@openclaw/ai 2026.7.2-beta.7 → 2026.8.1-beta.1

package/dist/transports.mjs

@@ -1,364 +1,45 @@
 import { n as getEnvApiKey } from "./env-api-keys-DrgeBuva.mjs";
-import { d as resolveClaudeSonnet5ModelIdentity, g as supportsClaudeNativeXhighEffort, h as supportsClaudeNativeMaxEffort, l as resolveClaudeNativeThinkingLevelMap, p as supportsClaudeAdaptiveThinking, u as resolveClaudeOpus5ModelIdentity } from "./src-QkygScBs.mjs";
-import { r as createAssistantMessageEventStream } from "./event-stream-D8n2uFee.mjs";
-import { a as applyClaudeRequestContract, c as requiresClaudeAdaptiveThinking, d as usesClaudeStreamingRefusalContract, f as hasNonEmptyString, g as isRecord, h as readStringValue, n as getAiTransportHost, o as defaultsClaudeAdaptiveThinking, p as normalizeLowercaseStringOrEmpty, r as resolveAiTransportHeaderSentinels, s as prepareClaudeNoPrefillRequestContext, u as usesClaudeFable5MessagesContract } from "./host-Dog2WQiR.mjs";
+import { d as resolveClaudeSonnet5ModelIdentity, g as supportsClaudeNativeXhighEffort, p as supportsClaudeAdaptiveThinking, u as resolveClaudeOpus5ModelIdentity } from "./src-QkygScBs.mjs";
+import { r as createAssistantMessageEventStream } from "./event-stream-CEV9t6da.mjs";
+import { _ as normalizeLowercaseStringOrEmpty, a as ANTHROPIC_CLAUDE_CODE_BILLING_SYSTEM_BLOCK, b as isRecord, c as defaultsClaudeAdaptiveThinking, d as requiresClaudeAdaptiveThinking, f as resolveAnthropicThinkingEffort, g as hasNonEmptyString, h as usesClaudeStreamingRefusalContract, l as mapAnthropicStopReason, m as usesClaudeFable5MessagesContract, n as getAiTransportHost, o as ANTHROPIC_CLAUDE_CODE_VERSION, r as resolveAiTransportHeaderSentinels, s as applyClaudeRequestContract, u as prepareClaudeNoPrefillRequestContext, y as readStringValue } from "./host-Bl7Kgddo.mjs";
+import { n as calculateCost } from "./model-utils-Dau5dlgm.mjs";
 import { n as clampOpenAIPromptCacheKey } from "./openai-prompt-cache-mZTCdRPo.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
-import { a as isImageWithMediaPayload, d as stripSystemPromptCacheBoundary, m as sortPromptCacheToolsByName, n as extractToolResultBlockText, r as extractToolResultText, t as describeToolResultMediaPlaceholder, u as splitSystemPromptCacheBoundary } from "./tool-result-text-CTpIRbYd.mjs";
-import { a as resolveModelRequestTimeoutMs, c as resolveProviderRequestCapabilities, i as buildGuardedModelFetch, l as resolveProviderRequestPolicyConfig, n as reconcileOpenAICompletionsToolChoice, o as resolveOpenAIStrictToolSetting, r as reconcileOpenAIResponsesToolChoice, s as resolveProviderEndpoint, t as projectOpenAITools, u as transformTransportMessages } from "./openai-tool-projection-OhX64DoP.mjs";
-import { t as toErrorObject } from "./error-coercion-DgxlWC0n.mjs";
-import { S as canonicalizeBase64, _ as omitFoundryBearerCredentialHeaders, a as projectAnthropicTools, b as normalizeAnthropicInlineContent, c as ANTHROPIC_OMITTED_REASONING_TEXT, d as ANTHROPIC_SERVER_SIDE_FALLBACK_BETA, g as applyAnthropicRefusal, h as resolveAnthropicFallbackServingModelCost, i as readLastAnthropicIterationUsage, l as findActiveAnthropicToolTurnAssistantIndex, m as readAnthropicFallbackBoundary, n as readAnthropicPromptUsageSnapshot, o as reconcileAnthropicToolChoice, p as applyAnthropicFallbackBoundary, r as readAnthropicUsageTokenCount, s as resolveOriginalAnthropicToolName, u as ANTHROPIC_SERVER_SIDE_FALLBACKS, v as usesFoundryBearerAuth, x as resolveAnthropicImageMediaType, y as createAnthropicInlineImageBudget } from "./anthropic-usage-DWU-x8MI.mjs";
-import { a as parseStrictPositiveInteger, c as calculateCost, l as clampThinkingLevel, s as applyProviderReportedUsageCost } from "./number-coercion-DvG7SNMg.mjs";
-import { a as resolveOpenAICompletionsCompat, c as clearPendingCommentaryText, i as detectOpenAICompletionsCompat, l as rememberPendingCommentaryTags, n as mapOpenAIStopReason, o as resolveOpenAICompletionsResponseFormat, r as convertMessages, s as shouldOmitOllamaCompatResponseFormat, t as resolveOpenAIReasoningEffortMap, u as tagPendingCommentaryText } from "./openai-reasoning-compat-YgeLncHw.mjs";
+import { a as isImageWithMediaPayload, h as stripSystemPromptCacheBoundary, n as extractToolResultBlockText, r as extractToolResultText, s as adjustMaxTokensForThinking, t as describeToolResultMediaPlaceholder, v as sortPromptCacheToolsByName } from "./tool-result-text-Dvkp2Dus.mjs";
+import { a as resolveProviderEndpoint, c as transformTransportMessages, i as resolveOpenAIStrictToolSetting, n as buildGuardedModelFetch, r as resolveModelRequestTimeoutMs, s as resolveProviderRequestPolicyConfig } from "./tool-schema-json-projection-B1b-XCn5.mjs";
+import { A as applyAnthropicPayloadPolicyToParams, C as usesFoundryBearerAuth, D as canonicalizeBase64, E as resolveAnthropicImageMediaType, M as resolveAnthropicPayloadPolicy, O as applyAnthropicCacheControlToMessages, S as omitFoundryBearerCredentialHeaders, T as normalizeAnthropicInlineContent, b as resolveAnthropicFallbackServingModelCost, c as normalizeAnthropicToolChoice, d as resolveOriginalAnthropicToolName, f as toClaudeCodeToolName, g as ANTHROPIC_SERVER_SIDE_FALLBACK_BETA, h as ANTHROPIC_SERVER_SIDE_FALLBACKS, j as resolveAnthropicEphemeralCacheControl, k as applyAnthropicEphemeralCacheControlMarkers, l as projectAnthropicTools, m as findActiveAnthropicToolTurnAssistantIndex, n as applyAnthropicMessageStartUsage, p as ANTHROPIC_OMITTED_REASONING_TEXT, s as normalizeAnthropicToolCallId, t as applyAnthropicMessageDeltaUsage, u as reconcileAnthropicToolChoice, v as applyAnthropicFallbackBoundary, w as createAnthropicInlineImageBudget, x as applyAnthropicRefusal, y as readAnthropicFallbackBoundary } from "./anthropic-usage-Ma4iX6uG.mjs";
+import { n as toErrorObject, t as createResponsesPromptEgressObserver } from "./openai-responses-prompt-observer-internal-DDZPxRfr.mjs";
+import { a as convertMessages, c as resolveOpenAICompletionsResponseFormat, d as rememberPendingCommentaryTags, f as tagPendingCommentaryText, i as finalizeOpenAICompletionsToolCalls, l as shouldOmitOllamaCompatResponseFormat, n as mapOpenAIStopReason, o as detectOpenAICompletionsCompat, r as createOpenAICompletionsToolCallDeltaNormalizer, s as resolveOpenAICompletionsCompat, t as resolveOpenAIReasoningEffortMap, u as clearPendingCommentaryText } from "./openai-reasoning-compat-DKebnIBL.mjs";
 import { t as createDeferredEventBuffer } from "./deferred-event-buffer-DAvyP7qA.mjs";
 import { n as parseStreamingJson } from "./json-parse-BvXNt1-7.mjs";
 import { t as notifyLlmRequestActivity } from "./llm-request-activity-BjtkplhG.mjs";
-import { C as resolveSecretSentinel, S as resolveModelHeaderSentinels$1, T as supportsModelTools, _ as isGoogleGemini3ProModel, a as failTransportStream, c as mergeTransportMetadata, d as transportAbortError, f as MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE, g as isGoogleGemini3FlashModel, h as isCodeModeModelVisibleToolName, i as createWritableTransportEventStream, l as sanitizeNonEmptyTransportPayloadText, m as estimateStringChars, n as coerceTransportToolCallArguments, o as finalizeTransportStream, p as createAbortError$1, r as createEmptyTransportUsage, s as mergeTransportHeaders, t as assignTransportErrorDetails, u as sanitizeTransportPayloadText, v as parseRetryAfterSeconds, w as sha256Hex, y as readResponseTextSnippet } from "./transport-stream-shared-D81p90xq.mjs";
-import { A as normalizeResponsesFailedEvent, B as resolveReplayableResponsesMessageId, C as prepareOpenAIResponsesReasoningItemForReplay, D as applyServiceTierPricing, E as tagOpenAIResponsesReasoningReplayItem, F as summarizeResponsesFailedNoDetailsObservation, G as throwIfModelStreamAborted, H as createModelStreamCooperativeScheduler, I as summarizeResponsesPayload, L as summarizeResponsesTools, M as stringifyRedactedEvent, N as stringifyRedactedPayload, O as buildResponsesFailedNoDetailsObservation, P as summarizeOpenAITransportError, R as AZURE_RESPONSES_FIRST_EVENT_TIMEOUT_MS, S as isInvalidEncryptedContentError, T as stripResponsesRequestEncryptedContent, U as log, V as GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, W as resolvePromptCacheKey, Y as normalizeOpenAIStrictToolParameters, Z as resolveOpenAIProjectedToolsStrictToolFlag, b as convertResponsesMessages, bt as resolveModelPayloadDebugMode, dt as isOpenAIGpt55Model, ft as isOpenAIGpt56Model, gt as supportsOpenAIReasoningEffort, j as safeDebugValue, k as logResponsesFailedNoDetails, mt as resolveOpenAIReasoningEffortForModel, n as processResponsesStream, o as observeResponsesStream, pt as normalizeOpenAIReasoningEffort, q as findOpenAIStrictToolProjectionDiagnostics, t as ResponsesStreamFailure, ut as isOpenAIGpt54MiniModel, v as buildOpenAIResponsesReasoningReplayMetadata, vt as uniqueStrings, w as resolveAzureOpenAIApiVersion, x as createResponsesStreamWithEncryptedContentRetry, xt as resolveModelSseDebugMode, y as buildResponsesInputMessage, yt as emitModelTransportDebug, z as OPENAI_CODEX_RESPONSES_DEFAULT_INSTRUCTIONS } from "./openai-responses-stream-internal-Cw5txaGW.mjs";
+import { A as normalizeResponsesFailedEvent, C as prepareOpenAIResponsesReasoningItemForReplay, D as applyServiceTierPricing, E as tagOpenAIResponsesReasoningReplayItem, F as summarizeResponsesFailedNoDetailsObservation, G as normalizeOpenAIStrictToolParameters, I as summarizeResponsesPayload, L as summarizeResponsesTools, M as stringifyRedactedEvent, N as stringifyRedactedPayload, O as buildResponsesFailedNoDetailsObservation, P as summarizeOpenAITransportError, R as AZURE_RESPONSES_FIRST_EVENT_TIMEOUT_MS, S as isInvalidEncryptedContentError, T as stripResponsesRequestEncryptedContent, U as findOpenAIStrictToolProjectionDiagnostics, V as resolveReplayableResponsesMessageId, _t as resolveModelSseDebugMode, b as convertResponsesMessages, bt as quoteUnsafeIntegerLiterals, ct as isOpenAIGpt56Model, ft as supportsOpenAIReasoningEffort, gt as resolveModelPayloadDebugMode, ht as emitModelTransportDebug, j as safeDebugValue, k as logResponsesFailedNoDetails, lt as normalizeOpenAIReasoningEffort, mt as uniqueStrings, n as processResponsesStream, o as observeResponsesStream, ot as isOpenAIGpt54MiniModel, q as resolveOpenAIProjectedToolsStrictToolFlag, st as isOpenAIGpt55Model, t as ResponsesStreamFailure, ut as resolveOpenAIReasoningEffortForModel, v as buildOpenAIResponsesReasoningReplayMetadata, vt as parseJsonObjectPreservingUnsafeIntegers, w as resolveAzureOpenAIApiVersion, x as createResponsesStreamWithEncryptedContentRetry, y as buildResponsesInputMessage, yt as parseJsonPreservingUnsafeIntegers, z as OPENAI_CODEX_RESPONSES_DEFAULT_INSTRUCTIONS } from "./openai-responses-stream-internal-Bl8YyVtF.mjs";
