@librechat/agents 3.3.12 → 3.4.0

package/dist/esm/llm/stream/smoother.mjs

@@ -0,0 +1,365 @@
+//#region src/llm/stream/smoother.ts
+const DEFAULT_STREAM_DELAY = 25;
+const MAX_SMOOTH_ITEM_SEGMENT_CHARS = 4096;
+const STREAM_BOUNDARIES = new Set([
+	" ",
+	".",
+	",",
+	"!",
+	"?",
+	";",
+	":"
+]);
+const STREAM_ABORT_MESSAGE = "AbortError: User aborted the request.";
+/**
+* How long generator teardown waits for the background producer to observe a
+* consumer close before abandoning it. Well-behaved streams settle in
+* microseconds (the next enqueue throws); a stalled provider that ignores
+* aborts otherwise blocks teardown — and abort propagation — indefinitely.
+* An abandoned producer still self-terminates on its next enqueue attempt.
+*/
+const PRODUCER_CLOSE_GRACE_MS = 1e3;
+/**
+* Resolves a configured stream delay to its effective value (default 25ms;
+* 0 disables smoothing). Non-finite inputs (NaN from a malformed config
+* value, ±Infinity) normalize to the default rather than poisoning piece
+* arithmetic downstream.
+*/
+function resolveStreamDelay(delay) {
+	if (delay == null || !Number.isFinite(delay)) return 25;
+	return Math.max(0, delay);
+}
+function isSignalAborted(signal) {
+	return signal?.aborted === true;
+}
+function findStreamChunkBoundary(text, minSize) {
+	if (minSize >= text.length) return text.length;
+	const scanEnd = Math.min(text.length, minSize + 64);
+	for (let position = minSize; position < scanEnd; position++) if (STREAM_BOUNDARIES.has(text[position])) return position + 1;
+	return scanEnd;
+}
+/**
+* Backlog-proportional piece sizing: emit enough per tick that the current
+* backlog drains in ~`targetLatencyMs`, so render lag stays pinned near the
+* target regardless of how fast the provider streams. Token-sized arrivals
+* never exceed the minimum piece, matching the legacy fixed-size splitter.
+*/
+function computeAdaptivePieceSize(bufferedTextLength, tickMs, targetLatencyMs = 250) {
+	if (bufferedTextLength <= 0) return 4;
+	if (tickMs <= 0 || targetLatencyMs <= 0) return bufferedTextLength;
+	return Math.max(4, Math.ceil(bufferedTextLength * tickMs / targetLatencyMs));
+}
+/**
+* A cadence, not an additive sleep: time the consumer already spent since the
+* last visible emission counts against the target delay, so slow downstream
+* handlers never compound latency.
+*/
+function getCadencedStreamDelay({ targetDelay, lastVisibleTextAt, now }) {
+	if (targetDelay <= 0 || lastVisibleTextAt == null) return 0;
+	return Math.max(0, targetDelay - (now - lastVisibleTextAt));
+}
+/** Abort-aware sleep that resolves (never rejects) on abort; callers re-check the signal. */
+async function waitForStreamDelay(delay, signal) {
+	if (delay <= 0 || isSignalAborted(signal)) return;
+	await new Promise((resolve) => {
+		const timeoutRef = {};
+		const onAbort = () => {
+			if (timeoutRef.current) clearTimeout(timeoutRef.current);
+			signal?.removeEventListener("abort", onAbort);
+			resolve();
+		};
+		timeoutRef.current = setTimeout(() => {
+			signal?.removeEventListener("abort", onAbort);
+			resolve();
+		}, delay);
+		signal?.addEventListener("abort", onAbort, { once: true });
+		if (isSignalAborted(signal)) onAbort();
+	});
+}
+/**
+* Bounded producer/consumer smoothing engine.
+*
+* The producer drains `source` eagerly into a bounded queue (the buffer is the
+* backlog measurement adaptive sizing needs); at capacity it parks, applying
+* backpressure to the underlying stream. The consumer emits paced pieces,
+* decrementing the text budget and waking the producer *before* each cadenced
+* sleep so the provider stream keeps being read during pacing.
+*
+* `delayMs <= 0` disables smoothing entirely: every item passes through FIFO,
+* unsplit and undelayed.
+*/
+async function* smoothStream({ source, delayMs, signal, abortUpstream }) {
+	if (!(delayMs > 0)) {
+		/** Disabled smoothing preserves fully lazy streaming: no background
+		* producer, no read-ahead — each provider chunk is pulled only when the
+		* consumer asks, exactly like the pre-engine pass-through paths. */
+		for await (const item of source) {
+			if (isSignalAborted(signal)) {
+				abortUpstream?.();
+				throw new Error(STREAM_ABORT_MESSAGE);
+			}
+			yield item.emit({
+				text: item.text,
+				isFirst: true,
+				isLast: true
+			});
+		}
+		return;
+	}
+	const queuedItems = [];
+	const producerState = {
+		done: false,
+		failed: false
+	};
+	let queuedItemIndex = 0;
+	let bufferedTextLength = 0;
+	let consumerClosed = false;
+	let notifyConsumer;
+	let notifyProducer;
+	const notifyConsumerForItem = () => {
+		notifyConsumer?.();
+		notifyConsumer = void 0;
+	};
+	const notifyProducerForSpace = () => {
+		notifyProducer?.();
+		notifyProducer = void 0;
+	};
+	const hasQueuedItems = () => queuedItemIndex < queuedItems.length;
+	const getQueuedItemCount = () => queuedItems.length - queuedItemIndex;
+	const isQueueAtCapacity = () => getQueuedItemCount() >= 256 || bufferedTextLength >= 8192;
+	/** Abort-aware: a consumer parked on an empty queue must wake when the
+	* signal fires even if the provider stream never honors the abort — the
+	* loop's top-of-iteration check then throws the canonical error. */
+	const waitForNextItem = async () => {
+		if (hasQueuedItems() || producerState.done || producerState.failed || isSignalAborted(signal)) return;
+		await new Promise((resolve) => {
+			const onAbort = () => {
+				signal?.removeEventListener("abort", onAbort);
+				resolve();
+			};
+			notifyConsumer = () => {
+				signal?.removeEventListener("abort", onAbort);
+				resolve();
+			};
+			signal?.addEventListener("abort", onAbort, { once: true });
+			if (isSignalAborted(signal)) onAbort();
+		});
+	};
+	const waitForQueueSpace = async () => {
+		while (isQueueAtCapacity() && !consumerClosed && !isSignalAborted(signal)) await new Promise((resolve) => {
+			const onAbort = () => {
+				signal?.removeEventListener("abort", onAbort);
+				resolve();
+			};
+			const onSpace = () => {
+				signal?.removeEventListener("abort", onAbort);
+				resolve();
+			};
+			notifyProducer = onSpace;
+			signal?.addEventListener("abort", onAbort, { once: true });
+			if (isSignalAborted(signal)) onAbort();
+		});
+	};
+	const dequeue = () => {
+		if (!hasQueuedItems()) return;
+		const queuedItem = queuedItems[queuedItemIndex];
+		queuedItemIndex++;
+		if (queuedItemIndex > 128 && queuedItemIndex * 2 >= queuedItems.length) {
+			queuedItems.splice(0, queuedItemIndex);
+			queuedItemIndex = 0;
+		}
+		return queuedItem;
+	};
+	const throwAborted = () => {
+		abortUpstream?.();
+		throw new Error(STREAM_ABORT_MESSAGE);
+	};
+	const enqueue = async (item) => {
+		await waitForQueueSpace();
+		if (consumerClosed || isSignalAborted(signal)) throwAborted();
+		const textLength = item.smooth ? item.text.length : 0;
+		queuedItems.push({
+			item,
+			textLength
+		});
+		bufferedTextLength += textLength;
+		notifyConsumerForItem();
+	};
+	/**
+	* Oversized splittable items are segmented at admission so a single giant
+	* provider chunk cannot blow past the text budget: each segment re-checks
+	* capacity, so the producer parks mid-chunk once the buffer fills — the
+	* same bound the legacy split-before-enqueue queues enforced. The wrapped
+	* emit maps segment-local pieces back to chunk-global isFirst/isLast so
+	* provider clone contracts are unaffected.
+	*/
+	const enqueueSegmented = async (item) => {
+		if (!item.smooth || item.atomic === true || item.text.length <= 4096) {
+			await enqueue(item);
+			return;
+		}
+		const segments = [];
+		let offset = 0;
+		while (offset < item.text.length) {
+			const end = offset + findStreamChunkBoundary(item.text.slice(offset), MAX_SMOOTH_ITEM_SEGMENT_CHARS);
+			segments.push({
+				start: offset,
+				end
+			});
+			offset = end;
+		}
+		for (let i = 0; i < segments.length; i++) {
+			const isFirstSegment = i === 0;
+			const isLastSegment = i === segments.length - 1;
+			await enqueue({
+				text: item.text.slice(segments[i].start, segments[i].end),
+				smooth: true,
+				emit: (piece) => item.emit({
+					text: piece.text,
+					isFirst: isFirstSegment && piece.isFirst,
+					isLast: isLastSegment && piece.isLast
+				})
+			});
+		}
+	};
+	const producer = (async () => {
+		try {
+			for await (const item of source) {
+				if (isSignalAborted(signal)) throwAborted();
+				await enqueueSegmented(item);
+			}
+		} catch (error) {
+			producerState.failed = true;
+			producerState.error = error;
+		} finally {
+			producerState.done = true;
+			notifyConsumerForItem();
+		}
+	})();
+	let hasEmittedText = false;
+	let lastVisibleTextAt;
+	let drainTicksRemaining;
+	let current;
+	let headOffset = 0;
+	let keepStreaming = true;
+	try {
+		while (keepStreaming) {
+			if (isSignalAborted(signal)) throwAborted();
+			if (current == null) {
+				await waitForNextItem();
+				current = dequeue();
+				headOffset = 0;
+			}
+			if (current == null) {
+				if (producerState.failed) throw producerState.error ?? /* @__PURE__ */ new Error("Stream producer failed.");
+				if (producerState.done) keepStreaming = false;
+				continue;
+			}
+			const { item } = current;
+			if (!item.smooth) {
+				notifyProducerForSpace();
+				current = void 0;
+				yield item.emit({
+					text: item.text,
+					isFirst: true,
+					isLast: true
+				});
+				continue;
+			}
+			if (item.text === "") {
+				bufferedTextLength = Math.max(0, bufferedTextLength - current.textLength);
+				notifyProducerForSpace();
+				current = void 0;
+				continue;
+			}
+			/** Once the producer is done the backlog is final: drain it linearly
+			* across the remaining target window instead of letting the
+			* proportional formula decay geometrically and stretch the tail. */
+			if (producerState.done && drainTicksRemaining == null) drainTicksRemaining = Math.max(1, Math.floor(250 / delayMs));
+			const tickBudget = drainTicksRemaining != null ? Math.max(4, Math.ceil(bufferedTextLength / drainTicksRemaining)) : computeAdaptivePieceSize(bufferedTextLength, delayMs);
+			if (drainTicksRemaining != null && drainTicksRemaining > 1) drainTicksRemaining -= 1;
+			await waitForStreamDelay(getCadencedStreamDelay({
+				targetDelay: hasEmittedText ? delayMs : 0,
+				lastVisibleTextAt,
+				now: Date.now()
+			}), signal);
+			if (isSignalAborted(signal)) throwAborted();
+			hasEmittedText = true;
+			lastVisibleTextAt = Date.now();
+			if (item.atomic === true) {
+				bufferedTextLength = Math.max(0, bufferedTextLength - current.textLength);
+				notifyProducerForSpace();
+				current = void 0;
+				yield item.emit({
+					text: item.text,
+					isFirst: true,
+					isLast: true
+				});
+				continue;
+			}
+			/** One cadence tick drains up to the adaptive budget ACROSS queued
+			* items, so token-sized provider deltas coalesce instead of costing a
+			* full tick each; passthrough items flush free mid-batch (FIFO), and
+			* atomic items end the batch to take their own tick. */
+			let consumed = 0;
+			while (consumed < tickBudget) {
+				if (isSignalAborted(signal)) throwAborted();
+				if (current == null) {
+					if (!hasQueuedItems()) break;
+					current = dequeue();
+					headOffset = 0;
+					if (current == null) break;
+				}
+				const batchItem = current.item;
+				if (!batchItem.smooth) {
+					notifyProducerForSpace();
+					current = void 0;
+					yield batchItem.emit({
+						text: batchItem.text,
+						isFirst: true,
+						isLast: true
+					});
+					continue;
+				}
+				if (batchItem.text === "") {
+					bufferedTextLength = Math.max(0, bufferedTextLength - current.textLength);
+					notifyProducerForSpace();
+					current = void 0;
+					continue;
+				}
+				if (batchItem.atomic === true) break;
+				const pieceLength = findStreamChunkBoundary(batchItem.text.slice(headOffset), tickBudget - consumed);
+				const pieceEnd = headOffset + pieceLength;
+				const piece = batchItem.text.slice(headOffset, pieceEnd);
+				const isFirst = headOffset === 0;
+				const isLast = pieceEnd === batchItem.text.length;
+				bufferedTextLength = Math.max(0, bufferedTextLength - piece.length);
+				notifyProducerForSpace();
+				if (isLast) current = void 0;
+				else headOffset = pieceEnd;
+				consumed += piece.length;
+				yield batchItem.emit({
+					text: piece,
+					isFirst,
+					isLast
+				});
+			}
+		}
+	} finally {
+		consumerClosed = true;
+		if (producerState.done) await producer;
+		else {
+			abortUpstream?.();
+			notifyProducerForSpace();
+			const closing = source.return?.call(source, void 0);
+			if (closing != null) closing.then(() => void 0, () => void 0);
+			await Promise.race([producer, new Promise((resolve) => {
+				setTimeout(resolve, PRODUCER_CLOSE_GRACE_MS).unref();
+			})]);
+		}
+	}
+}
+//#endregion
+export { DEFAULT_STREAM_DELAY, computeAdaptivePieceSize, isSignalAborted, resolveStreamDelay, smoothStream };
+
+//# sourceMappingURL=smoother.mjs.map

