@openclaw/ai 2026.7.2-beta.2 → 2026.7.2-beta.4

package/README.md

@@ -14,15 +14,16 @@ registerBuiltInApiProviders(runtime.registry);
 
 Provider-neutral contracts, validation, diagnostics, and event streams are
 available from the package root and focused subpaths such as
-`@openclaw/ai/event-stream` and `@openclaw/ai/validation`. No second OpenClaw
-runtime package is required.
+`@openclaw/ai/event-stream`, `@openclaw/ai/transports`, and
+`@openclaw/ai/validation`. No second OpenClaw runtime package is required.
 
 Provider ids, credentials, model catalogs, retries, and failover remain
 application concerns. OpenClaw supplies those policies around this package.
 Host policy (request fetch guarding, secret redaction, strict-tool defaults,
-diagnostics logging) can be injected with `configureAiTransportHost`; the
-defaults are inert.
+provider plugin hooks, and diagnostics logging) can be injected with
+`configureAiTransportHost`; the defaults are inert.
 
-`@openclaw/ai/internal/*` subpaths exist for the OpenClaw application itself.
+The explicit `@openclaw/ai/internal/anthropic`, `openai`, `retry-after`,
+`runtime`, and `shared` subpaths exist for the OpenClaw application itself.
 They carry no semver guarantee and can change or disappear in any release; do
 not depend on them outside OpenClaw.

