@openclaw/ai 2026.8.1-beta.2 → 2026.8.1-beta.3

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 "./anthropic-JsNA5KCu.mjs";
-import { r as createAssistantMessageEventStream } from "./event-stream-uSMZJ3FA.mjs";
+import { d as resolveClaudeSonnet5ModelIdentity, g as supportsClaudeNativeXhighEffort, p as supportsClaudeAdaptiveThinking, u as resolveClaudeOpus5ModelIdentity } from "./anthropic-BQXQudmA.mjs";
 import { n as normalizeLowercaseStringOrEmpty, t as hasNonEmptyString } from "./string-coerce-fsri9iCu.mjs";
-import { r as calculateCost } from "./sanitize-unicode-BYqrYtC_.mjs";
-import { _ as resolveAnthropicThinkingEffort, a as describeToolResultMediaPlaceholder, b as usesClaudeStreamingRefusalContract, d as ANTHROPIC_CLAUDE_CODE_VERSION, f as applyClaudeRequestContract, g as requiresClaudeAdaptiveThinking, h as prepareClaudeNoPrefillRequestContext, l as isImageWithMediaPayload, m as mapAnthropicStopReason, n as getAiTransportHost, o as extractToolResultBlockText, p as defaultsClaudeAdaptiveThinking, r as resolveAiTransportHeaderSentinels, s as extractToolResultText, u as ANTHROPIC_CLAUDE_CODE_BILLING_SYSTEM_BLOCK, y as usesClaudeFable5MessagesContract } from "./host-DTqNc7ad.mjs";
+import { r as calculateCost } from "./sanitize-unicode-BP7HLLZn.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-DF_yxOsA.mjs";
 import { a as isRecord } from "./record-coerce-DdXsgUd_.mjs";
-import { i as canonicalizeBase64, n as projectProviderError, o as stableStringify } from "./provider-error-BUwEnjXq.mjs";
-import { n as toErrorObject, t as createResponsesPromptEgressObserver } from "./openai-responses-prompt-observer-internal-DgNTYnRY.mjs";
+import { i as canonicalizeBase64, o as stableStringify } from "./provider-error-C8mBRJDl.mjs";
+import { n as toErrorObject, t as createResponsesPromptEgressObserver } from "./openai-responses-prompt-observer-internal-BG-1AsBI.mjs";
 import { r as asNonNegativeFiniteNumber } from "./number-coercion-H9qHik3g.mjs";
-import { _ as isOpenAIGpt56Model, c as OpenAIResponsesWebSocketPostDispatchError, d as OpenAIResponsesWebSocketSafeRetryError, g as isOpenAIGpt55Model, h as isOpenAIGpt54MiniModel, i as OPENAI_RESPONSES_APIS, l as OpenAIResponsesWebSocketPreDispatchError, p as parseOpenAIResponsesWebSocketServerError, r as OPENAI_CODEX_RESPONSES_DEFAULT_INSTRUCTIONS, t as AZURE_RESPONSES_FIRST_EVENT_TIMEOUT_MS, u as OpenAIResponsesWebSocketResponseFailedError, v as normalizeOpenAIReasoningEffort, w as uniqueStrings, y as resolveOpenAIReasoningEffortForModel } from "./openai-responses-contracts-CyfIkQi5.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, w as uniqueStrings, y as resolveOpenAIReasoningEffortForModel } from "./openai-responses-contracts-DvILoqBk.mjs";
 import { n as clampOpenAIPromptCacheKey } from "./openai-prompt-cache-mZTCdRPo.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
-import { f as sortPromptCacheToolsByName, l as stripSystemPromptCacheBoundary, t as adjustMaxTokensForThinking } from "./simple-options-D58D5Kvw.mjs";
-import { a as resolveProviderEndpoint, c as transformTransportMessages, i as resolveOpenAIStrictToolSetting, n as buildGuardedModelFetch, r as resolveModelRequestTimeoutMs, s as resolveProviderRequestPolicyConfig } from "./tool-schema-json-projection-q5d7QX5c.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-8lJNKXOE.mjs";
-import { a as createOpenAICompletionsToolCallDeltaNormalizer, c as resolveOpenAICompletionsCompat, d as clearPendingCommentaryText, f as rememberPendingCommentaryTags, i as hasToolCallHistory, l as resolveOpenAICompletionsResponseFormat, n as resolveOpenAIReasoningEffortMap, o as finalizeOpenAICompletionsToolCalls, p as tagPendingCommentaryText, r as convertMessages, s as detectOpenAICompletionsCompat, t as mapOpenAIStopReason, u as shouldOmitOllamaCompatResponseFormat } from "./openai-stop-reason-BkFkqqK0.mjs";
+import { C as withProviderResponseHook, E as quoteUnsafeIntegerLiterals, P as sortPromptCacheToolsByName, S as withProviderAcceptanceObserver, T as parseJsonPreservingUnsafeIntegers, _ as notifyProviderStreamOpened, a as assignTransportErrorDetails, b as sanitizeTransportPayloadText, c as createEmptyTransportUsage, d as finalizeTerminalToolCallArguments, f as finalizeTransportStream, g as notifyProviderHttpResponse, h as notifyProviderHttpMetadata, j as stripSystemPromptCacheBoundary, l as createWritableTransportEventStream, m as mergeTransportMetadata, o as coerceTransportToolCallArguments, p as mergeTransportHeaders, s as copyProviderAcceptanceObserver, t as adjustMaxTokensForThinking, u as failTransportStream, v as parseTerminalToolCallArguments, w as parseJsonObjectPreservingUnsafeIntegers, x as transportAbortError, y as sanitizeNonEmptyTransportPayloadText } from "./simple-options-oqsXB6U3.mjs";
+import { a as resolveProviderEndpoint, c as transformTransportMessages, i as resolveOpenAIStrictToolSetting, n as buildGuardedModelFetch, r as resolveModelRequestTimeoutMs, s as resolveProviderRequestPolicyConfig } from "./tool-schema-json-projection-FEr-jOoZ.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-4Zx7-aCr.mjs";
+import { a as createOpenAICompletionsToolCallDeltaNormalizer, c as resolveOpenAICompletionsCompat, d as clearPendingCommentaryText, f as rememberPendingCommentaryTags, h as tagUnresolvedTextAsCommentary, i as hasToolCallHistory, l as resolveOpenAICompletionsResponseFormat, m as tagPendingCommentaryText, n as resolveOpenAIReasoningEffortMap, o as finalizeOpenAICompletionsToolCalls, p as tagInterruptedTextPhases, r as convertMessages, s as detectOpenAICompletionsCompat, t as mapOpenAIStopReason, u as shouldOmitOllamaCompatResponseFormat } from "./openai-stop-reason-9OONKyAm.mjs";
 import { t as createDeferredEventBuffer } from "./deferred-event-buffer-DAvyP7qA.mjs";
 import { n as parseStreamingJson } from "./json-parse-CDnesDM_.mjs";
-import { n as notifyLlmRequestActivity } from "./headers-DdOQtGuU.mjs";
