@openclaw/ai 2026.9.6 → 2026.9.7

package/dist/{openai-completions-stream-BpuJ0ZOA.mjs → openai-completions-stream-CDCuw3DP.mjs}

@@ -1,27 +1,27 @@
-import { r as appendAssistantThinking } from "./event-stream-D8PARQfL.mjs";
-import { o as isRecord } from "./record-coerce-DwRYMj3t.mjs";
-import { n as isOpenAIGpt55Model, r as isOpenAIGpt56Model, t as isOpenAIGpt54MiniModel } from "./openai-reasoning-effort-ash8dPrd.mjs";
+import { r as appendAssistantThinking } from "./event-stream-vd9UnCAU.mjs";
+import { o as isRecord } from "./record-coerce-CNDovvft.mjs";
+import { k as isImageWithMediaPayload, n as getAiTransportHost } from "./host-CHi3X87F.mjs";
+import { n as isOpenAIGpt55Model, r as isOpenAIGpt56Model, t as isOpenAIGpt54MiniModel } from "./openai-reasoning-effort-ConhwkeX.mjs";
 import { r as uniqueStrings } from "./string-normalization-J9ZiLfGO.mjs";
-import { _ as isImageWithMediaPayload, d as describeToolResultMediaPlaceholder, m as extractToolResultText, v as sanitizeSurrogates } from "./host-yTvBrYM2.mjs";
-import { t as truncateUtf16Safe } from "./utf16-slice-CvGodqok.mjs";
-import { r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-CjV_2IRI.mjs";
-import { r as emitModelTransportDebug } from "./diagnostics-CH8j2wIG.mjs";
-import { i as asNonNegativeFiniteNumber } from "./base64-CPmadnUf.mjs";
-import { c as supportsModelTools, r as estimateStringChars } from "./transport-utils-03Qi0CN2.mjs";
-import { c as shouldOmitOllamaCompatResponseFormat, n as isNativeOpenAIEndpoint, s as resolveOpenAICompletionsResponseFormat, t as detectOpenAICompletionsCompat } from "./openai-completions-compat-CUollwl6.mjs";
-import { i as resolveProviderEndpoint, r as resolveOpenAIStrictToolSetting } from "./host-policy-DZO3oPPk.mjs";
+import { o as truncateUtf16Safe } from "./provider-error-CwT_j2v0.mjs";
+import { r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-D2V7MqR7.mjs";
+import { r as emitModelTransportDebug } from "./diagnostics-CJaB6g8J.mjs";
+import { i as asNonNegativeFiniteNumber } from "./base64-C1PsYcRQ.mjs";
+import { c as supportsModelTools, r as estimateStringChars } from "./transport-utils-C-4bf2Ok.mjs";
+import { c as shouldOmitOllamaCompatResponseFormat, n as isNativeOpenAIEndpoint, s as resolveOpenAICompletionsResponseFormat, t as detectOpenAICompletionsCompat } from "./openai-completions-compat-BOVWBZDn.mjs";
+import { r as resolveProviderEndpoint } from "./host-policy-BOAwRg_M.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
-import { f as splitSystemPromptCacheBoundary, h as stripSystemPromptRelocatableBoundary, m as stripSystemPromptCacheBoundary, p as splitSystemPromptRelocatableBoundary, v as sortPromptCacheToolsByName } from "./simple-options-np1XPuhm.mjs";
-import { A as resolvePromptCacheKey, C as isOpenAICompletionsThinkingEnabled, D as readOpenAICompletionsContentDeltas, E as parseOpenAICompletionsUsage, O as readOpenAICompletionsReasoningBatch, T as measureUtf8AppendBytes, d as prepareOpenAITools, f as projectOpenAITools, g as resolveOpenAISimpleReasoningEffort, h as resolveOpenAIRequestReasoning, j as throwIfModelStreamAborted, p as reconcileOpenAICompletionsToolChoice, s as getCompat, u as resolveOpenAIStrictToolFlagWithDiagnostics, v as GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, w as log, y as createModelStreamCooperativeScheduler } from "./openai-transport-params-DZKOTSfH.mjs";
-import { E as parseStreamingJson, c as finalizeTerminalToolCallArguments, w as createToolArgumentPreviewSchedule } from "./transport-stream-shared-B5RioB_Q.mjs";
+import { E as createToolArgumentPreviewSchedule, O as parseStreamingJson, d as iterateModelStream, j as sanitizeSurrogates, l as finalizeTerminalToolCallArguments, y as throwIfModelStreamAborted } from "./transport-stream-shared-Dh5qKJ8e.mjs";
+import { C as stripSystemPromptCacheBoundary, D as sortPromptCacheToolsByName, S as splitSystemPromptRelocatableBoundary, d as describeToolResultMediaPlaceholder, m as extractToolResultText, w as stripSystemPromptRelocatableBoundary, x as splitSystemPromptCacheBoundary } from "./simple-options-CGAkCwtM.mjs";
+import { C as log, D as readOpenAICompletionsReasoningBatch, E as readOpenAICompletionsContentDeltas, S as isOpenAICompletionsThinkingEnabled, T as parseOpenAICompletionsUsage, d as prepareOpenAITools, f as projectOpenAITools, g as resolveOpenAISimpleReasoningEffort, h as resolveOpenAIRequestReasoning, k as resolvePromptCacheKey, p as reconcileOpenAICompletionsToolChoice, s as getCompat, u as resolveOpenAIStrictToolFlagWithDiagnostics, v as GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, w as measureUtf8AppendBytes } from "./openai-transport-params-BwlpzX-2.mjs";
 import { a as tagUnresolvedTextAsCommentary, i as tagPendingCommentaryText, n as rememberPendingCommentaryTags, r as tagInterruptedTextPhases, t as clearPendingCommentaryText } from "./assistant-text-phase-C20rxWwP.mjs";
 import { t as notifyLlmRequestActivity } from "./llm-request-activity-BjtkplhG.mjs";
-import { a as withFirstStreamEventTimeout } from "./stream-first-event-timeout-BeptfRL6.mjs";
-import { t as transformProviderMessages } from "./provider-transcript-transform-BTocjCRs.mjs";
-import { f as withPreparedToolSchemaNormalization, r as normalizeOpenAIStrictToolParameters } from "./openai-tool-schema-D4BMcj2D.mjs";
+import { t as transformProviderMessages } from "./provider-transcript-transform-JbFbHV8h.mjs";
+import { a as withFirstStreamEventTimeout } from "./stream-first-event-timeout-B9o0S4SZ.mjs";
+import { f as withPreparedToolSchemaNormalization, r as normalizeOpenAIStrictToolParameters } from "./openai-tool-schema-vsl-cE-n.mjs";
 import { isGoogleGemini3FlashModel, isGoogleGemini3ProModel } from "./internal/google-model-family.mjs";
