Different patch formats for increased performance

[jvm123]

Consider supporting different patch formats

The AlphaEvolve paper suggests to use diff patches that search & replace segments of code, as implemented currently in OpenEvolve.
This format may not be the most ideal for all models (?).

OpenAI trains its models on a line-based diff format: https://fabianhertwig.com/blog/coding-assistants-file-edits/
In the GPT 4.1 prompting cookbook they provide prompts and Python code for applying the patches.

Anthropic uses tools for string replacements, cmp. https://docs.anthropic.com/en/docs/agents-and-tools/tool-use/text-editor-tool#multiple-matches-for-replacement

Visual Studio Code now implements both the OpenAI and the Anthropic formats.

[BradKML]

What is the standard for OpenHands and Cline/RooCode for patch edits? normal LLMs seemed to have odd notation on what to modify