@openclaw/ai 2026.9.1-beta.1 → 2026.9.2

package/dist/{openai-completions-stream-DUlGwRMz.mjs → openai-completions-stream-BsrBe3Gg.mjs}

@@ -1,18 +1,18 @@
 import { c as normalizeLowercaseStringOrEmpty, i as asOptionalObjectRecord, o as isRecord, t as truncateUtf16Safe } from "./utf16-slice-qz3nsy87.mjs";
-import { t as sanitizeSurrogates } from "./sanitize-unicode-DCkltN94.mjs";
-import { a as describeToolResultMediaPlaceholder, c as extractToolResultText, d as isImageWithMediaPayload, n as getAiTransportHost } from "./host-D51fmkH6.mjs";
-import { n as clampOpenAIPromptCacheKey } from "./openai-prompt-cache-mZTCdRPo.mjs";
-import { l as stripSystemPromptCacheBoundary } from "./simple-options-CFj7x3J4.mjs";
-import { o as resolveProviderRequestCapabilities, r as resolveModelRequestTimeoutMs, s as resolveProviderRequestPolicyConfig } from "./tool-schema-json-projection-qvTEs5Am.mjs";
-import { i as isCodeModeModelVisibleToolName, m as sha256Hex } from "./transport-utils-LJ1_rbi-.mjs";
-import { s as finalizeTerminalToolCallArguments } from "./transport-stream-shared-ljShVIZB.mjs";
+import { t as sanitizeSurrogates } from "./sanitize-unicode-D6xUvZaS.mjs";
+import { a as describeToolResultMediaPlaceholder, c as extractToolResultText, d as isImageWithMediaPayload, n as getAiTransportHost } from "./host-CEvLw30U.mjs";
+import { t as createReasoningTagTextPartitioner } from "./reasoning-tag-text-partitioner-DLCNXki6.mjs";
+import { n as clampOpenAIPromptCacheKey } from "./openai-prompt-cache-CGnVB74a.mjs";
+import { d as stripSystemPromptCacheBoundary } from "./simple-options-BjHCCh4v.mjs";
+import { o as resolveProviderRequestCapabilities, r as resolveModelRequestTimeoutMs, s as resolveProviderRequestPolicyConfig } from "./tool-schema-json-projection-mJhXDcyz.mjs";
+import { i as isCodeModeModelVisibleToolName, p as sha256Hex } from "./transport-utils-7il795_9.mjs";
+import { s as finalizeTerminalToolCallArguments } from "./transport-stream-shared-CZqMhfIw.mjs";
 import { r as parseStreamingJson, t as createToolArgumentPreviewSchedule } from "./json-parse-BuAJEbdW.mjs";
 import { t as notifyLlmRequestActivity } from "./llm-request-activity-BjtkplhG.mjs";
-import { A as readOpenAICompletionsReasoningBatch, D as measureUtf8AppendBytes, N as throwIfModelStreamAborted, O as parseOpenAICompletionsUsage, S as createModelStreamCooperativeScheduler, T as isOpenAICompletionsThinkingEnabled, a as resolveOpenAIProjectedToolsStrictToolFlag, k as readOpenAICompletionsContentDeltas, t as findOpenAIStrictToolProjectionDiagnostics } from "./openai-tool-schema-ho-hIkQf.mjs";
-import { a as withFirstStreamEventTimeout } from "./stream-first-event-timeout-MK28puvq.mjs";
-import { t as transformProviderMessages } from "./provider-transcript-transform-C01bQ5a3.mjs";
+import { A as readOpenAICompletionsReasoningBatch, D as measureUtf8AppendBytes, N as throwIfModelStreamAborted, O as parseOpenAICompletionsUsage, S as createModelStreamCooperativeScheduler, T as isOpenAICompletionsThinkingEnabled, a as resolveOpenAIProjectedToolsStrictToolFlag, k as readOpenAICompletionsContentDeltas, t as findOpenAIStrictToolProjectionDiagnostics } from "./openai-tool-schema-_pTAJqKF.mjs";
+import { a as withFirstStreamEventTimeout } from "./stream-first-event-timeout-DcNjoFQE.mjs";
+import { t as transformProviderMessages } from "./provider-transcript-transform-BaMbI1hr.mjs";
 import { t as mapOpenAIStopReason } from "./openai-stop-reason-Drnn_6Qj.mjs";
-import { t as createReasoningTagTextPartitioner } from "./reasoning-tag-text-partitioner-5ygO2rZc.mjs";
 import { randomUUID } from "node:crypto";
 //#region packages/ai/src/utils/assistant-text-phase.ts
