@xiaoso/dsh-tool-plus 0.1.5 → 0.1.7-beta.0

package/lib/{openai-completions-Bq4WT47M.mjs → openai-completions-BK2f2dBv.mjs}

@@ -1,28 +1,49 @@
 import { installBunShim as __installBunShim } from "./bun-shim.mjs";
 __installBunShim();
 import { r as __exportAll } from "./fs-promises-shim-Bx0fpxuk.mjs";
-import { P as $env, T as warn, _ as parseStreamingJsonThrottled, g as parseStreamingJson } from "./src-BcD0_rQR.mjs";
-import { B as OpenAIHttpError, d as callWithCopilotModelRetry, l as finalize, n as StreamTimeoutError, s as ProviderResponseError, t as ConfigurationError, v as AbortError } from "./error-C-f_Wnw4.mjs";
-import { A as NO_STRICT, F as sanitizeSchemaForGrammar, O as flattenExclusiveRequiredRootUnion, P as normalizeSchemaForMoonshot, S as kStreamingLastParseLen, a as normalizeSystemPrompts, g as isDemotedThinking, j as adaptSchemaForStrict, k as toolWireSchema, m as transformMessages, s as resolveCacheRetention, w as renderDemotedThinking } from "./utils-Cp2Mxf_U.mjs";
-import { $ as isOpenRouterAnthropicModel, Ct as withEmptyCompletionRetry, I as disableStrictToolsForScope, J as getOpenAIStrictToolsScope, K as getOpenAIPromptCacheKey, N as createInitialResponsesAssistantMessage, O as calculateOpenAIUsageAccounting, P as createOpenAIStrictToolsState, St as hasVisibleAssistantContent, T as applyWireModelIdTransform, Z as isCompiledGrammarTooLargeStrictError, _ as applyOpenAIExtraBody, _t as isForcedToolChoice, b as applyOpenRouterReportedCost, ct as resolveOpenAICompatPolicy, dt as resolveOpenAIRequestSetup, et as isStrictToolsDisabledForScope, h as applyChatCompletionsToolStream, it as parseAzureDeploymentNameMap, k as clearOpenAIStrictToolsState, lt as resolveOpenAICompletionsOutputClamp, m as applyChatCompletionsCompatPolicy, n as getEnvApiKey, pt as shouldRetryWithoutStrictTools, ut as resolveOpenAIOutputTokenParam, v as applyOpenAIGatewayRouting, vt as mapToOpenAICompletionsToolChoice, y as applyOpenAIServiceTier } from "./stream-Bjq8wjfr.mjs";
-import { t as createAbortSourceTracker } from "./abort-Cwwm4Err.mjs";
-import { b as isKimiModelId, c as resolveWireModelId, t as calculateCost } from "./models-DRA2c_8U.mjs";
-import "./identity-HNO1Oe9B.mjs";
-import { n as isDashscopeCompatibleModeTextOnlyQwen, r as joinTextWithImagePlaceholder, t as NON_VISION_IMAGE_PLACEHOLDER } from "./vision-guard-mBmtAHx0.mjs";
-import { t as findStrictToolSchemaViolation } from "./strict-tool-validation-imVLUcbI.mjs";
-import { t as AssistantMessageEventStream } from "./event-stream-Dh2guZxS.mjs";
-import { n as getOpenAIStreamFirstEventTimeoutMs, o as iterateWithIdleTimeout, r as getOpenAIStreamIdleTimeoutMs, s as iterateWithTerminalGrace } from "./idle-iterator-CYAxUkSD.mjs";
-import { t as getKimiCommonHeaders } from "./kimi-DgcUpPtK.mjs";
-import { t as notifyProviderResponse } from "./provider-response-C6vcbQyO.mjs";
-import { t as StreamMarkupHealing } from "./stream-markup-healing-DZ6bVDxH.mjs";
-import { a as getOpenAIReasoningEffortFallback, c as postOpenAIStream, i as createOpenAIReasoningEffortFallbackState, n as clearOpenAIReasoningEffortFallbackState, o as rememberOpenAIReasoningEffortFallback, r as createOpenAIReasoningEffortFallbackKey, s as resolveOpenAIReasoningEffortFallback, t as applyOpenAIReasoningEffortFallback } from "./openai-reasoning-fallback-DXsfiSXO.mjs";
-//#region node_modules/.pnpm/@oh-my-pi+pi-ai@17.3.5/node_modules/@oh-my-pi/pi-ai/src/providers/openai-completions.ts
+import { C as warn, N as $env, _ as parseStreamingJson, v as parseStreamingJsonThrottled } from "./src-BwW51vA5.mjs";
+import { B as OpenAIHttpError, V as ProviderHttpError, _ as AbortError, c as finalize, n as StreamTimeoutError, s as ProviderResponseError, t as ConfigurationError } from "./error-CORgGg5a.mjs";
+import { A as toolWireSchema, C as kStreamingLastParseLen, F as normalizeSchemaForMoonshot, L as sanitizeSchemaForGrammar, M as adaptSchemaForStrict, T as renderDemotedThinking, _ as isDemotedThinking, c as resolveCacheRetention, h as transformMessages, j as NO_STRICT, k as flattenExclusiveRequiredRootUnion, o as normalizeSystemPrompts } from "./utils-CUPhpw9A.mjs";
+import { At as withReplaySafeStreamRetry, C as applyChatCompletionsToolStream, D as applyOpenAIServiceTier, E as applyOpenAIGatewayRouting, H as createOpenAIStrictToolsState, I as calculateOpenAIUsageAccounting, L as clearOpenAIStrictToolsState, N as applyWireModelIdTransform, O as applyProviderReportedCost, S as applyChatCompletionsReasoningParams, St as shouldRetryWithoutStrictTools, T as applyOpenAIExtraBody, Tt as mapToOpenAICompletionsToolChoice, V as createInitialResponsesAssistantMessage, W as disableStrictToolsForScope, _t as resolveOpenAIOutputTokenParam, at as isCompiledGrammarTooLargeStrictError, dt as parseAzureDeploymentNameMap, et as getOpenAIPromptCacheKey, gt as resolveOpenAICompletionsOutputClamp, ht as resolveOpenAICompatPolicy, kt as hasVisibleAssistantContent, n as getEnvApiKey, nt as getOpenAIStrictToolsScope, st as isStrictToolsDisabledForScope, vt as resolveOpenAIRequestSetup, wt as isForcedToolChoice, xt as shouldDropAutoToolChoiceForReasoning } from "./stream-B_7qN-bA.mjs";
+import { s as resolveWireModelId } from "./anthropic-wire-CKuhDwZ1.mjs";
