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

package/dist/{openai-completions-BPnt4Sml.mjs → openai-completions-D88orSJ9.mjs}

@@ -1,18 +1,17 @@
 import { n as getEnvApiKey, r as __exportAll } from "./env-api-keys-DrgeBuva.mjs";
 import { t as AssistantMessageEventStream } from "./event-stream-uSMZJ3FA.mjs";
-import { i as clampThinkingLevel } from "./sanitize-unicode-BYqrYtC_.mjs";
-import { n as getAiTransportHost } from "./host-DTqNc7ad.mjs";
-import { n as projectProviderError } from "./provider-error-BUwEnjXq.mjs";
+import { i as clampThinkingLevel } from "./sanitize-unicode-BP7HLLZn.mjs";
+import { n as getAiTransportHost } from "./host-DF_yxOsA.mjs";
+import { n as projectProviderError } from "./provider-error-C8mBRJDl.mjs";
 import { n as clampOpenAIPromptCacheKey } from "./openai-prompt-cache-mZTCdRPo.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
-import { c as splitSystemPromptCacheBoundary, f as sortPromptCacheToolsByName, n as buildBaseOptions } from "./simple-options-D58D5Kvw.mjs";
-import { a as createOpenAICompletionsToolCallDeltaNormalizer, c as resolveOpenAICompletionsCompat, d as clearPendingCommentaryText, f as rememberPendingCommentaryTags, i as hasToolCallHistory, l as resolveOpenAICompletionsResponseFormat, n as resolveOpenAIReasoningEffortMap, o as finalizeOpenAICompletionsToolCalls, p as tagPendingCommentaryText, r as convertMessages, t as mapOpenAIStopReason, u as shouldOmitOllamaCompatResponseFormat } from "./openai-stop-reason-BkFkqqK0.mjs";
+import { A as splitSystemPromptCacheBoundary, C as withProviderResponseHook, P as sortPromptCacheToolsByName, n as buildBaseOptions, x as transportAbortError } from "./simple-options-oqsXB6U3.mjs";
+import { a as createOpenAICompletionsToolCallDeltaNormalizer, c as resolveOpenAICompletionsCompat, d as clearPendingCommentaryText, f as rememberPendingCommentaryTags, h as tagUnresolvedTextAsCommentary, i as hasToolCallHistory, l as resolveOpenAICompletionsResponseFormat, m as tagPendingCommentaryText, n as resolveOpenAIReasoningEffortMap, o as finalizeOpenAICompletionsToolCalls, p as tagInterruptedTextPhases, r as convertMessages, t as mapOpenAIStopReason, u as shouldOmitOllamaCompatResponseFormat } from "./openai-stop-reason-9OONKyAm.mjs";
 import { n as parseStreamingJson } from "./json-parse-CDnesDM_.mjs";
-import { n as notifyLlmRequestActivity } from "./headers-DdOQtGuU.mjs";
-import { f as transportAbortError, p as withProviderResponseHook } from "./provider-transcript-transform-ePx-Bbfr.mjs";
-import { d as readOpenAICompletionsContentDeltas, n as reconcileOpenAICompletionsToolChoice, o as createOpenAIResponseHook, s as isOpenAICompletionsThinkingEnabled, t as projectOpenAITools, u as parseOpenAICompletionsUsage } from "./openai-tool-projection-CY04OcvQ.mjs";
+import { t as notifyLlmRequestActivity } from "./llm-request-activity-BjtkplhG.mjs";
+import { c as isOpenAICompletionsThinkingEnabled, d as parseOpenAICompletionsUsage, f as readOpenAICompletionsContentDeltas, n as reconcileOpenAICompletionsToolChoice, o as createOpenAIProviderAcceptanceHook, p as readOpenAICompletionsReasoningBatch, t as projectOpenAITools } from "./openai-tool-projection-CHGLt03m.mjs";
 import { a as withFirstStreamEventTimeout, i as getFirstStreamEventTimeoutMs, r as getFirstStreamEventTimeoutHandler, t as createFirstStreamEventAbortController } from "./stream-first-event-timeout-MK28puvq.mjs";
