- New
reviewExamples: the paraphrase case the deterministic detector refuses to guess at. Two examples teaching the same lesson in different words score around 0.54 on word overlap — close enough to two genuinely distinct examples (~0.20) that catching them by similarity would mean flagging examples that teach different things. Deciding that "arrived quickly" and "arrived fast" demonstrate the same pattern needs a model, so this sits behind the optional LLM layer, costs a call, and never runs during an ordinaryoptimize(). - Returns
nullbelow two examples, so the caller does not pay for a foregone answer. - The response is treated as untrusted input, because it is. Indices are range-checked against the examples that exist, self-references dropped, overlapping groups collapsed so the same tokens are never counted twice, and the model's stated reason truncated. A model answering with prose produces an empty review — not a crash, and not a saving the prompt could not deliver.
- A provider error still throws. A bad answer is the model's problem and gets absorbed; a broken endpoint is the caller's configuration and hiding it would waste their afternoon.
- The CLI runs it under
--llm, reports it as a suggestion to read rather than a change made, and includes it in--json. - Shortens the GitHub Action's description to 113 characters: the Marketplace rejects anything over 125, which blocked publishing.
- Tests grow from 164 to 179.
0.7.0
No date in the file