完善岗位详情、周期投递时段与拟人化投递节奏

src-tauri/src/agent/tasks.rs

@@ -11,6 +11,7 @@ use crate::agent::prompts;
 use crate::agent::run::AgentTask;
 use crate::config::{AppRuntimeConfig, RegexRule};
 use crate::error::AppError;
+use crate::job_description;
 use crate::llm::template;
 use crate::llm::JobSemanticMatch;
 use crate::rpa::common::RpaJob;
@@ -106,7 +107,7 @@ impl AgentTask for ReplyDecisionTask<'_> {
                 .map(|message| message.text.clone())
                 .unwrap_or_default(),
             "job_description": job
-                .map(|job| clip(&job.detail, LONG_TEXT_LIMIT))
+                .map(|job| clip(&job_description::clean_text(&job.detail, &job.platform), LONG_TEXT_LIMIT))
                 .unwrap_or_default(),
             "job_content": job
                 .map(|job| format!("{}｜{}｜{}", job.title, job.company_name, job.salary))
@@ -182,6 +183,10 @@ impl<'a> GreetTask<'a> {
     pub fn new(config: &'a AppRuntimeConfig, job: &'a RpaJob) -> Self {
         Self { config, job }
     }
+
+    fn job_description(&self) -> String {
+        job_description::clean_text(&self.job.detail, self.job.platform.as_str())
+    }
 }
 
 impl AgentTask for GreetTask<'_> {
@@ -205,7 +210,7 @@ impl AgentTask for GreetTask<'_> {
     fn params(&self) -> Result<Value, AppError> {
         let mut params = json!({
             "job_content": serde_json::to_string(self.job).unwrap_or_default(),
-            "job_description": clip(&self.job.detail, LONG_TEXT_LIMIT),
+            "job_description": clip(&self.job_description(), LONG_TEXT_LIMIT),
             "resume": resume_text(self.config),
             "resume_context": resume_text(self.config),
         });
@@ -275,6 +280,10 @@ impl<'a> JobMatchTask<'a> {
     pub fn new(config: &'a AppRuntimeConfig, job: &'a RpaJob) -> Self {
         Self { config, job }
     }
+
+    fn job_description(&self) -> String {
+        job_description::clean_text(&self.job.detail, self.job.platform.as_str())
+    }
 }
 
 impl AgentTask for JobMatchTask<'_> {
@@ -307,7 +316,7 @@ impl AgentTask for JobMatchTask<'_> {
             "job": {
                 "title": self.job.title,
                 "company": self.job.company_name,
-                "description": clip(&self.job.detail, LONG_TEXT_LIMIT),
+                "description": clip(&self.job_description(), LONG_TEXT_LIMIT),
             },
             "resume": resume_text(self.config),
         });

src-tauri/src/command/communicated_jobs.rs

@@ -12,7 +12,7 @@ use crate::{
     command::base::CommandResult,
     dao::{chat_message_dao, job_detail_dao, model::JobDetail},
     logger,
-    rpa::boss::handler::{parse_chat_messages, parse_encrypt_job_id},
+    rpa::boss::handler::{parse_chat_messages, parse_encrypt_job_id, MessageDirection},
     rpa::run_flow::{is_job_task_stop_requested, try_start_job_task, JobTaskRunningGuard},
     utils::salary::decode_salary,
 };
@@ -116,6 +116,8 @@ async fn collect_communicated_jobs_from_browser(
     let mut jobs = Vec::new();
     let mut seen_ids = HashSet::new();
     let mut messages_inserted = 0;
+    // 我方 uid 一轮内不变，跨会话攒着用，见 [`MessageDirection`]
+    let mut direction = MessageDirection::default();
 
     for page_no in 1..=page_limit {
         if is_job_task_stop_requested() {
@@ -158,7 +160,7 @@ async fn collect_communicated_jobs_from_browser(
                 continue;
             }
 
