[robobun]

Problem

• With the default targets (Safari 14 needs :-webkit-full-screen), .f:fullscreen{inset-inline-start:1px} prints .f:full-screen:dir(ltr){left:1px}. No browser has :full-screen, so the declarations are lost.
• Cause 1: downlevel_selectors widens the prefix set of :fullscreen in place to NONE | WEBKIT, and get_prefix (src/css/selectors/selector.rs:227) reads that as an explicit prefix. When a :dir() or a second widened set follows, it returns empty, and the rule loses its per-prefix copies.
• Cause 2: the :fullscreen, ::placeholder and ::file-selector-button printer arms pick the legacy name with contains(). A widened set that prints outside a prefix pass becomes :full-screen. So does a sibling after a nested rule: .p:fullscreen{& .k:fullscreen{…} & .j{…}} ends in .p:full-screen .j.

Fix

• get_prefix counts a set that contains NONE as unprefixed. Only Targets::prefixes builds such a set, and downlevel_selectors computes it again.
• The three arms compare with ==, as lightningcss does. Outside a pass, a widened set prints its unprefixed name. css: print the vendor prefix on downleveled :any-link, :read-only, :read-write, :placeholder-shown and :autofill #42131 has the same three lines.
• Verified: test/js/bun/css/css.test.ts (10 new cases, 9 fail on main). Also test/js/bun/css/, test/bundler/css/, test/bundler/esbuild/css.test.ts.
• Self-reviewed: 7 concerns raised, 7 addressed. One answer was replaced later (see Notes).

Background

• VendorPrefix is a bit set. :fullscreen parses to NONE, :-webkit-full-screen to WEBKIT.
• If get_prefix returns a set with NONE, update_prefix calls downlevel_selectors for the prefixes the targets need. StyleRule::to_css prints the rule once per prefix.
• A logical property such as inset-inline-start becomes two extra rules with left and right. They clone the rule's widened selectors and append :dir(ltr) or :dir(rtl).
• Adjacent @media blocks with one query merge. Their rules are minified again, with widened sets.

