From 7e03e3cd5f050894f11953308cb3337da9346cb3 Mon Sep 17 00:00:00 2001
From: JOJO <1498581755@qq.com>
Date: Mon, 6 Apr 2026 15:12:25 +0800
Subject: [PATCH] feat: stabilize context compression and deep-compact flow

---
 AGENTS.md                                     | 130 ++++++--
 CLAUDE.md                                     |  14 +
 core/main_terminal_parts/context.py           |  15 +-
 modules/personalization_manager.py            |  12 +
 server/chat_flow_task_main.py                 |  74 ++++-
 server/chat_flow_tool_loop.py                 |  76 ++++-
 server/conversation.py                        |  91 +++++-
 server/deep_compression.py                    | 295 ++++++++++++++++++
 server/status.py                              |   9 +
 server/tasks.py                               |  37 ++-
 static/src/app/computed.ts                    |   4 +-
 static/src/app/methods/conversation.ts        |  49 ++-
 static/src/app/methods/message.ts             |  67 ++++
 static/src/app/methods/resources.ts           |  12 +
 static/src/app/methods/taskPolling.ts         |  48 ++-
 static/src/app/methods/ui.ts                  |   8 +
 static/src/app/state.ts                       |   4 +
 .../personalization/PersonalizationDrawer.vue |  40 +++
 static/src/stores/personalization.ts          |  10 +-
 utils/context_manager.py                      | 186 ++++++++++-
 utils/conversation_manager.py                 | 162 +++++++---
 21 files changed, 1224 insertions(+), 119 deletions(-)
 create mode 100644 server/deep_compression.py

diff --git a/AGENTS.md b/AGENTS.md
index 8030c3a..8828ff1 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -1,39 +1,103 @@
-# Repository Guidelines
+# Repository Guidelines (Code-Verified)
 
-## Project Structure & Module Organization
-- `main.py` runs the CLI agent; `web_server.py` serves the Flask/SocketIO UI from `static/`.
-- `core/` contains terminal orchestration (`main_terminal.py`, `web_terminal.py`) and should stay I/O-light.
-- `modules/` collects reusable capabilities (file/search/memory/terminal); extend here before touching entrypoints.
-- `utils/` provides API clients, context helpers, and logging; import from here instead of adding globals.
-- `prompts/`, `data/`, and `project/` bundle prompt templates, fixtures, and demos used primarily for reference.
-- `test/` stores integration helpers (`api_interceptor_server.py`, `test_deepseek_output.py`); add new suites alongside them.
-- 前端：`static/src/components/chat/monitor` 是虚拟显示器动画层；涉及新场景请在 `MonitorDirector`/`monitor store` 内对齐数据与动画的状态同步。
+> Last verified against current codebase: 2026-04-06
+> Scope: `/Users/jojo/Desktop/agents/正在修复中/agents`
 
-## Build, Test, and Development Commands
-- `pip install -r requirements.txt` installs Python dependencies.
-- `python main.py` launches the CLI agent loop.
-- `python web_server.py` serves the web UI at `http://localhost:8091`.
-- `python test/api_interceptor_server.py` starts the mock proxy that logs traffic to `api_logs/`.
-- `python test_deepseek_output.py` exercises streaming tool-calls; populate `config.py` with valid API credentials first.
-- `npm install` then `npm run build` builds the Vue frontend to `static/dist`; `npm run dev` watches; `npm run lint` for TS/Vue linting.
+这份文档基于当前仓库实际代码重写。若与未来代码冲突，以代码为准并及时更新本文档。
 
-## Coding Style & Naming Conventions
-- Target Python 3.11+, 4-space indentation, `snake_case` functions, and `PascalCase` classes.
-- TypeScript 5 / Vue 3 单文件组件，保持 `<script setup lang="ts">` 风格；新增资源放入 `static/src` 对应子目录。
-- Add type hints和简短双语 docstring；日志统一用 `utils.logger.setup_logger`，仅在需要时使用 `print`。
-- 模块聚焦：新增能力优先放 `modules/`（或 `static/src/components/chat/monitor` 对应区域）而非入口脚本。
+## 1) 当前项目结构（按代码现状）
 
-## Testing Guidelines
-- Prefer `pytest` for automation and name files `test/test_<feature>.py`, reusing assets in `data/` when relevant.
-- For network flows, run `python test/api_interceptor_server.py` and point the app to the proxy to avoid external calls.
-- Capture manual verification steps in PRs and add replay scripts under `test/` for repeated workflows.
+- **后端入口**
+  - `main.py`: 统一启动入口（当前默认走 Web 模式 + thinking_mode=True）
+  - `server/app.py`: 推荐的 Web 服务入口（封装并转发到 `server/app_legacy.py`）
+  - `web_server.py`: 兼容入口，已标记 deprecated，但仍可启动
+- **后端核心目录**
+  - `server/`: Flask + Socket.IO 业务主线（chat flow、conversation、tasks、status、api_v1 等）
+  - `core/`: 终端与工具编排（`main_terminal.py`、`web_terminal.py`、`main_terminal_parts/*`）
+  - `modules/`: 可复用能力模块（terminal/file/memory/sub_agent/upload_security/user 等）
+  - `config/`: 配置拆分（`api.py`, `limits.py`, `terminal.py`, `paths.py` ...），由 `config/__init__.py` 聚合并加载 `.env`
+  - `utils/`: API client、日志、上下文与对话工具等公共函数
+- **前端目录**
+  - `static/src/`: Vue 3 + TS 前端
+  - 监控动画相关核心文件：
+    - `static/src/components/chat/monitor/MonitorDirector.ts`
+    - `static/src/stores/monitor.ts`
+    - `static/src/components/chat/monitor/*`
+- **其他子项目/资源**
+  - `android-webview-app/`: Android WebView 客户端工程
+  - `easyagent/`: 独立前端子目录（与根目录前端并存）
 
-## Commit & Pull Request Guidelines
-- With no history yet, adopt Conventional Commits (`feat:`, `fix:`, `refactor:`) and keep summaries under 72 characters.
-- Squash iterative commits; PRs should outline changes, touched modules, and validation.
-- Link issues and attach UI captures or log snippets when behavior shifts.
+## 2) 当前可用启动/构建命令（已按代码核对）
 
-## Security & Configuration Tips
-- Keep real API keys out of git; override defaults via environment variables.
-- Clean sensitive data from `api_logs/`, `logs/`, and `experiment_outputs/` before sharing.
-- Extend `.gitignore` when new tooling produces temporary artifacts.
+### Python / 服务端
+- 安装依赖：`pip install -r requirements.txt`
+- 推荐启动 Web：`python -m server.app`
+- 可选参数：`python -m server.app --path ./project --port 8091 --debug --thinking-mode`
+- 兼容启动方式：`python web_server.py`
+- 统一入口：`python main.py`（当前实现会默认进入 Web 启动流程）
+
+### Frontend（根目录）
+- 安装依赖：`npm install`
+- 构建：`npm run build`
+- 开发监听（当前脚本是 build watch）：`npm run dev`
+- Lint：`npm run lint`
+
+## 3) 测试现状（不要再写过时命令）
+
+- 当前仓库内可见自动化冒烟：`test/test_server_refactor_smoke.py`（`unittest`）
+  - 运行方式：`python -m unittest test.test_server_refactor_smoke`
+- `test_system_message.py` 依赖外部 `MOONSHOT_API_KEY` 与网络，不属于离线稳定 CI 用例。
+- 当前仓库未发现 `pytest.ini`/`pyproject.toml`/`tox.ini`；不要默认要求 `pytest` 作为唯一入口。
+
+## 4) 代码修改约定（实用版）
+
+- **最小改动原则**：只改与需求直接相关文件，避免顺手重构。
+- **后端改动优先级**：先改 `modules/`、`server/` 内对应模块，最后才动入口。
+- **前端改动优先级**：按 `static/src` 现有分层改（`app/`、`stores/`、`components/`、`composables/`）。
+- 涉及 monitor 动画/事件联动时，至少同步检查：
+  - `MonitorDirector.ts`（动画与场景执行）
+  - `stores/monitor.ts`（事件队列与状态机）
+
+## 5) 风格与质量
+
+- Python：4 空格、`snake_case` 函数、`PascalCase` 类，新增公共函数尽量补 type hints。
+- Vue/TS：保持现有代码风格，不做无关风格清洗。
+- 日志：优先复用现有 logger/日志路径，不引入大量临时 `print`。
+- 提交前至少做与改动相关的最小验证（命令输出或手工步骤要可复现）。
+
+## 6) 提交与 PR
+
+- 使用 Conventional Commits：`feat:` `fix:` `refactor:` `chore:` ...
+- PR 描述至少包含：
+  - 改了哪些文件
+  - 为什么改
+  - 如何验证
+  - 风险/回滚点
+
+## 7) 安全与仓库卫生
+
+- 严禁提交真实密钥（`.env`、token、cookie、用户隐私）。
+- `logs/`, `data/`, `users/`, `project/` 为运行态目录（见 `.gitignore`），分享前需脱敏。
+- 不要把本地构建产物（如 `static/dist/`、`node_modules/`）纳入提交。
+
+## 8) Android App 发布联动要求（重要）
+
+当修改前端并需要发布 Android WebView App 时，必须同步执行以下步骤（依据 `doc/android_app_release_and_update.md`）：
+
+1. 同时更新版本信息  
+   - `android-webview-app/app/build.gradle.kts`：递增 `versionCode`，更新 `versionName`
+2. 同时更新更新说明  
+   - `android-webview-app/APP_CHANGELOG.md`：在顶部新增当前版本说明
+3. 完成修改后提醒用户运行上传脚本  
+   - 在发布流程中运行：`bash /Users/jojo/Desktop/agents/正在修复中/agents/upload_android_apk.sh`
+
+> 禁止只改前端代码而不更新版本号/更新说明，否则会导致客户端更新提示与分发信息不一致。
+
+## 9) 给 Agent 的硬性要求
+
+- 先读当前代码再执行，不依赖历史文档记忆。
+- 输出结果必须带：
+  1. 修改文件列表
+  2. 核心变更点
+  3. 验证结果（已执行命令/未执行原因）
+- 如果发现本文档过时，直接更新 `AGENTS.md` 并在结果中说明依据。
diff --git a/CLAUDE.md b/CLAUDE.md
index 7aa28c6..c551ddf 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -211,3 +211,17 @@ agents/
 - 容器镜像定制：编辑 `docker/terminal.Dockerfile` 并重新构建
 - 前端开发：`npm install && npm run dev` (在 `static/` 目录)
 - 查看详细执行流程：监控 `logs/debug_stream.log`
+
+## Android Release Notes (Required for frontend update)
+
+当改动前端并准备发 Android WebView 壳版本时，必须同步完成：
+
+1. 更新版本号（`android-webview-app/app/build.gradle.kts`）
+   - `versionCode` 递增
+   - `versionName` 更新
+2. 更新版本说明（`android-webview-app/APP_CHANGELOG.md`）
+   - 新版本条目写在顶部
+3. 修改完成后提醒用户运行上传脚本：
+   - `bash /Users/jojo/Desktop/agents/正在修复中/agents/upload_android_apk.sh`
+
+参考文档：`doc/android_app_release_and_update.md`
diff --git a/core/main_terminal_parts/context.py b/core/main_terminal_parts/context.py
index d4ed3dc..caba145 100644
--- a/core/main_terminal_parts/context.py
+++ b/core/main_terminal_parts/context.py
@@ -72,7 +72,7 @@ except ImportError:
 from modules.container_monitor import collect_stats, inspect_state
 from core.tool_config import TOOL_CATEGORIES
 from utils.api_client import DeepSeekClient
-from utils.context_manager import ContextManager
+from utils.context_manager import ContextManager, AUTO_SHALLOW_PLACEHOLDER
 from utils.tool_result_formatter import format_tool_result_for_context
 from utils.logger import setup_logger
 from config.model_profiles import (
@@ -147,6 +147,7 @@ class MainTerminalContextMixin:
             ]
 