package/dist/{anthropic-BIRSg5x9.mjs → anthropic-CrUDIBpM.mjs}

@@ -1,17 +1,117 @@
-import { n as getEnvApiKey, r as __exportAll } from "./env-api-keys-DnhYpc27.mjs";
+import { n as getEnvApiKey, r as __exportAll } from "./env-api-keys-DrgeBuva.mjs";
 import { c as resolveClaudeNativeThinkingLevelMap, d as supportsClaudeNativeMaxEffort, f as supportsClaudeNativeXhighEffort, l as resolveClaudeSonnet5ModelIdentity, u as supportsClaudeAdaptiveThinking } from "./src-CXno1H5g.mjs";
 import { t as AssistantMessageEventStream } from "./event-stream-D8n2uFee.mjs";
-import { n as getAiTransportHost, r as resolveAiTransportHeaderSentinels } from "./host-4t713IeR.mjs";
-import { n as calculateCost, r as clampThinkingLevel } from "./model-utils-Q1LSRIdo.mjs";
+import { n as getAiTransportHost, r as resolveAiTransportHeaderSentinels } from "./host-XYGZcgO8.mjs";
+import { C as isImageWithMediaPayload, D as buildBaseOptions, E as adjustMaxTokensForThinking, O as clampMaxTokensToModel, T as isRecord, a as stripSystemPromptCacheBoundary, b as extractToolResultBlockText, c as transformMessages, d as prepareClaudeSonnet5RequestContext, f as requiresClaudeAdaptiveThinking, h as usesClaudeStreamingRefusalContract, i as splitSystemPromptCacheBoundary, l as applyClaudeRequestContract, m as usesClaudeFable5MessagesContract, x as extractToolResultText, y as describeToolResultMediaPlaceholder } from "./system-prompt-cache-boundary-CbHeV4_l.mjs";
+import { t as sanitizeSurrogates } from "./sanitize-unicode-DT5o51ur.mjs";
+import { n as calculateCost, r as clampThinkingLevel } from "./model-utils-1GiZ2_rr.mjs";
 import { r as createDeferredEventBuffer, t as notifyLlmRequestActivity } from "./llm-request-activity-CehVkZP-.mjs";
 import { t as headersToRecord } from "./headers-B_e4-1J0.mjs";
 import { n as parseStreamingJson, t as parseJsonWithRepair } from "./json-parse-BvXNt1-7.mjs";
-import { t as sanitizeSurrogates } from "./sanitize-unicode-DT5o51ur.mjs";
-import { C as stripSystemPromptCacheBoundary, S as splitSystemPromptCacheBoundary, _ as usesClaudeFable5MessagesContract, c as adjustMaxTokensForThinking, d as isRecord, f as applyClaudeRequestContract, h as requiresClaudeAdaptiveThinking, i as extractToolResultText, l as buildBaseOptions, m as prepareClaudeSonnet5RequestContext, n as describeToolResultMediaPlaceholder, o as isImageWithMediaPayload, r as extractToolResultBlockText, t as transformMessages, v as usesClaudeStreamingRefusalContract } from "./transform-messages-BmS70CP5.mjs";
 import { t as projectRuntimeToolInputSchema } from "./tool-schema-json-projection-BwNu3nDi.mjs";
 import { a as resolveCacheRetention, i as resolveCloudflareBaseUrl, n as hasCopilotVisionInput, t as buildCopilotDynamicHeaders } from "./github-copilot-headers-BsH5cqGj.mjs";
 import Anthropic from "@anthropic-ai/sdk";
 import { Stream } from "@anthropic-ai/sdk/core/streaming.js";
+//#region packages/media-core/src/base64.ts
+/** Estimates decoded bytes without allocating a cleaned copy of the base64 payload. */
+function estimateBase64DecodedBytes(base64) {
+	let effectiveLen = 0;
+	for (let i = 0; i < base64.length; i += 1) {
+		if (base64.charCodeAt(i) <= 32) continue;
+		effectiveLen += 1;
+	}
+	if (effectiveLen === 0) return 0;
+	let padding = 0;
+	let end = base64.length - 1;
+	while (end >= 0 && base64.charCodeAt(end) <= 32) end -= 1;
+	if (end >= 0 && base64[end] === "=") {
+		padding = 1;
+		end -= 1;
+		while (end >= 0 && base64.charCodeAt(end) <= 32) end -= 1;
+		if (end >= 0 && base64[end] === "=") padding = 2;
+	}
+	const estimated = Math.floor(effectiveLen * 3 / 4) - padding;
+	return Math.max(0, estimated);
+}
+const CANONICALIZE_BASE64_CHUNK_SIZE = 8192;
+function isBase64DataChar(code) {
+	return code >= 65 && code <= 90 || code >= 97 && code <= 122 || code >= 48 && code <= 57 || code === 43 || code === 47;
+}
+/**
+* Normalizes and validates a base64 string, returning canonical no-whitespace
+* base64 only when the input has valid alphabet, padding, and length.
+*/
+function canonicalizeBase64(base64) {
+	const chunks = [];
+	let current = "";
+	let cleanedLength = 0;
+	let padding = 0;
+	let sawPadding = false;
+	const append = (char) => {
+		current += char;
+		cleanedLength += 1;
+		if (current.length >= CANONICALIZE_BASE64_CHUNK_SIZE) {
+			chunks.push(current);
+			current = "";
+		}
+	};
+	for (let i = 0; i < base64.length; i += 1) {
+		const code = base64.charCodeAt(i);
+		if (code <= 32) continue;
+		if (code === 61) {
+			padding += 1;
+			if (padding > 2) return;
+			sawPadding = true;
+			append("=");
+			continue;
+		}
+		if (sawPadding || !isBase64DataChar(code)) return;
+		append(base64[i] ?? "");
+	}
+	if (cleanedLength === 0) return;
