AI tools

AI Diff Reviewer

Paste code before and after a change, or a unified diff, to see which changes deserve a second look. Each flag comes with a question to ask in review. A heuristic aid: it cannot judge correctness. Private: nothing is uploaded.

A heuristic aid, not a code review. The tool points at changes that usually deserve a second look, such as a removed permission check or a new eval, and gives a question to ask for each. It reads text only and cannot judge whether the change is correct. No flags does not mean safe. Everything runs on your device. The code is never run, shown as HTML, sent, stored or put in the URL.
Input
Ready. Paste the code before and after, or a unified diff. Nothing is sent or stored.

Result

What it does

Paste the code before and after a change, or one unified diff (for example from git diff or a pull request). The tool shows the changes and flags the ones that usually deserve a human look, with a short question to ask in review for each flag. It is meant for reviewing code you did not write yourself, such as the output of an AI coding assistant, where a quiet change can slip through a long diff.

What it checks

Removed lines:

  • authentication and authorization checks (auth, isAdmin, permission, requireAuth, verify, csrf)
  • input validation and sanitizing (validate, sanitize, escape, schema)
  • error handling (try, catch, except, finally, rollback)
  • tests and assertions (assert, expect(, it(, test(, def test_)
  • logging of security events

Added lines:

  • skipped or disabled tests (.skip, xit, xdescribe, @pytest.mark.skip)
  • lint and type suppression (eslint-disable, @ts-ignore, # type: ignore, # noqa, nosec)
  • empty catch blocks and except: pass
  • TODO and FIXME
  • new outbound calls (fetch, axios, XMLHttpRequest, requests., http.get, urllib)
  • new eval and exec
  • new dependency imports that are not in the code before
  • disabled TLS verification
  • wildcard permissions and CORS
  • changed numbers for timeouts, limits, ports and retries, and changed URLs and hostnames

Moved code

When a check is removed in one place and the same line is added somewhere else, the tool flags both: the removal, with a note that the line was added again, and the addition, marked as moved. That way you can see that the check may have moved, and ask whether it still runs before the code it protects.

What it cannot catch

  • Whether the change is correct. The tool matches patterns in text. It does not understand what the code does, so a flag can be harmless and a dangerous change can have no flag.
  • Renamed or indirect checks. A check with an unusual name, or one that moved into a helper function, is not recognised.
  • Behaviour. Changed conditions, removed branches and logic mistakes are not flagged unless they match one of the patterns above.
  • Other languages and files. The patterns are tuned for JavaScript, TypeScript, Python and shell. Other languages and configuration files are matched only where the same words appear.
  • Dependencies that exist or not. A new import is compared only with the code you pasted. The tool does not look it up anywhere.

Limits

Each side can be up to 4,000,000 characters and 200,000 lines. The work runs in a background Web Worker so the page stays responsive, and a 20,000-line diff is reviewed in well under 5 seconds. When two texts are very different the line diff is approximate, and the tool says so. At most 5,000 flags are kept, and the diff view shows 1,500 lines at a time. A review that takes longer than 20 seconds is stopped. A combined (merge) diff is not supported.

Example: a change that looks small

The built-in example shows an admin endpoint before and after an edit. The after version looks shorter and cleaner. The tool reports 14 flags in 1 file, with 11 lines removed and 9 added:

  • Authentication and authorization removed (3): the requireAuth middleware and the isAdmin check are gone.
  • Validation removed (1), error handling removed (2) and security logging removed (1).
  • Errors swallowed (1): catch (err) {} ignores every failure.
  • Tests skipped or disabled (1) and TODO added (1): it.skip and a TODO to add the tests back.
  • New outbound calls (2) and new dependency (1): data is now posted to an external address through a newly imported library.
  • Changed limit (1): a timeout constant went from 5000 to 60000.

Every flag comes with a question to ask in review, for example “Why was this authentication or authorization check removed? Is it enforced somewhere else?” You can copy the whole list as a Markdown checklist for a pull request.

Frequently asked questions

Is this a replacement for code review?

No. It is a heuristic pass that points a reviewer to the lines that deserve a second look. It does not run the code or understand intent.

Does it work only for AI-written changes?

No. It looks at what changed, not at who wrote it, so it is just as useful for a human pull request.

Is my code uploaded?

No. Both versions are read on your device.

Privacy

This tool runs entirely in your browser. What you paste is processed on your device and is not sent to a server. The code is read as text only: it is never run and never displayed as HTML. The page does not write your input to local storage, cookies or the URL, and the tool does not send your input anywhere.

Share this tool

Your data is not included in this link.