Edit formats

editkit supports the three formats real models actually emit. You can mix all three in one LLM response; editkit detects, parses, and applies them in source order.

SEARCH/REPLACE blocks

PATH/TO/FILE
<<<<<<< SEARCH
...exact lines from the existing file...
=======
...what they should be replaced with...
>>>>>>> REPLACE

• Best default for most coding-agent tasks.
• Compact, focused, easy for small models to emit correctly.
• The SEARCH section must be unique in the file — extend it with surrounding context if the function appears multiple times.
• Empty SEARCH = create a new file. Empty REPLACE = delete code.
• Multiple blocks per response are fine; later edits to the same file see the earlier ones.

Unified diff

--- a/PATH/TO/FILE
+++ b/PATH/TO/FILE
@@ -OLD_START,OLD_LINES +NEW_START,NEW_LINES @@
 unchanged context
-removed line
+added line
 unchanged context

• Best for large refactors with many small changes scattered across a file.
• The diff structure stops the model from emitting // ...rest unchanged placeholders — which is why it wins on framework-migration tasks like Next 13 → 15.
• Three lines of context before and after each change.
• Use /dev/null as the source path for file creation, or destination for deletion.

Whole-file

PATH/TO/FILE
```ts
... full file contents ...
```

• Best for very small files (under 50 lines) or smallest models.
• The fence language (ts, py, etc.) is informational.
• Output one fenced block per file. Don't omit any lines.

Mixing formats in one response

import { applyEdits } from "editkit";

// LLM response can use all three formats at once:
const results = await applyEdits(llmOutput, fileReader);

detectFormats only returns whole-file when no SEARCH/REPLACE or unified-diff markers are present. So a mixed response would silently drop the whole-file block under the default detector. Pass formats explicitly when you want all three:

const results = await applyEdits(llmOutput, fileReader, {
  formats: ["search-replace", "unified-diff", "whole-file"],
});

Restricting to one format

When you've prompted in one format and want hard-fails on the others:

const results = await applyEdits(llmOutput, files, {
  formats: ["search-replace"],
});

Anything that isn't a SEARCH/REPLACE block is ignored. If the model emits a unified diff anyway, you'll get zero results back and can re-prompt.

Picking a format