@openclaw/ai 2026.9.1-beta.1 → 2026.9.1

package/dist/transports.mjs

@@ -1,31 +1,30 @@
-import { n as getEnvApiKey } from "./env-api-keys-DrgeBuva.mjs";
-import { d as resolveClaudeSonnet5ModelIdentity, g as supportsClaudeNativeXhighEffort, p as supportsClaudeAdaptiveThinking, u as resolveClaudeOpus5ModelIdentity } from "./src-Bu3FiFAG.mjs";
+import { g as supportsClaudeAdaptiveThinking, m as resolveClaudeSonnet5ModelIdentity, p as resolveClaudeOpus5ModelIdentity, y as supportsClaudeNativeXhighEffort } from "./src-C8U7lkoa.mjs";
 import { c as normalizeLowercaseStringOrEmpty, o as isRecord, s as hasNonEmptyString } from "./utf16-slice-qz3nsy87.mjs";
-import { r as calculateCost } from "./sanitize-unicode-DCkltN94.mjs";
-import { S as usesClaudeStreamingRefusalContract, _ as prepareClaudeNoPrefillRequestContext, a as describeToolResultMediaPlaceholder, c as extractToolResultText, d as isImageWithMediaPayload, f as ANTHROPIC_CLAUDE_CODE_BILLING_SYSTEM_BLOCK, g as mapAnthropicStopReason, h as defaultsClaudeAdaptiveThinking, m as applyClaudeRequestContract, n as getAiTransportHost, p as ANTHROPIC_CLAUDE_CODE_VERSION, r as resolveAiTransportHeaderSentinels, s as extractToolResultBlockText, v as requiresClaudeAdaptiveThinking, x as usesClaudeFable5MessagesContract, y as resolveAnthropicThinkingEffort } from "./host-D51fmkH6.mjs";
-import { i as canonicalizeBase64, o as stableStringify } from "./provider-error-DDOs_Bz0.mjs";
-import { n as toErrorObject, t as createResponsesPromptEgressObserver } from "./openai-responses-prompt-observer-internal-Cd4hJG-S.mjs";
-import { r as asNonNegativeFiniteNumber } from "./number-coercion-H9qHik3g.mjs";
+import { r as calculateCost } from "./sanitize-unicode-S6binQG-.mjs";
+import { S as usesClaudeStreamingRefusalContract, _ as prepareClaudeNoPrefillRequestContext, a as describeToolResultMediaPlaceholder, c as extractToolResultText, d as isImageWithMediaPayload, f as ANTHROPIC_CLAUDE_CODE_BILLING_SYSTEM_BLOCK, g as mapAnthropicStopReason, h as defaultsClaudeAdaptiveThinking, m as applyClaudeRequestContract, n as getAiTransportHost, p as ANTHROPIC_CLAUDE_CODE_VERSION, r as resolveAiTransportHeaderSentinels, s as extractToolResultBlockText, v as requiresClaudeAdaptiveThinking, x as usesClaudeFable5MessagesContract, y as resolveAnthropicThinkingEffort } from "./host-BIaiBURL.mjs";
+import { i as stableStringify } from "./provider-error-B6EGq6gg.mjs";
+import { r as toErrorObject } from "./diagnostics-KhXK-QJI.mjs";
+import { a as asNonNegativeFiniteNumber } from "./base64-CEFBpSkN.mjs";
 import { n as uniqueStrings } from "./string-normalization--fwJ4S2q.mjs";
+import { A as applyAnthropicCacheControlToMessages, C as applyAnthropicFallbackBoundary, D as createAnthropicInlineImageBudget, E as applyAnthropicRefusal, F as resolveAnthropicServerCompactionPlan, I as isAnthropicOAuthApiKey, L as omitFoundryBearerCredentialHeaders, M as applyAnthropicPayloadPolicyToParams, N as resolveAnthropicEphemeralCacheControl, O as normalizeAnthropicInlineContent, P as resolveAnthropicPayloadPolicy, R as usesFoundryBearerAuth, T as resolveAnthropicFallbackServingModelCost, _ as toClaudeCodeToolName, a as suppressAnthropicCompaction, b as ANTHROPIC_SERVER_SIDE_FALLBACKS, f as normalizeAnthropicToolCallId, g as resolveOriginalAnthropicToolName, h as reconcileAnthropicToolChoice, i as resolveNewestAnthropicCompaction, j as applyAnthropicEphemeralCacheControlMarkers, k as resolveAnthropicImageMediaType, m as projectAnthropicTools, n as createCompactionCapture, o as applyAnthropicMessageDeltaUsage, p as normalizeAnthropicToolChoice, r as isAnthropicReplayRejection, s as applyAnthropicMessageStartUsage, t as buildAnthropicReplayPlan, v as ANTHROPIC_OMITTED_REASONING_TEXT, w as readAnthropicFallbackBoundary, x as ANTHROPIC_SERVER_SIDE_FALLBACK_BETA, y as findActiveAnthropicToolTurnAssistantIndex } from "./anthropic-compaction-replay-OMJZ0uyo.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
-import { f as sortPromptCacheToolsByName, l as stripSystemPromptCacheBoundary, t as adjustMaxTokensForThinking } from "./simple-options-CFj7x3J4.mjs";
-import { a as resolveProviderEndpoint, c as transformTransportMessages, i as resolveOpenAIStrictToolSetting, n as buildGuardedModelFetch } from "./tool-schema-json-projection-qvTEs5Am.mjs";
-import { a as isGoogleGemini3FlashModel, c as parseRetryAfterSeconds, f as resolveModelHeaderSentinels$1, h as supportsModelTools, i as isCodeModeModelVisibleToolName, l as readResponseTextSnippet, m as sha256Hex, n as createAbortError$1, o as isGoogleGemini3ProModel, p as resolveSecretSentinel, r as estimateStringChars, t as MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE, u as redactIdentifier } from "./transport-utils-LJ1_rbi-.mjs";
-import { A as resolveAnthropicImageMediaType, C as readAnthropicFallbackBoundary, D as usesFoundryBearerAuth, E as omitFoundryBearerCredentialHeaders, F as resolveAnthropicPayloadPolicy, I as resolveAnthropicServerCompactionPlan, M as applyAnthropicEphemeralCacheControlMarkers, N as applyAnthropicPayloadPolicyToParams, O as createAnthropicInlineImageBudget, P as resolveAnthropicEphemeralCacheControl, S as applyAnthropicFallbackBoundary, T as applyAnthropicRefusal, _ as ANTHROPIC_OMITTED_REASONING_TEXT, a as applyAnthropicMessageDeltaUsage, b as ANTHROPIC_SERVER_SIDE_FALLBACK_BETA, d as normalizeAnthropicToolCallId, f as normalizeAnthropicToolChoice, g as toClaudeCodeToolName, h as resolveOriginalAnthropicToolName, i as suppressAnthropicCompaction, j as applyAnthropicCacheControlToMessages, k as normalizeAnthropicInlineContent, m as reconcileAnthropicToolChoice, n as createCompactionCapture, o as applyAnthropicMessageStartUsage, p as projectAnthropicTools, r as isAnthropicReplayRejection, t as buildAnthropicReplayPlan, v as findActiveAnthropicToolTurnAssistantIndex, w as resolveAnthropicFallbackServingModelCost, y as ANTHROPIC_SERVER_SIDE_FALLBACKS } from "./anthropic-compaction-replay-nhKaCXu6.mjs";
-import { S as quoteUnsafeIntegerLiterals, _ as transportAbortError, a as createWritableTransportEventStream, b as parseJsonObjectPreservingUnsafeIntegers, c as finalizeTransportStream, d as notifyProviderHttpMetadata, f as notifyProviderHttpResponse, g as sanitizeTransportPayloadText, h as sanitizeNonEmptyTransportPayloadText, i as createEmptyTransportUsage, l as mergeTransportHeaders, m as parseTerminalToolCallArguments, n as coerceTransportToolCallArguments, o as failTransportStream, p as notifyProviderStreamOpened, r as copyProviderAcceptanceObserver, s as finalizeTerminalToolCallArguments, t as assignTransportErrorDetails, u as mergeTransportMetadata, v as withProviderAcceptanceObserver, x as parseJsonPreservingUnsafeIntegers, y as withProviderResponseHook } from "./transport-stream-shared-ljShVIZB.mjs";
-import { C as createDeepSeekTextFilter, E as tagUnresolvedTextAsCommentary, S as shouldOmitOllamaCompatResponseFormat, T as tagPendingCommentaryText, _ as hasToolCallHistory, a as buildOpenAISdkClientOptions, b as resolveOpenAICompletionsCompat, c as filterCodeModePayloadTools, d as readCodeModePayloadToolName, f as resolveCodeModeResponsesVisibleToolNames, g as convertMessages, h as resolveOpenAIReasoningEffortMap, i as buildOpenAIClientHeaders, l as getCompat, m as usesNativeOpenAICodexResponsesBackend, n as shouldEmitOpenAICompletionsReasoning, o as buildOpenAISdkRequestOptions, p as resolveOpenAIStrictToolFlagWithDiagnostics, r as assertCodeModeResponsesToolSurface, s as enforceCodeModeResponsesToolSurface, t as processCompletionsStream, u as isOpenAICodexResponsesModel, v as finalizeOpenAICompletionsToolCalls, x as resolveOpenAICompletionsResponseFormat, y as detectOpenAICompletionsCompat } from "./openai-completions-stream-DUlGwRMz.mjs";
