@openclaw/ai 2026.9.33 → 2026.9.35

package/dist/{openai-chatgpt-responses-L0Jc-0DF.mjs → openai-chatgpt-responses-Crt3yb4D.mjs}

@@ -3,22 +3,21 @@ import { t as AssistantMessageEventStream } from "./event-stream-vd9UnCAU.mjs";
 import { o as isRecord } from "./record-coerce-CNDovvft.mjs";
 import { S as headersToRecord, n as getAiTransportHost, r as resolveAiTransportHeaderSentinels } from "./host-DqhMLv8J.mjs";
 import { d as supportsOpenAITemperature } from "./openai-reasoning-effort-C_0p3psG.mjs";
-import { n as clampOpenAIPromptCacheKey } from "./openai-prompt-cache-DHxMpeRz.mjs";
+import { n as clampOpenAIPromptCacheKey } from "./openai-prompt-cache-BcwCks0h.mjs";
 import { _ as extractErrorCodeOrErrno, o as WEBSOCKET_NON_RETRYABLE_CLOSE_ERROR_CODE, v as toErrorObject } from "./diagnostics-CJaB6g8J.mjs";
 import { s as resolveTimerTimeoutMs } from "./base64-C1PsYcRQ.mjs";
-import { t as MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE } from "./transport-utils-CnxX8hPj.mjs";
+import { t as MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE } from "./transport-utils-B11GlmDO.mjs";
 import { parseRetryAfterHeadersSeconds } from "./internal/retry-after.mjs";
 import { S as withProviderResponseHook, b as transportAbortError, h as notifyProviderStreamOpened, n as assignTransportErrorDetails } from "./transport-stream-shared-Dt7VrJFd.mjs";
 import { C as stripSystemPromptCacheBoundary, n as buildBaseOptions } from "./simple-options-B9u6-CF3.mjs";
-import { b as createOpenAIResponseHook, g as resolveOpenAISimpleReasoningEffort, h as resolveOpenAIRequestReasoning, x as createResponseModelTracker, y as createOpenAIProviderAcceptanceHook } from "./openai-transport-params-C4eZAXf2.mjs";
+import { b as createOpenAIResponseHook, g as resolveOpenAISimpleReasoningEffort, h as resolveOpenAIRequestReasoning, x as createResponseModelTracker, y as createOpenAIProviderAcceptanceHook } from "./openai-transport-params-DzSQTRdN.mjs";
 import { n as getEnvApiKey } from "./env-api-keys-CyX4VKGO.mjs";
 import { n as createAssistantOutput } from "./provider-transcript-transform-Cg-36VSt.mjs";
 import { a as withFirstStreamEventTimeout, i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-B9o0S4SZ.mjs";
 import { h as responsesPromptObserver } from "./openai-responses-contracts-CLxnqS9p.mjs";
-import { A as recordResponsesContextUsage, P as convertProviderResponsesMessages, S as isInvalidEncryptedContentError, T as ResponsesStreamFailure, Z as suppressOpenAIResponsesCompaction, b as commitResponsesEncryptedContentAttempt, i as convertResponsesToolPayload, n as applyResponsesServiceTierPricing, o as readOpenAIMisalignmentReview, q as buildOpenAIResponsesReasoningReplayMetadata, s as processResponsesStream, w as resolveNextResponsesEncryptedContentAttempt } from "./openai-responses-shared-BpGnmnV8.mjs";
-import { t as createSseByteGuard } from "./streaming-byte-guard-CC-HMn_u.mjs";
+import { A as recordResponsesContextUsage, P as convertProviderResponsesMessages, S as isInvalidEncryptedContentError, T as ResponsesStreamFailure, Z as suppressOpenAIResponsesCompaction, b as commitResponsesEncryptedContentAttempt, i as convertResponsesToolPayload, n as applyResponsesServiceTierPricing, o as readOpenAIMisalignmentReview, q as buildOpenAIResponsesReasoningReplayMetadata, s as processResponsesStream, w as resolveNextResponsesEncryptedContentAttempt } from "./openai-responses-shared-DLz_xgps.mjs";
 import { n as registerSessionResourceCleanup } from "./session-resources-CkR4WWy1.mjs";
-import { i as resolveOpenAIResponsesTextFormat, t as createResponsesPromptEgressObserver } from "./openai-responses-prompt-observer-internal-C5WB2Kxf.mjs";
+import { i as resolveOpenAIResponsesTextFormat, t as createResponsesPromptEgressObserver } from "./openai-responses-prompt-observer-internal-CGhqoBH1.mjs";
 import { n as withResponsesRequestAcceptance, t as responsesRequestLifecycle } from "./openai-responses-request-lifecycle-Fsag-IGZ.mjs";
 import { n as resolveOpenAICodexAccountId } from "./openai-chatgpt-jwt-KWcgd0d_.mjs";
 //#region packages/ai/src/providers/openai-chatgpt-responses-events.ts
@@ -83,7 +82,7 @@ function normalizeCodexStatus(status) {
 }
 //#endregion
 //#region packages/ai/src/providers/openai-chatgpt-responses-protocol.ts
-const OPENAI_CHATGPT_RESPONSES_SUCCESS_BODY_MAX_BYTES = 16777216;
+const OPENAI_CHATGPT_RESPONSES_SSE_FRAME_MAX_BYTES = 16777216;
 var CodexProtocolError = class extends Error {
 	constructor(message, options) {
