low-level-code-review

Provide a git range such as uncommitted changes, a branch diff against master, or a specific commit range. The review scans only added lines for mechanical bugs and returns a worklist grouped by file with line numbers, original code, and concrete fixes.

It catches definite bugs like numeric overflow, off-by-one errors, null dereferences, uninitialized variables, resource leaks, use-after-free, and boolean logic errors. It also flags likely mistakes such as copy-paste errors, wrong variables, and signed/unsigned comparisons.

Split the review into pieces and run each piece as a subagent, for example one subagent per file or directory. Diffs with more than roughly 500 lines of additions should be split to keep each review focused and thorough.

Yes, it checks ACID semantics on disk by verifying write-to-temp-plus-rename patterns, fsync of the temp file before rename, and fsync of the directory after rename. It also examines in-memory atomicity, locking, and error rollback.

It never suggests refactoring, redesign, or architectural changes, and skips anything requiring broader system understanding. It also ignores issues a compiler warning would catch, intentional patterns, and lines not modified in the diff.