@reconcrap/boss-recommend-mcp 2.1.14 → 2.1.15

package/README.md

@@ -176,7 +176,7 @@ boss-recommend-mcp schedule-status --schedule-id <schedule_id>
 - 不会对每位候选人重复确认
 - 推荐页详情处理完成后，会强制关闭详情页并确认已关闭
 - 简历提取优先使用 Network 响应；没有可解析 Network CV 时，回退到完整滚动截图序列再交给多模态模型判断
-- recommend / search / chat 正式运行默认全部使用 `screening-config.json` 配置的 LLM 筛选；deterministic/local scorer 只保留给明确测试场景，必须显式传 `debug_test_mode=true` 且 `screening_mode=deterministic` 或 `use_llm=false`。
+- recommend / search / 带 `criteria` 的 chat 正式运行默认全部使用 `screening-config.json` 配置的 LLM 筛选；chat 的 `criteria` 留空时进入收集简历模式，不需要 LLM 配置。deterministic/local scorer 只保留给明确测试场景，必须显式传 `debug_test_mode=true` 且 `screening_mode=deterministic` 或 `use_llm=false`。
 - `detail_limit=0`、`no_filter`、`filter_enabled=false`、后置动作 dry-run、chat 求简历 dry-run 等调试路径不会在正式 live run 默认启用；需要测试时必须显式传 `debug_test_mode=true`。
 - 提供显式运维自愈工具：只在手动调用 `run_recommend_self_heal` 时运行，不会接入正常 run / doctor / preflight 自动链路
 - 运行前会自动做依赖体检（Node.js、Python、Pillow、`chrome-remote-interface`、`ws`），缺失时会在 `doctor` 与流水线失败诊断中明确提示
@@ -433,7 +433,7 @@ node src/cli.js chat prepare-run --slow-live --port 9222
 
 说明：
 
-- `criteria` / `start_from` / `target_count` 为必填
+- `start_from` / `target_count` 为必填；`criteria` 可选，留空时 chat run 进入收集简历模式：不触发 LLM 筛选，针对没有在线/附件简历的人选发送求简历消息并点击求简历按钮
 - `greeting_text` 可选（兼容 `greetingText`）
 - `profile` 可选，默认 `default`
 - `job` 与 `port` 继承 recommend run 已选岗位和调试端口
@@ -468,15 +468,15 @@ node src/cli.js chat prepare-run --slow-live --port 9222
 chat-only 交互建议：
 
 - 先调用一次 `list_boss_chat_jobs` 或 `prepare_boss_chat_run`（可不带参数），服务会先导航到 `https://www.zhipin.com/web/chat/index` 并返回 `NEED_INPUT`，其中包含岗位 `job_options` 与待补字段。
-- 然后基于 `job_options` 让用户选择 `job`，并补齐 `start_from`、`target_count`、`criteria` 后调用 `start_boss_chat_run` 启动任务。
+- 然后基于 `job_options` 让用户选择 `job`，并补齐 `start_from`、`target_count` 后调用 `start_boss_chat_run` 启动任务；若要筛选再求简历，提供 `criteria`，若只想收集缺失简历则留空。
 - `greeting_text` 可选；未传时使用 `screening-config.json.greetingMessage`，若未配置则使用默认招呼语（`Hi同学，能麻烦发下简历吗？`）。
 - `target_count` 支持正整数、`all`、`-1`；若用户给出 `全部候选人` / `所有候选人`，会自动按不限（扫到底）处理。
 
 Trae-CN / 长对话防循环建议：
 
-- 固定流程：`boss_chat_health_check` -> `list_boss_chat_jobs(空参可)` / `prepare_boss_chat_run(空参可)` -> 一次性补齐 `job/start_from/target_count/criteria` -> `start_boss_chat_run`。
+- 固定流程：`boss_chat_health_check` -> `list_boss_chat_jobs(空参可)` / `prepare_boss_chat_run(空参可)` -> 一次性补齐 `job/start_from/target_count`，筛选任务再带 `criteria` -> `start_boss_chat_run`。
 - chat-only 场景严禁调用 `list_recommend_jobs`、`run_recommend` 或 `start_recommend_pipeline_run`。
-- `start_boss_chat_run` 的工具 schema 已把 `job/start_from/target_count/criteria` 标记为必填；不要用它获取岗位列表。
+- `start_boss_chat_run` 的工具 schema 已把 `job/start_from/target_count` 标记为必填；`criteria` 留空会进入收集简历模式。不要用它获取岗位列表。
 - 若 `pending_questions` / UI 选项里出现“扫到底（必须传 `target_count="all"`）”，下一次工具调用请直接照抄 `"target_count": "all"`，不要只保留“扫到底”这层自然语言语义。
 - `start_boss_chat_run` 返回 `ACCEPTED` 后直接结束当前回合，不要自动轮询。
 - 缺参或校验失败时，一次性列出全部缺失/错误项，避免重复同一句提示触发宿主“陷入循环”保护。