             personalization_config = getattr(self.context_manager, "custom_personalization_config", None) or load_personalization_config(self.data_dir)
+            shallow_replace_enabled = bool(personalization_config.get("auto_shallow_compress_enabled", False)) if isinstance(personalization_config, dict) else False
             skills_catalog = get_skills_catalog()
             enabled_skills = merge_enabled_skills(
                 personalization_config.get("enabled_skills") if isinstance(personalization_config, dict) else None,
@@ -208,6 +209,7 @@ class MainTerminalContextMixin:
 
             # 添加对话历史（保留完整结构，包括tool_calls和tool消息）
             conversation = context["conversation"]
+            replaced_tool_count = 0
             for idx, conv in enumerate(conversation):
                 metadata = conv.get("metadata") or {}
                 if conv["role"] == "assistant":
@@ -226,6 +228,15 @@ class MainTerminalContextMixin:
                     messages.append(message)
 
                 elif conv["role"] == "tool":
+                    if shallow_replace_enabled and metadata.get("auto_shallow_compacted"):
+                        messages.append({
+                            "role": "tool",
+                            "content": AUTO_SHALLOW_PLACEHOLDER,
+                            "tool_call_id": conv.get("tool_call_id", ""),
+                            "name": conv.get("name", "")
+                        })
+                        replaced_tool_count += 1
+                        continue
                     # Tool消息需要保留完整结构
                     images = conv.get("images") or metadata.get("images") or []
                     videos = conv.get("videos") or metadata.get("videos") or []
@@ -281,6 +292,8 @@ class MainTerminalContextMixin:
                     "role": "system",
                     "content": disabled_notice
                 })
+            if shallow_replace_enabled:
+                print(f"[ContextCompression] build_messages 替换tool占位符: {replaced_tool_count} 条")
             return messages
 
     def load_prompt(self, name: str) -> str:
diff --git a/modules/personalization_manager.py b/modules/personalization_manager.py
index 14a3340..6f10471 100644
--- a/modules/personalization_manager.py
+++ b/modules/personalization_manager.py
@@ -42,6 +42,8 @@ DEFAULT_PERSONALIZATION_CONFIG: Dict[str, Any] = {
     "skill_hints_enabled": False,  # Skill 提示系统开关（默认关闭）
     "default_model": "kimi-k2.5",
     "image_compression": "original",  # original / 1080p / 720p / 540p
+    "auto_shallow_compress_enabled": False,
+    "auto_deep_compress_enabled": False,
     "silent_tool_disable": False,  # 禁用工具时不向模型插入提示
     "enhanced_tool_display": True,  # 增强工具显示
 }
@@ -186,6 +188,16 @@ def sanitize_personalization_payload(
     elif base.get("image_compression") not in allowed_image_modes:
         base["image_compression"] = "original"
 
+    if "auto_shallow_compress_enabled" in data:
+        base["auto_shallow_compress_enabled"] = bool(data.get("auto_shallow_compress_enabled"))
+    else:
+        base["auto_shallow_compress_enabled"] = bool(base.get("auto_shallow_compress_enabled", False))
+
+    if "auto_deep_compress_enabled" in data:
+        base["auto_deep_compress_enabled"] = bool(data.get("auto_deep_compress_enabled"))
+    else:
+        base["auto_deep_compress_enabled"] = bool(base.get("auto_deep_compress_enabled", False))
+
     # 静默禁用工具提示
     if "silent_tool_disable" in data:
         base["silent_tool_disable"] = bool(data.get("silent_tool_disable"))
diff --git a/server/chat_flow_task_main.py b/server/chat_flow_task_main.py
index 658a0e5..a8c44cf 100644
--- a/server/chat_flow_task_main.py
+++ b/server/chat_flow_task_main.py
@@ -129,6 +129,7 @@ from .chat_flow_runtime import (
 from .chat_flow_task_support import process_sub_agent_updates
 from .chat_flow_tool_loop import execute_tool_calls
 from .chat_flow_stream_loop import run_streaming_attempts
+from .deep_compression import run_deep_compression
 
 async def poll_sub_agent_completion(*, web_terminal, workspace, conversation_id, client_sid, username):
     """后台轮询子智能体完成状态，完成后触发新一轮对话"""
@@ -547,7 +548,10 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
         except Exception:
             personal_config = {}
         auto_title_enabled = personal_config.get("auto_generate_title", True)
-        if auto_title_enabled:
+        skip_auto_title_generation = bool(
+            (web_terminal.context_manager.conversation_metadata or {}).get("skip_auto_title_generation", False)
+        )
+        if auto_title_enabled and not skip_auto_title_generation:
             conv_id = getattr(web_terminal.context_manager, "current_conversation_id", None)
             socketio.start_background_task(
                 generate_conversation_title_background,
@@ -556,6 +560,55 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
                 message,
                 username
             )
+
+    # 自动深层压缩（用户输入后触发）
+    try:
+        personal_config = load_personalization_config(workspace.data_dir)
+    except Exception:
+        personal_config = {}
+    auto_deep_enabled = bool(personal_config.get("auto_deep_compress_enabled", False))
+    current_tokens_for_deep = web_terminal.context_manager.get_current_context_tokens(conversation_id)
+    if (
+        auto_deep_enabled
+        and current_tokens_for_deep > 150_000
+        and not web_terminal.context_manager.is_compression_in_progress()
+    ):
+        web_terminal.context_manager._set_meta_flag("is_ultra_long_conversation", True)
+        sender('compression_state', {
+            "conversation_id": conversation_id,
+            "in_progress": True,
+            "mode": "auto",
+            "stage": "queued"
+        })
+        deep_result = await run_deep_compression(
+            web_terminal=web_terminal,
+            workspace=workspace,
+            conversation_id=conversation_id,
+            mode="auto",
+            sender=sender,
+        )
+        if not deep_result.get("success"):
+            sender('error', {
+                "message": deep_result.get("error") or "自动深层压缩失败",
+                "conversation_id": conversation_id,
+            })
+        else:
+            guide_message = (deep_result.get("guide_message") or "").strip()
+            if guide_message:
+                finalize_user_work_timer()
+                await handle_task_with_sender(
+                    web_terminal,
+                    workspace,
+                    guide_message,
+                    [],
+                    sender,
+                    client_sid,
+                    username,
+                    []
+                )
+                return
+        finalize_user_work_timer()
+        return
     
     # === 移除：不在这里计算输入token，改为在每次API调用前计算 ===
     
@@ -911,11 +964,30 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
             mark_force_thinking=mark_force_thinking,
             get_stop_flag=get_stop_flag,
             clear_stop_flag=clear_stop_flag,
+            workspace=workspace,
         )
         last_tool_call_time = tool_loop_result.get("last_tool_call_time", last_tool_call_time)
         if tool_loop_result.get("stopped"):
             finalize_user_work_timer()
             return
+        if tool_loop_result.get("deep_compressed"):
+            deep_result = tool_loop_result.get("deep_result") or {}
+            guide_message = (deep_result.get("guide_message") or "").strip()
+            if deep_result.get("success") and guide_message:
+                finalize_user_work_timer()
+                await handle_task_with_sender(
+                    web_terminal,
+                    workspace,
+                    guide_message,
+                    [],
+                    sender,
+                    client_sid,
+                    username,
+                    []
+                )
+                return
+            finalize_user_work_timer()
+            return
         
         # 标记不再是第一次迭代
         is_first_iteration = False
diff --git a/server/chat_flow_tool_loop.py b/server/chat_flow_tool_loop.py
index 2ed3dcd..1248475 100644
--- a/server/chat_flow_tool_loop.py
+++ b/server/chat_flow_tool_loop.py
@@ -13,9 +13,11 @@ from .chat_flow_helpers import detect_tool_failure
 from .chat_flow_runner_helpers import resolve_monitor_path, resolve_monitor_memory, capture_monitor_snapshot
 from utils.tool_result_formatter import format_tool_result_for_context
 from config import TOOL_CALL_COOLDOWN
+from modules.personalization_manager import load_personalization_config
+from .deep_compression import run_deep_compression
 
 
-async def execute_tool_calls(*, web_terminal, tool_calls, sender, messages, client_sid: str, username: str, iteration: int, conversation_id: Optional[str], last_tool_call_time: float, process_sub_agent_updates, maybe_mark_failure_from_message, mark_force_thinking, get_stop_flag, clear_stop_flag):
+async def execute_tool_calls(*, web_terminal, tool_calls, sender, messages, client_sid: str, username: str, iteration: int, conversation_id: Optional[str], last_tool_call_time: float, process_sub_agent_updates, maybe_mark_failure_from_message, mark_force_thinking, get_stop_flag, clear_stop_flag, workspace=None):
     previous_tool_loop_active = getattr(web_terminal, "_tool_loop_active", False)
     web_terminal._tool_loop_active = True
     # 执行每个工具
@@ -415,17 +417,11 @@ async def execute_tool_calls(*, web_terminal, tool_calls, sender, messages, clie
             ):
                 img_path = inj.get("path") if isinstance(inj, dict) else None
                 if img_path:
-                    text_part = tool_result_content if isinstance(tool_result_content, str) else ""
-                    tool_message_content = web_terminal.context_manager._build_content_with_images(
-                        text_part,
-                        [img_path]
-                    )
-                    tool_images = [img_path]
                     if metadata_payload is None:
                         metadata_payload = {}
                     metadata_payload["tool_image_path"] = img_path
                     sender('system_message', {
-                        'content': f'系统已按模型请求将图片附加到工具结果: {img_path}'
+                        'content': f'系统已记录图片路径（不再附带二进制数据）: {img_path}'
                     })
 