-import { t as mapOpenAIStopReason } from "./openai-stop-reason-Drnn_6Qj.mjs";
-import { t as createReasoningTagTextPartitioner } from "./reasoning-tag-text-partitioner-Di38crhI.mjs";
+import { t as mapOpenAIStopReason } from "./openai-stop-reason-8s70vDXj.mjs";
+import { t as createReasoningTagTextPartitioner } from "./reasoning-tag-text-partitioner-B6FIIBt6.mjs";
 import { randomUUID } from "node:crypto";
 //#region packages/ai/src/transports/deepseek-dsml-grammar.ts
 const DEEPSEEK_DSML_MARKERS = [
@@ -30,6 +30,25 @@ const DEEPSEEK_DSML_MARKERS = [
 	"｜｜"
 ].map((bar) => `${bar}DSML${bar}`);
 const DEEPSEEK_DSML_MARKER_PATTERN = `(${DEEPSEEK_DSML_MARKERS.map((marker) => marker.replaceAll("|", "\\|")).join("|")})`;
+function findEarliestDsmlToken(text, tokens, fromIndex = 0) {
+	let best = null;
+	for (const token of tokens) {
+		const index = text.indexOf(token, fromIndex);
+		if (index !== -1 && (!best || index < best.index)) best = {
+			index,
+			token
+		};
+	}
+	return best;
+}
+function longestDsmlTokenPrefixSuffixLength(text, tokens, maxTokenLength) {
+	const maxLength = Math.min(text.length, maxTokenLength - 1);
+	for (let length = maxLength; length > 0; length--) {
+		const suffix = text.slice(text.length - length);
+		if (tokens.some((token) => token.startsWith(suffix))) return length;
+	}
+	return 0;
+}
 //#endregion
 //#region packages/ai/src/transports/deepseek-text-filter.ts
 /**
@@ -67,7 +86,7 @@ function createDeepSeekTextFilter() {
 				if (final) closeToken = void 0;
 				return output;
 			}
-			const open = findEarliestToken(buffer, DSML_OPEN_TOKENS);
+			const open = findEarliestDsmlToken(buffer, DSML_OPEN_TOKENS);
 			if (open) {
 				emit(buffer.slice(0, open.index));
 				buffer = buffer.slice(open.index + open.token.length);
@@ -79,7 +98,7 @@ function createDeepSeekTextFilter() {
 				buffer = "";
 				return output;
 			}
-			const keep = longestDsmlOpenPrefixSuffixLength(buffer);
+			const keep = longestDsmlTokenPrefixSuffixLength(buffer, DSML_OPEN_TOKENS, MAX_OPEN_TOKEN_LEN);
 			const emitLength = buffer.length - keep;
 			if (emitLength <= 0) return output;
 			emit(buffer.slice(0, emitLength));
@@ -98,25 +117,6 @@ function createDeepSeekTextFilter() {
 		}
 	};
 }
-function findEarliestToken(text, tokens) {
-	let best = null;
-	for (const token of tokens) {
-		const index = text.indexOf(token);
-		if (index !== -1 && (!best || index < best.index)) best = {
-			index,
-			token
-		};
-	}
-	return best;
-}
-function longestDsmlOpenPrefixSuffixLength(text) {
-	const maxLength = Math.min(text.length, MAX_OPEN_TOKEN_LEN - 1);
-	for (let length = maxLength; length > 0; length--) {
-		const suffix = text.slice(text.length - length);
-		if (DSML_OPEN_TOKENS.some((token) => token.startsWith(suffix))) return length;
-	}
-	return 0;
-}
 //#endregion
 //#region packages/ai/src/transports/model-max-tokens-params.ts
 /**
@@ -466,7 +466,7 @@ function convertMessages(model, context, compat, options = {}) {
 		if (model.provider === "openai") return id.length > 40 ? truncateUtf16Safe(id, 40) : id;
 		return id;
 	};
-	const transformedMessages = transformProviderMessages(context.messages, model, (id) => normalizeToolCallId(id));
+	const transformedMessages = transformProviderMessages(context.messages, model, normalizeToolCallId);
 	let relocatableSplit;
 	let systemParamIndex;
 	if (context.systemPrompt) {
@@ -493,15 +493,12 @@ function convertMessages(model, context, compat, options = {}) {
 			content: "I have processed the tool results."
 		});
 		if (msg.role === "user") {
-			const isRuntimeContextCarrier = msg.runtimeContextCarrier === true;
-			if (typeof msg.content === "string") {
-				const userParam = {
-					role: "user",
-					content: sanitizeSurrogates(msg.content)
-				};
-				if (isRuntimeContextCarrier) options.cacheOptOutIndexes?.add(params.length);
-				params.push(userParam);
-			} else {
+			let userParam;
+			if (typeof msg.content === "string") userParam = {
+				role: "user",
+				content: sanitizeSurrogates(msg.content)
+			};
+			else {
 				const content = msg.content.map((item) => {
 					if (item.type === "text") return {
 						type: "text",
@@ -517,13 +514,13 @@ function convertMessages(model, context, compat, options = {}) {
 					};
 				});
 				if (content.length === 0) continue;
-				const userParam = {
+				userParam = {
 					role: "user",
 					content
 				};
-				if (isRuntimeContextCarrier) options.cacheOptOutIndexes?.add(params.length);
-				params.push(userParam);
 			}
+			if (msg.runtimeContextCarrier === true) options.cacheOptOutIndexes?.add(params.length);
+			params.push(userParam);
 		} else if (msg.role === "assistant") {
 			const assistantMsg = {
 				role: "assistant",
@@ -872,19 +869,18 @@ function resolveOpenAICompletionsModelMaxTokens(model) {
 const OPENAI_COMPLETIONS_INPUT_TOKEN_SAFETY_MARGIN = 1.25;
 const OPENAI_COMPLETIONS_IMAGE_CHAR_ESTIMATE = 8e3;
 const MIN_USEFUL_OUTPUT_TOKENS = 16;
+function estimateJsonChars(value, fallback) {
+	try {
+		return estimateStringChars(JSON.stringify(value));
+	} catch {
+		return fallback;
+	}
+}
 function estimateOpenAICompletionsInputTokens(payload) {
 	let adjustedChars = 0;
 	adjustedChars += estimateOpenAICompletionsMessagesChars(payload.messages);
-	if (Array.isArray(payload.tools) && payload.tools.length > 0) try {
-		adjustedChars += estimateStringChars(JSON.stringify(payload.tools));
-	} catch {
-		adjustedChars += 1024;
-	}
-	if (payload.response_format !== void 0) try {
-		adjustedChars += estimateStringChars(JSON.stringify(payload.response_format));
-	} catch {
-		adjustedChars += 256;
-	}
+	if (Array.isArray(payload.tools) && payload.tools.length > 0) adjustedChars += estimateJsonChars(payload.tools, 1024);
