@openclaw/ai 2026.10.5 → 2026.10.10-beta.2

package/dist/{openai-responses-shared-afo9-s16.mjs → openai-responses-shared-Bv_1NtH9.mjs}

@@ -1,45 +1,31 @@
-import "./src-C44MKrW5.mjs";
-import { _ as runtimeContextContentToText, f as hasRuntimeContextMarker, p as isRuntimeContextMessage } from "./types-CQc6JNG4.mjs";
+import "./src-CyA0zJnt.mjs";
+import { _ as runtimeContextContentToText, f as hasRuntimeContextMarker, p as isRuntimeContextMessage } from "./types-BXd0kI8-.mjs";
 import { o as readProviderRefusalReview, t as appendAssistantMessageDiagnostic } from "./diagnostics-DzrF1h4J.mjs";
 import { r as appendAssistantThinking } from "./event-stream-X2gmsUWv.mjs";
 import { o as isRecord, r as asOptionalObjectRecord } from "./record-coerce-CNDovvft.mjs";
-import { a as normalizeOptionalString, c as readStringValue, n as normalizeLowercaseStringOrEmpty, o as normalizeStringifiedOptionalString } from "./string-coerce-CtgOVOSv.mjs";
-import { t as asFiniteNumber } from "./number-coercion-C3hztWoY.mjs";
-import { D as applyProviderReportedUsageCost, E as headersToRecord, N as isImageWithMediaPayload, O as calculateCost, a as resolveAiTransportHeaderSentinels, k as clampThinkingLevel, r as getAiTransportHost } from "./host-Bf3NjYNj.mjs";
-import { A as MODEL_CATALOG_THINKING_LEVELS, C as normalizeOpenAIReasoningEffort, E as resolveOpenAIReasoningEffortMapping, M as resolveOpenAIThinkingApi, O as supportsOpenAIReasoningEffort, T as resolveOpenAIReasoningEffortForModel, _ as isResponsesServiceTierRejection, a as OPENAI_RESPONSES_COMPACTION_REPLAY_TYPE, c as OPENAI_RESPONSES_RETAINED_COMPACTION_REPLAY_TYPE, f as RESPONSE_FAILED_NO_DETAILS_MESSAGE, g as responsesServiceTierObserver, i as OPENAI_RESPONSES_APIS, k as supportsOpenAITemperature, o as OPENAI_RESPONSES_REASONING_REPLAY_BLOCK_META_KEY, p as isPreviousResponseRejection, s as OPENAI_RESPONSES_REASONING_REPLAY_META_KEY, v as nextResponsesServiceTier, w as resolveOpenAIModelReasoningEfforts } from "./openai-responses-contracts-DlKMMTxK.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 { n as clampOpenAIPromptCacheKey } from "./openai-prompt-cache-r93BJhjm.mjs";
-import { _ as resolveResponsesTerminalStopReason, a as resolveModelPayloadDebugMode, d as hasOnlyResponsesFunctionTools, f as isResponsesProviderTool, g as readResponsesReasoningTokens, h as mapResponsesTerminalUsage, o as resolveModelSseDebugMode, r as emitModelTransportDebug, u as ResponsesOutputIdentityError } from "./diagnostics-tQasYct1.mjs";
-import { a as redactSensitiveText, r as isCodeModeModelVisibleToolName } from "./transport-utils-34BQoXq3.mjs";
-import { c as resolveOpenAIResponsesTextFormat, i as resolveOpenAICompletionsCompat, o as usesNativeOpenAICodexResponsesBackend } from "./openai-completions-compat-CYX18uDk.mjs";
-import { r as resolveModelRequestTimeoutMs } from "./host-policy-BzZluO6_.mjs";
-import { A as shortHash, C as parseJsonObjectPreservingUnsafeIntegers, E as createToolArgumentPreviewSchedule, O as parseStreamingJson, S as withProviderResponseHook, b as transportAbortError, c as failTransportStream, d as iterateModelStream, g as parseTerminalToolCallArguments, m as notifyProviderHttpResponse, t as IncompleteToolCallError, u as finalizeTransportStream, v as sanitizeTransportPayloadText } from "./transport-stream-shared-CWGOuLcA.mjs";
-import { C as stripSystemPromptCacheBoundary, D as sortPromptCacheToolsByName, d as describeToolResultMediaPlaceholder, m as extractToolResultText, r as clampMaxTokensToModel, s as prepareModelRequestBody } from "./simple-options-BAu4B54D.mjs";
-import { i as providerReplayContextMatches, n as buildProviderReplayContext, r as isProviderReplayContext, t as encodeTextSignatureV1 } from "./text-signature-CUXnvk_c.mjs";
-import { n as transformProviderMessages } from "./assistant-output-BuhRpLeD.mjs";
-import { n as projectRuntimeToolInputSchema, t as prepareRuntimeToolInputSchema } from "./tool-schema-json-projection-B8sG1I9v.mjs";
+import { l as readStringValue, o as normalizeOptionalString, r as normalizeLowercaseStringOrEmpty, s as normalizeStringifiedOptionalString } from "./string-coerce-Cj8sqWHx.mjs";
+import { t as asFiniteNumber } from "./number-coercion-37gmNkuA.mjs";
+import { D as applyProviderReportedUsageCost, E as headersToRecord, N as isImageWithMediaPayload, O as calculateCost, a as resolveAiTransportHeaderSentinels, k as clampThinkingLevel, r as getAiTransportHost } from "./host-B2gk-d1H.mjs";
+import { A as supportsOpenAITemperature, D as resolveOpenAIReasoningEffortMapping, E as resolveOpenAIReasoningEffortForModel, L as resolveOpenAIThinkingApi, P as MODEL_CATALOG_THINKING_LEVELS, T as resolveOpenAIModelReasoningEfforts, _ as responsesServiceTierObserver, a as OPENAI_RESPONSES_COMPACTION_REPLAY_TYPE, c as OPENAI_RESPONSES_RETAINED_COMPACTION_REPLAY_TYPE, f as RESPONSES_RETAINED_USER, i as OPENAI_RESPONSES_APIS, k as supportsOpenAIReasoningEffort, m as isPreviousResponseRejection, o as OPENAI_RESPONSES_REASONING_REPLAY_BLOCK_META_KEY, p as RESPONSE_FAILED_NO_DETAILS_MESSAGE, s as OPENAI_RESPONSES_REASONING_REPLAY_META_KEY, v as isResponsesServiceTierRejection, w as normalizeOpenAIReasoningEffort, 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 { n as clampOpenAIPromptCacheKey } from "./openai-prompt-cache-DJ_P143P.mjs";
+import { _ as resolveResponsesTerminalStopReason, a as resolveModelPayloadDebugMode, d as hasOnlyResponsesFunctionTools, f as isResponsesProviderTool, g as readResponsesReasoningTokens, h as mapResponsesTerminalUsage, r as emitModelTransportDebug, u as ResponsesOutputIdentityError } from "./diagnostics-Cv7cW3pS.mjs";
+import { c as redactSensitiveText, n as isProviderReplayContext, o as isCodeModeModelVisibleToolName, r as providerReplayContextMatches, t as buildProviderReplayContext } from "./provider-replay-context-J6CfU3m1.mjs";
+import { c as resolveOpenAIResponsesTextFormat, i as resolveOpenAICompletionsCompat, o as usesNativeOpenAICodexResponsesBackend } from "./openai-completions-compat-CFhQ3UVU.mjs";
+import { r as resolveModelRequestTimeoutMs } from "./host-policy-mpqGhbeM.mjs";
+import { r as sortPromptCacheToolsByName } from "./prompt-cache-stability-D_MeCQFI.mjs";
+import { d as extractToolResultBlockText, f as extractToolResultText, h as preparedRequestSuffix, l as describeToolResultMediaPlaceholder, r as clampMaxTokensToModel, s as prepareModelRequestBody, w as stripSystemPromptCacheBoundary } from "./simple-options-BYgC6kse.mjs";
+import { A as shortHash, C as parseJsonObjectPreservingUnsafeIntegers, E as createToolArgumentPreviewSchedule, O as parseStreamingJson, S as withProviderResponseHook, b as transportAbortError, c as failTransportStream, d as iterateModelStream, g as parseTerminalToolCallArguments, m as notifyProviderHttpResponse, t as IncompleteToolCallError, u as finalizeTransportStream, v as sanitizeTransportPayloadText } from "./transport-stream-shared-CzQlAzJH.mjs";
+import { C as redactIdentifier, n as prepareRuntimeToolInputSchema, r as projectRuntimeToolInputSchema, w as sha256Hex } from "./tool-schema-json-projection-B0hHDV7N.mjs";
+import { n as transformProviderMessages } 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 { a as withFirstStreamEventTimeout, i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-Dc42hKv6.mjs";
-import { a as resolveOpenAIProjectedToolsStrictToolFlag, f as withPreparedToolSchemaNormalization, r as normalizeOpenAIStrictToolParameters, t as findOpenAIStrictToolProjectionDiagnostics } from "./openai-tool-schema-DDPXsvJP.mjs";