 		super(message);
@@ -95,44 +94,53 @@ var CodexProtocolError = class extends Error {
 async function* parseOpenAIChatGptResponsesSse(response) {
 	if (!response.body) return;
 	const reader = response.body.getReader();
-	const guard = createSseByteGuard(reader, {
-		maxBytes: OPENAI_CHATGPT_RESPONSES_SUCCESS_BODY_MAX_BYTES,
-		onOverflow: ({ size, maxBytes }) => /* @__PURE__ */ new Error(`OpenAI ChatGPT Responses success body exceeded ${maxBytes} bytes (received ${size})`)
-	});
 	const decoder = new TextDecoder();
-	let buffer = "";
+	let buffer = /* @__PURE__ */ new Uint8Array(1024);
+	let size = 0;
+	let lineEmpty = true;
+	let skipLf = false;
+	let cancelReason;
 	try {
 		while (true) {
-			const { done, value } = await guard.read();
-			if (value) buffer += decoder.decode(value, { stream: true });
-			if (done) buffer += decoder.decode();
-			while (true) {
-				const searchable = !done && buffer.endsWith("\r") && !buffer.endsWith("\r\r") && !buffer.endsWith("\n\r") ? buffer.slice(0, -1) : buffer;
-				const boundary = /(?:\r\n|\r(?!\n)|\n)(?:\r\n|\r(?!\n)|\n)/.exec(searchable);
-				if (!boundary && (!done || buffer.length === 0)) break;
-				const chunk = boundary ? buffer.slice(0, boundary.index) : buffer;
-				buffer = boundary ? buffer.slice(boundary.index + boundary[0].length) : "";
-				const dataLines = chunk.split(/\r\n|\r|\n/).filter((line) => line.startsWith("data:")).map((line) => line.slice(5).trim());
-				if (dataLines.length > 0) {
-					const data = dataLines.join("\n").trim();
-					if (data && data !== "[DONE]") {
-						let event;
-						try {
-							event = JSON.parse(data);
-						} catch (cause) {
-							if (!(cause instanceof SyntaxError)) throw cause;
-							throw new CodexProtocolError(MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE, { cause });
-						}
-						yield event;
+			const { done, value } = await reader.read();
+			for (let index = 0; index < (value?.length ?? 0) + (done ? 1 : 0); index++) {
+				const byte = value?.[index];
+				if (byte !== void 0) {
+					const crlf = skipLf && byte === 10;
+					skipLf = byte === 13;
+					const lineEnding = byte === 10 || byte === 13;
+					const boundary = !crlf && lineEnding && lineEmpty;
+					lineEmpty = lineEnding;
+					if (crlf && size === 0) continue;
+					if (size >= OPENAI_CHATGPT_RESPONSES_SSE_FRAME_MAX_BYTES) throw new Error(`OpenAI ChatGPT Responses SSE frame exceeded ${OPENAI_CHATGPT_RESPONSES_SSE_FRAME_MAX_BYTES} bytes`);
+					if (size === buffer.length) {
+						const grown = new Uint8Array(Math.min(buffer.length * 2, OPENAI_CHATGPT_RESPONSES_SSE_FRAME_MAX_BYTES));
+						grown.set(buffer);
+						buffer = grown;
 					}
+					buffer[size++] = byte;
+					if (!boundary) continue;
+				}
+				const chunk = decoder.decode(buffer.subarray(0, size), { stream: !done });
+				size = 0;
+				const data = chunk.split(/\r\n|\r|\n/).filter((line) => line.startsWith("data:")).map((line) => line.slice(5).trim()).join("\n").trim();
+				if (!data || data === "[DONE]") continue;
+				let event;
+				try {
+					event = JSON.parse(data);
+				} catch (cause) {
+					if (!(cause instanceof SyntaxError)) throw cause;
+					throw new CodexProtocolError(MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE, { cause });
 				}
+				yield event;
 			}
 			if (done) break;
 		}
+	} catch (error) {
+		cancelReason = error;
+		throw error;
 	} finally {
-		try {
-			await guard.cancel();
-		} catch {}
+		reader.cancel(cancelReason).catch(() => void 0);
 		try {
 			reader.releaseLock();
 		} catch {}
@@ -163,6 +171,7 @@ const RETRYABLE_CLOSE_CODES = /* @__PURE__ */ new Set([
 ]);
 const WEBSOCKET_MESSAGE_TOO_BIG_CLOSE_CODE = 1009;
 const WEBSOCKET_CONNECTION_LIMIT_REACHED_CODE = "websocket_connection_limit_reached";
+const WEBSOCKET_REPLAY_REJECTION = "Codex error: Persisted response contains hosted-tool, compaction, or unverifiable hidden reasoning state that Rustponses cannot replay. Start a new response or use the Python Responses service for this continuation.";
 const OPENAI_CHATGPT_RESPONSES_ERROR_BODY_MAX_BYTES = 16384;
 function resolveRequestTimeoutMs(options) {
 	const timeoutMs = options?.timeoutMs;
@@ -241,19 +250,18 @@ const streamOpenAICodexResponses = (model, context, options) => {
 			if (transport !== "sse" && !websocketDisabledForSession) {
 				const websocketHeaders = buildWebSocketHeaders(modelHeaders, optionHeaders, accountId, apiKey, sessionId || createCodexRequestId());
