feat(storage): 历史压缩保留原消息，前端展示完整对话、LLM 用压缩摘要

将压缩流程从物理删除原消息改为标记位 is_compacted 分离两个视图：
- LLM 视角(load_messages_for_topic)：只看 is_compacted=0 的压缩摘要+新消息
- UI 视角(load_messages_for_topic_full)：看原消息(含 is_compacted=1)+新消息，排除摘要

新增 compact_topic_history：不删除原消息，仅打标记+插入摘要，
同事务内删除旧摘要避免累积。修复 token 统计因压缩丢失 usage 的问题。

src/command/handlers/get_current.rs

@@ -79,7 +79,7 @@ async fn handle_get_current_session(
     // 直读 DB 按话题加载消息（不依赖 SessionManager 内存，重启后也能正确获取历史）
     let messages = handler
         .store
-        .load_messages_for_topic(topic_id, Some(&topic.session_id))
+        .load_messages_for_topic_full(topic_id, Some(&topic.session_id))
         .map_err(|e| CommandError::new("LOAD_MESSAGES_ERROR", e.to_string()))?;
 
     let actual_message_count = messages.len();

src/command/handlers/save_topic.rs

@@ -254,7 +254,7 @@ async fn handle_save_topic(
 
     let messages = handler
         .store
-        .load_messages_for_topic(topic_id, Some(&topic_record.session_id))
+        .load_messages_for_topic_full(topic_id, Some(&topic_record.session_id))
         .map_err(|e| CommandError::new("LOAD_MESSAGES_ERROR", e.to_string()))?;
 
     tracing::debug!(

src/gateway/compaction.rs

@@ -52,17 +52,17 @@ pub(crate) async fn schedule_background_history_compaction(
         .compress_two_segment(&history, &provider_config)
         .await?;
 
-    // Replace only this topic's history in DB (not the entire session).
-    // This avoids clobbering other topics' messages during compaction.
+    // 保留原始消息（标记 is_compacted=1）+ 插入压缩摘要，不删除原消息，
+    // 从而让前端仍能展示完整原始对话，LLM 只看压缩后的精简历史。
     store
-        .replace_topic_history(&session_id, &topic_id, &compressed)
-        .map_err(|e| AgentError::Other(format!("replace_topic_history error: {}", e)))?;
+        .compact_topic_history(&session_id, &topic_id, &compressed)
+        .map_err(|e| AgentError::Other(format!("compact_topic_history error: {}", e)))?;
 
     tracing::info!(
         chat_id = %chat_id,
         topic_id = %topic_id,
         compressed_msg_count = compressed.len(),
-        "Two-segment compression committed"
+        "Two-segment compression committed (original messages retained)"
     );
 
     session_guard.reload_topic_history(&chat_id, &topic_id)?;

src/gateway/processor.rs

@@ -362,7 +362,7 @@ impl InboundProcessor {
                                 tokio::spawn(async move {
                                     // 从 DB 查询该 topic 的第一条用户消息作为描述生成的依据
                                     let first_user_message = store_clone
-                                        .load_messages_for_topic(&topic_id_clone, None)
+                                        .load_messages_for_topic_full(&topic_id_clone, None)
                                         .ok()
                                         .and_then(|msgs| {
                                             msgs.into_iter().find(|m| m.role == "user")

src/gateway/ws.rs

@@ -786,7 +786,7 @@ async fn send_topic_history(
     task_repository: &Arc<dyn TaskRepository>,
 ) -> Result<(), Box<dyn std::error::Error>> {
     // 加载话题消息，按 session_id 过滤，避免混入子智能体消息
-    let messages = store.load_messages_for_topic(topic_id, Some(session_id))?;
+    let messages = store.load_messages_for_topic_full(topic_id, Some(session_id))?;
 
     tracing::info!(topic_id = %topic_id, message_count = messages.len(), "Sending topic history");
 

src/storage/migrations.rs

@@ -71,6 +71,16 @@ pub(super) fn ensure_messages_schema(conn: &Connection) -> Result<(), StorageErr
         )?;
     }
 