+import { a as parseStrictPositiveInteger } from "./number-coercion-1Miyb5MO.mjs";
+import { C as resolveSecretSentinel, S as resolveModelHeaderSentinels$1, T as supportsModelTools, _ as isGoogleGemini3ProModel, a as failTransportStream, c as mergeTransportMetadata, d as transportAbortError, f as MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE, g as isGoogleGemini3FlashModel, h as isCodeModeModelVisibleToolName, i as createWritableTransportEventStream, l as sanitizeNonEmptyTransportPayloadText, m as estimateStringChars, n as coerceTransportToolCallArguments, o as finalizeTransportStream, p as createAbortError$1, r as createEmptyTransportUsage, s as mergeTransportHeaders, t as assignTransportErrorDetails, u as sanitizeTransportPayloadText, v as parseRetryAfterSeconds, w as sha256Hex, y as readResponseTextSnippet } from "./transport-stream-shared-CPNv7A3r.mjs";
+import { a as parseOpenAICompletionsUsage, c as throwIfModelStreamAborted, d as reconcileOpenAIResponsesToolChoice, i as log, l as projectOpenAITools, n as createModelStreamCooperativeScheduler, o as readOpenAICompletionsContentDeltas, r as isOpenAICompletionsThinkingEnabled, s as resolvePromptCacheKey, t as GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, u as reconcileOpenAICompletionsToolChoice } from "./openai-transport-shared-Cipt7egQ.mjs";
 import { t as createReasoningTagTextPartitioner } from "./reasoning-tag-text-partitioner-CGDyLWUR.mjs";
-import { a as withFirstStreamEventTimeout, i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-BBys9hSb.mjs";
+import { a as withFirstStreamEventTimeout, i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-BIBomOGq.mjs";
 import { i as resolveAzureDeploymentNameFromMap, t as isOpenAICompatibleAzureResponsesBaseUrl } from "./azure-openai-responses-client-compat-C7K7QfUE.mjs";
+import { t as headersToRecord } from "./headers-B_e4-1J0.mjs";
 import { randomUUID } from "node:crypto";
 import OpenAI, { AzureOpenAI } from "openai";
-//#region packages/ai/src/transports/anthropic-payload-policy.ts
-/**
-* Anthropic-family request payload policy helpers.
-* Applies service-tier and cache-control markers only when provider endpoint
-* capabilities allow them.
-*/
-const ANTHROPIC_CACHE_CONTROL_LIMIT = 4;
-function resolveBaseUrlHostname(baseUrl) {
-	try {
-		return new URL(baseUrl).hostname;
-	} catch {
-		return;
-	}
-}
-function isLongTtlEligibleEndpoint(baseUrl) {
-	if (typeof baseUrl !== "string") return false;
-	const hostname = resolveBaseUrlHostname(baseUrl);
-	if (!hostname) return false;
-	return hostname === "api.anthropic.com" || hostname === "aiplatform.googleapis.com" || hostname.endsWith("-aiplatform.googleapis.com");
-}
-/** Resolve Anthropic cache-control marker retention for a request endpoint. */
-function resolveAnthropicEphemeralCacheControl(baseUrl, cacheRetention) {
-	const retention = resolveCacheRetention(cacheRetention);
-	if (retention === "none") return;
-	const ttl = retention === "long" && (cacheRetention === "long" || isLongTtlEligibleEndpoint(baseUrl)) ? "1h" : void 0;
-	return {
-		type: "ephemeral",
-		...ttl ? { ttl } : {}
-	};
-}
-function applyAnthropicCacheControlToSystem(system, cacheControl) {
-	if (!Array.isArray(system)) return;
-	const normalizedBlocks = [];
-	for (const block of system) {
-		if (!block || typeof block !== "object") {
-			normalizedBlocks.push(block);
-			continue;
-		}
-		const record = block;
-		if (record.type !== "text" || typeof record.text !== "string") {
-			normalizedBlocks.push(block);
-			continue;
-		}
-		const split = splitSystemPromptCacheBoundary(record.text);
-		if (!split) {
-			if (record.cache_control === void 0) record.cache_control = cacheControl;
-			normalizedBlocks.push(record);
-			continue;
-		}
-		const { cache_control: existingCacheControl, ...rest } = record;
-		if (split.stablePrefix) normalizedBlocks.push({
-			...rest,
-			text: split.stablePrefix,
-			cache_control: existingCacheControl ?? cacheControl
-		});
-		if (split.dynamicSuffix) normalizedBlocks.push({
-			...rest,
-			text: split.dynamicSuffix
-		});
-	}
-	system.splice(0, system.length, ...normalizedBlocks);
-}
-function stripAnthropicSystemPromptBoundary(system) {
-	if (!Array.isArray(system)) return;
-	for (const block of system) {
-		if (!block || typeof block !== "object") continue;
-		const record = block;
-		if (record.type === "text" && typeof record.text === "string") record.text = stripSystemPromptCacheBoundary(record.text);
-	}
-}
-function applyAnthropicCacheControlToMessages(messages, cacheControl, markerLimit, cacheBreakpointOptOutMessageIndexes) {
-	if (!Array.isArray(messages) || messages.length === 0 || markerLimit <= 0) return;
-	let fallbackToolResult;
-	for (let i = messages.length - 1; i >= 0; i--) {
-		const message = messages[i];
-		if (!message || typeof message !== "object") continue;
-		const record = message;
-		if (record.role !== "user" || cacheBreakpointOptOutMessageIndexes.has(i)) continue;
-		const content = record.content;
-		if (typeof content === "string") {
-			if (fallbackToolResult && markerLimit === 1) {
-				fallbackToolResult.cache_control = cacheControl;
-				return;
-			}
-			record.content = [{
-				type: "text",
-				text: content,
-				cache_control: cacheControl
-			}];
-			if (fallbackToolResult && markerLimit > 1) fallbackToolResult.cache_control = cacheControl;
-			return;
-		}
-		if (!Array.isArray(content)) continue;
-		for (let j = content.length - 1; j >= 0; j--) {
-			const block = content[j];
-			if (!block || typeof block !== "object") continue;
-			const blockRecord = block;
-			if (blockRecord.type === "text" || blockRecord.type === "image") {
-				if (fallbackToolResult && markerLimit === 1) {
-					fallbackToolResult.cache_control = cacheControl;
-					return;
-				}
-				blockRecord.cache_control = cacheControl;
-				if (fallbackToolResult && markerLimit > 1) fallbackToolResult.cache_control = cacheControl;
-				return;
-			}
-			if (blockRecord.type === "tool_result" && fallbackToolResult === void 0) fallbackToolResult = blockRecord;
-		}
-	}
-	if (fallbackToolResult) fallbackToolResult.cache_control = cacheControl;
-}
-function countAnthropicCacheControlMarkers(blocks) {
-	if (!Array.isArray(blocks)) return 0;
-	let count = 0;
-	for (const block of blocks) if (block && typeof block === "object" && "cache_control" in block) count += 1;
-	return count;
-}
-/** @deprecated Anthropic-family provider payload helper; do not use from third-party plugins. */
-function resolveAnthropicPayloadPolicy(input) {
-	return {
-		allowsServiceTier: resolveProviderRequestCapabilities({
-			provider: input.provider,
-			api: input.api,
-			baseUrl: input.baseUrl,
-			capability: "llm",
-			transport: "stream"
-		}).allowsAnthropicServiceTier,
-		cacheControl: input.enableCacheControl === true ? resolveAnthropicEphemeralCacheControl(input.baseUrl, input.cacheRetention) : void 0,
-		serviceTier: input.serviceTier
-	};
-}
-/** @deprecated Anthropic-family provider payload helper; do not use from third-party plugins. */
-function applyAnthropicPayloadPolicyToParams(payloadObj, policy, cacheBreakpointOptOutMessageIndexes) {
-	if (policy.allowsServiceTier && policy.serviceTier !== void 0 && payloadObj.service_tier === void 0) payloadObj.service_tier = policy.serviceTier;
-	if (policy.cacheControl) applyAnthropicCacheControlToSystem(payloadObj.system, policy.cacheControl);
-	else stripAnthropicSystemPromptBoundary(payloadObj.system);
-	if (!policy.cacheControl) return;
-	const usedMarkers = countAnthropicCacheControlMarkers(payloadObj.system) + countAnthropicCacheControlMarkers(payloadObj.tools);
-	applyAnthropicCacheControlToMessages(payloadObj.messages, policy.cacheControl, ANTHROPIC_CACHE_CONTROL_LIMIT - usedMarkers, cacheBreakpointOptOutMessageIndexes);
-}
-/** @deprecated Anthropic-family provider payload helper; do not use from third-party plugins. */
-function applyAnthropicEphemeralCacheControlMarkers(payloadObj, cacheControl = { type: "ephemeral" }) {
-	const messages = payloadObj.messages;
-	if (!Array.isArray(messages)) return;
-	for (const message of messages) {
-		if (message.role === "system" || message.role === "developer") {
-			if (!cacheControl) continue;
-			if (typeof message.content === "string") {
-				message.content = [{
-					type: "text",
-					text: message.content,
-					cache_control: cacheControl
-				}];
-				continue;
-			}
-			if (Array.isArray(message.content) && message.content.length > 0) {
-				const last = message.content[message.content.length - 1];
-				if (last && typeof last === "object") {
-					const record = last;
-					if (record.type !== "thinking" && record.type !== "redacted_thinking") record.cache_control = cacheControl;
-				}
-			}
-			continue;
-		}
-		if (message.role === "assistant" && Array.isArray(message.content)) for (const block of message.content) {
-			if (!block || typeof block !== "object") continue;
-			const record = block;
-			if (record.type === "thinking" || record.type === "redacted_thinking") delete record.cache_control;
-		}
-	}
-}
-//#endregion
-//#region packages/ai/src/transports/json-unsafe-integers.ts
-/**
-* JSON parsing helpers that preserve integer literals larger than
-* Number.MAX_SAFE_INTEGER as strings before JSON.parse can round them.
