Insert newlines after streaming tool results

[jdormit]

This is a fix for #769. It ensures that we insert a gptel-response-separator after a tool result comes in in streaming mode. This prevents weird formatting when gptel-include-tool-results is nil and the tool result came in between two chat responses (see the linked issue for details).

[karthink]

Thanks for the PR.

This is trickier than it appears. I need to test some edge cases when I can before merging.

[jdormit]

Think I've run into one of those edge cases you were talking about @karthink: we should really be inserting the separator before a chat response, rather than after a tool response - otherwise multiple tool calls in a row leads to a bunch of newlines. Not sure if I'll have time to tackle that today but should be able to soon-ish.

[karthink]

Here's another edge case.

When you test Claude it usually responds with text ("I'll run this tool for you"), then a tool call, then more text ("The results of the tool are X"). This fix works fine for those cases.

Suppose the LLM responds instead with a tool call followed by a response, as Gemini and OpenAI models do. (No preamble before the tool call.)

Then you get a response that looks like this (with response-prefix Response: )

User: What is the time in Greece?

Response:

The time in Greece is ...

instead of

User: What is the time in Greece?

Response: The time in Greece is...

The extra gptel-response-separator between Response: and the result is from your :last-was-tool-call tracking.

[karthink]

Also, a similar fix might be needed for non-streaming responses.

[karthink]

@jdormit I see a couple of new commits from you here. Is the edge case from my comment above handled now?

[karthink]

@jdormit I'd like to merge this PR but I'm not sure all the edge cases are covered. If you're still interested in it, please let me know -- we'll need some tests to ensure that they are, and I can work on merging.

Please let me know if you've moved on as well, and I can close the PR.