+	if (payload.response_format !== void 0) adjustedChars += estimateJsonChars(payload.response_format, 256);
 	return Math.ceil(adjustedChars / 4 * OPENAI_COMPLETIONS_INPUT_TOKEN_SAFETY_MARGIN);
 }
 function estimateOpenAICompletionsMessagesChars(messages) {
@@ -895,11 +891,7 @@ function estimateOpenAICompletionsMessagesChars(messages) {
 		const record = message;
 		adjustedChars += estimateOpenAICompletionsContentChars(record.content);
 		for (const field of COMPLETIONS_REASONING_REPLAY_FIELDS) adjustedChars += estimateOpenAICompletionsContentChars(record[field]);
-		if (record.tool_calls !== void 0) try {
-			adjustedChars += estimateStringChars(JSON.stringify(record.tool_calls));
-		} catch {
-			adjustedChars += 256;
-		}
+		if (record.tool_calls !== void 0) adjustedChars += estimateJsonChars(record.tool_calls, 256);
 	}
 	return adjustedChars;
 }
@@ -919,11 +911,7 @@ function estimateOpenAICompletionsContentChars(value) {
 			adjustedChars += estimateStringChars(text);
 			continue;
 		}
-		try {
-			adjustedChars += estimateStringChars(JSON.stringify(block));
-		} catch {
-			adjustedChars += 256;
-		}
+		adjustedChars += estimateJsonChars(block, 256);
 	}
 	return adjustedChars;
 }
@@ -960,7 +948,7 @@ function convertTools(tools, compat, model, mode) {
 	const prepared = mode === "managed" ? prepareOpenAITools(tools) : void 0;
 	const projection = prepared?.projection ?? projectOpenAITools(tools);
 	const convert = () => {
-		const strict = mode === "direct" ? compat.supportsStrictMode ? false : void 0 : resolveOpenAIStrictToolFlagWithDiagnostics(projection, resolveOpenAIStrictToolSetting(model, {
+		const strict = mode === "direct" ? compat.supportsStrictMode ? false : void 0 : resolveOpenAIStrictToolFlagWithDiagnostics(projection, getAiTransportHost().resolveOpenAIStrictToolSetting(model, {
 			transport: "stream",
 			supportsStrictMode: compat?.supportsStrictMode
 		}), {
@@ -1061,6 +1049,15 @@ function buildOpenAICompletionsRequest(model, context, options, policy) {
 		const toolChoice = reconcileOpenAICompletionsToolChoice(options.toolChoice, directToolProjection ?? projectOpenAITools([]));
 		if (toolChoice !== void 0) params.tool_choice = toolChoice;
 	}
+	const isOpenRouter = compat.thinkingFormat === "openrouter";
+	const usesBinaryOpenRouterThinking = isOpenRouter && (model.compat?.supportsReasoningEffort === false || model.compat?.supportedReasoningEfforts?.length === 0);
+	const simpleReasoning = options?.reasoning;
+	const requestedEffort = policy.mode === "direct" ? options?.reasoningEffort : options?.reasoningEffort ?? (simpleReasoning === "none" ? "none" : resolveOpenAISimpleReasoningEffort({
+		...model,
+		compat: model.compat ?? void 0
+	}, simpleReasoning)) ?? (usesBinaryOpenRouterThinking ? void 0 : "high");
+	const reasoning = resolveOpenAIRequestReasoning(model, requestedEffort);
+	const { effort, thinkingEnabled } = reasoning;
 	{
 		const maxTokenBudget = policy.mode === "direct" ? {
 			maxTokens: options?.maxTokens,
@@ -1080,7 +1077,10 @@ function buildOpenAICompletionsRequest(model, context, options, policy) {
 			if (clampedMaxTokens > remainingBudget) {
 				clampedMaxTokens = remainingBudget;
 				emitModelTransportDebug(log, `[completions] clamp_max_tokens provider=${model.provider} api=${model.api} model=${model.id} requested=${effectiveMaxTokens} output=${clampedMaxTokens} effectiveContext=${effectiveContextTokens} estimatedInput=${estimatedInputTokens}`);
-				if (remainingBudget < MIN_USEFUL_OUTPUT_TOKENS) log.warn(`[completions] insufficient_output_budget provider=${model.provider} api=${model.api} model=${model.id} output=${clampedMaxTokens} effectiveContext=${effectiveContextTokens} estimatedInput=${estimatedInputTokens}`);
+				if (remainingBudget < MIN_USEFUL_OUTPUT_TOKENS) {
+					if (model.reasoning && thinkingEnabled !== false) throw Object.assign(/* @__PURE__ */ new Error(`Context window exceeded: estimated input ${estimatedInputTokens} leaves only ${remainingBudget} output tokens within the ${effectiveContextTokens}-token context.`), { code: "context_length_exceeded" });
+					log.warn(`[completions] insufficient_output_budget provider=${model.provider} api=${model.api} model=${model.id} output=${clampedMaxTokens} effectiveContext=${effectiveContextTokens} estimatedInput=${estimatedInputTokens}`);
+				}
 			}
 		}
 		if (policy.mode === "direct" ? options?.maxTokens : clampedMaxTokens) {
@@ -1088,15 +1088,6 @@ function buildOpenAICompletionsRequest(model, context, options, policy) {
 			else params.max_completion_tokens = clampedMaxTokens;
 		}
 	}
-	const isOpenRouter = compat.thinkingFormat === "openrouter";
-	const usesBinaryOpenRouterThinking = isOpenRouter && (model.compat?.supportsReasoningEffort === false || model.compat?.supportedReasoningEfforts?.length === 0);
-	const simpleReasoning = options?.reasoning;
-	const requestedEffort = policy.mode === "direct" ? options?.reasoningEffort : options?.reasoningEffort ?? (simpleReasoning === "none" ? "none" : resolveOpenAISimpleReasoningEffort({
-		...model,
-		compat: model.compat ?? void 0
-	}, simpleReasoning)) ?? (usesBinaryOpenRouterThinking ? void 0 : "high");
-	const reasoning = resolveOpenAIRequestReasoning(model, requestedEffort);
-	const { effort, thinkingEnabled } = reasoning;
 	if (isOpenRouter && model.reasoning) {
 		if (usesBinaryOpenRouterThinking && thinkingEnabled !== void 0) params.reasoning = { enabled: thinkingEnabled };
 		else if (effort !== void 0) params.reasoning = { effort };
@@ -1132,6 +1123,7 @@ const DEEPSEEK_DSML_TOOL_MAX_OPEN_TOKEN_LEN = Math.max(...DEEPSEEK_DSML_TOOL_OPE
 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 = 65536;
+const DSML_NAME_ATTRIBUTE_RE = /\bname=("([^"]*)"|'([^']*)'|([^\s>]+))/;
 function createDsmlRecoverer() {
