Run an endpoint

Admits one execution: the policy is checked, a hold is reserved for the price's maximum charge, and a run is created. Input is provider-native — construct it against the inspected input_schema. A sync endpoint always answers 200 with the terminal run inside its published timeout_ms — it never returns 202, so no polling code is needed. An async endpoint answers 200 when the run is terminal within a short server budget, otherwise 202 with the in-flight run, which you can poll or stop. Status reports whether a usable provider response won settlement, never whether the provider accepted the request: a provider 404 or 500 is a COMPLETED run whose provider_response carries the received status, charged by the price's NO_RESULT or PROVIDER_ERROR clause. The Idempotency-Key header is required — a retry with the same key returns the same run and is never charged twice, while the same key with a different request body returns 409. Balance and policy refusals happen before admission and create no run.

curl -X POST https://api.glasser.ai/v1/runs \
  -H "Authorization: Bearer gl_..." \
  -H "Idempotency-Key: 9f4c1c2e-your-uuid" \
  -H "Content-Type: application/json" \
  -d '{"provider":"...","endpoint":"..."}'

Authorization: Bearer gl_...