down versions for release

threepointone · threepointone · commit 87d8a2a3814e · 2026-02-11T00:48:10.000Z
diff --git a/demos/agent-scheduler/package.json b/demos/agent-scheduler/package.json
@@ -19,7 +19,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/agent-task-manager-human-in-the-loop/package.json b/demos/agent-task-manager-human-in-the-loop/package.json
@@ -19,7 +19,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/agent-task-manager/package.json b/demos/agent-task-manager/package.json
@@ -19,7 +19,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/evaluator-optimiser/package.json b/demos/evaluator-optimiser/package.json
@@ -19,7 +19,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/image-generation/package.json b/demos/image-generation/package.json
@@ -18,7 +18,7 @@
 	"dependencies": {
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/model-scraper/package.json b/demos/model-scraper/package.json
@@ -23,7 +23,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/orchestrator-workers/package.json b/demos/orchestrator-workers/package.json
@@ -19,7 +19,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/parallelisation/package.json b/demos/parallelisation/package.json
@@ -19,7 +19,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/prompt-chaining/package.json b/demos/prompt-chaining/package.json
@@ -19,7 +19,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/routing/package.json b/demos/routing/package.json
@@ -18,7 +18,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/structured-output-node/package.json b/demos/structured-output-node/package.json
@@ -11,7 +11,7 @@
 	},
 	"dependencies": {
 		"ai": "^6.0.78",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/structured-output/package.json b/demos/structured-output/package.json
@@ -20,7 +20,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/text-generation-stream/package.json b/demos/text-generation-stream/package.json
@@ -19,7 +19,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/text-generation/package.json b/demos/text-generation/package.json
@@ -19,7 +19,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/tool-calling-stream/package.json b/demos/tool-calling-stream/package.json
@@ -19,7 +19,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/tool-calling/package.json b/demos/tool-calling/package.json
@@ -20,7 +20,7 @@
 		"agents": "^0.4.0",
 		"ai": "^6.0.78",
 		"hono": "^4.11.9",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/demos/vision/package.json b/demos/vision/package.json
@@ -18,7 +18,7 @@
 		"hono": "^4.11.9",
 		"react": "^19.2.4",
 		"react-dom": "^19.2.4",
-		"workers-ai-provider": "^3.0.5",
+		"workers-ai-provider": "^3.0.4",
 		"zod": "^4.3.6"
 	},
 	"devDependencies": {
diff --git a/packages/workers-ai-provider/CHANGELOG.md b/packages/workers-ai-provider/CHANGELOG.md
@@ -6,75 +6,67 @@
 
 - [#393](https://github.com/cloudflare/ai/pull/393) [`91b32e0`](https://github.com/cloudflare/ai/commit/91b32e0b0ef543fd198ddf387b9521ac3bd9650a) Thanks [@threepointone](https://github.com/threepointone)! - Comprehensive cleanup of the workers-ai-provider package.
 
