Add support for commands in the CLI/TUI/API

[krissetto]

Proposal for adding support for commands in the CLI/TUI/API

Example yaml file

version: "2"

agents:
  root:
    model: openai/gpt-5-nano
    instruction: you are a helpful computer assistant
    commands:
      df: "check how much free space i have on my disk"
      ls: "list the files in the current directory"
    toolsets:
      - type: filesystem
      - type: shell

  another:
    model: openai/gpt-5-nano
    instruction: you are a helpful computer assistant
    commands:
      - fun: "tell me a joke about computers"
      - info: "tell me some information about computers"

Example CLI usage

# run a command
$ cagent run test.yaml -c df
... opens tui and sends the chosen command as the user message

$ cagent run test.yaml --command ls
... opens tui and sends the chosen command as the user message


# list commands
$ cagent run test.yaml -c

For any feedback, please visit: https://docker.qualtrics.com/jfe/form/SV_cNsCIg92nQemlfw

Available commands for agent 'root':
 - df: check how much free space i have on my disk
 - ls: list the files in the current directory

closes #128

[rumpl]

I think I would prefer if we could run a command with cagent run agent.yaml df instead of -c df

And maybe have cagent commands agent.yaml ?

Things got merged and this has conflicts now, sorry

[krissetto]

conflicts resolved. i'm fine without the -c, but do we still want to support messages? right now cagent run agent.yaml hello works and that conflicts with what you're saying.

i can add the commands command if we want that

[krissetto]

-c without a command currently does the listing like i assume you want cagent commands agent.yaml to do, just in case a user misuses the flag

[rumpl]

cagent run agent.yaml --command that lists commands definitely looks weird to me.

What about:

$ cagent command ls agent.yaml

$ cagent command run agent.yaml <command>

I'm only asking mind you :)

[krissetto]

Oh listing with --command/-c wasn't meant as the way to list available commands, but more as a ux thing for users who mess up so they don't have to type another command to see what commands are available.

I'm ok with either approach really, having a dedicated cagent command for commands (lol) feels a bit overkill to me right now for what we're trying to do, which is just have a default first message from the agent file sent automatically, but I can make those changes if that's what we agree upon. Would love to hear from the others as well but if I don't then I'll just make these changes and move on with other things :)

[krissetto]

Unless there are hard feelings around this, I'd honestly prefer if we merge this as is for now (meaning without a new top level CLI command) so this doesn't remain stuck here, and then iterate if we need to.

After using it a bit we can decide if we feel the need to make a top level cagent command command, especially if we feel the need for more features than currently implemented. To me it feels like a variation of the run command that's too small to warrant it's own top level command from the get-go

[rumpl]

Let's go