+import { t as createAbortSourceTracker } from "./abort-Caqff3M6.mjs";
+import { t as calculateCost } from "./models-DnXMYGcJ.mjs";
+import { i as resolveCopilotRequestIdentity, o as wrapFetchForCopilotFallback } from "./github-copilot-headers-DOh1gt_w.mjs";
+import { n as isOpenAICompletionsVisionSupported, r as joinTextWithImagePlaceholder, t as NON_VISION_IMAGE_PLACEHOLDER } from "./vision-guard-_XwsL1DF.mjs";
+import { t as findStrictToolSchemaViolation } from "./strict-tool-validation-CBTz1dAA.mjs";
+import { t as AssistantMessageEventStream } from "./event-stream-ulcDenZD.mjs";
+import { n as clinePassClientHeaders } from "./cline-pass-Bwg3esu9.mjs";
+import { n as getOpenAIStreamFirstEventTimeoutMs, o as iterateWithIdleTimeout, r as getOpenAIStreamIdleTimeoutMs, s as iterateWithTerminalGrace } from "./idle-iterator-DlUkEfzW.mjs";
+import { t as getKimiCommonHeaders } from "./kimi-C2c_Zq8A.mjs";
+import { t as notifyProviderResponse } from "./provider-response-SqKSkkUc.mjs";
+import { a as getOpenAIReasoningEffortFallback, c as postOpenAIStream, i as createOpenAIReasoningEffortFallbackState, n as clearOpenAIReasoningEffortFallbackState, o as rememberOpenAIReasoningEffortFallback, r as createOpenAIReasoningEffortFallbackKey, s as resolveOpenAIReasoningEffortFallback, t as applyOpenAIReasoningEffortFallback } from "./openai-reasoning-fallback-BdB3RCbj.mjs";
+import { t as StreamMarkupHealing } from "./stream-markup-healing-DKT1lnkW.mjs";
+//#region node_modules/.pnpm/@oh-my-pi+pi-ai@18.1.17/node_modules/@oh-my-pi/pi-ai/src/providers/openai-completions.ts
 var openai_completions_exports = /* @__PURE__ */ __exportAll({
 	convertMessages: () => convertMessages,
 	isOpenAICompletionsProgressChunk: () => isOpenAICompletionsProgressChunk,
 	parseChunkUsage: () => parseChunkUsage,
 	streamOpenAICompletions: () => streamOpenAICompletions
 });
+const GEMINI_THOUGHT_SIGNATURE_NAMESPACES = ["google", "vertex"];
+function getGeminiThoughtSignatureExtraContent(value) {
+	if (typeof value !== "object" || value === null) return void 0;
+	for (const namespace of GEMINI_THOUGHT_SIGNATURE_NAMESPACES) {
+		const providerContent = Reflect.get(value, namespace);
+		if (typeof providerContent !== "object" || providerContent === null) continue;
+		const thoughtSignature = Reflect.get(providerContent, "thought_signature");
+		if (typeof thoughtSignature !== "string" || thoughtSignature.length === 0) continue;
+		return namespace === "google" ? { google: { thought_signature: thoughtSignature } } : { vertex: { thought_signature: thoughtSignature } };
+	}
+}
+function parseGeminiThoughtSignatureExtraContent(thoughtSignature) {
+	if (!thoughtSignature) return void 0;
+	try {
+		return getGeminiThoughtSignatureExtraContent(JSON.parse(thoughtSignature));
+	} catch {
+		return;
+	}
+}
 function firstPositiveNumber(...values) {
 	for (const value of values) if (typeof value === "number" && value > 0) return value;
 	return 0;
@@ -31,6 +52,7 @@ function hasPositiveCacheReadTokenField(rawUsage) {
 	const usageLike = rawUsage;
 	if (typeof usageLike.cached_tokens === "number" && usageLike.cached_tokens > 0) return true;
 	if (typeof usageLike.prompt_cache_hit_tokens === "number" && usageLike.prompt_cache_hit_tokens > 0) return true;
+	if (typeof usageLike.cachedContentTokenCount === "number" && usageLike.cachedContentTokenCount > 0) return true;
 	const rawPromptTokenDetails = usageLike.prompt_tokens_details;
 	if (typeof rawPromptTokenDetails !== "object" || rawPromptTokenDetails === null) return false;
 	const promptTokenDetails = rawPromptTokenDetails;
@@ -261,6 +283,37 @@ function getTrailingPartialDeepseekToken(text) {
 }
 const OPENAI_COMPLETIONS_FIRST_EVENT_TIMEOUT_MESSAGE = "OpenAI completions stream timed out while waiting for the first event";
 const OPENAI_COMPLETIONS_POST_FINISH_GRACE_MS = 2500;
+const OPENAI_COMPLETIONS_ERROR_STATUS_BY_TYPE = {
+	SERVICE_UNAVAILABLE: 503,
+	TOO_MANY_REQUESTS: 429,
+	REQUEST_TIMEOUT: 408
+};
+function parseOpenAICompletionsErrorStatus(value) {
+	const status = typeof value === "number" ? value : typeof value === "string" && /^\d{3}$/.test(value.trim()) ? Number(value) : void 0;
+	return status !== void 0 && Number.isInteger(status) && status >= 400 && status <= 599 ? status : void 0;
+}
+function createOpenAICompletionsStreamError(chunk, provider) {
+	if (!chunk || typeof chunk !== "object") return void 0;
+	const error = Reflect.get(chunk, "error");
+	const flatMessage = Reflect.get(chunk, "message");
+	const structuredError = error !== null && typeof error === "object";
+	if (!structuredError && typeof error !== "string" && typeof flatMessage !== "string") return void 0;
+	const parsed = OpenAIHttpError.parseEnvelope(chunk, void 0);
+	const detail = parsed.detail ?? "Provider returned an in-band OpenAI completions stream error";
+	if (!structuredError) return new ProviderResponseError(detail, {
+		provider,
+		kind: "runtime"
+	});
+	const typeValue = Reflect.get(error, "type");
+	const codeValue = Reflect.get(error, "code");
+	const type = typeof typeValue === "string" ? typeValue.trim() : void 0;
