@openclaw/ai 2026.10.5-beta.3 → 2026.10.10-beta.1

package/dist/{openai-completions-transport-Dy4CZo9Q.mjs → openai-completions-transport-DviPEqw3.mjs}

@@ -1,38 +1,40 @@
-import { _ as runtimeContextContentToText, f as hasRuntimeContextMarker, p as isRuntimeContextMessage } from "./types-CQc6JNG4.mjs";
+import { t as createReasoningTagTextPartitioner } from "./reasoning-tag-text-partitioner-DzlRubwU.mjs";
+import { _ as runtimeContextContentToText, f as hasRuntimeContextMarker, p as isRuntimeContextMessage } from "./types-BXd0kI8-.mjs";
+import { t as appendAssistantMessageDiagnostic } from "./diagnostics-DzrF1h4J.mjs";
 import { r as appendAssistantThinking } from "./event-stream-X2gmsUWv.mjs";
 import { i as asOptionalRecord, o as isRecord, r as asOptionalObjectRecord } from "./record-coerce-CNDovvft.mjs";
-import { s as readNonEmptyStringPreservingWhitespace, t as hasNonEmptyString } from "./string-coerce-CtgOVOSv.mjs";
-import { n as asNonNegativeFiniteNumber, r as asPositiveFiniteNumber } from "./number-coercion-C3hztWoY.mjs";
-import { N as isImageWithMediaPayload, a as resolveAiTransportHeaderSentinels, r as getAiTransportHost } from "./host-Bf3NjYNj.mjs";
-import { M as resolveOpenAIThinkingApi, S as isOpenAIGpt6Model, _ as isResponsesServiceTierRejection, b as isOpenAIGpt55Model, d as OpenAIResponsesWebSocketSafeRetryError, g as responsesServiceTierObserver, k as supportsOpenAITemperature, l as OpenAIResponsesWebSocketPostDispatchError, m as parseOpenAIResponsesWebSocketServerError, r as OPENAI_CODEX_RESPONSES_DEFAULT_INSTRUCTIONS, t as AZURE_RESPONSES_FIRST_EVENT_TIMEOUT_MS, u as OpenAIResponsesWebSocketPreDispatchError, v as nextResponsesServiceTier, x as isOpenAIGpt56Model, y as isOpenAIGpt54MiniModel } from "./openai-responses-contracts-DlKMMTxK.mjs";
-import { r as uniqueStrings } from "./string-normalization-I3ux50s9.mjs";
-import { n as getSessionResourceOwnerId, r as registerSessionResourceCleanup } from "./session-resources-D1wgDQ8C.mjs";
-import { c as stableStringify, s as truncateUtf16Safe } from "./provider-error-CF9GWCs2.mjs";
-import { r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-r93BJhjm.mjs";
-import { d as hasOnlyResponsesFunctionTools, i as emitModelTransportError, r as emitModelTransportDebug, t as formatModelTransportDebugBaseUrl } from "./diagnostics-tQasYct1.mjs";
-import { f as supportsModelTools, i as providerReplayContextMatches, n as buildProviderReplayContext, o as estimateStringChars, r as isProviderReplayContext } from "./text-signature-Drb4Erb8.mjs";
-import { c as resolveOpenAIResponsesTextFormat, i as resolveOpenAICompletionsCompat, l as shouldOmitOllamaCompatResponseFormat, n as isNativeOpenAIEndpoint, o as usesNativeOpenAICodexResponsesBackend, r as isOpenAICodexResponsesModel, s as resolveOpenAICompletionsResponseFormat, t as detectOpenAICompletionsCompat } from "./openai-completions-compat-CYX18uDk.mjs";
-import { i as resolveProviderEndpoint, t as buildGuardedModelFetch } from "./host-policy-BzZluO6_.mjs";
-import { i as resolveCacheRetention, r as buildCopilotDynamicHeaders } from "./cloudflare-5qSGvIUH.mjs";
-import { E as createToolArgumentPreviewSchedule, M as racePromiseWithAbortSignal, O as parseStreamingJson, P as sanitizeSurrogates, S as withProviderResponseHook, b as transportAbortError, c as failTransportStream, d as iterateModelStream, h as notifyProviderStreamOpened, l as finalizeTerminalToolCallArguments, s as createWritableTransportEventStream, u as finalizeTransportStream, v as sanitizeTransportPayloadText, y as throwIfModelStreamAborted } from "./transport-stream-shared-CWGOuLcA.mjs";
-import { C as stripSystemPromptCacheBoundary, D as sortPromptCacheToolsByName, S as splitSystemPromptRelocatableBoundary, d as describeToolResultMediaPlaceholder, l as codeModeToolSurfaceObserver, m as extractToolResultText, s as prepareModelRequestBody, u as reasoningTagTextPolicy, w as stripSystemPromptRelocatableBoundary, x as splitSystemPromptCacheBoundary } from "./simple-options-BAu4B54D.mjs";
-import { $ as replayResponsesReasoningUpdates, A as summarizeResponsesPayload, At as measureUtf8AppendBytes, B as responsesInputFingerprint, C as readResponsesRecoveryEvent, Ct as toOpenAIResponsesToolChoice, D as logResponsesFailedNoDetails, Dt as createResponseModelTracker, E as ResponsesStreamFailure, Ft as resolvePromptCacheKey, G as responsesContinuationRequestFingerprint, H as recordResponsesContinuationState, I as convertResponsesMessages, It as redactIdentifier, Lt as sha256Hex, Mt as readOpenAICompletionsContentDeltas, Nt as readOpenAICompletionsReasoningBatch, O as safeDebugValue, Ot as isOpenAICompletionsThinkingEnabled, P as buildResponsesInputMessage, Pt as resolveOpenAIClientBaseUrl, Q as isConfigurationUpdate, R as readResponsesInputReplayState, S as isInvalidEncryptedContentError, St as reconcileOpenAIResponsesToolChoice, T as resolveNextResponsesEncryptedContentAttempt, Tt as createOpenAIProviderAcceptanceHook, U as resolveResponsesContinuationRequest, V as claimOpenAIResponsesHttpContinuation, W as responsesContinuationPrefixFingerprint, _t as resolveOpenAISimpleReasoningEffort, a as prepareResponsesTools, at as suppressOpenAIResponsesCompaction, bt as projectOpenAITools, c as observeResponsesStream, ct as buildOpenAISdkClientOptions, et as supportsResponsesReasoningUpdate, ft as getCompat, gt as resolveOpenAIRequestReasoning, ht as resolveOpenAIStrictToolFlagWithDiagnostics, j as recordResponsesContextUsage, jt as parseOpenAICompletionsUsage, k as summarizeOpenAITransportError, kt as log, lt as buildOpenAISdkRequestOptions, mt as resolveCodeModeResponsesVisibleToolNames, n as applyResponsesServiceTierPricing, nt as buildOpenAIResponsesReasoningReplayMetadata, ot as assertCodeModeResponsesToolSurface, s as processResponsesStream, st as buildOpenAIClientHeaders, ut as enforceCodeModeResponsesToolSurface, w as resolveAzureOpenAIApiVersion, wt as GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, x as createResponsesStreamWithRecovery, xt as reconcileOpenAICompletionsToolChoice, yt as prepareOpenAITools, z as recordResponsesInputReplay } from "./openai-responses-shared-D-aDYAkJ.mjs";
-import { n as getEnvApiKey } from "./env-api-keys-CrxwdylR.mjs";
-import { n as transformProviderMessages, t as createAssistantOutput } from "./assistant-output-BuhRpLeD.mjs";
-import { a as tagUnresolvedTextAsCommentary, i as tagPendingCommentaryText, n as rememberPendingCommentaryTags, r as tagInterruptedTextPhases, t as clearPendingCommentaryText } from "./assistant-text-phase-Bn1hbvK9.mjs";
+import { c as readNonEmptyStringPreservingWhitespace, t as hasNonEmptyString } from "./string-coerce-Cj8sqWHx.mjs";
+import { n as asNonNegativeFiniteNumber, r as asPositiveFiniteNumber } from "./number-coercion-37gmNkuA.mjs";
+import { N as isImageWithMediaPayload, a as resolveAiTransportHeaderSentinels, r as getAiTransportHost } from "./host-B2gk-d1H.mjs";
+import { A as supportsOpenAITemperature, C as isOpenAIGpt6Model, L as resolveOpenAIThinkingApi, M as uniqueStrings, S as isOpenAIGpt56Model, _ as responsesServiceTierObserver, b as isOpenAIGpt54MiniModel, d as OpenAIResponsesWebSocketSafeRetryError, h as parseOpenAIResponsesWebSocketServerError, l as OpenAIResponsesWebSocketPostDispatchError, r as OPENAI_CODEX_RESPONSES_DEFAULT_INSTRUCTIONS, t as AZURE_RESPONSES_FIRST_EVENT_TIMEOUT_MS, u as OpenAIResponsesWebSocketPreDispatchError, v as isResponsesServiceTierRejection, x as isOpenAIGpt55Model, y as nextResponsesServiceTier } from "./openai-responses-contracts-Bx-16mnf.mjs";
+import { n as getSessionResourceOwnerId, r as registerSessionResourceCleanup } from "./session-resources-vKvGIE3J.mjs";
+import { c as stableStringify, s as truncateUtf16Safe } from "./provider-error-qtZD2inE.mjs";
+import { r as resolveOpenAIPromptCacheParams } from "./openai-prompt-cache-DJ_P143P.mjs";
+import { d as hasOnlyResponsesFunctionTools, i as emitModelTransportError, r as emitModelTransportDebug, t as formatModelTransportDebugBaseUrl } from "./diagnostics-Cv7cW3pS.mjs";
+import { a as estimateStringChars, d as supportsModelTools, n as isProviderReplayContext, r as providerReplayContextMatches, t as buildProviderReplayContext } from "./provider-replay-context-J6CfU3m1.mjs";
+import { c as resolveOpenAIResponsesTextFormat, i as resolveOpenAICompletionsCompat, l as shouldOmitOllamaCompatResponseFormat, n as isNativeOpenAIEndpoint, o as usesNativeOpenAICodexResponsesBackend, r as isOpenAICodexResponsesModel, s as resolveOpenAICompletionsResponseFormat, t as detectOpenAICompletionsCompat } from "./openai-completions-compat-CFhQ3UVU.mjs";
+import { i as resolveProviderEndpoint, t as buildGuardedModelFetch } from "./host-policy-mpqGhbeM.mjs";
+import { i as resolveCacheRetention, r as buildCopilotDynamicHeaders } from "./cloudflare-Dttf8jKq.mjs";
+import { t as sanitizeSurrogates } from "./sanitize-unicode-CRGuMONI.mjs";
+import { r as sortPromptCacheToolsByName } from "./prompt-cache-stability-D_MeCQFI.mjs";
+import { C as splitSystemPromptRelocatableBoundary, S as splitSystemPromptCacheBoundary, T as stripSystemPromptRelocatableBoundary, f as extractToolResultText, g as reasoningTagTextPolicy, h as preparedRequestSuffix, l as describeToolResultMediaPlaceholder, m as codeModeToolSurfaceObserver, r as clampMaxTokensToModel, s as prepareModelRequestBody, w as stripSystemPromptCacheBoundary } from "./simple-options-BYgC6kse.mjs";