-import { t as createReasoningTagTextPartitioner } from "./reasoning-tag-text-partitioner-rnPwX2pg.mjs";
+import { t as createReasoningTagTextPartitioner } from "./reasoning-tag-text-partitioner-5ygO2rZc.mjs";
 import { i as resolveCloudflareBaseUrl, n as hasCopilotVisionInput, r as isCloudflareProvider, t as buildCopilotDynamicHeaders } from "./github-copilot-headers-NCJtz9i0.mjs";
 import OpenAI from "openai";
 //#region packages/ai/src/providers/openai-completions.ts
@@ -57,6 +56,8 @@ const streamOpenAICompletions = (model, context, options) => {
 		try {
 			const apiKey = options?.apiKey || getEnvApiKey(model.provider) || "";
 			const compat = resolveOpenAICompletionsCompat(model);
+			const visibleReasoningDetailTypes = new Set(compat.visibleReasoningDetailTypes);
+			const shouldEmitReasoning = Boolean(model.reasoning && options?.reasoningEffort && isOpenAICompletionsThinkingEnabled(options.reasoningEffort));
 			const cacheRetention = resolveCacheRetention(options?.cacheRetention);
 			const cacheSessionId = cacheRetention === "none" ? void 0 : options?.sessionId;
 			const client = createClient(model, context, apiKey, options?.headers, cacheSessionId, compat);
@@ -74,14 +75,17 @@ const streamOpenAICompletions = (model, context, options) => {
 				stream: openaiStream,
 				signal: firstEventAbort.signal,
 				abort: firstEventAbort.abort,
-				hook: createOpenAIResponseHook(options?.onResponse, response, model),
+				hook: createOpenAIProviderAcceptanceHook(options, response, model),
 				onReady: () => stream.push({
 					type: "start",
 					partial: output
 				})
 			});
 			let textBlock = null;
+			let textBlockSource;
 			let thinkingBlock = null;
+			let pendingInterruptedTextBlock = null;
+			let confirmedInterruptedTextBlock = null;
 			let hasFinishReason = false;
 			const toolCallBlocksByIndex = /* @__PURE__ */ new Map();
 			const toolCallBlocksById = /* @__PURE__ */ new Map();
@@ -91,6 +95,7 @@ const streamOpenAICompletions = (model, context, options) => {
 			const blocks = output.content;
 			const finishedBlocks = /* @__PURE__ */ new Set();
 			const contentIndices = /* @__PURE__ */ new WeakMap();
+			let explicitVisibleTextBlocks;
 			const appendBlock = (block) => {
 				contentIndices.set(block, blocks.length);
 				blocks.push(block);
@@ -128,12 +133,21 @@ const streamOpenAICompletions = (model, context, options) => {
 					partial: output
 				});
 			};
-			const ensureTextBlock = () => {
+			const finishTextBlock = () => {
+				if (!textBlock) return;
+				finishBlock(textBlock);
+				textBlock = null;
+				textBlockSource = void 0;
+			};
+			const ensureTextBlock = (source) => {
+				if (textBlock && textBlockSource !== source) finishTextBlock();
 				if (!textBlock) {
 					textBlock = {
 						type: "text",
 						text: ""
 					};
+					textBlockSource = source;
+					if (source === "reasoning_detail") (explicitVisibleTextBlocks ??= /* @__PURE__ */ new Set()).add(textBlock);
 					appendBlock(textBlock);
 					stream.push({
 						type: "text_start",
@@ -165,10 +179,14 @@ const streamOpenAICompletions = (model, context, options) => {
 					thinkingBlock = null;
 				}
 			};
-			const appendTextDelta = (delta) => {
+			const appendTextDelta = (delta, source) => {
