Investigate a failed, blocked, or expensive request

Start with a request ID or UTC incident window. Observability › Requests supplies the audit trail; its Request drawer and the full request page share the same recorded request evidence. A View in raw data link opens Signals scoped to that request.

Use the recorded status, route stages, attempts and cost before deciding how to change a key, model, policy or prompt. Missing evidence carries its cause instead of an inferred zero.

Open Observability, select the UTC range, then search by request ID, agent, model or environment, or narrow with Provider type, Model, Object, Requester, Status or API key. Failed requests and Most expensive are one-click shortcuts. The server sorts the time, token, list cost and provider latency headers and pages the result; Total and Displayed describe that result.

Pipeline stations and guardrail verdicts shows recorded stages and their timing. A failed Attempt says failed and sits below the rail. Candidates and recorded decision shows the operative selection, candidate ladder, exclusions and mechanism modes. A final error is displayed separately from an earlier selection.

Quality tiers, affinity, axes and confidence can be unrecorded. Genuine zero measurements remain zero; unknown values state their cause. A selected candidate is not proof that a provider completed the request.

The ledger projects server amounts without recomputing savings. Negative amounts keep their sign. Read Classification and the key's Guardrails profile when routing or security evidence explains the result.

Cost separates recorded list price, provider billable cost, estimate and overhead. Optimization and Cache show their own recorded evidence. Prompt Optimization distinguishes the requested choice from the prompt actually dispatched, and a cache or dedup response can involve no provider send.

On the full request page, the Overview tab ends with the Savings receipt, labelled Estimate: the reference model and its source, reference and actual cost, the routing, response cache and compression savings, known Cloud Score overhead and net savings, then the price card and whether the request was re-priced on reconcile. With no recorded reference model it claims no savings; View cost details opens the Cost & savings tab.

Reference-model savings, same-model prompt avoidance and provider cache-read discounts have different bases. Do not add them without a compatible server total. Cost increase (est.) is a negative saving; it remains visible in error ink. Unrecorded cache savings or token avoidance are unavailable, never fabricated zero.

Server error text is rendered as text. Keep the displayed request ID when reporting a failure; do not share credentials or prompt contents unnecessarily. Troubleshooting describes setup failures.