Skip to content

feat: add dsh (DeepSeek Harness) CLI backend #3

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
dorokuma merged 2 commits into from
Aug 29, 2026
Merged

feat: add dsh (DeepSeek Harness) CLI backend #3

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fc95187
feat: add dsh (DeepSeek Harness) CLI backend
KOP-XIAO Aug 29, 2026
8dd9b5d
fix: harden dsh backend before merge
dorokuma Aug 29, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 8 additions & 0 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,14 @@ All notable changes follow [Keep a Changelog](https://keepachangelog.com/) and

## [Unreleased]

### Added

- **dsh 后端**：新增 DeepSeek Harness CLI（`dsh`）作为第四个 CLI 后端（`/backend dsh`）。以 `dsh --profile headless -- <prompt>` 单轮运行，cwd 为每用户会话目录；从输出提取 `file:///` 与 markdown 文件链接作为可回传产物（路径字符白名单避免误吞紧邻中文）；`DSH_HOME` 显式配置时启用专用主目录及自动会话保留清理，未配置时复用宿主 `~/.dsh` 且跳过自动清理；对会话目录外的 `@绝对路径` mention 予以屏蔽拦截（替换为 `[blocked-path]`）；`/help`、`/clear`、`/new` 支持，模型/强度/人格类指令暂返回"暂不支持"。headless 每次新建会话，故为单轮模式。

### Tests

- 新增 `tests/test_dsh.py` + `tests/fake_dsh.py`：命令构造、产物提取、成功/失败/超时路径、未登录预检、slash 指令、`/backend dsh` 切换。

## [1.4.9] - 2026-08-16

### Fixed
Expand Down
38 changes: 27 additions & 11 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,21 +5,21 @@
![license](https://img.shields.io/badge/license-MIT-blue.svg)
![python](https://img.shields.io/badge/python-3.10+-blue.svg)

WeChatBridge connects a WeChat bot to agentic coding CLIs (Google's agy / Antigravity, xAI's Grok Build, or OpenAI's Codex). From WeChat you can send text, images, files, and voice-as-text to the active CLI, get replies back, and receive certain generated files over the WeChat CDN. Switch backends per user with `/backend` — no restart.
WeChatBridge connects a WeChat bot to agentic coding CLIs (Google's agy / Antigravity, xAI's Grok Build, OpenAI's Codex, or DeepSeek Harness' dsh). From WeChat you can send text, images, files, and voice-as-text to the active CLI, get replies back, and receive certain generated files over the WeChat CDN. Switch backends per user with `/backend` — no restart.

```
WeChat (phone) ⇄ iLink bot API ⇄ WeChatBridge ⇄ agy / grok / codex CLI
WeChat (phone) ⇄ iLink bot API ⇄ WeChatBridge ⇄ agy / grok / codex / dsh CLI
(this project) (runs tools)
```

The bridge process stays up and long-polls iLink. For prompts that go to a CLI, it spawns one `agy` or `grok` child (`-p` single-turn) and exits that child when done — the child does not stay resident. Many slash commands (`/help`, `/backend`, `/persona`, …) are handled inside the bridge and never start a CLI. Only artifacts the bridge can detect under the user's allowed session paths are pushed back via CDN.
The bridge process stays up and long-polls iLink. For prompts that go to a CLI, it spawns one `agy` / `grok` / `codex` / `dsh` child (single-turn) and exits that child when done — the child does not stay resident. Many slash commands (`/help`, `/backend`, `/persona`, …) are handled inside the bridge and never start a CLI. Only artifacts the bridge can detect under the user's allowed session paths are pushed back via CDN.