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@reconcrap/boss-recommend-mcp",
-  "version": "2.1.14",
+  "version": "2.1.15",
   "description": "Unified MCP pipeline for recommend-page filtering and screening on Boss Zhipin",
   "keywords": [
     "boss",

package/skills/boss-chat/README.md

@@ -17,7 +17,7 @@ Please run a Boss chat-only task (do not switch to recommend flow).
 Execution order:
 1) Call boss_chat_health_check.
 2) Call list_boss_chat_jobs once (empty params allowed), or prepare_boss_chat_run once, to fetch job_options and missing fields.
-3) Ask for these required fields in one shot: job, start_from (unread/all), target_count, criteria, rest_level (low/medium/high).
+3) Ask for these required fields in one shot: job, start_from (unread/all), target_count, rest_level (low/medium/high). Ask for criteria only when the user wants screening; leave it blank for collect-CV runs.
 4) After user reply, call start_boss_chat_run exactly once to start the run.
 5) If ACCEPTED, reply only with run_id and "task started"; no auto polling.
 
@@ -26,7 +26,7 @@ Anti-loop rules:
 - On validation errors, list all missing/invalid fields once.
 - Do not call list_recommend_jobs for chat-only tasks; it is recommend-page only and will switch the browser to /web/chat/recommend.
 - Do not call run_recommend or start_recommend_pipeline_run for chat-only tasks; use start_boss_chat_run.
-- Do not use start_boss_chat_run for preflight. It is only for the final start call and must include job/start_from/target_count/criteria.
+- Do not use start_boss_chat_run for preflight. It is only for the final start call and must include job/start_from/target_count. Include criteria for screening; leave criteria blank to collect CVs from candidates without an online/attachment CV.
 - Do not call start_boss_chat_run repeatedly in one turn.
 - Do not call get_boss_chat_run unless user explicitly asks for progress.
 - Do not choose rest_level for the user. Pass the user's choice as `human_behavior.restLevel`.

package/skills/boss-chat/SKILL.md

@@ -29,11 +29,11 @@ Trae/Trae-CN split-server config exposes these under the `boss-chat` MCP server.
 - `job`
 - `start_from`: `unread|all`
 - `target_count`
-- `criteria`
 - `rest_level`: `low|medium|high`
 
 可选：
 
+- `criteria`（留空时进入收集简历模式：不触发 LLM 筛选，只向没有在线/附件简历的人选求简历）
 - `profile`（默认 `default`）
 - `greeting_text`（兼容 `greetingText`，可选自定义首条打招呼消息）
 - `port`
@@ -62,7 +62,7 @@ Trae/Trae-CN split-server config exposes these under the `boss-chat` MCP server.
 
