copilot-api-plus 1.0.51 → 1.0.53

package/README.md

@@ -42,7 +42,7 @@
 | 🌐 **代理支持** | 支持 HTTP/HTTPS 代理，配置持久化 |
 | 🐳 **Docker 支持** | 提供完整的 Docker 部署方案 |
 | 🔑 **API Key 认证** | 可选的 API Key 鉴权，保护公开部署的服务 |
-| ✂️ **智能上下文压缩** | Prompt 超过模型 Token 限制时自动截断，保留系统消息和最近对话 |
+| ✂️ **上下文透传** | 全量透传上下文至上游 API，由客户端（如 Claude Code）自行管理压缩 |
 | 🔍 **智能模型匹配** | 自动处理模型名格式差异（日期后缀、dash/dot 版本号等） |
 | 🔁 **Antigravity 端点容错** | 双端点自动切换，按模型族追踪速率限制，指数退避重试 |
 
@@ -648,14 +648,12 @@ curl http://localhost:4141/v1/messages \
 
 ## 🔧 技术细节
 
-### 智能上下文压缩
+### 上下文管理
 
-当 Prompt Token 数量超过模型的上下文窗口限制时，代理层会自动截断消息以避免上游 API 返回 400 错误：
+代理层不做上下文截断，全量透传消息至上游 API。上下文压缩由客户端负责：
 
-- **保留系统/开发者消息**：system 和 developer 角色的消息始终保留
-- **保留最近对话**：优先丢弃最早的消息，保留最近的上下文
-- **工具调用分组**：assistant 的 tool_calls 和对应的 tool result 消息作为一组，不会被拆散
-- **5% 安全余量**：实际限制为模型上下文窗口的 95%，避免边界情况
+- **Claude Code**：通过 `/count_tokens` 端点获取当前 token 数，接近上限时自动触发 `/compact` 压缩
+- **其他客户端**：如果上游 API 返回 400（token 超限），客户端自行处理重试
 
 ### 智能模型名匹配
 
@@ -690,12 +688,6 @@ Google Antigravity 模式内置了可靠性保障：
 - `out` — 输出 token 数
 - `cache_read` — 缓存命中的 token 数（仅在有缓存时显示）
 
-触发上下文压缩时会额外输出一行：
-
-```
-Truncated: 190385 -> 117537 tokens (-59 msgs)
-```
-
 ### 网络重试
 
 对上游 API 的请求内置了瞬时网络错误重试（TLS 断开、连接重置等）：

package/dist/main.js

@@ -2608,6 +2608,194 @@ const awaitApproval = async () => {
 	if (!await consola.prompt(`Accept incoming request?`, { type: "confirm" })) throw new HTTPError("Request rejected", Response.json({ message: "Request rejected" }, { status: 403 }));
 };
 
