Lark Agent Bridge
npm package: @bihangchi9/lark-agent-bridge · git repo: dsh-lark-bridge (repo name kept; clone paths below are unchanged)
A Feishu / Lark bridge for local coding agents — one group, one conversation, one pinned runtime. Bridges dsh (in-process plugin), CLI agents (traex / codex, spawned by a daemon), an IDE window (attached over a socket), or a custom agent — behind one gateway.
中文 README
Send a message in a Feishu chat, and a real coding agent — with its own tools, its own project directory, and its own persistent conversation — answers you right there. Each group chat is an isolated workspace pinned to exactly one runtime, so a team can run several projects (and several agents) in parallel, one per group.
What it does
- Feishu ⇄ your agent. Inbound Feishu messages drive a live agent; the reply streams back onto a live-updating Feishu message. The agent is whichever runtime this chat is pinned to (dsh / CLI / IDE / custom).
- Four host classes, one contract. Every runtime implements the same
AgentAdapter: dsh runs in-process as a plugin; CLI spawns traex/codex; IDE attaches a running window; custom loads your own module. They never silently replace each other.
- One group, one conversation, one pinned runtime. Every chat id maps to a stable directory (
<workspaceRoot>/<chatId>) and one runtime chosen with /agent. Different groups never touch each other's files, and a message is never broadcast to several agents. If the pinned runtime is down (e.g. an IDE window closed), that chat fails closed instead of retargeting.
- Persistent per-chat sessions. A chat's conversation survives restarts (policy-fingerprint-gated resume-or-create;
/new really clears it).
- Files and images. Send them to the bot and the bridge stores each message in an isolated
.attachments/<messageId>/ folder, then gives the paths to the agent. Limits: 5 attachments per message, images ≤10 MB, other files ≤20 MB; names are sanitized and stale files are swept after 7 days. Whether an image can actually be interpreted depends on the selected model's vision support.
- Zero-config setup. On first launch, if no credentials exist, a QR registration wizard runs — scan it in the Feishu app and it connects automatically. No portal spelunking.
- Slash commands.
/help, /new, /where, /models, /agent, and /whoami manage each chat locally; owner-only /agent, /model, /preset, /allow, and /disallow change shared chat state. /agent pins this chat to one installed runtime and never silently retargets.
Architecture in one picture
① Feishu Open Platform ← register a bot here (auto QR wizard does it for you)
│ gives: app_id + app_secret
▼
② lark-agent-bridge gateway ← holds the keys, opens a WebSocket to Feishu,
│ turns each message into one agent turn,
│ routes each chat to its pinned runtime
▼
③ the pinned runtime ← one of:
• dsh — in-process Cordis plugin (`dsh web`)
• CLI — daemon spawns traex / codex
• IDE — daemon attaches a running window over a socket
• custom — daemon loads your own AgentAdapter module
The bot registration lives entirely on Feishu. The gateway connects out to Feishu over a long-lived WebSocket (so no public IP or callback URL is needed). For dsh the gateway is the plugin loaded by dsh web; for CLI / IDE / custom it is a separate daemon (node lib/daemon.js), independent of any dsh host.
pnpm build
node lib/daemon.js # spawn traex/codex on PATH
LARK_BRIDGE_RUNTIME=traex node lib/daemon.js
LARK_BRIDGE_IDE_SOCKET=/tmp/ide.sock node lib/daemon.js
LARK_BRIDGE_CUSTOM_ADAPTER=./examples/custom-adapter.mjs node lib/daemon.js
CLI spawns the binary. IDE attaches a Unix-socket JSONL sidecar owned by the current user and not writable by group/others (chmod 600 /path/to.sock; window closed ⇒ that line dies). Custom loads an AgentAdapter module (see examples/custom-adapter.mjs). One group is still one conversation pinned with /agent; a dead line is not retargeted.
ByteDance-only overlay (SSO, bytecli, extra presets) lives in a local internal/ directory that is gitignored. Do not publish it to this GitHub repo; ship it through the internal skill marketplace.
Requirements
- A working DeepSeek Harness (dsh) checkout you can launch with
dsh web.
- Node.js
^22.19.0 || >=24.0.0.
- A DeepSeek API key (set
DEEPSEEK_API_KEY, or configure it in your dsh credentials).
- A Feishu account to scan the QR code (the wizard creates the app for you).
Install
Option 1: npm package (recommended for users)
The released package contains compiled JavaScript, both access-tier presets,
the bundled dsh-tool-lark-cli package, and the setup scripts. Install it in
a stable directory — dsh links to that location when registering the
bundle.