-*/
-const MAX_SAFE_INTEGER_ABS_STR = String(Number.MAX_SAFE_INTEGER);
-function isAsciiDigit(ch) {
-	return ch !== void 0 && ch >= "0" && ch <= "9";
-}
-function parseJsonNumberToken(input, start) {
-	let idx = start;
-	if (input[idx] === "-") idx += 1;
-	if (idx >= input.length) return null;
-	if (input[idx] === "0") idx += 1;
-	else if (isAsciiDigit(input[idx]) && input[idx] !== "0") while (isAsciiDigit(input[idx])) idx += 1;
-	else return null;
-	let isInteger = true;
-	if (input[idx] === ".") {
-		isInteger = false;
-		idx += 1;
-		if (!isAsciiDigit(input[idx])) return null;
-		while (isAsciiDigit(input[idx])) idx += 1;
-	}
-	if (input[idx] === "e" || input[idx] === "E") {
-		isInteger = false;
-		idx += 1;
-		if (input[idx] === "+" || input[idx] === "-") idx += 1;
-		if (!isAsciiDigit(input[idx])) return null;
-		while (isAsciiDigit(input[idx])) idx += 1;
-	}
-	return {
-		token: input.slice(start, idx),
-		end: idx,
-		isInteger
-	};
-}
-function isUnsafeIntegerLiteral(token) {
-	const digits = token[0] === "-" ? token.slice(1) : token;
-	if (digits.length < MAX_SAFE_INTEGER_ABS_STR.length) return false;
-	if (digits.length > MAX_SAFE_INTEGER_ABS_STR.length) return true;
-	return digits > MAX_SAFE_INTEGER_ABS_STR;
-}
-/** Quotes integer literals above Number.MAX_SAFE_INTEGER before JSON.parse. */
-function quoteUnsafeIntegerLiterals(input) {
-	let out = "";
-	let inString = false;
-	let escaped = false;
-	let idx = 0;
-	while (idx < input.length) {
-		const ch = input[idx] ?? "";
-		if (inString) {
-			out += ch;
-			if (escaped) escaped = false;
-			else if (ch === "\\") escaped = true;
-			else if (ch === "\"") inString = false;
-			idx += 1;
-			continue;
-		}
-		if (ch === "\"") {
-			inString = true;
-			out += ch;
-			idx += 1;
-			continue;
-		}
-		if (ch === "-" || isAsciiDigit(ch)) {
-			const parsed = parseJsonNumberToken(input, idx);
-			if (parsed) {
-				if (parsed.isInteger && isUnsafeIntegerLiteral(parsed.token)) out += `"${parsed.token}"`;
-				else out += parsed.token;
-				idx = parsed.end;
-				continue;
-			}
-		}
-		out += ch;
-		idx += 1;
-	}
-	return out;
-}
-/** Parses JSON while preserving unsafe integer literals as strings. */
-function parseJsonPreservingUnsafeIntegers(input) {
-	return JSON.parse(quoteUnsafeIntegerLiterals(input));
-}
-/** Parses or accepts an object while preserving unsafe integer literals in string input. */
-function parseJsonObjectPreservingUnsafeIntegers(value) {
-	if (typeof value === "string") {
-		try {
-			const parsed = parseJsonPreservingUnsafeIntegers(value);
-			if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) return parsed;
-		} catch {
-			return null;
-		}
-		return null;
-	}
-	if (value && typeof value === "object" && !Array.isArray(value)) return value;
-	return null;
-}
-//#endregion
 //#region packages/ai/src/transports/anthropic-transport-stream.ts
 /**
 * Native Anthropic Messages streaming transport.
 * Converts OpenClaw contexts/tools into Anthropic payloads, streams SSE events
 * back into runtime output blocks, and applies provider request policy.
 */
-const CLAUDE_CODE_VERSION = "2.1.75";
-const CLAUDE_CODE_BILLING_SYSTEM_BLOCK = `x-anthropic-billing-header: cc_version=${CLAUDE_CODE_VERSION}; cc_entrypoint=sdk-cli;`;
 const ANTHROPIC_MESSAGES_ERROR_BODY_MAX_BYTES = 8 * 1024;
 const ANTHROPIC_MESSAGES_ERROR_BODY_MAX_CHARS = 400;
 const ANTHROPIC_MESSAGES_ERROR_BODY_READ_IDLE_TIMEOUT_MS = 1e4;
 const ANTHROPIC_MESSAGES_DEFAULT_MAX_TOKENS = 4096;
 const ANTHROPIC_MESSAGES_FALLBACK_CONTEXT_DIVISOR = 4;
 const ANTHROPIC_MESSAGES_SSE_PENDING_BUFFER_MAX_CHARS = 16 * 1024 * 1024;
-const CLAUDE_CODE_TOOL_LOOKUP = new Map([
-	"Read",
-	"Write",
-	"Edit",
-	"Bash",
-	"Grep",
-	"Glob",
-	"AskUserQuestion",
-	"EnterPlanMode",
-	"ExitPlanMode",
-	"KillShell",
-	"NotebookEdit",
-	"Skill",
-	"Task",
-	"TaskOutput",
-	"TodoWrite",
-	"WebFetch",
-	"WebSearch"
-].map((tool) => [normalizeLowercaseStringOrEmpty(tool), tool]));
 function resolveAnthropicRequestModelId(model) {
 	if (isDirectAnthropicModel(model) && /^anthropic\//i.test(model.id)) return model.id.replace(/^anthropic\//i, "");
 	return model.id;
 }
 const EMPTY_ANTHROPIC_MESSAGES_FALLBACK_TEXT = ".";
-function normalizeAnthropicToolChoice(thinkingEnabled, toolChoice) {
-	if (thinkingEnabled && (toolChoice === "any" || typeof toolChoice === "object" && toolChoice.type === "tool")) return { type: "auto" };
-	return typeof toolChoice === "string" ? { type: toolChoice } : toolChoice;
-}
-function supportsNativeXhighEffort(model) {
-	return supportsClaudeNativeXhighEffort(model);
-}
-function supportsAdaptiveThinking(model) {
-	return supportsClaudeAdaptiveThinking(model);
-}
-function mapThinkingLevelToEffort(level, model) {
-	const thinkingLevelMap = resolveClaudeNativeThinkingLevelMap(model);
-	const resolvedLevel = clampThinkingLevel({
-		...model,
-		...typeof model.params?.canonicalModelId === "string" ? { reasoning: true } : {},
-		...thinkingLevelMap ? { thinkingLevelMap } : {}
-	}, level);
-	const mapped = thinkingLevelMap?.[resolvedLevel];
-	if (typeof mapped === "string") return mapped;
-	switch (resolvedLevel) {
-		case "off":
-		case "minimal":
-		case "low": return "low";
-		case "medium": return "medium";
-		case "xhigh": return supportsNativeXhighEffort(model) ? "xhigh" : "high";
-		case "max": return supportsClaudeNativeMaxEffort(model) ? "max" : "high";
-		default: return "high";
-	}
-}
-function clampReasoningLevel(level) {
-	return level === "xhigh" || level === "max" ? "high" : level;
-}
 function resolvePositiveAnthropicTokenLimit(value) {
 	if (typeof value !== "number" || !Number.isFinite(value)) return;
 	const floored = Math.floor(value);
@@ -373,23 +54,6 @@ function resolveAnthropicMessagesMaxTokens(params) {
 	const contextWindow = resolvePositiveAnthropicTokenLimit(params.modelContextWindow);
 	return contextWindow === void 0 ? ANTHROPIC_MESSAGES_DEFAULT_MAX_TOKENS : Math.max(1, Math.min(ANTHROPIC_MESSAGES_DEFAULT_MAX_TOKENS, Math.floor(contextWindow / ANTHROPIC_MESSAGES_FALLBACK_CONTEXT_DIVISOR)));
 }
-function adjustMaxTokensForThinking(params) {
-	const budgets = {
-		minimal: 1024,
-		low: 2048,
-		medium: 8192,
-		high: 16384,
-		...params.customBudgets
-	};
-	const minOutputTokens = 1024;
-	let thinkingBudget = budgets[clampReasoningLevel(params.reasoningLevel)];
-	const maxTokens = Math.min(params.baseMaxTokens + thinkingBudget, params.modelMaxTokens);
-	if (maxTokens <= thinkingBudget) thinkingBudget = Math.max(0, maxTokens - minOutputTokens);
-	return {
-		maxTokens,
-		thinkingBudget
-	};
-}
 function isAnthropicOAuthToken(apiKey) {
 	return (resolveSecretSentinel(apiKey) ?? apiKey).includes("sk-ant-oat");
 }
@@ -416,9 +80,6 @@ function buildAnthropicBetaHeader(model, betaFeatures, params) {
 	if (!isDirectAnthropicModel(model)) return;
 	return params.oauth ? `claude-code-20250219,oauth-2025-04-20,${betaFeatures.join(",")}` : betaFeatures.join(",");
 }
-function toClaudeCodeName(name) {
-	return CLAUDE_CODE_TOOL_LOOKUP.get(normalizeLowercaseStringOrEmpty(name)) ?? name;
-}
 const NON_VISION_USER_IMAGE_PLACEHOLDER = "(image omitted: model does not support images)";
 async function convertContentBlocks(content, model, imageBudget) {
 	const text = extractToolResultText(content);
@@ -459,15 +120,12 @@ async function convertContentBlocks(content, model, imageBudget) {
 	});
 	return blocks;
 }
-function normalizeToolCallId(id) {
-	return id.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 64);
-}
 async function convertAnthropicMessages(messages, model, isOAuthToken, options) {
 	const params = [];
 	const imageBudget = createAnthropicInlineImageBudget();
 	const allowReasoningContentReplay = options.allowReasoningContentReplay === true;
 	const replayThinkingEnabled = options.replayThinkingEnabled !== false;
-	const transformedMessages = transformTransportMessages(messages, model, normalizeToolCallId);
+	const transformedMessages = transformTransportMessages(messages, model, normalizeAnthropicToolCallId);
 	const activeToolTurnAssistantIndex = replayThinkingEnabled ? -1 : findActiveAnthropicToolTurnAssistantIndex(transformedMessages);
 	for (let i = 0; i < transformedMessages.length; i += 1) {
 		const msg = transformedMessages[i];
@@ -566,7 +224,7 @@ async function convertAnthropicMessages(messages, model, isOAuthToken, options)
 				if (block.type === "toolCall") blocks.push({
 					type: "tool_use",
 					id: block.id,
-					name: isOAuthToken ? toClaudeCodeName(block.name) : block.name,
+					name: isOAuthToken ? toClaudeCodeToolName(block.name) : block.name,
 					input: coerceTransportToolCallArguments(block.arguments)