package/dist/esm/llm/stream/smoother.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"smoother.mjs","names":["iterator"],"sources":["../../../../src/llm/stream/smoother.ts"],"sourcesContent":["export const DEFAULT_STREAM_DELAY = 25;\nexport const SMOOTH_TARGET_LATENCY_MS = 250;\nexport const MAX_STREAM_QUEUE_CHUNKS = 256;\nexport const MAX_STREAM_QUEUE_TEXT_CHARS = 8192;\nexport const MAX_SMOOTH_ITEM_SEGMENT_CHARS = 4096;\nexport const STREAM_CHUNK_MIN_SIZE = 4;\nexport const STREAM_BOUNDARIES: ReadonlySet<string> = new Set([\n  ' ',\n  '.',\n  ',',\n  '!',\n  '?',\n  ';',\n  ':',\n]);\n\nexport const STREAM_ABORT_MESSAGE = 'AbortError: User aborted the request.';\nexport const STREAM_PRODUCER_FAILURE = 'Stream producer failed.';\n\n/**\n * How long generator teardown waits for the background producer to observe a\n * consumer close before abandoning it. Well-behaved streams settle in\n * microseconds (the next enqueue throws); a stalled provider that ignores\n * aborts otherwise blocks teardown — and abort propagation — indefinitely.\n * An abandoned producer still self-terminates on its next enqueue attempt.\n */\nexport const PRODUCER_CLOSE_GRACE_MS = 1000;\n\n/**\n * Resolves a configured stream delay to its effective value (default 25ms;\n * 0 disables smoothing). Non-finite inputs (NaN from a malformed config\n * value, ±Infinity) normalize to the default rather than poisoning piece\n * arithmetic downstream.\n */\nexport function resolveStreamDelay(delay?: number): number {\n  if (delay == null || !Number.isFinite(delay)) {\n    return DEFAULT_STREAM_DELAY;\n  }\n  return Math.max(0, delay);\n}\n\nexport function isSignalAborted(signal?: AbortSignal): boolean {\n  return signal?.aborted === true;\n}\n\n/**\n * How far past the target size the word-boundary search may extend before\n * hard-cutting. Natural language hits a boundary within a few characters;\n * boundary-free runs (base64, minified data, long identifiers) must not\n * stretch a piece — or an admission segment — arbitrarily far past its\n * budget.\n */\nexport const STREAM_BOUNDARY_LOOKAHEAD_CHARS = 64;\n\nexport function findStreamChunkBoundary(\n  text: string,\n  minSize: number\n): number {\n  if (minSize >= text.length) {\n    return text.length;\n  }\n\n  const scanEnd = Math.min(\n    text.length,\n    minSize + STREAM_BOUNDARY_LOOKAHEAD_CHARS\n  );\n  for (let position = minSize; position < scanEnd; position++) {\n    if (STREAM_BOUNDARIES.has(text[position])) {\n      return position + 1;\n    }\n  }\n\n  return scanEnd;\n}\n\n/**\n * Backlog-proportional piece sizing: emit enough per tick that the current\n * backlog drains in ~`targetLatencyMs`, so render lag stays pinned near the\n * target regardless of how fast the provider streams. Token-sized arrivals\n * never exceed the minimum piece, matching the legacy fixed-size splitter.\n */\nexport function computeAdaptivePieceSize(\n  bufferedTextLength: number,\n  tickMs: number,\n  targetLatencyMs: number = SMOOTH_TARGET_LATENCY_MS\n): number {\n  if (bufferedTextLength <= 0) {\n    return STREAM_CHUNK_MIN_SIZE;\n  }\n  if (tickMs <= 0 || targetLatencyMs <= 0) {\n    return bufferedTextLength;\n  }\n  return Math.max(\n    STREAM_CHUNK_MIN_SIZE,\n    Math.ceil((bufferedTextLength * tickMs) / targetLatencyMs)\n  );\n}\n\n/**\n * A cadence, not an additive sleep: time the consumer already spent since the\n * last visible emission counts against the target delay, so slow downstream\n * handlers never compound latency.\n */\nexport function getCadencedStreamDelay({\n  targetDelay,\n  lastVisibleTextAt,\n  now,\n}: {\n  targetDelay: number;\n  lastVisibleTextAt?: number;\n  now: number;\n}): number {\n  if (targetDelay <= 0 || lastVisibleTextAt == null) {\n    return 0;\n  }\n  return Math.max(0, targetDelay - (now - lastVisibleTextAt));\n}\n\n/** Abort-aware sleep that resolves (never rejects) on abort; callers re-check the signal. */\nexport async function waitForStreamDelay(\n  delay: number,\n  signal?: AbortSignal\n): Promise<void> {\n  if (delay <= 0 || isSignalAborted(signal)) {\n    return;\n  }\n  await new Promise<void>((resolve) => {\n    const timeoutRef: { current?: ReturnType<typeof setTimeout> } = {};\n    const onAbort = (): void => {\n      if (timeoutRef.current) {\n        clearTimeout(timeoutRef.current);\n      }\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    };\n    timeoutRef.current = setTimeout(() => {\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    }, delay);\n    signal?.addEventListener('abort', onAbort, { once: true });\n    if (isSignalAborted(signal)) {\n      onAbort();\n    }\n  });\n}\n\nexport type SmoothPiece = {\n  text: string;\n  isFirst: boolean;\n  isLast: boolean;\n};\n\n/**\n * One classified unit of provider stream output.\n *\n * - `smooth: true` — visible text, paced at the configured cadence and (unless\n *   `atomic`) sliced adaptively at dequeue time.\n * - `atomic: true` — paced as a single piece, never split (text-bearing chunks\n *   whose metadata cannot survive slicing, e.g. logprobs / finish_reason).\n * - `smooth: false` — passthrough: tool-call deltas, usage-only, id-only and\n *   seal chunks. Zero delay, strict FIFO with the text around them.\n *\n * `emit` builds the provider-specific output for one piece; `isFirst` lets\n * providers keep usage_metadata on only the first piece of a split.\n */\nexport type SmoothItem<TEmit> = {\n  text: string;\n  smooth: boolean;\n  atomic?: boolean;\n  emit: (piece: SmoothPiece) => TEmit;\n};\n\ntype ProducerState = {\n  done: boolean;\n  failed: boolean;\n  error?: unknown;\n};\n\ntype QueuedSmoothItem<TEmit> = {\n  item: SmoothItem<TEmit>;\n  textLength: number;\n};\n\n/**\n * Bounded producer/consumer smoothing engine.\n *\n * The producer drains `source` eagerly into a bounded queue (the buffer is the\n * backlog measurement adaptive sizing needs); at capacity it parks, applying\n * backpressure to the underlying stream. The consumer emits paced pieces,\n * decrementing the text budget and waking the producer *before* each cadenced\n * sleep so the provider stream keeps being read during pacing.\n *\n * `delayMs <= 0` disables smoothing entirely: every item passes through FIFO,\n * unsplit and undelayed.\n */\nexport async function* smoothStream<TEmit>({\n  source,\n  delayMs,\n  signal,\n  abortUpstream,\n}: {\n  source: AsyncIterable<SmoothItem<TEmit>>;\n  delayMs: number;\n  signal?: AbortSignal;\n  abortUpstream?: () => void;\n}): AsyncGenerator<TEmit> {\n  if (!(delayMs > 0)) {\n    /** Disabled smoothing preserves fully lazy streaming: no background\n     * producer, no read-ahead — each provider chunk is pulled only when the\n     * consumer asks, exactly like the pre-engine pass-through paths. */\n    for await (const item of source) {\n      if (isSignalAborted(signal)) {\n        abortUpstream?.();\n        throw new Error(STREAM_ABORT_MESSAGE);\n      }\n      yield item.emit({ text: item.text, isFirst: true, isLast: true });\n    }\n    return;\n  }\n\n  const queuedItems: QueuedSmoothItem<TEmit>[] = [];\n  const producerState: ProducerState = { done: false, failed: false };\n  let queuedItemIndex = 0;\n  let bufferedTextLength = 0;\n  let consumerClosed = false;\n  let notifyConsumer: (() => void) | undefined;\n  let notifyProducer: (() => void) | undefined;\n\n  const notifyConsumerForItem = (): void => {\n    notifyConsumer?.();\n    notifyConsumer = undefined;\n  };\n\n  const notifyProducerForSpace = (): void => {\n    notifyProducer?.();\n    notifyProducer = undefined;\n  };\n\n  const hasQueuedItems = (): boolean => queuedItemIndex < queuedItems.length;\n\n  const getQueuedItemCount = (): number =>\n    queuedItems.length - queuedItemIndex;\n\n  const isQueueAtCapacity = (): boolean =>\n    getQueuedItemCount() >= MAX_STREAM_QUEUE_CHUNKS ||\n    bufferedTextLength >= MAX_STREAM_QUEUE_TEXT_CHARS;\n\n  /** Abort-aware: a consumer parked on an empty queue must wake when the\n   * signal fires even if the provider stream never honors the abort — the\n   * loop's top-of-iteration check then throws the canonical error. */\n  const waitForNextItem = async (): Promise<void> => {\n    if (\n      hasQueuedItems() ||\n      producerState.done ||\n      producerState.failed ||\n      isSignalAborted(signal)\n    ) {\n      return;\n    }\n    await new Promise<void>((resolve) => {\n      const onAbort = (): void => {\n        signal?.removeEventListener('abort', onAbort);\n        resolve();\n      };\n      notifyConsumer = (): void => {\n        signal?.removeEventListener('abort', onAbort);\n        resolve();\n      };\n      signal?.addEventListener('abort', onAbort, { once: true });\n      if (isSignalAborted(signal)) {\n        onAbort();\n      }\n    });\n  };\n\n  const waitForQueueSpace = async (): Promise<void> => {\n    while (\n      isQueueAtCapacity() &&\n      !consumerClosed &&\n      !isSignalAborted(signal)\n    ) {\n      await new Promise<void>((resolve) => {\n        const onAbort = (): void => {\n          signal?.removeEventListener('abort', onAbort);\n          resolve();\n        };\n        const onSpace = (): void => {\n          signal?.removeEventListener('abort', onAbort);\n          resolve();\n        };\n        notifyProducer = onSpace;\n        signal?.addEventListener('abort', onAbort, { once: true });\n        if (isSignalAborted(signal)) {\n          onAbort();\n        }\n      });\n    }\n  };\n\n  const dequeue = (): QueuedSmoothItem<TEmit> | undefined => {\n    if (!hasQueuedItems()) {\n      return undefined;\n    }\n    const queuedItem = queuedItems[queuedItemIndex];\n    queuedItemIndex++;\n    if (queuedItemIndex > 128 && queuedItemIndex * 2 >= queuedItems.length) {\n      queuedItems.splice(0, queuedItemIndex);\n      queuedItemIndex = 0;\n    }\n    return queuedItem;\n  };\n\n  const throwAborted = (): never => {\n    abortUpstream?.();\n    throw new Error(STREAM_ABORT_MESSAGE);\n  };\n\n  const enqueue = async (item: SmoothItem<TEmit>): Promise<void> => {\n    await waitForQueueSpace();\n    if (consumerClosed || isSignalAborted(signal)) {\n      throwAborted();\n    }\n    const textLength = item.smooth ? item.text.length : 0;\n    queuedItems.push({ item, textLength });\n    bufferedTextLength += textLength;\n    notifyConsumerForItem();\n  };\n\n  /**\n   * Oversized splittable items are segmented at admission so a single giant\n   * provider chunk cannot blow past the text budget: each segment re-checks\n   * capacity, so the producer parks mid-chunk once the buffer fills — the\n   * same bound the legacy split-before-enqueue queues enforced. The wrapped\n   * emit maps segment-local pieces back to chunk-global isFirst/isLast so\n   * provider clone contracts are unaffected.\n   */\n  const enqueueSegmented = async (item: SmoothItem<TEmit>): Promise<void> => {\n    if (\n      !item.smooth ||\n      item.atomic === true ||\n      item.text.length <= MAX_SMOOTH_ITEM_SEGMENT_CHARS\n    ) {\n      await enqueue(item);\n      return;\n    }\n\n    const segments: { start: number; end: number }[] = [];\n    let offset = 0;\n    while (offset < item.text.length) {\n      const end =\n        offset +\n        findStreamChunkBoundary(\n          item.text.slice(offset),\n          MAX_SMOOTH_ITEM_SEGMENT_CHARS\n        );\n      segments.push({ start: offset, end });\n      offset = end;\n    }\n\n    for (let i = 0; i < segments.length; i++) {\n      const isFirstSegment = i === 0;\n      const isLastSegment = i === segments.length - 1;\n      await enqueue({\n        text: item.text.slice(segments[i].start, segments[i].end),\n        smooth: true,\n        emit: (piece) =>\n          item.emit({\n            text: piece.text,\n            isFirst: isFirstSegment && piece.isFirst,\n            isLast: isLastSegment && piece.isLast,\n          }),\n      });\n    }\n  };\n\n  const producer = (async (): Promise<void> => {\n    try {\n      for await (const item of source) {\n        if (isSignalAborted(signal)) {\n          throwAborted();\n        }\n        await enqueueSegmented(item);\n      }\n    } catch (error) {\n      producerState.failed = true;\n      producerState.error = error;\n    } finally {\n      producerState.done = true;\n      notifyConsumerForItem();\n    }\n  })();\n\n  let hasEmittedText = false;\n  let lastVisibleTextAt: number | undefined;\n  let drainTicksRemaining: number | undefined;\n  let current: QueuedSmoothItem<TEmit> | undefined;\n  let headOffset = 0;\n  let keepStreaming = true;\n  try {\n    while (keepStreaming) {\n      if (isSignalAborted(signal)) {\n        throwAborted();\n      }\n\n      if (current == null) {\n        await waitForNextItem();\n        current = dequeue();\n        headOffset = 0;\n      }\n\n      if (current == null) {\n        if (producerState.failed) {\n          throw producerState.error ?? new Error(STREAM_PRODUCER_FAILURE);\n        }\n        if (producerState.done) {\n          keepStreaming = false;\n        }\n        continue;\n      }\n\n      const { item } = current;\n\n      if (!item.smooth) {\n        notifyProducerForSpace();\n        current = undefined;\n        yield item.emit({ text: item.text, isFirst: true, isLast: true });\n        continue;\n      }\n\n      if (item.text === '') {\n        bufferedTextLength = Math.max(\n          0,\n          bufferedTextLength - current.textLength\n        );\n        notifyProducerForSpace();\n        current = undefined;\n        continue;\n      }\n\n      /** Once the producer is done the backlog is final: drain it linearly\n       * across the remaining target window instead of letting the\n       * proportional formula decay geometrically and stretch the tail. */\n      if (producerState.done && drainTicksRemaining == null) {\n        drainTicksRemaining = Math.max(\n          1,\n          Math.floor(SMOOTH_TARGET_LATENCY_MS / delayMs)\n        );\n      }\n      const tickBudget =\n        drainTicksRemaining != null\n          ? Math.max(\n            STREAM_CHUNK_MIN_SIZE,\n            Math.ceil(bufferedTextLength / drainTicksRemaining)\n          )\n          : computeAdaptivePieceSize(bufferedTextLength, delayMs);\n      if (drainTicksRemaining != null && drainTicksRemaining > 1) {\n        drainTicksRemaining -= 1;\n      }\n\n      await waitForStreamDelay(\n        getCadencedStreamDelay({\n          targetDelay: hasEmittedText ? delayMs : 0,\n          lastVisibleTextAt,\n          now: Date.now(),\n        }),\n        signal\n      );\n      if (isSignalAborted(signal)) {\n        throwAborted();\n      }\n      hasEmittedText = true;\n      lastVisibleTextAt = Date.now();\n\n      if (item.atomic === true) {\n        bufferedTextLength = Math.max(\n          0,\n          bufferedTextLength - current.textLength\n        );\n        notifyProducerForSpace();\n        current = undefined;\n        yield item.emit({ text: item.text, isFirst: true, isLast: true });\n        continue;\n      }\n\n      /** One cadence tick drains up to the adaptive budget ACROSS queued\n       * items, so token-sized provider deltas coalesce instead of costing a\n       * full tick each; passthrough items flush free mid-batch (FIFO), and\n       * atomic items end the batch to take their own tick. */\n      let consumed = 0;\n      while (consumed < tickBudget) {\n        if (isSignalAborted(signal)) {\n          throwAborted();\n        }\n        if (current == null) {\n          if (!hasQueuedItems()) {\n            break;\n          }\n          current = dequeue();\n          headOffset = 0;\n          if (current == null) {\n            break;\n          }\n        }\n\n        const batchItem = current.item;\n        if (!batchItem.smooth) {\n          notifyProducerForSpace();\n          current = undefined;\n          yield batchItem.emit({\n            text: batchItem.text,\n            isFirst: true,\n            isLast: true,\n          });\n          continue;\n        }\n        if (batchItem.text === '') {\n          bufferedTextLength = Math.max(\n            0,\n            bufferedTextLength - current.textLength\n          );\n          notifyProducerForSpace();\n          current = undefined;\n          continue;\n        }\n        if (batchItem.atomic === true) {\n          break;\n        }\n\n        const remainingText = batchItem.text.slice(headOffset);\n        const pieceLength = findStreamChunkBoundary(\n          remainingText,\n          tickBudget - consumed\n        );\n        const pieceEnd = headOffset + pieceLength;\n        const piece = batchItem.text.slice(headOffset, pieceEnd);\n        const isFirst = headOffset === 0;\n        const isLast = pieceEnd === batchItem.text.length;\n\n        bufferedTextLength = Math.max(0, bufferedTextLength - piece.length);\n        notifyProducerForSpace();\n        if (isLast) {\n          current = undefined;\n        } else {\n          headOffset = pieceEnd;\n        }\n        consumed += piece.length;\n        yield batchItem.emit({ text: piece, isFirst, isLast });\n      }\n    }\n  } finally {\n    consumerClosed = true;\n    if (producerState.done) {\n      await producer;\n    } else {\n      abortUpstream?.();\n      notifyProducerForSpace();\n      const iterator = source as Partial<AsyncGenerator<SmoothItem<TEmit>>>;\n      const closing = iterator.return?.call(source, undefined as never);\n      if (closing != null) {\n        void closing.then(\n          () => undefined,\n          () => undefined\n        );\n      }\n      await Promise.race([\n        producer,\n        new Promise<void>((resolve) => {\n          const timeout = setTimeout(resolve, PRODUCER_CLOSE_GRACE_MS);\n          timeout.unref();\n        }),\n      ]);\n    }\n  }\n}\n"],"mappings":";AAAA,MAAa,uBAAuB;AAIpC,MAAa,gCAAgC;AAE7C,MAAa,oBAAyC,IAAI,IAAI;CAC5D;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,MAAa,uBAAuB;;;;;;;;AAUpC,MAAa,0BAA0B;;;;;;;AAQvC,SAAgB,mBAAmB,OAAwB;CACzD,IAAI,SAAS,QAAQ,CAAC,OAAO,SAAS,KAAK,GACzC,OAAA;CAEF,OAAO,KAAK,IAAI,GAAG,KAAK;AAC1B;AAEA,SAAgB,gBAAgB,QAA+B;CAC7D,OAAO,QAAQ,YAAY;AAC7B;AAWA,SAAgB,wBACd,MACA,SACQ;CACR,IAAI,WAAW,KAAK,QAClB,OAAO,KAAK;CAGd,MAAM,UAAU,KAAK,IACnB,KAAK,QACL,UAAA,EACF;CACA,KAAK,IAAI,WAAW,SAAS,WAAW,SAAS,YAC/C,IAAI,kBAAkB,IAAI,KAAK,SAAS,GACtC,OAAO,WAAW;CAItB,OAAO;AACT;;;;;;;AAQA,SAAgB,yBACd,oBACA,QACA,kBAAA,KACQ;CACR,IAAI,sBAAsB,GACxB,OAAA;CAEF,IAAI,UAAU,KAAK,mBAAmB,GACpC,OAAO;CAET,OAAO,KAAK,IAAA,GAEV,KAAK,KAAM,qBAAqB,SAAU,eAAe,CAC3D;AACF;;;;;;AAOA,SAAgB,uBAAuB,EACrC,aACA,mBACA,OAKS;CACT,IAAI,eAAe,KAAK,qBAAqB,MAC3C,OAAO;CAET,OAAO,KAAK,IAAI,GAAG,eAAe,MAAM,kBAAkB;AAC5D;;AAGA,eAAsB,mBACpB,OACA,QACe;CACf,IAAI,SAAS,KAAK,gBAAgB,MAAM,GACtC;CAEF,MAAM,IAAI,SAAe,YAAY;EACnC,MAAM,aAA0D,CAAC;EACjE,MAAM,gBAAsB;GAC1B,IAAI,WAAW,SACb,aAAa,WAAW,OAAO;GAEjC,QAAQ,oBAAoB,SAAS,OAAO;GAC5C,QAAQ;EACV;EACA,WAAW,UAAU,iBAAiB;GACpC,QAAQ,oBAAoB,SAAS,OAAO;GAC5C,QAAQ;EACV,GAAG,KAAK;EACR,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;EACzD,IAAI,gBAAgB,MAAM,GACxB,QAAQ;CAEZ,CAAC;AACH;;;;;;;;;;;;;AAmDA,gBAAuB,aAAoB,EACzC,QACA,SACA,QACA,iBAMwB;CACxB,IAAI,EAAE,UAAU,IAAI;;;;EAIlB,WAAW,MAAM,QAAQ,QAAQ;GAC/B,IAAI,gBAAgB,MAAM,GAAG;IAC3B,gBAAgB;IAChB,MAAM,IAAI,MAAM,oBAAoB;GACtC;GACA,MAAM,KAAK,KAAK;IAAE,MAAM,KAAK;IAAM,SAAS;IAAM,QAAQ;GAAK,CAAC;EAClE;EACA;CACF;CAEA,MAAM,cAAyC,CAAC;CAChD,MAAM,gBAA+B;EAAE,MAAM;EAAO,QAAQ;CAAM;CAClE,IAAI,kBAAkB;CACtB,IAAI,qBAAqB;CACzB,IAAI,iBAAiB;CACrB,IAAI;CACJ,IAAI;CAEJ,MAAM,8BAAoC;EACxC,iBAAiB;EACjB,iBAAiB,KAAA;CACnB;CAEA,MAAM,+BAAqC;EACzC,iBAAiB;EACjB,iBAAiB,KAAA;CACnB;CAEA,MAAM,uBAAgC,kBAAkB,YAAY;CAEpE,MAAM,2BACJ,YAAY,SAAS;CAEvB,MAAM,0BACJ,mBAAmB,KAAA,OACnB,sBAAA;;;;CAKF,MAAM,kBAAkB,YAA2B;EACjD,IACE,eAAe,KACf,cAAc,QACd,cAAc,UACd,gBAAgB,MAAM,GAEtB;EAEF,MAAM,IAAI,SAAe,YAAY;GACnC,MAAM,gBAAsB;IAC1B,QAAQ,oBAAoB,SAAS,OAAO;IAC5C,QAAQ;GACV;GACA,uBAA6B;IAC3B,QAAQ,oBAAoB,SAAS,OAAO;IAC5C,QAAQ;GACV;GACA,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;GACzD,IAAI,gBAAgB,MAAM,GACxB,QAAQ;EAEZ,CAAC;CACH;CAEA,MAAM,oBAAoB,YAA2B;EACnD,OACE,kBAAkB,KAClB,CAAC,kBACD,CAAC,gBAAgB,MAAM,GAEvB,MAAM,IAAI,SAAe,YAAY;GACnC,MAAM,gBAAsB;IAC1B,QAAQ,oBAAoB,SAAS,OAAO;IAC5C,QAAQ;GACV;GACA,MAAM,gBAAsB;IAC1B,QAAQ,oBAAoB,SAAS,OAAO;IAC5C,QAAQ;GACV;GACA,iBAAiB;GACjB,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;GACzD,IAAI,gBAAgB,MAAM,GACxB,QAAQ;EAEZ,CAAC;CAEL;CAEA,MAAM,gBAAqD;EACzD,IAAI,CAAC,eAAe,GAClB;EAEF,MAAM,aAAa,YAAY;EAC/B;EACA,IAAI,kBAAkB,OAAO,kBAAkB,KAAK,YAAY,QAAQ;GACtE,YAAY,OAAO,GAAG,eAAe;GACrC,kBAAkB;EACpB;EACA,OAAO;CACT;CAEA,MAAM,qBAA4B;EAChC,gBAAgB;EAChB,MAAM,IAAI,MAAM,oBAAoB;CACtC;CAEA,MAAM,UAAU,OAAO,SAA2C;EAChE,MAAM,kBAAkB;EACxB,IAAI,kBAAkB,gBAAgB,MAAM,GAC1C,aAAa;EAEf,MAAM,aAAa,KAAK,SAAS,KAAK,KAAK,SAAS;EACpD,YAAY,KAAK;GAAE;GAAM;EAAW,CAAC;EACrC,sBAAsB;EACtB,sBAAsB;CACxB;;;;;;;;;CAUA,MAAM,mBAAmB,OAAO,SAA2C;EACzE,IACE,CAAC,KAAK,UACN,KAAK,WAAW,QAChB,KAAK,KAAK,UAAA,MACV;GACA,MAAM,QAAQ,IAAI;GAClB;EACF;EAEA,MAAM,WAA6C,CAAC;EACpD,IAAI,SAAS;EACb,OAAO,SAAS,KAAK,KAAK,QAAQ;GAChC,MAAM,MACJ,SACA,wBACE,KAAK,KAAK,MAAM,MAAM,GACtB,6BACF;GACF,SAAS,KAAK;IAAE,OAAO;IAAQ;GAAI,CAAC;GACpC,SAAS;EACX;EAEA,KAAK,IAAI,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK;GACxC,MAAM,iBAAiB,MAAM;GAC7B,MAAM,gBAAgB,MAAM,SAAS,SAAS;GAC9C,MAAM,QAAQ;IACZ,MAAM,KAAK,KAAK,MAAM,SAAS,EAAE,CAAC,OAAO,SAAS,EAAE,CAAC,GAAG;IACxD,QAAQ;IACR,OAAO,UACL,KAAK,KAAK;KACR,MAAM,MAAM;KACZ,SAAS,kBAAkB,MAAM;KACjC,QAAQ,iBAAiB,MAAM;IACjC,CAAC;GACL,CAAC;EACH;CACF;CAEA,MAAM,YAAY,YAA2B;EAC3C,IAAI;GACF,WAAW,MAAM,QAAQ,QAAQ;IAC/B,IAAI,gBAAgB,MAAM,GACxB,aAAa;IAEf,MAAM,iBAAiB,IAAI;GAC7B;EACF,SAAS,OAAO;GACd,cAAc,SAAS;GACvB,cAAc,QAAQ;EACxB,UAAU;GACR,cAAc,OAAO;GACrB,sBAAsB;EACxB;CACF,EAAA,CAAG;CAEH,IAAI,iBAAiB;CACrB,IAAI;CACJ,IAAI;CACJ,IAAI;CACJ,IAAI,aAAa;CACjB,IAAI,gBAAgB;CACpB,IAAI;EACF,OAAO,eAAe;GACpB,IAAI,gBAAgB,MAAM,GACxB,aAAa;GAGf,IAAI,WAAW,MAAM;IACnB,MAAM,gBAAgB;IACtB,UAAU,QAAQ;IAClB,aAAa;GACf;GAEA,IAAI,WAAW,MAAM;IACnB,IAAI,cAAc,QAChB,MAAM,cAAc,yBAAS,IAAI,MAAA,yBAA6B;IAEhE,IAAI,cAAc,MAChB,gBAAgB;IAElB;GACF;GAEA,MAAM,EAAE,SAAS;GAEjB,IAAI,CAAC,KAAK,QAAQ;IAChB,uBAAuB;IACvB,UAAU,KAAA;IACV,MAAM,KAAK,KAAK;KAAE,MAAM,KAAK;KAAM,SAAS;KAAM,QAAQ;IAAK,CAAC;IAChE;GACF;GAEA,IAAI,KAAK,SAAS,IAAI;IACpB,qBAAqB,KAAK,IACxB,GACA,qBAAqB,QAAQ,UAC/B;IACA,uBAAuB;IACvB,UAAU,KAAA;IACV;GACF;;;;GAKA,IAAI,cAAc,QAAQ,uBAAuB,MAC/C,sBAAsB,KAAK,IACzB,GACA,KAAK,MAAA,MAAiC,OAAO,CAC/C;GAEF,MAAM,aACJ,uBAAuB,OACnB,KAAK,IAAA,GAEL,KAAK,KAAK,qBAAqB,mBAAmB,CACpD,IACE,yBAAyB,oBAAoB,OAAO;GAC1D,IAAI,uBAAuB,QAAQ,sBAAsB,GACvD,uBAAuB;GAGzB,MAAM,mBACJ,uBAAuB;IACrB,aAAa,iBAAiB,UAAU;IACxC;IACA,KAAK,KAAK,IAAI;GAChB,CAAC,GACD,MACF;GACA,IAAI,gBAAgB,MAAM,GACxB,aAAa;GAEf,iBAAiB;GACjB,oBAAoB,KAAK,IAAI;GAE7B,IAAI,KAAK,WAAW,MAAM;IACxB,qBAAqB,KAAK,IACxB,GACA,qBAAqB,QAAQ,UAC/B;IACA,uBAAuB;IACvB,UAAU,KAAA;IACV,MAAM,KAAK,KAAK;KAAE,MAAM,KAAK;KAAM,SAAS;KAAM,QAAQ;IAAK,CAAC;IAChE;GACF;;;;;GAMA,IAAI,WAAW;GACf,OAAO,WAAW,YAAY;IAC5B,IAAI,gBAAgB,MAAM,GACxB,aAAa;IAEf,IAAI,WAAW,MAAM;KACnB,IAAI,CAAC,eAAe,GAClB;KAEF,UAAU,QAAQ;KAClB,aAAa;KACb,IAAI,WAAW,MACb;IAEJ;IAEA,MAAM,YAAY,QAAQ;IAC1B,IAAI,CAAC,UAAU,QAAQ;KACrB,uBAAuB;KACvB,UAAU,KAAA;KACV,MAAM,UAAU,KAAK;MACnB,MAAM,UAAU;MAChB,SAAS;MACT,QAAQ;KACV,CAAC;KACD;IACF;IACA,IAAI,UAAU,SAAS,IAAI;KACzB,qBAAqB,KAAK,IACxB,GACA,qBAAqB,QAAQ,UAC/B;KACA,uBAAuB;KACvB,UAAU,KAAA;KACV;IACF;IACA,IAAI,UAAU,WAAW,MACvB;IAIF,MAAM,cAAc,wBADE,UAAU,KAAK,MAAM,UAE7B,GACZ,aAAa,QACf;IACA,MAAM,WAAW,aAAa;IAC9B,MAAM,QAAQ,UAAU,KAAK,MAAM,YAAY,QAAQ;IACvD,MAAM,UAAU,eAAe;IAC/B,MAAM,SAAS,aAAa,UAAU,KAAK;IAE3C,qBAAqB,KAAK,IAAI,GAAG,qBAAqB,MAAM,MAAM;IAClE,uBAAuB;IACvB,IAAI,QACF,UAAU,KAAA;SAEV,aAAa;IAEf,YAAY,MAAM;IAClB,MAAM,UAAU,KAAK;KAAE,MAAM;KAAO;KAAS;IAAO,CAAC;GACvD;EACF;CACF,UAAU;EACR,iBAAiB;EACjB,IAAI,cAAc,MAChB,MAAM;OACD;GACL,gBAAgB;GAChB,uBAAuB;GAEvB,MAAM,UAAUA,OAAS,QAAQ,KAAK,QAAQ,KAAA,CAAkB;GAChE,IAAI,WAAW,MACb,QAAa,WACL,KAAA,SACA,KAAA,CACR;GAEF,MAAM,QAAQ,KAAK,CACjB,UACA,IAAI,SAAe,YAAY;IAE7B,WAD2B,SAAS,uBAC9B,CAAC,CAAC,MAAM;GAChB,CAAC,CACH,CAAC;EACH;CACF;AACF"}

package/dist/esm/llm/vertexai/index.mjs

@@ -1,4 +1,6 @@
 import { GOOGLE_STREAMED_TOOL_CALL_ADAPTER, STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY, STREAMED_TOOL_CALL_SEAL_METADATA_KEY } from "../../tools/streamedToolCallSeals.mjs";
+import { resolveStreamDelay } from "../stream/smoother.mjs";
+import { smoothGenerationChunks } from "../stream/chunkAdapters.mjs";
 import { AIMessageChunk, isAIMessage } from "@langchain/core/messages";
 import { ChatGoogle } from "@langchain/google-gauth";
 import { ChatConnection } from "@langchain/google-common";
@@ -396,6 +398,7 @@ var ChatVertexAI = class extends ChatGoogle {
 		"chat_models",
 		"vertexai"
 	];
+	_lc_stream_delay;
 	dynamicThinkingBudget = false;
 	thinkingConfig;
 	static lc_name() {
@@ -413,6 +416,7 @@ var ChatVertexAI = class extends ChatGoogle {
 		});
 		this.dynamicThinkingBudget = dynamicThinkingBudget;
 		this.thinkingConfig = fields?.thinkingConfig;
+		this._lc_stream_delay = resolveStreamDelay(fields?._lc_stream_delay);
 	}
 	invocationParams(options) {
 		const params = super.invocationParams(options);
@@ -420,8 +424,16 @@ var ChatVertexAI = class extends ChatGoogle {
 		return params;
 	}
 	async *_streamResponseChunks(messages, options, runManager) {
+		yield* smoothGenerationChunks({
+			chunks: this._streamRepairedChunks(messages, options),
+			delayMs: this._lc_stream_delay,
+			signal: options.signal,
+			runManager
+		});
+	}
+	async *_streamRepairedChunks(messages, options) {
 		let lastGoodUsage;
-		for await (const chunk of super._streamResponseChunks(messages, options, runManager)) {
+		for await (const chunk of super._streamResponseChunks(messages, options, void 0)) {
 			const genUsage = chunk.generationInfo?.usage_metadata;
 			if (genUsage) lastGoodUsage = genUsage;
 			if (chunk.message instanceof AIMessageChunk) {