 ## Hard Rules
 
-- LLM 配置必须复用 `boss-recommend-mcp` 的 `screening-config.json`；不要再向用户单独要 `baseUrl/apiKey/model`。
+- 有筛选 `criteria` 时，LLM 配置必须复用 `boss-recommend-mcp` 的 `screening-config.json`；不要再向用户单独要 `baseUrl/apiKey/model`。`criteria` 留空的收集简历模式不需要 LLM 配置。
 - 路由护栏（强制）：
   - 只在用户明确是 chat-only 任务时使用本 skill。
   - 只要用户提到推荐页、先找人后沟通、或需要推荐筛选阶段，禁止直接调用 `start_boss_chat_run`；必须先交给 `boss-recommend-pipeline` 完成推荐页任务。
@@ -70,7 +70,7 @@ Trae/Trae-CN split-server config exposes these under the `boss-chat` MCP server.
 - 启动或准备 chat run 时，若本机默认 `127.0.0.1:9222` Chrome DevTools 端口不可连，工具会自动打开 Chrome 并导航到 `https://www.zhipin.com/web/chat/index`。
 - 只有工具返回 `BOSS_LOGIN_REQUIRED` / `requires_login=true` 时，才要求用户在自动打开的 Chrome 窗口人工登录 Boss 后重试；不要把“没开 9222 Chrome”当作缺参。
 - 若本机找不到 Chrome，可提示用户设置 `BOSS_MCP_CHROME_PATH` 或 `BOSS_RECOMMEND_CHROME_PATH`；非本机 debug host 不自动启动。
-- `job` / `start_from` / `criteria` 缺一不可；缺参时只补缺口。
+- `job` / `start_from` 缺一不可；`criteria` 可留空，留空时表示收集缺失简历而不是筛选。缺参时只补缺口。
 - `target_count` 在 chat-only 启动前也是必填项，不能默认省略。
 - chat-only 岗位列表只能通过 `list_boss_chat_jobs` 或 `prepare_boss_chat_run` 获取；严禁调用 `list_recommend_jobs`，因为它会切到推荐页。
 - chat-only 启动只能调用 `start_boss_chat_run`；严禁调用 `run_recommend` 或 `start_recommend_pipeline_run`。
@@ -80,10 +80,10 @@ Trae/Trae-CN split-server config exposes these under the `boss-chat` MCP server.
 - 当用户选择“扫到底/全部候选人/所有候选人”时，调用参数优先写：`"target_count": "all"`；`-1` 只作为兼容输入和内部 CLI 表示。
 - `greeting_text` 是可选项，不能因为缺少它阻塞启动或追加必填追问。
 - 若工具或提问选项里出现“扫到底（必须传 `target_count=\"all\"`）”之类字样，下一次工具调用时必须直接照抄这个字面量，不要只保留“扫到底”语义。
-- 禁止 agent 自行补全 `job/start_from/criteria` 并直接执行，必须由用户明确给出或确认。
+- 禁止 agent 自行补全 `job/start_from` 并直接执行，必须由用户明确给出或确认。`criteria` 只有在用户要筛选时才必须明确给出；用户要收集简历时可留空。
 - chat-only 启动流程必须先进入聊天页并拉取岗位列表，再让用户从列表中选择 `job`。
 - 必须先用空参调用 `list_boss_chat_jobs` 或 `prepare_boss_chat_run` 获取 `job_options`；不要用 `start_boss_chat_run` 做预备调用。
-- `start_boss_chat_run` 只能用于真正启动，必须一次性传齐 `job` / `start_from` / `target_count` / `criteria`。
+- `start_boss_chat_run` 只能用于真正启动，必须一次性传齐 `job` / `start_from` / `target_count`；`criteria` 留空会进入收集简历模式。
 - 若 `start_boss_chat_run` 返回 `NEED_INPUT` 且 `missing_fields` 包含 `target_count`，说明你没有把用户选择写入工具参数；下一次调用必须照 `next_call_example` 原样补上 `"target_count": "all"` 或正整数，不要重复空调用。
 - 默认不自动轮询；只有用户要求查进度时才调用 `get_boss_chat_run`。
 - `start_boss_chat_run` 返回 `ACCEPTED` 后，默认立即结束当前回合，不得主动连续调用 `get_boss_chat_run`。
@@ -95,12 +95,12 @@ Trae/Trae-CN split-server config exposes these under the `boss-chat` MCP server.
 
 - 若用户先运行了 recommend 流水线，并在手动状态检查时确认 recommend 已完成，且用户目标是“立即进入聊天沟通”：
   - 先调用 `prepare_boss_chat_run` 获取聊天页岗位列表与缺参。
-  - 显式确认 `job/start_from/target_count/criteria` 后再调用 `start_boss_chat_run`。
+  - 显式确认 `job/start_from/target_count` 后再调用 `start_boss_chat_run`；若要继续筛选再求简历，还要确认 `criteria`。
   - 不要查找或依赖 `follow_up.chat`；该自动衔接路径属于 legacy-only 行为。
 
 ## Response Style
 
 - 用结构化中文。
 - 首轮建议先调用一次 `list_boss_chat_jobs`（可空参）或 `prepare_boss_chat_run`（可空参）获取 `job_options` 与 `pending_questions`。
-- 缺参时必须逐项确认：`job`（来自岗位列表）、`start_from`（`unread|all`）、`target_count`、`criteria`、`rest_level`。
+- 缺参时必须逐项确认：`job`（来自岗位列表）、`start_from`（`unread|all`）、`target_count`、`rest_level`；只有用户要筛选时才追问 `criteria`。
 - 若健康检查失败，明确提示共享配置文件 `screening-config.json` 不可用。

package/skills/boss-recruit-pipeline/SKILL.md

@@ -20,6 +20,8 @@ Trae/Trae-CN split-server config exposes these under the `boss-recruit` MCP serv
 - 继续：`resume_recruit_pipeline_run`
 - 取消：`cancel_recruit_pipeline_run`
 
