cloudflare/ai @cloudflare/tanstack-ai@0.2.0

on GitHub

Minor Changes

• #590 fe0d182 Thanks @threepointone! - - Add resumable streaming to the Workers AI adapter (coming soon — not
  generally available yet while the AI Gateway resume backend rolls out; treat as
  experimental): catalog models dispatch through the AI Gateway run path, so
  transient mid-stream drops reconnect transparently via cf-aig-run-id.
  Configure with resume / onResumeExpired (no-op + warning where no run id is
  available, e.g. REST).

  • Gain the gpt-oss forced tool-call salvage (#560) and non-SSE
    graceful-degradation, now shared with workers-ai-provider.
  • Bump @tanstack/ai and the @tanstack/ai-* adapter peers to current versions
    (adapts to the multimodal MediaPrompt API). @ai-sdk/* is intentionally not
    bumped.

• #594 12fb307 Thanks @threepointone! - Retry transient Workers AI failures and normalize errors across every adapter.

  • Chat: the binding shim now surfaces binding failures as HTTP responses
    (e.g. "out of capacity" 3040 → 429, "no such model" 5007 → 400) so the
    OpenAI SDK's status-based retry engages and honors Retry-After. Aborts and
    unrecognized errors propagate untouched. Non-OK gateway run-path responses are
    returned verbatim instead of being swallowed into an empty completion.
  • Non-chat adapters (embedding, image, TTS, transcription, summarize) gain a
    bounded exponential-backoff retry (the OpenAI SDK isn't in play for these) and
    normalize binding / REST / gateway failures into a single WorkersAiRequestError
    carrying the HTTP status (and the raw Workers AI code when recognized). The
    retry loop honors a server Retry-After header. Non-OK gateway responses are no
    longer swallowed.
  • Add a maxRetries option to the adapter config: forwarded to the OpenAI SDK on
    the chat path, and used by the non-chat retry loop. Defaults to 2; set to 0
    to disable.