Non-Blocking Execution

Non-Blocking Execution Skill

Agents often need to run servers (npm start), watchers (npm run watch), or long builds. Standard execution blocks the agent until the command finishes—which for a server is never.

1. The Async Pattern (Preventing the Hang)

The Problem: By default, run_command waits for the command to finish. For npm run dev, this means the agent hangs forever and never gets to Step 3.

The Solution: You MUST use the WaitMsBeforeAsync parameter.

• This parameter tells the tool: "Run for X milliseconds, then detach and give me back control."
• This is the ONLY way to run a server without killing it or hanging the agent.

Mandatory Parameters

• WaitMsBeforeAsync: Set to 2000 (2 seconds).
  • Effect: The tool runs the command, waits 2 seconds to catch startup errors, and then IMMEDIATELY returns a CommandId while the process keeps running in the background.
  • Result: You regain control to perform Step 2 & 3.

Example

// ❌ WRONG - The agent will HANG FOREVER here.
// It will never reach the next line of code.
run_command({ CommandLine: "npm run dev" })

// ✅ CORRECT - The agent waits 2 seconds, then wakes up.
// The server stays running in the background.
run_command({ 
    CommandLine: "npm run dev",
    WaitMsBeforeAsync: 2000 
})

2. The Verification Loop

Async commands return a CommandId. You MUST verify they are actually running.

1. Launch: Run with WaitMsBeforeAsync: 2000.
2. Wait: Sleep/Tokens (implicit in tool usage).
3. Check Status: Use command_status with the CommandId and WaitDurationSeconds: 5 (to peek at new output).
   • Status: Is it running?
   • Output: Does it say "Server started on localhost:3000"?
4. Confirm: Only proceed once output confirms success.

3. Clean Termination (Ctrl+C)

Trigger: You are done testing the server or need to stop a blocking process. Tool: send_command_input

• Action: Terminate: true
• Why?: Leaving zombie servers eats resources and blocks ports for future agents.
• Equivalent: This is exactly the same as pressing Ctrl+C in a terminal.

Example:

// Stop the server
send_command_input({
    CommandId: "previously-returned-uuid",
    Terminate: true
})

4. Troubleshooting Blocking Commands

If you accidentally run a blocking command (forgot WaitMsBeforeAsync):

1. You will likely timeout or be stuck.
2. In the next turn, IMMEDIATELY use send_command_input with Terminate: true on the blocking command if you can identify it, or ask the user to kill it.
3. Self-Correction: Restart the command with WaitMsBeforeAsync: 2000.

5. Common Blocking Commands (The Block List)

MANDATORY: If you see a command in this list OR a compound command containing one of these (e.g., npm install && npm run dev), you MUST use WaitMsBeforeAsync: 2000.

Compound Commands (Chains)

• Rule: If a command uses &&, ;, or | and any part of it is a blocking command, the entire command is blocking.
• Example: npm install && npm start -> BLOCKING. Use async pattern.
• Example: cd app && python manage.py runserver -> BLOCKING. Use async pattern.

Web & Node.js

• npm start, npm run start, npm run dev, npm run watch, npm run serve, npm run build:watch
• yarn start, yarn dev, yarn watch
• pnpm start, pnpm dev
• npx next dev, npx vite, npx webpack serve, npx nodemon
• node --watch

Mobile (iOS/Android/Cross-Platform)

• npx react-native start, npx expo start
• flutter run, flutter drive
• ./gradlew installDebug, ./gradlew bootRun
• xcodebuild -scheme <Schema> run
• adb logcat (unless piped/limited)

Backend & Systems

• Python: python manage.py runserver, uvicorn, flask run, celery worker
• .NET: dotnet watch, dotnet run
• Java/JVM: ./gradlew bootRun, mvn spring-boot:run
• Go: go run ., air (live reload)
• Rust: cargo run, cargo watch
• Ruby: rails server, bundle exec sidekiq
• PHP: php artisan serve, symfony server:start

Infrastructure & Tools

• Docker: docker-compose up (without -d), docker run (without -d if interactive service)
• Database Consoles: psql, mysql, mongo (interactive shells block)
• Terraform/Cloud: terraform apply (can be long/interactive), kubectl port-forward, kubectl logs -f