 	let buffer = "";
 	let bufferBytes = 0;
@@ -1150,7 +1142,7 @@ function createDsmlRecoverer() {
 			const open = activeOpenToken ? {
 				index: 0,
 				token: activeOpenToken
-			} : findEarliestStringToken(buffer, DEEPSEEK_DSML_TOOL_OPEN_TOKENS);
+			} : findEarliestDsmlToken(buffer, DEEPSEEK_DSML_TOOL_OPEN_TOKENS);
 			if (!open) {
 				resetBlockScan();
 				if (final) {
@@ -1163,7 +1155,7 @@ function createDsmlRecoverer() {
 					bufferEndsWithHighSurrogate = false;
 					return output;
 				}
-				const keep = longestDeepSeekDsmlToolOpenPrefixSuffixLength(buffer);
+				const keep = longestDsmlTokenPrefixSuffixLength(buffer, DEEPSEEK_DSML_TOOL_OPEN_TOKENS, DEEPSEEK_DSML_TOOL_MAX_OPEN_TOKEN_LEN);
 				const emitLength = buffer.length - keep;
 				if (emitLength > 0) {
 					const emitted = buffer.slice(0, emitLength);
@@ -1249,7 +1241,7 @@ function parseDeepSeekDsmlToolCallBlock(body) {
 		if (invokeCloseIndex === -1) break;
 		const invokeBody = body.slice(invokeBodyStart, invokeCloseIndex);
 		invokeOpenRegex.lastIndex = invokeCloseIndex + invokeCloseToken.length;
-		const invokeName = parseXmlAttribute(openMatch[2] ?? "", "name");
+		const invokeName = parseDsmlNameAttribute(openMatch[2] ?? "");
 		if (!invokeName) continue;
 		const parsedArguments = parseDeepSeekDsmlInvokeArguments(invokeBody);
 		if (!parsedArguments) continue;
@@ -1267,7 +1259,7 @@ function parseDeepSeekDsmlInvokeArguments(body) {
 	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[2] ?? "", "name");
+		const name = parseDsmlNameAttribute(parameterMatch[2] ?? "");
 		if (!name) continue;
 		const rawValue = parameterMatch[3] ?? "";
 		if (rawValue.length === 0) continue;
@@ -1284,34 +1276,14 @@ function parseDeepSeekDsmlInvokeArguments(body) {
 	}
 	return null;
 }
-const xmlAttributeRegexCache = /* @__PURE__ */ new Map();
-function xmlAttributeRegex(name) {
-	const cached = xmlAttributeRegexCache.get(name);
-	if (cached) return cached;
-	const escaped = name.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
-	const pattern = new RegExp(`\\b${escaped}=("([^"]*)"|'([^']*)'|([^\\s>]+))`);
-	xmlAttributeRegexCache.set(name, pattern);
-	return pattern;
-}
-function parseXmlAttribute(attributes, name) {
-	const match = xmlAttributeRegex(name).exec(attributes);
+function parseDsmlNameAttribute(attributes) {
+	const match = DSML_NAME_ATTRIBUTE_RE.exec(attributes);
 	const value = match?.[2] ?? match?.[3] ?? match?.[4];
 	return value ? decodeDeepSeekDsmlText(value) : null;
 }
 function decodeDeepSeekDsmlText(value) {
 	return value.replaceAll("&quot;", "\"").replaceAll("&apos;", "'").replaceAll("&lt;", "<").replaceAll("&gt;", ">").replaceAll("&amp;", "&");
 }
-function findEarliestStringToken(text, tokens, fromIndex = 0) {
-	let best = null;
-	for (const token of tokens) {
-		const index = text.indexOf(token, fromIndex);
-		if (index !== -1 && (!best || index < best.index)) best = {
-			index,
-			token
-		};
-	}
-	return best;
-}
 function scanDeepSeekDsmlToolBlock(text, closeToken, contentStartIndex, state) {
 	while (state.offset < text.length) {
 		if (state.invoke?.kind === "open") {
@@ -1350,9 +1322,9 @@ function scanDeepSeekDsmlToolBlock(text, closeToken, contentStartIndex, state) {
 			state.invoke = void 0;
 			continue;
 		}
-		const toolOpen = findEarliestStringToken(text, DEEPSEEK_DSML_TOOL_OPEN_TOKENS, state.offset);
+		const toolOpen = findEarliestDsmlToken(text, DEEPSEEK_DSML_TOOL_OPEN_TOKENS, state.offset);
 		const toolCloseIndex = text.indexOf(closeToken, state.offset);
-		const invokeOpen = findEarliestStringToken(text, DEEPSEEK_DSML_INVOKE_OPEN_PREFIXES, state.offset);
+		const invokeOpen = findEarliestDsmlToken(text, DEEPSEEK_DSML_INVOKE_OPEN_PREFIXES, state.offset);
 		const next = [
 			toolOpen ? {
 				kind: "nested-open",
@@ -1384,14 +1356,6 @@ function scanDeepSeekDsmlToolBlock(text, closeToken, contentStartIndex, state) {
 	}
 	return { kind: "incomplete" };
 }
-function longestDeepSeekDsmlToolOpenPrefixSuffixLength(text) {
-	const maxLength = Math.min(text.length, DEEPSEEK_DSML_TOOL_MAX_OPEN_TOKEN_LEN - 1);
-	for (let length = maxLength; length > 0; length -= 1) {
-		const suffix = text.slice(text.length - length);
-		if (DEEPSEEK_DSML_TOOL_OPEN_TOKENS.some((token) => token.startsWith(suffix))) return length;
-	}
-	return 0;
-}
 //#endregion
 //#region packages/ai/src/transports/openai-completions-stream.ts
 function extractToolCallThoughtSignature(toolCall) {
@@ -1436,29 +1400,20 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 	let finishReason;
 	let sawNativeToolCallDelta = false;
 	const blockIndex = () => directMode && currentBlock && currentBlock.type !== "toolCall" ? contentBlockIndices.get(currentBlock) ?? output.content.length - 1 : output.content.length - 1;
-	const measureUtf8Bytes = (text) => Buffer.byteLength(text, "utf8");
 	let chunkPushedEvent = false;
 	const pushStreamEvent = (event) => {
 		chunkPushedEvent = true;
 		stream.push(event);
 	};
 	const queuePostToolCallDelta = (next) => {
-		const nextBytes = measureUtf8Bytes(next.text);
+		const nextBytes = Buffer.byteLength(next.text, "utf8");
 		if (pendingPostToolCallBytes + nextBytes > MAX_POST_TOOL_CALL_BUFFER_BYTES) throw new Error("Exceeded post-tool-call delta buffer limit");
 		pendingPostToolCallBytes += nextBytes;
 		const previous = pendingPostToolCallDeltas[pendingPostToolCallDeltas.length - 1];
-		if (!previous || previous.kind !== next.kind || previous.kind === "text" && next.kind === "text" && previous.source !== next.source) {
+		if (!previous || previous.kind !== next.kind || previous.kind === "text" && next.kind === "text" && previous.source !== next.source || previous.kind === "thinking" && next.kind === "thinking" && previous.signature !== next.signature) {