+import { A as shortHash, E as createToolArgumentPreviewSchedule, M as racePromiseWithAbortSignal, O as parseStreamingJson, S as withProviderResponseHook, b as transportAbortError, c as failTransportStream, d as iterateModelStream, h as notifyProviderStreamOpened, l as finalizeTerminalToolCallArguments, s as createWritableTransportEventStream, u as finalizeTransportStream, v as sanitizeTransportPayloadText, y as throwIfModelStreamAborted } from "./transport-stream-shared-CzQlAzJH.mjs";
+import { C as redactIdentifier, w as sha256Hex } from "./tool-schema-json-projection-B0hHDV7N.mjs";
+import { n as getEnvApiKey } from "./env-api-keys-Xd8dm_aV.mjs";
+import { n as transformProviderMessages, t as createAssistantOutput } from "./assistant-output-6iFA1nbC.mjs";
 import { t as notifyLlmRequestActivity } from "./llm-request-activity-CfMuh5cC.mjs";
 import { n as resolveOpencodeSessionHeaders } from "./session-affinity-DA4WHv1h.mjs";
+import { isGoogleGemini3ThinkingLevelModel } from "./internal/google-model-family.mjs";
 import "./types.mjs";
-import { i as createOpenAIProviderClient, n as resolveProviderSimpleCompletionHeaders, r as resolveProviderTransportTurnState, t as filterProviderTurnHeadersForExplicitOpencodeSession } from "./provider-transport-turn-state-BOpxgm-J.mjs";
-import { a as withFirstStreamEventTimeout, i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-Dc42hKv6.mjs";
+import { $ as isConfigurationUpdate, A as summarizeResponsesPayload, At as log, B as recordResponsesInputReplay, C as readResponsesRecoveryEvent, Ct as reconcileOpenAIResponsesToolChoice, D as logResponsesFailedNoDetails, E as ResponsesStreamFailure, Et as createOpenAIProviderAcceptanceHook, Ft as resolveOpenAIClientBaseUrl, G as responsesContinuationPrefixFingerprint, H as claimOpenAIResponsesHttpContinuation, I as convertResponsesMessages, It as resolvePromptCacheKey, K as responsesContinuationRequestFingerprint, Lt as withOpenAIErrorResponseHook, Mt as parseOpenAICompletionsUsage, Nt as readOpenAICompletionsContentDeltas, O as safeDebugValue, Ot as createResponseModelTracker, P as buildResponsesInputMessage, Pt as readOpenAICompletionsReasoningBatch, S as isInvalidEncryptedContentError, St as reconcileOpenAICompletionsToolChoice, T as resolveNextResponsesEncryptedContentAttempt, Tt as GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP, U as recordResponsesContinuationState, V as responsesInputFingerprint, W as resolveResponsesContinuationRequest, _t as resolveOpenAIRequestReasoning, a as prepareResponsesTools, bt as prepareOpenAITools, c as observeResponsesStream, ct as buildOpenAIClientHeaders, dt as enforceCodeModeResponsesToolSurface, et as replayResponsesReasoningUpdates, gt as resolveOpenAIStrictToolFlagWithDiagnostics, ht as resolveCodeModeResponsesVisibleToolNames, j as recordResponsesContextUsage, jt as measureUtf8AppendBytes, k as summarizeOpenAITransportError, kt as isOpenAICompletionsThinkingEnabled, lt as buildOpenAISdkClientOptions, n as applyResponsesServiceTierPricing, ot as suppressOpenAIResponsesCompaction, pt as getCompat, rt as buildOpenAIResponsesReasoningReplayMetadata, s as processResponsesStream, st as assertCodeModeResponsesToolSurface, tt as supportsResponsesReasoningUpdate, ut as buildOpenAISdkRequestOptions, vt as resolveOpenAISimpleReasoningEffort, w as resolveAzureOpenAIApiVersion, wt as toOpenAIResponsesToolChoice, x as createResponsesStreamWithRecovery, xt as projectOpenAITools, z as readResponsesInputReplayState } from "./openai-responses-shared-Bv_1NtH9.mjs";
+import { i as createOpenAIProviderClient, n as resolveProviderSimpleCompletionHeaders, r as resolveProviderTransportTurnState, t as filterProviderTurnHeadersForExplicitOpencodeSession } from "./provider-transport-turn-state-Jxppzuc2.mjs";
+import { a as withFirstStreamEventTimeout, i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-D_TfJHRJ.mjs";
 import { t as boundResponseBody } from "./streaming-byte-guard-DBXr1XYn.mjs";
-import { f as withPreparedToolSchemaNormalization, r as normalizeOpenAIStrictToolParameters } from "./openai-tool-schema-DDPXsvJP.mjs";
-import { isGoogleGemini3ThinkingLevelModel } from "./internal/google-model-family.mjs";
+import { b as prepareOpenAIFunctionTool, f as normalizeToolSchema, p as withPreparedToolSchemaNormalization, r as normalizeOpenAIStrictToolParameters } from "./openai-tool-schema-Buv0GXFb.mjs";
 import { t as mapOpenAIStopReason } from "./openai-stop-reason-8s70vDXj.mjs";
-import { t as createReasoningTagTextPartitioner } from "./reasoning-tag-text-partitioner-CpOy9-IQ.mjs";
 import { i as resolveAzureDeploymentNameFromMap, t as isOpenAICompatibleAzureResponsesBaseUrl } from "./azure-openai-responses-client-compat-D0dvJMbZ.mjs";
-import { i as isOpenAIResponsesCompactionOutput, l as resolveOpenAIResponsesPayloadPolicy, n as supportsNativeOpenAIResponsesEndpoint, o as sanitizeResponsesImagePayload, r as createBoundedOpenAIResponsesCompactionFetch, s as applyOpenAIResponsesPayloadPolicy, t as isOfficialOpenAIResponsesBaseUrl } from "./openai-responses-endpoint-CinGh4VZ.mjs";
-import { n as withResponsesRequestAcceptance, r as createResponsesPromptEgressObserver, t as responsesRequestLifecycle } from "./openai-responses-request-lifecycle-NSIJo9jY.mjs";
+import { a as isOpenAIResponsesCompactionOutput, c as applyOpenAIResponsesPayloadPolicy, d as resolveOpenAIResponsesPayloadPolicy, i as getOpenAIResponsesCompactionOutputError, n as supportsNativeOpenAIResponsesEndpoint, r as createBoundedOpenAIResponsesCompactionFetch, s as sanitizeResponsesImagePayload, t as isOfficialOpenAIResponsesBaseUrl } from "./openai-responses-endpoint-D4v-eeYy.mjs";
+import { a as createResponsesV2CompactionCollector, i as createResponsesPromptEgressObserver, n as responsesRequestLifecycle, o as claimResponsesCompactRequest, r as withResponsesRequestAcceptance, t as independentResponsesRequest } from "./openai-responses-request-lifecycle-CTdAb2v1.mjs";
 import { randomUUID } from "node:crypto";
 import OpenAI, { AzureOpenAI } from "openai";
 import { ResponsesWS } from "openai/resources/responses/ws.js";
@@ -207,9 +209,21 @@ function applyCompletionsAnthropicCacheControl(payload, cacheControl = { type: "
 	if (shapedPayloads.has(payload)) return;
 	shapedPayloads.add(payload);
 	const messages = Array.isArray(payload.messages) ? payload.messages : [];
-	const tools = Array.isArray(payload.tools) ? payload.tools.filter(isRecord) : [];
+	const tools = Array.isArray(payload.tools) ? payload.tools : [];
+	let lastToolIndex = -1;
+	for (let index = 0; index < tools.length; index++) {
+		const tool = tools[index];
+		if (isRecord(tool)) {
+			lastToolIndex = index;
+			if (Object.hasOwn(tool, "cache_control")) {
+				const unmarked = { ...tool };
+				delete unmarked.cache_control;
+				tools[index] = unmarked;
+			}
+		}
+	}
 	const blocks = messages.filter(isRecord).flatMap((message) => Array.isArray(message.content) ? message.content.filter(isRecord) : []);
-	for (const block of [...tools, ...blocks]) delete block.cache_control;
+	for (const block of blocks) delete block.cache_control;
 	if (!cacheControl) return;
 	const markText = (message, splitBoundary) => {
 		if (typeof message.content === "string" && message.content) message.content = [{
@@ -237,8 +251,10 @@ function applyCompletionsAnthropicCacheControl(payload, cacheControl = { type: "
 		}
 		return false;
 	};
-	const lastTool = tools.at(-1);
-	if (markTools && lastTool) lastTool.cache_control = cacheControl;
+	if (markTools && lastToolIndex >= 0) tools[lastToolIndex] = {
+		...tools[lastToolIndex],
+		cache_control: cacheControl
+	};
 	if (!markMessages) return;
 	const system = messages.find((message) => isRecord(message) && (message.role === "system" || message.role === "developer"));
 	if (isRecord(system)) markText(system, true);
@@ -451,12 +467,12 @@ function finalizeOpenAICompletionsToolCalls(output, options = {}) {
 		output.errorMessage = "Provider returned an incomplete or malformed tool call";
 		output.content = output.content.filter((candidate) => !isToolCall(candidate));
 	};
-	if (toolCalls.some((call) => typeof call.name !== "string" || call.name.trim().length === 0 || typeof call.partialArgs !== "string" || call.partialArgs.trim().length === 0)) {
+	if (toolCalls.some((call) => typeof call.name !== "string" || call.name.trim().length === 0 || typeof call.partialJson !== "string" || call.partialJson.trim().length === 0)) {
 		rejectToolCalls();
 		return;
 	}
 	try {
-		finalizeTerminalToolCallArguments(toolCalls, (call) => call.partialArgs);
+		finalizeTerminalToolCallArguments(toolCalls, (call) => call.partialJson);
 	} catch {
 		rejectToolCalls();
 		return;
@@ -464,12 +480,75 @@ function finalizeOpenAICompletionsToolCalls(output, options = {}) {
 	for (let contentIndex = 0; contentIndex < output.content.length; contentIndex += 1) {
 		const block = output.content[contentIndex];
 		if (!block || !isToolCall(block)) continue;
-		delete block.partialArgs;
+		delete block.partialJson;
 		delete block.streamIndex;
 		options.onConfirmedToolCall?.(block, contentIndex);
 	}
 }
 //#endregion
+//#region packages/ai/src/transports/openai-completions-context-budget-buffer.ts
+const MAX_CONTEXT_BUDGET_BUFFER_CHUNKS = 8192;
+const MAX_CONTEXT_BUDGET_BUFFER_BYTES = 4194304;