-import { a as readOpenAIResponsesCompactionWindow, i as isOpenAIResponsesCompactionOutput, n as supportsNativeOpenAIResponsesEndpoint } from "./openai-responses-endpoint-CinGh4VZ.mjs";
-import { hash, randomUUID } from "node:crypto";
-//#region packages/normalization-core/src/node-crypto.ts
-function sha256Hex(input) {
-	return hash("sha256", input, "hex");
-}
-function sha256HexPrefixCore(input, length) {
-	return sha256Hex(input).slice(0, length);
-}
-/** Redacts an identifier to a stable hash label, or "-" for missing values. */
-function redactIdentifier(value, opts) {
-	const trimmed = normalizeOptionalString(value);
-	if (!trimmed) return "-";
-	const len = opts?.len ?? 12;
-	return `sha256:${sha256HexPrefixCore(trimmed, Number.isFinite(len) ? Math.max(1, Math.floor(len)) : 12)}`;
-}
-//#endregion
+import { a as withFirstStreamEventTimeout, i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-D_TfJHRJ.mjs";
+import { a as resolveOpenAIProjectedToolsStrictToolFlag, b as prepareOpenAIFunctionTool, p as withPreparedToolSchemaNormalization, r as normalizeOpenAIStrictToolParameters, t as findOpenAIStrictToolProjectionDiagnostics } from "./openai-tool-schema-Buv0GXFb.mjs";
+import { a as isOpenAIResponsesCompactionOutput, n as supportsNativeOpenAIResponsesEndpoint, o as readOpenAIResponsesCompactionWindow } from "./openai-responses-endpoint-D4v-eeYy.mjs";
+import { randomUUID } from "node:crypto";
+import { Stream } from "openai/core/streaming";
 //#region packages/ai/src/transports/openai-transport-shared.ts
 const OPENAI_RESPONSE_MODEL_HEADER_NAMES = /* @__PURE__ */ new Set(["openai-model", "x-openai-model"]);
 const OPENAI_RESPONSE_MODEL_EVENT_TYPES = /* @__PURE__ */ new Set([
@@ -199,21 +185,11 @@ function readOpenAICompletionsReasoningBatch(delta, visibleReasoningDetailTypes)
 		hasVisibleText: mirroredThinking.length < deltas.length
 	};
 }
-function detachPartialSnapshot(event) {
-	const source = event;
-	if (!source.partial) return event;
-	const partial = source.partial;
-	return {
-		...event,
-		partial: {
-			...partial,
-			content: partial.content.map((block) => block.type === "text" ? Object.assign({}, block) : block)
-		}
-	};
-}
 function parseOpenAICompletionsUsage(rawUsage, model, options) {
-	const cacheRead = rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens ?? 0;
-	const cacheWrite = rawUsage.prompt_tokens_details?.cache_write_tokens ?? rawUsage.prompt_tokens_details?.cache_creation_input_tokens ?? rawUsage.cache_creation_input_tokens ?? 0;
+	const reportedCacheRead = rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens ?? rawUsage.cached_tokens;
+	const reportedCacheWrite = rawUsage.prompt_tokens_details?.cache_write_tokens ?? rawUsage.prompt_tokens_details?.cache_creation_input_tokens ?? rawUsage.cache_creation_input_tokens;
+	const cacheRead = reportedCacheRead ?? 0;
+	const cacheWrite = reportedCacheWrite ?? 0;
 	const input = Math.max(0, (rawUsage.prompt_tokens || 0) - cacheRead - cacheWrite);
 	const output = rawUsage.completion_tokens || 0;
 	const reasoningTokens = asFiniteNumber(rawUsage.completion_tokens_details?.reasoning_tokens);
@@ -229,6 +205,7 @@ function parseOpenAICompletionsUsage(rawUsage, model, options) {
 		output,
 		cacheRead,
 		cacheWrite,
+		cacheTelemetry: { state: reportedCacheRead !== void 0 || reportedCacheWrite !== void 0 ? "available" : "unavailable" },
 		...options?.includeReasoningTokens !== false && reasoningTokens !== void 0 ? { reasoningTokens } : {},
 		contextUsage: hasCoherentContext ? {
 			state: "available",
@@ -261,6 +238,18 @@ function createOpenAIProviderAcceptanceHook(options, response, model) {
 		model
 	});
 }
+function withOpenAIErrorResponseHook(fetch, onResponse, model) {
+	return async (input, init) => {
+		const response = await fetch(input, init);
+		if (!response.ok) await notifyProviderHttpResponse({
+			options: { onResponse },
+			response,
+			model,
+			signal: init?.signal ?? (input instanceof Request ? input.signal : void 0)
+		});
+		return response;
+	};
+}
 /** Measure one UTF-8 append without double-counting a surrogate pair split across chunks. */
 function measureUtf8AppendBytes(bufferEndsWithHighSurrogate, chunk) {
 	let bytes = Buffer.byteLength(chunk, "utf8");
@@ -400,8 +389,8 @@ function projectOpenAIToolDescriptors(tools, schemas) {
 			});
 			continue;
 		}
-		const prepared = schemas ? prepareRuntimeToolInputSchema(parameters ?? {}, `${name}.parameters`) : void 0;
-		const schemaProjection = prepared?.projection ?? projectRuntimeToolInputSchema(parameters ?? {}, `${name}.parameters`);
+		const prepared = schemas ? prepareRuntimeToolInputSchema(parameters ?? {}, `${name}.parameters`, name) : void 0;
+		const schemaProjection = prepared?.projection ?? projectRuntimeToolInputSchema(parameters ?? {}, `${name}.parameters`, name);
 		if (!isRecord(schemaProjection.schema) || schemaProjection.violations.length > 0) {
 			diagnostics.push({
 				toolIndex,
@@ -560,7 +549,7 @@ function resolveOpenAIRequestReasoning(model, reasoning, offDefault = "none") {
 const MAX_OPENAI_STRICT_TOOL_DOWNGRADE_DIAGNOSTIC_KEYS = 256;
 const loggedOpenAIStrictToolDowngradeDiagnosticKeys = /* @__PURE__ */ new Set();
 const OPENAI_COMPLETIONS_APIS = /* @__PURE__ */ new Set(["openai-completions", "openclaw-openai-completions-transport"]);
-function readToolPayloadField$1(record, field) {
+function readToolPayloadField(record, field) {
 	try {
 		return Object.hasOwn(record, field) ? record[field] : void 0;
 	} catch {
@@ -569,16 +558,16 @@ function readToolPayloadField$1(record, field) {
 }
 function readCodeModePayloadToolName(tool) {
 	if (!isRecord(tool)) return;
-	const name = readToolPayloadField$1(tool, "name");
+	const name = readToolPayloadField(tool, "name");
 	if (typeof name === "string") return name;
-	const fn = readToolPayloadField$1(tool, "function");
+	const fn = readToolPayloadField(tool, "function");
 	if (!isRecord(fn)) return;
-	const fnName = readToolPayloadField$1(fn, "name");
+	const fnName = readToolPayloadField(fn, "name");
 	return typeof fnName === "string" ? fnName : void 0;
 }
 function readCodeModePayloadToolIdentity(tool, visibleToolNames, allowedHostedToolTypes) {
 	if (!isRecord(tool)) return;
-	const type = readToolPayloadField$1(tool, "type");
+	const type = readToolPayloadField(tool, "type");
 	if (typeof type === "string" && allowedHostedToolTypes?.has(type)) {
 		try {
 			if (Object.hasOwn(tool, "name") || Object.hasOwn(tool, "function") || Object.hasOwn(tool, "functionDeclarations") || Object.hasOwn(tool, "function_declarations")) return false;
@@ -592,7 +581,7 @@ function readCodeModePayloadToolIdentity(tool, visibleToolNames, allowedHostedTo
 }
 function readCodeModePayloadToolIdentities(payload) {
 	if (!isRecord(payload)) return [];
-	const tools = readToolPayloadField$1(payload, "tools");
+	const tools = readToolPayloadField(payload, "tools");
 	if (!Array.isArray(tools)) return [];
 	return tools.flatMap((tool) => {
 		if (!isRecord(tool)) return [];
@@ -600,21 +589,21 @@ function readCodeModePayloadToolIdentities(payload) {
 		const name = readCodeModePayloadToolName(tool);
 		if (name) identities.push(`client:${name}`);
 		for (const key of ["functionDeclarations", "function_declarations"]) {
-			const declarations = readToolPayloadField$1(tool, key);
+			const declarations = readToolPayloadField(tool, key);
 			if (!Array.isArray(declarations)) continue;
