CI & teams

The CI pattern

Check existing behaviors strictly, but auto-record brand new ones so a freshly added test doesn't fail the build before anyone has reviewed its baseline. That's exactly what --behaviorci-record-missing does.

# .github/workflows/behavior.yml
name: Behavior
on: [push, pull_request]

jobs:
  behavior:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"

      - name: Cache the embedding model
        uses: actions/cache@v4
        with:
          path: ~/.cache/huggingface
          key: ${{ runner.os }}-behaviorci-model

      - run: pip install "behaviorci[local] @ git+https://github.com/0-uddeshya-0/BehaviorCI.git" pytest

      - run: pytest --behaviorci-record-missing --behaviorci-report bci.json

      - uses: actions/upload-artifact@v4
        with:
          name: behaviorci-report
          path: bci.json

Caching ~/.cache/huggingface keeps the one-time model download (~80 MB) out of every run.

Reading the result in automation

The --behaviorci-report JSON is the integration point. A follow-up step can turn it into a PR comment, a status check, or a metrics push:

python - <<'PY'
import json
r = json.load(open("bci.json"))
s = r["summary"]
print(f"BehaviorCI: {s['passed']}/{s['total']} passed, {s['failed']} regressions")
for item in r["results"]:
    if not item["passed"]:
        print(f"  ✗ {item['behavior_id']} — similarity {item.get('similarity')}")
PY

Team workflows for the binary database

Baselines live in a single SQLite file (.behaviorci/behaviorci.db). Commit it to version your behavior alongside your code. Because it's binary, Git can't merge two copies, so concurrent baseline edits on different branches can conflict.

A workflow that avoids that:

• CI stays read-only. pytest --behaviorci on pull requests is always safe and never writes.
• Record on a known branch. Run --behaviorci-record / --behaviorci-update on main, or let a single maintainer own baseline updates.
• Ignore the sidecars. WAL files (*.db-wal, *.db-shm) are throwaway and already excluded by .gitignore.

Parallel runs

The database uses WAL mode with per-thread connections, so pytest -n auto (pytest-xdist) is supported out of the box — no database is locked errors.

A Git-friendly backend is on the roadmap

A JSON snapshot backend that diffs and merges cleanly in Git is planned; until then, the single-writer workflow above keeps things simple.