+async function bufferContextLimitedCompletions(source, signal) {
+	const iterator = source[Symbol.asyncIterator]();
+	const chunks = [];
+	let bufferedBytes = 0;
+	let overflow;
+	let finished = false;
+	let failed = false;
+	let sourceError;
+	let finishReason;
+	try {
+		for (;;) {
+			const next = await iterator.next();
+			if (next.done) {
+				finished = true;
+				break;
+			}
+			const chunk = next.value;
+			const chunkBytes = Buffer.byteLength(JSON.stringify(chunk) ?? "", "utf8");
+			if (chunks.length >= MAX_CONTEXT_BUDGET_BUFFER_CHUNKS || bufferedBytes + chunkBytes > MAX_CONTEXT_BUDGET_BUFFER_BYTES) {
+				overflow = next;
+				break;
+			}
+			chunks.push(chunk);
+			bufferedBytes += chunkBytes;
+			const chunkFinishReason = chunk?.choices?.[0]?.finish_reason;
+			if (chunkFinishReason) finishReason = chunkFinishReason;
+			signal?.throwIfAborted();
+		}
+	} catch (error) {
+		failed = true;
+		sourceError = error;
+		iterator.return?.().catch(() => void 0);
+	}
+	return {
+		bounded: overflow === void 0,
+		failed,
+		finishReason,
+		stream: { async *[Symbol.asyncIterator]() {
+			try {
+				yield* chunks;
+				if (failed) throw sourceError;
+				if (overflow !== void 0) {
+					yield overflow.value;
+					for (;;) {
+						const next = await iterator.next();
+						if (next.done) {
+							finished = true;
+							return;
+						}
+						yield next.value;
+					}
+				}
+			} finally {
+				chunks.length = 0;
+				if (!finished && !failed) iterator.return?.().catch(() => void 0);
+			}
+		} }
+	};
+}
+//#endregion
 //#region packages/ai/src/transports/openai-completions-host.ts
 /**
 * Chat Completions accepts Azure AI Foundry hosts in addition to traditional
@@ -488,6 +567,7 @@ function hasToolCallHistory(messages) {
 /** Convert a normalized transcript to OpenAI Chat Completions messages. */
 function convertMessages(model, context, compat, options = {}) {
 	const params = [];
+	const supportsTools = options.supportsTools !== false;
 	const normalizeToolCallId = (id) => {
 		if (id.includes("|")) return id.slice(0, id.indexOf("|")).replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 40);
 		if (model.provider === "openai") return id.length > 40 ? truncateUtf16Safe(id, 40) : id;
@@ -565,7 +645,10 @@ function convertMessages(model, context, compat, options = {}) {
 			msg.content.forEach((block) => {
 				if (block.type === "text" && block.text.trim().length > 0) assistantTexts.push(sanitizeSurrogates(block.text));
 				else if (block.type === "thinking" && block.thinking.trim().length > 0) nonEmptyThinkingBlocks.push(block);
-				else if (block.type === "toolCall") toolCalls.push(block);
+				else if (block.type === "toolCall") {
+					if (supportsTools) toolCalls.push(block);
+					else assistantTexts.push(sanitizeSurrogates(`[tool call id=${block.id} name=${block.name}] ${JSON.stringify(block.arguments)}`));
+				}
 			});
 			if (nonEmptyThinkingBlocks.length > 0 && compat.requiresThinkingAsText) assistantMsg.content = [{
 				type: "text",
@@ -617,13 +700,28 @@ function convertMessages(model, context, compat, options = {}) {
 				const textResult = extractToolResultText(toolMsg.content);
 				const mediaPlaceholder = describeToolResultMediaPlaceholder(toolMsg.content);
 				const images = toolMsg.content.filter(isImageWithMediaPayload);
-				const toolResultMsg = {
-					role: "tool",
-					content: textResult.trim() ? textResult : mediaPlaceholder ?? "(no output)",
-					tool_call_id: toolMsg.toolCallId
-				};
-				if (compat.requiresToolResultName && toolMsg.toolName) toolResultMsg.name = toolMsg.toolName;
-				params.push(toolResultMsg);
+				const content = textResult.trim() ? textResult : mediaPlaceholder ?? "(no output)";
+				if (supportsTools) {
+					const toolResultMsg = {
+						role: "tool",
+						content,
+						tool_call_id: preparedRequestSuffix.has(context) ? preparedRequestSuffix.toolCallId(context, toolMsg.toolCallId, normalizeToolCallId(toolMsg.toolCallId)) : toolMsg.toolCallId
+					};
+					if (compat.requiresToolResultName && toolMsg.toolName) toolResultMsg.name = toolMsg.toolName;
+					params.push(toolResultMsg);
+				} else {
+					const text = sanitizeSurrogates(`[tool result id=${toolMsg.toolCallId} name=${toolMsg.toolName}] ${content}`);
+					const previous = params.at(-1);
+					if (previous?.role !== "assistant") params.push({
+						role: "assistant",
+						content: text
+					});
+					else if (Array.isArray(previous.content)) previous.content.push({
+						type: "text",
+						text
+					});
+					else previous.content = previous.content ? `${previous.content}\n${text}` : text;
+				}
 				if (images.length > 0 && model.input.includes("image")) {
 					const boundedToolName = sanitizeSurrogates(truncateUtf16Safe(toolMsg.toolName ?? "", 64));
 					imageContentParts.push({
@@ -639,7 +737,7 @@ function convertMessages(model, context, compat, options = {}) {
 			}
 			i = j - 1;
 			if (imageContentParts.length > 0) {
-				if (compat.requiresAssistantAfterToolResult) params.push({
+				if (supportsTools && compat.requiresAssistantAfterToolResult) params.push({
 					role: "assistant",
 					content: "I have processed the tool results."
 				});
@@ -648,7 +746,7 @@ function convertMessages(model, context, compat, options = {}) {
 					content: imageContentParts
 				});
 				lastRole = "user";
-			} else lastRole = "toolResult";
+			} else lastRole = supportsTools ? "toolResult" : "assistant";
 			continue;
 		}
 		lastRole = msg.role;
@@ -685,7 +783,7 @@ function applyCompletionsReasoningAndRouting(params, model, reasoning, compat, m
 	const direct = mode === "direct";
 	const nativeEffort = compat.thinkingFormat === "openrouter" ? void 0 : reasoning.effort;
 	const enabled = reasoning.thinkingEnabled ?? false;
-	if (model.reasoning) {
+	if (model.reasoning && (!compat.thinkingOffControlUndeclared || reasoning.thinkingEnabled !== void 0 || !reasoning.thinkingExplicit)) {
 		let allowScalarEffort = true;
 		if (direct && compat.thinkingFormat === "zai") {
 			params.thinking = enabled ? {
@@ -697,9 +795,11 @@ function applyCompletionsReasoningAndRouting(params, model, reasoning, compat, m
 			params.enable_thinking = enabled;
 			allowScalarEffort = false;
 		} else if (compat.thinkingFormat === "qwen-chat-template") {
+			const chatTemplateReasoningEffort = enabled && reasoning.thinkingExplicit && model.compat?.supportedReasoningEfforts?.length ? reasoning.effort : void 0;
 			params.chat_template_kwargs = {
 				enable_thinking: enabled,
-				...direct ? { preserve_thinking: true } : {}
+				...direct ? { preserve_thinking: true } : {},
+				...chatTemplateReasoningEffort ? { reasoning_effort: chatTemplateReasoningEffort } : {}
 			};
 			allowScalarEffort = false;
 		} else if (compat.thinkingFormat === "together" || direct && compat.thinkingFormat === "deepseek") {
@@ -878,10 +978,6 @@ function resolveOpenAICompletionsMaxTokens(model, options) {
 		clampToModelMaxTokens: false
 	};
 }
-function resolveOpenAICompletionsModelMaxTokens(model) {
-	const maxTokens = asPositiveFiniteNumber(model.maxTokens);
-	return maxTokens === void 0 ? void 0 : Math.floor(maxTokens);
-}
 const OPENAI_COMPLETIONS_INPUT_TOKEN_SAFETY_MARGIN = 1.25;
 const OPENAI_COMPLETIONS_IMAGE_CHAR_ESTIMATE = 8e3;
 const MIN_USEFUL_OUTPUT_TOKENS = 16;
@@ -893,8 +989,7 @@ function estimateJsonChars(value, fallback) {
 	}
 }
 function estimateOpenAICompletionsInputTokens(payload) {
-	let adjustedChars = 0;
-	adjustedChars += estimateOpenAICompletionsMessagesChars(payload.messages);
+	let adjustedChars = estimateOpenAICompletionsMessagesChars(payload.messages);
 	if (payload.tools?.length) adjustedChars += estimateJsonChars(payload.tools, 1024);
 	if (payload.response_format !== void 0) adjustedChars += estimateJsonChars(payload.response_format, 256);
 	return Math.ceil(adjustedChars / 4 * OPENAI_COMPLETIONS_INPUT_TOKEN_SAFETY_MARGIN);
@@ -944,16 +1039,8 @@ function convertTools(tools, compat, model, mode) {
 		return {
 			projection,
 			tools: sortPromptCacheToolsByName(projection.tools).map((tool) => {
-				const functionTool = {
-					name: tool.name,
-					description: tool.description,
-					parameters: mode === "direct" ? tool.parameters : normalizeOpenAIStrictToolParameters(tool.parameters, strict === true, model.compat)
-				};
-				if (strict !== void 0) functionTool.strict = strict;
-				return {
-					type: "function",
-					function: functionTool
-				};
+				const parameters = compat.requiresTypedUnionBranches ? normalizeToolSchema(tool.parameters, "moonshot") : tool.parameters;
+				return prepareOpenAIFunctionTool(tool.name, tool.description, mode === "direct" ? parameters : normalizeOpenAIStrictToolParameters(parameters, strict === true, model.compat), strict).completions;
 			})
 		};
 	};
@@ -962,7 +1049,22 @@ function convertTools(tools, compat, model, mode) {
 function buildOpenAICompletionsParams(model, context, options) {
 	return buildOpenAICompletionsRequest(model, context, options, { mode: "managed" });
 }
+const contextOutputBudgets = /* @__PURE__ */ new WeakMap();
+function completionsInputHash(payload) {
+	return shortHash(JSON.stringify([
+		payload.messages,
+		payload.tools,
+		payload.response_format
+	]));
+}
+function resolveCompletionsContextOutputBudget(request, payload) {
+	const budget = contextOutputBudgets.get(request);