 			for (const declaration of declarations) {
 				const declarationName = readCodeModePayloadToolName(declaration);
 				if (declarationName) identities.push(`client:${declarationName}`);
 			}
 		}
-		const type = readToolPayloadField$1(tool, "type");
+		const type = readToolPayloadField(tool, "type");
 		if (typeof type === "string" && type !== "function") identities.push(`hosted:${type}`);
 		return identities;
 	});
 }
 function filterCodeModePayloadTools(payload, visibleToolNames, allowedHostedToolTypes, observer) {
 	if (!isRecord(payload)) return;
-	const tools = readToolPayloadField$1(payload, "tools");
+	const tools = readToolPayloadField(payload, "tools");
 	if (!Array.isArray(tools)) return;
 	const beforeToolIdentities = observer ? readCodeModePayloadToolIdentities(payload) : void 0;
 	payload.tools = tools.flatMap((tool) => {
@@ -624,7 +613,7 @@ function filterCodeModePayloadTools(payload, visibleToolNames, allowedHostedTool
 		if (!isRecord(tool)) return [];
 		const filteredGroups = {};
 		for (const key of ["functionDeclarations", "function_declarations"]) {
-			const declarations = readToolPayloadField$1(tool, key);
+			const declarations = readToolPayloadField(tool, key);
 			if (!Array.isArray(declarations)) continue;
 			const filtered = declarations.filter((declaration) => {
 				const declarationName = readCodeModePayloadToolName(declaration);
@@ -644,7 +633,7 @@ function resolveCodeModeResponsesVisibleToolNames(context) {
 }
 function enforceCodeModeResponsesToolSurface(payload, visibleToolNames, allowedHostedToolTypes, observer) {
 	if (!isRecord(payload)) return;
-	const tools = readToolPayloadField$1(payload, "tools");
+	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)));
@@ -654,7 +643,7 @@ function enforceCodeModeResponsesToolSurface(payload, visibleToolNames, allowedH
 	});
 }
 function assertCodeModeResponsesToolSurface(payload, visibleToolNames, allowedHostedToolTypes) {
-	const tools = isRecord(payload) ? readToolPayloadField$1(payload, "tools") : void 0;
+	const tools = isRecord(payload) ? readToolPayloadField(payload, "tools") : void 0;
 	if (!Array.isArray(tools)) throw new Error("Code mode payload tool surface violation: expected exec,wait; got no tools");
 	const identities = tools.map((tool) => readCodeModePayloadToolIdentity(tool, visibleToolNames, allowedHostedToolTypes));
 	const names = identities.flatMap((identity) => typeof identity === "string" && identity.startsWith("client:") ? [identity.slice(7)] : []).toSorted((left, right) => left.localeCompare(right));
@@ -773,7 +762,7 @@ function isOpenAIResponsesCompactionState(state) {
 function readOpenAIResponsesCompactionReplayState(value) {
 	return isRecord(value) && isProviderReplayContext(value) && isOpenAIResponsesCompactionState(value) ? value : void 0;
 }
-function captureOpenAIResponsesCompaction(output, item, boundary, model, captureMetadata, compactedOutput) {
+function captureOpenAIResponsesCompaction(output, item, boundary, model, captureMetadata, compactedOutput, outputTokens) {
 	const metadata = captureMetadata ?? buildOpenAIResponsesReasoningReplayMetadata(model);
 	if (!item.encrypted_content) return;
 	if (!metadata?.baseUrlHash) {
@@ -799,7 +788,8 @@ function captureOpenAIResponsesCompaction(output, item, boundary, model, capture
 		...metadata.authProfileHash ? { authProfileHash: metadata.authProfileHash } : {},
 		...compactedOutput ? { compactedWindow: {
 			state: "ready",
-			output: JSON.stringify(compactedOutput)
+			output: JSON.stringify(compactedOutput),
+			...outputTokens !== void 0 ? { outputTokens } : {}
 		} } : {}
 	};
 	if (compactedOutput && !readOpenAIResponsesCompactionWindow(replay, model)) throw new Error("Responses compact endpoint checkpoint is invalid or exceeds 16 MiB");
@@ -863,6 +853,7 @@ function resolveNewestOpenAIResponsesCompactionReplay(messages, model, options)
 				owner: message,
 				mode: "complete-window",
 				output,
+				...replay.compactedWindow?.state === "ready" && replay.compactedWindow.outputTokens !== void 0 ? { outputTokens: replay.compactedWindow.outputTokens } : {},
 				item,
 				replayIndex: replay.type === "openai-responses-retained-compaction" ? message.content.length : replay.replayIndex ?? 0
 			};
@@ -1153,15 +1144,6 @@ function continuationHistoryMatches(previousInput, currentInput) {
 		}
 		if (previousItem.type === "function_call_output" && currentItem.type === "function_call_output" && typeof previousItem.call_id === "string" && typeof currentItem.call_id === "string" && currentItem.call_id !== previousItem.call_id && currentItem.call_id !== canonicalizeToolCallId(previousItem.call_id, void 0)) return false;
 	}
-	const previousCalls = responseToolCalls(previousInput);
-	const currentCalls = responseToolCalls(currentInput);
-	if (previousCalls.length !== currentCalls.length) return false;
-	for (const [index, previousCall] of previousCalls.entries()) {
-		const currentCall = currentCalls[index];
-		if (!currentCall) return false;
-		const previousShapes = toolCallReplayShapes(previousCall);
-		if (![...toolCallReplayShapes(currentCall)].some((shape) => previousShapes.has(shape))) return false;
-	}
 	const normalizedPrevious = normalizeContinuationHistory(previousInput);
 	const normalizedCurrent = normalizeContinuationHistory(currentInput);
 	return normalizedPrevious !== void 0 && normalizedCurrent !== void 0 && jsonValuesEqual(normalizedCurrent, normalizedPrevious);
@@ -1312,7 +1294,6 @@ function claimOpenAIResponsesHttpContinuation(params) {
 				if (httpContinuationEntries.get(key) !== claimed) return;
 				const state = recordResponsesContinuationState(previous?.state, effectiveRequest, response, previous !== void 0 && dispatchedPreviousResponseId === previous.state.lastResponseId);
 				const retainedBytes = Buffer.byteLength(JSON.stringify(state), "utf8");
-				if (httpContinuationEntries.get(key) !== claimed) return;
 				if (retainedBytes > MAX_HTTP_CONTINUATION_RETAINED_BYTES) {
 					deleteHttpContinuationIfOwned(key, claimed);
 					return;
@@ -1408,6 +1389,29 @@ function resolveReplayableResponsesMessageId(params) {
 	if (!params.textSignatureId) return params.fallbackOrdinal === 0 ? params.fallbackId : `${params.fallbackId}_${params.fallbackOrdinal}`;
 	return params.previousReplayItemWasReasoning ? params.textSignatureId : void 0;
 }
+function normalizeOpenAIResponsesReasoningReplayItem(item, api) {
+	if (api === "openai-chatgpt-responses") return Array.isArray(item.summary) ? item : {
+		...item,
+		summary: []
+	};
+	return {
+		id: item.id,
+		type: "reasoning",
+		summary: [],
+		...Array.isArray(item.content) && item.content.length === 0 ? { content: [] } : {},
+		status: item.status,
+		encrypted_content: item.encrypted_content
+	};
+}
+//#endregion
+//#region packages/ai/src/utils/text-signature.ts
+function encodeTextSignatureV1(id, phase) {
+	return JSON.stringify({
+		v: 1,
+		id,
+		...phase ? { phase } : {}
+	});
+}
 //#endregion
 //#region packages/ai/src/transports/openai-responses-replay-messages-internal.ts
 function resolveResponsesInstructionRole(model) {
@@ -1448,33 +1452,27 @@ function readOpenAIResponsesReasoningReplayBlockMetadata(block) {
 	const value = block[OPENAI_RESPONSES_REASONING_REPLAY_BLOCK_META_KEY];
 	return isOpenAIResponsesReasoningReplayMetadata(value) ? value : null;
 }
-function normalizeOpenAIResponsesReasoningReplayItem(item) {
-	return Array.isArray(item.summary) ? item : {
-		...item,
