ai-retry 1.9.0 → 1.9.1

package/README.md

@@ -1183,6 +1183,13 @@ In the second case, errors during stream processing will not always be retried,
 > [!IMPORTANT]
 > **Streaming limitation:** Retries and fallbacks only apply before the first content chunk is emitted. Once streaming begins delivering content, the response is committed to the current model. Mid-stream errors will propagate to the caller rather than triggering a fallback. If reliable retries are critical for your use case, consider using `generateText` instead of `streamText`.
 
+#### Preamble buffering
+
+Every stream begins with a non-content preamble (`stream-start`, then optionally `response-metadata` and `text-start` / `reasoning-start`) that providers emit as soon as the response headers arrive, before any content flows. Because a retry can still happen during this window, `ai-retry` does not forward the preamble immediately. It buffers the leading non-content parts and flushes them only when the first content chunk arrives (or when the stream finishes with no content). If a retry fires before any content, the buffered preamble is discarded and replaced by the fallback's, so the consumer always sees exactly one preamble — the one belonging to the model that actually produced the output, with its own `warnings` and `response-metadata`. Without this, a fallback's `stream-start` would be emitted a second time after the primary's, which some consumers (e.g. `streamText`) reject.
+
+> [!NOTE]
+> One side effect: the consumer's "stream started" signal now arrives at first-content time rather than when the response headers arrive (typically a sub-second difference). For UIs that show a typing indicator off `stream-start` this is negligible.
+
 ### API Reference
 
 #### `createRetryable(options: RetryableModelOptions): LanguageModelV3 | EmbeddingModelV3 | ImageModelV3`

package/dist/{create-retryable-model-VgywojZt.mjs → create-retryable-model-Ch2cqC3Z.mjs}

@@ -1202,6 +1202,16 @@ var RetryableLanguageModel = class extends BaseRetryableModel {
 					let capturedWarnings = [];
 					let capturedResponseMetadata = {};
 					/**
+					* Buffer for the leading non-content parts (`stream-start`,
+					* `response-metadata`, `text-start`, `reasoning-start`, …) of this
+					* attempt. While no content has been forwarded the preamble is held
+					* here rather than enqueued, so a pre-content retry can discard it
+					* and the consumer sees exactly one preamble — the one belonging to
+					* the model that actually produced the output. Reset per attempt;
+					* flushed on the first content part or at completion.
+					*/
+					let preambleBuffer = [];
+					/**
 					* Set when a `finish` part triggers a retry decision. Causes the
 					* inner read loop to exit without enqueuing the finish part, and
 					* the outer loop to re-stream against the next model.
@@ -1270,11 +1280,25 @@ var RetryableLanguageModel = class extends BaseRetryableModel {
 								}
 							}
 							/**
-							* Mark that streaming has started once we receive actual content
+							* Mark that streaming has started once we receive actual
+							* content. On the first content part, flush this attempt's
+							* buffered preamble (in order) ahead of the content, then
+							* forward normally from here on.
 							*/
-							if (isStreamContentPart(value)) isStreaming = true;
-							/**
-							* Enqueue the chunk to the consumer of the stream
+							if (isStreamContentPart(value)) {
+								isStreaming = true;
+								for (const buffered of preambleBuffer) controller.enqueue(buffered);
+								preambleBuffer = [];
+								controller.enqueue(value);
+							} else if (!isStreaming)
+ /**
+							* Pre-content part: buffer it so a pre-content retry can
+							* replace it with the next attempt's preamble.
+							*/
+							preambleBuffer.push(value);
+							else
+ /**
+							* Content already flowing: forward directly.
 							*/
 							controller.enqueue(value);
 						}
@@ -1306,7 +1330,13 @@ var RetryableLanguageModel = class extends BaseRetryableModel {
 								currentRetry,
 								recorder
 							});
-							await reader?.cancel();
+							/**
+							* Cancelling a reader whose stream has already errored (e.g.
+							* a mid-stream `controller.error`) rejects with that stored
+							* error. Swallow it: the retry already succeeded and that
+							* rejection must not abort the wrapped stream.
+							*/
+							await reader?.cancel().catch(() => {});
 							result = retriedResult.result;
 							attempts = retriedResult.attempts;
 							finalCallOptions = retriedResult.callOptions;
@@ -1318,6 +1348,13 @@ var RetryableLanguageModel = class extends BaseRetryableModel {
 							outcome: "success",
 							finishReason: streamFinishReason
 						});
+						/**
+						* A stream that completes with no content part still has its
+						* preamble buffered. Flush it so a zero-content completion emits
+						* its `stream-start` (and any metadata/finish) before closing.
+						*/
+						for (const buffered of preambleBuffer) controller.enqueue(buffered);
+						preambleBuffer = [];
 						controller.close();
 						break;
 					} catch (error) {
@@ -1408,9 +1445,13 @@ var RetryableLanguageModel = class extends BaseRetryableModel {
 							recorder
 						});
 						/**
-						* Cancel the previous reader and stream if we are retrying
+						* Cancel the previous reader and stream if we are retrying.
+						* Cancelling a reader whose stream has already errored (e.g. a
+						* mid-stream `controller.error`) rejects with that stored
+						* error. Swallow it: the retry already succeeded and that
+						* rejection must not abort the wrapped stream.
 						*/
-						await reader?.cancel();
+						await reader?.cancel().catch(() => {});
 						result = retriedResult.result;
 						attempts = retriedResult.attempts;
 						finalCallOptions = retriedResult.callOptions;

package/dist/experimental/embedding-model/index.mjs

@@ -1,4 +1,4 @@
-import { t as createRetryable$1 } from "../../create-retryable-model-VgywojZt.mjs";
+import { t as createRetryable$1 } from "../../create-retryable-model-Ch2cqC3Z.mjs";
 import "../../error-CaTT-xX8.mjs";
 import { aborted, error, httpStatus, timeout } from "./retryables/index.mjs";
 

package/dist/experimental/image-model/index.mjs

@@ -1,4 +1,4 @@
-import { t as createRetryable$1 } from "../../create-retryable-model-VgywojZt.mjs";
+import { t as createRetryable$1 } from "../../create-retryable-model-Ch2cqC3Z.mjs";
 import "../../error-CaTT-xX8.mjs";
 import { a as noImage, i as timeout, n as error, r as httpStatus, t as aborted } from "../../retryables-CPAbu_M3.mjs";
 

package/dist/experimental/language-model/index.mjs

@@ -1,4 +1,4 @@
-import { t as createRetryable$1 } from "../../create-retryable-model-VgywojZt.mjs";
+import { t as createRetryable$1 } from "../../create-retryable-model-Ch2cqC3Z.mjs";
 import "../../error-CaTT-xX8.mjs";
 import { a as result, i as httpStatus, n as error, o as schemaInvalid, r as finishReason, s as timeout, t as aborted } from "../../retryables-M5l_6w9k.mjs";
 

package/dist/index.mjs

@@ -1,4 +1,4 @@
-import { n as getModelKey, t as createRetryable } from "./create-retryable-model-VgywojZt.mjs";
+import { n as getModelKey, t as createRetryable } from "./create-retryable-model-Ch2cqC3Z.mjs";
 import { n as isErrorAttempt, s as isResultAttempt } from "./guards-D8UJtxDK.mjs";
 
 export { createRetryable, getModelKey, isErrorAttempt, isResultAttempt };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ai-retry",
-  "version": "1.9.0",
+  "version": "1.9.1",
   "description": "Retry and fallback mechanisms for AI SDK",
   "keywords": [
     "ai",