+import { a as codeModeToolSurfaceObserver, d as stripSystemPromptCacheBoundary, m as sortPromptCacheToolsByName, o as reasoningTagTextPolicy, t as adjustMaxTokensForThinking } from "./simple-options-0PLDyJ-d.mjs";
+import { a as resolveProviderEndpoint, c as transformTransportMessages, i as resolveOpenAIStrictToolSetting, n as buildGuardedModelFetch } from "./tool-schema-json-projection-BtZiml7r.mjs";
+import { a as isGoogleGemini3FlashModel, c as parseRetryAfterSeconds, f as resolveModelHeaderSentinels$1, i as isCodeModeModelVisibleToolName, l as readResponseTextSnippet, m as supportsModelTools, n as createAbortError$1, o as isGoogleGemini3ProModel, p as sha256Hex, r as estimateStringChars, t as MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE, u as redactIdentifier } from "./transport-utils-CtuS1Upe.mjs";
+import { n as getEnvApiKey } from "./env-api-keys-bktO00EJ.mjs";
+import { S as quoteUnsafeIntegerLiterals, _ as transportAbortError, a as createWritableTransportEventStream, b as parseJsonObjectPreservingUnsafeIntegers, c as finalizeTransportStream, d as notifyProviderHttpMetadata, f as notifyProviderHttpResponse, g as sanitizeTransportPayloadText, h as sanitizeNonEmptyTransportPayloadText, i as createEmptyTransportUsage, l as mergeTransportHeaders, m as parseTerminalToolCallArguments, n as coerceTransportToolCallArguments, o as failTransportStream, p as notifyProviderStreamOpened, r as copyProviderAcceptanceObserver, s as finalizeTerminalToolCallArguments, t as assignTransportErrorDetails, u as mergeTransportMetadata, v as withProviderAcceptanceObserver, x as parseJsonPreservingUnsafeIntegers, y as withProviderResponseHook } from "./transport-stream-shared-CytPVLIg.mjs";
+import { C as createDeepSeekTextFilter, E as tagUnresolvedTextAsCommentary, S as shouldOmitOllamaCompatResponseFormat, T as tagPendingCommentaryText, _ as hasToolCallHistory, a as buildOpenAISdkClientOptions, b as resolveOpenAICompletionsCompat, c as filterCodeModePayloadTools, d as readCodeModePayloadToolName, f as resolveCodeModeResponsesVisibleToolNames, g as convertMessages, h as resolveOpenAIReasoningEffortMap, i as buildOpenAIClientHeaders, l as getCompat, m as usesNativeOpenAICodexResponsesBackend, n as shouldEmitOpenAICompletionsReasoning, o as buildOpenAISdkRequestOptions, p as resolveOpenAIStrictToolFlagWithDiagnostics, r as assertCodeModeResponsesToolSurface, s as enforceCodeModeResponsesToolSurface, t as processCompletionsStream, u as isOpenAICodexResponsesModel, v as finalizeOpenAICompletionsToolCalls, x as resolveOpenAICompletionsResponseFormat, y as detectOpenAICompletionsCompat } from "./openai-completions-stream-DZjwK8vp.mjs";
 import { t as createDeferredEventBuffer } from "./deferred-event-buffer-DAvyP7qA.mjs";
 import { r as parseStreamingJson, t as createToolArgumentPreviewSchedule } from "./json-parse-BuAJEbdW.mjs";
 import { t as notifyLlmRequestActivity } from "./llm-request-activity-BjtkplhG.mjs";
 import { t as googleFlashSupportsMinimalThinking } from "./google-thinking-level-C-V3tecN.mjs";
-import { B as buildResponsesInputMessage, C as summarizeResponsesTools, G as suppressOpenAIResponsesCompaction, H as createOpenAIResponsesAssistantOutput, I as createResponsesStreamWithEncryptedContentRetry, J as resolveModelPayloadDebugMode, K as resolveReplayableResponsesMessageId, L as isInvalidEncryptedContentError, R as resolveAzureOpenAIApiVersion, S as summarizeResponsesPayload, U as buildOpenAIResponsesReasoningReplayMetadata, V as convertResponsesMessages, W as captureOpenAIResponsesCompaction, Y as resolveModelSseDebugMode, _ as safeDebugValue, b as summarizeOpenAITransportError, c as processResponsesStream, f as observeResponsesStream, g as normalizeResponsesFailedEvent, h as logResponsesFailedNoDetails, m as buildResponsesFailedNoDetailsObservation, n as applyResponsesServiceTierPricing, p as ResponsesStreamFailure, q as emitModelTransportDebug, v as stringifyRedactedEvent, x as summarizeResponsesFailedNoDetailsObservation, y as stringifyRedactedPayload, z as resolveNextResponsesEncryptedContentAttempt } from "./openai-responses-shared-DJThc6jW.mjs";
-import { t as codeModeToolSurfaceObserver } from "./provider-options-AvldWZt8.mjs";
-import { A as readOpenAICompletionsReasoningBatch, C as createOpenAIProviderAcceptanceHook, D as measureUtf8AppendBytes, E as log, M as resolvePromptCacheKey, N as throwIfModelStreamAborted, O as parseOpenAICompletionsUsage, S as createModelStreamCooperativeScheduler, T as isOpenAICompletionsThinkingEnabled, b as reconcileOpenAIResponsesToolChoice, j as resolveOpenAIClientBaseUrl, k as readOpenAICompletionsContentDeltas, r as normalizeOpenAIStrictToolParameters, v as projectOpenAITools, w as createOpenAIResponseHook, x as GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, y as reconcileOpenAICompletionsToolChoice } from "./openai-tool-schema-ho-hIkQf.mjs";
-import { i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-MK28puvq.mjs";
-import { _ as isOpenAIGpt56Model, d as OpenAIResponsesWebSocketSafeRetryError, g as isOpenAIGpt55Model, h as isOpenAIGpt54MiniModel, i as OPENAI_RESPONSES_APIS, l as OpenAIResponsesWebSocketPostDispatchError, p as parseOpenAIResponsesWebSocketServerError, r as OPENAI_CODEX_RESPONSES_DEFAULT_INSTRUCTIONS, t as AZURE_RESPONSES_FIRST_EVENT_TIMEOUT_MS, u as OpenAIResponsesWebSocketPreDispatchError, v as normalizeOpenAIReasoningEffort, y as resolveOpenAIReasoningEffortForModel } from "./openai-responses-contracts-QUY12X0B.mjs";
-import { i as resolveOpenAIResponsesServerCompactionPlan, n as resolveOpenAIResponsesCompactEndpointPlan, r as resolveOpenAIResponsesPayloadPolicy, t as applyOpenAIResponsesPayloadPolicy } from "./openai-responses-payload-policy-D8EdimYe.mjs";
+import { B as buildResponsesInputMessage, C as summarizeResponsesTools, G as captureOpenAIResponsesCompaction, H as createOpenAIResponsesAssistantOutput, I as createResponsesStreamWithEncryptedContentRetry, J as resolveReplayableResponsesMessageId, K as resolveNewestOpenAIResponsesCompactionReplay, L as isInvalidEncryptedContentError, R as resolveAzureOpenAIApiVersion, S as summarizeResponsesPayload, U as CompactionReplayRefreshRequiredError, V as convertResponsesMessages, W as buildOpenAIResponsesReasoningReplayMetadata, X as resolveModelPayloadDebugMode, Y as emitModelTransportDebug, Z as resolveModelSseDebugMode, _ as safeDebugValue, b as summarizeOpenAITransportError, c as processResponsesStream, f as observeResponsesStream, g as normalizeResponsesFailedEvent, h as logResponsesFailedNoDetails, m as buildResponsesFailedNoDetailsObservation, n as applyResponsesServiceTierPricing, p as ResponsesStreamFailure, q as suppressOpenAIResponsesCompaction, v as stringifyRedactedEvent, x as summarizeResponsesFailedNoDetailsObservation, y as stringifyRedactedPayload, z as resolveNextResponsesEncryptedContentAttempt } from "./openai-responses-shared-DiAdpNAM.mjs";
+import { A as readOpenAICompletionsReasoningBatch, C as createOpenAIProviderAcceptanceHook, D as measureUtf8AppendBytes, E as log, M as resolvePromptCacheKey, N as throwIfModelStreamAborted, O as parseOpenAICompletionsUsage, S as createModelStreamCooperativeScheduler, T as isOpenAICompletionsThinkingEnabled, b as reconcileOpenAIResponsesToolChoice, j as resolveOpenAIClientBaseUrl, k as readOpenAICompletionsContentDeltas, r as normalizeOpenAIStrictToolParameters, v as projectOpenAITools, w as createOpenAIResponseHook, x as GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, y as reconcileOpenAICompletionsToolChoice } from "./openai-tool-schema-BMiHFH36.mjs";
+import { i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-DcNjoFQE.mjs";
