What Is a DSH Lark Bridge (and Why Use One)

DeepSeek Harness (the deepseek-ai/deepseek-harness repository) is a plugin-architecture coding-agent framework built around the motto “Everything is a Plugin.” It sits around a ~204k-star ecosystem and sees active development pushes through late August 2026. Because DSH is a local, self-hosted agent harness, it historically lived on your workstation or a private server — reachable only from a terminal.

A DSH Lark bridge changes that. It publishes a lightweight bot into Feishu (China) or Lark (international) and maps the messaging surface onto your harness:

• A group or topic becomes a task inbox. Each Feishu/Lark group maps to a Project, and each topic or thread maps to an isolated Session.
• A message becomes one DSH Agent task. Send a request, and the harness picks it up and executes it.
• A single native card carries the whole lifecycle. The bridge updates one card as work progresses, routes tool-call approvals and structured ask_user_question prompts back to the reader, and relays files and images.

The practical payoff is that you can drive a serious, long-running coding agent from a phone or a group chat — approve a risky file write, answer a clarifying question, watch intermediate tool calls stream in — without ever opening a terminal. For a solo developer with a laptop, or a small team sharing one harness host, this collapses the distance between “I had an idea” and “the agent did the work.”

Prerequisites: DeepSeek Harness, Node.js, API Key, and a Feishu/Lark Account

Before you can bridge DSH into Feishu/Lark, you need the pieces the harness and the plugins all assume are already running. The requirements are consistent across the leading bridge plugins:

The single most common onboarding blocker is not the model key — it is the networking story. Historically, any bridge that wants to receive events from Feishu needed a public endpoint. The modern generation of DSH–Feishu bridges sidesteps this entirely by using Feishu’s WebSocket long-connection API through the official @larksuiteoapi/node-sdk, so no public IP, domain, webhook, or NAT tunnel is required — a detail we will unpack in the WebSocket section below.

The 6 Leading DSH↔Feishu/Lark Bridge Plugins Compared

As of August 2026 the ecosystem has produced at least half a dozen serious bridge projects. They differ meaningfully in setup friction, security model, reliability guarantees, and how aggressively they isolate workspaces. Here is the comparison matrix.

The registries that validate these installs (dsharness.org, dshfind.com, dshbase.com, dsh-plugin.org) point to a fast-growing plugin ecosystem around DSH↔Feishu bridges in 2026.

Choosing the Right Bridge for Your Use Case

Your pick should hinge on three questions: who will install it, how much isolation you need, and whether you trust the reliability layer.

• Choose imetn/dsh-lark-bridge when you want a controller-grade bridge with a --brand flag for ByteDance larkoffice.com tenants and polished card presets (compact / standard / developer). It is tested against DSH 0.1.0-rc.6, so expect developer-preview rough edges.
• Choose bihangchi9-creator/dsh-lark-bridge when strict file isolation is non-negotiable — each chat gets its own <workspaceRoot>/<chatId> folder and per-message .attachments/ storage, so groups never touch each other’s files. Remember the critical rule: build before install, because lib/ is git-ignored.
• Choose amlyczz/dsh-lark-link when you need at-least-once delivery guarantees and a real diagnostic path. Its outbound Outbox (JSONL, idempotent) plus inbound WAL replay and /doctor ZIP diagnostics make it the strongest reliability story.
• Choose ailoushu666/dsh-feishu-bridge when you want the leanest setup (official dsh plugin add, official @larksuiteoapi/node-sdk) and live tool-call visibility with interactive option questions.
• Choose PlutoKeating/dsh-lark-bot when you want parallel multi-task in one group, multi-role agents via /role, and a crash safety-net that keeps replying even if the harness dies.

Step-by-Step Setup with a QR Scan (No Public Network Needed)

The lowest-friction path is the QR-scan onboarding shared by amlyczz/dsh-lark-link and PlutoKeating/dsh-lark-bot. The “30-second setup” story works because the bridge creates the Feishu app for you — it auto-subscribes to message events, group mention, and reaction permissions — so there is no portal spelunking.

1. Install the plugin.

   # amlyczz/dsh-lark-link
   npx dsh-lark-link setup --project "$PWD"  # or the registry install for your repo
   
   # PlutoKeating/dsh-lark-bot
   npx dsh-lark-bot@latest setup --profile dsh-lark
   

   Both require Node.js ≥ 24 (lark-bot accepts ≥ 22) and @deepseek-ai/dsh installed.

