aicodeswitch 4.0.4 → 5.0.0

package/dist/server/conversions/compact.js

@@ -0,0 +1,613 @@
+"use strict";
+/**
+ * Conversation Compaction for the Responses API.
+ *
+ * Provides the core logic for `/v1/responses/compact`:
+ * - Extract conversation text from Responses API `input` array
+ * - Build compaction prompt
+ * - Build upstream requests via the conversion system (transformRequest)
+ * - Extract summary from upstream responses (transformResponse)
+ * - Build the final Responses API compaction response
+ *
+ * Two usage levels:
+ * - **High-level**: `prepareCompactRequest` + `processCompactResponse` — unified API
+ * - **Low-level**: Individual functions for fine-grained control
+ */
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.COMPACTION_SYSTEM_PROMPT = void 0;
+exports.extractConversationText = extractConversationText;
+exports.extractMessageContent = extractMessageContent;
+exports.isClaudeCompactRequest = isClaudeCompactRequest;
+exports.isLastClaudeMessageCompact = isLastClaudeMessageCompact;
+exports.isCodexCompactRequest = isCodexCompactRequest;
+exports.sanitizeClaudeMessagesForCompact = sanitizeClaudeMessagesForCompact;
+exports.flattenClaudeToolBlocksForCompact = flattenClaudeToolBlocksForCompact;
+exports.normalizeClaudeCompactRequestBody = normalizeClaudeCompactRequestBody;
+exports.stripClaudeCompactResponseContent = stripClaudeCompactResponseContent;
+exports.countUnpairedClaudeToolUses = countUnpairedClaudeToolUses;
+exports.summarizeClaudeMessagesForDebug = summarizeClaudeMessagesForDebug;
+exports.collectCompactPayloadDebugInfo = collectCompactPayloadDebugInfo;
+exports.collectClaudeToolUseDiagnostics = collectClaudeToolUseDiagnostics;
+exports.buildCompactionPrompt = buildCompactionPrompt;
+exports.buildCompactUpstreamRequest = buildCompactUpstreamRequest;
+exports.extractSummaryFromResponse = extractSummaryFromResponse;
+exports.buildCompactedResponse = buildCompactedResponse;
+exports.prepareCompactRequest = prepareCompactRequest;
+exports.processCompactResponse = processCompactResponse;
+const crypto_1 = __importDefault(require("crypto"));
+const index_js_1 = require("./index.js");
+// ============================================================
+// Conversation Text Extraction
+// ============================================================
+/**
+ * Extract conversation text from a Responses API `input` array.
+ * Converts each item (message, function_call, function_call_output) into
+ * readable text suitable for a compaction prompt.
+ * Compaction items are skipped (encrypted_content is opaque).
+ */
+function extractConversationText(input) {
+    const parts = [];
+    for (const item of input) {
+        if (!item || typeof item !== 'object') {
+            if (typeof item === 'string')
+                parts.push(item);
+            continue;
+        }
+        // Skip compaction items — their encrypted_content is not readable
+        if (item.type === 'compaction')
+            continue;
+        // Skip item references
+        if (item.type === 'item_reference')
+            continue;
+        // Message items (EasyInputMessage, ResponseOutputMessage, etc.)
+        if (item.type === 'message' || item.role) {
+            const role = item.role || 'unknown';
+            const content = extractMessageContent(item.content);
+            if (content) {
+                parts.push(`[${role}]: ${content}`);
+            }
+        }
+        // Function call items
+        if (item.type === 'function_call') {
+            const name = item.name || 'unknown';
+            const args = item.arguments || item.input || '';
+            const callId = item.call_id || '';
+            parts.push(`[function_call${callId ? ` (${callId})` : ''} -> ${name}]: ${typeof args === 'string' ? args : JSON.stringify(args)}`);
+        }
+        // Function call output items
+        if (item.type === 'function_call_output') {
+            const callId = item.call_id || '';
+            const output = item.output || '';
+            parts.push(`[function_call_output${callId ? ` (${callId})` : ''}]: ${typeof output === 'string' ? output : JSON.stringify(output)}`);
+        }
+    }
+    return parts.join('\n\n');
+}
+/**
+ * Extract text content from a Responses API message's content field.
