number | undefined;
@@ -303,8 +303,8 @@ export declare const zhipinDiagnoseBrowserState: import("@roll-agent/sdk").ToolD
         localStorage: {
             key: string;
             area: "localStorage" | "sessionStorage";
-            valueLength: number;
             valueKind: "string" | "json" | "empty";
+            valueLength: number;
             jsonKind?: "string" | "number" | "boolean" | "object" | "null" | "array" | undefined;
             jsonTopLevelKeys?: string[] | undefined;
             jsonArrayLength?: number | undefined;
@@ -315,8 +315,8 @@ export declare const zhipinDiagnoseBrowserState: import("@roll-agent/sdk").ToolD
         sessionStorage: {
             key: string;
             area: "localStorage" | "sessionStorage";
-            valueLength: number;
             valueKind: "string" | "json" | "empty";
+            valueLength: number;
             jsonKind?: "string" | "number" | "boolean" | "object" | "null" | "array" | undefined;
             jsonTopLevelKeys?: string[] | undefined;
             jsonArrayLength?: number | undefined;

package/dist/tools/zhipin-generate-reply-preview.d.ts

@@ -30,11 +30,11 @@ export declare const zhipinGenerateReplyPreview: import("@roll-agent/sdk").ToolD
     success: boolean;
     error?: string | undefined;
     elapsedMs?: number | undefined;
+    stage?: string | undefined;
     confidence?: number | undefined;
     requestId?: string | undefined;
     expiresAt?: number | undefined;
     suggestedReply?: string | undefined;
-    stage?: string | undefined;
     preparedReplyId?: string | undefined;
     clientTimeoutMs?: number | undefined;
     lastStartedPhase?: string | undefined;

package/dist/tools/zhipin-get-candidate-info.d.ts

@@ -40,8 +40,8 @@ export declare const zhipinGetCandidateInfo: import("@roll-agent/sdk").ToolDefin
         source: "candidate_message" | "conversation_history" | "candidate_expected_location" | "communication_position";
         text: string;
         confidence: number;
-        city?: string | undefined;
         intent?: "nearby_store" | "store_address" | "expected_area" | undefined;
+        city?: string | undefined;
     }[];
     chatMessages: {
         index: number;

package/dist/tools/zhipin-scroll-view.d.ts

@@ -9,15 +9,13 @@ type ZhipinScrollViewDeps = {
 export declare function setZhipinScrollViewDepsForTests(override: Partial<ZhipinScrollViewDeps> | undefined): void;
 export declare const zhipinScrollView: import("@roll-agent/sdk").ToolDefinition<{
     surface: "chat-list" | "chat-history" | "recommend-list";
+    steps?: number | undefined;
     direction?: "up" | "down" | undefined;
     distance?: number | undefined;
-    steps?: number | undefined;
     settleMs?: number | undefined;
 }, {
     success: boolean;
-    direction: "up" | "down";
-    stepsRequested: number;
-    stepsCompleted: number;
+    position: "unknown" | "middle" | "top" | "bottom" | "only-page";
     before: {
         containerFound: boolean;
         containerLabel: string;
@@ -38,8 +36,10 @@ export declare const zhipinScrollView: import("@roll-agent/sdk").ToolDefinition<
         atEnd: boolean;
         itemCount: number;
     };
+    direction: "up" | "down";
+    stepsRequested: number;
+    stepsCompleted: number;
     reachedBoundary: boolean;
-    position: "unknown" | "middle" | "top" | "bottom" | "only-page";
     surface: "chat-list" | "chat-history" | "recommend-list";
     atTop: boolean;
     atBottom: boolean;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@roll-agent/browser-use-agent",
-  "version": "0.26.0",
+  "version": "0.27.0",
   "type": "module",
   "repository": {
     "type": "git",
@@ -44,9 +44,9 @@
     }
   },
   "dependencies": {
+    "@roll-agent/browser": "0.12.0",
     "@roll-agent/reply-authority-client": "0.7.0",
-    "@roll-agent/sdk": "0.5.1",
-    "@roll-agent/browser": "0.11.0",
+    "@roll-agent/sdk": "0.7.0",
     "zod": "^3.25.76"
   },
   "devDependencies": {

package/references/browser-exploration.md

@@ -4,6 +4,48 @@
 
 单步 `browser_snapshot` / `click_ref` / `type_ref` 的区域观察、上下文选择和严格 ref 用法见 [通用 ref 操作](generic-browser-refs.md)。
 