+	const limits = [payload.max_tokens, payload.max_completion_tokens].filter((value) => value !== void 0);
+	if (!budget || payload.model !== budget.model || !Array.isArray(payload.messages) || payload.tools !== void 0 && !Array.isArray(payload.tools) || limits.length === 0 || limits.some((value) => value !== budget.cap) || completionsInputHash(payload) !== budget.inputHash) return;
+	return budget.cap;
+}
 function buildOpenAICompletionsRequest(model, context, options, policy) {
+	let contextOutputCap;
 	const resolvedPolicy = policy.mode === "direct" ? policy : {
 		...policy,
 		compat: getCompat(model)
@@ -980,8 +1082,10 @@ function buildOpenAICompletionsRequest(model, context, options, policy) {
 		"qwen"
 	].includes(model.provider));
 	const cacheOptOutIndexes = /* @__PURE__ */ new Set();
+	const supportsTools = policy.mode === "direct" || supportsModelTools(model);
 	const convertedMessages = convertMessages(model, context, compat, {
 		cacheOptOutIndexes,
+		supportsTools,
 		preserveSystemPromptCacheBoundary: cacheControl !== void 0 && !managedCompat?.requiresStringContent
 	});
 	let messages = convertedMessages;
@@ -1014,7 +1118,7 @@ function buildOpenAICompletionsRequest(model, context, options, policy) {
 	if (policy.mode === "managed" && options?.seed !== void 0) params.seed = options.seed;
 	if (options?.stop !== void 0 && options.stop.length > 0) params.stop = options.stop;
 	let directToolProjection;
-	if (policy.mode === "direct" || supportsModelTools(model)) {
+	if (supportsTools) {
 		if (context.tools) {
 			const converted = convertTools(context.tools, compat, model, policy.mode);
 			if (policy.mode === "direct") directToolProjection = converted.projection;
@@ -1043,7 +1147,11 @@ function buildOpenAICompletionsRequest(model, context, options, policy) {
 		...model,
 		compat: model.compat ?? void 0
 	}, simpleReasoning)) ?? (usesBinaryOpenRouterThinking ? void 0 : "high");
-	const reasoning = resolveOpenAIRequestReasoning(model, requestedEffort, compat.reasoningEffortForOff);
+	const thinkingExplicit = options?.openclawThinkingExplicit ?? (options?.reasoningEffort !== void 0 || options?.reasoning !== void 0);
+	const reasoning = {
+		...resolveOpenAIRequestReasoning(model, requestedEffort, compat.reasoningEffortForOff),
+		thinkingExplicit
+	};
 	const { effort, thinkingEnabled } = reasoning;
 	{
 		const maxTokenBudget = policy.mode === "direct" ? {
@@ -1053,7 +1161,8 @@ function buildOpenAICompletionsRequest(model, context, options, policy) {
 		const effectiveMaxTokens = maxTokenBudget.maxTokens;
 		const effectiveContextTokens = asPositiveFiniteNumber(model.contextTokens) ?? asPositiveFiniteNumber(model.contextWindow);
 		let clampedMaxTokens = effectiveMaxTokens;
-		const modelMaxTokens = resolveOpenAICompletionsModelMaxTokens(model);
+		const modelOutputLimit = asPositiveFiniteNumber(model.maxTokens);
+		const modelMaxTokens = modelOutputLimit === void 0 ? void 0 : Math.floor(modelOutputLimit);
 		if (maxTokenBudget.clampToModelMaxTokens && clampedMaxTokens !== void 0 && modelMaxTokens !== void 0 && clampedMaxTokens > modelMaxTokens) {
 			clampedMaxTokens = modelMaxTokens;
 			if (policy.mode === "managed") emitModelTransportDebug(log, `[completions] clamp_max_tokens provider=${model.provider} api=${model.api} model=${model.id} requested=${effectiveMaxTokens} output=${clampedMaxTokens} modelMaxTokens=${modelMaxTokens}`);
@@ -1062,8 +1171,9 @@ function buildOpenAICompletionsRequest(model, context, options, policy) {
 			const estimatedInputTokens = estimateOpenAICompletionsInputTokens(params);
 			const remainingBudget = Math.max(0, effectiveContextTokens - estimatedInputTokens - 1);
 			if (clampedMaxTokens > remainingBudget) {
-				if (remainingBudget < MIN_USEFUL_OUTPUT_TOKENS) throw Object.assign(/* @__PURE__ */ new Error(`Context window exceeded: estimated input ${estimatedInputTokens} leaves only ${remainingBudget} output tokens within the ${effectiveContextTokens}-token context.`), { code: "context_length_exceeded" });
+				if (remainingBudget < (params.tools?.length && params.tool_choice !== "none" ? Math.max(MIN_USEFUL_OUTPUT_TOKENS, Math.min(2048, Math.ceil((modelMaxTokens ?? 0) / 8))) : MIN_USEFUL_OUTPUT_TOKENS)) throw Object.assign(/* @__PURE__ */ new Error(`Context window exceeded: estimated input ${estimatedInputTokens} leaves only ${remainingBudget} output tokens within the ${effectiveContextTokens}-token context.`), { code: "context_length_exceeded" });
 				clampedMaxTokens = remainingBudget;
+				contextOutputCap = remainingBudget;
 				emitModelTransportDebug(log, `[completions] clamp_max_tokens provider=${model.provider} api=${model.api} model=${model.id} requested=${effectiveMaxTokens} output=${clampedMaxTokens} effectiveContext=${effectiveContextTokens} estimatedInput=${estimatedInputTokens}`);
 			}
 		}
@@ -1082,6 +1192,11 @@ function buildOpenAICompletionsRequest(model, context, options, policy) {
 		if (isOpenAIGpt56Model(model)) params.reasoning_effort = "none";
 		else if (isOpenAIGpt54MiniModel(model) || isOpenAIGpt55Model(model)) delete params.reasoning_effort;
 	}
+	if (contextOutputCap !== void 0) contextOutputBudgets.set(params, {
+		model: params.model,
+		cap: contextOutputCap,
+		inputHash: completionsInputHash(params)
+	});
 	return params;
 }
 //#endregion
@@ -1218,7 +1333,7 @@ function parseDeepSeekDsmlToolCallBlock(body) {
 			kind: "toolCall",
 			name: invokeName,
 			arguments: parsedArguments,
-			partialArgs: JSON.stringify(parsedArguments)
+			partialJson: JSON.stringify(parsedArguments)
 		});
 	}
 	return toolCalls;
@@ -1379,7 +1494,7 @@ function recoverGemmaCalls(text) {
 				kind: "toolCall",
 				name,
 				arguments: argumentsValue,
-				partialArgs: JSON.stringify(argumentsValue)
+				partialJson: JSON.stringify(argumentsValue)
 			});
 		} catch {
 			return;
@@ -1437,6 +1552,38 @@ function extractToolCallThoughtSignature(toolCall) {
 	const extra = tc.extra_content?.google;
 	return readNonEmptyStringPreservingWhitespace(extra?.thought_signature) ?? readNonEmptyStringPreservingWhitespace(tc.function?.thought_signature) ?? readNonEmptyStringPreservingWhitespace(tc.thought_signature);
 }
+async function* observeOpenAICompletionsProgress(responseStream, signal) {
+	const maxUsageTokens = {
+		prompt: 0,
+		completion: 0,
+		total: 0,
+		reasoning: 0
+	};
+	for await (const rawChunk of responseStream) {
+		throwIfModelStreamAborted(signal);
+		if (rawChunk && typeof rawChunk === "object") {
+			const chunk = rawChunk;
+			const choice = Array.isArray(chunk.choices) ? chunk.choices[0] : void 0;
+			const usage = chunk.usage || choice?.usage;
+			let hasUsageProgress = false;
+			for (const [counter, reported] of [
+				["prompt", usage?.prompt_tokens],
+				["completion", usage?.completion_tokens],
+				["total", usage?.total_tokens],
+				["reasoning", usage?.completion_tokens_details?.reasoning_tokens]
+			]) {
+				const tokens = asPositiveFiniteNumber(reported) ?? 0;
+				if (tokens > maxUsageTokens[counter]) {
+					maxUsageTokens[counter] = tokens;
+					hasUsageProgress = true;
+				}
+			}
+			const rawChoiceDelta = choice?.delta ?? choice?.message;
+			notifyLlmRequestActivity(signal, Boolean(hasUsageProgress || choice?.finish_reason || rawChoiceDelta && (rawChoiceDelta.tool_calls?.length || hasOpenAICompletionsDeltaContent(rawChoiceDelta))));
+		}
+		yield rawChunk;
+	}
+}
 async function processCompletionsStream(responseStream, output, model, stream, options) {
 	const MAX_POST_TOOL_CALL_BUFFER_BYTES = 256e3;
 	const directMode = options?.mode === "direct";
@@ -1453,17 +1600,14 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 	let currentBlock = null;
 	const directContent = { block: null };
 	let currentTextSource;
-	let pendingInterruptedTextBlock = null;
-	let confirmedInterruptedTextBlock = null;
+	let sawReasoning = false;
 	let pendingPostToolCallDeltas = [];
 	let pendingPostToolCallBytes = 0;
 	const toolCallBlocksByIndex = /* @__PURE__ */ new Map();
 	const toolCallBlocksById = /* @__PURE__ */ new Map();
 	const encryptedReasoning = createOpenAIEncryptedToolCallReasoningTracker();
 	const toolArgumentPreviewSchedules = /* @__PURE__ */ new WeakMap();
-	const provisionalCommentaryTags = directMode ? options.provisionalCommentaryTags : /* @__PURE__ */ new Map();
 	const blockIndices = /* @__PURE__ */ new WeakMap();
-	let explicitVisibleTextBlocks;
 	const normalizeToolCallDeltas = createOpenAICompletionsToolCallDeltaNormalizer();
 	let finishReason;
 	let sawNativeToolCallDelta = false;
@@ -1505,7 +1649,6 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 					text: ""
 				};
 				currentTextSource = delta.source;
-				if (delta.source === "reasoning_detail") (explicitVisibleTextBlocks ??= /* @__PURE__ */ new Set()).add(currentBlock);
 			} else currentBlock = {
 				type: "thinking",
 				thinking: "",
@@ -1521,13 +1664,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 			});
 		}
 		if (currentBlock.type === "thinking") appendAssistantThinking(currentBlock, delta.text);
-		else {
-			currentBlock.text += delta.text;
-			if (pendingInterruptedTextBlock && delta.text.trim()) {
-				confirmedInterruptedTextBlock = pendingInterruptedTextBlock;
-				pendingInterruptedTextBlock = null;
-			}
-		}
+		else currentBlock.text += delta.text;
 		const event = {
 			contentIndex: blockIndex(),
 			delta: delta.text
