fix(git): report a detached HEAD from git_checkout

[CryoThrust]

Summary

git_checkout validates its argument with rev_parse, which resolves any revision — a sha, a tag, HEAD~1, a remote-tracking ref, a full refs/... name. For everything that is not a branch, git checkout detaches HEAD, but the tool still returned Switched to branch '<name>'.

Report the resulting state instead:

repo.git.checkout(branch_name)
if repo.head.is_detached:
    return f"HEAD is now detached at {repo.head.commit.hexsha[:7]}"
return f"Switched to branch '{repo.active_branch.name}'"

The reply is the only thing the calling model sees — content is a single text block, structuredContent is null and isError is false — so a wrong success sentence is not corrected downstream. The branch wording is now taken from active_branch too, so it cannot disagree with the ref that was actually checked out.

Rejecting non-branch revisions would also close the issue, but it removes a capability some callers use deliberately (checking out a tag or a remote ref), so I kept it and made the message truthful instead.

Tests

Added three cases to src/git/tests/test_server.py:

• a sha detaches HEAD and the reply says so, without "Switched to branch"
• a tag behaves the same
• a real branch still returns Switched to branch '<name>' and leaves HEAD attached

On main the first two fail (assert 'detached' in "Switched to branch 'v1'").

Validation: uv run pytest in src/git — 46 passed; pyright — 0 errors.

Closes #4804.

[yangshuaiyufalv]

[MCP demo] 测试评审（只读组织，预期权限错误）。

[buyan201430-code]

Reviewed as the reporter of #4804. I ran the head (d37bd1e) locally: pytest in src/git gives 46 passed, matching your note.

The fix also covers the three rows of the #4804 table that the new tests don't exercise. Checked each against this branch:

HEAD~1              -> HEAD is now detached at <sha>   (head.is_detached: True)
refs/heads/feature  -> HEAD is now detached at <sha>   (head.is_detached: True)
origin/main         -> HEAD is now detached at <sha>   (head.is_detached: True)

A name that is both a branch and a tag (dup) still attaches to the branch and returns Switched to branch 'dup', so taking the wording from active_branch holds there too.

One non-blocking suggestion: origin/main and refs/heads/<branch> are the two an agent is most likely to send, and they are the only ones not pinned by a test. This case passes on this branch and fails on main (3 failed with the old return line); ruff format --check is clean:

@pytest.mark.parametrize("revision", ["HEAD~1", "refs/heads/feature", "origin/main"])
def test_git_checkout_non_branch_revisions_report_detached_head(
    test_repository, tmp_path, revision
):
    test_repository.git.branch("feature")
    test_repository.index.commit("second commit")
    git.Repo.init(tmp_path / "remote.git", bare=True)
    test_repository.create_remote("origin", str(tmp_path / "remote.git"))
    test_repository.git.push("origin", "HEAD:refs/heads/main")

    result = git_checkout(test_repository, revision)

    assert test_repository.head.is_detached
    assert result.startswith("HEAD is now detached at ")

Otherwise this looks right to me — keeping non-branch checkouts and making the reply truthful is a better trade than rejecting them.

[CryoThrust]

Thanks for running it against the head and for checking the rows the tests didn't cover — that's exactly the gap I left, so I took the suggestion verbatim in 723b128.

test_git_checkout_non_branch_revisions_report_detached_head is parametrized over HEAD~1, refs/heads/feature and origin/main, with the bare-remote scaffolding from your snippet. Confirmed the signal both ways: with the old unconditional return all three fail, and with the fix they pass and assert result.startswith("HEAD is now detached at ") rather than just the is_detached flag.

src/git: 49 passed   (46 before, +3 from the parametrization)

Your point about the branch-and-tag name resolving to the branch is a good one to have on record too — that's the case where taking the wording from active_branch matters most, since echoing the argument would have been ambiguous there.