Diff Edits vs. Full Rewrites in Cursor: Why Passing Tests Still Doesn't Prove Behavior Held

Every AI edit in the Cursor code editor ends the same way: text lands on disk. Whether the model rewrote the whole file or emitted a few targeted hunks, the result looks plausible in the diff viewer. A tutorial published November 3, 2026 by Tech Insider starts from that point and asks what a diff viewer cannot answer: did behavior change anywhere you did not ask for?

The two edit shapes fail differently. In a rewrite, the model emits the complete new file and the editor swaps it in. Nothing has to match, so it applies easily, but the model regenerates every line, including ones you never mentioned. Models tidy. They swap a comparison operator, change a rounding mode, drop a constant that looks unused, or reflow a block. Each is a silent behavior change buried in a large diff. In the diff shape, the model emits only the changed regions, usually as search/replace blocks in the format the Aider project popularized. Fewer lines are touched and review is easier, but the edit fails loudly when the search text matches zero or several places, and can fail quietly when the replacement lands in the wrong spot. Line-numbered unified diffs add a third problem: once an earlier hunk shifts the file, later line numbers go stale.

What is actually known about Cursor's internals is narrow. The article says Cursor publishes no current specification of its apply step, and anything claimed about how it chooses between rewrite and diff should be treated as unconfirmed. On record: a Cursor team reply on the company forum dated August 11, 2026 said the team is working to train models to apply targeted diffs more accurately rather than rewrite whole files. A forum report from October 28 described an agent that showed a diff for only the first edit in a file while later edits applied without highlighting; the team said it knew of the issue but gave no fix date. Release notes through late October reportedly added a Last Agent Turn scope, better handling of large and renamed-file diffs, and a safeguard against Apply writing empty files when a large cloud-agent diff arrives incomplete. On the business side, SpaceX bought Cursor in a deal AI Magazine put at US$60 billion, and plans were rebuilt around the X Subscription in October.

The tutorial's answer is differential testing. You freeze the old file, load old and new side by side, and let Hypothesis feed both hundreds of generated inputs. The author argues this beats reading diffs: a better diff algorithm still asks a human to spot a changed rounding mode in 70 lines, while a test asks a machine to spot it in 500 generated orders. The two stack, with the diff judging intent and the tests judging consequences.

The lab guards one realistic change in a checkout module, pricing.py, which handles bulk discounts, coupons, free shipping over $100, and regional tax. The feature request is a loyalty discount that grows with each full year of membership up to a cap, applied after coupons and before shipping and tax. The baseline uses Decimal throughout, banker's rounding for the final amount, a $50 coupon cap, and a strict greater-than comparison for the free-shipping line. Each is a place where a casual rewrite can drift.

Three checks run cheapest first. A scope checker uses Python's ast module to list every top-level name an edit changed, failing before any test if an unauthorized function was touched. The differential test loads the frozen baseline and the candidate and asserts identical results, with the new feature off, for generated orders. Spec tests then check the cap, the ordering, and error handling. The baseline is extracted with git show from the last commit, so it is the exact committed bytes; the author advises refreshing it before every new AI task, since a stale baseline makes the test pass for the wrong reasons. The strict applier requires each SEARCH text to match exactly once.

Three candidates are run: a correct diff edit, a tidy-looking whole-file rewrite that drifts in four places, and a diff edit with a logic-order bug. Output comes from real runs on Python 3.12.3 with Hypothesis 6.170.14 and pytest 9.2.0, and one full verification took 4.5 seconds. The candidates are hand-written fixtures modeled on common drift patterns, not captured Cursor output, and the tutorial says so. The lab runs offline; Cursor enters only in Part 5, where the harness is pointed at real edits. The open question is whether behavior-level checks become routine beside diff review.