@@ -1579,20 +1716,19 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 			currentBlock = null;
 			flushPendingPostToolCallDeltas();
 		}
-		rememberPendingCommentaryTags(provisionalCommentaryTags, tagPendingCommentaryText(output.content));
 		const block = {
 			type: "toolCall",
 			id: `call_${randomUUID().replaceAll("-", "").slice(0, 24)}`,
 			name: toolCall.name,
 			arguments: toolCall.arguments,
-			partialArgs: toolCall.partialArgs
+			partialJson: toolCall.partialJson
 		};
 		currentBlock = block;
 		appendToolCallBlock(block);
 		pushStreamEvent({
 			type: "toolcall_delta",
 			contentIndex: blockIndices.get(block) ?? -1,
-			delta: toolCall.partialArgs,
+			delta: toolCall.partialJson,
 			partial: output
 		});
 	};
@@ -1623,7 +1759,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		});
 	};
 	const flushReasoningTagTextPartitioner = (allowRecovery = true) => {
-		const recoverUnclosed = allowRecovery && output.stopReason !== "length" && output.stopReason !== "error" && output.stopReason !== "aborted";
+		const recoverUnclosed = allowRecovery && !sawReasoning && output.stopReason !== "length" && output.stopReason !== "error" && output.stopReason !== "aborted";
 		for (const delta of reasoningTagTextPartitioner.flush({ recoverUnclosed })) appendPartitionedVisibleDelta(delta);
 	};
 	const flushGemmaToolCallRecoverer = (allowRecovery = true) => {
@@ -1634,36 +1770,17 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		if (currentBlock?.type !== "text" && !reasoningTagTextPartitioner.hasPending()) return;
 		flushReasoningTagTextPartitioner();
 		if (currentBlock?.type !== "text") return;
-		if (currentTextSource !== "reasoning_detail" && currentBlock.text.trim()) pendingInterruptedTextBlock = currentBlock;
 		currentBlock = null;
 		if (directMode) directContent.block = 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();
+	const beginReasoning = (hasFollowingVisibleText) => {
+		sawReasoning = true;
 		if (!hasFollowingVisibleText || !reasoningTagTextPartitioner.hasPendingSyntax()) sealTextBeforeReasoning();
 	};
-	const guardedStream = withFirstStreamEventTimeout(responseStream, {
-		provider: model.provider,
-		api: model.api,
-		model: model.id,
-		timeoutMs: options?.firstEventTimeoutMs ?? 0,
-		stage: "completions",
-		abort: options?.abortFirstEventStream,
-		onTimeout: options?.onFirstEventTimeout,
-		hint: "The provider may be stalled while parsing the tool payload; retry with a smaller tool surface or enable OPENCLAW_DEBUG_MODEL_PAYLOAD=tools to inspect exposed tools."
-	});
-	const events = directMode ? guardedStream : iterateModelStream(guardedStream, options?.signal);
-	const maxUsageTokens = {
-		prompt: 0,
-		completion: 0,
-		total: 0,
-		reasoning: 0
-	};
+	const events = directMode ? responseStream : iterateModelStream(responseStream, options?.signal);
+	let maxReasoningTokens = 0;
+	let providerProcessingMs;
 	for await (const rawChunk of events) {
 		throwIfModelStreamAborted(options?.signal);
 		chunkPushedEvent = false;
@@ -1673,24 +1790,21 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 		if (typeof chunk.model === "string" && chunk.model.length > 0 && chunk.model !== model.id) output.responseModel ||= chunk.model;
 		const choice = Array.isArray(chunk.choices) ? chunk.choices[0] : void 0;
 		const usage = chunk.usage || choice?.usage;
-		let hasUsageProgress = false;
-		let hasReasoningUsageActivity = false;
-		for (const [counter, reported] of [
-			["prompt", usage?.prompt_tokens],
-			["completion", usage?.completion_tokens],
-			["total", usage?.total_tokens],
-			["reasoning", usage?.completion_tokens_details?.reasoning_tokens]
-		]) {
-			const tokens = asPositiveFiniteNumber(reported) ?? 0;
-			if (tokens > maxUsageTokens[counter]) {
-				maxUsageTokens[counter] = tokens;
-				hasUsageProgress = true;
-				hasReasoningUsageActivity ||= counter === "reasoning";
-			}
+		if (chunk.timings !== void 0 && (choice?.finish_reason || Array.isArray(chunk.choices) && chunk.choices.length === 0 && chunk.usage)) {
+			const promptMs = asNonNegativeFiniteNumber(chunk.timings?.prompt_ms);
+			const predictedMs = asNonNegativeFiniteNumber(chunk.timings?.predicted_ms);
+			providerProcessingMs = promptMs !== void 0 && predictedMs !== void 0 ? asNonNegativeFiniteNumber(promptMs + predictedMs) : void 0;
+		}
+		const reasoningTokens = asPositiveFiniteNumber(usage?.completion_tokens_details?.reasoning_tokens) ?? 0;
+		const hasReasoningUsageActivity = reasoningTokens > maxReasoningTokens;
+		maxReasoningTokens = Math.max(maxReasoningTokens, reasoningTokens);
+		if (usage) {
+			output.usage = parseOpenAICompletionsUsage(usage, model, { includeReasoningTokens: !directMode });
+			const terminalUsage = Boolean(finishReason || choice?.finish_reason || chunk.choices?.length === 0);
+			const completeCounts = [usage.prompt_tokens, usage.completion_tokens].every((value) => typeof value === "number" && Number.isFinite(value) && value >= 0);
+			output.usage.usageReport = { state: terminalUsage && completeCounts ? "available" : "unavailable" };
 		}
-		if (usage) output.usage = parseOpenAICompletionsUsage(usage, model, { includeReasoningTokens: !directMode });
 		const rawChoiceDelta = choice?.delta ?? choice?.message;
-		notifyLlmRequestActivity(options?.signal, Boolean(hasUsageProgress || choice?.finish_reason || rawChoiceDelta && (rawChoiceDelta.tool_calls?.length || hasOpenAICompletionsDeltaContent(rawChoiceDelta))));
 		if (!choice) {
 			emitReasoningUsageActivity(hasReasoningUsageActivity);
 			continue;
@@ -1715,15 +1829,12 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 			const lastVisibleTextIndex = contentDeltas.findLastIndex((delta) => delta.kind === "text");
 			const hasSameChunkVisibleText = reasoningBatch.hasVisibleText || lastVisibleTextIndex !== -1;
 			if (hasReasoningThinking) {
-				beginReasoning(hasSameChunkVisibleText, true);
+				beginReasoning(hasSameChunkVisibleText);
 				appendReasoningDeltas(reasoningDeltas);
 			}
 			for (const [contentDeltaIndex, contentDelta] of contentDeltas.entries()) if (contentDelta.kind === "text") {
 				const parts = gemmaToolCallRecoverer?.push(contentDelta.text) ?? [contentDelta];
-				for (const part of parts) {
-					const routedDeltas = hasReasoningThinking ? reasoningTagTextPartitioner.push(part.text) : reasoningTagTextPartitioner.pushVisible(part.text);
-					for (const routedDelta of routedDeltas) appendPartitionedVisibleDelta(routedDelta);
-				}
+				for (const part of parts) for (const routedDelta of reasoningTagTextPartitioner.pushVisible(part.text)) appendPartitionedVisibleDelta(routedDelta);
 			} else {
 				beginReasoning(contentDeltaIndex < lastVisibleTextIndex);
 				if (emitReasoning) {
@@ -1737,7 +1848,6 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 				flushGemmaToolCallRecoverer(false);
 				sawNativeToolCallDelta = true;
 				flushReasoningTagTextPartitioner(false);
-				rememberPendingCommentaryTags(provisionalCommentaryTags, tagPendingCommentaryText(output.content));
 				for (const toolCall of toolCallDeltas) {
 					const streamIndex = typeof toolCall.index === "number" ? toolCall.index : void 0;
 					let block = streamIndex !== void 0 ? toolCallBlocksByIndex.get(streamIndex) : void 0;
@@ -1755,7 +1865,7 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 							id: toolCall.id || "",
 							name: toolCall.function?.name || "",
 							arguments: {},
-							partialArgs: "",
+							partialJson: "",
 							...initialSig ? { thoughtSignature: initialSig } : {}
 						};
 						encryptedReasoning.rememberToolCall(block.id, block);
@@ -1776,8 +1886,8 @@ async function processCompletionsStream(responseStream, output, model, stream, o
 					if (deltaSig) block.thoughtSignature = deltaSig;
 					const toolArgumentsDelta = toolCall.function?.arguments;
 					if (toolArgumentsDelta) {
-						block.partialArgs += toolArgumentsDelta;
-						if (toolArgumentPreviewSchedules.get(block)?.(block.partialArgs.length)) block.arguments = parseStreamingJson(block.partialArgs);
+						block.partialJson += toolArgumentsDelta;
+						if (toolArgumentPreviewSchedules.get(block)?.(block.partialJson.length)) block.arguments = parseStreamingJson(block.partialJson);
 					}
 					if (toolArgumentsDelta || directMode) pushStreamEvent({
 						type: "toolcall_delta",
@@ -1812,14 +1922,16 @@ 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);
-	if (!output.usage.contextUsage && !options?.signal?.aborted && output.stopReason !== "error" && output.stopReason !== "aborted") {
+	output.usage.usageReport ??= { state: "unavailable" };
+	if ((!output.usage.contextUsage || output.usage.usageReport.state === "unavailable") && !options?.signal?.aborted && output.stopReason !== "error" && output.stopReason !== "aborted") {
 		output.usage.contextUsage = { state: "unavailable" };
 		if (!compat.supportsUsageInStreaming) warnMissingStreamingUsage(openAIModel);
 	}
+	if (providerProcessingMs !== void 0) appendAssistantMessageDiagnostic(output, {
+		type: "provider_timing",
+		timestamp: Date.now(),
+		details: { processingMs: providerProcessingMs }
+	});
 }
 const MAX_MISSING_USAGE_HINT_KEYS = 256;
 const missingUsageHintKeys = /* @__PURE__ */ new Set();
@@ -1920,7 +2032,7 @@ function ensureOpenAIResponsesNonEmptyInput(messages, context) {
 		text: " "
 	}]));
 }