 			pendingPostToolCallDeltas.push(next);
 			return;
 		}
-		if (next.kind === "thinking" && previous.kind === "thinking") {
-			if (previous.signature !== next.signature) {
-				pendingPostToolCallDeltas.push(next);
-				return;
-			}
-			previous.text += next.text;
-			return;
-		}
 		previous.text += next.text;
 	};
 	const appendThinkingDeltaInternal = (reasoningDelta) => {
@@ -1595,11 +1550,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 			partial: output
 		});
 	};
-	const appendFilteredVisibleTextDelta = (text) => {
-		const recoveredParts = deepSeekToolCallRecoverer?.push(text) ?? [{
-			kind: "text",
-			text
-		}];
+	const appendRecoveredParts = (recoveredParts) => {
 		for (const recoveredPart of recoveredParts) {
 			if (recoveredPart.kind === "toolCall") {
 				appendRecoveredToolCall(recoveredPart);
@@ -1609,31 +1560,11 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 			for (const part of parts) appendVisibleTextDelta(part);
 		}
 	};
-	const flushDeepSeekToolCallRecovererAtEnd = () => {
-		const recoveredParts = deepSeekToolCallRecoverer?.flush();
-		if (!recoveredParts) return;
-		for (const recoveredPart of recoveredParts) {
-			if (recoveredPart.kind === "toolCall") {
-				appendRecoveredToolCall(recoveredPart);
-				continue;
-			}
-			const parts = deepSeekTextFilter?.push(recoveredPart.text) ?? [recoveredPart.text];
-			for (const part of parts) appendVisibleTextDelta(part);
-		}
-	};
-	const flushDeepSeekTextFilterAtEnd = () => {
-		const parts = deepSeekTextFilter?.flush();
-		if (!parts) return;
-		for (const part of parts) appendVisibleTextDelta(part);
-	};
-	const appendRoutedContentDelta = (delta) => {
-		if (delta.kind === "text") {
-			appendFilteredVisibleTextDelta(delta.text);
-			return;
-		}
-		if (!emitReasoning) return;
-		if (currentBlock?.type === "toolCall" && !directMode) queuePostToolCallDelta(delta);
-		else appendThinkingDelta(delta);
+	const appendFilteredVisibleTextDelta = (text) => {
+		appendRecoveredParts(deepSeekToolCallRecoverer?.push(text) ?? [{
+			kind: "text",
+			text
+		}]);
 	};
 	const appendPartitionedVisibleDelta = (delta) => {
 		if (delta.kind === "text") appendFilteredVisibleTextDelta(delta.text);
@@ -1665,7 +1596,6 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		if (forceStrict || reasoningTagTextPartitioner.hasPending()) reasoningTagTextPartitioner.markStrict();
 		if (!hasFollowingVisibleText || !reasoningTagTextPartitioner.hasPendingSyntax()) sealTextBeforeReasoning();
 	};
-	const cooperativeScheduler = directMode ? void 0 : createModelStreamCooperativeScheduler(options?.signal);
 	const guardedStream = withFirstStreamEventTimeout(responseStream, {
 		provider: model.provider,
 		api: model.api,
@@ -1676,13 +1606,11 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		onTimeout: options?.onFirstEventTimeout,
 		hint: "The provider may be stalled while parsing the tool payload; retry with a smaller tool surface or enable OPENCLAW_DEBUG_MODEL_PAYLOAD=tools to inspect exposed tools."
 	});
-	for await (const rawChunk of guardedStream) {
+	const events = directMode ? guardedStream : iterateModelStream(guardedStream, options?.signal);
+	for await (const rawChunk of events) {
 		throwIfModelStreamAborted(options?.signal);
 		chunkPushedEvent = false;
-		if (!rawChunk || typeof rawChunk !== "object") {
-			if (cooperativeScheduler) await cooperativeScheduler.afterEvent();
-			continue;
-		}
+		if (!rawChunk || typeof rawChunk !== "object") continue;
 		notifyLlmRequestActivity(options?.signal);
 		const chunk = rawChunk;
 		output.responseId ||= chunk.id;
@@ -1695,7 +1623,6 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		const choice = Array.isArray(chunk.choices) ? chunk.choices[0] : void 0;
 		if (!choice) {
 			emitReasoningUsageActivity(hasReasoningUsageActivity);
-			if (cooperativeScheduler) await cooperativeScheduler.afterEvent();
 			continue;
 		}
 		const choiceUsage = choice.usage;
@@ -1712,7 +1639,6 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		const rawChoiceDelta = choice.delta ?? choice.message;
 		if (!rawChoiceDelta) {
 			emitReasoningUsageActivity(hasReasoningUsageActivity);
-			if (cooperativeScheduler) await cooperativeScheduler.afterEvent();
 			continue;
 		}
 		for (const normalizedDelta of normalizeToolCallDeltas(rawChoiceDelta, choice.finish_reason)) {
@@ -1733,7 +1659,10 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 				for (const routedDelta of routedDeltas) appendPartitionedVisibleDelta(routedDelta);
 			} else {
 				beginReasoning(contentDeltaIndex < lastVisibleTextIndex);
-				appendRoutedContentDelta(contentDelta);
+				if (emitReasoning) {
+					if (currentBlock?.type === "toolCall" && !directMode) queuePostToolCallDelta(contentDelta);
+					else appendThinkingDelta(contentDelta);
+				}
 			}
 			if (!hasReasoningThinking) appendReasoningDeltas(reasoningDeltas);
 			const toolCallDeltas = normalizedDelta.toolCalls;
@@ -1800,13 +1729,12 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		}
 		flushPendingPostToolCallDeltas();
 		emitReasoningUsageActivity(hasReasoningUsageActivity);
-		if (cooperativeScheduler) await cooperativeScheduler.afterEvent();
 	}
 	throwIfModelStreamAborted(options?.signal);
 	if (!finishReason && (directMode || options?.sawStreamDONE?.() === false)) throw new Error("Stream ended without finish_reason");
 	flushReasoningTagTextPartitioner();
-	flushDeepSeekToolCallRecovererAtEnd();
-	flushDeepSeekTextFilterAtEnd();
+	appendRecoveredParts(deepSeekToolCallRecoverer?.flush() ?? []);
+	for (const part of deepSeekTextFilter?.flush() ?? []) appendVisibleTextDelta(part);
 	currentBlock = null;
 	flushPendingPostToolCallDeltas();
 	finalizeOpenAICompletionsToolCalls(output, {

