Git diffs that speak in functions, not line numbers

What it does

sem is a Rust CLI that wraps Git with tree-sitter parsing across 27 languages. It extracts functions, classes, methods, and other entities, then reports diffs, blame, and history at that level—“function validateToken modified” rather than “lines 12–18 changed.” It installs into existing repos with no configuration and can even hijack git diff output globally via sem setup.

The interesting bit

The matching logic is the quiet star: exact ID matches, structural hashing (to detect renames and moves while ignoring whitespace), and fuzzy token overlap for probable renames. This means sem knows when you renamed a function, not just that something vaguely similar disappeared and appeared. It also token-budgets context for LLMs, trimming dependency trees to fit a strict limit and flagging when even the target signature won’t fit.

Key highlights

• Entity-native commands: diff, blame, log, impact (cross-file dependency blast radius), and context (token-budgeted LLM context) all operate on named code entities.
• Agent plumbing: JSON output throughout, plus an MCP server exposing six tools (sem_diff, sem_impact, etc.) for direct AI agent integration.
• Broad language coverage: 27 programming languages plus structured formats (JSON, YAML, TOML, CSV, Markdown), with fallback chunk-based diffing for everything else.
• Rename/move detection: Three-phase matching catches structural renames and moves that line-based diffing misses entirely.
• GNU Parallel collision: The binary name sem conflicts with GNU Parallel’s symlink; the README documents workarounds but it’s a genuine papercut.

Caveats

• The project is explicitly part of a larger “agent-native” stack (Ataraxy Labs) with companion tools like weave and inspect—useful context, but also means sem is positioned as infrastructure rather than a standalone end-user product.
• 2,590 stars is respectable but early; some surface area (e.g., exact accuracy of impact analysis across dynamic languages) is hard to verify from README claims alone.

Verdict

Worth a look if you’re building coding agents, CI pipelines, or developer tools that need to reason about code semantically. Skip it if you just want prettier line diffs—delta already solved that.