-function buildOpenAIResponsesParams(model, context, options, metadata, replayMode = "checkpoint") {
+function buildOpenAIResponsesParams(model, context, options, metadata, replayMode = "checkpoint", retainUserProvenance = false) {
 	const payloadPolicy = resolveOpenAIResponsesPayloadPolicy(model, { storeMode: "transport-default" });
 	const policyAllowsReplayIds = payloadPolicy.explicitStore !== false && !payloadPolicy.shouldStripStore;
 	const replayResponsesItemIds = !usesNativeOpenAICodexResponsesBackend(model) && (options?.replayResponsesItemIds ?? policyAllowsReplayIds);
@@ -1930,7 +2042,8 @@ function buildOpenAIResponsesParams(model, context, options, metadata, replayMod
 		replayResponsesItemIds,
 		authProfileId: options?.authProfileId,
 		sessionId: options?.sessionId,
-		replayMode
+		replayMode,
+		retainUserProvenance
 	});
 	ensureOpenAIResponsesNonEmptyInput(messages, context);
 	const cacheRetention = resolveCacheRetention(options?.cacheRetention);
@@ -1947,7 +2060,7 @@ function buildOpenAIResponsesParams(model, context, options, metadata, replayMod
 		...metadata ? { metadata } : {}
 	};
 	const effectiveMaxTokens = options?.maxTokens || model.maxTokens;
-	if (effectiveMaxTokens) params.max_output_tokens = Math.max(effectiveMaxTokens, 16);
+	if (effectiveMaxTokens) params.max_output_tokens = clampMaxTokensToModel(model, Math.max(effectiveMaxTokens, 16));
 	if (options?.temperature !== void 0 && supportsOpenAITemperature(model)) params.temperature = options.temperature;
 	if (options?.topP !== void 0 && (!isOpenAIGpt6Model(model) || resolveOpenAIThinkingApi(model.api) === "azure-openai-responses")) params.top_p = options.topP;
 	if (options?.responseFormat !== void 0) params.text = {
@@ -1992,26 +2105,31 @@ function buildOpenAIResponsesParams(model, context, options, metadata, replayMod
 //#endregion
 //#region packages/ai/src/transports/openai-responses-compact-client.ts
 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 instructions = typeof params.request.instructions === "string" && params.request.instructions.length > 0 ? params.request.instructions : void 0;
+	const body = instructions && !supportsNativeOpenAIResponsesEndpoint(params.model) ? {
+		model: params.request.model,
+		input: [buildOpenAIResponsesCompactSystemMessage(params.model, instructions), ...params.request.input ?? []]
+	} : {
+		model: params.request.model,
+		input: params.request.input,
+		...instructions ? { instructions } : {}
+	};
 	const response = await params.client.post("/responses/compact", {
 		...buildOpenAISdkRequestOptions(params.model, params.options?.signal, { timeoutMs: params.options?.timeoutMs }),
-		body: {
-			model: params.request.model,
-			input: compactInput
-		}
+		body
 	});
-	const output = isRecord(response) && Array.isArray(response.output) ? response.output : [];
-	const item = output.at(-1);
-	const retainedItems = output.slice(0, -1);
-	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" || !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");
+	if (!isRecord(response) || response.object !== "response.compaction") throw new Error("Responses compact endpoint must return a response.compaction object");
+	const output = response.output;
+	if (!isOpenAIResponsesCompactionOutput(output, params.model)) throw new Error(`Responses compact endpoint: ${getOpenAIResponsesCompactionOutputError(output, params.model)}`);
+	const window = output;
+	const item = window.at(-1);
+	if (window.length > 1 && !supportsNativeOpenAIResponsesEndpoint(params.model)) throw new Error("Responses compact endpoint: retained messages require native OpenAI");
+	const usage = isRecord(response.usage) ? response.usage : void 0;
+	if (!usage || typeof usage.input_tokens !== "number" || typeof usage.output_tokens !== "number") throw new Error("Responses compact endpoint: usage must include input_tokens and output_tokens");
 	return {
-		output,
+		output: window,
 		item,
-		historyMode: retainedUserMessageCount > 0 ? "retained-users" : "compacted-prefix",
+		historyMode: window.length > 1 ? "retained-users" : "compacted-prefix",
 		usage,
 		model: params.model,
 		replayMetadata: buildOpenAIResponsesReasoningReplayMetadata(params.model, {
@@ -2021,39 +2139,6 @@ async function postOpenAIResponsesCompaction(params) {
 	};
 }
 //#endregion
-//#region packages/ai/src/transports/openai-responses-compact-request.ts
-const COMPACT_REQUEST = Symbol("openaiResponsesCompactRequest");
-function claimResponsesCompactRequest(options) {
-	const controller = options ? Reflect.get(options, COMPACT_REQUEST) : void 0;
-	if (controller?.claimed === false) {
-		controller.claimed = true;
-		return controller;
-	}
-}
-/** Run a compact-endpoint request through the session's prepared stream stack. */
-async function requestPreparedOpenAIResponsesCompaction(streamFn, model, context, options) {
-	const preparedOptions = { ...options };
-	let resolveResult;
-	let rejectResult;
-	const result = new Promise((resolve, reject) => {
-		resolveResult = resolve;
-		rejectResult = reject;
-	});
-	const controller = {
-		claimed: false,
-		resolve: resolveResult,
-		reject: rejectResult
-	};
-	Reflect.set(preparedOptions, COMPACT_REQUEST, controller);
-	const stream = await Promise.resolve(streamFn(model, context, preparedOptions));
-	if (!controller.claimed) throw new Error("Prepared stream did not reach an OpenAI Responses transport");
-	try {
-		return await result;
-	} finally {
-		await stream.result().catch(() => void 0);
-	}
-}
-//#endregion
 //#region packages/ai/src/transports/openai-responses-reasoning-state.ts
 /** Save only admitted settings and hashes, never another copy of the conversation. */
 function recordResponsesReasoningState(message, model, identity, request, output) {
@@ -2314,6 +2399,16 @@ function createRecoverableResponsesWebSocketStream(params) {
 const SESSION_WEBSOCKET_CACHE_TTL_MS = 3e5;
 const SESSION_WEBSOCKET_MAX_AGE_MS = 33e5;
 const WEBSOCKET_OPEN_STATE = 1;
+function resolveNativeOpenAIResponsesWebSocketMode(model, transport) {
+	if (transport !== "websocket" && transport !== "websocket-cached" && transport !== "auto") return;
+	if (getAiTransportHost().requiresManagedTransport(model)) return;
+	return supportsNativeOpenAIResponsesEndpoint(model) ? transport : void 0;
+}
+function combineWebSocketTimeoutSignal(signal, model, timeoutMs) {
+	const resolvedTimeoutMs = timeoutMs !== void 0 && Number.isFinite(timeoutMs) && timeoutMs > 0 ? timeoutMs : getAiTransportHost().resolveModelRequestTimeoutMs(model);
+	if (resolvedTimeoutMs === void 0 || !Number.isFinite(resolvedTimeoutMs)) return signal;
+	return AbortSignal.any([signal, AbortSignal.timeout(Math.max(1, resolvedTimeoutMs))]);
+}
 const websocketSessionCache = /* @__PURE__ */ new Map();
 const degradedWebSocketConnections = /* @__PURE__ */ new Map();
 function closeWebSocketSilently(socket, reason = "done") {
@@ -2675,16 +2770,6 @@ function closeOpenAIResponsesWebSocketSessions(sessionId, owner) {
 registerSessionResourceCleanup(closeOpenAIResponsesWebSocketSessions);
 //#endregion
 //#region packages/ai/src/transports/openai-responses-client.ts
-function resolveNativeOpenAIResponsesWebSocketMode(model, transport) {
-	if (transport !== "websocket" && transport !== "websocket-cached" && transport !== "auto") return;
-	if (getAiTransportHost().requiresManagedTransport(model)) return;
-	return supportsNativeOpenAIResponsesEndpoint(model) ? transport : void 0;
-}
-function combineWebSocketTimeoutSignal(signal, model, timeoutMs) {
-	const resolvedTimeoutMs = timeoutMs !== void 0 && Number.isFinite(timeoutMs) && timeoutMs > 0 ? timeoutMs : getAiTransportHost().resolveModelRequestTimeoutMs(model);
-	if (resolvedTimeoutMs === void 0 || !Number.isFinite(resolvedTimeoutMs)) return signal;
-	return AbortSignal.any([signal, AbortSignal.timeout(Math.max(1, resolvedTimeoutMs))]);
-}
 function createOpenAIResponsesClient(model, apiKey, defaultHeaders, fetchOverride) {
 	return new OpenAI({
 		apiKey,
@@ -2698,7 +2783,9 @@ function createOpenAIResponsesClient(model, apiKey, defaultHeaders, fetchOverrid
 function createResponsesTransportExecutor(config) {
 	return (model, context, options) => {
 		const responsesOptions = options;
+		const independent = independentResponsesRequest.has(options);
 		const compactRequest = claimResponsesCompactRequest(responsesOptions);
+		const v2Compaction = compactRequest?.mode === "v2" ? createResponsesV2CompactionCollector(compactRequest.replayBudget) : void 0;
 		const { eventStream, stream } = createWritableTransportEventStream();
 		(async () => {
 			const output = createAssistantOutput(model, config.outputApi);
@@ -2706,9 +2793,10 @@ function createResponsesTransportExecutor(config) {
 			let continuationClaim;
 			const requestLifecycle = responsesRequestLifecycle.get(options);
 			try {
+				if (compactRequest && requestLifecycle) throw new Error("Provider review continuation cannot replace its reviewed input with compaction");
 				const apiKey = options?.apiKey || getEnvApiKey(model.provider) || "";
-				const websocketMode = resolveNativeOpenAIResponsesWebSocketMode(model, responsesOptions?.transport);
-				const encodeBody = prepareModelRequestBody(websocketMode || compactRequest ? void 0 : options);
+				const websocketMode = independent ? void 0 : resolveNativeOpenAIResponsesWebSocketMode(model, responsesOptions?.transport);