package/dist/{openai-prompt-cache-C4RHdN4n.d.mts → openai-prompt-cache-CouErLh7.d.mts}

@@ -1,5 +1,5 @@
-import { M as OpenAICompletionsCompat, O as Model, l as CacheRetention } from "./types-qZFbaDbK.mjs";
-import "./types-DTvddOfm.mjs";
+import { M as OpenAICompletionsCompat, O as Model, l as CacheRetention } from "./types-CLLR3eB4.mjs";
+import "./types-CB-_vNyI.mjs";
 //#region packages/ai/src/providers/openai-prompt-cache.d.ts
 /** Selects documented lifetime fields shared by Responses and Chat Completions. */
 declare function resolveOpenAIPromptCacheParams(model: Pick<Model, "id" | "provider" | "baseUrl">, cacheRetention: CacheRetention, compat: Required<Pick<OpenAICompletionsCompat, "supportsPromptCacheKey" | "supportsLongCacheRetention">>): {

package/dist/{openai-prompt-cache-CjV_2IRI.mjs → openai-prompt-cache-D2V7MqR7.mjs}

@@ -1,4 +1,4 @@
-import { n as isNativeOpenAIEndpoint } from "./openai-completions-compat-CUollwl6.mjs";
+import { n as isNativeOpenAIEndpoint } from "./openai-completions-compat-BOVWBZDn.mjs";
 //#region packages/normalization-core/src/code-points.ts
 /** Truncates to a nonnegative code-point budget; grapheme clusters may be split. */
 function truncateCodePoints(text, maxCodePoints) {

