| name | smart-dispatch |
| description | 智能调度层——接到用户自然语言指令后，自动拆解步骤、从技能库中选择最快工具组合执行。使用时机：用户说"帮我XX"时。| Smart dispatch — decomposes natural language commands into steps, picks the fastest tool for each step from the skill library. Use when: user says "help me do X". |
Smart Dispatch — Claude Code Skill
This is a meta-skill. It doesn't add new commands — it teaches the AI how to compose existing skills in the fastest way.
Principle
Every user command is decomposed into tool-independent steps. For each step, pick the fastest available tool from this priority ladder:
| Priority | Method | Speed | Use when |
|---|
| 1 | content query/insert | 0.5s | DB access (contacts, SMS, calendar) |
| 2 | am start / am force-stop | 1-2s | Open app, kill app, open URL |
| 3 | mimic dump --filter interactive | 0.5s | UI element tree on all apps with accessibility data |
| 4 | uiautomator dump (elements.py) | 1s | Standard Android apps (native Views only) |
| 5 | input tap/keyevent/text | 0.3s | Execute click/type after coordinates found |
| 6 | cmd clipboard + paste | 0.5s | Long Chinese text input |
| 7 | dumpsys (battery/power/wifi) | 0.5s | Read system state |
| 8 | screencap + vision AI | 8s | 🚨 Last resort only |
Golden rule: mimic first (0.5s, all compliant apps), uiautomator second (1s, AOSP apps), vision last (8s). Never jump to vision without trying the first two.
App Compatibility Matrix (vivo S19 verified, 2026-07-22)
| App | mimic | uiautomator | Notes |
|---|
| 系统设置 | ✅ 0.5s | ✅ 1s | 完美 |
| Termux | ✅ 0.5s | ✅ 1s | 完美 |
| 微信 | ❌ | ❌ | 腾讯不填无障碍数据，自研渲染引擎 |
| 抖音 | ⚠️ 未实测 | ⚠️ 未实测 | monkey/am start 被 vivo 双开拦截，待测 |
| 大部分原生 App | ✅ | ✅ | 标准 Android View 的都能读 |
Available Tool Catalog
Communication
| Task | Command | Speed |
|---|
| Find contact number | content query --uri content://contacts/phones | 0.5s |
| Send SMS | content insert --uri content://sms ... then am start broadcast | 2s |
| Read SMS inbox | content query --uri content://sms | 0.5s |
| Open WeChat to chat | am start com.tencent.mm → elements.py --text "搜索" --tap | 3s |
System
| Task | Command | Speed |
|---|
| Battery | dumpsys battery | 0.5s |
| WiFi on/off | svc wifi enable/disable | 0.5s |
| Flashlight | cmd flashlight on/off | 0.3s |
| Storage | df -h /sdcard | 0.3s |
| Uptime | cat /proc/uptime | 0.2s |
Media
| Task | Command | Speed |
|---|
| Screenshot | screencap -p /sdcard/screenshot.png | 1s |
| Play/pause | input keyevent 85 | 0.3s |
| Next track | input keyevent 87 | 0.3s |
| Open camera | am start -a android.media.action.STILL_IMAGE_CAMERA | 1s |
UI Interaction
| Task | Command | Speed |
|---|
| Find button by text | python3 elements.py --text "确认" | 1s |
| Find + tap in one step | python3 elements.py --text "确认" --tap | 1s |
| Type Chinese text | cmd clipboard set-text "内容" → input keyevent 279 | 0.5s |
| Tap coordinates | input tap X Y | 0.3s |
| Scroll | input swipe X1 Y1 X2 Y2 300 | 0.5s |
| Back | input keyevent 4 | 0.3s |
| Home | input keyevent 3 | 0.3s |
App Management
| Task | Command | Speed |
|---|
| Open app | am start <package> or monkey -p <package> 1 | 1s |
| Force stop | am force-stop <package> | 0.5s |
| List installed | pm list packages | 1s |
| Uninstall | pm uninstall <package> | 2s |
File System
| Task | Command | Speed |
|---|
| Recent photos | `ls -lt /sdcard/DCIM/Camera/ | head -10` |
| Downloads | `ls -lt /sdcard/Download/ | head -10` |
| Open file | am start -d "file:///path" -t "mime/type" | 2s |
Composite Workflow Templates
"Send WeChat to X saying Y"
1. am start com.tencent.mm (launch WeChat)
2. python3 elements.py --text "搜索" --tap (tap search bar)
3. cmd clipboard set-text "X"; input keyevent 279 (paste contact name)
4. python3 elements.py --text "X" --tap (tap contact)
5. python3 elements.py --id "input" --tap (tap input field)
6. cmd clipboard set-text "Y"; input keyevent 279 (paste message)
7. python3 elements.py --text "发送" --tap (send)
Total: ~5s
"Send SMS to X saying Y"
1. content query --uri content://contacts/phones (find X's number)
2. am start -a android.intent.action.SENDTO -d sms:NUMBER
3. input text "Y"; input keyevent 66 (send)
Total: ~3s
"Take a screenshot and tell me what's on screen"
1. screencap -p /sdcard/screenshot.png
2. python3 vision.py /sdcard/screenshot.png "what's on screen?"
Total: ~8s (vision is the bottleneck)
"Open file /sdcard/Download/report.pdf"
1. am start -d "file:///sdcard/Download/report.pdf" -t "application/pdf"
Total: ~2s
"Music moment — spontaneous song when headphones connected"
# Shizuku
1. rish -c 'dumpsys audio' | grep bt_a2dp (confirm headphones active)
2. rish -c 'input keyevent 87' (next track → auto-resumes any music app)
3. rish -c 'dumpsys media_session' | grep metadata (read title + artist)
# ADB fallback
1. adb -s 127.0.0.1:5555 shell dumpsys audio | grep bt_a2dp
2. adb -s 127.0.0.1:5555 shell input keyevent 87
3. adb -s 127.0.0.1:5555 shell dumpsys media_session | grep metadata
# Then:
4. WebSearch "artist song lyrics" (find lyrics)
5. Continue the conversation about the song (natural chat)
Total: ~5s
Decision Logic
When the user says "do X", follow this decision tree:
- Is X a single system command? → Run it directly (battery check, WiFi toggle, etc.)
- Is the target app WeChat? → Skip uiautomator/mimic, go straight to vision fallback (Tencent blocks accessibility data)
- Does X involve finding a UI element? → Try mimic first (0.5s), then
elements.py (1s). Never use vision when either works.
- Does X involve typing Chinese? → Use clipboard paste, not
input text
- Does X span multiple apps? → Chain the fastest method for each step
- Is the screen layout unknown/novel? → Fall back to screenshot + vision (8s last resort)
Always report: what you did, which tools you used, and how long it took.