 				sealNativeReasoningBeforeText();
-				const block = ensureTextBlock();
+				const block = ensureTextBlock(source);
 				block.text += delta;
+				if (pendingInterruptedTextBlock && delta.trim()) {
+					confirmedInterruptedTextBlock = pendingInterruptedTextBlock;
+					pendingInterruptedTextBlock = null;
+				}
 				stream.push({
 					type: "text_delta",
 					contentIndex: getContentIndex(block),
@@ -186,6 +204,15 @@ const streamOpenAICompletions = (model, context, options) => {
 					partial: output
 				});
 			};
+			const appendReasoningDeltas = (reasoningDeltas) => {
+				for (const reasoningDelta of reasoningDeltas) if (reasoningDelta.kind === "thinking") {
+					if (!shouldEmitReasoning) continue;
+					finishTextBlock();
+					const signature = reasoningDelta.signature;
+					const thinkingSignature = model.provider === "opencode-go" && signature === "reasoning" ? "reasoning_content" : signature;
+					appendThinkingDelta(thinkingSignature, reasoningDelta.text);
+				} else appendTextDelta(reasoningDelta.text, reasoningDelta.source);
+			};
 			const ensureToolCallBlock = (toolCall) => {
 				const streamIndex = typeof toolCall.index === "number" ? toolCall.index : void 0;
 				let block = streamIndex !== void 0 ? toolCallBlocksByIndex.get(streamIndex) : void 0;
@@ -226,6 +253,21 @@ const streamOpenAICompletions = (model, context, options) => {
 			const flushPartitionedContent = () => {
 				for (const delta of reasoningTagTextPartitioner.flush()) if (delta.kind === "text") appendTextDelta(delta.text);
 			};
+			const sealTextBeforeReasoning = () => {
+				if (!textBlock && !reasoningTagTextPartitioner.hasPending()) return;
+				flushPartitionedContent();
+				if (!textBlock) return;
+				if (textBlockSource !== "reasoning_detail" && textBlock.text.trim()) pendingInterruptedTextBlock = textBlock;
+				finishTextBlock();
+			};
+			const beginReasoning = (hasFollowingVisibleText, forceStrict = false) => {
+				if (!output.openclawDelivery?.textPhaseRequiresTerminal) output.openclawDelivery = {
+					...output.openclawDelivery,
+					textPhaseRequiresTerminal: true
+				};
+				if (forceStrict || reasoningTagTextPartitioner.hasPending()) reasoningTagTextPartitioner.markStrict();
+				if (!hasFollowingVisibleText || !reasoningTagTextPartitioner.hasPendingSyntax()) sealTextBeforeReasoning();
+			};
 			const guardedOpenaiStream = withFirstStreamEventTimeout(hookedOpenAIStream, {
 				provider: model.provider,
 				api: model.api,
@@ -255,30 +297,21 @@ const streamOpenAICompletions = (model, context, options) => {
 				const rawChoiceDelta = choice.delta ?? choice.message;
 				if (rawChoiceDelta) for (const normalizedDelta of normalizeToolCallDeltas(rawChoiceDelta, choice.finish_reason)) {
 					const choiceDelta = normalizedDelta.delta;
-					const reasoningFields = [
-						"reasoning_content",
-						"reasoning",
-						"reasoning_text"
-					];
-					const deltaFields = choiceDelta;
-					const shouldEmitReasoning = Boolean(model.reasoning && options?.reasoningEffort && isOpenAICompletionsThinkingEnabled(options.reasoningEffort));
-					let foundReasoningField = null;
-					for (const field of reasoningFields) {
-						const value = deltaFields[field];
-						if (typeof value === "string" && value.length > 0) {
-							foundReasoningField = field;
-							break;