-		summary: []
-	};
-}
 function prepareOpenAIResponsesReasoningItemForReplay(item, context, blockMetadata, options) {
 	const hasRawMetadata = Object.hasOwn(item, OPENAI_RESPONSES_REASONING_REPLAY_META_KEY);
 	const { [OPENAI_RESPONSES_REASONING_REPLAY_META_KEY]: rawMetadata, ...rest } = item;
-	if (!("encrypted_content" in rest)) return normalizeOpenAIResponsesReasoningReplayItem(rest);
+	if (!("encrypted_content" in rest)) return normalizeOpenAIResponsesReasoningReplayItem(rest, context.api);
 	const metadata = blockMetadata !== void 0 ? blockMetadata ?? void 0 : isOpenAIResponsesReasoningReplayMetadata(rawMetadata) ? rawMetadata : void 0;
-	if (blockMetadata === void 0 && !hasRawMetadata && options?.preserveUnattributedEncryptedContent === true || metadata && providerReplayContextMatches(metadata, context)) return normalizeOpenAIResponsesReasoningReplayItem(rest);
-	return normalizeOpenAIResponsesReasoningReplayItem(stripEncryptedReasoningContentFields(rest));
+	if (blockMetadata === void 0 && !hasRawMetadata && options?.preserveUnattributedEncryptedContent === true || metadata && providerReplayContextMatches(metadata, context)) return normalizeOpenAIResponsesReasoningReplayItem(rest, context.api);
+	return normalizeOpenAIResponsesReasoningReplayItem(stripEncryptedReasoningContentFields(rest), context.api);
 }
 function normalizeResponsesReplayItemId(id) {
 	if (!id) return;
 	if (id.length <= 64) return id;
 	return `msg_${shortHash(id)}`;
 }
+function responsesAssistantTurnKey(message) {
+	const id = message.turnId || message.responseId;
+	return id ? `${message.provider}:${message.api}:${message.model}:${id}` : void 0;
+}
 function orderResponsesAsyncToolResults(source) {
-	const turnKey = (message) => {
-		const id = message.turnId || message.responseId;
-		return id ? `${message.provider}:${message.api}:${message.model}:${id}` : void 0;
-	};
 	const lastAssistant = /* @__PURE__ */ new Map();
 	for (const [index, message] of source.entries()) if (message.role === "assistant") {
-		const key = turnKey(message);
+		const key = responsesAssistantTurnKey(message);
 		if (key) lastAssistant.set(key, index);
 	}
 	const owners = /* @__PURE__ */ new Map();
@@ -1482,7 +1480,7 @@ function orderResponsesAsyncToolResults(source) {
 	const ordered = [];
 	for (const [index, message] of source.entries()) {
 		if (message.role === "assistant") {
-			const key = turnKey(message);
+			const key = responsesAssistantTurnKey(message);
 			if (key) {
 				for (const block of message.content) if (block.type === "toolCall" && block.async) owners.set(block.id, key);
 			}
@@ -1578,14 +1576,17 @@ function convertResponsesMessagesWithStyle(model, context, allowedToolCallProvid
 	});
 	const transformMessages = (source) => providerStyle ? transformProviderMessages(source, model, normalizeToolCallId) : getAiTransportHost().transformTransportMessages(source, model, normalizeToolCallId, {
 		normalizeSameModelToolCallIds: shouldNormalizeSameModelToolCallIds,
-		preserveUnframedToolResults: replayPlan.preserveUnframedToolResults
+		preserveUnframedToolResults: replayPlan.preserveUnframedToolResults || preparedRequestSuffix.has(context)
 	});
 	const transformedMessages = orderResponsesAsyncToolResults(transformMessages(replayPlan.messages));
 	if ((options?.includeSystemPrompt ?? true) && context.systemPrompt) messages.push(buildResponsesInputMessage(resolveResponsesInstructionRole(model), [{
 		type: "input_text",
 		text: sanitizeTransportPayloadText(stripSystemPromptCacheBoundary(context.systemPrompt))
 	}]));
-	if (replayPlan.compactedWindow) messages.push(...replayPlan.compactedWindow);
+	if (replayPlan.compactedWindow) messages.push(...replayPlan.compactedWindow.map((item) => {
+		if (options?.retainUserProvenance && item.type === "message" && item.role === "user") return Object.assign({}, item, { [RESPONSES_RETAINED_USER]: true });
+		return item;
+	}));
 	let replayMessages = replayPlan.compaction ? [replayPlan.compaction, ...transformedMessages] : transformedMessages;
 	const isCarrier = (message) => "role" in message && hasRuntimeContextMarker(message);
 	if (replayMessages.some(isCarrier)) {
@@ -1626,12 +1627,18 @@ function convertResponsesMessagesWithStyle(model, context, allowedToolCallProvid
 				detail: "auto",
 				image_url: `data:${item.mimeType};base64,${item.data}`
 			}).filter((item) => providerStyle || model.input.includes("image") || item.type !== "input_image");
-			if (content.length > 0) messages.push(buildResponsesInputMessage(role, content, msg));
-			else if (providerStyle) continue;
+			if (content.length > 0) {
+				const input = buildResponsesInputMessage(role, content, msg);
+				if (options?.retainUserProvenance && role === "user" && !msg.synthetic && !msg.operatorMessage && !hasRuntimeContextMarker(msg)) Object.assign(input, { [RESPONSES_RETAINED_USER]: true });
+				messages.push(input);
+			} else if (providerStyle) continue;
 		} else if (msg.role === "assistant") {
 			const output = [];
+			const sourceTurnKey = responsesAssistantTurnKey(msg);
+			const sourceFragmentId = msg.content.find((block) => block.type === "toolCall")?.id;
 			let textFallbackOrdinal = 0;
 			let previousReplayItemWasReasoning = false;
+			let seenReplayableReasoning = false;
 			const isDifferentModel = msg.model !== model.id && msg.provider === model.provider && msg.api === model.api;
 			for (const block of msg.content) if (block.type === "thinking") {
 				if (shouldReplayReasoningItems && block.thinkingSignature && (providerStyle || block.thinkingSignature.startsWith("{"))) {
@@ -1647,13 +1654,14 @@ function convertResponsesMessagesWithStyle(model, context, allowedToolCallProvid
 					if (shouldReplayResponsesItemIds && !providerStyle && model.provider === "github-copilot" && !isSafeResponsesReplayItemId(replayableReasoningItem.id)) continue;
 					output.push(replayableReasoningItem);
 					previousReplayItemWasReasoning = true;
+					seenReplayableReasoning ||= typeof replayableReasoningItem.id === "string" && replayableReasoningItem.id.startsWith("rs_");
 				}
 			} else if (block.type === "text") {
 				const textSignature = parseOpenAIResponsesTextSignature(block.textSignature);
 				let msgId = resolveReplayableResponsesMessageId({
 					replayResponsesItemIds: shouldReplayResponsesItemIds,
 					textSignatureId: textSignature?.id,
-					fallbackId: `msg_${msgIndex}`,
+					fallbackId: sourceTurnKey ? `msg_${shortHash(JSON.stringify([sourceTurnKey, sourceFragmentId]))}` : `msg_${msgIndex}`,
 					fallbackOrdinal: textFallbackOrdinal,
 					previousReplayItemWasReasoning
 				});
@@ -1677,7 +1685,7 @@ function convertResponsesMessagesWithStyle(model, context, allowedToolCallProvid
 				const separatorIndex = block.id.indexOf("|");
 				const callId = separatorIndex === -1 ? block.id : block.id.slice(0, separatorIndex);
 				const itemIdRaw = separatorIndex === -1 ? void 0 : block.id.slice(separatorIndex + 1);
-				const itemId = shouldReplayResponsesItemIds && !(isDifferentModel && itemIdRaw?.startsWith("fc_")) ? itemIdRaw : void 0;
+				const itemId = shouldReplayResponsesItemIds && (!itemIdRaw?.startsWith("fc_") || seenReplayableReasoning && !isDifferentModel) ? itemIdRaw : void 0;
 				output.push({
 					type: "function_call",
 					...itemId ? { id: itemId } : {},
@@ -1701,21 +1709,44 @@ function convertResponsesMessagesWithStyle(model, context, allowedToolCallProvid
 			const mediaPlaceholder = describeToolResultMediaPlaceholder(msg.content);