-import { B as buildResponsesInputMessage, C as summarizeResponsesTools, G as suppressOpenAIResponsesCompaction, H as createOpenAIResponsesAssistantOutput, I as createResponsesStreamWithEncryptedContentRetry, K as resolveReplayableResponsesMessageId, L as isInvalidEncryptedContentError, R as resolveAzureOpenAIApiVersion, S as summarizeResponsesPayload, U as buildOpenAIResponsesReasoningReplayMetadata, V as convertResponsesMessages, W as captureOpenAIResponsesCompaction, Y as normalizeOpenAIStrictToolParameters, Z as resolveOpenAIProjectedToolsStrictToolFlag, _ as safeDebugValue, b as summarizeOpenAITransportError, c as processResponsesStream, dt as resolveModelPayloadDebugMode, f as observeResponsesStream, ft as resolveModelSseDebugMode, g as normalizeResponsesFailedEvent, h as logResponsesFailedNoDetails, ht as quoteUnsafeIntegerLiterals, m as buildResponsesFailedNoDetailsObservation, mt as parseJsonPreservingUnsafeIntegers, n as applyResponsesServiceTierPricing, p as ResponsesStreamFailure, pt as parseJsonObjectPreservingUnsafeIntegers, q as findOpenAIStrictToolProjectionDiagnostics, ut as emitModelTransportDebug, v as stringifyRedactedEvent, x as summarizeResponsesFailedNoDetailsObservation, y as stringifyRedactedPayload, z as resolveNextResponsesEncryptedContentAttempt } from "./openai-responses-shared-DXIt3iY5.mjs";
-import { a as createWritableTransportEventStream, c as mergeTransportHeaders, d as sanitizeTransportPayloadText, f as transportAbortError, i as createEmptyTransportUsage, l as mergeTransportMetadata, n as assignTransportErrorDetails, o as failTransportStream, p as withProviderResponseHook, r as coerceTransportToolCallArguments, s as finalizeTransportStream, u as sanitizeNonEmptyTransportPayloadText } from "./provider-transcript-transform-ePx-Bbfr.mjs";
-import { a as isGoogleGemini3FlashModel, c as readResponseTextSnippet, d as resolveModelHeaderSentinels$1, f as resolveSecretSentinel, i as isCodeModeModelVisibleToolName, l as redactIdentifier, m as supportsModelTools, n as createAbortError$1, o as isGoogleGemini3ProModel, p as sha256Hex, r as estimateStringChars, s as parseRetryAfterSeconds, t as MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE } from "./transport-utils-DJqkxbhC.mjs";
+import { t as notifyLlmRequestActivity } from "./llm-request-activity-BjtkplhG.mjs";
+import { a as isGoogleGemini3FlashModel, c as readResponseTextSnippet, d as resolveModelHeaderSentinels$1, f as resolveSecretSentinel, i as isCodeModeModelVisibleToolName, l as redactIdentifier, m as supportsModelTools, n as createAbortError$1, o as isGoogleGemini3ProModel, p as sha256Hex, r as estimateStringChars, s as parseRetryAfterSeconds, t as MALFORMED_STREAMING_FRAGMENT_ERROR_MESSAGE } from "./transport-utils-Bv0FHkd0.mjs";
 import { t as googleFlashSupportsMinimalThinking } from "./google-thinking-level-C-V3tecN.mjs";
-import { a as createModelStreamCooperativeScheduler, c as log, d as readOpenAICompletionsContentDeltas, f as resolvePromptCacheKey, i as GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, l as measureUtf8AppendBytes, n as reconcileOpenAICompletionsToolChoice, o as createOpenAIResponseHook, p as throwIfModelStreamAborted, r as reconcileOpenAIResponsesToolChoice, s as isOpenAICompletionsThinkingEnabled, t as projectOpenAITools, u as parseOpenAICompletionsUsage } from "./openai-tool-projection-CY04OcvQ.mjs";
+import { B as buildResponsesInputMessage, C as summarizeResponsesTools, G as suppressOpenAIResponsesCompaction, H as createOpenAIResponsesAssistantOutput, I as createResponsesStreamWithEncryptedContentRetry, K as resolveReplayableResponsesMessageId, L as isInvalidEncryptedContentError, R as resolveAzureOpenAIApiVersion, S as summarizeResponsesPayload, U as buildOpenAIResponsesReasoningReplayMetadata, V as convertResponsesMessages, W as captureOpenAIResponsesCompaction, Y as normalizeOpenAIStrictToolParameters, Z as resolveOpenAIProjectedToolsStrictToolFlag, _ as safeDebugValue, b as summarizeOpenAITransportError, c as processResponsesStream, dt as resolveModelPayloadDebugMode, f as observeResponsesStream, ft as resolveModelSseDebugMode, g as normalizeResponsesFailedEvent, h as logResponsesFailedNoDetails, m as buildResponsesFailedNoDetailsObservation, n as applyResponsesServiceTierPricing, p as ResponsesStreamFailure, q as findOpenAIStrictToolProjectionDiagnostics, ut as emitModelTransportDebug, v as stringifyRedactedEvent, x as summarizeResponsesFailedNoDetailsObservation, y as stringifyRedactedPayload, z as resolveNextResponsesEncryptedContentAttempt } from "./openai-responses-shared-Cvt5MgEm.mjs";
+import { t as codeModeToolSurfaceObserver } from "./provider-options-AvldWZt8.mjs";
+import { a as createModelStreamCooperativeScheduler, c as isOpenAICompletionsThinkingEnabled, d as parseOpenAICompletionsUsage, f as readOpenAICompletionsContentDeltas, h as throwIfModelStreamAborted, i as GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, l as log, m as resolvePromptCacheKey, n as reconcileOpenAICompletionsToolChoice, o as createOpenAIProviderAcceptanceHook, p as readOpenAICompletionsReasoningBatch, r as reconcileOpenAIResponsesToolChoice, s as createOpenAIResponseHook, t as projectOpenAITools, u as measureUtf8AppendBytes } from "./openai-tool-projection-CHGLt03m.mjs";
 import { a as withFirstStreamEventTimeout, i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-MK28puvq.mjs";
-import { t as createReasoningTagTextPartitioner } from "./reasoning-tag-text-partitioner-rnPwX2pg.mjs";
-import { i as resolveOpenAIResponsesServerCompactionPlan, n as resolveOpenAIResponsesCompactEndpointPlan, r as resolveOpenAIResponsesPayloadPolicy, t as applyOpenAIResponsesPayloadPolicy } from "./openai-responses-payload-policy-BDxV-W0c.mjs";
+import { t as createReasoningTagTextPartitioner } from "./reasoning-tag-text-partitioner-5ygO2rZc.mjs";
+import { i as resolveOpenAIResponsesServerCompactionPlan, n as resolveOpenAIResponsesCompactEndpointPlan, r as resolveOpenAIResponsesPayloadPolicy, t as applyOpenAIResponsesPayloadPolicy } from "./openai-responses-payload-policy-BDo8CQsx.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 { randomUUID } from "node:crypto";
@@ -67,7 +66,7 @@ function isAnthropicOAuthToken(apiKey) {
 }
 function isDirectAnthropicModel(model) {
 	if (normalizeLowercaseStringOrEmpty(model.provider) !== "anthropic") return false;
-	const endpointClass = resolveProviderEndpoint(model.baseUrl).endpointClass;
+	const endpointClass = resolveProviderEndpoint(model).endpointClass;
 	return endpointClass === "default" || endpointClass === "anthropic-public";
 }
 function isKimiAnthropicProvider(provider) {
@@ -82,7 +81,7 @@ function useAnthropicServerSideFallback(model) {
 	return (usesClaudeFable5MessagesContract(model) || resolveClaudeOpus5ModelIdentity(model) !== void 0) && isDirectAnthropicModel(model);
 }
 function supportsReasoningContentReplay(model) {
