@taskproof/adapter-browser-use

The second runner adapter — and the proof of the moat: it drives browser-use (a Python agent) through a thin FastAPI sidecar, yet emits the identical RunArtifact and grades through the same @taskproof/grader as the Claude adapter. The only difference between the two is which agent drove the browser.

How it stays uniform: the sidecar runs browser-use and reports raw run data plus probe results (final URL, network log, and a DOM-probe result per spec selector — evaluated via CDP since browser-use 0.13 dropped Playwright). The TS adapter maps that into the shared artifact and grades it. Grading logic lives in one place; the sidecar never sees the spec or artifact schema.

Running it

The sidecar is a uv project under sidecar/. One-time setup:

cd sidecar
uv sync                      # installs browser-use 0.13.1 + FastAPI (pinned)
uv run browser-use install   # fetches browser-use's Chromium (separate from pip)
export ANTHROPIC_API_KEY=…    # browser-use drives Claude as its LLM
uv run uvicorn taskproof_sidecar.app:app --port 8765

Then point the adapter at it (default http://127.0.0.1:8765, override with TASKPROOF_BROWSER_USE_URL or sidecarUrl) and run:

taskproof run tasks/*.yaml --models claude-opus-4-8,browser-use

browser-use (or browser-use:claude-sonnet-4-6) selects this adapter; the Claude model named is the LLM browser-use uses. Comparing claude-opus-4-8 vs browser-use on the same tasks is the headline comparison: same model, two different agent harnesses.

Pre-release. Live-validated 2026-06-13 against browser-use 0.13.1: a real taskproof run --models browser-use reached the target, the CDP DOM probe and the url assertion both graded correctly through the shared grader, and token cost was extracted (~$0.24 for a 3-step task). Verified API details now in runner.py: omit temperature (Opus 4.8 rejects it), read the version via importlib.metadata, get usage from agent.token_cost_service.get_usage_summary(), and set keep_alive=True so the session survives agent.run() for the post-run DOM probe.

Network capture (HAR-based): uses browser-use's HarRecordingWatchdog (record_har_path), validated live — it captures same-origin HTTPS traffic (the site's own API calls, the common network-assertion case). Remaining edge: the watchdog only enables CDP Network on the initial session, so a cross-origin navigation to a new target is missed; full coverage would need per-target Network.enable via Target.attachedToTarget. HTTP (non-TLS) isn't captured.

Budget cap (maxCostUsd / --max-cost) is NOT enforced mid-run here. Unlike the Claude adapter, which gates each turn against the cap before paying, taskproof can't stop a browser-use run partway — it runs to its own completion or maxSteps. The cap is therefore advisory for this adapter: cost may exceed it (the report shows the real figure), and the CLI warns when you pass --max-cost with a browser-use model. maxSteps is the hard bound on browser-use spend — lower it to bound cost.