welsione/dsh-mmx-bridge

dsh-mmx-bridge

One tool = all MiniMax multimodal capabilities. Give DeepSeek Harness (DSH) the ability to see images, generate art, create videos, speak, sing, search the web, and more.

English · 中文 README

---

Overview

DSH is text-only by default — no images, no speech, no video. dsh-mmx-bridge plugs in MiniMax's full multimodal stack through a single mmx_bridge tool. Install once, get 8 capabilities:

v1.0.5+: just drop an image to send it — your input stays untouched. Dropping or pasting an image into the chat input works with text-only models: your message (image + prompt) is displayed and stored exactly as you wrote it. Behind the scenes the plugin saves the image to a temp dir (default /tmp/mmx-out/) and replaces it with "image URL + local path" text — the Agent then automatically calls read_image / mmx_bridge(describe) to look at it. Models that genuinely support image input pass through untouched.

v1.0.7+: image recognition cache (embedded JSON, on by default). Recognition results are written back into the image itself as standard imgjson blocks (PNG tEXt / JPEG COM). The same image + the same question is served straight from the embedded cache on later reads — zero VLM calls. Follow-up questions are cached per prompt layer, never overwriting each other; when the image is re-encoded the cache invalidates and rebuilds automatically.

Compatibility

Install / Uninstall

Prerequisites

1. DSH v0.1.0-rc.7+
2. mmx-cli installed & logged in: npm i -g mmx-cli && mmx auth login

Install

dsh plugin --profile web add dsh-mmx-bridge

⚠️ npm unreachable? Use dsh plugin --profile web add github:welsione/dsh-mmx-bridge

Restart dsh after installing (server-side ESM cache does not hot-reload), then refresh the Web GUI. See AGENT.md for details.

Have your agent install it (paste this prompt to your agent)

Install the DSH plugin dsh-mmx-bridge for me (repo https://github.com/welsione/dsh-mmx-bridge ):

Run `dsh plugin --profile web add dsh-mmx-bridge` to install into the web GUI profile; if npm is unreachable, use `dsh plugin --profile web add github:welsione/dsh-mmx-bridge` instead. Verify the plugin is mounted after installing, then remind me to restart dsh (the settings-page card only appears after a restart).

Uninstall

dsh plugin --profile web rm dsh-mmx-bridge

Quick Start

One tool, the whole multimodal family. mmx_bridge dispatches on action:

Video params (since 1.0.10): duration / ratio are only supported by MiniMax-H3; passing either auto-selects H3 (or set model explicitly: MiniMax-Hailuo-2.3 default / MiniMax-Hailuo-2.3-Fast (i2v only) / MiniMax-H3 / MiniMax-H3-Max). Note that MiniMax Credits / Token Plan accounts do not cover the H3 series — H3 requests fail with error 2013; the bridge translates that into an actionable message. On Credits, omit duration/ratio: the Hailuo-2.3 default yields ~6 s, 16:9.

Just ask the agent in plain language: "describe this image", "generate a cyberpunk cat", "turn this text into speech".

Configuration

Everything is tunable via the Settings → Plugins → Plugin config card or the control file.

Control file (default /tmp/dsh-vision-control.json)

Missing keys fall back to defaults (set false to disable explicitly). Settings-page toggles write to the same file.

mmx environment management (Settings card → "Environment")

The plugin closes the full mmx-cli discovery / config / install / login loop:

1. Auto-scan: control mmxBin > env MMX_BIN > auto-scan (command -v mmx / where mmx, /usr/local/bin/mmx, /opt/homebrew/bin/mmx, npm global dir, …);
2. Not found → configure: type a valid path in the card's "mmx path" field (validated for existence); save empty to clear and rescan;
3. Not installed → one-click install: runs npm install -g mmx-cli (uses the system npm config), then rescans automatically;
4. Not logged in → api-key login: paste a MiniMax API Key and hit Login (runs mmx auth login --api-key internally); the key is never persisted, logged, or echoed; "Login status" button queries mmx auth status live.
5. Model self-healing (no manual config needed): a new mmx_env tool (status / install / login / set-path) is exposed to the agent. When mmx_bridge reports "mmx not found / not logged in / command error", the error message includes a fix hint; the agent can status to diagnose, then install (one-click), login (ask the user for an API Key — never echoed back), or set-path to fix it on the spot, and confirm with status — the user just keeps chatting.

