Overview

7 days30 days

Latest CI Pipeline Executions

Fuzzy

Succeeded
36514
e5ed5d42 test(core): cover the FIFO the config read's O_NONBLOCK exists for `readOwnedFileSync` opens with `O_NOFOLLOW | O_NONBLOCK`. The symlink case covers the first flag; the second had no executable coverage, so dropping it would have left every test passing while restoring a hang. It is worth being explicit about the failure mode, because it is unusual: if `O_NONBLOCK` is dropped, this test hangs rather than failing, until the suite timeout kills it. A hung git-utils suite means look at the open flags in `readOwnedFileSync`. That is a worse signal than a red assertion and a much better one than none, since the flag is what keeps a planted FIFO from blocking a read that happens at module scope of `cache-directory.ts` -- every command in that workspace would hang before printing anything. `child_process` is mocked in this file, so the FIFO is created through `vi.importActual`; the `isFIFO()` assertion is there because without it a mocked `mkfifo` silently creates nothing and the test passes vacuously.
by Craigory C...
Failed
36514
6d0986af fix(core): refuse a git config that is not ours before reading it Builds on the previous commit's checks rather than replacing them. Shape tells us a directory is a repository; it does not tell us whose. A real repository owned by another user has `HEAD` and `objects` too, and the walk goes up past the workspace into directories the caller may not own -- `/tmp` is mode 1777, and the workspace need only not be a git repository itself. Git refuses a foreign-owned repository for exactly this reason (`safe.directory`, CVE-2022-24765), so deferring to git lands on the same refusal rather than a different answer. - `.git` directories go through `isOwnedRealDirectory`, which also `lstat`s, so a symlink standing in for `.git` is refused. - The linked-worktree branch gets the same treatment. `gitdir:` and `commondir` are paths read out of file contents, so that branch could be pointed at any directory; it was returning one unchecked. - Config, the `.git` pointer and `commondir` are read through one helper that opens with `O_NOFOLLOW | O_NONBLOCK` and takes the file type and the owner from `fstat` on the descriptor it then reads. `O_NOFOLLOW` keeps the symlink refusal the previous commit's `lstat` provided, `O_NONBLOCK` keeps a FIFO from blocking the open, and reading through the descriptor closes the window between the check and the read that an `lstat`-then-open leaves. Also drops the `NX_CACHE_DIRECTORY` guard in `migrate-ui-api.spec.ts`. The `vi.clearAllMocks()` added alongside it is sufficient on its own -- measured by removing each in turn -- and the env guard coupled the spec to the order of the guards inside `computeSharedDataLocation`.
by Craigory C...
Canceled
36514
6d0986af fix(core): refuse a git config that is not ours before reading it Builds on the previous commit's checks rather than replacing them. Shape tells us a directory is a repository; it does not tell us whose. A real repository owned by another user has `HEAD` and `objects` too, and the walk goes up past the workspace into directories the caller may not own -- `/tmp` is mode 1777, and the workspace need only not be a git repository itself. Git refuses a foreign-owned repository for exactly this reason (`safe.directory`, CVE-2022-24765), so deferring to git lands on the same refusal rather than a different answer. - `.git` directories go through `isOwnedRealDirectory`, which also `lstat`s, so a symlink standing in for `.git` is refused. - The linked-worktree branch gets the same treatment. `gitdir:` and `commondir` are paths read out of file contents, so that branch could be pointed at any directory; it was returning one unchecked. - Config, the `.git` pointer and `commondir` are read through one helper that opens with `O_NOFOLLOW | O_NONBLOCK` and takes the file type and the owner from `fstat` on the descriptor it then reads. `O_NOFOLLOW` keeps the symlink refusal the previous commit's `lstat` provided, `O_NONBLOCK` keeps a FIFO from blocking the open, and reading through the descriptor closes the window between the check and the read that an `lstat`-then-open leaves. Also drops the `NX_CACHE_DIRECTORY` guard in `migrate-ui-api.spec.ts`. The `vi.clearAllMocks()` added alongside it is sufficient on its own -- measured by removing each in turn -- and the env guard coupled the spec to the order of the guards inside `computeSharedDataLocation`.
by Craigory C...
