@roll-agent/browser-use-agent 0.17.0 → 0.19.0

package/dist/pages/zhipin/job-signals.d.ts

@@ -1,6 +1,5 @@
-import type { AgentLLM, AgentLogger } from "@roll-agent/sdk";
-import { type CandidateLocationSignal } from "@roll-agent/reply-authority-client";
-export declare const LocationSignalAnalysisPathValues: readonly ["llm", "fallback", "profile_only"];
+import type { CandidateLocationSignal } from "@roll-agent/reply-authority-client";
+export declare const LocationSignalAnalysisPathValues: readonly ["llm", "fallback", "profile_only", "speculative", "none"];
 export type LocationSignalAnalysisPath = (typeof LocationSignalAnalysisPathValues)[number];
 export declare const LocationSignalInquiryTypeValues: readonly ["location_inquiry", "non_location_inquiry"];
 export type LocationSignalInquiryType = (typeof LocationSignalInquiryTypeValues)[number];
@@ -9,37 +8,13 @@ export type LocationSignalResolution = {
     readonly analysisPath: LocationSignalAnalysisPath;
     readonly inquiryType?: LocationSignalInquiryType;
 };
-export type LocationSignalMessage = {
-    readonly index: number;
-    readonly sender: "candidate" | "recruiter" | "system";
-    readonly content: string;
-};
-type LocationSignalEvidenceInput = {
-    readonly latestCandidateMessage: string;
-    readonly recentMessages: readonly LocationSignalMessage[];
-    readonly expectedLocation: string;
-    readonly communicationPosition: string;
-};
-type ResolveLocationSignalsInput = {
-    readonly llm: AgentLLM;
-    readonly logger?: Pick<AgentLogger, "warn"> | undefined;
-    readonly messages: readonly LocationSignalMessage[];
-    readonly expectedLocation: string;
-    readonly communicationPosition: string;
-    readonly timeoutMs?: number | undefined;
-};
-export declare function resolveLocationSignalsFromLlmText(input: LocationSignalEvidenceInput & {
-    readonly llmText: string;
-}): CandidateLocationSignal[];
-export declare function resolveLocationSignals(input: ResolveLocationSignalsInput): Promise<LocationSignalResolution>;
-export declare function clearLocationSignalResolutionCacheForTest(): void;
-export declare function resolveLocationSignalsWithLlm(input: ResolveLocationSignalsInput): Promise<CandidateLocationSignal[]>;
+export declare function formatLocationSignalsVisualLabel(resolution: LocationSignalResolution | readonly CandidateLocationSignal[]): string;
+export declare function resolvePreferredBrandId(communicationPosition: string): number | undefined;
 export declare function resolvePreferredBrand(communicationPosition: string): string | undefined;
 export declare function resolveExpectedSignals(expectedJobText: string): {
     expectedLocation: string;
     expectedPosition: string;
 };
-export declare function formatLocationSignalsVisualLabel(resolution: LocationSignalResolution | readonly CandidateLocationSignal[]): string;
 export declare function resolveConversationSignals(input: {
     communicationPosition: string;
     expectedJobText: string;
@@ -48,5 +23,5 @@ export declare function resolveConversationSignals(input: {
     expectedLocation: string;
     expectedPosition: string;
     preferredBrand?: string;
+    preferredBrandId?: number;
 };
-export {};

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

@@ -29,9 +29,17 @@ export declare const zhipinGenerateReplyPreview: import("@roll-agent/sdk").ToolD
     error?: string | undefined;
     confidence?: number | undefined;
     requestId?: string | undefined;
+    expiresAt?: number | undefined;
     suggestedReply?: string | undefined;
     stage?: string | undefined;
     preparedReplyId?: string | undefined;
-    expiresAt?: number | undefined;
+    timing?: {
+        totalLatencyMs: number;
+        preparedContextHit: boolean;
+        replyLatencyMs?: number | undefined;
+        turnPlanningLatencyMs?: number | undefined;
+        contextBuildingLatencyMs?: number | undefined;
+    } | undefined;
+    gateRewritten?: boolean | undefined;
 }>;
 export {};