-  **Bug fixes:**
-
-  - Fixed phantom dependency on `fetch-event-stream` that caused runtime crashes when installed outside the monorepo. Replaced with a built-in SSE parser.
-  - Fixed streaming buffering: responses now stream token-by-token instead of arriving all at once. The root cause was twofold — an eager `ReadableStream` `start()` pattern that buffered all chunks, and a heuristic that silently fell back to non-streaming `doGenerate` whenever tools were defined. Both are fixed. Streaming now uses a proper `TransformStream` pipeline with backpressure.
-  - Fixed `reasoning-delta` ID mismatch in simulated streaming — was using `generateId()` instead of the `reasoningId` from the preceding `reasoning-start` event, causing the AI SDK to drop reasoning content.
-  - Fixed REST API client (`createRun`) silently swallowing HTTP errors. Non-200 responses now throw with status code and response body.
-  - Fixed `response_format` being sent as `undefined` on every non-JSON request. Now only included when actually set.
-  - Fixed `json_schema` field evaluating to `false` (a boolean) instead of `undefined` when schema was missing.
-
-  **Workers AI quirk workarounds:**
-
-  - Added `sanitizeToolCallId()` — strips non-alphanumeric characters and pads/truncates to 9 chars, fixing tool call round-trips through the binding which rejects its own generated IDs.
-  - Added `normalizeMessagesForBinding()` — converts `content: null` to `""` and sanitizes tool call IDs before every binding call. Only applied on the binding path (REST preserves original IDs).
-  - Added null-finalization chunk filtering for streaming tool calls.
-  - Added numeric value coercion in native-format streams (Workers AI sometimes returns numbers instead of strings for the `response` field).
-  - Improved image model to handle all output types from `binding.run()`: `ReadableStream`, `Uint8Array`, `ArrayBuffer`, `Response`, and `{ image: base64 }` objects.
-  - Graceful degradation: if `binding.run()` returns a non-streaming response despite `stream: true`, it wraps the complete response as a simulated stream instead of throwing.
-
-  **Premature stream termination detection:**
-
-  - Streams that end without a `[DONE]` sentinel now report `finishReason: "error"` with `raw: "stream-truncated"` instead of silently reporting `"stop"`.
-  - Stream read errors are caught and emit `finishReason: "error"` with `raw: "stream-error"`.
-
-  **AI Search (formerly AutoRAG):**
-
-  - Added `createAISearch` and `AISearchChatLanguageModel` as the canonical exports, reflecting the rename from AutoRAG to AI Search.
-  - `createAutoRAG` still works but emits a one-time deprecation warning pointing to `createAISearch`.
-  - `createAutoRAG` preserves `"autorag.chat"` as the provider name for backward compatibility.
-  - AI Search now warns when tools or JSON response format are provided (unsupported by the `aiSearch` API).
-  - Simplified AI Search internals — removed dead tool/response-format processing code.
-
-  **Code quality:**
-
-  - Removed dead code: `workersai-error.ts` (never imported), `workersai-image-config.ts` (inlined).
-  - Consistent file naming: renamed `workers-ai-embedding-model.ts` to `workersai-embedding-model.ts`.
-  - Replaced `StringLike` catch-all index signatures with `[key: string]: unknown` on settings types.
-  - Replaced `any` types with proper interfaces (`FlatToolCall`, `OpenAIToolCall`, `PartialToolCall`).
-  - Tightened `processToolCall` format detection to check `function.name` instead of just the presence of a `function` property.
-  - Removed `@ai-sdk/provider-utils` and `zod` peer dependencies (no longer used in source).
-  - Added `imageModel` to the `WorkersAI` interface type for consistency.
-
-  **Tests:**
-
-  - 149 unit tests across 10 test files (up from 82).
-  - New test coverage: `sanitizeToolCallId`, `normalizeMessagesForBinding`, `prepareToolsAndToolChoice`, `processText`, `mapWorkersAIUsage`, image model output types, streaming error scenarios (malformed SSE, premature termination, empty stream), backpressure verification, graceful degradation (non-streaming fallback with text/tools/reasoning), REST API error handling (401/404/500), AI Search warnings, embedding `TooManyEmbeddingValuesForCallError`, message conversion with images and reasoning.
-  - Integration tests for REST API and binding across 12 models and 7 categories (chat, streaming, multi-turn, tool calling, tool round-trip, structured output, image generation, embeddings).
-  - All tests use the AI SDK's public APIs (`generateText`, `streamText`, `generateImage`, `embedMany`) instead of internal `.doGenerate()`/`.doStream()` methods.
-
-  **README:**
-
-  - Rewritten from scratch with concise examples, model recommendations, configuration guide, and known limitations section.
-  - Updated to use current AI SDK v6 APIs (`generateText` + `Output.object` instead of deprecated `generateObject`, `generateImage` instead of `experimental_generateImage`, `stopWhen: stepCountIs(2)` instead of `maxSteps`).
-  - Added sections for tool calling, structured output, embeddings, image generation, and AI Search.
-  - Uses `wrangler.jsonc` format for configuration examples.
+    **Bug fixes:**
