diff --git a/crates/df-storage/src/crud/message_repo.rs b/crates/df-storage/src/crud/message_repo.rs
index 8eb7dde..39ae826 100644
--- a/crates/df-storage/src/crud/message_repo.rs
+++ b/crates/df-storage/src/crud/message_repo.rs
@@ -265,6 +265,12 @@ impl AiMessageRepo {
     /// 全量重写语义——以入参 records 为该对话的**唯一真相**,先删该 conv 全部旧行再批量插。
     /// 单事务保证「删 + 插」原子,无中间空窗(reload 不会读到半删半插的中间态)。
     ///
+    /// created_at 保出生时间(2026-10-17 修复批量污染):created_at 语义 = 消息首次入库时间,
+    /// 全量重写只该变内容不变出生时间。调用方(save_conversation)对整批 records 统一盖
+    /// 对话级 created_at(老对话 None 时 now 兜底),直插会把全对话消息 created_at 抹成同一秒。
+    /// 故 DELETE 前先取存量 `id → created_at` 映射,重插时同 id 行回填原值;新 id(真新消息)
+    /// 保留入参值。ai_messages 表无 updated_at 列,内容变更无单独时间戳(不在此列语义内)。
+    ///
     /// 设计权衡(非 dirty 增量,留 P2.1 优化):
     /// - 内存 ContextManager 是运行时真相源,save 是同步点,每轮 save 全量回写简单可靠;
     /// - INSERT OR IGNORE 幂等:records 内 id 重复或与残留行(理论不应有,事务已 DELETE)冲突跳过;
@@ -274,7 +280,7 @@ impl AiMessageRepo {
     pub async fn replace_conversation(
         &self,
         conv_id: &str,
-        records: Vec<AiMessageRecord>,
+        mut records: Vec<AiMessageRecord>,
     ) -> Result<()> {
         let conn = self.conn.clone();
         let conv_id = conv_id.to_owned();
@@ -282,7 +288,29 @@ impl AiMessageRepo {
             let mut guard = conn.blocking_lock();
             let tx = guard.transaction().map_err(storage_err)?;
             {
-                // 先删该 conv 全部旧行(全量重写语义)
+                // 先取存量 id → created_at 映射(DELETE 前留证,重插回填保出生时间)
+                {
+                    let mut stmt = tx
+                        .prepare("SELECT id, created_at FROM ai_messages WHERE conversation_id = ?1")
+                        .map_err(storage_err)?;
+                    let rows = stmt
+                        .query_map(params![conv_id], |r| {
+                            Ok((r.get::<_, String>("id")?, r.get::<_, String>("created_at")?))
+                        })
+                        .map_err(storage_err)?;
+                    let mut original: std::collections::HashMap<String, String> =
+                        std::collections::HashMap::new();
+                    for r in rows {
+                        let (id, created_at) = r.map_err(storage_err)?;
+                        original.insert(id, created_at);
+                    }
+                    for rec in &mut records {
+                        if let Some(orig) = original.get(&rec.id) {
+                            rec.created_at = orig.clone();
+                        }
+                    }
+                }
+                // 再删该 conv 全部旧行(全量重写语义)
                 tx.execute(
                     "DELETE FROM ai_messages WHERE conversation_id = ?1",
                     params![conv_id],
@@ -693,6 +721,62 @@ mod tests {
         assert_eq!(got.len(), 1, "重复 replace 不应叠加");
     }
 
+    /// replace_conversation 保 created_at 出生时间(2026-10-17 批量污染回归):
+    /// 调用方对整批统一盖一个 created_at(如对话级时间或 now),重写后既有 id 行
+    /// 的 created_at 必须逐行保留原入库值,不得被统一时间抹平;新 id 行保留入参值。
+    #[tokio::test]
+    async fn replace_conversation_preserves_created_at() {
+        let db = Database::open_in_memory().await.expect("open_in_memory");
+        let repo = AiMessageRepo::new(&db);
+
+        // 预置两行,created_at 各异(模拟真实逐条入库的出生时间)
+        let mut m0 = mk_msg("m0", "c", 0, "user", "v1");
+        m0.created_at = "2026-10-01T10:00:00.000".into();
+        let mut m1 = mk_msg("m1", "c", 1, "assistant", "v1");
+        m1.created_at = "2026-10-02T11:30:00.000".into();
+        repo.insert_batch(vec![m0, m1]).await.expect("insert");
+
+        // 重写:既有 id 内容变了、created_at 被调用方统一盖成同一时间;
+        // 另带一条新 id 消息(真新消息,该保留入参时间)。
+        let uniform_ts = "2026-10-17T21:25:01.000";
+        let rewrite = |id: &str, seq: i64, role: &str, content: &str| {
+            let mut m = mk_msg(id, "c", seq, role, content);
+            m.created_at = uniform_ts.into();
+            m
+        };
+        repo.replace_conversation(
+            "c",
+            vec![
+                rewrite("m0", 0, "user", "v2-改写内容"),
+                rewrite("m1", 1, "assistant", "v2-改写内容"),
+                rewrite("m2_new", 2, "user", "新消息"),
+            ],
+        )
+        .await
+        .expect("replace");
+
+        let got = repo.list_by_conversation("c").await.expect("list");
+        assert_eq!(got.len(), 3);
+        // 逐行断言:既有 id 的 created_at 保持原值,内容确实变了
+        assert_eq!(got[0].id, "m0");
+        assert_eq!(got[0].created_at, "2026-10-01T10:00:00.000", "既有行 created_at 不得被重写抹平");
+        assert_eq!(got[0].content, "v2-改写内容");
+        assert_eq!(got[1].id, "m1");
+        assert_eq!(got[1].created_at, "2026-10-02T11:30:00.000", "既有行 created_at 不得被重写抹平");
+        assert_eq!(got[1].content, "v2-改写内容");
+        // 新 id(首次入库)保留入参 created_at
+        assert_eq!(got[2].id, "m2_new");
+        assert_eq!(got[2].created_at, uniform_ts, "新消息保留入参 created_at");
+
+        // 二次重写再验:出生时间已固化,连续重写不漂移
+        repo.replace_conversation("c", vec![rewrite("m0", 0, "user", "v3")])
+            .await
+            .expect("replace again");
+        let got = repo.list_by_conversation("c").await.expect("list");
+        assert_eq!(got.len(), 1);
+        assert_eq!(got[0].created_at, "2026-10-01T10:00:00.000", "连续重写 created_at 不漂移");
+    }
+
     /// limit=0 契约:返空集(直传 SQLite LIMIT 0),不再旧版 max(1) 钳成 1 条。
     /// 锁死语义防回归:请求 0 条却回 1 条会让调用方误判页内有数据。
     #[tokio::test]
diff --git a/crates/df-storage/src/crud/project_repo.rs b/crates/df-storage/src/crud/project_repo.rs
index 0368de3..54112da 100644
--- a/crates/df-storage/src/crud/project_repo.rs
+++ b/crates/df-storage/src/crud/project_repo.rs
@@ -87,9 +87,9 @@ fn build_order_clause(order_by: Option<&str>) -> Result<String> {
 
