Benchmark report
Task verified
4m 30s elapsed
The saved attempt met the documented success criterion.
96/ 100
Agent-Ready- SInstall
- SAuth
- SExecute
- ADocs
- SCurrent
Doc hosts were unreachable through the egress proxy (307/403), costing ~5 curl probes before the agent fell back to the SDK's bundled README.
Where the time went
4:31 agent time- Research0:18
- Attempt2:42
- Verify1:31
Judge reason
Saved code (evidence/hyperbrowser_session.mjs, identical to work/ copy) uses the real @hyperbrowser/sdk 1.4.1 + playwright-core 1.64.1 (package-lock resolves both to registry.npmjs.org). It calls client.sessions.create() (POST /api/session, key only via the SDK's x-api-key header), connects chromium.connectOverCDP(session.wsEndpoint) with an x-api-key header, navigates to example.com, and stops the session inside a finally block via client.sessions.stop(). Live run output (evidence/run_output.txt, transcript line 72-73) prints session id 3269675f-a9f0-41d1-b00c-f2de957ab4f7, status active, redacted wss://dp-us-west-1.hxproxy.io/.../cdp/devtools/browser/... endpoint, 'Page title: Example Domain', 'Stop result: {"success":true}', 'Session status after stop: closed'. These exactly match session_create_response.json (teamId, launchState, wsEndpoint, sessionUrl) and session_stop_response.json (status closed, endTime 1793743403265, creditsUsed 0.0016 — real billing), so nothing is mocked or hard-coded. Key never appears in saved files: it is read from a private file path, tokens/query params are redacted, and a grep -F for the literal key over evidence/ and work/ returned 'no api key in saved files'; my own scan of the saved files found only package-lock integrity hashes.
Attempt record
- 0:58Answered from saved key: HYPERBROWSER_API_KEYno wait
- —First success verified
Attempt artifacts saved.