 const EMPTY_ASSISTANT_TEXT_BLOCK_SET = /* @__PURE__ */ new Set();
@@ -34,7 +34,7 @@ function tagUnphasedText(content, phase, idPrefix) {
 	const tagged = /* @__PURE__ */ new Map();
 	for (const block of textBlocks) {
 		if (block.text.trim().length === 0 || block.textSignature !== void 0) continue;
-		const signature = encodeAssistantTextSignatureV1(`${idPrefix}-${phaseIndex}`, phase);
+		const signature = encodeAssistantTextSignatureV1(`${idPrefix}-${phaseIndex}-${randomUUID().replaceAll("-", "").slice(0, 24)}`, phase);
 		block.textSignature = signature;
 		tagged.set(block, signature);
 		phaseIndex += 1;
@@ -67,6 +67,14 @@ function rememberPendingCommentaryTags(target, tagged) {
 	for (const [block, signature] of tagged) target.set(block, signature);
 }
 //#endregion
+//#region packages/ai/src/transports/deepseek-dsml-grammar.ts
+const DEEPSEEK_DSML_MARKERS = [
+	"|",
+	"｜",
+	"｜｜"
+].map((bar) => `${bar}DSML${bar}`);
+const DEEPSEEK_DSML_MARKER_PATTERN = `(${DEEPSEEK_DSML_MARKERS.map((marker) => marker.replaceAll("|", "\\|")).join("|")})`;
+//#endregion
 //#region packages/ai/src/transports/deepseek-text-filter.ts
 /**
 * DeepSeek DSML streaming text filter.
@@ -79,38 +87,35 @@ const DSML_KINDS = [
 	"tool_call",
 	"function_calls"
 ];
-const DSML_BARS = ["|", "｜"];
-const DSML_OPEN_TOKENS = DSML_BARS.flatMap((bar) => DSML_KINDS.map((kind) => `<${bar}DSML${bar}${kind}>`));
-const DSML_CLOSE_TOKENS = DSML_BARS.flatMap((bar) => DSML_KINDS.map((kind) => `</${bar}DSML${bar}${kind}>`));
+const DSML_OPEN_TOKENS = DEEPSEEK_DSML_MARKERS.flatMap((marker) => DSML_KINDS.map((kind) => `<${marker}${kind}>`));
 const MAX_OPEN_TOKEN_LEN = Math.max(...DSML_OPEN_TOKENS.map((token) => token.length));
-const MAX_CLOSE_TOKEN_LEN = Math.max(...DSML_CLOSE_TOKENS.map((token) => token.length));
 /** Create an incremental text filter that strips DeepSeek DSML tool blocks. */
 function createDeepSeekTextFilter() {
 	let buffer = "";
-	let insideDsml = false;
+	let closeToken;
 	const consume = (final) => {
 		const output = [];
 		const emit = (text) => {
 			if (text) output.push(text);
 		};
 		while (buffer) {
-			if (insideDsml) {
-				const close = findEarliestToken(buffer, DSML_CLOSE_TOKENS);
-				if (close) {
-					buffer = buffer.slice(close.index + close.token.length);
-					insideDsml = false;
+			if (closeToken) {
+				const closeIndex = buffer.indexOf(closeToken);
+				if (closeIndex !== -1) {
+					buffer = buffer.slice(closeIndex + closeToken.length);
+					closeToken = void 0;
 					continue;
 				}
-				const keep = final ? 0 : Math.min(buffer.length, MAX_CLOSE_TOKEN_LEN - 1);
+				const keep = final ? 0 : Math.min(buffer.length, closeToken.length - 1);
 				buffer = buffer.slice(buffer.length - keep);
-				if (final) insideDsml = false;
+				if (final) closeToken = void 0;
 				return output;
 			}
 			const open = findEarliestToken(buffer, DSML_OPEN_TOKENS);
 			if (open) {
 				emit(buffer.slice(0, open.index));
 				buffer = buffer.slice(open.index + open.token.length);
-				insideDsml = true;
+				closeToken = open.token.replace("<", "</");
 				continue;
 			}
 			if (final) {
@@ -215,7 +220,7 @@ function resolveOpenAICompletionsCompatDefaults(input) {
 	const isTogether = knownProviderFamily === "together" || input.baseUrl?.includes("api.together.ai") === true || input.baseUrl?.includes("api.together.xyz") === true || isDefaultRoute && isDefaultRouteProvider(input.provider, "together");
 	const isCloudflareAiGateway = provider === "cloudflare-ai-gateway" || input.baseUrl?.includes("gateway.ai.cloudflare.com") === true;
 	const isXiaomi = endpointClass === "xiaomi-native" || isDefaultRoute && isDefaultRouteProvider(input.provider, "xiaomi");
