What Is CallDiff and Why Do Call-Stack Diffs Matter for Agentic Review?

When an AI coding agent rewires call flow across a codebase, plain line diffs bury the shape of the change. You see dozens of removed and added lines, but the important question — did the agent actually reroute this function to the intended new callee? — stays hidden. CallDiff answers that question by diffing the call graph itself: which callees appeared, which disappeared, and which moved under an entrypoint.

The tool was created on 2026-08-07 by tanishqkancharla and already sits at 296 stars with 14 forks as of research on 2026-08-11 (GitHub API, api.github.com/repos/tanishqkancharla/calldiff). It is MIT-licensed, written in TypeScript, and installable via npx calldiff@latest or npm install -g calldiff.

The philosophy is an extension of the “shape-of-the-change” idea popularized by stacked-diff workflows. Jackson Gabbard’s widely-cited post “Stacked Diffs Versus Pull Requests” argues that diff-shaped review beats PR-shaped review because engineers review the shape of a change, not just the file layout. CallDiff applies that same insight one level deeper: to the call graph, not the file diff.

How Do You Install CallDiff and What Are the Three Commands?

Installation is a one-liner. With Node.js installed, run:

npx calldiff@latest

or globally:

npm install -g calldiff

On first use, CallDiff installs the tree-sitter grammars it needs into ~/.cache/calldiff/grammars. You can override this location with the CALLDIFF_GRAMMAR_CACHE environment variable if you need a custom or shared cache path.

CallDiff exposes three commands:

The diff command is the core. It compares two git refs (for example main and feature) and produces an output shaped like git diff: - means a callee that is gone from the “from” commit, and + means a callee that is new in the “to” commit.

How Do You Read a Call-Diff Output: ASCII Call Trees and +/- Semantics?

CallDiff prints colored ASCII call-stack trees by default. Each tree is rooted at an entrypoint, and the branches show which functions call which. When you run a diff, you get side-by-side or sequential trees where the +/- markers tell you precisely what changed in the call flow.

A minimal read: if a - appears under an entrypoint, that callee existed before and is now gone. If a + appears, a new callee was introduced. If a branch “moved,” you’ll see a function appear under a different parent — the same symbol relocated in the call graph. That moved-branch signal is exactly what a line diff cannot show you, and it is the most valuable output for verifying an agent’s work.

For agents and scripts, you can switch the output format with --format json|yaml|md|jsonl. JSON output is the most useful for programmatic consumption: you can pipe it into CI checks, feed it into a review agent, or write assertions that a specific callee did or did not appear in the new call stack.

CallDiff vs Codiff: 22 Languages vs 2

The closest direct competitor is codiff, a structural call-graph diff tool by issahammoud that is also built for coding agents. Both solve the same problem — showing what changed at the function/call level instead of the line level — but they diverge sharply on language support and ecosystem.

The single strongest objective advantage is breadth: 22 languages versus codiff’s 2. In a polyglot 2026 codebase, that is the difference between a tool you can standardize on and one you can only use in isolated corners. Codiff compensates with Mermaid diagram output for PR descriptions and a ready-made GitHub Action, and both are fully offline and syntactic. But if your team touches Python and TypeScript and Go, CallDiff’s 22-language reach is decisive.

Is CallDiff Agent-First? JSON, –llms, skills add, and MCP

CallDiff is designed to be consumed by AI agents, not just humans. That agent-first orientation shows up in several concrete features:

• --format json|yaml|md|jsonl: machine-readable output so an agent or CI script can parse the call-flow diff directly.
• --llms: an incur-native flag that emits an LLM-friendly rendering of the tool’s help and behavior.
• skills add and mcp add: CallDiff builds on the incur CLI framework (wevm/incur, 594 stars, TypeScript), which lets you register CallDiff as an MCP server or add it as an agent skill.
• CTAs after diffs: incur prints call-to-action prompts after diffs, guiding the next review step.

The underlying framework, wevm/incur, describes itself as a “CLI framework for agents and humans.” That dual target is the whole point: CallDiff is built to slot into agent toolchains as easily as it sits in a human developer’s terminal. The tool even nudges users to have agents walk through their changes — the prompt is “dearest clod, walk me through the code changes you made using npx calldiff@latest.”