+    - Fixed phantom dependency on `fetch-event-stream` that caused runtime crashes when installed outside the monorepo. Replaced with a built-in SSE parser.
+    - Fixed streaming buffering: responses now stream token-by-token instead of arriving all at once. The root cause was twofold — an eager `ReadableStream` `start()` pattern that buffered all chunks, and a heuristic that silently fell back to non-streaming `doGenerate` whenever tools were defined. Both are fixed. Streaming now uses a proper `TransformStream` pipeline with backpressure.
+    - Fixed `reasoning-delta` ID mismatch in simulated streaming — was using `generateId()` instead of the `reasoningId` from the preceding `reasoning-start` event, causing the AI SDK to drop reasoning content.
+    - Fixed REST API client (`createRun`) silently swallowing HTTP errors. Non-200 responses now throw with status code and response body.
+    - Fixed `response_format` being sent as `undefined` on every non-JSON request. Now only included when actually set.
+    - Fixed `json_schema` field evaluating to `false` (a boolean) instead of `undefined` when schema was missing.
+
+    **Workers AI quirk workarounds:**
+    - Added `sanitizeToolCallId()` — strips non-alphanumeric characters and pads/truncates to 9 chars, fixing tool call round-trips through the binding which rejects its own generated IDs.
+    - Added `normalizeMessagesForBinding()` — converts `content: null` to `""` and sanitizes tool call IDs before every binding call. Only applied on the binding path (REST preserves original IDs).
+    - Added null-finalization chunk filtering for streaming tool calls.
+    - Added numeric value coercion in native-format streams (Workers AI sometimes returns numbers instead of strings for the `response` field).
+    - Improved image model to handle all output types from `binding.run()`: `ReadableStream`, `Uint8Array`, `ArrayBuffer`, `Response`, and `{ image: base64 }` objects.
+    - Graceful degradation: if `binding.run()` returns a non-streaming response despite `stream: true`, it wraps the complete response as a simulated stream instead of throwing.
+
+    **Premature stream termination detection:**
+    - Streams that end without a `[DONE]` sentinel now report `finishReason: "error"` with `raw: "stream-truncated"` instead of silently reporting `"stop"`.
+    - Stream read errors are caught and emit `finishReason: "error"` with `raw: "stream-error"`.
+
+    **AI Search (formerly AutoRAG):**
+    - Added `createAISearch` and `AISearchChatLanguageModel` as the canonical exports, reflecting the rename from AutoRAG to AI Search.
+    - `createAutoRAG` still works but emits a one-time deprecation warning pointing to `createAISearch`.
+    - `createAutoRAG` preserves `"autorag.chat"` as the provider name for backward compatibility.
+    - AI Search now warns when tools or JSON response format are provided (unsupported by the `aiSearch` API).
+    - Simplified AI Search internals — removed dead tool/response-format processing code.
+
+    **Code quality:**
+    - Removed dead code: `workersai-error.ts` (never imported), `workersai-image-config.ts` (inlined).
+    - Consistent file naming: renamed `workers-ai-embedding-model.ts` to `workersai-embedding-model.ts`.
+    - Replaced `StringLike` catch-all index signatures with `[key: string]: unknown` on settings types.
+    - Replaced `any` types with proper interfaces (`FlatToolCall`, `OpenAIToolCall`, `PartialToolCall`).
+    - Tightened `processToolCall` format detection to check `function.name` instead of just the presence of a `function` property.
+    - Removed `@ai-sdk/provider-utils` and `zod` peer dependencies (no longer used in source).
+    - Added `imageModel` to the `WorkersAI` interface type for consistency.
+
+    **Tests:**
+    - 149 unit tests across 10 test files (up from 82).
+    - New test coverage: `sanitizeToolCallId`, `normalizeMessagesForBinding`, `prepareToolsAndToolChoice`, `processText`, `mapWorkersAIUsage`, image model output types, streaming error scenarios (malformed SSE, premature termination, empty stream), backpressure verification, graceful degradation (non-streaming fallback with text/tools/reasoning), REST API error handling (401/404/500), AI Search warnings, embedding `TooManyEmbeddingValuesForCallError`, message conversion with images and reasoning.
+    - Integration tests for REST API and binding across 12 models and 7 categories (chat, streaming, multi-turn, tool calling, tool round-trip, structured output, image generation, embeddings).
+    - All tests use the AI SDK's public APIs (`generateText`, `streamText`, `generateImage`, `embedMany`) instead of internal `.doGenerate()`/`.doStream()` methods.
+
+    **README:**
+    - Rewritten from scratch with concise examples, model recommendations, configuration guide, and known limitations section.
+    - Updated to use current AI SDK v6 APIs (`generateText` + `Output.object` instead of deprecated `generateObject`, `generateImage` instead of `experimental_generateImage`, `stopWhen: stepCountIs(2)` instead of `maxSteps`).
+    - Added sections for tool calling, structured output, embeddings, image generation, and AI Search.
+    - Uses `wrangler.jsonc` format for configuration examples.
 