         # view_video: 将视频直接附加到 tool 结果中（不再插入 user 消息）
@@ -439,19 +435,11 @@ async def execute_tool_calls(*, web_terminal, tool_calls, sender, messages, clie
             ):
                 video_path = inj.get("path") if isinstance(inj, dict) else None
                 if video_path:
-                    text_part = tool_result_content if isinstance(tool_result_content, str) else ""
-                    video_payload = [video_path]
-                    tool_message_content = web_terminal.context_manager._build_content_with_images(
-                        text_part,
-                        [],
-                        video_payload
-                    )
-                    tool_videos = [video_path]
                     if metadata_payload is None:
                         metadata_payload = {}
                     metadata_payload["tool_video_path"] = video_path
                     sender('system_message', {
-                        'content': f'系统已按模型请求将视频附加到工具结果: {video_path}'
+                        'content': f'系统已记录视频路径（不再附带二进制数据）: {video_path}'
                     })
 
         # 立即保存工具结果
@@ -464,6 +452,26 @@ async def execute_tool_calls(*, web_terminal, tool_calls, sender, messages, clie
             images=tool_images,
             videos=tool_videos
         )
+        try:
+            workspace_data_dir = getattr(workspace, "data_dir", None) if workspace else None
+            personal_config = load_personalization_config(workspace_data_dir) if workspace_data_dir else {}
+        except Exception:
+            personal_config = {}
+        auto_shallow_enabled = bool(personal_config.get("auto_shallow_compress_enabled", False))
+        auto_deep_enabled = bool(personal_config.get("auto_deep_compress_enabled", False))
+        compressed_count = 0
+        try:
+            compressed_count = int(
+                web_terminal.context_manager.on_tool_call_finished(function_name, enable_shallow=auto_shallow_enabled) or 0
+            )
+        except Exception as exc:
+            debug_log(f"[ContextCompression] 工具后浅压缩检测失败: {exc}")
+        if compressed_count > 0:
+            sender('shallow_compression', {
+                "conversation_id": conversation_id,
+                "compressed_count": compressed_count,
+                "keep_recent_tools": 15,
+            })
         debug_log(f"💾 增量保存：工具结果 {function_name}")
         if function_name != "wait_sub_agent":
             system_message = result_data.get("system_message") if isinstance(result_data, dict) else None
@@ -471,6 +479,40 @@ async def execute_tool_calls(*, web_terminal, tool_calls, sender, messages, clie
                 web_terminal._record_sub_agent_message(system_message, result_data.get("task_id"), inline=False)
                 maybe_mark_failure_from_message(web_terminal, system_message)
 
+        # 自动深层压缩（工具调用后触发）
+        current_context_tokens = web_terminal.context_manager.get_current_context_tokens(conversation_id)
+        if (
+            auto_deep_enabled
+            and current_context_tokens > 150_000
+            and not web_terminal.context_manager.is_compression_in_progress()
+        ):
+            web_terminal.context_manager._set_meta_flag("is_ultra_long_conversation", True)
+            sender('compression_state', {
+                "conversation_id": conversation_id,
+                "in_progress": True,
+                "mode": "auto",
+                "stage": "queued"
+            })
+            deep_result = await run_deep_compression(
+                web_terminal=web_terminal,
+                workspace=workspace,
+                conversation_id=conversation_id,
+                mode="auto",
+                sender=sender,
+            )
+            if not deep_result.get("success"):
+                sender('error', {
+                    "message": deep_result.get("error") or "自动深层压缩失败",
+                    "conversation_id": conversation_id,
+                })
+            web_terminal._tool_loop_active = previous_tool_loop_active
+            return {
+                "stopped": False,
+                "deep_compressed": True,
+                "deep_result": deep_result,
+                "last_tool_call_time": last_tool_call_time
+            }
+
         # 添加到消息历史（用于API继续对话）
         messages.append({
             "role": "tool",
diff --git a/server/conversation.py b/server/conversation.py
index ab884ae..521de18 100644
--- a/server/conversation.py
+++ b/server/conversation.py
@@ -82,6 +82,7 @@ from .conversation_stats import (
     collect_upload_events,
     summarize_upload_events,
 )
+from .deep_compression import run_deep_compression
 
 conversation_bp = Blueprint('conversation', __name__)
 
@@ -452,16 +453,24 @@ def get_conversation_messages(conversation_id, terminal: WebTerminal, workspace:
 @api_login_required
 @with_terminal
 def compress_conversation(conversation_id, terminal: WebTerminal, workspace: UserWorkspace, username: str):
-    """压缩指定对话的大体积消息，生成压缩版新对话"""
+    """深层压缩指定对话，生成 compact 文件并切换到新对话。"""
     try:
         policy = resolve_admin_policy(get_current_user_record())
         if policy.get("ui_blocks", {}).get("block_compress_conversation"):
             return jsonify({"success": False, "error": "压缩对话已被管理员禁用"}), 403
         normalized_id = conversation_id if conversation_id.startswith('conv_') else f"conv_{conversation_id}"
-        result = terminal.context_manager.compress_conversation(normalized_id)
+        result = asyncio.run(
+            run_deep_compression(
+                web_terminal=terminal,
+                workspace=workspace,
+                conversation_id=normalized_id,
+                mode="manual",
+                sender=None,
+            )
+        )
 
         if not result.get("success"):
-            status_code = 404 if "不存在" in result.get("error", "") else 400
+            status_code = 404 if "不存在" in result.get("error", "") else (409 if result.get("in_progress") else 400)
             return jsonify(result), status_code
 
         new_conversation_id = result["compressed_conversation_id"]
@@ -486,11 +495,36 @@ def compress_conversation(conversation_id, terminal: WebTerminal, workspace: Use
         response_payload = {
             "success": True,
             "compressed_conversation_id": new_conversation_id,
-            "compressed_types": result.get("compressed_types", []),
-            "system_message": result.get("system_message"),
+            "compact_file": result.get("compact_file"),
+            "summary_failed": result.get("summary_failed", False),
+            "guide_message": result.get("guide_message"),
             "load_result": load_result
         }
 
+        guide_message = (result.get("guide_message") or "").strip()
+        if guide_message:
+            try:
+                from .tasks import task_manager
+                workspace_id = session.get("workspace_id") or "default"
+                task_rec = task_manager.create_chat_task(
+                    username,
+                    workspace_id,
+                    guide_message,
+                    images=[],
+                    conversation_id=new_conversation_id,
+                    videos=[],
+                )
+                response_payload["auto_task_started"] = True
+                response_payload["auto_task_id"] = task_rec.task_id
+                response_payload["auto_task_status"] = task_rec.status
+            except Exception as exc:
+                # 任务启动失败时前端会回退到手动发送 guide_message
+                debug_log(f"[Compression] 自动发送引导消息失败: {exc}")
+                response_payload["auto_task_started"] = False
+                response_payload["auto_task_error"] = str(exc)
+        else:
+            response_payload["auto_task_started"] = False
+
         return jsonify(response_payload)
 
     except Exception as e:
@@ -502,6 +536,53 @@ def compress_conversation(conversation_id, terminal: WebTerminal, workspace: Use
         }), 500
 