Notes

• A fuzzer found this (fuzz ledger rows 58203 and 59607). No user reported it. It is not a regression: 1.3.14, 1.4.2 and 1.4.3 print the same bytes.
• Repro on 1.4.3: printf '.f:fullscreen { inset-inline-start: 1px; color: red }\n' > in.css && bun build ./in.css --minify prints .f:-webkit-full-screen{color:red}.f:fullscreen{color:red}.f:full-screen:dir(ltr){left:1px}.f:full-screen:dir(rtl){right:1px}. After: the :dir() rules print as .f:-webkit-full-screen:not(:-webkit-any(:lang(ae),…)), .f:fullscreen:not(:-moz-any(…)) and .f:fullscreen:not(:is(…)), plus the three rtl forms. A hand-written .f:fullscreen:dir(ltr) already printed that way.
• Second repro, no logical property: @media (min-width:1px){.a:fullscreen .b:fullscreen{color:red}} @media (min-width:1px){.c{color:blue}} prints .a:full-screen .b:full-screen{color:red}. The merged @media rule list is minified again, and both components are already widened. The same path skipped the :is() to :-webkit-any() lowering for .a:autofill:is(.p,.q).
• Why only some selectors: Selector::append puts :dir() at the end of the last compound. .f:fullscreen gives the order widened, then NONE, which get_prefix rejects. .f:fullscreen .x and .f::file-selector-button give NONE first, which it accepts. Those already printed correctly.
• Nested siblings: a nested rule with its own passes prints only in the last pass of its parent (css: bound vendor-prefix fan-out when serializing nested rules #31642), and StyleRule::to_css then resets the printer's prefix to empty. Every later sibling prints outside any pass. With only the get_prefix change, the generated :dir() rules gain their own passes and reach that reset, so .p:fullscreen{& .k:fullscreen{inset-inline-start:1px} & .j{…}} would turn from .p:fullscreen .j into .p:full-screen .j. The self-review found that.
• History of the second change: the first version of this PR restored the printer's prefix after a nested rule's passes. A 20,000-stylesheet fuzz comparison then showed 2 stylesheets where a later sibling with a hand-written prefix and no pass of its own (.j:is(.x,.y)::-moz-selection, .j:not(.x):-webkit-any(a,b)) printed ::selection and :is(a,b) in the last pass. Main prints those as written in that position. The == comparison fixes the same invalid names and leaves such a sibling as written, so it replaced the restore. A new test pins it, and each of the three arms has a case that fails with contains().
• Differential runs against 1.4.3 with cssInternals.prefixTest. Flat corpus, 133,950 combinations of selector, declaration, wrapper and targets: 12,618 outputs change. Outputs with an invalid pseudo name (:full-screen, ::input-placeholder, ::file-upload-button, ::browse) go from 7,038 to 0, and none is new. A rule inside merged @media blocks prints the same as outside them in 26,790 of 26,790 cases (2,408 differ on 1.4.3). Nested corpus, 30,240 outputs with plain and hand-prefixed later siblings: invalid names go from 744 to 0, and no output loses a hand-written prefix (the restore version lost one in 2,701).
• css: print the vendor prefix on downleveled :any-link, :read-only, :read-write, :placeholder-shown and :autofill #42131 changes | to & in write_prefixed and has the same three == lines. The two branches merge without a conflict, and both test sets pass on the merge. With only css: print the vendor prefix on downleveled :any-link, :read-only, :read-write, :placeholder-shown and :autofill #42131 the reported input prints .f:fullscreen:dir(ltr): valid in new browsers, but Safari 14 and Chrome 80 to 119 get no -webkit- copy and no lowered :dir().
• Until css: print the vendor prefix on downleveled :any-link, :read-only, :read-write, :placeholder-shown and :autofill #42131 lands, .a:autofill .b:autofill inside merged @media blocks prints a duplicate rule, because the -webkit- pass prints :autofill unprefixed. That is harmless. Outside merged @media blocks main already prints that duplicate.
• css: print explicitly prefixed selectors as written in every vendor prefix pass #42106 reworks explicit prefix printing. Its new get_prefix has the same rule for sets that contain NONE. css: keep a selector list whole when a member is vendor prefixed #42100 changes update_prefix and to_css_base. Both merge with this branch without a conflict, and both test sets pass on each merge. So do css: keep the selectors on rules added for color fallbacks and logical properties #43240 and css: bump the default browser targets to Vite's baseline widely available set #40369 (merge check only). Hand-written prefix ladders (.a:-webkit-full-screen{…}.a:fullscreen{…}), which lose a rule when they are minified again, are the subject of css: print explicitly prefixed selectors as written in every vendor prefix pass #42106, not of this PR.
• Main with only css: keep a selector list whole when a member is vendor prefixed #42100 prints .a:full-screen{color:red} for a single .a:fullscreen rule inside merged @media blocks, which main prints correctly today. Its new guard in update_prefix reads the widened set that get_prefix returns on the second run as a hand-written mix. With this branch merged in, the rule prints correctly. So this PR should land first.
• Behaviour change: a rule that mixes a widened pseudo-class with an explicit prefix of another vendor (.a:autofill:-moz-full-screen) inside merged @media blocks now prints the same three passes as outside them. Before, it printed once as written. The first-run behaviour for explicit prefixes is the subject of css: print explicitly prefixed selectors as written in every vendor prefix pass #42106 and css: keep a selector list whole when a member is vendor prefixed #42100.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository: oven-sh/bun/.coderabbit.yaml

Review profile: ASSERTIVE

Plan: Essentials

Run ID: 96b4bad2-b22a-4bb3-8afe-781d92602035

📥 Commits

Reviewing files that changed from the base of the PR and between 367d939 and bcd307c.

📒 Files selected for processing (2)

• src/css/selectors/selector.rs
• test/js/bun/css/css.test.ts

Included review availability: Your plan provides up to 10 included reviews per hour; 0 remain after this review.

---

Walkthrough

The CSS selector serializer now normalizes mixed prefix sets and applies legacy names only for exact vendor prefixes. Tests cover fullscreen, logical-direction, nested-rule, media-block, pseudo-element, and sibling-prefix output.

Changes

CSS prefix preservation

Layer / File(s)	Summary
Prefix state handling src/css/selectors/selector.rs	VendorPrefix::NONE is recognized in combined prefix sets. Legacy fullscreen, placeholder, and file-selector-button names now require exact supported vendor prefixes.
Prefix regression coverage test/js/bun/css/css.test.ts	Added regression tests for fullscreen expansion, logical-direction selectors, merged media blocks, nested rules, pseudo-elements, and preservation of hand-written sibling prefixes.