package/dist/{openai-provider-client-BP8J1aeg.mjs → openai-provider-client-CwfsGBbA.mjs}

@@ -1,5 +1,5 @@
-import { n as getAiTransportHost } from "./host-yTvBrYM2.mjs";
-import { k as resolveOpenAIClientBaseUrl } from "./openai-transport-params-DZKOTSfH.mjs";
+import { n as getAiTransportHost } from "./host-CHi3X87F.mjs";
+import { O as resolveOpenAIClientBaseUrl } from "./openai-transport-params-BwlpzX-2.mjs";
 import { i as resolveCloudflareBaseUrl, r as isCloudflareProvider } from "./github-copilot-headers-B37TB3rB.mjs";
 import OpenAI from "openai";
 //#region packages/ai/src/providers/openai-provider-client.ts

package/dist/{openai-reasoning-effort-ash8dPrd.mjs → openai-reasoning-effort-ConhwkeX.mjs}

@@ -1,5 +1,5 @@
+import { i as asOptionalRecord } from "./record-coerce-CNDovvft.mjs";
 import { n as normalizeLowercaseStringOrEmpty } from "./string-coerce-fsri9iCu.mjs";
-import { i as asOptionalRecord } from "./record-coerce-DwRYMj3t.mjs";
 import { n as normalizeUniqueTrimmedStringList } from "./string-normalization-J9ZiLfGO.mjs";
 //#region packages/llm-core/src/model-data.ts
 const MODEL_DATA_APIS = [
@@ -8,6 +8,7 @@ const MODEL_DATA_APIS = [
 	"openai-chatgpt-responses",
 	"anthropic-messages",
 	"google-generative-ai",
+	"google-interactions",
 	"google-vertex",
 	"github-copilot",
 	"bedrock-converse-stream",