package/dist/esm/llm/vertexai/index.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.mjs","names":[],"sources":["../../../../src/llm/vertexai/index.ts"],"sourcesContent":["import { ChatGoogle } from '@langchain/google-gauth';\nimport { ChatConnection } from '@langchain/google-common';\nimport { AIMessageChunk, isAIMessage } from '@langchain/core/messages';\nimport type {\n  GeminiContent,\n  GeminiRequest,\n  GoogleAIModelRequestParams,\n  GoogleAbstractedClient,\n} from '@langchain/google-common';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { GoogleThinkingConfig, VertexAIClientOptions } from '@/types';\nimport {\n  STREAMED_TOOL_CALL_SEAL_METADATA_KEY,\n  STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY,\n  GOOGLE_STREAMED_TOOL_CALL_ADAPTER,\n} from '@/tools/streamedToolCallSeals';\n\n/**\n * `@langchain/google-common`'s `_streamResponseChunks` emits usage on TWO\n * different paths within the same stream:\n *\n *   - Streaming chunks set `chunk.generationInfo.usage_metadata` via\n *     `responseToUsageMetadata`, which correctly sums\n *     `candidatesTokenCount + thoughtsTokenCount` and includes\n *     `output_token_details.reasoning`.\n *   - The trailing fallback chunk (emitted after the API stream exhausts)\n *     attaches its own `chunk.message.usage_metadata` built inline as\n *     `output_tokens = candidatesTokenCount` only — dropping\n *     `thoughtsTokenCount` and `output_token_details` entirely.\n *\n * After `AIMessageChunk.concat`, only `message.usage_metadata` survives —\n * which is the buggy fallback value. This breaks the documented\n * `total_tokens === input_tokens + output_tokens` invariant and silently\n * undercharges thinking models for reasoning tokens.\n *\n * The repair: track the last `generationInfo.usage_metadata` we see, and\n * when the fallback chunk arrives with its buggy `message.usage_metadata`,\n * replace it with the tracked good value. `CustomChatGoogleGenerativeAI`\n * solves the same problem for the Google API path differently — by\n * overriding `_convertToUsageMetadata`.\n */\nexport function repairStreamUsageMetadata(\n  current: UsageMetadata | undefined,\n  generationInfoUsage: UsageMetadata | undefined\n): UsageMetadata | undefined {\n  if (!current) return current;\n  if (!generationInfoUsage) return current;\n  if (generationInfoUsage.total_tokens !== current.total_tokens) return current;\n  if (generationInfoUsage.output_tokens <= current.output_tokens)\n    return current;\n  return generationInfoUsage;\n}\n\n/**\n * The Gemini API delivers function calls as complete objects — never as\n * partial arg deltas. `@langchain/google-common` pre-parses each streamed\n * functionCall part into `tool_calls` (invalid args land in\n * `invalid_tool_calls` instead), so a chunk whose tool-call chunks all parsed\n * cleanly is sealed on arrival for eager tool execution. Anything that fails\n * the parse check is left unstamped and falls back to the lazy path.\n */\nexport function sealCompleteStreamedToolCalls(message: AIMessageChunk): void {\n  const chunkCount = message.tool_call_chunks?.length ?? 0;\n  if (\n    chunkCount === 0 ||\n    (message.invalid_tool_calls?.length ?? 0) > 0 ||\n    (message.tool_calls?.length ?? 0) !== chunkCount\n  ) {\n    return;\n  }\n  message.response_metadata = {\n    ...message.response_metadata,\n    [STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY]:\n      GOOGLE_STREAMED_TOOL_CALL_ADAPTER,\n    [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: { kind: 'all' },\n  };\n}\n\ntype AdditionalKwargs =\n  | undefined\n  | (BaseMessage['additional_kwargs'] & {\n      signatures?: Array<string | undefined>;\n    });\n\n/**\n * Fixes thought signatures on functionCall parts in the formatted Gemini request.\n *\n * `@langchain/google-common` stores signatures as a flat array in\n * `additional_kwargs.signatures` (one per response part) and re-attaches them\n * by index only when `signatures.length === parts.length`. This fails when:\n * - The API omits a signature (length mismatch)\n * - Streaming chunks merge with different part counts\n * - The signature for a functionCall part is an empty string\n *\n * This function correlates each \"model\" content block in the formatted request\n * back to its originating AI message by *position*, then re-attaches non-empty\n * signatures that the library failed to apply. AI messages without signatures\n * still consume their slot — filtering them out shifted later messages onto\n * the wrong content block and dropped real signatures on the floor.\n */\nexport function fixThoughtSignatures(\n  contents: GeminiContent[],\n  input: BaseMessage[]\n): void {\n  // All AI messages, in order — non-signature ones still consume positional\n  // slots so later messages line up with their model content blocks.\n  const aiMessages = input.filter(isAIMessage);\n  const modelContents = contents.filter((c) => c.role === 'model');\n\n  const count = Math.min(aiMessages.length, modelContents.length);\n  for (let i = 0; i < count; i++) {\n    const signatures = (aiMessages[i].additional_kwargs as AdditionalKwargs)\n      ?.signatures;\n    if (!Array.isArray(signatures) || signatures.length === 0) continue;\n\n    const content = modelContents[i];\n    const attachedSignatures = new Set(\n      content.parts\n        .map((p) => p.thoughtSignature)\n        .filter((s): s is string => s != null && s !== '')\n    );\n    const availableSignatures = signatures.filter(\n      (s): s is string => s != null && s !== '' && !attachedSignatures.has(s)\n    );\n\n    let sigIdx = 0;\n    for (const part of content.parts) {\n      if (\n        'functionCall' in part &&\n        (part.thoughtSignature == null || part.thoughtSignature === '') &&\n        sigIdx < availableSignatures.length\n      ) {\n        part.thoughtSignature = availableSignatures[sigIdx];\n        sigIdx++;\n      }\n    }\n  }\n}\n\nclass CustomChatConnection extends ChatConnection<VertexAIClientOptions> {\n  thinkingConfig?: GoogleThinkingConfig;\n\n  async formatData(\n    input: BaseMessage[],\n    parameters: GoogleAIModelRequestParams\n  ): Promise<unknown> {\n    const formattedData = (await super.formatData(\n      input,\n      parameters\n    )) as GeminiRequest;\n    if (formattedData.generationConfig?.thinkingConfig?.thinkingBudget === -1) {\n      // -1 means \"let the model decide\" - delete the property so the API doesn't receive an invalid value\n      if (\n        formattedData.generationConfig.thinkingConfig.includeThoughts === false\n      ) {\n        formattedData.generationConfig.thinkingConfig.includeThoughts = true;\n      }\n      delete formattedData.generationConfig.thinkingConfig.thinkingBudget;\n    }\n    if (this.thinkingConfig?.thinkingLevel != null) {\n      formattedData.generationConfig ??= {};\n      // thinkingLevel and thinkingBudget cannot coexist — the API rejects the request.\n      // Remove thinkingBudget when thinkingLevel is set.\n      const { thinkingBudget: _, ...existingThinkingConfig } =\n        (formattedData.generationConfig.thinkingConfig as\n          | Record<string, unknown>\n          | undefined) ?? {};\n      (\n        formattedData.generationConfig as Record<string, unknown>\n      ).thinkingConfig = {\n        ...existingThinkingConfig,\n        thinkingLevel: this.thinkingConfig.thinkingLevel,\n        ...(this.thinkingConfig.includeThoughts != null && {\n          includeThoughts: this.thinkingConfig.includeThoughts,\n        }),\n      };\n    }\n    if (formattedData.contents) {\n      fixThoughtSignatures(formattedData.contents, input);\n      // gemini-3.1+ models reject role=\"function\"; convert to role=\"user\"\n      for (const content of formattedData.contents) {\n        if (content.role === 'function') {\n          (content as { role: string }).role = 'user';\n        }\n      }\n    }\n    return formattedData;\n  }\n}\n\n/**\n * Integration with Google Vertex AI chat models.\n *\n * Setup:\n * Install `@langchain/google-vertexai` and set your stringified\n * Vertex AI credentials as an environment variable named `GOOGLE_APPLICATION_CREDENTIALS`.\n *\n * ```bash\n * npm install @langchain/google-vertexai\n * export GOOGLE_APPLICATION_CREDENTIALS=\"path/to/credentials\"\n * ```\n *\n * ## [Constructor args](https://api.js.langchain.com/classes/_langchain_google_vertexai.index.ChatVertexAI.html#constructor.new_ChatVertexAI)\n *\n * ## [Runtime args](https://api.js.langchain.com/interfaces/langchain_google_common_types.GoogleAIBaseLanguageModelCallOptions.html)\n *\n * Runtime args can be passed as the second argument to any of the base runnable methods `.invoke`. `.stream`, `.batch`, etc.\n * They can also be passed via `.withConfig`, or the second arg in `.bindTools`, like shown in the examples below:\n *\n * ```typescript\n * // When calling `.withConfig`, call options should be passed via the first argument\n * const llmWithArgsBound = llm.withConfig({\n *   stop: [\"\\n\"],\n *   tools: [...],\n * });\n *\n * // When calling `.bindTools`, call options should be passed via the second argument\n * const llmWithTools = llm.bindTools(\n *   [...],\n *   {\n *     tool_choice: \"auto\",\n *   }\n * );\n * ```\n *\n * ## Examples\n *\n * <details open>\n * <summary><strong>Instantiate</strong></summary>\n *\n * ```typescript\n * import { ChatVertexAI } from '@langchain/google-vertexai';\n *\n * const llm = new ChatVertexAI({\n *   model: \"gemini-1.5-pro\",\n *   temperature: 0,\n *   // other params...\n * });\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Invoking</strong></summary>\n *\n * ```typescript\n * const input = `Translate \"I love programming\" into French.`;\n *\n * // Models also accept a list of chat messages or a formatted prompt\n * const result = await llm.invoke(input);\n * console.log(result);\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"J'adore programmer\\\" \\n\\nHere's why this is the best translation:\\n\\n* **J'adore** means \\\"I love\\\" and conveys a strong passion.\\n* **Programmer** is the French verb for \\\"to program.\\\"\\n\\nThis translation is natural and idiomatic in French. \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 63,\n *     \"total_tokens\": 72\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Streaming Chunks</strong></summary>\n *\n * ```typescript\n * for await (const chunk of await llm.stream(input)) {\n *   console.log(chunk);\n * }\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"J'adore programmer\\\" \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {\n *     \"finishReason\": \"stop\"\n *   },\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 8,\n *     \"total_tokens\": 17\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Aggregate Streamed Chunks</strong></summary>\n *\n * ```typescript\n * import { AIMessageChunk } from '@langchain/core/messages';\n * import { concat } from '@langchain/core/utils/stream';\n *\n * const stream = await llm.stream(input);\n * let full: AIMessageChunk | undefined;\n * for await (const chunk of stream) {\n *   full = !full ? chunk : concat(full, chunk);\n * }\n * console.log(full);\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"J'adore programmer\\\" \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {\n *     \"finishReason\": \"stop\"\n *   },\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 8,\n *     \"total_tokens\": 17\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Bind tools</strong></summary>\n *\n * ```typescript\n * import { z } from 'zod';\n *\n * const GetWeather = {\n *   name: \"GetWeather\",\n *   description: \"Get the current weather in a given location\",\n *   schema: z.object({\n *     location: z.string().describe(\"The city and state, e.g. San Francisco, CA\")\n *   }),\n * }\n *\n * const GetPopulation = {\n *   name: \"GetPopulation\",\n *   description: \"Get the current population in a given location\",\n *   schema: z.object({\n *     location: z.string().describe(\"The city and state, e.g. San Francisco, CA\")\n *   }),\n * }\n *\n * const llmWithTools = llm.bindTools([GetWeather, GetPopulation]);\n * const aiMsg = await llmWithTools.invoke(\n *   \"Which city is hotter today and which is bigger: LA or NY?\"\n * );\n * console.log(aiMsg.tool_calls);\n * ```\n *\n * ```txt\n * [\n *   {\n *     name: 'GetPopulation',\n *     args: { location: 'New York City, NY' },\n *     id: '33c1c1f47e2f492799c77d2800a43912',\n *     type: 'tool_call'\n *   }\n * ]\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Structured Output</strong></summary>\n *\n * ```typescript\n * import { z } from 'zod';\n *\n * const Joke = z.object({\n *   setup: z.string().describe(\"The setup of the joke\"),\n *   punchline: z.string().describe(\"The punchline to the joke\"),\n *   rating: z.number().optional().describe(\"How funny the joke is, from 1 to 10\")\n * }).describe('Joke to tell user.');\n *\n * const structuredLlm = llm.withStructuredOutput(Joke, { name: \"Joke\" });\n * const jokeResult = await structuredLlm.invoke(\"Tell me a joke about cats\");\n * console.log(jokeResult);\n * ```\n *\n * ```txt\n * {\n *   setup: 'What do you call a cat that loves to bowl?',\n *   punchline: 'An alley cat!'