 			const hasImages = msg.content.some(isImageWithMediaPayload);
 			const separatorIndex = msg.toolCallId.indexOf("|");
-			const callId = separatorIndex === -1 ? msg.toolCallId : msg.toolCallId.slice(0, separatorIndex);
-			messages.push({
-				type: "function_call_output",
-				call_id: callId,
-				output: hasImages && model.input.includes("image") ? [...hasText ? [{
+			const rawCallId = separatorIndex === -1 ? msg.toolCallId : msg.toolCallId.slice(0, separatorIndex);
+			const callId = preparedRequestSuffix.has(context) ? preparedRequestSuffix.toolCallId(context, rawCallId, normalizeToolCallId(rawCallId, model, model)) : rawCallId;
+			let output = hasText ? textResult : mediaPlaceholder ?? "(no output)";
+			if (hasImages && model.input.includes("image")) {
+				const parts = [];
+				let hasExplicitText = false;
+				for (const item of msg.content) if (isImageWithMediaPayload(item)) parts.push({
+					type: "input_image",
+					detail: "auto",
+					image_url: `data:${item.mimeType};base64,${item.data}`
+				});
+				else if (hasText && item.type === "text") {
+					const text = extractToolResultBlockText(item);
+					if (text) {
+						const joinedText = `${hasExplicitText ? "\n" : ""}${text}`;
+						hasExplicitText = true;
+						const previous = parts.at(-1);
+						if (previous?.type === "input_text") previous.text += joinedText;
+						else parts.push({
+							type: "input_text",
+							text: joinedText
+						});
+					}
+				}
+				if (hasText && !hasExplicitText) parts.unshift({
 					type: "input_text",
 					text: textResult
-				}] : mediaPlaceholder === "(see attached media)" ? [{
+				});
+				else if (!hasText && mediaPlaceholder === "(see attached media)") parts.unshift({
 					type: "input_text",
 					text: mediaPlaceholder
-				}] : [], ...msg.content.filter(isImageWithMediaPayload).map((item) => ({
-					type: "input_image",
-					detail: "auto",
-					image_url: `data:${item.mimeType};base64,${item.data}`
-				}))] : hasText ? textResult : mediaPlaceholder ?? "(no output)"
+				});
+				output = parts;
+			}
+			messages.push({
+				type: "function_call_output",
+				call_id: callId,
+				output
 			});
 		}
 		msgIndex += 1;
@@ -1856,24 +1887,17 @@ function summarizeResponsesCompactionItems(input) {
 		`compactionInputIndexes=${compactions.map((item) => item.inputIndex).join(",") || "none"}`
 	];
 }
-function readToolPayloadField(record, field) {
-	try {
-		return record[field];
-	} catch {
-		return;
-	}
-}
 function readResponsesToolDisplayName(tool) {
 	if (!tool || typeof tool !== "object") return "";
 	const record = tool;
-	const name = readToolPayloadField(record, "name");
+	const name = record.name;
 	if (typeof name === "string") return name;
-	const fn = readToolPayloadField(record, "function");
+	const fn = record.function;
 	if (fn && typeof fn === "object") {
-		const fnName = readToolPayloadField(fn, "name");
+		const fnName = fn.name;
 		if (typeof fnName === "string") return fnName;
 	}
-	const type = readToolPayloadField(record, "type");
+	const type = record.type;
 	return typeof type === "string" && type !== "function" ? type : "";
 }
 function summarizeResponsesTools(tools) {
@@ -2138,6 +2162,59 @@ async function resolveNextResponsesEncryptedContentAttempt(attempt, error, optio
 	};
 }
 async function createResponsesStreamWithRecovery(params) {
+	async function* streamWithTerminalDrain(response, body) {
+		const reader = body.getReader();
+		let terminal = false;
+		try {
+			const forwarded = new ReadableStream({
+				async pull(controller) {
+					const { done, value } = await reader.read();
+					if (done) controller.close();
+					else controller.enqueue(value);
+				},
+				cancel() {}
+			}, { highWaterMark: 0 });
+			const stream = Stream.fromSSEResponse(new Response(forwarded, response), new AbortController(), params.client);
+			for await (const event of stream) {
+				terminal ||= event?.type === "response.completed" || event?.type === "response.incomplete";
+				yield event;
+			}
+		} finally {
+			if (!terminal) {
+				await reader.cancel().catch(() => void 0);
+				reader.releaseLock();
+			} else {
+				const controller = new AbortController();
+				const unregister = registerSessionResourceCleanup((sessionId) => {
+					if (!sessionId) controller.abort();
+				});
+				const signal = AbortSignal.any([
+					controller.signal,
+					AbortSignal.timeout(250),
+					...params.requestOptions?.signal ? [params.requestOptions.signal] : []
+				]);
+				const cancel = () => {
+					reader.cancel().catch(() => void 0);
+				};
+				signal.addEventListener("abort", cancel, { once: true });
+				(async () => {
+					try {
+						let bytes = 0;
+						while (!signal.aborted && bytes < 65536) {
+							const { done, value } = await reader.read();
+							if (done) break;
+							bytes += value.byteLength;
+						}
+					} catch {} finally {
+						cancel();
+						reader.releaseLock();
+						signal.removeEventListener("abort", cancel);
+						unregister();
+					}
+				})();
+			}
+		}
+	}
 	const encodeBody = params.encodeBody ?? prepareModelRequestBody(void 0);
 	let tierCeiling;
 	const limitTier = (request) => tierCeiling && (request.service_tier === "ultrafast" || tierCeiling === "default" && request.service_tier === "priority") ? {
@@ -2190,7 +2267,7 @@ async function createResponsesStreamWithRecovery(params) {
 				});
 				if (!isAsyncIterable(data)) throw new Error("OpenAI Responses streaming request returned a non-stream response");
 				return {
-					stream: data,
+					stream: data instanceof Stream && response.body ? streamWithTerminalDrain(response, response.body) : data,
 					response,
 					attempt
 				};
@@ -2329,7 +2406,7 @@ function createResponsesOutputTracker(params) {
 		trackStream(sink) {
 			return { push(event) {
 				visibleTextEmitted ||= event.type === "text_delta" && event.delta.length > 0 || event.type === "text_end" && event.content.length > 0;
-				sink.push(detachPartialSnapshot(event));
+				sink.push(event);
 			} };
 		},
 		observeEvent(event) {
@@ -2508,23 +2585,43 @@ async function* adaptResponsesStream(stream, signal) {
 		yield event;
 	}
 }
-async function* observeResponsesStream(stream, model, requestStartedAt) {
+async function* observeResponsesStream(stream, model, requestStartedAt, request, transport = "sse") {
 	const startedAt = Date.now();
 	const eventTypes = /* @__PURE__ */ new Map();
-	const debugMode = resolveModelSseDebugMode();
 	let eventCount = 0;
+	let responseId;
+	let servedTier;
+	let outputTokens;
+	let reasoningTokens;
+	let terminalType;
+	let incompleteReason;
+	const requestDetails = `transport=${transport} requestedTier=${request?.service_tier ?? "unspecified"} max_output_tokens=${request?.max_output_tokens ?? "provider-default"}`;
+	emitModelTransportDebug(log, `[responses] stream_start provider=${model.provider} api=${model.api} model=${model.id} ${requestDetails}`);
 	try {
 		for await (const event of stream) {
-			const type = isRecord(event) && typeof event.type === "string" ? event.type : "unknown";
+			const record = isRecord(event) ? event : void 0;
+			const type = typeof record?.type === "string" ? record.type : "unknown";
+			const response = isRecord(record?.response) ? record.response : void 0;
+			if (typeof response?.id === "string") responseId = response.id;
+			if (typeof response?.service_tier === "string") servedTier = response.service_tier;
+			if (isRecord(response?.usage)) {