-	const isNonStandard = endpointClass === "cerebras-native" || endpointClass === "chutes-native" || endpointClass === "deepseek-native" || endpointClass === "mistral-public" || endpointClass === "opencode-native" || endpointClass === "xai-native" || isXiaomi || isZai || isDefaultRoute && isDefaultRouteProvider(input.provider, "cerebras", "chutes", "deepseek", "opencode", "xai");
+	const isNonStandard = endpointClass === "cerebras-native" || endpointClass === "chutes-native" || endpointClass === "deepseek-native" || endpointClass === "mistral-public" || endpointClass === "opencode-native" || endpointClass === "opencode-go-native" || endpointClass === "xai-native" || isXiaomi || isZai || isDefaultRoute && isDefaultRouteProvider(input.provider, "cerebras", "chutes", "deepseek", "opencode", "xai");
 	const isOpenRouterLike = input.provider === "openrouter" || endpointClass === "openrouter";
 	const isLocalEndpoint = endpointClass === "local";
 	const usesMaxTokens = endpointClass === "chutes-native" || endpointClass === "mistral-public" || knownProviderFamily === "mistral" || isMoonshot || isCloudflareAiGateway || isZai || isTogether || isDefaultRoute && isDefaultRouteProvider(provider, "chutes");
@@ -231,9 +236,9 @@ function resolveOpenAICompletionsCompatDefaults(input) {
 		supportsJsonSchemaResponseFormat: (endpointClass === "openai-public" || isDefaultRoute && isDefaultRouteProvider(provider, "openai")) && isKnownOpenAIJsonSchemaModelId(modelId),
 		requiresReasoningContentOnAssistantMessages: isDeepSeek || isXiaomi,
 		requiresNonEmptyUserOrAssistantMessage: isModelStudioLike,
-		cacheControlFormat: provider === "openrouter" && modelId?.startsWith("anthropic/") === true ? "anthropic" : void 0,
+		cacheControlFormat: isModelStudioLike && endpointClass !== "custom" || provider === "openrouter" && modelId?.startsWith("anthropic/") === true ? "anthropic" : void 0,
 		sessionAffinityFormat: isOpenRouterLike ? "openrouter" : "openai",
-		supportsLongCacheRetention: provider !== "cloudflare-workers-ai" && provider !== "cloudflare-ai-gateway" && knownProviderFamily !== "together" && !input.baseUrl?.includes("api.cloudflare.com") && !input.baseUrl?.includes("gateway.ai.cloudflare.com") && !input.baseUrl?.includes("api.together.ai") && !input.baseUrl?.includes("api.together.xyz")
+		supportsLongCacheRetention: !isModelStudioLike && provider !== "cloudflare-workers-ai" && provider !== "cloudflare-ai-gateway" && knownProviderFamily !== "together" && !input.baseUrl?.includes("api.cloudflare.com") && !input.baseUrl?.includes("gateway.ai.cloudflare.com") && !input.baseUrl?.includes("api.together.ai") && !input.baseUrl?.includes("api.together.xyz")
 	};
 }
 function resolveOpenAICompletionsCompatDefaultsFromCapabilities(input) {
@@ -391,11 +396,13 @@ function createOpenAICompletionsToolCallDeltaNormalizer() {
 		}];
 		const hadPendingLegacyCall = pendingLegacyToolCall !== void 0;
 		const leadingDeltas = [];
-		if (hasOrdinaryContent(ordinaryDelta)) if (hadPendingLegacyCall) bufferFollowingDelta(ordinaryDelta);
-		else leadingDeltas.push({
-			delta: ordinaryDelta,
-			toolCalls: []
-		});
+		if (hasOrdinaryContent(ordinaryDelta)) {
+			if (hadPendingLegacyCall) bufferFollowingDelta(ordinaryDelta);
+			else leadingDeltas.push({
+				delta: ordinaryDelta,
+				toolCalls: []
+			});
+		}
 		if (functionCall && (functionCall.name || functionCall.arguments)) {
 			const nextFunctionName = pendingLegacyToolCall?.function?.name ? void 0 : functionCall.name;
 			const nextArgumentBytes = Buffer.byteLength(functionCall.arguments ?? "", "utf8") + Buffer.byteLength(nextFunctionName ?? "", "utf8");
@@ -555,23 +562,24 @@ function convertMessages(model, context, compat, options = {}) {