## Features

- Text, image, file, and voice (WeChat server-side transcription only) go to the **active** backend (`agy`, `grok`, or `codex`)
- Text, image, file, and voice (WeChat server-side transcription only) go to the **active** backend (`agy`, `grok`, `codex`, or `dsh`)
- Detected CLI artifacts under the per-user allowed tree can be sent back (size-capped); not every file the CLI touches
- Each WeChat user gets an isolated workspace; model / effort / mode are remembered **per backend**
- Runtime backend switch: `/backend agy`, `/backend grok`, or `/backend codex` (clears that backend's continuation state — the agy/grok continuation flag and the codex `thread_id`/resume state — so the next CLI turn starts a fresh session; history files on disk are not wiped immediately)
- Runtime backend switch: `/backend agy`, `/backend grok`, `/backend codex`, or `/backend dsh` (clears that backend's continuation state — the agy/grok continuation flag and the codex `thread_id`/resume state — so the next CLI turn starts a fresh session; history files on disk are not wiped immediately)
- Slash commands for model, session reset, persona, and more (see below)
- Dangerous-prompt gate: a **keyword list** of concrete destructive patterns asks for confirmation before run
- Sender whitelist (`WECHATBRIDGE_ALLOWED_SENDERS`; empty = allow all)
Expand All @@ -41,8 +41,17 @@ Default data paths expand from `~` (e.g. `~/.local/share/wechatbridge/<instance>
- **agy** (default) — Google Antigravity CLI
- **grok** — xAI Grok Build CLI
- **codex** — OpenAI Codex CLI
- **dsh** — DeepSeek Harness CLI (one-shot `headless` profile)

Per-user switch: `/backend agy`, `/backend grok`, or `/backend codex`. Each backend keeps its own model / effort / mode memory and persona file layout. Global default is `WECHATBRIDGE_BACKEND`.
Per-user switch: `/backend agy`, `/backend grok`, `/backend codex`, or `/backend dsh`. Each backend keeps its own model / effort / mode memory and persona file layout. Global default is `WECHATBRIDGE_BACKEND`.

### dsh backend notes

- **Single-turn:** the `headless` profile always creates a fresh session per invocation (`session-<uuid>`), so every WeChat message starts a new dsh session. `/clear` / `/new` are accepted but are no-ops, and `/model`, `/fast`, `/planning`, `/persona`, `/add-dir` are not wired yet (they return a short "not supported" notice).
- Runs `dsh --profile headless -- <prompt>` with `cwd` = the per-user session directory (per-user workspace; model-created files land there and can be sent back via CDN).
- Image and file attachments are merged into the prompt text as `@/absolute/path` mentions only; whether the `headless` profile reads them has not been verified. The bridge blocks `@/absolute/path` mentions pointing outside the session directory (replaced with `[blocked-path]`).
- Auth / profiles are **machine-wide** (`~/.dsh`, same model as grok's machine-wide login): the child's `HOME` points at the per-user session dir, so the bridge always passes `DSH_HOME` explicitly. Set `WECHATBRIDGE_DSH_HOME` to configure a dedicated service home with automatic session retention cleanup; when unset, it reuses the host `~/.dsh` without automatic session cleanup (managed by the operator). `DSH_BIN_PATH`, `DSH_PROFILE`, and `DSH_TIMEOUT` are configurable.
- Status: implemented from the published `dsh` CLI contract (headless bundle source) plus a fake CLI in the test suite; final acceptance depends on a real `dsh login` + headless run.

### Grok backend notes

Expand All @@ -62,7 +71,8 @@ Per-user switch: `/backend agy`, `/backend grok`, or `/backend codex`. Each back
- **agy** on `PATH`, or set `AGY_BIN_PATH`
- **and/or grok** on `PATH`, or set `GROK_BIN_PATH`
- **and/or codex** on `PATH`, or set `CODEX_BIN_PATH`
- Antigravity is Google's terminal agentic coding CLI (successor to Gemini CLI). Grok Build is xAI's counterpart; Codex is OpenAI's terminal agentic coding CLI.
- **and/or dsh** on `PATH`, or set `DSH_BIN_PATH` (DeepSeek Harness, `dsh login` required)
- Antigravity is Google's terminal agentic coding CLI (successor to Gemini CLI). Grok Build is xAI's counterpart; Codex is OpenAI's terminal agentic coding CLI; dsh is DeepSeek Harness' CLI.
- A WeChat account with a [ClawBot / iLink](https://ilinkai.weixin.qq.com) bot (QR bind on first run)
- Python 3.10+

