Protocol Documentation

Truncated salted fingerprint — never the raw key.

When the key was issued, in unix seconds UTC.

When the key should stop working. Zero = no expiry.

RFC-3339 string with nanosecond resolution. String form keeps the wire format readable in text logs and survives clock-skew debates between services.

Salted hash of the input query — never the plaintext.

Response length in characters, for volume metrics. Zero on halt.

Structured policy violation codes.

Wall time the decision took end-to-end at the gateway.

Model name actually routed to (may differ from requested model when the gateway falls back). Empty on pure heuristic scoring.

Safety hook decisions attached to this audit row.

Delta content for streaming chunks. Assistants should populate `message.content` for non-stream and `delta_content` for stream.

Sampling temperature. Zero signals "unset"; gateways forward the upstream default rather than zero-inject.

Tenant binding. Empty = use the tenant bound to the API key.

Opaque per-request ID for correlation. Gateways must not mutate.

Director-AI augmentation. Present on both stream chunks and final non-stream responses.

OpenAI allows a stable name per message, used for multi-user conversations and tool responses. Empty = not set.

Scores below `hard_limit` halt the stream. Carried on the wire so a thin client can log the operating point without querying the server for its configuration.

Confidence intervals for `score`, if computed. Zero-zero means "not available" — consumers must not infer tight certainty from an absent CI.

Evidence trail for the verdict. Unused in heuristic mode.

Free-form human string for logging and UI hover. Machine consumers should branch on `halt_reason`, not parse this.

Per-call override of the server default. Zero = use server.

Server-measured wall time, independent of any client-side timer.

Accumulated assistant output to this point (not a delta). Mirrors the `accumulated = "".join(output_buffer) + token` shape on the Python side.

Candidate token about to be emitted. Empty on the final "flush" request that closes the stream.

When the tenant was created, in unix seconds UTC.

Per-tenant rate limits. Zero = inherit gateway default.

Fingerprint identifier of the API key bound to this tenant, as emitted by audit_salt.get_audit_salt + truncated SHA-256. Raw keys never cross this wire.