acp-deepseek-adapter

🤖 DeepSeek TUI ↔ cc-connect

Chat with DeepSeek TUI from your phone — Feishu, WeChat, QQ, Discord, Telegram. 把 DeepSeek TUI 接入飞书、微信、QQ、Discord、Telegram——手机就是终端。

Full tool-backed ACP — read, write, execute, git, all from IM. 完整工具调用支持——读文件、写代码、跑命令、git 操作，全在聊天框里。

---

---

English | 中文

How it works / 工作原理

Feishu / WeChat / QQ / Discord / Telegram
       │
       ▼
  cc-connect (multi-IM bridge)
       │
       ▼  ACP protocol (stdin/stdout)
       │
  acp_deepseek_adapter.py  ←──  YOU ARE HERE
       │
       ▼  deepseek exec --auto
       │
  DeepSeek TUI (agent mode)
       │
       ├─ read_file   ── read project files
       ├─ write_file  ── create / edit files
       ├─ exec_shell  ── run terminal commands
       ├─ git_*       ── version control
       ├─ grep_files  ── search codebase
       └─ ...         ── full tool suite
       │
       ▼
  Response → filtered, batched, styled → back to IM

Quick Start

1. Configure cc-connect

Add to ~/.cc-connect/config.toml:

[[projects]]
name = "deepseek"

[projects.agent]
type = "acp"

[projects.agent.options]
command = "python3"
args = ["acp_deepseek_adapter.py"]
work_dir = "."
env = { ADAPTER_HIDE_TOOLS = "1", ADAPTER_STRIP_THINKING = "0" }

[[projects.platforms]]
type = "feishu"    # also: weixin, qq, discord, telegram

[projects.platforms.options]
app_id = "your-feishu-app-id"
app_secret = "your-feishu-app-secret"

[projects.auto_compress]
enabled = true
max_tokens = 100000
min_gap_mins = 15

2. Environment

Variable	Default	Description
DEEPSEEK_BIN	~/deepseek	Path to deepseek binary
DEEPSEEK_WORKDIR	$HOME	Working directory
ADAPTER_LOG_FILE	/tmp/deepseek-ccconnect.log	Log file path
ADAPTER_STRIP_THINKING	1	Strip thinking tokens (0 = keep)
ADAPTER_HIDE_TOOLS	1	Hide tool call output (0 = debug)
KIMI_API_KEY	(none)	Kimi For Coding API key

3. Start

cc-connect

cc-connect manages the adapter process automatically. Restart after config changes.

快速开始

1. 配置 cc-connect

在 ~/.cc-connect/config.toml 里加一个 project：

[[projects]]
name = "deepseek"

[projects.agent]
type = "acp"

[projects.agent.options]
command = "python3"
args = ["acp_deepseek_adapter.py"]
work_dir = "."
env = { ADAPTER_HIDE_TOOLS = "1", ADAPTER_STRIP_THINKING = "0" }

[[projects.platforms]]
type = "feishu"    # 也支持 weixin, qq, discord, telegram

[projects.platforms.options]
app_id = "your-feishu-app-id"
app_secret = "your-feishu-app-secret"

[projects.auto_compress]
enabled = true
max_tokens = 100000
min_gap_mins = 15

2. 环境变量

变量	默认值	说明
DEEPSEEK_BIN	~/deepseek	DeepSeek TUI 二进制路径
DEEPSEEK_WORKDIR	$HOME	工作目录
ADAPTER_LOG_FILE	/tmp/deepseek-ccconnect.log	日志文件
ADAPTER_STRIP_THINKING	1	去掉思考过程 (0 = 保留)
ADAPTER_HIDE_TOOLS	1	隐藏工具调用输出 (0 = 调试)
KIMI_API_KEY	(无)	Kimi 编程模型密钥

3. 启动

cc-connect

cc-connect 自动管理适配器进程，改完配置重启即可。

Slash Commands / 斜杠命令

Command	Description
/dir <path>	Switch working directory. /dir - goes back / 切换工作目录
/mode default|yolo	Change permission mode / 权限模式切换
/model v4-pro|v4-flash	Switch DeepSeek model / 模型切换
/new [name]	New session / 新建会话
/list	List sessions / 列出会话
/current	Current session info / 当前会话
/memory	Read/write AGENTS.md / 读写记忆文件
/compact	Context usage & compaction status / 上下文压缩

Filter Architecture / 过滤架构

Raw output from deepseek exec
       │
       ▼
  ┌─ Phase 1: Tool depth ───── suppress lines between tool: / tool X completed
  │
  ├─ Phase 2: Regex patterns ─ catch diff, git log, JSON, source headers
  │     ├─ _STDOUT_TOOL_FILTER_RE   (tool:, ---, +++, @@, Author:, etc.)
  │     ├─ _JSON_METADATA_RE        (checklist_write, sessionUpdate, etc.)
  │     └─ _SOURCE_HEADER_RE        (shebang, git status, README headers)
  │
  └─ Phase 3: CJK purity ────── drop lines without Chinese characters
       (after first tool call, pure-ASCII lines = internal noise)
       │
       ▼
  Clean output → batched into paragraphs → sent to IM

Background / 背景

DeepSeek TUI already supports deepseek serve --acp (issue #782), but the tool_call notification pipeline isn’t exposed yet. This adapter bridges that gap using deepseek exec --auto until native support lands.

DeepSeek TUI 官方已支持 serve --acp，但工具调用链路还没暴露。本项目在官方做完之前，用 deepseek exec --auto 自己搭了一套完整的 ACP 桥接。

All pitfalls documented in issue #1092. 踩过的坑都记在 issue #1092。

Changelog / 更新日志

v3.9.2

• Code-noise regex v2: _CODE_NOISE_RE 补上 ps 输出、进程命令行泄漏两个关键模式，切断历史污染反馈循环
• History repair: 清理已污染的历史文件条目，防止源码碎片在后续轮次回传

v3.9.1

• Auto-compress: cc-connect 项目配置加入 auto_compress，上下文达 100k token 自动触发压缩（15 分钟间隔），与终端设置同步

v3.9.0

• Streaming output: 输出分段缓冲，首段即时响应，后续内容合并为 2-3 段，减少 IM 消息碎片
• Section separators: 段落间插入 --- 分隔线，飞书渲染为可见分段
• Triple-layer filter: 三层过滤体系（正则 + 工具深度 + CJK 字符检测），彻底拦截源代码/diff/shell/git log 泄漏
• CJK purity filter: 工具调用后，不含中文字符的行自动丢弃，杜绝英文代码穿透
• History cleanup: 防污染机制，已知泄漏模式不存入对话历史，切断自循环
• Footer: 恢复 [ctx: ~X%] 上下文用量显示
• Regex expansion: 新增 shebang、git status short、git log、grep -n 等 10+ 条过滤规则

Links / 链接

adapter repo / 适配器仓库 · DeepSeek TUI · cc-connect

---

_{Built by talking to DeepSeek TUI through Feishu, overnight, by someone who doesn't code.
本项目由 DeepSeek TUI 自身辅助开发完成——一个不懂代码的小白，在飞书上对话一个通宵就做出来了。}