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

@@ -1,3 +1,4 @@
+import { prepareReplyContext } from "@roll-agent/reply-authority-client";
 import { NativeVisualActivitySession } from "../native-visual-activity-session.ts";
 import { openZhipinNativePagePort } from "../pages/zhipin/native-page.ts";
 import type { ZhipinNativePagePort } from "../pages/zhipin/native-page.ts";
@@ -5,7 +6,9 @@ type NativeVisualActivitySessionLike = Pick<NativeVisualActivitySession, "begin"
 type ZhipinGetCandidateInfoDeps = {
     readonly openNativePagePort: typeof openZhipinNativePagePort;
     readonly createNativeVisualActivitySession: (page: ZhipinNativePagePort) => NativeVisualActivitySessionLike;
+    readonly prepareReplyContext: typeof prepareReplyContext;
 };
+export declare function resetPrepareReplyContextCooldownForTests(): void;
 export declare function setZhipinGetCandidateInfoDepsForTests(override: Partial<ZhipinGetCandidateInfoDeps> | undefined): void;
 export declare const zhipinGetCandidateInfo: import("@roll-agent/sdk").ToolDefinition<{
     index?: number | undefined;
@@ -50,5 +53,6 @@ export declare const zhipinGetCandidateInfo: import("@roll-agent/sdk").ToolDefin
     formattedHistory: string[];
     error?: string | undefined;
     preferredBrand?: string | undefined;
+    preferredBrandId?: number | undefined;
 }>;
 export {};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@roll-agent/browser-use-agent",
-  "version": "0.17.0",
+  "version": "0.19.0",
   "type": "module",
   "repository": {
     "type": "git",
@@ -45,8 +45,8 @@
   },
   "dependencies": {
     "zod": "^3.25.76",
-    "@roll-agent/reply-authority-client": "0.2.0",
     "@roll-agent/sdk": "0.2.0",
+    "@roll-agent/reply-authority-client": "0.4.0",
     "@roll-agent/browser": "0.9.0"
   },
   "devDependencies": {

package/references/zhipin-workflows.md

@@ -125,38 +125,45 @@ recommend-list  -> 推荐牛人列表，默认向下滚动，去重主键 candid
    - `no-new-items`：连续滚动没有新去重项。
    - `max-steps`：达到滚动步数上限。
 
-## preferredBrand
+## preferredBrand / preferredBrandId
 
-`zhipin_get_candidate_info` 会从 `candidateInfo.communicationPosition` 中解析：
+`zhipin_get_candidate_info` 会从页面信号中解析：
 
-- `expectedLocation`
-- `expectedPosition`
-- `preferredBrand`
+- `expectedLocation` / `expectedPosition`（来自期望职位文本）
+- `preferredBrandId` 或 `preferredBrand`（来自 `candidateInfo.communicationPosition`，互斥，最多输出一个）
 
-`preferredBrand` 规则：
+沟通职位有两种命名格式，解析规则：
 
-- 仅当 `communicationPosition` 含连字符类分隔符时输出：`-` / `－` / `—` / `–`。
-- 取分隔符前的第一段，例如 `肯德基-服务员` -> `preferredBrand: "肯德基"`。
-- 没有分隔符时不输出。
-- 禁止用通用岗位名或 `zhipin_get_candidate_list.company` 伪造品牌。
+- 新格式 `自定义描述…[品牌ID]`（末尾方括号包数字，兼容全角 `【】` / `［］` 与括号内空白）：输出 `preferredBrandId`，例如 `咖啡早班店员-接受小白-免费咖啡[10027]` -> `preferredBrandId: 10027`。此时**不输出** `preferredBrand`——新格式第一段是岗位描述，不是品牌名。
+- 老格式 `品牌名-职位`（含连字符类分隔符 `-` / `－` / `—` / `–`，且无 ID 尾缀）：取第一段输出 `preferredBrand`，例如 `肯德基-服务员` -> `preferredBrand: "肯德基"`。
+- 两者都不满足时不输出任何品牌字段。
+- 编排器原样透传工具输出即可，不要自行解析 `communicationPosition`；Reply Authority 服务端会以 `candidateInfo.communicationPosition` 为权威再次提取品牌 ID 并与 `preferredBrandId` 对账（不一致以服务端提取为准）。
+- 禁止用通用岗位名或 `zhipin_get_candidate_list.company` 伪造品牌；禁止把新格式的第一段当品牌名。
 