What Are CallDiff’s Realistic Limitations?

CallDiff is syntactic (AST-based via tree-sitter), not a full typechecker. That design choice buys speed, error tolerance, and 22-language breadth, but it carries real limits you should know before trusting it blindly:

• Dynamic calls won’t resolve. If a function is called through a variable, dispatch table, or reflection, the AST has no static binding to follow, so CallDiff cannot trace it.
• Entrypoint inference is conservative. CallDiff only reports entrypoints whose expanded call trees changed, and it catches exported functions — not every internal helper.
• Grammars install on first use. The first run can be slow while tree-sitter grammars download into the cache; plan for that in CI.

These are the standard trade-offs of static call-graph extraction. CallDiff is a review aid, not a runtime profiler. It tells you what the code says is called at a structural level; it does not tell you what is actually executed under load. For verifying agent rewiring, that is usually exactly what you want — a fast, deterministic, offline check of call-flow intent.

A Walkthrough: Verifying an Agent Rewired a Call Flow Across Commits

Here is the practical workflow that makes CallDiff shine. Imagine an agent claimed it moved authentication logic from a legacy auth/ module into a new iam/ service, and you want to verify the rewiring without reading 200 lines of diff.

# 1. Install
npx calldiff@latest

# 2. Diff the two commits you care about
calldiff diff main feature

# 3. Machine-check the result in CI or from an agent
calldiff diff main feature --format json | jq '.entrypoints[].callees'

In step 2 you read the ASCII call tree: you should see - markers under the old auth.authenticate entrypoint and + markers under the new iam.authorize entrypoint, with the callee branches relocated. That relocation is the proof the agent actually rerouted the flow. In step 3 you can automate the assertion: check that no - remains under iam.authorize and that the expected new callee appears.

For a full review loop, hand the tool to an agent: paste “dearest clod, walk me through the code changes you made using npx calldiff@latest” and let it produce a call-flow walkthrough from the structured output.

Who Should Use CallDiff and How Does It Fit a 2026 Review Stack?

CallDiff is for two audiences:

1. Teams reviewing AI-agent changes. If your workflow includes agents rewriting call flow, line diffs are no longer enough. CallDiff surfaces whether the call graph was rewired as intended — the check that matters most.
2. Polyglot teams that want one review tool. With 22 languages, you can standardize call-graph review across Python, TypeScript, Go, Rust, Java, and more, where codiff only covers Python and TypeScript.

It fits naturally alongside your existing review stack: use git diff for the file-level shape, CallDiff for the call-graph shape, and CI assertions over the JSON output for automated verification of agent changes. It is MIT-licensed, offline, syntactic, and free — the only costs are the first-use grammar download and the honest caveat that dynamic calls stay invisible.

If you are reviewing agent-produced rewiring in 2026, CallDiff is the fastest way to see who-calls-whom change between commits.

FAQ

What is CallDiff? CallDiff is an open-source CLI that diffs function call stacks between two git commits, “like git diff, but for who-calls-whom.” It uses tree-sitter AST parsing to compare call graphs across 22 languages and is built for agentic code review.

How do I install CallDiff? Run npx calldiff@latest or npm install -g calldiff. It is MIT-licensed and written in TypeScript. On first use it downloads the tree-sitter grammars it needs into ~/.cache/calldiff/grammars.

What languages does CallDiff support? 22 languages: TypeScript, TSX, JavaScript, JSX, Python, Go, Rust, Java, Ruby, C, C++, C#, PHP, Kotlin, Swift, Scala, Lua, Elixir, Bash, Haskell, Zig, Solidity, and OCaml.

How is CallDiff different from codiff? CallDiff supports 22 languages versus codiff’s 2 (Python and TypeScript), and at research time had 296 stars versus codiff’s 7. Codiff adds Mermaid output and a GitHub Action; CallDiff adds JSON/YAML/MD/JSONL output, tree and reach commands, and MCP/skills integration via the incur framework.

Does CallDiff resolve dynamic calls? No. CallDiff is syntactic (AST-based), not a full typechecker, so dynamic or dispatched calls won’t resolve. It’s best used as a fast, offline check of static call-flow intent, not as a runtime profiler.