 				role: "assistant",
 				content: compat.requiresAssistantAfterToolResult ? "" : null
 			};
-			const assistantTextParts = msg.content.filter(isTextContentBlock).filter((block) => block.text.trim().length > 0).map((block) => ({
-				type: "text",
-				text: sanitizeSurrogates(block.text)
-			}));
-			const assistantText = assistantTextParts.map((part) => part.text).join("\n");
+			const assistantTexts = msg.content.filter(isTextContentBlock).filter((block) => block.text.trim().length > 0).map((block) => sanitizeSurrogates(block.text));
+			const assistantText = assistantTexts.join("\n");
 			const nonEmptyThinkingBlocks = msg.content.filter(isThinkingContentBlock).filter((block) => block.thinking.trim().length > 0);
-			if (nonEmptyThinkingBlocks.length > 0) if (compat.requiresThinkingAsText) assistantMsg.content = [{
-				type: "text",
-				text: nonEmptyThinkingBlocks.map((block) => sanitizeSurrogates(block.thinking)).join("\n\n")
-			}, ...assistantTextParts];
-			else {
-				if (assistantText.length > 0) assistantMsg.content = assistantText;
-				let signature = nonEmptyThinkingBlocks.at(0)?.thinkingSignature;
-				if (model.provider === "opencode-go" && signature === "reasoning") signature = "reasoning_content";
-				if (signature && signature.length > 0) assistantMsg[signature] = nonEmptyThinkingBlocks.map((block) => block.thinking).join("\n");
-			}
-			else if (assistantText.length > 0) assistantMsg.content = assistantText;
+			if (nonEmptyThinkingBlocks.length > 0) {
+				if (compat.requiresThinkingAsText) assistantMsg.content = [{
+					type: "text",
+					text: nonEmptyThinkingBlocks.map((block) => sanitizeSurrogates(block.thinking)).join("\n\n")
+				}, ...assistantTexts.map((text) => ({
+					type: "text",
+					text
+				}))];
+				else {
+					if (assistantText.length > 0) assistantMsg.content = assistantText;
+					let signature = nonEmptyThinkingBlocks.at(0)?.thinkingSignature;
+					if (model.provider === "opencode-go" && signature === "reasoning") signature = "reasoning_content";
+					if (signature && signature.length > 0) assistantMsg[signature] = nonEmptyThinkingBlocks.map((block) => block.thinking).join("\n");
+				}
+			} else if (assistantText.length > 0) assistantMsg.content = assistantText;
 			const toolCalls = msg.content.filter(isToolCallBlock);
 			if (toolCalls.length > 0) {
 				assistantMsg.tool_calls = toolCalls.map((toolCall) => ({
@@ -868,17 +876,20 @@ function resolveOpenAISdkTimeoutMs(model, timeoutMs) {
 }
 function buildOpenAISdkClientOptions(model) {
 	const timeout = resolveOpenAISdkTimeoutMs(model);
-	return timeout === void 0 ? {} : { timeout };
+	return {
+		...timeout === void 0 ? {} : { timeout },
+		maxRetries: 0
+	};
 }
 function buildOpenAISdkRequestOptions(model, signal, options) {
 	const timeout = resolveOpenAISdkTimeoutMs(model, options?.timeoutMs);
 	const headers = options?.stream === true && usesNativeOpenAICodexResponsesBackend(model) ? { Accept: "text/event-stream" } : void 0;
-	if (timeout === void 0 && options?.maxRetries === void 0 && !signal && !headers) return;
+	if (timeout === void 0 && !signal && !headers) return;
 	return {
 		...headers ? { headers } : {},
 		...signal ? { signal } : {},
 		...timeout !== void 0 ? { timeout } : {},
-		...options?.maxRetries !== void 0 ? { maxRetries: options.maxRetries } : {}
+		maxRetries: 0
 	};
 }
 function getCompat(model) {
@@ -896,39 +907,29 @@ function getCompat(model) {
 }
 //#endregion
 //#region packages/ai/src/transports/openai-completions-dsml.ts
-const DEEPSEEK_DSML_BARS = ["|", "｜"];
 const DEEPSEEK_DSML_TOOL_KINDS = [
 	"tool_calls",
 	"tool_call",
 	"function_calls"