+    // is_compacted: 1 表示该消息是被压缩消费掉的原始消息（前端可见、LLM 不可见）。
+    // 压缩摘要消息 is_compacted=0（LLM 可见），通过 system_context='history_compaction*'
+    // 在前端查询中被排除。保留的原消息（system_guards / 最新 user）is_compacted=0，不重复。
+    if !has_column(conn, "messages", "is_compacted")? {
+        add_column_if_missing(
+            conn,
+            "ALTER TABLE messages ADD COLUMN is_compacted INTEGER NOT NULL DEFAULT 0",
+        )?;
+    }
+
     // 创建 topic_id 索引（如果不存在）
     conn.execute(
         "CREATE INDEX IF NOT EXISTS idx_messages_topic_seq ON messages(topic_id, seq) WHERE topic_id IS NOT NULL",

src/storage/mod.rs

@@ -958,6 +958,107 @@ impl SessionStore {
         Ok(())
     }
 
+    /// 压缩该 topic 的历史：保留原始消息（标记 is_compacted=1，前端可见、LLM 不可见），
+    /// 并插入压缩摘要消息（is_compacted=0，LLM 可见，前端通过 system_context 过滤排除）。
+    ///
+    /// 与 `replace_topic_history` 的区别：不删除原消息，仅打标记，从而让前端仍能展示
+    /// 完整原始对话，同时 LLM 只看到压缩后的精简历史。
+    ///
+    /// `new_messages` 是 `compress_two_segment` 的输出，包含：
+    ///   - 保留原样的消息（system_guards / 最新 user，保留原 ID）
+    ///   - 压缩摘要消息（system_context = history_compaction_*，新 ID）
+    pub fn compact_topic_history(
+        &self,
+        session_id: &str,
+        topic_id: &str,
+        new_messages: &[ChatMessage],
+    ) -> Result<(), StorageError> {
+        let mut conn = self.pool.get()?;
+        let tx = conn.transaction_with_behavior(TransactionBehavior::Immediate)?;
+        let now = current_timestamp();
+
+        // 分离摘要消息与保留消息（保留消息携带原 ID，摘要消息是新构造的）
+        let (summaries, preserved): (Vec<&ChatMessage>, Vec<&ChatMessage>) = new_messages
+            .iter()
+            .partition(|m| {
+                m.system_context
+                    .as_deref()
+                    .map_or(false, |sc| sc.starts_with("history_compaction"))
+            });
+
+        // 先删除该 topic 下已有的旧压缩摘要（system_context LIKE 'history_compaction%'）。
+        // 旧摘要已被新摘要替代，保留它们只会累积垃圾行（前端和 LLM 都看不到，但占存储）。
+        tx.execute(
+            "DELETE FROM messages \
+             WHERE topic_id = ?1 AND session_id = ?2 \
+             AND system_context LIKE 'history_compaction%'",
+            params![topic_id, session_id],
+        )?;
+
+        // 将该 topic 中未被保留的原消息标记为 is_compacted=1（仅更新尚未标记的行，避免重复写）。
+        // 保留消息（system_guards / 最新 user）保持 is_compacted=0，不重复插入。
+        let preserved_ids: Vec<String> =
+            preserved.iter().map(|m| m.id.clone()).collect();
+        if preserved_ids.is_empty() {
+            tx.execute(
+                "UPDATE messages SET is_compacted = 1 \
+                 WHERE topic_id = ?1 AND session_id = ?2 AND is_compacted = 0",