\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Usage Metadata</strong></summary>\n *\n * ```typescript\n * const aiMsgForMetadata = await llm.invoke(input);\n * console.log(aiMsgForMetadata.usage_metadata);\n * ```\n *\n * ```txt\n * { input_tokens: 9, output_tokens: 8, total_tokens: 17 }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Stream Usage Metadata</strong></summary>\n *\n * ```typescript\n * const streamForMetadata = await llm.stream(\n *   input,\n *   {\n *     streamUsage: true\n *   }\n * );\n * let fullForMetadata: AIMessageChunk | undefined;\n * for await (const chunk of streamForMetadata) {\n *   fullForMetadata = !fullForMetadata ? chunk : concat(fullForMetadata, chunk);\n * }\n * console.log(fullForMetadata?.usage_metadata);\n * ```\n *\n * ```txt\n * { input_tokens: 9, output_tokens: 8, total_tokens: 17 }\n * ```\n * </details>\n *\n * <br />\n */\nexport class ChatVertexAI extends ChatGoogle {\n  lc_namespace = ['langchain', 'chat_models', 'vertexai'];\n  dynamicThinkingBudget = false;\n  thinkingConfig?: GoogleThinkingConfig;\n\n  static lc_name(): 'LibreChatVertexAI' {\n    return 'LibreChatVertexAI';\n  }\n\n  constructor(model: string, fields?: Omit<VertexAIClientOptions, 'model'>);\n  constructor(fields?: VertexAIClientOptions);\n  constructor(\n    modelOrFields?: string | VertexAIClientOptions,\n    params?: Omit<VertexAIClientOptions, 'model'>\n  ) {\n    const fields =\n      typeof modelOrFields === 'string'\n        ? { ...(params ?? {}), model: modelOrFields }\n        : modelOrFields;\n    const dynamicThinkingBudget = fields?.thinkingBudget === -1;\n    super({\n      ...fields,\n      platformType: 'gcp',\n    });\n    this.dynamicThinkingBudget = dynamicThinkingBudget;\n    this.thinkingConfig = fields?.thinkingConfig;\n  }\n  invocationParams(\n    options?: this['ParsedCallOptions'] | undefined\n  ): GoogleAIModelRequestParams {\n    const params = super.invocationParams(options);\n    if (this.dynamicThinkingBudget) {\n      params.maxReasoningTokens = -1;\n    }\n    return params;\n  }\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    let lastGoodUsage: UsageMetadata | undefined;\n    for await (const chunk of super._streamResponseChunks(\n      messages,\n      options,\n      runManager\n    )) {\n      const genUsage = (\n        chunk.generationInfo as { usage_metadata?: UsageMetadata } | undefined\n      )?.usage_metadata;\n      if (genUsage) {\n        lastGoodUsage = genUsage;\n      }\n      if (chunk.message instanceof AIMessageChunk) {\n        const repaired = repairStreamUsageMetadata(\n          chunk.message.usage_metadata,\n          lastGoodUsage\n        );\n        if (repaired !== chunk.message.usage_metadata) {\n          chunk.message.usage_metadata = repaired;\n        }\n        sealCompleteStreamedToolCalls(chunk.message);\n      }\n      yield chunk;\n    }\n  }\n  buildConnection(\n    fields: VertexAIClientOptions | undefined,\n    client: GoogleAbstractedClient\n  ): void {\n    // Note: buildConnection is called from super() BEFORE this.thinkingConfig is set,\n    // so we must read thinkingConfig from `fields` directly.\n    const thinkingConfig = fields?.thinkingConfig ?? this.thinkingConfig;\n\n    const connection = new CustomChatConnection(\n      { ...fields, ...this },\n      this.caller,\n      client,\n      false\n    );\n    connection.thinkingConfig = thinkingConfig;\n    this.connection = connection;\n\n    const streamedConnection = new CustomChatConnection(\n      { ...fields, ...this },\n      this.caller,\n      client,\n      true\n    );\n    streamedConnection.thinkingConfig = thinkingConfig;\n    this.streamedConnection = streamedConnection;\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2CA,SAAgB,0BACd,SACA,qBAC2B;CAC3B,IAAI,CAAC,SAAS,OAAO;CACrB,IAAI,CAAC,qBAAqB,OAAO;CACjC,IAAI,oBAAoB,iBAAiB,QAAQ,cAAc,OAAO;CACtE,IAAI,oBAAoB,iBAAiB,QAAQ,eAC/C,OAAO;CACT,OAAO;AACT;;;;;;;;;AAUA,SAAgB,8BAA8B,SAA+B;CAC3E,MAAM,aAAa,QAAQ,kBAAkB,UAAU;CACvD,IACE,eAAe,MACd,QAAQ,oBAAoB,UAAU,KAAK,MAC3C,QAAQ,YAAY,UAAU,OAAO,YAEtC;CAEF,QAAQ,oBAAoB;EAC1B,GAAG,QAAQ;GACV,0CACC;GACD,uCAAuC,EAAE,MAAM,MAAM;CACxD;AACF;;;;;;;;;;;;;;;;;AAwBA,SAAgB,qBACd,UACA,OACM;CAGN,MAAM,aAAa,MAAM,OAAO,WAAW;CAC3C,MAAM,gBAAgB,SAAS,QAAQ,MAAM,EAAE,SAAS,OAAO;CAE/D,MAAM,QAAQ,KAAK,IAAI,WAAW,QAAQ,cAAc,MAAM;CAC9D,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;EAC9B,MAAM,aAAc,WAAW,EAAE,CAAC,mBAC9B;EACJ,IAAI,CAAC,MAAM,QAAQ,UAAU,KAAK,WAAW,WAAW,GAAG;EAE3D,MAAM,UAAU,cAAc;EAC9B,MAAM,qBAAqB,IAAI,IAC7B,QAAQ,MACL,KAAK,MAAM,EAAE,gBAAgB,CAAC,CAC9B,QAAQ,MAAmB,KAAK,QAAQ,MAAM,EAAE,CACrD;EACA,MAAM,sBAAsB,WAAW,QACpC,MAAmB,KAAK,QAAQ,MAAM,MAAM,CAAC,mBAAmB,IAAI,CAAC,CACxE;EAEA,IAAI,SAAS;EACb,KAAK,MAAM,QAAQ,QAAQ,OACzB,IACE,kBAAkB,SACjB,KAAK,oBAAoB,QAAQ,KAAK,qBAAqB,OAC5D,SAAS,oBAAoB,QAC7B;GACA,KAAK,mBAAmB,oBAAoB;GAC5C;EACF;CAEJ;AACF;AAEA,IAAM,uBAAN,cAAmC,eAAsC;CACvE;CAEA,MAAM,WACJ,OACA,YACkB;EAClB,MAAM,gBAAiB,MAAM,MAAM,WACjC,OACA,UACF;EACA,IAAI,cAAc,kBAAkB,gBAAgB,mBAAmB,IAAI;GAEzE,IACE,cAAc,iBAAiB,eAAe,oBAAoB,OAElE,cAAc,iBAAiB,eAAe,kBAAkB;GAElE,OAAO,cAAc,iBAAiB,eAAe;EACvD;EACA,IAAI,KAAK,gBAAgB,iBAAiB,MAAM;GAC9C,cAAc,qBAAqB,CAAC;GAGpC,MAAM,EAAE,gBAAgB,GAAG,GAAG,2BAC3B,cAAc,iBAAiB,kBAEd,CAAC;GACrB,cACgB,iBACd,iBAAiB;IACjB,GAAG;IACH,eAAe,KAAK,eAAe;IACnC,GAAI,KAAK,eAAe,mBAAmB,QAAQ,EACjD,iBAAiB,KAAK,eAAe,gBACvC;GACF;EACF;EACA,IAAI,cAAc,UAAU;GAC1B,qBAAqB,cAAc,UAAU,KAAK;GAElD,KAAK,MAAM,WAAW,cAAc,UAClC,IAAI,QAAQ,SAAS,YACnB,QAA8B,OAAO;EAG3C;EACA,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4RA,IAAa,eAAb,cAAkC,WAAW;CAC3C,eAAe;EAAC;EAAa;EAAe;CAAU;CACtD,wBAAwB;CACxB;CAEA,OAAO,UAA+B;EACpC,OAAO;CACT;CAIA,YACE,eACA,QACA;EACA,MAAM,SACJ,OAAO,kBAAkB,WACrB;GAAE,GAAI,UAAU,CAAC;GAAI,OAAO;EAAc,IAC1C;EACN,MAAM,wBAAwB,QAAQ,mBAAmB;EACzD,MAAM;GACJ,GAAG;GACH,cAAc;EAChB,CAAC;EACD,KAAK,wBAAwB;EAC7B,KAAK,iBAAiB,QAAQ;CAChC;CACA,iBACE,SAC4B;EAC5B,MAAM,SAAS,MAAM,iBAAiB,OAAO;EAC7C,IAAI,KAAK,uBACP,OAAO,qBAAqB;EAE9B,OAAO;CACT;CACA,OAAO,sBACL,UACA,SACA,YACqC;EACrC,IAAI;EACJ,WAAW,MAAM,SAAS,MAAM,sBAC9B,UACA,SACA,UACF,GAAG;GACD,MAAM,WACJ,MAAM,gBACL;GACH,IAAI,UACF,gBAAgB;GAElB,IAAI,MAAM,mBAAmB,gBAAgB;IAC3C,MAAM,WAAW,0BACf,MAAM,QAAQ,gBACd,aACF;IACA,IAAI,aAAa,MAAM,QAAQ,gBAC7B,MAAM,QAAQ,iBAAiB;IAEjC,8BAA8B,MAAM,OAAO;GAC7C;GACA,MAAM;EACR;CACF;CACA,gBACE,QACA,QACM;EAGN,MAAM,iBAAiB,QAAQ,kBAAkB,KAAK;EAEtD,MAAM,aAAa,IAAI,qBACrB;GAAE,GAAG;GAAQ,GAAG;EAAK,GACrB,KAAK,QACL,QACA,KACF;EACA,WAAW,iBAAiB;EAC5B,KAAK,aAAa;EAElB,MAAM,qBAAqB,IAAI,qBAC7B;GAAE,GAAG;GAAQ,GAAG;EAAK,GACrB,KAAK,QACL,QACA,IACF;EACA,mBAAmB,iBAAiB;EACpC,KAAK,qBAAqB;CAC5B;AACF"}
+{"version":3,"file":"index.mjs","names":[],"sources":["../../../../src/llm/vertexai/index.ts"],"sourcesContent":["import { ChatGoogle } from '@langchain/google-gauth';\nimport { ChatConnection } from '@langchain/google-common';\nimport { AIMessageChunk, isAIMessage } from '@langchain/core/messages';\nimport type {\n  GeminiContent,\n  GeminiRequest,\n  GoogleAIModelRequestParams,\n  GoogleAbstractedClient,\n} from '@langchain/google-common';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { GoogleThinkingConfig, VertexAIClientOptions } from '@/types';\nimport {\n  STREAMED_TOOL_CALL_SEAL_METADATA_KEY,\n  STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY,\n  GOOGLE_STREAMED_TOOL_CALL_ADAPTER,\n} from '@/tools/streamedToolCallSeals';\nimport { smoothGenerationChunks } from '@/llm/stream/chunkAdapters';\nimport { resolveStreamDelay } from '@/llm/stream/smoother';\n\n/**\n * `@langchain/google-common`'s `_streamResponseChunks` emits usage on TWO\n * different paths within the same stream:\n *\n *   - Streaming chunks set `chunk.generationInfo.usage_metadata` via\n *     `responseToUsageMetadata`, which correctly sums\n *     `candidatesTokenCount + thoughtsTokenCount` and includes\n *     `output_token_details.reasoning`.\n *   - The trailing fallback chunk (emitted after the API stream exhausts)\n *     attaches its own `chunk.message.usage_metadata` built inline as\n *     `output_tokens = candidatesTokenCount` only — dropping\n *     `thoughtsTokenCount` and `output_token_details` entirely.\n *\n * After `AIMessageChunk.concat`, only `message.usage_metadata` survives —\n * which is the buggy fallback value. This breaks the documented\n * `total_tokens === input_tokens + output_tokens` invariant and silently\n * undercharges thinking models for reasoning tokens.\n *\n * The repair: track the last `generationInfo.usage_metadata` we see, and\n * when the fallback chunk arrives with its buggy `message.usage_metadata`,\n * replace it with the tracked good value. `CustomChatGoogleGenerativeAI`\n * solves the same problem for the Google API path differently — by\n * overriding `_convertToUsageMetadata`.\n */\nexport function repairStreamUsageMetadata(\n  current: UsageMetadata | undefined,\n  generationInfoUsage: UsageMetadata | undefined\n): UsageMetadata | undefined {\n  if (!current) return current;\n  if (!generationInfoUsage) return current;\n  if (generationInfoUsage.total_tokens !== current.total_tokens) return current;\n  if (generationInfoUsage.output_tokens <= current.output_tokens)\n    return current;\n  return generationInfoUsage;\n}\n\n/**\n * The Gemini API delivers function calls as complete objects — never as\n * partial arg deltas. `@langchain/google-common` pre-parses each streamed\n * functionCall part into `tool_calls` (invalid args land in\n * `invalid_tool_calls` instead), so a chunk whose tool-call chunks all parsed\n * cleanly is sealed on arrival for eager tool execution. Anything that fails\n * the parse check is left unstamped and falls back to the lazy path.\n */\nexport function sealCompleteStreamedToolCalls(message: AIMessageChunk): void {\n  const chunkCount = message.tool_call_chunks?.length ?? 0;\n  if (\n    chunkCount === 0 ||\n    (message.invalid_tool_calls?.length ?? 0) > 0 ||\n    (message.tool_calls?.length ?? 0) !== chunkCount\n  ) {\n    return;\n  }\n  message.response_metadata = {\n    ...message.response_metadata,\n    [STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY]:\n      GOOGLE_STREAMED_TOOL_CALL_ADAPTER,\n    [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: { kind: 'all' },\n  };\n}\n\ntype AdditionalKwargs =\n  | undefined\n  | (BaseMessage['additional_kwargs'] & {\n      signatures?: Array<string | undefined>;\n    });\n\n/**\n * Fixes thought signatures on functionCall parts in the formatted Gemini request.\n *\n * `@langchain/google-common` stores signatures as a flat array in\n * `additional_kwargs.signatures` (one per response part) and re-attaches them\n * by index only when `signatures.length === parts.length`. This fails when:\n * - The API omits a signature (length mismatch)\n * - Streaming chunks merge with different part counts\n * - The signature for a functionCall part is an empty string\n *\n * This function correlates each \"model\" content block in the formatted request\n * back to its originating AI message by *position*, then re-attaches non-empty\n * signatures that the library failed to apply. AI messages without signatures\n * still consume their slot — filtering them out shifted later messages onto\n * the wrong content block and dropped real signatures on the floor.\n */\nexport function fixThoughtSignatures(\n  contents: GeminiContent[],\n  input: BaseMessage[]\n): void {\n  // All AI messages, in order — non-signature ones still consume positional\n  // slots so later messages line up with their model content blocks.\n  const aiMessages = input.filter(isAIMessage);\n  const modelContents = contents.filter((c) => c.role === 'model');\n\n  const count = Math.min(aiMessages.length, modelContents.length);\n  for (let i = 0; i < count; i++) {\n    const signatures = (aiMessages[i].additional_kwargs as AdditionalKwargs)\n      ?.signatures;\n    if (!Array.isArray(signatures) || signatures.length === 0) continue;\n\n    const content = modelContents[i];\n    const attachedSignatures = new Set(\n      content.parts\n        .map((p) => p.thoughtSignature)\n        .filter((s): s is string => s != null && s !== '')\n    );\n    const availableSignatures = signatures.filter(\n      (s): s is string => s != null && s !== '' && !attachedSignatures.has(s)\n    );\n\n    let sigIdx = 0;\n    for (const part of content.parts) {\n      if (\n        'functionCall' in part &&\n        (part.thoughtSignature == null || part.thoughtSignature === '') &&\n        sigIdx < availableSignatures.length\n      ) {\n        part.thoughtSignature = availableSignatures[sigIdx];\n        sigIdx++;\n      }\n    }\n  }\n}\n\nclass CustomChatConnection extends ChatConnection<VertexAIClientOptions> {\n  thinkingConfig?: GoogleThinkingConfig;\n\n  async formatData(\n    input: BaseMessage[],\n    parameters: GoogleAIModelRequestParams\n  ): Promise<unknown> {\n    const formattedData = (await super.formatData(\n      input,\n      parameters\n    )) as GeminiRequest;\n    if (formattedData.generationConfig?.thinkingConfig?.thinkingBudget === -1) {\n      // -1 means \"let the model decide\" - delete the property so the API doesn't receive an invalid value\n      if (\n        formattedData.generationConfig.thinkingConfig.includeThoughts === false\n      ) {\n        formattedData.generationConfig.thinkingConfig.includeThoughts = true;\n      }\n      delete formattedData.generationConfig.thinkingConfig.thinkingBudget;\n    }\n    if (this.thinkingConfig?.thinkingLevel != null) {\n      formattedData.generationConfig ??= {};\n      // thinkingLevel and thinkingBudget cannot coexist — the API rejects the request.