+	const status = parseOpenAICompletionsErrorStatus(codeValue) ?? (type ? OPENAI_COMPLETIONS_ERROR_STATUS_BY_TYPE[type.toUpperCase()] : void 0);
+	if (status === void 0) return new ProviderResponseError(detail, {
+		provider,
+		kind: "runtime"
+	});
+	return new ProviderHttpError(`${status} ${detail}`, status, { code: parsed.code });
+}
 const streamOpenAICompletionsOnce = (model, context, options) => {
 	const stream = new AssistantMessageEventStream();
 	(async () => {
@@ -273,17 +326,21 @@ const streamOpenAICompletionsOnce = (model, context, options) => {
 		const firstEventTimeoutAbortError = new StreamTimeoutError(OPENAI_COMPLETIONS_FIRST_EVENT_TIMEOUT_MESSAGE);
 		const { requestAbortController, requestSignal } = abortTracker;
 		const onSseEvent = options?.onSseEvent;
-		const rawSseObserver = onSseEvent ? (event) => {
-			if (!event.event && event.data && event.data !== "[DONE]") try {
-				const parsed = JSON.parse(event.data);
-				const resolvedEvent = typeof parsed.type === "string" ? parsed.type : typeof parsed.object === "string" ? parsed.object : null;
-				if (resolvedEvent) {
-					event.event = resolvedEvent;
-					event.raw = [`event: ${resolvedEvent}`, ...event.raw];
-				}
-			} catch {}
-			onSseEvent(event, model);
-		} : void 0;
+		let sawDoneSentinel = false;
+		const rawSseObserver = (event) => {
+			if (event.data === "[DONE]") sawDoneSentinel = true;
+			if (onSseEvent) {
+				if (!event.event && event.data && event.data !== "[DONE]") try {
+					const parsed = JSON.parse(event.data);
+					const resolvedEvent = typeof parsed.type === "string" ? parsed.type : typeof parsed.object === "string" ? parsed.object : null;
+					if (resolvedEvent) {
+						event.event = resolvedEvent;
+						event.raw = [`event: ${resolvedEvent}`, ...event.raw];
+					}
+				} catch {}
+				onSseEvent(event, model);
+			}
+		};
 		let finishOpenBlocksOnError = () => {};
 		try {
 			const apiKey = options?.apiKey || getEnvApiKey(model.provider) || "";
@@ -291,7 +348,7 @@ const streamOpenAICompletionsOnce = (model, context, options) => {
 			const idleTimeoutMs = options?.streamIdleTimeoutMs ?? getOpenAIStreamIdleTimeoutMs(idleTimeoutFallbackMs);
 			const firstEventTimeoutMs = options?.streamFirstEventTimeoutMs ?? getOpenAIStreamFirstEventTimeoutMs(idleTimeoutMs, model.compat.streamFirstEventTimeoutMs);
 			const requestTimeoutMs = firstEventTimeoutMs !== void 0 && firstEventTimeoutMs > 0 ? firstEventTimeoutMs : void 0;
-			const { copilotPremiumRequests, baseUrl, headers, query, requestHeaders } = createRequestSetup(model, context, apiKey, options?.headers, options?.initiatorOverride, getOpenAIPromptCacheKey(options));
+			const { copilotPremiumRequests, baseUrl, headers, query, requestHeaders } = createRequestSetup(model, context, apiKey, options?.headers, options?.initiatorOverride, getOpenAIPromptCacheKey(options), options?.sessionId);
 			const premiumRequestsTotal = copilotPremiumRequests;
 			let appliedStrictTools = false;
 			const requestReasoningEffortFallbacks = /* @__PURE__ */ new Map();
@@ -304,14 +361,16 @@ const streamOpenAICompletionsOnce = (model, context, options) => {
 			const trimmedBaseUrl = baseUrl.replace(/\/+$/, "");
 			const completionsUrl = query ? `${trimmedBaseUrl}/chat/completions?${new URLSearchParams(query)}` : `${trimmedBaseUrl}/chat/completions`;
 			const createCompletionsStream = async (toolStrictModeOverride) => {
-				const { params, strictToolsApplied } = buildParams(model, context, options, disableStrictTools ? "none" : toolStrictModeOverride);
-				appliedStrictTools = strictToolsApplied;
+				const builtParams = buildParams(model, context, options, disableStrictTools ? "none" : toolStrictModeOverride);
+				appliedStrictTools = builtParams.strictToolsApplied;
+				let params = builtParams.params;
 				const reasoningEffortFallbackKey = createOpenAIReasoningEffortFallbackKey("chat-completions", trimmedBaseUrl, params.model);
 				const requestReasoningEffortFallback = requestReasoningEffortFallbacks.has(reasoningEffortFallbackKey) ? requestReasoningEffortFallbacks.get(reasoningEffortFallbackKey) : getOpenAIReasoningEffortFallback(providerSessionState, reasoningEffortFallbackKey);
 				if (requestReasoningEffortFallback !== void 0) applyOpenAIReasoningEffortFallback(params, requestReasoningEffortFallback);
 				activeReasoningEffortFallbackKey = reasoningEffortFallbackKey;
+				const replacedParams = await options?.onPayload?.(params, model);
+				if (replacedParams !== void 0) params = replacedParams;
 				activeRequestParams = params;
-				options?.onPayload?.(params, model);
 				rawRequestDump = {
 					provider: model.provider,
 					api: output.api,
@@ -331,32 +390,30 @@ const streamOpenAICompletionsOnce = (model, context, options) => {
 						headers: headersWithTimeout,
 						body: params,
 						signal: requestSignal,
-						fetch: options?.fetch,
+						fetch: wrapFetchForCopilotFallback(options?.fetch, model.provider === "github-copilot", resolveCopilotRequestIdentity(options?.headers)),
 						onSseEvent: rawSseObserver
 					});
 					await notifyProviderResponse(options, response, model, requestId);
 					return events;
 				} finally {
-					if (requestTimeout !== void 0) clearTimeout(requestTimeout);