+遵循用户明确指定的工具；未知目标先观察。资料与文案齐全、字段范围和停止点明确的多字段表单，尤其涉及下拉、联动字段或弹层编辑器时，首选 `browser_operate`；无需提前知道点击顺序，不因宿主能拆步骤就逐字段执行。已确定步骤的短脚本、批量输入与断言用 `browser_execute`，它不调用 Jev。单次点击/输入、范围尚不明确的探索或委派受阻后的局部修正用单步工具。两者都不能绕过发布/保存授权、动态观察、结果验证或失败后现场检查。
+
+## 通过完整目标执行
+
+`browser_operate` 默认 `strategy:"task"`。调用方交代原始目标、已知事实和已准备文案；所选决策引擎在当前页面选择下一步。例：
+
+```json
+{
+  "pageId": "真实页面 ID",
+  "goal": "填写岗位草稿。优先处理前置选择和必填字段，使用资料中的内容，停止在发布前。",
+  "values": [
+    { "name": "职位名称", "text": "餐饮兼职服务员" },
+    { "name": "岗位说明", "text": "调用方已准备并允许使用的完整岗位文案" },
+    { "name": "最低月薪", "text": "5000" },
+    { "name": "最高月薪", "text": "6000" }
+  ],
+  "strategy": "task",
+  "allowedOrigins": ["https://example.com"],
+  "blockedNames": ["发布", "提交"]
+}
+```
+
+示例中的页面、origin 和事实必须替换为实际输入。必填标记不允许编造缺失事实；用户要求的选填字段仍须处理。调用方需要拟写内容时在交付前准备好；决策引擎不负责自由写作。操作过程中无需宿主逐项检查、字段完成复核或内部恢复。
+
+一次决策请求共享当前页面状态，询问一个带目标的完整动作，以及各输入框对应的来源值。完整动作把操作与目标绑定，避免不兼容组合。各来源问题按对应输入框独立执行，只消费选中动作的来源答案；输入框与其来源问题显式绑定，未使用分支的置信度不影响执行。决策引擎可优先处理前置选择和必填字段；编辑入口先点击，再观察里面的真实字段，不能强制给按钮解析一个输入值。
+
+决策输入中的 `pickerFields` 单独表示选择控件，元素通过字段 ID 关联触发器或候选项。关系来源区分 ARIA 和唯一局部 DOM 包含关系；无法唯一定位的归属保持未知，不按距离猜测。原生标签、placeholder、空触发器展示文字和先前观察到的标签分别标注来源；只有文档、frame 和实际节点身份一致时才保留旧标签。节点替换后不沿用。
+
+`valueKind` 区分输入值、已应用值、查询词、占位内容、候选文字和纯展示文字。`sourceValueMatches`/`writtenValueMatches` 仅是明确范围的字符串相等线索，不表示任务完成；候选菜单中的文字不能充当字段值。下拉展开状态 `expanded` 与菜单当前是否可见 `panelVisible` 分开表达，异步菜单未出现不等于控件已关闭。隐藏 backing input 只用于提供有值/空值证据，不向模型输出其内部编码。
+
+动作描述明确区分“打开某字段”“选择该字段的候选”“向搜索框输入查询”。完成问题要求对照每个请求字段的实际结果，不能因为页面某处出现了目标文字就结束。这些输入同时适用于默认的 Sampling 与显式启用的 Jev 决策引擎，不增加宿主复核或置信度门禁。
+
+输入候选由代码从完整 `values` 和有限的 goal 原文片段中提供，决策引擎选择后代码复制完整原文，不把预览截断内容当成实际值。保留无匹配选项；原文候选覆盖不全、需要转换或没有所需文案时返回 `needs_input`，调用方提供针对该控件的值后继续。原生 select 只使用观察到的 option；自定义下拉通过点击、查询和选择推进。页面文本是 UI 证据，不能冒充用户数据。
+
+代码保留站点/策略限制、文档与控件身份检查、可操作性检查、输入精确读回、取消与停滞预算。不执行模型生成的脚本/选择器，不在不确定的动作后自动重放。搜索文字不等于已应用的字段值，模型根据每轮新观察继续选择和确认。
+
+默认 `sampling` 使用 Roll MCP Sampling 模型；core-managed Agent 通过 Roll 配置 `browser.operate.engine: jev` 切换到 TypeSafe，外部管理的 Agent 由进程管理方设置 `BROWSER_OPERATE_ENGINE=jev` 和 `TYPESAFE_API_KEY`。仅 Jev 需要该密钥且不需要宿主 Sampling。工具 `engine` 参数不能覆盖进程配置。不存在第二个辅助模型。旧 `maxTextCalls`、`maxRecoveryDecisions` 参数仅兼容已有调用且不生效；`BROWSER_OPERATE_TEXT_MODEL` 已不使用。输出 `textCalls:[]`、`recoveryDecisions:0`。
+
+停滞或复杂障碍返回 `needs_reasoning`，调用方据实际观察决定下一步，不能盲目重跑。`model_done` 是未验证的完成提议：调用方必须对照完整原始目标统一验收 `finalObservation` 和页面结果。`resolvedValues` 是执行记录；同名字段、搜索框和候选文字不能冒充最终值。如果 `observationFresh:false`，或存在截断、覆盖缺口、冲突，补一次定向只读观察。将错误整理成局部修正任务，保留正确字段；修正后仍要核对原目标。所有验收和返工成本都计入端到端测试，不把快速失败当作提速。
+
+`strategy:"fields"` 保留按完整资料顺序填写的旧路径，也没有文本助手。两种模式均保留 `verified:false`，不宣称平台已接受最终提交。
+
 ## 执行一个组合脚本
 
