Repository navigation
Description
Written by an AI agent (Claude) on behalf of @jasonkuhrt.
Problem
With automatic output tracking, the default when cache.output is omitted or holds { auto: true }, a task that writes its output somewhere else and then renames it into place gets a cache entry without that output. A cache hit then restores nothing.
Writing to a temporary path and renaming is the standard way to replace a file or a folder atomically. TanStack Router's generator writes its route tree this way: a temp file, then rename (safeFileWrite, router-generator/src/generator.ts L1203-L1277). Our own code generators stage a whole folder and rename it over the old one.
Reproduction on vite-plus 1.4.0, 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: stages the output, then renames it into place
import { mkdirSync, renameSync, rmSync, writeFileSync } from 'node:fs';
rmSync('.staging', { recursive: true, force: true });
mkdirSync('.staging');
writeFileSync('.staging/a.ts', 'export const a = 1;\n');
rmSync('generated', { recursive: true, force: true });
renameSync('.staging', 'generated');$ vp run generate
$ node generate.mjs
$ ls generated
a.ts
$ rm -rf generated
$ vp run generate
vp run: cache hit, 39ms saved.
$ ls generated
ls: generated: No such file or directory # the hit restored nothing
Replacing the renameSync line with a direct writeFileSync('generated/a.ts', …) makes the same hit restore generated/a.ts.
Cause
fspy records the write to .staging/a.ts. It has no rename interception (fspy_preload_unix/src/interceptions/mod.rs covers access, dirent, open, spawn and stat), so generated/a.ts is never recorded. When outputs are archived, the recorded .staging/a.ts no longer exists and is skipped (archive.rs L29-L34), which leaves the entry empty.
The workaround today is a positive cache.output glob for the destination. But a glob archives whatever matches on disk, including files a previous run left there, so it gives up what automatic tracking gets right.
Proposal: cover the common case, leave the long tail to explicit config
Record a rename's destination as a write:
- A file renamed to
dst:dstis written. - A folder renamed to
dst: every file underdstis written. These can be collected by walkingdstwhen outputs are archived, instead of tracking each file inside.
Covering rename, renameat and renameat2 on Linux and rename and renameat on macOS should catch nearly every atomic-write tool, since that is how Node's fs.rename and most tools' atomic writers move files.
Deliberately out of scope, where users keep declaring cache.output:
RENAME_EXCHANGE;- moving files into place with
link; - platforms or syscalls the interception layer doesn't reach yet.
Cross-filesystem moves need nothing, since they copy and so are already recorded as writes.
Related: #904 covers the cache-hit side, where a hit leaves behind files the cached run did not write.
Metadata
Metadata
Assignees
Labels
Type
Fields
Priority
Activity