Tools
Listen Mode
By An OpenClaw plugin that teaches AI to shut up and listen first. 让 AI 学会闭嘴，先听完再说。
Install
npm install @im_dd/openclaw-listen-modeREADME
# @im_dd/openclaw-listen-mode
> An OpenClaw plugin that teaches AI to shut up and listen first.
> 让 AI 学会闭嘴，先听完再说。
[](https://www.npmjs.com/package/@im_dd/openclaw-listen-mode)
[](LICENSE)
## What is this?
When users vent to an AI chatbot, the AI typically responds after every single message. This creates an awkward, interrupted experience — like talking to someone who keeps cutting you off.
**Listen Mode** fixes this. When activated, the AI:
1. **Shuts up** and lets the user talk
2. **Buffers** all messages during the silence
3. **Waits** intelligently (dynamic timeout based on punctuation, typing speed, message length)
4. **Responds** with a single, comprehensive empathic reply that addresses everything the user said
```
User: 我朋友跟我说了件事
User: 她男朋友劈腿了
User: 被她闺蜜发现的
User: 然后她男朋友还死不承认
AI: (listening... 🎧)
User: 她现在特别难过
AI: 天呐这也太过分了吧... 被闺蜜发现的还死不承认，你朋友现在肯定又气又伤心
<<<SPLIT>>>
她现在怎么样了？有没有人陪着她
```
## Features
- **State Machine**: `NORMAL → LISTENING → RESPONDING` with stay-in-mode support
- **Dynamic Silence Timeout**: Rule-based 4-layer calculation — punctuation, message length, typing speed, voice detection
- **LLM Completeness Check**: Asks a lightweight LLM "has the user finished talking?" before responding
- **Emotion-Aware Acks**: 6 emotion categories (anger, sadness, anxiety, helpless, gossip, rant) with keyword classification
- **Silence-Based Exit**: 5 minutes of no messages → warm farewell → exit (resets on every message)
- **Intent Classification**: LLM-based task/vent/chat classification with keyword fallback
- **Multi-Round Sessions**: Responds but stays in listen mode — user's conversation is continuous
- **Persona System**: Switchable personas with preset and custom definitions
- **i18n**: Chinese and English support
- **135 Tests**: Comprehensive unit and integration test coverage
## Install
```bash
npm install @im_dd/openclaw-listen-mode
```
## Quick Start
```typescript
import { ListenModePlugin } from '@im_dd/openclaw-listen-mode';
const plugin = new ListenModePlugin();
// or with custom config:
const plugin = new ListenModePlugin({
silenceTimeoutMs: 20000,
maxListenTimeMs: 300000, // 5min silence exit
completenessCheck: {
enabled: true, // LLM completeness check (default: on)
},
});
```
## How It Works
### Trigger → Listen → Respond → Stay
```
User says "听我说" (trigger phrase)
→ AI: "我在，你说 🫶" (entry ack)
→ Enter LISTENING mode
User sends messages...
→ Buffer all messages
→ Send emotion-aware acks every 5-7 messages
→ Dynamic silence timer resets on each message
User stops typing...
→ Silence timer fires (7-30s dynamic)
→ LLM completeness check: "User finished?"
→ No → extend 7s, check again (max 2x)
→ Yes → merge buffer → send to AI agent
→ AI responds comprehensively
→ Stay in LISTENING mode (wait for more)
User says "不聊了" (exit phrase)
→ AI: "好嘞，随时来找我" (farewell)
→ Exit to NORMAL mode
-- OR --
5 minutes of silence after last message
→ AI: "我先在这等你哈，回来了随时继续说 ☺️"
→ Exit to NORMAL mode
```
### Dynamic Silence Timeout
The silence timeout adapts to the user's behavior in real-time:
```
timeout = (base + lengthBonus + voiceBonus) × punctuation × trend
```
| Layer | Rule | Effect |
|-------|------|--------|
| Base | 7 seconds | Default wait time |
| Length | Message > 30 chars | +5s bonus |
| Voice | Last message is voice | +15s bonus |
| Punctuation | `？` question mark | ×0.3 (respond fast) |
| | `。` period | ×0.5 (sentence done) |
| | `...` `，` ellipsis/comma | ×1.5 (still thinking) |
| Trend | Typing faster | ×1.3 (wait longer) |
| | Typing slower | ×0.7 (almost done) |
Clamped to **[2s, 30s]**.
