dsh-mimo-agent-tools

mimo-agent-tools

DSH (DeepSeek Harness) Cordis plugin that turns the Xiaomi MiMo API into model tools for an agent: web search, image/audio/video understanding, speech-to-text, and text-to-speech.

Backed by the OpenAI-compatible endpoint https://api.xiaomimimo.com/v1.

Tools

audio-tools skill

The plugin registers an audio-tools skill that guides when to use the four audio tools (mimo_asr, mimo_tts, mimo_voiceclone, mimo_audio). The tools themselves are always registered — they are lightweight pure-API calls — so the skill only teaches usage, it does not gate the tools.

Requirements

• A MiMo API key — the web_search tool requires the connected-search plugin enabled in the MiMo console.
• python3 on the host (used by driver/mimo_driver.py).
• DSH host with the shell and sandboxPolicy services.

Configuration

The API key is resolved at tool-call time from the DSH credentials service first (key name XIAOMI_API_KEY — the web Models page writes keys there), falling back to the environment. Nothing is hardcoded:

Other options (environment, read at apply time):

Install the driver at the default path (or point MIMO_DRIVER at it):

mkdir -p ~/.local/lib/mimo-agent-tools/driver
cp driver/mimo_driver.py ~/.local/lib/mimo-agent-tools/driver/

Why a python driver?

Multimodal payloads are multi-megabyte base64 strings. Agent shells commonly cap captured stdout (DSH's bash seam caps at 64KB), which silently truncates large payloads. The driver therefore does ALL file reading, body assembly and HTTP POSTing inside one python3 process, using spec/response files on disk — nothing large ever crosses the shell.

Install (DSH bundle)

Standard DSH bundle — install with the official plugin command (auto-inits the profile, pnpm-installs, and appends the bundle layer per dsh.bundle.patch):

# From a local checkout, or via git/npm:
dsh plugin --profile web add /path/to/dsh-mimo-agent-tools
# or: dsh plugin --profile web add github:you/dsh-mimo-agent-tools

# Install the python driver to the default path (MIMO_DRIVER points at it):
mkdir -p ~/.local/lib/mimo-agent-tools/driver
cp driver/mimo_driver.py ~/.local/lib/mimo-agent-tools/driver/

# Restart dsh web; the tools mount automatically.

Dependencies are declared as peerDependencies (ecosystem convention — @deepseek-ai/dsh-tools is already loaded in the DSH process, so nothing is duplicated). dsh plugin add installs the bundle into the profile's node_modules where peer deps resolve against the running harness.

Notes on the MiMo API (from the official docs)

• TTS target text goes in the assistant message; the voice description (voicedesign model) goes in the user message; a style instruction rides the user message for preset voices and becomes an inline (风格) tag prefix for voicedesign voices.
• mimo-v2.5-tts-voicedesign does not accept an audio.voice field — it uses optimize_text_preview instead.
• ASR (mimo-v2.5-asr) must not receive a thinking field.
• input_audio.data / video_url.url accept either a public URL or a data:<mime>;base64,... data URL (video base64 capped at 50MB per the docs).
• Web search costs per keyword round (max_keyword, default 3) — see MiMo pricing. force_search (default true) trades freshness against cost; user_location biases results, e.g. {"type":"approximate","country":"China","region":"Hubei","city":"Wuhan"}.
• Deep thinking (mimo_think) returns reasoning_content + content; in multi-turn agent conversations with tool calls, reasoning_content from earlier turns must be echoed back or the API returns 400.
• Structured output (mimo_json) needs an explicit JSON shape description in the prompt (fields, types, nesting); keep max_completion_tokens generous so the JSON is not truncated mid-document.

Tests

python3 tests/test_driver.py   # driver request-body assembly (12 cases)
node --test tests/tools.test.mjs  # tool registration surface (9 cases)

License

MIT