Skip to content

Turbocharge Claude Code, Cursor, Codex, Gemini & every coding agent: faster, cheaper, with contextual understanding specific to your codebase.

Let the Windows leg fail the build

anirudhkumar-nanonetsfba562dacc23
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 43ba7ff2..358f49f7 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -15,23 +15,15 @@ jobs:
# Windows is not redundant coverage here: graft stores repo-relative paths and
# parses them with `/` by hand, so a platform-separator leak matches nothing
# rather than erroring. That class of bug is invisible to a posix-only matrix
- # (it shipped in 0.8.2 and took three user reports to find). `fail-fast: false`
- # so a Windows-only failure still reports the Linux result.
+ # (it shipped in 0.8.2 and took three user reports to find). This leg gates, and
+ # earned it immediately — it caught npx detection never firing on Windows and a
+ # seeded worktree silently losing its freshness record. `fail-fast: false` so a
+ # Windows-only failure still reports the Linux result.
strategy:
fail-fast: false
matrix:
os: [ubuntu-latest, windows-latest]
runs-on: ${{ matrix.os }}
- # The Windows leg reports but does not yet gate. It currently fails 21 tests
- # that predate it and have nothing to do with path handling: `chmod 000` does
- # not deny access on Windows (so the "unreadable file" and exec-bit tests
- # cannot hold as written), several tests assert `/` in paths that are
- # deliberately printed with the native separator, and a few spawn child
- # processes that exit early. Blocking on those would mean either reverting the
- # leg or rushing 9 unrelated test files; leaving it red-but-blocking would
- # train everyone to ignore CI. Once the 21 are cleared, delete this line —
- # the leg is only worth having if it can fail the build.
- continue-on-error: ${{ matrix.os == 'windows-latest' }}
steps:
- name: Checkout code
uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4
diff --git a/CHANGELOG.md b/CHANGELOG.md
index f9a5b0b2..83ed5866 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,28 @@
### Fixed
+- **Windows: `graft upgrade` no longer reinstalls over an npx run.** The npx-cache check
+ matched `/_npx/` against a path that arrives with the platform separator, so it was
+ always false on Windows and `graft upgrade` ran `npm install -g` instead of explaining
+ that npx already fetches the latest build on every run.
+
+- **Windows: a git worktree kept the graph it was seeded with, but not the record that
+ makes it cheap.** The seed's copy filter dropped every sidecar next to the graph — the
+ freshness fingerprint included — so the query right behind the seed found no
+ fingerprint, could not diff against the parent checkout, and re-parsed the whole repo.
+ It answered correctly the whole time, which is why nothing reported it; the only
+ visible trace was a `(? files changed)` note instead of a count.
+
+- **`graft init` printed one path with two separators** on Windows (`~\.codex/`), from a
+ `/` concatenated onto an otherwise native display path.
+
+- The `windows-latest` CI leg now **gates** rather than merely reporting. The 21 failures
+ it shipped with are resolved: two were the real bugs above, most of the rest were tests
+ asserting `/` in paths that are deliberately printed with the native separator or
+ pointing a child process at `HOME` (Windows reads `USERPROFILE`), and four are now
+ explicit named skips — nothing in Node's `fs` can deny a *read* on Windows, and there
+ is no exec bit or `SIGTERM` to test.
+
- **Windows: path scoping and `map` work again.** graft stores a repo-relative path
for every indexed file — in node ids, `node.path`, the extract cache, the freshness
fingerprint — and it was produced with `relative()`, which returns the *platform*