+If the visible tool surface only offers `boss-recommend/*` for a search/recruit task, stop and report a tool-surface/config error. Do not call `boss-recommend/list_recommend_jobs`, `boss-recommend/run_recommend`, or `boss-recommend/start_recommend_pipeline_run` as a fallback for search.
+
 ## Hard Rules
 
 - 只在用户明确说搜索页、search、recruit、招聘搜索、`/web/chat/search` 时使用本 skill。
@@ -31,9 +33,14 @@ Trae/Trae-CN split-server config exposes these under the `boss-recruit` MCP serv
 - 只有工具返回 `BOSS_LOGIN_REQUIRED` / `requires_login=true` 时，才要求用户在自动打开的 Chrome 窗口人工登录 Boss 后重试；不要把“没开 9222 Chrome”当作缺参。
 - 若本机找不到 Chrome，可提示用户设置 `BOSS_MCP_CHROME_PATH` 或 `BOSS_RECOMMEND_CHROME_PATH`；非本机 debug host 不自动启动。
 - 若用户未提供岗位，必须先询问岗位。搜索页岗位选择在关键词输入框旁边；不要猜测默认岗位。
+- 搜索页任务不要调用 `list_recommend_jobs` 获取岗位；推荐页岗位列表和搜索页岗位选择不是同一个工具面。用户已经给出岗位时直接传 `overrides.job`。
 - 若用户提供城市、学历、学校、关键词、过滤已看、人选目标数、筛选条件等参数，必须逐项传入或确认。
 - 搜索页和推荐页一样支持多选筛选条件；不要把多选降级成单选。
 - 每次 run 必须明确询问用户本次休息强度 `rest_level`：`low`（旧策略）/ `medium`（约 5 小时或 700 人累计休息 30 分钟）/ `high`（约 5 小时或 700 人累计休息 1 小时）；不得默认使用配置文件里的值替用户决定。
+- 启动前展示一次包含岗位、关键词、城市、学历、学校标签、是否过滤已看、criteria、目标人数、后置动作和休息强度的总确认；用户确认后，`confirmation` 只需要 `{ "final_confirmed": true }`。
+- 不要让工具重写用户的 `criteria`。用户给出 `筛选条件` / `筛选标准` / `硬条件` 时，逐字写入 `overrides.criteria`；不要传系统简化版。
+- 用户说学校类型“不限”时，在 `overrides.school_tag` 显式传 `"不限"` 或在 `overrides.schools` 传 `[]`；不要因为 criteria 里出现 `985/211/双一流` 就把它们当作搜索页学校过滤器。
+- 用户说只看未查看“不限”时，在 `overrides.recent_not_view` 显式传 `"不限"` 或在 `overrides.filter_recent_viewed` 传 `false`。
 
 ## Required Inputs
 
@@ -51,10 +58,25 @@ Trae/Trae-CN split-server config exposes these under the `boss-recruit` MCP serv
 - `recent_not_view`
 - `port`
 
+## Confirmation Flow
+
+1. 先从用户原始消息里提取参数；缺什么只问什么。不要先调用 MCP 让工具猜缺参。
+2. `criteria` 必须来自用户明确提供的筛选条件，或来自你向用户追问后得到的完整自然语言标准。禁止把岗位、关键词、学历等字段拼成 criteria。
+3. 如果用户一开始已经给齐 `job`、`keyword`、`criteria`、`target_count`、`rest_level`，并且可选筛选项也已明确或确认不限，直接展示一次总确认。
+4. 用户确认后调用 `boss-recruit/start_recruit_pipeline_run` 或 `boss-recruit/run_recruit_pipeline`，传：
+   - `confirmation: { "final_confirmed": true }`
+   - 所有规范化字段放在 `overrides`
+   - 完整 criteria 放在 `overrides.criteria`
+   - 本次休息强度放在 `human_behavior.restLevel`
+5. 工具返回 `NEED_INPUT` 时，按 `pending_questions` 只补具体缺口；不要接受或转述工具生成的简化 criteria。
+
 启动工具时，把用户确认的休息强度写入 `human_behavior.restLevel`，例如：
 
 ```json
-{ "human_behavior": { "restLevel": "medium" } }
+{
+  "confirmation": { "final_confirmed": true },
+  "human_behavior": { "restLevel": "medium" }
+}
 ```
 
 ## Response Style

package/src/chat-mcp.js

@@ -66,12 +66,11 @@ const RUN_MODE_ASYNC = "async";
 const DETACHED_WORKER_SCRIPT = fileURLToPath(new URL("./detached-worker.js", import.meta.url));
 const DETACHED_WORKER_POLL_MS = 1000;
 
