Arbiter

Agent attribution for code quality.
Install arbiter-score, run arbiter score, and quickly see which AI agent or human produced your highest-risk code.

pip install "arbiter-score[analyzers]"
arbiter score .

Example output:

┌──────────────────────┬──────────┬────────┐
│ Agent                │ Quality  │ Grade  │
├──────────────────────┼──────────┼────────┤
│ claude               │ 72.3%    │ C      │
│ gpt-4o               │ 88.1%    │ B      │
│ copilot              │ 91.4%    │ A      │
│ human (you)          │ 94.2%    │ A      │
└──────────────────────┴──────────┴────────┘

Every team using AI coding tools has a "who wrote this mess?" moment. Arbiter makes that visible -- per file, per commit, and per author -- then grades the code quality signal.

pip install "arbiter-score[analyzers]"    # one-time install
arbiter score .                         # grade everything
arbiter diff . --base main              # only changed files in this branch
arbiter agents                          # author leaderboard
arbiter trend --days 30                 # quality over time
arbiter worst --limit 20                # highest-risk files
arbiter audit-fleet ./repos             # score every repo in a directory
arbiter serve                           # start web dashboard

Why Arbiter exists

AI-written code often optimizes for speed over maintainability. The result can include unused imports, high complexity, duplicated logic, and risky patterns. Standard linters catch some symptoms; Arbiter also attributes risk back to the contributor model/actor so you can review the right places first.

In practice, teams tune gates per agent tier (for example: stricter thresholds for autonomous agents, looser for exploratory branches) and keep human review focused on high-risk commits.

What makes Arbiter different

Feature	Traditional tools	Arbiter
Agent attribution	None	First-class: tracks contribution source (AI and humans)
Per-commit scoring	Repository-wide only	Scores changed files by commit and author
Diff analysis	Optional / limited	Scores only what changed in PR/branch workflows
Review workflows	Static report	Supports agent-specific gate thresholds
Dashboard output	SaaS + build steps	Local, single-file HTML dashboard endpoint
Core runtime	Multi-tool runtime	Lightweight core (LibCST only); optional analyzer dependencies

Scoring model

Analyzer	Weight	Finds
Lint (ruff)	35%	Style violations, import errors, bugbear patterns
Security (bandit)	30%	Hardcoded secrets, shell injection, risky command patterns
Complexity (radon)	35%	Cyclomatic complexity by function (A-F)
Dead code (vulture)	penalty	Unused functions, imports, variables
Duplication (AST hash)	penalty	Near-duplicate function bodies

Arbiter computes a normalized score from analyzer penalties and total lines of code.

Common format:

100 - (penalty / LOC) * normalization

Grades: A (90+), B (80+), C (70+), D (60+), F (<60).

Install

From PyPI (recommended)

pip install arbiter-score                    # core (LibCST runtime)
pip install "arbiter-score[analyzers]"       # + ruff, radon, vulture, bandit

Note: The PyPI package is arbiter-score (not arbiter, which is different). The CLI command is arbiter.

From source

git clone https://github.com/hummbl-dev/arbiter.git && cd arbiter
pip install -e ".[analyzers]"

Tests

pip install -e ".[test]"
python -m pytest tests/ -v

Requirements

• Python 3.11+
• git

Quality gates

Arbiter enforces quality checks in CI.

arbiter score . --fail-under 90 --json
arbiter diff . --base main --fail-under 80

Common gotchas

• Use arbiter-score package name to avoid installing the wrong package.
• Run from repository root for consistent path reporting and results.
• Optional analyzers improve signal quality, but many core checks run without the analyzer extras installed.

License

Apache 2.0 — see LICENSE.

---

Built by HUMMBL.