+//#endregion
+//#region src/lib/rate-limit.ts
+async function checkRateLimit(state$1) {
+	if (state$1.rateLimitSeconds === void 0) return;
+	const now = Date.now();
+	if (!state$1.lastRequestTimestamp) {
+		state$1.lastRequestTimestamp = now;
+		return;
+	}
+	const elapsedSeconds = (now - state$1.lastRequestTimestamp) / 1e3;
+	if (elapsedSeconds > state$1.rateLimitSeconds) {
+		state$1.lastRequestTimestamp = now;
+		return;
+	}
+	const waitTimeSeconds = Math.ceil(state$1.rateLimitSeconds - elapsedSeconds);
+	if (!state$1.rateLimitWait) {
+		consola.warn(`Rate limit exceeded. Need to wait ${waitTimeSeconds} more seconds.`);
+		throw new HTTPError("Rate limit exceeded", Response.json({ message: "Rate limit exceeded" }, { status: 429 }));
+	}
+	const waitTimeMs = waitTimeSeconds * 1e3;
+	consola.warn(`Rate limit reached. Waiting ${waitTimeSeconds} seconds before proceeding...`);
+	await sleep(waitTimeMs);
+	state$1.lastRequestTimestamp = now;
+	consola.info("Rate limit wait completed, proceeding with request");
+}
+
+//#endregion
+//#region src/services/copilot/create-chat-completions.ts
+const createChatCompletions = async (payload) => {
+	if (!state.copilotToken) throw new Error("Copilot token not found");
+	const enableVision = payload.messages.some((x) => typeof x.content !== "string" && x.content?.some((x$1) => x$1.type === "image_url"));
+	const isAgentCall = payload.messages.some((msg) => ["assistant", "tool"].includes(msg.role));
+	const buildHeaders = () => ({
+		...copilotHeaders(state, enableVision),
+		"X-Initiator": isAgentCall ? "agent" : "user"
+	});
+	consola.debug("Sending request to Copilot:", {
+		model: payload.model,
+		endpoint: `${copilotBaseUrl(state)}/chat/completions`
+	});
+	const url = `${copilotBaseUrl(state)}/chat/completions`;
+	const body = payload.stream ? {
+		...payload,
+		stream_options: { include_usage: true }
+	} : payload;
+	const bodyString = JSON.stringify(body);
+	const maxRetries = 2;
+	let lastError;
+	let response;
+	for (let attempt = 0; attempt <= maxRetries; attempt++) try {
+		response = await fetch(url, {
+			method: "POST",
+			headers: buildHeaders(),
+			body: bodyString
+		});
+		break;
+	} catch (error) {
+		lastError = error;
+		if (attempt < maxRetries) {
+			const delay = 1e3 * (attempt + 1);
+			consola.warn(`Network error on attempt ${attempt + 1}/${maxRetries + 1}, retrying in ${delay}ms:`, error instanceof Error ? error.message : error);
+			await new Promise((r) => setTimeout(r, delay));
+		}
+	}
+	if (!response) throw lastError;
+	if (response.status === 401) {
+		consola.warn("Copilot token expired, refreshing and retrying...");
+		try {
+			await refreshCopilotToken();
+			response = await fetch(url, {
+				method: "POST",
+				headers: buildHeaders(),
+				body: bodyString
+			});
+		} catch (refreshError) {
+			consola.error("Failed to refresh token:", refreshError);
+		}
+	}
+	if (!response.ok) {
+		const errorBody = await response.text();
+		consola.error("Failed to create chat completions", {
+			status: response.status,
+			statusText: response.statusText,
+			body: errorBody
+		});
+		throw new HTTPError(`Failed to create chat completions: ${response.status} ${errorBody}`, response);
+	}
+	if (payload.stream) return events(response);
