pre-commit hook fails the test suite: git exports GIT_DIR into hooks, which overrides `git -C` in tests/test_pipeline.c

[metehanulusoy]

What happens

CONTRIBUTING.md tells contributors to install the repo hooks (git config core.hooksPath scripts/hooks). The pre-commit hook runs the full test suite. Git exports GIT_DIR into hook environments, GIT_DIR overrides git -C, and the tests in tests/test_pipeline.c that shell out to git therefore operate on the real repository instead of their temp fixture.

The result: with the documented hook installed, git commit fails the suite every time.

Reproduction — no hook needed, one variable

$ scripts/test.sh --suites pipeline
  264 passed

$ GIT_DIR="$PWD/.git" scripts/test.sh --suites pipeline
  git_context_linked_worktree   FAIL tests/test_pipeline.c:7444: run_cmd(cmd) == 32768, expected 0 == 0
  261 passed, 3 failed

Verified on main with no local changes (3627eff), macOS 15 arm64, git 2.50.1.

Line 7444 is:

snprintf(cmd, sizeof(cmd), "git -C \"%s\" checkout -b main >%s 2>&1", repo, null_dev);
ASSERT_EQ(run_cmd(cmd), 0);

With GIT_DIR set, -C no longer selects the repository: the command runs against the real checkout, where main already exists, so git exits 128.

It is not only a failing assert — the suite writes to the real repository

While tracing this I ran the test's own worktree add line with GIT_DIR set:

$ GIT_DIR=/path/to/codebase-memory-mcp/.git \
    git -C "$TMP/repo with space" worktree add -b feat2 "$TMP/wt with space"
Preparing worktree (new branch 'feat2')
Updating files: 100% (2118/2118), done.
HEAD is now at 17786374 Merge pull request #1993 from DeusData/fix/unify-type-short-name-index

It checked out 2118 files of the actual repository into the temp path and created a branch in the contributor's clone. So a contributor who follows the documented setup gets branches and worktree registrations created in their working repo by the test suite. git worktree prune plus git branch -D cleans it up, but nothing warns them it happened.

Why it is easy to miss

canonical_root_linked_worktree guards itself and reports SKIP (platform: git worktree add unavailable (git 2.5+ required)) under the same conditions — so the environment problem is rendered as a platform capability message rather than as the env-var collision it is. git_context_linked_worktree has no such guard and fails outright.

Suggested fix

Clear the git-supplied environment at the top of the hook — or in scripts/test.sh, so the suite is immune however it is invoked:

unset GIT_DIR GIT_WORK_TREE GIT_INDEX_FILE GIT_PREFIX GIT_OBJECT_DIRECTORY

Belt-and-braces alternative for the tests themselves: build the git commands as env -u GIT_DIR -u GIT_WORK_TREE git -C ..., so a stray GIT_DIR in any caller's environment can never redirect a test at the real repository. git_available() at tests/test_pipeline.c:7380 is the natural place to also assert a clean env once, rather than each site defending itself.

I am happy to open a PR for whichever of the two you prefer — the hook/test.sh unset is the one-line version, the env -u wrapper is the one that holds no matter who invokes the suite. Flagging first per CONTRIBUTING, since this touches build/CI configuration.

Environment

• main @ 3627eff, built from source with scripts/build.sh
• macOS 15 (Darwin 25.6.0), arm64, clang, git 2.50.1