# macOS / Linux
mkdir -p ~/lark-agent-bridge && cd ~/lark-agent-bridge
npm init -y
npm install @bihangchi9/lark-agent-bridge
bash node_modules/@bihangchi9/lark-agent-bridge/scripts/setup.sh
# Windows PowerShell
New-Item -ItemType Directory -Force -Path "$HOME\lark-agent-bridge" | Out-Null
cd "$HOME\lark-agent-bridge"
npm init -y
npm install "@bihangchi9/lark-agent-bridge"
powershell -ExecutionPolicy Bypass -File node_modules\@bihangchi9\lark-agent-bridge\scripts\setup.ps1
The script preflights Node, installs the lark-workspace / lark-readonly
presets, and registers both the bridge bundle and its dsh-tool-lark-cli
dependency. When dsh is available it uses the official dsh plugin command,
which initializes a missing web or headless profile automatically.
After setup, launch dsh with no --patch flag:
# macOS / Linux
DSH_PERMISSION_MODE=danger-full-access dsh web
# Windows PowerShell
$env:DSH_PERMISSION_MODE = "danger-full-access"; dsh web
Different profile / dsh home:
DSH_PROFILE=headless DSH_HOME=/path/.dsh bash node_modules/@bihangchi9/lark-agent-bridge/scripts/setup.sh
For the standalone CLI daemon only, a dsh profile is not required:
npx -p @bihangchi9/lark-agent-bridge lark-agent-register
npx -p @bihangchi9/lark-agent-bridge lark-agent-bridge
Option 2: source checkout one-command setup (contributors)
Build first. Git contains TypeScript source, while compiled lib/ is
git-ignored. A fresh source checkout therefore has no build output; the
plugin entry is lib/index.js, so registering it without building gives dsh
an empty package and the host fails to load it. pnpm setup builds for
you.
git clone https://github.com/bihangchi9-creator/dsh-lark-bridge.git
cd dsh-lark-bridge
pnpm setup # macOS / Linux (scripts/setup.sh) — builds, links, registers
pnpm setup:win # Windows (scripts/setup.ps1)
The script preflights your Node version, builds the plugin (fails loudly if
the build fails), installs the access-tier presets, and registers both the
bridge and its dsh-tool-lark-cli dependency. When dsh is available it uses
the official dsh plugin command, which initializes a missing web or
headless profile automatically — no preliminary dsh web launch is
needed. After that, launch dsh directly with no --patch flag:
# macOS / Linux
DSH_PERMISSION_MODE=danger-full-access dsh web
# Windows PowerShell
$env:DSH_PERMISSION_MODE = "danger-full-access"; dsh web
Different profile: DSH_PROFILE=headless pnpm setup; custom dsh home:
DSH_HOME=/path/.dsh pnpm setup (both env vars work on Windows too). If no
dsh command is available, setup can only use its manual fallback and
therefore requires an already-initialized profile; it fails before building
or copying presets, so it leaves no partial installation.
Option 3: official dsh plugin command from source (build first!)
git clone https://github.com/bihangchi9-creator/dsh-lark-bridge.git
cd dsh-lark-bridge
pnpm install && pnpm build # REQUIRED — link installs pull lib/ from this dir
# then, from your dsh checkout:
dsh plugin --profile web add link:/path/to/dsh-lark-bridge
dsh plugin runs pnpm add in the profile directory and auto-reconciles
dsh.profile.bundles: a package that declares dsh.bundle joins the layer
stack automatically. Remove/update with the same family:
dsh plugin --profile web remove @bihangchi9/lark-agent-bridge /
dsh plugin --profile web update @bihangchi9/lark-agent-bridge.
⚠️ A link: install points the profile dependency at THIS directory. If you
later move or delete it, the next dsh web cannot resolve the bundle and
fails to boot. Keep the clone in place, or use Option 1.
Option 4: manual install from source (contributors / offline fallback)
Use this only when neither the npm setup script nor the official dsh plugin
command fits. It installs from source alongside your dsh checkout.
# 1. Clone next to your dsh checkout; install & build
git clone https://github.com/bihangchi9-creator/dsh-lark-bridge.git
cd dsh-lark-bridge
pnpm install
pnpm build # compiles src/ -> lib/ (REQUIRED before the plugin can load)
Then register it as a dsh bundle (once it's in the profile, dsh web loads it automatically — no --patch):
# 2. Link it into the profile's node_modules (bundle resolution anchor)
# macOS / Linux:
mkdir -p ~/.dsh/profiles/web/node_modules/@bihangchi9
ln -s "$(pwd)" ~/.dsh/profiles/web/node_modules/@bihangchi9/lark-agent-bridge
# Windows PowerShell (directory junction — no admin rights needed):
# New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.dsh\profiles\web\node_modules\@bihangchi9"
# New-Item -ItemType Junction -Path "$env:USERPROFILE\.dsh\profiles\web\node_modules\@bihangchi9\lark-agent-bridge" -Target (Get-Location).Path
# 3. Append the package name to dsh.profile.bundles in ~/.dsh/profiles/web/package.json:
# "bundles": ["@deepseek-ai/dsh-base", "@deepseek-ai/dsh-web-app", "@bihangchi9/lark-agent-bridge"]
Launch dsh (the bundle loads the plugin automatically):
# from your dsh checkout
DSH_PERMISSION_MODE=danger-full-access dsh web
DSH_PERMISSION_MODE=danger-full-access makes the agent's approval policy never. This is needed because Feishu users cannot click through a local approval prompt. Only use it in an environment you trust.