\n      // Remove thinkingBudget when thinkingLevel is set.\n      const { thinkingBudget: _, ...existingThinkingConfig } =\n        (formattedData.generationConfig.thinkingConfig as\n          | Record<string, unknown>\n          | undefined) ?? {};\n      (\n        formattedData.generationConfig as Record<string, unknown>\n      ).thinkingConfig = {\n        ...existingThinkingConfig,\n        thinkingLevel: this.thinkingConfig.thinkingLevel,\n        ...(this.thinkingConfig.includeThoughts != null && {\n          includeThoughts: this.thinkingConfig.includeThoughts,\n        }),\n      };\n    }\n    if (formattedData.contents) {\n      fixThoughtSignatures(formattedData.contents, input);\n      // gemini-3.1+ models reject role=\"function\"; convert to role=\"user\"\n      for (const content of formattedData.contents) {\n        if (content.role === 'function') {\n          (content as { role: string }).role = 'user';\n        }\n      }\n    }\n    return formattedData;\n  }\n}\n\n/**\n * Integration with Google Vertex AI chat models.\n *\n * Setup:\n * Install `@langchain/google-vertexai` and set your stringified\n * Vertex AI credentials as an environment variable named `GOOGLE_APPLICATION_CREDENTIALS`.\n *\n * ```bash\n * npm install @langchain/google-vertexai\n * export GOOGLE_APPLICATION_CREDENTIALS=\"path/to/credentials\"\n * ```\n *\n * ## [Constructor args](https://api.js.langchain.com/classes/_langchain_google_vertexai.index.ChatVertexAI.html#constructor.new_ChatVertexAI)\n *\n * ## [Runtime args](https://api.js.langchain.com/interfaces/langchain_google_common_types.GoogleAIBaseLanguageModelCallOptions.html)\n *\n * Runtime args can be passed as the second argument to any of the base runnable methods `.invoke`. `.stream`, `.batch`, etc.\n * They can also be passed via `.withConfig`, or the second arg in `.bindTools`, like shown in the examples below:\n *\n * ```typescript\n * // When calling `.withConfig`, call options should be passed via the first argument\n * const llmWithArgsBound = llm.withConfig({\n *   stop: [\"\\n\"],\n *   tools: [...],\n * });\n *\n * // When calling `.bindTools`, call options should be passed via the second argument\n * const llmWithTools = llm.bindTools(\n *   [...],\n *   {\n *     tool_choice: \"auto\",\n *   }\n * );\n * ```\n *\n * ## Examples\n *\n * <details open>\n * <summary><strong>Instantiate</strong></summary>\n *\n * ```typescript\n * import { ChatVertexAI } from '@langchain/google-vertexai';\n *\n * const llm = new ChatVertexAI({\n *   model: \"gemini-1.5-pro\",\n *   temperature: 0,\n *   // other params...\n * });\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Invoking</strong></summary>\n *\n * ```typescript\n * const input = `Translate \"I love programming\" into French.`;\n *\n * // Models also accept a list of chat messages or a formatted prompt\n * const result = await llm.invoke(input);\n * console.log(result);\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"J'adore programmer\\\" \\n\\nHere's why this is the best translation:\\n\\n* **J'adore** means \\\"I love\\\" and conveys a strong passion.\\n* **Programmer** is the French verb for \\\"to program.\\\"\\n\\nThis translation is natural and idiomatic in French. \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 63,\n *     \"total_tokens\": 72\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Streaming Chunks</strong></summary>\n *\n * ```typescript\n * for await (const chunk of await llm.stream(input)) {\n *   console.log(chunk);\n * }\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"J'adore programmer\\\" \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {},\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": []\n * }\n * AIMessageChunk {\n *   \"content\": \"\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {\n *     \"finishReason\": \"stop\"\n *   },\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 8,\n *     \"total_tokens\": 17\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Aggregate Streamed Chunks</strong></summary>\n *\n * ```typescript\n * import { AIMessageChunk } from '@langchain/core/messages';\n * import { concat } from '@langchain/core/utils/stream';\n *\n * const stream = await llm.stream(input);\n * let full: AIMessageChunk | undefined;\n * for await (const chunk of stream) {\n *   full = !full ? chunk : concat(full, chunk);\n * }\n * console.log(full);\n * ```\n *\n * ```txt\n * AIMessageChunk {\n *   \"content\": \"\\\"J'adore programmer\\\" \\n\",\n *   \"additional_kwargs\": {},\n *   \"response_metadata\": {\n *     \"finishReason\": \"stop\"\n *   },\n *   \"tool_calls\": [],\n *   \"tool_call_chunks\": [],\n *   \"invalid_tool_calls\": [],\n *   \"usage_metadata\": {\n *     \"input_tokens\": 9,\n *     \"output_tokens\": 8,\n *     \"total_tokens\": 17\n *   }\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Bind tools</strong></summary>\n *\n * ```typescript\n * import { z } from 'zod';\n *\n * const GetWeather = {\n *   name: \"GetWeather\",\n *   description: \"Get the current weather in a given location\",\n *   schema: z.object({\n *     location: z.string().describe(\"The city and state, e.g. San Francisco, CA\")\n *   }),\n * }\n *\n * const GetPopulation = {\n *   name: \"GetPopulation\",\n *   description: \"Get the current population in a given location\",\n *   schema: z.object({\n *     location: z.string().describe(\"The city and state, e.g. San Francisco, CA\")\n *   }),\n * }\n *\n * const llmWithTools = llm.bindTools([GetWeather, GetPopulation]);\n * const aiMsg = await llmWithTools.invoke(\n *   \"Which city is hotter today and which is bigger: LA or NY?\"\n * );\n * console.log(aiMsg.tool_calls);\n * ```\n *\n * ```txt\n * [\n *   {\n *     name: 'GetPopulation',\n *     args: { location: 'New York City, NY' },\n *     id: '33c1c1f47e2f492799c77d2800a43912',\n *     type: 'tool_call'\n *   }\n * ]\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Structured Output</strong></summary>\n *\n * ```typescript\n * import { z } from 'zod';\n *\n * const Joke = z.object({\n *   setup: z.string().describe(\"The setup of the joke\"),\n *   punchline: z.string().describe(\"The punchline to the joke\"),\n *   rating: z.number().optional().describe(\"How funny the joke is, from 1 to 10\")\n * }).describe('Joke to tell user.');\n *\n * const structuredLlm = llm.withStructuredOutput(Joke, { name: \"Joke\" });\n * const jokeResult = await structuredLlm.invoke(\"Tell me a joke about cats\");\n * console.log(jokeResult);\n * ```\n *\n * ```txt\n * {\n *   setup: 'What do you call a cat that loves to bowl?',\n *   punchline: 'An alley cat!'\n * }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Usage Metadata</strong></summary>\n *\n * ```typescript\n * const aiMsgForMetadata = await llm.invoke(input);\n * console.log(aiMsgForMetadata.usage_metadata);\n * ```\n *\n * ```txt\n * { input_tokens: 9, output_tokens: 8, total_tokens: 17 }\n * ```\n * </details>\n *\n * <br />\n *\n * <details>\n * <summary><strong>Stream Usage Metadata</strong></summary>\n *\n * ```typescript\n * const streamForMetadata = await llm.stream(\n *   input,\n *   {\n *     streamUsage: true\n *   }\n * );\n * let fullForMetadata: AIMessageChunk | undefined;\n * for await (const chunk of streamForMetadata) {\n *   fullForMetadata = !fullForMetadata ? chunk : concat(fullForMetadata, chunk);\n * }\n * console.log(fullForMetadata?.usage_metadata);\n * ```\n *\n * ```txt\n * { input_tokens: 9, output_tokens: 8, total_tokens: 17 }\n * ```\n * </details>\n *\n * <br />\n */\nexport class ChatVertexAI extends ChatGoogle {\n  lc_namespace = ['langchain', 'chat_models', 'vertexai'];\n  _lc_stream_delay: number;\n  dynamicThinkingBudget = false;\n  thinkingConfig?: GoogleThinkingConfig;\n\n  static lc_name(): 'LibreChatVertexAI' {\n    return 'LibreChatVertexAI';\n  }\n\n  constructor(model: string, fields?: Omit<VertexAIClientOptions, 'model'>);\n  constructor(fields?: VertexAIClientOptions);\n  constructor(\n    modelOrFields?: string | VertexAIClientOptions,\n    params?: Omit<VertexAIClientOptions, 'model'>\n  ) {\n    const fields =\n      typeof modelOrFields === 'string'\n        ? { ...(params ?? {}), model: modelOrFields }\n        : modelOrFields;\n    const dynamicThinkingBudget = fields?.thinkingBudget === -1;\n    super({\n      ...fields,\n      platformType: 'gcp',\n    });\n    this.dynamicThinkingBudget = dynamicThinkingBudget;\n    this.thinkingConfig = fields?.thinkingConfig;\n    this._lc_stream_delay = resolveStreamDelay(fields?._lc_stream_delay);\n  }\n  invocationParams(\n    options?: this['ParsedCallOptions'] | undefined\n  ): GoogleAIModelRequestParams {\n    const params = super.invocationParams(options);\n    if (this.dynamicThinkingBudget) {\n      params.maxReasoningTokens = -1;\n    }\n    return params;\n  }\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    yield* smoothGenerationChunks({\n      chunks: this._streamRepairedChunks(messages, options),\n      delayMs: this._lc_stream_delay,\n      signal: options.signal,\n      runManager,\n    });\n  }\n\n  private async *_streamRepairedChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions']\n  ): AsyncGenerator<ChatGenerationChunk> {\n    let lastGoodUsage: UsageMetadata | undefined;\n    for await (const chunk of super._streamResponseChunks(\n      messages,\n      options,\n      undefined\n    )) {\n      const genUsage = (\n        chunk.generationInfo as { usage_metadata?: UsageMetadata } | undefined\n      )?.usage_metadata;\n      if (genUsage) {\n        lastGoodUsage = genUsage;\n      }\n      if (chunk.message instanceof AIMessageChunk) {\n        const repaired = repairStreamUsageMetadata(\n          chunk.message.usage_metadata,\n          lastGoodUsage\n        );\n        if (repaired !== chunk.message.usage_metadata) {\n          chunk.message.usage_metadata = repaired;\n        }\n        sealCompleteStreamedToolCalls(chunk.message);\n      }\n      yield chunk;\n    }\n  }\n  buildConnection(\n    fields: VertexAIClientOptions | undefined,\n    client: GoogleAbstractedClient\n  ): void {\n    // Note: buildConnection is called from super() BEFORE this.thinkingConfig is set,\n    // so we must read thinkingConfig from `fields` directly.\n    const thinkingConfig = fields?.thinkingConfig ?? this.thinkingConfig;\n\n    const connection = new CustomChatConnection(\n      { ...fields, ...this },\n      this.caller,\n      client,\n      false\n    );\n    connection.thinkingConfig = thinkingConfig;\n    this.connection = connection;\n\n    const streamedConnection = new CustomChatConnection(\n      { ...fields, ...this },\n      this.caller,\n      client,\n      true\n    );\n    streamedConnection.thinkingConfig = thinkingConfig;\n    this.streamedConnection = streamedConnection;\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6CA,SAAgB,0BACd,SACA,qBAC2B;CAC3B,IAAI,CAAC,SAAS,OAAO;CACrB,IAAI,CAAC,qBAAqB,OAAO;CACjC,IAAI,oBAAoB,iBAAiB,QAAQ,cAAc,OAAO;CACtE,IAAI,oBAAoB,iBAAiB,QAAQ,eAC/C,OAAO;CACT,OAAO;AACT;;;;;;;;;AAUA,SAAgB,8BAA8B,SAA+B;CAC3E,MAAM,aAAa,QAAQ,kBAAkB,UAAU;CACvD,IACE,eAAe,MACd,QAAQ,oBAAoB,UAAU,KAAK,MAC3C,QAAQ,YAAY,UAAU,OAAO,YAEtC;CAEF,QAAQ,oBAAoB;EAC1B,GAAG,QAAQ;GACV,0CACC;GACD,uCAAuC,EAAE,MAAM,MAAM;CACxD;AACF;;;;;;;;;;;;;;;;;AAwBA,SAAgB,qBACd,UACA,OACM;CAGN,MAAM,aAAa,MAAM,OAAO,WAAW;CAC3C,MAAM,gBAAgB,SAAS,QAAQ,MAAM,EAAE,SAAS,OAAO;CAE/D,MAAM,QAAQ,KAAK,IAAI,WAAW,QAAQ,cAAc,MAAM;CAC9D,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;EAC9B,MAAM,aAAc,WAAW,EAAE,CAAC,mBAC9B;EACJ,IAAI,CAAC,MAAM,QAAQ,UAAU,KAAK,WAAW,WAAW,GAAG;EAE3D,MAAM,UAAU,cAAc;EAC9B,MAAM,qBAAqB,IAAI,IAC7B,QAAQ,MACL,KAAK,MAAM,EAAE,gBAAgB,CAAC,CAC9B,QAAQ,MAAmB,KAAK,QAAQ,MAAM,EAAE,CACrD;EACA,MAAM,sBAAsB,WAAW,QACpC,MAAmB,KAAK,QAAQ,MAAM,MAAM,CAAC,mBAAmB,IAAI,CAAC,CACxE;EAEA,IAAI,SAAS;EACb,KAAK,MAAM,QAAQ,QAAQ,OACzB,IACE,kBAAkB,SACjB,KAAK,oBAAoB,QAAQ,KAAK,qBAAqB,OAC5D,SAAS,oBAAoB,QAC7B;GACA,KAAK,mBAAmB,oBAAoB;GAC5C;EACF;CAEJ;AACF;AAEA,IAAM,uBAAN,cAAmC,eAAsC;CACvE;CAEA,MAAM,WACJ,OACA,YACkB;EAClB,MAAM,gBAAiB,MAAM,MAAM,WACjC,OACA,UACF;EACA,IAAI,cAAc,kBAAkB,gBAAgB,mBAAmB,IAAI;GAEzE,IACE,cAAc,iBAAiB,eAAe,oBAAoB,OAElE,cAAc,iBAAiB,eAAe,kBAAkB;GAElE,OAAO,cAAc,iBAAiB,eAAe;EACvD;EACA,IAAI,KAAK,gBAAgB,iBAAiB,MAAM;GAC9C,cAAc,qBAAqB,CAAC;GAGpC,MAAM,EAAE,gBAAgB,GAAG,GAAG,2BAC3B,cAAc,iBAAiB,kBAEd,CAAC;GACrB,cACgB,iBACd,iBAAiB;IACjB,GAAG;IACH,eAAe,KAAK,eAAe;IACnC,GAAI,KAAK,eAAe,mBAAmB,QAAQ,EACjD,iBAAiB,KAAK,eAAe,gBACvC;GACF;EACF;EACA,IAAI,cAAc,UAAU;GAC1B,qBAAqB,cAAc,UAAU,KAAK;GAElD,KAAK,MAAM,WAAW,cAAc,UAClC,IAAI,QAAQ,SAAS,YACnB,QAA8B,OAAO;EAG3C;EACA,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4RA,IAAa,eAAb,cAAkC,WAAW;CAC3C,eAAe;EAAC;EAAa;EAAe;CAAU;CACtD;CACA,wBAAwB;CACxB;CAEA,OAAO,UAA+B;EACpC,OAAO;CACT;CAIA,YACE,eACA,QACA;EACA,MAAM,SACJ,OAAO,kBAAkB,WACrB;GAAE,GAAI,UAAU,CAAC;GAAI,OAAO;EAAc,IAC1C;EACN,MAAM,wBAAwB,QAAQ,mBAAmB;EACzD,MAAM;GACJ,GAAG;GACH,cAAc;EAChB,CAAC;EACD,KAAK,wBAAwB;EAC7B,KAAK,iBAAiB,QAAQ;EAC9B,KAAK,mBAAmB,mBAAmB,QAAQ,gBAAgB;CACrE;CACA,iBACE,SAC4B;EAC5B,MAAM,SAAS,MAAM,iBAAiB,OAAO;EAC7C,IAAI,KAAK,uBACP,OAAO,qBAAqB;EAE9B,OAAO;CACT;CACA,OAAO,sBACL,UACA,SACA,YACqC;EACrC,OAAO,uBAAuB;GAC5B,QAAQ,KAAK,sBAAsB,UAAU,OAAO;GACpD,SAAS,KAAK;GACd,QAAQ,QAAQ;GAChB;EACF,CAAC;CACH;CAEA,OAAe,sBACb,UACA,SACqC;EACrC,IAAI;EACJ,WAAW,MAAM,SAAS,MAAM,sBAC9B,UACA,SACA,KAAA,CACF,GAAG;GACD,MAAM,WACJ,MAAM,gBACL;GACH,IAAI,UACF,gBAAgB;GAElB,IAAI,MAAM,mBAAmB,gBAAgB;IAC3C,MAAM,WAAW,0BACf,MAAM,QAAQ,gBACd,aACF;IACA,IAAI,aAAa,MAAM,QAAQ,gBAC7B,MAAM,QAAQ,iBAAiB;IAEjC,8BAA8B,MAAM,OAAO;GAC7C;GACA,MAAM;EACR;CACF;CACA,gBACE,QACA,QACM;EAGN,MAAM,iBAAiB,QAAQ,kBAAkB,KAAK;EAEtD,MAAM,aAAa,IAAI,qBACrB;GAAE,GAAG;GAAQ,GAAG;EAAK,GACrB,KAAK,QACL,QACA,KACF;EACA,WAAW,iBAAiB;EAC5B,KAAK,aAAa;EAElB,MAAM,qBAAqB,IAAI,qBAC7B;GAAE,GAAG;GAAQ,GAAG;EAAK,GACrB,KAAK,QACL,QACA,IACF;EACA,mBAAmB,iBAAiB;EACpC,KAAK,qBAAqB;CAC5B;AACF"}