+	const remainder = cleanedLength % 4;
+	if (remainder !== 0) {
+		if (sawPadding || remainder === 1) return;
+		current += "=".repeat(4 - remainder);
+	}
+	if (current) chunks.push(current);
+	return chunks.join("");
+}
+//#endregion
+//#region packages/ai/src/internal/anthropic-inline-images.ts
+const ANTHROPIC_IMAGE_MEDIA_TYPE_SET = /* @__PURE__ */ new Set([
+	"image/jpeg",
+	"image/png",
+	"image/gif",
+	"image/webp"
+]);
+const ANTHROPIC_INLINE_IMAGES_DECODE_SAFETY_BYTES = 64 * 1024 * 1024;
+function createAnthropicInlineImageBudget() {
+	return { totalBytes: 0 };
+}
+function resolveAnthropicImageMediaType(value) {
+	if (ANTHROPIC_IMAGE_MEDIA_TYPE_SET.has(value)) return value;
+	throw new Error(`Unsupported Anthropic image media type after normalization: ${value}`);
+}
+async function normalizeAnthropicInlineContent(content, budget) {
+	if (!content.some((block) => block.type === "image")) return content.filter((block) => block.type === "text");
+	const inputBytes = content.reduce((total, block) => block.type === "image" ? total + estimateBase64DecodedBytes(block.data) : total, 0);
+	if (budget.totalBytes + inputBytes > ANTHROPIC_INLINE_IMAGES_DECODE_SAFETY_BYTES) throw new Error("Anthropic inline images exceed the 64 MB aggregate decoded safety limit.");
+	const normalized = [];
+	for (const block of content) {
+		if (block.type !== "image") {
+			normalized.push(block);
+			continue;
+		}
+		const normalizedBlocks = await getAiTransportHost().normalizeAnthropicInlineContentBlocks([block]);
+		const outputBytes = normalizedBlocks.reduce((total, normalizedBlock) => normalizedBlock.type === "image" ? total + estimateBase64DecodedBytes(normalizedBlock.data) : total, 0);
+		if (budget.totalBytes + outputBytes > ANTHROPIC_INLINE_IMAGES_DECODE_SAFETY_BYTES) throw new Error("Anthropic inline images exceed the 64 MB aggregate decoded safety limit.");
+		budget.totalBytes += outputBytes;
+		normalized.push(...normalizedBlocks);
+	}
+	return normalized;
+}
+//#endregion
 //#region packages/ai/src/providers/anthropic-auth-headers.ts
 function usesFoundryBearerAuth(model) {
 	return model.provider === "microsoft-foundry" && (model.authHeader === true || hasBearerAuthorizationHeader(model.headers));
@@ -295,6 +395,16 @@ function resolveOriginalAnthropicToolName(name, projection) {
 function readAnthropicUsageTokenCount(value) {
 	return typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : void 0;
 }
+function readAnthropicCacheWriteUsage(usage) {
+	if (!usage.cache_creation || typeof usage.cache_creation !== "object") return {};
+	const cacheCreation = usage.cache_creation;
+	const cacheWrite5m = readAnthropicUsageTokenCount(cacheCreation.ephemeral_5m_input_tokens);
+	const cacheWrite1h = readAnthropicUsageTokenCount(cacheCreation.ephemeral_1h_input_tokens);
+	return {
+		...cacheWrite5m !== void 0 ? { cacheWrite5m } : {},
+		...cacheWrite1h !== void 0 ? { cacheWrite1h } : {}
+	};
+}
 function readAnthropicPromptUsageSnapshot(usage) {
 	const input = readAnthropicUsageTokenCount(usage.input_tokens);
 	const cacheRead = usage.cache_read_input_tokens == null ? 0 : readAnthropicUsageTokenCount(usage.cache_read_input_tokens);
@@ -371,7 +481,7 @@ const toClaudeCodeName = (name) => ccToolLookup.get(name.toLowerCase()) ?? name;
 /**
 * Convert content blocks to Anthropic API format
 */
-function convertContentBlocks(content, isError) {
+async function convertContentBlocks(content, isError, imageBudget) {
 	const text = extractToolResultText(content);
 	const mediaPlaceholder = describeToolResultMediaPlaceholder(content);
 	if (!content.some(isImageWithMediaPayload)) {
@@ -392,12 +502,18 @@ function convertContentBlocks(content, isError) {
 			hasTextBlock = true;
 		}
 		if (!isImageWithMediaPayload(record)) continue;
+		const [normalizedImage] = await normalizeAnthropicInlineContent([{
+			type: "image",
+			data: typeof record.data === "string" ? record.data : "",
+			mimeType: typeof record.mimeType === "string" ? record.mimeType : "image/jpeg"
+		}], imageBudget);
+		if (normalizedImage?.type !== "image") continue;
 		blocks.push({
 			type: "image",
 			source: {
 				type: "base64",
-				media_type: typeof record.mimeType === "string" ? record.mimeType : "image/jpeg",
-				data: record.data
+				media_type: resolveAnthropicImageMediaType(normalizedImage.mimeType),
+				data: normalizedImage.data
 			}
 		});
 	}
@@ -417,7 +533,8 @@ function getAnthropicCompat(model) {
 		supportsEagerToolInputStreaming: model.compat?.supportsEagerToolInputStreaming ?? !isFireworks,
 		supportsLongCacheRetention: model.compat?.supportsLongCacheRetention ?? !isFireworks,
 		sendSessionAffinityHeaders: model.compat?.sendSessionAffinityHeaders ?? (isFireworks || isCloudflareAiGatewayAnthropic),