+import { C as isOpenAIGpt54MiniModel, D as resolveOpenAIReasoningEffortForModel, E as normalizeOpenAIReasoningEffort, T as isOpenAIGpt56Model, _ as OpenAIResponsesWebSocketPostDispatchError, a as applyOpenAIResponsesPayloadPolicy, c as resolveOpenAIResponsesServerCompactionPlan, d as OPENAI_CODEX_RESPONSES_DEFAULT_INSTRUCTIONS, f as OPENAI_RESPONSES_APIS, i as sanitizeResponsesImagePayload, l as AZURE_RESPONSES_FIRST_EVENT_TIMEOUT_MS, n as isOpenAIResponsesCompactionOutput, o as resolveOpenAIResponsesCompactEndpointPlan, s as resolveOpenAIResponsesPayloadPolicy, t as createBoundedOpenAIResponsesCompactionFetch, v as OpenAIResponsesWebSocketPreDispatchError, w as isOpenAIGpt55Model, x as parseOpenAIResponsesWebSocketServerError, y as OpenAIResponsesWebSocketSafeRetryError } from "./openai-responses-compaction-window-DO7yV7az.mjs";
 import { i as resolveAzureDeploymentNameFromMap, t as isOpenAICompatibleAzureResponsesBaseUrl } from "./azure-openai-responses-client-compat-C7K7QfUE.mjs";
 import { n as registerSessionResourceCleanup } from "./session-resources-CkR4WWy1.mjs";
+import { t as createResponsesPromptEgressObserver } from "./openai-responses-prompt-observer-internal-C7x4IK7r.mjs";
 import { randomUUID } from "node:crypto";
 import OpenAI, { AzureOpenAI } from "openai";
 import { ResponsesWS } from "openai/resources/responses/ws.js";
@@ -35,12 +34,12 @@ import { ResponsesWS } from "openai/resources/responses/ws.js";
 * Converts OpenClaw contexts/tools into Anthropic payloads, streams SSE events
 * back into runtime output blocks, and applies provider request policy.
 */
-const ANTHROPIC_MESSAGES_ERROR_BODY_MAX_BYTES = 8 * 1024;
+const ANTHROPIC_MESSAGES_ERROR_BODY_MAX_BYTES = 8192;
 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 ANTHROPIC_MESSAGES_SSE_PENDING_BUFFER_MAX_CHARS = 16777216;
 function resolveAnthropicRequestModelId(model) {
 	if (isDirectAnthropicModel(model) && /^anthropic\//i.test(model.id)) return model.id.replace(/^anthropic\//i, "");
 	return model.id;
@@ -60,9 +59,6 @@ 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 isAnthropicOAuthToken(apiKey) {
-	return (resolveSecretSentinel(apiKey) ?? apiKey).includes("sk-ant-oat");
-}
 function isDirectAnthropicModel(model) {
 	if (normalizeLowercaseStringOrEmpty(model.provider) !== "anthropic") return false;
 	const endpointClass = resolveProviderEndpoint(model).endpointClass;
@@ -482,7 +478,7 @@ function createAnthropicTransportClient(params) {
 	}
 	const betaFeatures = ["fine-grained-tool-streaming-2025-05-14"];
 	if (needsInterleavedBeta) betaFeatures.push("interleaved-thinking-2025-05-14");
-	if (isAnthropicOAuthToken(apiKey)) {
+	if (isAnthropicOAuthApiKey(apiKey)) {
 		const betaHeader = buildAnthropicBetaHeader(model, betaFeatures, { oauth: true });
 		return {
 			client: createAnthropicMessagesClient({
@@ -579,18 +575,19 @@ async function buildAnthropicParams(model, context, isOAuthToken, options) {
 		if (convertedTools.tools.length > 0) params.tools = convertedTools.tools;
 	}
 	if (mandatoryAdaptiveThinking || model.reasoning || supportsClaudeAdaptiveThinking(model)) {
-		if (mandatoryAdaptiveThinking || options?.thinkingEnabled) if (supportsClaudeAdaptiveThinking(model)) {
-			params.thinking = {
-				type: "adaptive",
-				display: options?.thinkingDisplay ?? "summarized"
+		if (mandatoryAdaptiveThinking || options?.thinkingEnabled) {
+			if (supportsClaudeAdaptiveThinking(model)) {
+				params.thinking = {
+					type: "adaptive",
+					display: options?.thinkingDisplay ?? "summarized"
+				};
+				const effort = options?.effort ?? (mandatoryAdaptiveThinking ? "high" : void 0);
+				if (effort) params.output_config = { effort };
+			} else params.thinking = {
+				type: "enabled",
+				budget_tokens: options?.thinkingBudgetTokens ?? 1024
 			};
-			const effort = options?.effort ?? (mandatoryAdaptiveThinking ? "high" : void 0);
-			if (effort) params.output_config = { effort };
-		} else params.thinking = {
-			type: "enabled",
-			budget_tokens: options?.thinkingBudgetTokens ?? 1024
-		};
-		else if (options?.thinkingEnabled === false) params.thinking = { type: "disabled" };
+		} else if (options?.thinkingEnabled === false) params.thinking = { type: "disabled" };
 	}
 	if (options?.metadata && typeof options.metadata.user_id === "string") params.metadata = { user_id: options.metadata.user_id };
 	if (options?.toolChoice) {
@@ -987,16 +984,18 @@ function createAnthropicMessagesTransportStreamFn() {
 							let appendedContent = false;
 							if (!hasNativeAnthropicDelta && typeof delta?.content === "string" && delta.content.length > 0) {
 								const text = sanitizeTransportPayloadText(delta.content);
-								if (text.length > 0) if (block?.type === "text" && index !== void 0) {
-									block.text += text;
-									eventSink.push({
-										type: "text_delta",
-										contentIndex: index,
-										delta: text,
-										partial: output
-									});
-									appendedContent = true;
-								} else appendedContent = appendReasoningContentTextDelta(event.index, text);
+								if (text.length > 0) {
+									if (block?.type === "text" && index !== void 0) {
+										block.text += text;
+										eventSink.push({
+											type: "text_delta",
+											contentIndex: index,
+											delta: text,
+											partial: output
+										});
+										appendedContent = true;
+									} else appendedContent = appendReasoningContentTextDelta(event.index, text);
+								}
 							}
 							if ((appendedThinking || appendedContent) && !hasNativeAnthropicDelta) continue;
 						}
@@ -1104,8 +1103,10 @@ function createAnthropicMessagesTransportStreamFn() {
 					if (event.type === "message_delta") {
 						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 = mapAnthropicStopReason(delta.stop_reason);
+						if (delta?.stop_reason) {
+							if (delta.stop_reason === "refusal") applyAnthropicRefusal(output, delta.stop_details, model.provider);
+							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);
@@ -1134,17 +1135,17 @@ function createAnthropicMessagesTransportStreamFn() {
 					output
 				});
 			} catch (error) {
-				if (refusalBuffer) {
-					refusalBuffer.discard();
-					output.content = [];
-				} else output.content = output.content.filter((block) => block.type !== "toolCall");
-				if (usedCompactionReplay && isAnthropicReplayRejection(error)) suppressAnthropicCompaction(output, model, options);
 				failTransportStream({
 					stream,
 					output,
 					signal: options?.signal,
 					error,
 					cleanup: () => {
+						if (refusalBuffer) {
+							refusalBuffer.discard();
+							output.content = [];
+						} else output.content = output.content.filter((block) => block.type !== "toolCall");
+						if (usedCompactionReplay && isAnthropicReplayRejection(error)) suppressAnthropicCompaction(output, model, options);
 						for (const block of output.content) delete block.index;
 					}
 				});
@@ -1595,10 +1596,11 @@ function convertTools(tools, compat, model) {
 function buildOpenAICompletionsParams(model, context, options) {
 	const compat = getCompat(model);
 	const compatDetection = detectOpenAICompletionsCompat(model);
-	let messages = convertMessages(model, context.systemPrompt ? {
+	const completionsContext = context.systemPrompt ? {
 		...context,
 		systemPrompt: stripSystemPromptCacheBoundary(context.systemPrompt)
-	} : context, compat);
+	} : context;
+	let messages = convertMessages(model, completionsContext, compat);
 	applyCompletionsReplay(messages, context, model, compat);
 	if (compat.strictMessageKeys) messages = stripCompletionMessagesToRoleContent(messages);
 	const cacheRetention = resolveCacheRetention(options?.cacheRetention);
@@ -1659,8 +1661,10 @@ function buildOpenAICompletionsParams(model, context, options) {
