Command
build
Is this a regression?
No. Present since the esbuild-based builder merged stylesheet metafiles this way.
Description
Build output is non-deterministic: two production builds of identical sources write different stats.json files, while every emitted JS, CSS and media file is byte-identical.
The cause is one loop. In packages/angular/build/src/tools/esbuild/angular/compiler-plugin.ts, build.onEnd merges each component stylesheet's esbuild metafile into the main metafile with Object.assign while iterating additionalResults.values(). That map fills as each stylesheet finishes bundling, so its order changes from build to build. When two stylesheets produce the same output file name, two components with the same file name and byte-identical SCSS, or two identical images referenced from different stylesheets, each metafile claims the output with its own inputs, and the last one merged wins. outputs[<file>].inputs then names one source on one build and the other on the next.
statsJson: true makes it visible, but the merged metafile is the one esbuild hands to every onEnd callback, so any consumer of result.metafile sees the same flip.
Minimal Reproduction
Repository: https://github.com/sdjayna/angular-metafile-merge-order-repro
npm ci
scripts/repro.sh # builds twice, compares stats.json, hashes dist/repro/browser
scripts/repro-loop.sh 5 # repeats the pair 5 times, counts flips
The app is a fresh ng new repro --style=scss with statsJson: true and outputHashing: "all" in the production configuration, plus 60 generated standalone components: 12 pairs share a file name and byte-identical SCSS across src/app/a and src/app/b, and 4 pairs reference byte-identical PNG files from two asset folders.
Expected: two production builds of identical sources write identical stats.json.
Actual: every build pair differed in 7 to 10 of the 54 outputs entries (5 of 5 pairs). Emitted files under dist/repro/browser had identical SHA-256 hashes in every pair.
build 1: 4 s
build 2: 2 s
emitted files under dist/repro/browser: identical (8 files)
stats.json: differs
outputs: 54
outputs whose inputs differ: 9
dup-0-4P5FWS4E.css
run1: src/app/a/dup-0/dup-0.scss
run2: src/app/b/dup-0/dup-0.scss
dup-10-FB2SWP2Q.css
run1: src/app/b/dup-10/dup-10.scss
run2: src/app/a/dup-10/dup-10.scss
media/pic-0-Q45TVVQX.png
run1: angular:css-resource:src/app/a/img/pic-0.png
run2: angular:css-resource:src/app/b/img/pic-0.png
Sorting additionalResults by key before the merge removes it: with that one-line change applied to node_modules/@angular/build, 0 of 3 pairs differed and stats.json was byte-identical. Reverting it brought the flip back. PR to follow.
Your Environment
Angular CLI : 22.1.9
Angular : 22.2.0
Node.js : 26.8.1
Package Manager : npm 11.19.0
Operating System : darwin arm64
Package Installed Version
@angular/build 22.1.9
@angular/cli 22.1.9
@angular/common 22.2.0
@angular/compiler 22.2.0
@angular/compiler-cli 22.2.0
@angular/core 22.2.0
typescript 6.0.3
Anything else relevant?
Reproduced on macOS 26.7 (arm64, 16 cores). The same loop exists unchanged on main (additionalResults.values() in compiler-plugin.ts), where the file is dist/browser-stats.json.
Command
build
Is this a regression?
No. Present since the esbuild-based builder merged stylesheet metafiles this way.
Description
Build output is non-deterministic: two production builds of identical sources write different
stats.jsonfiles, while every emitted JS, CSS and media file is byte-identical.The cause is one loop. In
packages/angular/build/src/tools/esbuild/angular/compiler-plugin.ts,build.onEndmerges each component stylesheet's esbuild metafile into the main metafile withObject.assignwhile iteratingadditionalResults.values(). That map fills as each stylesheet finishes bundling, so its order changes from build to build. When two stylesheets produce the same output file name, two components with the same file name and byte-identical SCSS, or two identical images referenced from different stylesheets, each metafile claims the output with its owninputs, and the last one merged wins.outputs[<file>].inputsthen names one source on one build and the other on the next.statsJson: truemakes it visible, but the merged metafile is the one esbuild hands to everyonEndcallback, so any consumer ofresult.metafilesees the same flip.Minimal Reproduction
Repository: https://github.com/sdjayna/angular-metafile-merge-order-repro
The app is a fresh
ng new repro --style=scsswithstatsJson: trueandoutputHashing: "all"in the production configuration, plus 60 generated standalone components: 12 pairs share a file name and byte-identical SCSS acrosssrc/app/aandsrc/app/b, and 4 pairs reference byte-identical PNG files from two asset folders.Expected: two production builds of identical sources write identical
stats.json.Actual: every build pair differed in 7 to 10 of the 54
outputsentries (5 of 5 pairs). Emitted files underdist/repro/browserhad identical SHA-256 hashes in every pair.Sorting
additionalResultsby key before the merge removes it: with that one-line change applied tonode_modules/@angular/build, 0 of 3 pairs differed andstats.jsonwas byte-identical. Reverting it brought the flip back. PR to follow.Your Environment
Anything else relevant?
Reproduced on macOS 26.7 (arm64, 16 cores). The same loop exists unchanged on
main(additionalResults.values()incompiler-plugin.ts), where the file isdist/browser-stats.json.