cloudflare/ai workers-ai-provider@3.3.0 on GitHub

Minor Changes

#590 fe0d182 Thanks @threepointone! - The AI Gateway delegate gains cross-vendor server-side fallback
(fallback: { mode: "server" }) — multiple vendors in one gateway run, with the
winner selected via cf-aig-step.
The gateway delegate now reaches header parity with the run path: the gateway
path forwards cacheKey, eventId, requestTimeoutMs, and retries from the
gateway options as cf-aig-* headers, and DelegateCallOptions gains two new
universal-endpoint controls — byokAlias (cf-aig-byok-alias, select a stored
BYOK key by alias) and zdr (cf-aig-zdr, per-request Zero Data Retention
override for Unified Billing, applied on both transports).
Internally, the provider registry, cf-aig-* header building, resumable-stream
engine, and Workers AI SSE helpers are now shared across the Cloudflare AI
packages (bundled inline — no new dependency for you to install).
#593 1c6afd0 Thanks @threepointone! - Native Workers AI failures are now surfaced as AI SDK APICallErrors so the AI
SDK's built-in retry (maxRetries) can engage on transient errors.
Previously the binding path (env.AI.run) threw plain Errors and the REST
path threw a generic Error, so the AI SDK never retried them — most notably
the common "out of capacity" failure (internal code 3040, HTTP 429) and
other 5xx blips just failed the call outright.
- Binding path: errors thrown by env.AI.run are normalized into an
  APICallError across every Workers AI model — chat, embedding, image, speech,
  transcription, and reranking. The Workers AI internal error code is parsed from
  the message (or a numeric code property) and mapped to the documented HTTP
  status (e.g. 3040/3036 → 429, 3007/3008 → 408, 5007 → 400), and
  APICallError derives isRetryable from that status (retryable on
  408/409/429/5xx). Unrecognized errors get no status and stay non-retryable
  (prior behavior). AbortError/TimeoutError cancellations propagate
  unchanged.
- REST path: non-OK responses now throw an APICallError carrying the real
  statusCode, response headers (so Retry-After is honored), and body, instead
  of a generic Error. The error message keeps the same
  Workers AI API error (<status> <statusText>): <body> shape.
This means transient capacity/5xx errors are now automatically retried with
exponential backoff by generateText/streamText (default 2 retries; tune via
maxRetries). Set maxRetries: 0 to opt out.