pi-semantic-edit

Drop-in replacement for Pi's built-in edit. Same edits[] contract, so the model keeps writing the calls it already writes. What changes is what happens after the call: the matcher absorbs the drift models actually produce, and failures say what to fix instead of bouncing the model into a blind retry.

{ "path": "src/foo.ts", "edits": [{ "oldText": "let x = 1;", "newText": "let x = 2;" }] }

Why you'd install it

The built-in edit is strict. oldText has to match the file byte for byte, and when it doesn't the model retries in the dark. This one runs the text through a 10-pass fuzzy chain (line trim, indentation, escaped sequences, Unicode quotes and dashes) before giving up. Every pass re-verifies its match against the actual file, so tolerance never turns into an edit at the wrong location.

Two behaviors the built-in lacks:

• replaceAll: true on an edit replaces every occurrence. That's what renames actually need; without it a multi-match fails with the line positions and you add context.
• Stale-read rejection. If the file changed since the model last read it, the edit is refused with a re-read message instead of silently landing on stale content.

On no-match the error quotes the closest real text plus a similarity percentage, so the retry targets what's actually in the file, not what the model remembers.

Install

pi install npm:pi-semantic-edit

That's it. edit now resolves to this tool.

Legacy input

Aider-format patch blocks still work for sessions that started on older versions, but they're deprecated. New sessions should use edits[].

Docs

• Reference: input format, matcher passes, safety behaviors, errors
• Explanation: why the matcher is built this way

License

MIT, see LICENSE.