ppy-web/dsh-plugin-xiaomi-mimo-tts
ppy-web/dsh-plugin-xiaomi-mimo-tts为 DSH 提供小米 MiMo、豆包和浏览器本地语音朗读,支持自动播报、自定义音色与可选 UI 音效。
Your DeepSeek Harness agent rings your actual phone. It asks the question out loud, you answer out loud, and what you said goes straight back into the run.
English | 中文
Every other reachability plugin sends you a notification. This one places a phone call. Your phone rings through CallKit like any other call, you pick it up, a voice reads the agent's question, you say "yes, ship it, but hold the migration", and the transcript of that sentence is what the agent reads next. No tab to come back to, no app to open, no keyboard.
you walk away -> agent finishes -> your phone rings -> you answer out loud -> the run continuesdsh plugin --profile web add github:radres/dsh-plugin-call-mePlain JavaScript, no build step, so a git install needs no allowBuilds permission. Then pair a phone:
cordis.patch.yml:- insert:
- id: call-me
name: dsh-plugin-call-me
config:
number: '5551234567'If this machine already runs /call-me for another agent, skip step 3: the number in ~/.aiphone/config.json is picked up automatically, and both agents reach the same phone.
Nothing rings until a number resolves. An unpaired install is inert on purpose.
curl -sS https://serdaroztetik.com/aiphone/ring \
-H 'content-type: application/json' \
-d '{"to":"<YOUR_10_DIGITS>","text":"Can you hear me?","from":"dsh"}'That is the same call the plugin makes. It blocks, rings your phone, and prints what you said back.
Two tools the model can call on purpose
Tool · What it does call_me · Rings your phone, speaks one question, waits, returns your spoken answer transcribed. Blocking, up to 5 minutes. text_me · One-way text. Returns immediately, and tells the model when your phone did not actually show it.
Three things that happen without the model asking
line quoting what the agent last said. Type anything in that window and your phone stays quiet, so being at the keyboard costs you nothing. Set turnEnd.mode: call and it calls instead, then feeds your answer back with agent.steer(): a finished run picks itself up from what you said.
(approval.mode: answer) ring you and take the decision from your voice. Anything that is not a clear yes denies, and an unanswered call hands the question back to whoever is at the keyboard. Silence never approves.
hour later and it lands in that session: an idle run wakes up (agent.followup()), a busy one picks it up at its next step (agent.inject()).
Every field has a default that works. Add only what you want to change, and remember that a patch row replaces the whole config value, so restate the keys you need.
Field · Default · Meaning number · '' · The 10-digit number the app shows. Empty means read ~/.aiphone/config.json, then $CALLME_USER_NUMBER. label · DSH: · Thread name on the phone. One project keeps one conversation. callTimeoutSeconds · 300 · How long a call waits for an answer (30 to 300). quietSeconds · 900 · Minimum gap between UNSOLICITED contacts. Tools the model calls are never throttled. turnEnd.mode · text · off, text, or call when a run stops. turnEnd.graceSeconds · 120 · Wait this long first; typing cancels it. Use 0 for one-shot headless runs. turnEnd.reasons · completed, blocked, error · Which turn endings are worth a phone. Also available: max-tokens, aborted. approval.mode · text · off, text, or answer (decide by voice). inbound.enabled · true · Deliver texts you send from the phone into the running session.
Ring-the-phone-when-I-stop, and let me answer by voice:
- insert:
- id: call-me
name: dsh-plugin-call-me
config:
number: '5551234567'
turnEnd:
mode: call
graceSeconds: 60
approval:
mode: answerNo core is patched. Everything is an ordinary Cordis plugin on a documented extension point:
Feature · Mechanism call_me, text_me · ctx.tools.register() with raw JSON-Schema tool definitions when to use them · one ctx.systemPrompt.section(), resolved per assembly so it names your live number turn-end reachability · session/event (turn/end to arm, user/message to stand down, assistant/message for what to quote) resuming a finished run · agent.followup() on an idle agent, agent.steer() on a busy one approvals · the approval/request waterfall, which returns allowed-once or rejected, or delegates with next() replies from the phone · a long poll on the /call-me event stream, stopped on agent/disposed
Three rules the code keeps, because they are what makes a plugin like this safe to leave installed:
approval/request and the turn-end path sit inyour agent's own control flow. A phone network blip must not end a run.
every turn gets uninstalled by lunchtime.
/call-me is a hosted service plus an iPhone app. Calls arrive over CallKit and VoIP push, what you say is transcribed live and not stored, and the agent never learns your real phone number: it dials a 10-digit /call-me number that belongs to the app.
Known limits, stated plainly:
That is the cost of taking the decision by voice; the other modes hold nothing.
and denies everything else, but do not point it at commands you would not want decided by a sentence.
npm install
npm test # 24 offline tests: a fake Cordis context and a recording fetch, no phone rings
npm run check # syntaxMIT. Issues and PRs welcome.