+	return await response.json();
+};
+
+//#endregion
+//#region src/routes/chat-completions/handler.ts
+/**
+* Set max_tokens from model limits if not already provided in the payload.
+*/
+function applyMaxTokens(payload) {
+	if (!isNullish(payload.max_tokens)) return payload;
+	const selectedModel = findModel(payload.model);
+	if (!selectedModel) return payload;
+	const maxTokens = selectedModel.capabilities.limits.max_output_tokens;
+	if (maxTokens) {
+		consola.debug("Set max_tokens to:", maxTokens);
+		return {
+			...payload,
+			max_tokens: maxTokens
+		};
+	}
+	return payload;
+}
+async function handleCompletion$1(c) {
+	await checkRateLimit(state);
+	const rawPayload = await c.req.json();
+	consola.debug("Request payload:", JSON.stringify(rawPayload).slice(-400));
+	const payload = applyMaxTokens(rawPayload);
+	if (state.manualApprove) await awaitApproval();
+	const response = await createChatCompletions(payload);
+	if (isNonStreaming$1(response)) {
+		consola.debug("Non-streaming response:", JSON.stringify(response));
+		if (response.usage) setTokenUsage({
+			inputTokens: response.usage.prompt_tokens,
+			outputTokens: response.usage.completion_tokens,
+			cacheReadTokens: response.usage.prompt_tokens_details?.cached_tokens
+		});
+		return c.json(response);
+	}
+	consola.debug("Streaming response");
+	return streamSSE(c, async (stream) => {
+		for await (const chunk of response) {
+			consola.debug("Streaming chunk:", JSON.stringify(chunk));
+			try {
+				const sseChunk = chunk;
+				if (sseChunk.data && sseChunk.data !== "[DONE]") {
+					const parsed = JSON.parse(sseChunk.data);
+					if (parsed.usage) {
+						const usage = {
+							inputTokens: parsed.usage.prompt_tokens ?? 0,
+							outputTokens: parsed.usage.completion_tokens ?? 0,
+							cacheReadTokens: parsed.usage.prompt_tokens_details?.cached_tokens
+						};
+						setTokenUsage(usage);
+					}
+				}
+			} catch {}
+			await stream.writeSSE(chunk);
+		}
+		signalStreamDone();
+	});
+}
+const isNonStreaming$1 = (response) => Object.hasOwn(response, "choices");
+
+//#endregion
+//#region src/routes/chat-completions/route.ts
+const completionRoutes = new Hono();
+completionRoutes.post("/", async (c) => {
+	try {
+		return await handleCompletion$1(c);
+	} catch (error) {
+		return await forwardError(c, error);
+	}
+});
+
+//#endregion
+//#region src/services/copilot/create-embeddings.ts
+const createEmbeddings = async (payload) => {
+	if (!state.copilotToken) throw new Error("Copilot token not found");
+	const response = await fetch(`${copilotBaseUrl(state)}/embeddings`, {
+		method: "POST",
+		headers: copilotHeaders(state),
+		body: JSON.stringify(payload)
+	});
+	if (!response.ok) throw new HTTPError("Failed to create embeddings", response);
+	return await response.json();
+};
+
+//#endregion
+//#region src/routes/embeddings/route.ts
+const embeddingRoutes = new Hono();
+embeddingRoutes.post("/", async (c) => {
+	try {
+		const paylod = await c.req.json();
+		const response = await createEmbeddings(paylod);
+		return c.json(response);
+	} catch (error) {
+		return await forwardError(c, error);
+	}
+});
+
 //#endregion
 //#region src/lib/tokenizer.ts
 const ENCODING_MAP = {
@@ -2801,337 +2989,6 @@ const getTokenCount = async (payload, model) => {
 	};
 };
 
