📱 DSH Feishu Remote

> Control your DeepSeek Harness agent from Feishu / Lark on your phone. Send a task by DM — get the result back in chat. Fully working closed loop.

  

✨ What it does

• 📨 Feishu → Agent → Feishu: DM your bot a task (e.g. 2+3等于几？), a DSH agent runs it with your configured LLM, and the answer comes back to the chat.
• 🖼️ Image understanding: send a screenshot or photo — it's downloaded and analyzed by a vision model (Qwen3-VL), then the agent replies with what it sees.
• 🎨 Image generation round-trip: the agent can generate an image (generate_image), download it locally (feishu_download), and send the actual image back to your Feishu chat (feishu_send_image) — not just a link.
• 📤 Agent → Feishu: the model gets feishu_send / feishu_send_image / feishu_download tools to push results, files, and generated images to any user or chat.
• 🧹 Retention cleanup: downloaded images are kept for 7 days (configurable via FEISHU_IMG_RETENTION_DAYS), then auto-deleted — no unbounded disk growth.
• 🤖 Long connection: uses lark-cli's WebSocket event bus — no public webhook server needed, works on localhost/LAN/private servers.
• 🔒 Secure: reuses lark-cli's OS-keychain credential storage and permission system; event listener runs unsandboxed by design (it must hold the WebSocket).

🚀 Install

0. Prerequisites

1. A Feishu/Lark self-built app with: - Bot ability enabled - Event subscription im.message.receive_v1 (long-connection mode) - Permissions: im:message, im:message:send_as_bot, im:message.p2p_msg:readonly, im:chat:read, im:resource - A published version - (Setup in the Feishu developer console — the CLI can enable the bot ability via API, but events/permissions need the console.)

2. lark-cli installed & authenticated once:

npm i -g @larksuite/cli
lark-cli config init            # paste your App ID / Secret
lark-cli auth login --recommend # scan QR to authorize

1. Install the plugin

dsh plugin --profile demo add github:ShiXiangYu2/dsh-feishu-remote

The bundle contains two pieces:

• index.js — the Cordis plugin: registers the feishu_send model tool and attempts an in-process event listener.
• feishu-resident.mjs — the recommended resident launcher: boots the web profile and runs the long-connection event loop in a detached process (see below).

2. Configure the model

The DSH profile must have a working LLM route (e.g. DeepSeek via SiliconFlow):

# profile cordis.patch.yml
- id: llm-deepseek
  config:
    apiKeyEnv: SILICONFLOW_API_KEY
    baseURL: https://api.siliconflow.cn/v1
    thinking: disabled
    reasoningEffort: off
    models:
      - id: deepseek-ai/DeepSeek-V3.2
        name: DeepSeek-V3.2 (via SiliconFlow)
        contextWindow: 65536
        maxTokens: 8192

- id: agent-default-model
  config:
    provider: deepseek-official
    model: deepseek-ai/DeepSeek-V3.2

3. Run the resident (recommended)

The closed loop must live in a long-lived process. Use the included resident launcher:

# Adjust the absolute paths in feishu-resident.mjs (LARK_HOME, CLI) to your setup.
DSH_HOME=~/.dsh SILICONFLOW_API_KEY=sk-... \
  node --import tsx/esm feishu-resident.mjs

It boots the web profile, spawns lark-cli event consume as a detached process (holding stdin open via a tail -f /dev/null pipe so the listener never exits on EOF), and for each inbound DM: ack → create agent → run task → extract final text → reply.

4. Use it

DM your Feishu bot anything, e.g. 帮我总结一下 ~/projects 的 README — the agent runs and the result comes back to the chat.

🛠 Tools

🔌 How it works

Feishu DM ──► lark-cli event consume (WebSocket long-connection, detached process)
                  │  NDJSON event on stdout
                  ▼
        feishu-resident.mjs (long-lived process)
                  │  image? → download (messages-resources-download)
                  │          → vision describe (Qwen3-VL via SiliconFlow)
                  │  agents.create + followup(task) + whenIdle()
                  ▼
        final assistant text (ev.data.message.content)
                  │  lark-cli im +messages-send
                  ▼
        Feishu chat reply

Image handling

• The event content for an image arrives as [Image: img_v3_xxx].
• The resident detects that pattern, downloads the resource via

lark-cli im +messages-resources-download (using the event's real message_id + the image_key), then asks a SiliconFlow vision model (Qwen/Qwen3-VL-8B-Instruct, overridable with FEISHU_VISION_MODEL) to describe the picture.

• The description is prepended to the user's message and fed to the DSH agent,

so the agent can reason about the image and reply in Feishu.

• Downloaded images land in the IMG_DIR (/root/dsh /feishu-images by

default; the download command requires a relative --output path, so the resident cds into that directory first).

⚠️ Notes

• Why a resident process? DSH's shell service binds background processes to the calling plugin fiber; a listener started inside a plugin's apply() is killed when the fiber settles. The resident launcher owns the listener in its own process, so it survives.
• Text extraction: the final answer is read from the session log's assistant/message events (ev.data.message.content, mirroring the official headless summarize()).
• Long tasks: replies are truncated to the final text block; very long runs may exceed Feishu message limits.
• lark-cli event output streams as NDJSON on stdout (stderr carries [event] log lines) — both are parsed.

📄 License

MIT