 ];
-const DEEPSEEK_DSML_TOOL_OPEN_TOKENS = DEEPSEEK_DSML_BARS.flatMap((bar) => DEEPSEEK_DSML_TOOL_KINDS.map((kind) => `<${bar}DSML${bar}${kind}>`));
-const DEEPSEEK_DSML_TOOL_CLOSE_TOKENS = DEEPSEEK_DSML_BARS.flatMap((bar) => DEEPSEEK_DSML_TOOL_KINDS.map((kind) => `</${bar}DSML${bar}${kind}>`));
-const DEEPSEEK_DSML_INVOKE_OPEN_PREFIXES = DEEPSEEK_DSML_BARS.map((bar) => `<${bar}DSML${bar}invoke`);
-const DEEPSEEK_DSML_INVOKE_CLOSE_TOKENS = DEEPSEEK_DSML_BARS.map((bar) => `</${bar}DSML${bar}invoke>`);
+const DEEPSEEK_DSML_TOOL_OPEN_TOKENS = DEEPSEEK_DSML_MARKERS.flatMap((marker) => DEEPSEEK_DSML_TOOL_KINDS.map((kind) => `<${marker}${kind}>`));
+const DEEPSEEK_DSML_INVOKE_OPEN_PREFIXES = DEEPSEEK_DSML_MARKERS.map((marker) => `<${marker}invoke`);
+const DEEPSEEK_DSML_INVOKE_OPEN_TAG = new RegExp(`^<${DEEPSEEK_DSML_MARKER_PATTERN}invoke\\b[^<>]*>$`);
 const DEEPSEEK_DSML_TOOL_MAX_OPEN_TOKEN_LEN = Math.max(...DEEPSEEK_DSML_TOOL_OPEN_TOKENS.map((token) => token.length));
-const DEEPSEEK_DSML_RECOVERY_MAX_BOUNDARY_LEN = Math.max(...DEEPSEEK_DSML_TOOL_OPEN_TOKENS.map((token) => token.length), ...DEEPSEEK_DSML_TOOL_CLOSE_TOKENS.map((token) => token.length), ...DEEPSEEK_DSML_INVOKE_OPEN_PREFIXES.map((token) => token.length), ...DEEPSEEK_DSML_INVOKE_CLOSE_TOKENS.map((token) => token.length));
+const DEEPSEEK_DSML_RECOVERY_MAX_BOUNDARY_LEN = Math.max(...DEEPSEEK_DSML_TOOL_OPEN_TOKENS.map((token) => token.length + 1), ...DEEPSEEK_DSML_INVOKE_OPEN_PREFIXES.map((token) => token.length + 2));
 const MAX_DSML_RECOVERY_BUFFER_BYTES = 256e3;
-const DEEPSEEK_DSML_SCAN_BATCH_CHARS = 64 * 1024;
+const DEEPSEEK_DSML_SCAN_BATCH_CHARS = 65536;
 function createDsmlRecoverer() {
 	let buffer = "";
 	let bufferBytes = 0;
 	let bufferEndsWithHighSurrogate = false;
 	let pendingScanChars = 0;
 	let activeOpenToken = null;
-	let blockScanState = {
-		offset: 0,
-		mode: "outer",
-		invokeOpenStart: -1
-	};
+	let blockScanState = { offset: 0 };
 	const resetBlockScan = () => {
 		activeOpenToken = null;
 		pendingScanChars = 0;
-		blockScanState = {
-			offset: 0,
-			mode: "outer",
-			invokeOpenStart: -1
-		};
+		blockScanState = { offset: 0 };
 	};
 	const consume = (final) => {
 		const output = [];
@@ -1026,15 +1027,16 @@ function createDsmlRecoverer() {
 }
 function parseDeepSeekDsmlToolCallBlock(body) {
 	const toolCalls = [];
-	const invokeOpenRegex = /<[|｜]DSML[|｜]invoke\b([^<>]*)>/g;
+	const invokeOpenRegex = new RegExp(`<${DEEPSEEK_DSML_MARKER_PATTERN}invoke\\b([^<>]*)>`, "g");
 	let openMatch;
 	while ((openMatch = invokeOpenRegex.exec(body)) !== null) {
 		const invokeBodyStart = openMatch.index + openMatch[0].length;
-		const invokeClose = findEarliestStringToken(body.slice(invokeBodyStart), ["</|DSML|invoke>", "</｜DSML｜invoke>"]);
-		if (!invokeClose) break;
-		const invokeBody = body.slice(invokeBodyStart, invokeBodyStart + invokeClose.index);
-		invokeOpenRegex.lastIndex = invokeBodyStart + invokeClose.index + invokeClose.token.length;
-		const invokeName = parseXmlAttribute(openMatch[1] ?? "", "name");
+		const invokeCloseToken = `</${openMatch[1]}invoke>`;
+		const invokeCloseIndex = body.indexOf(invokeCloseToken, invokeBodyStart);
+		if (invokeCloseIndex === -1) break;
+		const invokeBody = body.slice(invokeBodyStart, invokeCloseIndex);
+		invokeOpenRegex.lastIndex = invokeCloseIndex + invokeCloseToken.length;
+		const invokeName = parseXmlAttribute(openMatch[2] ?? "", "name");
 		if (!invokeName) continue;
 		const parsedArguments = parseDeepSeekDsmlInvokeArguments(invokeBody);
 		if (!parsedArguments) continue;
@@ -1049,12 +1051,12 @@ function parseDeepSeekDsmlToolCallBlock(body) {
 }
 function parseDeepSeekDsmlInvokeArguments(body) {