 
+@conversation_bp.route('/api/conversations/<conversation_id>/compression_status', methods=['GET'])
+@api_login_required
+@with_terminal
+def get_conversation_compression_status(conversation_id, terminal: WebTerminal, workspace: UserWorkspace, username: str):
+    try:
+        normalized_id = conversation_id if conversation_id.startswith('conv_') else f"conv_{conversation_id}"
+        data = terminal.context_manager.conversation_manager.load_conversation(normalized_id) or {}
+        meta = data.get("metadata", {}) or {}
+        return jsonify({
+            "success": True,
+            "data": {
+                "conversation_id": normalized_id,
+                "compression_in_progress": bool(meta.get("compression_in_progress", False)),
+                "compression_mode": meta.get("compression_mode"),
+                "compression_stage": meta.get("compression_stage"),
+                "compression_error": meta.get("compression_error"),
+                "compression_count": int(meta.get("compression_count", 0) or 0),
+                "compression_job_id": meta.get("compression_job_id"),
+            }
+        })
+    except Exception as e:
+        return jsonify({"success": False, "error": str(e)}), 500
+
+
+@conversation_bp.route('/api/conversations/<conversation_id>/compression_cancel', methods=['POST'])
+@api_login_required
+@with_terminal
+def cancel_conversation_compression(conversation_id, terminal: WebTerminal, workspace: UserWorkspace, username: str):
+    try:
+        normalized_id = conversation_id if conversation_id.startswith('conv_') else f"conv_{conversation_id}"
+        ok = terminal.context_manager.conversation_manager.update_conversation_metadata(
+            normalized_id,
+            {
+                "compression_in_progress": False,
+                "compression_mode": None,
+                "compression_stage": None,
+                "compression_resume_payload": None,
+                "compression_error": "用户切换对话导致压缩取消",
+            }
+        )
+        if not ok:
+            return jsonify({"success": False, "error": "对话不存在或取消失败"}), 404
+        return jsonify({"success": True, "conversation_id": normalized_id})
+    except Exception as e:
+        return jsonify({"success": False, "error": str(e)}), 500
+
+
 @conversation_bp.route('/api/sub_agents', methods=['GET'])
 @api_login_required
 @with_terminal
diff --git a/server/deep_compression.py b/server/deep_compression.py
new file mode 100644
index 0000000..b07a5e0
--- /dev/null
+++ b/server/deep_compression.py
@@ -0,0 +1,295 @@
+from __future__ import annotations
+
+import asyncio
+import json
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
+
+
+SUMMARY_PROMPT = (
+    "由于当前对话过长，系统正在自动压缩。请你基于已有上下文输出一份可继续执行的工作总结，要求：\n"
+    "1) 任务目标与用户真实诉求\n"
+    "2) 已完成工作（按时间顺序）\n"
+    "3) 关键决策与原因\n"
+    "4) 修改过的文件与核心变更点\n"
+    "5) 工具调用中的重要结果/错误与修复\n"
+    "6) 当前未完成事项与下一步计划（可直接执行）\n"
+    "7) 风险与注意事项\n"
+    "请使用中文，结构清晰，尽量具体，不要省略关键上下文。"
+)
+
+GUIDE_USER_MESSAGE_TEMPLATE = "当前对话已经被自动压缩，请阅读{path}并继续工作"
+
+
+def _emit(sender, event_type: str, payload: Dict[str, Any]):
+    if not callable(sender):
+        return
+    try:
+        sender(event_type, payload)
+    except Exception:
+        pass
+
+
+def _extract_text_only(content: Any) -> str:
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts: List[str] = []
+        for item in content:
+            if not isinstance(item, dict):
+                continue
+            if item.get("type") == "text":
+                txt = item.get("text")
+                if isinstance(txt, str) and txt.strip():
+                    parts.append(txt)
+        return "\n".join(parts).strip()
+    return str(content)
+
+
+def _extract_tool_arg_map(messages: List[Dict[str, Any]]) -> Dict[str, Dict[str, Any]]:
+    tool_map: Dict[str, Dict[str, Any]] = {}
+    for msg in messages:
+        if msg.get("role") != "assistant":
+            continue
+        for tc in msg.get("tool_calls") or []:
+            tc_id = tc.get("id") or tc.get("tool_call_id")
+            if not tc_id:
+                continue
+            func = tc.get("function") or {}
+            name = func.get("name") or tc.get("name") or "unknown_tool"
+            args_raw = func.get("arguments")
+            args_obj: Any = args_raw
+            if isinstance(args_raw, str):
+                try:
+                    args_obj = json.loads(args_raw)
+                except Exception:
+                    args_obj = args_raw
+            tool_map[tc_id] = {"name": name, "arguments": args_obj}
+    return tool_map
+
+
+def _collect_last_tool_entries(messages: List[Dict[str, Any]], limit: int = 5, max_content_chars: int = 3000) -> List[Dict[str, Any]]:
+    tool_arg_map = _extract_tool_arg_map(messages)
+    entries: List[Dict[str, Any]] = []
+    for msg in messages:
+        if msg.get("role") != "tool":
+            continue
+        tc_id = msg.get("tool_call_id") or msg.get("id")
+        mapping = tool_arg_map.get(tc_id, {})
+        content_text = _extract_text_only(msg.get("content"))
+        if len(content_text) > max_content_chars:
+            content_text = content_text[:max_content_chars] + "\n...(已截断)"
+        entries.append({
+            "tool_call_id": tc_id,
+            "tool_name": msg.get("name") or mapping.get("name") or "unknown_tool",
+            "arguments": mapping.get("arguments"),
+            "content": content_text,
+        })
+    return entries[-max(1, limit):]
+
+
+def _collect_user_texts(messages: List[Dict[str, Any]]) -> List[str]:
+    result: List[str] = []
+    for msg in messages:
+        if msg.get("role") != "user":
+            continue
+        text = _extract_text_only(msg.get("content"))
+        if text.strip():
+            result.append(text.strip())
+    return result
+
+
+async def _generate_summary(web_terminal, prompt: str, retries: int = 5) -> Tuple[str, Optional[str]]:
+    last_reason: Optional[str] = None
+    context = web_terminal.build_context()
+    messages = web_terminal.build_messages(context, prompt)
+    # build_messages 不会自动附加 user_input，压缩总结必须显式注入
+    if prompt and isinstance(prompt, str):
+        messages = list(messages) + [{"role": "user", "content": prompt}]
+    for _ in range(max(1, retries)):
+        try:
+            response_text = ""
+            async for chunk in web_terminal.api_client.chat(messages, tools=None, stream=False):
+                if not isinstance(chunk, dict):
+                    continue
+                choices = chunk.get("choices") or []
+                if choices:
+                    msg = choices[0].get("message") or {}
+                    content = msg.get("content")
+                    if isinstance(content, str):
+                        response_text = content
+            if response_text.strip():
+                return response_text.strip(), None
+            last_reason = "模型返回空内容"
+        except Exception as exc:
+            last_reason = str(exc)
+        await asyncio.sleep(0.2)
+    return f"生成总结失败（{last_reason or '未知原因'}）", last_reason
+
+
+def _write_compact_file(project_path: Path, *, compression_index: int, summary_text: str, user_inputs: List[str], last_tools: List[Dict[str, Any]]) -> str:
+    compact_dir = project_path / "compact_result"
+    compact_dir.mkdir(parents=True, exist_ok=True)
+    filename = f"compact_{datetime.now().strftime('%Y%m%d_%H%M%S')}_{compression_index:03d}.md"
+    file_path = compact_dir / filename
+    lines: List[str] = [
+        f"# 对话已被第{compression_index}次压缩",
+        "",
+        "## 工作总结",
+        summary_text or "生成总结失败",
+        "",
+        "## 用户的所有输入（仅文字）",
+    ]
+    if user_inputs:
+        for idx, text in enumerate(user_inputs, start=1):
+            lines.append(f"{idx}. {text}")
+    else:
+        lines.append("- （无）")
+    lines.extend(["", "## 最近5条工具调用（参数 + 结果）"])
+    if last_tools:
+        for idx, item in enumerate(last_tools, start=1):
+            lines.extend([
+                f"### {idx}. {item.get('tool_name')}",
+                f"- tool_call_id: {item.get('tool_call_id')}",
+                "- 参数：",
+                "```json",
+                json.dumps(item.get("arguments"), ensure_ascii=False, indent=2),
+                "```",
+                "- 结果：",
+                "```text",
+                str(item.get("content") or ""),
+                "```",
+                "",
+            ])
+    else:
+        lines.append("- （无）")
+    file_path.write_text("\n".join(lines).strip() + "\n", encoding="utf-8")
+    return str(file_path.relative_to(project_path))
+
+
+async def run_deep_compression(
+    *,
+    web_terminal,
+    workspace,
+    conversation_id: str,
+    mode: str,
+    sender=None,
+) -> Dict[str, Any]:
+    cm = web_terminal.context_manager
+    conv_data = cm.conversation_manager.load_conversation(conversation_id)
+    if not conv_data:
+        return {"success": False, "error": f"对话不存在: {conversation_id}"}
+
+    metadata = conv_data.get("metadata", {}) or {}
+    if metadata.get("compression_in_progress"):
+        return {"success": False, "error": "对话正在压缩中", "in_progress": True}
+
+    old_title = conv_data.get("title") or "未命名"
+    compression_count = int(metadata.get("compression_count", 0) or 0)
+    target_count = compression_count + 1
+    job_id = f"cmp_{datetime.now().strftime('%Y%m%d_%H%M%S_%f')}"
+    cm.set_compression_state(
+        in_progress=True,
+        mode=mode,
+        stage="generating_summary",
+        job_id=job_id,
+        resume_payload={"conversation_id": conversation_id, "mode": mode},
+    )
+    _emit(sender, "compression_state", {
+        "conversation_id": conversation_id,
+        "in_progress": True,
+        "mode": mode,
+        "stage": "generating_summary",
+        "job_id": job_id,
+    })
+
+    summary_text, summary_fail_reason = await _generate_summary(web_terminal, SUMMARY_PROMPT, retries=5)
+    if summary_fail_reason:
+        _emit(sender, "system_message", {"content": f"自动压缩总结失败，将使用失败占位文本：{summary_fail_reason}"})
+
+    cm.set_compression_state(
+        in_progress=True,
+        mode=mode,
+        stage="writing_compact",
+        job_id=job_id,
+    )
+    _emit(sender, "compression_state", {
+        "conversation_id": conversation_id,
+        "in_progress": True,
+        "mode": mode,
+        "stage": "writing_compact",
+        "job_id": job_id,
+    })
+
+    messages = conv_data.get("messages") or []
+    user_inputs = _collect_user_texts(messages)
+    last_tools = _collect_last_tool_entries(messages, limit=5, max_content_chars=3000)
+    relative_compact_path = _write_compact_file(
+        Path(workspace.project_path),
+        compression_index=target_count,
+        summary_text=summary_text,
+        user_inputs=user_inputs,
+        last_tools=last_tools,
+    )
+
+    cm.set_compression_state(
+        in_progress=True,
+        mode=mode,
+        stage="creating_new_conversation",
+        job_id=job_id,
+    )
+
+    new_conv_id = cm.conversation_manager.create_conversation(
+        project_path=str(workspace.project_path),
+        thinking_mode=bool(metadata.get("thinking_mode", False)),
+        run_mode=metadata.get("run_mode") or ("thinking" if metadata.get("thinking_mode") else "fast"),
+        initial_messages=[],
+        model_key=metadata.get("model_key"),
+        has_images=False,
+        has_videos=False,
+        metadata_overrides={
+            "compression_count": target_count,
+            "title_locked": True,
+            "skip_auto_title_generation": True,
+        }
+    )
+    cm.conversation_manager.update_conversation_title(new_conv_id, f"{old_title}对话 压缩后")
+    web_terminal.load_conversation(new_conv_id)
+    guide_message = GUIDE_USER_MESSAGE_TEMPLATE.format(path=relative_compact_path)
+
+    # 清理旧对话压缩状态
+    cm.conversation_manager.update_conversation_metadata(
+        conversation_id,
+        {
+            "compression_in_progress": False,
+            "compression_mode": None,
+            "compression_stage": None,
+            "compression_job_id": None,
+            "compression_error": summary_fail_reason,
+            "compression_resume_payload": None,
+            "is_ultra_long_conversation": True,
+        },
+    )
+    _emit(sender, "conversation_resolved", {
+        "conversation_id": new_conv_id,
+        "title": f"{old_title}对话 压缩后",
+        "created": True,
+    })
+    _emit(sender, "compression_finished", {
+        "source_conversation_id": conversation_id,
+        "conversation_id": new_conv_id,
+        "in_progress": False,
+        "compact_file": relative_compact_path,
+        "job_id": job_id,
+    })
+
+    return {
+        "success": True,
+        "compressed_conversation_id": new_conv_id,
+        "compact_file": relative_compact_path,
+        "summary_failed": bool(summary_fail_reason),
+        "guide_message": guide_message,
+    }
diff --git a/server/status.py b/server/status.py
index c1d88d4..9a91ab4 100644
--- a/server/status.py
+++ b/server/status.py
@@ -71,6 +71,15 @@ def get_status(terminal, workspace, username):
                 status['conversation']['title'] = current_conv_data.get('title', '未知对话')
                 status['conversation']['created_at'] = current_conv_data.get('created_at')
                 status['conversation']['updated_at'] = current_conv_data.get('updated_at')
+                meta = current_conv_data.get("metadata", {}) or {}
+                status['conversation']['compression'] = {
+                    "in_progress": bool(meta.get("compression_in_progress", False)),
+                    "mode": meta.get("compression_mode"),
+                    "stage": meta.get("compression_stage"),
+                    "error": meta.get("compression_error"),
+                    "count": int(meta.get("compression_count", 0) or 0),
+                    "job_id": meta.get("compression_job_id"),
+                }
     except Exception as exc:
         print(f"[Status] 获取当前对话信息失败: {exc}")
     status['project_path'] = str(workspace.project_path)
diff --git a/server/tasks.py b/server/tasks.py
index 8932abc..8cc65e8 100644
--- a/server/tasks.py
+++ b/server/tasks.py
@@ -262,17 +262,34 @@ class TaskManager:
                 except Exception:
                     pass
 