 				});
 			}
@@ -625,7 +283,7 @@ function ensureNonEmptyAnthropicMessages(messages) {
 	}];
 }
 function convertAnthropicTools(tools, isOAuthToken) {
-	const projection = projectAnthropicTools(tools ?? [], (name) => isOAuthToken ? toClaudeCodeName(name) : name);
+	const projection = projectAnthropicTools(tools ?? [], (name) => isOAuthToken ? toClaudeCodeToolName(name) : name);
 	const converted = [];
 	for (const tool of projection.tools) converted.push({
 		name: tool.wireName,
@@ -640,18 +298,6 @@ function convertAnthropicTools(tools, isOAuthToken) {
 function parseAnthropicToolCallArguments(inputJson) {
 	return parseJsonObjectPreservingUnsafeIntegers(inputJson) ?? parseStreamingJson(inputJson);
 }
-function mapStopReason(reason) {
-	switch (reason) {
-		case "end_turn": return "stop";
-		case "max_tokens": return "length";
-		case "tool_use": return "toolUse";
-		case "pause_turn": return "stop";
-		case "refusal":
-		case "sensitive": return "error";
-		case "stop_sequence": return "stop";
-		default: throw new Error(`Unhandled stop reason: ${String(reason)}`);
-	}
-}
 const DEFAULT_ANTHROPIC_BASE_URL = "https://api.anthropic.com";
 /** Resolve the effective Anthropic API base URL from model or environment. */
 function resolveAnthropicBaseUrl(baseUrl) {
@@ -794,7 +440,7 @@ async function readAnthropicMessagesErrorBodySnippet(response) {
 }
 function createAnthropicTransportClient(params) {
 	const { model, context, apiKey, options } = params;
-	const needsInterleavedBeta = (options?.interleavedThinking ?? true) && !supportsAdaptiveThinking(model);
+	const needsInterleavedBeta = (options?.interleavedThinking ?? true) && !supportsClaudeAdaptiveThinking(model);
 	const fetch = isKimiAnthropicProvider(model.provider) && options?.thinkingEnabled === true ? buildGuardedModelFetch(model, void 0, { sanitizeSse: false }) : buildGuardedModelFetch(model);
 	if (model.provider === "github-copilot") {
 		const betaFeatures = needsInterleavedBeta ? ["interleaved-thinking-2025-05-14"] : [];
@@ -843,7 +489,7 @@ function createAnthropicTransportClient(params) {
 					accept: "application/json",
 					"anthropic-dangerous-direct-browser-access": "true",
 					...betaHeader ? { "anthropic-beta": betaHeader } : {},
-					"user-agent": `claude-cli/${CLAUDE_CODE_VERSION}`,
+					"user-agent": `claude-cli/${ANTHROPIC_CLAUDE_CODE_VERSION}`,
 					"x-app": "cli"
 				}, model.headers, options?.headers),
 				fetch
@@ -899,7 +545,7 @@ async function buildAnthropicParams(model, context, isOAuthToken, options) {
 	if (isOAuthToken) params.system = [
 		{
 			type: "text",
-			text: CLAUDE_CODE_BILLING_SYSTEM_BLOCK
+			text: ANTHROPIC_CLAUDE_CODE_BILLING_SYSTEM_BLOCK
 		},
 		{
 			type: "text",
@@ -914,7 +560,7 @@ async function buildAnthropicParams(model, context, isOAuthToken, options) {
 		type: "text",
 		text: sanitizeTransportPayloadText(context.systemPrompt)
 	}];
-	if (options?.temperature !== void 0 && !options.thinkingEnabled && !supportsNativeXhighEffort(model)) params.temperature = options.temperature;
+	if (options?.temperature !== void 0 && !options.thinkingEnabled && !supportsClaudeNativeXhighEffort(model)) params.temperature = options.temperature;
 	if (options?.stop !== void 0 && options.stop.length > 0) params.stop_sequences = options.stop;
 	let toolProjection;
 	if (context.tools) {
@@ -922,8 +568,8 @@ async function buildAnthropicParams(model, context, isOAuthToken, options) {
 		toolProjection = convertedTools.projection;
 		if (convertedTools.tools.length > 0) params.tools = convertedTools.tools;
 	}
-	if (mandatoryAdaptiveThinking || model.reasoning || supportsAdaptiveThinking(model)) {
-		if (mandatoryAdaptiveThinking || options?.thinkingEnabled) if (supportsAdaptiveThinking(model)) {
+	if (mandatoryAdaptiveThinking || model.reasoning || supportsClaudeAdaptiveThinking(model)) {
+		if (mandatoryAdaptiveThinking || options?.thinkingEnabled) if (supportsClaudeAdaptiveThinking(model)) {
 			params.thinking = {
 				type: "adaptive",
 				display: options?.thinkingDisplay ?? "summarized"
@@ -985,17 +631,12 @@ function resolveAnthropicTransportOptions(model, options, apiKey) {
 		if (resolved.thinkingEnabled) resolved.effort = "high";
 		return resolved;
 	}
-	if (supportsAdaptiveThinking(model)) {
+	if (supportsClaudeAdaptiveThinking(model)) {
 		resolved.thinkingEnabled = true;
-		resolved.effort = mapThinkingLevelToEffort(reasoning, model);
+		resolved.effort = resolveAnthropicThinkingEffort(model, reasoning);
 		return resolved;
 	}
-	const adjusted = adjustMaxTokensForThinking({
-		baseMaxTokens,
-		modelMaxTokens: reasoningModelMaxTokens,
-		reasoningLevel: reasoning,
-		customBudgets: options?.thinkingBudgets
-	});
+	const adjusted = adjustMaxTokensForThinking(baseMaxTokens, reasoningModelMaxTokens, reasoning === "max" ? "high" : reasoning, options?.thinkingBudgets);
 	const thinkingEnabled = adjusted.thinkingBudget >= 1024;
 	resolved.maxTokens = adjusted.maxTokens;
 	resolved.thinkingEnabled = thinkingEnabled;
@@ -1153,23 +794,7 @@ function createAnthropicMessagesTransportStreamFn() {
 						const usage = message?.usage ?? {};
 						output.responseId = typeof message?.id === "string" ? message.id : void 0;
 						output.responseModel = typeof message?.model === "string" ? message.model : void 0;
-						const promptUsage = readAnthropicPromptUsageSnapshot(usage);
-						const messageStartPromptTokens = promptUsage ? promptUsage.input + promptUsage.cacheRead + promptUsage.cacheWrite : 0;
-						messageStartPromptUsage = messageStartPromptTokens > 0 ? promptUsage : void 0;
-						const inputTokens = readAnthropicUsageTokenCount(usage.input_tokens);
-						if (inputTokens !== void 0) output.usage.input = inputTokens;
-						const outputTokens = readAnthropicUsageTokenCount(usage.output_tokens);
-						if (outputTokens !== void 0) output.usage.output = outputTokens;
-						const cacheReadTokens = usage.cache_read_input_tokens == null ? 0 : readAnthropicUsageTokenCount(usage.cache_read_input_tokens);
-						if (cacheReadTokens !== void 0) output.usage.cacheRead = cacheReadTokens;
-						const cacheWriteTokens = usage.cache_creation_input_tokens == null ? 0 : readAnthropicUsageTokenCount(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;
-						if (messageStartPromptUsage && outputTokens !== void 0) output.usage.contextUsage = {
-							state: "available",
-							promptTokens: messageStartPromptTokens,
-							totalTokens: messageStartPromptTokens + output.usage.output
-						};
+						messageStartPromptUsage = applyAnthropicMessageStartUsage(output.usage, usage);
 						calculateCost(costModel, output.usage);
 						eventSink.push({
 							type: "start",
@@ -1447,31 +1072,8 @@ function createAnthropicMessagesTransportStreamFn() {
 						const delta = event.delta;
 						const usage = event.usage;
 						if (delta?.stop_reason) if (delta.stop_reason === "refusal") applyAnthropicRefusal(output, delta.stop_details, model.provider);
-						else output.stopReason = mapStopReason(delta.stop_reason);
-						const inputTokens = readAnthropicUsageTokenCount(usage?.input_tokens);
-						if (inputTokens !== void 0) output.usage.input = inputTokens;
-						const outputTokens = readAnthropicUsageTokenCount(usage?.output_tokens);
-						if (outputTokens !== void 0) output.usage.output = outputTokens;
-						const cacheReadTokens = readAnthropicUsageTokenCount(usage?.cache_read_input_tokens);
-						if (cacheReadTokens !== void 0) output.usage.cacheRead = cacheReadTokens;
-						const cacheWriteTokens = readAnthropicUsageTokenCount(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(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 = {
-								state: "available",
-								promptTokens,
-								totalTokens: promptTokens + output.usage.output
-							};
-						} else output.usage.contextUsage = { state: "unavailable" };
+						else output.stopReason = mapAnthropicStopReason(delta.stop_reason);
+						applyAnthropicMessageDeltaUsage(output.usage, usage, messageStartPromptUsage);
 						calculateCost(costModel, output.usage);
 						if (output.stopReason === "toolUse" || output.content.some((block) => block.type === "toolCall")) tagPendingCommentaryText(output.content);
 						flushPendingTextEnds();
@@ -2098,7 +1700,11 @@ function createOpenAICompletionsTransportStreamFn() {
 					stream,
 					output,
 					signal: options?.signal,
-					error
+					error,
+					cleanup: () => {
+						output.stopReason = options?.signal?.aborted ? "aborted" : "error";
+						finalizeOpenAICompletionsToolCalls(output, { allowSilentToolCallPromotion: false });
+					}
 				});
 			} finally {