-		supportsCacheControlOnTools: model.compat?.supportsCacheControlOnTools ?? !isFireworks
+		supportsCacheControlOnTools: model.compat?.supportsCacheControlOnTools ?? !isFireworks,
+		allowEmptySignature: model.compat?.allowEmptySignature ?? false
 	};
 }
 function mergeHeaders(...headerSources) {
@@ -507,7 +624,7 @@ const streamAnthropic = (model, context, options) => {
 				isOAuth = created.isOAuthToken;
 				serverSideFallback = created.serverSideFallback;
 			}
-			const builtParams = buildParams(model, requestContext, isOAuth, requestOptions, serverSideFallback);
+			const builtParams = await buildParams(model, requestContext, isOAuth, requestOptions, serverSideFallback);
 			let params = builtParams.params;
 			const toolProjection = builtParams.toolProjection;
 			const nextParams = await requestOptions?.onPayload?.(params, model);
@@ -516,7 +633,7 @@ const streamAnthropic = (model, context, options) => {
 			const sdkRequestOptions = {
 				...requestOptions?.signal ? { signal: requestOptions.signal } : {},
 				...requestOptions?.timeoutMs !== void 0 ? { timeout: requestOptions.timeoutMs } : {},
-				...requestOptions?.maxRetries !== void 0 ? { maxRetries: requestOptions.maxRetries } : {}
+				maxRetries: requestOptions?.maxRetries ?? 0
 			};
 			const response = await client.messages.create({
 				...params,
@@ -542,6 +659,8 @@ const streamAnthropic = (model, context, options) => {
 				if (cacheReadTokens !== void 0) output.usage.cacheRead = cacheReadTokens;
 				const cacheWriteTokens = event.message.usage.cache_creation_input_tokens == null ? 0 : readAnthropicUsageTokenCount(event.message.usage.cache_creation_input_tokens);
 				if (cacheWriteTokens !== void 0) output.usage.cacheWrite = cacheWriteTokens;
+				const cacheWriteUsage = readAnthropicCacheWriteUsage(event.message.usage);
+				if (cacheWriteUsage.cacheWrite1h !== void 0) output.usage.cacheWrite1h = cacheWriteUsage.cacheWrite1h;
 				output.usage.totalTokens = output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
 				if (messageStartPromptUsage && outputTokens !== void 0) output.usage.contextUsage = {
 					state: "available",
@@ -730,30 +849,34 @@ const streamAnthropic = (model, context, options) => {
 			} else if (event.type === "message_delta") {
 				if (event.delta.stop_reason) if (event.delta.stop_reason === "refusal") applyAnthropicRefusal(output, event.delta.stop_details, model.provider);
 				else output.stopReason = mapStopReason(event.delta.stop_reason);
-				const inputTokens = readAnthropicUsageTokenCount(event.usage.input_tokens);
-				if (inputTokens !== void 0) output.usage.input = inputTokens;
-				const outputTokens = readAnthropicUsageTokenCount(event.usage.output_tokens);
-				if (outputTokens !== void 0) output.usage.output = outputTokens;
-				const cacheReadTokens = readAnthropicUsageTokenCount(event.usage.cache_read_input_tokens);
-				if (cacheReadTokens !== void 0) output.usage.cacheRead = cacheReadTokens;
-				const cacheWriteTokens = readAnthropicUsageTokenCount(event.usage.cache_creation_input_tokens);
-				if (cacheWriteTokens !== void 0) output.usage.cacheWrite = cacheWriteTokens;
-				output.usage.totalTokens = output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
-				const iterationUsage = readLastAnthropicIterationUsage(event.usage);
-				if (iterationUsage.state === "valid") output.usage.contextUsage = {
-					state: "available",
-					promptTokens: iterationUsage.usage.contextPromptTokens,
-					totalTokens: iterationUsage.usage.totalTokens
-				};
-				else if (iterationUsage.state === "invalid") output.usage.contextUsage = { state: "unavailable" };
-				else if (outputTokens !== void 0 && (messageStartPromptUsage !== void 0 || inputTokens !== void 0 && cacheReadTokens !== void 0 && cacheWriteTokens !== void 0)) {
-					const promptTokens = output.usage.input + output.usage.cacheRead + output.usage.cacheWrite;
-					output.usage.contextUsage = {
+				if (event.usage) {
+					const inputTokens = readAnthropicUsageTokenCount(event.usage.input_tokens);
+					if (inputTokens !== void 0) output.usage.input = inputTokens;
+					const outputTokens = readAnthropicUsageTokenCount(event.usage.output_tokens);
+					if (outputTokens !== void 0) output.usage.output = outputTokens;
+					const cacheReadTokens = readAnthropicUsageTokenCount(event.usage.cache_read_input_tokens);
+					if (cacheReadTokens !== void 0) output.usage.cacheRead = cacheReadTokens;
+					const cacheWriteTokens = readAnthropicUsageTokenCount(event.usage.cache_creation_input_tokens);
+					if (cacheWriteTokens !== void 0) output.usage.cacheWrite = cacheWriteTokens;
+					const cacheWriteUsage = readAnthropicCacheWriteUsage(event.usage);