package/dist/esm/main.mjs

@@ -3,7 +3,7 @@ import { ANTHROPIC_TOOL_TOKEN_MULTIPLIER, DEFAULT_MAX_SEALS, DEFAULT_RECURSION_L
 import { CODE_EXECUTION_TOOLS, Callback, CommonEvents, Constants, ContentTypes, EnvVar, GraphEvents, GraphNodeActions, GraphNodeKeys, LOCAL_CODING_BUNDLE_NAMES, LOCAL_CODING_TOOL_NAMES, Providers, StepTypes, TitleMethod, ToolCallTypes } from "./common/enum.mjs";
 import "./common/index.mjs";
 import { HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, truncateToolInput, truncateToolResultContent } from "./utils/truncation.mjs";
-import { IMAGE_TOKEN_SAFETY_MARGIN, TokenEncoderManager, apportionTokenCounts, createTokenCounter, encodingForModel, estimateAnthropicImageTokens, estimateDocumentBlockTokens, estimateImageBlockTokens, estimateOpenAIImageTokens, estimateTimedMediaBlockTokens, extractImageDimensions, getTokenCountForMessage, hasUnsafeStructuredSerialization } from "./utils/tokens.mjs";
+import { IMAGE_TOKEN_SAFETY_MARGIN, TokenEncoderManager, UnsafeTokenMeasurementError, apportionTokenCounts, createTokenCounter, encodingForModel, estimateAnthropicImageTokens, estimateDocumentBlockTokens, estimateImageBlockTokens, estimateOpenAIImageTokens, estimateTimedMediaBlockTokens, extractImageDimensions, getTokenCountForMessage, hasUnsafeStructuredSerialization } from "./utils/tokens.mjs";
 import { DEFAULT_PROMPT_CACHE_TTL, addBedrockCacheControl, addBedrockTailCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, addTailCacheControl, buildAnthropicCacheControl, buildBedrockCachePoint, cloneMessage, resolveBedrockPromptCacheTtl, resolvePromptCacheTtl, stripAnthropicCacheControl, stripBedrockCacheControl, supportsBedrockToolCache } from "./messages/cache.mjs";
 import { OPENAI_RESPONSES_REPLAY_POSITIONS_KEY, convertMessagesToContent, findLastIndex, formatAnthropicArtifactContent, formatAnthropicMessage, formatArtifactPayload, getConverseOverrideMessage, modifyDeltaProperties, projectAnthropicArtifactContent, projectArtifactPayload, projectCacheControlledToolOutputsToText, projectComputerCallOutputsToText, projectOpenAIChatToolMessageContent, projectOpenAIResponsesToolMessageContent, projectOpenAIToolMessageContent, projectOpenRouterToolMessageContent, projectSingleTextToolOutputsToText, projectStructuredToolOutputsToText, projectToolStreamContentForProvider } from "./messages/core.mjs";
 import { getMessageId } from "./messages/ids.mjs";
@@ -26,6 +26,8 @@ import { DEFAULT_MAX_TOOL_CALL_ARG_BYTES, StreamLimitExceededError, resolveStrea
 import { INTENT_ARG, INTENT_DESCRIPTION, INTENT_LABEL_MARKER, INTENT_PROPERTY, applyOutcome, isIntentLabelProperty, outcomeFieldsFromResult, readIntent, readOutcomeFields, resolveToolOutcome, stripIntent, withIntent, withoutIntent } from "./tools/intentArg.mjs";
 import { isAnthropicLike, isGoogleLike, isOpenAILike } from "./utils/llm.mjs";
 import { ChatModelStreamHandler, SDK_STREAM_DISPATCH, createContentAggregator, dispatchesChatModelStream, getChunkContent } from "./stream.mjs";
+import { DEFAULT_STREAM_DELAY, computeAdaptivePieceSize, resolveStreamDelay, smoothStream } from "./llm/stream/smoother.mjs";
+import { CustomChatMistralAI } from "./llm/mistral/index.mjs";
 import { CustomOpenAIClient } from "./llm/openai/index.mjs";
 import { ChatOpenRouter } from "./llm/openrouter/index.mjs";
 import { getChatModelClass } from "./llm/providers.mjs";
@@ -104,4 +106,4 @@ import { Runnable, RunnableLambda, RunnableSequence } from "./langchain/runnable
 import { DynamicStructuredTool, StructuredTool, Tool, tool } from "./langchain/tools.mjs";
 import "./langchain/index.mjs";
 import { BaseCheckpointSaver, Command, INTERRUPT, MemorySaver, interrupt, isInterrupted } from "@langchain/langgraph";
-export { AIMessage, AIMessageChunk, ANTHROPIC_TOOL_TOKEN_MULTIPLIER, AgentSession, BASH_SHELL_GUIDANCE, BaseCheckpointSaver, BaseMessage, BaseMessageChunk, BashExecutionToolDefinition, BashExecutionToolDescription, BashExecutionToolName, BashExecutionToolSchema, BashProgrammaticToolCallingDefinition, BashProgrammaticToolCallingDescription, BashProgrammaticToolCallingName, BashProgrammaticToolCallingSchema, BashToolOutputReferencesGuide, CALIBRATION_RATIO_MAX, CALIBRATION_RATIO_MIN, CLOUDFLARE_BASH_CODING_TOOL_NAMES, CLOUDFLARE_CODING_TOOL_NAMES, CODE_API_AUTHORIZATION_ERROR_MESSAGE, CODE_API_EXECUTION_FAILED_ERROR_MESSAGE, CODE_API_INVALID_REQUEST_ERROR_MESSAGE, CODE_API_RATE_LIMITED_ERROR_MESSAGE, CODE_API_UNAVAILABLE_ERROR_MESSAGE, CODE_ARTIFACT_PATH_GUIDANCE, CODE_EXECUTION_TOOLS, Calculator, CalculatorSchema, CalculatorToolDefinition, CalculatorToolDescription, CalculatorToolName, Callback, ChatModelStreamHandler, ChatOpenRouter, CloudflareBashExecutionToolDescription, CloudflareCodeExecutionToolDescription, CodeApiRequestError, CodeExecutionToolDefinition, CodeExecutionToolDescription, CodeExecutionToolName, CodeExecutionToolSchema, Command, CommonEvents, CompileCheckToolName, Constants, ContentTypes, CustomOpenAIClient, DATE_RANGE, DEFAULT_CONTEXT_PRUNING_SETTINGS, DEFAULT_COUNTRY_DESCRIPTION, DEFAULT_HOOK_TIMEOUT_MS, DEFAULT_MAX_SEALS, DEFAULT_MAX_TOOL_CALL_ARG_BYTES, DEFAULT_PROMPT_CACHE_TTL, DEFAULT_QUERY_DESCRIPTION, DEFAULT_RECURSION_LIMIT, DEFAULT_RESERVE_RATIO, DEFAULT_RETAIN_RECENT_TURNS, DEFAULT_TOOL_TOKEN_MULTIPLIER, DynamicStructuredTool, EnvVar, FAILED_EXECUTION_FILE_REMINDER, FakeChatModel, Graph, GraphEvents, GraphNodeActions, GraphNodeKeys, HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, HOOK_EVENTS, HOOK_INJECTED_MESSAGES_CAPABLE, HOOK_PREEMPT_BOUNDARY_CAPABLE, HandlerRegistry, HookRegistry, HumanMessage, IMAGE_TOKEN_SAFETY_MARGIN, INTENT_ARG, INTENT_DESCRIPTION, INTENT_LABEL_MARKER, INTENT_PROPERTY, INTERRUPT, JsonlSessionStore, LLMStreamHandler, LOCAL_CODING_BUNDLE_NAMES, LOCAL_CODING_TOOL_NAMES, LOCAL_SPAWN_TIMEOUT_MS, LocalBashExecutionToolDescription, LocalCodeExecutionToolDescription, LocalEditFileToolName, LocalEditFileToolSchema, LocalFileCheckpointerImpl, LocalGlobSearchToolName, LocalGlobSearchToolSchema, LocalGrepSearchToolName, LocalGrepSearchToolSchema, LocalListDirectoryToolName, LocalListDirectoryToolSchema, LocalReadFileToolSchema, LocalWriteFileToolName, LocalWriteFileToolSchema, MAX_CACHE_SIZE, MAX_PATTERN_LENGTH, MemorySaver, ModelEndHandler, MultiAgentGraph, OPENAI_RESPONSES_REPLAY_POSITIONS_KEY, ORIGINAL_CONTENT_MAX_CHARS, PREDECESSOR_HANDOFF_CUE, PREEMPT_BOUNDARY_HOOK_TIMEOUT_MS, ProgrammaticToolCallingDefinition, ProgrammaticToolCallingDescription, ProgrammaticToolCallingName, ProgrammaticToolCallingSchema, PromptTemplate, Providers, REMOVE_ALL_MESSAGES, REPLY_PRIMER_TOKENS, ReadFileToolDefinition, ReadFileToolDescription, ReadFileToolName, ReadFileToolSchema, Run, Runnable, RunnableCallable, RunnableLambda, RunnableSequence, SDK_STREAM_DISPATCH, STATEFUL_BASH_NOTE, STATEFUL_ENV_NOTE, SessionManager, SkillToolDefinition, SkillToolDescription, SkillToolName, SkillToolSchema, StandardGraph, StatefulBashExecutionToolDescription, StatefulCodeExecutionToolDescription, StepTypes, StreamLimitExceededError, StructuredTool, SubagentExecutor, SubagentToolDefinition, SubagentToolDescription, SubagentToolName, SubagentToolSchema, SystemMessage, TMP_SCRATCH_OUTPUT_REMINDER, TestChatStreamHandler, TestLLMStreamHandler, TitleMethod, TokenEncoderManager, Tool, ToolCallTypes, ToolEndHandler, ToolMessage, ToolNode, ToolSearchToolDefinition, ToolSearchToolDescription, ToolSearchToolName, ToolSearchToolSchema, WebSearchToolDefinition, WebSearchToolDescription, WebSearchToolName, WebSearchToolSchema, _createBashProgramForTests, _resetLocalEngineWarningsForTests, _resetRipgrepCacheForTests, _resetSyntaxCheckProbeCacheForTests, _resetUnrecognizedTriggerWarnings, addBedrockCacheControl, addBedrockTailCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, addTailCacheControl, appendCodeSessionFileSummary, appendFailedExecutionFileReminder, appendPredecessorHandoffCue, appendTmpScratchReminder, applyContextPruning, applyEdit, applyOutcome, applyPreToolUseHooksForBridge, apportionTokenCounts, askUserQuestion, attemptInvoke, bashAstFindingsToErrors, buildAnthropicCacheControl, buildBashExecutionToolDescription, buildBashExecutionToolSchema, buildBedrockCachePoint, buildChildInputs, buildCodeApiExecutionErrorMessage, buildCodeApiHttpErrorMessage, buildCodeExecutionToolDescription, buildCodeExecutionToolSchema, buildSandboxRuntimeConfig, buildSubagentToolParams, calculateMaxToolCallInputChars, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, calculateTotalTokens, canSealPreempt, checkValidNumber, clampCalibrationRatio, classifyAttachment, clientExecTimeoutMs, clientFsTimeoutMs, cloneMessage, coalesceAdjacentUserTurns, composeAbortSignals, composeEventHandlers, convertInjectedMessages, convertMessagesToContent, countNestedGroups, countrySchema, createAgentSession, createBashExecutionTool, createBashProgrammaticToolCallingSchema, createBashProgrammaticToolCallingTool, createCloudflareBashExecutionTool, createCloudflareBashProgrammaticToolCallingTool, createCloudflareBridgeRuntime, createCloudflareCodeExecutionTool, createCloudflareCodingToolBundle, createCloudflareCodingTools, createCloudflareExecutionTool, createCloudflareLocalExecutionConfig, createCloudflareProgrammaticToolCallingTool, createCloudflareWorkspaceFS, createCodeExecutionTool, createCompileCheckTool, createCompileCheckToolDefinition, createContentAggregator, createFakeStreamingLLM, createHandlers, createLocalBashExecutionTool, createLocalBashProgrammaticToolCallingTool, createLocalCodeExecutionTool, createLocalCodingToolBundle, createLocalCodingToolDefinitions, createLocalCodingToolRegistry, createLocalCodingTools, createLocalEditFileTool, createLocalFileCheckpointer, createLocalGlobSearchTool, createLocalGrepSearchTool, createLocalListDirectoryTool, createLocalProgrammaticToolCallingTool, createLocalReadFileTool, createLocalWriteFileTool, createMetadataAggregator, createProgrammaticToolCallingSchema, createProgrammaticToolCallingTool, createPruneMessages, createRemoveAllMessage, createRunHandlers, createSchemaOnlyTool, createSchemaOnlyTools, createSearchTool, createSubagentToolDefinition, createTokenCounter, createToolErrorOwnership, createToolPolicyHook, createToolSearch, createWorkspacePolicyHook, dateSchema, decodeFile, defaultOmitOptions, deserializeMessage, dispatchesChatModelStream, emptyOutputMessage, encodeFile, encodingForModel, enforceOriginalContentCap, ensureThinkingBlockInMessages, escapeRegexSpecialChars, estimateAnthropicImageTokens, estimateDocumentBlockTokens, estimateImageBlockTokens, estimateOpenAIImageTokens, estimateTimedMediaBlockTokens, execWithClientTimeout, executeCloudflareBash, executeCloudflareCode, executeHooks, executeLocalBash, executeLocalBashWithArgs, executeLocalCode, executeParallelSearches, executeTools, extractErrorMessage, extractImageDimensions, extractMcpServerName, extractTextFromContent, extractToolDiscoveries, extractUsedBashToolNames, extractUsedToolNames, fetchSessionFiles, filterBashToolsByUsage, filterSubagentResult, filterToolsByUsage, findLastIndex, foldToolBlocksForToollessAgent, formatAgentMessages, formatAnthropicArtifactContent, formatAnthropicMessage, formatArtifactPayload, formatCloudflareOutput, formatCompletedResponse, formatContentStrings, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, formatServerListing, formatSkillCatalog, getAvailableMcpServers, getBaseToolName, getBufferString, getChatModelClass, getChunkContent, getCloudflareWorkspaceRoot, getCodeBaseURL, getContextOverflowInfo, getConverseOverrideMessage, getDeferredToolsListing, getLocalCwd, getLocalSessionId, getMaxOutputTokensKey, getMessageId, getMessagesWithinTokenLimit, getReadRoots, getSpawn, getTokenCountForMessage, getWorkspaceFS, getWorkspaceRoots, getWriteRoots, handleServerToolResult, handleToolCallChunks, handleToolCalls, hasNestedQuantifier, hasNestedQuantifiers, hasToolSearchInCurrentTurn, hasUnsafeStructuredSerialization, imageAttachmentContent, imagesSchema, initializeModel, interrupt, isAIMessage, isAnthropicLike, isBaseMessage, isContextOverflowError, isDangerousPattern, isFromAnyMcpServer, isFromMcpServer, isGoogleLike, isIntentLabelProperty, isInterrupted, isLegacyConvertible, isLikelyContextOverflowError, isOpenAILike, isPresent, isSyntheticProviderContextMessage, isThinkingEnabled, isToolMessage, isZodSchema, joinKeys, labelContentByAgent, locateEdit, makeIsDeferred, makeRequest, maskConsumedToolResults, matchesQuery, messagesStateReducer, modifyDeltaProperties, newsSchema, normalizeBashToolResultsForReplay, normalizeCodeApiRequestError, normalizeServerFilter, normalizeToBashIdentifier, normalizeToPythonIdentifier, outcomeFieldsFromResult, parseBooleanEnv, partitionAndMarkAnthropicToolCache, performLocalSearch, preFlightTruncateToolCallInputs, preFlightTruncateToolResults, projectAgentContextUsage, projectAnthropicArtifactContent, projectArtifactPayload, projectCacheControlledToolOutputsToText, projectComputerCallOutputsToText, projectOpenAIChatToolMessageContent, projectOpenAIResponsesToolMessageContent, projectOpenAIToolMessageContent, projectOpenRouterToolMessageContent, projectSingleTextToolOutputsToText, projectStructuredToolOutputsToText, projectToolCallInputs, projectToolStreamContentForProvider, querySchema, readIntent, readOutcomeFields, removePredecessorHandoffCue, repairOrphanedToolMessages, resetIfNotEmpty, resolveBedrockPromptCacheTtl, resolveCloudflareSandbox, resolveCodeApiAuthHeaders, resolveContextPruningSettings, resolveFetchProxyAgent, resolveLocalExecutionConfig, resolveLocalExecutionTools, resolveLocalToolRegistry, resolveLocalToolsForBinding, resolvePromptCacheTtl, resolveSearchOutcome, resolveStreamLimits, resolveSubagentConfigs, resolveToolOutcome, resolveWorkspacePath, resolveWorkspacePathSafe, runBashAstChecks, runPostEditSyntaxCheck, sanitizeOrphanToolBlocks, sanitizeRegex, serializeMessage, serializeToolCallInput, shellQuote, shiftIndexTokenCountMap, shouldBypassProxy, shouldTriggerSummarization, sleep, spawnLocalProcess, splitAtRecencyBoundary, strictAlternationProviders, stripAnthropicCacheControl, stripBedrockCacheControl, stripCodeSessionFileSummary, stripIntent, summarizeEvent, supportsBedrockToolCache, syncBudgetDerivedFields, toJsonSchema, tool, toolResultTypes, toolsCondition, truncateLocalOutput, truncateToolInput, truncateToolResultContent, tryFallbackProviders, unescapeObject, unwrapToolResponse, validateBashCommand, validateCloudflareBashCommand, videosSchema, withClientTimeout, withIntent, withMessageRole, withoutIntent };
+export { AIMessage, AIMessageChunk, ANTHROPIC_TOOL_TOKEN_MULTIPLIER, AgentSession, BASH_SHELL_GUIDANCE, BaseCheckpointSaver, BaseMessage, BaseMessageChunk, BashExecutionToolDefinition, BashExecutionToolDescription, BashExecutionToolName, BashExecutionToolSchema, BashProgrammaticToolCallingDefinition, BashProgrammaticToolCallingDescription, BashProgrammaticToolCallingName, BashProgrammaticToolCallingSchema, BashToolOutputReferencesGuide, CALIBRATION_RATIO_MAX, CALIBRATION_RATIO_MIN, CLOUDFLARE_BASH_CODING_TOOL_NAMES, CLOUDFLARE_CODING_TOOL_NAMES, CODE_API_AUTHORIZATION_ERROR_MESSAGE, CODE_API_EXECUTION_FAILED_ERROR_MESSAGE, CODE_API_INVALID_REQUEST_ERROR_MESSAGE, CODE_API_RATE_LIMITED_ERROR_MESSAGE, CODE_API_UNAVAILABLE_ERROR_MESSAGE, CODE_ARTIFACT_PATH_GUIDANCE, CODE_EXECUTION_TOOLS, Calculator, CalculatorSchema, CalculatorToolDefinition, CalculatorToolDescription, CalculatorToolName, Callback, ChatModelStreamHandler, ChatOpenRouter, CloudflareBashExecutionToolDescription, CloudflareCodeExecutionToolDescription, CodeApiRequestError, CodeExecutionToolDefinition, CodeExecutionToolDescription, CodeExecutionToolName, CodeExecutionToolSchema, Command, CommonEvents, CompileCheckToolName, Constants, ContentTypes, CustomChatMistralAI, CustomOpenAIClient, DATE_RANGE, DEFAULT_CONTEXT_PRUNING_SETTINGS, DEFAULT_COUNTRY_DESCRIPTION, DEFAULT_HOOK_TIMEOUT_MS, DEFAULT_MAX_SEALS, DEFAULT_MAX_TOOL_CALL_ARG_BYTES, DEFAULT_PROMPT_CACHE_TTL, DEFAULT_QUERY_DESCRIPTION, DEFAULT_RECURSION_LIMIT, DEFAULT_RESERVE_RATIO, DEFAULT_RETAIN_RECENT_TURNS, DEFAULT_STREAM_DELAY, DEFAULT_TOOL_TOKEN_MULTIPLIER, DynamicStructuredTool, EnvVar, FAILED_EXECUTION_FILE_REMINDER, FakeChatModel, Graph, GraphEvents, GraphNodeActions, GraphNodeKeys, HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, HOOK_EVENTS, HOOK_INJECTED_MESSAGES_CAPABLE, HOOK_PREEMPT_BOUNDARY_CAPABLE, HandlerRegistry, HookRegistry, HumanMessage, IMAGE_TOKEN_SAFETY_MARGIN, INTENT_ARG, INTENT_DESCRIPTION, INTENT_LABEL_MARKER, INTENT_PROPERTY, INTERRUPT, JsonlSessionStore, LLMStreamHandler, LOCAL_CODING_BUNDLE_NAMES, LOCAL_CODING_TOOL_NAMES, LOCAL_SPAWN_TIMEOUT_MS, LocalBashExecutionToolDescription, LocalCodeExecutionToolDescription, LocalEditFileToolName, LocalEditFileToolSchema, LocalFileCheckpointerImpl, LocalGlobSearchToolName, LocalGlobSearchToolSchema, LocalGrepSearchToolName, LocalGrepSearchToolSchema, LocalListDirectoryToolName, LocalListDirectoryToolSchema, LocalReadFileToolSchema, LocalWriteFileToolName, LocalWriteFileToolSchema, MAX_CACHE_SIZE, MAX_PATTERN_LENGTH, MemorySaver, ModelEndHandler, MultiAgentGraph, OPENAI_RESPONSES_REPLAY_POSITIONS_KEY, ORIGINAL_CONTENT_MAX_CHARS, PREDECESSOR_HANDOFF_CUE, PREEMPT_BOUNDARY_HOOK_TIMEOUT_MS, ProgrammaticToolCallingDefinition, ProgrammaticToolCallingDescription, ProgrammaticToolCallingName, ProgrammaticToolCallingSchema, PromptTemplate, Providers, REMOVE_ALL_MESSAGES, REPLY_PRIMER_TOKENS, ReadFileToolDefinition, ReadFileToolDescription, ReadFileToolName, ReadFileToolSchema, Run, Runnable, RunnableCallable, RunnableLambda, RunnableSequence, SDK_STREAM_DISPATCH, STATEFUL_BASH_NOTE, STATEFUL_ENV_NOTE, SessionManager, SkillToolDefinition, SkillToolDescription, SkillToolName, SkillToolSchema, StandardGraph, StatefulBashExecutionToolDescription, StatefulCodeExecutionToolDescription, StepTypes, StreamLimitExceededError, StructuredTool, SubagentExecutor, SubagentToolDefinition, SubagentToolDescription, SubagentToolName, SubagentToolSchema, SystemMessage, TMP_SCRATCH_OUTPUT_REMINDER, TestChatStreamHandler, TestLLMStreamHandler, TitleMethod, TokenEncoderManager, Tool, ToolCallTypes, ToolEndHandler, ToolMessage, ToolNode, ToolSearchToolDefinition, ToolSearchToolDescription, ToolSearchToolName, ToolSearchToolSchema, UnsafeTokenMeasurementError, WebSearchToolDefinition, WebSearchToolDescription, WebSearchToolName, WebSearchToolSchema, _createBashProgramForTests, _resetLocalEngineWarningsForTests, _resetRipgrepCacheForTests, _resetSyntaxCheckProbeCacheForTests, _resetUnrecognizedTriggerWarnings, addBedrockCacheControl, addBedrockTailCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, addTailCacheControl, appendCodeSessionFileSummary, appendFailedExecutionFileReminder, appendPredecessorHandoffCue, appendTmpScratchReminder, applyContextPruning, applyEdit, applyOutcome, applyPreToolUseHooksForBridge, apportionTokenCounts, askUserQuestion, attemptInvoke, bashAstFindingsToErrors, buildAnthropicCacheControl, buildBashExecutionToolDescription, buildBashExecutionToolSchema, buildBedrockCachePoint, buildChildInputs, buildCodeApiExecutionErrorMessage, buildCodeApiHttpErrorMessage, buildCodeExecutionToolDescription, buildCodeExecutionToolSchema, buildSandboxRuntimeConfig, buildSubagentToolParams, calculateMaxToolCallInputChars, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, calculateTotalTokens, canSealPreempt, checkValidNumber, clampCalibrationRatio, classifyAttachment, clientExecTimeoutMs, clientFsTimeoutMs, cloneMessage, coalesceAdjacentUserTurns, composeAbortSignals, composeEventHandlers, computeAdaptivePieceSize, convertInjectedMessages, convertMessagesToContent, countNestedGroups, countrySchema, createAgentSession, createBashExecutionTool, createBashProgrammaticToolCallingSchema, createBashProgrammaticToolCallingTool, createCloudflareBashExecutionTool, createCloudflareBashProgrammaticToolCallingTool, createCloudflareBridgeRuntime, createCloudflareCodeExecutionTool, createCloudflareCodingToolBundle, createCloudflareCodingTools, createCloudflareExecutionTool, createCloudflareLocalExecutionConfig, createCloudflareProgrammaticToolCallingTool, createCloudflareWorkspaceFS, createCodeExecutionTool, createCompileCheckTool, createCompileCheckToolDefinition, createContentAggregator, createFakeStreamingLLM, createHandlers, createLocalBashExecutionTool, createLocalBashProgrammaticToolCallingTool, createLocalCodeExecutionTool, createLocalCodingToolBundle, createLocalCodingToolDefinitions, createLocalCodingToolRegistry, createLocalCodingTools, createLocalEditFileTool, createLocalFileCheckpointer, createLocalGlobSearchTool, createLocalGrepSearchTool, createLocalListDirectoryTool, createLocalProgrammaticToolCallingTool, createLocalReadFileTool, createLocalWriteFileTool, createMetadataAggregator, createProgrammaticToolCallingSchema, createProgrammaticToolCallingTool, createPruneMessages, createRemoveAllMessage, createRunHandlers, createSchemaOnlyTool, createSchemaOnlyTools, createSearchTool, createSubagentToolDefinition, createTokenCounter, createToolErrorOwnership, createToolPolicyHook, createToolSearch, createWorkspacePolicyHook, dateSchema, decodeFile, defaultOmitOptions, deserializeMessage, dispatchesChatModelStream, emptyOutputMessage, encodeFile, encodingForModel, enforceOriginalContentCap, ensureThinkingBlockInMessages, escapeRegexSpecialChars, estimateAnthropicImageTokens, estimateDocumentBlockTokens, estimateImageBlockTokens, estimateOpenAIImageTokens, estimateTimedMediaBlockTokens, execWithClientTimeout, executeCloudflareBash, executeCloudflareCode, executeHooks, executeLocalBash, executeLocalBashWithArgs, executeLocalCode, executeParallelSearches, executeTools, extractErrorMessage, extractImageDimensions, extractMcpServerName, extractTextFromContent, extractToolDiscoveries, extractUsedBashToolNames, extractUsedToolNames, fetchSessionFiles, filterBashToolsByUsage, filterSubagentResult, filterToolsByUsage, findLastIndex, foldToolBlocksForToollessAgent, formatAgentMessages, formatAnthropicArtifactContent, formatAnthropicMessage, formatArtifactPayload, formatCloudflareOutput, formatCompletedResponse, formatContentStrings, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, formatServerListing, formatSkillCatalog, getAvailableMcpServers, getBaseToolName, getBufferString, getChatModelClass, getChunkContent, getCloudflareWorkspaceRoot, getCodeBaseURL, getContextOverflowInfo, getConverseOverrideMessage, getDeferredToolsListing, getLocalCwd, getLocalSessionId, getMaxOutputTokensKey, getMessageId, getMessagesWithinTokenLimit, getReadRoots, getSpawn, getTokenCountForMessage, getWorkspaceFS, getWorkspaceRoots, getWriteRoots, handleServerToolResult, handleToolCallChunks, handleToolCalls, hasNestedQuantifier, hasNestedQuantifiers, hasToolSearchInCurrentTurn, hasUnsafeStructuredSerialization, imageAttachmentContent, imagesSchema, initializeModel, interrupt, isAIMessage, isAnthropicLike, isBaseMessage, isContextOverflowError, isDangerousPattern, isFromAnyMcpServer, isFromMcpServer, isGoogleLike, isIntentLabelProperty, isInterrupted, isLegacyConvertible, isLikelyContextOverflowError, isOpenAILike, isPresent, isSyntheticProviderContextMessage, isThinkingEnabled, isToolMessage, isZodSchema, joinKeys, labelContentByAgent, locateEdit, makeIsDeferred, makeRequest, maskConsumedToolResults, matchesQuery, messagesStateReducer, modifyDeltaProperties, newsSchema, normalizeBashToolResultsForReplay, normalizeCodeApiRequestError, normalizeServerFilter, normalizeToBashIdentifier, normalizeToPythonIdentifier, outcomeFieldsFromResult, parseBooleanEnv, partitionAndMarkAnthropicToolCache, performLocalSearch, preFlightTruncateToolCallInputs, preFlightTruncateToolResults, projectAgentContextUsage, projectAnthropicArtifactContent, projectArtifactPayload, projectCacheControlledToolOutputsToText, projectComputerCallOutputsToText, projectOpenAIChatToolMessageContent, projectOpenAIResponsesToolMessageContent, projectOpenAIToolMessageContent, projectOpenRouterToolMessageContent, projectSingleTextToolOutputsToText, projectStructuredToolOutputsToText, projectToolCallInputs, projectToolStreamContentForProvider, querySchema, readIntent, readOutcomeFields, removePredecessorHandoffCue, repairOrphanedToolMessages, resetIfNotEmpty, resolveBedrockPromptCacheTtl, resolveCloudflareSandbox, resolveCodeApiAuthHeaders, resolveContextPruningSettings, resolveFetchProxyAgent, resolveLocalExecutionConfig, resolveLocalExecutionTools, resolveLocalToolRegistry, resolveLocalToolsForBinding, resolvePromptCacheTtl, resolveSearchOutcome, resolveStreamDelay, resolveStreamLimits, resolveSubagentConfigs, resolveToolOutcome, resolveWorkspacePath, resolveWorkspacePathSafe, runBashAstChecks, runPostEditSyntaxCheck, sanitizeOrphanToolBlocks, sanitizeRegex, serializeMessage, serializeToolCallInput, shellQuote, shiftIndexTokenCountMap, shouldBypassProxy, shouldTriggerSummarization, sleep, smoothStream, spawnLocalProcess, splitAtRecencyBoundary, strictAlternationProviders, stripAnthropicCacheControl, stripBedrockCacheControl, stripCodeSessionFileSummary, stripIntent, summarizeEvent, supportsBedrockToolCache, syncBudgetDerivedFields, toJsonSchema, tool, toolResultTypes, toolsCondition, truncateLocalOutput, truncateToolInput, truncateToolResultContent, tryFallbackProviders, unescapeObject, unwrapToolResponse, validateBashCommand, validateCloudflareBashCommand, videosSchema, withClientTimeout, withIntent, withMessageRole, withoutIntent };