-const CHAT_REQUIRED_FIELDS = Object.freeze([
-  "job",
-  "start_from",
-  "target_count",
-  "criteria"
-]);
+const CHAT_REQUIRED_FIELDS = Object.freeze([
+  "job",
+  "start_from",
+  "target_count"
+]);
 
 const TERMINAL_STATUSES = new Set([
   RUN_STATUS_COMPLETED,
@@ -1097,14 +1096,13 @@ function buildChatNextCallExample(args, missingFields, normalized) {
   return Object.keys(example).length ? example : null;
 }
 
-function getMissingChatStartFields(args = {}, normalized = normalizeChatStartInput(args)) {
-  const missing = [];
-  if (!normalized.job) missing.push("job");
-  if (!["unread", "all"].includes(normalized.startFrom)) missing.push("start_from");
-  if (!normalized.target.provided || normalized.target.parseError) missing.push("target_count");
-  if (!normalized.criteria) missing.push("criteria");
-  return missing;
-}
+function getMissingChatStartFields(args = {}, normalized = normalizeChatStartInput(args)) {
+  const missing = [];
+  if (!normalized.job) missing.push("job");
+  if (!["unread", "all"].includes(normalized.startFrom)) missing.push("start_from");
+  if (!normalized.target.provided || normalized.target.parseError) missing.push("target_count");
+  return missing;
+}
 
 function buildTargetCountDiagnostics(args, missingFields, normalized) {
   if (!missingFields.includes("target_count")) return {};
@@ -1168,17 +1166,10 @@ function buildPendingChatQuestions({ args, missingFields, normalized, jobOptions
         parse_error: normalized.target.parseError || null
       };
     }
-    if (field === "criteria") {
-      return {
-        field,
-        question: "请提供自然语言筛选 criteria。",
-        value: normalized.criteria || null
-      };
-    }
-    return {
-      field,
-      question: `请提供 ${field}。`,
-      value: null
+    return {
+      field,
+      question: `请提供 ${field}。`,
+      value: null
     };
   });
 }
@@ -1186,19 +1177,21 @@ function buildPendingChatQuestions({ args, missingFields, normalized, jobOptions
 async function buildNeedInputResponse({ args, missingFields, normalized }) {
   const diagnostics = buildTargetCountDiagnostics(args, missingFields, normalized);
   return {
-    status: "NEED_INPUT",
-    required_fields: CHAT_REQUIRED_FIELDS.slice(),
-    missing_fields: missingFields,
-    ...diagnostics,
+    status: "NEED_INPUT",
+    required_fields: CHAT_REQUIRED_FIELDS.slice(),
+    missing_fields: missingFields,
+    criteria_optional: true,
+    empty_criteria_mode: "collect_cv",
+    ...diagnostics,
     pending_questions: buildPendingChatQuestions({ args, missingFields, normalized }),
     job_options: [],
-    error: {
-      code: "MISSING_REQUIRED_FIELDS",
-      message: "缺少必要字段。请补齐 job、start_from、target_count、criteria 后再启动 Boss chat CDP-only run。",
-      retryable: true
-    }
-  };
-}
+    error: {
+      code: "MISSING_REQUIRED_FIELDS",
+      message: "缺少必要字段。请补齐 job、start_from、target_count 后再启动 Boss chat CDP-only run。criteria 可留空；留空时会进入收集简历模式。",
+      retryable: true
+    }
+  };
+}
 
 function shouldRequestChatResume(args = {}, context = {}) {
   const action = normalizeText(args.post_action || args.action).toLowerCase();
@@ -1232,38 +1225,40 @@ function isDebugTestMode(args = {}) {
   return args.debug_test_mode === true || args.allow_debug_test_mode === true;
 }
 