package/dist/{openai-responses-BJ_IEoMU.mjs → openai-responses-CKCBEc2W.mjs}

@@ -1,15 +1,16 @@
 import { t as __exportAll } from "./rolldown-runtime-BhDjJH2R.mjs";
-import { t as AssistantMessageEventStream } from "./event-stream-D8PARQfL.mjs";
-import { n as clampOpenAIPromptCacheKey, r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-CjV_2IRI.mjs";
-import { i as resolveOpenAICompletionsCompat } from "./openai-completions-compat-CUollwl6.mjs";
+import { t as AssistantMessageEventStream } from "./event-stream-vd9UnCAU.mjs";
+import { n as clampOpenAIPromptCacheKey, r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-D2V7MqR7.mjs";
+import { i as resolveOpenAICompletionsCompat } from "./openai-completions-compat-BOVWBZDn.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
-import { n as buildBaseOptions } from "./simple-options-np1XPuhm.mjs";
-import { g as resolveOpenAISimpleReasoningEffort } from "./openai-transport-params-DZKOTSfH.mjs";
-import { n as getEnvApiKey } from "./env-api-keys-bktO00EJ.mjs";
-import { n as resolveProviderSimpleCompletionHeaders } from "./provider-transport-turn-state-lrXEFkYr.mjs";
-import { a as runResponsesStreamLifecycle, i as createResponsesAssistantOutput, n as applyResponsesServiceTierPricing, r as convertResponsesMessages, t as applyCommonResponsesParams } from "./openai-responses-shared-DvSqtJ6_.mjs";
+import { n as buildBaseOptions } from "./simple-options-CGAkCwtM.mjs";
+import { g as resolveOpenAISimpleReasoningEffort } from "./openai-transport-params-BwlpzX-2.mjs";
+import { n as getEnvApiKey } from "./env-api-keys-CyX4VKGO.mjs";
+import { n as createAssistantOutput } from "./provider-transcript-transform-JbFbHV8h.mjs";
+import { n as resolveProviderSimpleCompletionHeaders } from "./provider-transport-turn-state-BheUi_hl.mjs";
+import { P as convertProviderResponsesMessages, n as applyResponsesServiceTierPricing, r as runResponsesStreamLifecycle, t as applyCommonResponsesParams } from "./openai-responses-shared-DL4ulQn7.mjs";
 import { n as hasCopilotVisionInput, t as buildCopilotDynamicHeaders } from "./github-copilot-headers-B37TB3rB.mjs";
-import { t as createOpenAIProviderClient } from "./openai-provider-client-BP8J1aeg.mjs";
+import { t as createOpenAIProviderClient } from "./openai-provider-client-CwfsGBbA.mjs";
 //#region packages/ai/src/providers/openai-responses.ts
 var openai_responses_exports = /* @__PURE__ */ __exportAll({
 	streamOpenAIResponses: () => streamOpenAIResponses,
@@ -24,7 +25,7 @@ function getCompat(model) {
 }
 const streamOpenAIResponses = (model, context, options) => {
 	const stream = new AssistantMessageEventStream();
-	const output = createResponsesAssistantOutput(model);
+	const output = createAssistantOutput(model);
 	runResponsesStreamLifecycle({
 		stream,
 		model,
@@ -74,7 +75,7 @@ function createClient(model, context, apiKey, optionsHeaders, sessionId) {
 	return createOpenAIProviderClient(model, apiKey, headers, optionsHeaders);
 }
 function buildParams(model, context, options, replayMode = "checkpoint") {
-	const messages = convertResponsesMessages(model, context, OPENAI_TOOL_CALL_PROVIDERS, {
+	const messages = convertProviderResponsesMessages(model, context, OPENAI_TOOL_CALL_PROVIDERS, {
 		replayResponsesItemIds: options?.replayResponsesItemIds ?? false,
 		sessionId: options?.sessionId,
 		authProfileId: options?.authProfileId,

package/dist/openai-responses-compaction-replay-DiYn-gFM.d.mts

@@ -0,0 +1,12 @@
+import "./index-D6KsSE74.mjs";
+import { O as Model, a as AssistantMessage } from "./types-CLLR3eB4.mjs";
+import "./provider-options-Df7ov0K1.mjs";
+import { i as ReplayableResponseCompactionItem, r as OpenAIResponsesReasoningReplayMetadata, s as OpenAIResponsesCompactionOutput } from "./openai-responses-contracts-D1iZ0njF.mjs";
+import "openai/resources/responses/responses.js";
+//#region packages/ai/src/transports/openai-responses-compaction-replay.d.ts
+declare function captureOpenAIResponsesCompaction(output: Pick<AssistantMessage, "providerReplay">, item: ReplayableResponseCompactionItem, boundary: number | "retained-users", model: Model, captureMetadata?: OpenAIResponsesReasoningReplayMetadata, compactedOutput?: OpenAIResponsesCompactionOutput): void;
+declare class CompactionReplayRefreshRequiredError extends Error {
+  constructor();
+}
+//#endregion
+export { captureOpenAIResponsesCompaction as n, CompactionReplayRefreshRequiredError as t };

package/dist/{openai-responses-compaction-window-BiNCT0bS.mjs → openai-responses-compaction-window-UNK1GDpy.mjs}

@@ -1,9 +1,9 @@
+import { o as isRecord } from "./record-coerce-CNDovvft.mjs";
 import { i as normalizeOptionalLowercaseString, n as normalizeLowercaseStringOrEmpty, s as readStringValue } from "./string-coerce-fsri9iCu.mjs";
-import { o as isRecord } from "./record-coerce-DwRYMj3t.mjs";
-import { u as supportsOpenAIReasoningEffort } from "./openai-reasoning-effort-ash8dPrd.mjs";
-import { t as canonicalizeBase64 } from "./base64-CPmadnUf.mjs";
-import { t as parsePositiveInteger } from "./positive-integer-BiUxrd71.mjs";
-import { i as OPENAI_RESPONSES_APIS } from "./openai-responses-contracts-DWrfMODE.mjs";
+import { u as supportsOpenAIReasoningEffort } from "./openai-reasoning-effort-ConhwkeX.mjs";
+import { t as canonicalizeBase64 } from "./base64-C1PsYcRQ.mjs";
+import { t as parsePositiveInteger } from "./positive-integer-CjN5OJbH.mjs";
+import { i as OPENAI_RESPONSES_APIS } from "./openai-responses-contracts-CLxnqS9p.mjs";
 import { t as createSseByteGuard } from "./streaming-byte-guard-CC-HMn_u.mjs";
 //#region packages/model-catalog-core/src/provider-id.ts
 function normalizeProviderId(provider) {