 /// 带「最近活跃时间」的项目记录(list_active_with_activity 专用返回结构)。
 ///
-/// `last_active_at` = COALESCE(project_events 最新事件 created_at, projects.updated_at):
-/// 反映业务活跃(任务/灵感/状态推进等事件),无事件项目回退 updated_at。供前端「最近活跃排序」
-/// 展示,语义比 projects.updated_at(随元信息修改如改 description 也会刷新)更准确。
+/// `last_active_at` = MAX(tasks.updated_at 最新值, project_events 最新 created_at,
+/// projects.updated_at):任务活跃为主信号(任务面是真实工作发生地),事件表已停写仅保留维度,
+/// updated_at 兜底。供前端「最近活跃排序」展示与 system prompt 项目清单排序。
 ///
 /// Serialize 供 IPC 层直接序列化回前端(serde 字段名 snake_case,对齐 ProjectRecord)。
 #[derive(Debug, Clone, serde::Serialize)]
@@ -243,29 +243,35 @@ impl ProjectRepo {
         .map_err(storage_err)?
     }
 
-    /// 列出未删除项目,**按「最近活跃」排序**(LEFT JOIN project_events 取最新事件时间,
-    /// COALESCE 回退 updated_at)。
+    /// 列出未删除项目,**按「最近活跃」排序**(三源取大:任务活跃为主信号)。
     ///
-    /// 业务语义:项目「最近活跃」应反映业务事件(任务/灵感/状态推进),而非 projects.updated_at
-    /// (后者随元信息修改如改 description/path 也会刷新,不反映真实业务活跃)。本方法通过子查询
-    /// 取每个项目 project_events 最新 created_at,无事件项目回退 updated_at,排序列统一可比。
+    /// 业务语义:last_active_at = MAX(tasks.updated_at 最新值, project_events 最新 created_at,
+    /// projects.updated_at)。**任务活跃为主信号**——ai-chat 会话/任务面是真实工作发生地,
+    /// 任务停更即项目停更;project_events 已停写(现网空表,保留维度向后兼容);
+    /// projects.updated_at 仅元信息修改(改 description/path 也会刷新,不反映真实业务活跃),
+    /// 仅作兜底。无任务无事件的项目回退 updated_at,排序列统一可比。
     ///
-    /// 子查询而非 JOIN:project_events 每项目可能 0..N 条,JOIN 会展开需 DISTINCT,子查询
-    /// `(SELECT MAX(created_at) FROM project_events WHERE project_id = projects.id)` 每行一次
-    /// 聚合,语义清晰无笛卡尔积风险。单用户桌面应用项目数小,无性能压力。
+    /// 子查询而非 JOIN:tasks/project_events 每项目可能 0..N 条,JOIN 会展开需 DISTINCT,
+    /// 标量子查询每行一次聚合,语义清晰无笛卡尔积风险(SQLite 多参 MAX 为标量取大,
+    /// IFNULL 兜空子查询 NULL)。单用户桌面应用项目数小,无性能压力。
     ///
-    /// 返回 `ProjectActivityRecord`(ProjectRecord + last_active_at 字段),供前端排序展示。
+    /// 返回 `ProjectActivityRecord`(ProjectRecord + last_active_at 字段),供前端排序展示
+    /// 与 system prompt「当前项目」清单(build_context_lists)按活跃度排序。
     pub async fn list_active_with_activity(&self) -> Result<Vec<ProjectActivityRecord>> {
         let conn = self.conn.clone();
         tokio::task::spawn_blocking(move || {
             let guard = conn.blocking_lock();
-            // COALESCE(子查询最新事件, projects.updated_at):无事件项目回退 updated_at,
-            // 保证所有项目有可比活跃时间。ORDER BY 该列 DESC,同时间 created_at DESC 兜底稳定。
+            // 三源取大(SQLite 多参 MAX=标量取大,IFNULL 兜空子查询):任务活跃为主信号,
+            // 事件表已停写仅保留维度,updated_at 兜底。ORDER BY 该列 DESC,同时间 created_at DESC
+            // 兜底稳定。
             let mut stmt = guard
                 .prepare(
                     "SELECT id, name, description, status, idea_id, path, stack, archived, created_at, updated_at, \
-                     COALESCE((SELECT MAX(created_at) FROM project_events WHERE project_id = projects.id), \
-                              updated_at) AS last_active_at \
+                     MAX(\
+                       IFNULL((SELECT MAX(updated_at) FROM tasks WHERE project_id = projects.id AND deleted_at IS NULL), 0), \
+                       IFNULL((SELECT MAX(created_at) FROM project_events WHERE project_id = projects.id), 0), \
+                       updated_at\
+                     ) AS last_active_at \
                      FROM projects WHERE deleted_at IS NULL AND archived = 0 \
                      ORDER BY last_active_at DESC, created_at DESC",
                 )
@@ -840,4 +846,62 @@ mod tests {
         assert_eq!(after.name, "原项目", "回收站项目内容不应被改写");
         assert_eq!(after.updated_at, current.updated_at, "updated_at 不应抖动");
     }
+
+    #[tokio::test]
+    async fn list_active_with_activity_orders_by_task_activity_over_created_at() {
+        // 回归(事故形状):项目 A created_at 较晚(创建序排前)但无任务/停更;项目 B
+        // created_at 较早但有一条 tasks.updated_at 最新 → 活跃序 B 排 A 前
+        // (任务活跃为主信号,治创建序误导 AI 把停摆项目当最高优先锚点)。
+        use crate::crud::task_repo::TaskRepo;
+        use crate::models::TaskRecord;
+        use df_types::types::TaskStatus;
+
+        let db = Database::open_in_memory().await.expect("open_in_memory");
+        let projects = ProjectRepo::new(&db);
+        let tasks = TaskRepo::new(&db);
+
+        // B 先建(早),A 后建(晚)——创建序会排 A 前
+        let mut early = prec("pb", "早建但活跃");
+        early.created_at = "1700000000000".to_string();
+        early.updated_at = "1700000000000".to_string();
+        let mut late = prec("pa", "晚建但停更");
+        late.created_at = "1700000010000".to_string();
+        late.updated_at = "1700000010000".to_string();
+        projects.insert(early).await.unwrap();
+        projects.insert(late).await.unwrap();
+
+        // B 挂一条最新任务;A 无任务 → B 活跃时间取任务 updated_at,A 回退自身 updated_at
+        let task = TaskRecord {
+            id: "tb".to_string(),
+            project_id: "pb".to_string(),
+            title: "B 的任务".to_string(),
+            description: String::new(),
+            status: TaskStatus::InProgress,
+            priority: 0,
+            branch_name: None,
+            assignee: None,
+            workflow_def_id: None,
+            base_branch: None,
+            review_rounds: 0,
+            output_json: None,
+            idea_id: None,
+            queue: "todo".to_string(),
+            parent_id: None,
+            content_json: None,
+            module_id: None,
+            archived: false,
+            created_at: "1700000000000".to_string(),
+            updated_at: "1790000000000".to_string(),
+        };
+        tasks.insert(task).await.unwrap();
+
+        let listed = projects.list_active_with_activity().await.unwrap();
+        assert_eq!(listed.len(), 2);
+        assert_eq!(listed[0].record.id, "pb", "任务活跃最新的早建项目应排前");
+        assert_eq!(listed[1].record.id, "pa", "无任务的晚建项目应排后");
+        assert_eq!(
+            listed[0].last_active_at, "1790000000000",
+            "活跃时间应取该项目任务 updated_at 最新值"
+        );
+    }
 }
diff --git a/src-tauri/src/commands/ai/commands/conversation.rs b/src-tauri/src/commands/ai/commands/conversation.rs
index 0aa4b8b..b01f25f 100644
--- a/src-tauri/src/commands/ai/commands/conversation.rs
+++ b/src-tauri/src/commands/ai/commands/conversation.rs
@@ -232,14 +232,24 @@ pub async fn ai_conversation_create(
 ///
 /// 截断按组独立:活跃/归档各自取 limit,归档项不挤占活跃名额(侧栏归档分组数据源依赖
 /// include_archived=true,见 useAiConversations.loadConversations)。
+///
+/// F4-1(2026-10-17)分页:active_offset/archived_offset 按组跳过前 N 条(默认 0=第一页,
+/// 旧调用方零改动)。治「会话列表+搜索硬上限 50」——实测 115 条活跃主会话 57% 在桌面
+/// 端零入口;侧栏首屏仍一页(50/组)保启动性能,「加载更多」按 offset 续拉,搜索前前端
+/// 逐页拉满(搜索跑在已加载列表上,截断列表=搜索失灵)。offset 超组长度返回该组空页
+/// (自然收敛,无需显式 has_more 字段:页内某组满额 limit 即可能还有下一页)。
 #[tauri::command]
 pub async fn ai_conversation_list(
     state: State<'_, AppState>,
     limit: Option<usize>,
     include_archived: Option<bool>,
+    active_offset: Option<usize>,
+    archived_offset: Option<usize>,
 ) -> Result<Vec<serde_json::Value>, String> {
     let limit = limit.unwrap_or(50);
     let include_archived = include_archived.unwrap_or(false);
+    let active_offset = active_offset.unwrap_or(0);
+    let archived_offset = archived_offset.unwrap_or(0);
     let records = state.ai_conversations.list_all().await.map_err(err_str)?;
     // list_all 已按 created_at DESC(最新在前）；默认排除归档 + 截断 limit。
     // origin 非空(子代理轨迹,V44)不进侧栏列表——轨迹在库可追溯,展示面只留主会话。
@@ -273,20 +283,25 @@ pub async fn ai_conversation_list(
                 "loop_state": r.loop_state,
             })
         }).collect();
-    // 分组截断:活跃/归档各自独立取 limit,防归档项挤占活跃名额(前端侧栏活跃列表被截空)。
-    // 原单 take(limit) 在 include_archived=true 时归档项混排会占满 50 名额。
-    let mut active = Vec::with_capacity(limit);
-    let mut archived = Vec::with_capacity(limit);
-    for s in summaries {
-        if s.get("archived").and_then(|v| v.as_bool()) == Some(true) {
-            if archived.len() < limit { archived.push(s); }
-        } else if active.len() < limit {
-            active.push(s);
-        }
-    }
-    // 活跃在前(侧栏主区),归档在后(归档分组);各自保持 created_at DESC 原序。
-    active.extend(archived);
-    Ok(active)
+    Ok(paginate_by_group(summaries, limit, active_offset, archived_offset))
+}
+
+/// F4-1:摘要按组独立分页——活跃/归档各自 skip(offset)+take(limit),活跃在前归档在后,
+/// 组内保持传入原序(created_at DESC)。offset=0 与旧「分组截断」口径逐字节等价
+/// (回归保护见 paginate_first_page_matches_legacy_grouped_truncation)。
+/// 抽纯函数供单测:分组正确性/offset 续拉/越界空页/归档不挤占活跃名额。
+fn paginate_by_group(
+    summaries: Vec<serde_json::Value>,
+    limit: usize,
+    active_offset: usize,
+    archived_offset: usize,
+) -> Vec<serde_json::Value> {
+    let (active, archived): (Vec<_>, Vec<_>) = summaries
+        .into_iter()
+        .partition(|s| s.get("archived").and_then(|v| v.as_bool()) != Some(true));
+    let mut page: Vec<serde_json::Value> = active.into_iter().skip(active_offset).take(limit).collect();
+    page.extend(archived.into_iter().skip(archived_offset).take(limit));
+    page
 }
 