-						}
-					}
-					if (foundReasoningField) reasoningTagTextPartitioner.markStrict();
-					if (shouldEmitReasoning && foundReasoningField) {
-						const delta = deltaFields[foundReasoningField];
-						if (typeof delta === "string" && delta.length > 0) appendThinkingDelta(model.provider === "opencode-go" && foundReasoningField === "reasoning" ? "reasoning_content" : foundReasoningField, delta);
+					const reasoningBatch = readOpenAICompletionsReasoningBatch(choiceDelta, visibleReasoningDetailTypes);
+					const reasoningDeltas = reasoningBatch.deltas;
+					const hasReasoningThinking = reasoningBatch.hasThinking;
+					const contentDeltas = readOpenAICompletionsContentDeltas(choiceDelta.content, choiceDelta.refusal, reasoningBatch.mirroredThinking);
+					const lastVisibleTextIndex = contentDeltas.findLastIndex((delta) => delta.kind === "text");
+					const hasSameChunkVisibleText = reasoningBatch.hasVisibleText || lastVisibleTextIndex !== -1;
+					if (hasReasoningThinking) {
+						beginReasoning(hasSameChunkVisibleText, true);
+						appendReasoningDeltas(reasoningDeltas);
 					}
-					for (const contentDelta of readOpenAICompletionsContentDeltas(choiceDelta.content, choiceDelta.refusal, foundReasoningField ? [deltaFields[foundReasoningField]] : [])) if (contentDelta.kind === "thinking") {
-						if (reasoningTagTextPartitioner.hasPending()) reasoningTagTextPartitioner.markStrict();
+					for (const [contentDeltaIndex, contentDelta] of contentDeltas.entries()) if (contentDelta.kind === "thinking") {
+						beginReasoning(contentDeltaIndex < lastVisibleTextIndex);
 						if (shouldEmitReasoning) appendThinkingDelta(contentDelta.signature, contentDelta.text);
-					} else appendPartitionedContent(contentDelta.text, Boolean(foundReasoningField));
+					} else appendPartitionedContent(contentDelta.text, hasReasoningThinking);
+					if (!hasReasoningThinking) appendReasoningDeltas(reasoningDeltas);
 					const toolCallDeltas = normalizedDelta.toolCalls;
 					if (toolCallDeltas.length > 0) {
 						flushPartitionedContent();
@@ -332,6 +365,7 @@ const streamOpenAICompletions = (model, context, options) => {
 				for (const block of blocks) if (block.type !== "toolCall") finishBlock(block);
 				throw new Error(output.errorMessage || (output.stopReason === "aborted" ? "Request was aborted" : "Provider returned an invalid tool call"));
 			}
+			if (output.stopReason !== "toolUse" && confirmedInterruptedTextBlock) tagInterruptedTextPhases(output.content, confirmedInterruptedTextBlock, explicitVisibleTextBlocks);
 			for (const block of blocks) if (block.type !== "toolCall" || output.stopReason === "toolUse") finishBlock(block);
 			if (output.stopReason !== "toolUse") clearPendingCommentaryText(provisionalCommentaryTags);
 			if (output.stopReason === "toolUse") tagPendingCommentaryText(output.content);
@@ -345,6 +379,8 @@ const streamOpenAICompletions = (model, context, options) => {
 			const terminal = projectProviderError(error, options?.signal);
 			Object.assign(output, terminal);
 			finalizeOpenAICompletionsToolCalls(output, { allowSilentToolCallPromotion: false });
+			clearPendingCommentaryText(provisionalCommentaryTags);
+			tagUnresolvedTextAsCommentary(output);
 			for (const block of output.content) {
 				delete block.index;
 				delete block.partialArgs;