 				let websocketStarted = false;
-				let websocketRequestSent = false;
-				let retriedWebSocketConnectionLimit = false;
+				let retriedWebSocketRejection = false;
 				while (true) {
 					const activeAttempt = semanticAttempt;
+					const dispatch = {};
 					websocketStarted = false;
-					websocketRequestSent = false;
 					try {
 						const terminal = await processWebSocketStream(resolveCodexWebSocketUrl(model.baseUrl), activeAttempt.request, websocketHeaders, output, stream, model, () => {
 							websocketStarted = true;
 						}, () => {
 							commitSemanticAttempt(activeAttempt);
-						}, requestOptions, firstEventAbort.abort, observePromptEgress, activeAttempt.kind, () => {
-							websocketRequestSent = true;
+						}, requestOptions, firstEventAbort.abort, observePromptEgress, activeAttempt.kind, (request) => {
+							dispatch.request = request;
 						}, options?.signal);
 						if (activeSignal?.aborted) throw transportAbortError(activeSignal);
 						if (output.stopReason === "aborted" || output.stopReason === "error") throw new CodexApiError(output.errorMessage ?? "An unknown error occurred");
@@ -267,19 +275,20 @@ const streamOpenAICodexResponses = (model, context, options) => {
 						return;
 					} catch (error) {
 						const aborted = activeSignal?.aborted;
-						const nextSemanticAttempt = !aborted && websocketRequestSent && !websocketStarted && error instanceof CodexApiError && isInvalidEncryptedContentError(error) ? await resolveNextResponsesEncryptedContentAttempt(activeAttempt, error, { buildFullHistoryRequest: () => buildBody("full-history") }) : void 0;
+						const nextSemanticAttempt = !aborted && dispatch.request && !websocketStarted && error instanceof CodexApiError && isInvalidEncryptedContentError(error) ? await resolveNextResponsesEncryptedContentAttempt(activeAttempt, error, { buildFullHistoryRequest: () => buildBody("full-history") }) : void 0;
 						if (nextSemanticAttempt) {
 							semanticAttempt = nextSemanticAttempt;
-							retriedWebSocketConnectionLimit = false;
+							retriedWebSocketRejection = false;
 							continue;
 						}
-						const connectionLimitBeforeStart = !websocketStarted && isWebSocketConnectionLimitReachedError(error);
-						if (!aborted && connectionLimitBeforeStart && !retriedWebSocketConnectionLimit) {
-							retriedWebSocketConnectionLimit = true;
+						const reconnectBeforeStart = !websocketStarted && isWebSocketConnectionLimitReachedError(error) || !websocketStarted && dispatch.request?.previous_response_id && error instanceof CodexApiError && error.message === WEBSOCKET_REPLAY_REJECTION;
+						if (!aborted && reconnectBeforeStart && !retriedWebSocketRejection) {
+							retriedWebSocketRejection = true;
 							continue;
 						}
-						if (aborted || isCodexNonTransportError(error) && !connectionLimitBeforeStart) throw error;
-						appendAssistantMessageDiagnostic(output, createAssistantMessageDiagnostic("provider_transport_failure", error, {
+						const transportError = isWebSocketConnectionLimitReachedError(error) ? Object.assign(new Error(error.message, { cause: error }), { code: WEBSOCKET_TRANSPORT_ERROR_CODE }) : error;
+						if (aborted || isCodexNonTransportError(transportError)) throw transportError;
+						appendAssistantMessageDiagnostic(output, createAssistantMessageDiagnostic("provider_transport_failure", transportError, {
 							configuredTransport: transport,
 							fallbackTransport: transport === "auto" && !websocketStarted ? "sse" : void 0,
 							eventsEmitted: websocketStarted,
@@ -287,7 +296,7 @@ const streamOpenAICodexResponses = (model, context, options) => {
 							requestBytes: new TextEncoder().encode(JSON.stringify(activeAttempt.request)).byteLength
 						}));
 						if (transport === "auto" && options?.sessionId) websocketSseFallbackSessions.add(options.sessionId);
-						if (websocketStarted || transport !== "auto") throw error;
+						if (websocketStarted || transport !== "auto") throw transportError;
 						break;
 					}
 				}
@@ -926,7 +935,7 @@ async function processWebSocketStream(url, body, headers, output, stream, model,
 			type: "response.create",
 			...requestBody
 		}));
-		onRequestSent?.();
+		onRequestSent?.(requestBody);
 		const terminal = await processResponsesStream(startWebSocketOutputOnFirstEvent(mapCodexEvents(parseWebSocket(socket, options?.signal), void 0, options), output, stream, onFirstProviderEvent, () => notifyProviderStreamOpened({
 			options,
 			cancelStream: () => {