+ * Handles string content, content arrays (input_text, output_text, text, etc.),
+ * and nested structures.
+ */
+function extractMessageContent(content) {
+    if (!content)
+        return '';
+    if (typeof content === 'string')
+        return content;
+    if (Array.isArray(content)) {
+        return content
+            .map((block) => {
+            if (!block || typeof block !== 'object')
+                return String(block || '');
+            if (typeof block === 'string')
+                return block;
+            // text, input_text, output_text
+            if (block.text)
+                return block.text;
+            // input_image, input_file — skip
+            if (block.type === 'input_image' || block.type === 'input_file')
+                return '[media content]';
+            return '';
+        })
+            .filter(Boolean)
+            .join('\n');
+    }
+    return String(content);
+}
+// ============================================================
+// Compact Request Detection
+// ============================================================
+/**
+ * 检测消息是否为 Claude Code 的 compact 命令请求。
+ *
+ * Compact 命令触发时，Claude Code 会在 messages 末尾插入一条特殊指令：
+ * - role 为 "user"
+ * - content 为数组，包含一个 text 块
+ * - text 内容以 "CRITICAL: Respond with TEXT ONLY" 开头
+ * - 包含对话摘要生成指令，要求输出 <analysis> 和 <summary> 结构
+ */
+function isClaudeCompactRequest(message) {
+    if (!message || message.role !== 'user') {
+        return false;
+    }
+    const content = message.content;
+    if (!Array.isArray(content)) {
+        return false;
+    }
+    for (const block of content) {
+        if ((block === null || block === void 0 ? void 0 : block.type) === 'text' && typeof block.text === 'string') {
+            const text = block.text;
+            if (text.includes('CRITICAL: Respond with TEXT ONLY') &&
+                text.includes('create a detailed summary of the conversation') &&
+                text.includes('<analysis>') &&
+                text.includes('<summary>')) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
+/**
+ * 检测消息列表中的最后一条消息是否为 Claude Code compact 请求。
+ */
+function isLastClaudeMessageCompact(messages) {
+    if (!Array.isArray(messages) || messages.length === 0) {
+        return false;
+    }
+    return isClaudeCompactRequest(messages[messages.length - 1]);
+}
+/**
+ * 检测请求是否为 Codex 的 compact（压缩）请求。
+ *
+ * Codex 基于 OpenAI Responses API，compact 操作走独立端点：
+ * - POST /v1/responses/compact
+ */
+function isCodexCompactRequest(path) {
+    if (!path || typeof path !== 'string') {
+        return false;
+    }
+    const normalizedPath = path.split('?')[0];
+    return /\/v1\/responses\/compact\/?$/.test(normalizedPath);
+}
+// ============================================================
+// Claude Messages Sanitization
+// ============================================================
+/**
+ * 判断内容块是否为工具调用类型（包括 tool_use 和 server_tool_use）。
+ * Claude Code 的内置工具（如 webReader）使用 server_tool_use 类型，
+ * 上游 Claude 兼容 API 同样要求其紧邻的下一条 user 消息包含对应的 tool_result。
+ */
+function isToolUseBlock(block) {
+    return ((block === null || block === void 0 ? void 0 : block.type) === 'tool_use' || (block === null || block === void 0 ? void 0 : block.type) === 'server_tool_use') && typeof block.id === 'string' && block.id;
+}
+/**
+ * 清理 Claude Messages API 格式的 messages，确保所有 tool_use/server_tool_use 都有对应的 tool_result。
+ * 对于没有对应 tool_result 的工具调用，添加合成的 tool_result 块。
+ *
+ * 这解决了 Claude Code compact 请求中 assistant 消息末尾有 tool_use/server_tool_use 但
+ * 下一条 user 消息（compact 指令）不含对应 tool_result 导致上游 400 错误的问题。
+ */
+function sanitizeClaudeMessagesForCompact(messages) {
+    if (!Array.isArray(messages) || messages.length === 0)
+        return messages;
+    const result = [...messages];
+    for (let i = 0; i < result.length; i++) {
+        const msg = result[i];
+        if ((msg === null || msg === void 0 ? void 0 : msg.role) !== 'assistant' || !Array.isArray(msg.content))