 				firstEventAbort?.dispose();
@@ -2124,6 +1730,7 @@ async function processOpenAICompletionsStream(responseStream, output, model, str
 	const provisionalCommentaryTags = /* @__PURE__ */ new Map();
 	const toolCallBlockBytes = /* @__PURE__ */ new WeakMap();
 	const toolCallBlockIndices = /* @__PURE__ */ new WeakMap();
+	const normalizeToolCallDeltas = createOpenAICompletionsToolCallDeltaNormalizer();
 	let sawStopFinishReason = false;
 	let sawNativeToolCallDelta = false;
 	const blockIndex = () => output.content.length - 1;
@@ -2296,10 +1903,7 @@ async function processOpenAICompletionsStream(responseStream, output, model, str
 		const latestBlock = output.content[output.content.length - 1];
 		if (currentBlock?.type === "text" || currentBlock?.type === "toolCall") return;
 		if (latestBlock?.type === "text" || latestBlock?.type === "toolCall") return;
-		appendThinkingDelta({
-			signature: "",
-			text: ""
-		});
+		appendThinkingDelta({ text: "" });
 	};
 	const flushReasoningTagTextPartitionerAtEnd = () => {
 		for (const delta of reasoningTagTextPartitioner.flush()) appendPartitionedVisibleDelta(delta);
@@ -2327,7 +1931,7 @@ async function processOpenAICompletionsStream(responseStream, output, model, str
 		output.responseId ||= chunk.id;
 		let hasReasoningUsageActivity = false;
 		if (chunk.usage) {
-			output.usage = parseTransportChunkUsage(chunk.usage, model);
+			output.usage = parseOpenAICompletionsUsage(chunk.usage, model);
 			hasReasoningUsageActivity = hasOpenAICompletionsReasoningUsageActivity(chunk.usage);
 		}
 		const choice = Array.isArray(chunk.choices) ? chunk.choices[0] : void 0;
@@ -2338,7 +1942,7 @@ async function processOpenAICompletionsStream(responseStream, output, model, str
 		}
 		const choiceUsage = choice.usage;
 		if (!chunk.usage && choiceUsage) {
-			output.usage = parseTransportChunkUsage(choiceUsage, model);
+			output.usage = parseOpenAICompletionsUsage(choiceUsage, model);
 			hasReasoningUsageActivity = hasOpenAICompletionsReasoningUsageActivity(choiceUsage);
 		}
 		if (choice.finish_reason) {
@@ -2347,91 +1951,92 @@ async function processOpenAICompletionsStream(responseStream, output, model, str
 			if (finishReasonResult.stopReason === "stop") sawStopFinishReason = true;
 			if (finishReasonResult.errorMessage) output.errorMessage = finishReasonResult.errorMessage;
 		}
-		const choiceDelta = choice.delta ?? choice.message;
-		if (!choiceDelta) {
+		const rawChoiceDelta = choice.delta ?? choice.message;
+		if (!rawChoiceDelta) {
 			emitReasoningUsageActivity(hasReasoningUsageActivity);
 			await cooperativeScheduler.afterEvent();
 			continue;
 		}
-		const reasoningDeltas = getCompletionsReasoningDeltas(choiceDelta, compat.visibleReasoningDetailTypes);
-		const hasMirroredReasoning = reasoningDeltas.some((delta) => delta.kind === "thinking");
-		if (hasMirroredReasoning) reasoningTagTextPartitioner.markStrict();
-		if (choiceDelta.content) {
-			const contentDeltas = getCompletionsContentDeltas(choiceDelta.content);
+		for (const normalizedDelta of normalizeToolCallDeltas(rawChoiceDelta, choice.finish_reason)) {
+			const choiceDelta = normalizedDelta.delta;
+			const reasoningDeltas = getCompletionsReasoningDeltas(choiceDelta, compat.visibleReasoningDetailTypes);
+			const hasMirroredReasoning = reasoningDeltas.some((delta) => delta.kind === "thinking");
+			if (hasMirroredReasoning) reasoningTagTextPartitioner.markStrict();
+			const contentDeltas = readOpenAICompletionsContentDeltas(choiceDelta.content, choiceDelta.refusal, reasoningDeltas.filter((reasoningDelta) => reasoningDelta.kind === "thinking").map((reasoningDelta) => reasoningDelta.text));
+			const appendReasoningDeltas = () => {
+				for (const reasoningDelta of reasoningDeltas) {
+					if (reasoningDelta.kind === "thinking" && !emitReasoning) continue;
+					if (currentBlock?.type === "toolCall") {
+						queuePostToolCallDelta({ ...reasoningDelta });
+						continue;
+					}
+					if (reasoningDelta.kind === "text") appendTextDelta(reasoningDelta.text);
+					else if (emitReasoning) appendThinkingDelta(reasoningDelta);
+				}
+			};
+			if (hasMirroredReasoning) appendReasoningDeltas();
 			for (const contentDelta of contentDeltas) if (contentDelta.kind === "text") {
 				const routedDeltas = hasMirroredReasoning ? reasoningTagTextPartitioner.push(contentDelta.text) : reasoningTagTextPartitioner.pushVisible(contentDelta.text);
 				for (const routedDelta of routedDeltas) appendPartitionedVisibleDelta(routedDelta);
 			} else {
-				reasoningTagTextPartitioner.markStrict();
+				if (reasoningTagTextPartitioner.hasPending()) reasoningTagTextPartitioner.markStrict();
 				appendRoutedContentDelta(contentDelta);
 			}
-		}
-		const refusalText = typeof choiceDelta.refusal === "string" ? choiceDelta.refusal : "";
-		if (refusalText) {
-			const routedDeltas = hasMirroredReasoning ? reasoningTagTextPartitioner.push(refusalText) : reasoningTagTextPartitioner.pushVisible(refusalText);
-			for (const routedDelta of routedDeltas) appendPartitionedVisibleDelta(routedDelta);
-		}
-		for (const reasoningDelta of reasoningDeltas) {
-			if (reasoningDelta.kind === "thinking" && !emitReasoning) continue;
-			if (currentBlock?.type === "toolCall") {
-				queuePostToolCallDelta({ ...reasoningDelta });
-				continue;
-			}
-			if (reasoningDelta.kind === "text") appendTextDelta(reasoningDelta.text);
-			else if (emitReasoning) appendThinkingDelta(reasoningDelta);
-		}
-		if (choiceDelta.tool_calls && choiceDelta.tool_calls.length > 0) {
-			sawNativeToolCallDelta = true;
-			flushReasoningTagTextPartitionerAtEnd();
-			rememberPendingCommentaryTags(provisionalCommentaryTags, tagPendingCommentaryText(output.content));
-			for (const toolCall of choiceDelta.tool_calls) {
-				const streamIndex = typeof toolCall.index === "number" ? toolCall.index : void 0;
-				let block = streamIndex !== void 0 ? toolCallBlocksByIndex.get(streamIndex) : void 0;
-				if (!block && toolCall.id) block = toolCallBlocksById.get(toolCall.id);
-				if (!block) {
-					if (currentBlock?.type === "toolCall") {
-						currentBlock = null;
-						flushPendingPostToolCallDeltas();
+			if (!hasMirroredReasoning) appendReasoningDeltas();
+			const toolCallDeltas = normalizedDelta.toolCalls;
+			if (toolCallDeltas.length > 0) {
+				sawNativeToolCallDelta = true;
+				flushReasoningTagTextPartitionerAtEnd();
+				rememberPendingCommentaryTags(provisionalCommentaryTags, tagPendingCommentaryText(output.content));
+				for (const toolCall of toolCallDeltas) {
+					const streamIndex = typeof toolCall.index === "number" ? toolCall.index : void 0;
+					let block = streamIndex !== void 0 ? toolCallBlocksByIndex.get(streamIndex) : void 0;
+					if (!block && toolCall.id) block = toolCallBlocksById.get(toolCall.id);
+					if (!block) {
+						if (currentBlock?.type === "toolCall") {
+							currentBlock = null;
+							flushPendingPostToolCallDeltas();
+						}
+						const initialSig = extractGoogleThoughtSignature(toolCall);
+						block = {
+							type: "toolCall",
+							id: toolCall.id || "",
+							name: toolCall.function?.name || "",
+							arguments: {},
+							partialArgs: "",
+							...initialSig ? { thoughtSignature: initialSig } : {}
+						};
+						output.content.push(block);
+						toolCallBlockIndices.set(block, output.content.length - 1);
+						pushStreamEvent({
+							type: "toolcall_start",
+							contentIndex: toolCallBlockIndices.get(block) ?? -1,
+							partial: output
+						});
+					}
+					if (streamIndex !== void 0 && !toolCallBlocksByIndex.has(streamIndex)) toolCallBlocksByIndex.set(streamIndex, block);
+					if (toolCall.id) {
+						block.id = toolCall.id;
+						toolCallBlocksById.set(toolCall.id, block);
+					}
+					currentBlock = block;
+					if (toolCall.function?.name) block.name = toolCall.function.name;
+					const deltaSig = extractGoogleThoughtSignature(toolCall);
+					if (deltaSig) block.thoughtSignature = deltaSig;
+					if (toolCall.function?.arguments) {
+						const nextArgumentBytes = measureUtf8Bytes(toolCall.function.arguments);
+						const currentBlockArgBytes = toolCallBlockBytes.get(block) ?? 0;
+						if (currentBlockArgBytes + nextArgumentBytes > MAX_TOOL_CALL_ARGUMENT_BUFFER_BYTES) throw new Error("Exceeded tool-call argument buffer limit");
+						toolCallBlockBytes.set(block, currentBlockArgBytes + nextArgumentBytes);
+						block.partialArgs += toolCall.function.arguments;
+						block.arguments = parseStreamingJson(block.partialArgs);
+						pushStreamEvent({
+							type: "toolcall_delta",
+							contentIndex: toolCallBlockIndices.get(block) ?? -1,