package/dist/{openai-completions-gyTpI3xG.mjs → openai-completions-BwcqZo-Y.mjs}

@@ -1,19 +1,19 @@
 import { t as __exportAll } from "./rolldown-runtime-BhDjJH2R.mjs";
 import { t as AssistantMessageEventStream } from "./event-stream-vd9UnCAU.mjs";
 import { T as clampThinkingLevel } from "./host-DqhMLv8J.mjs";
-import { i as resolveOpenAICompletionsCompat } from "./openai-completions-compat-D1q02qdx.mjs";
+import { i as resolveOpenAICompletionsCompat } from "./openai-completions-compat-BuszNIRF.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
 import { S as withProviderResponseHook, b as transportAbortError, n as assignTransportErrorDetails } from "./transport-stream-shared-Dt7VrJFd.mjs";
 import { n as buildBaseOptions, s as prepareModelRequestBody, u as reasoningTagTextPolicy } from "./simple-options-B9u6-CF3.mjs";
-import { S as isOpenAICompletionsThinkingEnabled, y as createOpenAIProviderAcceptanceHook } from "./openai-transport-params-C4eZAXf2.mjs";
+import { C as isOpenAICompletionsThinkingEnabled, y as createOpenAIProviderAcceptanceHook } from "./openai-transport-params-DzSQTRdN.mjs";
 import { n as getEnvApiKey } from "./env-api-keys-CyX4VKGO.mjs";
 import { a as tagUnresolvedTextAsCommentary, t as clearPendingCommentaryText } from "./assistant-text-phase-C20rxWwP.mjs";
 import { n as createAssistantOutput } from "./provider-transcript-transform-Cg-36VSt.mjs";
-import { a as convertMessages, i as buildOpenAICompletionsRequest, s as finalizeOpenAICompletionsToolCalls, t as processCompletionsStream } from "./openai-completions-stream-BrsoGafc.mjs";
+import { a as convertMessages, i as buildOpenAICompletionsRequest, s as finalizeOpenAICompletionsToolCalls, t as processCompletionsStream } from "./openai-completions-stream-D--YONza.mjs";
 import { i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-B9o0S4SZ.mjs";
 import { n as resolveProviderSimpleCompletionHeaders } from "./provider-transport-turn-state-CUUfdJPN.mjs";
 import { n as hasCopilotVisionInput, t as buildCopilotDynamicHeaders } from "./github-copilot-headers-B37TB3rB.mjs";
-import { t as createOpenAIProviderClient } from "./openai-provider-client-D1gd0dfR.mjs";
+import { t as createOpenAIProviderClient } from "./openai-provider-client-Co1u7I3J.mjs";
 //#region packages/ai/src/providers/openai-completions.ts
 var openai_completions_exports = /* @__PURE__ */ __exportAll({
 	convertMessages: () => convertMessages,

package/dist/{openai-completions-compat-D1q02qdx.mjs → openai-completions-compat-BuszNIRF.mjs}