+					if (cacheWriteUsage.cacheWrite1h !== void 0) output.usage.cacheWrite1h = cacheWriteUsage.cacheWrite1h;
+					output.usage.totalTokens = output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
+					const iterationUsage = readLastAnthropicIterationUsage(event.usage);
+					if (iterationUsage.state === "valid") output.usage.contextUsage = {
 						state: "available",
-						promptTokens,
-						totalTokens: promptTokens + output.usage.output
+						promptTokens: iterationUsage.usage.contextPromptTokens,
+						totalTokens: iterationUsage.usage.totalTokens
 					};
-				} else output.usage.contextUsage = { state: "unavailable" };
+					else if (iterationUsage.state === "invalid") output.usage.contextUsage = { state: "unavailable" };
+					else if (outputTokens !== void 0 && (messageStartPromptUsage !== void 0 || inputTokens !== void 0 && cacheReadTokens !== void 0 && cacheWriteTokens !== void 0)) {
+						const promptTokens = output.usage.input + output.usage.cacheRead + output.usage.cacheWrite;
+						output.usage.contextUsage = {
+							state: "available",
+							promptTokens,
+							totalTokens: promptTokens + output.usage.output
+						};
+					} else output.usage.contextUsage = { state: "unavailable" };
+				}
 				calculateCost(costModel, output.usage);
 			}
 			if (requestOptions?.signal?.aborted) throw new Error("Request was aborted");
@@ -842,6 +965,7 @@ const streamSimpleAnthropic = (model, context, options) => {
 	if (!apiKey) throw new Error(`No API key for provider: ${model.provider}`);
 	const base = {
 		...buildBaseOptions(model, options, apiKey),
+		maxTokens: clampMaxTokensToModel(model, options?.maxTokens ?? model.maxTokens),
 		toolChoice: options?.toolChoice
 	};
 	const mandatoryAdaptiveThinking = requiresClaudeAdaptiveThinking(model);
@@ -869,10 +993,11 @@ const streamSimpleAnthropic = (model, context, options) => {
 		});
 	}
 	const adjusted = adjustMaxTokensForThinking(base.maxTokens, model.maxTokens, reasoning, options?.thinkingBudgets);
-	const thinkingEnabled = adjusted.thinkingBudget >= 1024;
+	const thinkingEnabled = adjusted.thinkingBudget >= ANTHROPIC_MIN_THINKING_BUDGET_TOKENS;
+	const maxTokens = thinkingEnabled ? adjusted.maxTokens : clampMaxTokensToModel(model, options?.maxTokens ?? model.maxTokens);
 	return streamAnthropic(model, context, {
 		...base,
-		maxTokens: adjusted.maxTokens,
+		maxTokens,
 		thinkingEnabled,
 		thinkingBudgetTokens: thinkingEnabled ? adjusted.thinkingBudget : void 0
 	});
@@ -998,13 +1123,13 @@ function createClient(model, apiKey, thinkingEnabled, interleavedThinking, useFi
 		serverSideFallback
 	};
 }
-function buildParams(model, context, isOAuthTokenResult, options, serverSideFallback = false) {
+async function buildParams(model, context, isOAuthTokenResult, options, serverSideFallback = false) {
 	const mandatoryAdaptiveThinking = requiresClaudeAdaptiveThinking(model);
 	const replayThinkingEnabled = mandatoryAdaptiveThinking || options?.thinkingEnabled === true;
 	const { cacheControl } = getCacheControl(model, options?.cacheRetention);
 	const system = buildAnthropicSystemBlocks(context.systemPrompt, isOAuthTokenResult, cacheControl);
-	const compat = context.tools ? getAnthropicCompat(model) : void 0;
-	const convertedTools = context.tools && compat ? convertTools(context.tools, isOAuthTokenResult, compat.supportsEagerToolInputStreaming, compat.supportsCacheControlOnTools ? cacheControl : void 0) : void 0;
+	const compat = getAnthropicCompat(model);
+	const convertedTools = context.tools ? convertTools(context.tools, isOAuthTokenResult, compat.supportsEagerToolInputStreaming, compat.supportsCacheControlOnTools ? cacheControl : void 0) : void 0;
 	const tools = convertedTools?.tools;
 	const toolProjection = convertedTools?.projection;
 	const systemCacheControlCount = countNativeCacheControlMarkers(system);
@@ -1012,7 +1137,7 @@ function buildParams(model, context, isOAuthTokenResult, options, serverSideFall
 	const messageCacheControlLimit = Math.max(0, ANTHROPIC_CACHE_CONTROL_LIMIT - systemCacheControlCount - toolCacheControlCount);
 	const params = {
 		model: model.id,
-		messages: convertMessages(context.messages, model, isOAuthTokenResult, cacheControl, messageCacheControlLimit, replayThinkingEnabled),
+		messages: await convertMessages(context.messages, model, isOAuthTokenResult, cacheControl, messageCacheControlLimit, replayThinkingEnabled, compat.allowEmptySignature),
 		max_tokens: options?.maxTokens ?? model.maxTokens,
 		stream: true
 	};
@@ -1055,8 +1180,9 @@ function buildParams(model, context, isOAuthTokenResult, options, serverSideFall
 function normalizeToolCallId(id) {
 	return id.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 64);
 }