+            continue;
+        // 收集当前 assistant 消息中所有 tool_use/server_tool_use 的 id
+        const toolUseIds = new Set();
+        for (const block of msg.content) {
+            if (isToolUseBlock(block)) {
+                toolUseIds.add(block.id);
+            }
+        }
+        if (toolUseIds.size === 0)
+            continue;
+        // 检查下一条 user 消息中是否有对应的 tool_result
+        const nextMsg = result[i + 1];
+        if (!nextMsg || nextMsg.role !== 'user') {
+            // 若下一条不是 user，直接插入一条合成 user(tool_result) 消息，
+            // 强制满足 Claude 要求的“紧邻下一条消息必须含 tool_result”约束。
+            result.splice(i + 1, 0, {
+                role: 'user',
+                content: [...toolUseIds].map(id => ({
+                    type: 'tool_result',
+                    tool_use_id: id,
+                    content: '[Result omitted for compaction]',
+                })),
+            });
+            i += 1;
+            continue;
+        }
+        const userContent = Array.isArray(nextMsg.content)
+            ? [...nextMsg.content]
+            : (typeof nextMsg.content === 'string' && nextMsg.content.trim())
+                ? [{ type: 'text', text: nextMsg.content }]
+                : [];
+        const toolResultBlocks = userContent.filter((block) => (block === null || block === void 0 ? void 0 : block.type) === 'tool_result' && typeof block.tool_use_id === 'string' && block.tool_use_id);
+        const providedResultIds = new Set(toolResultBlocks.map((block) => block.tool_use_id));
+        const nonToolResultBlocks = userContent.filter((block) => (block === null || block === void 0 ? void 0 : block.type) !== 'tool_result');
+        // 找到未配对的 tool_use id
+        const missingIds = [...toolUseIds].filter(id => !providedResultIds.has(id));
+        // 为未配对的 tool_use 补充合成的 tool_result
+        const syntheticResults = missingIds.map(id => ({
+            type: 'tool_result',
+            tool_use_id: id,
+            content: '[Result omitted for compaction]',
+        }));
+        const orderedToolResults = [...toolResultBlocks, ...syntheticResults];
+        // 某些 Claude 兼容实现要求 tool_result 独占“下一条消息”，
+        // 不能与后续 compact 指令文本混在同一个 user message 中。
+        if (nonToolResultBlocks.length > 0) {
+            result.splice(i + 1, 1, {
+                role: 'user',
+                content: orderedToolResults,
+            }, Object.assign(Object.assign({}, nextMsg), { content: nonToolResultBlocks }));
+            i += 1;
+            continue;
+        }
+        result[i + 1] = Object.assign(Object.assign({}, nextMsg), { content: orderedToolResults });
+    }
+    return result;
+}
+/**
+ * 将 Claude 历史中的 tool_use / tool_result 块降级为普通文本块。
+ *
+ * compact 请求的目标只是生成摘要，不需要保留严格的工具调用协议语义。
+ * 对某些 Claude 兼容实现（如第三方 Claude 标准接口）而言，历史中出现
+ * tool_use/tool_result 往往会触发比 Anthropic 官方更严格的校验。
+ * 因此在 compact 场景下，将工具块平铺为文本是更稳妥的上游输入形式。
+ */
+function flattenClaudeToolBlocksForCompact(messages) {
+    if (!Array.isArray(messages) || messages.length === 0)
+        return messages;
+    return messages.map((message) => {
+        if (!Array.isArray(message === null || message === void 0 ? void 0 : message.content)) {
+            return message;
+        }
+        const flattenedContent = message.content.map((block) => {
+            if ((block === null || block === void 0 ? void 0 : block.type) === 'tool_use' || (block === null || block === void 0 ? void 0 : block.type) === 'server_tool_use') {
+                const toolName = typeof block.name === 'string' && block.name ? block.name : 'tool';
+                const toolInput = block.input ? JSON.stringify(block.input) : '';
+                return {
+                    type: 'text',
+                    text: `[Tool use: ${toolName}${toolInput ? ` ${toolInput}` : ''}]`,
+                };
+            }
+            if ((block === null || block === void 0 ? void 0 : block.type) === 'tool_result') {
+                let resultText = '';
