cgg — Python bindings

Offline, deterministic call graphs for 44 languages, in-process.

import cgg

g = cgg.analyze("./src")
print(g.to_mermaid())

No network calls, no language servers, no build artifacts required. The analysis is the same Rust pipeline the cgg command-line tool runs, in the same order, so the two cannot disagree — there is a parity test that compares this module's JSON output against the binary's on the same tree.

Install

pip install cgg-callgraphgenerator

import cgg

The distribution is cgg-callgraphgenerator; the import is cgg. PyPI's cgg belongs to an unrelated GGUF tool, so the short name was not available. Python separates these two names routinely — pip install pillow gives you import PIL.

One caveat, because the other package also installs a top-level cgg module: do not install both into the same environment. Both write to site-packages/cgg/, pip will not stop you, and whichever lands second overwrites the first. If you already have pip install cgg (the GGUF tool), use a separate virtualenv.

The extension is built against the stable ABI (abi3-py39), so a single wheel per platform covers every CPython ≥ 3.9 — no per-version builds.

Prebuilt, that is one platform today: manylinux_2_17_x86_64. 0.6.3 also publishes an sdist, so pip install still works on macOS, Windows and aarch64 Linux — pip falls back to building from source there, which needs a Rust toolchain (≥ 1.85) and takes a few minutes. Wheels for those platforms build in CI and ship with the first tagged release.

Usage

import cgg

# Whole tree.
g = cgg.analyze("./src")

# A neighbourhood around what you care about.
g = cgg.analyze("./src", filter=[r"handle_request$"], hops=2)

# Several trees, one graph.
g = cgg.analyze(["./api", "./worker"], lang=["python", "go"])

g.to_mermaid()        # str — what agents read; byte-identical to `cgg -t mermaid`
g.to_json()           # str — `cgg -t json`, bar the per-run timings it embeds
g.to_dot()            # str — Graphviz
g.to_graphml()        # str — Gephi / yEd / networkx
g.to_dict()           # dict — the escape hatch

len(g)                # callable count
g.callables           # tuple[Callable, ...]
g.edges               # tuple[Edge, ...]
g.files               # tuple[File, ...]
g.metrics             # run counters
g.notices             # what the CLI would print to stderr
g.jobs                # worker threads the run actually used

g.callable("mypkg.mod.func")     # Callable | None
g.callers_of("mypkg.mod.func")   # list[Callable]
g.callees_of("mypkg.mod.func")   # list[Callable]

Finding code nothing calls

g = cgg.analyze("./src", dead_code=True, dead_code_confidence="high")
paths = {f.id: f.path for f in g.files}
for c in g.callables:
    if c.unreferenced:
        print(f"{c.unreferenced:6} {c.qualified_name}  {paths[c.file]}:{c.start_line}")

BEST EFFORT. Every finding is a hypothesis. It means cgg could not find a caller, not that none exists — reflection, FFI, a framework cgg has no rules for, and dynamic dispatch all produce callers it cannot see.

Filtering by trust

Every edge carries how it was established and how much cgg trusts it, so you can narrow to what you are willing to rely on:

solid = [e for e in g.edges if e.confidence == "high" and e.via == "direct"]

Two things worth knowing

Renderers never build Python objects. to_mermaid() and friends render straight from the Rust graph; g.callables constructs one Python object per callable, once, then caches. Measured on cgg's own crates/ (2,019 callables): to_mermaid() produces 180 KB in 1.5 ms, the first .callables access costs 0.84 ms, and every access after it costs 0.2 µs. Both are small here and both scale with the graph, so on a repository an order of magnitude larger the attribute path is what you would notice. Reach for the renderer when a string is what you want.

Concurrent analyze() calls actually run concurrently. The GIL is released (py.detach) and there is no internal lock, so a thread pool scales. N analyses of crates/cgg-lang/src/plugins from a ThreadPoolExecutor(N), against one analysis alone (56 ms) — medians of four repetitions, 32-core host, jobs at its default of 8:

Four analyses for 1.34x the wall clock of one; eight for 1.57x. Absolute numbers are machine-specific and each analysis is already internally parallel — regenerate them rather than trusting them.

Earlier builds would have had to take a process-wide lock for the whole of analyze, because extraction read two process-global switches (DEADCODE_SIGNALS and EXTRA_REGISTRAR_VERBS) that a second concurrent call would corrupt. Those now travel in a per-run cgg_lang::ExtractCtx, so there is no lock and no shared cell. Raising jobs on one call still works and is simpler if you only have one tree to analyze.

Not in this release

--why-live proofs, the --write-roots baseline and the audit event stream are reachable from the Rust API but have no keyword and no Graph attribute here. Use the CLI for those.

The framework-coverage table is not missing — it arrives rendered, as one of the strings in g.notices, naming both what cgg recognised and what it saw without rules. What is missing is a structured object; parse the notice or use cgg --framework-coverage if you need fields.

Building from source

scripts/build-python.sh          # from the repository root

Needs cargo (Rust ≥ 1.85), uv and git — the script checks for all three up front and stops if one is missing.

cargo build compiles the .so, but only maturin can make it importable — it writes the wheel metadata and puts the library where Python will find it. That is all build-python.sh does, plus provisioning an interpreter, since abi3-py39 rules out anything older than 3.9 and a system python3 often is.

The crate is an ordinary workspace member. It builds without a Python interpreter present at all: abi3 fixes the ABI at compile time and extension-module means libpython is never linked, so Py_* resolves at load time from whichever interpreter imports the module.

License

Apache-2.0 OR MIT.