agentgap
v0.1.0-alpha.4
Published
Read-only reliability and safety auditor for AI agent repositories
Maintainers
Readme
AgentHarness (npm wrapper)
AgentHarness is a read-only auditor for AI agent repositories. It finds the reliability and safety gaps that agent projects share — unbounded provider retries, unguarded side effects, swallowed exceptions, leaked secrets — and reports them with file, line, and symbol evidence.
It never edits your source. Everything it writes lives in a .agentharness/ directory.
AgentHarness is a Python tool. This package is a thin npm wrapper for teams whose toolchain is
already Node-based: installing it creates a private virtual environment inside the package
directory and installs the matching agentharness wheel from PyPI into it. The harness command
then forwards every argument and exit code to that interpreter.
Requirements
- Node.js 18 or newer
- Python 3.12, 3.13, or 3.14 available on
PATH
Installation fails with instructions if no supported interpreter is found. Set
AGENTHARNESS_SKIP_INSTALL=1 to skip the Python install step, for example when vendoring
node_modules in an image build that installs the Python package separately.
Install
npm install -g agentgap
harness --versionThe package is agentgap; the command it installs is harness.
If you do not need the Node entry point, install the Python package directly instead:
uv tool install agentgap # or: pipx install agentgapUsage
Run it with no arguments inside any repository. This reads your code and writes nothing to it.
harness # read-only audit of the current directory
harness report # render the findings as Markdown for review or for a coding agent
harness improve # audit, stage a diff, ask for approval, apply, then verifyFor CI, harness check . --fail-on high exits 1 when an active finding meets the threshold and
0 otherwise.
What it finds
| Rule | Severity | What it detects |
| --- | --- | --- |
| AH-R101 | high | A model-provider call with no detected reliability instrumentation |
| AH-R102 | medium | A broad except Exception that can hide provider or tool failures |
| AH-S101 | high | A shell, process, or filesystem write with no enforceable boundary |
| AH-S201 | medium | A file containing a credential-shaped value, excluded from analysis |
| AH-U101 | low | Dynamic import or lookup that static analysis cannot resolve |
Generation exists for AH-R101 only, and it writes new scaffolding files rather than editing your
code. Every other finding is reported and stops there. Findings are static pattern matches: treat
them as a review queue, not a work order.
See the project README for the full command reference, supported scope, and safety position.
License
MIT