+                if (typeof block.content === 'string') {
+                    resultText = block.content;
+                }
+                else if (Array.isArray(block.content)) {
+                    resultText = block.content
+                        .map((item) => {
+                        if (typeof item === 'string')
+                            return item;
+                        if (typeof (item === null || item === void 0 ? void 0 : item.text) === 'string')
+                            return item.text;
+                        return '';
+                    })
+                        .filter(Boolean)
+                        .join('\n');
+                }
+                return {
+                    type: 'text',
+                    text: `[Tool result${resultText ? `: ${resultText}` : ''}]`,
+                };
+            }
+            return block;
+        });
+        return Object.assign(Object.assign({}, message), { content: flattenedContent });
+    });
+}
+/**
+ * Normalize Claude compact request payload so the upstream model can only
+ * produce a plain-text summary instead of reasoning/tool output.
+ */
+function normalizeClaudeCompactRequestBody(body) {
+    if (!body || typeof body !== 'object') {
+        return body;
+    }
+    const normalized = Object.assign({}, body);
+    delete normalized.thinking;
+    delete normalized.tools;
+    delete normalized.tool_choice;
+    delete normalized.mcp_servers;
+    return normalized;
+}
+/**
+ * Strip non-text assistant content from Claude compact responses.
+ * Claude Code compact expects a plain text summary, so thinking/tool blocks
+ * should not be sent back downstream even if the upstream model produced them.
+ */
+function stripClaudeCompactResponseContent(response) {
+    if (!response || typeof response !== 'object' || !Array.isArray(response.content)) {
+        return response;
+    }
+    const filteredContent = response.content.filter((block) => (block === null || block === void 0 ? void 0 : block.type) === 'text');
+    return Object.assign(Object.assign({}, response), { content: filteredContent, stop_reason: response.stop_reason === 'tool_use' ? 'end_turn' : response.stop_reason });
+}
+function countUnpairedClaudeToolUses(messages) {
+    if (!Array.isArray(messages) || messages.length === 0)
+        return 0;
+    let unpairedCount = 0;
+    for (let i = 0; i < messages.length; i++) {
+        const msg = messages[i];
+        if ((msg === null || msg === void 0 ? void 0 : msg.role) !== 'assistant' || !Array.isArray(msg.content))
+            continue;
+        const toolUseIds = msg.content
+            .filter((block) => isToolUseBlock(block))
+            .map((block) => block.id);
+        if (toolUseIds.length === 0)
+            continue;
+        const nextMsg = messages[i + 1];
+        if (!nextMsg || nextMsg.role !== 'user') {
+            unpairedCount += toolUseIds.length;
+            continue;
+        }
+        const nextContent = Array.isArray(nextMsg.content) ? nextMsg.content : [];
+        const isPureToolResultMessage = nextContent.length > 0 && nextContent.every((block) => (block === null || block === void 0 ? void 0 : block.type) === 'tool_result' && typeof block.tool_use_id === 'string' && block.tool_use_id);
+        if (!isPureToolResultMessage) {
+            unpairedCount += toolUseIds.length;
+            continue;
+        }
+        const resultIds = new Set(nextContent.map((block) => block.tool_use_id));
+        unpairedCount += toolUseIds.filter((id) => !resultIds.has(id)).length;
+    }
+    return unpairedCount;
+}
+function summarizeClaudeMessagesForDebug(messages, startIndex, endIndex) {
+    if (!Array.isArray(messages) || messages.length === 0)
+        return [];
+    const start = Math.max(0, startIndex);
+    const end = Math.min(messages.length - 1, endIndex);
+    const summary = [];
+    for (let i = start; i <= end; i++) {
+        const msg = messages[i];
