Benchmark report
Task verified
4m 8s elapsed
The saved attempt met the documented success criterion.
96/ 100
Agent-Ready- SInstall
- SAuth
- SExecute
- ADocs
- SCurrent
WebFetch truncated the OpenAPI page so GET /usage was not visible, forcing an extra curl of the raw 162KB spec to confirm the endpoint.
Where the time went
4:09 agent time- Research0:23
- Attempt2:27
- Verify1:18
Judge reason
Saved code (code/reconcile_credits.py) runs GET /api/v1/usage -> POST /api/v1/enrich_company {"url":"https://www.linkedin.com/company/stripe"} -> GET /api/v1/usage against https://api.richapi.ai/api/v1 with the x-api-key header; transcript tool output and evidence/*.status.txt show HTTP 200 for all three, and exactly one enrichment call. Saved responses are self-consistent and live-looking: before total_credits 133.69 / credits_remaining 4966.31 with enrich_company 19 calls/19.0 credits, after 134.69 / 4965.31 with 20 calls/20.0 credits, i.e. a 1.0-credit delta; enrich response is a real Stripe LinkedIn record (followers 1813532, objectUrn 2135371). The documented price is confirmed from the agent's own saved openapi.json: POST /enrich_company description contains '**Pricing:** 1 credit per call', and GET /usage is documented as 'Return the supplied API key's team-wide credit consumption and current balance.' Report states agreement (1.0 observed = 1 documented, no delay). The LinkedIn-URL substitution is correct: the spec's required body field is a 'LinkedIn company page URL'. Credential came from benchmark-request as a file path; a grep for rk_live/rk_test/bearer/access_token patterns across /workspace/home and the transcript found no key value, and the script redacts before writing. No mocking or hard-coding.
Attempt record
- 1:13Answered from saved key: RichAPI API key (x-api-key, rk_live_...)no wait
- —First success verified
Attempt artifacts saved.