-//#endregion
-//#region src/lib/context-compression.ts
-/**
-* Get the maximum prompt token limit for a model.
-* Prefers max_prompt_tokens, falls back to max_context_window_tokens minus max_output_tokens.
-*/
-const getMaxPromptTokens = (model) => {
-	const limits = model.capabilities.limits;
-	if (limits.max_prompt_tokens) return limits.max_prompt_tokens;
-	if (limits.max_context_window_tokens) {
-		const outputReserve = limits.max_output_tokens ?? 4096;
-		return limits.max_context_window_tokens - outputReserve;
-	}
-};
-/**
-* Check if a message is a tool-related message (tool call or tool result).
-* Tool messages must be kept together with their paired assistant message.
-*/
-const isToolMessage = (message) => {
-	return message.role === "tool";
-};
-/**
-* Check if an assistant message contains tool calls.
-*/
-const hasToolCalls = (message) => {
-	return message.role === "assistant" && Array.isArray(message.tool_calls) && message.tool_calls.length > 0;
-};
-const groupMessages = (messages) => {
-	const groups = [];
-	let i = 0;
-	while (i < messages.length) {
-		const message = messages[i];
-		if (message.role === "system" || message.role === "developer") {
-			groups.push({
-				messages: [message],
-				isSystem: true,
-				isRecent: false
-			});
-			i++;
-			continue;
-		}
-		if (hasToolCalls(message)) {
-			const group = [message];
-			let j = i + 1;
-			while (j < messages.length && isToolMessage(messages[j])) {
-				group.push(messages[j]);
-				j++;
-			}
-			groups.push({
-				messages: group,
-				isSystem: false,
-				isRecent: false
-			});
-			i = j;
-			continue;
-		}
-		groups.push({
-			messages: [message],
-			isSystem: false,
-			isRecent: false
-		});
-		i++;
-	}
-	return groups;
-};
-/**
-* Create a truncation notice message to inform the model that earlier context was removed.
-*/
-const createTruncationNotice = () => ({
-	role: "user",
-	content: "[Note: Earlier conversation history was automatically truncated to fit within the model's context window. The most recent messages have been preserved.]"
-});
-/**
-* Intelligently truncate messages to fit within the model's token limit.
-*
-* Strategy:
-* 1. Always preserve system/developer messages (they contain critical instructions)
-* 2. Always preserve the most recent messages (they contain the current task context)
-* 3. Remove middle conversation messages, oldest first
-* 4. Insert a truncation notice where messages were removed
-* 5. Keep tool call/result pairs together (never split them)
-*
-* Safety margin: keeps 5% below the limit to account for token counting inaccuracies.
-*/
-const truncateMessages = async (payload, model) => {
-	const maxPromptTokens = getMaxPromptTokens(model);
-	if (!maxPromptTokens) {
-		consola.debug("No token limit found for model, skipping truncation");
-		return payload;
-	}
-	const tokenCount = await getTokenCount(payload, model);
-	const safeLimit = Math.floor(maxPromptTokens * .95);
-	if (tokenCount.input <= safeLimit) return payload;
-	const groups = groupMessages(payload.messages);
-	const systemGroups = groups.filter((g) => g.isSystem);
-	const conversationGroups = groups.filter((g) => !g.isSystem);
-	if (conversationGroups.length === 0) {
-		consola.warn("No conversation messages to truncate, only system messages");
-		return payload;
-	}
-	let truncatedPayload = payload;
-	let dropCount = 0;
-	const maxDrop = Math.max(0, conversationGroups.length - 1);
-	while (dropCount <= maxDrop) {
-		const keptConversationGroups = conversationGroups.slice(dropCount);
-		const truncationNotice = dropCount > 0 ? [createTruncationNotice()] : [];
-		const newMessages = [
-			...systemGroups.flatMap((g) => g.messages),
-			...truncationNotice,
-			...keptConversationGroups.flatMap((g) => g.messages)
-		];
-		truncatedPayload = {
-			...payload,
-			messages: newMessages
-		};
-		const newTokenCount = await getTokenCount(truncatedPayload, model);
-		if (newTokenCount.input <= safeLimit) {
-			if (dropCount > 0) {
-				const droppedMessages = conversationGroups.slice(0, dropCount).reduce((sum, g) => sum + g.messages.length, 0);
-				console.log(`Truncated: ${tokenCount.input} -> ${newTokenCount.input} tokens (-${droppedMessages} msgs)`);
-			}
-			return truncatedPayload;
-		}
-		dropCount++;
-	}
-	const finalTokenCount = await getTokenCount(truncatedPayload, model);
-	consola.warn(`Could not reduce tokens below limit even after maximum truncation. Current: ${finalTokenCount.input}, limit: ${maxPromptTokens}. System messages or the last message may be too large.`);