 /// B3(无人值守可靠性专项):列出可恢复会话(loop_state='parked')。
@@ -391,11 +406,25 @@ pub(crate) async fn recover_parked_inner(
     .await
 }
 
+/// 恢复产物前缀:恢复链自身生成的引导消息以它开头(恢复产物非用户意图,任何情况下
+/// 都不该作为有效目标回放——2026-10-15 目标语义重构约定)。递归套娃根因:恢复消息
+/// 经 begin_turn(chat.rs:950/2461)被记为 WorkingContext Active 目标落库,下次恢复
+/// 回放时被当有效目标嵌套进新恢复消息,逐层加深(实证 conv 76730ce8 seq 250:
+/// 目标文本=上一条恢复消息原文截 200 字符)。防线收敛在恢复链单点:
+/// parse_pinned_goals 反解剔除 + build_recovery_message 回放剔除,同判据。
+pub(crate) const RECOVERY_ARTIFACT_PREFIX: &str = "[系统恢复]";
+
+/// 恢复产物判定:目标文本以恢复前缀开头即恢复链自产物,不作为有效目标。
+fn is_recovery_artifact(text: &str) -> bool {
+    text.starts_with(RECOVERY_ARTIFACT_PREFIX)
+}
+
 /// 目标快照:按最终有效意图生成 parked 恢复引导消息。
 ///
 /// 「按快照续不重放历史目标流」:取 WorkingContext.goals 与 pinned_goals 中仍 Active 的
-/// 目标 + unresolved 未解项;中途被调整作废(Cancelled)或已完成(Completed)的目标不复活。
-/// 快照全空(旧会话无快照)回退 [`PARKED_RECOVERY_MESSAGE`],行为同今日。
+/// 目标 + unresolved 未解项;中途被调整作废(Cancelled)或已完成(Completed)的目标不复活;
+/// 恢复产物([系统恢复] 前缀)剔除——回放它即递归套娃,历史污染目标在此单点拦截。
+/// 快照全空(旧会话无快照/全被过滤)回退 [`PARKED_RECOVERY_MESSAGE`],行为同今日。
 /// 不另立存储:快照即 working_context_json + pinned_goals 两列的语义化读取(单一真值来源)。
 pub(crate) fn build_recovery_message(
     working_context: &df_ai::context_helpers::WorkingContext,
@@ -405,12 +434,18 @@ pub(crate) fn build_recovery_message(
     let mut seen = std::collections::HashSet::new();
     let mut active: Vec<&str> = Vec::new();
     for g in &working_context.goals {
-        if matches!(g.status, GoalStatus::Active) && seen.insert(g.text.clone()) {
+        if matches!(g.status, GoalStatus::Active)
+            && !is_recovery_artifact(&g.text)
+            && seen.insert(g.text.clone())
+        {
             active.push(&g.text);
         }
     }
     for g in pinned_goals {
-        if matches!(g.status, crate::commands::ai::GoalStatus::Active) && seen.insert(g.text.clone()) {
+        if matches!(g.status, crate::commands::ai::GoalStatus::Active)
+            && !is_recovery_artifact(&g.text)
+            && seen.insert(g.text.clone())
+        {
             active.push(&g.text);
         }
     }
@@ -442,7 +477,8 @@ pub(crate) fn build_recovery_message(
 