+					clearTimeout(requestTimeout);
 				}
 			};
 			let openaiStream;
 			try {
-				openaiStream = await callWithCopilotModelRetry(() => createCompletionsStream(), {
-					provider: model.provider,
-					signal: requestSignal
-				});
+				openaiStream = await createCompletionsStream();
 			} catch (error) {
 				const capturedErrorResponse = error instanceof OpenAIHttpError ? error.captured : void 0;
-				const reasoningEffortFallback = activeReasoningEffortFallbackKey && activeRequestParams && !requestSignal.aborted ? resolveOpenAIReasoningEffortFallback(error, capturedErrorResponse, activeRequestParams, { explicitDisable: options?.disableReasoning === true && options.reasoning === void 0 }) : void 0;
+				const isExplicitDisable = options?.disableReasoning === true;
+				const reasoningEffortFallback = activeReasoningEffortFallbackKey && activeRequestParams && !requestSignal.aborted ? resolveOpenAIReasoningEffortFallback(error, capturedErrorResponse, activeRequestParams, { explicitDisable: isExplicitDisable }) : void 0;
 				if (reasoningEffortFallback !== void 0 && activeReasoningEffortFallbackKey) {
 					const retryMarker = `${activeReasoningEffortFallbackKey}:${String(reasoningEffortFallback)}`;
 					if (attemptedReasoningEffortFallbacks.has(retryMarker)) throw error;
 					attemptedReasoningEffortFallbacks.add(retryMarker);
 					requestReasoningEffortFallbacks.set(activeReasoningEffortFallbackKey, reasoningEffortFallback);
 					openaiStream = await createCompletionsStream();
-					rememberOpenAIReasoningEffortFallback(providerSessionState, activeReasoningEffortFallbackKey, reasoningEffortFallback);
-				} else if (isOpenRouterAnthropicModel(model) && !disableStrictTools && isCompiledGrammarTooLargeStrictError(error, capturedErrorResponse)) {
+					if (!isExplicitDisable) rememberOpenAIReasoningEffortFallback(providerSessionState, activeReasoningEffortFallbackKey, reasoningEffortFallback);
+				} else if (model.compat.retryWithoutStrictOnGrammarError && !disableStrictTools && isCompiledGrammarTooLargeStrictError(error, capturedErrorResponse)) {
 					disableStrictToolsForScope(providerSessionState, strictToolsScope);
 					disableStrictTools = true;
 					openaiStream = await createCompletionsStream("none");
@@ -418,7 +475,7 @@ const streamOpenAICompletionsOnce = (model, context, options) => {
 				});
 			};
 			const finishPendingToolCallBlocks = () => {
-				for (const block of [...pendingToolCallBlocks]) finishToolCallBlock(block);
+				for (const block of Array.from(pendingToolCallBlocks)) finishToolCallBlock(block);
 			};
 			const finishCurrentBlock = (block) => {
 				if (!block) return;
@@ -582,7 +639,7 @@ const streamOpenAICompletionsOnce = (model, context, options) => {
 			let sawUsagePayload = false;
 			let awaitTrailingUsageDetails = false;
 			const applyUsagePayload = (rawUsage) => {
-				output.usage = parseChunkUsage(rawUsage, model, premiumRequestsTotal);
+				output.usage = parseChunkUsage(rawUsage, model, premiumRequestsTotal, output.timestamp);
 				sawUsagePayload = true;
 				awaitTrailingUsageDetails = !hasPositiveCacheReadTokenField(rawUsage);
 			};
@@ -603,6 +660,8 @@ const streamOpenAICompletionsOnce = (model, context, options) => {
 			});
 			for await (const chunk of terminalAwareStream) {
 				if (!chunk || typeof chunk !== "object") continue;
+				const streamError = createOpenAICompletionsStreamError(chunk, model.provider);
+				if (streamError) throw streamError;
 				output.responseId ||= chunk.id;
 				if (!output.upstreamProvider) {
 					const upstreamProvider = chunk.provider;
@@ -698,6 +757,8 @@ const streamOpenAICompletionsOnce = (model, context, options) => {
 							}
 							if (toolCall.id) block.id = toolCall.id;
 							if (incomingName) block.name = incomingName;
+							const extraContent = getGeminiThoughtSignatureExtraContent(Reflect.get(toolCall, "extra_content"));
+							if (extraContent) block.thoughtSignature = JSON.stringify(extraContent);
 							let delta = "";
 							const rawArgs = toolCall.function?.arguments;
 							if (typeof rawArgs === "string") {
@@ -742,7 +803,7 @@ const streamOpenAICompletionsOnce = (model, context, options) => {
 				if (healedToolCallEmitted && output.stopReason === "stop") output.stopReason = "toolUse";
 			}
 			if (stripDeepseekChatTemplateTokens) flushDeepseekStripBuffer(true);
-			if (streamFinishedAt === void 0 && output.content.length > 0) throw new ProviderResponseError("OpenAI completions stream closed before a finish_reason was received", {
+			if (streamFinishedAt === void 0 && !sawDoneSentinel && output.content.length > 0) throw new ProviderResponseError("OpenAI completions stream closed before a finish_reason was received", {
 				provider: model.provider,
 				kind: "incomplete-stream"
 			});
@@ -804,13 +865,15 @@ const streamOpenAICompletionsOnce = (model, context, options) => {
 	return stream;
 };
 /**
-* Public entry: wrap the single-attempt streamer with bounded empty-completion
-* retries — flaky gateways occasionally 200 with `delta: {}` + `finish_reason:
-* "stop"` and no usage, which would otherwise stall the agent loop. Shared with
-* the Anthropic provider via `withEmptyCompletionRetry`.
