calldiff
calldiff answers "who calls whom", read straight from a git tree. It is
syntactic (tree-sitter, no type checking), so it is fast and needs no build,
and it will miss dynamic dispatch. Treat its output as evidence that narrows
attention — never as proof that nothing else is affected.
Installed globally: run calldiff, not npx calldiff.
The three modes worth using
| command | use for |
|---|---|
calldiff reach --entry <fn> --to <symbol> --locs | blast radius. Every call path from an entrypoint down to a target, each hop with file:line. Use before editing a shared function, and to answer "can this change reach X". |
calldiff diff --file <path> | structural review of one changed file: which calls this change added, removed, or swapped, shown in tree context. |
calldiff tree --file <path> | a call outline of one module, when you need its shape before reading it. |
Add --max-depth <n> when a tree runs long. --locs is cheap and almost
always worth it — line numbers make the output actionable.
Do not run bare calldiff diff
It reprints whole unchanged call trees around every change. Measured on a
mid-size React monorepo: 36k tokens, versus ~18k for the raw git diff —
double the cost for less information. Positional path scoping (diff HEAD <path>) silently returned nothing in the same test; do not rely on it.
Scope with --file or --entry, one target at a time. When output size is
uncertain, preflight it:
calldiff diff --file <path> --token-count
Where this earns its place
- Before editing a shared helper —
reachfrom the app entrypoint tells you which callers exist and where, in seconds, instead of a Grep sweep. - Scoping a review — drive
diff --filefrom the changed-file list (git diff --name-only), not across the whole repo. - Framework noise is real. In React code every
useState()anduseRef()counts as a call, so component trees are mostly noise. calldiff is most useful on plain modules of named functions, least useful on component trees.
Where it does not help
Dynamic dispatch, higher-order functions, DI containers, and hook indirection do not resolve. If the answer must be complete, calldiff is a starting point and not the answer. Say so when you report what it found.
Other output formats (--format json|md) and an MCP mode exist; the CLI text
output is the right default for reading and for pasting into a review.