-            match collect_conversation_messages(page, &snapshot, &id) {
+            match collect_conversation_messages(page, &snapshot, &id, &mut direction) {
                 Ok(inserted) => {
                     messages_inserted += inserted;
                     if inserted > 0 {
@@ -263,6 +265,7 @@ fn collect_conversation_messages(
     page: &ChromiumPage,
     snapshot: &CommunicatedJobCardSnapshot,
     fallback_job_id: &str,
+    direction: &mut MessageDirection,
 ) -> Result<usize, anyhow::Error> {
     let history_listener = page.listen_url("zpchat/geek/historyMsg")?;
     let boss_data_listener = page.listen_url("zpchat/geek/getBossData")?;
@@ -287,7 +290,7 @@ fn collect_conversation_messages(
     };
 
     let body_str = String::from_utf8(body).context("historyMsg 响应非 UTF-8 编码")?;
-    let chat_messages = parse_chat_messages(&body_str)?;
+    let chat_messages = parse_chat_messages(&body_str, direction)?;
     // BOSS 的 encryptJobId 同时充当会话标识与岗位标识
     let saved = chat_message_dao::upsert_incremental(
         chat_message_dao::ConversationKey {
@@ -520,26 +523,15 @@ fn extract_detail_text(page: &Page) -> Result<String, anyhow::Error> {
     })()
     "#;
     let value = page.run_js_await(script)?;
-    let raw_text = extract_remote_value(value)
+    // 原样落库：清洗规则跟着平台页面结构走，统一放在读取侧的
+    // [`crate::job_description`] 里。写入时就洗掉，规则一改旧数据就再也救不回来了
+    Ok(extract_remote_value(value)
         .as_str()
         .unwrap_or("")
+        .replace("\r\n", "\n")
+        .replace('\r', "\n")
         .trim()
-        .to_string();
-
-    Ok(sanitize_boss_job_detail(&raw_text))
-}
-
-fn sanitize_boss_job_detail(text: &str) -> String {
-    let normalized = text.replace("\r\n", "\n").replace('\r', "\n");
-    let marker = "职位描述";
-    let body = normalized
-        .find(marker)
-        .map(|index| &normalized[index + marker.len()..])
-        .unwrap_or(normalized.as_str());
-
-    let cleaned = body.trim().trim_start_matches([':', '：', '-', '—']).trim();
-
-    cleaned.replace("\n\n\n", "\n\n")
+        .to_string())
 }
 
 fn scroll_page_once(page: &ChromiumPage) -> Result<(), anyhow::Error> {
@@ -640,13 +632,17 @@ mod tests {
         );
     }
 
+    /// 抓到的原文原样落库，「职位描述」这类标题留给读取侧切小节用
     #[test]
-    fn sanitizes_boss_detail_text_before_job_description_heading() {
+    fn keeps_scraped_detail_intact_for_the_read_side_parser() {
         let raw = "微信扫码分享\n举\n报\n职位描述\n负责推荐系统后端开发。\n要求熟悉 Rust。";
+        let parsed = crate::job_description::parse(raw, "boss");
 
+        assert_eq!(parsed.sections.len(), 1);
+        assert_eq!(parsed.sections[0].title, "职位描述");
         assert_eq!(
-            sanitize_boss_job_detail(raw),
-            "负责推荐系统后端开发。\n要求熟悉 Rust。"
+            parsed.sections[0].items,
+            vec!["负责推荐系统后端开发。", "要求熟悉 Rust。"]
         );
     }
 

src-tauri/src/command/config.rs

@@ -10,13 +10,17 @@ pub fn load_app_config(app_handle: tauri::AppHandle) -> CommandResult<AppRuntime
     }
 }
 
+/// 保存配置，并把**落盘后**的那份交回前端。
+///
+/// 保存路径上会做迁移、夹取、补生成人格种子，落盘内容与提交内容并不相同。
+/// 前端拿这个返回值刷新自己手里的配置，才不会带着旧值继续编辑
 #[tauri::command]
 pub fn save_app_config(
     app_handle: tauri::AppHandle,
     config: AppRuntimeConfig,
-) -> CommandResult<()> {
+) -> CommandResult<AppRuntimeConfig> {
     match config::save_app_config_inner(app_handle, config) {
-        Ok(()) => CommandResult::ok(()),
+        Ok(saved) => CommandResult::ok(saved),
         Err(err) => CommandResult::err(err),
     }
 }

src-tauri/src/command/job.rs

@@ -2,6 +2,7 @@ use crate::command::base::CommandResult;
 use crate::config;
 use crate::dao::model::{ChatMessageRecord, InterviewJobAnalysis, JobDetail};
 use crate::dao::{analysis_dao, chat_message_dao, job_detail_dao};
+use crate::job_description::ParsedJobDescription;
 use chrono::{Duration, Local, NaiveDate, NaiveDateTime, TimeZone};
 use serde::{Deserialize, Serialize};
 use std::collections::{HashMap, HashSet};
@@ -630,6 +631,22 @@ pub fn job_delete(id: String) -> CommandResult<()> {
     }
 }
 
