浏览器操作

export interface ElementRef {  ref: string;  role: string;  name: string;  elementFingerprint: string;}export interface ArtifactRef {  path: string;  sha256: string;  bytes: number;  mediaType: string;}export interface PageObservation {  observationId: string;  url: string;  title: string;  visibleText: { text: string; truncated: boolean };  elements: ElementRef[];  screenshot: ArtifactRef | null;  observedAt: string;}

export type BrowserAction = { risk: "low" | "high"; replayable: boolean } & (  | { kind: "navigate"; target: string }  | { kind: "click"; ref: string }  | { kind: "type"; ref: string; text: string }  | { kind: "select"; ref: string; value: string }  | { kind: "press"; ref: string; key: string }  | { kind: "wait"; ref: string; forText: string });export function isReplayable(action: BrowserAction): boolean {  return action.replayable;}

export interface OpenOptions {  startUrl: string;  headless: boolean;  credentialRef?: string;}export interface BrowserSession {  open(options: OpenOptions, signal: AbortSignal): Promise<void>;  observe(signal: AbortSignal): Promise<PageObservation>;  act(action: BrowserAction, signal: AbortSignal): Promise<ActionOutcome>;  close(): Promise<void>;}export interface ActionOutcome {  ok: boolean;  error?: "stale_element_ref" | "action_not_replayable" | "navigation_failed";}

export function sessionEventDetail(input: {  action: BrowserAction;  observationId: string;  outcome: ActionOutcome;}): string {  const { risk, replayable } = input.action;  return JSON.stringify({    kind: input.action.kind,    target: "ref" in input.action ? input.action.ref : input.action.target,    risk,    replayable,    observationId: input.observationId,    outcome: input.outcome.ok ? "ok" : input.outcome.error,  });}

export async function resolveRef(  session: BrowserSession,  target: ElementRef,  signal: AbortSignal,): Promise<ActionOutcome | ElementRef> {  const current = await session.observe(signal);  const matched = current.elements.find(    (element) =>      element.ref === target.ref && element.elementFingerprint === target.elementFingerprint,  );  if (!matched) return { ok: false, error: "stale_element_ref" };  return matched;}

import { authorize } from "../policy.js";export function authorizeBrowserAction(  tool: RegisteredTool,  action: BrowserAction,  context: PolicyContext,  session: BrowserSession,): PolicyDecision {  if (action.risk === "low") {    return { type: "allow", scope: "browser_low_risk" };  }  return {    type: "ask",    request: createApprovalRequest({      tool,      input: action,      cwd: context.cwd,      network: context.network,      summary: `对当前页面执行高风险动作：${action.kind}`,      risks: [        "外部系统状态不可撤销",        isReplayable(action) ? "重复执行可能产生双份结果" : "该动作不可重放，失败后必须重新观测",        action.kind === "type" ? "输入内容可能包含敏感数据" : "目标元素可能已失效",      ],    }),  };}

const recorded = appendEvent(state, "approval_requested", request.actionDigest);state = transitionState(recorded, "waiting", "approval_required");

export const browserActTool = defineTool({  name: "browser_act",  description: "Execute one BrowserAction against the current browser session",  effect: "external",  schema: z.object({    kind: z.enum(["navigate", "click", "type", "select", "press", "wait"]),    target: z.string().min(1),    value: z.string().nullable(),    risk: z.enum(["low", "high"]),  }).strict(),  async execute(input, context) {    const session = requireBrowserSession(context.cwd, context.signal);    const action = toBrowserAction(input, session.currentObservationId);    const decision = authorizeBrowserAction(browserActTool, action, policyContext, session);    if (decision.type === "deny") return { ok: false, error: decision.reason };    if (decision.type === "ask") return { pending: decision.request.actionDigest };    return session.act(action, context.signal);  },});

expect(await runScenario({ mutateAfterObserve: true })).toMatchObject({  error: "stale_element_ref",});expect(await runScenario({ approveHighRisk: false })).toMatchObject({  status: "waiting",  requestReason: "approval_required",});const detail = sessionEventDetail({ action, observationId, outcome });expect(detail).not.toContain("hunter2");expect(contextForModel).not.toContain("sessionId=");expect(observation.screenshot?.sha256).toMatch(/^[0-9a-f]{64}$/);