+							delta: toolCall.function.arguments,
+							partial: output
+						});
 					}
-					const initialSig = extractGoogleThoughtSignature(toolCall);
-					block = {
-						type: "toolCall",
-						id: toolCall.id || "",
-						name: toolCall.function?.name || "",
-						arguments: {},
-						partialArgs: "",
-						...initialSig ? { thoughtSignature: initialSig } : {}
-					};
-					output.content.push(block);
-					toolCallBlockIndices.set(block, output.content.length - 1);
-					pushStreamEvent({
-						type: "toolcall_start",
-						contentIndex: toolCallBlockIndices.get(block) ?? -1,
-						partial: output
-					});
-				}
-				if (streamIndex !== void 0 && !toolCallBlocksByIndex.has(streamIndex)) toolCallBlocksByIndex.set(streamIndex, block);
-				if (toolCall.id) {
-					block.id = toolCall.id;
-					toolCallBlocksById.set(toolCall.id, block);
-				}
-				currentBlock = block;
-				if (toolCall.function?.name) block.name = toolCall.function.name;
-				const deltaSig = extractGoogleThoughtSignature(toolCall);
-				if (deltaSig) block.thoughtSignature = deltaSig;
-				if (toolCall.function?.arguments) {
-					const nextArgumentBytes = measureUtf8Bytes(toolCall.function.arguments);
-					const currentBlockArgBytes = toolCallBlockBytes.get(block) ?? 0;
-					if (currentBlockArgBytes + nextArgumentBytes > MAX_TOOL_CALL_ARGUMENT_BUFFER_BYTES) throw new Error("Exceeded tool-call argument buffer limit");
-					toolCallBlockBytes.set(block, currentBlockArgBytes + nextArgumentBytes);
-					block.partialArgs += toolCall.function.arguments;
-					block.arguments = parseStreamingJson(block.partialArgs);
-					pushStreamEvent({
-						type: "toolcall_delta",
-						contentIndex: toolCallBlockIndices.get(block) ?? -1,
-						delta: toolCall.function.arguments,
-						partial: output
-					});
 				}
 			}
 		}
@@ -2444,11 +2049,17 @@ async function processOpenAICompletionsStream(responseStream, output, model, str
 	flushDeepSeekTextFilterAtEnd();
 	currentBlock = null;
 	flushPendingPostToolCallDeltas();
-	const hasToolCalls = output.content.some((block) => block.type === "toolCall");
-	const hasVisibleText = output.content.some((block) => block.type === "text" && typeof block.text === "string" && block.text.trim().length > 0);
-	if (output.stopReason === "toolUse" && !hasToolCalls) output.stopReason = "stop";
-	if (output.stopReason === "stop" && hasToolCalls && !hasVisibleText && (sawStopFinishReason || sawNativeToolCallDelta && (options?.sawStreamDONE?.() ?? false))) output.stopReason = "toolUse";
-	if (hasToolCalls && output.stopReason !== "toolUse") output.content = output.content.filter((block) => block.type !== "toolCall");
+	finalizeOpenAICompletionsToolCalls(output, {
+		allowSilentToolCallPromotion: sawStopFinishReason || sawNativeToolCallDelta && (options?.sawStreamDONE?.() ?? false),
+		onConfirmedToolCall(block, contentIndex) {
+			pushStreamEvent({
+				type: "toolcall_end",
+				contentIndex,
+				toolCall: block,
+				partial: output
+			});
+		}
+	});
 	if (output.stopReason !== "toolUse") clearPendingCommentaryText(provisionalCommentaryTags);
 	if (output.stopReason === "toolUse") tagPendingCommentaryText(output.content);
 }
@@ -2463,42 +2074,80 @@ const DEEPSEEK_DSML_TOOL_KINDS = [
 ];
 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_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 MAX_DSML_RECOVERY_BUFFER_BYTES = 256e3;
+const DEEPSEEK_DSML_SCAN_BATCH_CHARS = 64 * 1024;
 function createDeepSeekDsmlToolCallRecoverer() {
 	let buffer = "";
+	let bufferBytes = 0;
+	let bufferEndsWithHighSurrogate = false;
+	let pendingScanChars = 0;
+	let activeOpenToken = null;
+	let blockScanState = {
+		offset: 0,
+		mode: "outer",
+		invokeOpenStart: -1
+	};
+	const resetBlockScan = () => {
+		activeOpenToken = null;
+		pendingScanChars = 0;
+		blockScanState = {
+			offset: 0,
+			mode: "outer",
+			invokeOpenStart: -1
+		};
+	};
 	const consume = (final) => {
 		const output = [];
 		while (buffer) {
-			const open = findEarliestStringToken(buffer, DEEPSEEK_DSML_TOOL_OPEN_TOKENS);
+			const open = activeOpenToken ? {
+				index: 0,
+				token: activeOpenToken
+			} : findEarliestStringToken(buffer, DEEPSEEK_DSML_TOOL_OPEN_TOKENS);
 			if (!open) {
+				resetBlockScan();
 				if (final) {
 					output.push({
 						kind: "text",
 						text: buffer
 					});
 					buffer = "";
+					bufferBytes = 0;
+					bufferEndsWithHighSurrogate = false;
 					return output;
 				}
 				const keep = longestDeepSeekDsmlToolOpenPrefixSuffixLength(buffer);
 				const emitLength = buffer.length - keep;
 				if (emitLength > 0) {
+					const emitted = buffer.slice(0, emitLength);
 					output.push({
 						kind: "text",
-						text: buffer.slice(0, emitLength)
+						text: emitted
 					});
-					buffer = buffer.slice(emitLength);
+					bufferBytes -= Buffer.byteLength(emitted, "utf8");
+					buffer = buffer.slice(emitted.length);
+					if (!buffer) bufferEndsWithHighSurrogate = false;
 				}
 				return output;
 			}
 			if (open.index > 0) {
+				const prefix = buffer.slice(0, open.index);
 				output.push({
 					kind: "text",
-					text: buffer.slice(0, open.index)
+					text: prefix
 				});
-				buffer = buffer.slice(open.index);
+				bufferBytes -= Buffer.byteLength(prefix, "utf8");
+				buffer = buffer.slice(prefix.length);
+				resetBlockScan();
 			}
-			const afterOpen = buffer.slice(open.token.length);
-			const close = findEarliestStringToken(afterOpen, DEEPSEEK_DSML_TOOL_CLOSE_TOKENS);
+			activeOpenToken = open.token;
+			if (blockScanState.offset === 0) blockScanState.offset = open.token.length;
+			const blockScan = scanDeepSeekDsmlToolBlock(buffer, open.token.replace("<", "</"), open.token.length, blockScanState);
+			if (blockScan.kind === "nested-open") throw new Error("Nested DeepSeek DSML recovery wrappers are not supported");
+			const close = blockScan.kind === "close" ? blockScan : null;
 			if (!close) {
 				if (final) {
 					output.push({
@@ -2506,24 +2155,38 @@ function createDeepSeekDsmlToolCallRecoverer() {
 						text: buffer
 					});
 					buffer = "";
+					bufferBytes = 0;
+					bufferEndsWithHighSurrogate = false;
+					return output;
 				}
+				if (bufferBytes > MAX_DSML_RECOVERY_BUFFER_BYTES) throw new Error("Exceeded DeepSeek DSML recovery buffer limit");
 				return output;
 			}
-			const body = afterOpen.slice(0, close.index);
-			const blockLength = open.token.length + close.index + close.token.length;
+			resetBlockScan();
+			const body = buffer.slice(open.token.length, close.index);
+			const blockText = buffer.slice(0, close.index + close.token.length);
+			if (Buffer.byteLength(blockText, "utf8") > MAX_DSML_RECOVERY_BUFFER_BYTES) throw new Error("Exceeded DeepSeek DSML recovery buffer limit");
 			const recoveredToolCalls = parseDeepSeekDsmlToolCallBlock(body);
 			if (recoveredToolCalls.length > 0) output.push(...recoveredToolCalls);
 			else output.push({
 				kind: "text",
-				text: buffer.slice(0, blockLength)
+				text: blockText
 			});
-			buffer = buffer.slice(blockLength);
+			bufferBytes -= Buffer.byteLength(blockText, "utf8");
+			buffer = buffer.slice(blockText.length);
+			if (!buffer) bufferEndsWithHighSurrogate = false;
 		}
 		return output;
 	};
 	return {
 		push(chunk) {
+			const append = utf8ByteLengthForAppend(bufferEndsWithHighSurrogate, chunk);
+			bufferBytes += append.bytes;
+			bufferEndsWithHighSurrogate = append.endsWithHighSurrogate;
 			buffer += chunk;
+			pendingScanChars += chunk.length;
+			if (activeOpenToken && pendingScanChars < DEEPSEEK_DSML_SCAN_BATCH_CHARS && !chunk.includes("<") && !chunk.includes(">") && bufferBytes <= MAX_DSML_RECOVERY_BUFFER_BYTES) return [];
+			pendingScanChars = 0;
 			return consume(false);
 		},
 		flush() {
@@ -2533,16 +2196,16 @@ function createDeepSeekDsmlToolCallRecoverer() {
 }
 function parseDeepSeekDsmlToolCallBlock(body) {
 	const toolCalls = [];
-	const invokeOpenRegex = /<[|｜]DSML[|｜]invoke\b([^>]*)>/g;
+	const invokeOpenRegex = /<[|｜]DSML[|｜]invoke\b([^<>]*)>/g;
 	let openMatch;
 	while ((openMatch = invokeOpenRegex.exec(body)) !== null) {
-		const invokeName = parseXmlAttribute(openMatch[1] ?? "", "name");
-		if (!invokeName) continue;
 		const invokeBodyStart = openMatch.index + openMatch[0].length;
 		const invokeClose = findEarliestStringToken(body.slice(invokeBodyStart), ["</|DSML|invoke>", "</｜DSML｜invoke>"]);
-		if (!invokeClose) continue;
+		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");
+		if (!invokeName) continue;
 		const parsedArguments = parseDeepSeekDsmlInvokeArguments(invokeBody);
 		if (!parsedArguments) continue;
 		toolCalls.push({
