agent2playwright

Convert browser-agent trajectories into certified deterministic pytest-Playwright tests.

agent2playwright (CLI: a2pw) compiles noisy browser-agent runs — browser-use, Playwright MCP, workflow-use style histories — into CI-ready pytest-Playwright suites using multi-run consensus, Playwright locator policy, secret sanitization, and a certification scorecard. No LLM is required at replay time.

browser-use / MCP / OTIR runs
            │
            ▼
     adapt → OTIR (open IR)
            │
            ▼
 clean → consensus → harden → sanitize → assert
            │
            ▼
  pytest-playwright + POM + SCORECARD.md

Install

pip install agent2playwright
# or from source:
pip install -e ".[dev]"
python -m playwright install chromium

Quickstart

# Compile 3 agent runs of the same task into one certified test
a2pw compile \
  examples/trajectories/login_run_a.json \
  examples/trajectories/login_run_b.json \
  examples/trajectories/login_run_c.json \
  -o a2pw-out \
  --test-name test_login_dashboard

from agent2playwright import compile_trajectories
from agent2playwright.adapters import detect_and_load
from agent2playwright.models import CompileOptions

runs = [detect_and_load(p) for p in [
    "examples/trajectories/login_run_a.json",
    "examples/trajectories/login_run_b.json",
    "examples/trajectories/login_run_c.json",
]]
result = compile_trajectories(runs, CompileOptions(test_name="test_login_dashboard"))
print(result.scorecard.summary)
Path("test_login_dashboard.py").write_text(result.test_code)

Why agent2playwright?

Live verification

python benchmarks/live_multisite.py

How do I convert browser-use agent history to Playwright pytest?

1. Save agent history JSON from browser-use (or export OTIR).
2. Run the same task 2–3 times for consensus.
3. Compile:

a2pw compile run_a.json run_b.json run_c.json -o tests_out --test-name test_flow
export A2PW_PASSWORD='...'
pytest tests_out/tests --browser chromium

How is agent2playwright different from browse-to-test / recast / flow2skill?

Use agent2playwright when you already have browser-agent runs and need CI-stable pytest with a fail/pass certification gate.

Can I compile multiple agent trajectories into one stable test?

Yes. Pass multiple JSON files to a2pw compile. Steps must agree across enough runs (default 67%) to survive consensus. Retries, wrong clicks, and one-off detours are dropped.

OTIR

Open Trajectory Intermediate Representation — driver-agnostic JSON (src/agent2playwright/schema/otir.schema.json).

a2pw adapt raw-browser-use.json -o otir.json
a2pw certify otir.json

CLI

a2pw adapt <trajectory.json> -o otir.json
a2pw compile <run1.json> [run2.json ...] -o out/ --require-cert
a2pw certify <trajectory.json>

Exit codes: 0 certified, 1 compiled but not certified, 2 hard failure with --require-cert.

License

Apache-2.0