Skip to main content
Status: The partner-preview gateway can return standard HTTP protection and capacity failures. A dedicated, documented praxa_sk_ per-key rate-tier product is not yet available, so no per-key quota numbers are promised here.
Do not infer a production limit from Praxa’s unrelated in-app or custom-agent Cloudflare bindings. Those surfaces have their own counters and fail posture. The public execution gateway must document a stable tier before clients can depend on exact requests-per-minute values.

Handling responses

Problem bodies are best-effort RFC 9457-style application/problem+json. Proxies can fail without a valid body, so fall back to the HTTP status and conservative backoff.

Idempotent retry guidance

  • Send one Idempotency-Key per logical POST /v1/execute request.
  • Reuse that key only with the identical normalized request.
  • If the connection drops after submission, prefer the known run URL or retry the identical request with the same key; do not mint a new key blindly.
  • Resume SSE with the last durably processed numeric Last-Event-ID.
  • A cancellation request is cooperative and is not rollback.
  • During uncertain-outcome reconciliation, obey this exact message: The run is reconciling an uncertain execution outcome. Do not retry or resubmit it.
The cancellation-reconciliation migration and compatible Workers are deployed, and its final states have passed separate production canaries. If a terminal run_outcome_unknown appears, do not retry automatically; reconcile the external system first.

See also

Last modified on August 14, 2026