newb 0.3.2

Test your package through the eyes of a newbie agent — a fresh AI agent reads only your docs/skills and tries to use your package.

newb

Test your package through the eyes of a newbie agent.

A fresh AI agent reads only your _skills/ (or equivalent docs) and tries to use your package. If it succeeds — your docs work. If it fails — your CI tells you why.

Install

pip install newb

Use

newb run ./src/mypkg/_skills/mypkg
newb run ./_skills --format markdown >> README.md

newb run spins up a clean docker container with only your skills mounted (no host ~/.claude leak), then asks a fresh Claude agent four canonical questions:

1. Identity — "What is this package for?" / "What problems does it solve?"
2. Usage — "Show a working example" / "When should I NOT use this?"
3. Boundary — author-supplied red tests in _red_tests.yaml ("Can this do ?" → must redirect, not hallucinate)

Output: JSON (for CI) or markdown (for README injection).

Requirements

• Docker on PATH (for the agent sandbox).
• ANTHROPIC_API_KEY in env (used inside the container).
• Python 3.10+.

Library API

import newb

# Module-callable shortcut — for quick scripts:
report = newb("./src/mypkg/_skills/mypkg")

# Explicit form — identical behaviour, clearer intent:
report = newb.run("./src/mypkg/_skills/mypkg")

# Render the report as a README-ready markdown block:
print(newb.render_markdown(report))

newb.verify and newb.self_explain are backward-compat aliases for newb.run (kept through one minor release; removed in 1.0).

The verb mirrors pytest.main() — neutral, importable, no implication that the agent's success "proves" anything beyond what the asserts say.

Why no aggregate "score"?

Principle #1: No verification without specification. If you want a score, you must define what counts as correct.

Today newb returns the agent's actual answers (text) and per-test boundary results (boolean from _red_tests.yaml). It does not emit an aggregate "0.85" score because:

• A single number invites gaming (people optimise for the score, not for actually-better docs).
• Different tasks (description, usage, boundary) measure different things; averaging them is dishonest.
• Without an explicit expected answer per question, the "score" is whatever the LLM judge feels like that minute — non-reproducible.

A scoring system based on author-provided expected answers (pytest- style discovery: tests_newb.py with def test_X(agent): assert ...) is planned for v0.3.0. Until then newb deliberately gives you the raw evidence and lets you decide what "good enough" means for your package.

Aliases

Also available as pip install newbie-test and pip install agentic-test (same package, defensive name reservations that depend on newb).

Heritage

newb was extracted from scitex-dev where the canonical integration still lives:

scitex-dev skills self-explain <package-name>

License

AGPL-3.0-only. Same as the SciTeX ecosystem from which newb was extracted.