Suggested reviewers: jarred-sumner

Priority: ➖ Normal

Merge Risk: ⚪ Minimal · up to bcd30

The prefix-state and selector-serialization changes have no identified merge-blocking risk.

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly describes the main fix: preserving vendor-prefix passes for cloned and re-minified rules while preventing invalid bare :full-screen output. It is specific and related to the change…
Description check	✅ Passed	The description explains the problem, root causes, fix, affected scenarios, verification steps, test results, and related issues. It does not use the exact template headings, but it provides the requi…

---

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

[robobun]

Status

Reproduced on 1.4.3 and on a debug build of main (367d939):

printf '.f:fullscreen { inset-inline-start: 1px; color: red }\n' > in.css && bun build ./in.css --minify

It prints .f:full-screen:dir(ltr){left:1px}.f:full-screen:dir(rtl){right:1px} after the two color rules. With this branch it prints the six :-webkit-full-screen / :fullscreen rules with :dir() lowered.

bun bd test test/js/bun/css/css.test.ts: 9 of the 10 new cases fail with the src/ of main, and all pass with this branch (1302 pass, 0 fail). The tenth case is the hand-written .f:fullscreen:dir(ltr) reference, which main already prints correctly.

[claude]

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

Additional findings (outside the current diff — GitHub can't attach inline comments there):

• 🟣 src/css/rules/style.rs — Pre-existing, in the function this PR reworks: a nested rule whose own prefix set is only NONE still loses the parent's -webkit- pass, and prefixed nested rules still print out of source order. style.rs:290 defers every nested rule with a non-empty set to the parent's final pass, and that rule then prints only its own passes. Fix: defer a nested rule only when its set covers every prefix of the parent's set, otherwise print it in each parent pass with dest.vendor_prefix from that pass; with that in place the order of siblings also stays as written. The new test at css.test.ts:5852 certifies the reordered output. [also at: src/css/rules/style.rs:303 - pre-existing: Safari 14 users get a reversed cascade between a prefixed nested rule and a later unprefixed sibling when nesting is compiled away.]

  Extended reasoning...

  Trigger: .p:fullscreen { & input:placeholder-shown { color: red } } with targets safari 14. :fullscreen is widened to NONE|WEBKIT, so .p prints two passes. input:placeholder-shown gets targets.prefixes(NONE, PseudoClassPlaceholderShown) which is NONE for Safari (prefixes.rs:947 has no safari arm), so the nested rule's vendor_prefix is NONE, non-empty. In the parent's WEBKIT pass style.rs:290 sets skip_prefixed_nested_rules, and mod.rs:436 skips the nested rule because mod.rs:412 treats any non-empty set as deferred. In the parent's NONE pass the nested rule runs its own loop at style.rs:108 with only NONE, so & expands with dest.vendor_prefix = NONE and prints .p:fullscreen input:placeholder-shown{color:red} once. No .p:-webkit-full-screen input:placeholder-shown is ever emitted, so Safari 14 through 16.0 (fullscreen WEBKIT range at prefixes.rs:1008) never matches the rule. The base branch prints the same. Second consequence, same mechanism: .p:fullscreen { & .k:fullscreen { color: red } & .j { color: blue } } prints .p:-webkit-full-screen .j before…

  Verification: pre-existing (base fails by the same route; the PR touches the pass machinery in StyleRule::to_css but not the deferral predicate). Triggering condition: nesting compiled away for targets that widen a parent pseudo (default targets include safari14, which needs :-webkit-full-screen), with a nested rule whose own prefix set is exactly NONE. Mechanism verified by tracing the code (no debug…

• 🟣 src/css/selectors/selector.rs — Safari 14 users now lose rules such as .f:fullscreen:is(.a .b){inset-inline-start:1px} in the new -webkit- pass, because the :is() arm at selector.rs:879-886 prints :-webkit-any(.a .b) whenever the pass is WEBKIT, and :-webkit-any() rejects complex selectors. The get_prefix change at selector.rs:252 makes :dir() clones and re-minified merged @ media rules reach a WEBKIT pass that the base never printed for them. Fix: print :-webkit-any/:-moz-any only for an :is() that downlevel_component actually converted (no combinators, selector.rs:132-141), and keep :is( otherwise in every pass.

  Extended reasoning...