+/// 岗位描述的结构化视图。
+///
+/// 前端不再自己洗一遍 JD：清洗规则跟着平台页面结构走，前后端各存一份必然漂移。
+/// 页面要展示什么就从这个出口取，和喂给模型的是同一份文本
+#[tauri::command]
+pub fn job_description_view(job_id: String) -> CommandResult<ParsedJobDescription> {
+    match job_detail_dao::get_by_id(&job_id) {
+        Ok(Some(job)) => CommandResult::ok(crate::job_description::parse(
+            &job.detail,
+            &job.platform,
+        )),
+        Ok(None) => CommandResult::err(format!("岗位不存在: {}", job_id)),
+        Err(e) => CommandResult::err(e.to_string()),
+    }
+}
+
 #[tauri::command]
 pub fn chat_messages_by_job(job_id: String) -> CommandResult<Vec<ChatMessageRecord>> {
     match chat_message_dao::find_by_job_id(&job_id) {
@@ -675,6 +692,9 @@ struct LlmAnalysisOutput {
 }
 
 fn build_analysis_prompt(job: &JobDetail) -> String {
+    // 抓下来的 JD 混着反爬注入的样式代码和噪声词，原样喂进去既占额度又干扰判断，
+    // 统一走 job_description 这个出口洗一遍
+    let detail = crate::job_description::clean_text(&job.detail, &job.platform);
     // 只填岗位骨架；resume_context / background_context / chat_context 这些
     // 业务变量留给后续的模板渲染，两层的替换时机不同不能混做
     crate::agent::prompts::compose(
@@ -684,7 +704,7 @@ fn build_analysis_prompt(job: &JobDetail) -> String {
             ("JOB_COMPANY", &job.company_name),
             ("JOB_SALARY", &job.salary),
             ("JOB_LOCATION", job.location.as_deref().unwrap_or("-")),
-            ("JOB_DETAIL", &job.detail),
+            ("JOB_DETAIL", &detail),
         ],
     )
 }
@@ -745,7 +765,11 @@ pub async fn job_analyze_batch(
         Err(e) => return CommandResult::err(format!("加载配置失败: {}", e)),
     };
     if app_config.llm_chain().is_empty() {
-        return CommandResult::err("请先配置大模型服务".to_string());
+        return CommandResult::err(if app_config.llm_configured() {
+            "大模型已停用，请先启用模型服务".to_string()
+        } else {
+            "请先配置大模型服务".to_string()
+        });
     }
     let skip_analyzed = skip_analyzed.unwrap_or(true);
 

src-tauri/src/command/llm_provider.rs

@@ -129,8 +129,8 @@ where
 }
 
 /// 在降级链中按标识定位一个服务。
-/// 链由已保存的配置推导而来，所以「找不到」既可能是标识写错，
-/// 也可能是用户刚在界面上填完还没保存，错误文案需要同时覆盖这两种情况。
+/// 链只收录填写完整且处于启用状态的服务，所以「找不到」通常是缺服务地址或模型名，
+/// 其次才是该服务被停用了。
 fn find_chain_link<'a>(
     chain: &'a [LlmChainLink],
     entry_id: &str,
@@ -139,9 +139,7 @@ fn find_chain_link<'a>(
         .iter()
         .find(|link| link.id == entry_id)
         .ok_or_else(|| {
-            AppError::configuration(
-                "未找到该大模型服务，请先保存配置后再测试；未保存或已停用的服务无法测试",
-            )
+            AppError::configuration("该大模型服务尚不可用，请补齐服务地址和模型名称并确认它已启用")
         })
 }
 
