@roll-agent/browser-use-agent 0.22.2 → 0.23.0

package/dist/reply-authority/prepared-reply-decision.d.ts

@@ -0,0 +1,56 @@
+import { z } from "zod";
+export declare const PreparedReplyOptionValues: readonly ["option_1", "option_2"];
+export declare const PreparedReplyDecisionSourceValues: readonly ["judge", "orchestrator"];
+export declare const PreparedReplySendDecisionSourceValues: readonly ["judge", "orchestrator", "service_recommended_fallback", "explicit_no_judge"];
+export declare const PreparedReplyFallbackReasons: {
+    readonly RUBRIC_FETCH_FAILED: "rubric_fetch_failed";
+    readonly RUBRIC_MISMATCH: "rubric_mismatch";
+    readonly INVALID_VARIANT_SHAPE: "invalid_variant_shape";
+    readonly JUDGE_SAMPLING_FAILED: "judge_sampling_failed";
+    readonly JUDGE_OUTPUT_INVALID: "judge_output_invalid";
+};
+export declare const PreparedReplyVariantDecisionSchema: z.ZodObject<{
+    chosenOption: z.ZodEnum<["option_1", "option_2"]>;
+    reason: z.ZodString;
+    confirmedFindingCodes: z.ZodOptional<z.ZodEffects<z.ZodArray<z.ZodEnum<["too_many_questions", "audit_tone", "reply_overpacked", "off_axis_fact_disclosure"]>, "many">, ("too_many_questions" | "audit_tone" | "reply_overpacked" | "off_axis_fact_disclosure")[], ("too_many_questions" | "audit_tone" | "reply_overpacked" | "off_axis_fact_disclosure")[]>>;
+    judgeModel: z.ZodOptional<z.ZodString>;
+}, "strip", z.ZodTypeAny, {
+    reason: string;
+    chosenOption: "option_1" | "option_2";
+    confirmedFindingCodes?: ("too_many_questions" | "audit_tone" | "reply_overpacked" | "off_axis_fact_disclosure")[] | undefined;
+    judgeModel?: string | undefined;
+}, {
+    reason: string;
+    chosenOption: "option_1" | "option_2";
+    confirmedFindingCodes?: ("too_many_questions" | "audit_tone" | "reply_overpacked" | "off_axis_fact_disclosure")[] | undefined;
+    judgeModel?: string | undefined;
+}>;
+export declare const JudgeModelOutputSchema: z.ZodObject<{
+    reason: z.ZodString;
+    chosenOption: z.ZodEnum<["option_1", "option_2"]>;
+} & {
+    confirmedFindingCodes: z.ZodEffects<z.ZodArray<z.ZodEnum<["too_many_questions", "audit_tone", "reply_overpacked", "off_axis_fact_disclosure"]>, "many">, ("too_many_questions" | "audit_tone" | "reply_overpacked" | "off_axis_fact_disclosure")[], ("too_many_questions" | "audit_tone" | "reply_overpacked" | "off_axis_fact_disclosure")[]>;
+}, "strip", z.ZodTypeAny, {
+    reason: string;
+    confirmedFindingCodes: ("too_many_questions" | "audit_tone" | "reply_overpacked" | "off_axis_fact_disclosure")[];
+    chosenOption: "option_1" | "option_2";
+}, {
+    reason: string;
+    confirmedFindingCodes: ("too_many_questions" | "audit_tone" | "reply_overpacked" | "off_axis_fact_disclosure")[];
+    chosenOption: "option_1" | "option_2";
+}>;
+export type PreparedReplyOption = (typeof PreparedReplyOptionValues)[number];
+export type PreparedReplyDecisionSource = (typeof PreparedReplyDecisionSourceValues)[number];
+export type PreparedReplySendDecisionSource = (typeof PreparedReplySendDecisionSourceValues)[number];
+export type PreparedReplyFallbackReason = (typeof PreparedReplyFallbackReasons)[keyof typeof PreparedReplyFallbackReasons];
+export type PreparedReplyVariantDecision = z.infer<typeof PreparedReplyVariantDecisionSchema>;
+export type PreparedReplyJudgement = {
+    readonly kind: "decision";
+    readonly source: PreparedReplyDecisionSource;
+    readonly decision: PreparedReplyVariantDecision;
+} | {
+    readonly kind: "fallback";
+    readonly source: "service_recommended_fallback";
+    readonly recommendedOption: PreparedReplyOption;
+    readonly reason: PreparedReplyFallbackReason;
+};