-function convertMessages(messages, model, isOAuthTokenValue, cacheControl, messageCacheControlLimit = 4, replayThinkingEnabled = true) {
+async function convertMessages(messages, model, isOAuthTokenValue, cacheControl, messageCacheControlLimit = 4, replayThinkingEnabled = true, allowEmptySignature = false) {
 	const params = [];
+	const imageBudget = createAnthropicInlineImageBudget();
 	const cacheBreakpointOptOutParamIndexes = /* @__PURE__ */ new Set();
 	const transformedMessages = transformMessages(messages, model, normalizeToolCallId);
 	const activeToolTurnAssistantIndex = replayThinkingEnabled ? -1 : findActiveAnthropicToolTurnAssistantIndex(transformedMessages);
@@ -1074,7 +1200,7 @@ function convertMessages(messages, model, isOAuthTokenValue, cacheControl, messa
 					});
 				}
 			} else {
-				const filteredBlocks = msg.content.map((item) => {
+				const filteredBlocks = (await normalizeAnthropicInlineContent(msg.content, imageBudget)).map((item) => {
 					if (item.type === "text") return {
 						type: "text",
 						text: sanitizeSurrogates(item.text)
@@ -1083,7 +1209,7 @@ function convertMessages(messages, model, isOAuthTokenValue, cacheControl, messa
 						type: "image",
 						source: {
 							type: "base64",
-							media_type: item.mimeType,
+							media_type: resolveAnthropicImageMediaType(item.mimeType),
 							data: item.data
 						}
 					};
@@ -1123,7 +1249,7 @@ function convertMessages(messages, model, isOAuthTokenValue, cacheControl, messa
 				const thinkingSignature = block.thinkingSignature?.trim();
 				const hasNativeThinkingSignature = Boolean(thinkingSignature) && thinkingSignature !== "reasoning_content";
 				if (block.thinking.trim().length === 0 && !hasNativeThinkingSignature) continue;
-				if (!thinkingSignature) blocks.push({
+				if (!thinkingSignature && !allowEmptySignature) blocks.push({
 					type: "text",
 					text: sanitizeSurrogates(block.thinking)
 				});
@@ -1132,7 +1258,7 @@ function convertMessages(messages, model, isOAuthTokenValue, cacheControl, messa
 					blocks.push({
 						type: "thinking",
 						thinking: block.thinking,
-						signature: thinkingSignature
+						signature: thinkingSignature ?? ""
 					});
 				}
 			} else if (block.type === "toolCall") blocks.push({
@@ -1155,7 +1281,7 @@ function convertMessages(messages, model, isOAuthTokenValue, cacheControl, messa
 			toolResults.push({
 				type: "tool_result",
 				tool_use_id: msg.toolCallId,
-				content: convertContentBlocks(msg.content, msg.isError),
+				content: await convertContentBlocks(msg.content, msg.isError, imageBudget),
 				is_error: msg.isError
 			});
 			let j = i + 1;
@@ -1165,7 +1291,7 @@ function convertMessages(messages, model, isOAuthTokenValue, cacheControl, messa
 				toolResults.push({
 					type: "tool_result",
 					tool_use_id: nextMsg.toolCallId,
-					content: convertContentBlocks(nextMsg.content, nextMsg.isError),
+					content: await convertContentBlocks(nextMsg.content, nextMsg.isError, imageBudget),
 					is_error: nextMsg.isError
 				});
 				j++;
@@ -1302,4 +1428,4 @@ function mapStopReason(reason) {
 	}
 }
 //#endregion
-export { readAnthropicFallbackBoundary as _, readAnthropicUsageTokenCount as a, usesFoundryBearerAuth as b, reconcileAnthropicToolChoice as c, findActiveAnthropicToolTurnAssistantIndex as d, ANTHROPIC_SERVER_SIDE_FALLBACK_BETA as f, buildAnthropicServerSideFallbacks as g, applyAnthropicFallbackBoundary as h, readAnthropicPromptUsageSnapshot as i, resolveOriginalAnthropicToolName as l, CLAUDE_FABLE_5_FALLBACK_MODEL_COST as m, streamAnthropic as n, readLastAnthropicIterationUsage as o, CLAUDE_FABLE_5_FALLBACK_MODEL as p, streamSimpleAnthropic as r, projectAnthropicTools as s, anthropic_exports as t, ANTHROPIC_OMITTED_REASONING_TEXT as u, applyAnthropicRefusal as v, omitFoundryBearerCredentialHeaders as y };
+export { normalizeAnthropicInlineContent as C, createAnthropicInlineImageBudget as S, canonicalizeBase64 as T, buildAnthropicServerSideFallbacks as _, readAnthropicPromptUsageSnapshot as a, omitFoundryBearerCredentialHeaders as b, projectAnthropicTools as c, ANTHROPIC_OMITTED_REASONING_TEXT as d, findActiveAnthropicToolTurnAssistantIndex as f, applyAnthropicFallbackBoundary as g, CLAUDE_FABLE_5_FALLBACK_MODEL_COST as h, readAnthropicCacheWriteUsage as i, reconcileAnthropicToolChoice as l, CLAUDE_FABLE_5_FALLBACK_MODEL as m, streamAnthropic as n, readAnthropicUsageTokenCount as o, ANTHROPIC_SERVER_SIDE_FALLBACK_BETA as p, streamSimpleAnthropic as r, readLastAnthropicIterationUsage as s, anthropic_exports as t, resolveOriginalAnthropicToolName as u, readAnthropicFallbackBoundary as v, resolveAnthropicImageMediaType as w, usesFoundryBearerAuth as x, applyAnthropicRefusal as y };