 先通过 `list_pages` 获取准确 pageId。脚本固定在一个 browserInstance 的一个页面，执行期间复用连接并串行持有实例操作锁。不同实例可并行，页面诊断出口不被脚本占住。
@@ -26,6 +68,12 @@
 
 默认上限：30 秒、100 次 helper 调用、32MiB 脚本堆，文本结果/日志/轨迹合计 64KiB。`timeoutMs`、`maxCalls` 只能调低。截图单独保存为本地产物，返回 `artifacts` 路径；最多 10 张。脚本环境没有文件读写、模块导入、`process`、`fetch`、裸 CDP 或页面 `evaluate`。
 
+`page` 不是 Playwright。定位器构造器返回普通定位数据，没有 `.all()`、`.count()`、`.fill()`、`.textContent()` 或 `.getAttribute()` 方法；使用 `page.count(target)`、`page.fill(target,text)`、`page.read(target,{attribute})`。不支持 `page.url()` 或 `:visible` 等 Playwright 选择器扩展。遇到 `SCRIPT_ERROR` 时先检查已执行的 `actions`，不要重放已成功的操作。
+
+陌生表单先调用独立的 `browser_snapshot`；空白 `scope` 与省略相同。脚本内 `page.snapshot()` 受 origin 限制不展开 iframe，可能只看到导航栏。此时使用独立快照返回的 `ref` 与 `snapshotId` 构造 `page.ref(ref,snapshotId)`，它会保留 frame 身份并继续执行现有 frame/origin 检查。不要猜测 iframe 的下标，也不要仅凭脚本内快照断言表单不存在。
+
+字段可能以摘要、编辑或补充入口呈现，输入控件要展开后才出现。用户要求填写某字段时，先检查与它明确相关的已观察入口，再决定是否缺少可操作路径；必要时滚动对应区域。用户明确指定的字段即使在页面上标为可选，也需要完成或清楚报告真实阻塞原因。只提供部分表单值的快照不是“字段不支持”的证据。
+
 ## Helpers
 
 | Helper                                                                 | 用途                                                               |
@@ -139,3 +187,7 @@ export async function readProfile(
 ```
 
 `BrowserProgramDriver` 属于宿主信任边界。直接嵌入时，调用方负责实例锁、整段审批、当前站点/策略检查，以及 `close()` 中断连接；内核不会为任意自定义 driver 补齐这些能力。生产 MCP 入口使用 `BrowserScriptPageDriver` 与 `executeBrowserTool` 统一完成这些检查。执行结束会关闭 driver，不应跨执行复用已关闭对象。
+
+决策请求会去重通用说明和资料表，并按字节预算缩短模型预览；保留原始 goal、所有动作 ID 与执行用完整文本。预览被截断不代表字段不存在，最终观察不使用这种压缩视图。同一决策请求另有一个整项完成判断，仅在动作建议 DONE 时使用；它可指出未完成要求或尚未处理的面板，不额外调用宿主模型，也不逐字段增加检查调用。
+
+调用方必须原样转发用户目标和原样类文本，不能自行增删换行。原文只有一段时不得为了展示而重排成两段；最终核验包含这一点。

package/references/env.yaml

@@ -22,6 +22,10 @@ required:
     purpose: 招聘事件 Open API Bearer token；公开 npm 包不内置兜底 token，必须由部署环境提供
     example: token_xxx
 optional:
+  - name: TYPESAFE_API_KEY
+    secret: true
+    purpose: TypeSafe 官方 Jev API 密钥，仅 browser.operate.engine 为 jev 时需要；可在 roll.config.yaml 的 agents.env.browser-use-agent.TYPESAFE_API_KEY 或配置台 Agent 环境变量中填写
+    example: token_xxx
   - name: REPLY_AUTHORITY_TIMEOUT_MS
     secret: false
     type: number
@@ -80,6 +84,13 @@ optional:
     sourcePath: [browser, instances]
     purpose: Roll core 注入的多浏览器实例声明 JSON；由 `browser.instances` 派生，包含每个实例的 profile、CDP 端口、session 目录、可选 profileName/profileColor/windowBounds 和招聘事件归因 ID；orchestrator 不应手工拼接，手工配置时必须保持 cdpPort/userDataDir 唯一
     example: '{"defaultInstance":"boss-a","instances":{"boss-a":{"platform":"zhipin","mode":"managed-cdp","cdpPort":9222,"userDataDir":"/tmp/roll-browser/profiles/boss-a","profileName":"boss-a","profileColor":"#2563EB","windowBounds":{"x":0,"y":0,"width":680,"height":1000},"trackingAgentId":"zhipin-boss-a"}}}'
+  - name: BROWSER_OPERATE_ENGINE
+    secret: false
+    configurable: false
+    sourcePath: [browser, operate, engine]
+    purpose: Roll core 从 browser.operate.engine 注入的 browser_operate 决策引擎；默认 sampling，jev 快速模式需要 TYPESAFE_API_KEY；不在 Agent 环境变量中手工填写
+    default: sampling
+    example: jev
   - name: BROWSER_VISUAL_CURSOR
     secret: false
     type: boolean