 /// pinned_goals 落库反解(新旧格式兼容):新格式 [{text,status}] 保 status,
 /// 旧格式 ["text"] 转 Active;非数组/坏 JSON 一律空(降级不炸)。switch 恢复与
-/// parked 恢复共用(DRY,单一事实源)。
+/// parked 恢复共用(DRY,单一事实源)。恢复产物([系统恢复] 前缀)剔除——存量
+/// 污染目标不进内存,switch 载入后下轮 save 重写 pinned_goals 即自愈落库。
 pub(crate) fn parse_pinned_goals(raw: Option<&str>) -> Vec<crate::commands::ai::GoalEntry> {
     let parsed: Option<serde_json::Value> = raw.and_then(|s| serde_json::from_str(s).ok());
     match parsed {
@@ -468,6 +504,7 @@ pub(crate) fn parse_pinned_goals(raw: Option<&str>) -> Vec<crate::commands::ai::
                     Some(crate::commands::ai::GoalEntry::new(text))
                 }
             })
+            .filter(|e| !is_recovery_artifact(&e.text))
             .collect(),
         _ => vec![],
     }
@@ -1477,6 +1514,73 @@ mod tests {
         assert_eq!(back.timestamp, Some(123));
     }
 
+    // ============================================================
+    // F4-1 会话列表分页(paginate_by_group:活跃/归档各自 offset+limit)
+    // ============================================================
+
+    fn conv_val(id: &str, archived: bool) -> serde_json::Value {
+        serde_json::json!({ "id": id, "archived": archived })
+    }
+
+    /// 首页(offset=0)与旧「分组截断」口径等价:各组前 limit 条,活跃在前归档在后。
+    #[test]
+    fn paginate_first_page_matches_legacy_grouped_truncation() {
+        // 60 活跃 + 60 归档,limit 50:旧口径各取 50(归档不挤占活跃名额)
+        let summaries: Vec<_> = (0..60).map(|i| conv_val(&format!("a{i}"), false))
+            .chain((0..60).map(|i| conv_val(&format!("r{i}"), true)))
+            .collect();
+        let page = paginate_by_group(summaries, 50, 0, 0);
+        assert_eq!(page.len(), 100, "各组独立取满 50");
+        assert!(page.iter().take(50).all(|s| s["archived"] == false), "前 50 全活跃");
+        assert!(page.iter().skip(50).all(|s| s["archived"] == true), "后 50 全归档");
+        assert_eq!(page[0]["id"], "a0");
+        assert_eq!(page[49]["id"], "a49");
+        assert_eq!(page[50]["id"], "r0");
+        assert_eq!(page[99]["id"], "r49");
+    }
+
+    /// offset 续拉:第二页从各组 offset 处继续,两组独立推进互不影响
+    /// (115 活跃场景:第一页 a0..a49,第二页 active_offset=50 拉 a50..a99)。
+    #[test]
+    fn paginate_offset_advances_each_group_independently() {
+        let summaries: Vec<_> = (0..115).map(|i| conv_val(&format!("a{i}"), false))
+            .chain((0..3).map(|i| conv_val(&format!("r{i}"), true)))
+            .collect();
+        // 第二页:活跃 offset 50(满页 50 条),归档 offset 0(仅 3 条全回)
+        let page2 = paginate_by_group(summaries.clone(), 50, 50, 0);
+        assert_eq!(page2.len(), 53);
+        assert_eq!(page2[0]["id"], "a50");
+        assert_eq!(page2[49]["id"], "a99");
+        assert_eq!(page2[50]["id"], "r0", "归档组从 0 起独立分页,不被活跃 offset 影响");
+        // 第三页:活跃 offset 100(尾页 15 条),归档 offset 3(越界空)
+        let page3 = paginate_by_group(summaries, 50, 100, 3);
+        assert_eq!(page3.len(), 15);
+        assert_eq!(page3[0]["id"], "a100");
+        assert_eq!(page3[14]["id"], "a114");
+        assert!(page3.iter().all(|s| s["archived"] == false), "归档越界返回该组空页");
+    }
+
+    /// offset 超组长度的组返回空页(自然收敛);两组同时越界返回空 Vec。
+    #[test]
+    fn paginate_offset_beyond_length_yields_empty() {
+        let summaries = vec![conv_val("a0", false), conv_val("r0", true)];
+        assert!(paginate_by_group(summaries.clone(), 50, 5, 0).iter().all(|s| s["archived"] == true));
+        assert!(paginate_by_group(summaries.clone(), 50, 0, 5).iter().all(|s| s["archived"] == false));
+        assert!(paginate_by_group(summaries, 50, 5, 5).is_empty());
+    }
+
+    /// 输入混排(归档穿插在活跃间,list_all 全量 DESC 实际形态)时 partition 仍按组归位,
+    /// 组内保持传入原序。
+    #[test]
+    fn paginate_partition_from_interleaved_input() {
+        let summaries = vec![
+            conv_val("a0", false), conv_val("r0", true), conv_val("a1", false), conv_val("r1", true),
+        ];
+        let page = paginate_by_group(summaries, 50, 0, 0);
+        let ids: Vec<&str> = page.iter().map(|s| s["id"].as_str().unwrap()).collect();
+        assert_eq!(ids, vec!["a0", "a1", "r0", "r1"], "活跃全量在前(原序),归档全量在后(原序)");
+    }
+
     // ============================================================
     // F-260619-03 批次 B 端到端 round-trip(写路径全链:映射 → replace → list → 还原)
     // ============================================================
@@ -1789,4 +1893,72 @@ mod tests {
         assert!(parse_pinned_goals(Some("not-json")).is_empty());
         assert!(parse_pinned_goals(None).is_empty());
     }
+
+    // ── 恢复产物过滤(递归套娃防线,conv 76730ce8 seq 250 实证) ──
+
+    /// 构造一条上一代恢复消息全文(截 200 字符,复刻 begin_turn 误记目标的现场形态)。
+    fn prev_recovery_text() -> String {
+        let full = format!(
+            "{} 应用重启导致上次任务中断,已完成进度已保留。请按以下目标快照继续完成未竟任务:\n\n有效目标:\n- 检查 各个 u-claw 实例状态\n\n若快照不足以判断进度,先向用户确认已完成步骤再继续。",
+            RECOVERY_ARTIFACT_PREFIX
+        );
+        full.chars().take(200).collect()
+    }
+
+    #[test]
+    fn recovery_message_zero_nesting_with_polluted_goals() {
+        use df_ai::context_helpers::GoalStatus;
+        // 嵌套恢复现场:WorkingContext.goals 与 pinned_goals 都躺着上一条恢复消息全文,
+        // 两侧 Active——未过滤时它会被当有效目标嵌进新恢复消息(套娃逐层加深)。
+        let wc = wc_with_goals(&[
+            (&prev_recovery_text(), GoalStatus::Active),
+            ("交付 v2 登录接口", GoalStatus::Active),
+        ]);
+        let pinned = vec![
+            crate::commands::ai::GoalEntry::new(prev_recovery_text()),
+            crate::commands::ai::GoalEntry::new("补齐单测覆盖".to_string()),
+        ];
+        let msg = build_recovery_message(&wc, &pinned);
+        // 输出只允许保留消息头一处恢复前缀(零套娃),正常用户目标不受影响
+        assert_eq!(
+            msg.matches(RECOVERY_ARTIFACT_PREFIX).count(),
+            1,
+            "恢复产物目标必须被剔除,输出仅消息头一处前缀: {msg}"
+        );
+        assert!(msg.contains("交付 v2 登录接口"), "正常用户目标须保留: {msg}");
