Operayde Gateway API

Generates a model response for the given chat conversation. Supports streaming (SSE) and non-streaming modes, tool/function calling, and per-tenant model routing. Requests pass through OPA policy evaluation, ingress/egress guardrails, PII redaction, spend-cap checks, and rate limiting before reaching the upstream provider.

Anthropic Messages API surface. The gateway translates the request into its internal chat format, applies the same policy/guardrail/ routing pipeline as /v1/chat/completions, and translates the response back to the Anthropic shape. Used by the Anthropic SDKs and Claude Code (which send the virtual key in x-api-key).

Mirrors Anthropic's messages/count_tokens endpoint — the same body shape as /v1/messages but without max_tokens. For routes that resolve to a native Anthropic provider the request is proxied upstream (their tokenizer is authoritative); for every other provider the gateway's internal estimator is used, consistent with pre-dispatch budget checks and billing.

Creates an embedding vector representing the input text. Routes to the configured provider for the requested model. Subject to OPA policy evaluation and virtual-key scope enforcement (embed:invoke).

Returns the models available to the authenticated caller. Results are scoped to the tenant's configured model routes and further filtered by the virtual key's models_allow list. Requires the models:list scope.

Returns the tenant's knowledge scoping catalog (companies / functions / projects) plus the caller's own memberships, assembled from the config workspace tables. Strictly tenant-scoped — staff and customer-realm principals are rejected with 403.

Opens (or resumes, by idempotent session_key) a bulk ingest session on the tenant's resolved workspace-runtime. The gateway reverse-proxies the request to the appliance or central runtime — the same routing decision the memory enricher makes.

Uploads one batch of pre-embedded vector points into an open session. Batches carry a monotonically increasing batch_seq for resumability; point batches can be a few MB (forwarding timeout ~120s).

Finalises an open session, making its points visible to retrieval.

Returns the status of an ingest session — used to resume an interrupted upload (next_seq) or confirm a commit.

Returns 200 if the gateway process is running.

HEAD variant of the liveness probe for load balancers.

Checks downstream dependencies (Postgres, OPA). Returns 200 when all are healthy, 503 when any are degraded. Component names are omitted from the response to avoid leaking infrastructure topology.

Flushes a virtual key from the gateway's in-memory VK cache, by raw bearer or by key hash. Called by config-service after a key is revoked or updated so the change takes effect immediately.

Re-reads the model routing table from the database. Returns status "noop" when the gateway loaded static routes at boot and has nothing to reload.

Returns the current state of all provider circuit breakers.

Resets a single provider's breaker (body {"provider": "..."}) or all breakers (empty body).

Process-lifetime snapshot assembled only from in-process state: breaker states, cache hit/miss counters, and optimization counters. Request/error totals live in Prometheus, not here.

Returns a sanitized snapshot of the gateway boot configuration — enabled providers and endpoints, live hot-sync flag values. API keys, DB credentials, and other secrets are stripped.

Returns the memory subsystem mode (disabled / local / remote / per_tenant) and per-layer status (working / semantic / episodic).

Per-deployment in-flight counts, latency EWMA, totals, and deployment breaker states for multi-deployment routing.

Runs the wire-format + capability auto-detect flow against an upstream endpoint and returns what the operator should store in config.model_routes. 30s timeout.