product: assert audience: test-developer authority: normative

Python Dependencies and Offline Environments

Production test stations are frequently deployed in air-gapped networks with no internet access. The Python runner Docker image ships with only the gRPC transport libraries — nothing else. Third-party packages MUST be made available before a test executes.

How pipPackages works

Declare dependencies in package.json:

{
  "requirements": {
    "minPythonVersion": "3.11",
    "pipPackages": ["requests>=2.32.0"]
  }
}

The Python runner automatically installs these packages at startup by scanning every installed package and reading pipPackages from each package.json.

These are tried in order — they are a fallback chain, not alternatives:

Vendoring wheels for production

Step 1 — Download wheels on a machine with internet access:

# Helper script (reads pipPackages from package.json)
.\scripts\vendor-wheels.ps1 -PackagePath "path\to\my-package"

# Or manually
pip download requests>=2.32.0 \
  --dest wheels/ \
  --platform manylinux2014_x86_64 \
  --python-version 3.13 \
  --only-binary=:all:

The Python runner runs on python:3.13-slim (Debian Linux, x86_64). Always download wheels for manylinux2014_x86_64 and Python 3.13.

This matters for compiled wheels. pip silently ignores a wheel whose ABI tag (cp311, cp313, …) does not match the interpreter, then reports ERROR: Could not find a version that satisfies the requirement <pkg> (from versions: none) — which reads like the wheel was never vendored. Pure-Python wheels (py3-none-any) are unaffected, so a mismatch typically singles out one compiled package (numpy, matplotlib, pillow, cffi …). The runner detects this case and says so explicitly in the step error; re-vendor with the matching --python-version. Reinstalling the package will NOT fix it.

runner_type: python3.11 in test YAML is an unrelated routing token — it selects the Python runner, it does not pin an interpreter version.

Step 2 — Commit wheels into wheels/ in your package. The runner installs them automatically at startup — no manual docker exec required.

How the runner finds Python modules

The runner adds to sys.path at startup:

1. /app/scripts — legacy mount for ad-hoc scripts.
2. <package>/python_modules/ for every installed package — picked up automatically on every step invocation.

Development vs production

• Development: Set the station config key AllowOnlinePip = true to install from PyPI. This is a station config key, not an environment variable — the runner reads it when the first step executes, so a station can be switched online without a restart.
• Production: MUST vendor wheels. MUST NOT rely on internet access.