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 — plus, since the single-repo merge
(docs/adr/0001, absorbing dsh-voice-mimo), the voice UI: a 🔊 read-aloud
button on every assistant reply, in-conversation speech strips for stored
mimo_tts output (play / download / regenerate), and a configurable voice
map in the Plugins settings tab.

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:;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 (14 cases)
node --test tests/tools.test.mjs        # tool registration surface
node --test tests/mimo.test.mjs         # ./mimo TTS/ASR client (mocked fetch)
node --test tests/speech.test.mjs       # speechToText provider surface
node --test tests/package-surface.test.mjs # exports map + dsh.client contract
node --test tests/audio-store.test.mjs  # audio store (manifest/retention)
node --test tests/web.test.mjs          # host routes (speak/audio/regenerate)
node --test tests/tts-store.test.mjs    # mimo_tts store mode + Config wiring
# or simply: node --test   (discovers all of the above)

Reusable MiMo client (./mimo)

mimo.js exports a ctx-light MiMo TTS/ASR client — plain node fetch,
no shell/sandbox/credentials coupling (pass the apiKey yourself; inject
fetchImpl in tests). The python-driver tools above are untouched; this
entry exists for host-side consumers such as the merged voice UI's read-aloud
routes (single MiMo transport — see docs/adr/0001; the ./client export is
reserved for the browser UI entry per the DSH client-bundler convention):

import { createMiMoClient } from 'dsh-mimo-agent-tools/mimo'
const mimo = createMiMoClient({ apiKey: '...' })
const { bytes } = await mimo.speak({ voice: 'Mia', text: '你好' })          // preset
await mimo.speak({ voice: '温柔的女声', text: 'hi', style: '轻快' })         // voicedesign
await mimo.speak({ voice: 'me', voiceMap, reference: dataUrl, text: '...' }) // voiceclone
await mimo.transcribe({ audio: base64Wav, language: 'zh' })

Voice-map semantics (preset / voicedesign / voiceclone resolution via
resolveTtsTarget, style channeling via applyStyle) stay
caller-configurable — pass your own voiceMap.

Voice UI (merged from dsh-voice-mimo)

The ./client entry is the browser half (DSH client-bundler convention,
dsh.client in package.json):

• 🔊 read-aloud — a speaker button on every assistant reply; the host
  synthesizes through the shared ./mimo client and the browser plays the
  same-origin audio.
• Speech strips — mimo_tts with store: true stores the artifact
  (/cache/dsh-mimo-agent-tools/long/ + manifest) and renders a
  play / download / regenerate strip in the conversation.
• Retention — artifacts of archived sessions are cleaned (client watcher
• host startup sweep); a loose count/days fallback bounds the store. The
  manifest record survives cleanup so 「↻ 重新生成」 can replay the exact
  request.

Configuration (voice map, 朗读音色/朗读语气, audio dir, retention, inline
threshold) lives in the plugin Config, surfaced through the Plugins
settings tab.

Dropped with the merge (issue #4): the 🎤 mic button (covered by the official
speechToText provider in ./speech), the understand-audio composer button
(mimo_audio is the agent-side path), the voice_transcribe / voice_understand
/ voice_speak tools (functional duplicates of mimo_asr / mimo_audio /
mimo_tts), and auto read-aloud notifications (never used).

Official voice input (speechToText provider)

speech.js is a separate plugin entry that plugs Xiaomi MiMo ASR
(mimo-v2.5-asr) into dsh's official voice-input stack as a cloud
provider: the built-in 🎤 capture → RPC → provider-registry pipeline stays
untouched; recordings are transcribed by MiMo and the transcript lands in
the input box.

• The entry stays dormant (inject: ['speechToText', ...]) on deployments
  without @deepseek-ai/dsh-experimental-voice-input-bundle — the agent
  tools are unaffected either way.
• The bundle patch overrides the speech-to-text row's defaultProvider
  to mimo. Bundle order matters: this package must compose AFTER the
  voice-input bundle in dsh.profile.bundles (the override only lands on an
  already-inserted row). SenseVoice stays registered and can be re-selected
  in the official voice settings.
• Credentials: the same XIAOMI_API_KEY entry in the DSH credentials
  service the tools use.

License

MIT