Rewrite ai-gateway-provider: new API, backward-compatible deprecations, comprehensive tests

.changeset/ai-gateway-provider-rewrite.md

@@ -0,0 +1,32 @@
+---
+"ai-gateway-provider": minor
+---
+
+Rewrote the AI Gateway provider with a new API. `createAIGateway` now wraps any standard AI SDK provider — import directly from `@ai-sdk/*` instead of `ai-gateway-provider/providers/*`.
+
+**New features:**
+
+- `createAIGateway({ provider, binding/accountId })` — wraps any AI SDK provider
+- `createAIGatewayFallback({ models, binding/accountId })` — cross-provider fallback
+- `providerName` — explicit provider name for custom base URLs
+- `byok: true` — strips provider auth headers for BYOK/Unified Billing
+- `byokAlias` and `zdr` gateway options
+- 22 providers auto-detected from URL (up from 13)
+- `resolveProvider` exported as a utility
+
+**Bug fixes:**
+
+- Updated stale cache header names (`cf-skip-cache` → `cf-aig-skip-cache`, `cf-cache-ttl` → `cf-aig-cache-ttl`)
+- Fixed BYOK leaking dummy API keys (`Authorization: Bearer unused` was forwarded to the provider)
+- Fixed fetch mutation not being restored after gateway calls
+- Removed unanchored Google Vertex regex that could match unintended URLs
+- Fixed query parameter loss for custom URLs with `providerName`
+- Propagate abort signals to binding and fetch calls
+
+**Deprecated (all still work, with a one-time console warning):**
+
+- `createAiGateway()` — use `createAIGateway()` (wraps a provider) or `createAIGatewayFallback()` (cross-provider fallback) instead
+- `ai-gateway-provider/providers/*` subpath imports — import directly from `@ai-sdk/*` and use `createAIGateway()` instead
+- Old type names: `AiGatewaySettings` → `AiGatewayConfig`, `AiGatewayAPISettings` → `AiGatewayAPIConfig`, `AiGatewayBindingSettings` → `AiGatewayBindingConfig`, `AiGatewayReties` → `AiGatewayRetries`
+
+All deprecated APIs will be removed in the next major version.

.changeset/workers-ai-gateway-tests.md

@@ -0,0 +1,5 @@
+---
+"workers-ai-provider": patch
+---
+
+Added AI Gateway e2e tests for the binding path. The test worker now accepts a `?gateway=<name>` query parameter to route requests through a specified AI Gateway, validating that the built-in `gateway` option works for chat, streaming, multi-turn, tool calling, structured output, embeddings, and image generation.

.oxlintrc.json

@@ -19,7 +19,8 @@
 		"no-await-in-loop": "off",
 		"react/react-in-jsx-scope": "off",
 		"@typescript-eslint/no-non-null-assertion": "off",
-		"@typescript-eslint/no-explicit-any": "off"
+		"@typescript-eslint/no-explicit-any": "off",
+		"eslint(no-shadow)": "off"
 	},
 	"ignorePatterns": ["**/worker-configuration.d.ts", "types/workerd.d.ts"]
 }

package.json

@@ -62,8 +62,6 @@
 			"sharp",
 			"workerd"
 		],
-		"patchedDependencies": {
-			"@ai-sdk/google-vertex": "patches/@ai-sdk__google-vertex.patch"
-		}
+		"patchedDependencies": {}
 	}
 }