-	return resolveProviderEndpoint(model.baseUrl).endpointClass === "xiaomi-native";
+	return resolveProviderEndpoint(model).endpointClass === "xiaomi-native";
 }
 function buildAnthropicBetaHeader(model, betaFeatures, params) {
 	if (!isDirectAnthropicModel(model)) return;
@@ -304,7 +303,7 @@ function convertAnthropicTools(tools, isOAuthToken) {
 	};
 }
 function parseAnthropicToolCallArguments(inputJson) {
-	return parseJsonObjectPreservingUnsafeIntegers(inputJson) ?? parseStreamingJson(inputJson);
+	return coerceTransportToolCallArguments(parseJsonObjectPreservingUnsafeIntegers(inputJson) ?? parseStreamingJson(inputJson));
 }
 const DEFAULT_ANTHROPIC_BASE_URL = "https://api.anthropic.com";
 /** Resolve the effective Anthropic API base URL from model or environment. */
@@ -407,7 +406,7 @@ async function* parseAnthropicSseBody(body, signal) {
 }
 function createAnthropicMessagesClient(params) {
 	const url = resolveAnthropicMessagesUrl(params.baseURL);
-	return { messages: { async *stream(body, options) {
+	return { messages: { async stream(body, options) {
 		const headers = mergeTransportHeaders({
 			"content-type": "application/json",
 			"anthropic-version": "2023-06-01",
@@ -420,12 +419,10 @@ function createAnthropicMessagesClient(params) {
 			body: JSON.stringify(body),
 			signal: options?.signal
 		});
-		if (!response.ok) {
-			const detail = await readAnthropicMessagesErrorBodySnippet(response);
-			throw new Error(formatAnthropicMessagesHttpError(response, detail));
-		}
-		if (!response.body) return;
-		yield* parseAnthropicSseBody(response.body, options?.signal);
+		return {
+			response,
+			stream: response.body ? parseAnthropicSseBody(response.body, options?.signal) : []
+		};
 	} } };
 }
 function formatAnthropicMessagesHttpError(response, detail) {
@@ -536,7 +533,7 @@ async function buildAnthropicParams(model, context, isOAuthToken, options) {
 		baseUrl: model.baseUrl,
 		cacheRetention: options?.cacheRetention,
 		enableCacheControl: true
-	});
+	}, model);
 	const cacheBreakpointOptOutMessageIndexes = /* @__PURE__ */ new Set();
 	const replayPlan = buildAnthropicReplayPlan(context.messages, model, {
 		enabled: !isOAuthToken && options?.anthropicServerCompaction === true,
@@ -620,7 +617,7 @@ function resolveAnthropicTransportOptions(model, options, apiKey) {
 	const reasoningModelMaxTokens = resolvePositiveAnthropicTokenLimit(model.maxTokens) ?? baseMaxTokens;
 	const mandatoryAdaptiveThinking = requiresClaudeAdaptiveThinking(model);
 	const reasoning = options?.reasoning === "off" && mandatoryAdaptiveThinking ? "low" : options?.reasoning;
-	const resolved = {
+	const resolved = copyProviderAcceptanceObserver(options, {
 		temperature: options?.temperature,
 		stop: options?.stop,
 		maxTokens: baseMaxTokens,
@@ -630,6 +627,7 @@ function resolveAnthropicTransportOptions(model, options, apiKey) {
 		sessionId: options?.sessionId,
 		headers: options?.headers,
 		onPayload: options?.onPayload,
+		onResponse: options?.onResponse,
 		maxRetryDelayMs: options?.maxRetryDelayMs,
 		metadata: options?.metadata,
 		interleavedThinking: options?.interleavedThinking,
@@ -638,7 +636,7 @@ function resolveAnthropicTransportOptions(model, options, apiKey) {
 		reasoning,
 		...options?.anthropicServerCompaction === true ? { anthropicServerCompaction: true } : {},
 		...options?.authProfileId ? { authProfileId: options.authProfileId } : {}
-	};
+	});
 	if (reasoning === "off") {
 		resolved.thinkingEnabled = false;
 		return resolved;
@@ -700,12 +698,22 @@ function createAnthropicMessagesTransportStreamFn() {
 				const nextParams = await transportOptions.onPayload?.(params, model);
 				if (nextParams !== void 0) params = nextParams;
 				applyClaudeRequestContract(params, model);
-				const anthropicStream = client.messages.stream({
+				const { response, stream: anthropicStream } = await client.messages.stream({
 					...params,
 					stream: true
 				}, transportOptions.signal ? { signal: transportOptions.signal } : void 0);
+				await notifyProviderHttpResponse({
+					options: transportOptions,
+					response,
+					model
+				});
+				if (!response.ok) {
+					const detail = await readAnthropicMessagesErrorBodySnippet(response);
+					throw new Error(formatAnthropicMessagesHttpError(response, detail));
+				}
 				const blocks = output.content;
 				const blockIndexes = /* @__PURE__ */ new Map();
+				const sealedToolCalls = [];
 				const compactionCapture = createCompactionCapture(output, model, transportOptions);
 				const pendingThinkingSignatures = /* @__PURE__ */ new Map();
 				const allowReasoningContentReplay = supportsReasoningContentReplay(model);
@@ -740,6 +748,7 @@ function createAnthropicMessagesTransportStreamFn() {
 							partial: output
 						});
 					}
+					if (contentIndex === void 0) return false;
 					block.thinking += text;
 					block.thinkingSignature = "reasoning_content";
 					eventSink.push({
@@ -771,6 +780,7 @@ function createAnthropicMessagesTransportStreamFn() {
 							partial: output
 						});
 					}
+					if (contentIndex === void 0) return false;
 					block.text += text;
 					eventSink.push({
 						type: "text_delta",
@@ -833,6 +843,7 @@ function createAnthropicMessagesTransportStreamFn() {
 						const fallbackBoundary = refusalBuffer ? readAnthropicFallbackBoundary(contentBlock) : null;
 						if (fallbackBoundary) {
 							refusalBuffer?.discard();
+							sealedToolCalls.length = 0;
 							pendingTextEnds.length = 0;
 							blockIndexes.clear();
 							pendingThinkingSignatures.clear();
@@ -975,7 +986,7 @@ 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") {
+								if (text.length > 0) if (block?.type === "text" && index !== void 0) {
 									block.text += text;
 									eventSink.push({
 										type: "text_delta",
@@ -1003,6 +1014,7 @@ function createAnthropicMessagesTransportStreamFn() {
 								partial: output
 							});
 						}
+						if (index === void 0) continue;
 						if (block?.type === "text" && delta?.type === "text_delta" && typeof delta.text === "string") {
 							block.text += delta.text;
 							eventSink.push({
@@ -1080,12 +1092,9 @@ function createAnthropicMessagesTransportStreamFn() {
 							continue;
 						}
 						if (block.type === "toolCall") {
-							delete block.partialJson;
-							eventSink.push({
-								type: "toolcall_end",
-								contentIndex: index,
-								toolCall: block,
