deepseek-harness-lark

DeepSeek Harness Lark / Feishu plugin

A DeepSeek Harness channel powered by the official @larksuiteoapi/node-sdk. It uses the official WebSocket transport, so no public callback server is required, and supports both Feishu and international Lark.

Features

• Official WebSocket connection and automatic reconnection
• Direct and group access policies; groups require a bot mention by default
• Text, rich-text, image, and ordinary-file input
• Durable Harness attachments with model-aware multimodal input
• Ordinary files are downloaded into .dsh-lark/inbox/ under the Agent workspace so filesystem tools can inspect them
• Model text replies, generated-image upload, and workspace-file delivery
• Persistent, isolated Harness sessions per chat; /new retains old history and switches to a blank session
• Agent presets are mounted and recorded for both creation and resume; the default is standard
• Registered Harness slash commands execute directly instead of being sent to the model
• /bot-ping, /bot-help, /bot-image-test, /bot-file-test, /bot-status, /bot-cancel
• App Secret resolution through the Harness credential service
• Dormant startup when App ID or App Secret is not configured, so installation alone never blocks Harness Web

Install

pnpm dsh plugin --profile web add github:sliverp/DeepSeek-harness-lark

For a local checkout:

pnpm dsh plugin --profile web add /absolute/path/to/DeepSeek-harness-lark

Configure the Lark application

1. Create a custom application in the Feishu developer console and enable its bot capability.
2. Grant im:message.p2p_msg:readonly, im:message.group_at_msg:readonly, im:message:readonly, im:message:send_as_bot, and im:resource. The im:message:readonly scope is required by Feishu's message-resource download endpoint for user-sent images and files.
3. Select long-connection event delivery and subscribe to im.message.receive_v1.
4. Publish an application version and add the bot to the required chats.
5. Put the App ID in LARK_APP_ID and store the App Secret under the Harness credential reference LARK_APP_SECRET.

Environment injection is supported for development:

export LARK_APP_ID='cli_your-app-id'
export LARK_APP_SECRET='your-app-secret'
pnpm dsh --profile web

For durable use, put the App ID in ~/.dsh/.env and store the App Secret through the Harness credential settings surface. Never commit credentials.

Bundle configuration

- id: lark-channel
  name: deepseek-harness-lark
  config:
    appId: !!js process.env.LARK_APP_ID
    appSecretRef: LARK_APP_SECRET
    cwd: !!js process.env.DSH_LARK_CWD ?? process.cwd()

International Lark and restrictive policy example:

    domain: lark
    singlePolicy: allowlist
    singleAllowFrom: [ou_xxx]
    groupPolicy: allowlist
    groupAllowChats: [oc_xxx]
    groupRequireMention: true
    imageInputMode: auto
    maxInboundFiles: 10
    maxInboundFileBytes: 52428800
    maxInboundMessageFileBytes: 104857600
    maxReplyFiles: 5
    maxOutboundFileBytes: 31457280
    agentPreset: standard

Access policies accept open, allowlist, or disabled. Use allowlists and least-privilege Harness permissions in production.

Inbound images and ordinary files are downloaded through the message-scoped resource endpoint using their matching message_id and resource key; this requires im:message:readonly (or the broader im:message). The separate im:resource scope remains necessary for bot-side image/file upload. Filenames are reduced to safe leaf names, files are created without overwriting existing paths, and each message gets a private directory below <cwd>/.dsh-lark/inbox/. The defaults allow 10 files, 50 MiB per file, and 100 MiB total per message. The model receives the saved path and must use its normal filesystem tools to read the file; file contents are not executed or silently injected into the prompt.

When the model intentionally returns a regular workspace file, it places an explicit Markdown link to that file in the final visible answer. The plugin resolves the canonical path, rejects missing files, directories, symlink escapes, and every target outside the session cwd, then uploads the bounded bytes with Lark's file API. Only the final assistant message is inspected; intermediate tool output cannot trigger a file send. By default, one reply may send up to 5 non-empty files of at most 30 MiB each.

agentPreset is the explicit fallback for new sessions and persisted records without a preset. The tool-loop fix advances the plugin session namespace from lark-v1-* to lark-v2-*: existing files are neither deleted nor rewritten, but the plugin no longer appends to potentially DSML-contaminated lark-v1-* records. They remain available for inspection in the Web UI.

Verify

Send /bot-ping, /bot-image-test, /bot-file-test, and /new. The bot should return pong, a blue diagnostic image, a downloadable text file, and confirmation that it switched to a blank session. Send ordinary text, an image, and a CSV file; ask the bot to inspect the CSV and verify that it uses the downloaded workspace path. Then ask it to create and send a CSV; the final reply should contain the file attachment.

Develop

pnpm install
pnpm run check
pnpm pack

The repository uses PNPM 10.33.4. Plugin runtime requires Node.js >=22.19 and does not require PNPM 11.

Protocol and configuration behavior were cross-checked against larksuite/openclaw-lark. Transport and media operations use the official Lark Node SDK. MIT licensed.