Per-provider connection settings, merged over the built-in defaults.
Force first-party auth (e.g. an OAuth store) for a provider:
```json
{ "auth": "default" }
```
Point a provider at a custom base URL with auth disabled:
```json
{ "url": "https://gateway.internal/v1", "auth": "none" }
```
Request-target routing keeps the catalog model identity and its capabilities while sending a different provider-specific model or deployment name on the wire:
```json
{ "requestModel": "production-coding-deployment" }
```
Use `url` for providers with resource-specific endpoints, such as Microsoft Foundry. A trailing slash is normalized before `/chat/completions` is appended. A provider with `requestModel` cannot appear more than once in an alloy because the target would be ambiguous.
Give a slow local model more time to produce its first token. A request fails with a retryable timeout once the provider has been silent for `idleTimeoutSecs` (default 300), before responding or between stream events:
```json
{ "idleTimeoutSecs": 900 }
```
Pin a Bedrock application inference profile:
```json
{ "inferenceProfileArn": "arn:aws:bedrock:us-west-2:000000000000:application-inference-profile/abc" }
```