 				emitModelTransportDebug(log, `[completions] clamp_max_tokens provider=${model.provider} api=${model.api} model=${model.id} requested=${effectiveMaxTokens} output=${clampedMaxTokens} effectiveContext=${effectiveContextTokens} estimatedInput=${estimatedInputTokens}`);
 			}
 		}
-		if (clampedMaxTokens) if (compat.maxTokensField === "max_tokens") params.max_tokens = clampedMaxTokens;
-		else params.max_completion_tokens = clampedMaxTokens;
+		if (clampedMaxTokens) {
+			if (compat.maxTokensField === "max_tokens") params.max_tokens = clampedMaxTokens;
+			else params.max_completion_tokens = clampedMaxTokens;
+		}
 	}
 	const completionsReasoningEffort = resolveOpenAICompletionsReasoningEffort(options);
 	const resolvedCompletionsReasoningEffort = completionsReasoningEffort ? resolveOpenAIReasoningEffortForModel({
@@ -1831,11 +1835,8 @@ function createOpenAICompletionsTransportStreamFn() {
 				if (getCompat(model).requiresNonEmptyUserOrAssistantMessage) assertOpenAICompletionsPayloadHasConversationTurn(params, model);
 				const emitReasoning = shouldEmitOpenAICompletionsReasoning(model, options);
 				firstEventAbort = createFirstStreamEventAbortController(options?.signal);
-				const { data: responseStream, response } = await client.chat.completions.create(params, buildOpenAISdkRequestOptions(model, firstEventAbort.signal, {
-					timeoutMs: options?.timeoutMs,
-					maxRetries: options?.maxRetries
-				})).withResponse();
-				await processCompletionsStream(withProviderResponseHook({
+				const { data: responseStream, response } = await client.chat.completions.create(params, buildOpenAISdkRequestOptions(model, firstEventAbort.signal, { timeoutMs: options?.timeoutMs })).withResponse();
+				const hookedResponseStream = withProviderResponseHook({
 					stream: responseStream,
 					signal: firstEventAbort.signal,
 					abort: firstEventAbort.abort,
@@ -1844,9 +1845,11 @@ function createOpenAICompletionsTransportStreamFn() {
 						type: "start",
 						partial: output
 					})
-				}), output, model, stream, {
+				});
+				await processCompletionsStream(hookedResponseStream, output, model, stream, {
 					signal: options?.signal,
 					emitReasoning,
+					strictReasoningTags: reasoningTagTextPolicy.isStrict(options),
 					firstEventTimeoutMs: getFirstStreamEventTimeoutMs(options),
 					abortFirstEventStream: firstEventAbort.abort,
 					onFirstEventTimeout: getFirstStreamEventTimeoutHandler(options),
@@ -1911,7 +1914,7 @@ async function requestPreparedOpenAIResponsesCompaction(streamFn, model, context
 }
 //#endregion
 //#region packages/ai/src/transports/openai-responses-continuation.ts
-const HTTP_CONTINUATION_IDLE_TTL_MS = 300 * 1e3;
+const HTTP_CONTINUATION_IDLE_TTL_MS = 3e5;
 const TURN_HEADERS = /* @__PURE__ */ new Set([
 	"traceparent",
 	"x-openclaw-turn-id",
@@ -1921,7 +1924,7 @@ function jsonValuesEqual(left, right) {
 	return stableStringify(JSON.parse(JSON.stringify(left))) === stableStringify(JSON.parse(JSON.stringify(right)));
 }
 function requestWithoutInput(request) {
-	const { input: _input, previous_response_id: _previousResponseId, ...rest } = request;
+	const { input: _input, previous_response_id: _previousResponseId, instructions: _instructions, tools: _tools, ...rest } = request;
 	if (!isRecord(rest.metadata)) return rest;
 	const metadata = Object.fromEntries(Object.entries(rest.metadata).filter(([key]) => key !== "openclaw_turn_id" && key !== "openclaw_turn_attempt"));
 	return {
@@ -1929,12 +1932,16 @@ function requestWithoutInput(request) {
 		metadata
 	};
 }
-function normalizeAssistantReplayInput(input) {
+function normalizeAssistantReplayInput(input, fromResponse = false) {
 	return input.map((item) => {
 		if (!isRecord(item)) return item;
 		if (item.type === "reasoning") return { type: "reasoning" };
 		if (item.type !== "function_call" && !(item.type === "message" && item.role === "assistant")) return item;
 		const { id: _id, status: _status, ...stableItem } = item;
+		if (fromResponse && item.type === "function_call") {
+			const args = parseJsonObjectPreservingUnsafeIntegers(stableItem.arguments);
+			stableItem.arguments = args ? JSON.stringify(args) : stableItem.arguments;
+		}
 		if (item.type === "message" && Array.isArray(stableItem.content)) stableItem.content = stableItem.content.map((part) => {
 			if (!isRecord(part) || part.type !== "output_text") return part;
 			const { annotations: _annotations, logprobs: _logprobs, ...stablePart } = part;
@@ -1963,7 +1970,7 @@ function resolveResponsesContinuationRequest(continuation, request) {
 		request,
 		continuationStatus: "history_shorter"
 	};
-	if (!jsonValuesEqual(normalizeAssistantReplayInput(currentInput.slice(0, previousInput.length)), normalizeAssistantReplayInput(previousInput)) || !jsonValuesEqual(normalizeAssistantReplayInput(currentInput.slice(previousInput.length, baselineLength)), normalizeAssistantReplayInput(continuation.lastResponseItems))) return {
+	if (!jsonValuesEqual(normalizeAssistantReplayInput(currentInput.slice(0, previousInput.length)), normalizeAssistantReplayInput(previousInput)) || !jsonValuesEqual(normalizeAssistantReplayInput(currentInput.slice(previousInput.length, baselineLength)), normalizeAssistantReplayInput(continuation.lastResponseItems, true))) return {
 		request,
 		continuationStatus: "history_changed"
 	};
@@ -2112,18 +2119,28 @@ function sanitizeOpenAICodexResponsesParams(model, params) {
 	stripOpenAICodexResponsesUnsupportedTextFields(params);
 	return params;
 }
-function buildOpenAICodexResponsesInstructions(context) {
+function buildOpenAIResponsesInstructionsText(context) {
 	if (!context.systemPrompt) return;
 	return sanitizeTransportPayloadText(stripSystemPromptCacheBoundary(context.systemPrompt));
 }
-function resolveOpenAICodexResponsesInstructions(model, context) {
-	const instructions = buildOpenAICodexResponsesInstructions(context);
+function resolveOpenAIResponsesInstructions(model, context, usesInstructionsField) {
+	if (!usesInstructionsField) return;
+	const instructions = buildOpenAIResponsesInstructionsText(context);
 	if (instructions && instructions.trim().length > 0) return instructions;
 	return usesNativeOpenAICodexResponsesBackend(model) ? OPENAI_CODEX_RESPONSES_DEFAULT_INSTRUCTIONS : void 0;
 }
-function ensureOpenAICodexResponsesInput(messages, context) {
+function buildOpenAIResponsesCompactSystemMessage(model, instructions) {
+	const compat = getCompat(model);
+	const supportsDeveloperRole = typeof compat.supportsDeveloperRole === "boolean" ? compat.supportsDeveloperRole : void 0;
+	const role = model.reasoning && supportsDeveloperRole !== false ? "developer" : "system";
+	return buildResponsesInputMessage(role, [{
+		type: "input_text",
+		text: instructions
+	}]);
+}
+function ensureOpenAIResponsesNonEmptyInput(messages, context) {
 	if (messages.length > 0 || !context.systemPrompt) return;
-	if (!buildOpenAICodexResponsesInstructions(context)) throw new Error("OpenAI Codex Responses requires non-empty input when only systemPrompt is provided.");
+	if (!buildOpenAIResponsesInstructionsText(context)) throw new Error("OpenAI Responses requires non-empty input when only systemPrompt is provided.");
 	messages.push(buildResponsesInputMessage("user", [{
 		type: "input_text",
 		text: " "
@@ -2137,16 +2154,12 @@ function resolveOpenAIResponsesTextFormat(responseFormat) {
 	return responseFormat;
 }
 function convertOpenAIResponsesMessagesForRequest(model, context, options, replayMode) {
-	const isCodexResponses = isOpenAICodexResponsesModel(model);
 	const isNativeCodexResponses = usesNativeOpenAICodexResponsesBackend(model);
-	const compat = getCompat(model);
-	const supportsDeveloperRole = typeof compat.supportsDeveloperRole === "boolean" ? compat.supportsDeveloperRole : void 0;