-								partial: output
+							sealedToolCalls.push({
+								block,
+								contentIndex: index
 							});
 							finishReasoningContentSidecars(event.index);
 						}
@@ -1105,6 +1114,17 @@ function createAnthropicMessagesTransportStreamFn() {
 				if (isDirectAnthropicModel(model) && !sawMessageStop) throw new Error("Anthropic stream ended before message_stop");
 				if (transportOptions.signal?.aborted) throw transportAbortError(transportOptions.signal);
 				if (output.stopReason === "aborted" || output.stopReason === "error") throw new Error(output.errorMessage ?? "An unknown error occurred");
+				if ([...blockIndexes.values()].some((index) => blocks[index]?.type === "toolCall")) throw new Error("Provider completed stream with an incomplete tool call");
+				finalizeTerminalToolCallArguments(sealedToolCalls.map(({ block }) => block), (block) => block.partialJson && block.partialJson.length > 0 ? block.partialJson : block.arguments);
+				for (const sealed of sealedToolCalls) {
+					delete sealed.block.partialJson;
+					eventSink.push({
+						type: "toolcall_end",
+						contentIndex: sealed.contentIndex,
+						toolCall: sealed.block,
+						partial: output
+					});
+				}
 				refusalBuffer?.flush();
 				if (output.stopReason === "toolUse" || output.content.some((block) => block.type === "toolCall")) tagPendingCommentaryText(output.content);
 				flushPendingTextEnds();
@@ -1116,7 +1136,7 @@ function createAnthropicMessagesTransportStreamFn() {
 				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,
@@ -1555,10 +1575,33 @@ function readCodeModePayloadToolIdentity(tool, visibleToolNames, allowedHostedTo
 	const name = readCodeModePayloadToolName(tool);
 	return typeof name === "string" && isCodeModeModelVisibleToolName(name, visibleToolNames) ? `client:${name}` : void 0;
 }
-function filterCodeModePayloadTools(payload, visibleToolNames, allowedHostedToolTypes) {
+function readCodeModePayloadToolIdentities(payload) {
+	if (!isRecord(payload)) return [];
+	const tools = readToolPayloadField(payload, "tools");
+	if (!Array.isArray(tools)) return [];
+	return tools.flatMap((tool) => {
+		if (!isRecord(tool)) return [];
+		const identities = [];
+		const name = readCodeModePayloadToolName(tool);
+		if (name) identities.push(`client:${name}`);
+		for (const key of ["functionDeclarations", "function_declarations"]) {
+			const declarations = readToolPayloadField(tool, key);
+			if (!Array.isArray(declarations)) continue;
+			for (const declaration of declarations) {
+				const declarationName = readCodeModePayloadToolName(declaration);
+				if (declarationName) identities.push(`client:${declarationName}`);
+			}
+		}
+		const type = readToolPayloadField(tool, "type");
+		if (typeof type === "string" && type !== "function") identities.push(`hosted:${type}`);
+		return identities;
+	});
+}
+function filterCodeModePayloadTools(payload, visibleToolNames, allowedHostedToolTypes, observer) {
 	if (!isRecord(payload)) return;
 	const tools = readToolPayloadField(payload, "tools");
 	if (!Array.isArray(tools)) return;
+	const beforeToolIdentities = observer ? readCodeModePayloadToolIdentities(payload) : void 0;
 	payload.tools = tools.flatMap((tool) => {
 		const identity = readCodeModePayloadToolIdentity(tool, visibleToolNames, allowedHostedToolTypes);
 		if (identity) return [tool];
@@ -1576,15 +1619,24 @@ function filterCodeModePayloadTools(payload, visibleToolNames, allowedHostedTool
 		}
 		return Object.keys(filteredGroups).length > 0 ? [filteredGroups] : [];
 	});
+	if (beforeToolIdentities) observer?.({
+		beforeToolIdentities,
+		afterToolIdentities: readCodeModePayloadToolIdentities(payload)
+	});
 }
 function resolveCodeModeResponsesVisibleToolNames(context) {
 	return new Set((context.tools ?? []).map(readCodeModePayloadToolName).filter((name) => typeof name === "string"));
 }
-function enforceCodeModeResponsesToolSurface(payload, visibleToolNames, allowedHostedToolTypes) {
+function enforceCodeModeResponsesToolSurface(payload, visibleToolNames, allowedHostedToolTypes, observer) {
 	if (!isRecord(payload)) return;
 	const tools = readToolPayloadField(payload, "tools");
 	if (!Array.isArray(tools)) return;
+	const beforeToolIdentities = observer ? readCodeModePayloadToolIdentities(payload) : void 0;
 	payload.tools = tools.filter((tool) => Boolean(readCodeModePayloadToolIdentity(tool, visibleToolNames, allowedHostedToolTypes)));
+	if (beforeToolIdentities) observer?.({
+		beforeToolIdentities,
+		afterToolIdentities: readCodeModePayloadToolIdentities(payload)
+	});
 }
 function assertCodeModeResponsesToolSurface(payload, visibleToolNames, allowedHostedToolTypes) {
 	const tools = isRecord(payload) ? readToolPayloadField(payload, "tools") : void 0;
@@ -1668,7 +1720,7 @@ function buildOpenAIClientHeaders(model, context, optionHeaders, turnHeaders, se
 		...optionHeaders,
 		...turnHeaders
 	};
-	const resolvedHeaders = resolveProviderRequestPolicyConfig({
+	const resolvedHeaders = resolveProviderRequestPolicyConfig(model, {
 		provider: model.provider,
 		api: model.api,
 		baseUrl: model.baseUrl,
@@ -1716,7 +1768,7 @@ function getCompat(model) {
 //#region packages/ai/src/transports/openai-completions-params.ts
 function isKnownOpenAICompletionsEndpoint(model) {
 	if (!model.baseUrl.trim()) return true;
-	const endpointClass = resolveProviderEndpoint(model.baseUrl).endpointClass;
+	const endpointClass = resolveProviderEndpoint(model).endpointClass;
 	if (endpointClass === "openai-public" || endpointClass === "azure-openai") return true;
 	try {
 		return isAzureOpenAICompatibleHost(new URL(model.baseUrl).hostname.toLowerCase());
@@ -2244,11 +2296,15 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 	const MAX_POST_TOOL_CALL_BUFFER_BYTES = 256e3;
 	const emitReasoning = options?.emitReasoning ?? true;
 	const compat = getCompat(model);
+	const visibleReasoningDetailTypes = new Set(compat.visibleReasoningDetailTypes);
 	const shouldFilterDeepSeekDsmlText = compat.thinkingFormat === "deepseek";
 	const deepSeekTextFilter = shouldFilterDeepSeekDsmlText ? createDeepSeekTextFilter() : null;
 	const deepSeekToolCallRecoverer = shouldFilterDeepSeekDsmlText ? createDsmlRecoverer() : null;
 	const reasoningTagTextPartitioner = createReasoningTagTextPartitioner();
 	let currentBlock = null;
+	let currentTextSource;
+	let pendingInterruptedTextBlock = null;
+	let confirmedInterruptedTextBlock = null;