Succeeded
36514
6d0986af fix(core): refuse a git config that is not ours before reading it Builds on the previous commit's checks rather than replacing them. Shape tells us a directory is a repository; it does not tell us whose. A real repository owned by another user has `HEAD` and `objects` too, and the walk goes up past the workspace into directories the caller may not own -- `/tmp` is mode 1777, and the workspace need only not be a git repository itself. Git refuses a foreign-owned repository for exactly this reason (`safe.directory`, CVE-2022-24765), so deferring to git lands on the same refusal rather than a different answer. - `.git` directories go through `isOwnedRealDirectory`, which also `lstat`s, so a symlink standing in for `.git` is refused. - The linked-worktree branch gets the same treatment. `gitdir:` and `commondir` are paths read out of file contents, so that branch could be pointed at any directory; it was returning one unchecked. - Config, the `.git` pointer and `commondir` are read through one helper that opens with `O_NOFOLLOW | O_NONBLOCK` and takes the file type and the owner from `fstat` on the descriptor it then reads. `O_NOFOLLOW` keeps the symlink refusal the previous commit's `lstat` provided, `O_NONBLOCK` keeps a FIFO from blocking the open, and reading through the descriptor closes the window between the check and the read that an `lstat`-then-open leaves. Also drops the `NX_CACHE_DIRECTORY` guard in `migrate-ui-api.spec.ts`. The `vi.clearAllMocks()` added alongside it is sufficient on its own -- measured by removing each in turn -- and the env guard coupled the spec to the order of the guards inside `computeSharedDataLocation`.
by Craigory C...
Canceled
36514
4d37c4d5 fix(core): defer to git wherever the config parser cannot match it The `.git/config` fast path answered confidently in four cases where git answers differently, and its result feeds both the shared cache directory and the repository name sent to Nx Cloud onboarding. Git ends a value at an unquoted `#`/`;` and honours `\` escapes, so `url = git@github.com:acme/app.git # mirror` parsed as the slug `acme/app.git # mirror`. It also resolves `[url] insteadOf` before reporting a remote, so a cross-host rewrite reported the mirror rather than the real host, which silently disabled the GitHub checks in `performance-report` and `connect-to-nx-cloud`. Both now hand the file to git, matching the existing `include`/`includeIf` behaviour. `locateGitDir` accepted any directory named `.git`, where git requires a repository: a `.git` holding only a config file, planted in a writable ancestor such as `/tmp`, decided the identity of every workspace beneath it. It now checks for `HEAD` and `objects` first. Reading the config blind also followed a symlink out of the repository, and would have blocked forever on a FIFO -- this resolves at module scope of `cache-directory.ts`, so every command in that workspace would hang before printing anything. An `lstat` for a regular file covers both. Every case degrades to spawning git rather than to a wrong answer.
by FrozenPandaz
Succeeded
36514
065f7cf4 perf(core): read the git remote from .git/config instead of spawning git `cacheDir` resolves at module scope and derives the workspace identity, so every Nx process paid two subprocesses on its import path -- one CLI run, one per isolated plugin worker, and again under NX_DAEMON=false. Measured on this repo: `git remote -v` 23.09 ms (it spawns a shell as well as git), `git rev-parse --show-toplevel` 16.09 ms. Locating `.git` by walking up answers both questions at once, so neither subprocess is needed on the happy path. Measured after: 0.390 ms per derivation, and byte-identical with git removed from PATH, which is what proves nothing is spawned. - locateGitDir() handles the linked worktree and submodule case, where `.git` is a FILE holding `gitdir:` and the remotes live in the shared common dir that `commondir` names. - The config parser returns null on include/includeIf rather than answering from a partial view, so git decides those. - Both paths fall back to git: config read, then `git remote -v`, then the first-commit SHA. A failed parse must not skip to that last one, which is an unbounded history walk. getGitRootPath() no longer shells out either, so it now reports the root as the caller referred to it rather than the realpath. Both its callers are better for it on macOS, where mkdtemp and process.cwd() give /var/folders while git reports /private/var/folders: - deriveRepoKey compared the two and never produced '' for a repo root, so the repo key embedded a traversal segment. Four workspace-id specs were red on every macOS machine and green on Linux CI. The shared directory name changes once for a symlink-crossing workspace -- one cold cache. - GitRepository.root feeds `relative(destinationGitClient.root, absDestination)` in nx import, where absDestination is join(process.cwd(), destination). Mixing a realpathed root with an unresolved destination produced the same traversal garbage; both operands now come from the same namespace. The migrate-ui-api spec mocks fs wholesale, so the config fast path cannot read there and it falls back to spawning git, which its no-subprocess assertion would attribute to undoMigration. Settling the cache location up front keeps that assertion blunt rather than narrowing it.
by Craigory C...
