Benchmark bun fmt against Prettier

Part of #649. Runs after #654 closes (ruling 3 on #649).

Benchmark

Field	Value
Metric	(1) Exact: files that `bun fmt` formats differently from Prettier, among files with a Prettier parser, by file type. (2) Format gate wall time, in seconds.
Before	Not measured yet (agent in the repo, Step 2b, before any change). CI reference, from the 2026-11-11 milestone review read of the Run Report: `🔍 Lint, format & styles` step median 45 s (27–51 s, n = 7, 2026-11-07 to 2026-11-11). That step also runs lint, so the format share is not split out.
After	Not measured yet.
Condition	A `git archive HEAD` copy. Named laptop (model, cores, RAM, OS). Bun 1.4.4. Prettier version from `bun.lock` (3.9.10 at HEAD `0b302b222`). Warm runs, no other load. Record whether the #554 Q11 pool is on `main`. If it lands mid-benchmark, re-take Step 2b.
How to run	Step 2b: the gate's Prettier command from `scripts/check-ci.ts` (the `format` block) on the gate's file list, n = 5; and `bun run check:ci`, n = 5. Step 4: `bun fmt --list-different` on the same list, one exact count. Step 5: `bun fmt` on the same list, n = 5. Report median and min–max.
Pass rule	Adopt only if 0 files differ and the gate saves at least 10 % of the measured `check:ci` median. Any counted file: do not adopt; take the Decision below. Timing runs in every case.
Guard	`bun fmt --doctor`: every file type formats natively and the Tailwind sort comes from `.prettierrc.json`. If any file goes back to Prettier, the timing is void. `bun fmt --debug-check` runs once.
Owner	Agent in the repo: Steps 1 to 6 and the Measured comment. Maintainer: the reformat decision.
Claim	Adopt: "Formatting the whole docs.plus repo fell from s to s (median of 5, )." Kill: "bun fmt formatted of files differently from Prettier, so docs.plus stays on Prettier."

Why

Bun 1.4.4 added bun fmt, a formatter that reads our Prettier config. Vendor numbers: 286x faster than Prettier 3.9.10 on the Prettier repo, and 9,648 of 9,731 Prettier tests identical. On real projects only 57 of 74 runs (77%) matched Prettier exactly.

Our Prettier is ^3.9.10 (root package.json), the version Bun tests against. Record the resolved version from bun.lock in Setup. It runs in three places:

the format gate in scripts/check-ci.ts:166-207, on tracked files;
bun run format (prettier --check .) in .github/workflows/quality-lint.yml:26;
.lintstagedrc.js, on staged files.

.prettierrc.json:14 loads prettier-plugin-tailwindcss. With that plugin listed, the bun fmt Tailwind sort is already on (Bun docs).

Steps

The working tree is often dirty, so work in a copy: git archive HEAD | tar -x -C "$T", then bun install --frozen-lockfile --ignore-scripts in the copy. Never use a branch or a worktree.

1. Build the gate's file list in the source repo: git ls-files -z, without symlinks, as scripts/check-ci.ts does with lstat. Check that the count equals the gate's file count.
2. In the copy, run the gate's own Prettier command (scripts/check-ci.ts:184-197) on that list. It must pass.
2b. Before, always. Time that Prettier command 5 times on the gate's list. Also time bun run check:ci 5 times. Report median and min–max for both.
3. Run bun fmt --doctor once. Every file type must be formatted natively, and the Tailwind sort must come from .prettierrc.json. If any file is sent back to Prettier, the timing is void.
4. Pipe the list to xargs -0 bun fmt --list-different --ignore-unknown --ignore-path .prettierignore -- in the copy.
- Count a file only if prettier --file-info <file> gives a non-null inferredParser.
- List the others, such as the .toml files and bun.lock, as out of scope, not as differences.
- Group the count by file type.
5. In every case, time bun fmt on the same list, 5 warm runs, median and min–max. Label it "not adoptable" if any file is counted. Run bun fmt --debug-check once.
6. Post a Measured comment in the [Tracking] Bun 1.4.x: release-age guard, bunfig cleanup, and measured bun check, bun fmt and bun test numbers #649 format. State the gate saving and the check:ci median.