@@ -2593,10 +2256,10 @@ function parseXmlAttribute(attributes, name) {
 function decodeDeepSeekDsmlText(value) {
 	return value.replaceAll("&quot;", "\"").replaceAll("&apos;", "'").replaceAll("&lt;", "<").replaceAll("&gt;", ">").replaceAll("&amp;", "&");
 }
-function findEarliestStringToken(text, tokens) {
+function findEarliestStringToken(text, tokens, fromIndex = 0) {
 	let best = null;
 	for (const token of tokens) {
-		const index = text.indexOf(token);
+		const index = text.indexOf(token, fromIndex);
 		if (index !== -1 && (!best || index < best.index)) best = {
 			index,
 			token
@@ -2604,6 +2267,89 @@ function findEarliestStringToken(text, tokens) {
 	}
 	return best;
 }
+function scanDeepSeekDsmlToolBlock(text, closeToken, contentStartIndex, state) {
+	while (state.offset < text.length) {
+		if (state.mode === "invoke-open") {
+			const nextOpen = text.indexOf("<", state.offset);
+			const nextClose = text.indexOf(">", state.offset);
+			if (nextClose === -1 && nextOpen === -1) {
+				state.offset = text.length;
+				return { kind: "incomplete" };
+			}
+			if (nextOpen !== -1 && (nextClose === -1 || nextOpen < nextClose)) {
+				state.mode = "outer";
+				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;
+				continue;
+			}
+			state.mode = "invoke-body";
+			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) {
+				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;
+			continue;
+		}
+		const toolOpen = findEarliestStringToken(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 next = [
+			toolOpen ? {
+				kind: "nested-open",
+				...toolOpen
+			} : null,
+			toolCloseIndex === -1 ? null : {
+				kind: "close",
+				index: toolCloseIndex,
+				token: closeToken
+			},
+			invokeOpen ? {
+				kind: "invoke-open",
+				...invokeOpen
+			} : null
+		].filter((candidate) => candidate !== null).toSorted((left, right) => left.index - right.index)[0];
+		if (!next) {
+			state.offset = Math.max(contentStartIndex, text.length - DEEPSEEK_DSML_RECOVERY_MAX_BOUNDARY_LEN + 1);
+			return { kind: "incomplete" };
+		}
+		if (next.kind === "invoke-open") {
+			state.mode = "invoke-open";
+			state.invokeOpenStart = next.index;
+			state.offset = next.index + next.token.length;
+			continue;
+		}
+		return next;
+	}
+	return { kind: "incomplete" };
+}
+function utf8ByteLengthForAppend(bufferEndsWithHighSurrogate, chunk) {
+	let bytes = Buffer.byteLength(chunk, "utf8");
+	if (!chunk) return {
+		bytes,
+		endsWithHighSurrogate: bufferEndsWithHighSurrogate
+	};
+	const nextCodeUnit = chunk.charCodeAt(0);
+	if (bufferEndsWithHighSurrogate && nextCodeUnit >= 56320 && nextCodeUnit <= 57343) bytes -= 2;
+	const finalCodeUnit = chunk.charCodeAt(chunk.length - 1);
+	return {
+		bytes,
+		endsWithHighSurrogate: finalCodeUnit >= 55296 && finalCodeUnit <= 56319
+	};
+}
 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) {
@@ -2612,37 +2358,6 @@ function longestDeepSeekDsmlToolOpenPrefixSuffixLength(text) {
 	}
 	return 0;
 }
-function getCompletionsContentDeltas(content) {
-	if (typeof content === "string") return content ? [{
-		kind: "text",
-		text: content
-	}] : [];
-	if (Array.isArray(content)) return content.flatMap((item) => getCompletionsContentDeltas(item));
-	if (!content || typeof content !== "object") return [];
-	const record = content;
-	const type = typeof record.type === "string" ? record.type.toLowerCase() : "";
-	const extractText = (value) => {
-		if (typeof value === "string") return value;
-		if (Array.isArray(value)) return value.map((item) => extractText(item)).join("");
-		if (value && typeof value === "object") {
-			const nested = value;
-			return extractText(nested.text ?? nested.content ?? nested.thinking);
-		}
-		return "";
-	};
-	const text = extractText(record.text ?? record.content ?? record.thinking);
-	if (!text) return [];
-	if (type.includes("thinking") || type.includes("reasoning")) return [{
-		kind: "thinking",
-		signature: "content",
-		text
-	}];
-	if (type === "text" || type === "output_text" || type.endsWith(".output_text")) return [{
-		kind: "text",
-		text
-	}];
-	return [];
-}
 function getCompletionsReasoningDeltas(delta, visibleReasoningDetailTypes) {
 	const output = [];
 	const pushDelta = (next) => {
@@ -2795,10 +2510,6 @@ function resolveOpenAICompletionsEffectiveContextTokens(model) {
 function isQwenOpenAICompletionsThinkingFormat(format) {
 	return format === "qwen" || format === "qwen-chat-template";
 }
-function isOpenAICompletionsThinkingEnabled(effort) {
-	const normalized = effort.trim().toLowerCase();
-	return normalized !== "off" && normalized !== "none";
-}
 function setQwenChatTemplateThinking(params, enabled) {
 	const existing = params.chat_template_kwargs;
 	params.chat_template_kwargs = existing && typeof existing === "object" && !Array.isArray(existing) ? {
@@ -3097,32 +2808,6 @@ function buildOpenAICompletionsParams(model, context, options) {
 	else if (resolvedCompletionsReasoningEffort && model.reasoning && compat.supportsReasoningEffort && !handledQwenThinkingFormat && !omitChatCompletionsToolReasoningEffort) params.reasoning_effort = resolvedCompletionsReasoningEffort;
 	return params;
 }
-function parseTransportChunkUsage(rawUsage, model) {
-	const cacheWriteTokens = rawUsage.prompt_tokens_details?.cache_write_tokens || 0;
-	const cachedTokens = rawUsage.prompt_tokens_details?.cached_tokens || 0;
-	const promptTokens = rawUsage.prompt_tokens || 0;
-	const input = Math.max(0, promptTokens - cachedTokens - cacheWriteTokens);
-	const outputTokens = rawUsage.completion_tokens || 0;
-	const reasoningTokens = rawUsage.completion_tokens_details?.reasoning_tokens;
-	const usage = {
-		input,
-		output: outputTokens,
-		cacheRead: cachedTokens,
-		cacheWrite: cacheWriteTokens,
-		...typeof reasoningTokens === "number" && Number.isFinite(reasoningTokens) ? { reasoningTokens } : {},
-		totalTokens: input + outputTokens + cachedTokens + cacheWriteTokens,
-		cost: {
-			input: 0,
-			output: 0,
-			cacheRead: 0,
-			cacheWrite: 0,
-			total: 0
-		}
-	};
-	calculateCost(model, usage);
-	applyProviderReportedUsageCost(usage, rawUsage.cost);
-	return usage;
-}
 function hasOpenAICompletionsReasoningUsageActivity(rawUsage) {
 	const reasoningTokens = rawUsage.completion_tokens_details?.reasoning_tokens;
 	return typeof reasoningTokens === "number" && Number.isFinite(reasoningTokens) && reasoningTokens > 0;
@@ -3132,7 +2817,7 @@ const completionsTesting = {
 	createSseDoneDetector,
 	createOpenAICompletionsClient,
 	buildOpenAICompletionsClientConfig,
-	parseTransportChunkUsage,
+	parseTransportChunkUsage: parseOpenAICompletionsUsage,
 	processOpenAICompletionsStream,
 	shouldEmitOpenAICompletionsReasoningForModel
 };
@@ -3731,6 +3416,7 @@ function createResponsesTransportExecutor(config) {
 					enforceCodeModeResponsesToolSurface(params, visibleToolNames);
 					assertCodeModeResponsesToolSurface(params, visibleToolNames);
 				}
+				const observePrompt = createResponsesPromptEgressObserver(responsesOptions, context.systemPrompt);
 				const requestStartedAt = Date.now();
 				firstEventAbort = createFirstStreamEventAbortController(options?.signal);
 				const requestOptions = buildOpenAISdkRequestOptions(model, firstEventAbort.signal, {
@@ -3739,12 +3425,17 @@ function createResponsesTransportExecutor(config) {
 					maxRetries: options?.maxRetries
 				});
 				emitModelTransportDebug(log, `[responses] start provider=${model.provider} api=${model.api} model=${model.id} baseUrl=${formatModelTransportDebugBaseUrl(model.baseUrl)} timeoutMs=${safeDebugValue(requestOptions?.timeout)} apiKey=${apiKey ? "present" : "missing"} ${summarizeResponsesPayload(params)}`);
-				const responseStream = await config.createResponseStream({
+				const { stream: responseStream, response } = await config.createResponseStream({
 					client,
 					request: params,
 					requestOptions,
-					model
+					model,
+					observePrompt
 				});
+				await options?.onResponse?.({
+					status: response.status,
+					headers: headersToRecord(response.headers)
+				}, model);
 				emitModelTransportDebug(log, `[responses] headers provider=${model.provider} api=${model.api} model=${model.id} elapsedMs=${Date.now() - requestStartedAt}`);
 				stream.push({
 					type: "start",
@@ -3804,7 +3495,17 @@ function createAzureOpenAIResponsesTransportStreamFn() {
 		firstEventTimeoutMs: AZURE_RESPONSES_FIRST_EVENT_TIMEOUT_MS,
 		createClient: createAzureOpenAIClient,
 		buildRequest: (model, context, options, metadata) => buildAzureOpenAIResponsesParams(model, context, options, resolveAzureDeploymentName(model), metadata),
