copilot-relay 0.4.0 → 0.4.1

package/dist/main.js

@@ -2299,6 +2299,22 @@ async function boundModelOutputTokens(config, model, requested) {
 
 //#endregion
 //#region src/claude/utils.ts
+const isEffortOnlyControl = (control) => control !== null && typeof control === "object" && !Array.isArray(control) && Object.keys(control).length === 1 && "effort" in control && isConfiguredReasoningEffort(control.effort);
+function getClaudeTurnEffort(payload) {
+	let requested = getRequestReasoningEffort(payload);
+	let pending;
+	if (Array.isArray(payload.messages)) {
+		for (const message of payload.messages) if (message?.role === "system" && message.clear_at === void 0 && isEffortOnlyControl(message.output_config)) pending = message.output_config.effort;
+		else if (message?.role === "user" && pending !== void 0) {
+			requested = pending;
+			pending = void 0;
+		}
+	}
+	return {
+		requested,
+		effective: resolveReasoningEffort(requested)
+	};
+}
 function mapOpenAIStopReasonToClaude(finishReason) {
 	if (finishReason === null) return null;
 	return {
@@ -2379,8 +2395,8 @@ function translateToOpenAI(payload, _settings, toolNameMapper) {
 	const model = translateModelName(payload.model);
 	const mapper = toolNameMapper ?? createClaudeToolNameMapper(payload.tools, { ...getToolNameMapperOptionsForModel(model) });
 	const tools = translateClaudeToolsToOpenAI(payload.tools, mapper);
-	const effort = resolveReasoningEffort(getRequestReasoningEffort(payload));
-	validateClaudeMessages(payload.messages, false, effort);
+	const { effective: effort } = getClaudeTurnEffort(payload);
+	validateClaudeMessages(payload.messages);
 	return {
 		model,
 		messages: normalizeFinalAssistantPrefill$1(translateClaudeMessagesToOpenAI(payload.messages, payload.system, mapper)),
@@ -2395,7 +2411,7 @@ function translateToOpenAI(payload, _settings, toolNameMapper) {
 		tool_choice: tools && tools.length > 0 ? translateClaudeToolChoiceToOpenAI(payload.tool_choice, mapper) : void 0
 	};
 }
-function validateClaudeMessages(messages, native = false, effort) {
+function validateClaudeMessages(messages, native = false) {
 	if (!Array.isArray(messages)) throw invalidMessage("Messages must be an array.");
 	for (const message of messages) {
 		if (!message || ![
@@ -2405,8 +2421,7 @@ function validateClaudeMessages(messages, native = false, effort) {
 		].includes(message.role)) throw invalidMessage("Unsupported message role.");
 		if (message.role !== "system") continue;
 		const control = message.output_config;
-		const redundantEffort = control !== null && typeof control === "object" && !Array.isArray(control) && Object.keys(control).length === 1 && isConfiguredReasoningEffort(control.effort) && control.effort === effort;
-		if (!native && (message.clear_at !== void 0 || control !== void 0 && !redundantEffort)) throw invalidMessage("Per-message system controls require the native Messages route unless effort exactly matches the request.");
+		if (!native && (message.clear_at !== void 0 || control !== void 0 && !isEffortOnlyControl(control))) throw invalidMessage("Translated system controls support only output_config.effort with low, medium, high, xhigh, or max; other controls require the native Messages route.");
 		if (typeof message.content !== "string" && (!Array.isArray(message.content) || message.content.some((block) => !block || block.type !== "text" || typeof block.text !== "string"))) throw invalidMessage("System message content must contain only text.");
 	}
 }
@@ -2416,7 +2431,9 @@ function translateClaudeMessagesToOpenAI(claudeMessages, system, toolNameMapper)
 		switch (message.role) {
 			case "user": return handleUserMessage(message);
 			case "assistant": return handleAssistantMessage(message, toolNameMapper);
-			case "system": return handleSystemPrompt(message.content);
+			case "system":
+				if (message.output_config !== void 0 && (typeof message.content === "string" ? message.content.length === 0 : message.content.every((block) => block.text.length === 0))) return [];
+				return handleSystemPrompt(message.content);
 			default: throw invalidMessage("Unsupported message role.");
 		}
 	});
@@ -2854,7 +2871,11 @@ const getClaudeWebSearchToolCallFromChatResponse = (response, toolNameMapper) =>
 };
 const buildSearchInput = (payload, requestedQuery) => {
 	const systemText = typeof payload.system === "string" ? payload.system : Array.isArray(payload.system) ? payload.system.map((block) => block.text).join("\n\n") : "";
-	const messages = payload.messages.map((message) => `${message.role}: ${textFromMessageContent(message.content)}`).join("\n\n");
+	const messages = payload.messages.flatMap((message) => {
+		const text$1 = textFromMessageContent(message.content);
+		if (message.role === "system" && message.output_config !== void 0 && (typeof message.content === "string" ? message.content.length === 0 : message.content.every((block) => block.text.length === 0))) return [];