+        assert!(msg.contains("补齐单测覆盖"), "正常钉扎目标须保留: {msg}");
+        assert!(
+            !msg.contains("有效目标:\n- [系统恢复]"),
+            "上一代恢复消息不得作为目标行回放: {msg}"
+        );
+    }
+
+    #[test]
+    fn recovery_message_all_polluted_falls_back_to_constant() {
+        use df_ai::context_helpers::GoalStatus;
+        // 快照只剩恢复产物(历史最深套娃态):全部过滤后回退常量,套娃链在此截断
+        let wc = wc_with_goals(&[(&prev_recovery_text(), GoalStatus::Active)]);
+        let pinned = vec![crate::commands::ai::GoalEntry::new(prev_recovery_text())];
+        assert_eq!(
+            build_recovery_message(&wc, &pinned),
+            PARKED_RECOVERY_MESSAGE,
+            "全污染快照过滤后按空快照回退常量"
+        );
+    }
+
+    #[test]
+    fn parse_pinned_goals_drops_recovery_artifacts_idempotent() {
+        // 存量污染反解剔除:恢复前缀条目不进内存(幂等——清洗后再解析结果不变)
+        let polluted = serde_json::to_string(&vec![
+            crate::commands::ai::GoalEntry::new(prev_recovery_text()),
+            crate::commands::ai::GoalEntry::new("正常目标保留".to_string()),
+        ])
+        .unwrap();
+        let parsed = parse_pinned_goals(Some(&polluted));
+        assert_eq!(parsed.len(), 1, "恢复产物条目须剔除: {parsed:?}");
+        assert_eq!(parsed[0].text, "正常目标保留");
+        // 幂等:清洗结果重解析不变
+        let cleaned = serde_json::to_string(&parsed).unwrap();
+        assert_eq!(parse_pinned_goals(Some(&cleaned)).len(), 1);
+    }
 }
diff --git a/src-tauri/src/commands/ai/prompt.rs b/src-tauri/src/commands/ai/prompt.rs
index f0a3dec..f2115e3 100644
--- a/src-tauri/src/commands/ai/prompt.rs
+++ b/src-tauri/src/commands/ai/prompt.rs
@@ -127,6 +127,7 @@ fn system_prompt_parts(lang: &str) -> (String, &'static str, &'static str) {
                  - {}\n\
                  - {}\n\
                  - {}\n\
+                 - Response style: lead with the conclusion and the action/result; omit restating the user's question, pleasantries, and summary closers; prefer bullet points over paragraphs; do not expand on what was not asked. Explain whenever explanation is needed — cut filler, not substance.\n\
                  ## Focus\n\
                  - Always center your response on the core goal of the user's current request; the previous round's topic is only background, not the current task.\n\
                  - When the user switches topics (a clear new intent), follow the latest request; do not drag the old topic into the new answer.\n\
@@ -160,6 +161,7 @@ fn system_prompt_parts(lang: &str) -> (String, &'static str, &'static str) {
                  - {}\n\
                  - {}\n\
                  - {}\n\
+                 - 回复风格：结论先行，直接给动作/结果；不复述用户问题，不做客套开场和总结性收尾；要点优先于段落；用户没问的不展开。需要解释时照常解释——省的是空话，不是内容。\n\
                  ## 聚焦准则\n\
                  - 始终围绕用户当前请求的核心目标回答；上一轮的主题只是背景，不是当前任务。\n\
                  - 用户切换话题（明显的新意图）时，以最新请求为准，不要把旧话题带进新回答。\n\
@@ -445,7 +447,10 @@ async fn entity_fingerprint(db: &Arc<Database>, sql: &'static str) -> Option<Str
 /// - 日期:env_profile_line 注入当日日期,跨天必须失效(否则缓存串日期过期)。
 /// - lang:system_prompt_parts / 各策略段 / 注明语随 lang 分支。
 /// - custom_prompt:与 build 内读的同一设置,变更即失效。
-/// - 活跃 project/task 的 id+updated_at:新增/软删/改 updated_at 均反映到指纹(比 MAX 更稳)。
+/// - 活跃 project/task 指纹:task 取 id+updated_at;project 取 id+活跃时间(三源取大:
+///   tasks.updated_at 最新/project_events 最新 created_at/projects.updated_at,仅未归档未删除,
+///   与「当前项目」清单的 list_active_with_activity 排序输入对齐)——任务更新/归档变化均
+///   反映到指纹,清单排序变了缓存串必失效。
 /// - 去重排除集:被@实体走 augmentation 精准投影,排除集变化会改变清单内容,须纳入指纹。
 async fn system_prompt_fingerprint(
     state: &AppState,
@@ -463,7 +468,11 @@ async fn system_prompt_fingerprint(
         .unwrap_or_default();
     let proj_fp = entity_fingerprint(
         &state.db,
-        "SELECT id, updated_at FROM projects WHERE deleted_at IS NULL",
+        "SELECT p.id, MAX(\
+           IFNULL((SELECT MAX(t.updated_at) FROM tasks t WHERE t.project_id = p.id AND t.deleted_at IS NULL), 0), \
+           IFNULL((SELECT MAX(e.created_at) FROM project_events e WHERE e.project_id = p.id), 0), \
+           IFNULL(p.updated_at, 0)\
+         ) FROM projects p WHERE p.deleted_at IS NULL AND p.archived = 0",
     )
     .await?;
     let task_fp = entity_fingerprint(
@@ -586,24 +595,25 @@ async fn build_context_lists(
 ) -> String {
     let (_, proj_label, task_label) = system_prompt_parts(lang);
     let mut out = String::new();
-    // 附加当前数据上下文
-    if let Ok(projects) = state.projects.list_active().await {
+    // 附加当前数据上下文。项目段按最近活跃排序(任务活跃为主信号),治创建序误导项目绑定
+    // (创建序会把早停摆项目排在清单首位,AI 误当最高优先项目锚点)。
+    if let Ok(projects) = state.projects.list_active_with_activity().await {
         if !projects.is_empty() {
             out.push_str(proj_label);
             // system prompt 前缀克制：仅最近 20 个项目，防 context 膨胀
             // 去重:被 @ 的项目跳过(已有 augmentation 精准投影,清单再现致同一实体两次入 prompt)
             for p in projects.iter().take(20) {
-                if excluded_project_ids.iter().any(|id| id == &p.id) {
+                if excluded_project_ids.iter().any(|id| id == &p.record.id) {
                     continue;
                 }
                 out.push_str(&format!(
                     "- {} (id: {}) ({}): {}\n",
-                    p.name,
-                    p.id,
-                    p.status.as_str(),
-                    truncate_desc(&p.description, CONTEXT_DESC_MAX_CHARS)
+                    p.record.name,
+                    p.record.id,
+                    p.record.status.as_str(),
+                    truncate_desc(&p.record.description, CONTEXT_DESC_MAX_CHARS)
                 ));
-                if let Some(ref dir) = p.path {
+                if let Some(ref dir) = p.record.path {
                     out.push_str(&format!("  目录: {}\n", dir));
                 }
             }
@@ -921,6 +931,32 @@ mod tests {
         assert!(prefix.contains("聚焦准则"));
     }
 
