LeetCode Rating Analytics

LeetCode tells you a problem is "Medium". The community rating dataset knows it was Q3 of weekly contest 106 and came out at 1711. This puts that number where you are already looking: on your profile, on every problem page, and on every row of the problem list.

Version 2 is a rewrite. What it does that v1 did not, and why, is below.

What you get

On your profile. Coverage by rating band, which is the chart the whole thing exists for: not how many problems you solved in each band, but what share of the problems that exist in each band you have done. A raw count peaks in the middle for everyone, because LeetCode has far more 1500s than 2500s. Dividing by what is available leaves the actual shape of one person's ability, and the point where that curve crosses 50% is a defensible answer to "what rating am I".

Also: a topic by rating-band grid, the topics you are behind on measured against your own coverage rather than against other people, contest slot coverage (Q1 through Q4, which separates "my topics are weak" from "I never finish a contest"), your recent solves plotted against time, a revision list, your contest rating next to your practice level, and CSV or JSON export.

On a problem page. A chip next to the difficulty: rating, contest, slot, and whether that sits above or below your own level.

In the problem list. The rating inline on every row, colour-banded.

What it costs you

Nothing leaves your browser. There is no server, no analytics and no account.

A default install declares exactly one permission, storage. The rating dataset ships inside the extension, so a fresh install makes zero third-party requests. The only calls it ever makes are to leetcode.com's own GraphQL endpoint, signed in as you, from a leetcode.com page, which is the same call the site makes to draw its own problem list.

There is one optional extra: refreshing the rating dataset from GitHub weekly. It is off, it asks for the host permission from the options page and never at install, and the permission is re-checked on every single call so revoking it at chrome://extensions stops the traffic immediately.

Install

1. chrome://extensions (or brave://extensions)
2. Developer mode on
3. Load unpacked, select this folder
4. Open any LeetCode page

After changing code: reload with the circular arrow on the extension card, then refresh the LeetCode tab.

Develop

node test/run.cjs        all test suites
node tools/check.cjs     the build check, run before loading
node tools/build-data.cjs path/to/ratings.txt   refresh the bundled dataset

tools/check.cjs exists because of v1. Its manifest listed libs/chart.min.js and the file on disk was libs/chart.umd.min.js, so Chrome refused to load the extension and the repo as committed did not run at all. The check verifies every file the manifest names exists, that every module in src/ is actually loaded by the manifest, that the permission list has not grown, that the shipped dataset holds the rows it claims, and that no CSS rule is unnamespaced enough to restyle leetcode.com itself.

What changed from v1, and why

It loads. See above.

The cookie plumbing is gone, and the cookies permission with it. v1 asked the service worker for LEETCODE_SESSION and csrftoken, then set a Cookie header on fetch. Cookie is a forbidden header name: the browser strips it silently. The requests worked anyway, because credentials: "include" on a same-origin request already sends the real cookies. So the entire pipeline, and the scariest permission in the manifest, did nothing.

It no longer fires 74 simultaneous requests. v1 paginated the whole problem set 50 at a time inside one Promise.all. On a 3,700 problem set that is about 74 concurrent POSTs to an endpoint that rate limits, and the first 429 threw a TypeError because the response was read without a check. v2 uses pages of 100, three at a time, with exponential backoff, Retry-After support, and a shape check on every response.

It caches. v1 re-downloaded the dataset and re-walked the API on every single profile view. v2 syncs once, keeps the result for 12 hours, and after that a profile view costs zero requests.

Public profiles are honest. v1 fetched tags one problem at a time, up to 500 round trips, for data identical for every user on earth. It also claimed "public profile mode" based on a condition that was almost never true. LeetCode only returns another account's 20 most recent accepted submissions, so v2 says exactly that and shows only what it can actually compute.

Chart.js is gone. 200K of UMD bundle loaded into every leetcode.com page, including the problem editor, to draw two charts. v2 draws SVG by hand: no dependency, themes for free because inline SVG inherits CSS custom properties, and chart code that is a pure string function and therefore testable.

It follows your theme. v1 hardcoded a dark palette and painted a dark card into the middle of LeetCode's light mode.

Histograms have no holes. v1 built its bands from the brackets that happened to be present, so a band you have never solved vanished and the two either side were drawn adjacent. A cliff looked like a smooth slope.

/problemset/all/ is no longer a user called "all". v1 took the last path segment as a username and excluded six reserved words.

Data

Problem ratings come from zerotrac/leetcode_problem_rating, which fits a rating to each contest problem from contest results. Only problems that appeared in a rated contest have one: about 2,600 of LeetCode's ~3,700. Everything else is counted and reported separately, never folded in silently.

The bundled snapshot date is printed on the panel.