+				const encodeBody = prepareModelRequestBody(websocketMode || compactRequest?.mode === "endpoint" ? void 0 : options);
 				const turnState = resolveProviderTransportTurnState(model, {
 					sessionId: options?.sessionId,
 					turnId: randomUUID(),
@@ -2720,21 +2808,26 @@ function createResponsesTransportExecutor(config) {
 				const websocketTurnHeaders = filterProviderTurnHeadersForExplicitOpencodeSession(model, options, websocketSessionPolicy?.headers);
 				const websocketHeaders = websocketMode ? buildOpenAIClientHeaders(model, context, options?.headers, websocketTurnHeaders, options?.sessionId, options?.cacheRetention) : void 0;
 				const httpHeaders = buildOpenAIClientHeaders(model, context, options?.headers, httpTurnHeaders, options?.sessionId, options?.cacheRetention);
-				const fetchOverride = createResponsesRequestFetch(model, {
-					compact: Boolean(compactRequest),
+				const fetchOverride = withOpenAIErrorResponseHook(createResponsesRequestFetch(model, {
+					compact: compactRequest?.mode === "endpoint",
 					stream: config.streamRequest,
 					lifecycle: requestLifecycle,
 					onSseComment: () => notifyLlmRequestActivity(options?.signal, false)
-				});
+				}) ?? buildGuardedModelFetch(model), options?.onResponse, model);
 				const client = config.createClient(model, apiKey, httpHeaders, fetchOverride);
 				const nativeAstra = model.id === "gpt-6-astra" && supportsNativeOpenAIResponsesEndpoint(model);
 				const asyncToolExecutionEligible = nativeAstra && options?.asyncToolExecution === true && !responsesOptions?.openclawCodeModeToolSurface;
 				const prepareRequest = async (request) => {
 					let params = request;
+					const preserveUsers = v2Compaction?.preserveUsers(request.input ?? []);
 					const nextParams = await options?.onPayload?.(params, model);
 					if (nextParams !== void 0) params = nextParams;
+					if (preserveUsers) params = {
+						...params,
+						input: preserveUsers(params.input ?? [])
+					};
 					params = sanitizeOpenAICodexResponsesParams(model, params);
-					params = sanitizeResponsesImagePayload(params);
+					params = v2Compaction ? v2Compaction.sanitizeImages(params) : sanitizeResponsesImagePayload(params);
 					if (responsesOptions?.openclawCodeModeToolSurface === true) {
 						const visibleToolNames = resolveCodeModeResponsesVisibleToolNames(context);
 						const allowedHostedToolTypes = responsesOptions?.openclawCodeModeAllowedHostedToolTypes;
@@ -2748,12 +2841,16 @@ function createResponsesTransportExecutor(config) {
 							async: true
 						} : tool);
 					}
+					if (v2Compaction) {
+						params.input = [...params.input ?? [], { type: "compaction_trigger" }];
+						v2Compaction.assertReplayFits(params.input, model);
+					}
 					return params;
 				};
-				const buildRequest = (replayMode, requestContext = context) => prepareRequest(config.buildRequest(model, requestContext, responsesOptions, turnState?.metadata, replayMode));
+				const buildRequest = (replayMode, requestContext = context) => prepareRequest(config.buildRequest(model, requestContext, responsesOptions, turnState?.metadata, replayMode, Boolean(v2Compaction)));
 				let params = await buildRequest("checkpoint");
-				const asyncTools = asyncToolExecutionEligible && params.model === "gpt-6-astra" && params.multi_agent?.enabled !== true;
-				if (compactRequest) {
+				const asyncTools = !compactRequest && asyncToolExecutionEligible && params.model === "gpt-6-astra" && params.multi_agent?.enabled !== true;
+				if (compactRequest?.mode === "endpoint") {
 					const compacted = await postOpenAIResponsesCompaction({
 						client,
 						model,
@@ -2763,6 +2860,7 @@ function createResponsesTransportExecutor(config) {
 					output.usage.input = compacted.usage.input_tokens;
 					output.usage.output = compacted.usage.output_tokens;
 					output.usage.totalTokens = compacted.usage.input_tokens + compacted.usage.output_tokens;
+					output.usage.usageReport = { state: "available" };
 					compactRequest.resolve(compacted);
 					finalizeTransportStream({
 						stream,
@@ -2772,16 +2870,17 @@ function createResponsesTransportExecutor(config) {
 					return;
 				}
 				const sessionId = options?.sessionId;
-				if (config.httpContinuation && !websocketMode && !getAiTransportHost().requiresManagedTransport(model) && (supportsNativeOpenAIResponsesEndpoint(model) || resolveOpenAIResponsesPayloadPolicy(model).explicitContinuationOptIn) && sessionId && (params.store === true || supportsResponsesReasoningUpdate(params)) && !params.previous_response_id) {
+				if (config.httpContinuation && !compactRequest && !independent && !websocketMode && !getAiTransportHost().requiresManagedTransport(model) && (supportsNativeOpenAIResponsesEndpoint(model) || resolveOpenAIResponsesPayloadPolicy(model).explicitContinuationOptIn) && sessionId && (params.store === true || supportsResponsesReasoningUpdate(params)) && !params.previous_response_id) {
+					const restoreRequest = () => restoreResponsesReasoningState(context, model, responsesOptions, params);
 					continuationClaim = claimOpenAIResponsesHttpContinuation({
 						sessionId,
 						apiKey,
 						baseUrl: model.baseUrl,
 						headers: httpHeaders,
 						request: params,
-						restoreRequest: () => restoreResponsesReasoningState(context, model, responsesOptions, params)
+						restoreRequest
 					});
-					if (continuationClaim) params = continuationClaim.fullRequest;
+					params = continuationClaim?.fullRequest ?? restoreRequest();
 				}
 				const observePrompt = createResponsesPromptEgressObserver(responsesOptions, context.systemPrompt);
 				const requestStartedAt = Date.now();
@@ -2829,7 +2928,7 @@ function createResponsesTransportExecutor(config) {
 							continuationBaseline = attempt.request.previous_response_id ? params : attempt.request;
 							const trackedResponseStream = responseModelTracker.track(response, rawResponseStream);
 							return withProviderResponseHook({
-								stream: observeResponsesStream(trackedResponseStream, model, requestStartedAt),
+								stream: observeResponsesStream(v2Compaction ? v2Compaction.observe(trackedResponseStream) : trackedResponseStream, model, requestStartedAt, attempt.request),
 								signal: firstEvent.signal,
 								abort: firstEvent.abort,
 								hook: createOpenAIProviderAcceptanceHook(options, response, model),
@@ -2864,7 +2963,7 @@ function createResponsesTransportExecutor(config) {
 						signal: websocketSignal,
 						callerSignal: options?.signal,
 						degradeCooldownMs: websocketSessionPolicy?.degradeCooldownMs,
-						onActiveResponse: nativeAstra && params.model === "gpt-6-astra" ? options?.onActiveResponse : void 0,
+						onActiveResponse: !compactRequest && nativeAstra && params.model === "gpt-6-astra" ? options?.onActiveResponse : void 0,
 						steeringInput: (messages) => {
 							contextUsageEligible = false;
 							return projectResponsesSteeringInput(params, () => buildRequest("checkpoint", {
@@ -2885,7 +2984,7 @@ 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 = createRecoverableResponsesWebSocketStream({
-						trackedWebSocketStream: responseModelTracker.track(void 0, websocket.stream),
+						trackedWebSocketStream: observeResponsesStream(v2Compaction ? v2Compaction.observe(responseModelTracker.track(void 0, websocket.stream)) : responseModelTracker.track(void 0, websocket.stream), model, requestStartedAt, websocket.request, "websocket"),
 						websocket,
 						websocketSignal,
 						options,
@@ -2911,7 +3010,7 @@ function createResponsesTransportExecutor(config) {
 						canRetryIdentityConflict: () => hasOnlyResponsesFunctionTools(transport === "websocket" ? websocketBaseline : continuationBaseline),
 						...config.pricingOptions?.(responsesOptions, model),
 						resolveServiceTier: (responseTier, originalTier) => responseTier ?? (requestedTier === "priority" || requestedTier === "default" ? requestedTier : originalTier),
-						onServiceTier: (responseTier) => responsesServiceTierObserver.observe(options, requestedTier, responseTier),
+						onServiceTier: (tier, response) => responsesServiceTierObserver.observe(options, requestedTier, tier, response),
 						firstEventTimeoutMs: getFirstStreamEventTimeoutMs(options) ?? config.firstEventTimeoutMs,
 						abortFirstEventStream: firstEvent.abort,
 						onFirstEventTimeout: getFirstStreamEventTimeoutHandler(options),
@@ -2926,7 +3025,25 @@ function createResponsesTransportExecutor(config) {
 					finishWebSocket?.();
 					if (options?.signal?.aborted) throw transportAbortError(options.signal);
 					if (output.stopReason === "aborted" || output.stopReason === "error") throw new Error(output.errorMessage ?? "An unknown error occurred");
-					const admitted = transport === "websocket" ? websocketBaseline : continuationBaseline;
+					const admitted = !independent && (transport === "websocket" ? websocketBaseline : continuationBaseline);
+					if (v2Compaction && compactRequest) {
+						const compacted = v2Compaction.finish({
+							terminal,
+							input: params.input ?? [],
+							model,
+							options: responsesOptions
+						});
+						compactRequest.resolve({
+							...compacted,
+							modelUsage: output.usage
+						});
+						finalizeTransportStream({
+							stream,
+							output,
+							signal: options?.signal
+						});
+						return;
+					}