Platform notes (Windows vs macOS/Linux)
| Item | macOS / Linux | Windows |
|---|
| npm setup (users) | bash node_modules/@bihangchi9/lark-agent-bridge/scripts/setup.sh | powershell -ExecutionPolicy Bypass -File node_modules\@bihangchi9\lark-agent-bridge\scripts\setup.ps1 |
| Source setup (contributors) | pnpm setup (scripts/setup.sh) | pnpm setup:win (scripts/setup.ps1) |
| dsh home directory | ~/.dsh (i.e. $HOME/.dsh) | %USERPROFILE%\.dsh |
| Directory link | ln -s (symlink) | New-Item -ItemType Junction (junction — no admin rights needed) |
| Env var syntax | DSH_PERMISSION_MODE=danger-full-access dsh web | PowerShell: $env:DSH_PERMISSION_MODE="danger-full-access"; dsh web; cmd: set DSH_PERMISSION_MODE=danger-full-access && dsh web |
| Registration URL file | ~/.dsh-lark-bridge/register-url.txt | %USERPROFILE%\.dsh-lark-bridge\register-url.txt |
| Run as a background service | launchd (macOS) / systemd (Linux) | Task Scheduler (schtasks) |
| QR registration / build / chat commands | identical everywhere | identical everywhere |
The two setup scripts behave identically and are idempotent: preflight → build → link → register bundle.
First run: register your bot
Every user registers their own Feishu bot — you cannot share an app_secret, as that hands over control of your bot.
On the first launch with no credentials, the plugin prints a QR code in the terminal (and writes the raw URL to ~/.dsh-lark-bridge/register-url.txt for backgrounded runs). Steps:
- Open the Feishu mobile app, scan the QR code.
- Confirm creating a self-built app on your phone.
- The plugin receives the credentials, saves them to
~/.dsh-lark-bridge/credentials.json, and connects automatically.
- Add the bot to a group (or DM it) and start talking.
Prefer to do it manually / re-register / switch accounts? Run the standalone wizard:
pnpm register # source checkout
npx -p @bihangchi9/lark-agent-bridge lark-agent-register # npm package
Already have credentials? Skip the wizard entirely by exporting them:
export LARK_APP_ID=cli_xxx
export LARK_APP_SECRET=yyy
export LARK_TENANT=feishu # or `lark` for larksuite.com
Using it in a chat
| Command | What it does |
|---|
| (any text) | A prompt to this chat's agent |
/help | Show help |
/new | Start a fresh session (clears this chat's context) |
/where | Show this chat's project directory |
/models | List available providers/models |
/model [provider/model] | (owner only) Show or switch the model for this chat |
/preset [workspace|read-only|full] | (owner only) Show or switch this chat's access tier |
/agent [id] | Show this chat's runtime; owner pins it (dsh today). A dead line is not retargeted |
/whoami | Show identity, this chat's runtime, and authorization |
/allow | (owner only, group chats) Authorize this chat to use the bot |
/disallow | (owner only, group chats) Revoke this chat's authorization |
In a group chat, @-mention the bot to trigger it (unless mention is disabled). In a DM, just send a message.
Configuration
Every field can come from the plugin config: block or an environment variable (env is the friendlier default).
| Config | Env var | Default | Meaning |
|---|
appId | LARK_APP_ID | — | Feishu app id (cli_...) |
appSecret | LARK_APP_SECRET | — | Feishu app secret |
tenant | LARK_TENANT | feishu | feishu (feishu.cn) or lark (larksuite.com) |
provider | DSH_LARK_PROVIDER | dsh default | LLM provider route |
model | DSH_LARK_MODEL | dsh default | Model for created agents |
workspaceRoot | DSH_LARK_WORKSPACE_ROOT | ~/dsh-lark-workspaces | Root for per-chat folders |
allowDm | DSH_LARK_ALLOW_DM | true | Respond in direct messages |
requireMention | DSH_LARK_REQUIRE_MENTION | true | In groups, require an @-mention |
turnTimeoutMs | DSH_LARK_TURN_TIMEOUT_MS | 600000 | Hard deadline for one agent turn; timeout disposes the stuck session |
allowedChats | DSH_LARK_ALLOWED_CHATS | [] | Chat ids allowed to use the bot (comma-separated). Empty = no group allowed (fail-closed) |
allowedUsers | DSH_LARK_ALLOWED_USERS | [] | User open_ids allowed to DM the bot (comma-separated). Empty = only the owner may DM |
accessMode | DSH_LARK_ACCESS_MODE | workspace | Default access tier: read-only, workspace, or full |
|
Credentials are read in this order: inline config → environment variables → the file written by the registration wizard.
