feat: agent supports dynamic instructions

[sigoden]

This PR implement agent dynamic instructions.

A jsonviewer agent at sigoden/llm-functions#134

close #1019

[einarpersson]

Hi!

Hmm interesting. Since my original proposal was rejected I resorted to creating a wrapper function. The wrapper does the following:

1. Sources a script with some common exported env vars
2. Optionally sources a similar file for the specific role being in use
3. populate any env vars in the role markdown file with envsubst with actual values
4. create a temp role file and call aichat binary with a temp role folder specified

If I were to adopt this new way then I guess I need to transition from roles to agents (I actually liked writing instructions in markdown files, but I guess I can create a util if i would like to keep providing instructions for agents in a INSTRUCTIONS.md file i really wanted to)...

EDIT: Hmm, but agents still can't use shared tools right? Or can they use toolsets now? Don't see anything about it in the docs...

EDIT 2: I just realized another limitation. By using my solution I can make the shell integration role a lot smarter and more reliable by increasing the context given. However, this new feature does only apply to agents as I understand it. Can we make it so that you can override the default shell integration with a %shell% agent? (instead of just a role?)

[sigoden]

Hmm, but agents still can't use shared tools right? Or can they use toolsets now?

1. Agent has tools.txt. For example, https://github.com/sigoden/llm-functions/blob/main/agents/coder/tools.txt
2. Agent configuration has use_tools field that adds addition tools,

   aichat/config.agent.example.yaml

   Line 6 in 6b6413b

   use_tools: null # Which additional tools to use by agent. (e.g. 'fs,web_search')

Can we make it so that you can override the default shell integration with a %shell% agent?

No. Here are the reasons:

1. It's not necessary. For most people, %shell% role is good enough.
2. %shell% role can be easy to customize, try .edit role.
3. The agent needs to install llm-functions.

[einarpersson]

Thanks for pointing out the shared tools feature, I have been reading the docs but somehow didn't see it.

Hmm, don't know really what .edit role is supposed to do

temp/assistant) .role %shell%                                                                                                                              0

temp/%shell%) .edit role
Error: No role

but I imagine that it still won't be able to provide the dynamic information I am after.

I have to think about this. I would of course like to just your builtin feature rather than my home-made wrapper (but I have to admit really enjoyed just being able to put in arbitrary env vars in the markdown files, it felt very convenient. And I do feel it is very useful to have dynamic context available for the shell assistant).

Hmm, wait - I guess I could just replace the shell integration to use my shell-agent and everything will be fine?

PS. Hope you don't read my comments as criticism, I am just thinking out loud and sharing experiences. Thank you once again for this project.