 	const payloadPolicy = resolveOpenAIResponsesPayloadPolicy(model, { storeMode: "disable" });
 	const policyAllowsReplayIds = payloadPolicy.explicitStore !== false && !payloadPolicy.shouldStripStore;
 	const replayResponsesItemIds = !isNativeCodexResponses && (options?.replayResponsesItemIds ?? policyAllowsReplayIds);
 	return convertResponsesMessages(model, context, OPENAI_RESPONSES_TOOL_CALL_PROVIDERS, {
-		includeSystemPrompt: !isCodexResponses,
-		supportsDeveloperRole,
+		includeSystemPrompt: !payloadPolicy.usesInstructionsField,
 		replayReasoningItems: true,
 		replayResponsesItemIds,
 		authProfileId: options?.authProfileId,
@@ -2155,19 +2168,19 @@ function convertOpenAIResponsesMessagesForRequest(model, context, options, repla
 	});
 }
 function buildOpenAIResponsesParams(model, context, options, metadata, replayMode = "checkpoint") {
-	const isCodexResponses = isOpenAICodexResponsesModel(model);
 	const payloadPolicy = resolveOpenAIResponsesPayloadPolicy(model, { storeMode: "disable" });
 	const messages = convertOpenAIResponsesMessagesForRequest(model, context, options, replayMode);
-	if (isCodexResponses) ensureOpenAICodexResponsesInput(messages, context);
+	ensureOpenAIResponsesNonEmptyInput(messages, context);
 	const cacheRetention = resolveCacheRetention(options?.cacheRetention);
 	const promptCacheKey = resolvePromptCacheKey(options, cacheRetention);
+	const instructions = resolveOpenAIResponsesInstructions(model, context, payloadPolicy.usesInstructionsField);
 	const params = {
 		model: model.id,
 		input: messages,
 		stream: true,
 		prompt_cache_key: promptCacheKey,
 		prompt_cache_retention: getPromptCacheRetention(model.baseUrl, cacheRetention),
-		...isCodexResponses ? { instructions: resolveOpenAICodexResponsesInstructions(model, context) } : {},
+		...instructions ? { instructions } : {},
 		...metadata ? { metadata } : {}
 	};
 	const effectiveMaxTokens = options?.maxTokens || model.maxTokens;
@@ -2189,9 +2202,10 @@ function buildOpenAIResponsesParams(model, context, options, metadata, replayMod
 	}
 	if (model.reasoning) {
 		if (options?.reasoningEffort || options?.reasoning || options?.reasoningSummary) {
+			const requestedReasoningEffort = resolveOpenAIReasoningEffort(options);
 			const resolvedReasoningEffort = resolveOpenAIReasoningEffortForModel({
 				model,
-				effort: resolveOpenAIReasoningEffort(options)
+				effort: requestedReasoningEffort
 			});
 			const reasoningEffort = resolvedReasoningEffort ? raiseMinimalReasoningForResponsesWebSearch({
 				model,
@@ -2218,8 +2232,8 @@ function buildOpenAIResponsesParams(model, context, options, metadata, replayMod
 }
 //#endregion
 //#region packages/ai/src/transports/openai-responses-websocket.ts
-const SESSION_WEBSOCKET_CACHE_TTL_MS = 300 * 1e3;
-const SESSION_WEBSOCKET_MAX_AGE_MS = 3300 * 1e3;
+const SESSION_WEBSOCKET_CACHE_TTL_MS = 3e5;
+const SESSION_WEBSOCKET_MAX_AGE_MS = 33e5;
 const WEBSOCKET_OPEN_STATE = 1;
 const websocketSessionCache = /* @__PURE__ */ new Map();
 const degradedWebSocketConnections = /* @__PURE__ */ new Map();
@@ -2472,149 +2486,6 @@ function closeOpenAIResponsesWebSocketSessions(sessionId) {
 }
 registerSessionResourceCleanup(closeOpenAIResponsesWebSocketSessions);
 //#endregion
-//#region packages/media-core/src/inline-image-data-url.ts
-/** Prefix used to distinguish inline data URLs from remote/local image references. */
-const INLINE_IMAGE_DATA_URL_PREFIX = "data:";
-const IMAGE_SIGNATURES = [
-	{
-		mime: "image/png",
-		matches: (buffer) => buffer.length >= 8 && buffer[0] === 137 && buffer[1] === 80 && buffer[2] === 78 && buffer[3] === 71 && buffer[4] === 13 && buffer[5] === 10 && buffer[6] === 26 && buffer[7] === 10
-	},
-	{
-		mime: "image/jpeg",
-		matches: (buffer) => buffer.length >= 3 && buffer[0] === 255 && buffer[1] === 216 && buffer[2] === 255
-	},
-	{
-		mime: "image/webp",
-		matches: (buffer) => buffer.length >= 12 && buffer.subarray(0, 4).toString("ascii") === "RIFF" && buffer.subarray(8, 12).toString("ascii") === "WEBP"
-	},
-	{
-		mime: "image/gif",
-		matches: (buffer) => buffer.length >= 6 && (buffer.subarray(0, 6).toString("ascii") === "GIF87a" || buffer.subarray(0, 6).toString("ascii") === "GIF89a")
-	},
-	{
-		mime: "image/bmp",
-		matches: (buffer) => buffer.length >= 2 && buffer[0] === 66 && buffer[1] === 77
-	}
-];
-const HEIC_BRANDS = /* @__PURE__ */ new Set([
-	"heic",
-	"heix",
-	"hevc",
-	"hevx",
-	"heis",
-	"heim",
-	"hevm",
-	"hevs"
-]);
-const HEIF_BRANDS = /* @__PURE__ */ new Set(["mif1", "msf1"]);
-const IMAGE_SIGNATURE_PREFIX_BASE64_CHARS = 128;
-const INLINE_IMAGE_DATA_URL_MIMES = /* @__PURE__ */ new Set([
-	"image/png",
-	"image/jpeg",
-	"image/webp",
-	"image/gif"
-]);
-function startsWithDataUrl(value) {
-	return value.slice(0, 5).toLowerCase() === INLINE_IMAGE_DATA_URL_PREFIX;
-}
-function sniffIsoBmffImageMime(buffer) {
-	if (buffer.length < 12 || buffer.subarray(4, 8).toString("ascii") !== "ftyp") return;
-	const brands = [buffer.subarray(8, 12).toString("ascii")];
-	for (let offset = 16; offset + 4 <= buffer.length; offset += 4) brands.push(buffer.subarray(offset, offset + 4).toString("ascii"));
-	if (brands.some((brand) => HEIC_BRANDS.has(brand))) return "image/heic";
-	if (brands.some((brand) => HEIF_BRANDS.has(brand))) return "image/heif";
-}
-/** Sniffs supported inline image formats from decoded bytes. */
-function sniffInlineImageMime(buffer) {
-	return IMAGE_SIGNATURES.find((signature) => signature.matches(buffer))?.mime ?? sniffIsoBmffImageMime(buffer);
-}
-function isImageMimeType(value) {
-	return value.trim().toLowerCase().startsWith("image/");
-}
-/** Canonicalizes trusted inline image base64 and rejects malformed or non-image payloads. */
-function sanitizeInlineImageBase64(params) {
-	if (!isImageMimeType(params.mimeType)) return;
-	const canonicalPayload = canonicalizeBase64(params.base64);
-	if (!canonicalPayload) return;
-	const sniffedMimeType = sniffInlineImageMime(Buffer.from(canonicalPayload.slice(0, IMAGE_SIGNATURE_PREFIX_BASE64_CHARS), "base64"));
-	if (!sniffedMimeType) return;
-	return {
-		mimeType: sniffedMimeType,
-		base64: canonicalPayload
-	};
-}
-function parseInlineImageDataUrl(value) {
-	if (!startsWithDataUrl(value)) return {
-		metadata: [],
-		payload: value
-	};
-	const commaIndex = value.indexOf(",");
-	if (commaIndex < 0) return;
-	return {
-		metadata: value.slice(5, commaIndex).split(";").map((part) => part.trim()),
-		payload: value.slice(commaIndex + 1)
-	};
-}
-function metadataAllowsImageBase64(metadata) {
-	const [mimeType, ...options] = metadata;
-	return mimeType !== void 0 && isImageMimeType(mimeType) && options.some((part) => part.toLowerCase() === "base64");
-}
-function sanitizeInlineImageDataUrlWithAllowedMimes(imageUrl, allowedMimes) {
-	const parsed = parseInlineImageDataUrl(imageUrl);
-	if (!parsed) return;
-	if (parsed.metadata.length === 0) return imageUrl;
-	if (!metadataAllowsImageBase64(parsed.metadata)) return;
-	const [mimeType] = parsed.metadata;
-	const sanitized = sanitizeInlineImageBase64({
-		mimeType: mimeType ?? "",
-		base64: parsed.payload
-	});
-	if (!sanitized) return;
-	if (allowedMimes && !allowedMimes.has(sanitized.mimeType)) return;
-	return `data:${sanitized.mimeType};base64,${sanitized.base64}`;
-}
-/** Canonicalizes provider-safe inline image data URLs and rejects unsupported formats. */
-function sanitizeInlineImageDataUrl(imageUrl) {
-	return sanitizeInlineImageDataUrlWithAllowedMimes(imageUrl, INLINE_IMAGE_DATA_URL_MIMES);
-}
-//#endregion
-//#region packages/ai/src/transports/responses-image-payload-sanitizer.ts
-/**
-* Sanitizes OpenAI Responses payloads before transport. Invalid inline images
-* are replaced with text placeholders so the request remains valid and
-* auditable.
