Documentation

Production readiness

A production Northsun (formerly Mnemix) integration is ready when authentication is server-side, the recall/write-back lifecycle is observable, failure modes preserve the agent experience, and every public claim matches behavior you have actually observed.

Before launch

  • Store MNEMIX_API_KEY in a managed secret store and confirm it never reaches browser or mobile code. The variable name is the one the shipped clients read; it is a runtime identifier, not a brand.
  • Pin the npm package version you tested; do not assume the repository version is the published version.
  • Use E.164 phone numbers and URL-encode path parameters.
  • Set an end-to-end request deadline before the first network call.
  • Treat missing enrichment as a partial response, not a failed conversation.
  • Select context deliberately: recall_and_enrich returns stored caller memory in recency order — most recently updated first — not ranked by relevance to the current call. Your application decides what reaches the prompt.
  • Handle 400, 401, 404, 422, 429, timeouts, and temporary 5xx responses separately.
  • Add capped backoff with jitter only for retryable failures.
  • Deduplicate voice-platform webhook retries by the platform call identifier.
  • Confirm calls/end is called once after hangup, treat 204 as submission acknowledgement, and prove durable write-back with a later recall.
  • Record endpoint, status, trace ID, and safe correlation metadata without logging secrets or PII.

Prove the lifecycle

Use a dedicated test tenant and number:

  1. Call recall_and_enrich and retain its trace ID.
  2. Call calls/end with a short synthetic transcript and the same session_id.
  3. Confirm the 204 response — no wait is required before the next read.
  4. Call recall_and_enrich again and verify the expected history is present.
  5. Call GET /v1/caller/{phone_number} and verify the read-only profile.
  6. Exercise one validation failure and one rate-limit/retry path in a non-production environment.

This proves your mapping and lifecycle. It does not prove a universal latency benchmark, customer outcome, compliance posture, or every dependency state.

Voice fail-open behavior

Design the surrounding agent so unavailable optional context does not prevent a call from continuing. A safe fallback greeting should not reveal whether enrichment failed, and it should never guess a person's identity.

Change management

  • Follow the changelog for contract and client changes.
  • Pin package versions and update deliberately.
  • Re-run the lifecycle proof after changing voice-platform webhooks, key configuration, phone normalization, or client versions.
  • Keep a kill switch for any autonomous workflow that sends, posts, spends, or performs an irreversible action.

Next steps