 	let pendingPostToolCallDeltas = [];
 	let pendingPostToolCallBytes = 0;
 	let isFlushingPendingPostToolCallDeltas = false;
@@ -2256,6 +2312,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 	const toolCallBlocksById = /* @__PURE__ */ new Map();
 	const provisionalCommentaryTags = /* @__PURE__ */ new Map();
 	const toolCallBlockIndices = /* @__PURE__ */ new WeakMap();
+	let explicitVisibleTextBlocks;
 	const normalizeToolCallDeltas = createOpenAICompletionsToolCallDeltaNormalizer();
 	let sawStopFinishReason = false;
 	let sawNativeToolCallDelta = false;
@@ -2271,7 +2328,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		if (pendingPostToolCallBytes + nextBytes > MAX_POST_TOOL_CALL_BUFFER_BYTES) throw new Error("Exceeded post-tool-call delta buffer limit");
 		pendingPostToolCallBytes += nextBytes;
 		const previous = pendingPostToolCallDeltas[pendingPostToolCallDeltas.length - 1];
-		if (!previous || previous.kind !== next.kind) {
+		if (!previous || previous.kind !== next.kind || previous.kind === "text" && next.kind === "text" && previous.source !== next.source) {
 			pendingPostToolCallDeltas.push(next);
 			return;
 		}
@@ -2307,12 +2364,15 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 			partial: output
 		});
 	};
-	const appendTextDeltaInternal = (text) => {
+	const appendTextDeltaInternal = (text, source) => {
+		if (currentBlock?.type === "text" && currentTextSource !== source) currentBlock = null;
 		if (!currentBlock || currentBlock.type !== "text") {
 			currentBlock = {
 				type: "text",
 				text: ""
 			};
+			currentTextSource = source;
+			if (source === "reasoning_detail") (explicitVisibleTextBlocks ??= /* @__PURE__ */ new Set()).add(currentBlock);
 			output.content.push(currentBlock);
 			pushStreamEvent({
 				type: "text_start",
@@ -2321,6 +2381,10 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 			});
 		}
 		currentBlock.text += text;
+		if (pendingInterruptedTextBlock && text.trim()) {
+			confirmedInterruptedTextBlock = pendingInterruptedTextBlock;
+			pendingInterruptedTextBlock = null;
+		}
 		pushStreamEvent({
 			type: "text_delta",
 			contentIndex: blockIndex(),
@@ -2333,7 +2397,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		const bufferedDeltas = pendingPostToolCallDeltas;
 		pendingPostToolCallDeltas = [];
 		pendingPostToolCallBytes = 0;
-		for (const delta of bufferedDeltas) if (delta.kind === "text") appendTextDeltaInternal(delta.text);
+		for (const delta of bufferedDeltas) if (delta.kind === "text") appendTextDeltaInternal(delta.text, delta.source);
 		else if (emitReasoning) appendThinkingDeltaInternal(delta);
 		isFlushingPendingPostToolCallDeltas = false;
 	};
@@ -2341,9 +2405,9 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		flushPendingPostToolCallDeltas();
 		appendThinkingDeltaInternal(reasoningDelta);
 	};
-	const appendTextDelta = (text) => {
+	const appendTextDelta = (text, source) => {
 		flushPendingPostToolCallDeltas();
-		appendTextDeltaInternal(text);
+		appendTextDeltaInternal(text, source);
 	};
 	const appendVisibleTextDelta = (text) => {
 		if (!text) return;
@@ -2353,6 +2417,17 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		});
 		else appendTextDelta(text);
 	};
+	const appendReasoningDeltas = (reasoningDeltas) => {
+		for (const reasoningDelta of reasoningDeltas) {
+			if (reasoningDelta.kind === "thinking" && !emitReasoning) continue;
+			if (currentBlock?.type === "toolCall") {
+				queuePostToolCallDelta({ ...reasoningDelta });
+				continue;
+			}
+			if (reasoningDelta.kind === "text") appendTextDelta(reasoningDelta.text, reasoningDelta.source);
+			else if (emitReasoning) appendThinkingDelta(reasoningDelta);
+		}
+	};
 	const appendRecoveredToolCall = (toolCall) => {
 		if (currentBlock?.type === "toolCall") {
 			currentBlock = null;
@@ -2431,9 +2506,25 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		if (latestBlock?.type === "text" || latestBlock?.type === "toolCall") return;
 		appendThinkingDelta({ text: "" });
 	};
-	const flushReasoningTagTextPartitionerAtEnd = () => {
+	const flushReasoningTagTextPartitioner = () => {
 		for (const delta of reasoningTagTextPartitioner.flush()) appendPartitionedVisibleDelta(delta);
 	};
+	const sealTextBeforeReasoning = () => {
+		if (currentBlock?.type !== "text" && !reasoningTagTextPartitioner.hasPending()) return;
+		flushReasoningTagTextPartitioner();
+		if (currentBlock?.type !== "text") return;
+		if (currentTextSource !== "reasoning_detail" && currentBlock.text.trim()) pendingInterruptedTextBlock = currentBlock;
+		currentBlock = null;
+		currentTextSource = void 0;
+	};
+	const beginReasoning = (hasFollowingVisibleText, forceStrict = false) => {
+		if (!output.openclawDelivery?.textPhaseRequiresTerminal) output.openclawDelivery = {
+			...output.openclawDelivery,
+			textPhaseRequiresTerminal: true
+		};
+		if (forceStrict || reasoningTagTextPartitioner.hasPending()) reasoningTagTextPartitioner.markStrict();
+		if (!hasFollowingVisibleText || !reasoningTagTextPartitioner.hasPendingSyntax()) sealTextBeforeReasoning();
+	};
 	const cooperativeScheduler = createModelStreamCooperativeScheduler(options?.signal);
 	const guardedStream = withFirstStreamEventTimeout(responseStream, {
 		provider: model.provider,
@@ -2485,34 +2576,28 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		}
 		for (const normalizedDelta of normalizeToolCallDeltas(rawChoiceDelta, choice.finish_reason)) {
 			const choiceDelta = normalizedDelta.delta;
-			const reasoningDeltas = getCompletionsReasoningDeltas(choiceDelta, compat.visibleReasoningDetailTypes);
-			const hasMirroredReasoning = reasoningDeltas.some((delta) => delta.kind === "thinking");
-			if (hasMirroredReasoning) reasoningTagTextPartitioner.markStrict();
-			const contentDeltas = readOpenAICompletionsContentDeltas(choiceDelta.content, choiceDelta.refusal, reasoningDeltas.filter((reasoningDelta) => reasoningDelta.kind === "thinking").map((reasoningDelta) => reasoningDelta.text));
-			const appendReasoningDeltas = () => {
-				for (const reasoningDelta of reasoningDeltas) {
-					if (reasoningDelta.kind === "thinking" && !emitReasoning) continue;
-					if (currentBlock?.type === "toolCall") {
-						queuePostToolCallDelta({ ...reasoningDelta });
-						continue;
-					}
-					if (reasoningDelta.kind === "text") appendTextDelta(reasoningDelta.text);