+		return [`${message.role}: ${text$1}`];
+	}).join("\n\n");
 	return [
 		"You are fulfilling an Anthropic web_search server tool request for Claude Code.",
 		"Search the web using the provided web_search_preview tool.",
@@ -2870,7 +2891,7 @@ const buildWebSearchRequestPayload = (payload, requestedQuery, model) => ({
 	model,
 	input: buildSearchInput(payload, requestedQuery),
 	tools: [{ type: "web_search_preview" }],
-	reasoning: { effort: resolveReasoningEffort(getRequestReasoningEffort(payload)) },
+	reasoning: { effort: getClaudeTurnEffort(payload).effective },
 	max_output_tokens: Math.max(256, Math.min(payload.max_tokens ?? 1024, 1200)),
 	temperature: payload.temperature,
 	top_p: payload.top_p
@@ -3096,7 +3117,7 @@ const createClaudeWebSearchExecution = async (config, payload, requestedQuery, o
 	} catch (error) {
 		if (!(error instanceof SyntaxError)) throw error;
 	}
-	return interpretSearchResponse(upstream, config, requestedQuery, request, getRequestReasoningEffort(payload) ?? "unset", options.requestId);
+	return interpretSearchResponse(upstream, config, requestedQuery, request, getClaudeTurnEffort(payload).requested ?? "unset", options.requestId);
 };
 const buildSearchResultBlock = (toolUseId, results) => ({
 	type: "web_search_tool_result",
@@ -4346,7 +4367,7 @@ async function createNativeMessages(config, payload, options) {
 	};
 	const beta = options.headers?.get("anthropic-beta");
 	if (beta) headers["anthropic-beta"] = beta;
-	log.info(sanitizeTerminalString(`request_id=${options.requestId} Model request client=claude requested_model=${payload.model} upstream_model=${payload.model} upstream_api=messages effective_think_effort=${body.output_config.effort}`));
+	log.info(sanitizeTerminalString(`request_id=${options.requestId} Model request client=claude requested_model=${payload.model} upstream_model=${payload.model} upstream_api=messages effective_think_effort=${getClaudeTurnEffort(body).effective}`));
 	const response = await fetchCopilot(getCopilotProviderContext(config), "/v1/messages", {
 		method: "POST",
 		headers,
@@ -4560,7 +4581,10 @@ async function handleNativeMessages(config, payload, options, write) {
 	}
 	const searchCall = searches[0];
 	if (!searchCall.id || !searchCall.name || typeof searchCall.input?.query !== "string" || !searchCall.input.query.trim()) throw new Error("Invalid native search call.");
-	const search = await createClaudeWebSearchExecution(config, payload, searchCall.input.query, {
+	const search = await createClaudeWebSearchExecution(config, {
+		...payload,
+		messages: normalizeNativeHistory(payload.messages)
+	}, searchCall.input.query, {
 		...options,
 		signal
 	});
@@ -4672,7 +4696,7 @@ const createTokenCountModel = (modelId, tokenizer = "o200k_base") => ({
 	id: modelId
 });
 const isNonStreamingResponse = (response) => typeof response === "object" && response !== null && Object.hasOwn(response, "choices");
-const getClaudeRequestedThinkEffort = (payload) => getRequestReasoningEffort(payload) ?? "unset";
+const getClaudeRequestedThinkEffort = (payload) => getClaudeTurnEffort(payload).requested ?? "unset";
 const getClaudeRequestedThinking = (payload) => {
 	if (!payload.thinking) return "none";
 	return [`type:${payload.thinking.type}`, `budget:${payload.thinking.budget_tokens ?? "none"}`].join(",");
@@ -4877,10 +4901,7 @@ const handleClaudeMessageRequest = async (config, claudePayload, requestSignal,
 		const webSearchToolCall = shouldLetModelDecideWebSearch ? getClaudeWebSearchToolCallFromChatResponse(effectiveResponse, toolNameMapper) : void 0;
 		let claudeResponse;
 		if (webSearchToolCall) {
-			const search = await createClaudeWebSearchExecution(config, {
-				...claudePayload,
-				reasoning_effort: openAIPayload.reasoning_effort
-			}, webSearchToolCall.query, {
+			const search = await createClaudeWebSearchExecution(config, claudePayload, webSearchToolCall.query, {
 				requestId,
 				signal: requestSignal,
 				timeoutMs: config.upstreamTimeoutMs
@@ -4995,11 +5016,11 @@ claudeRoutes.post("/messages", async (c) => {
 	const claudePayload = await c.req.json();
 	const requestSignal = createCopilotRequestSignal(c.req.raw.signal, config.upstreamTimeoutMs);
 	try {
-		const effort = resolveReasoningEffort(getRequestReasoningEffort(claudePayload));
+		getRequestReasoningEffort(claudePayload);
 		const upstreamModel = translateModelName(claudePayload.model);
 		const validate = () => {
 			if (shouldUseNativeMessages(config, upstreamModel)) validateNativeMessages(claudePayload);
-			else validateClaudeMessages(claudePayload.messages, false, effort);
+			else validateClaudeMessages(claudePayload.messages);
 		};
 		validateClaudeMessages(claudePayload.messages, true);
 		if (config.claudeUpstreamApi !== "auto") validate();