+    // M4 回复风格纪律:中文版行为准则含「回复风格」条款(反口水话,不反内容)
+    #[test]
+    fn system_prompt_zh_has_response_style_clause() {
+        let (prefix, _, _) = system_prompt_parts("zh-CN");
+        assert!(
+            prefix.contains("回复风格"),
+            "中文 system prompt 应含回复风格条款,实际: {}",
+            prefix
+        );
+        // 核心条款抽样:结论先行 + 「省的是空话,不是内容」的度(需要解释时照常解释)
+        assert!(prefix.contains("结论先行"));
+        assert!(prefix.contains("省的是空话，不是内容"));
+    }
+
+    // M4 回复风格纪律:英文版 Guidelines 含 Response style 条款
+    #[test]
+    fn system_prompt_en_has_response_style_clause() {
+        let (prefix, _, _) = system_prompt_parts("en");
+        assert!(
+            prefix.contains("Response style"),
+            "英文 system prompt 应含 Response style 条款,实际: {}",
+            prefix
+        );
+        assert!(prefix.contains("cut filler, not substance"));
+    }
+
     // 文档探索策略段(中/英)存在且含关键引导
     #[test]
     fn doc_exploration_strategy_section_present_zh_en() {
diff --git a/src-tauri/src/commands/ai/tools/file.rs b/src-tauri/src/commands/ai/tools/file.rs
index 706a1ce..464496a 100644
--- a/src-tauri/src/commands/ai/tools/file.rs
+++ b/src-tauri/src/commands/ai/tools/file.rs
@@ -1098,7 +1098,7 @@ pub fn register(
             }
 
             // output_mode 分派返回结构
-            let result = match output_mode {
+            let mut result = match output_mode {
                 "files_with_matches" => {
                     // 仅返命中文件路径列表(去重,顺序保留首次命中)
                     let files: Vec<String> = matches_out.iter()
@@ -1183,6 +1183,11 @@ pub fn register(
                     })
                 }
             };
+            // 0 命中「阴性存疑」提示:路径/glob 写法/大小写任一不对都静默 0 命中,
+            // LLM 拿空结果易直接当「不存在」下结论(实测假阴性被当事实断言 3 轮),强制附一行复核提示。
+            if total == 0 {
+                result["hint"] = serde_json::json!(negative_hit_hint("路径/glob 写法/大小写 -i"));
+            }
             Ok(result)
         }
     );
@@ -1228,10 +1233,13 @@ pub fn register(
             let mut out = serde_json::json!({
                 "path": path, "pattern": pattern, "results": page_results, "total": total, "has_more": has_more
             });
-            // total=0 时空结果引导(治 LLM 搜不到就放弃/盲目重试)。
+            // total=0 时空结果引导(治 LLM 搜不到就放弃/盲目重试)+ 阴性存疑提示(0 命中≠不存在)。
             // 注:search_files 已大小写不敏感(内部 pattern.to_lowercase + 包含匹配),文案勿写"换大小写"。
             if total == 0 {
-                out["hint"] = serde_json::json!("未找到匹配文件。可尝试: ① 放宽关键词(更短/字符更少); ② 用 grep 工具做跨文件内容正则搜索(pattern 支持正则, -i 大小写不敏感)。注: search_files 为文件名大小写不敏感包含匹配");
+                out["hint"] = serde_json::json!(format!(
+                    "{}\n未找到匹配文件。可尝试: ① 补 recursive=true 搜子目录(不传只搜当前层); ② 放宽关键词(更短/字符更少); ③ 用 grep 工具做跨文件内容正则搜索(pattern 支持正则, -i 大小写不敏感)。注: search_files 为文件名大小写不敏感包含匹配",
+                    negative_hit_hint("路径/recursive")
+                ));
             }
             Ok(out)
         }
@@ -1675,6 +1683,16 @@ fn file_not_found_error(path: &str) -> anyhow::Error {
     anyhow::anyhow!(msg)
 }
 
+/// 搜索工具(search_files/grep)0 命中「阴性存疑」提示。
+///
+/// 治勘察假阴性被当事实断言:search_files 未传 recursive 只搜根目录,
+/// 0 命中被 LLM 直接当「功能未实施」连续断言 3 轮。路径/glob/recursive 等范围
+/// 参数写错时同样静默 0 命中,故空结果强制附本提示让 LLM 复核后再下结论。
+/// 措辞克制一行;`scope` 按工具传入各自的复核点。
+fn negative_hit_hint(scope: &str) -> String {
+    format!("⚠ 0 命中不一定代表不存在——请复核搜索范围({scope})后再下结论")
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -2003,6 +2021,23 @@ mod tests {
         assert!(msg.contains("相近文件候选"), "应含相近文件候选,实际: {msg}");
         assert!(msg.contains("target.rs"), "候选应含 target.rs,实际: {msg}");
     }
+
+    // ============================================================
+    // negative_hit_hint — 搜索工具 0 命中阴性存疑提示(治假阴性被当事实断言)
+    // ============================================================
+
+    /// 两搜索工具 0 命中提示共用句式:声明「0 命中≠不存在」+ 按工具各自的复核点。
+    #[test]
+    fn negative_hit_hint_states_doubt_and_scope() {
+        // grep 侧复核点:路径/glob 写法/大小写
+        let grep_hint = negative_hit_hint("路径/glob 写法/大小写 -i");
+        assert!(grep_hint.contains("0 命中不一定代表不存在"), "应声明阴性存疑,实际: {grep_hint}");
+        assert!(grep_hint.contains("复核搜索范围"), "应引导复核,实际: {grep_hint}");
+        assert!(grep_hint.contains("glob"), "grep 侧应含 glob 复核点,实际: {grep_hint}");
+        // search_files 侧复核点:recursive(实测根因是未传 recursive 只搜根目录)
+        let search_hint = negative_hit_hint("路径/recursive");
+        assert!(search_hint.contains("recursive"), "search_files 侧应含 recursive 复核点,实际: {search_hint}");
+    }
 }
 
 
diff --git a/src/api/types.ts b/src/api/types.ts
index 4ccd0c3..7331941 100644
--- a/src/api/types.ts
+++ b/src/api/types.ts
@@ -571,7 +571,7 @@ export interface AiMessage {
    * 消息来源(958 溯源/961 轮次归属):undefined=null=桌面本端发起;'inject'=运行中
    * 补充注入(961② 轮次分桶判据,归属当前轮不开新轮);'feishu'=飞书入站发起。
    */
-  source?: { channel: string; message_id: string; chat_id?: string; sender?: string } | null
+  source?: { channel: string; message_id?: string; chat_id?: string; sender?: string } | null
   /**
    * 推理过程文本(DeepSeek reasoning_content / Anthropic thinking 同通道,后端落库
    * ai_messages.reasoning_content)。undefined=该轮无思考内容。历史消息 reload 时
diff --git a/src/composables/ai/useAiConversations.ts b/src/composables/ai/useAiConversations.ts
index f8f41da..b98c478 100644
--- a/src/composables/ai/useAiConversations.ts
+++ b/src/composables/ai/useAiConversations.ts
@@ -4,7 +4,8 @@
 //! - 多处调 useAiEvents.notifyConversationChanged
 //! - newConversation/switchConversation 写 appSettings 持久化活跃会话 id
 