package/dist/{openai-completions-compat-Dt3dcawL.d.mts → openai-completions-compat-B0KTleXp.d.mts}

@@ -1,5 +1,5 @@
-import { E as Model, O as OpenAICompletionsCompat } from "./types-BHNrPS1l.mjs";
-import { a as AiProviderRequestPolicyInput, i as AiProviderRequestCapabilities } from "./host-vWgMMhiJ.mjs";
+import { D as Model, k as OpenAICompletionsCompat } from "./types-DTAP49HF.mjs";
+import { a as AiProviderRequestPolicyInput, i as AiProviderRequestCapabilities } from "./host-DJq_FGcz.mjs";
 
 //#region packages/ai/src/transports/openai-completions-compat.d.ts
 type ProviderRequestCapabilities = AiProviderRequestCapabilities;

package/dist/{openai-responses-DhIKtOup.mjs → openai-responses-BXEcr-Tp.mjs}

@@ -1,11 +1,11 @@
 import { n as getEnvApiKey, r as __exportAll } from "./env-api-keys-DrgeBuva.mjs";
 import { t as AssistantMessageEventStream } from "./event-stream-uSMZJ3FA.mjs";
-import { n as getAiTransportHost } from "./host-DTqNc7ad.mjs";
-import { S as supportsOpenAITemperature } from "./openai-responses-contracts-CyfIkQi5.mjs";
+import { n as getAiTransportHost } from "./host-DF_yxOsA.mjs";
+import { S as supportsOpenAITemperature } from "./openai-responses-contracts-DvILoqBk.mjs";
 import { n as clampOpenAIPromptCacheKey } from "./openai-prompt-cache-mZTCdRPo.mjs";
 import { t as resolveCacheRetention } from "./cache-retention-0x979a5V.mjs";
-import { n as buildBaseOptions } from "./simple-options-D58D5Kvw.mjs";
-import { a as resolveResponsesReasoningEffort, i as createResponsesAssistantOutput, n as applyResponsesServiceTierPricing, o as runResponsesStreamLifecycle, r as convertResponsesMessages, t as applyCommonResponsesParams } from "./openai-responses-shared-DXIt3iY5.mjs";
+import { n as buildBaseOptions } from "./simple-options-oqsXB6U3.mjs";
+import { a as resolveResponsesReasoningEffort, i as createResponsesAssistantOutput, n as applyResponsesServiceTierPricing, o as runResponsesStreamLifecycle, r as convertResponsesMessages, t as applyCommonResponsesParams } from "./openai-responses-shared-Cvt5MgEm.mjs";
 import { i as resolveCloudflareBaseUrl, n as hasCopilotVisionInput, r as isCloudflareProvider, t as buildCopilotDynamicHeaders } from "./github-copilot-headers-NCJtz9i0.mjs";
 import OpenAI from "openai";
 //#region packages/ai/src/providers/openai-responses.ts

package/dist/{openai-responses-contracts-XpZJxrRG.d.mts → openai-responses-contracts-CD75pQ0m.d.mts}

@@ -1,5 +1,5 @@
-import { n as Api } from "./types-BHNrPS1l.mjs";
-import { f as OpenAIApiReasoningEffort, i as BaseOpenAIStreamOptions, p as OpenAIReasoningEffort } from "./provider-options-B96RdNpH.mjs";
+import { n as Api } from "./types-DTAP49HF.mjs";
+import { h as OpenAIReasoningEffort, i as BaseOpenAIStreamOptions, m as OpenAIApiReasoningEffort } from "./provider-options-Cc5ZQ6Aa.mjs";
 import { FunctionTool, ResponseCompactionItem, ResponseCreateParamsStreaming, ResponseInput } from "openai/resources/responses/responses.js";
 
 //#region packages/ai/src/transports/openai-responses-contracts.d.ts

package/dist/{openai-responses-contracts-CyfIkQi5.mjs → openai-responses-contracts-DvILoqBk.mjs}

