DSH Plugins Marketplace

DSH Plugins

Plugins

/

minimax-image

m

minimax-image

Manifest valid

MiniMax H3 image generation and editing for DeepSeek Harness: composer permission chip, text-to-image and reference-image editing tools, settings page.

UI (client)hasBundlePatch

minimax-image

English | 中文

MiniMax H3 image generation and editing for DeepSeek Harness, as an installable DSH bundle.

  • Composer chip — an Image: Off / Image: On chip beside the permission and mode chips in the composer tool row; the visible words follow the client locale. The default is Off, and the Host remembers each conversation's choice.

  • Two model tools while it is on — switching the chip to On adds both:

    • generate_image — text to image (POST /v1/images/generations);
    • edit_image — edit or combine 1–16 reference images with an instruction (POST /v1/images/edits).

    The model chooses aspect_ratio and area per call; the plugin does not force defaults from settings.

  • Bundled prompt skill — minimax-image-prompting, a detailed playbook that interviews the user for the missing style decisions, locks one art direction across a series, and turns weak results into an actionable next edit.

  • Settings page — configures the endpoint, the model (MiniMax-H3-Image by default, or MiniMax-H3-Image-Highspeed), and the API key.

Install

Requires a DSH profile that composes the Web surface bundle (@deepseek-ai/dsh-web-app), because the chip and the settings page are browser surfaces and the Host route lives on the web server.

From GitHub

dsh plugin --profile desktop add github:yyxcnasd/minimax-image

The package ships plain ESM with no build script, so pnpm runs nothing at install time and no allowBuilds entry is needed. Pin a commit if you want immutability:

dsh plugin --profile desktop add github:yyxcnasd/minimax-image#<commit-sha>

A git install shells out to git ls-remote against github.com:443, so on a network that blocks that host use the npm package or a packed tarball instead (dsh plugin --profile desktop add ./minimax-image-0.4.0.tgz).

From npm

dsh plugin --profile desktop add minimax-image

From a local checkout

dsh plugin --profile desktop add ./minimax-image

dsh plugin --profile <name> add ... forwards to pnpm inside $DSH_HOME/profiles/<name>, then appends the bundle to dsh.profile.bundles. Use web instead of desktop for a web profile — the same bundle serves both.

Installing applies immediately when the profile can recompose; if the Loader reports restart-required, restart DeepSeek Harness once. A browser page refresh loads the client half either way.

Uninstall with:

dsh plugin --profile desktop remove minimax-image

The bundled skill

skills/minimax-image-prompting.md is registered as a model-invocable skill at activation. It is written for the model, not for the user, and it covers:

  1. The interview — at most two questions, each with a recommended default, and only for decisions that change the picture (use and aspect ratio, moment and mood, series style); everything else the model decides and states.
  2. The prompt skeleton — seven slots in order (subject and action, environment, light, camera and composition, medium, colour and texture, mood and detail) with a length budget and a five-item avoid clause.
  3. Aspect ratio and area — a table from use to ratio, and 1K / 2K / 4K guidance with the reporting habit.
  4. Style locking — a reusable six-line style block so a set of images shares one art direction, plus the rule to repeat it verbatim instead of referring to earlier prompts.
  5. Edit discipline — one or two changes per pass, naming what must stay, and how to combine references.
  6. One-lever iteration — a symptom-to-lever table for flat light, cheap colour, plastic texture, busy framing, garbled text, and content-safety refusals.
  7. Templates — product shot, portrait, landscape, food, character sheet, Chinese-text poster, illustration, each with a ratio and area suggestion.
  8. Worked upgrades, working habits, and rights guidance — weak-versus-strong prompt pairs, reporting ratio and area, never reusing a person's likeness without the right to it.

The skill reaches the model through the skill registry (@deepseek-ai/dsh-skill) and the skill tool (@deepseek-ai/dsh-tool-skill); a profile without them still gets the tools and simply skips the registration. If you prefer a filesystem skill, copy the markdown straight into a skills directory — its frontmatter is kept for that purpose, and the registry path strips it.

Configuration

  1. Settings → MiniMax image generation: set the endpoint (https://api.minimaxi.com), the model, and paste your MiniMax API key (sk-api-…). Save.
  2. In a conversation, switch the composer chip to On.
  3. Ask for an image, or ask to change one you generated or attached.

Endpoint and model are stored in $DSH_HOME/minimax-image/settings.json and apply to the next request. The API key is stored in the profile credential store under the reference MINIMAX_IMAGE_API_KEY — never in plugin configuration, never in the settings file. The row in the profile patch only supplies the initial endpoint, model, and credential reference:

- insert:
    - id: minimax-image
      name: minimax-image
      config:
        baseUrl: https://api.minimaxi.com
        model: MiniMax-H3-Image
        apiKeyEnv: MINIMAX_IMAGE_API_KEY

Behaviour notes

  • generate_image posts model, prompt, n, response_format: "url", area, and aspect_ratio (omitted when the ratio is auto). edit_image adds images: [{ image_url }]; local files are inlined as data URIs, capped at 20 MB each and 45 MB per request, so pass an http(s) URL for anything larger.
  • The service derives width and height from the ratio and area and reports them per image as data[].size; the tool result records that exact size beside each saved file.
  • A partial failure is still HTTP 200: every successful image is saved and the failed ones are listed in the tool result. When nothing succeeds, the tool fails with the service's own message.
  • Files land in <session workspace>/generated-images/ as minimax-<timestamp>-<n>.<ext> (generated) or minimax-edit-<timestamp>-<n>.<ext> (edited).
  • Results are text plus paths; open a path with the built-in read_image tool to display the picture inline.
  • The menu is the plugin's own card (portaled to document.body), styled from the shipped menu tokens: ArrowUp/ArrowDown move focus, Enter/Space picks, Escape closes, a pointer press outside closes, and the chevron rotates while it is open.

Privacy

  • Per-conversation switches: $DSH_HOME/minimax-image/sessions.json.
  • Endpoint and model: $DSH_HOME/minimax-image/settings.json.
  • API key: the profile credential store ($DSH_HOME/.credentials.yaml, owner-only permissions).
  • Generated images: the conversation working directory, under generated-images/.
  • Requests go only to the configured endpoint; nothing else is transmitted.

Limitations

  • Streaming (stream: true) is not used; both tools wait for the complete response.
  • HEIC/HEIF references are recognised by extension only.
  • The API key is stored per credential reference, not per model: switching the model keeps the same key.
  • Editing the package's files in place does not hot-swap the running Host module; a code update needs a restart.

Releasing

The registry name is exactly minimax-image — that is what dsh plugin add minimax-image resolves. There is no build step: index.js and client.js are the shipped files, and files already lists every runtime asset.

npm publish
gh release create v0.4.0 --generate-notes   # optional

License

MIT

Comments

Loading…