feature: minimal diff

[will-lynas]

Feature request

The LLMs tend to suggest replacement of large blocks of code, even when most of the block is unchanged. This makes it harder to review the changes it wants to make. Would it be possible to generate a minimal diff and apply that inside the merge conflict markers instead?

Example: This is generated

<<<<<<< HEAD
CREATE TABLE serbian_scores (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    user_id INTEGER,
    test_id INTEGER,
    FOREIGN KEY (user_id) REFERENCES users(id)
    ON DELETE CASCADE
);
=======
CREATE TABLE serbian_scores (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    user_id INTEGER,
    test_id INTEGER,
    FOREIGN KEY (user_id) REFERENCES users(id)
        ON DELETE CASCADE,
    FOREIGN KEY (test_id) REFERENCES serbian_tests(id)
        ON DELETE CASCADE
);
>>>>>>> Snippet

But it could just be

<<<<<<< HEAD
=======
    FOREIGN KEY (test_id) REFERENCES serbian_tests(id)
        ON DELETE CASCADE
>>>>>>> Snippet

Motivation

No response

Other

No response

[aarnphm]

This is syntatic diff you are suggesting. I think that in the context of code it is better to dump the whole thing.

What we can do is to have a semantic diff support, but seems like too much work atm :))

[will-lynas]

Can I make a PR to add it as an option? Just find lines that are the same in both sections and delete them?

[will-lynas]

Maybe just stripping duplicates off the ends of the block to avoid turning one block into multiple

[will-lynas]

I've had a go at making it. Seems to work from my testing

CREATE TABLE serbian_scores (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    user_id INTEGER,
    test_id INTEGER,
    FOREIGN KEY (user_id) REFERENCES users(id)
<<<<<<< HEAD
    ON DELETE CASCADE
=======
            ON DELETE CASCADE,
        FOREIGN KEY (test_id) REFERENCES serbian_tests(id)
            ON DELETE CASCADE
>>>>>>> Snippet
);

[aarnphm]

The indentation doesn't seem correct?

[will-lynas]

Yeah. Something is defo wrong. I'll investigate and update in the PR

[reaink]

‌‌Perhaps it's better to have the smallest replacement when returning in the llm uplift prompt?