Skip to main content
Proxy-generated errors use this shape:
Every response includes x-proxy-request-id. Include it when reporting a problem.

Common failures

The most common project mistake

OpenAI-Project takes the project’s Weights & Biases project name, not its project alias. The two differ when the project was created with a project name other than the alias. In that case, the alias in OpenAI-Project returns 404 with No routing configuration for project. Copy the value from the project header in Studio:

A model ID error is not a proxy 404

The proxy validates the provider name, not every upstream model ID. If the provider exists but the model does not, the selected provider returns its own error. This distinction helps separate routing configuration from provider catalog drift.

Tracing failures

Trace writes are asynchronous and do not turn a successful provider response into a client failure. Operators monitor structured write-failure events with the same proxy request ID. Deterministic trace-input problems are rejected before provider dispatch when possible.
Provide the agent with the request URL, the OpenAI-Project header value, a redacted request body, the HTTP status, the full response body, and x-proxy-request-id. Do not share the bearer key or provider credential.An agent should check the following in order:
  1. The OpenAI-Project value, and whether its project is the Weights & Biases project name rather than the project alias.
  2. Whether model is router or a model override, and which version forge.project_version selects.
  3. Whether the provider and routing deployment are applied.
  4. Whether the response body came from the proxy or the upstream provider.
For error response schemas, see the Chat Completions API reference.
Last modified on November 3, 2026