package/dist/{api-registry-Byvoz8Ha.d.mts → api-registry-BMphkFf1.d.mts}

@@ -1,5 +1,4 @@
-import { E as Model, F as SimpleStreamOptions, R as StreamFunction, n as Api, o as AssistantMessageEventStreamContract, u as Context, z as StreamOptions } from "./types-Cx2zJtyz.mjs";
-
+import { E as Model, F as SimpleStreamOptions, R as StreamFunction, n as Api, o as AssistantMessageEventStreamContract, u as Context, z as StreamOptions } from "./types-AFwwWium.mjs";
 //#region packages/ai/src/api-registry.d.ts
 /** Runtime stream adapter signature stored in the API provider registry. */
 type ApiStreamFunction = (model: Model, context: Context, options?: StreamOptions) => AssistantMessageEventStreamContract;

package/dist/{azure-openai-responses-CHi7Wo4A.mjs → azure-openai-responses-CbrpVeEv.mjs}

@@ -1,10 +1,10 @@
-import { n as getEnvApiKey } from "./env-api-keys-DnhYpc27.mjs";
+import { n as getEnvApiKey } from "./env-api-keys-DrgeBuva.mjs";
 import { t as AssistantMessageEventStream } from "./event-stream-D8n2uFee.mjs";
-import { n as getAiTransportHost } from "./host-4t713IeR.mjs";
-import { l as buildBaseOptions } from "./transform-messages-BmS70CP5.mjs";
-import { a as resolveResponsesReasoningEffort, n as convertResponsesMessages, o as runResponsesStreamLifecycle, r as createResponsesAssistantOutput, t as applyCommonResponsesParams } from "./openai-responses-shared-Befb4D6R.mjs";
+import { n as getAiTransportHost } from "./host-XYGZcgO8.mjs";
+import { D as buildBaseOptions } from "./system-prompt-cache-boundary-CbHeV4_l.mjs";
+import { a as resolveResponsesReasoningEffort, n as convertResponsesMessages, o as runResponsesStreamLifecycle, r as createResponsesAssistantOutput, t as applyCommonResponsesParams } from "./openai-responses-shared-CzCurmY1.mjs";
 import { i as resolveAzureDeploymentNameFromMap, t as isOpenAICompatibleAzureResponsesBaseUrl } from "./azure-openai-responses-client-compat-C7K7QfUE.mjs";
-import { a as clampOpenAIPromptCacheKey } from "./openai-tool-projection-CFqm42J2.mjs";
+import { a as clampOpenAIPromptCacheKey } from "./openai-tool-projection-ITOU9bG1.mjs";
 import OpenAI, { AzureOpenAI } from "openai";
 //#region packages/ai/src/providers/azure-openai-responses.ts
 const DEFAULT_AZURE_API_VERSION = "v1";
@@ -94,11 +94,8 @@ function resolveAzureConfig(model, options) {
 	};
 }
 function createClient(model, apiKeyInput, options) {
-	let apiKey = apiKeyInput;
-	if (!apiKey) {
-		if (!process.env.AZURE_OPENAI_API_KEY) throw new Error("Azure OpenAI API key is required. Set AZURE_OPENAI_API_KEY environment variable or pass it as an argument.");
-		apiKey = process.env.AZURE_OPENAI_API_KEY;
-	}
+	const apiKey = apiKeyInput.trim();
+	if (!apiKey) throw new Error("Azure OpenAI API key is required. Set AZURE_OPENAI_API_KEY environment variable or pass it as an argument.");
 	const headers = { ...model.headers };
 	if (options?.headers) Object.assign(headers, options.headers);
 	const { baseUrl, apiVersion } = resolveAzureConfig(model, options);
@@ -127,7 +124,8 @@ function buildParams(model, context, options, deploymentName) {
 		model: deploymentName,
 		input: convertResponsesMessages(model, context, AZURE_TOOL_CALL_PROVIDERS),
 		stream: true,
-		prompt_cache_key: options?.cacheRetention === "none" ? void 0 : clampOpenAIPromptCacheKey(options?.promptCacheKey ?? options?.sessionId)
+		prompt_cache_key: options?.cacheRetention === "none" ? void 0 : clampOpenAIPromptCacheKey(options?.promptCacheKey ?? options?.sessionId),
+		store: false
 	};
 	applyCommonResponsesParams(params, model, context, options);
 	return params;

package/dist/{env-api-keys-DnhYpc27.mjs → env-api-keys-DrgeBuva.mjs}

@@ -82,7 +82,7 @@ function getProcEnv(key) {
 	return procEnvCache.get(key);
 }
 function getEnvValue(key) {
-	return getProcessEnv()?.[key] || getProcEnv(key);
+	return (getProcessEnv()?.[key] || getProcEnv(key))?.trim() || void 0;
 }
 let cachedVertexAdcCredentialsExists = null;
 function hasVertexAdcCredentials() {

package/dist/{event-stream-uQspnL8S.d.mts → event-stream-Douf9dob.d.mts}

@@ -1,4 +1,4 @@
-import { a as AssistantMessageEvent, i as AssistantMessage, o as AssistantMessageEventStreamContract } from "./types-Cx2zJtyz.mjs";
+import { a as AssistantMessageEvent, i as AssistantMessage, o as AssistantMessageEventStreamContract } from "./types-AFwwWium.mjs";
 