package/dist/reply-authority/prepared-reply-judge-context.d.ts

@@ -0,0 +1,75 @@
+import { type GenerateReplyToolInput } from "@roll-agent/reply-authority-client";
+import { z } from "zod";
+export declare const PreparedReplyJudgeContextSchema: z.ZodObject<{
+    candidateMessage: z.ZodString;
+    recentConversation: z.ZodArray<z.ZodString, "many">;
+    candidateInfo: z.ZodOptional<z.ZodObject<Pick<{
+        name: z.ZodOptional<z.ZodString>;
+        position: z.ZodOptional<z.ZodString>;
+        expectedPosition: z.ZodOptional<z.ZodString>;
+        communicationPosition: z.ZodOptional<z.ZodString>;
+        age: z.ZodOptional<z.ZodString>;
+        gender: z.ZodOptional<z.ZodString>;
+        experience: z.ZodOptional<z.ZodString>;
+        education: z.ZodOptional<z.ZodString>;
+        expectedSalary: z.ZodOptional<z.ZodString>;
+        expectedLocation: z.ZodOptional<z.ZodString>;
+        jobAddress: z.ZodOptional<z.ZodString>;
+        height: z.ZodOptional<z.ZodString>;
+        weight: z.ZodOptional<z.ZodString>;
+        healthCertificate: z.ZodOptional<z.ZodBoolean>;
+        activeTime: z.ZodOptional<z.ZodString>;
+        info: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
+        fullText: z.ZodOptional<z.ZodString>;
+    }, "info" | "experience" | "age" | "education" | "expectedLocation" | "expectedPosition" | "expectedSalary" | "communicationPosition">, "strip", z.ZodTypeAny, {
+        info?: string[] | undefined;
+        experience?: string | undefined;
+        age?: string | undefined;
+        education?: string | undefined;
+        expectedLocation?: string | undefined;
+        expectedPosition?: string | undefined;
+        expectedSalary?: string | undefined;
+        communicationPosition?: string | undefined;
+    }, {
+        info?: string[] | undefined;
+        experience?: string | undefined;
+        age?: string | undefined;
+        education?: string | undefined;
+        expectedLocation?: string | undefined;
+        expectedPosition?: string | undefined;
+        expectedSalary?: string | undefined;
+        communicationPosition?: string | undefined;
+    }>>;
+    preferredBrand: z.ZodOptional<z.ZodString>;
+}, "strip", z.ZodTypeAny, {
+    candidateMessage: string;
+    recentConversation: string[];
+    candidateInfo?: {
+        info?: string[] | undefined;
+        experience?: string | undefined;
+        age?: string | undefined;
+        education?: string | undefined;
+        expectedLocation?: string | undefined;
+        expectedPosition?: string | undefined;
+        expectedSalary?: string | undefined;
+        communicationPosition?: string | undefined;
+    } | undefined;
+    preferredBrand?: string | undefined;
+}, {
+    candidateMessage: string;
+    recentConversation: string[];
+    candidateInfo?: {
+        info?: string[] | undefined;
+        experience?: string | undefined;
+        age?: string | undefined;
+        education?: string | undefined;
+        expectedLocation?: string | undefined;
+        expectedPosition?: string | undefined;
+        expectedSalary?: string | undefined;
+        communicationPosition?: string | undefined;
+    } | undefined;
+    preferredBrand?: string | undefined;
+}>;
+export type PreparedReplyJudgeContext = z.infer<typeof PreparedReplyJudgeContextSchema>;
+export declare function redactPreparedReplyJudgeText(value: string): string;
+export declare function buildPreparedReplyJudgeContext(input: GenerateReplyToolInput): PreparedReplyJudgeContext;