-*/
-const IMAGE_OMITTED_TEXT = "omitted image payload: invalid inline image data";
-function invalidSnakeImage() {
-	return {
-		type: "input_text",
-		text: `[${IMAGE_OMITTED_TEXT}]`
-	};
-}
-function sanitizeValue(value) {
-	if (Array.isArray(value)) return value.map(sanitizeValue);
-	if (!isRecord(value)) return value;
-	if (value.type === "input_image" && typeof value.image_url === "string") {
-		const imageUrl = sanitizeInlineImageDataUrl(value.image_url);
-		return imageUrl ? {
-			...value,
-			image_url: imageUrl
-		} : invalidSnakeImage();
-	}
-	const next = {};
-	for (const [key, child] of Object.entries(value)) next[key] = sanitizeValue(child);
-	return next;
-}
-/** Sanitize inline image fields inside a Responses API payload. */
-function sanitizeResponsesImagePayload(params) {
-	if (!Array.isArray(params.input)) return params;
-	return {
-		...params,
-		input: sanitizeValue(params.input)
-	};
-}
-//#endregion
 //#region packages/ai/src/transports/openai-responses-client.ts
 function resolveNativeOpenAIResponsesWebSocketMode(model, transport) {
 	if (transport !== "websocket" && transport !== "websocket-cached" && transport !== "auto") return;
@@ -2648,37 +2519,35 @@ function resolveProviderTransportTurnState(model, params) {
 		}
 	});
 }
-function createOpenAIResponsesClient(model, context, apiKey, optionHeaders, turnHeaders, sessionId) {
+function createOpenAIResponsesClient(model, context, apiKey, optionHeaders, turnHeaders, sessionId, fetchOverride) {
 	return new OpenAI({
 		apiKey,
 		baseURL: resolveOpenAIClientBaseUrl(model),
 		dangerouslyAllowBrowser: true,
 		defaultHeaders: buildOpenAIClientHeaders(model, context, optionHeaders, turnHeaders, sessionId),
-		fetch: buildGuardedModelFetch(model),
+		fetch: fetchOverride ?? buildGuardedModelFetch(model),
 		...buildOpenAISdkClientOptions(model)
 	});
 }
 async function postOpenAIResponsesCompaction(params) {
+	const compactInput = typeof params.request.instructions === "string" && params.request.instructions.length > 0 ? [buildOpenAIResponsesCompactSystemMessage(params.model, params.request.instructions), ...params.request.input ?? []] : params.request.input;
 	const response = await params.client.post("/responses/compact", {
-		...buildOpenAISdkRequestOptions(params.model, params.options?.signal, {
-			timeoutMs: params.options?.timeoutMs,
-			maxRetries: params.options?.maxRetries
-		}),
+		...buildOpenAISdkRequestOptions(params.model, params.options?.signal, { timeoutMs: params.options?.timeoutMs }),
 		body: {
 			model: params.request.model,
-			input: params.request.input
+			input: compactInput
 		}
 	});
 	const output = isRecord(response) && Array.isArray(response.output) ? response.output : [];
 	const item = output.at(-1);
 	const retainedItems = output.slice(0, -1);
-	const retainedMessagesAreValid = retainedItems.every((candidate) => isRecord(candidate) && candidate.type === "message" && (candidate.role === "user" || candidate.role === "developer" || candidate.role === "system") && Array.isArray(candidate.content));
 	const retainedUserMessageCount = retainedItems.filter((candidate) => isRecord(candidate) && candidate.type === "message" && candidate.role === "user" && Array.isArray(candidate.content)).length;
 	const inputUserMessageCount = Array.isArray(params.request.input) ? params.request.input.filter((candidate) => isRecord(candidate) && candidate.type === "message" && candidate.role === "user").length : 0;
 	const retainedMessagePrefixSupported = supportsNativeOpenAIResponsesEndpoint(params.model);
 	const usage = isRecord(response) && isRecord(response.usage) ? response.usage : void 0;
-	if (!isRecord(response) || response.object !== "response.compaction" || !retainedMessagesAreValid || retainedItems.length > 0 && (!retainedMessagePrefixSupported || retainedUserMessageCount !== inputUserMessageCount) || !isRecord(item) || item.type !== "compaction" || typeof item.encrypted_content !== "string" || item.encrypted_content.length === 0 || !usage || typeof usage.input_tokens !== "number" || typeof usage.output_tokens !== "number") throw new Error("Responses compact endpoint did not return one trailing compaction item");
+	if (!isRecord(response) || response.object !== "response.compaction" || !isOpenAIResponsesCompactionOutput(output, params.model) || retainedItems.length > 0 && (!retainedMessagePrefixSupported || retainedUserMessageCount !== inputUserMessageCount) || !isRecord(item) || item.type !== "compaction" || typeof item.encrypted_content !== "string" || item.encrypted_content.length === 0 || !usage || typeof usage.input_tokens !== "number" || typeof usage.output_tokens !== "number") throw new Error("Responses compact endpoint did not return one trailing compaction item");
 	return {
+		output,
 		item,
 		historyMode: retainedUserMessageCount > 0 ? "retained-users" : "compacted-prefix",
 		usage,
@@ -2709,7 +2578,7 @@ function createResponsesTransportExecutor(config) {
 				});
 				const websocketSessionPolicy = websocketMode ? turnState?.websocket : void 0;
 				const websocketHeaders = websocketMode ? buildOpenAIClientHeaders(model, context, options?.headers, websocketSessionPolicy?.headers, options?.sessionId) : void 0;
-				const client = config.createClient(model, context, apiKey, options?.headers, turnState?.headers, options?.sessionId);
+				const client = config.createClient(model, context, apiKey, options?.headers, turnState?.headers, options?.sessionId, compactRequest ? createBoundedOpenAIResponsesCompactionFetch(buildGuardedModelFetch(model)) : void 0);
 				const buildRequest = async (replayMode) => {
 					let params = config.buildRequest(model, context, responsesOptions, turnState?.metadata, replayMode);
 					const nextParams = await options?.onPayload?.(params, model);
@@ -2772,14 +2641,13 @@ function createResponsesTransportExecutor(config) {
 				firstEventAbort = firstEvent;
 				const requestOptions = buildOpenAISdkRequestOptions(model, firstEvent.signal, {
 					stream: config.streamRequest,
-					timeoutMs: options?.timeoutMs,
-					maxRetries: options?.maxRetries
+					timeoutMs: options?.timeoutMs
 				});
 				const websocketSignal = combineWebSocketTimeoutSignal(firstEvent.signal, model, requestOptions?.timeout);
 				emitModelTransportDebug(log, `[responses] start provider=${model.provider} api=${model.api} model=${model.id} requestIdHash=${redactIdentifier(options?.requestId, { len: 64 })} baseUrl=${formatModelTransportDebugBaseUrl(model.baseUrl)} timeoutMs=${safeDebugValue(requestOptions?.timeout)} apiKey=${apiKey ? "present" : "missing"} ${summarizeResponsesPayload(params)}`);
 				let continuationBaseline;
 				const createSseStream = async (initialRequest = continuationClaim?.request ?? params, initialAttemptKind = "initial", initialRejectedCompaction) => {
-					const { stream: responseStream } = await config.createResponseStream({
+					const { stream: responseStream } = await createResponsesStreamWithEncryptedContentRetry({