+* Retries benign empty completions and transient provider failures only before
+* assistant output commits the attempt.
 */
-const streamOpenAICompletions = (model, context, options) => withEmptyCompletionRetry(model, context, options, streamOpenAICompletionsOnce);
-function createRequestSetup(model, context, apiKey, extraHeaders, initiatorOverride, promptCacheSessionId) {
+const streamOpenAICompletions = (model, context, options) => withReplaySafeStreamRetry(model, context, options, streamOpenAICompletionsOnce, {
+	retryEmptyCompletion: true,
+	retryProviderErrors: true,
+	maxProviderErrorRetries: 1
+});
+function createRequestSetup(model, context, apiKey, extraHeaders, initiatorOverride, promptCacheSessionId, sessionId) {
 	const apiVersion = $env.AZURE_OPENAI_API_VERSION || "2024-10-21";
 	const deploymentName = parseAzureDeploymentNameMap($env.AZURE_OPENAI_DEPLOYMENT_NAME_MAP).get(model.id) ?? model.id;
 	const setup = resolveOpenAIRequestSetup(model, {
@@ -818,9 +881,10 @@ function createRequestSetup(model, context, apiKey, extraHeaders, initiatorOverr
 		extraHeaders,
 		initiatorOverride,
 		promptCacheSessionId,
+		sessionId,
 		messages: context.messages,
 		defaultBaseUrl: "https://api.openai.com/v1",
-		prependHeaders: model.provider === "kimi-code" ? getKimiCommonHeaders : void 0,
+		prependHeaders: model.provider === "kimi-code" ? getKimiCommonHeaders : model.provider === "cline-pass" ? () => clinePassClientHeaders(promptCacheSessionId) : void 0,
 		alibabaCodingPlanAuth: true,
 		azureChatCompletions: {
 			apiVersion,
@@ -838,18 +902,12 @@ function resolveOpenAICompatForRequest(model, options) {
 		toolChoice: mapToOpenAICompletionsToolChoice(options?.toolChoice)
 	});
 }
-function dropOpenRouterKimiForcedToolReasoning(params, model, policy) {
-	if (policy.reasoning.disableReason === "forced-tool-choice" && policy.reasoning.disableMode === "openrouter-enabled-false" && policy.compat.isOpenRouterHost && isKimiModelId(model.id)) delete params.reasoning;
+function dropOpenRouterKimiForcedToolReasoning(params, _model, policy) {
+	if (policy.reasoning.disableReason === "forced-tool-choice" && policy.reasoning.disableMode === "openrouter-enabled-false" && policy.compat.isOpenRouterHost) delete params.reasoning;
 }
 function hasActiveNativeKimiK3Reasoning(model, options) {
-	if (model.provider !== "kimi-code" || model.id.toLowerCase() !== "k3" || !model.reasoning) return false;
-	if (options?.reasoning === void 0 || options.disableReasoning) return false;
-	try {
-		const url = new URL(model.baseUrl);
-		return url.hostname === "api.kimi.com" && (url.pathname === "/coding" || url.pathname.startsWith("/coding/"));
-	} catch {
-		return false;
-	}
+	if (!model.compat.nativeKimiK3Reasoning || !model.reasoning) return false;
+	return options?.reasoning !== void 0 && !options.disableReasoning;
 }
 function isChatCompletionsPromptCacheableContentBlock(block) {
 	if (typeof block !== "object" || block === null || !("type" in block)) return false;
@@ -890,7 +948,7 @@ function markLatestStableChatCompletionsCacheBreakpoint(messages) {
 }
 function applyOpenAIChatCompletionsPromptCachePolicy(params, model, options) {
 	const promptCacheKey = getOpenAIPromptCacheKey(options);
-	if (model.provider === "kimi-code" && promptCacheKey !== void 0) params.prompt_cache_key = promptCacheKey;
+	if (model.compat.supportsPromptCacheKey && promptCacheKey !== void 0) params.prompt_cache_key = promptCacheKey;
 	const promptCache = options?.promptCache;
 	if (!promptCache || resolveCacheRetention(options?.cacheRetention) === "none") return;
 	if (!model.compat.supportsPromptCacheBreakpoints) {
@@ -933,7 +991,7 @@ function buildParams(model, context, options, toolStrictModeOverride) {
 	}
 	applyOpenAIServiceTier(params, options?.serviceTier, model);
 	if (context.tools?.length) {
-		const builtTools = convertTools(context.tools, initialCompat, toolStrictModeOverride, model.provider);
+		const builtTools = convertTools(context.tools, initialCompat, toolStrictModeOverride);
 		params.tools = builtTools.tools;
 		toolStrictMode = builtTools.toolStrictMode;
 		strictToolsApplied = builtTools.strictToolsApplied;
@@ -948,6 +1006,7 @@ function buildParams(model, context, options, toolStrictModeOverride) {
 	if (isForcedToolChoice(params.tool_choice) && !initialCompat.supportsForcedToolChoice) params.tool_choice = "auto";
 	if ((!Array.isArray(params.tools) || params.tools.length === 0) && (params.tool_choice === "none" || isForcedToolChoice(params.tool_choice))) delete params.tool_choice;
 	if (forcedToolName !== void 0 && (!Array.isArray(params.tools) || !params.tools.some((tool) => tool.type === "function" && tool.function.name === forcedToolName))) delete params.tool_choice;
+	if (shouldDropAutoToolChoiceForReasoning(model, initialCompat, params.tool_choice, options)) delete params.tool_choice;
