Tempdir out of the project; TS parse cache; fast local test default - #34
Open
brainkim wants to merge 7 commits into
Open
Tempdir out of the project; TS parse cache; fast local test default#34brainkim wants to merge 7 commits into
brainkim wants to merge 7 commits into
Conversation
…ject Three cost/annoyance fixes: - dts: TypeScript lib files and node_modules types are parsed ONCE per process instead of once per createProgram. Every build paid the full lib.d.ts parse; libuild's own suite runs 136 builds, making this the suite's dominant compute. Fixture files are never cached; SourceFiles are immutable and all programs use identical options, so sharing is safe. The three dts-heaviest test files pass unchanged. - test bundles live in a per-run mkdtemp under the OS temp dir, not <cwd>/.libuild-test: no gitignore noise, no litter after crashed runs, and two simultaneous runs in one project no longer share a directory and delete each other's bundles (real latent collision). Externalized-package resolution is preserved by mirroring the cwd's node_modules ancestry into the temp dir with symlinks (junction on Windows), one nested level per ancestor - verified e2e on plain consumers (bun+node) and a hoisted workspace member with no local node_modules. --debug keeps and prints the dir. - npm test is fast raw `bun test` again; the self-hosted dual-platform gate is test:all, run by CI (release.yml). Local machines stop paying for what CI exists to do. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
bun defaults to 5s per test, so a legitimately slow test died inside the shard long before the per-file budget mattered - "--timeout 60000" silently didn't govern the failures users actually saw (fold's corpus tests, 6-10s each, failed on bun for this alone). The file budget now passes through as bun's per-test timeout too; one number serves both, since no single test can legitimately exceed its file's budget. node has no default per-test cap and its --test-timeout flag postdates our supported floor, so node stays governed by the file budget alone. Also adds the README section the runner never had: usage, flags with the real timeout semantics, and runtime notes (bun's no-nested-test() limitation and the ESLint RuleTester pattern that trips it, per fold's migration report; the ESM-only policy). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Both steps of the CI-staging bootstrap - npm trust AND npm stage publish - require the package to already exist, so the first publish is the unique moment the setup becomes possible, and the stage-only grant (--allow-stage-publish) is the part nobody discovers unaided. After a publish of a previously-unpublished name, libuild now prints the exact next command, with owner/repo parsed from the manifest's repository field (verified against npm trust's documented syntax). Suggested by the fold session after its stage preflight - which it triggered by accident via a shell-quoting slip - turned out to be the clearest statement anywhere of the package-must-exist constraint. The existence check is non-blocking (offline skips the hint), and the hint also prints on --dry-run of an unpublished name, matching the existing dry-run "Published!" looseness. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The prepublishOnly-script guard accreted three mechanisms around one
idea: an executable shell one-liner in the manifest, substring
detection of it ("exit 1"), and installation coupled to --save's
manifest rewrite. Meanwhile a consumer hand-rolling the same guard had
a quoting slip that command-substituted and EXECUTED a release command
mid-edit - a safety feature complex enough to fire the gun it guards.
`private: true` is the entire mechanism now: npm itself refuses to
publish private packages. One word, nothing executable, no detection
heuristics needed (the substring check stays only to honor existing
script guards). The missing-guard warning names the one-word fix.
The dist manifest is unaffected - it never carried the guard, which is
why real publishes work.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This reverts commit f751d2a.
'Use libuild publish instead' was wrong guidance for anyone whose release path is stage - which is why fold forked the message, and forks drift. Brian's wording covers both paths in fewer words. Applied to our own root manifest too. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Three responses to "running tests is way too expensive" + ".libuild-test is annoying":
No more
.libuild-testin the projectBundles live in a per-run
mkdtempunder OS tmp. Fixes two real defects beyond the annoyance: crashed runs no longer litter the project, and two simultaneouslibuild testruns in one project no longer share the directory and delete each other's bundles. Externalized-package resolution is preserved by symlink-mirroring the cwd'snode_modulesancestry into the temp dir (one nested level per ancestor with anode_modules, outermost first) — so hoisted workspace members resolve root-hoisted deps exactly as before. Verified e2e: plain consumers on bun+node, hoisted-workspace member, and post-run project cleanliness.--debugkeeps and prints the dir.TS parse cache (the suite's furnace)
Every
build()created a fresh TS program, re-parsinglib.d.tsand all node_modules declarations — 136 times per suite run. Immutable inputs are now parsed once per process via a caching CompilerHost; fixture files are never cached. This speeds every consumer's builds, not just our suite. The three dts-heaviest test files (40 tests incl. declaration-content assertions) pass unchanged in 23.5s.Local test economics
npm test= rawbun test(fast dev loop).test:all= the self-hosted dual-platform gate, run by CI on every release. Dogfooding stays a hard gate; laptops stop paying for it.Verification note
Deliberately NOT fully suite-verified locally (the point of the change is to stop doing that); targeted files green, e2e consumers green, and CI's
test:allis the gate before this can ship in a release.🤖 Generated with Claude Code
https://claude.ai/code/session_01QAc3iDuBb3zN41HrepErUw