 	const args = {};
-	const parameterRegex = /<[|｜]DSML[|｜]parameter\b([^>]*)>([\s\S]*?)<\/[|｜]DSML[|｜]parameter>/g;
+	const parameterRegex = new RegExp(`<${DEEPSEEK_DSML_MARKER_PATTERN}parameter\\b([^>]*)>([\\s\\S]*?)</\\1parameter>`, "g");
 	let parameterMatch;
 	while ((parameterMatch = parameterRegex.exec(body)) !== null) {
-		const name = parseXmlAttribute(parameterMatch[1] ?? "", "name");
+		const name = parseXmlAttribute(parameterMatch[2] ?? "", "name");
 		if (!name) continue;
-		const rawValue = parameterMatch[2] ?? "";
+		const rawValue = parameterMatch[3] ?? "";
 		if (rawValue.length === 0) continue;
 		args[name] = decodeDeepSeekDsmlText(rawValue);
 	}
@@ -1099,7 +1101,7 @@ function findEarliestStringToken(text, tokens, fromIndex = 0) {
 }
 function scanDeepSeekDsmlToolBlock(text, closeToken, contentStartIndex, state) {
 	while (state.offset < text.length) {
-		if (state.mode === "invoke-open") {
+		if (state.invoke?.kind === "open") {
 			const nextOpen = text.indexOf("<", state.offset);
 			const nextClose = text.indexOf(">", state.offset);
 			if (nextClose === -1 && nextOpen === -1) {
@@ -1107,31 +1109,32 @@ function scanDeepSeekDsmlToolBlock(text, closeToken, contentStartIndex, state) {
 				return { kind: "incomplete" };
 			}
 			if (nextOpen !== -1 && (nextClose === -1 || nextOpen < nextClose)) {
-				state.mode = "outer";
+				state.invoke = void 0;
 				state.offset = nextOpen;
-				state.invokeOpenStart = -1;
 				continue;
 			}
-			const invokeOpenTag = text.slice(state.invokeOpenStart, nextClose + 1);
-			if (!/^<[|｜]DSML[|｜]invoke\b[^<>]*>$/.test(invokeOpenTag)) {
-				state.mode = "outer";
-				state.offset = state.invokeOpenStart + 1;
-				state.invokeOpenStart = -1;
+			const invokeOpenTag = text.slice(state.invoke.start, nextClose + 1);
+			const open = DEEPSEEK_DSML_INVOKE_OPEN_TAG.exec(invokeOpenTag);
+			if (!open) {
+				state.offset = state.invoke.start + 1;
+				state.invoke = void 0;
 				continue;
 			}
-			state.mode = "invoke-body";
+			state.invoke = {
+				kind: "body",
+				closeToken: `</${open[1]}invoke>`
+			};
 			state.offset = nextClose + 1;
-			state.invokeOpenStart = -1;
 			continue;
 		}
-		if (state.mode === "invoke-body") {
-			const invokeClose = findEarliestStringToken(text, DEEPSEEK_DSML_INVOKE_CLOSE_TOKENS, state.offset);
-			if (!invokeClose) {
+		if (state.invoke?.kind === "body") {
+			const invokeCloseIndex = text.indexOf(state.invoke.closeToken, state.offset);
+			if (invokeCloseIndex === -1) {
 				state.offset = Math.max(0, text.length - DEEPSEEK_DSML_RECOVERY_MAX_BOUNDARY_LEN + 1);
 				return { kind: "incomplete" };
 			}
-			state.mode = "outer";
-			state.offset = invokeClose.index + invokeClose.token.length;
+			state.offset = invokeCloseIndex + state.invoke.closeToken.length;
+			state.invoke = void 0;
 			continue;
 		}
 		const toolOpen = findEarliestStringToken(text, DEEPSEEK_DSML_TOOL_OPEN_TOKENS, state.offset);
@@ -1157,8 +1160,10 @@ function scanDeepSeekDsmlToolBlock(text, closeToken, contentStartIndex, state) {
 			return { kind: "incomplete" };
 		}
 		if (next.kind === "invoke-open") {
-			state.mode = "invoke-open";
-			state.invokeOpenStart = next.index;
+			state.invoke = {
+				kind: "open",
+				start: next.index
+			};
 			state.offset = next.index + next.token.length;
 			continue;
 		}
@@ -1196,6 +1201,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 	const deepSeekTextFilter = shouldFilterDeepSeekDsmlText ? createDeepSeekTextFilter() : null;
 	const deepSeekToolCallRecoverer = shouldFilterDeepSeekDsmlText ? createDsmlRecoverer() : null;
 	const reasoningTagTextPartitioner = createReasoningTagTextPartitioner();