-function normalizeScreeningModeArg(args = {}) {
-  const raw = normalizeText(args.screening_mode || args.screeningMode || "");
-  if (args.use_llm === false) return "deterministic";
-  return ["deterministic", "local", "local_scorer"].includes(raw.toLowerCase())
-    ? "deterministic"
-    : "llm";
-}
-
-function collectChatDebugTestOptions(args = {}) {
-  const reasons = [];
-  if (normalizeScreeningModeArg(args) === "deterministic") reasons.push("deterministic_screening");
-  if (parseNonNegativeInteger(args.detail_limit, null) === 0) reasons.push("detail_limit=0");
-  if (args.dry_run === true || args.dry_run_request_cv === true) reasons.push("dry_run_request_cv");
-  return reasons;
-}
-
-function shouldUseChatLlm(args = {}) {
-  return normalizeScreeningModeArg(args) !== "deterministic";
-}
-
-function getRunOptions(args, normalized, session, { workspaceRoot = "", configResolution = null } = {}) {
+function normalizeScreeningModeArg(args = {}, normalized = normalizeChatStartInput(args)) {
+  if (!normalized.criteria) return "collect_cv";
+  const raw = normalizeText(args.screening_mode || args.screeningMode || "");
+  if (args.use_llm === false) return "deterministic";
+  return ["deterministic", "local", "local_scorer"].includes(raw.toLowerCase())
+    ? "deterministic"
+    : "llm";
+}
+
+function collectChatDebugTestOptions(args = {}) {
+  const reasons = [];
+  if (normalizeScreeningModeArg(args) === "deterministic") reasons.push("deterministic_screening");
+  if (parseNonNegativeInteger(args.detail_limit, null) === 0) reasons.push("detail_limit=0");
+  if (args.dry_run === true || args.dry_run_request_cv === true) reasons.push("dry_run_request_cv");
+  return reasons;
+}
+
+function shouldUseChatLlm(args = {}, normalized = normalizeChatStartInput(args)) {
+  return normalizeScreeningModeArg(args, normalized) === "llm";
+}
+
+function getRunOptions(args, normalized, session, { workspaceRoot = "", configResolution = null } = {}) {
   const slowLive = args.slow_live === true;
   const isAllTarget = normalized.publicTargetCount === "all";
   const processedLimit = parsePositiveInteger(
     args.max_candidates,
     isAllTarget ? CHAT_ALL_MAX_CANDIDATES : CHAT_ALL_MAX_CANDIDATES
-  );
-  const shouldRequestResume = shouldRequestChatResume(args);
-  const useLlm = shouldUseChatLlm(args);
-  const resolvedConfig = configResolution || (useLlm ? resolveBossScreeningConfig(workspaceRoot) : { ok: false });
-  const humanBehavior = resolveHumanBehaviorForRun(args, resolvedConfig?.config || {});
-  return {
+  );
+  const shouldRequestResume = shouldRequestChatResume(args);
+  const screeningMode = normalizeScreeningModeArg(args, normalized);
+  const useLlm = shouldUseChatLlm(args, normalized);
+  const resolvedConfig = configResolution || (useLlm ? resolveBossScreeningConfig(workspaceRoot) : { ok: false });
+  const humanBehavior = resolveHumanBehaviorForRun(args, resolvedConfig?.config || {});
+  return {
     client: session.client,
     targetUrl: CHAT_TARGET_URL,
     job: normalized.job,
@@ -1302,7 +1297,7 @@ function getRunOptions(args, normalized, session, { workspaceRoot = "", configRe
     llmImageDetail: normalizeText(
       args.llm_image_detail || resolvedConfig.config?.llmImageDetail || resolvedConfig.config?.imageDetail
     ) || "low",
-    screeningMode: normalizeScreeningModeArg(args),
+    screeningMode,
     listMaxScrolls: parsePositiveInteger(args.list_max_scrolls, 200),
     listStableSignatureLimit: parsePositiveInteger(args.list_stable_signature_limit, 2),
     listWheelDeltaY: parsePositiveInteger(args.list_wheel_delta_y, 850),
@@ -1373,7 +1368,7 @@ async function startBossChatRunInternal(args = {}, { workspaceRoot = "", runId =
   }
 