  Trigger: .f:fullscreen:is(.a .b){inset-inline-start:1px} with targets that need :-webkit-full-screen (defaults), minify.
  The logical property creates a :dir(ltr) clone (context.rs:236-254); it is minified at mod.rs:877.
  Base: get_prefix returns empty for the clone (widened set then NONE), so it prints once as .f:full-screen:is(.a .b):dir(ltr).
  After the diff: get_prefix returns NONE, downlevel_selectors gives NONE|WEBKIT, and the clone prints a WEBKIT pass.
  In that pass selector.rs:127-148 did not convert the :is() because .a .b has a combinator, but the printer at selector.rs:879-886 still checks dest.vendor_prefix and writes :-webkit-any(.a .b).
  :-webkit-any() only accepts compound selectors, so the WEBKIT copy is invalid and dropped by the browser; the NONE copy remains for modern browsers.
  The same applies to the :not(:-webkit-any(:lang(ae),...)) produced from :dir() inside any WEBKIT pass. The finder marked it pre-existing because the printer arm is untouched, but the population that reaches the WEBKIT pass with an unconverted :is() grows with this…

  Verification: pre-existing. Mechanism verified, but the base branch already drops the same rule for Safari 14 and the defective code is untouched by this diff. Triggering condition: a rule whose selector contains :fullscreen (or any pseudo the targets widen to NONE|WEBKIT) together with an :is() holding a complex selector, printed in a WEBKIT pass. The printer arm at… | pre-existing. Mechanism verified,…

[robobun]

Reply to the two findings outside the diff. Both are on main, and this PR does not change them.

• A nested rule whose own prefix set is only NONE: confirmed. .p:fullscreen { & input:placeholder-shown { color: red } } at safari 14 prints only .p:fullscreen input:placeholder-shown on main and on this branch. css: print explicitly prefixed selectors as written in every vendor prefix pass #42106 covers it. With this branch merged into css: print explicitly prefixed selectors as written in every vendor prefix pass #42106, the output also has .p:-webkit-full-screen input:placeholder-shown.
• The order of a deferred nested rule and a later sibling is a separate effect of the deferral from css: bound vendor-prefix fan-out when serializing nested rules #31642, and css: print explicitly prefixed selectors as written in every vendor prefix pass #42106 keeps it. .p:fullscreen { & .k:is(.a, .b) { color: red } & .k.c { color: blue } } at chrome 80 and safari 14 prints .p:-webkit-full-screen .k.c before .p:-webkit-full-screen .k:-webkit-any(.a, .b), so the cascade is reversed for browsers that match only the -webkit- forms. I passed that on as a separate fix. The two new nested cases here pin the last rule (.p:fullscreen .j). The order in them is the order that main prints today.
• :-webkit-any() around a complex selector in a -webkit- pass: confirmed, and the plain rule has it on main too. .f:fullscreen:is(.a .b){color:red} prints .f:-webkit-full-screen:-webkit-any(.a .b). With css: print explicitly prefixed selectors as written in every vendor prefix pass #42106 it prints .f:-webkit-full-screen:is(.a .b).

[robobun]

^{Updated 4:15 PM PT - Sep 18th, 2026}

✅ @robobun, your commit bcd307cc10c9b97a4954fa31f437f98cd6209a47 passed in Build #118039! 🎉

---

🧪   To try this PR locally:

bunx bun-pr 43339

That installs a local version of the PR into your bun-43339 executable, so you can run:

bun-43339 --bun

[claude]

Code review completed

Nothing new to post: everything this review found is already covered by existing comments on this pull request or didn't merit a separate one.

[claude]

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

[claude]

Code review completed

Nothing new to post: everything this review found is already covered by existing comments on this pull request or didn't merit a separate one.

[Jarred-Sumner]

Landed in #47025 (7942008).

Keep target-added vendor prefix passes on cloned and re-minified rules, and never print a bare :full-screen. 12 real stylesheets

[robobun]

Confirmed on main at 7942008. Both source changes of this PR are there: the NONE rule in get_prefix and == in the three printer arms. On a debug build of main, the ten cases from the last head of this PR pass, and bun build --minify on the reported input prints .f:-webkit-full-screen:not(:-webkit-any(…)) and .f:fullscreen:not(:is(…)), with no :full-screen.

One gap is left: main has no test for the :fullscreen arm. With that arm back on contains(), the CSS suites pass (3,667 tests) and .p:fullscreen{& .k:fullscreen{…}& .j{…}} ends in .p:full-screen .j again. #47115 adds one row for it.