@@ -179,19 +177,43 @@ pub async fn test_llm_entry_connection(
     }
 }
 
-/// 列出降级链中某个服务可用的模型
+/// 界面上正在编辑的服务参数。两项齐备才算数：只传一半无从判断该配哪个客户端，
+/// 与其猜一个，不如退回已落盘的配置。
+fn draft_override(
+    provider: Option<LlmProviderPreset>,
+    base_url: Option<String>,
+) -> Option<(LlmProviderPreset, String)> {
+    let base_url = base_url.map(|url| url.trim().to_string())?;
+    if base_url.is_empty() {
+        return None;
+    }
+    Some((provider?, base_url))
+}
+
+/// 列出降级链中某个服务可用的模型。
+///
+/// `provider` / `base_url` 是界面上正在编辑的草稿，传了就以草稿为准。
+/// 备用服务的模型名为空时整份配置根本保存不了，若这里坚持只读已落盘的配置，
+/// 就成了「保存要模型、取模型要先保存」的死锁——新建的备用服务永远拿不到模型列表。
+/// 密钥始终按 `entry_id` 从凭据库读取，明文不经过前端。
 #[tauri::command]
 pub async fn list_llm_models_for(
     app_handle: tauri::AppHandle,
     entry_id: String,
+    provider: Option<LlmProviderPreset>,
+    base_url: Option<String>,
 ) -> CommandResult<Vec<String>> {
-    let (link, credential) = match resolve_chain_entry(app_handle, &entry_id) {
-        Ok(resolved) => resolved,
-        Err(error) => return CommandResult::err(error),
+    let (provider, base_url, credential) = match draft_override(provider, base_url) {
+        Some((provider, base_url)) => match credential::resolve_for_entry(&entry_id) {
+            Ok(credential) => (provider, base_url, credential),
+            Err(error) => return CommandResult::err(error),
+        },
+        None => match resolve_chain_entry(app_handle, &entry_id) {
+            Ok((link, credential)) => (link.provider, link.base_url, credential),
+            Err(error) => return CommandResult::err(error),
+        },
     };
-    to_command_result(
-        fetch_model_list_with_credential(link.provider, &link.base_url, &credential).await,
-    )
+    to_command_result(fetch_model_list_with_credential(provider, &base_url, &credential).await)
 }
 
 fn validate_llm_base_url(base_url: &str) -> Result<(), AppError> {
@@ -463,23 +485,54 @@ mod tests {
     }
 
     #[test]
-    fn find_chain_link_reports_unsaved_service_clearly() {
+    fn find_chain_link_reports_an_unusable_service_clearly() {
         let chain = chain();
 
-        // 界面上刚填完还没保存就点测试，是最常见的误用，错误文案必须点明这一点
-        let error = find_chain_link(&chain, "backup-unsaved").unwrap_err();
+        // 草稿也会落盘，所以链里没有它多半是缺地址或模型名，错误文案要直接指向该补什么
+        let error = find_chain_link(&chain, "backup-incomplete").unwrap_err();
 
         assert_eq!(error.code, AppErrorCode::Configuration);
-        assert!(error.message.contains("未找到该大模型服务"));
-        assert!(error.message.contains("保存"));
+        assert!(error.message.contains("尚不可用"));
+        assert!(error.message.contains("模型名称"));
     }
 
     #[test]
     fn find_chain_link_on_empty_chain_reports_the_same_error() {
         let error = find_chain_link(&[], PRIMARY_LLM_ENTRY_ID).unwrap_err();
 
         assert_eq!(error.code, AppErrorCode::Configuration);
-        assert!(error.message.contains("未找到该大模型服务"));
+        assert!(error.message.contains("尚不可用"));
+    }
+
+    #[test]
+    fn draft_override_takes_the_editing_values_and_trims_them() {
+        let draft = draft_override(
+            Some(LlmProviderPreset::DeepSeek),
+            Some("  https://api.deepseek.com  ".to_string()),
+        );
+
+        assert_eq!(
+            draft,
+            Some((
+                LlmProviderPreset::DeepSeek,
+                "https://api.deepseek.com".to_string()
+            ))
+        );
+    }
+
+    #[test]
+    fn draft_override_falls_back_to_saved_config_when_incomplete() {
+        // 只传一半、空串、全不传，都退回已落盘的配置，不去猜另一半
+        assert_eq!(draft_override(Some(LlmProviderPreset::OpenAi), None), None);
+        assert_eq!(
+            draft_override(None, Some("https://api.openai.com/v1".to_string())),
+            None
+        );
+        assert_eq!(
+            draft_override(Some(LlmProviderPreset::OpenAi), Some("   ".to_string())),
+            None
+        );
+        assert_eq!(draft_override(None, None), None);
     }
 
     #[test]