 //#region packages/llm-core/src/utils/event-stream.d.ts
 /** Generic async-iterable event stream with a separately awaited final result. */

package/dist/event-stream.d.mts

@@ -1,2 +1,2 @@
-import { n as EventStream, r as createAssistantMessageEventStream, t as AssistantMessageEventStream } from "./event-stream-uQspnL8S.mjs";
+import { n as EventStream, r as createAssistantMessageEventStream, t as AssistantMessageEventStream } from "./event-stream-Douf9dob.mjs";
 export { AssistantMessageEventStream, EventStream, createAssistantMessageEventStream };

package/dist/{google-BmsqsNwR.mjs → google-Cs62KA7t.mjs}

@@ -1,8 +1,8 @@
-import { n as getEnvApiKey } from "./env-api-keys-DnhYpc27.mjs";
+import { n as getEnvApiKey } from "./env-api-keys-DrgeBuva.mjs";
 import { t as AssistantMessageEventStream } from "./event-stream-D8n2uFee.mjs";
-import { n as getAiTransportHost, r as resolveAiTransportHeaderSentinels } from "./host-4t713IeR.mjs";
-import { l as buildBaseOptions } from "./transform-messages-BmS70CP5.mjs";
-import { a as runGoogleGenerateContentLifecycle, i as getDisabledGoogleThinkingConfig, n as buildGoogleSimpleThinking, r as createGoogleAssistantOutput, t as buildGoogleGenerateContentParams } from "./google-shared-Dx2aba47.mjs";
+import { n as getAiTransportHost, r as resolveAiTransportHeaderSentinels } from "./host-XYGZcgO8.mjs";
+import { D as buildBaseOptions } from "./system-prompt-cache-boundary-CbHeV4_l.mjs";
+import { a as runGoogleGenerateContentLifecycle, i as getDisabledGoogleThinkingConfig, n as buildGoogleSimpleThinking, r as createGoogleAssistantOutput, t as buildGoogleGenerateContentParams } from "./google-shared-CMLI-tCZ.mjs";
 import { GoogleGenAI } from "@google/genai";
 //#region packages/ai/src/providers/google.ts
 let toolCallCounter = 0;

package/dist/{google-shared-Dx2aba47.mjs → google-shared-CMLI-tCZ.mjs}

@@ -1,6 +1,7 @@
-import { n as calculateCost, r as clampThinkingLevel } from "./model-utils-Q1LSRIdo.mjs";
+import { C as isImageWithMediaPayload, a as stripSystemPromptCacheBoundary, c as transformMessages, x as extractToolResultText, y as describeToolResultMediaPlaceholder } from "./system-prompt-cache-boundary-CbHeV4_l.mjs";
 import { t as sanitizeSurrogates } from "./sanitize-unicode-DT5o51ur.mjs";
-import { C as stripSystemPromptCacheBoundary, i as extractToolResultText, n as describeToolResultMediaPlaceholder, o as isImageWithMediaPayload, t as transformMessages } from "./transform-messages-BmS70CP5.mjs";
+import { n as calculateCost, r as clampThinkingLevel } from "./model-utils-1GiZ2_rr.mjs";
+import { t as formatProviderError } from "./provider-error-C4VvV_3t.mjs";
 import { FinishReason, FunctionCallingConfigMode, ThinkingLevel } from "@google/genai";
 //#region packages/ai/src/providers/google-shared.ts
 /**
@@ -273,7 +274,7 @@ async function runGoogleGenerateContentLifecycle(params) {
 	} catch (error) {
 		for (const block of output.content) if ("index" in block) delete block.index;
 		output.stopReason = options?.signal?.aborted ? "aborted" : "error";
-		output.errorMessage = error instanceof Error ? error.message : JSON.stringify(error);
+		output.errorMessage = formatProviderError(error);
 		stream.push({
 			type: "error",
 			reason: output.stopReason,

package/dist/{google-vertex-HZ_0rTwS.mjs → google-vertex-B4SD3U0f.mjs}

@@ -1,7 +1,7 @@
 import { t as AssistantMessageEventStream } from "./event-stream-D8n2uFee.mjs";
-import { n as getAiTransportHost, r as resolveAiTransportHeaderSentinels } from "./host-4t713IeR.mjs";
-import { l as buildBaseOptions } from "./transform-messages-BmS70CP5.mjs";
-import { a as runGoogleGenerateContentLifecycle, n as buildGoogleSimpleThinking, r as createGoogleAssistantOutput, t as buildGoogleGenerateContentParams } from "./google-shared-Dx2aba47.mjs";
+import { n as getAiTransportHost, r as resolveAiTransportHeaderSentinels } from "./host-XYGZcgO8.mjs";
+import { D as buildBaseOptions } from "./system-prompt-cache-boundary-CbHeV4_l.mjs";
+import { a as runGoogleGenerateContentLifecycle, n as buildGoogleSimpleThinking, r as createGoogleAssistantOutput, t as buildGoogleGenerateContentParams } from "./google-shared-CMLI-tCZ.mjs";
 import { GoogleGenAI, ResourceScope } from "@google/genai";
 //#region packages/ai/src/providers/google-vertex.ts
 const API_VERSION = "v1";