Traces & payloads

Every request produces two artefacts: a small trace record (always) and, when payload storage is enabled, a full payload archive with the original caller request, upstream request, upstream response, and the final caller response.

Trace record

A compact metadata record, persisted for every request and returned inline when the caller sends Routeur-Trace: true.

trace record

{
  "request_id":        "01K...",
  "organization_id":   "org_42",
  "api_key_id":        "key_01J...",
  "provider":          "openai",
  "model":             "gpt-4o-mini",
  "requested_model":   "auto",
  "route_reason":      "default",
  "latency_ms":        2939,
  "prompt_tokens":     13,
  "completion_tokens": 7,
  "total_tokens":      20,
  "cost_usd":          0.00000615,
  "status_code":       200,
  "redactions":        1,
  "dlp_applied":       true,
  "streamed":          false,
  "gcs_path":          "gs://<project>-routeur-payloads/01K..."
}

Trace fields

request_idstring

Stable id used across response metadata, trace records, and payload archives.

route_reasonstring

default or rule:<slug> depending on how routing resolved.

redactionsinteger

Number of input substitutions DLP made before the upstream call.

dlp_applied / shield_applied / moderation_appliedboolean

Whether each safety layer ran on this request, regardless of outcome. These drive per-feature metering.

scoresobject

The auto-router's 0–1 capability and safety axes from the single prompt-inspection pass.

streamedboolean

True when the response was delivered as SSE.

ttfb_msnumber

Milliseconds from request start until the first response byte left the gateway. The headline streaming metric.

stream_abortedboolean

True when a stream was cut after delivery began — a mid-stream upstream failure or a client disconnect.

disclosed_bytesinteger

Bytes of model output already delivered when chunked moderation aborted a stream. The audit trail for that trade-off.

gcs_pathstringoptional

Location of the full payload archive, when payload capture is enabled for your organisation.

cost_usdnumber

Calculated estimate from the pricing table for the selected provider/model.

The per-stage timings (auth_ms, dlp_in_ms, input_mod_ms, routing_ms, upstream_ms, moderation_ms) are also recorded, alongside client_ip and user_agent.

Payload archive

Payload capture is off by default and enabled per organisation. When enabled, routeur.ai writes the full 4-layer payload to archive storage and records its location on the trace as gcs_path. The archive bundles every view of the request:

caller_request — raw request body from the application.
upstream_request — what routeur.ai actually sent (post-redaction, post-rule).
upstream_response — raw response from the provider.
caller_response — the response body actually returned to the caller.

Audit-friendly. The four layers together reconstruct any single request — useful for security reviews of what actually crossed the gateway boundary.

← Output moderation

Presets reference →