 ## 3.0.4
 
 ### Patch Changes
 
 - [#390](https://github.com/cloudflare/ai/pull/390) [`41b92a3`](https://github.com/cloudflare/ai/commit/41b92a34ce4d9dffba8bb42b4933bbc06e4b1aaa) Thanks [@mchenco](https://github.com/mchenco)! - fix(workers-ai-provider): extract actual finish reason in streaming instead of hardcoded "stop"
 
-  Previously, the streaming implementation always returned `finishReason: "stop"` regardless of the actual completion reason. This caused:
-
-  - Tool calling scenarios to incorrectly report "stop" instead of "tool-calls"
-  - Multi-turn tool conversations to fail because the AI SDK couldn't detect when tools were requested
-  - Length limit scenarios to show "stop" instead of "length"
-  - Error scenarios to show "stop" instead of "error"
+    Previously, the streaming implementation always returned `finishReason: "stop"` regardless of the actual completion reason. This caused:
+    - Tool calling scenarios to incorrectly report "stop" instead of "tool-calls"
+    - Multi-turn tool conversations to fail because the AI SDK couldn't detect when tools were requested
+    - Length limit scenarios to show "stop" instead of "length"
+    - Error scenarios to show "stop" instead of "error"
 
-  The fix extracts the actual `finish_reason` from streaming chunks and uses the existing `mapWorkersAIFinishReason()` function to properly map it to the AI SDK's finish reason format. This enables proper multi-turn tool calling and accurate completion status reporting.
+    The fix extracts the actual `finish_reason` from streaming chunks and uses the existing `mapWorkersAIFinishReason()` function to properly map it to the AI SDK's finish reason format. This enables proper multi-turn tool calling and accurate completion status reporting.
 
 ## 3.0.3
 
@@ -118,7 +110,7 @@
 
 - [#256](https://github.com/cloudflare/ai/pull/256) [`a538901`](https://github.com/cloudflare/ai/commit/a5389013b9a512707fb1de1501a1547fce20c014) Thanks [@jahands](https://github.com/jahands)! - feat: Migrate to AI SDK v5
 
-  This updates workers-ai-provider and ai-gateway-provider to use the AI SDK v5. Please refer to the official migration guide to migrate your code https://ai-sdk.dev/docs/migration-guides/migration-guide-5-0
+    This updates workers-ai-provider and ai-gateway-provider to use the AI SDK v5. Please refer to the official migration guide to migrate your code https://ai-sdk.dev/docs/migration-guides/migration-guide-5-0
 
 ### Patch Changes
 