@@ -1,4 +1,4 @@
-import { n as PROVIDER_POST_DISPATCH_AMBIGUITY_ERROR_CODE, t as PROVIDER_FAILURE_WITH_OUTPUT_ERROR_CODE } from "./types-BDdaOVi2.mjs";
+import { n as PROVIDER_POST_DISPATCH_AMBIGUITY_ERROR_CODE } from "./types-BDdaOVi2.mjs";
 import "./src-D2H6yKkH.mjs";
 import { a as normalizeOptionalString, n as normalizeLowercaseStringOrEmpty } from "./string-coerce-fsri9iCu.mjs";
 import { a as isRecord } from "./record-coerce-DdXsgUd_.mjs";
@@ -177,6 +177,7 @@ const RESPONSE_FAILED_NO_DETAILS_MESSAGE = "Unknown error (no error details in r
 const OPENAI_RESPONSES_REASONING_REPLAY_META_KEY = "__openclaw_replay";
 const OPENAI_RESPONSES_REASONING_REPLAY_BLOCK_META_KEY = "openclawReasoningReplay";
 const OPENAI_RESPONSES_COMPACTION_REPLAY_TYPE = "openai-responses-compaction";
+const OPENAI_RESPONSES_RETAINED_COMPACTION_REPLAY_TYPE = "openai-responses-retained-compaction";
 const OPENAI_RESPONSES_APIS = /* @__PURE__ */ new Set([
 	"openai-responses",
 	"azure-openai-responses",
@@ -185,13 +186,6 @@ const OPENAI_RESPONSES_APIS = /* @__PURE__ */ new Set([
 	"openclaw-openai-chatgpt-responses-transport",
 	"openclaw-azure-openai-responses-transport"
 ]);
-var OpenAIResponsesWebSocketResponseFailedError = class extends Error {
-	constructor(hasOutput) {
-		super("OpenAI Responses WebSocket returned response.failed");
-		this.name = "OpenAIResponsesWebSocketResponseFailedError";
-		this.code = hasOutput ? PROVIDER_FAILURE_WITH_OUTPUT_ERROR_CODE : PROVIDER_POST_DISPATCH_AMBIGUITY_ERROR_CODE;
-	}
-};
 var OpenAIResponsesWebSocketPreDispatchError = class extends Error {
 	constructor(cause) {
 		super("OpenAI Responses WebSocket failed before request dispatch", { cause });
@@ -250,4 +244,4 @@ const responsesPromptObserver = {
 	}
 };
 //#endregion
-export { normalizeStringEntries as C, supportsOpenAITemperature as S, uniqueValues as T, isOpenAIGpt56Model as _, OPENAI_RESPONSES_COMPACTION_REPLAY_TYPE as a, resolveOpenAISupportedReasoningEfforts as b, OpenAIResponsesWebSocketPostDispatchError as c, OpenAIResponsesWebSocketSafeRetryError as d, RESPONSE_FAILED_NO_DETAILS_MESSAGE as f, isOpenAIGpt55Model as g, isOpenAIGpt54MiniModel as h, OPENAI_RESPONSES_APIS as i, OpenAIResponsesWebSocketPreDispatchError as l, responsesPromptObserver as m, DEFAULT_AZURE_OPENAI_API_VERSION as n, OPENAI_RESPONSES_REASONING_REPLAY_BLOCK_META_KEY as o, parseOpenAIResponsesWebSocketServerError as p, OPENAI_CODEX_RESPONSES_DEFAULT_INSTRUCTIONS as r, OPENAI_RESPONSES_REASONING_REPLAY_META_KEY as s, AZURE_RESPONSES_FIRST_EVENT_TIMEOUT_MS as t, OpenAIResponsesWebSocketResponseFailedError as u, normalizeOpenAIReasoningEffort as v, uniqueStrings as w, supportsOpenAIReasoningEffort as x, resolveOpenAIReasoningEffortForModel as y };
+export { normalizeStringEntries as C, supportsOpenAITemperature as S, uniqueValues as T, isOpenAIGpt56Model as _, OPENAI_RESPONSES_COMPACTION_REPLAY_TYPE as a, resolveOpenAISupportedReasoningEfforts as b, OPENAI_RESPONSES_RETAINED_COMPACTION_REPLAY_TYPE as c, OpenAIResponsesWebSocketSafeRetryError as d, RESPONSE_FAILED_NO_DETAILS_MESSAGE as f, isOpenAIGpt55Model as g, isOpenAIGpt54MiniModel as h, OPENAI_RESPONSES_APIS as i, OpenAIResponsesWebSocketPostDispatchError as l, responsesPromptObserver as m, DEFAULT_AZURE_OPENAI_API_VERSION as n, OPENAI_RESPONSES_REASONING_REPLAY_BLOCK_META_KEY as o, parseOpenAIResponsesWebSocketServerError as p, OPENAI_CODEX_RESPONSES_DEFAULT_INSTRUCTIONS as r, OPENAI_RESPONSES_REASONING_REPLAY_META_KEY as s, AZURE_RESPONSES_FIRST_EVENT_TIMEOUT_MS as t, OpenAIResponsesWebSocketPreDispatchError as u, normalizeOpenAIReasoningEffort as v, uniqueStrings as w, supportsOpenAIReasoningEffort as x, resolveOpenAIReasoningEffortForModel as y };