Failed
36514
f35f988a fix(core): clear mocks between migrate-ui-api git invocation tests Resolving `cacheDir` now derives the workspace identity, which shells out to git at module load. Those calls land on the `child_process` mocks before the first test runs, so `should pass the ref to git as an argument rather than through a shell` saw a `git remote -v` it did not make and failed its `expect(execSyncMock).not.toHaveBeenCalled()`. That assertion is the regression guard for #36379 -- it is what proves `undoMigration` never routes a `migrations.json` ref through a shell -- so it is worth keeping intact rather than loosening.
by FrozenPandaz
Canceled
36514
3315d152 fix(core): clear mocks between migrate-ui-api git invocation tests Resolving `cacheDir` now derives the workspace identity, which shells out to git at module load. Those calls land on the `child_process` mocks before the first test runs, so `should pass the ref to git as an argument rather than through a shell` saw a `git remote -v` it did not make and failed its `expect(execSyncMock).not.toHaveBeenCalled()`. That assertion is the regression guard for #36379 -- it is what proves `undoMigration` never routes a `migrations.json` ref through a shell -- so it is worth keeping intact rather than loosening.
by FrozenPandaz
Failed
36514
31066c8a docs(core): correct what NX_WORKSPACE_DATA_DIRECTORY relocates The row claimed daemon logs and the file map "stay in `.nx/workspace-data` either way" -- false precisely for the reader who sets the variable it documents. `workspaceDataDirectoryForWorkspace` resolves `NX_WORKSPACE_DATA_DIRECTORY ?? NX_PROJECT_GRAPH_CACHE_DIRECTORY` ahead of the default, and `DAEMON_DIR_FOR_CURRENT_WORKSPACE`, `nxFileMap`, `nxProjectGraph` and `nxSourceMaps` all derive from it, so setting the variable moves every one of them and nothing is written to `.nx/workspace-data` at all. Someone debugging a daemon problem would have followed the row to `<root>/.nx/workspace-data/d/daemon.log` and found nothing. Says instead what the checkout-local directory holds, and that this variable relocates all of it -- the sharing it opts out of is the project graph database.
by Craigory C...
Failed
36514
223423ef docs(core): correct which checkouts share the cache Three pages said the directory is "shared by every checkout of the repository". Two clones of one repository sit at different paths, hash differently and do not share; only a checkout and the worktrees added to it do. The hash keys the checkout, not the repository, so the intro to the cache-location page was wrong on both counts. Also updates the sample error on the unknown-local-cache page, which still quoted a `.nx/cache/...` artifact path.
by Craigory C...
Failed
36514
ee971d45 refactor(core): drop the perWorkspace callback from sharedDataDirectory The callback existed because the sharing decision lived in `worktree-data-root.ts`, which could not import `cache-directory` without a cycle, so each caller had to inject its own unshared resolver. Folding the two modules together removed the cycle and put both resolvers in the same file as the function taking the callback, leaving a parameter with nothing behind it -- and a doc comment still claiming it avoided an import it no longer needed. Switching on `kind` says the same thing in one place, and the call sites lose an argument that only ever had one correct value each. Claude-Session: https://claude.ai/code/session_01Sr9WDU54dksQvnDxB6csPZ
by Craigory C...
Failed
36514
9f04017b refactor(core): drop the perWorkspace callback from sharedDataDirectory The callback existed because the sharing decision lived in `worktree-data-root.ts`, which could not import `cache-directory` without a cycle, so each caller had to inject its own unshared resolver. Folding the two modules together removed the cycle and put both resolvers in the same file as the function taking the callback, leaving a parameter with nothing behind it -- and a doc comment still claiming it avoided an import it no longer needed. Switching on `kind` says the same thing in one place, and the call sites lose an argument that only ever had one correct value each. Claude-Session: https://claude.ai/code/session_01Sr9WDU54dksQvnDxB6csPZ
by Craigory C...
Failed
36514
d4564206 refactor(core): drop the perWorkspace callback from sharedDataDirectory The callback existed because the sharing decision lived in `worktree-data-root.ts`, which could not import `cache-directory` without a cycle, so each caller had to inject its own unshared resolver. Folding the two modules together removed the cycle and put both resolvers in the same file as the function taking the callback, leaving a parameter with nothing behind it -- and a doc comment still claiming it avoided an import it no longer needed. Switching on `kind` says the same thing in one place, and the call sites lose an argument that only ever had one correct value each. Claude-Session: https://claude.ai/code/session_01Sr9WDU54dksQvnDxB6csPZ
by Craigory C...
