Allow configuring a default model for cell magics (and line error magic)

[krassowski]

• Fixes Default model for %%ai magic command #372
• Fixes Cell magic uses incorrect IPython shell #952 (this was required to add tests here)

The name and help for default_language_model is copied over from the one used in equivalent jupyter-ai traitlet.

[haesleinhuepf]

Awesome @krassowski! Fantastic to see this happening.

Just out of curiosity: Would it be possible to make the syntax more intuitive, e.g. instead of this:

%config AiMagics.default_language_model = "anthropic:claude-v1.2"

One could also say

%ai set default llm to anthropic:claude-v1.2

That's human readable and easy to memorize, even though could be even shorter...

[krassowski]

I would argue that %config is preferred as this is the standard way to change configuration in IPython, see:

• https://ipython.readthedocs.io/en/stable/config/intro.html#the-config-magic
• https://ipython.readthedocs.io/en/stable/interactive/magics.html#magic-config

%ai magic uses click parser and syntax that you propose would not be easy to implement. I think it might be worth considering as a separate enhancement.

[srdas]

@krassowski - Looks great! Have wanted this for some time - TY. I tried various things to break it - but it's all good.

Note: Why not default to the Text Completion LLM set in Chat always (if there is one set), and only use the %config magic if you want to override and use a different default LLM for the magics?

The documentation add-ons are also good.

Code to correct cell magic use of shell looks good.

[krassowski]

Note: Why not default to the Text Completion LLM set in Chat always (if there is one set), and only use the %config magic if you want to override and use a different default LLM for the magics?

Maybe in a future PR. It is rather complex quest because the chat and ai magics do not necessarily live in the same environment (may be physically different computers). While chat imports a copy of the magics package, it is potentially a different copy than the one used in the kernel. There are ways to solve this, I explored some of the ideas in the linked issue.

[srdas]

A few more checks on the PR:

1. Check that setting the LLM with magics %config does not overwrite the specified chat LLM.✅
2. Does setting the default LLM in magics also carry over to a new notebook in the same Jupyter session? No it does not, which is good so that each notebook can have it's own default LLM for magics. ✅
3. If you put in a non-existent LLM as default, check that it fails. Then if we use the %%ai magics with a different LLM and then go back to default, make sure that the default is not overwritten. ✅
   
   All looks good.

[3coins]

@krassowski
I am not deeply familiar with magics, curious if this change has any implications or is a drop-in replacement for the existing statement?

[krassowski]

It is drop-in replacement, except that it now can be properly tested and used outside of IPython interpreter. There is a reason why Magics constructor receives the shell argument. get_ipython() should never be used in magics.