+				if (typeof response.usage.output_tokens === "number") outputTokens = response.usage.output_tokens;
+				const details = response.usage.output_tokens_details;
+				if (isRecord(details) && typeof details.reasoning_tokens === "number") reasoningTokens = details.reasoning_tokens;
+			}
+			if (type === "response.completed" || type === "response.incomplete" || type === "response.failed") {
+				terminalType = type;
+				const details = response?.incomplete_details;
+				if (isRecord(details) && typeof details.reason === "string") incompleteReason = details.reason;
+			}
 			eventCount += 1;
 			eventTypes.set(type, (eventTypes.get(type) ?? 0) + 1);
-			if (eventCount === 1) emitModelTransportDebug(log, `[responses] first_event provider=${model.provider} api=${model.api} model=${model.id} elapsedMs=${Date.now() - (requestStartedAt ?? startedAt)} headersToEventMs=${Date.now() - startedAt} type=${type}`);
-			if (debugMode === "peek" && eventCount <= 5) emitModelTransportDebug(log, `[responses] event_peek provider=${model.provider} api=${model.api} model=${model.id} index=${eventCount} type=${type} event=${stringifyRedactedEvent(event)}`);
+			if (eventCount === 1) emitModelTransportDebug(log, `[responses] first_event provider=${model.provider} api=${model.api} model=${model.id} elapsedMs=${Date.now() - (requestStartedAt ?? startedAt)} headersToEventMs=${Date.now() - startedAt} type=${type} responseId=${responseId ?? "unknown"}`);
 			yield event;
 		}
 	} finally {
 		const types = [...eventTypes].map(([type, count]) => `${type}:${count}`).join(",");
-		emitModelTransportDebug(log, `[responses] stream_done provider=${model.provider} api=${model.api} model=${model.id} elapsedMs=${Date.now() - startedAt} events=${eventCount} types=${types}`);
+		emitModelTransportDebug(log, `[responses] stream_done provider=${model.provider} api=${model.api} model=${model.id} ${requestDetails} responseId=${responseId ?? "unknown"} servedTier=${servedTier ?? "unknown"} outputTokens=${outputTokens ?? "unknown"} reasoningTokens=${reasoningTokens ?? "unknown"} streamDurationMs=${Date.now() - startedAt} elapsedMs=${Date.now() - (requestStartedAt ?? startedAt)} terminal=${terminalType ?? "none"} incompleteReason=${incompleteReason ?? "none"} events=${eventCount} types=${types}`);
 	}
 }
 //#endregion
@@ -2556,22 +2653,28 @@ function resolveCompletedResponsesToolCall(item, streamed) {
 function createResponsesTerminalController(params) {
 	const { output, stream, model, options } = params;
 	const blocks = output.content;
+	const pendingReasoning = /* @__PURE__ */ new Set();
+	const deferredToolCompletions = [];
+	const recordPendingReasoning = (contentIndex) => {
+		pendingReasoning.add(contentIndex);
+	};
+	const releaseDeferredToolCompletions = () => {
+		for (const event of deferredToolCompletions) stream.push(event);
+		deferredToolCompletions.length = 0;
+		pendingReasoning.clear();
+	};
 	let rejectedToolCallId;
 	const recordIncompleteToolCall = (event, item) => {
 		rejectedToolCallId = params.toolCalls.resolve(event, readResponsesToolCallItemIdentity(item))?.block.id ?? (item.call_id ? resolveResponsesToolCallId(item) : void 0);
 	};
-	const backfillReasoning = (items) => {
+	const settleReasoning = (items) => {
 		for (const [outputIndex, item] of items.entries()) {
-			if (item.type !== "reasoning" || !item.encrypted_content) continue;
+			if (item.type !== "reasoning") continue;
 			const tracked = params.outputs.get(item, outputIndex);
 			const block = tracked && blocks[tracked.contentIndex];
-			if (block?.type !== "thinking" || !block.thinkingSignature) continue;
-			const stored = JSON.parse(block.thinkingSignature);
-			if (!stored.encrypted_content) block.thinkingSignature = JSON.stringify({
-				...stored,
-				encrypted_content: item.encrypted_content
-			});
-			if (options?.reasoningReplayMetadata) Object.assign(block, { [OPENAI_RESPONSES_REASONING_REPLAY_BLOCK_META_KEY]: options.reasoningReplayMetadata });
+			if (block?.type !== "thinking") continue;
+			block.thinkingSignature = JSON.stringify(normalizeOpenAIResponsesReasoningReplayItem(item, model.api));
+			if (item.encrypted_content && options?.reasoningReplayMetadata) Object.assign(block, { [OPENAI_RESPONSES_REASONING_REPLAY_BLOCK_META_KEY]: options.reasoningReplayMetadata });
 		}
 	};
 	const appendText = (item, initialIndex) => {
@@ -2663,12 +2766,19 @@ function createResponsesTerminalController(params) {
 			});
 		}
 		params.outputs.set(item, contentIndex, outputIndex, true);
-		stream.push({
+		const completion = {
 			type: "toolcall_end",
 			contentIndex,
 			toolCall,
 			partial: output
-		});
+		};
+		if (toolCall.async) {
+			for (const index of pendingReasoning) if (index < contentIndex) {
+				deferredToolCompletions.push(completion);
+				return;
+			}
+		}
+		stream.push(completion);
 	};
 	const recoverTerminalOutput = (items, completeToolCall) => {
 		let hasCompletedLaterOutput = false;
@@ -2712,7 +2822,7 @@ function createResponsesTerminalController(params) {
 		}
 	};
 	const finalizeTerminalFacts = (response, responseId = response.id) => {
-		options?.onServiceTier?.(response.service_tier);
+		options?.onServiceTier?.(response.service_tier, response);
 		output.responseId = responseId || output.responseId;
 		output.responseModel = options?.resolveResponseModel ? options.resolveResponseModel()?.trim() || void 0 : response.model?.trim() || void 0;
 		const usage = mapResponsesTerminalUsage(response.usage);
@@ -2720,6 +2830,7 @@ function createResponsesTerminalController(params) {
 		if (usage) output.usage = {
 			...usage,
 			...reasoningTokens === void 0 ? {} : { reasoningTokens },
+			usageReport: { state: "available" },
 			cost: {
 				input: 0,
 				output: 0,
@@ -2728,14 +2839,15 @@ function createResponsesTerminalController(params) {
 				total: 0
 			}
 		};
+		else output.usage.usageReport = { state: "unavailable" };
 		calculateCost(model, output.usage);
 		if (options?.applyServiceTierPricing) {
 			const tier = options.resolveServiceTier ? options.resolveServiceTier(response.service_tier, options.serviceTier) : response.service_tier ?? options.serviceTier;
 			options.applyServiceTierPricing(output.usage, tier);
 		}
 	};
-	const finalizeResponse = (response, terminalEventType, hasRejectedToolCall) => {
-		backfillReasoning(response.output ?? []);
+	const finalizeResponse = (response, terminalEventType, rejectedToolCall) => {
+		settleReasoning(response.output ?? []);
 		finalizeTerminalFacts(response);
 		const terminal = resolveResponsesTerminalStopReason({
 			status: response.status,
@@ -2744,11 +2856,11 @@ function createResponsesTerminalController(params) {
 			hasToolCall: blocks.some((block) => block.type === "toolCall")
 		});
 		output.stopReason = terminal.stopReason;
-		output.errorMessage = terminal.errorMessage;
+		if (terminal.errorMessage !== void 0) output.errorMessage = terminal.errorMessage;
 		if (terminalEventType === "response.completed" && typeof response.end_turn === "boolean") output.endTurn = response.end_turn;
 		const incompleteReason = response.incomplete_details?.reason;
 		const incompleteCall = response.output?.find((item) => item.type === "function_call" && item.status && item.status !== "completed");
-		const incompleteToolCallId = rejectedToolCallId ?? (terminalEventType === "response.incomplete" && !hasRejectedToolCall ? params.toolCalls.values()[0]?.block.id : void 0) ?? (incompleteCall?.type === "function_call" && incompleteCall.call_id ? resolveResponsesToolCallId(incompleteCall) : void 0);
