libgaze 0.2.1

See what your code does to the world before it runs.

libgaze

See what your code does to the world before it runs.

$ libgaze check agent_tool.py

agent_tool.py  can Fs, Net, Unsafe

  restricted_import:75  can Unsafe
    99 | return __import__(name, custom_globals, custom_locals, fromlist or (), level)
  exec:119  can Unsafe
    126 | exec(code, {"__builtins__": SandboxPython.safe_builtins()}, locals)
  _run:194  can Fs, Net, Unsafe
    calls run_code_unsafe (line 347)
    calls run_code_safety (line 281)
  run_code_unsafe:347  can Unsafe
    365 | os.system(f"pip install {library}")
    370 | exec(code, {}, exec_locals)

2/13 functions are pure.

libgaze is a static effect analyzer for Python. It scans your code and reports which of 10 effects each function performs. Effects propagate through the call graph: if function A calls function B, A inherits B's effects.

Install

pip install libgaze

Usage

# Check a file
libgaze check myfile.py

# JSON output (for CI pipelines and agents)
libgaze check myfile.py --json

# Scan a directory
libgaze scan src/tools/

# Fail if denied effects are found (CI gate)
libgaze scan src/ --deny Unsafe,Db

# Check against a policy file
libgaze policy myfile.py --policy .gazepolicy

Policy files

A .gazepolicy file declares which effects are allowed or denied:

{
    "deny": ["Unsafe", "Db"],
    "functions": {
        "transform": { "allow": [] }
    }
}

The ten effects

Effect	What it means
Net	Touches the network
Fs	Reads or writes files
Db	Queries or mutates a database
Console	Reads or writes the terminal
Env	Reads environment variables
Time	Reads the clock or sleeps
Rand	Generates random numbers
Async	Spawns concurrent tasks
Unsafe	Subprocess, exec, eval, FFI, deserialization
Fail	Can fail (sys.exit, etc.)

How it works

Two-pass analysis:

1. AST walk. Detect direct effects from known functions (open, subprocess.run, os.getenv, etc.) and module imports (requests, sqlite3, docker, etc.).

2. Call graph propagation. Trace self.method(), ClassName.method(), and bare function() calls within the same file. If a callee has effects, the caller inherits them. Iterate until stable.

This catches the common patterns. It does not catch effects hidden behind getattr, dynamic imports, method calls on injected objects, or metaclass magic. Silence doesn't mean safe.

Benchmark

Two benchmarks: a labeled accuracy suite and an unlabeled scale scan.

Accuracy (labeled, 101 functions)

8 files, 101 functions with human-labeled # EXPECT: ground truth. Stdlib patterns, class method propagation, async, network, edge cases, pure code, and real-world agent tools from CrewAI.

precision  100.0%
recall     100.0%
F1         100.0%

uv run --extra dev python bench/run.py

100% on a benchmark you wrote yourself is expected. It proves the tool works on the patterns it was designed for and catches regressions. Add a .py file to bench/ with # EXPECT: comments to extend it.

Scale (unlabeled, 12,511 functions)

1,604 files across 4 real agent framework repos: CrewAI Tools, LangChain Community, LangChain Core, and AutoGPT.

1,604 files, 12,511 functions
8,251 pure (66%), 4,260 effectful (34%)
71 functions with Unsafe + other effects

Effect frequency across all effectful functions:

Effect	Count	% of effectful
Async	2,125	50%
Net	1,621	38%
Fs	342	8%
Time	250	6%
Db	238	6%
Console	220	5%
Env	168	4%
Unsafe	106	2%
Rand	20	<1%

Spot-checked against source code: every flagged effect verified as real. Known false negatives occur on method calls to injected objects (e.g., self._firecrawl.scrape_url() where _firecrawl is a FirecrawlApp instance set in __init__). These are correctly reported as pure because libgaze does single-file analysis and cannot resolve the type.

uv run --extra dev python bench/scan_repos.py

Limitations

libgaze uses static AST analysis. Python is dynamic. The things it catches:

• Direct calls to stdlib functions (open, os.system, subprocess.run)
• Imports of known modules (requests, docker, sqlite3)
• Intra-module call propagation (self.method(), ClassName.method())
• Builtins (print, input, exec, eval)

The things it misses:

• Method calls on injected objects (container.exec_run())
• Dynamic dispatch (getattr(obj, name)())
• Metaclass magic, monkey-patching, *args/**kwargs forwarding
• Cross-file analysis (effects from imported user modules)

It's a first line of defense, not a proof.

Part of Gaze

libgaze is the Python effect analyzer from the Gaze project. The same ten effects are used in:

• Gaze language — compiler-enforced effects (Rust)
• libgaze — static analysis for Python (this package)
• libgaze-ts — static analysis for TypeScript

The vocabulary is language-independent. It was tested against 15,293 functions across Python and TypeScript without modification.