 	const finalPolicy = resolveOpenAICompatPolicy(model, {
 		endpoint: "chat-completions",
 		reasoning: options?.reasoning,
@@ -973,7 +1032,10 @@ function buildParams(model, context, options, toolStrictModeOverride) {
 		else if (outputToken.field === "max_completion_tokens") params.max_completion_tokens = outputToken.value;
 	}
 	applyChatCompletionsToolStream(params, model, compat);
-	applyChatCompletionsCompatPolicy(params, finalPolicy);
+	applyChatCompletionsReasoningParams(params, model, compat, {
+		...options,
+		toolChoice: params.tool_choice
+	});
 	dropOpenRouterKimiForcedToolReasoning(params, model, finalPolicy);
 	applyOpenAIGatewayRouting(params, compat, cacheRetention !== "none");
 	applyOpenAIExtraBody(params, compat.extraBody, { dropThinkingWhenReasoningEffort: compat.dropThinkingWhenReasoningEffort });
@@ -984,7 +1046,7 @@ function buildParams(model, context, options, toolStrictModeOverride) {
 		strictToolsApplied
 	};
 }
-function parseChunkUsage(rawUsage, model, premiumRequests) {
+function parseChunkUsage(rawUsage, model, premiumRequests, timestamp) {
 	const usageLike = rawUsage;
 	const rawPromptTokenDetails = usageLike.prompt_tokens_details;
 	const promptTokenDetails = typeof rawPromptTokenDetails === "object" && rawPromptTokenDetails !== null ? rawPromptTokenDetails : void 0;
@@ -996,13 +1058,14 @@ function parseChunkUsage(rawUsage, model, premiumRequests) {
 	const promptCacheHitTokens = usageLike.prompt_cache_hit_tokens;
 	const promptCacheMissTokens = usageLike.prompt_cache_miss_tokens;
 	const promptTokenCachedTokens = promptTokenDetails?.cached_tokens;
+	const cachedContentTokenCount = usageLike.cachedContentTokenCount;
 	const completionReasoningTokens = completionTokenDetails?.reasoning_tokens;
 	const cacheWriteTokens = promptTokenDetails?.cache_write_tokens;
 	const usage = {
 		...calculateOpenAIUsageAccounting({
 			promptTokens: typeof promptTokens === "number" ? promptTokens : 0,
 			outputTokens: typeof completionTokens === "number" ? completionTokens : 0,
-			cachedTokens: firstPositiveNumber(cachedTokens, promptCacheHitTokens, promptTokenCachedTokens),
+			cachedTokens: firstPositiveNumber(cachedTokens, promptCacheHitTokens, promptTokenCachedTokens, cachedContentTokenCount),
 			reasoningTokens: typeof completionReasoningTokens === "number" ? completionReasoningTokens : 0,
 			cacheWriteOpenRouter: typeof cacheWriteTokens === "number" ? cacheWriteTokens : void 0,
 			cacheWriteDeepSeek: typeof promptCacheMissTokens === "number" ? promptCacheMissTokens : void 0,
@@ -1017,8 +1080,8 @@ function parseChunkUsage(rawUsage, model, premiumRequests) {
 		},
 		...premiumRequests !== void 0 ? { premiumRequests } : {}
 	};
-	calculateCost(model, usage);
-	applyOpenRouterReportedCost(model, usage, rawUsage);
+	calculateCost(model, usage, timestamp);
+	applyProviderReportedCost(model, usage, rawUsage);
 	return usage;
 }
 function maybeAddAnthropicCacheControl(compat, messages) {
@@ -1115,7 +1178,7 @@ function convertMessages(model, context, compat) {
 					content: text
 				});
 			} else {
-				const supportsImages = model.input.includes("image") && !isDashscopeCompatibleModeTextOnlyQwen(model);
+				const supportsImages = isOpenAICompletionsVisionSupported(model);
 				const content = [];
 				let omittedImages = false;
 				for (const item of msg.content) if (item.type === "text") {
@@ -1128,7 +1191,7 @@ function convertMessages(model, context, compat) {
 				} else if (supportsImages) content.push({
 					type: "image_url",
 					image_url: {
-						url: `data:${item.mimeType};base64,${item.data}`,
+						url: item.url ?? `data:${item.mimeType};base64,${item.data}`,
 						...item.detail && item.detail !== "original" ? { detail: item.detail } : {}
 					}
 				});
@@ -1211,7 +1274,7 @@ function convertMessages(model, context, compat) {
 				assistantMsg.tool_calls = toolCalls.map((tc, toolCallIndex) => {
 					const toolCallId = ensureToolCallId(tc.id, `${i}:${toolCallIndex}:${tc.name}`, msg);
 					rememberToolCallId(tc.id, toolCallId);
-					return {
+					const replayedToolCall = {
 						id: normalizeMistralToolId(toolCallId, compat.requiresMistralToolIds),
 						type: "function",
 						function: {
@@ -1219,14 +1282,20 @@ function convertMessages(model, context, compat) {
 							arguments: serializeToolArguments(tc.arguments)
 						}
 					};
+					const extraContent = parseGeminiThoughtSignatureExtraContent(tc.thoughtSignature);
+					if (extraContent) replayedToolCall.extra_content = extraContent;
+					return replayedToolCall;
 				});
-				const reasoningDetails = toolCalls.filter((tc) => tc.thoughtSignature).map((tc) => {
+				const reasoningDetails = toolCalls.flatMap((tc) => {
+					const thoughtSignature = tc.thoughtSignature;
