feat: improve badge width estimation

[t128n]

🔗 Linked issue

Resolves#1601

🧭 Context

In production, measuring text with canvas is crashing, resulting in inaccurate fallback measurements:

Introduced a character lookup table of "manually" measured character widths for more
accurate lookups.

📚 Description

Tweaked the estimateTextWidth function to use a on-character level lookup table for more accurate results when the Canvas API from @napi-rs/canvas is unavailable.
Lookup table derived from running

// ...constctx=createCanvas(1,1).getContext('2d')constchars=' !"#$%&\'()*+,-./:;<=>?@[\\]^_`{|}~0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'ctx.font=BADGE_FONT_SHORTHANDconstentries=[...chars].map(ch=>`'${ch==="'" ? "\\'" : ch==='\\' ? '\\\\' : ch}': ${Math.ceil(ctx.measureText(ch).width)}`)console.log('default: {\n '+entries.join(', ')+'\n}')

locally, where fonts and the API are available.

Using a generous fallback for characters that aren't in the map, such as emojis.

📷 Comparison

Overflow / Long Labels

Case	Prod	PR
Long label
Long value
Long package name

Shields.io Style (?style=shieldsio)

Case	Prod	PR
Normal
Long label
Emoji label

Emojis

Case	Prod	PR
Emoji in label
Emoji in value
Multiple emojis
Emoji-only label

CJK Characters

Case	Prod	PR
Chinese label
Japanese label
Korean label
Mixed CJK + ASCII

Special & Punctuation Characters

Case	Prod	PR
Symbols
Narrow chars (iiiii)
Wide chars (WWWWW)
Mixed widths

Badge Types Smoke Test

Type	Prod	PR
version
license
downloads
size
types
deprecated

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
docs.npmx.dev	Ready	Preview, Comment	Apr 19, 2026 10:30pm
npmx.dev	Ready	Preview, Comment	Apr 19, 2026 10:30pm

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
npmx-lunaria	Ignored		Apr 19, 2026 10:30pm

[coderabbitai]

📝 Walkthrough

Walkthrough

Replaced heuristic, bucket-based text width estimation with explicit per-character width lookup tables for supported badge fonts and a fixed fallback width for unmapped characters; updated estimateTextWidth to use font and sum per-character widths.

Changes

Cohort / File(s)	Summary
Badge Text-Width Measurement server/api/registry/badge/[type]/[...pkg].get.ts	Removed regex/set-based categorisation and FALLBACK_WIDTHS/NARROW_CHARS/MEDIUM_CHARS buckets. Added CHAR_WIDTHS for default and shieldsio fonts, CHAR_WIDTH_FALLBACK for unknown characters, and refactored estimateTextWidth(font, text) to sum per-character widths with fallback.

Possibly related PRs

• style: revise the badge char width #908: Also replaces badge text-width measurement with per-character lookup approach using scalar widths.
• fix: improve badge text measurement accuracy #1486: Alternatively modifies badge text-width measurement in the same route using Canvas-based measurement instead of lookup tables.
• fix: fix fallback not triggered for 0 measuredWidth #1584: Tightens measurement validity check for the measured width in the same badge measurement logic.

Suggested reviewers

• danielroe

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description check	✅ Passed	The description is comprehensive and clearly related to the changeset, detailing the problem (Canvas measurement failures on Vercel), the solution (per-character lookup table), and providing extensive visual comparisons demonstrating the fix.
Linked Issues check	✅ Passed	The PR successfully addresses issue #1601 by implementing a per-character lookup table for accurate text-width measurement when Canvas-based measurement fails on Vercel, preventing badge overflow and ensuring consistent rendering across environments.
Out of Scope Changes check	✅ Passed	All changes are directly scoped to fixing text-width estimation for badges by replacing the heuristic approach with a lookup table; no unrelated modifications or refactoring are present.
Title check	✅ Passed	The title 'feat: improve badge width estimation' is directly related to the main change: replacing a heuristic text width estimator with a deterministic per-character lookup table to improve badge width estimation accuracy.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ All tests successful. No failed tests found.

📢 Thoughts on this report? Let us know!

[ghostdevv]

this doesn't break on the last test case from #2199 so we can yolo this :p

[ghostdevv]

actually, rq why is it "when canvas not available"? @t128n

[t128n]

actually, rq why is it "when canvas not available"? @t128n

Text is usually measured using createCanvasContext. On Vercel, createCanvasContext can’t be created (returns null, likely due to missing native node.js addons in the serveless runtime), so it falls back to estimateTextWidth. estimateTextWidth makes a best guess for characters’ widths and has edge cases with wider characters like w.
That's the whole context.

So Canvas not available was referring to the createCanvasContext (or what its name is - currently on mobile and it's hard to look outside the PRs boundary to find the imports).

[ghostdevv]

actually, rq why is it "when canvas not available"? @t128n

Text is usually measured using createCanvasContext. On Vercel, createCanvasContext can’t be created (returns null, likely due to missing native node.js addons in the serveless runtime), so it falls back to estimateTextWidth. estimateTextWidth makes a best guess for characters’ widths and has edge cases with wider characters like w. That's the whole context.

So Canvas not available was referring to the createCanvasContext (or what its name is - currently on mobile and it's hard to look outside the PRs boundary to find the imports).

ah but that's always been true right? like that fact hasn't changed in this PR?

also that's why we have the napi-rs canvas? or no?

[t128n]

Yeah, my assumptions were flawed - I thought it would be about the whole canvas not being available, but I made some further tests real quick.

So, @napi-rs/canvas & canvasContext generally are available, but if I'm not missing anything major, it's producing 0 widths because Geist as a font is not available on the Vercel serverless function (my assumption based on the testing).

functionmeasureTextWidth(text: string,font: string): number|null{constcontext=getCanvasContext()if(context){context.font=fontconstmeasuredWidth=context.measureText(text).widthif(Number.isFinite(measuredWidth)&&measuredWidth>0){// Returns false as font not available and nothing can be measuredreturnMath.ceil(measuredWidth)}}returnnull// yields null for canvas context and then falls back to estimateTextWidth }

So it's rather: use pre-measured widths for a lookup table rather than use table for badges when canvas not available

[wojtekmaj]

ah but that's always been true right? like that fact hasn't changed in this PR?

Yea, it is my flawed implementation that I wasn't able to debug because I didn't know what errors were happening on Vercel, and it is not reproducible locally at all.

[ghostdevv]

I think we can create an issue to try and load the font and see what the perf is like on that, if we can get it working that is. If there are Vercel issues then we can help with debugging that this time around as we should now have access to logs

[serhalp]

@ghostdevv what's the status of this PR? I see you enabled auto-merge but you still have changes requested 🤔🤔🤔

[t128n]

@ghostdevv what's the status of this PR? I see you enabled auto-merge but you still have changes requested 🤔🤔🤔

There was #2587 created to fix the actual bug (Geist font not available, and therefore the endpoint falling back to the measuring algorithm) instead of working around it by improving the text-measurement algorithm

[ghostdevv]

actually, rq why is it "when canvas not available"? @t128n sorry this is what was in the box when I went to approve - GitHub fully functioning software! xD I think this should be good and we can do the follow up later...