Failed
36514
Fix ready→
09403711 refactor(core): drop the perWorkspace callback from sharedDataDirectory The callback existed because the sharing decision lived in `worktree-data-root.ts`, which could not import `cache-directory` without a cycle, so each caller had to inject its own unshared resolver. Folding the two modules together removed the cycle and put both resolvers in the same file as the function taking the callback, leaving a parameter with nothing behind it -- and a doc comment still claiming it avoided an import it no longer needed. Switching on `kind` says the same thing in one place, and the call sites lose an argument that only ever had one correct value each. Claude-Session: https://claude.ai/code/session_01Sr9WDU54dksQvnDxB6csPZ
by Craigory C...
Canceled
36514
f64bbd4e refactor(core): drop the perWorkspace callback from sharedDataDirectory The callback existed because the sharing decision lived in `worktree-data-root.ts`, which could not import `cache-directory` without a cycle, so each caller had to inject its own unshared resolver. Folding the two modules together removed the cycle and put both resolvers in the same file as the function taking the callback, leaving a parameter with nothing behind it -- and a doc comment still claiming it avoided an import it no longer needed. Switching on `kind` says the same thing in one place, and the call sites lose an argument that only ever had one correct value each. Claude-Session: https://claude.ai/code/session_01Sr9WDU54dksQvnDxB6csPZ
by Craigory C...
Canceled
36514
c400d2a2 docs(misc): describe the cache where Nx now writes it `.nx/cache` was named as the default in five places. It is now `~/.nx/<hash>/cache` for every checkout, so each of those was about to be wrong, and a reader caching `.nx/cache` in CI would have had no warning. Two corrections to the worktree page beyond the relocation: "Nx decides this by attempting the directory rather than by recognizing the agent, so any sandbox is covered" was only true on a first run. Once any unsandboxed run had created the directories, the probe was four no-ops and a write-denied agent was told to use the shared root anyway. The probe now writes a marker, so the prose is true as written. "`nx reset` clears the shared directory for the current repository" only held from inside a worktree, since a main checkout resolved to no shared directory at all. It holds everywhere now. Claude-Session: https://claude.ai/code/session_01Sr9WDU54dksQvnDxB6csPZ
by Craigory C...
Canceled
36514
c400d2a2 docs(misc): describe the cache where Nx now writes it `.nx/cache` was named as the default in five places. It is now `~/.nx/<hash>/cache` for every checkout, so each of those was about to be wrong, and a reader caching `.nx/cache` in CI would have had no warning. Two corrections to the worktree page beyond the relocation: "Nx decides this by attempting the directory rather than by recognizing the agent, so any sandbox is covered" was only true on a first run. Once any unsandboxed run had created the directories, the probe was four no-ops and a write-denied agent was told to use the shared root anyway. The probe now writes a marker, so the prose is true as written. "`nx reset` clears the shared directory for the current repository" only held from inside a worktree, since a main checkout resolved to no shared directory at all. It holds everywhere now. Claude-Session: https://claude.ai/code/session_01Sr9WDU54dksQvnDxB6csPZ
by Craigory C...
Succeeded
36514
1e1820fd chore(core): move the worktree-data specs to vitest and keep the analytics gate packages/nx moved to vitest in #36754, which has no `jest` global, so worktree-data-root.spec.ts would have thrown at runtime rather than conflicting. Restacking onto the rebased #36586 also picked up #36663's analytics opt-in, so the sandbox block keeps its gate: the socket and filesystem grants are written either way, only the egress follows the opt-in.
by Craigory C...
Failed
36514
Flaky task→
1e1820fd chore(core): move the worktree-data specs to vitest and keep the analytics gate packages/nx moved to vitest in #36754, which has no `jest` global, so worktree-data-root.spec.ts would have thrown at runtime rather than conflicting. Restacking onto the rebased #36586 also picked up #36663's analytics opt-in, so the sandbox block keeps its gate: the socket and filesystem grants are written either way, only the egress follows the opt-in.
by Craigory C...
Succeeded
36514
737a56be refactor(core): nest worktree data under one directory per repository Per Jason: `~/.nx/<hash>/{cache,workspace-data}` rather than `~/.nx/{cache,workspace-data}/<hash>`. One repository's shared data is now one directory instead of two siblings under separate roots. That also shortens the establish list from five directories to four, and gives the reclaim work in NXC-4875 a single unlink per repository rather than one per kind. The layout is pinned with a literal expected path rather than one derived from the same `join` the implementation uses, so flipping the nesting back has to be a deliberate edit to the test. RED-verified: restoring the old order fails four rows, including that pin.
by Craigory C...