-		createResponseStream: async ({ client, request, requestOptions }) => await client.responses.create(request, requestOptions)
+		createResponseStream: async ({ client, request, requestOptions, observePrompt }) => {
+			observePrompt?.(request, {
+				egress: "responses-sdk",
+				payloadVariant: "initial"
+			});
+			const { data, response } = await client.responses.create(request, requestOptions).withResponse();
+			return {
+				stream: data,
+				response
+			};
+		}
 	});
 }
 function normalizeAzureBaseUrl(baseUrl) {
@@ -3971,6 +3672,7 @@ function buildTransportAwareSimpleStreamFn(model, ctx) {
 //#endregion
 //#region packages/ai/src/transports/simple-completion-transport.ts
 const PROVIDER_SIMPLE_COMPLETION_API_PREFIX = "openclaw-provider-simple:";
+const PROVIDER_STREAM_API_PREFIX = "openclaw-provider-stream:";
 const INVALID_CODEX_BASE_URL_MESSAGE = "OpenAI Codex Responses baseUrl must not include query parameters or fragments";
 function registerCustomApi(registry, api, streamFn) {
 	getAiTransportHost().registerCustomApi(registry, api, streamFn);
@@ -4023,6 +3725,15 @@ function resolveProviderSimpleCompletionApi(model) {
 	];
 	return `${PROVIDER_SIMPLE_COMPLETION_API_PREFIX}${parts.map((part) => encodeURIComponent(part)).join(":")}`;
 }
+function resolveProviderStreamApi(model) {
+	const parts = [
+		model.provider,
+		model.id,
+		model.api,
+		model.baseUrl || "default"
+	];
+	return `${PROVIDER_STREAM_API_PREFIX}${parts.map((part) => encodeURIComponent(part)).join(":")}`;
+}
 function applyProviderSimpleCompletionWrapper(registry, model, cfg) {
 	if (model.api.startsWith(PROVIDER_SIMPLE_COMPLETION_API_PREFIX)) return model;
 	const sourceProvider = registry.getApiProvider(model.api);
@@ -4069,7 +3780,8 @@ function wrapPluginProviderStream(streamFn) {
 		});
 	};
 }
-function registerProviderStreamForModel(params) {
+function prepareProviderStreamModel(params) {
+	if (params.model.api === "google-generative-ai") return;
 	const pluginModel = resolveModelHeaderSentinels(params.model);
 	const providerStreamFn = getAiTransportHost().plugin.resolveProviderStream({
 		provider: params.model.provider,
@@ -4083,15 +3795,19 @@ function registerProviderStreamForModel(params) {
 	});
 	const transportFallback = providerStreamFn ? void 0 : createTransportAwareStreamFnForModel(params.model.api === "google-generative-ai" ? pluginModel : params.model, { cfg: params.cfg });
 	const streamFn = providerStreamFn ? wrapPluginProviderStream(providerStreamFn) : transportFallback && params.model.api === "google-generative-ai" ? wrapPluginProviderStream(transportFallback) : transportFallback;
-	return streamFn && registerCustomApi(params.apiRegistry, params.model.api, streamFn) ? streamFn : void 0;
+	if (!streamFn) return;
+	const api = params.apiRegistry.getApiProvider(params.model.api) ? resolveProviderStreamApi(params.model) : params.model.api;
+	if (!registerCustomApi(params.apiRegistry, api, streamFn)) return;
+	return api === params.model.api ? params.model : projectModel(params.model, { api });
 }
 function prepareModelForSimpleCompletion(params) {
 	const { apiRegistry, model, cfg } = params;
-	if (!apiRegistry.getApiProvider(model.api) && registerProviderStreamForModel({
+	const providerStreamModel = prepareProviderStreamModel({
 		model,
 		cfg,
 		apiRegistry
-	})) return applyProviderSimpleCompletionWrapper(apiRegistry, model, cfg);
+	});
+	if (providerStreamModel) return applyProviderSimpleCompletionWrapper(apiRegistry, providerStreamModel, cfg);
 	const codexTransportModel = prepareCodexSimpleTransportModel(apiRegistry, model, cfg);
 	if (codexTransportModel) return applyProviderSimpleCompletionWrapper(apiRegistry, codexTransportModel, cfg);
 	const transportAwareModel = prepareTransportAwareSimpleModel(model, { cfg });
@@ -4107,4 +3823,4 @@ function prepareModelForSimpleCompletion(params) {
 	return applyProviderSimpleCompletionWrapper(apiRegistry, model, cfg);
 }
 //#endregion
-export { GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, applyAnthropicEphemeralCacheControlMarkers, applyAnthropicPayloadPolicyToParams, applyOpenAIResponsesPayloadPolicy, assertCodeModeResponsesToolSurface, assignTransportErrorDetails, buildOpenAIClientHeaders, buildOpenAICompletionsParams, buildOpenAISdkClientOptions, buildOpenAISdkRequestOptions, buildTransportAwareSimpleStreamFn, canonicalizeMaxTokensParam, coerceTransportToolCallArguments, createAnthropicMessagesTransportStreamFn, createAzureOpenAIResponsesTransportStreamFn, createBoundaryAwareStreamFnForModel, createDeepSeekTextFilter, createEmptyTransportUsage, createModelStreamCooperativeScheduler, createOpenAICompletionsTransportStreamFn, createOpenAIResponsesTransportStreamFn, createOpenClawTransportStreamFnForModel, createTransportAwareStreamFnForModel, createWritableTransportEventStream, detectOpenAICompletionsCompat, emitModelTransportDebug, enforceCodeModeResponsesToolSurface, failTransportStream, filterCodeModePayloadTools, finalizeTransportStream, flattenCompletionMessagesToStringContent, formatModelTransportDebugBaseUrl, formatModelTransportDebugUrl, getCompat, hasOpenAICompatibleConversationTurn, isCodeModeModelVisibleToolName, isOpenAICodexResponsesModel, log, mergeTransportHeaders, mergeTransportMetadata, normalizeCodexResponsesBaseUrlForOpenAISdk, parseJsonObjectPreservingUnsafeIntegers, parseJsonPreservingUnsafeIntegers, prepareModelForSimpleCompletion, prepareTransportAwareSimpleModel, quoteUnsafeIntegerLiterals, readCodeModePayloadToolName, resolveAnthropicEphemeralCacheControl, resolveAnthropicMessagesUrl, resolveAnthropicPayloadPolicy, resolveCodeModeResponsesVisibleToolNames, resolveMaxTokensParam, resolveModelPayloadDebugMode, resolveModelSseDebugMode, resolveOpenAICompletionsCompat, resolveOpenAIReasoningEffortMap, resolveOpenAIResponsesPayloadPolicy, resolveOpenAIStrictToolFlagWithDiagnostics, resolvePromptCacheKey, resolveReplayableResponsesMessageId, resolveTransportAwareSimpleApi, sanitizeNonEmptyTransportPayloadText, sanitizeResponsesImagePayload, sanitizeTransportPayloadText, sortPromptCacheToolsByName as sortTransportToolsByName, stripCompletionMessagesToRoleContent, throwIfModelStreamAborted, transportAbortError, usesNativeOpenAICodexResponsesBackend };
+export { GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, applyAnthropicCacheControlToMessages, applyAnthropicEphemeralCacheControlMarkers, applyAnthropicPayloadPolicyToParams, applyOpenAIResponsesPayloadPolicy, assertCodeModeResponsesToolSurface, assignTransportErrorDetails, buildOpenAIClientHeaders, buildOpenAICompletionsParams, buildOpenAISdkClientOptions, buildOpenAISdkRequestOptions, buildTransportAwareSimpleStreamFn, canonicalizeMaxTokensParam, coerceTransportToolCallArguments, createAnthropicMessagesTransportStreamFn, createAzureOpenAIResponsesTransportStreamFn, createBoundaryAwareStreamFnForModel, createDeepSeekTextFilter, createEmptyTransportUsage, createModelStreamCooperativeScheduler, createOpenAICompletionsTransportStreamFn, createOpenAIResponsesTransportStreamFn, createOpenClawTransportStreamFnForModel, createTransportAwareStreamFnForModel, createWritableTransportEventStream, detectOpenAICompletionsCompat, emitModelTransportDebug, enforceCodeModeResponsesToolSurface, failTransportStream, filterCodeModePayloadTools, finalizeTransportStream, flattenCompletionMessagesToStringContent, formatModelTransportDebugBaseUrl, formatModelTransportDebugUrl, getCompat, hasOpenAICompatibleConversationTurn, isCodeModeModelVisibleToolName, isOpenAICodexResponsesModel, isOpenAICompletionsThinkingEnabled, log, mergeTransportHeaders, mergeTransportMetadata, normalizeCodexResponsesBaseUrlForOpenAISdk, parseJsonObjectPreservingUnsafeIntegers, parseJsonPreservingUnsafeIntegers, parseOpenAICompletionsUsage, prepareModelForSimpleCompletion, prepareTransportAwareSimpleModel, quoteUnsafeIntegerLiterals, readCodeModePayloadToolName, readOpenAICompletionsContentDeltas, resolveAnthropicEphemeralCacheControl, resolveAnthropicMessagesUrl, resolveAnthropicPayloadPolicy, resolveCodeModeResponsesVisibleToolNames, resolveMaxTokensParam, resolveModelPayloadDebugMode, resolveModelSseDebugMode, resolveOpenAICompletionsCompat, resolveOpenAIReasoningEffortMap, resolveOpenAIResponsesPayloadPolicy, resolveOpenAIStrictToolFlagWithDiagnostics, resolvePromptCacheKey, resolveReplayableResponsesMessageId, resolveTransportAwareSimpleApi, sanitizeNonEmptyTransportPayloadText, sanitizeResponsesImagePayload, sanitizeTransportPayloadText, sortPromptCacheToolsByName as sortTransportToolsByName, stripCompletionMessagesToRoleContent, throwIfModelStreamAborted, transportAbortError, usesNativeOpenAICodexResponsesBackend };