Kill rules

The kill rules decide adoption only. Steps 2b and 5 run in every case.

More than 0 counted files: post the list, and take the decision below.
0 counted files, but the gate saves under 10 % of the measured check:ci median: close. State both numbers in Measured.
This replaces the fixed 16 s line. That line was 10% of the 163 s local check:ci that Upgrade Bun to 1.4 and measure test plus backend wins #171 measured on Bun 1.4.0. Type-check packages/email-templates and packages/pad-url in the typecheck gate #658 changed check:ci since. [Tracking] Faster, change-aware CI/CD pipeline #554 Q11 (a pool of 4) is not filed or built yet; if it lands mid-benchmark, re-take Step 2b.

Decision (only if files are counted)

Accept a one-time repo-wide reformat to move to bun fmt? Recommended: no. Stay on Prettier until the real-project match improves, because a reformat conflicts with every open change.

If adopted later

Change the format scripts, .lintstagedrc.js and the scripts/check-ci.ts gate in one change. quality-lint.yml calls bun run format, so the script change covers it. Remove prettier-plugin-tailwindcss only after that.

Measured, 2026-11-12

Question: Does bun fmt format any gate file differently from Prettier, and does the format gate save at least 10 % of the measured check:ci median?

Setup

Field	Value
Commit	`cfac48d87` (the issue names `0b302b222`; HEAD moved before the run). Every copy was pinned to it.
Machine	Laptop, Mac17,6, 18 cores, 48 GiB, macOS 27.0.1
Bun	1.4.4
Prettier	3.9.10, with `prettier-plugin-tailwindcss` 0.8.1 (both from `bun.lock`)
Tree	A scratch clone of `cfac48d87` outside the repo, then `bun install --frozen-lockfile --ignore-scripts`, `bun x prisma generate` and `bash scripts/build-extensions.sh`. No env files.
Why a clone	`scripts/check-ci.ts` calls `git ls-files`. A bare `git archive` copy fails the format gate and skips both `build:ci` gates. The clone has the same files as `git archive cfac48d87`, plus `.git`.
Git state	`origin/main` set to `cfac48d87` in the clone, so the extension matrix is "none". The tree stayed clean before and after every run.
#554 Q11 pool	Not on `main`. `check-ci.ts` runs its gates one after another.
Load	Other work shared the laptop. A run counts only if the 1-minute load average was under 6 at its start. All counted runs started at 3.99 to 5.96.
Date	2026-11-12, 11:52 to 12:10 UTC

The #654 Measured draft came first, as ruling 3 on #649 sets the order.

Method

Step 1, file list: git ls-tree -r -z cfac48d87, without mode 120000 entries. That gives 3,108 of 3,116 tracked files (8 symlinks out). It equals git ls-files in the clone without lstat symlinks, file for file.
Step 2, gate command: xargs -0 node_modules/.bin/prettier --check --ignore-unknown --ignore-path .prettierignore -- on that list. It passed.
Step 2b, before: that command, then bun run check:ci. 1 discarded warm-up each, then 5 timed runs with /usr/bin/time -l. Every check:ci run printed PASS with both build:ci gates passing.
Step 3, guard: bun fmt --doctor --no-script, once.
Step 4, count: xargs -0 bun fmt --list-different --ignore-unknown --ignore-path .prettierignore --no-script -- on the list. Prettier's getFileInfo (the API form of --file-info) set the scope: files that are not ignored and have a parser.
Step 5, timing: xargs -0 bun fmt --check --ignore-unknown --ignore-path .prettierignore --no-script -- on the list. It ran in turn with the Prettier command, 1 discarded warm-up, then 5 runs. Then bun fmt --debug-check once.
Step 5 in the issue says "time bun fmt". With no flag, bun fmt writes files, so the timed command is --check, to match the gate's prettier --check.
Positive control: a second scratch clone got one format fault in each of 6 files. The files were .ts, .tsx (Tailwind class order), .md, package.json, .scss and .yml.