+	if (options?.strictReasoningTags) reasoningTagTextPartitioner.markStrict();
 	let currentBlock = null;
 	let directTextBlock = null;
 	let directThinkingBlock = null;
@@ -1559,7 +1565,8 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 						if (block.id === toolCall.id) encryptedReasoning?.rememberToolCall(toolCall.id, block);
 					}
 					currentBlock = block;
-					if (toolCall.function?.name && (!directMode || !block.name)) block.name = toolCall.function.name;
+					const conflictingId = directMode && block.id && toolCall.id && block.id !== toolCall.id;
+					if (toolCall.function?.name && !conflictingId) block.name = toolCall.function.name;
 					const deltaSig = directMode ? void 0 : extractToolCallThoughtSignature(toolCall);
 					if (deltaSig) block.thoughtSignature = deltaSig;
 					const toolArgumentsDelta = toolCall.function?.arguments;

package/dist/openai-prompt-cache-CGnVB74a.mjs

@@ -0,0 +1,21 @@
+//#region packages/ai/src/providers/openai-prompt-cache.ts
+/** Selects the documented cache lifetime fields for a Responses request. */
+function resolveOpenAIResponsesCacheParams(model, cacheRetention, supportsLongCacheRetention) {
+	if (cacheRetention === "none") return {};
+	if (model.id === "gpt-6-astra" && model.api === "openai-responses") {
+		const endpoint = URL.parse(model.baseUrl ?? "https://api.openai.com/v1");
+		if (endpoint?.protocol === "https:" && (endpoint.hostname === "api.openai.com" || endpoint.hostname.endsWith(".api.openai.com"))) return { prompt_cache_options: { ttl: "30m" } };
+	}
+	return cacheRetention === "long" && supportsLongCacheRetention ? { prompt_cache_retention: "24h" } : {};
+}
+/** Maximum prompt cache key length accepted by OpenAI-compatible request metadata. */
+const OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH = 64;
+/** Truncates a prompt cache key by Unicode code point count. */
+function clampOpenAIPromptCacheKey(key) {
+	if (key === void 0) return;
+	const chars = Array.from(key);
+	if (chars.length <= 64) return key;
+	return chars.slice(0, 64).join("");
+}
+//#endregion
+export { clampOpenAIPromptCacheKey as n, resolveOpenAIResponsesCacheParams as r, OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH as t };

package/dist/openai-prompt-cache-CNoIfHYC.d.mts

@@ -0,0 +1,16 @@
+import { D as Model, l as CacheRetention } from "./types-CJ1-Ht7A.mjs";
+import "./types-3Lnm-QSJ.mjs";
+//#region packages/ai/src/providers/openai-prompt-cache.d.ts
+/** Selects the documented cache lifetime fields for a Responses request. */
+declare function resolveOpenAIResponsesCacheParams(model: Pick<Model, "id" | "api" | "baseUrl">, cacheRetention: CacheRetention, supportsLongCacheRetention: boolean): {
+  prompt_cache_retention?: "24h";
+  prompt_cache_options?: {
+    ttl: "30m";
+  };
+};
+/** Maximum prompt cache key length accepted by OpenAI-compatible request metadata. */
+declare const OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH = 64;
+/** Truncates a prompt cache key by Unicode code point count. */
+declare function clampOpenAIPromptCacheKey(key: string | undefined): string | undefined;
+//#endregion
+export { clampOpenAIPromptCacheKey as n, resolveOpenAIResponsesCacheParams as r, OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH as t };

package/dist/{openai-responses-BogP8OjX.mjs → openai-responses-CCL41ALM.mjs}

@@ -1,12 +1,13 @@
-import { n as getEnvApiKey, r as __exportAll } from "./env-api-keys-DrgeBuva.mjs";
-import { t as AssistantMessageEventStream } from "./event-stream-uSMZJ3FA.mjs";
-import { n as getAiTransportHost } from "./host-D51fmkH6.mjs";
-import { n as clampOpenAIPromptCacheKey } from "./openai-prompt-cache-mZTCdRPo.mjs";