+            # 轮询模式需要把 context_manager 的回调切到当前任务 sender，
+            # 否则 token_update 等事件只走 websocket，前端任务轮询拿不到实时更新。
+            previous_ctx_callback = None
+            try:
+                if terminal and getattr(terminal, "context_manager", None):
+                    previous_ctx_callback = getattr(terminal.context_manager, "_web_terminal_callback", None)
+                    terminal.context_manager.set_web_terminal_callback(sender)
+            except Exception as exc:
+                debug_log(f"[Task] 设置上下文回调失败: {exc}")
+
             # 将 task_id 作为 client_sid，供 stop_flags 检测
-            run_chat_task_sync(
-                terminal=terminal,
-                message=rec.message,
-                images=images,
-                sender=sender,
-                client_sid=rec.task_id,
-                workspace=workspace,
-                username=username,
-                videos=videos,
-            )
+            try:
+                run_chat_task_sync(
+                    terminal=terminal,
+                    message=rec.message,
+                    images=images,
+                    sender=sender,
+                    client_sid=rec.task_id,
+                    workspace=workspace,
+                    username=username,
+                    videos=videos,
+                )
+            finally:
+                try:
+                    if terminal and getattr(terminal, "context_manager", None):
+                        terminal.context_manager.set_web_terminal_callback(previous_ctx_callback)
+                except Exception as exc:
+                    debug_log(f"[Task] 恢复上下文回调失败: {exc}")
 
             # 结束状态
             canceled_flag = rec.stop_requested or stop_hint or bool(stop_flags.get(rec.task_id, {}).get("stop"))
diff --git a/static/src/app/computed.ts b/static/src/app/computed.ts
index f86bbb5..e62a71e 100644
--- a/static/src/app/computed.ts
+++ b/static/src/app/computed.ts
@@ -155,14 +155,14 @@ export const computed = {
         return !!this.policyUiBlocks.block_virtual_monitor;
     },
     displayLockEngaged() {
-        return false;
+        return !!this.compressionInProgress || !!this.compressing;
     },
     streamingUi() {
         return this.streamingMessage || this.hasPendingToolActions();
     },
     composerBusy() {
         const monitorLock = this.monitorIsLocked && this.chatDisplayMode === 'monitor';
-        return this.streamingUi || this.taskInProgress || monitorLock || this.stopRequested;
+        return this.streamingUi || this.taskInProgress || monitorLock || this.stopRequested || this.compressionInProgress || this.compressing;
     },
     composerHeroActive() {
         return (this.blankHeroActive || this.blankHeroExiting) && !this.composerInteractionActive;
diff --git a/static/src/app/methods/conversation.ts b/static/src/app/methods/conversation.ts
index 2ada225..28acc80 100644
--- a/static/src/app/methods/conversation.ts
+++ b/static/src/app/methods/conversation.ts
@@ -212,6 +212,31 @@ export const conversationMethods = {
             return;
         }
 
+        if ((this.compressionInProgress || this.compressing) && !force) {
+            const confirmed = await this.confirmAction({
+                title: '压缩进行中',
+                message: '对话正在压缩中，切换对话会导致压缩失败，确认要继续吗？',
+                confirmText: '确认',
+                cancelText: '取消'
+            });
+            if (!confirmed) {
+                this.suppressTitleTyping = false;
+                this.titleReady = true;
+                return;
+            }
+            try {
+                if (this.currentConversationId) {
+                    await fetch(`/api/conversations/${this.currentConversationId}/compression_cancel`, { method: 'POST' });
+                }
+            } catch (e) {
+                console.warn('取消压缩失败:', e);
+            }
+            this.compressionInProgress = false;
+            this.compressing = false;
+            this.compressionMode = '';
+            this.compressionStage = '';
+        }
+
         // 应用个性化设置中的默认模型和思考模式
         try {
             const { usePersonalizationStore } = await import('../../stores/personalization');
@@ -354,6 +379,28 @@ export const conversationMethods = {
 
     async createNewConversation() {
         console.log('[DEBUG_AWAITING] ===== createNewConversation =====');
+        if (this.compressionInProgress || this.compressing) {
+            const confirmed = await this.confirmAction({
+                title: '压缩进行中',
+                message: '对话正在压缩中，切换对话会导致压缩失败，确认要继续吗？',
+                confirmText: '确认',
+                cancelText: '取消'
+            });
+            if (!confirmed) {
+                return;
+            }
+            try {
+                if (this.currentConversationId) {
+                    await fetch(`/api/conversations/${this.currentConversationId}/compression_cancel`, { method: 'POST' });
+                }
+            } catch (e) {
+                console.warn('取消压缩失败:', e);
+            }
+            this.compressionInProgress = false;
+            this.compressing = false;
+            this.compressionMode = '';
+            this.compressionStage = '';
+        }
 
         debugLog('创建新对话...');
         traceLog('createNewConversation:start', {
@@ -585,7 +632,7 @@ export const conversationMethods = {
             if (response.ok && result.success) {
                 const newId = result.duplicate_conversation_id;
                 if (newId) {
-                    this.currentConversationId = newId;
+                    await this.loadConversation(newId, { force: true });
                 }
 
                 this.conversationsOffset = 0;
diff --git a/static/src/app/methods/message.ts b/static/src/app/methods/message.ts
index 1c16acc..89b348d 100644
--- a/static/src/app/methods/message.ts
+++ b/static/src/app/methods/message.ts
@@ -102,6 +102,14 @@ export const messageMethods = {
     },
 
     handleSendOrStop() {
+        if (this.compressionInProgress) {
+            this.uiPushToast({
+                title: '对话自动压缩中',
+                message: '当前不可发送/停止，请等待压缩完成',
+                type: 'warning'
+            });
+            return;
+        }
         if (this.composerBusy) {
             this.stopTask();
         } else {
@@ -112,6 +120,14 @@ export const messageMethods = {
     async sendMessage() {
         console.log('[DEBUG_AWAITING] ===== sendMessage =====');
 
+        if (this.compressionInProgress) {
+            this.uiPushToast({
+                title: '对话自动压缩中',
+                message: '压缩完成后才能继续发送消息',
+                type: 'warning'
+            });
+            return;
+        }
         if (this.streamingUi) {
             return;
         }
@@ -301,6 +317,14 @@ export const messageMethods = {
     // 新增：停止任务方法
     async stopTask() {
         console.log('[DEBUG_AWAITING] ===== stopTask =====');
+        if (this.compressionInProgress) {
+            this.uiPushToast({
+                title: '对话自动压缩中',
+                message: '压缩进行中，当前不可停止任务',
+                type: 'warning'
+            });
+            return;
+        }
 
         const canStop = this.composerBusy && !this.stopRequested;
         if (!canStop) {
@@ -428,6 +452,16 @@ export const messageMethods = {
         }
 
         this.compressing = true;
+        this.compressionInProgress = true;
+        this.compressionMode = 'manual';
+        this.compressionStage = 'requesting';
+        this.compressionError = '';
+        this.uiPushToast({
+            title: '压缩中',
+            message: '对话正在压缩，请稍候…',
+            type: 'info',
+            duration: 2500
+        });
 
         try {
             const response = await fetch(`/api/conversations/${this.currentConversationId}/compress`, {
@@ -437,17 +471,46 @@ export const messageMethods = {
             const result = await response.json();
 
             if (response.ok && result.success) {
+                this.compressionStage = 'switching';
                 const newId = result.compressed_conversation_id;
                 if (newId) {
                     await this.loadConversation(newId, { force: true });
                 }
+                const guideMessage = (result.guide_message || '').trim();
+                const autoTaskStarted = !!result.auto_task_started;
+                const autoTaskId = result.auto_task_id;
+                if (autoTaskStarted && autoTaskId) {
+                    const { useTaskStore } = await import('../../stores/task');
+                    const taskStore = useTaskStore();
+                    if (typeof this.clearProcessedEvents === 'function') {
+                        this.clearProcessedEvents();
+                    }
+                    taskStore.resumeTask(autoTaskId, {
+                        status: result.auto_task_status || 'running',
+                        resetOffset: true,
+                        eventHandler: (event) => this.handleTaskEvent(event)
+                    });
+                    this.taskInProgress = true;
+                    if (typeof this.monitorShowPendingReply === 'function') {
+                        this.monitorShowPendingReply();
+                    }
+                } else if (guideMessage) {
+                    await this.sendAutoUserMessage(guideMessage);
+                }
 
                 this.conversationsOffset = 0;
                 await this.loadConversationsList();
 
                 debugLog('对话压缩完成:', result);
+                this.uiPushToast({
+                    title: '压缩完成',
+                    message: '已生成压缩后的新对话',
+                    type: 'success',
+                    duration: 2200
+                });
             } else {
                 const message = result.message || result.error || '压缩失败';
+                this.compressionError = message;
                 this.uiPushToast({
                     title: '压缩失败',
                     message,
@@ -456,6 +519,7 @@ export const messageMethods = {
             }
         } catch (error) {
             console.error('压缩对话异常:', error);
+            this.compressionError = error.message || '请稍后重试';
             this.uiPushToast({
                 title: '压缩对话异常',
                 message: error.message || '请稍后重试',
@@ -463,6 +527,9 @@ export const messageMethods = {
             });
         } finally {
             this.compressing = false;
+            this.compressionInProgress = false;
+            this.compressionMode = '';
+            this.compressionStage = '';
         }
     },
 
diff --git a/static/src/app/methods/resources.ts b/static/src/app/methods/resources.ts
index 5844e38..2946e8f 100644
--- a/static/src/app/methods/resources.ts
+++ b/static/src/app/methods/resources.ts
@@ -145,6 +145,18 @@ export const resourceMethods = {
         if (status && typeof status.has_videos !== 'undefined') {
             this.conversationHasVideos = !!status.has_videos;
         }
+        const compression = status?.conversation?.compression;
+        if (compression && typeof compression === 'object') {
+            this.compressionInProgress = !!compression.in_progress;
+            this.compressionMode = compression.mode || '';
+            this.compressionStage = compression.stage || '';
+            this.compressionError = compression.error || '';
+        } else {
+            this.compressionInProgress = false;
+            this.compressionMode = '';
+            this.compressionStage = '';
+            this.compressionError = '';
+        }
     },
 
     updateContainerStatus(status) {
diff --git a/static/src/app/methods/taskPolling.ts b/static/src/app/methods/taskPolling.ts
index 48bcbe3..bdb22f3 100644
--- a/static/src/app/methods/taskPolling.ts
+++ b/static/src/app/methods/taskPolling.ts
@@ -51,7 +51,7 @@ export const taskPollingMethods = {
 