+        const content = Array.isArray(msg === null || msg === void 0 ? void 0 : msg.content)
+            ? msg.content.map((block, blockIndex) => ({
+                blockIndex,
+                type: block === null || block === void 0 ? void 0 : block.type,
+                id: typeof (block === null || block === void 0 ? void 0 : block.id) === 'string' ? block.id : undefined,
+                tool_use_id: typeof (block === null || block === void 0 ? void 0 : block.tool_use_id) === 'string' ? block.tool_use_id : undefined,
+                name: typeof (block === null || block === void 0 ? void 0 : block.name) === 'string' ? block.name : undefined,
+                text: typeof (block === null || block === void 0 ? void 0 : block.text) === 'string' ? block.text.slice(0, 120) : undefined,
+                content: typeof (block === null || block === void 0 ? void 0 : block.content) === 'string' ? block.content.slice(0, 120) : undefined,
+            }))
+            : msg === null || msg === void 0 ? void 0 : msg.content;
+        summary.push({
+            index: i,
+            role: msg === null || msg === void 0 ? void 0 : msg.role,
+            content,
+        });
+    }
+    return summary;
+}
+function collectCompactPayloadDebugInfo(body, targetCallId) {
+    const messages = Array.isArray(body === null || body === void 0 ? void 0 : body.messages) ? body.messages : [];
+    const serialized = (() => {
+        try {
+            return JSON.stringify(body);
+        }
+        catch (_a) {
+            return '';
+        }
+    })();
+    return {
+        messageCount: messages.length,
+        hasToolUseLiteral: serialized.includes('"tool_use"'),
+        hasToolResultLiteral: serialized.includes('"tool_result"'),
+        hasTargetCallId: targetCallId ? serialized.includes(targetCallId) : false,
+        window47to49: summarizeClaudeMessagesForDebug(messages, 47, 49),
+    };
+}
+function summarizeClaudeMessageContent(content) {
+    if (!Array.isArray(content))
+        return typeof content;
+    return content.map((block, index) => {
+        const type = (block === null || block === void 0 ? void 0 : block.type) || 'unknown';
+        const id = (block === null || block === void 0 ? void 0 : block.id) ? `#${block.id}` : '';
+        const toolUseId = (block === null || block === void 0 ? void 0 : block.tool_use_id) ? `->${block.tool_use_id}` : '';
+        const text = typeof (block === null || block === void 0 ? void 0 : block.text) === 'string'
+            ? `:${block.text.slice(0, 40).replace(/\s+/g, ' ')}`
+            : '';
+        return `${index}:${type}${id}${toolUseId}${text}`;
+    }).join(' | ');
+}
+function collectClaudeToolUseDiagnostics(messages) {
+    if (!Array.isArray(messages) || messages.length === 0)
+        return [];
+    const diagnostics = [];
+    for (let i = 0; i < messages.length; i++) {
+        const msg = messages[i];
+        if ((msg === null || msg === void 0 ? void 0 : msg.role) !== 'assistant' || !Array.isArray(msg.content))
+            continue;
+        const toolUseIds = msg.content
+            .filter((block) => isToolUseBlock(block))
+            .map((block) => block.id);
+        if (toolUseIds.length === 0)
+            continue;
+        const nextMsg = messages[i + 1];
+        diagnostics.push(`[${i}] assistant tool_use_ids=${toolUseIds.join(', ')} blocks=${summarizeClaudeMessageContent(msg.content)}`);
+        diagnostics.push(`[${i + 1}] ${(nextMsg === null || nextMsg === void 0 ? void 0 : nextMsg.role) || 'missing'} blocks=${summarizeClaudeMessageContent(nextMsg === null || nextMsg === void 0 ? void 0 : nextMsg.content)}`);
+    }
+    return diagnostics;
+}
+// ============================================================
+// Compaction Prompt
+// ============================================================
+/**
+ * The system prompt used to instruct the upstream model to compact a conversation.