package/dist/{openai-responses-payload-policy-BDxV-W0c.mjs → openai-responses-payload-policy-BDo8CQsx.mjs}

@@ -1,6 +1,6 @@
 import { i as normalizeOptionalLowercaseString, n as normalizeLowercaseStringOrEmpty, s as readStringValue } from "./string-coerce-fsri9iCu.mjs";
 import { s as parseStrictPositiveInteger } from "./number-coercion-H9qHik3g.mjs";
-import { i as OPENAI_RESPONSES_APIS, x as supportsOpenAIReasoningEffort } from "./openai-responses-contracts-CyfIkQi5.mjs";
+import { i as OPENAI_RESPONSES_APIS, x as supportsOpenAIReasoningEffort } from "./openai-responses-contracts-DvILoqBk.mjs";
 //#region packages/model-catalog-core/src/provider-id.ts
 function normalizeProviderId(provider) {
 	return normalizeLowercaseStringOrEmpty(provider);
@@ -132,8 +132,10 @@ function parsePositiveInteger(value) {
 	if (typeof value === "string") return parseStrictPositiveInteger(value);
 }
 function resolveOpenAIResponsesCompactThreshold(model) {
+	const contextTokens = parsePositiveInteger(model.contextTokens);
 	const contextWindow = parsePositiveInteger(model.contextWindow);
-	if (contextWindow) return Math.max(1e3, Math.floor(contextWindow * .7));
+	const effectiveBudget = contextTokens && contextWindow ? Math.min(contextTokens, contextWindow) : contextTokens || contextWindow;
+	if (effectiveBudget) return Math.max(1e3, Math.floor(effectiveBudget * .7));
 	return 8e4;
 }
 /** Resolve the server-compaction gate and effective threshold for a Responses route. */

package/dist/{openai-responses-payload-policy-BSs371VM.d.mts → openai-responses-payload-policy-CNGFT9zr.d.mts}

@@ -4,6 +4,7 @@ type OpenAIResponsesPayloadModel = {
   baseUrl?: unknown;
   id?: unknown;
   provider?: unknown;
+  contextTokens?: unknown;
   contextWindow?: unknown;
   compat?: unknown;
 };

package/dist/{openai-responses-prompt-observer-internal-DgNTYnRY.mjs → openai-responses-prompt-observer-internal-BG-1AsBI.mjs}

@@ -1,6 +1,5 @@
-import { m as responsesPromptObserver } from "./openai-responses-contracts-CyfIkQi5.mjs";
-import { l as stripSystemPromptCacheBoundary } from "./simple-options-D58D5Kvw.mjs";
-import { d as sanitizeTransportPayloadText } from "./provider-transcript-transform-ePx-Bbfr.mjs";
+import { m as responsesPromptObserver } from "./openai-responses-contracts-DvILoqBk.mjs";
+import { b as sanitizeTransportPayloadText, j as stripSystemPromptCacheBoundary } from "./simple-options-oqsXB6U3.mjs";
 //#region packages/normalization-core/src/error-coercion.ts
 /**
 * Normalizes an unknown thrown value into an Error. Non-Error objects become