Consider using unified diff without line numbers as output format

[rubengmartins]

Some people point out that asking for code in JSON format may lead to worse results.

Unified diff is an alternative but the LLM always messes up the line numbers.

Another alternative would be something like: https://aider.chat/docs/unified-diffs.html#unified-diff-editing-format

We could encourage a high-level diff where we ask the LLM the reduce the number of blocks changed and then do a string replacement in the code.

Pros:

• Potential better code generation
• We will be able to tackle very large functions since the patch is usually small

Cons:

• The format may fail (sometimes the LLM does not generate - or + when removing or adding a line; sometimes comments are omitted; etc)
• If the format fails we cannot apply the patch