         // 检查事件的 conversation_id 是否匹配当前对话
         // 如果不匹配，忽略该事件（避免切换对话后旧任务的事件显示到新对话中）
-        if (eventData.conversation_id && this.currentConversationId) {
+        if (eventType !== 'shallow_compression' && eventData.conversation_id && this.currentConversationId) {
             if (eventData.conversation_id !== this.currentConversationId) {
                 console.log(`[DEBUG_AWAITING] 忽略不匹配的事件`, {
                     eventType,
@@ -151,6 +151,15 @@ export const taskPollingMethods = {
             case 'conversation_resolved':
                 this.handleConversationResolved(eventData, eventIdx);
                 break;
+            case 'compression_state':
+                this.handleCompressionState(eventData, eventIdx);
+                break;
+            case 'compression_finished':
+                this.handleCompressionFinished(eventData, eventIdx);
+                break;
+            case 'shallow_compression':
+                this.handleShallowCompression(eventData, eventIdx);
+                break;
 
             case 'user_message':
                 this.handleUserMessage(eventData, eventIdx);
@@ -922,6 +931,43 @@ export const taskPollingMethods = {
         }
     },
 
+    handleCompressionState(data: any) {
+        if (!data || typeof data !== 'object') {
+            return;
+        }
+        this.compressionInProgress = !!data.in_progress;
+        this.compressionMode = data.mode || '';
+        this.compressionStage = data.stage || '';
+        if (!this.compressionInProgress) {
+            this.compressionError = data.error || '';
+        }
+    },
+
+    async handleCompressionFinished(data: any) {
+        this.compressionInProgress = false;
+        this.compressionMode = '';
+        this.compressionStage = '';
+        this.compressionError = '';
+        const newId = data?.conversation_id;
+        if (newId && newId !== this.currentConversationId) {
+            await this.loadConversation(newId, { force: true });
+        }
+    },
+
+    handleShallowCompression(data: any, eventIdx?: number) {
+        const count = Number(data?.compressed_count || 0);
+        if (count <= 0) {
+            return;
+        }
+        debugLog('[TaskPolling] 自动浅层压缩触发, idx:', eventIdx, data);
+        this.uiPushToast({
+            title: '自动浅层压缩',
+            message: `已自动压缩 ${count} 条较早工具结果`,
+            type: 'info',
+            duration: 2500
+        });
+    },
+
     /**
      * 清理已处理的事件索引
      */
diff --git a/static/src/app/methods/ui.ts b/static/src/app/methods/ui.ts
index 05592ff..9fe78a6 100644
--- a/static/src/app/methods/ui.ts
+++ b/static/src/app/methods/ui.ts
@@ -534,6 +534,14 @@ export const uiMethods = {
         if (this.compressing || this.streamingMessage || !this.isConnected) {
             return;
         }
+        if (this.compressionInProgress) {
+            this.uiPushToast({
+                title: '对话自动压缩中',
+                message: '当前对话正在压缩，请稍后再试',
+                type: 'warning'
+            });
+            return;
+        }
         if (this.isPolicyBlocked('block_compress_conversation', '压缩对话已被管理员禁用')) {
             return;
         }
diff --git a/static/src/app/state.ts b/static/src/app/state.ts
index b807f38..612feb7 100644
--- a/static/src/app/state.ts
+++ b/static/src/app/state.ts
@@ -40,6 +40,10 @@ export function dataState() {
 
         // 对话压缩状态
         compressing: false,
+        compressionInProgress: false,
+        compressionMode: '',
+        compressionStage: '',
+        compressionError: '',
         skipConversationLoadedEvent: false,
         skipConversationHistoryReload: false,
         _scrollListenerReady: false,
diff --git a/static/src/components/personalization/PersonalizationDrawer.vue b/static/src/components/personalization/PersonalizationDrawer.vue
index 5686f29..2d72903 100644
--- a/static/src/components/personalization/PersonalizationDrawer.vue
+++ b/static/src/components/personalization/PersonalizationDrawer.vue
@@ -483,6 +483,46 @@
                           <span>显示格式化的工具结果（更清晰易读）</span>
                         </label>
                       </div>
+                      <div class="behavior-field">
+                        <div class="behavior-field-header">
+                          <span class="field-title">上下文压缩策略</span>
+                          <p class="field-desc">可分别控制自动浅层压缩与自动深层压缩。</p>
+                        </div>
+                        <label class="toggle-row">
+                          <input
+                            type="checkbox"
+                            :checked="form.auto_shallow_compress_enabled"
+                            @change="personalization.updateField({ key: 'auto_shallow_compress_enabled', value: $event.target.checked })"
+                          />
+                          <span class="fancy-check" aria-hidden="true">
+                            <svg viewBox="0 0 64 64">
+                              <path
+                                d="M 0 16 V 56 A 8 8 90 0 0 8 64 H 56 A 8 8 90 0 0 64 56 V 8 A 8 8 90 0 0 56 0 H 8 A 8 8 90 0 0 0 8 V 16 L 32 48 L 64 16 V 8 A 8 8 90 0 0 56 0 H 8 A 8 8 90 0 0 0 8 V 56 A 8 8 90 0 0 8 64 H 56 A 8 8 90 0 0 64 56 V 16"
+                                pathLength="575.0541381835938"
+                                class="fancy-path"
+                              ></path>
+                            </svg>
+                          </span>
+                          <span>自动浅层压缩</span>
+                        </label>
+                        <label class="toggle-row">
+                          <input
+                            type="checkbox"
+                            :checked="form.auto_deep_compress_enabled"
+                            @change="personalization.updateField({ key: 'auto_deep_compress_enabled', value: $event.target.checked })"
+                          />
+                          <span class="fancy-check" aria-hidden="true">
+                            <svg viewBox="0 0 64 64">
+                              <path
+                                d="M 0 16 V 56 A 8 8 90 0 0 8 64 H 56 A 8 8 90 0 0 64 56 V 8 A 8 8 90 0 0 56 0 H 8 A 8 8 90 0 0 0 8 V 16 L 32 48 L 64 16 V 8 A 8 8 90 0 0 56 0 H 8 A 8 8 90 0 0 0 8 V 56 A 8 8 90 0 0 8 64 H 56 A 8 8 90 0 0 64 56 V 16"
+                                pathLength="575.0541381835938"
+                                class="fancy-path"
+                              ></path>
+                            </svg>
+                          </span>
+                          <span>自动深层压缩</span>
+                        </label>
+                      </div>
                       <div class="behavior-field">
                         <div class="behavior-field-header">
                           <span class="field-title">堆叠块显示模式</span>
diff --git a/static/src/stores/personalization.ts b/static/src/stores/personalization.ts
index 94c4916..2821280 100644
--- a/static/src/stores/personalization.ts
+++ b/static/src/stores/personalization.ts
@@ -22,6 +22,8 @@ interface PersonalForm {
   default_run_mode: RunMode | null;
   default_model: string | null;
   image_compression: string;
+  auto_shallow_compress_enabled: boolean;
+  auto_deep_compress_enabled: boolean;
 }
 
 interface LiquidGlassPosition {
@@ -79,7 +81,9 @@ const defaultForm = (): PersonalForm => ({
   disabled_tool_categories: [],
   default_run_mode: null,
   default_model: 'kimi-k2.5',
-  image_compression: 'original'
+  image_compression: 'original',
+  auto_shallow_compress_enabled: false,
+  auto_deep_compress_enabled: false
 });
 
 const defaultExperimentState = (): ExperimentState => ({
@@ -228,7 +232,9 @@ export const usePersonalizationStore = defineStore('personalization', {
             ? data.default_run_mode as RunMode
             : null,
         default_model: typeof data.default_model === 'string' ? data.default_model : fallbackModel,
-        image_compression: typeof data.image_compression === 'string' ? data.image_compression : 'original'
+        image_compression: typeof data.image_compression === 'string' ? data.image_compression : 'original',
+        auto_shallow_compress_enabled: !!data.auto_shallow_compress_enabled,
+        auto_deep_compress_enabled: !!data.auto_deep_compress_enabled
       };
       this.clearFeedback();
     },
diff --git a/utils/context_manager.py b/utils/context_manager.py
index 871adbb..e7f8447 100644
--- a/utils/context_manager.py
+++ b/utils/context_manager.py
@@ -45,6 +45,21 @@ except ImportError:
     from config.model_profiles import get_model_prompt_replacements
 from utils.conversation_manager import ConversationManager
 
+AUTO_SHALLOW_PLACEHOLDER = "过早的工具结果已经被自动压缩"
+AUTO_SHALLOW_TOOL_WHITELIST = {
+    "write_file",
+    "read_file",
+    "edit_file",
+    "terminal_input",
+    "terminal_snapshot",
+    "web_search",
+    "extract_webpage",
+    "run_python",
+    "run_command",
+    "view_image",
+    "view_video",
+}
+
 class ContextManager:
     def __init__(self, project_path: str, data_dir: Optional[str] = None):
         self.project_path = Path(project_path).resolve()
@@ -66,6 +81,7 @@ class ContextManager:
         self.conversation_metadata: Dict[str, Any] = {}
         self.project_snapshot: Optional[Dict[str, Any]] = None
         self._host_runtime_cache: Optional[Dict[str, str]] = None
+        self._shallow_compact_round: int = 0
         
