fix(server): git 2.56 broken-gitfile worktrees count as non-repositories again#33423

Problem

Git 2.56 changed what it prints when a worktree's .git gitfile points at a directory that no longer exists. Older releases print fatal: not a git repository: /missing-git-directory. Git 2.56 prints fatal: gitfile does not point to a valid repository: <path>/.git, from READ_GITFILE_ERR_NOT_A_REPO in git's setup.c.

The server decides "this directory is not a repository" by looking for not a git repository in stderr. Three places make that check:

• isNonRepositoryGitStderr in apps/server/src/vcs/GitVcsDriverCore.ts
• an inline copy in StorageCleanup.branchChanged (apps/server/src/storageCleanup.ts)
• an inline copy in isNotGitRepositoryError (apps/server/src/git/GitManager.ts)

With git 2.56, none of the three recognises a worktree whose gitdir has gone. Expected behaviour: storage cleanup keeps such a managed worktree as "Repository or branch changed". Actual behaviour: it reports the worktree as failed, with the raw gitfile does not point to a valid repository line as the reason. GitManager has the same problem: its status and PR lookups fail on that folder instead of using their existing non-repository fallback.

Reproduction: the existing live test in storageCleanup.test.ts corrupts a worktree's .git to gitdir: /missing-git-directory and expects cleanup to keep the worktree. The test passes with older git and fails with git 2.56.0, which GitHub-hosted ubuntu-24.04 runners now ship. We hit this failure on hosted CI in a fork.

Change

• isNonRepositoryGitStderr is exported and now matches either phrase, ignoring case.
• storageCleanup.ts and GitManager.ts call the shared classifier instead of keeping their own copies. The three checks then always agree, and all three change together if git rewords the message again.
• Nothing changes for real repositories. The only new case treated as non-repository is the one where git itself says the gitfile does not point to a valid repository. Other gitfile errors such as invalid gitfile format still fail, and a test row covers that.

Scope and approval

This is a small, focused fix for an obvious bug, so it qualifies without a prior issue. It is one defect: git reworded a message the server already classifies. There is no new behaviour; git 2.56 simply reaches the existing non-repository handling again. All four touched files serve that one fix: the classifier, its two inline duplicates and its test.

Verification

• New table test isNonRepositoryGitStderr in GitVcsDriverCore.test.ts has 4 rows that must classify as non-repository (both wordings, plus upper case) and 4 that must not (empty, unknown revision, index.lock exists, invalid gitfile format).
  • vp test run src/vcs/GitVcsDriverCore.test.ts -t isNonRepositoryGitStderr: 8 passed. Without the fix, the 2 gitfile rows fail.
• Regression comparison on Windows 11 with git 2.52, current main: vp test run src/vcs/GitVcsDriverCore.test.ts src/storageCleanup.test.ts src/git/GitManager.test.ts
  • on main (b8c2cfad6b): 37 failed / 512 passed
  • with this change: 37 failed / 520 passed (the 8 new rows)
  • The 37 failures are identical by name in both runs. They are pre-existing live-git failures in this Windows environment and unrelated to this change.
• pnpm run typecheck (apps/server) passes, and the commit hooks' vp fmt, vp lint and knip pass.
• Not checked locally: git 2.52 cannot produce the 2.56 message, so the end-to-end storageCleanup worktree test with git 2.56 is checked by CI on Linux. The same change, without the GitManager reuse, passed all 25 checks on GitHub-hosted runners with git 2.56 in the fork (Ali-Al1/t3code#4), where the storageCleanup test had failed before it.

Code written by Claude Opus 5.5 in Claude Code.

🤖 Generated with Claude Code