@@ -1,4 +1,5 @@
 import { o as isRecord } from "./record-coerce-CNDovvft.mjs";
+import { c as resolveOpenAIReasoningEffortMapping, u as supportsOpenAIReasoningEffort } from "./openai-reasoning-effort-C_0p3psG.mjs";
 import { i as resolveProviderRequestCapabilities } from "./host-policy-B3DuH3Le.mjs";
 //#region packages/ai/src/providers/openai-response-format.ts
 const JSON_SCHEMA_RESPONSE_FORMAT_NAME = "openclaw_response";
@@ -140,14 +141,17 @@ function resolveSessionAffinity(model, detectedFormat) {
 }
 /** Applies explicit model overrides once on top of the canonical transport defaults. */
 function resolveOpenAICompletionsCompat(model, resolveCapabilities) {
-	const { defaults } = detectOpenAICompletionsCompat(model, resolveCapabilities);
+	const { defaults, capabilities } = detectOpenAICompletionsCompat(model, resolveCapabilities);
 	const configured = model.compat;
+	const thinkingFormat = configured?.thinkingFormat ?? defaults.thinkingFormat;
+	const thinkingOffControlUndeclared = capabilities.usesExplicitProxyLikeEndpoint && thinkingFormat === "zai" && configured?.thinkingFormat === void 0 && !supportsOpenAIReasoningEffort({ compat: configured }, "none") && !supportsOpenAIReasoningEffort({ compat: configured }, "off") && resolveOpenAIReasoningEffortMapping("off", configured?.reasoningEffortMap) === void 0;
 	return {
 		supportsStore: configured?.supportsStore ?? defaults.supportsStore,
 		supportsDeveloperRole: configured?.supportsDeveloperRole ?? defaults.supportsDeveloperRole,
 		supportsReasoningEffort: configured?.supportsReasoningEffort ?? defaults.supportsReasoningEffort,
 		supportedReasoningEfforts: configured?.supportedReasoningEfforts,
 		reasoningEffortMap: configured?.reasoningEffortMap,
+		thinkingOffControlUndeclared,
 		supportsUsageInStreaming: configured?.supportsUsageInStreaming ?? defaults.supportsUsageInStreaming,
 		maxTokensField: configured?.maxTokensField ?? defaults.maxTokensField,
 		requiresToolResultName: configured?.requiresToolResultName ?? false,

package/dist/{openai-completions-compat-CqWP1fn5.d.mts → openai-completions-compat-WTcM-PUc.d.mts}

@@ -1,6 +1,6 @@
-import "./index-DHeFfpcz.mjs";
-import { M as OpenAICompletionsCompat, O as Model } from "./types-CLLR3eB4.mjs";
-import { a as AiProviderRequestPolicyInput, i as AiProviderRequestCapabilities } from "./host-COwfRiSL.mjs";
+import "./index-Iw8l620S.mjs";
+import { M as OpenAICompletionsCompat, O as Model } from "./types-B89w6ly2.mjs";
+import { a as AiProviderRequestPolicyInput, i as AiProviderRequestCapabilities } from "./host-BhpDO5Sy.mjs";
 //#region packages/ai/src/transports/openai-completions-compat.d.ts
 type ProviderRequestCapabilities = AiProviderRequestCapabilities;
 type OpenAICompletionsSessionAffinity = "none" | "openai" | "openrouter";
@@ -31,6 +31,7 @@ type ResolvedOpenAICompletionsCompat = Omit<Required<OpenAICompletionsCompat>, "
   visibleReasoningDetailTypes: string[];
   requiresNonEmptyUserOrAssistantMessage: boolean;
   configuredSupportsLongCacheRetention?: boolean;
+  thinkingOffControlUndeclared?: boolean;
 };
 /** Native OpenAI defaults never apply to a configured proxy endpoint. */
 declare function isNativeOpenAIEndpoint(model: {

package/dist/{openai-completions-stream-BrsoGafc.mjs → openai-completions-stream-D--YONza.mjs}

