Introduction

The ecosystem philosophy of DeepSeek Harness is “everything is a plugin.” For developers who need to handle complex video generation tasks, ComfyUI provides high flexibility, but using its native API directly often lacks persistent workflow UI management, and multi-server orchestration can be cumbersome. This plugin aims to solve this problem by importing ComfyUI API JSON into long-lived dynamic workflow pages, and by providing cross-server task queue management and output archiving.

Plugin Overview

dsh-comfy-video-studio is an open-source workflow plugin developed by a1303845406 under the MIT license. It no longer includes specific model templates such as MiniMax and Wan. Instead, it supports importing any JSON in ComfyUI API format, converting it into a visual workflow page, and providing runtime compatibility checks, multi-server concurrency management, and video output management capabilities.

Core Features

  1. API JSON Import and Parsing
    The plugin supports importing JSON generated by ComfyUI’s “Export (API Format)”. It automatically identifies constant parameters within nodes, image or video upload inputs, enum types, model references, and SaveVideo output nodes.

  2. Runtime Compatibility Checks
    Before import or execution, the plugin calls the /object_info endpoint of the target ComfyUI server to check for missing nodes, missing models, invalid enum values, and port types, preventing errors when the task is executed remotely.

  3. Workflow Page Management
    After a successful import, users can save and rename workflow pages, and edit fields. Fields can be modified for display labels, groups, order, and advanced parameter marking.

  4. Multi-Server and Task Queues
    The plugin supports dynamically saving multiple ComfyUI servers. It can detect each server’s latency, device status, VRAM usage, and remote queue status. Each server can be configured independently with 1 to 3 active plugin slots. Tasks that exceed the active slots are kept in a local FIFO queue and wait for execution.

  5. Output Management and Persistence
    All video outputs are saved. Users can designate one of them as the main preview for a task. Workflow, connection, task, and queue data are persisted, ensuring that pending tasks and monitoring states with existing prompt IDs can be recovered after a Harness restart.

Usage Flow

  1. Add or verify the target ComfyUI server on the “Server Management” page.
  2. Upload a JSON file in API format on the “Import” page.
  3. Review the runtime compatibility report generated by the plugin.
  4. In the workflow definition, select the main SaveVideo output node and check the parameters that need to be displayed in the interface.
  5. Save the workflow page, then submit tasks through the dynamic page.

Configuration and Data Storage

The plugin supports basic configuration through a YAML configuration file. The default persistence directory is $DSH_HOME/storages/dsh-comfy-video-studio/, which contains connection settings, history, raw JSON, and page definition files.

Configuration example:

- insert:
    - id: comfy-video-studio
      name: dsh-comfy-video-studio
      config:
        baseUrl: http://127.0.0.1:8188
        maxUploadMb: 100
        maxWorkflowMb: 5
        requestTimeoutMs: 30000
        historyLimit: 200

Notes

  1. JSON Format Limitation: The plugin only supports JSON generated by “Export (API Format)”. This format requires that top-level keys are node IDs, and each node contains class_type and inputs. Canvas JSON will be rejected, and users will be prompted to re-export it.
  2. Authentication: Server addresses accept only HTTP(S) URLs without embedded account credentials. If ComfyUI requires authentication, it should be handled by a reverse proxy.
  3. Dependency Handling: Missing node or model dependencies only prevent task execution; they do not prevent saving workflow pages. The plugin does not perform online lookups or automatically install nodes and models.
  4. Concurrent Execution: A standard single-GPU ComfyUI server usually still executes remote prompts serially. Configuring 2–3 active slots controls the number of tasks the plugin submits in advance, and does not guarantee true hardware-level parallel execution on the GPU.

Conclusion

This plugin provides a standardized workflow entry point and queue management solution for integrating ComfyUI in DeepSeek Harness. Developers can obtain the source code and more details through the community catalog page or the GitHub repository.