Repository navigation
Description
Written by an AI agent (Claude) on behalf of @jasonkuhrt.
Problem
On macOS, a task that copies files into place with cp gets a cache entry without them, so a cache hit restores nothing.
fspy already swaps /bin/cp for its bundled coreutils so that copies can be tracked (spawn/macos.rs, COREUTILS_FUNCTIONS), and the swap does happen: inside vp run, cp --version prints cp (uutils coreutils) 0.4.0. But cp's writes still aren't recorded, while those of other swapped tools are.
Reproduction on vite-plus 1.4.0, macOS, APFS, with no remote cache:
// vite.config.ts
import { defineConfig } from 'vite-plus';
export default defineConfig({
run: {
tasks: {
generate: {
command: 'node generate.mjs',
cache: { input: ['generate.mjs'] },
},
},
},
});// generate.mjs: makes a file in the OS temp folder, then copies it into place
import { mkdirSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs';
import { execFileSync } from 'node:child_process';
import { tmpdir } from 'node:os';
import { join } from 'node:path';
const t = mkdtempSync(join(tmpdir(), 'gen-'));
writeFileSync(join(t, 'a.ts'), 'export const a = 1;\n');
rmSync('generated', { recursive: true, force: true });
mkdirSync('generated');
execFileSync('cp', [join(t, 'a.ts'), 'generated/a.ts']);$ vp run generate # generated/a.ts now exists
$ ls node_modules/.vite/task-cache/v18/*.tar.zst
ls: no matches found # no output archive was stored
$ rm -rf generated
$ vp run generate
vp run: cache hit, 103ms saved.
$ ls generated
ls: generated: No such file or directory
The same task, with only the line that puts generated/a.ts in place changed:
How generated/a.ts gets there |
Recorded as output |
|---|---|
fs.copyFileSync (Node) |
yes |
sh -c 'cat "$0" > generated/a.ts' (swapped shell) |
yes |
mkdir + touch (swapped coreutils) |
yes |
Homebrew python3 shutil.copytree |
yes |
cp src generated/a.ts (swapped coreutils) |
no |
cp -R src_dir generated |
no |
cp -R --reflink=never src_dir generated |
no |
Cause
The swapped cp creates its destination with clonefile(src, dst, 0), which the preload doesn't intercept; std::fs::copy's fallback, fclonefileat, isn't intercepted either. Evidence and experiments: see the first comment.
Proposal
Record the source of the macOS clone and copy calls as a read and the destination as a write: clonefile, clonefileat, fclonefileat, and copyfile. That fixes the swapped cp, and any other tool that copies the same way.
Related: #908 (files renamed into place aren't recorded either).
Metadata
Metadata
Assignees
Labels
Type
Fields
Priority
Activity
jasonkuhrt commentedon Oct 29, 2026
Cause confirmed: the swapped
cpcreates its destination withclonefile(src, dst, 0), and the preload doesn't interceptclonefile.clonefilewithdlsym(RTLD_NEXT, …)and calls it before it looks at the reflink mode (cp/src/platform/macos.rs#L50-L64), so--reflink=nevertakes the same path._dyld_get_image_namelists fspy's preload, andcpandtouchare the same multicall executable. A test interposer using__DATA,__interposecaught thedlsym'd call, so fspy'sintercept!can too.crates/fspy/build.rs; it was built with Rust 1.85.0.Each call below is a task's only write, with a fresh cache each time (macOS arm64, APFS, vite-plus 1.4.0,
VP_REMOTE_CACHE=off):open+writeclonefile, also throughdlsym(RTLD_NEXT, …)clonefileatfclonefileatcopyfile(COPYFILE_DATA), with or withoutCOPYFILE_CLONEopen+fcopyfileopenForcing
clonefileto fail sends uutils tostd::fs::copy, which on macOS triesfclonefileatnext (std/src/sys/pal/unix/fs.rs#L2052-L2084): also unrecorded. Only when both clones fail does itopenthe destination and callfcopyfile, and then the output is archived.Suggested fix: interpose
clonefile,clonefileatandfclonefileatnext to the existing interceptions (interceptions/mod.rs). Record the source as a read and the destination as a write, resolving the dirfd-relative forms withPathAtasopenatdoes (open.rs#L43-L61).copyfileneeds the same, with care for paths supplied through its state,COPYFILE_CHECKqueries, and recursive copies.Not tested: Linux, and a patched preload.