 						client,
 						request: initialRequest,
 						requestOptions,
@@ -2933,7 +2801,6 @@ function createOpenAIResponsesTransportStreamFn() {
 		httpContinuation: true,
 		createClient: createOpenAIResponsesClient,
 		buildRequest: buildOpenAIResponsesParams,
-		createResponseStream: createResponsesStreamWithEncryptedContentRetry,
 		pricingOptions: (options, model) => ({
 			serviceTier: options?.serviceTier,
 			applyServiceTierPricing: (usage, serviceTier) => applyResponsesServiceTierPricing(usage, serviceTier, model)
@@ -2945,27 +2812,23 @@ function createAzureOpenAIResponsesTransportStreamFn() {
 		outputApi: "azure-openai-responses",
 		firstEventTimeoutMs: AZURE_RESPONSES_FIRST_EVENT_TIMEOUT_MS,
 		createClient: createAzureOpenAIClient,
-		buildRequest: (model, context, options, metadata, replayMode) => buildAzureOpenAIResponsesParams(model, context, options, resolveAzureDeploymentName(model), metadata, replayMode),
-		createResponseStream: createResponsesStreamWithEncryptedContentRetry
+		buildRequest: (model, context, options, metadata, replayMode) => buildAzureOpenAIResponsesParams(model, context, options, resolveAzureDeploymentName(model), metadata, replayMode)
 	});
 }
-function normalizeAzureBaseUrl(baseUrl) {
-	return baseUrl.replace(/\/+$/, "");
-}
 function resolveAzureDeploymentName(model) {
 	return resolveAzureDeploymentNameFromMap({
 		modelId: model.id,
 		deploymentMap: process.env.AZURE_OPENAI_DEPLOYMENT_NAME_MAP
 	});
 }
-function createAzureOpenAIClient(model, context, apiKey, optionHeaders, turnHeaders) {
-	const baseURL = normalizeAzureBaseUrl(model.baseUrl);
+function createAzureOpenAIClient(model, context, apiKey, optionHeaders, turnHeaders, _sessionId, fetchOverride) {
+	const baseURL = model.baseUrl.replace(/\/+$/, "");
 	const clientOptions = {
 		apiKey,
 		dangerouslyAllowBrowser: true,
 		defaultHeaders: buildOpenAIClientHeaders(model, context, optionHeaders, turnHeaders),
 		baseURL,
-		fetch: buildGuardedModelFetch(model),
+		fetch: fetchOverride ?? buildGuardedModelFetch(model),
 		...buildOpenAISdkClientOptions(model)
 	};
 	if (isOpenAICompatibleAzureResponsesBaseUrl(baseURL)) return new OpenAI(clientOptions);
@@ -3011,6 +2874,112 @@ const responsesTesting = {
 if (process.env.VITEST || false) globalThis.openclawOpenAIResponsesTransportTestApi = responsesTesting;
 //#endregion
 //#region packages/ai/src/transports/provider-compaction-replay.ts
+function isAssistantReplayMessage(message) {
+	return message.role === "assistant";
+}
+/** Resolve from prepared transport facts; never retain or return credential material. */
+function resolveCompactionReplayEligibility(model, options) {
+	if (OPENAI_RESPONSES_APIS.has(model.api)) return true;
+	return Boolean(options.apiKey?.trim()) && resolveAnthropicServerCompactionPlan(model, options.extraParams, options.apiKey).enabled;
+}
+function resolveCompactionSource(messages, model, identity) {
+	if (identity.enabled === false) return;
+	const assistants = messages.filter(isAssistantReplayMessage);
+	const selected = (() => {
+		if (OPENAI_RESPONSES_APIS.has(model.api)) {
+			const responses = resolveNewestOpenAIResponsesCompactionReplay(assistants, model, identity);
+			return responses ? {
+				family: "responses",
+				...responses
+			} : void 0;
+		}
+		if (identity.enabled !== true) return;
+		const anthropic = resolveNewestAnthropicCompaction(assistants, model, identity);
+		return anthropic ? {
+			family: "anthropic",
+			...anthropic
+		} : void 0;
+	})();
+	if (!selected) return;
+	const owner = assistants.find((message) => message === selected.owner);
+	return owner ? {
+		...selected,
+		owner
+	} : void 0;
+}
+/** Manual recovery uses the durable client compactor, never a guessed retained-user prefix. */
+function requiresCompactionReplayRefresh(messages, model, identity) {
+	const checkpoint = resolveCompactionSource(messages, model, identity);
+	return checkpoint?.family === "responses" && checkpoint.mode === "refresh-required";
+}
+/** Carry a checkpoint immediately after reference-preserving history limiting, before repair. */
+function preserveCompactionReplayWindow(source, windowed, model, identity) {
+	const checkpoint = resolveCompactionSource(source, model, identity);
+	if (!checkpoint || windowed.some((message) => message === checkpoint.owner)) return windowed;
+	const owner = checkpoint.owner;
+	const replay = checkpoint.owner.providerReplay;
+	const providerReplay = replay?.type === "openai-responses-retained-compaction" ? replay : {
+		...replay,
+		replayIndex: 0
+	};
+	const suffix = new Set(source.slice(source.indexOf(owner) + 1));
+	return [{
+		...owner,
+		content: [],
+		providerReplay
+	}, ...windowed.filter((message) => suffix.has(message))];
+}
+function estimateResponsesWindow(checkpoint, estimate) {
+	return checkpoint.output.reduce((tokens, entry) => {
+		if (entry.type === "compaction") {
+			const { encrypted_content, ...metadata } = entry;
+			return tokens + estimate.text(encrypted_content) + estimate.json(metadata);
+		}
+		const { content, ...metadata } = entry;
+		return tokens + estimate.json(metadata) + content.reduce((sum, block) => {
+			if (block.type === "input_text") {
+				const { text, ...fields } = block;
+				return sum + estimate.text(text) + estimate.json(fields);
+			}
+			if (block.type === "input_image") {
+				const { image_url: _imageUrl, ...fields } = block;
+				return sum + estimate.image() + estimate.json(block.image_url?.startsWith("data:") ? fields : block);
+			}
+			return sum + estimate.json(block);
+		}, 0);
+	}, 0);
+}
+/** Estimate the canonical prefix and its tail once, independent of unbound usage snapshots. */
+function resolveCompactionReplayPressure(messages, model, identity, estimate) {
+	const checkpoint = resolveCompactionSource(messages, model, identity);
+	if (!checkpoint) return;
+	if (checkpoint.family === "responses" && checkpoint.mode === "refresh-required") throw new CompactionReplayRefreshRequiredError();
+	const ownerIndex = messages.findIndex((message) => message === checkpoint.owner);
+	const owner = messages[ownerIndex];
+	if (!owner) return;
+	const prefixTokens = checkpoint.family === "anthropic" ? estimate.text(checkpoint.summary) : checkpoint.mode === "complete-window" ? estimateResponsesWindow(checkpoint, estimate) : estimate.text(checkpoint.item.encrypted_content);
+	const { contextUsage: _staleContextUsage, ...usage } = checkpoint.owner.usage;
+	const tail = [];
+	for (const message of messages.slice(ownerIndex + 1)) {
+		if (!isAssistantReplayMessage(message) || message.usage.contextUsage === void 0) {
+			tail.push(message);
+			continue;
+		}
+		const { contextUsage: _unboundContextUsage, ...billedUsage } = message.usage;
+		tail.push({
+			...message,
+			usage: billedUsage
+		});
+	}
+	return {
+		messages: [{
+			...owner,
+			content: checkpoint.owner.content.slice(checkpoint.replayIndex),
+			usage
+		}, ...tail],