+                params![topic_id, session_id],
+            )?;
+        } else {
+            let placeholders = (0..preserved_ids.len())
+                .map(|_| "?")
+                .collect::<Vec<_>>()
+                .join(",");
+            let sql = format!(
+                "UPDATE messages SET is_compacted = 1 \
+                 WHERE topic_id = ? AND session_id = ? AND is_compacted = 0 \
+                 AND id NOT IN ({})",
+                placeholders
+            );
+            let mut params_vec: Vec<String> = vec![topic_id.to_string(), session_id.to_string()];
+            params_vec.extend(preserved_ids.iter().cloned());
+            tx.execute(&sql, rusqlite::params_from_iter(params_vec))?;
+        }
+
+        // 插入压缩摘要消息（is_compacted=0，由列默认值保证）
+        let start_seq: i64 = tx.query_row(
+            "SELECT COALESCE(MAX(seq), 0) + 1 FROM messages WHERE session_id = ?1",
+            params![session_id],
+            |row| row.get(0),
+        )?;
+        for (i, message) in summaries.iter().enumerate() {
+            let seq = start_seq + i as i64;
+            insert_message_with_topic_seq(&tx, session_id, topic_id, seq, message)?;
+        }
+
+        // 更新 topic / session 计数（基于该 topic 全部消息，含被压缩的原始消息）
+        let topic_count: i64 = tx.query_row(
+            "SELECT COUNT(*) FROM messages WHERE session_id = ?1 AND topic_id = ?2",
+            params![session_id, topic_id],
+            |row| row.get(0),
+        )?;
+        tx.execute(
+            "UPDATE topics SET message_count = ?2, last_active_at = ?3, updated_at = ?3 WHERE id = ?1",
+            params![topic_id, topic_count, now],
+        )?;
+        let (total_count, user_turn_count): (i64, i64) = tx.query_row(
+            "SELECT COUNT(*), COALESCE(SUM(CASE WHEN role = 'user' THEN 1 ELSE 0 END), 0) \
+             FROM messages WHERE session_id = ?1",
+            params![session_id],
+            |row| Ok((row.get(0)?, row.get(1)?)),
+        )?;
+        tx.execute(
+            "UPDATE sessions SET message_count = ?2, user_turn_count = ?3, \
+             updated_at = ?4, last_active_at = ?4, archived_at = NULL \
+             WHERE id = ?1 AND deleted_at IS NULL",
+            params![session_id, total_count, user_turn_count, now],
+        )?;
+
+        tx.commit()?;
+        Ok(())
+    }
+
     pub fn mark_agent_prompt_reinjected(&self, session_id: &str) -> Result<(), StorageError> {
         let now = current_timestamp();
         let conn = self.pool.get()?;
@@ -1556,6 +1657,8 @@ impl SessionStore {
         load_messages_after(&conn, session_id, 0)
     }
 