+					if (!thoughtSignature) return [];
 					try {
-						return JSON.parse(tc.thoughtSignature);
+						const parsed = JSON.parse(thoughtSignature);
+						return getGeminiThoughtSignatureExtraContent(parsed) ? [] : [parsed];
 					} catch {
-						return null;
+						return [];
 					}
-				}).filter(Boolean);
+				});
 				if (reasoningDetails.length > 0) assistantMsg.reasoning_details = reasoningDetails;
 			}
 			if (assistantMsg.content === null && (hasReasoningField || assistantMsg.tool_calls)) assistantMsg.content = "";
@@ -1241,7 +1310,7 @@ function convertMessages(model, context, compat) {
 			for (; j < transformedMessages.length && transformedMessages[j].role === "toolResult"; j++) {
 				const toolMsg = transformedMessages[j];
 				const textResult = toolMsg.content.filter((c) => c.type === "text").map((c) => c.text).join("\n");
-				const supportsImages = model.input.includes("image") && !isDashscopeCompatibleModeTextOnlyQwen(model);
+				const supportsImages = isOpenAICompletionsVisionSupported(model);
 				const hasImages = toolMsg.content.some((c) => c.type === "image");
 				const omittedImages = hasImages && !supportsImages;
 				const hasText = textResult.length > 0;
@@ -1256,7 +1325,7 @@ function convertMessages(model, context, compat) {
 				if (hasImages && supportsImages) {
 					for (const block of toolMsg.content) if (block.type === "image") imageBlocks.push({
 						type: "image_url",
-						image_url: { url: `data:${block.mimeType};base64,${block.data}` }
+						image_url: { url: block.url ?? `data:${block.mimeType};base64,${block.data}` }
 					});
 				}
 			}
@@ -1281,11 +1350,11 @@ function convertMessages(model, context, compat) {
 	}
 	return params;
 }
-function convertTools(tools, compat, toolStrictModeOverride, provider) {
-	const rejectXaiRootObjectUnion = provider === "xai" || provider === "xai-oauth";
+function convertTools(tools, compat, toolStrictModeOverride) {
+	const rejectRootObjectUnion = compat.rejectRootObjectUnion;
 	const adaptedTools = tools.map((tool) => {
 		const strict = !NO_STRICT && compat.supportsStrictMode !== false && tool.strict !== false;
-		const baseParameters = rejectXaiRootObjectUnion ? flattenExclusiveRequiredRootUnion(toolWireSchema(tool)) : toolWireSchema(tool);
+		const baseParameters = rejectRootObjectUnion ? flattenExclusiveRequiredRootUnion(toolWireSchema(tool)) : toolWireSchema(tool);
 		const adapted = adaptSchemaForStrict(baseParameters, strict);
 		return {
 			tool,
@@ -1303,7 +1372,7 @@ function convertTools(tools, compat, toolStrictModeOverride, provider) {
 		const includeExplicitFalse = !includeStrict && tool.strict === false && toolStrictMode === "mixed" && compat.supportsStrictMode !== false;
 		const wireParameters = includeStrict ? parameters : baseParameters;
 		const emittedParameters = compat.toolSchemaFlavor === "moonshot-mfjs" ? normalizeSchemaForMoonshot(wireParameters) : compat.toolSchemaFlavor === "grammar" ? sanitizeSchemaForGrammar(wireParameters) : wireParameters;
-		const violation = findStrictToolSchemaViolation(emittedParameters, "#", { rejectXaiRootObjectUnion });
+		const violation = findStrictToolSchemaViolation(emittedParameters, "#", { rejectRootObjectUnion });
 		if (violation) {
 			warn(`Tool "${tool.name}" omitted from the openai-completions request: its parameter schema is invalid for this provider at ${violation} (an enum/const value cannot match its declared type, or leftover xAI object-root union). Other tools are unaffected.`);
 			continue;
@@ -1328,10 +1397,11 @@ function convertTools(tools, compat, toolStrictModeOverride, provider) {
 const EMPTY_OLLAMA_LENGTH_COMPLETION_MESSAGE = "Model returned no content: prompt filled the context window; raise Ollama num_ctx or shorten the prompt.";
 function mapStopReason(reason) {
 	if (reason === null) return { stopReason: "stop" };
-	switch (reason) {
+	switch (typeof reason === "string" ? reason.toLowerCase() : reason) {
 		case "stop":
 		case "end": return { stopReason: "stop" };
-		case "length": return { stopReason: "length" };
+		case "length":
+		case "max_tokens": return { stopReason: "length" };
 		case "function_call":
 		case "tool_calls": return { stopReason: "toolUse" };
 		case "content_filter": return {

package/lib/openai-configuration-update-xMYUjaK_.mjs

@@ -0,0 +1,150 @@
+import { installBunShim as __installBunShim } from "./bun-shim.mjs";
+__installBunShim();
+//#region node_modules/.pnpm/@oh-my-pi+pi-ai@18.1.17/node_modules/@oh-my-pi/pi-ai/src/providers/grammar.ts
+function compactGrammarDefinition(syntax, definition) {
+	if (syntax !== "lark") return definition;
+	return compactLarkGrammarDefinition(definition);
+}
+function compactLarkGrammarDefinition(definition) {
+	const lines = [];
+	for (const line of definition.split(/\r?\n/)) {
