perf(cli): read literal config metadata without loading Vite

[fengmk2]

The check process initializes Vite even when vite.config.* exports only literal metadata. Read these configs with vp_static_config before calling the JavaScript resolver.

Accept a single export default object with literal lint, fmt, check, run, or staged fields. Imports, config functions, plugin hooks, Vite settings, and executable statements retain runtime resolution. Each call reads the file again so later commands observe changes.

Tool processes still resolve their own configs. The existing benchmark measures static exports and the defineConfig fallback separately.

Single-run comparison with time

Before: 76ce17ced. After: 0afe9b5c9. Each case ran once per version without warmup. Both versions used debug native bindings on Apple M4 Max, macOS 26.2, and Node v22.23.2.

The fixture contains three unformatted TypeScript files in packages/app/src. Config cases use export default { lint: {}, fmt: {} };. Each invocation starts with identical files and an empty NODE_COMPILE_CACHE directory.

The columns show real / user / sys from /usr/bin/time -p, in seconds. The no-config and standalone-tool cases are controls; their variation also shows single-run noise.

Case and CLI arguments	Before (s)	After (s)	Real-time change
Root, no config: check --fix	0.58 / 0.32 / 0.10	0.46 / 0.31 / 0.07	-20.7%
Root, with config: check --fix	0.77 / 0.59 / 0.13	0.69 / 0.52 / 0.13	-10.4%
Root, with config: fmt	0.31 / 0.24 / 0.05	0.31 / 0.24 / 0.05	+0.0%
Root, with config: lint --fix	0.28 / 0.24 / 0.05	0.27 / 0.25 / 0.05	-3.6%
Package, with config: check --fix	0.80 / 0.62 / 0.16	0.64 / 0.50 / 0.11	-20.0%

Example invocation from the fixture root:

NODE_COMPILE_CACHE="$(mktemp -d)" /usr/bin/time -p \
  node "$CHECKOUT/packages/cli/bin/vp" check --fix packages/app/src/*.ts

Set $CHECKOUT to the before or after worktree. Point the fixture's node_modules/vite-plus link to that worktree's packages/cli. Reset the input files before each invocation. Package cases run from packages/app with src/*.ts as the file arguments.

Controlled Linux comparison

CI run compares native revisions 76ce17ced and 0afe9b5c9 on one runner. Both use release bindings, Node v22.18.0, identical JavaScript, and the original CI fixtures.

The following medians use 21 samples per version and case. CPU time is user + sys from /usr/bin/time -p.

Case	Elapsed before → after (ms)	Change	CPU before → after (s)
root/check/minimal	495.6 → 417.8	-15.7%	0.48 → 0.38
root/check/defineConfig	496.2 → 501.8	+1.1%	0.50 → 0.49
package/check/minimal	480.5 → 417.3	-13.2%	0.47 → 0.38
package/check/blocks	495.9 → 429.5	-13.4%	0.49 → 0.39
root/fmt/minimal	171.2 → 172.1	+0.5%	0.15 → 0.15
root/staged/minimal	611.8 → 543.3	-11.2%	0.64 → 0.55

A third CLI snapshot uses the same optimized native binary. Its median differences stay within 1.2%. Each snapshot uses a separate warm compile cache.

Historical measurements from different runners do not isolate the code change. Use the comparison above to assess the second optimization.

Stacked on #2782. Related to #2698.

[github-actions]

CLI artifact sizes (68fe5af)

Final release artifacts built by the canonical build-upstream and build-windows-cli actions.
The dist rows use the Linux build. The core total excludes .node files to match the release artifact.

Artifact	Format	Base	PR	Change
packages/cli/dist	Directory total	2.12 MiB	2.12 MiB	0 B (0.00%)
packages/core/dist	Directory total	3.95 MiB	3.95 MiB	0 B (0.00%)
Combined package dist	Directory total	6.07 MiB	6.07 MiB	0 B (0.00%)
vp (Linux x64)	Binary	11.28 MiB	11.28 MiB	0 B (0.00%)
vp (Linux x64)	gzip -9	4.88 MiB	4.88 MiB	0 B (0.00%)
NAPI (Linux x64)	Binary	32.11 MiB	32.13 MiB	+16.00 KiB (+0.05%)
NAPI (Linux x64)	gzip -9	12.74 MiB	12.74 MiB	+4.38 KiB (+0.03%)
vp (macOS ARM64)	Binary	8.41 MiB	8.41 MiB	0 B (0.00%)
vp (macOS ARM64)	gzip -9	4.26 MiB	4.26 MiB	0 B (0.00%)
NAPI (macOS ARM64)	Binary	39.72 MiB	39.72 MiB	+16 B (+0.00%)
NAPI (macOS ARM64)	gzip -9	17.04 MiB	17.05 MiB	+7.02 KiB (+0.04%)
vp (Windows x64)	Binary	9.16 MiB	9.16 MiB	0 B (0.00%)
vp (Windows x64)	gzip -9	4.00 MiB	4.00 MiB	0 B (0.00%)
NAPI (Windows x64)	Binary	27.07 MiB	27.09 MiB	+18.00 KiB (+0.06%)
NAPI (Windows x64)	gzip -9	10.84 MiB	10.84 MiB	+7.33 KiB (+0.07%)
Trampoline (Windows x64)	Binary	13.50 KiB	13.50 KiB	0 B (0.00%)
Trampoline (Windows x64)	gzip -9	7.09 KiB	7.09 KiB	-1 B (-0.01%)
Installer (Windows x64)	Binary	4.56 MiB	4.56 MiB	0 B (0.00%)
Installer (Windows x64)	gzip -9	2.13 MiB	2.13 MiB	0 B (0.00%)

[github-actions]

Benchmark run for PR head 68fe5afb7c22d2bd6adaee9bbd4762f393ad6f99.

