[quantizor]

Fixes

Follow-up to #427.

Minification dropping rules. A line comment after a string containing an apostrophe, escaped quote, or parenthesis survived minification. styled-components then treated the rest of the joined template as a comment and dropped it:

const Quote = styled.p`
  content: "it's"; // apostrophe
  color: red;      // ← this and everything below was dropped
`

The same miscount could strip spaces inside a later string (' a ; b ' became ' a;b '). The minifier now follows strings and parentheses the way the styled-components runtime does.

Finding the styled-components import. Several shapes were missed or mishandled. Each is now covered by a fixture:

• An import inside a TypeScript or Flow declare module block was treated as real and could hide the actual import.
• import * as styled lost styled.default.x decoration once the file also used the css prop or a default import.
• styled.default(Component), through a namespace import or require, emitted broken code (withConfig.withConfig(...)).
• import type / import typeof counted as the real import.
• The css prop crashed in CommonJS script files without a styled-components import, or with the require after the prop.
• import { styled as sc } and const sc = require('styled-components') weren't recognized under their real names. A bare global styled beside a require is still recognized, so existing componentIds don't move.

Faster transforms

The plugin's own time per file drops about 2–4× on typical files, and much more on CommonJS files with many css props (one large generated file went from about 1.8 s to about 65 ms). Emitted code is unchanged except that a file whose css props need styled-components injected now gets one import instead of one alias per prop.

• Import lookup reads the file's top-level statements instead of walking the whole file each time.
• css props share one injected default import per file.
• Whether a call is styled is checked once per call instead of up to four times.
• topLevelImportPaths patterns compile once per config.

pnpm bench compares any git ref against the working tree over generated inputs and fails if the outputs differ. pnpm bench:profile prints a CPU profile summary.

Tooling

• pnpm 12 with the global virtual store for local worktrees. CI uses the per-project store, because the global one fails to install on Windows runners.
• Tests now also run on Windows, and .gitattributes keeps checkouts LF.
• Fixture error snapshots hold only the message and code frame, with the working directory shown as <cwd> on every OS. This comes from a small patch to babel-test, and a test fails if any fixture contains an absolute path.
• Jest is scoped to test/.
• AGENTS.md is refreshed for contributors' tools.

[changeset-bot]

🦋 Changeset detected

Latest commit: 33c6157

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
babel-plugin-styled-components	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[pullfrog]

✅ No new issues found.

Reviewed changes — the minifier's string and paren tracking is rewritten to match the styled-components runtime, plus a pnpm 12 migration.

• String-aware comment stripping — stripLineComment is now a single pass that tracks the open quote (honoring backslash escapes) and counts parens only outside strings. A line comment after content: "it's" or content: "(" is now stripped instead of swallowing the rest of the joined template. Paren depth still resets per line, so an unmatched ) from a previous line is ignored, as before.
• String-aware symbol compression — compressSymbols tracks quote state across fragments via the new quoteOpenAfter helper instead of counting quote characters on the accumulated output, so whitespace inside a later string (' a ; b ') is preserved. This also drops the O(n²) rescan.
• Tests — adds regression coverage for other-quote-type strings, escaped quotes, parens inside strings, multi-line selector and at-rule preludes, and compressSymbols quote tracking.
• Tooling — pnpm 12.14.0 with virtualStoreType: global, pnpm/action-setup@v6.1.0, and Jest scoped to test/.

I verified the full suite passes (76 tests) and pnpm build compiles. I also fuzzed the old and new implementations of both functions across quote/paren/backslash/symbol alphabets: every divergence involves escaped quotes or backslashes outside a string, where the new behavior is the more CSS-correct one, and none is a regression on realistic input.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes — the seven commits since the prior pullfrog review are performance refactors to import lookup and the css-prop transform, plus regression fixtures, sitting alongside the already-reviewed minify fix.

• Faster import lookup — importLocalName reads Program.body directly for top-level imports and only falls back to a full Program.traverse when a TS/Flow module declaration (TSModuleDeclaration/DeclareModule, including export-wrapped) is present.
• One injected import per file — transpileCssProp caches the injected styled-components default import on state.file, so a file whose css props need an injection now gets one import instead of one alias per prop. This is the only emitted-code shape change and is runtime-equivalent.
• Single isStyled check per call — displayNameAndId curries isStyled once per plugin and extracts calleeNeedsConfig, replacing four repeated isStyled calls and the inline OR chain with an equivalent predicate.
• Matchers compiled once — useTopLevelImportPathMatchers memoizes compiled picomatch matchers per topLevelImportPaths value in a WeakMap.
• Regression fixtures — pin import lookup for ambient module declarations and exported ambient namespaces, the single injected import, and multiple css props under a namespace import.

I checked the refactors for behavior drift by compiling a 25-snippet corpus (tagged/call/attrs/withConfig shapes, namespace and rebind forms, require/interop/sequence shapes, TS ambient modules, and css props) across five configs with the plugin built at ffa23b0 versus c032536. The only output divergences are the intended one-import-per-file cleanup; every other result is byte-identical. The full suite (80 tests) and pnpm build both pass.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

ℹ️ No critical issues — one minor documentation consistency note.

Reviewed changes — docs-only delta on top of the previously reviewed minify fix and perf refactors.

• Added .changeset/faster-transforms.md describing the transform speedup.
• Added an AGENTS.md testing rule noting the css-prop fixture family needs one fixture built without @babel/preset-env (the CommonJS transform folds repeated imports into one require, hiding import-shape bugs).

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes — the six commits since the prior pullfrog review fix two import-detection bugs and record them.

• Ambient imports ignored — importLocalName now scans only the top level of Program.body, so a styled-components import nested in a TypeScript or Flow declare module / declare namespace body no longer masquerades as a real binding. Components in files with such ambient imports get their displayName/componentId back.
• Namespace imports tracked separately — the new importNamespaceNames records import * as X from 'styled-components' bindings, so X.default.y is recognized even when a default import is also present or injected by the css-prop transform.
• Regression fixtures — pin ambient imports in TypeScript declare module, Flow declare module, exported ambient namespaces, and import-free files; namespace-alongside-default and aliased-namespace-with-css-prop; three existing fixture outputs shift counter values where previously-unrecognized components now receive config, which is the intended effect.
• Docs — two patch changesets for the fixes, shortened changeset wording, and an AGENTS.md clarification that a changeset may state its own release's measured speedup.

I ran the full suite (84 passed) and pnpm build (17 files compiled). The ambient fix reads only top-level declarations and the namespace arm is gated behind the existing X.default. member shape, so neither widens recognition beyond genuine styled-components bindings.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes — the two commits since the prior pullfrog review add a benchmark/profiler toolchain and document it.

• A/B benchmark — scripts/bench/index.js transforms generated inputs (small/medium/large/large-cjs) with the plugin built at a git ref (--base, default origin/main) and the working tree, interleaves runs, and reports each variant's median time minus the no-plugin baseline. It exits non-zero when outputs differ, and hints at --normalize-aliases when only the injected _styledN aliases differ.
• Input generator and build helpers — gen.js synthesizes realistic styled, css-prop, and CommonJS inputs; build-ref.js extracts src at a ref via git archive and compiles it in a temp dir linked against the repo's node_modules, never writing the tracked lib/.
• Profiler — scripts/bench/profile.js records a .cpuprofile over one generated input and ranks self time, inclusive time, and the nearest plugin callers of a target frame.
• Docs — AGENTS.md gains the pnpm bench and pnpm bench:profile commands; the performance rule now points at pnpm bench.

I smoke-tested both. pnpm bench --rounds 1 flags the intended one-import-per-file output change on large-cjs.js with the alias hint, and --normalize-aliases clears it while showing current far below base; pnpm bench:profile --input small --runs 3 prints the ranked frames. The full suite and pnpm build still pass.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes — the four commits since the prior pullfrog review make the fixture snapshots machine-path-free and run CI on Windows.

• Machine-path-free error snapshots — patches/babel-test@0.2.4.patch makes babel-test snapshot the error message (which carries Babel's code frame) instead of the stack, replaces the working directory with <cwd> in either separator style and any drive-letter case, and rewrites the rest of each such path with forward slashes. I confirmed the patch applies cleanly to the pinned babel-test@0.2.4, that its sha256 matches the patchedDependencies hash in pnpm-lock.yaml, and that the normalization collapses both C:\... and /home/... cwds to <cwd>/....
• Absolute-path leak test — test/index.test.js scans every output.js/error.js for POSIX, drive-letter, and UNC absolute paths and fails the suite on any leak.
• Cross-platform CI — the test workflow gains a windows-latest/Node 24 job beside Ubuntu 22/24, sets PNPM_CONFIG_VIRTUAL_STORE_TYPE=project (pnpm's documented env override for virtualStoreType, which otherwise fails on Windows with the global layout), and a new .gitattributes (* text=auto eol=lf) keeps checkouts LF so snapshots match. The same env is set on the release workflow.
• Docs — AGENTS.md records the no-machine-paths snapshot rule and the CI virtual-store override.

I ran the suite with the patch applied (86 passed) and dry-run verified the patch against the installed babel-test. The Windows job on this head was still pending at review time, so the Windows snapshot path is verified by the normalization behavior rather than a green run.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes — the six commits added since the prior pullfrog review (the branch was rebased; the net delta is four import-detection fixes plus formatting and docs) each close a real detection gap and ship a fixture.

• Type-only imports ignored — importLocalName and importNamespaceNames skip importKind: 'type' | 'typeof' at both declaration and specifier level, so import type { styled } no longer counts as the real binding (and the css prop no longer emits a call to an erased identifier).
• Real local names — a named styled import now resolves to specifier.local.name, so import { styled as sc } is recognized; importLocalName('default') starts from the require variable name, and a globalStyledAlias keeps a bare global styled resolving beside a require.
• styled.default(Component) call forms — isNamespaceDefault recognizes ns.default on the default local, the global alias, or a namespace local across member, call, and (0, ns.default)(...) sequence shapes, so these no longer emit withConfig.withConfig(...).
• Script-mode css prop — the injected default import is tracked on state.file and a require() binding is treated as "no callable default", so a CommonJS file with the require after the prop no longer crashes and reuses one injected import.
• Formatting and docs — src/ reformatted with Prettier (no semantic change) and AGENTS.md refreshed.

I rebuilt the base plugin (origin/main) and the working tree through scripts/bench/build-ref and diffed them across the new shapes: the only divergences are the intended fixes (named aliases now decorated, call forms no longer doubled, type-only imports left alone, script-mode css prop injecting a real default). I also checked the require + global styled + css-prop interaction and it is byte-identical to main, so the injected-name override does not regress it. Full suite passes (98 passed) and pnpm build compiles.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}