   const shouldRequestResume = shouldRequestChatResume(args);
-  const useLlm = shouldUseChatLlm(args);
+  const useLlm = shouldUseChatLlm(args, normalized);
   const debugTestOptions = collectChatDebugTestOptions(args);
   if (debugTestOptions.length && !isDebugTestMode(args)) {
     return {
@@ -1538,9 +1533,11 @@ export async function prepareBossChatRunTool({ workspaceRoot = "", args = {} } =
       status: missingFields.length ? "NEED_INPUT" : "READY",
       stage: "chat_run_setup",
       page_url: session.navigation?.url || session.target?.url || CHAT_TARGET_URL,
-      required_fields: CHAT_REQUIRED_FIELDS.slice(),
-      missing_fields: missingFields,
-      job_options: jobOptions,
+      required_fields: CHAT_REQUIRED_FIELDS.slice(),
+      missing_fields: missingFields,
+      criteria_optional: true,
+      empty_criteria_mode: "collect_cv",
+      job_options: jobOptions,
       selected_job: selectedJob,
       selected_job_label: jobs?.selected_label || selectedJob?.label || "",
       job_options_source: jobs?.source || "",
@@ -1552,10 +1549,10 @@ export async function prepareBossChatRunTool({ workspaceRoot = "", args = {} } =
         jobOptions
       }),
       ...diagnostics,
-      ...(nextCallExample ? { next_call_example: nextCallExample } : {}),
-      message: missingFields.length
-        ? "已通过 CDP-only 读取 Boss 聊天页岗位列表，请补齐 job / start_from / target_count / criteria。"
-        : "Boss chat CDP-only preflight is ready. Use start_boss_chat_run to start screening.",
+      ...(nextCallExample ? { next_call_example: nextCallExample } : {}),
+      message: missingFields.length
+        ? "已通过 CDP-only 读取 Boss 聊天页岗位列表，请补齐 job / start_from / target_count。criteria 可留空；留空会进入收集简历模式。"
+        : "Boss chat CDP-only preflight is ready. Use start_boss_chat_run to start screening or collect CVs.",
       runtime_evaluate_used: false,
       method_summary: methodSummary(session.methodLog || []),
       method_log: session.methodLog || [],
@@ -1723,7 +1720,7 @@ export async function startBossChatDetachedRunTool({ workspaceRoot = "", args =
     });
   }
 
-  const useLlm = shouldUseChatLlm(args);
+  const useLlm = shouldUseChatLlm(args, normalized);
   const debugTestOptions = collectChatDebugTestOptions(args);
   if (debugTestOptions.length && !isDebugTestMode(args)) {
     return {

package/src/domains/chat/detail.js

@@ -51,7 +51,8 @@ import {
   makeForbiddenChatResumeNavigationError
 } from "./page-guard.js";
 
-export const CHAT_UNSAFE_ONLINE_RESUME_LINK_CODE = "CHAT_UNSAFE_ONLINE_RESUME_LINK";
+export const CHAT_UNSAFE_ONLINE_RESUME_LINK_CODE = "CHAT_UNSAFE_ONLINE_RESUME_LINK";
+export const CHAT_ONLINE_RESUME_MODAL_NOT_OPEN_CODE = "CHAT_ONLINE_RESUME_MODAL_NOT_OPEN";
 
 const CHAT_CONVERSATION_CONTROL_SCOPE_SELECTORS = Object.freeze([
   ".conversation-main",
@@ -128,10 +129,15 @@ export function makeUnsafeChatOnlineResumeLinkError(target = {}, buttonHTML = ""
   return error;
 }
 
-export function isUnsafeChatOnlineResumeLinkError(error) {
-  return error?.code === CHAT_UNSAFE_ONLINE_RESUME_LINK_CODE
-    || /CHAT_UNSAFE_ONLINE_RESUME_LINK/i.test(String(error?.message || error || ""));
-}
+export function isUnsafeChatOnlineResumeLinkError(error) {
+  return error?.code === CHAT_UNSAFE_ONLINE_RESUME_LINK_CODE
+    || /CHAT_UNSAFE_ONLINE_RESUME_LINK/i.test(String(error?.message || error || ""));
+}
+
+export function isChatOnlineResumeModalOpenFailureError(error) {
+  return error?.code === CHAT_ONLINE_RESUME_MODAL_NOT_OPEN_CODE
+    || /Chat online resume modal did not open/i.test(String(error?.message || error || ""));
+}
 
 export function createChatProfileNetworkRecorder(client) {
   const events = [];
@@ -242,9 +248,9 @@ function chatCandidateIdFromAttributes(attributes = {}) {
   );
 }
 
-async function hydrateActiveChatCandidate(client, activeCandidate = null) {
-  if (!activeCandidate?.node_id) return activeCandidate;
-  let attributes = {};
+async function hydrateActiveChatCandidate(client, activeCandidate = null) {
+  if (!activeCandidate?.node_id) return activeCandidate;
+  let attributes = {};
   let outerHTML = "";
   try {
     [attributes, outerHTML] = await Promise.all([
@@ -258,12 +264,25 @@ async function hydrateActiveChatCandidate(client, activeCandidate = null) {
     candidate_id: chatCandidateIdFromAttributes(attributes) || null,
     label: normalizeDetailText(htmlToText(outerHTML)),
     outer_html_length: outerHTML.length
-  };
-}
-
-export async function waitForChatOnlineResumeButton(client, {
-  timeoutMs = 12000,
-  intervalMs = 250,
+  };
+}
+
+export async function readChatActiveCandidateState(client) {
+  const rootState = await getChatRoots(client);
+  const activeCandidate = await queryFirstAcrossChatRoots(
+    client,
+    rootState.roots,
+    CHAT_ACTIVE_CANDIDATE_SELECTORS
+  );
+  return {
+    roots: rootState.roots,
+    active_candidate: await hydrateActiveChatCandidate(client, activeCandidate)
+  };
+}
+
+export async function waitForChatOnlineResumeButton(client, {
+  timeoutMs = 12000,
+  intervalMs = 250,
   expectedCandidateId = ""
 } = {}) {
   const started = Date.now();
@@ -982,10 +1001,12 @@ export async function openChatOnlineResume(client, {
     }
   }
 