+    /// LLM 视角：只返回 is_compacted = 0 的消息（压缩摘要 + 未被压缩的新消息）。
+    /// 被压缩消费掉的原始消息（is_compacted = 1）对 LLM 不可见，以节省 context。
     pub fn load_messages_for_topic(
         &self,
         topic_id: &str,
@@ -1563,12 +1666,56 @@ impl SessionStore {
     ) -> Result<Vec<ChatMessage>, StorageError> {
         let conn = self.pool.get()?;
 
+        if let Some(sid) = session_id {
+            let mut stmt = conn.prepare(
+                "
+                SELECT id, role, content, system_context, reasoning_content, media_refs_json, created_at, tool_call_id, tool_name, tool_calls_json, tool_duration_ms, prompt_tokens, completion_tokens, total_tokens, context_window_tokens
+                FROM messages
+                WHERE topic_id = ?1 AND session_id = ?2 AND is_compacted = 0
+                ORDER BY seq ASC
+                ",
+            )?;
+            let rows = stmt.query_map(params![topic_id, sid], map_chat_message_row)?;
+            let mut messages = Vec::new();
+            for row in rows {
+                messages.push(row?);
+            }
+            Ok(messages)
+        } else {
+            let mut stmt = conn.prepare(
+                "
+                SELECT id, role, content, system_context, reasoning_content, media_refs_json, created_at, tool_call_id, tool_name, tool_calls_json, tool_duration_ms, prompt_tokens, completion_tokens, total_tokens, context_window_tokens
+                FROM messages
+                WHERE topic_id = ?1 AND is_compacted = 0
+                ORDER BY seq ASC
+                ",
+            )?;
+            let rows = stmt.query_map(params![topic_id], map_chat_message_row)?;
+            let mut messages = Vec::new();
+            for row in rows {
+                messages.push(row?);
+            }
+            Ok(messages)
+        }
+    }
+
+    /// UI 视角：返回原始消息（含被压缩消费的 is_compacted=1 消息）+ 未压缩新消息，
+    /// 排除压缩摘要消息（system_context LIKE 'history_compaction%'）。
+    /// 用于前端历史展示、/current、/save-topic、topic 描述生成等场景。
+    pub fn load_messages_for_topic_full(
+        &self,
+        topic_id: &str,
+        session_id: Option<&str>,
+    ) -> Result<Vec<ChatMessage>, StorageError> {
+        let conn = self.pool.get()?;
+
         if let Some(sid) = session_id {
             let mut stmt = conn.prepare(
                 "
                 SELECT id, role, content, system_context, reasoning_content, media_refs_json, created_at, tool_call_id, tool_name, tool_calls_json, tool_duration_ms, prompt_tokens, completion_tokens, total_tokens, context_window_tokens
                 FROM messages
                 WHERE topic_id = ?1 AND session_id = ?2
+                  AND (system_context IS NULL OR system_context NOT LIKE 'history_compaction%')
                 ORDER BY seq ASC
                 ",
             )?;
@@ -1584,6 +1731,7 @@ impl SessionStore {
                 SELECT id, role, content, system_context, reasoning_content, media_refs_json, created_at, tool_call_id, tool_name, tool_calls_json, tool_duration_ms, prompt_tokens, completion_tokens, total_tokens, context_window_tokens
                 FROM messages
                 WHERE topic_id = ?1
+                  AND (system_context IS NULL OR system_context NOT LIKE 'history_compaction%')
                 ORDER BY seq ASC
                 ",
             )?;

src/storage/ports.rs

@@ -28,6 +28,14 @@ pub trait ConversationRepository: Send + Sync + 'static {
         session_id: Option<&str>,
     ) -> Result<Vec<ChatMessage>, StorageError>;
 
+    /// UI 视角：返回原始消息（含被压缩消费的 is_compacted=1 消息）+ 未压缩新消息，
+    /// 排除压缩摘要消息（system_context LIKE 'history_compaction%'）。
+    fn load_messages_for_topic_full(
+        &self,
+        topic_id: &str,
+        session_id: Option<&str>,
+    ) -> Result<Vec<ChatMessage>, StorageError>;
+
     fn append_message(&self, session_id: &str, message: &ChatMessage) -> Result<(), StorageError>;
 
     fn append_message_with_topic(
@@ -74,6 +82,15 @@ pub trait ConversationRepository: Send + Sync + 'static {
         topic_id: &str,
         messages: &[ChatMessage],
     ) -> Result<(), StorageError>;
+
+    /// 压缩该 topic 的历史：保留原始消息（标记 is_compacted=1）+ 插入压缩摘要。
+    /// 不删除原消息，让前端仍能展示完整原始对话，LLM 只看压缩后的精简历史。
+    fn compact_topic_history(
+        &self,
+        session_id: &str,
+        topic_id: &str,
+        new_messages: &[ChatMessage],
+    ) -> Result<(), StorageError>;
 }
 
 pub trait PromptInjectionRepository: Send + Sync + 'static {
@@ -272,6 +289,23 @@ impl ConversationRepository for super::SessionStore {
     ) -> Result<(), StorageError> {
         super::SessionStore::replace_topic_history(self, session_id, topic_id, messages)
     }
+
+    fn load_messages_for_topic_full(
+        &self,
+        topic_id: &str,
+        session_id: Option<&str>,
+    ) -> Result<Vec<ChatMessage>, StorageError> {
+        super::SessionStore::load_messages_for_topic_full(self, topic_id, session_id)
+    }
+
+    fn compact_topic_history(
+        &self,
+        session_id: &str,
+        topic_id: &str,
+        new_messages: &[ChatMessage],
+    ) -> Result<(), StorageError> {
+        super::SessionStore::compact_topic_history(self, session_id, topic_id, new_messages)
+    }
 }
 
 impl PromptInjectionRepository for super::SessionStore {