 					if (terminal && admitted && supportsNativeOpenAIResponsesEndpoint(model)) recordResponsesReasoningState(output, model, responsesOptions, admitted, terminal.output);
 					if (continuationClaim && continuationBaseline && terminal) continuationClaim.commit(continuationBaseline, terminal, dispatchedPreviousResponseId);
 					if (terminal && admitted && contextUsageEligible) recordResponsesContextUsage(output, model, responsesOptions, admitted, terminal.output, "transport");
@@ -3158,7 +3275,7 @@ function streamOpenAICompletionsRequest(model, context, options, mode) {
 	const { eventStream, stream } = createWritableTransportEventStream();
 	(async () => {
 		const output = createAssistantOutput(model);
-		const provisionalCommentaryTags = /* @__PURE__ */ new Map();
+		let discardCandidate = false;
 		let firstEventAbort;
 		try {
 			const apiKey = options?.apiKey || getEnvApiKey(model.provider) || "";
@@ -3170,10 +3287,11 @@ function streamOpenAICompletionsRequest(model, context, options, mode) {
 				cacheRetention
 			} : { mode };
 			const { client, sawStreamDONE } = policy.mode === "direct" ? {
-				client: createDirectCompletionsClient(model, context, apiKey, resolveProviderSimpleCompletionHeaders(model, options), cacheRetention === "none" ? void 0 : options?.sessionId, policy.compat),
+				client: createDirectCompletionsClient(model, context, apiKey, options, cacheRetention === "none" ? void 0 : options?.sessionId, policy.compat),
 				sawStreamDONE: void 0
 			} : createManagedCompletionsClient(model, context, options, apiKey, cacheRetention);
-			let params = buildOpenAICompletionsRequest(model, context, options, policy);
+			const builtParams = buildOpenAICompletionsRequest(model, context, options, policy);
+			let params = builtParams;
 			const encodeBody = prepareModelRequestBody(options);
 			const nextParams = await options?.onPayload?.(params, model);
 			if (nextParams !== void 0) params = nextParams;
@@ -3185,6 +3303,7 @@ function streamOpenAICompletionsRequest(model, context, options, mode) {
 				}
 				if (getCompat(model).requiresNonEmptyUserOrAssistantMessage) assertOpenAICompletionsPayloadHasConversationTurn(params, model);
 			}
+			const contextOutputBudget = resolveCompletionsContextOutputBudget(builtParams, params);
 			const emitReasoning = mode === "direct" ? directEmitReasoning : shouldEmitOpenAICompletionsReasoning(model, options);
 			firstEventAbort = createFirstStreamEventAbortController(options?.signal);
 			const requestOptions = mode === "direct" ? {
@@ -3214,27 +3333,42 @@ function streamOpenAICompletionsRequest(model, context, options, mode) {
 				signal: firstEventAbort.signal,
 				abort: firstEventAbort.abort,
 				hook: createOpenAIProviderAcceptanceHook(options, response, model),
-				onReady: () => stream.push({
-					type: "start",
-					partial: output
-				})
+				onReady: () => {
+					if (contextOutputBudget === void 0) stream.push({
+						type: "start",
+						partial: output
+					});
+				}
+			});
+			const providerStream = observeOpenAICompletionsProgress(withFirstStreamEventTimeout(hookedResponseStream, {
+				provider: model.provider,
+				api: model.api,
+				model: model.id,
+				timeoutMs: getFirstStreamEventTimeoutMs(options) ?? 0,
+				stage: "completions",
+				abort: firstEventAbort.abort,
+				onTimeout: getFirstStreamEventTimeoutHandler(options),
+				hint: "The provider may be stalled while parsing the tool payload; retry with a smaller tool surface or enable OPENCLAW_DEBUG_MODEL_PAYLOAD=tools to inspect exposed tools."
+			}), options?.signal);
+			const buffered = contextOutputBudget === void 0 ? void 0 : await bufferContextLimitedCompletions(providerStream, options?.signal);
+			discardCandidate = Boolean(buffered?.bounded && (buffered.failed || buffered.finishReason === "length"));
+			if (buffered && !discardCandidate) stream.push({
+				type: "start",
+				partial: output
 			});
 			const directEvents = mode === "direct" ? createDirectCompletionsEventStream(output, stream) : void 0;
 			try {
-				await processCompletionsStream(hookedResponseStream, output, model, directEvents?.stream ?? stream, {
+				await processCompletionsStream(buffered?.stream ?? providerStream, output, model, discardCandidate ? { push() {} } : directEvents?.stream ?? stream, {
 					...directEvents ? {
 						mode: "direct",
-						beforeContentBlock: directEvents.beforeContentBlock,
-						provisionalCommentaryTags
+						beforeContentBlock: directEvents.beforeContentBlock
 					} : { mode: "managed" },
-					signal: options?.signal,
+					signal: discardCandidate ? void 0 : options?.signal,
 					emitReasoning,
 					strictReasoningTags: reasoningTagTextPolicy.isStrict(options),
-					firstEventTimeoutMs: getFirstStreamEventTimeoutMs(options),
-					abortFirstEventStream: firstEventAbort.abort,
-					onFirstEventTimeout: getFirstStreamEventTimeoutHandler(options),
 					sawStreamDONE
 				});
+				if (discardCandidate) throw new Error(`Context length exceeded: automatic output budget of ${contextOutputBudget} tokens exhausted before completion.`);
 				if (directEvents) {
 					if (options?.signal?.aborted) throw transportAbortError(options.signal);
 					if (output.stopReason === "aborted" || output.stopReason === "error") throw new Error(output.errorMessage || (output.stopReason === "aborted" ? "Request was aborted" : "Provider returned an invalid tool call"));
@@ -3250,6 +3384,7 @@ function streamOpenAICompletionsRequest(model, context, options, mode) {
 				signal: options?.signal
 			});
 		} catch (error) {
+			if (discardCandidate) output.content = [];
 			failTransportStream({
 				stream,
 				output,
@@ -3257,11 +3392,9 @@ function streamOpenAICompletionsRequest(model, context, options, mode) {
 				error,
 				cleanup: () => {
 					finalizeOpenAICompletionsToolCalls(output, { allowSilentToolCallPromotion: false });
-					clearPendingCommentaryText(provisionalCommentaryTags);
-					tagUnresolvedTextAsCommentary(output);
 					if (mode === "direct") for (const block of output.content) {
 						delete block.index;
-						delete block.partialArgs;
+						if (block.type === "toolCall") delete block.partialJson;
 						delete block.streamIndex;
 					}
 				}
@@ -3272,7 +3405,7 @@ function streamOpenAICompletionsRequest(model, context, options, mode) {
 	})();
 	return eventStream;
 }
-function createDirectCompletionsClient(model, context, apiKey, optionsHeaders, sessionId, compat) {
+function createDirectCompletionsClient(model, context, apiKey, options, sessionId, compat) {
 	if (!apiKey) throw new Error(`No API key for provider: ${model.provider}`);
 	const headers = { ...model.headers };
 	if (model.provider === "github-copilot") Object.assign(headers, buildCopilotDynamicHeaders(context.messages));
@@ -3284,7 +3417,7 @@ function createDirectCompletionsClient(model, context, apiKey, optionsHeaders, s
 			headers["x-session-affinity"] = sessionId;
 		}
 	}
-	return createOpenAIProviderClient(model, apiKey, headers, optionsHeaders);
+	return createOpenAIProviderClient(model, apiKey, headers, resolveProviderSimpleCompletionHeaders(model, options), options?.onResponse);
 }
 function createManagedCompletionsClient(model, context, options, apiKey, cacheRetention) {
 	const turnState = resolveProviderTransportTurnState(model, {
@@ -3296,7 +3429,7 @@ function createManagedCompletionsClient(model, context, options, apiKey, cacheRe
 	const optionHeaders = resolveOpencodeSessionHeaders(model, options);
 	const turnHeaders = filterProviderTurnHeadersForExplicitOpencodeSession(model, options, turnState?.headers);
 	const doneDetector = createSseDoneDetector();
-	const baseFetch = buildGuardedModelFetch(model, void 0, { onSseComment: () => notifyLlmRequestActivity(options?.signal, false) });
+	const baseFetch = withOpenAIErrorResponseHook(buildGuardedModelFetch(model, void 0, { onSseComment: () => notifyLlmRequestActivity(options?.signal, false) }), options?.onResponse, model);
 	const doneDetectingFetch = async (url, init) => {
 		const response = await baseFetch(url, init);
 		if (!response.body || !response.ok) return response;
@@ -3330,4 +3463,4 @@ function createManagedCompletionsClient(model, context, options, apiKey, cacheRe
 	};
 }
 //#endregion
-export { requestPreparedOpenAIResponsesCompaction as a, convertMessages as c, applyCompletionsAnthropicCacheControl as d, hasOpenAICompatibleConversationTurn as f, createDeepSeekTextFilter as h, createOpenAIResponsesTransportStreamFn as i, flattenCompletionMessagesToStringContent as l, resolveMaxTokensParam as m, streamOpenAICompletionsRequest as n, hasResponsesWebSearchTool as o, canonicalizeMaxTokensParam as p, createAzureOpenAIResponsesTransportStreamFn as r, buildOpenAICompletionsParams as s, createOpenAICompletionsTransportStreamFn as t, stripCompletionMessagesToRoleContent as u };
+export { hasResponsesWebSearchTool as a, flattenCompletionMessagesToStringContent as c, hasOpenAICompatibleConversationTurn as d, canonicalizeMaxTokensParam as f, createOpenAIResponsesTransportStreamFn as i, stripCompletionMessagesToRoleContent as l, createDeepSeekTextFilter as m, streamOpenAICompletionsRequest as n, buildOpenAICompletionsParams as o, resolveMaxTokensParam as p, createAzureOpenAIResponsesTransportStreamFn as r, convertMessages as s, createOpenAICompletionsTransportStreamFn as t, applyCompletionsAnthropicCacheControl as u };