+		const uncommented = stripLarkLineComment(line).trimEnd();
+		if (uncommented.trim()) lines.push(uncommented);
+	}
+	return lines.join("\n");
+}
+function stripLarkLineComment(line) {
+	let inString;
+	let inRegex = false;
+	let escaped = false;
+	for (let i = 0; i < line.length; i++) {
+		const char = line[i];
+		const next = line[i + 1];
+		if (escaped) {
+			escaped = false;
+			continue;
+		}
+		if (char === "\\") {
+			escaped = true;
+			continue;
+		}
+		if (inString) {
+			if (char === inString) inString = void 0;
+			continue;
+		}
+		if (inRegex) {
+			if (char === "/") inRegex = false;
+			continue;
+		}
+		if (char === "/" && next === "/") return line.slice(0, i);
+		if (char === "\"" || char === "'") {
+			inString = char;
+			continue;
+		}
+		if (char === "/") inRegex = true;
+	}
+	return line;
+}
+//#endregion
+//#region node_modules/.pnpm/@oh-my-pi+pi-ai@18.1.17/node_modules/@oh-my-pi/pi-ai/src/providers/openai-configuration-update.ts
+function createOpenAIEffortControlState() {
+	return { transitions: [] };
+}
+const MAX_EFFORT_CONTROL_STATES = 16;
+/**
+* Fetch (or create) the control state for one conversation from a provider's
+* bounded per-session map, refreshing its LRU slot.
+*/
+function getOpenAIEffortControlState(states, key) {
+	const existing = states.get(key);
+	if (existing) {
+		states.delete(key);
+		states.set(key, existing);
+		return existing;
+	}
+	const created = createOpenAIEffortControlState();
+	states.set(key, created);
+	if (states.size > MAX_EFFORT_CONTROL_STATES) {
+		const oldest = states.keys().next().value;
+		if (oldest !== void 0) states.delete(oldest);
+	}
+	return created;
+}
+/**
+* Fingerprint of the item a transition sits after. Output-only lifecycle
+* fields are excluded: a live response item carries `id`/`status` that the
+* sanitized replay of the same item drops.
+*/
+function effortControlAnchor(input, index) {
+	if (index === 0) return "";
+	const item = input[index - 1];
+	if (!item) return "";
+	const { id: _id, status: _status, ...stable } = item;
+	return String(Bun.hash(JSON.stringify(stable)));
+}
+function resetOpenAIEffortControlState(state) {
+	state.baseEffort = void 0;
+	state.currentEffort = void 0;
+	state.transitions = [];
+}
+/**
+* Discard the baseline when the request no longer continues the conversation
+* it was captured for: a wire history that shrank or was rewritten under a
+* recorded transition (compaction, branch switch, `/clear`). The next request
+* re-baselines from its own effort, which is what the API asks for after
+* compaction anyway.
+*/
+function syncOpenAIEffortControlState(state, input) {
+	for (const transition of state.transitions) if (transition.index > input.length || transition.anchor !== effortControlAnchor(input, transition.index)) {
+		resetOpenAIEffortControlState(state);
+		return;
+	}
+}
+/**
+* Pin the request-level effort to the session baseline and splice pending
+* `configuration_update` items into `input` (mutated in place).
+*
+* `input` is the freshly built transcript for this request, without any
+* `configuration_update` items. `requested` is the wire effort the caller
+* would otherwise send at the request level. Returns the effort to send at the
+* request level (`requested` on the first request, the baseline afterwards).
+*/
+function planStableOpenAIEffort(state, input, requested) {
+	syncOpenAIEffortControlState(state, input);
+	if (state.baseEffort === void 0) {
+		state.baseEffort = requested;
+		state.currentEffort = requested;
+		return requested;
+	}
+	if (state.currentEffort !== requested) {
+		const last = input[input.length - 1];
+		const index = last && "role" in last && last.role === "user" ? input.length - 1 : input.length;
+		const existing = state.transitions.find((transition) => transition.index === index);
+		if (existing) existing.effort = requested;
+		else state.transitions.push({
+			index,
+			anchor: effortControlAnchor(input, index),
+			effort: requested
+		});
+		let preceding = state.baseEffort;
+		let precedingIndex = -1;
+		for (const transition of state.transitions) if (transition.index < index && transition.index > precedingIndex) {
+			preceding = transition.effort;
+			precedingIndex = transition.index;
+		}
+		if (requested === preceding) state.transitions = state.transitions.filter((transition) => transition.index !== index);
+		state.currentEffort = requested;
+	}
+	state.transitions.sort((a, b) => a.index - b.index);
+	let offset = 0;
+	for (const transition of state.transitions) {
+		input.splice(transition.index + offset, 0, {
+			type: "configuration_update",
+			reasoning: { effort: transition.effort }
+		});
+		offset++;
+	}
+	return state.baseEffort;
+}
+//#endregion
+export { planStableOpenAIEffort as n, compactGrammarDefinition as r, getOpenAIEffortControlState as t };