Description
Currently, git diff captures pure text changes, which may include mostly irrelevant changes.
This consumes LLM tokens and creates commit message comments that bloat the message, potentially hiding important changes in the toll pile of garage messages.
Suggested Solution
I’d like to have a --semantic-diff parameter when enabled with oco, it would use diffsitter that is based on AST code changes.
https://github.com/afnanenayet/diffsitter
Alternatives
No response
Additional Context
This idea could be extended:
By still leveraging the remaining less important changes that are captured as usual git diff, but then clearly identifying and separating them as less important, that we could ask LLM to pay less attention to and produce less detailed summary.
So I could get full picture of the changes, but have LLm focus with on summarizing with more details important chances. While keeping semantically not important formatting and spend less output LLM tokens on those (potentially even using another cheaper model).
Description
Currently, git diff captures pure text changes, which may include mostly irrelevant changes.
This consumes LLM tokens and creates commit message comments that bloat the message, potentially hiding important changes in the toll pile of garage messages.
Suggested Solution
I’d like to have a
--semantic-diffparameter when enabled with oco, it would use diffsitter that is based on AST code changes.https://github.com/afnanenayet/diffsitter
Alternatives
No response
Additional Context
This idea could be extended:
By still leveraging the remaining less important changes that are captured as usual git diff, but then clearly identifying and separating them as less important, that we could ask LLM to pay less attention to and produce less detailed summary.
So I could get full picture of the changes, but have LLm focus with on summarizing with more details important chances. While keeping semantically not important formatting and spend less output LLM tokens on those (potentially even using another cheaper model).