-import { ref } from 'vue'
+import { ref, reactive } from 'vue'
+import { invoke } from '@tauri-apps/api/core'
 import { getCurrentWebviewWindow } from '@tauri-apps/api/webviewWindow'
 import { aiApi } from '@/api'
 import { useAppSettingsStore } from '@/stores/appSettings'
@@ -14,7 +15,7 @@ import { setStreaming } from './streamingGuard'
 import { emitDetachedOpenConversation, registerDetachedOpenConvListener } from './useAiWindow'
 import { nextMsgId, getConvState, clearConvStreamState, clearLastDelta, startApprovalTimer, clearAllApprovalTimers, switchingConvs, getConvStreamState, setConvCurrentText, notifyConversationChanged, setConvPendingApprovals, discardConvPendingApprovals, backfillConvAgentFinished, bumpToolEpoch } from './aiShared'
 import { t } from '@/i18n/i18n-helpers'
-import type { AiConversationDetail, AiMessage, AiToolCallInfo, AiToolCallStatus, ConvId, MessageId } from '@/api/types'
+import type { AiConversationDetail, AiConversationSummary, AiMessage, AiToolCallInfo, AiToolCallStatus, ConvId, MessageId } from '@/api/types'
 
 // ── V11c(2026-10-05)活跃会话 KV 按窗口域分流 ──
 // 主窗与分离窗是独立 JS context,却共用同一 SQLite KV:两窗各自切会话都写 'df-ai-active-conv'
@@ -35,17 +36,78 @@ export function activeConvKvKey(): string {
 
 const appSettings = useAppSettingsStore()
 
-/** 拉取会话列表;首次加载时若 appSettings 仍有有效 active-conv 且当前无活跃对话则恢复 */
+// ============================================================
+// F4-1(2026-10-17)会话列表分页:后端 ai_conversation_list 活跃/归档各自
+// offset+limit 分页(每页 CONV_PAGE_SIZE/组),首屏一页保启动性能;侧栏「加载更多」
+// 按游标续拉,搜索输入时拉满(搜索过滤跑在已加载列表上,截断列表=搜索失灵)。
+// 治实测 115 条活跃主会话 57% 在桌面端零入口(硬 50 截断)。
+// ============================================================
+
+/** 每页条数(每组独立;对齐后端 limit.unwrap_or(50) 首屏口径) */
+const CONV_PAGE_SIZE = 50
+/** 单次「拉满」安全页数上限(防异常数据下无限循环;50*40=2000/组远超真实规模) */
+const CONV_MAX_PAGES = 40
+
+/** 分页游标(模块级单例,主/分离窗各自实例):已拉取条数按组计数 + 是否还有下一页。
+ *  reactive 供侧栏「加载更多」按钮直接绑显隐/loading 态。 */
+export const convListPaging = reactive({
+  activeLoaded: 0,
+  archivedLoaded: 0,
+  hasMore: false,
+  loading: false,
+})
+
+/** 拉一页会话摘要。直接 invoke:api 层 listConversations 封装不带 offset 参数,
+ *  分页需透传 activeOffset/archivedOffset(composable 直连 invoke 有 useAiSend 等先例)。 */
+function fetchConvPage(activeOffset: number, archivedOffset: number): Promise<AiConversationSummary[]> {
+  return invoke<AiConversationSummary[]>('ai_conversation_list', {
+    limit: CONV_PAGE_SIZE,
+    includeArchived: true,
+    activeOffset,
+    archivedOffset,
+  })
+}
+
+/** 拉取会话列表;首次加载时若 appSettings 仍有有效 active-conv 且当前无活跃对话则恢复。
+ *  F4-1 分页:从第一页重拉到覆盖既有加载深度(convListPaging)或拉尽——事件后刷新
+ *  (scheduleConversationsRefresh)不把「加载更多」拉出的深度回缩成首屏 50;首启深度 0
+ *  即只拉一页,首屏性能不变。 */
 async function loadConversations() {
+  if (convListPaging.loading) return
+  convListPaging.loading = true
   try {
+    const depthActive = convListPaging.activeLoaded
+    const depthArchived = convListPaging.archivedLoaded
+    const accActive: AiConversationSummary[] = []
+    const accArchived: AiConversationSummary[] = []
+    const seen = new Set<string>()
+    let aOff = 0
+    let rOff = 0
     // includeArchived=true:侧栏归档分组(ConversationSidebar archivedConvs)数据源即本内存列表,
     // 若拉取时排除归档,重启后已归档会话不可见且无恢复入口(2026-10-05 实锤 bug)。
-    // 活跃/归档分组由侧栏 groupedActive/archivedConvs 各自分离渲染;后端已分组截断,归档不挤占活跃名额。
-    state.conversations = await aiApi.listConversations(50, true)
+    // 活跃/归档分组由侧栏 groupedActive/archivedConvs 各自分离渲染;后端分组分页,归档不挤占活跃名额。
+    for (let i = 0; i < CONV_MAX_PAGES; i++) {
+      const page = await fetchConvPage(aOff, rOff)
+      const act = page.filter(c => !c.archived)
+      const arc = page.filter(c => c.archived)
+      for (const c of act) if (!seen.has(c.id)) { seen.add(c.id); accActive.push(c) }
+      for (const c of arc) if (!seen.has(c.id)) { seen.add(c.id); accArchived.push(c) }
+      aOff += act.length
+      rOff += arc.length
+      // 某组满额一页即可能还有下一页(后端无显式 has_more,满额启发式同 load_more 消息分页)
+      convListPaging.hasMore = act.length === CONV_PAGE_SIZE || arc.length === CONV_PAGE_SIZE
+      if (!convListPaging.hasMore) break
+      // 已覆盖此前加载深度即止(刷新场景),首启深度 0 第一页后即止
+      if (aOff >= depthActive && rOff >= depthArchived) break
+    }
+    state.conversations = [...accActive, ...accArchived]
+    convListPaging.activeLoaded = aOff
+    convListPaging.archivedLoaded = rOff
     // 虚拟项保底:活跃会话未落库(懒创建——首条消息前 DB 无记录)时,侧栏显示虚拟占位项。
     // 后端 ai_conversation_create 仅生成 id 存内存不落库,首条消息发送后 save_conversation 才写库。
     // 此处 active 会话不在 DB 列表时补一个虚拟项,实现「点新建立即出现在侧栏,不发消息不落库」。
     // 首条消息落库后 AiCompleted/AiError 触发的 loadConversations 拉到同 id 真实记录,虚拟项被替代无重复。
+    // (虚拟项不计入分页游标——游标只数 DB 真实拉取条数。)
     if (state.activeConversationId
       && !state.conversations.some(c => c.id === state.activeConversationId)) {
       const now = String(Date.now())
@@ -94,6 +156,41 @@ async function loadConversations() {
       } as AiMessage)
       notifyConversationChanged()
     }
+  } finally {
+    convListPaging.loading = false
+  }
+}
+
+/** F4-1「加载更多」:从当前游标续拉一页并追加(不重拉首页,避免重复传输+闪烁);
+ *  返回是否实际新增条目(侧栏据此决定是否继续,数据漂移致空页时自然停)。 */
+async function loadMoreConversations(): Promise<boolean> {
+  if (convListPaging.loading || !convListPaging.hasMore) return false
+  convListPaging.loading = true
+  try {
+    const page = await fetchConvPage(convListPaging.activeLoaded, convListPaging.archivedLoaded)
+    const act = page.filter(c => !c.archived)
+    const arc = page.filter(c => c.archived)
+    const existing = new Set(state.conversations.map(c => c.id))
+    const fresh = page.filter(c => !existing.has(c.id))
+    if (fresh.length) state.conversations = [...state.conversations, ...fresh]
+    convListPaging.activeLoaded += act.length
+    convListPaging.archivedLoaded += arc.length
+    convListPaging.hasMore = act.length === CONV_PAGE_SIZE || arc.length === CONV_PAGE_SIZE
+    return fresh.length > 0
+  } catch (e) {
+    console.error('[AI] 加载更多会话失败:', e)
+    return false
+  } finally {
+    convListPaging.loading = false
+  }
+}
+
+/** F4-1 搜索全量兜底:把列表拉满。搜索过滤(stores.filteredConversations)跑在已加载
+ *  列表上,未拉满时老会话搜不到——搜索本就该扫全量,输入触发时一次拉尽。 */
+async function loadAllConversations(): Promise<void> {
+  for (let i = 0; i < CONV_MAX_PAGES; i++) {
+    if (!convListPaging.hasMore) break
+    if (!(await loadMoreConversations())) break
   }
 }
 
