add full file_edits_to_llm_message + tests #481
Conversation
|
@PCSwingle can you take a look at this |
| @@ -62,6 +100,12 @@ def parse_string(self, git_diff: str) -> ParsedLLMResponse: | |||
| is_deletion=is_deletion, | |||
| rename_file_path=new_name, | |||
| ) | |||
| if not is_creation: | |||
| file_edit.previous_file_lines = ( | |||
There was a problem hiding this comment.
This is used specifically for undo's (when the fileedit is applied, it sets this so that the fileedit can be undone). It feels a bit weird using it here, but I guess it works.
There was a problem hiding this comment.
In the next PR, the parsedLLMResponse will be saved to conversation history, and the sampler will use these lines to generate a git diff for the conversation history.
I made a class MentatAssistantMessageParam that lets us attach the parsedLLMResponse, like I referenced in an old PR.
mentat/parsers/git_parser.py
Outdated
| async for chunk in response: | ||
| for content in chunk_to_lines(chunk): | ||
| string += content | ||
| return self.parse_string(string) | ||
| # Wait for a full line |
There was a problem hiding this comment.
Why are we parsing it this way? If we aren't streaming it, this just adds a bunch of complexity for no reason.
mentat/parsers/git_parser.py
Outdated
| if conversation: | ||
| while not conversation.endswith("\n\n"): | ||
| conversation += "\n" | ||
| parsedLLMResponse.full_response = ( |
There was a problem hiding this comment.
Why don't we just have parse_string return a complete parsedLLMResponse (complete with conversation and full_response) instead of tacking it on up here? I think that would make more sense
This adds the missing
GitParser.file_edits_to_llm_messagemethod, which we'll use to convert file edits from your conversation history into git diffs for Sampler v.0.2.0.Pull Request Checklist