-	return truncatedPayload;
-};
-
-//#endregion
-//#region src/lib/rate-limit.ts
-async function checkRateLimit(state$1) {
-	if (state$1.rateLimitSeconds === void 0) return;
-	const now = Date.now();
-	if (!state$1.lastRequestTimestamp) {
-		state$1.lastRequestTimestamp = now;
-		return;
-	}
-	const elapsedSeconds = (now - state$1.lastRequestTimestamp) / 1e3;
-	if (elapsedSeconds > state$1.rateLimitSeconds) {
-		state$1.lastRequestTimestamp = now;
-		return;
-	}
-	const waitTimeSeconds = Math.ceil(state$1.rateLimitSeconds - elapsedSeconds);
-	if (!state$1.rateLimitWait) {
-		consola.warn(`Rate limit exceeded. Need to wait ${waitTimeSeconds} more seconds.`);
-		throw new HTTPError("Rate limit exceeded", Response.json({ message: "Rate limit exceeded" }, { status: 429 }));
-	}
-	const waitTimeMs = waitTimeSeconds * 1e3;
-	consola.warn(`Rate limit reached. Waiting ${waitTimeSeconds} seconds before proceeding...`);
-	await sleep(waitTimeMs);
-	state$1.lastRequestTimestamp = now;
-	consola.info("Rate limit wait completed, proceeding with request");
-}
-
-//#endregion
-//#region src/services/copilot/create-chat-completions.ts
-const createChatCompletions = async (payload) => {
-	if (!state.copilotToken) throw new Error("Copilot token not found");
-	const enableVision = payload.messages.some((x) => typeof x.content !== "string" && x.content?.some((x$1) => x$1.type === "image_url"));
-	const isAgentCall = payload.messages.some((msg) => ["assistant", "tool"].includes(msg.role));
-	const buildHeaders = () => ({
-		...copilotHeaders(state, enableVision),
-		"X-Initiator": isAgentCall ? "agent" : "user"
-	});
-	consola.debug("Sending request to Copilot:", {
-		model: payload.model,
-		endpoint: `${copilotBaseUrl(state)}/chat/completions`
-	});
-	const url = `${copilotBaseUrl(state)}/chat/completions`;
-	const body = payload.stream ? {
-		...payload,
-		stream_options: { include_usage: true }
-	} : payload;
-	const bodyString = JSON.stringify(body);
-	const maxRetries = 2;
-	let lastError;
-	let response;
-	for (let attempt = 0; attempt <= maxRetries; attempt++) try {
-		response = await fetch(url, {
-			method: "POST",
-			headers: buildHeaders(),
-			body: bodyString
-		});
-		break;
-	} catch (error) {
-		lastError = error;
-		if (attempt < maxRetries) {
-			const delay = 1e3 * (attempt + 1);
-			consola.warn(`Network error on attempt ${attempt + 1}/${maxRetries + 1}, retrying in ${delay}ms:`, error instanceof Error ? error.message : error);
-			await new Promise((r) => setTimeout(r, delay));
-		}
-	}
-	if (!response) throw lastError;
-	if (response.status === 401) {
-		consola.warn("Copilot token expired, refreshing and retrying...");
-		try {
-			await refreshCopilotToken();
-			response = await fetch(url, {
-				method: "POST",
-				headers: buildHeaders(),
-				body: bodyString
-			});
-		} catch (refreshError) {
-			consola.error("Failed to refresh token:", refreshError);
-		}
-	}
-	if (!response.ok) {
-		const errorBody = await response.text();
-		consola.error("Failed to create chat completions", {
-			status: response.status,
-			statusText: response.statusText,
-			body: errorBody
-		});
-		throw new HTTPError(`Failed to create chat completions: ${response.status} ${errorBody}`, response);
-	}
-	if (payload.stream) return events(response);
-	return await response.json();
-};
-
-//#endregion
-//#region src/routes/chat-completions/handler.ts
-/**
-* Calculate token count, log it, and auto-truncate if needed.
-*
-* Uses multi-strategy exact matching via findModel() to handle
-* mismatches between requested and available model names.
-*/
-async function processPayloadTokens(payload) {
-	const selectedModel = findModel(payload.model);
-	if (!selectedModel) {
-		consola.warn("No model selected, skipping token count calculation");
-		return payload;
-	}
-	try {
-		const tokenCount = await getTokenCount(payload, selectedModel);
-		consola.debug("Current token count:", tokenCount);
-		const truncated = await truncateMessages(payload, selectedModel);
-		if (isNullish(truncated.max_tokens)) {
-			const withMaxTokens = {
-				...truncated,
-				max_tokens: selectedModel.capabilities.limits.max_output_tokens
-			};
-			consola.debug("Set max_tokens to:", JSON.stringify(withMaxTokens.max_tokens));
-			return withMaxTokens;
-		}
-		return truncated;
-	} catch (error) {
-		consola.warn("Failed to calculate token count:", error);
