@lingxinsuixing/dsh-agnes

Agnes AI image generation and video generation for DeepSeek Harness — plus an optional image recognition that can be switched off.

Tool results are text-only. Generated images and videos are downloaded to a local folder; the tool returns the file path and the URL, never the image bytes.

Install

dsh plugin add github:lingxinsuixing/dsh-agnes

This is the same command the DSH plugin marketplaces print. The package has no build step, so nothing has to be approved before it installs.

Configure

Open Plugins → @lingxinsuixing/dsh-agnes in the Web UI and paste your Agnes API key, or use the DSH credential store:

# ~/.dsh/credentials.yaml  (or the credential store in the Web UI)
AGNES_API_KEY: your-key

The key is looked up in this order: the plugin config (config.apiKey) → the credential store entry AGNES_API_KEY → the environment variable AGNES_API_KEY.

The key you type in the settings page is written to your own profile as config.apiKey in plain text. It is never written into the package, and the package ships without any credential.

Tools

Tool	What it does	Key parameters
agnes_image_generate (main)	Generates or edits an image from a text prompt	prompt, size, ratio, references, saveTo
agnes_video_generate (main)	Generates a video from a text prompt	prompt, mode, seconds, size, aspectRatio, images, saveTo
agnes_image_recognize (optional)	Describes or answers questions about an image	image (local path or http(s) URL), prompt

Every tool returns text: a path, a URL and a short summary. Nothing binary is put into the conversation.

Note on size: the API maps an exact size such as 1920x1080 to the nearest tier, so the size in the result is the real pixel size of the file on disk, not necessarily what was asked for.

Samples

Real output from agnes_image_generate (2K, 16:9), produced with this plugin:

Optional: image recognition

Image recognition is optional. Set enableRecognition: false in the plugin config and the agnes_image_recognize tool is not registered at all — only image and video generation stay. Tools are registered while the plugin loads, so switching it off takes effect after one DSH restart.

Statement. Recognition works without changing the provider: this tool is independent of whether DSH talks to Agnes over openai-responses or openai-completions. Only DSH’s own read_image requires the provider to be switched to openai-completions; without that switch, do not give the model image input, because one read_image bricks the whole conversation.

If your provider talks to Agnes over api: openai-completions, declare reasoningEfforts in the model entry as well. Without it DSH does not send reasoning_effort, the gateway returns no reasoning_content, and the model writes its whole reasoning process into the message body — the thinking chain ends up in the conversation.

Do not use read_image inside the main conversation

Once a model is marked as supporting images, the built-in read_image puts an image content block into the conversation. The Agnes gateway /v1/responses endpoint does not understand input_image (its responses→chat bridge falls back to unknown), so the next request fails with a 400. A 400 is not retried, and every turn rebuilds the whole history from the session log — one bad message bricks the entire conversation and you have to start a new one. Use agnes_image_recognize instead: the image goes to the Agnes vision model and only text comes back into the conversation.

Where output goes

Generated media is written to the folder configured as outputDir (default: media/ under the DSH home directory: $DSH_HOME/media, or ~/.dsh/media when $DSH_HOME is not set). Each call returns the absolute path of the file it wrote. You can also change it from the plugin page (plugin → this package → page → “Output folder and read scope”), together with allowedReadRoots — the folders recognition is allowed to read from (empty = no restriction).

Free models are available: the whole image model line is currently free, and the video line has a free tier (agnes-video-2.5-flash, 720P only). Model list and prices: https://www.agnes-ai.cn/zh-Hans/docs/overview

A saveTo argument must point to an existing directory: the plugin never creates directories at arbitrary paths, and file names are generated by the plugin, so nothing of yours is overwritten.

Local images read by the recognition tool can be restricted with allowedReadRoots (a list of absolute folders). Empty = any readable image file is accepted.

Requirements

• DeepSeek Harness with @deepseek-ai/schemastery >=3.18.1
• An Agnes API key
• Agnes image generation is currently free; video generation is billed by Agnes per generated second (the -flash variant is currently limited to 720P)

License

MIT