+ */
+exports.COMPACTION_SYSTEM_PROMPT = `You are a conversation compaction assistant. Your task is to create a comprehensive but concise summary of the following programming conversation, preserving all essential context needed to continue the conversation without any loss of critical information.
+
+The summary MUST include:
+1. **Current Task/Goal**: What the user is trying to accomplish
+2. **Decisions Made**: Key decisions and their rationale
+3. **Code State**: Current state of relevant files, including file paths, function names, and recent changes
+4. **Errors Encountered**: Any errors or issues and their resolution status
+5. **Pending Items**: Tasks or actions that are still in progress or yet to be done
+6. **Context Details**: Important variables, configurations, or technical details discussed
+
+Output a well-structured summary that can fully replace the original conversation. Be thorough but concise. Use markdown formatting for clarity.`;
+/**
+ * Build the full compaction system prompt text.
+ */
+function buildCompactionPrompt(instructions) {
+    let prompt = exports.COMPACTION_SYSTEM_PROMPT;
+    if (instructions) {
+        prompt += `\n\nAdditional instructions from the user:\n${instructions}`;
+    }
+    return prompt;
+}
+// ============================================================
+// Upstream Request / Response
+// ============================================================
+/**
+ * Build an upstream request for compaction using the conversion system.
+ *
+ * Instead of manually constructing requests for each format, this builds a
+ * standard Responses API request body and uses `transformRequest()` to convert
+ * it to the target upstream format.
+ */
+function buildCompactUpstreamRequest(conversationText, instructions, toFormat, model) {
+    const systemPrompt = buildCompactionPrompt(instructions);
+    // Build a standard Responses API request that represents the compact task
+    const responsesBody = {
+        model,
+        instructions: systemPrompt,
+        input: [
+            {
+                type: 'message',
+                role: 'user',
+                content: conversationText,
+            },
+        ],
+        max_output_tokens: 8192,
+        stream: false,
+    };
+    const result = (0, index_js_1.transformRequest)({
+        fromFormat: 'responses',
+        toFormat,
+        body: responsesBody,
+    });
+    return { body: result.body };
+}
+/**
+ * Extract the summary text from an upstream response.
+ *
+ * Uses `transformResponse()` to convert the upstream response back to
+ * Responses API format, then extracts the text content.
+ */
+function extractSummaryFromResponse(response, fromFormat) {
+    if (!response)
+        return '';
+    // If already in responses format, extract directly
+    if (fromFormat === 'responses') {
+        return extractSummaryFromResponsesFormat(response);
+    }
+    // Convert upstream response to responses format, then extract
+    const responsesResponse = (0, index_js_1.transformResponse)({
+        fromFormat,
+        toFormat: 'responses',
+        response,
+    });
+    return extractSummaryFromResponsesFormat(responsesResponse);
+}
+/**
+ * Extract summary text from a Responses API format response.
+ */
+function extractSummaryFromResponsesFormat(response) {
+    if (!response)
+        return '';
+    // Responses API: output array with message items containing text
+    if (Array.isArray(response.output)) {
+        return response.output
+            .flatMap((item) => {
+            if (item.type === 'message' && Array.isArray(item.content)) {
+                return item.content
+                    .filter((block) => block.type === 'output_text' || block.type === 'text')
+                    .map((block) => block.text || '');
+            }
+            return [];
+        })
+            .filter(Boolean)
+            .join('\n');
+    }
+    // Fallback for raw string
+    if (typeof response === 'string')
+        return response;
+    return '';
+}
+// ============================================================
+// Compacted Response Builder
+// ============================================================
+/**
+ * Build a CompactedResponse object in the OpenAI Responses API format.