package/dist/reply-authority/prepared-reply-store.d.ts

@@ -1,6 +1,8 @@
 import type { ReplyGateAdvisoryCode, ReplyVariantKind } from "@roll-agent/reply-authority-client";
-export declare const PreparedReplyOptionValues: readonly ["option_1", "option_2"];
-export type PreparedReplyOption = (typeof PreparedReplyOptionValues)[number];
+import { PreparedReplyOptionValues, type PreparedReplyFallbackReason, type PreparedReplyJudgement, type PreparedReplyOption } from "./prepared-reply-decision.ts";
+import type { PreparedReplyJudgeContext } from "./prepared-reply-judge-context.ts";
+export { PreparedReplyOptionValues };
+export type { PreparedReplyOption };
 export type PreparedReplyVariantOption = {
     readonly option: PreparedReplyOption;
     readonly variant: ReplyVariantKind;
@@ -12,19 +14,28 @@ export type PreparedReplyVariantFinding = {
     readonly code: ReplyGateAdvisoryCode;
     readonly description: string;
 };
-export type PreparedReplyVariantGroup = {
+type PreparedReplyVariantGroupBase = {
     readonly groupId: string;
-    readonly options: readonly PreparedReplyVariantOption[];
-    readonly findings: readonly PreparedReplyVariantFinding[];
     readonly rubricVersion: string;
     readonly rubricHash: string;
+    readonly feedbackExpiresAt?: number;
     readonly target: {
         readonly platform: "zhipin";
         readonly tenantId: string;
         readonly conversationId: string;
     };
-    readonly recommendedOption: PreparedReplyOption;
 };
+export type PreparedReplyVariantGroup = (PreparedReplyVariantGroupBase & {
+    readonly state: "judge_ready";
+    readonly options: readonly PreparedReplyVariantOption[];
+    readonly findings: readonly PreparedReplyVariantFinding[];
+    readonly recommendedOption: PreparedReplyOption;
+    readonly judgeContext: PreparedReplyJudgeContext;
+}) | (PreparedReplyVariantGroupBase & {
+    readonly state: "not_learned";
+    readonly chosenVariant: "draft";
+    readonly reason: PreparedReplyFallbackReason;
+});
 export type PreparedReplyRecord = {
     readonly preparedReplyId: string;
     readonly signedEnvelope: string;
@@ -35,6 +46,7 @@ export type PreparedReplyRecord = {
     readonly requestId?: string;
     readonly unreadCountBeforeReply?: number;
     readonly variantGroup?: PreparedReplyVariantGroup;
+    readonly judgement?: PreparedReplyJudgement;
 };
 export type PreparedReplyConsumeResult = {
     readonly ok: true;
@@ -44,7 +56,9 @@ export type PreparedReplyConsumeResult = {
     readonly reason: "not_found" | "expired" | "consumed";
 };
 export type PreparedReplyInspectResult = PreparedReplyConsumeResult;
+export type PreparedReplyJudgementUpdateResult = PreparedReplyConsumeResult;
 export declare function savePreparedReply(input: Omit<PreparedReplyRecord, "preparedReplyId">, nowSeconds?: number): PreparedReplyRecord;
 export declare function consumePreparedReply(preparedReplyId: string, nowSeconds?: number): PreparedReplyConsumeResult;
 export declare function inspectPreparedReply(preparedReplyId: string, nowSeconds?: number): PreparedReplyInspectResult;
+export declare function setPreparedReplyJudgement(preparedReplyId: string, judgement: PreparedReplyJudgement, nowSeconds?: number): PreparedReplyJudgementUpdateResult;
 export declare function resetPreparedReplyStoreForTests(): void;