+		const incompleteToolCallId = rejectedToolCallId ?? (terminalEventType === "response.incomplete" && !rejectedToolCall ? params.toolCalls.values()[0]?.block.id : void 0) ?? (incompleteCall?.type === "function_call" && incompleteCall.call_id ? resolveResponsesToolCallId(incompleteCall) : void 0);
 		appendAssistantMessageDiagnostic(output, {
 			type: "openai_responses_terminal",
 			timestamp: Date.now(),
@@ -2760,14 +2872,17 @@ function createResponsesTerminalController(params) {
 				endTurn: typeof response.end_turn === "boolean" ? response.end_turn : response.end_turn === void 0 ? "absent" : "invalid"
 			}
 		});
-		if (!hasRejectedToolCall && terminalEventType === "response.incomplete" && params.toolCalls.hasActive()) throw output.errorMessage ? new Error(output.errorMessage) : new IncompleteToolCallError("Responses stream completed with unresolved tool calls");
+		const failure = rejectedToolCall ?? (terminalEventType === "response.incomplete" && params.toolCalls.hasActive() ? { error: new IncompleteToolCallError("Responses stream completed with unresolved tool calls") } : void 0);
+		if (failure) throw output.errorMessage ? new Error(output.errorMessage) : failure.error;
 	};
 	return {
 		finalizeResponse,
 		finalizeFailedResponse: finalizeTerminalFacts,
 		recordIncompleteToolCall,
 		recoverTerminalOutput,
-		emitToolCallCompletion
+		emitToolCallCompletion,
+		recordPendingReasoning,
+		releaseDeferredToolCompletions
 	};
 }
 //#endregion
@@ -2792,17 +2907,19 @@ async function processResponsesStream(openaiStream, output, sink, model, options
 	});
 	const createOutputSlot = (event, item) => {
 		let slot;
-		if (item.type === "reasoning") slot = {
-			type: "thinking",
-			item,
-			block: {
+		if (item.type === "reasoning") {
+			slot = {
 				type: "thinking",
-				thinking: ""
-			},
-			contentIndex: blocks.length,
-			outputIndex: readResponsesOutputIndex(event)
-		};
-		else if (item.type === "message") {
+				item,
+				block: {
+					type: "thinking",
+					thinking: ""
+				},
+				contentIndex: blocks.length,
+				outputIndex: readResponsesOutputIndex(event)
+			};
+			terminal.recordPendingReasoning(slot.contentIndex);
+		} else if (item.type === "message") {
 			const collapseCandidate = lastTextBlock;
 			const block = collapseCandidate ? null : createTextBlock(item, "");
 			slot = {
@@ -2916,7 +3033,7 @@ async function processResponsesStream(openaiStream, output, sink, model, options
 			if (tracked?.completed) continue;
 			const state = streamingToolCalls.resolve({ output_index: outputIndex }, readResponsesToolCallItemIdentity(item), allowUnmatchedIdentity);
 			if (tracked && !state) throw new Error("Responses stream completed with unresolved tool calls");
-			const validated = resolveCompletedResponsesToolCall(item, { name: state?.block.name });
+			const validated = resolveCompletedResponsesToolCall(item, { name: state?.providerName });
 			if (state) recovered.push(state);
 			prepared.set(outputIndex, () => finalizeToolCall(item, outputIndex, state, validated));
 		}
@@ -2966,6 +3083,7 @@ async function processResponsesStream(openaiStream, output, sink, model, options
 					const contentIndex = output.content.length;
 					const toolCallState = {
 						block: toolCallBlock,
+						providerName: toolCallBlock.name,
 						contentIndex,
 						argumentStreamReliable: true,
 						argumentsStreamed: false,
@@ -3076,7 +3194,7 @@ async function processResponsesStream(openaiStream, output, sink, model, options
 					const summaryText = item.summary?.map((s) => s.text).join("\n\n") || "";
 					const contentText = item.content?.map((c) => c.text).join("\n\n") || "";
 					outputSlot.block.thinking = summaryText || contentText || outputSlot.block.thinking;
-					outputSlot.block.thinkingSignature = JSON.stringify(item);
+					outputSlot.block.thinkingSignature = JSON.stringify(normalizeOpenAIResponsesReasoningReplayItem(item, model.api));
 					outputs.set(item, outputSlot.contentIndex, readResponsesOutputIndex(event) ?? outputSlot.outputIndex, true);
 					if (item.encrypted_content && options?.reasoningReplayMetadata) outputSlot.block[OPENAI_RESPONSES_REASONING_REPLAY_BLOCK_META_KEY] = options.reasoningReplayMetadata;
 					stream.push({
@@ -3150,7 +3268,7 @@ async function processResponsesStream(openaiStream, output, sink, model, options
 					let validated;
 					try {
 						validated = resolveCompletedResponsesToolCall(item, {
-							name: streamingToolCall?.block.name,
+							name: streamingToolCall?.providerName,
 							arguments: parsedStreamedArguments ?? (completedArguments || streamedArguments)
 						});
 					} catch (error) {
@@ -3160,8 +3278,7 @@ async function processResponsesStream(openaiStream, output, sink, model, options
 					finalizeToolCall(item, readResponsesOutputIndex(event), streamingToolCall, validated);
 				}
 			} else if (event.type === "response.completed" || event.type === "response.incomplete") {
-				terminal.finalizeResponse(event.response, event.type, Boolean(rejectedToolCall));
-				if (rejectedToolCall) throw output.errorMessage ? new Error(output.errorMessage) : rejectedToolCall.error;
+				terminal.finalizeResponse(event.response, event.type, rejectedToolCall);
 				if (event.type === "response.completed" || output.stopReason === "length") {
 					const items = event.response.output ?? [];
 					const completeToolCall = event.type === "response.completed" ? prepareTerminalToolCalls(items) : void 0;
@@ -3190,7 +3307,8 @@ async function processResponsesStream(openaiStream, output, sink, model, options
 		if (terminalResponse === void 0) throw new Error("OpenAI Responses stream ended before a terminal response event");
 		return terminalResponse ?? void 0;
 	} finally {
-		for (const block of output.content) delete block.partialJson;
+		for (const block of output.content) if (block.type === "toolCall") delete block.partialJson;
+		terminal.releaseDeferredToolCompletions();
 	}
 }
 //#endregion
@@ -3226,14 +3344,7 @@ function convertPreparedResponsesTools(prepared, strictSetting, model) {
 			model
 		}) : resolveOpenAIProjectedToolsStrictToolFlag(projection, strictSetting);
 		return sortPromptCacheToolsByName(projection.tools).map((tool) => {
-			const result = {
-				type: "function",
-				name: tool.name,
-				description: tool.description,
-				parameters: normalizeOpenAIStrictToolParameters(tool.parameters, strict === true, model?.compat)
-			};
-			if (strict !== void 0) result.strict = strict;
-			return result;
+			return prepareOpenAIFunctionTool(tool.name, tool.description, normalizeOpenAIStrictToolParameters(tool.parameters, strict === true, model?.compat), strict).responses;
 		});
 	});
 }
@@ -3282,7 +3393,7 @@ function applyCommonResponsesParams(params, model, context, options, config) {
 function cleanStreamingScratchBuffers(output) {
 	for (const block of output.content) {
 		delete block.index;
-		delete block.partialJson;
+		if (block.type === "toolCall") delete block.partialJson;
 	}
 }
 async function runResponsesStreamLifecycle(params) {
@@ -3353,7 +3464,7 @@ async function runResponsesStreamLifecycle(params) {
 				const tier = requestedTier === "priority" || requestedTier === "default" ? requestedTier : originalTier;
 				return processStreamOptions?.resolveServiceTier ? processStreamOptions.resolveServiceTier(responseTier, tier) : responseTier ?? tier;