### LLM Completeness Check
When the silence timer fires, instead of responding immediately:
```
Timer fires → LLM: "Has user finished their thought?"
→ {"done": false} → extend 7s, try again (max 2 extensions)
→ {"done": true} → trigger response
→ LLM error → fallback to immediate response
```
Cost: ~140 tokens per check, max 2 checks per round. Negligible.
## Configuration
All options with defaults:
```typescript
{
// Trigger
triggerMode: 'both', // 'manual' | 'auto' | 'both'
sensitivity: 'high', // 'low' | 'medium' | 'high'
languages: ['zh', 'en'],
// Timing
silenceTimeoutMs: 20000, // Base silence timeout (dynamic overrides this)
maxListenTimeMs: 300000, // 5min silence → exit
maxBufferMessages: 20, // Max messages before forced response
// Acknowledgments
ack: {
entryReply: true, // Send "我在，你说" on enter
intervalMessages: [5, 7], // Ack every 5-7 messages
useLLM: true, // Use LLM for contextual acks
llmModel: 'haiku', // Cheap model for acks
},
// Reply splitting (for WeChat multi-bubble)
reply: {
splitEnabled: true,
maxCharsPerMessage: 50,
delayBaseMs: 1500, // Human-like typing delay
},
// Intelligence
intelligence: {
intentClassification: true, // LLM intent detection
fallbackToKeywords: true, // Keyword fallback if LLM fails
},
// Dynamic timeout
dynamicTimeout: {
enabled: true,
minMs: 10000,
maxMs: 30000,
voiceExtraMs: 15000,
},
// LLM completeness check
completenessCheck: {
enabled: true,
model: 'auto', // Use gateway default model
timeoutMs: 1500,
maxExtensions: 2, // Max 2 extensions (2×7s = 14s extra)
extensionMs: 7000, // 7s per extension
},
// Persona
persona: {
default: 'warm-friend',
allowSwitch: true,
},
// Session
session: {
enabled: true,
maxDurationMs: 600000, // 10min session cap
maxRounds: 5,
},
}
```
## Project Structure
```
src/
├── ack/ # Acknowledgment system
│ ├── emotion-category.ts # 6-category emotion classifier
│ ├── entry-responder.ts # Entry/exit/farewell messages
│ ├── fallback-pool.ts # Emotion-specific ack pools
│ ├── interim-responder.ts # Mid-listen ack logic
│ └── llm-ack-generator.ts # LLM-powered contextual acks
├── config/ # Configuration
│ ├── defaults.ts # Default values
│ ├── merger.ts # Deep merge user config
│ └── schema.ts # Zod validation
├── core/ # Core logic
│ ├── state-machine.ts # Main state machine (NORMAL/LISTENING/RESPONDING)
│ ├── dynamic-timeout.ts # 4-layer dynamic silence timeout
│ ├── completeness-checker.ts # LLM "user finished?" check
│ ├── exit-detector.ts # Exit condition detection
│ ├── intent-classifier.ts # LLM intent classification
│ ├── emotion-scorer.ts # Emotion intensity scoring
│ ├── message-buffer.ts # Per-user message buffering
│ ├── message-merger.ts # Buffer → single prompt
│ ├── persona.ts # Persona management
│ ├── listen-session.ts # Multi-round session tracking
│ └── analytics.ts # Session analytics events
├── delivery/ # Message delivery
│ ├── reply-splitter.ts # <<<SPLIT>>> handling
│ ├── delivery-queue.ts # Ordered delivery with delays
│ └── delay-calculator.ts # Human-like typing delays
├── i18n/ # Internationalization
│ ├── keywords-zh.ts # Chinese triggers/exits/tasks
│ ├── keywords-en.ts # English triggers/exits/tasks
│ └── responses.ts # Ack/farewell response pools
├── types/ # TypeScript types
└── plugin.ts # Plugin entry point
```
## Testing
```bash
npm test # Run all 135 tests
npm run test:watch # Watch mode
```
## License
MIT
tools
Comments
Sign in to leave a comment