Observability
Request Detail
Every field of a single request, its errors, and its stored payloads.
Request detail is the full record of one request, opened by selecting its request id in Traffic. The header shows the request id, status, model and provider, and a row of tiles gives the Latency, Cost, Tokens, time to first token, Attempts and Router overhead. Sections below give every recorded field, and when the router stored payloads, Full Request Payload opens the request and response themselves.
Request
The Request section identifies the request and its caller:
| Field | Shows |
|---|---|
| Request ID | The request id, such as req_7Kd2Qx9WmB4Tn6Ls, also returned in x-mr-request-id. |
| Timestamp | When the request arrived. |
| Organization | The organization id. |
| Token | The id of the API key that made the request, such as tok_4Gm8Rq2VxN. |
| Subject type | The kind of Machine Identity workload, when one made the request. |
| Machine issuer | The identity provider of the workload. |
| Machine subject | The workload identity. |
| Machine binding | The machine binding the workload matched. |
| Correlation ID | The caller's x-mr-correlation-id. |
| Customer | The caller's customer id. |
| Feature | The caller's feature label from X-MR-Feature-Id, such as support/triage. |
| Session | The caller's session id. |
| Peer IP | The address the request came from. |
Routing
The Routing section shows how the request was served:
| Field | Shows |
|---|---|
| Alias | The alias the request used, linked to the router. |
| Router ID | The router that served the request, linked to its page. For a fallback or weighted router, this is the member that served it. |
| Model requested | The model named in the request. |
| Model served | The catalog model that answered, such as anthropic/claude-sonnet-4.5. |
| Vendor | The provider that served the request. |
| Variant | The offering variant, such as standard or geo. |
| Region | The region that served the request. |
| Wire id | The provider's own model id, such as claude-sonnet-4-5-20250929. |
Prompt binding
The Prompt binding section shows the bound prompt the router applied, with dashes when it applied none. It shows the Prompt ID, File path and Revision used; for a GitHub prompt it also shows the Repository, and the revision is the commit, linked to the file at that commit on GitHub. See Prompt Binding.
Latency
The Latency section splits the request's total time into its parts, and the Provider response chart shows the same parts as a timeline:
| Field | Measures |
|---|---|
| Total | From the request arriving to the response finishing. |
| TTFT | Time to first token: from the request arriving to the first streamed output. |
| Request upload | Receiving the request body from the caller. |
| Router overhead | Multi-Router preparing the request before sending it to the provider. |
| Upstream | The provider handling the request. |
| Router response | Multi-Router finishing the response after the provider. |
Tokens
The Tokens section shows Prompt tokens, with how many were read from the provider's prompt cache and how many were written to it, by cache lifetime, and Completion tokens. It also shows the request and response sizes in bytes.
Cost
The Cost section shows the Total and its parts: Input, Output, Cache read and Cache write, by cache lifetime, and Web search and File search when the request used those tools. See Cost.
Errors
A failed request shows an error card with its category, a readable description and hint, the provider's status code when the provider answered, the error kind and the error message. The categories are Request error, Config error, Upstream error and Other error. Requests refused by Multi-Router's per-key rate limit or by a budget are answered before a record is written, so they appear in the caller's logs and not in Traffic. The error kinds recorded are:
| Error kind | Label | Meaning |
|---|---|---|
invalid_request | Invalid request | The request was malformed, or Multi-Router refused it before sending, such as for a model the router does not serve. |
auth_error | Authentication error | The provider refused the credential. |
model_not_found | Model not found | The provider reported that the requested model is not available. |
config_error | Configuration error | The router is set up in a way that cannot serve the request, such as a missing provider key. |
rate_limit_error | Rate limited | The provider's rate limit was reached. |
timeout | Timeout | The provider took too long to respond. |
upstream_error | Upstream error | The provider returned an error. |
cancelled | Cancelled | The request was cancelled before it finished. |
client_disconnect | client_disconnect | The caller closed the connection before the response finished. |
metering_failed | metering_failed | The response finished, and its token usage could not be read. |
internal_error | internal_error | Multi-Router hit an internal error while serving the request. |
Payloads
Payloads are the full request and response, stored when the router's Request Logging settings capture them. Full Request Payload in the detail header opens them; a newly finished request shows Full Request Payload available in a few minutes, with a countdown, while its payload is being stored. The payload viewer shows:
| Part | Contents |
|---|---|
| Client request | The request exactly as the caller sent it. |
| Provider Request | The request as Multi-Router sent it to the provider, with any bound prompt applied. Shown when it differs from the client request. |
| Response | The response returned to the caller. |
Each part opens as a tree or as raw text, Split view shows two parts side by side, and Headers shows the headers of each part, with credentials such as Authorization replaced by [redacted]. Copy copies a body, and a badge marks a body that was decompressed or truncated. See Request Logging.