feat(bugfix-detect): broaden classification beyond Conventional Commit prefix - #18

The logic for detecting cross-repo issue references (e.g., owner/repo#123) is currently broken.

1. The loop at line 116 correctly consumes the repository prefix, but the condition at line 123 bytes[idx - 1] == b'/' is incorrect for standard owner/repo#123 syntax (where the character before # is part of the repo name, not a slash).
2. Repository names can also contain dots (.), which are currently not included in the allowed character set in the loop.
3. The GH- check is redundant because the input string has already been converted to lowercase.

I suggest updating the logic to check if the loop stopped at a # character and adding . to the allowed set.

fn followed_by_issue_ref(rest: &str) -> bool {
    // Allow optional : and whitespace, then either #NNN, gh-NNN, or
    // <owner>/<repo>#NNN. We only need a quick check, not a full parser.
    let trimmed = rest.trim_start_matches([':', ' ', '\t']);
    let bytes = trimmed.as_bytes();

    // Skip an optional <owner>/<repo> prefix (cross-repo references).
    let mut idx = 0;
    while idx < bytes.len()
        && (bytes[idx].is_ascii_alphanumeric()
            || bytes[idx] == b'-'
            || bytes[idx] == b'_'
            || bytes[idx] == b'/'
            || bytes[idx] == b'.')
    {
        idx += 1;
    }

    // If we stopped at a '#', we likely skipped a repository prefix.
    let after_org = if idx < bytes.len() && bytes[idx] == b'#' {
        &trimmed[idx..]
    } else {
        trimmed
    };
    let bytes = after_org.as_bytes();

    if bytes.starts_with(b"#") && bytes.len() > 1 && bytes[1].is_ascii_digit() {
        return true;
    }
    if bytes.starts_with(b"gh-")
        && bytes.len() > 3
        && bytes[3].is_ascii_digit()
    {
        return true;
    }
    false
}

Fixed in c3d902c. Rewrote followed_by_issue_ref(): find # directly, then validate the prefix splits on / into a non-empty owner and repo with only legal repo-name characters (alphanumeric, dots, dashes, underscores). Also dropped the redundant uppercase GH- check (input is already lowercased) and added cross-repo cases to the test suite (agent-sh/agnix#900, repos with dots and underscores, malformed refs).

The current test suite for bug-fix detection does not verify cross-repo issue references that lack a separate keyword (like Fix). Adding such a case would ensure the has_issue_closure logic is working correctly independently of the keyword heuristic.

        let delta = make_delta(vec![
            make_commit("alice", "fix: explicit conventional fix", file()),
            make_commit("alice", "Fix race condition in worker pool", file()),
            make_commit("alice", "Resolves #42", file()),
            make_commit("alice", "Resolves agent-sh/agnix#900", file()),
            make_commit("alice", "hotfix for prod outage", file()),
            make_commit("alice", "feat: add login flow", file()),
            make_commit("alice", "Add prefix support to parser", file()),
        ]);

        merge_delta(&mut map, &delta);

        let activity = &map.file_activity["src/lib.rs"];
        assert_eq!(
            activity.bug_fix_changes, 5,
            "expected 5 bug-fix commits (conventional + race + resolves + cross-repo + hotfix)"
        );

Done in c3d902c — added Closes agent-sh/agnix#900 (no fix keyword) to the aggregator integration test, bumping the expected count to 5 and exercising the cross-repo branch independently of the keyword heuristic.

Cross-repo issue reference parsing never matches

Medium Severity

followed_by_issue_ref fails to detect cross-repo references like owner/repo#NNN. The while loop consumes all alphanumeric, -, _, and / characters, so for "agent-sh/agnix#900" it eats agent-sh/agnix (stopping at # with idx=14). Then bytes[idx - 1] == b'/' checks the last consumed char, which is x (not /), so after_org falls back to the full string — which doesn't start with #. A subject like "Closes org/repo#42" (no fix keyword) silently returns false. The test on line 221 is masked because "Fix" triggers the keyword path, not the closure path.

 

^{Reviewed by Cursor Bugbot for commit 7056eee. Configure here.}

Fixed in c3d902c — replaced the broken character-eating loop with: find # directly, then validate prefix contains / and both segments use only repo-name-legal characters. Closes agent-sh/agnix#900 and Closes org/repo#42 (no fix keyword) now register correctly.