Environment variables (MMX_*)

Permissions & Data

• Generated/bridged files: images, videos, audio are saved to MMX_OUT_DIR (default /tmp/mmx-out/) and served same-origin via /mmx-files/ (HTTP Range, path-traversal guarded).
• Attachment reads: bridge reads image bytes from the DSH attachment store (attachments/v1) by content address — read-only, never written.
• Cache writes: the plugin only writes back to its own bridge-* copies inside the out dir; it never modifies the immutable originals in the DSH attachment store.
• Cache is plaintext: the embedded JSON is not encrypted — anyone who knows the file format can read it. Do not embed sensitive information.
• Re-encoding invalidates the cache: after social-platform re-encoding/compression/screenshots, the embedded cache fails its sha256 check and is reported as stale instead of silently serving old data.

Troubleshooting

Q: Don't see the mmx_bridge tool after install?

Verify DSH version ≥ 0.1.0-rc.7 and mmx-cli is installed (mmx --version). Refresh the Web GUI page.

Q: "Current model does not support images" when sending a picture?

Upgrade to v1.0.5+ and restart dsh (server-side ESM cache does not hot-reload), hard-refresh the page, then check Settings → Plugins → Plugin config → make sure "Image bridge" is enabled. Models that genuinely support images pass through untouched.

Q: How does the recognition cache hit/expire?

Recognition results are embedded into the bridge copy as standard imgjson blocks. The same image + same question (read_image or mmx_bridge(describe)) is served from the embedded cache with cached:true — zero VLM calls. Different questions (follow-ups) get their own cached layer; re-encoding invalidates the cache explicitly and it rebuilds automatically.

Q: Don't see the dsh-mmx-bridge card in Settings → Plugins → Plugin config?

Since DSH rc.7, the settings page dispatches cards by the server-registered settings namespace, so plugin ≥ 1.0.4 is required. Confirm the installed version, restart dsh (server-side ESM cache does not hot-reload), then hard-refresh the Web GUI page.

Q: "API key not found" when generating images?

Run mmx auth login first. If using Token Plan, ensure your subscription is active.

Q: Network issues in mainland China?

mmx-cli connects directly to MiniMax API (api.minimax.chat), usually reachable from China without a proxy. Check your proxy settings if you still have trouble.

Development

git clone https://github.com/welsione/dsh-mmx-bridge.git
cd dsh-mmx-bridge
npm test                 # unit tests (image cache / read_image wrapper, real PNG/JPEG)
npm run check            # syntax checks (4 lib files)
dsh plugin --profile web add .   # local install for testing

Architecture

User chat → DSH Agent → mmx_bridge tool → mmx-cli → MiniMax API
                                                  ↓
                                           /mmx-files/ (same-origin)
                                          (image preview / audio & video players)
Recognition result → imgjson block embedded back into the image (PNG tEXt / JPEG COM)
                  → same image + same question reuses it on the next read

• No third-party runtime deps — Node builtins + @deepseek-ai ecosystem packages (host-provided)
• Same-origin media — generated files served via /mmx-files/ with inline preview
• Web GUI enhancement — image preview, audio/video players, settings card auto-load

License & Security

• MIT: LICENSE
• No telemetry; no system-credential reads; never modifies the DSH attachment store
• Embedded recognition cache is plaintext — encrypt it yourself before writing back if it must stay private

---

Related

If this plugin helps you, please consider giving it a ⭐ Star!