         # 新增：对话持久化管理器
         self.conversation_manager = ConversationManager(base_dir=self.data_dir)
@@ -429,8 +445,10 @@ class ContextManager:
             prompt_tokens = int(usage.get("prompt_tokens") or 0)
             completion_tokens = int(usage.get("completion_tokens") or 0)
             total_tokens = int(usage.get("total_tokens") or (prompt_tokens + completion_tokens))
+            # 当前上下文长度优先取专用字段；缺失时回退到 prompt_tokens
+            current_context_tokens = int(usage.get("current_context_tokens") or prompt_tokens)
         except (TypeError, ValueError):
-            prompt_tokens = completion_tokens = total_tokens = 0
+            prompt_tokens = completion_tokens = total_tokens = current_context_tokens = 0
 
         try:
             self._increment_workspace_token_totals(prompt_tokens, completion_tokens, total_tokens)
@@ -446,7 +464,8 @@ class ContextManager:
                 self.current_conversation_id,
                 prompt_tokens,
                 completion_tokens,
-                total_tokens
+                total_tokens,
+                current_context_tokens=current_context_tokens,
             )
             
             if success:
@@ -481,6 +500,144 @@ class ContextManager:
         if not stats:
             return 0
         return stats.get("current_context_tokens", 0)
+
+    def _get_meta_flag(self, key: str, default: Any = None) -> Any:
+        return (self.conversation_metadata or {}).get(key, default)
+
+    def _set_meta_flag(self, key: str, value: Any, save: bool = True):
+        self.conversation_metadata[key] = value
+        if save and self.current_conversation_id:
+            try:
+                self.conversation_manager.update_conversation_metadata(
+                    self.current_conversation_id,
+                    {key: value}
+                )
+            except Exception as exc:
+                print(f"[ContextCompression] 保存 metadata 失败 {key}: {exc}")
+
+    def is_compression_in_progress(self) -> bool:
+        return bool(self._get_meta_flag("compression_in_progress", False))
+
+    def set_compression_state(
+        self,
+        *,
+        in_progress: bool,
+        mode: Optional[str] = None,
+        stage: Optional[str] = None,
+        error: Optional[str] = None,
+        resume_payload: Optional[Dict[str, Any]] = None,
+        job_id: Optional[str] = None,
+    ):
+        updates: Dict[str, Any] = {
+            "compression_in_progress": bool(in_progress),
+            "compression_mode": mode if in_progress else None,
+            "compression_stage": stage if in_progress else None,
+            "compression_error": error if in_progress else None,
+            "compression_resume_payload": resume_payload if in_progress else None,
+            "compression_job_id": job_id if in_progress else None,
+        }
+        for k, v in updates.items():
+            self.conversation_metadata[k] = v
+        if self.current_conversation_id:
+            try:
+                self.conversation_manager.update_conversation_metadata(self.current_conversation_id, updates)
+            except Exception as exc:
+                print(f"[ContextCompression] 更新压缩状态失败: {exc}")
+
+    def on_tool_call_finished(self, tool_name: Optional[str] = None, *, enable_shallow: bool = True) -> int:
+        """每次工具调用完成后触发：更新计数并执行自动浅压缩。返回本轮浅压缩条数。"""
+        if not self.current_conversation_id:
+            return 0
+        # 兼容历史对话：tool_call_count 可能从 0 开始，但历史里已存在大量 tool 结果。
+        # 这里优先对齐到“当前历史中的 tool 消息数量”，避免长期无法触发浅压缩。
+        meta_tool_count = int(self._get_meta_flag("tool_call_count", 0) or 0)
+        history_tool_count = 0
+        try:
+            history_tool_count = sum(
+                1 for msg in (self.conversation_history or [])
+                if isinstance(msg, dict) and msg.get("role") == "tool"
+            )
+        except Exception:
+            history_tool_count = 0
+        tool_count = max(meta_tool_count + 1, history_tool_count)
+        self._set_meta_flag("tool_call_count", tool_count, save=False)
+
+        current_tokens = self.get_current_context_tokens(self.current_conversation_id)
+        if current_tokens > 80_000 and not self._get_meta_flag("is_long_conversation", False):
+            self._set_meta_flag("is_long_conversation", True, save=False)
+        if current_tokens > 150_000 and not self._get_meta_flag("is_ultra_long_conversation", False):
+            self._set_meta_flag("is_ultra_long_conversation", True, save=False)
+
+        # 每 10 次工具调用触发一次浅压缩（当 long 已标记）
+        last_checkpoint = int(self._get_meta_flag("last_shallow_compress_tool_count", 0) or 0)
+        should_try_shallow = (
+            bool(enable_shallow)
+            and bool(self._get_meta_flag("is_long_conversation", False))
+            and (tool_count - last_checkpoint >= 10)
+        )
+        changed = False
+        compressed_count = 0
+        if should_try_shallow:
+            compressed = self._run_auto_shallow_compression(batch_size=10, keep_recent_tools=15)
+            self._set_meta_flag("last_shallow_compress_tool_count", tool_count, save=False)
+            compressed_count = max(0, int(compressed or 0))
+            changed = compressed > 0
+
+        if self.current_conversation_id:
+            updates = {
+                "tool_call_count": tool_count,
+                "is_long_conversation": bool(self._get_meta_flag("is_long_conversation", False)),
+                "is_ultra_long_conversation": bool(self._get_meta_flag("is_ultra_long_conversation", False)),
+                "last_shallow_compress_tool_count": int(self._get_meta_flag("last_shallow_compress_tool_count", 0) or 0),
+            }
+            try:
+                self.conversation_manager.update_conversation_metadata(self.current_conversation_id, updates)
+            except Exception as exc:
+                print(f"[ContextCompression] 写入计数失败: {exc}")
+        if changed:
+            self.auto_save_conversation(force=True)
+        return compressed_count
+
+    def _run_auto_shallow_compression(self, batch_size: int = 10, keep_recent_tools: int = 15) -> int:
+        """浅压缩：仅打标记，不修改原文。返回本轮标记条数。"""
+        history = self.conversation_history or []
+        if not history:
+            return 0
+
+        tool_indices = [idx for idx, msg in enumerate(history) if msg.get("role") == "tool"]
+        protected_indices = set(tool_indices[-max(0, keep_recent_tools):]) if keep_recent_tools > 0 else set()
+
+        candidates: List[int] = []
+        for idx, msg in enumerate(history):
+            if msg.get("role") != "tool":
+                continue
+            if idx in protected_indices:
+                continue
+            tool_name = (msg.get("name") or "").strip()
+            if tool_name not in AUTO_SHALLOW_TOOL_WHITELIST:
+                continue
+            metadata = msg.get("metadata") or {}
+            if metadata.get("auto_shallow_compacted"):
+                continue
+            candidates.append(idx)
+            if len(candidates) >= max(1, batch_size):
+                break
+
+        if not candidates:
+            return 0
+
+        self._shallow_compact_round += 1
+        now = datetime.now().isoformat()
+        for idx in candidates:
+            msg = history[idx]
+            metadata = msg.get("metadata") or {}
+            metadata["auto_shallow_compacted"] = True
+            metadata["auto_shallow_compacted_at"] = now
+            metadata["auto_shallow_compact_round"] = self._shallow_compact_round
+            msg["metadata"] = metadata
+            history[idx] = msg
+        self.conversation_history = history
+        return len(candidates)
     
     # ===========================================
     # 新增：对话持久化相关方法
@@ -1765,6 +1922,18 @@ class ContextManager:
         if workspace_system:
             messages.append({"role": "system", "content": workspace_system})
 
+        # 浅压缩替换开关：仅当个人空间开启自动浅层压缩时，才将已打标 tool 结果替换为占位符
+        shallow_replace_enabled = False
+        try:
+            personalization_config = (
+                getattr(self, "custom_personalization_config", None)
+                or load_personalization_config(self.data_dir)
+            )
+            shallow_replace_enabled = bool(personalization_config.get("auto_shallow_compress_enabled", False))
+        except Exception:
+            shallow_replace_enabled = False
+
+        replaced_tool_count = 0
         # 添加对话历史
         for conv in context["conversation"]:
             if conv["role"] == "assistant":
@@ -1779,6 +1948,17 @@ class ContextManager:
                     message["tool_calls"] = conv["tool_calls"]
                 messages.append(message)
             elif conv["role"] == "tool":
+                conv_meta = conv.get("metadata") or {}
+                if shallow_replace_enabled and conv_meta.get("auto_shallow_compacted"):
+                    message = {
+                        "role": "tool",
+                        "content": AUTO_SHALLOW_PLACEHOLDER,
+                        "tool_call_id": conv.get("tool_call_id", ""),
+                        "name": conv.get("name", "")
+                    }
+                    messages.append(message)
+                    replaced_tool_count += 1
+                    continue
                 images = conv.get("images") or (conv.get("metadata") or {}).get("images") or []
                 videos = conv.get("videos") or (conv.get("metadata") or {}).get("videos") or []
                 content_value = conv.get("content")
@@ -1803,6 +1983,8 @@ class ContextManager:
                     "role": conv["role"],
                     "content": content_payload
                 })
+        if shallow_replace_enabled:
+            print(f"[ContextCompression] build_messages 替换tool占位符: {replaced_tool_count} 条")
         
         # 添加终端内容（如果有的话）
         # 这里需要从参数传入或获取
diff --git a/utils/conversation_manager.py b/utils/conversation_manager.py
index 60c38ea..21e80ee 100644
--- a/utils/conversation_manager.py
+++ b/utils/conversation_manager.py
@@ -3,6 +3,8 @@
 import json
 import os
 import time
+import tempfile
+import threading
 from datetime import datetime
 from pathlib import Path
 from typing import Dict, List, Optional, Any
@@ -42,6 +44,7 @@ class ConversationManager:
         self.base_dir = Path(base_dir).expanduser().resolve() if base_dir else Path(DATA_DIR).resolve()
         self.conversations_dir = self.base_dir / "conversations"
         self.index_file = self.conversations_dir / "index.json"
+        self._io_lock = threading.RLock()
         self.current_conversation_id: Optional[str] = None
         self.workspace_root = Path(__file__).resolve().parents[1]
         self._ensure_directories()
@@ -60,11 +63,47 @@ class ConversationManager:
     
     def _iter_conversation_files(self, sort_by_mtime: bool = True):
         """遍历对话文件（排除索引文件），可按修改时间降序排序。"""
-        files = [p for p in self.conversations_dir.glob("*.json") if p != self.index_file]
+        files = []
+        for p in self.conversations_dir.glob("*.json"):
+            if p == self.index_file:
+                continue
+            stem = p.stem or ""
+            # 跳过索引备份/损坏文件，避免被误当作对话文件参与重建
+            if stem == "index" or stem.startswith("index_corrupt_"):
+                continue
+            # 仅纳入标准对话文件，避免其他 json 干扰索引
+            if not stem.startswith("conv_"):
+                continue
+            files.append(p)
         if sort_by_mtime:
             files.sort(key=lambda p: p.stat().st_mtime, reverse=True)
         return files
 