package/dist/reply-authority/reply-feedback-outbox.d.ts

@@ -0,0 +1,24 @@
+import { type ReplyFeedbackBody, type ReplyFeedbackResponse } from "@roll-agent/reply-authority-client";
+import { type AgentLogger } from "@roll-agent/sdk";
+export declare const REPLY_FEEDBACK_SUBMISSION_STATUS_VALUES: readonly ["accepted", "duplicate", "queued", "failed"];
+export type ReplyFeedbackSubmissionStatus = (typeof REPLY_FEEDBACK_SUBMISSION_STATUS_VALUES)[number];
+export interface ReplyFeedbackSubmissionResult {
+    readonly status: ReplyFeedbackSubmissionStatus;
+    readonly error?: string;
+}
+export interface SubmitReplyFeedbackOptions {
+    readonly feedbackExpiresAt?: number;
+}
+export type ReplyFeedbackDeliver = (body: ReplyFeedbackBody) => Promise<ReplyFeedbackResponse>;
+export interface InitializeReplyFeedbackOutboxOptions {
+    readonly dbPath?: string;
+    readonly retentionSeconds?: number;
+    readonly flushIntervalMs?: number;
+    readonly retryBaseDelayMs?: number;
+    readonly authRetryDelayMs?: number;
+    readonly deliver?: ReplyFeedbackDeliver;
+    readonly logger?: AgentLogger;
+}
+export declare function initializeReplyFeedbackOutbox(options?: InitializeReplyFeedbackOutboxOptions): void;
+export declare function shutdownReplyFeedbackOutbox(): Promise<void>;
+export declare function submitReplyFeedback(body: ReplyFeedbackBody, deliver: ReplyFeedbackDeliver, logger: AgentLogger, options?: SubmitReplyFeedbackOptions): Promise<ReplyFeedbackSubmissionResult>;

package/dist/tools/zhipin-judge-prepared-reply.d.ts

@@ -1,8 +1,12 @@
+import { type AgentContext } from "@roll-agent/sdk";
 import { fetchReplyFeedbackRubric } from "@roll-agent/reply-authority-client";
+import { type PreparedReplyJudgement } from "../reply-authority/prepared-reply-decision.ts";
+import { type PreparedReplyRecord } from "../reply-authority/prepared-reply-store.ts";
 type ZhipinJudgePreparedReplyDeps = {
     readonly fetchReplyFeedbackRubric: typeof fetchReplyFeedbackRubric;
 };
 export declare function setZhipinJudgePreparedReplyDepsForTests(override: Partial<ZhipinJudgePreparedReplyDeps> | undefined): void;
+export declare function ensurePreparedReplyJudgement(record: PreparedReplyRecord, ctx: AgentContext, judgeModel?: string): Promise<PreparedReplyJudgement>;
 export declare const zhipinJudgePreparedReply: import("@roll-agent/sdk").ToolDefinition<{
     preparedReplyId: string;
     judgeModel?: string | undefined;
@@ -10,6 +14,7 @@ export declare const zhipinJudgePreparedReply: import("@roll-agent/sdk").ToolDef
     success: boolean;
     error?: string | undefined;
     fallback?: boolean | undefined;
+    decisionSource?: "judge" | "orchestrator" | "service_recommended_fallback" | "explicit_no_judge" | undefined;
     recommendedOption?: "option_1" | "option_2" | undefined;
     variantDecision?: {
         reason: string;