@@ -4,22 +4,22 @@ import { k as isImageWithMediaPayload, n as getAiTransportHost } from "./host-Dq
 import { n as isOpenAIGpt55Model, r as isOpenAIGpt56Model, t as isOpenAIGpt54MiniModel } from "./openai-reasoning-effort-C_0p3psG.mjs";
 import { r as uniqueStrings } from "./string-normalization-J9ZiLfGO.mjs";
 import { o as truncateUtf16Safe } from "./provider-error-CwT_j2v0.mjs";
-import { r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-DHxMpeRz.mjs";
+import { r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-BcwCks0h.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-CnxX8hPj.mjs";
-import { c as shouldOmitOllamaCompatResponseFormat, n as isNativeOpenAIEndpoint, s as resolveOpenAICompletionsResponseFormat, t as detectOpenAICompletionsCompat } from "./openai-completions-compat-D1q02qdx.mjs";
+import { l as supportsModelTools, r as estimateStringChars } from "./transport-utils-B11GlmDO.mjs";
+import { c as shouldOmitOllamaCompatResponseFormat, n as isNativeOpenAIEndpoint, s as resolveOpenAICompletionsResponseFormat, t as detectOpenAICompletionsCompat } from "./openai-completions-compat-BuszNIRF.mjs";
 import { r as resolveProviderEndpoint } from "./host-policy-B3DuH3Le.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
 import { E as createToolArgumentPreviewSchedule, O as parseStreamingJson, d as iterateModelStream, j as sanitizeSurrogates, l as finalizeTerminalToolCallArguments, y as throwIfModelStreamAborted } from "./transport-stream-shared-Dt7VrJFd.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-B9u6-CF3.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-C4eZAXf2.mjs";
+import { A as resolvePromptCacheKey, C as isOpenAICompletionsThinkingEnabled, D as readOpenAICompletionsContentDeltas, E as parseOpenAICompletionsUsage, O as readOpenAICompletionsReasoningBatch, S as detachPartialSnapshot, T as measureUtf8AppendBytes, d as prepareOpenAITools, f as projectOpenAITools, g as resolveOpenAISimpleReasoningEffort, h as resolveOpenAIRequestReasoning, p as reconcileOpenAICompletionsToolChoice, s as getCompat, u as resolveOpenAIStrictToolFlagWithDiagnostics, v as GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, w as log } from "./openai-transport-params-DzSQTRdN.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 { t as transformProviderMessages } from "./provider-transcript-transform-Cg-36VSt.mjs";
+import { isGoogleGemini3ThinkingLevelModel } from "./internal/google-model-family.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 { f as withPreparedToolSchemaNormalization, r as normalizeOpenAIStrictToolParameters } from "./openai-tool-schema-DSAxYnHW.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";
@@ -651,7 +651,7 @@ function relocateNonBehavioralRegion(args) {
 function applyDirectCompletionsReasoningAndRouting(params, model, reasoning, compat) {
 	const nativeEffort = compat.thinkingFormat === "openrouter" ? void 0 : reasoning.effort;
 	const reasoningEnabled = reasoning.thinkingEnabled ?? false;
-	if (compat.thinkingFormat === "zai" && model.reasoning) params.thinking = reasoningEnabled ? {
+	if (compat.thinkingFormat === "zai" && model.reasoning && compat.thinkingOffControlUndeclared && reasoning.thinkingEnabled === false) {} else if (compat.thinkingFormat === "zai" && model.reasoning) params.thinking = reasoningEnabled ? {
 		type: "enabled",
 		clear_thinking: false
 	} : { type: "disabled" };
@@ -685,7 +685,7 @@ function isGoogleOpenAICompatModel(model) {
 	return model.provider === "google" || endpointClass === "google-generative-ai" || endpointClass === "google-vertex";
 }
 function requiresGoogleCompatToolCallThoughtSignature(model) {
-	return isGoogleGemini3ProModel(model.id) || isGoogleGemini3FlashModel(model.id);
+	return isGoogleGemini3ThinkingLevelModel(model.id);
 }
 const GOOGLE_COMPAT_THOUGHT_SIGNATURE_ELLIPSIS_RE = /[\u2026]|\.\.\./;
 const GOOGLE_COMPAT_THOUGHT_SIGNATURE_BASE64_RE = /^[A-Za-z0-9+/=]+$/;
@@ -767,6 +767,8 @@ const REASONING_CONTENT_REPLAY_MODEL_IDS = /* @__PURE__ */ new Set([
 	"kimi-k2.7-code",
 	"kimi-k2.7-code-highspeed",
 	"kimi-k3",
+	"kimi-k3.1-pro",
+	"kimi-k3.1-pro-highspeed",
 	"kimi-k2-thinking",
 	"kimi-k2-thinking-turbo",
 	"mimo-v2-pro",
@@ -1397,7 +1399,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 	let chunkPushedEvent = false;
 	const pushStreamEvent = (event) => {
 		chunkPushedEvent = true;
-		stream.push(event);
+		stream.push(detachPartialSnapshot(event));
 	};
 	const queuePostToolCallDelta = (next) => {
 		const nextBytes = Buffer.byteLength(next.text, "utf8");
@@ -1570,7 +1572,8 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		if (latestBlock?.type === "text" || latestBlock?.type === "toolCall") return;
 		appendThinkingDelta({ text: "" });
 	};
-	const flushReasoningTagTextPartitioner = () => {
+	const flushReasoningTagTextPartitioner = (allowRecovery = true) => {
+		if (!(allowRecovery && output.stopReason !== "length" && output.stopReason !== "error" && output.stopReason !== "aborted")) reasoningTagTextPartitioner.markStrict();
 		for (const delta of reasoningTagTextPartitioner.flush()) appendPartitionedVisibleDelta(delta);
 	};
 	const sealTextBeforeReasoning = () => {
@@ -1662,7 +1665,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 			const toolCallDeltas = normalizedDelta.toolCalls;
 			if (toolCallDeltas.length > 0) {
 				sawNativeToolCallDelta = true;
-				flushReasoningTagTextPartitioner();
+				flushReasoningTagTextPartitioner(false);
 				rememberPendingCommentaryTags(provisionalCommentaryTags, tagPendingCommentaryText(output.content));
 				for (const toolCall of toolCallDeltas) {
 					const streamIndex = typeof toolCall.index === "number" ? toolCall.index : void 0;

package/dist/{openai-prompt-cache-DHxMpeRz.mjs → openai-prompt-cache-BcwCks0h.mjs}

