agent-uniformity

Reference implementation for the Agent Almanac Code Uniformity benchmark. Same code path that produced the published numbers — install it, point it at any of the 48 sampled repos, get matching results.

What this is

A Python package that implements every step of the Agent Almanac Code Uniformity benchmark: function extraction (Python ast + tree-sitter for TS/JS/Go/Rust), AI authorship detection via git blame against Claude-tagged commits, semble-driven similarity scoring, the six per-repo evaluators, and the multi-repo aggregation that produces the headline hypothesis tests.

The locked task list (48 repos with HEAD SHAs from the inaugural run) ships with the package. You can re-run any single repo against the locked SHA and verify the numbers we published, or re-run the whole sample sequentially and compare your aggregates against ours.

Install

pip install agent-uniformity

Requires Python 3.11 or 3.12 (semble's tree-sitter pin doesn't currently work on 3.13). Brings in semble, radon, tree-sitter-language-pack (pinned to <1.8 — newer versions have a broken Python API), and pydantic.

Quick verification — reproduce one repo

agent-uniformity tasks                                     # list the 48 task IDs
agent-uniformity run-one davila7-claude-code-templates --output ./out

This clones the repo at the SHA in tasks/q2-2026.json, runs the analysis pipeline, and writes ./out/davila7-claude-code-templates.json containing every per-function metric (similarity scores, AI ratio, complexity, etc.) and per-repo evaluator scores.

Compare key numbers against agent-uniformity-q2-2026/analysis/summary.csv. Expected variance: ~1% from semble's BM25 non-determinism.

Run the full benchmark

agent-uniformity run-all --output ./out         # ~6-8 hours sequential on a laptop
agent-uniformity aggregate ./out                # writes ./out/analysis/*.csv
agent-uniformity deep ./out                     # writes H4 + H5 CSVs

Use as a library

from pathlib import Path
from agent_uniformity import analyze_repo, runner

# Run one task by ID
tasks = {t.task_id: t for t in runner.load_tasks()}
result = runner.run_one(tasks["dora-rs-dora"], Path("./out"))
print(result.output.function_count, result.output.repo_ai_ratio_observed)

# Or analyze an arbitrary repo you've already cloned
out = analyze_repo(
    repo=Path("./my-repo"),
    repo_slug="myorg/my-repo",
    base_sha="HEAD",
)
print(out.function_count)

What's in this repo

agent-uniformity/
├── methodology.md             frozen v0.1.0 — pre-registered hypotheses + metrics
├── tasks/q2-2026.json         the 48 sampled repos with locked HEAD SHAs
├── agent_uniformity/
│   ├── extract.py             function extraction (Python ast, tree-sitter)
│   ├── blame.py               git blame + AI commit detection
│   ├── analyze.py             per-repo metrics + 6 evaluators
│   ├── aggregate.py           multi-repo summary CSVs + hypothesis tests
│   ├── deep.py                H4 + H5 deep-pass analyses
│   ├── runner.py              sequential runner (clone → checkout → analyze)
│   ├── schema.py              FunctionFact, RepoOutput pydantic models
│   └── cli.py                 click-based CLI
└── tests/

Versioning policy

The package version (currently 0.1.0) tracks the methodology version. A change to:

• the metric definitions
• the evaluator formulas
• the hypothesis-test logic
• the function-extraction filters
• the AI-commit detection signals

…requires a methodology version bump and a corresponding package release. Old methodology versions remain installable from PyPI; old reports cite specific package versions.

Related repos

• saucam/agent-uniformity-q2-2026 — the published report (methodology, sampling, analysis CSVs, hypothesis tests)
• huggingface.co/datasets/saucam/agent-uniformity-q2-2026 — raw per-repo partials (~120 MB) and report.json
• agentalmanac.com — Agent Almanac publication

License

MIT — see LICENSE. Each upstream repository referenced in tasks/q2-2026.json retains its original license.

Citation

Datta, Y. (saucam). (2026). agent-uniformity: reference implementation for the
Agent Almanac Code Uniformity benchmark. https://github.com/saucam/agent-uniformity