-					else if (emitReasoning) appendThinkingDelta(reasoningDelta);
-				}
-			};
-			if (hasMirroredReasoning) appendReasoningDeltas();
-			for (const contentDelta of contentDeltas) if (contentDelta.kind === "text") {
-				const routedDeltas = hasMirroredReasoning ? reasoningTagTextPartitioner.push(contentDelta.text) : reasoningTagTextPartitioner.pushVisible(contentDelta.text);
+			const reasoningBatch = readOpenAICompletionsReasoningBatch(choiceDelta, visibleReasoningDetailTypes);
+			const reasoningDeltas = reasoningBatch.deltas;
+			const hasReasoningThinking = reasoningBatch.hasThinking;
+			const contentDeltas = readOpenAICompletionsContentDeltas(choiceDelta.content, choiceDelta.refusal, reasoningBatch.mirroredThinking);
+			const lastVisibleTextIndex = contentDeltas.findLastIndex((delta) => delta.kind === "text");
+			const hasSameChunkVisibleText = reasoningBatch.hasVisibleText || lastVisibleTextIndex !== -1;
+			if (hasReasoningThinking) {
+				beginReasoning(hasSameChunkVisibleText, true);
+				appendReasoningDeltas(reasoningDeltas);
+			}
+			for (const [contentDeltaIndex, contentDelta] of contentDeltas.entries()) if (contentDelta.kind === "text") {
+				const routedDeltas = hasReasoningThinking ? reasoningTagTextPartitioner.push(contentDelta.text) : reasoningTagTextPartitioner.pushVisible(contentDelta.text);
 				for (const routedDelta of routedDeltas) appendPartitionedVisibleDelta(routedDelta);
 			} else {
-				if (reasoningTagTextPartitioner.hasPending()) reasoningTagTextPartitioner.markStrict();
+				beginReasoning(contentDeltaIndex < lastVisibleTextIndex);
 				appendRoutedContentDelta(contentDelta);
 			}
-			if (!hasMirroredReasoning) appendReasoningDeltas();
+			if (!hasReasoningThinking) appendReasoningDeltas(reasoningDeltas);
 			const toolCallDeltas = normalizedDelta.toolCalls;
 			if (toolCallDeltas.length > 0) {
 				sawNativeToolCallDelta = true;
-				flushReasoningTagTextPartitionerAtEnd();
+				flushReasoningTagTextPartitioner();
 				rememberPendingCommentaryTags(provisionalCommentaryTags, tagPendingCommentaryText(output.content));
 				for (const toolCall of toolCallDeltas) {
 					const streamIndex = typeof toolCall.index === "number" ? toolCall.index : void 0;
@@ -2566,7 +2651,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		emitReasoningUsageActivity(hasReasoningUsageActivity);
 		await cooperativeScheduler.afterEvent();
 	}
-	flushReasoningTagTextPartitionerAtEnd();
+	flushReasoningTagTextPartitioner();
 	flushDeepSeekToolCallRecovererAtEnd();
 	flushDeepSeekTextFilterAtEnd();
 	currentBlock = null;
@@ -2574,6 +2659,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 	finalizeOpenAICompletionsToolCalls(output, {
 		allowSilentToolCallPromotion: sawStopFinishReason || sawNativeToolCallDelta && (options?.sawStreamDONE?.() ?? false),
 		onConfirmedToolCall(block, contentIndex) {
+			if (block.type !== "toolCall") return;
 			pushStreamEvent({
 				type: "toolcall_end",
 				contentIndex,
@@ -2582,66 +2668,11 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 			});
 		}
 	});
+	if (confirmedInterruptedTextBlock && output.stopReason !== "toolUse" && output.stopReason !== "error" && output.stopReason !== "aborted") tagInterruptedTextPhases(output.content, confirmedInterruptedTextBlock, explicitVisibleTextBlocks);
 	if (output.stopReason !== "toolUse") clearPendingCommentaryText(provisionalCommentaryTags);
+	if (output.stopReason === "error" || output.stopReason === "aborted") tagUnresolvedTextAsCommentary(output);
 	if (output.stopReason === "toolUse") tagPendingCommentaryText(output.content);
 }
-function getCompletionsReasoningDeltas(delta, visibleReasoningDetailTypes) {
-	const output = [];
-	const pushDelta = (next) => {
-		const previous = output[output.length - 1];
-		if (!previous || previous.kind !== next.kind) {
-			output.push(next);
-			return;
-		}
-		if (next.kind === "thinking" && previous.kind === "thinking") {
-			if (previous.signature !== next.signature) {
-				output.push(next);
-				return;
-			}
-			previous.text += next.text;
-			return;
-		}
-		previous.text += next.text;
-	};
-	const reasoningDetails = delta.reasoning_details;
-	let usedReasoningThinkingDetails = false;
-	if (Array.isArray(reasoningDetails)) {
-		const visibleTypes = new Set(visibleReasoningDetailTypes);
-		for (const item of reasoningDetails) {
-			const detail = item;
-			if (typeof detail.text !== "string" || !detail.text) continue;
-			if (detail.type === "reasoning.text") {
-				usedReasoningThinkingDetails = true;
-				pushDelta({
-					kind: "thinking",
-					signature: "reasoning_details",
-					text: detail.text
-				});
-				continue;
-			}
-			if (typeof detail.type === "string" && visibleTypes.has(detail.type)) pushDelta({
-				kind: "text",
-				text: detail.text
-			});
-		}
-	}
-	if (!usedReasoningThinkingDetails) for (const field of [
-		"reasoning_content",
-		"reasoning",
-		"reasoning_text"
-	]) {
-		const value = delta[field];
-		if (typeof value === "string" && value.length > 0) {
-			pushDelta({
-				kind: "thinking",
-				signature: field,
-				text: value
-			});
-			break;
-		}
-	}
-	return output;
-}
 function resolveOpenAICompletionsReasoningEffort(options) {
 	return options?.reasoningEffort ?? options?.reasoning ?? "high";
 }
@@ -2738,8 +2769,7 @@ function buildOpenAICompletionsClientConfig(model, context, optionHeaders) {
 }
 function createOpenAICompletionsTransportStreamFn() {
 	return (model, context, options) => {
-		const eventStream = createAssistantMessageEventStream();
-		const stream = eventStream;
+		const { eventStream, stream } = createWritableTransportEventStream();
 		(async () => {
 			const output = {
 				role: "assistant",
@@ -2794,7 +2824,7 @@ function createOpenAICompletionsTransportStreamFn() {
 				if (nextParams !== void 0) params = nextParams;
 				if (options?.openclawCodeModeToolSurface === true) {
 					const visibleToolNames = resolveCodeModeResponsesVisibleToolNames(context);
-					enforceCodeModeResponsesToolSurface(params, visibleToolNames);
+					enforceCodeModeResponsesToolSurface(params, visibleToolNames, void 0, codeModeToolSurfaceObserver.get(options));
 					assertCodeModeResponsesToolSurface(params, visibleToolNames);
 				}
 				if (getCompat(model).requiresNonEmptyUserOrAssistantMessage) assertOpenAICompletionsPayloadHasConversationTurn(params, model);
@@ -2808,7 +2838,7 @@ function createOpenAICompletionsTransportStreamFn() {