@@ -1,4 +1,4 @@
-import { n as isNativeOpenAIEndpoint } from "./openai-completions-compat-D1q02qdx.mjs";
+import { n as isNativeOpenAIEndpoint } from "./openai-completions-compat-BuszNIRF.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-prompt-cache-YTmJ_iMS.d.mts → openai-prompt-cache-CuY5QDIc.d.mts}

@@ -1,5 +1,5 @@
-import { M as OpenAICompletionsCompat, O as Model, l as CacheRetention } from "./types-CLLR3eB4.mjs";
-import "./types-CoL-LJ7j.mjs";
+import { M as OpenAICompletionsCompat, O as Model, l as CacheRetention } from "./types-B89w6ly2.mjs";
+import "./types-DbmNRKvK.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-provider-client-D1gd0dfR.mjs → openai-provider-client-Co1u7I3J.mjs}

@@ -1,5 +1,5 @@
 import { n as getAiTransportHost } from "./host-DqhMLv8J.mjs";
-import { O as resolveOpenAIClientBaseUrl } from "./openai-transport-params-C4eZAXf2.mjs";
+import { k as resolveOpenAIClientBaseUrl } from "./openai-transport-params-DzSQTRdN.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-responses-DluFZzn_.mjs → openai-responses-C7B8_pLG.mjs}

@@ -1,16 +1,16 @@
 import { t as __exportAll } from "./rolldown-runtime-BhDjJH2R.mjs";
 import { t as AssistantMessageEventStream } from "./event-stream-vd9UnCAU.mjs";
-import { n as clampOpenAIPromptCacheKey, r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-DHxMpeRz.mjs";
-import { i as resolveOpenAICompletionsCompat } from "./openai-completions-compat-D1q02qdx.mjs";
+import { n as clampOpenAIPromptCacheKey, r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-BcwCks0h.mjs";
+import { i as resolveOpenAICompletionsCompat } from "./openai-completions-compat-BuszNIRF.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
 import { n as buildBaseOptions } from "./simple-options-B9u6-CF3.mjs";
-import { g as resolveOpenAISimpleReasoningEffort } from "./openai-transport-params-C4eZAXf2.mjs";
+import { g as resolveOpenAISimpleReasoningEffort } from "./openai-transport-params-DzSQTRdN.mjs";
 import { n as getEnvApiKey } from "./env-api-keys-CyX4VKGO.mjs";
 import { n as createAssistantOutput } from "./provider-transcript-transform-Cg-36VSt.mjs";
 import { n as resolveProviderSimpleCompletionHeaders } from "./provider-transport-turn-state-CUUfdJPN.mjs";
-import { P as convertProviderResponsesMessages, n as applyResponsesServiceTierPricing, r as runResponsesStreamLifecycle, t as applyCommonResponsesParams } from "./openai-responses-shared-BpGnmnV8.mjs";
+import { P as convertProviderResponsesMessages, n as applyResponsesServiceTierPricing, r as runResponsesStreamLifecycle, t as applyCommonResponsesParams } from "./openai-responses-shared-DLz_xgps.mjs";
 import { n as hasCopilotVisionInput, t as buildCopilotDynamicHeaders } from "./github-copilot-headers-B37TB3rB.mjs";
-import { t as createOpenAIProviderClient } from "./openai-provider-client-D1gd0dfR.mjs";
+import { t as createOpenAIProviderClient } from "./openai-provider-client-Co1u7I3J.mjs";
 //#region packages/ai/src/providers/openai-responses.ts
 var openai_responses_exports = /* @__PURE__ */ __exportAll({
 	streamOpenAIResponses: () => streamOpenAIResponses,

package/dist/{openai-responses-compaction-replay-CQHR2isp.d.mts → openai-responses-compaction-replay-DiGjwK4d.d.mts}

@@ -1,7 +1,7 @@
-import "./index-DHeFfpcz.mjs";
-import { O as Model, a as AssistantMessage } from "./types-CLLR3eB4.mjs";
-import "./provider-options-yDbWMoEr.mjs";
-import { i as ReplayableResponseCompactionItem, r as OpenAIResponsesReasoningReplayMetadata, s as OpenAIResponsesCompactionOutput } from "./openai-responses-contracts-C2H8pffb.mjs";
+import "./index-Iw8l620S.mjs";
+import { O as Model, a as AssistantMessage } from "./types-B89w6ly2.mjs";
+import "./provider-options-CFiaIG4K.mjs";
+import { i as ReplayableResponseCompactionItem, r as OpenAIResponsesReasoningReplayMetadata, s as OpenAIResponsesCompactionOutput } from "./openai-responses-contracts-CYScOS7J.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;

package/dist/{openai-responses-contracts-C2H8pffb.d.mts → openai-responses-contracts-CYScOS7J.d.mts}

