[xzyangha]

Issue for this PR

Closes #58371

Type of change

• Bug fix

What does this PR do?

Snapshot capture aborts when the work tree contains a path that is not a regular file. Git.tree.capture lists the changed and untracked paths and hands them to git update-index --add --remove --replace, and Git refuses sockets, FIFOs and device nodes (fatal: Unable to process path …). One such path fails the whole update-index call, so the capture throws and no snapshot exists for that step.

The change stats the candidates once and leaves out the paths Git refuses, so the rest of the tree is still captured. A path that no longer exists has no stat information and is still passed through, because --remove handles deletions; directories are passed through as well, since a tracked path replaced by an embedded repository is recorded as a gitlink.

fs.stat follows links, so a symlink to a socket reports the socket's type. Git records a link itself and never opens its target, so those paths are kept: the ones with another type are only dropped when readLink shows they are not links.

This is driven by the path type rather than by which list produced it, so it covers both a tracked path that became a socket and an untracked socket that Git lists on Windows. Snapshot docs now list non-regular paths among what capture leaves out.

How did you verify your code works?

• New cases in packages/core/test/git.test.ts: a tracked file replaced by a unix socket is left out of the tree while capture succeeds and records an unrelated edit made in the same capture; a symlink whose target is a socket is still recorded as a link (120000 blob). On dev the first fails with fatal: Unable to process path a.txt, the error from the issue.
• bun test test/git.test.ts test/snapshot.test.ts in packages/core, plus bun run typecheck:tsc and oxlint on the changed files.

Screenshots / recordings

Not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

---

Added by the opencode bot (AI-generated).

Related: #48657

[opencode-agent]

This fixes the reported failure, but it introduces one regression that should be fixed before merge.

I checked it by running the new test against the base code: it fails there with fatal: Unable to process path a.txt, the error from the issue, and passes with this change. test/snapshot.test.ts passes. tsc -b for packages/core succeeds. I tested on Linux at the Git capture layer only, not on Windows, where the issue was reported.

Public API: unchanged.

The regression: fs.stat follows symlinks, so a symlink that points to a socket (for example docker.sock -> /var/run/docker.sock) is now dropped from the snapshot, though Git could record it as a link before. I confirmed this with a temporary test: the base captures link.sock and this change doesn't. Details are inline.

Related: #48657

[opencode-agent]

The symlink regression from my last review is fixed: paths whose type is neither file nor directory are now checked with readLink, and symlinks are kept, so a symlink to a socket is captured as a link again. The new test checks that it is stored as a link.

In packages/core, test/git.test.ts and test/snapshot.test.ts pass (29 tests), and tsc -b succeeds. The socket test fails on the base with the error from the issue and passes here. I tested on Linux only, not on Windows, where the issue was reported.

Public API: unchanged.

No new concerns.

[xzyangha]

@jlongster this one is ready for a look when you have a moment.

It fixes #58371: Git.tree.capture hands every changed and untracked path to git update-index --add --remove --replace, and Git refuses sockets, FIFOs and device nodes, so a single one anywhere in the work tree aborts the whole call and that step ends up with no snapshot. The change stats the candidates once and leaves the refused paths out by type, so the rest of the tree still gets captured; symlinks are kept, since Git records the link itself without opening its target. Three files: the fix, two tests (a tracked file replaced by a unix socket; a symlink whose target is a socket), and a docs line.

The bot review found a symlink regression on the first revision and reports no new concerns on the current head. check-compliance and check-standards are green, but check / test / nix-eval have not started: this is a fork PR, so they need an "Approve and run workflows" from a maintainer first.

[opencode-agent]

The new push leaves the code exactly as it was at my last review: the tree is identical, and only the commit was rewritten. My earlier results still apply. Symlinks are now kept, which fixes my earlier point, the tests pass, and tsc -b succeeds. I tested on Linux only, not on Windows, where the issue was reported.

Public API: unchanged.

No new concerns.

[Dextheking1]

one stat driving both checks is clean. symlinks kept, the rest skipped. good.

[xzyangha]

@rekram1-node could you take a look at this one when you get a chance?

It fixes #58371: Git.tree.capture hands every changed and untracked path to git update-index --add --remove --replace, and Git refuses sockets, FIFOs and device nodes, so a single one anywhere in the work tree aborts the whole call and that step ends up with no snapshot. The change stats the candidates once and leaves the refused paths out by type; symlinks are kept, since Git records the link itself without opening its target. Three files: the fix, two tests, one docs line.

The review bot reports no new concerns — it caught a symlink regression on the first revision, which is fixed in the current head. check-compliance and check-standards are green, but check, test and nix-eval are sitting in action_required: as a fork PR they need an "Approve and run workflows" from a maintainer before they start.

[xzyangha]

@rekram1-node — this one still applies cleanly after #59419 and #59417: updateIndex is unchanged on current dev, so the stat-once + non-regular-file filter goes in exactly as written, and GitHub reports the branch mergeable.

Status right now: check-compliance and check-standards pass, the bot review has no new concerns (the symlink regression it caught on the first revision is fixed — symlinks are kept and covered by a test), and check / test / nix-eval are in action_required because this is a fork PR, so they need an "Approve and run workflows" before they can start.

If you'd rather have it without the docs line, or with the two socket tests merged into one, say which and I'll adjust.