@@ -754,6 +851,8 @@ export function useAiConversations() {
   return {
     loadConversations,
     scheduleConversationsRefresh,
+    loadMoreConversations,
+    loadAllConversations,
     newConversation,
     switchConversation,
     deleteConversation,
@@ -768,4 +867,4 @@ export function useAiConversations() {
   }
 }
 
-export { loadConversations, scheduleConversationsRefresh, loadMoreHistory, getLoadMoreState }
+export { loadConversations, scheduleConversationsRefresh, loadMoreConversations, loadAllConversations, loadMoreHistory, getLoadMoreState }
diff --git a/src/composables/ai/useAiSend.ts b/src/composables/ai/useAiSend.ts
index c7d02bb..696f06b 100644
--- a/src/composables/ai/useAiSend.ts
+++ b/src/composables/ai/useAiSend.ts
@@ -409,6 +409,9 @@ async function sendMessage(text: string, skill?: string, forceMode = false, part
               role: 'user',
               content: text.trim(),
               timestamp: Date.now(),
+              // 与后端落库同标记(chat.rs 落库带 source.channel='inject'):本地气泡
+              // 不带则轮次分桶把它当「发起消息」多计一轮,重载后才归位
+              source: { channel: 'inject' },
             })
           }
           void emit('ai-toast', { msg: t('ai.injectMerged'), type: 'success' })
diff --git a/src/i18n/en/aiChat.ts b/src/i18n/en/aiChat.ts
index c7ec4a5..ce350a6 100644
--- a/src/i18n/en/aiChat.ts
+++ b/src/i18n/en/aiChat.ts
@@ -133,6 +133,8 @@ export default {
     queueSave: 'Save',
     queueCancel: 'Cancel',
     queueSendNow: 'Send now',
+    // UX walk panel-2 F2-1: send button title/aria (was mis-bound to queueSendNow)
+    send: 'Send',
     // 3c UX batch addition (I4 queue backfill: pull a queued item back into the input box)
     queueBackfill: 'Backfill',
 
@@ -185,6 +187,10 @@ export default {
     copyMsg: 'Copy message',
     copied: 'Copied',
     copyFailed: 'Copy failed',
+    // F1-1: code block copy button (injected at v-html render layer via decorateCodeHtml)
+    copyCode: 'Copy',
+    // F1-1: collapsed long-message summary toggle (expanded state reuses collapse key)
+    foldedSummary: '{chars} chars collapsed / ~{tokens} tokens — click to expand',
     // Copy bubble as image (html2canvas → clipboard; falls back to download)
     copyAsImage: 'Copy as image',
     copiedAsImage: 'Copied as image',
@@ -287,6 +293,8 @@ export default {
     recoverNow: 'Resume',
     viewHistory: 'View history messages',
     viewSummary: 'View context summary',
+    // UX walk panel-6 F6-1: inline hint when a history/summary jump misses (target inside a collapsed segment)
+    historyLocateInSegment: 'This message is inside a collapsed segment (archived/compressed). Expand the divider in the message flow first',
     viewContext: 'View AI context',
     contextSystemPrompt: 'System Prompt',
     contextEnvironment: 'Environment',
@@ -342,6 +350,8 @@ export default {
     editEmptyBlocked: 'Edit content is empty, submission cancelled',
     skillNoMatchQuery: 'No skill matches "{q}", send blocked',
     backfillOverwriteConfirm: 'The input box has an unsent draft; backfilling will overwrite it. Continue?',
+    // UX walk panel-2 F2-4: example-card overwrite confirm (aligned with backfill confirm)
+    exampleOverwriteConfirm: 'The input box has an unsent draft; the example will overwrite it. Continue?',
 
     // ── Outbox / scheduled messages (OutboxDrawer + ChatInput clock entry, design 2026-10-10 §3.4) ──
     outboxEntry: 'Scheduled messages',
diff --git a/src/i18n/zh-CN/aiChat.ts b/src/i18n/zh-CN/aiChat.ts
index e020ff9..9d6e9d1 100644
--- a/src/i18n/zh-CN/aiChat.ts
+++ b/src/i18n/zh-CN/aiChat.ts
@@ -133,6 +133,8 @@ export default {
     queueSave: '保存',
     queueCancel: '取消',
     queueSendNow: '立即发送',
+    // 走查-信息UX 面二 F2-1:发送钮 title/aria(原误绑 queueSendNow,语义是插队而非发送)
+    send: '发送',
     // 3c UX 批追加(I4 队列回填:取出排队项填回输入框重新编辑)
     queueBackfill: '回填',
 
@@ -186,6 +188,10 @@ export default {
     copyMsg: '复制消息',
     copied: '已复制',
     copyFailed: '复制失败',
+    // F1-1:代码块复制按钮(v-html 渲染层注入,decorateCodeHtml 传参)
+    copyCode: '复制',
+    // F1-1:长文折叠条带量按钮(展开态复用 collapse 键)
+    foldedSummary: '已折叠 {chars} 字符 / 约 {tokens} tokens，点击展开',
     // 气泡复制为图片(html2canvas 渲染 → 剪贴板;失败降级下载)
     copyAsImage: '复制为图片',
     copiedAsImage: '已复制为图片',
@@ -288,6 +294,8 @@ export default {
     recoverNow: '恢复',
     viewHistory: '查看历史消息',
     viewSummary: '查看上下文摘要',
+    // 走查-信息UX 面六 F6-1:历史/摘要项点击定位未命中(目标在折叠段内未渲染 DOM)的行内提示
+    historyLocateInSegment: '该消息在折叠段(已归档/已压缩)内,请先在消息流中展开对应分隔条',
     viewContext: '查看 AI 上下文',
     contextSystemPrompt: '系统提示',
     contextEnvironment: '环境',
@@ -343,6 +351,8 @@ export default {
     editEmptyBlocked: '编辑内容为空,已取消提交',
     skillNoMatchQuery: '未找到技能「{q}」,已阻止发送',
     backfillOverwriteConfirm: '输入框已有未发送草稿,回填将覆盖,确定继续?',
+    // 走查-信息UX 面二 F2-4:示例问题卡覆盖草稿二次确认(对齐回填确认语义)
+    exampleOverwriteConfirm: '输入框已有未发送草稿,点击示例将覆盖,确定继续?',
 
     // ── 待发箱/定时消息(OutboxDrawer + ChatInput ⏰ 入口,设计 2026-10-10 §3.4)──
     outboxEntry: '定时消息',