 					stream: responseStream,
 					signal: firstEventAbort.signal,
 					abort: firstEventAbort.abort,
-					hook: createOpenAIResponseHook(options?.onResponse, response, model),
+					hook: createOpenAIProviderAcceptanceHook(options, response, model),
 					onReady: () => stream.push({
 						type: "start",
 						partial: output
@@ -2835,6 +2865,7 @@ function createOpenAICompletionsTransportStreamFn() {
 					cleanup: () => {
 						output.stopReason = options?.signal?.aborted ? "aborted" : "error";
 						finalizeOpenAICompletionsToolCalls(output, { allowSilentToolCallPromotion: false });
+						tagUnresolvedTextAsCommentary(output);
 					}
 				});
 			} finally {
@@ -3405,9 +3436,8 @@ function createOpenAIResponsesWebSocketStream(params) {
 					}
 					const event = readServerEvent(next.value);
 					if (!event) continue;
-					if (event.type === "response.failed") throw new OpenAIResponsesWebSocketResponseFailedError(event.response.output.length > 0);
 					if (event.type === "response.completed") terminalResponse = event.response;
-					terminalReceived = event.type === "response.completed" || event.type === "response.incomplete";
+					terminalReceived = event.type === "response.completed" || event.type === "response.incomplete" || event.type === "response.failed";
 					yield event;
 					if (terminalReceived) {
 						degradedWebSocketConnections.delete(degradationKey);
@@ -3419,7 +3449,7 @@ function createOpenAIResponsesWebSocketStream(params) {
 				const safeRetry = error instanceof OpenAIResponsesWebSocketSafeRetryError;
 				if (!params.callerSignal?.aborted && !safeRetry) markDegraded();
 				if (!requestDispatched && !params.signal?.aborted) throw new OpenAIResponsesWebSocketPreDispatchError(error);
-				if (!requestDispatched || params.callerSignal?.aborted || error instanceof OpenAIResponsesWebSocketResponseFailedError || safeRetry) throw error;
+				if (!requestDispatched || params.callerSignal?.aborted || safeRetry) throw error;
 				throw new OpenAIResponsesWebSocketPostDispatchError(error);
 			} finally {
 				await iterator.return?.().catch(() => void 0);
@@ -3639,11 +3669,17 @@ async function postOpenAIResponsesCompaction(params) {
 		}
 	});
 	const output = isRecord(response) && Array.isArray(response.output) ? response.output : [];
-	const item = output[0];
+	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" || output.length !== 1 || !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 exactly one compaction item");
+	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");
 	return {
 		item,
+		historyMode: retainedUserMessageCount > 0 ? "retained-users" : "compacted-prefix",
 		usage,
 		model: params.model,
 		replayMetadata: buildOpenAIResponsesReasoningReplayMetadata(params.model, {
@@ -3656,8 +3692,7 @@ function createResponsesTransportExecutor(config) {
 	return (model, context, options) => {
 		const responsesOptions = options;
 		const compactRequest = claimResponsesCompactRequest(responsesOptions);
-		const eventStream = createAssistantMessageEventStream();
-		const stream = eventStream;
+		const { eventStream, stream } = createWritableTransportEventStream();
 		(async () => {
 			const output = createOpenAIResponsesAssistantOutput(model, config.outputApi);
 			let firstEventAbort;
@@ -3684,7 +3719,7 @@ function createResponsesTransportExecutor(config) {
 					if (options?.openclawCodeModeToolSurface === true) {
 						const visibleToolNames = resolveCodeModeResponsesVisibleToolNames(context);
 						const allowedHostedToolTypes = responsesOptions?.openclawCodeModeAllowedHostedToolTypes;
-						enforceCodeModeResponsesToolSurface(params, visibleToolNames, allowedHostedToolTypes);
+						enforceCodeModeResponsesToolSurface(params, visibleToolNames, allowedHostedToolTypes, codeModeToolSurfaceObserver.get(options));
 						assertCodeModeResponsesToolSurface(params, visibleToolNames, allowedHostedToolTypes);
 					}
 					return params;
@@ -3701,12 +3736,11 @@ function createResponsesTransportExecutor(config) {
 					output.usage.output = compacted.usage.output_tokens;
 					output.usage.totalTokens = compacted.usage.input_tokens + compacted.usage.output_tokens;
 					compactRequest.resolve(compacted);
-					stream.push({
-						type: "done",
-						reason: output.stopReason,
-						message: output
+					finalizeTransportStream({
+						stream,
+						output,
+						signal: options?.signal
 					});
-					stream.end();
 					return;
 				}
 				const sessionId = options?.sessionId;
@@ -3760,7 +3794,7 @@ function createResponsesTransportExecutor(config) {
 						stream: observeResponsesStream(rawResponseStream, model, requestStartedAt),
 						signal: firstEvent.signal,
 						abort: firstEvent.abort,
-						hook: createOpenAIResponseHook(options?.onResponse, response, model),
+						hook: createOpenAIProviderAcceptanceHook(options, response, model),
 						onReady: () => {
 							emitModelTransportDebug(log, `[responses] headers provider=${model.provider} api=${model.api} model=${model.id} transport=sse elapsedMs=${Date.now() - requestStartedAt}`);
 							startStream();
@@ -3796,8 +3830,16 @@ function createResponsesTransportExecutor(config) {
 					transport = "websocket";
 					emitModelTransportDebug(log, `[responses] websocket_selected provider=${model.provider} api=${model.api} model=${model.id} mode=${websocketMode} reused=${websocket.reusedConnection} continuation=${websocket.continuationStatus === "continued"} continuationStatus=${websocket.continuationStatus} sessionIdHash=${redactIdentifier(options?.sessionId)} headersHash=${redactIdentifier(JSON.stringify(Object.entries(websocketHeaders ?? {}).toSorted(([a], [b]) => a.localeCompare(b))))}`);
 					responseStream = { async *[Symbol.asyncIterator]() {
+						let providerAccepted = false;
 						try {
 							for await (const event of websocket.stream) {
+								if (!providerAccepted) {
+									providerAccepted = true;
+									await notifyProviderStreamOpened({
+										options,
+										cancelStream: () => websocket.finish({ keep: false })
+									});
+								}
 								startStream();
 								yield event;
 							}
@@ -3848,33 +3890,30 @@ function createResponsesTransportExecutor(config) {
 					throw error;
 				}
 				emitModelTransportDebug(log, `[responses] completed provider=${model.provider} api=${model.api} model=${model.id} transport=${transport} elapsedMs=${Date.now() - requestStartedAt}`);
