gemini --experimental-acp hangs, how to debug?

[tisonkun]

As title. gemini -y works well.

Any method I can leverage, like debug logs or a self-check list

[revanthsaii]

The --experimental-acp flag can hang due to several reasons. Here are debugging methods and solutions:

Debugging Methods:

1. Run with verbose logging:
   geminiacli --experimental-acp --debug

Or set DEBUG environment variable:
export DEBUG=*
geminiacli --experimental-acp

2. Add timeout to prevent indefinite hanging:
   timeout 30 geminiacli --experimental-acp

This will terminate if it hangs for more than 30 seconds.

Common Causes & Solutions:

1. Slow API Response

   • Check your internet connection
   • Verify GEMINI_API_KEY is valid
   • Try testing with a simpler prompt first

2. Authentication Issues

   • Verify your API credentials are set correctly:
     export GEMINI_API_KEY=your_actual_key
   • Test basic gemini -y command first

3. Plugin/Extension Loading Delay

   • ACP may be loading custom plugins
   • Check for any local plugin configurations
   • Try disabling custom plugins temporarily

4. Memory or Resource Issues

   • Check system resources (RAM, disk space)
   • Try running in isolation:
     node --max-old-space-size=4096 geminiacli --experimental-acp

Quick Self-Check List:

• Is GEMINI_API_KEY set and valid?
• Does gemini -y work without hanging?
• Is your internet connection stable?
• Are you running latest version (v0.24.0+)?
• Is there enough free disk space and RAM?
• Are there any firewall/proxy issues?
• Try clearing cache: rm -rf ~/.geminiacli/cache

Try these steps and let me know which debug output you get. That will help identify the exact issue!

[udayvalera]

@tisonkun I don't know if this solves your issue or not. But when I was experimenting with the acp initially, the acp would hang indefinitely for no apparent reason. But what I found out after painstaking debugging was that, the agent (gemini-cli in this case) would send some user approval message in-between, or some tool call that required user input indirectly. That waits for the user input for indefinitely.

In my case, when the agent try to create a new react application through the gemini-cli in-build shell tool.

Shell Tool prompt for the user input "inside" of Shell Tool. For example, select language like javascript or typescript.
ACP is non-interactive, there is no way to give this user input.
ACP hangs indefinitely.

To test this try adding --yolo flag, and re-run or simulate the exact flow to check. Although, yolo mode does not bypass the user input required inside of a tool call.