Benchmark report
Task verified
4m 43s elapsed
The saved attempt met the documented success criterion.
96/ 100
Agent-Ready- SInstall
- SAuth
- SExecute
- ADocs
- SCurrent
Three of four guessed Kernel api-reference doc pages returned 404, forcing a detour to the 803KB OpenAPI spec.json to find the browser operations.
Where the time went
4:43 agent time- Research0:27
- Attempt2:50
- Verify1:26
Judge reason
Saved code (code/kernel_session.py, pure urllib — no browser automation lib) and seven saved response files show a real lifecycle against api.onkernel.com with Authorization: Bearer only. POST /browsers with explicit {"timeout_seconds":300} returned HTTP 200 and session_id vvtsc8gwbz1iv6n985eumf4b with timeout_seconds echoed 300 (live headers X-Request-Id n8mdj7m27hvga8sdvhjpittu, X-Ratelimit-Remaining 249). GET /browsers/{id} returned 200 with no deleted_at; GET /browsers?status=active returned a list containing exactly that session_id; DELETE /browsers/{id}, executed inside a finally block, returned the documented 204 with empty body; the re-fetch returned 404 {"code":"not_found"} and the include_deleted=true fetch returned 200 with deleted_at=2026-11-03T05:28:29.035181Z, with the post-delete active list empty. Kernel's OpenAPI Browser schema has no literal status field (transcript jq for .Browser.properties.status returned null), so active/ended were derived from the documented status filter and deleted_at description rather than invented. Console log values match the saved JSON exactly; grep for the key and for bearer/JWT patterns across saved files found nothing (signed-URL jwts redacted).
Attempt record
- 1:18Answered from saved key: Kernel API keyno wait
- —First success verified
Attempt artifacts saved.