+		prefixTokens
+	};
+}
 /** Whether provider replay state is a prefix-bound server compaction checkpoint. */
 function isCompactionReplayCheckpoint(replay) {
 	const type = replay && typeof replay === "object" ? replay.type : void 0;
@@ -3306,4 +3275,4 @@ function prepareModelForSimpleCompletion(params) {
 	return applyProviderSimpleCompletionWrapper(apiRegistry, model, cfg);
 }
 //#endregion
-export { GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE, applyAnthropicCacheControlToMessages, applyAnthropicEphemeralCacheControlMarkers, applyAnthropicPayloadPolicyToParams, applyOpenAIResponsesPayloadPolicy, assertCodeModeResponsesToolSurface, assignTransportErrorDetails, buildOpenAIClientHeaders, buildOpenAICompletionsParams, buildOpenAISdkClientOptions, buildOpenAISdkRequestOptions, buildTransportAwareSimpleStreamFn, canonicalizeMaxTokensParam, captureOpenAIResponsesCompaction, coerceTransportToolCallArguments, copyProviderAcceptanceObserver, createAnthropicMessagesTransportStreamFn, createAzureOpenAIResponsesTransportStreamFn, createBoundaryAwareStreamFnForModel, createDeepSeekTextFilter, createEmptyTransportUsage, createModelStreamCooperativeScheduler, createOpenAICompletionsTransportStreamFn, createOpenAIProviderAcceptanceHook, createOpenAIResponseHook, createOpenAIResponsesTransportStreamFn, createOpenClawTransportStreamFnForModel, createTransportAwareStreamFnForModel, createWritableTransportEventStream, detectOpenAICompletionsCompat, emitModelTransportDebug, enforceCodeModeResponsesToolSurface, failTransportStream, filterCodeModePayloadTools, finalizeTerminalToolCallArguments, finalizeTransportStream, flattenCompletionMessagesToStringContent, formatModelTransportDebugBaseUrl, formatModelTransportDebugUrl, getCompat, googleFlashSupportsMinimalThinking, hasOpenAICompatibleConversationTurn, isCodeModeModelVisibleToolName, isCompactionReplayCheckpoint, isOpenAICodexResponsesModel, isOpenAICompletionsThinkingEnabled, log, measureUtf8AppendBytes, mergeTransportHeaders, mergeTransportMetadata, normalizeCodexResponsesBaseUrlForOpenAISdk, notifyProviderHttpMetadata, notifyProviderHttpResponse, notifyProviderStreamOpened, parseJsonObjectPreservingUnsafeIntegers, parseJsonPreservingUnsafeIntegers, parseOpenAICompletionsUsage, parseTerminalToolCallArguments, prepareModelForSimpleCompletion, prepareTransportAwareSimpleModel, quoteUnsafeIntegerLiterals, readCodeModePayloadToolName, readOpenAICompletionsContentDeltas, readOpenAICompletionsReasoningBatch, replaceCompactionReplayOwnerContent, requestPreparedOpenAIResponsesCompaction, resolveAnthropicEphemeralCacheControl, resolveAnthropicMessagesUrl, resolveAnthropicPayloadPolicy, resolveAnthropicServerCompactionPlan, resolveCodeModeResponsesVisibleToolNames, resolveMaxTokensParam, resolveModelPayloadDebugMode, resolveModelSseDebugMode, resolveOpenAIClientBaseUrl, resolveOpenAICompletionsCompat, resolveOpenAIReasoningEffortMap, resolveOpenAIResponsesCompactEndpointPlan, resolveOpenAIResponsesPayloadPolicy, resolveOpenAIResponsesServerCompactionPlan, resolveOpenAIStrictToolFlagWithDiagnostics, resolvePromptCacheKey, resolveReplayableResponsesMessageId, resolveTransportAwareSimpleApi, sanitizeNonEmptyTransportPayloadText, sanitizeResponsesImagePayload, sanitizeTransportPayloadText, sortPromptCacheToolsByName as sortTransportToolsByName, stripCompactionReplayCheckpoint, stripCompactionReplayCheckpointInPlace, stripCompletionMessagesToRoleContent, throwIfModelStreamAborted, transportAbortError, usesNativeOpenAICodexResponsesBackend, withProviderAcceptanceObserver, withProviderResponseHook };
+export { CompactionReplayRefreshRequiredError, GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE, applyAnthropicCacheControlToMessages, applyAnthropicEphemeralCacheControlMarkers, applyAnthropicPayloadPolicyToParams, applyOpenAIResponsesPayloadPolicy, assertCodeModeResponsesToolSurface, assignTransportErrorDetails, buildOpenAIClientHeaders, buildOpenAICompletionsParams, buildOpenAISdkClientOptions, buildOpenAISdkRequestOptions, buildTransportAwareSimpleStreamFn, canonicalizeMaxTokensParam, captureOpenAIResponsesCompaction, coerceTransportToolCallArguments, copyProviderAcceptanceObserver, createAnthropicMessagesTransportStreamFn, createAzureOpenAIResponsesTransportStreamFn, createBoundaryAwareStreamFnForModel, createDeepSeekTextFilter, createEmptyTransportUsage, createModelStreamCooperativeScheduler, createOpenAICompletionsTransportStreamFn, createOpenAIProviderAcceptanceHook, createOpenAIResponseHook, createOpenAIResponsesTransportStreamFn, createOpenClawTransportStreamFnForModel, createTransportAwareStreamFnForModel, createWritableTransportEventStream, detectOpenAICompletionsCompat, emitModelTransportDebug, enforceCodeModeResponsesToolSurface, failTransportStream, filterCodeModePayloadTools, finalizeTerminalToolCallArguments, finalizeTransportStream, flattenCompletionMessagesToStringContent, formatModelTransportDebugBaseUrl, formatModelTransportDebugUrl, getCompat, googleFlashSupportsMinimalThinking, hasOpenAICompatibleConversationTurn, isCodeModeModelVisibleToolName, isCompactionReplayCheckpoint, isOpenAICodexResponsesModel, isOpenAICompletionsThinkingEnabled, log, measureUtf8AppendBytes, mergeTransportHeaders, mergeTransportMetadata, normalizeCodexResponsesBaseUrlForOpenAISdk, notifyProviderHttpMetadata, notifyProviderHttpResponse, notifyProviderStreamOpened, parseJsonObjectPreservingUnsafeIntegers, parseJsonPreservingUnsafeIntegers, parseOpenAICompletionsUsage, parseTerminalToolCallArguments, prepareModelForSimpleCompletion, prepareTransportAwareSimpleModel, preserveCompactionReplayWindow, quoteUnsafeIntegerLiterals, readCodeModePayloadToolName, readOpenAICompletionsContentDeltas, readOpenAICompletionsReasoningBatch, replaceCompactionReplayOwnerContent, requestPreparedOpenAIResponsesCompaction, requiresCompactionReplayRefresh, resolveAnthropicEphemeralCacheControl, resolveAnthropicMessagesUrl, resolveAnthropicPayloadPolicy, resolveAnthropicServerCompactionPlan, resolveCodeModeResponsesVisibleToolNames, resolveCompactionReplayEligibility, resolveCompactionReplayPressure, resolveMaxTokensParam, resolveModelPayloadDebugMode, resolveModelSseDebugMode, resolveOpenAIClientBaseUrl, resolveOpenAICompletionsCompat, resolveOpenAIReasoningEffortMap, resolveOpenAIResponsesCompactEndpointPlan, resolveOpenAIResponsesPayloadPolicy, resolveOpenAIResponsesServerCompactionPlan, resolveOpenAIStrictToolFlagWithDiagnostics, resolvePromptCacheKey, resolveReplayableResponsesMessageId, resolveTransportAwareSimpleApi, sanitizeNonEmptyTransportPayloadText, sanitizeResponsesImagePayload, sanitizeTransportPayloadText, sortPromptCacheToolsByName as sortTransportToolsByName, stripCompactionReplayCheckpoint, stripCompactionReplayCheckpointInPlace, stripCompletionMessagesToRoleContent, throwIfModelStreamAborted, transportAbortError, usesNativeOpenAICodexResponsesBackend, withProviderAcceptanceObserver, withProviderResponseHook };