6 case(s) changed by more than ±5% in median time. Negative changes are faster; positive changes are slower.

Config performance

Baseline revision: daef3076e332d6f0074b2484be29cecc612d9b34.

Current revision: 41ce6bd004e01e4d8ab00e132637a298b4a07fa6. Node v22.18.0; linux/x64; AMD EPYC.

Case	Median (ms), baseline → current	p95 (ms), baseline → current	Median change	Config evaluations (current, separate probe)
root/check/no-config	223.0 → 252.4	244.0 → 269.6	+29.5 ms (+13.2%)	0
root/check/minimal	472.5 → 485.7	557.1 → 552.5	+13.2 ms (+2.8%)	check: 1, oxfmt: 1, oxlint: 1, oxfmt: 1
root/check/defineConfig	479.5 → 569.4	730.7 → 688.5	+89.9 ms (+18.8%)	check: 1, oxfmt: 1, oxlint: 1, oxfmt: 1
package/check/minimal	472.7 → 495.3	507.1 → 583.4	+22.6 ms (+4.8%)	check: 1, oxfmt: 1, oxlint: 1, oxfmt: 1
package/check/blocks	476.3 → 479.4	537.0 → 540.8	+3.1 ms (+0.7%)	check: 1, oxfmt: 1, oxlint: 1, oxfmt: 1
root/fmt/minimal	167.3 → 202.0	195.8 → 222.3	+34.7 ms (+20.7%)	oxfmt: 1
root/lint/minimal	138.3 → 163.4	158.4 → 203.4	+25.1 ms (+18.2%)	oxlint: 1
root/staged/minimal	609.8 → 660.8	680.2 → 758.8	+50.9 ms (+8.4%)	staged: 1, check: 1, oxfmt: 1, oxlint: 1, oxfmt: 1
root/staged/noop	146.3 → 193.6	170.0 → 215.3	+47.3 ms (+32.3%)	staged: 1

Fresh processes, warm filesystem and Node compile caches. Config logging is disabled during timing.
A timing regression requires >20% and >40 ms median growth, with current p25 above baseline p75.

• Regression: root/staged/noop: 146.3 → 193.6 ms (+32.3%)

[cloudflare-workers-and-pages]

🚀 Deploying Preview to Cloudflare 🚀

Preview URL: https://perf-static-config-metadata-viteplus-dev.voidzero-docs.workers.dev (commit 68fe5af)

This URL reflects your latest Preview deployment

Preview Deployments by commit

Status	Deployment URL	Commit	Updated (UTC)	See this deployment's details
Build: Success ✅ Deployment: Success ✅ View logs ↗	https://7a24ebc5-viteplus-dev.voidzero-docs.workers.dev	68fe5af	2026-09-22T15:11:55.276Z	Visit the dashboard ↗
Build: Success ✅ Deployment: Success ✅ View logs ↗	https://4f06d2de-viteplus-dev.voidzero-docs.workers.dev	0afe9b5	2026-09-22T14:10:18.844Z	Visit the dashboard ↗

[fengmk2]

The same-runner comparison confirms that the second optimization reduces config loading time. The earlier cross-run report did not isolate the code change.

Controlled CI run: Linux x64, Node v22.18.0, release native bindings, 21 samples per version and case. The CLI snapshots use identical JavaScript, dependencies, and fixtures. Only the native revision changes from 76ce17ced to 0afe9b5c9.

Case	Before → after (ms)	Change	CPU time before → after (s)
root/check/no-config	222.8 → 221.8	-0.4%	0.12 → 0.12
root/check/minimal	495.6 → 417.8	-15.7%	0.48 → 0.38
root/check/defineConfig	496.2 → 501.8	+1.1%	0.50 → 0.49
package/check/minimal	480.5 → 417.3	-13.2%	0.47 → 0.38
package/check/blocks	495.9 → 429.5	-13.4%	0.49 → 0.39
root/fmt/minimal	171.2 → 172.1	+0.5%	0.15 → 0.15
root/lint/minimal	139.7 → 132.7	-5.0%	0.15 → 0.15
root/staged/minimal	611.8 → 543.3	-11.2%	0.64 → 0.55
root/staged/noop	141.4 → 141.4	+0.0%	0.17 → 0.18

The elapsed-time column shows medians. CPU time is the median of user + sys from /usr/bin/time -p. A third CLI snapshot uses the same optimized binary; its median differences stay within 1.2%. The lint difference is not credited to static config resolution, which it does not use.

The original reports used different runners and compared historical medians. Matching AMD EPYC, CPU count and Node version did not control runner conditions. Unchanged paths became 15–20% slower in the earlier run. A rerun of unchanged product code now becomes 5.8–18.4% faster across the cases. These cross-job changes can hide the optimization.

Both CI source revisions differ only by this optimization. Dependency changes and unrelated merge commits are excluded. The original jobs lack CPU scheduling and frequency telemetry, so the precise host-level cause cannot be recovered. Their cache-hit versus fresh-build difference is not independently proven causal.

The exact-fixture local repeat also improves literal checks by 7.6–10.1%. Local debug and CI release builds explain why their absolute times and percentages differ.

The benchmark should compare base and head within one job, with balanced execution order and separate warm compile caches. Keep ±5% comments based on those measurements; use historical artifacts for trends. The temporary diagnostic instrumentation will be removed from the product PR.