package/dist/tools/zhipin-send-prepared-reply.d.ts

@@ -1,8 +1,10 @@
 import { postReplyFeedback } from "@roll-agent/reply-authority-client";
+import { submitReplyFeedback } from "../reply-authority/reply-feedback-outbox.ts";
 import { sendSignedZhipinReply } from "./zhipin-send-reply.ts";
 type ZhipinSendPreparedReplyDeps = {
     readonly sendSignedZhipinReply: typeof sendSignedZhipinReply;
     readonly postReplyFeedback: typeof postReplyFeedback;
+    readonly submitReplyFeedback: typeof submitReplyFeedback;
 };
 export declare function setZhipinSendPreparedReplyDepsForTests(override: Partial<ZhipinSendPreparedReplyDeps> | undefined): void;
 export declare const zhipinSendPreparedReply: import("@roll-agent/sdk").ToolDefinition<{
@@ -19,11 +21,17 @@ export declare const zhipinSendPreparedReply: import("@roll-agent/sdk").ToolDefi
     toolActionApproval?: {
         id: string;
     } | undefined;
+    skipVariantJudge?: boolean | undefined;
 }, {
     success: boolean;
     sentMessage: string;
     error?: string | undefined;
-    feedbackStatus?: "failed" | "accepted" | "duplicate" | "skipped" | undefined;
+    decisionSource?: "judge" | "orchestrator" | "service_recommended_fallback" | "explicit_no_judge" | undefined;
+    judgeModel?: string | undefined;
+    chosenOption?: "option_1" | "option_2" | undefined;
+    decisionReason?: string | undefined;
+    feedbackExpected?: boolean | undefined;
+    feedbackStatus?: "failed" | "accepted" | "duplicate" | "queued" | "skipped" | undefined;
     feedbackError?: string | undefined;
 }>;
 export {};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@roll-agent/browser-use-agent",
-  "version": "0.22.2",
+  "version": "0.23.0",
   "type": "module",
   "repository": {
     "type": "git",
@@ -16,7 +16,7 @@
     "references"
   ],
   "engines": {
-    "node": ">=22.6.0"
+    "node": ">=22.13.0"
   },
   "exports": {
     ".": {
@@ -45,7 +45,7 @@
   },
   "dependencies": {
     "zod": "^3.25.76",
-    "@roll-agent/reply-authority-client": "0.5.0",
+    "@roll-agent/reply-authority-client": "0.6.0",
     "@roll-agent/sdk": "0.3.0",
     "@roll-agent/browser": "0.9.1"
   },

package/references/env.yaml

@@ -6,7 +6,7 @@ required:
     example: https://reply-authority.duliday.com
   - name: REPLY_AUTHORITY_BEARER_TOKEN
     secret: true
-    purpose: 调用 Reply Authority Service `POST /generate-signed-reply` 的 Bearer token；只用于生成 prepared reply，不会返回给 orchestrator
+    purpose: 调用 Reply Authority Service `POST /generate-signed-reply`、读取冻结 feedback rubric 和 `POST /reply-feedback` 的 Bearer token；必须能访问目标租户并具备 `reply-feedback:write`，不会返回给 orchestrator
     example: token_xxx
   - name: REPLY_AUTHORITY_KEYS_URL
     secret: false
@@ -22,6 +22,23 @@ 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: REPLY_FEEDBACK_OUTBOX_DB_PATH
+    secret: false
+    purpose: browser-use-agent 双稿 feedback SQLite outbox 路径；独立于浏览器 session，容器部署时应挂载到持久卷
+    default: ~/.roll-agent/browser/reply-feedback-outbox.sqlite
+    example: /data/roll-agent/reply-feedback-outbox.sqlite
+  - name: REPLY_FEEDBACK_OUTBOX_RETENTION_SECONDS
+    secret: false
+    type: number
+    purpose: feedback 临时失败后的本地最长重试时间；双稿闭环要求 Reply Authority RFC V3，实际截止时间取服务端 feedbackExpiresAt 与本值的较早者
+    default: "604800"
+    example: "604800"
+  - name: REPLY_FEEDBACK_OUTBOX_FLUSH_INTERVAL_MS
+    secret: false
+    type: number
+    purpose: feedback outbox 后台扫描待重试记录的间隔毫秒数
+    default: "30000"
+    example: "30000"
   - name: RECRUITMENT_EVENTS_ENABLED
     secret: false
     type: boolean

package/references/zhipin-workflows.md

@@ -185,7 +185,7 @@ recommend-list  -> 推荐牛人列表，默认向下滚动，去重主键 candid
 
 ## Reply Authority
 
-`zhipin_send_prepared_reply` 只接受 `zhipin_generate_reply_preview` 生成的 `preparedReplyId`，内部再取回 Reply Authority Service 签发的 envelope 并完成验签。
+`zhipin_send_prepared_reply` 只接受 `zhipin_generate_reply_preview` 生成的 `preparedReplyId`，从 prepared store 读取预览阶段保存的 Reply Authority 签名 envelope 并完成验签。
 
 发送前校验：
 
@@ -218,7 +218,46 @@ recommend-list  -> 推荐牛人列表，默认向下滚动，去重主键 candid
 }
 ```
 
-第二种代理模式由 `smart-reply-agent` 调用 Reply Authority resolver 解析 `tenantId + recruiterBinding`。
+第二种代理模式由当前 Agent 使用的共享 `reply-authority-client` 先调用 Reply Authority resolver，解析出 `tenantId + recruiterBinding`；`browser-use-agent` 与 `smart-reply-agent` 都能独立完成该解析，不需要为了 resolver 在两者之间跳转。
+
+### RFC V3 双稿反馈闭环
+
+当前 browser-use 双稿链路要求 Reply Authority RFC V3；部署顺序必须是 **Reply Authority V3 → 新版 browser-use-agent**。调用方 token 必须能访问目标租户并具备 `reply-feedback:write`。
+
+```text
+preview 保存 group / rubric / feedbackExpiresAt
+  -> send 内部 Judge，或复用独立 Judge / orchestrator 显式选择
+  -> 发送一份签名 option
+  -> 将 selected 或 not_learned 终态写入 SQLite outbox
+  -> POST /reply-feedback
+```
+
+| `feedbackOutcome` | `decisionSource` | Beta 学习 | Pending |
+| --- | --- | --- | --- |
+| `selected` | `judge` / `orchestrator` | 更新 | 闭合 |
+| `not_learned` | `service_recommended_fallback` / `explicit_no_judge` | 不更新 | 闭合 |
+
+`zhipin_judge_prepared_reply` 只是可选预览工具；默认直接调用 send 即可，send 会在双稿缺少 `variantDecision` 时确定性执行并缓存 Judge。若 preview 阶段 rubric 拉取、hash 或双稿结构异常，prepared store 会保留 group 生命周期元数据；send 发送顶层推荐稿成功后提交 `not_learned`，不会让服务端 Pending 静默过期。
+
+发送后的 feedback 状态：
+
+| `feedbackStatus` | 用户消息 | feedback 闭环 | 后续动作 |
+| --- | --- | --- | --- |
+| `accepted` / `duplicate` | 已发送 | 已闭合 | 无需处理 |
+| `queued` | 已发送 | 等待 outbox 重试 | 不重调 send；等待 outbox 仅重试 POST |
+| `failed` | 已发送 | 有缺口 | 不重调 send；修复 token、服务或 outbox 后单独排查 |
+
+`feedbackExpected:false` 只表示 fallback/no-Judge 不进入 Beta 学习，不表示无需终态。outbox 的实际重试截止时间取服务端 `feedbackExpiresAt` 与本地 retention 的较早值。
+
+跨服务 fallback 只使用稳定安全码，原始 provider / HTTP / parser 错误只留在 browser-use-agent 本地日志：
+
+| Code | 含义 |
+| --- | --- |
+| `rubric_fetch_failed` | rubric 获取失败 |
+| `rubric_mismatch` | rubric version/hash 不一致 |
+| `invalid_variant_shape` | 服务端双稿结构无法安全判定 |
+| `judge_sampling_failed` | MCP Sampling 调用失败 |
+| `judge_output_invalid` | Judge 输出无法验证 |
 
 ## 推荐候选人链路