@@ -1,7 +1,7 @@
-import "./index-DHeFfpcz.mjs";
-import { O as Model, j as ModelThinkingLevel, n as Api } from "./types-CLLR3eB4.mjs";
-import { a as BaseOpenAIStreamOptions, f as OpenAIReasoningEffort } from "./provider-options-yDbWMoEr.mjs";
-import "./types-CoL-LJ7j.mjs";
+import "./index-Iw8l620S.mjs";
+import { O as Model, j as ModelThinkingLevel, n as Api } from "./types-B89w6ly2.mjs";
+import { a as BaseOpenAIStreamOptions, f as OpenAIReasoningEffort } from "./provider-options-CFiaIG4K.mjs";
+import "./types-DbmNRKvK.mjs";
 import { ResponseCompactionItem, ResponseCreateParamsStreaming, ResponseInputItem } from "openai/resources/responses/responses.js";
 //#region packages/ai/src/providers/openai-request-reasoning.d.ts
 type OpenAIRequestReasoningEffort = ModelThinkingLevel | "none";

package/dist/{openai-responses-prompt-observer-internal-C5WB2Kxf.mjs → openai-responses-prompt-observer-internal-CGhqoBH1.mjs}

@@ -1,15 +1,15 @@
 import { o as isRecord } from "./record-coerce-CNDovvft.mjs";
 import { n as getAiTransportHost } from "./host-DqhMLv8J.mjs";
 import { d as supportsOpenAITemperature, i as isOpenAIGpt6Model, m as resolveOpenAIThinkingApi } from "./openai-reasoning-effort-C_0p3psG.mjs";
-import { r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-DHxMpeRz.mjs";
-import { o as usesNativeOpenAICodexResponsesBackend } from "./openai-completions-compat-D1q02qdx.mjs";
+import { r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-BcwCks0h.mjs";
+import { o as usesNativeOpenAICodexResponsesBackend } from "./openai-completions-compat-BuszNIRF.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
 import { v as sanitizeTransportPayloadText } from "./transport-stream-shared-Dt7VrJFd.mjs";
-import { C as stripSystemPromptCacheBoundary } from "./simple-options-B9u6-CF3.mjs";
-import { g as resolveOpenAISimpleReasoningEffort, h as resolveOpenAIRequestReasoning, k as resolvePromptCacheKey, m as reconcileOpenAIResponsesToolChoice, s as getCompat } from "./openai-transport-params-C4eZAXf2.mjs";
+import { C as stripSystemPromptCacheBoundary, r as clampMaxTokensToModel } from "./simple-options-B9u6-CF3.mjs";
+import { A as resolvePromptCacheKey, g as resolveOpenAISimpleReasoningEffort, h as resolveOpenAIRequestReasoning, m as reconcileOpenAIResponsesToolChoice, s as getCompat } from "./openai-transport-params-DzSQTRdN.mjs";
 import { h as responsesPromptObserver, r as OPENAI_CODEX_RESPONSES_DEFAULT_INSTRUCTIONS } from "./openai-responses-contracts-CLxnqS9p.mjs";
 import { a as applyOpenAIResponsesPayloadPolicy, s as resolveOpenAIResponsesPayloadPolicy } from "./openai-responses-compaction-window-CYGCjAQz.mjs";
-import { F as convertResponsesMessages, N as buildResponsesInputMessage, a as prepareResponsesTools } from "./openai-responses-shared-BpGnmnV8.mjs";
+import { F as convertResponsesMessages, N as buildResponsesInputMessage, a as prepareResponsesTools } from "./openai-responses-shared-DLz_xgps.mjs";
 //#region packages/ai/src/transports/openai-responses-params-internal.ts
 const OPENAI_RESPONSES_TOOL_CALL_PROVIDERS = /* @__PURE__ */ new Set([
 	"openai",
@@ -124,7 +124,7 @@ function buildOpenAIResponsesParams(model, context, options, metadata, replayMod
 		...metadata ? { metadata } : {}
 	};
 	const effectiveMaxTokens = options?.maxTokens || model.maxTokens;
-	if (effectiveMaxTokens) params.max_output_tokens = Math.max(effectiveMaxTokens, 16);
+	if (effectiveMaxTokens) params.max_output_tokens = clampMaxTokensToModel(model, Math.max(effectiveMaxTokens, 16));
 	if (options?.temperature !== void 0 && supportsOpenAITemperature(model)) params.temperature = options.temperature;
 	if (options?.topP !== void 0 && (!isOpenAIGpt6Model(model) || resolveOpenAIThinkingApi(model.api) === "azure-openai-responses")) params.top_p = options.topP;
 	if (options?.responseFormat !== void 0) params.text = {