Include paths from diff/pr-diff if no paths given

[granawkins]

Addresses #239: Use --diff or --pr-diff to set paths if no paths arg given. We used to do this but it must've gotten lost at some point.

Another issue: If the diff/pr-diff is too big for context, it raises a MentatError but doesn't close Mentat. Seems to be a lifecycle issue - the shutdown listeners maybe not yet in place. I'd defer to @waydegg on the best way to address this. Might be related to #237 too - handling errors in the session startup phase.

[jakethekoenig]

I wanted to add a test and it was easier if the logic was in code_context.set_paths then the session constructor. You can verify this test fails before this PR (and passed without the assert, it didn't require any paths to be passed to the python client).

I'm okay with this PR but will wait for someone else to LGTM since I made the last change.

I'm not entirely sure the behavior is what users will expect. I personally would expect all the changed files to always be added to the context if --diff was included. This is a more complicated change but maybe it would be best to include intervals around the actual changes so much larger changes would have a hope of fitting in context?

[jakethekoenig]

conversation_viewer.json
Oh one more thing. I was sort of unhappy with this conversation with mentat (rename to .html, github doesn't like you uploading htmls). I don't like how it said The added lines are marked with a "+" symbol at the beginning.. Not sure how common such a sentence is in diff explanations. Maybe a fluke or we should do some prompt engineering to suppress it.

[biobootloader]

Thanks for the fix @granawkins and the test @jakethekoenig

I'm not entirely sure the behavior is what users will expect. I personally would expect all the changed files to always be added to the context if --diff was included. This is a more complicated change but maybe it would be best to include intervals around the actual changes so much larger changes would have a hope of fitting in context?

Yeah that seems like it'd be the more expected behavior. I'll make an issue to discuss and potentially change this, it'll be easier once we have function-level features instead of whole files, so even diffs touching a lot of files can be added.

Oh one more thing. I was sort of unhappy with this conversation with mentat (rename to .html, github doesn't like you uploading htmls). I don't like how it said The added lines are marked with a "+" symbol at the beginning.. Not sure how common such a sentence is in diff explanations. Maybe a fluke or we should do some prompt engineering to suppress it.

I'll also make an issue for this