Positron Notebooks: Inline chat

[seeM]

We should support basic inline chat in Positron notebooks.

[rodrigosf672]

@seeM I have some note to add about this.

Yesterday, @cindyytong and I were usability testing Positron on each other and we realized there is currently a confusing experience with Notebooks + Assistant's inline chat. Here's a brief description of what happened:

1. Upon executing script below in a notebook cell, user obtained an error (due to summary being misspelled as summar.

# 1D. Intentional dependency: this will FAIL if you didn't run 1A–1C first.
assert 'summary' in globals(), "Run cells 1A–1C first (or try **Run All Above**)"
print("✅ Dependent step finished. Summary:", summar)

2. Then, user clicked on the error message, obtaining immediate visual feedback upon clicking (i.e., rectangular box around error message turned into a blue outline, confirming that the error message was "selected").

3. User then pressed Cmd+I and asked help from the Assistant, which mistakenly opened a new cell and responded that there was no code for it to analyze.

4. Then, the user highlighted the error message with the cursor, and then asked for help from the Assistant; once again, the Assistant didn't acknowledge the presence of the error message due to it only considering the context of a new, blank cell instead of the output of the previous cell.

5. User then copied and pasted the output error message to the Assistant. Now, it recognized the type for summary, but it did NOT suggest change to fix it.

6. Finally, user was told that they should click inside the cell of interest and then ask help from the Assistant.

Overall, the experience seemed confusing and dissatisfying. I understand there will still be changes and enhancements, hence why you logged this here. Within the context of notebooks, I wonder the following:
1. Do we want to have the Assistant to consider all of the notebook's context or only the specific cell's context?
2. If only the specific cell's context, do we want it to consider the output's context as well?
3. Is there a way to have the output bound to the cell such that, when the Assistant goes over it, it considers both the cell and output's context together?
4. If the output is extremely long/loopy (see script exemplified in #8997), how would we protect/warn the user against excessive use of their LLM tokens?

Please let me know what you think, or we can discuss this further in a future Notebooks meeting. cc @testlabauto @jonvanausdeln @juliasilge

[seeM]

Thanks @rodrigosf672 and @cindyytong, that's valuable feedback! We definitely need to take this into consideration. IMO it's also worth collecting user feedback like this in a Google doc or Miro board. Somewhere where it's easier to collaborate, include images, comment on specific portions, etc.

[jlang-311]

@juliasilge @seeM can we merge this today and before the branch is cut? It is a large testing effort but necessary for the release.

[seeM]

@jlang-311 I wasn't intending on merging before the branch is cut since this is specific to the new notebook UI which is still feature flagged.

EDIT: ... and I needed to rework the implementation.
EDIT2: Bumped the milestone!