+import { t as __exportAll } from "./rolldown-runtime-BhDjJH2R.mjs";
+import { t as AssistantMessageEventStream } from "./event-stream-BgDvQeum.mjs";
+import { n as getAiTransportHost } from "./host-CEvLw30U.mjs";
+import { n as clampOpenAIPromptCacheKey, r as resolveOpenAIResponsesCacheParams } from "./openai-prompt-cache-CGnVB74a.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
-import { n as buildBaseOptions } from "./simple-options-CFj7x3J4.mjs";
-import { a as resolveResponsesReasoningEffort, i as createResponsesAssistantOutput, n as applyResponsesServiceTierPricing, o as runResponsesStreamLifecycle, r as convertResponsesMessages, t as applyCommonResponsesParams } from "./openai-responses-shared-DJThc6jW.mjs";
-import { j as resolveOpenAIClientBaseUrl } from "./openai-tool-schema-ho-hIkQf.mjs";
-import { S as supportsOpenAITemperature } from "./openai-responses-contracts-QUY12X0B.mjs";
+import { n as buildBaseOptions } from "./simple-options-BjHCCh4v.mjs";
+import { n as getEnvApiKey } from "./env-api-keys-bktO00EJ.mjs";
+import { a as resolveResponsesReasoningEffort, i as createResponsesAssistantOutput, n as applyResponsesServiceTierPricing, r as convertResponsesMessages, s as runResponsesStreamLifecycle, t as applyCommonResponsesParams } from "./openai-responses-shared-Bmma_3Qc.mjs";
+import { j as resolveOpenAIClientBaseUrl } from "./openai-tool-schema-_pTAJqKF.mjs";
+import { j as supportsOpenAITemperature } from "./openai-responses-compaction-window-BMVHFOtq.mjs";
 import { i as resolveCloudflareBaseUrl, n as hasCopilotVisionInput, r as isCloudflareProvider, t as buildCopilotDynamicHeaders } from "./github-copilot-headers-NCJtz9i0.mjs";
 import OpenAI from "openai";
 //#region packages/ai/src/providers/openai-responses.ts
@@ -21,18 +22,16 @@ function getCompat(model) {
 		supportsLongCacheRetention: model.compat?.supportsLongCacheRetention ?? true
 	};
 }
-function getPromptCacheRetention(compat, cacheRetention) {
-	return cacheRetention === "long" && compat.supportsLongCacheRetention ? "24h" : void 0;
-}
 /**
 * Generate function for OpenAI Responses API
 */
 const streamOpenAIResponses = (model, context, options) => {
 	const stream = new AssistantMessageEventStream();
+	const output = createResponsesAssistantOutput(model);
 	runResponsesStreamLifecycle({
 		stream,
 		model,
-		output: createResponsesAssistantOutput(model),
+		output,
 		options,
 		createClient: () => {
 			const apiKey = options?.apiKey || getEnvApiKey(model.provider) || "";
@@ -81,11 +80,13 @@ function createClient(model, context, apiKey, optionsHeaders, sessionId) {
 		Authorization: headers.Authorization ?? null,
 		"cf-aig-authorization": `Bearer ${apiKey}`
 	} : headers;
+	const baseUrl = isCloudflareProvider(model.provider) ? resolveCloudflareBaseUrl(model) : model.baseUrl;
 	return new OpenAI({
 		apiKey,
-		baseURL: resolveOpenAIClientBaseUrl(model, isCloudflareProvider(model.provider) ? resolveCloudflareBaseUrl(model) : model.baseUrl),
+		baseURL: resolveOpenAIClientBaseUrl(model, baseUrl),
 		dangerouslyAllowBrowser: true,
 		defaultHeaders,
+		maxRetries: 0,
 		fetch: getAiTransportHost().buildModelFetch(model)
 	});
 }
@@ -103,7 +104,7 @@ function buildParams(model, context, options, replayMode = "checkpoint") {
 		input: messages,
 		stream: true,
 		prompt_cache_key: cacheRetention === "none" ? void 0 : clampOpenAIPromptCacheKey(options?.promptCacheKey ?? options?.sessionId),
-		prompt_cache_retention: getPromptCacheRetention(compat, cacheRetention),
+		...resolveOpenAIResponsesCacheParams(model, cacheRetention, compat.supportsLongCacheRetention),
 		store: false
 	};
 	if (options?.maxTokens) params.max_output_tokens = options?.maxTokens;