-## locationSignals
+## 地点证据（服务端提取）
 
-`zhipin_get_candidate_info` 和 `zhipin_generate_reply_preview` 会在读取聊天记录后输出/透传结构化地点证据 `locationSignals`，供下游 Reply Authority 做 geocode 与门店/岗位距离匹配。
+地点证据提取已收编进 Reply Authority 服务端的 turn_planning 阶段（RFC pipeline-latency-restructure M1）：browser-use 只上送原始对话与候选人资料，服务端 `planTurn` 合并提取地点信号并做逐字证据校验，再交给 geocode 与门店/岗位距离匹配。
 
-字段形状（与 `@roll-agent/reply-authority-client` 契约一致）：
+`zhipin_generate_reply_preview` 通过 SSE `location.resolved` 事件回显服务端提取结果：
 
 ```json
 {
-  "text": "人民广场",
-  "source": "candidate_message",
-  "city": "上海",
-  "intent": "nearby_store",
-  "confidence": 0.93
+  "type": "location.resolved",
+  "inquiryType": "location_inquiry",
+  "analysisPath": "llm",
+  "signals": [
+    {
+      "text": "人民广场",
+      "source": "candidate_message",
+      "city": "上海",
+      "intent": "nearby_store",
+      "confidence": 0.93
+    }
+  ]
 }
 ```
 
-`source` 取值：
+`source` 取值（与 `@roll-agent/reply-authority-client` 契约一致）：
 
 | `source` | 含义 |
 | --- | --- |
@@ -165,23 +172,16 @@ recommend-list  -> 推荐牛人列表，默认向下滚动，去重主键 candid
 | `candidate_expected_location` | 候选人资料里的期望城市/区域（弱信号） |
 | `communication_position` | 沟通岗位文本中可确认的城市/门店片段 |
 
-`intent` 常见取值：`nearby_store`（附近门店/岗位）、`store_address`（门店地址）、`expected_area`（期望区域）。
+`intent` 常见取值：`nearby_store`（附近门店/岗位）、`store_address`（门店地址）、`expected_area`（期望区域）。`analysisPath` 取值：`llm`（planTurn 合并提取）、`speculative`（词典投机命中）、`none`（无地点信号）。
 
-抽取规则：
+规则与边界：
 
-1. browser-use 通过指挥官 LLM 抽取，并在超时或失败时回退到规则匹配；`text` 必须来自输入原文，不改写成下游门店名。
-2. 多轮对话会扫描全部候选人消息（去重），避免「历史问过地点、最新一句非地点」时丢失证据。
-3. `candidateInfo.expectedLocation` 只作城市边界弱信号（confidence 通常 ≤ 0.6），不能替代候选人本轮追问的具体 POI。
-4. 问候/职位介绍类消息可能跳过 LLM 抽取，仅保留 `expectedLocation` 弱信号。
-5. orchestrator 不要自行构造或改写 `locationSignals`；需要地点证据时调用 `zhipin_get_candidate_info` 或 `zhipin_generate_reply_preview`，由工具自动填充。
-
-边界：
-
-- browser-use **不**调用外部地图、**不**做 geocode、**不**计算门店距离。
-- 岗位/门店事实源仍是下游同步的岗位配置；外部地图 POI 不能当作门店或岗位事实。
-- `smart-reply-agent.generate_reply` 只校验 schema 并原样转发 `locationSignals`，不做地点推断。
-
-`zhipin_generate_reply_preview` 会把非空 `locationSignals` 写入 Reply Authority stream input；`zhipin_get_candidate_info` 在 output 中返回同结构数组，便于编排器先读证据再决定是否生成回复。
+1. browser-use 不抽取地点、不调用外部地图、不做 geocode、不计算门店距离；上送的对话原文是服务端提取的唯一证据源。
+2. `candidateInfo.expectedLocation` 只作城市边界弱信号（服务端 confidence 封顶 0.6），不能替代候选人本轮追问的具体 POI。
+3. orchestrator 不要自行构造或改写地点信号：`generate_reply` 请求中的 `locationSignals` 字段已废弃，服务端过渡期内会逐字校验后合并，两个版本周期后忽略。
+4. `zhipin_get_candidate_info` 输出中的 `locationSignals` 恒为空数组，仅为输出契约兼容保留，不要依赖。
+5. 岗位/门店事实源仍是下游同步的岗位配置；外部地图 POI 不能当作门店或岗位事实。
+6. `smart-reply-agent.generate_reply` 只校验 schema 并原样转发（deprecated 字段同样适用），不做地点推断。
 
 ## Reply Authority