Expand Down Expand Up @@ -133,10 +143,14 @@ Key variables (all have defaults):
| `AGY_BIN_PATH` | `agy` | path to the agy binary |
| `GROK_BIN_PATH` | `grok` | path to the grok binary |
| `CODEX_BIN_PATH` | `codex` | path to the codex binary |
| `WECHATBRIDGE_BACKEND` | `agy` | global default backend (`agy` / `grok` / `codex`; overridable per user via `/backend`) |
| `DSH_BIN_PATH` | `dsh` | path to the dsh binary |
| `DSH_PROFILE` | `headless` | dsh profile booted for one-shot tasks |
| `DSH_TIMEOUT` | `600` | dsh CLI run timeout in seconds |
| `WECHATBRIDGE_DSH_HOME` | _empty_ | explicit `DSH_HOME` passed to the dsh child. Explicitly set = dedicated home + auto session cleanup; unset = reuse host `~/.dsh` without auto cleanup |
| `WECHATBRIDGE_BACKEND` | `agy` | global default backend (`agy` / `grok` / `codex` / `dsh`; overridable per user via `/backend`) |
| `WECHATBRIDGE_INSTANCE` | `default` | instance name; state / session / QR paths derive from it |
| `WECHATBRIDGE_ALLOWED_SENDERS` | _empty_ | comma-separated WeChat IDs (empty = allow all) |
| `AGY_TIMEOUT` | `600` | CLI run timeout in seconds (all three backends) |
| `AGY_TIMEOUT` | `600` | CLI run timeout in seconds (agy / grok / codex backends) |
| `WECHATBRIDGE_MAX_OUTBOUND_BYTES` | `104857600` | max file size sent back to WeChat (100 MB) |
| `WECHATBRIDGE_MAX_INBOUND_BYTES` | `20971520` | max inbound image/file after download (20 MB) |
| `WECHATBRIDGE_MAX_CONCURRENT` | `4` | global concurrent process slots; same user serial (queue does not hold a slot); extras get a busy reply |
Expand Down Expand Up @@ -222,7 +236,7 @@ See [`deploy/wechatbridge-windows.md`](deploy/wechatbridge-windows.md).
| Command | Action |
|---|---|
| `/help` | list supported commands for the active backend |
| `/backend <agy\|grok\|codex>` | switch CLI backend for this WeChat user (on real change: clears that backend's continuation state — agy/grok flag and codex `thread_id`/resume — so the next turn starts a fresh session; history files may remain until retention cleanup) |
| `/backend <agy\|grok\|codex\|dsh>` | switch CLI backend for this WeChat user (on real change: clears that backend's continuation state — agy/grok flag and codex `thread_id`/resume — so the next turn starts a fresh session; history files may remain until retention cleanup) |
| `/clear` or `/new` | drop continue flag so the next CLI turn is a new conversation (does not instantly delete history files) |
| `/model <name>` | set model (all backends validate against a live list: agy/grok via CLI `models`; codex via `codex debug models` [then `--bundled`]; unknown name or list-fetch failure refuse and do not write prefs; see `/models`) |
| `/models` | list models — agy/grok/codex all query the live CLI (codex: `debug models`; falls back to a built-in reference note only if the live list cannot be fetched) |
Expand Down Expand Up @@ -254,8 +268,10 @@ Other `/…` commands are either rejected (e.g. `/exit`), reported as unsupporte

## Limitations

- Not a standalone agent — requires agy and/or grok and/or codex.
- Not a standalone agent — requires agy and/or grok and/or codex and/or dsh.
- The **codex** backend is not yet verified against a real Codex subscription/CLI; it is validated by source research, a JSONL fixture, and a fake CLI in tests. Treat it as community-tested until a real user confirms.
- The **dsh** backend is single-turn only (the `headless` profile always starts a fresh session) and has not yet been verified against a real `dsh login` + headless run; it is validated against the published headless bundle contract and a fake CLI in tests.
- dsh model/effort/mode/persona slash commands are not wired yet; `/model`, `/fast`, `/planning`, `/persona`, `/add-dir` return a "not supported" notice on the dsh backend.
- Voice is WeChat speech-to-text only; no local ASR; empty transcript → “type instead”.
- No video send/receive; no native WeChat voice-bubble replies (no silk encode).
- One WeChat binding per process; multiple accounts need multiple instances (`WECHATBRIDGE_INSTANCE`).
Expand Down
Loading