+    def _atomic_write_json(self, target_file: Path, payload: Dict[str, Any]):
+        """原子写入 JSON：使用唯一临时文件，避免并发写同名 .tmp 产生竞态。"""
+        target_file.parent.mkdir(parents=True, exist_ok=True)
+        temp_path: Optional[Path] = None
+        try:
+            with tempfile.NamedTemporaryFile(
+                mode="w",
+                encoding="utf-8",
+                dir=str(target_file.parent),
+                prefix=f".{target_file.name}.",
+                suffix=".tmp",
+                delete=False,
+            ) as fh:
+                temp_path = Path(fh.name)
+                json.dump(payload, fh, ensure_ascii=False, indent=2)
+                fh.flush()
+                os.fsync(fh.fileno())
+            os.replace(str(temp_path), str(target_file))
+        finally:
+            if temp_path and temp_path.exists():
+                try:
+                    temp_path.unlink()
+                except Exception:
+                    pass
+
     def _rebuild_index_from_files(self, max_count: Optional[int] = None) -> Dict:
         """
         从现有对话文件重建索引。
@@ -156,7 +195,7 @@ class ConversationManager:
         except (json.JSONDecodeError, Exception) as e:
             print(f"⚠️ 加载对话索引失败，将尝试重建: {e}")
             backup_path = self.index_file.with_name(
-                f"{self.index_file.stem}_corrupt_{int(time.time())}{self.index_file.suffix}"
+                f"{self.index_file.stem}_corrupt_{int(time.time() * 1000)}{self.index_file.suffix}"
             )
             try:
                 if self.index_file.exists():
@@ -164,7 +203,8 @@ class ConversationManager:
                     print(f"🗄️ 已备份损坏的索引文件到: {backup_path.name}")
             except Exception as backup_exc:
                 print(f"⚠️ 备份损坏索引文件失败: {backup_exc}")
-            rebuilt = self._rebuild_index_from_files(max_count=max_rebuild)
+            # 索引损坏场景优先全量重建，避免仅重建部分导致“对话丢失”错觉
+            rebuilt = self._rebuild_index_from_files(max_count=None)
             if rebuilt:
                 self._save_index(rebuilt)
                 if ensure_integrity:
@@ -174,17 +214,10 @@ class ConversationManager:
     
     def _save_index(self, index: Dict):
         """保存对话索引"""
-        temp_file = self.index_file.with_suffix(self.index_file.suffix + ".tmp")
         try:
-            with open(temp_file, 'w', encoding='utf-8') as f:
-                json.dump(index, f, ensure_ascii=False, indent=2)
-            os.replace(temp_file, self.index_file)
+            with self._io_lock:
+                self._atomic_write_json(self.index_file, index)
         except Exception as e:
-            try:
-                if temp_file.exists():
-                    temp_file.unlink()
-            except Exception:
-                pass
             print(f"⌘ 保存对话索引失败: {e}")
 
     def _ensure_index_covering(self, limit: int, offset: int) -> Dict:
@@ -237,14 +270,12 @@ class ConversationManager:
         return "新对话"
     
     def _count_tools_in_messages(self, messages: List[Dict]) -> int:
-        """统计消息中的工具调用数量"""
+        """统计消息中的工具调用数量（仅统计 assistant.tool_calls）。"""
         tool_count = 0
         for msg in messages:
             if msg.get("role") == "assistant" and "tool_calls" in msg:
                 tool_calls = msg.get("tool_calls", [])
                 tool_count += len(tool_calls) if isinstance(tool_calls, list) else 0
-            elif msg.get("role") == "tool":
-                tool_count += 1
         return tool_count
 
     def _prepare_project_path_metadata(self, project_path: Optional[str]) -> Dict[str, Optional[str]]:
@@ -317,7 +348,8 @@ class ConversationManager:
         initial_messages: List[Dict] = None,
         model_key: Optional[str] = None,
         has_images: bool = False,
-        has_videos: bool = False
+        has_videos: bool = False,
+        metadata_overrides: Optional[Dict[str, Any]] = None
     ) -> str:
         """
         创建新对话
@@ -336,6 +368,38 @@ class ConversationManager:
         # 创建对话数据
         path_metadata = self._prepare_project_path_metadata(project_path)
         normalized_mode = run_mode if run_mode in {"fast", "thinking", "deep"} else ("thinking" if thinking_mode else "fast")
+        metadata = {
+            "project_path": path_metadata["project_path"],
+            "project_relative_path": path_metadata["project_relative_path"],
+            "thinking_mode": thinking_mode,
+            "run_mode": normalized_mode,
+            "model_key": model_key,
+            "has_images": has_images,
+            "has_videos": has_videos,
+            # 首次对话尚未生成文件树快照，待首次用户消息时填充
+            "project_file_tree": None,
+            "project_statistics": None,
+            "project_snapshot_at": None,
+            "total_messages": len(messages),
+            "total_tools": self._count_tools_in_messages(messages),
+            "status": "active",
+            # 压缩相关字段
+            "compression_count": 0,
+            "is_long_conversation": False,
+            "is_ultra_long_conversation": False,
+            "tool_call_count": 0,
+            "last_shallow_compress_tool_count": 0,
+            "compression_in_progress": False,
+            "compression_mode": None,
+            "compression_stage": None,
+            "compression_job_id": None,
+            "compression_error": None,
+            "compression_resume_payload": None,
+            "skip_auto_title_generation": False,
+        }
+        if isinstance(metadata_overrides, dict):
+            metadata.update(metadata_overrides)
+
         conversation_data = {
             "id": conversation_id,
             "title": self._extract_title_from_messages(messages),
@@ -343,23 +407,7 @@ class ConversationManager:
             "updated_at": datetime.now().isoformat(),
             "messages": messages,
             "todo_list": None,
-            "metadata": {
-                "project_path": path_metadata["project_path"],
-                "project_relative_path": path_metadata["project_relative_path"],
-                "thinking_mode": thinking_mode,
-                "run_mode": normalized_mode,
-                "model_key": model_key,
-                "has_images": has_images,
-                "has_videos": has_videos,
-                "has_videos": has_videos,
-                # 首次对话尚未生成文件树快照，待首次用户消息时填充
-                "project_file_tree": None,
-                "project_statistics": None,
-                "project_snapshot_at": None,
-                "total_messages": len(messages),
-                "total_tools": self._count_tools_in_messages(messages),
-                "status": "active"
-            },
+            "metadata": metadata,
             "token_statistics": self._initialize_token_statistics()  # 新增
         }
         
@@ -374,6 +422,25 @@ class ConversationManager:
         
         return conversation_id
 
+    def update_conversation_metadata(self, conversation_id: str, updates: Dict[str, Any]) -> bool:
+        """合并更新对话 metadata。"""
+        if not conversation_id or not isinstance(updates, dict):
+            return False
+        try:
+            data = self.load_conversation(conversation_id)
+            if not data:
+                return False
+            metadata = data.get("metadata", {}) or {}
+            metadata.update(updates)
+            data["metadata"] = metadata
+            data["updated_at"] = datetime.now().isoformat()
+            self._save_conversation_file(conversation_id, data)
+            self._update_index(conversation_id, data)
+            return True
+        except Exception as exc:
+            print(f"⚠️ 更新对话 metadata 失败 {conversation_id}: {exc}")
+            return False
+
     def update_conversation_title(self, conversation_id: str, title: str) -> bool:
         """更新对话标题并刷新索引。"""
         if not conversation_id or not title:
@@ -399,19 +466,12 @@ class ConversationManager:
     def _save_conversation_file(self, conversation_id: str, data: Dict):
         """保存对话文件"""
         file_path = self._get_conversation_file_path(conversation_id)
-        temp_file = file_path.with_suffix(file_path.suffix + ".tmp")
         try:
             # 确保Token统计数据有效
             data = self._validate_token_statistics(data)
-            with open(temp_file, 'w', encoding='utf-8') as f:
-                json.dump(data, f, ensure_ascii=False, indent=2)
-            os.replace(temp_file, file_path)
+            with self._io_lock:
+                self._atomic_write_json(file_path, data)
         except Exception as e:
-            try:
-                if temp_file.exists():
-                    temp_file.unlink()
-            except Exception:
-                pass
             print(f"⌘ 保存对话文件失败 {conversation_id}: {e}")
     
     def _update_index(self, conversation_id: str, conversation_data: Dict):
@@ -662,6 +722,15 @@ class ConversationManager:
                     data["metadata"] = metadata
                     self._save_conversation_file(conversation_id, data)
                     print(f"🔧 为对话 {conversation_id} 补齐项目快照字段")
+
+                # 兼容历史口径：旧版本 total_tools 可能把 role=tool 也计入，导致翻倍。
+                expected_total_tools = self._count_tools_in_messages(data.get("messages") or [])
+                if int(metadata.get("total_tools", 0) or 0) != expected_total_tools:
+                    metadata["total_tools"] = expected_total_tools
+                    data["metadata"] = metadata
+                    self._save_conversation_file(conversation_id, data)
+                    self._update_index(conversation_id, data)
+                    print(f"🔧 修正对话 {conversation_id} 的 total_tools 统计为 {expected_total_tools}")
                 
                 # 回填缺失的模型字段：从最近的助手消息元数据推断
                 if metadata.get("model_key") is None:
@@ -688,7 +757,8 @@ class ConversationManager:
         conversation_id: str,
         input_tokens: int,
         output_tokens: int,
-        total_tokens: int
+        total_tokens: int,
+        current_context_tokens: Optional[int] = None,
     ) -> bool:
         """
         更新对话的Token统计
@@ -698,6 +768,7 @@ class ConversationManager:
             input_tokens: 输入Token数量
             output_tokens: 输出Token数量
             total_tokens: 本次请求的总Token数量（prompt+completion）
+            current_context_tokens: 当前上下文长度（用于压缩阈值判断）
         
         Returns:
             bool: 更新是否成功
@@ -717,7 +788,10 @@ class ConversationManager:
             token_stats["total_input_tokens"] = token_stats.get("total_input_tokens", 0) + input_tokens
             token_stats["total_output_tokens"] = token_stats.get("total_output_tokens", 0) + output_tokens
             token_stats["total_tokens"] = token_stats.get("total_tokens", 0) + total_tokens
-            token_stats["current_context_tokens"] = total_tokens
+            if current_context_tokens is None:
+                # 兼容旧调用：未显式传入时，默认以输入 token 作为当前上下文长度
+                current_context_tokens = input_tokens
+            token_stats["current_context_tokens"] = max(0, int(current_context_tokens or 0))
             token_stats["updated_at"] = datetime.now().isoformat()
             
             # 保存更新