Closes #2420

Problem

vp migrate derives several independent boolean signals from a package's source tree — browser mode, retained upstream vitest module references, and one source-scan per opt-in browser provider. Each ran its own full recursive traversal via sourceTreeMatches, reading every eligible TS/JS file once per signal.

On the complete-miss path (no browser/provider references, no tsconfig retaining Vitest types, no webdriverio dependency to short-circuit) that is 5 reads per source file for a standalone project, exactly as the issue reports.

Changes

• sourceTreeMatches is generalized to sourceTreeMatchesEach, which evaluates several content predicates in a single pass: each file is read once and offered to every predicate that has not yet matched, unwinding as soon as all are decided. sourceTreeMatches becomes a one-predicate wrapper over it, so there is still a single traversal implementation.
• New collectPackageSourceScanSignals enrolls all the per-package signals — browser mode, both provider scans, and the retained-vitest-module scan — in that one pass.
• Both per-package call sites use it: rewriteStandaloneProject and rewriteMonorepoProject.
• The standalone path stops rescanning the tree for the webdriverio provider. As the issue notes, providerSourceModes already holds that result.

Two deliberate limits:

• The tsconfig short-circuit is preserved. sourceTreeReferencesRetainedVitestModule returns early when a tsconfig already settles the signal; in the combined scan that predicate is simply not enrolled in that case, so the tree is not walked for it.
• Results are not shared between the workspace and package phases of a monorepo migration. The issue points out that rewriteMonorepoProject performs config merging between them, which may create or modify vite.config.ts, so cross-phase reuse would need a separate correctness argument. This PR consolidates within each phase only — the fallback the issue describes as safe. The workspace phase's own per-package scans are left for a follow-up.

Per-predicate semantics are unchanged: same traversal order, same skip directories, same nested-package boundary, same "unreadable file is ignored and scanning continues" behaviour.

Performance

Measured on the full-miss path — the worst case, and the one #2420 reports: a standalone project with no browser/provider references anywhere, no tsconfig retaining Vitest types, and no webdriverio dependency to short-circuit a scan.

Each cell is the median of 15 timed iterations after 3 warmups, and each row is the median of three such rounds; run-to-run spread stayed under 8%. Node v22.22.2, Linux x64, 4-core Xeon @ 2.10 GHz, warm page cache.

What is timed is the per-package scan phase exactly as each revision's rewriteStandaloneProject performs it:

• before — collectProviderSourceModes (one traversal per opt-in provider = 2) + usesVitestBrowserMode (1) + sourceTreeReferencesRetainedVitestModule (1) + the redundant usesWebdriverioProvider rescan in the package.json callback (1) = 5 traversals;
• after — collectPackageSourceScanSignals = 1 traversal.

The measured read counts came out at exactly 5 per source file before and 1 after, independently reproducing the amplification the issue reports and matching the regression test's expected 5 to be 1.

Wall time falls ~3.3× rather than the full 5× because only the read and UTF-8 decode per file are removed — each file's content is still matched against the same predicates either way. These are warm-page-cache numbers, so they understate the gain on a cold cache.

source-scan.ts cannot be imported on its own in a bare checkout: it imports projectUsesVitestDirectly from the ../migrator.ts barrel, which transitively reaches the native NAPI binding. That import — along with readPackageJsonIfExists and WorkspacePackage — is used only by workspaceUsesVitestDirectly, which is not on the per-package scan path.

So the benchmark loads each revision's source-scan.ts with its first 40 lines (the import block plus that one function) replaced by a prelude that inlines the three constants it needs from shared.ts and a JSON-parsing stand-in for hasVitestTypesInTsconfig. Everything from line 41 onward is byte-identical to the shipped module at that revision (verified with diff), so every measured function is the real one. The hasVitestTypesInTsconfig stand-in is never actually invoked, because the fixture contains no tsconfig.

The fixture is generated: N plausible ~1.4 KB TS modules, 20 per directory under src/, none containing any scan hint string. Before timing, each run asserts that every signal came back false, so a fixture that accidentally short-circuited a traversal would fail rather than silently produce a flattering number.

Testing

New regression test reads each source file once during a standalone migration counts fs.readFileSync calls against a full-miss fixture and asserts each source file is read once. It also asserts every file is read at least once, so a fixture that stopped exercising the scan would fail rather than silently pass.

Verified in both directions:

• with source-scan.ts and orchestrators.ts reverted to main and the test kept, it fails: AssertionError: expected 5 to be 1 — independently reproducing the issue's measured standalone count;
• with the change applied, it passes.

Checks actually run on this branch:

• vitest run packages/cli/src/migration — 409 passed across 10 files
• vitest run (full unit suite) — 1023 passed, 1 skipped; 7 pre-existing snapshot failures in packages/prompts/src/__tests__/render.spec.ts, identical on unmodified main in this environment (verified by re-running on a clean tree: 1022 passed with the same 7 failures)
• tsgo -b tsconfig.json — error set byte-identical to unmodified main here (87 pre-existing errors from unbuilt docs/workspace packages, none in the touched files)
• vp lint --type-aware --type-check — clean for the touched files
• vp fmt --check — clean

Not run: the PTY snapshot suite and ecosystem e2e. This change alters no CLI output, only how many times the same files are read.

AI assistance

Claude Opus 5 wrote the implementation, the test, the benchmark, and this description. The change is agent-authored and has not had a separate human review. The results quoted above are from actual runs, not estimates.

Generated by Claude Code