-		return payload;
-	}
-}
-async function handleCompletion$1(c) {
-	await checkRateLimit(state);
-	const rawPayload = await c.req.json();
-	consola.debug("Request payload:", JSON.stringify(rawPayload).slice(-400));
-	const payload = await processPayloadTokens(rawPayload);
-	if (state.manualApprove) await awaitApproval();
-	const response = await createChatCompletions(payload);
-	if (isNonStreaming$1(response)) {
-		consola.debug("Non-streaming response:", JSON.stringify(response));
-		if (response.usage) setTokenUsage({
-			inputTokens: response.usage.prompt_tokens,
-			outputTokens: response.usage.completion_tokens,
-			cacheReadTokens: response.usage.prompt_tokens_details?.cached_tokens
-		});
-		return c.json(response);
-	}
-	consola.debug("Streaming response");
-	return streamSSE(c, async (stream) => {
-		for await (const chunk of response) {
-			consola.debug("Streaming chunk:", JSON.stringify(chunk));
-			try {
-				const sseChunk = chunk;
-				if (sseChunk.data && sseChunk.data !== "[DONE]") {
-					const parsed = JSON.parse(sseChunk.data);
-					if (parsed.usage) {
-						const usage = {
-							inputTokens: parsed.usage.prompt_tokens ?? 0,
-							outputTokens: parsed.usage.completion_tokens ?? 0,
-							cacheReadTokens: parsed.usage.prompt_tokens_details?.cached_tokens
-						};
-						setTokenUsage(usage);
-					}
-				}
-			} catch {}
-			await stream.writeSSE(chunk);
-		}
-		signalStreamDone();
-	});
-}
-const isNonStreaming$1 = (response) => Object.hasOwn(response, "choices");
-
-//#endregion
-//#region src/routes/chat-completions/route.ts
-const completionRoutes = new Hono();
-completionRoutes.post("/", async (c) => {
-	try {
-		return await handleCompletion$1(c);
-	} catch (error) {
-		return await forwardError(c, error);
-	}
-});
-
-//#endregion
-//#region src/services/copilot/create-embeddings.ts
-const createEmbeddings = async (payload) => {
-	if (!state.copilotToken) throw new Error("Copilot token not found");
-	const response = await fetch(`${copilotBaseUrl(state)}/embeddings`, {
-		method: "POST",
-		headers: copilotHeaders(state),
-		body: JSON.stringify(payload)
-	});
-	if (!response.ok) throw new HTTPError("Failed to create embeddings", response);
-	return await response.json();
-};
-
-//#endregion
-//#region src/routes/embeddings/route.ts
-const embeddingRoutes = new Hono();
-embeddingRoutes.post("/", async (c) => {
-	try {
-		const paylod = await c.req.json();
-		const response = await createEmbeddings(paylod);
-		return c.json(response);
-	} catch (error) {
-		return await forwardError(c, error);
-	}
-});
-
 //#endregion
 //#region src/routes/messages/utils.ts
 function mapOpenAIStopReasonToAnthropic(finishReason) {
@@ -3373,7 +3230,7 @@ async function handleCountTokens(c) {
 		let finalTokenCount = tokenCount.input + tokenCount.output;
 		if (anthropicPayload.model.startsWith("claude")) finalTokenCount = Math.round(finalTokenCount * 1.15);
 		else if (anthropicPayload.model.startsWith("grok")) finalTokenCount = Math.round(finalTokenCount * 1.03);
-		consola.debug("Token count:", finalTokenCount);
+		console.log("Token count:", finalTokenCount);
 		return c.json({ input_tokens: finalTokenCount });
 	} catch (error) {
 		consola.error("Error counting tokens:", error);
@@ -3507,30 +3364,10 @@ function translateChunkToAnthropicEvents(chunk, state$1) {
 
 //#endregion
 //#region src/routes/messages/handler.ts
-/**
-* Auto-truncate OpenAI payload if prompt tokens exceed model limit.
-*
-* Uses multi-strategy exact matching via findModel() to handle
-* mismatches between Anthropic and Copilot model naming conventions.
-*/
-async function autoTruncatePayload(payload) {
-	const selectedModel = findModel(payload.model);
-	if (!selectedModel) {
-		consola.warn("No model selected for Anthropic endpoint, skipping auto-truncation");
-		return payload;
-	}
-	try {
-		return await truncateMessages(payload, selectedModel);
-	} catch (error) {
-		consola.warn("Failed to auto-truncate context:", error);
-		return payload;
-	}
-}
 async function handleCompletion(c) {
 	await checkRateLimit(state);
 	const anthropicPayload = await c.req.json();
-	const rawOpenAIPayload = translateToOpenAI(anthropicPayload);
-	const openAIPayload = await autoTruncatePayload(rawOpenAIPayload);
+	const openAIPayload = translateToOpenAI(anthropicPayload);
 	if (state.manualApprove) await awaitApproval();
 	const response = await createChatCompletions(openAIPayload);
 	if (isNonStreaming(response)) {