2. Scan the QR to create the Feishu app. The wizard auto-creates the bot and grants the message-event subscription perms it needs, so you never touch the Feishu Open Platform console for the common path.

3. Start the harness with the bridge profile.

   dsh --profile dsh-lark
   

4. Add the bot to a group (or DM it) and confirm. Send /help or /lark to verify the connection, then send a real task.

5. Run /doctor or /where if anything is off — the lark-link tool bundles a diagnostic ZIP (session log + redacted config) that is far faster to share than a wall of terminal output.

How the WebSocket Connection Eliminates Public IP, Domain and Tunnels

The #1 reason self-hosted agent bridges stall is networking: “I don’t want to expose my box.” The DSH bridge ecosystem solved this at the platform level. Feishu’s Open Platform supports WebSocket long connections for receiving events, and the plugins implement them via the official @larksuiteoapi/node-sdk.

On a WebSocket long-connection model, your client initiates the connection outbound to open.feishu.cn or open.larksuite.com and keeps it alive. Feishu pushes events (messages, mentions, reactions) down that existing connection. Because the connection is outbound, there is:

• No public IP — your harness host stays behind your NAT/Firewall.
• No domain — nothing to buy, configure, or renew.
• No webhook endpoint — no ingress route to secure.
• No NAT tunnel / ngrok-style relay — one less moving part and one less attack surface.

The bidirectional plugins layer the sending side on the same outbound channel. imetn/dsh-lark-bridge, for example, is explicitly designed as “WebSocket long connections, no public webhook server.” If your use case is Feishu (China), the base URL is https://open.feishu.cn/open-apis; for Lark (international) it is https://open.larksuite.com/open-apis. ByteDance larkoffice.com tenants instead pass --brand larkoffice, and chenxin105/dsh-lark-bridge exposes a baseURL switch for exactly this split.

Managing Sessions, Projects and Workspace Isolation

How a bridge scopes work determines whether five teammates can share one harness safely. The plugins converge on a few patterns:

• Group → Project; topic/thread → Session. imetn/dsh-lark-bridge maps each group to a project and each topic to an isolated session, so parallel conversations do not bleed into each other.
• One chat = one project folder. bihangchi9-creator/dsh-lark-bridge pins each chat id to a stable <workspaceRoot>/<chatId> directory. Groups literally cannot touch each other’s files. Files and images land in isolated .attachments/<messageId>/ per message, with sane limits (5 attachments, images ≤ 10 MB, other files ≤ 20 MB) and a 7-day sweep.
• Session reuse with thread independence. ailoushu666/dsh-feishu-bridge reuses the same DSH Session across messages in a chat to preserve context, while different threads stay independent.
• Per-session git worktrees. PlutoKeating/dsh-lark-bot isolates each session in its own git worktree, with /session, /archive, and /retention commands to prune old workspaces.
• Persistent per-chat sessions. Sessions in bihangchi9-creator/dsh-lark-bridge survive restarts; a /new command clears the current one. This is the difference between a stateless bot that forgets everything and a durable workspace that resumes mid-task.

If you are running parallel, long-lived agent jobs in shared groups, favor the folder-worktree isolation models over a single shared-session bridge.

Approving Tool Calls and Answering Agent Questions from Cards

A coding agent without a human-in-the-loop approval gate is a liability. The bridge plugins build that gate into the Feishu/Lark card itself:

• Card buttons approve or reject tool calls. imetn/dsh-lark-bridge renders card buttons that let you approve a pending tool invocation (or decline it) without touching a keyboard.
• Structured Agent questions resolve in-chat. When the harness calls ask_user_question, ailoushu666/dsh-feishu-bridge surfaces an interactive option list — reply with the option number or text and the agent proceeds. amlyczz/dsh-lark-link uses intent-confirm cards for the same purpose, so an ambiguous decision is confirmed before any side effect.
• Execution is visible. Real-time tool-call hints (🔧) and intermediate replies stream through a throttle queue so you are not staring at an opaque “thinking…” card.
• Replies link back to the triggering message (DM and group @bot by default), keeping the thread readable.

The practical rule: choose a plugin whose approval/rejection flow runs on the card. If your agents can write to files or run commands, you want that gate to be one tap in Feishu, not a return trip to a terminal.

Security: Credentials, Permission Tiers, and Allowlists

Feishu/Lark bridge plugins handle two classes of secret: the platform app credential and tool permissions for the agent. The ecosystem’s approach varies, so read the security model before choosing.