 			},
-			onServiceTier: (responseTier) => responsesServiceTierObserver.observe(options, requestedTier, responseTier),
+			onServiceTier: (responseTier, response) => responsesServiceTierObserver.observe(options, requestedTier, responseTier, response),
 			canRetryIdentityConflict: () => hasOnlyResponsesFunctionTools(admittedRequest),
 			reasoningReplayMetadata: buildOpenAIResponsesReasoningReplayMetadata(model, {
 				sessionId: options?.sessionId,
@@ -3393,4 +3504,4 @@ async function runResponsesStreamLifecycle(params) {
 	}
 }
 //#endregion
-export { replayResponsesReasoningUpdates as $, summarizeResponsesPayload as A, log as At, responsesInputFingerprint as B, readResponsesRecoveryEvent as C, toOpenAIResponsesToolChoice as Ct, logResponsesFailedNoDetails as D, createResponseModelTracker as Dt, ResponsesStreamFailure as E, createOpenAIResponseHook as Et, convertProviderResponsesMessages as F, resolveOpenAIClientBaseUrl as Ft, responsesContinuationRequestFingerprint as G, recordResponsesContinuationState as H, convertResponsesMessages as I, resolvePromptCacheKey as It, OPENAI_RESPONSES_ID_MAX_LENGTH as J, OPENAI_RESPONSES_CALL_ID_RE as K, resolveReplayableResponsesMessageId as L, redactIdentifier as Lt, resolveResponsesContextUsageBoundary as M, parseOpenAICompletionsUsage as Mt, bindResponsesInputMessage as N, readOpenAICompletionsContentDeltas as Nt, safeDebugValue as O, detachPartialSnapshot as Ot, buildResponsesInputMessage as P, readOpenAICompletionsReasoningBatch as Pt, isConfigurationUpdate as Q, readResponsesInputReplayState as R, sha256Hex as Rt, isInvalidEncryptedContentError as S, reconcileOpenAIResponsesToolChoice as St, resolveNextResponsesEncryptedContentAttempt as T, createOpenAIProviderAcceptanceHook as Tt, resolveResponsesContinuationRequest as U, claimOpenAIResponsesHttpContinuation as V, responsesContinuationPrefixFingerprint as W, shouldNormalizeOpenAIResponsesToolCallId as X, normalizeOpenAIResponsesFunctionCallId as Y, splitOpenAIFunctionCallPairing as Z, isResponsesTextContentPartType as _, resolveOpenAISimpleReasoningEffort as _t, prepareResponsesTools as a, suppressOpenAIResponsesCompaction as at, commitResponsesEncryptedContentAttempt as b, projectOpenAITools as bt, observeResponsesStream as c, buildOpenAISdkClientOptions as ct, AZURE_RESPONSES_TEXT_CONTENT_PART_TYPE as d, filterCodeModePayloadTools as dt, supportsResponsesReasoningUpdate as et, AZURE_RESPONSES_TEXT_DELTA_EVENT_TYPE as f, getCompat as ft, isAzureResponsesTextDeltaEventType as g, resolveOpenAIRequestReasoning as gt, isAzureResponsesTextDeltaEvent as h, resolveOpenAIStrictToolFlagWithDiagnostics as ht, convertResponsesToolPayload as i, resolveNewestOpenAIResponsesCompactionReplay as it, recordResponsesContextUsage as j, measureUtf8AppendBytes as jt, summarizeOpenAITransportError as k, isOpenAICompletionsThinkingEnabled as kt, createResponsesToolCallTracker as l, buildOpenAISdkRequestOptions as lt, OPENAI_RESPONSES_OUTPUT_TEXT_DELTA_EVENT_TYPE as m, resolveCodeModeResponsesVisibleToolNames as mt, applyResponsesServiceTierPricing as n, buildOpenAIResponsesReasoningReplayMetadata as nt, readOpenAIMisalignmentReview as o, assertCodeModeResponsesToolSurface as ot, OPENAI_RESPONSES_OUTPUT_TEXT_CONTENT_PART_TYPE as p, readCodeModePayloadToolName as pt, OPENAI_RESPONSES_FUNCTION_CALL_ITEM_ID_RE as q, runResponsesStreamLifecycle as r, captureOpenAIResponsesCompaction as rt, processResponsesStream as s, buildOpenAIClientHeaders as st, applyCommonResponsesParams as t, CompactionReplayRefreshRequiredError as tt, readResponsesToolCallItemIdentity as u, enforceCodeModeResponsesToolSurface as ut, isResponsesTextDeltaEventType as v, resolveOpenAIReasoningEffortMap as vt, resolveAzureOpenAIApiVersion as w, GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP as wt, createResponsesStreamWithRecovery as x, reconcileOpenAICompletionsToolChoice as xt, resolveResponsesMessageSnapshotCollapse as y, prepareOpenAITools as yt, recordResponsesInputReplay as z };
+export { isConfigurationUpdate as $, summarizeResponsesPayload as A, log as At, recordResponsesInputReplay as B, readResponsesRecoveryEvent as C, reconcileOpenAIResponsesToolChoice as Ct, logResponsesFailedNoDetails as D, createOpenAIResponseHook as Dt, ResponsesStreamFailure as E, createOpenAIProviderAcceptanceHook as Et, convertProviderResponsesMessages as F, resolveOpenAIClientBaseUrl as Ft, responsesContinuationPrefixFingerprint as G, claimOpenAIResponsesHttpContinuation as H, convertResponsesMessages as I, resolvePromptCacheKey as It, OPENAI_RESPONSES_FUNCTION_CALL_ITEM_ID_RE as J, responsesContinuationRequestFingerprint as K, normalizeOpenAIResponsesReasoningReplayItem as L, withOpenAIErrorResponseHook as Lt, resolveResponsesContextUsageBoundary as M, parseOpenAICompletionsUsage as Mt, bindResponsesInputMessage as N, readOpenAICompletionsContentDeltas as Nt, safeDebugValue as O, createResponseModelTracker as Ot, buildResponsesInputMessage as P, readOpenAICompletionsReasoningBatch as Pt, splitOpenAIFunctionCallPairing as Q, resolveReplayableResponsesMessageId as R, isInvalidEncryptedContentError as S, reconcileOpenAICompletionsToolChoice as St, resolveNextResponsesEncryptedContentAttempt as T, GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP as Tt, recordResponsesContinuationState as U, responsesInputFingerprint as V, resolveResponsesContinuationRequest as W, normalizeOpenAIResponsesFunctionCallId as X, OPENAI_RESPONSES_ID_MAX_LENGTH as Y, shouldNormalizeOpenAIResponsesToolCallId as Z, isResponsesTextContentPartType as _, resolveOpenAIRequestReasoning as _t, prepareResponsesTools as a, resolveNewestOpenAIResponsesCompactionReplay as at, commitResponsesEncryptedContentAttempt as b, prepareOpenAITools as bt, observeResponsesStream as c, buildOpenAIClientHeaders as ct, AZURE_RESPONSES_TEXT_CONTENT_PART_TYPE as d, enforceCodeModeResponsesToolSurface as dt, replayResponsesReasoningUpdates as et, AZURE_RESPONSES_TEXT_DELTA_EVENT_TYPE as f, filterCodeModePayloadTools as ft, isAzureResponsesTextDeltaEventType as g, resolveOpenAIStrictToolFlagWithDiagnostics as gt, isAzureResponsesTextDeltaEvent as h, resolveCodeModeResponsesVisibleToolNames as ht, convertResponsesToolPayload as i, captureOpenAIResponsesCompaction as it, recordResponsesContextUsage as j, measureUtf8AppendBytes as jt, summarizeOpenAITransportError as k, isOpenAICompletionsThinkingEnabled as kt, createResponsesToolCallTracker as l, buildOpenAISdkClientOptions as lt, OPENAI_RESPONSES_OUTPUT_TEXT_DELTA_EVENT_TYPE as m, readCodeModePayloadToolName as mt, applyResponsesServiceTierPricing as n, CompactionReplayRefreshRequiredError as nt, readOpenAIMisalignmentReview as o, suppressOpenAIResponsesCompaction as ot, OPENAI_RESPONSES_OUTPUT_TEXT_CONTENT_PART_TYPE as p, getCompat as pt, OPENAI_RESPONSES_CALL_ID_RE as q, runResponsesStreamLifecycle as r, buildOpenAIResponsesReasoningReplayMetadata as rt, processResponsesStream as s, assertCodeModeResponsesToolSurface as st, applyCommonResponsesParams as t, supportsResponsesReasoningUpdate as tt, readResponsesToolCallItemIdentity as u, buildOpenAISdkRequestOptions as ut, isResponsesTextDeltaEventType as v, resolveOpenAISimpleReasoningEffort as vt, resolveAzureOpenAIApiVersion as w, toOpenAIResponsesToolChoice as wt, createResponsesStreamWithRecovery as x, projectOpenAITools as xt, resolveResponsesMessageSnapshotCollapse as y, resolveOpenAIReasoningEffortMap as yt, readResponsesInputReplayState as z };