Observability Logs And Traces

Agent invokes Portkey 5 times in a single user turn (planner → tool call → summarizer); operator wants all 5 calls grouped in one trace.

Set x-portkey-trace-id to the same operator-generated id on each call, or read x-portkey-trace-id from the first response and propagate. Logs and dashboards group by trace_id for end-to-end agent latency / cost attribution. Do not let each call get an independent trace.

Dashboard shows overall p95 latency at 3.2s but operator can't tell which provider is slow.

Break down latency by x-portkey-provider (the actual upstream that served the request) so fallback-tail latency is attributable. Confirm the trace log shows per-attempt upstream timings. Do not assume gateway latency equals upstream latency — guards and cache lookups add overhead.

Operator wants cost-per-tenant in the Portkey dashboard.

Send x-portkey-metadata as a JSON-string header with reserved key _user (or operator-defined keys like tenant_id). Dashboards filter and aggregate by metadata fields. Treat metadata as PII-bearing and keep values to opaque ids — never put email/phone in metadata.