Result: differences (Steps 3 and 4)

Item	Count
Gate list	3,108 files
Ignored by `.prettierignore`	216
No Prettier parser (out of scope)	351
In scope (Prettier checks them)	2,541
`bun fmt` files selected	2,541, the same set
`bun fmt --doctor`: formatted natively	2,541; 0 sent back to Prettier
Files that `bun fmt` formats differently	0 of 2,541
`bun fmt --debug-check`	exit 0, 2,541 files

In scope by Prettier parser, 0 differences in each:

Parser	Files	Differ
typescript (`.ts`, `.tsx`)	2,123	0
markdown	153	0
babel (`.js`, `.mjs`, `.cjs`)	88	0
json	52	0
yaml	45	0
scss	30	0
json-stringify (`package.json` and similar)	25	0
html	13	0
css	12	0

Out of scope, not differences: 351 files with no Prettier parser. They are 189 .sql, 55 .png, 25 .sh, 4 .toml, bun.lock and 77 others, such as images, fonts, Git hooks and ignore files.
Guard: the doctor reads the Prettier config through package.json. It runs prettier-plugin-tailwindcss as a built-in plugin and sends no file to Prettier. The timing is valid.
Positive control: both tools listed the same 6 seeded files and exited 1. The formatted output of each file matched byte for byte. Both sorted className="p-4 flex" to "flex p-4".
A write run of bun fmt in the second clone printed "Formatted 0 files, 2541 unchanged", before the seeds.

Result: timing (Steps 2b and 5)

n = 5 each, 1 warm-up discarded:

Command	Median (s)	Min–max (s)	Peak RSS, median (MiB)
Gate Prettier command, 3,108-file list	10.61	10.03–11.41	570
`bun fmt --check`, same list	0.15	0.14–0.16	88
`bun run check:ci` (Prettier gate)	90.34	89.47–92.59	3,687 (largest single process)

Gate saving, median against median: 10.61 − 0.15 = 10.46 s.
10 % of the check:ci median: 9.03 s. The saving is 11.6 % of 90.34 s.
Worst case across the ranges: 10.03 − 0.16 = 9.87 s against 10 % of 92.59 s (9.26 s). It still clears the line.
The 9.03 s line replaces the fixed 16 s line, as the kill rules say.

Raw data

Attached: 655-format.jsonl, 655-checkci.jsonl, 655-fileinfo.json, 655-doctor.json, 655-list-different.txt (empty), 655-positive-control.txt. Each timing line holds the load average and free memory at its start.

Not measured

bun run format (prettier --check .) in quality-lint.yml, and .lintstagedrc.js on staged files.
A check:ci run with bun fmt in the gate. The saving comes from the two standalone commands.
bun fmt in CI. The CI reference step also runs lint, so it cannot isolate the format share.
Files outside the gate list, such as untracked files.

Verdict

Adopt rule met: 0 of 2,541 files differ, and the gate saves 10.46 s, 11.6 % of the 90.34 s check:ci median. The maintainer rules the swap.

Ruling, 2026-11-14: adopt

The maintainer ruled the swap at the v2.3.6 release audit. The adopt rule passed: 0 of 2,541 files differ, and the gate saves 10.46 s, 11.6 % of the 90.34 s check:ci median (Measured).

The swap is #795, a child of #554. It changes the format scripts, .lintstagedrc.js and the scripts/check-ci.ts gate in one commit. Prettier stays installed until a later change removes it.

Closing as completed.

Navigation Menu

Repository navigation

Benchmark bun fmt against Prettier #655

Benchmark

Why

Steps

Kill rules

Decision (only if files are counted)

If adopted later

Measured, 2026-11-12

Setup

Method

Result: differences (Steps 3 and 4)

Result: timing (Steps 2b and 5)

Raw data

Not measured

Verdict

Ruling, 2026-11-14: adopt

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Benchmark bun fmt against Prettier #655

Description

Benchmark

Why

Steps

Kill rules

Decision (only if files are counted)

If adopted later

Activity

HMarzban commented on Nov 12, 2026

Measured, 2026-11-12

Setup

Method

Result: differences (Steps 3 and 4)

Result: timing (Steps 2b and 5)

Raw data

Not measured

Verdict

HMarzban commented on Nov 14, 2026

Ruling, 2026-11-14: adopt

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions