cloudflare/ai workers-ai-provider@3.0.5

on GitHub

Patch Changes

• #393 91b32e0 Thanks @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.