-				stream.push({
-					type: "done",
-					reason: output.stopReason,
-					message: output
+				finalizeTransportStream({
+					stream,
+					output,
+					signal: options?.signal
 				});
-				stream.end();
 			} catch (error) {
 				if (compactRequest) {
 					compactRequest.reject(error);
-					Object.assign(output, projectProviderError(error, options?.signal));
-					stream.push({
-						type: "error",
-						reason: output.stopReason,
-						error: output
+					failTransportStream({
+						stream,
+						output,
+						signal: options?.signal,
+						error
 					});
-					stream.end();
 					return;
 				}
 				if (error instanceof ResponsesStreamFailure && error.observation) logResponsesFailedNoDetails(error.observation);
 				log.warn(`[responses] error provider=${model.provider} api=${model.api} model=${model.id} ` + summarizeOpenAITransportError(error));
-				Object.assign(output, projectProviderError(error, options?.signal));
-				stream.push({
-					type: "error",
-					reason: output.stopReason,
-					error: output
+				failTransportStream({
+					stream,
+					output,
+					signal: options?.signal,
+					error
 				});
-				stream.end();
 			} finally {
 				continuationClaim?.release();
 				firstEventAbort?.dispose();
@@ -3970,7 +4009,7 @@ if (process.env.VITEST || false) globalThis.openclawOpenAIResponsesTransportTest
 /** Whether provider replay state is a prefix-bound server compaction checkpoint. */
 function isCompactionReplayCheckpoint(replay) {
 	const type = replay && typeof replay === "object" ? replay.type : void 0;
-	return type === "anthropic-compaction" || type === "openai-responses-compaction";
+	return type === "anthropic-compaction" || type === "openai-responses-compaction" || type === "openai-responses-retained-compaction";
 }
 /** Strip prefix-bound checkpoints after local history rewrites. */
 function stripCompactionReplayCheckpoint(message) {
@@ -3991,8 +4030,9 @@ function replaceCompactionReplayOwnerContent(message, content) {
 	};
 	const replay = message.providerReplay;
 	if (!isCompactionReplayCheckpoint(replay)) return next;
+	if (replay.type === "openai-responses-retained-compaction") return next;
 	const replayIndex = replay.replayIndex ?? 0;
-	if (content.length === 0 || replayIndex > message.content.length) return stripCompactionReplayCheckpoint(next);
+	if (content.length === 0 && message.content.length > 0 || replayIndex > message.content.length) return stripCompactionReplayCheckpoint(next);
 	let sourceIndex = 0;
 	let nextReplayIndex = 0;
 	if (!content.every((block) => {
@@ -4261,4 +4301,4 @@ function prepareModelForSimpleCompletion(params) {
 	return applyProviderSimpleCompletionWrapper(apiRegistry, model, cfg);
 }
 //#endregion
-export { GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, applyAnthropicCacheControlToMessages, applyAnthropicEphemeralCacheControlMarkers, applyAnthropicPayloadPolicyToParams, applyOpenAIResponsesPayloadPolicy, assertCodeModeResponsesToolSurface, assignTransportErrorDetails, buildOpenAIClientHeaders, buildOpenAICompletionsParams, buildOpenAISdkClientOptions, buildOpenAISdkRequestOptions, buildTransportAwareSimpleStreamFn, canonicalizeMaxTokensParam, captureOpenAIResponsesCompaction, coerceTransportToolCallArguments, createAnthropicMessagesTransportStreamFn, createAzureOpenAIResponsesTransportStreamFn, createBoundaryAwareStreamFnForModel, createDeepSeekTextFilter, createEmptyTransportUsage, createModelStreamCooperativeScheduler, createOpenAICompletionsTransportStreamFn, createOpenAIResponseHook, createOpenAIResponsesTransportStreamFn, createOpenClawTransportStreamFnForModel, createTransportAwareStreamFnForModel, createWritableTransportEventStream, detectOpenAICompletionsCompat, emitModelTransportDebug, enforceCodeModeResponsesToolSurface, failTransportStream, filterCodeModePayloadTools, finalizeTransportStream, flattenCompletionMessagesToStringContent, formatModelTransportDebugBaseUrl, formatModelTransportDebugUrl, getCompat, googleFlashSupportsMinimalThinking, hasOpenAICompatibleConversationTurn, isCodeModeModelVisibleToolName, isCompactionReplayCheckpoint, isOpenAICodexResponsesModel, isOpenAICompletionsThinkingEnabled, log, measureUtf8AppendBytes, mergeTransportHeaders, mergeTransportMetadata, normalizeCodexResponsesBaseUrlForOpenAISdk, parseJsonObjectPreservingUnsafeIntegers, parseJsonPreservingUnsafeIntegers, parseOpenAICompletionsUsage, prepareModelForSimpleCompletion, prepareTransportAwareSimpleModel, quoteUnsafeIntegerLiterals, readCodeModePayloadToolName, readOpenAICompletionsContentDeltas, replaceCompactionReplayOwnerContent, requestPreparedOpenAIResponsesCompaction, resolveAnthropicEphemeralCacheControl, resolveAnthropicMessagesUrl, resolveAnthropicPayloadPolicy, resolveAnthropicServerCompactionPlan, resolveCodeModeResponsesVisibleToolNames, resolveMaxTokensParam, resolveModelPayloadDebugMode, resolveModelSseDebugMode, resolveOpenAICompletionsCompat, resolveOpenAIReasoningEffortMap, resolveOpenAIResponsesCompactEndpointPlan, resolveOpenAIResponsesPayloadPolicy, resolveOpenAIResponsesServerCompactionPlan, resolveOpenAIStrictToolFlagWithDiagnostics, resolvePromptCacheKey, resolveReplayableResponsesMessageId, resolveTransportAwareSimpleApi, sanitizeNonEmptyTransportPayloadText, sanitizeResponsesImagePayload, sanitizeTransportPayloadText, sortPromptCacheToolsByName as sortTransportToolsByName, stripCompactionReplayCheckpoint, stripCompactionReplayCheckpointInPlace, stripCompletionMessagesToRoleContent, throwIfModelStreamAborted, transportAbortError, usesNativeOpenAICodexResponsesBackend, withProviderResponseHook };
+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, resolveOpenAICompletionsCompat, resolveOpenAIReasoningEffortMap, resolveOpenAIResponsesCompactEndpointPlan, resolveOpenAIResponsesPayloadPolicy, resolveOpenAIResponsesServerCompactionPlan, resolveOpenAIStrictToolFlagWithDiagnostics, resolvePromptCacheKey, resolveReplayableResponsesMessageId, resolveTransportAwareSimpleApi, sanitizeNonEmptyTransportPayloadText, sanitizeResponsesImagePayload, sanitizeTransportPayloadText, sortPromptCacheToolsByName as sortTransportToolsByName, stripCompactionReplayCheckpoint, stripCompactionReplayCheckpointInPlace, stripCompletionMessagesToRoleContent, throwIfModelStreamAborted, transportAbortError, usesNativeOpenAICodexResponsesBackend, withProviderAcceptanceObserver, withProviderResponseHook };