Access control (security model)
The bot's security boundary is exactly "who may send it a message": every
message becomes an agent turn with host-level permissions, so access is
deny-by-default:
- The owner always passes. The person who scanned the QR at registration
is the owner (open_id stored in
credentials.json); older installs are
backfilled automatically at startup via the app-info API.
- Groups: only chat ids listed in
DSH_LARK_ALLOWED_CHATS may use the bot.
- DMs: only open_ids listed in
DSH_LARK_ALLOWED_USERS may use the bot
(owner always allowed).
- Fail-closed: with an unknown owner and empty allowlists, every message
is denied — the denial reply includes the chat id so you can configure it.
Example:
# Allow groups oc_xxx1, oc_xxx2 and let ou_friend DM the bot
export DSH_LARK_ALLOWED_CHATS="oc_xxx1,oc_xxx2"
export DSH_LARK_ALLOWED_USERS="ou_friend"
Runtime owner resolution needs the application-info scope; the
registration wizard captures the open_id directly, so usually nothing extra
is needed. Any deployment reachable by people outside your team should
configure the allowlists.
Access tiers (blast radius)
Even after the access gate passes, what an agent may touch is tiered
(DSH_LARK_ACCESS_MODE, default workspace):
| Tier | Preset | What the agent can do |
|---|
read-only | lark-readonly | search/read files only — no writes, no shell, no network |
workspace (default) | lark-workspace | read/write/edit files; no shell, no network, no subagents (no arbitrary code execution). Ships with a lark_cli tool: Feishu operations (IM, docs, sheets, calendar, …) via the host's authenticated lark-cli, spawned as an argv array with a timeout and output cap — Feishu power without opening a shell |
full | deployment default | everything the host offers (shell, network, subagents) |
Presets are dsh's toolset compositions: the host sandbox is identical for
every preset, so the enforceable difference between tiers is which tools
exist. The workspace tier removes the crown jewels of the attack surface —
arbitrary code execution, network egress, and delegation.
The setup scripts (pnpm setup / pnpm setup:win) install the presets into
the harness-home user root automatically. Manual install (discovery is
uncached):
mkdir -p ~/.dsh/.agent-presets
cp -r presets/lark-workspace presets/lark-readonly ~/.dsh/.agent-presets/
Further host-level hardening: dsh's workspace-write permission preset
(sandbox = writes inside the workspace, wider operations require approval)
hard-bounds fs writes — but for remote users "wider" means "denied" (nobody
can click the approval prompt), and it changes shell behavior, so verify it
in your target deployment before enabling. The plugin-level tiers already
remove shell/web/subagents, which is the highest value-per-risk change.
Pairs well with lark-cli
If you already use lark-cli / the Lark skills to drive Feishu (docs, sheets, IM, calendar…), this plugin slots in beside it: keep using lark-cli for structured Feishu operations, and let Lark Agent Bridge be the conversational coding agent living in your group chats. You're very welcome to combine the two — for example, ask the agent in a group to draft something, then use lark-cli skills to push it into a Feishu doc.
Troubleshooting
- Bot says nothing / "(no output)" — make sure a model is resolvable (dsh's default model service must be configured, or set
DSH_LARK_MODEL).
- "missing Feishu credentials" — the wizard didn't complete; re-run
pnpm register or export LARK_APP_ID / LARK_APP_SECRET.
- QR not visible (backgrounded dsh) — open the URL saved at
~/.dsh-lark-bridge/register-url.txt in a browser.
- Group messages ignored — you must
@-mention the bot, or set DSH_LARK_REQUIRE_MENTION=false.
Credits
Lark Agent Bridge (npm @bihangchi9/lark-agent-bridge, repo dsh-lark-bridge) is a creative extension of lark-coding-agent-bridge (originally feishu-claude-code-bridge) by zarazhangrui, by way of trae-to-lark. This project is a native DeepSeek Harness plugin reimplementation. All original work remains under its MIT license; see LICENSE and NOTICE for the full copyright chain.
License
MIT