-  const error = new Error("Chat online resume modal did not open");
-  error.attempts = attempts;
-  throw error;
-}
+  const error = new Error("Chat online resume modal did not open");
+  error.code = CHAT_ONLINE_RESUME_MODAL_NOT_OPEN_CODE;
+  error.retryable = true;
+  error.attempts = attempts;
+  throw error;
+}
 
 export async function readChatConversationReadyState(client) {
   const rootState = await getChatRoots(client);
@@ -1173,20 +1194,21 @@ export async function sendChatMessage(client, expectedText = "", {
   };
 }
 
-export async function clickChatAskResume(client, {
-  timeoutMs = 8000,
-  settleMs = 700
-} = {}) {
+export async function clickChatAskResume(client, {
+  timeoutMs = 8000,
+  settleMs = 700,
+  skipWhenAttachmentResumeAvailable = true
+} = {}) {
   const started = Date.now();
   let lastState = null;
   let lastDisabledAskResume = null;
   while (Date.now() - started <= timeoutMs) {
     const state = await readChatConversationReadyState(client);
     lastState = state;
-    if (state.attachment_resume_enabled) {
-      return {
-        ok: true,
-        already_requested: true,
+    if (skipWhenAttachmentResumeAvailable && state.attachment_resume_enabled) {
+      return {
+        ok: true,
+        already_requested: true,
         attachment_resume_available: true,
         control: state.attachment_resume
       };
@@ -1368,23 +1390,32 @@ export async function waitForChatResumeRequestMessage(client, {
   };
 }
 
-export async function requestChatResumeForPassedCandidate(client, {
-  greetingText = "Hi同学，能麻烦发下简历吗？",
-  maxAttempts = 3,
-  askResumeTimeoutMs = 8000,
-  dryRun = false
-} = {}) {
-  const effectiveGreetingText = normalizeDetailText(greetingText) || "Hi同学，能麻烦发下简历吗？";
-  const initialState = await readChatConversationReadyState(client);
-  if (initialState.attachment_resume_enabled) {
-    return {
-      requested: false,
-      skipped: true,
-      reason: "attachment_resume_already_available",
-      initial_state: initialState
-    };
-  }
-  if (dryRun) {
+export async function requestChatResumeForPassedCandidate(client, {
+  greetingText = "Hi同学，能麻烦发下简历吗？",
+  maxAttempts = 3,
+  askResumeTimeoutMs = 8000,
+  dryRun = false,
+  skipWhenAttachmentResumeAvailable = true
+} = {}) {
+  const effectiveGreetingText = normalizeDetailText(greetingText) || "Hi同学，能麻烦发下简历吗？";
+  const initialState = await readChatConversationReadyState(client);
+  if (skipWhenAttachmentResumeAvailable && initialState.attachment_resume_enabled) {
+    return {
+      requested: false,
+      skipped: true,
+      reason: "attachment_resume_already_available",
+      initial_state: initialState
+    };
+  }
+  if (initialState.already_requested_resume) {
+    return {
+      requested: false,
+      skipped: true,
+      reason: "resume_request_already_pending",
+      initial_state: initialState
+    };
+  }
+  if (dryRun) {
     return {
       requested: false,
       skipped: false,
@@ -1416,10 +1447,11 @@ export async function requestChatResumeForPassedCandidate(client, {
 
   const attempts = [];
   for (let attempt = 0; attempt < maxAttempts; attempt += 1) {
-    const before = await getChatResumeRequestMessageState(client);
-    const askResult = await clickChatAskResume(client, {
-      timeoutMs: askResumeTimeoutMs
-    });
+    const before = await getChatResumeRequestMessageState(client);
+    const askResult = await clickChatAskResume(client, {
+      timeoutMs: askResumeTimeoutMs,
+      skipWhenAttachmentResumeAvailable
+    });
     let confirmResult = {
       confirmed: false,
       assumed_requested: Boolean(askResult.already_requested),