• Credential-resolved tenant_access_token with auto-refresh. chenxin105/dsh-lark-bridge resolves tokens through an auth provider, refreshing automatically, and keeps each tenant’s token fresh. App ID/Secret are read from env (FEISHU_APP_ID / FEISHU_APP_SECRET) rather than hard-coded, and the bridge stores them through a credentials domain after a one-time web UI onboarding dialog.
• role('secret') fields stay out of describe(). Sensitive config fields carry an explicit secret role so they never leak into a harness describe() response or diagnostics bundle.
• Allowlists and permission tiers. amlyczz/dsh-lark-link exposes an open_id allowlist, a groupPolicy (open / mention / keywords / reply), and a denyList fallback — so a bot only responds to the right people in the right contexts. Several plugins add explicit permission tiers (read / write / full) and owner-only admin commands (/allow, /disallow, /model, /preset).
• No public ingress = smaller attack surface. Because the bridge connects outbound over WebSocket, there is no public webhook URL for attackers to discover or abuse.

Before connecting a real repository, set an allowlist, keep the bot’s group permissions narrow, and confirm which of your agent’s tools require approval.

Reliability: Outbox, WAL Replay, and Connection Self-Healing

Agent bridges fail in unglamorous ways: a dropped WebSocket, a crashed harness, a reply lost between the model and the card. The strongest plugins engineer around each failure mode.

• Outbound Outbox (at-least-once, idempotent). amlyczz/dsh-lark-link appends outbound messages to a JSONL outbox before sending, then de-duplicates on re-delivery — so a crash between “I answered” and “Feishu delivered” does not lose the message, and a retry does not double-send it.
• Inbound WAL replay. The same plugin logs inbound events to a write-ahead log and replays on crash, so a task sent right before a restart is not silently dropped.
• Connection self-heal with a QuotaGovernor. Long connections inevitably drop; a QuotaGovernor re-establishes them and paces reconnects so you do not hammer Feishu’s rate limits during an outage.
• Crash safety-net. PlutoKeating/dsh-lark-bot keeps replying in Feishu even if dsh itself crashes; /safemode boots a core-only self-healing config, and /jobs auto-retries queued tasks.
• Reaction receipts. Confirmed delivery is surfaced as reaction receipts (received, then DONE ✅), so you see at a glance whether the agent actually accepted the job.

For anything production-shaped, prefer a plugin with explicit outbox/WAL guarantees over one that fires a message and hopes.

Common Pitfalls and Quick Troubleshooting (/doctor, build-before-install, brand flags)

Even the smoothest bridge has a few sharp edges. These are the four that trip up first-time users.

Remember that DeepSeek Harness 0.1.0-rc.6 is a developer preview — API surfaces move. Pin your plugin and harness versions together.

FAQ: Feishu vs Lark, Costs, and When to Self-Host vs Use a Plugin Registry

What is the difference between Feishu and Lark for this bridge? Feishu is ByteDance’s collaboration app for the China market; Lark is the international build. They share the same Open Platform API but different base URLs: https://open.feishu.cn/open-apis for Feishu and https://open.larksuite.com/open-apis for Lark. ByteDance larkoffice.com tenants use a separate --brand larkoffice flag. Choose your region’s platform and match the plugin’s brand/baseURL accordingly.

Does the bridge need a public IP, domain, or webhook? No. The modern plugins use Feishu’s WebSocket long-connection API through @larksuiteoapi/node-sdk. Your harness connects outbound to the Open Platform and receives events over that persistent connection, so no public endpoint, domain, webhook, or NAT tunnel is required.

Does it cost money to run a DSH Lark bridge? There is usually a fee to enable Feishu/Lark bot or long-connection capabilities in some open-platform tiers, plus you pay for the model calls your DeepSeek Harness agents make via DEEPSEEK_API_KEY. The bridge software itself is open source; you self-host it on your own Node.js ≥ 22 / ≥ 24 machine.

When should I self-host a branch from source vs install from a plugin registry? Use a registry install (dsh plugin add, npx) for convenience and faster updates when the plugin validates on registries like dsharness.org, dshfind.com, dshbase.com, or dsh-plugin.org. Pull and build from source (e.g., with pnpm setup) when you need to patch behavior, need strict build-before-install guarantees, or the plugin has not been packaged.

Which bridge should I pick for first contact? For a first install, choose a QR-scan plugin that auto-creates the Feishu app (amlyczz/dsh-lark-link for its reliability and /doctor, or PlutoKeating/dsh-lark-bot for multi-task and crash safety). Both eliminate the Open Platform console work, connect over WebSocket with no public network, and get you to a working “chat with an agent” in under a few minutes.