+ */
+function buildCompactedResponse(summary, _model, inputTokens = 0, outputTokens = 0) {
+    const randHex = () => crypto_1.default.randomUUID().replace(/-/g, '').substring(0, 24);
+    return {
+        id: `resp_compact_${randHex()}`,
+        created_at: Math.floor(Date.now() / 1000),
+        object: 'response.compaction',
+        output: [
+            {
+                type: 'message',
+                id: `msg_${randHex()}`,
+                role: 'assistant',
+                content: [
+                    {
+                        type: 'output_text',
+                        text: summary,
+                    },
+                ],
+                status: 'completed',
+            },
+        ],
+        usage: {
+            input_tokens: inputTokens,
+            input_tokens_details: { cached_tokens: 0 },
+            output_tokens: outputTokens,
+            output_tokens_details: { reasoning_tokens: 0 },
+            total_tokens: inputTokens + outputTokens,
+        },
+    };
+}
+/**
+ * 构建 compact 上游请求（统一入口）。
+ *
+ * - 目标格式是 `responses` → 直接转发原始请求体（passthrough）
+ * - 目标格式是其他格式 → 提取对话文本 → 构造压缩提示词 → 转换为目标格式
+ *
+ * 外部只需传入目标格式和模型名称即可，无需关心内部处理逻辑。
+ */
+function prepareCompactRequest(options) {
+    const { body, toFormat, model } = options;
+    // 目标格式是 responses：直接转发原始请求
+    if (toFormat === 'responses') {
+        return { body, isPassthrough: true };
+    }
+    // 其他格式：提取对话文本 → 构造压缩提示词 → 转换
+    const conversationText = extractConversationText(body.input);
+    const { body: transformedBody } = buildCompactUpstreamRequest(conversationText, body.instructions, toFormat, model);
+    return { body: transformedBody, isPassthrough: false };
+}
+/**
+ * 处理 compact 上游响应（统一入口）。
+ *
+ * - `isPassthrough: true` → 直接返回原始响应（上游已经是 Responses compact 格式）
+ * - `isPassthrough: false` → 提取摘要 → 构建 Responses API compact 响应
+ */
+function processCompactResponse(response, fromFormat, model, isPassthrough, usage) {
+    // Passthrough：上游已经是 Responses compact 格式，直接返回
+    if (isPassthrough)
+        return response;
+    // 非 passthrough：提取摘要 → 构建 compact 响应
+    const summary = extractSummaryFromResponse(response, fromFormat);
+    return buildCompactedResponse(summary, model, usage === null || usage === void 0 ? void 0 : usage.inputTokens, usage === null || usage === void 0 ? void 0 : usage.outputTokens);
+}

package/dist/server/conversions/detector.js

@@ -0,0 +1,70 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.detectRequestFormat = detectRequestFormat;
+exports.sourceTypeToFormat = sourceTypeToFormat;
+/**
+ * Detect the format of the incoming request based on path and body structure.
+ * Enhanced version that distinguishes between Chat Completions and Responses API.
+ */
+function detectRequestFormat(path, body) {
+    // Path-based detection (highest priority)
+    if (path.includes('/v1/messages') || path.includes('/messages')) {
+        return 'claude';
+    }
+    if (path.includes('/v1/chat/completions') || path.includes('/chat/completions')) {
+        return 'completions';
+    }
+    // Responses API: /v1/responses but NOT /v1/responses/compact
+    if (path.includes('/v1/responses') || path.includes('/responses')) {
+        if (!path.includes('/responses/compact') && !path.includes('/responses/compact')) {
+            return 'responses';
+        }
+    }
+    // Body-based detection
+    if (body) {
+        // Claude Messages format: messages array with content blocks
+        if (body.messages && Array.isArray(body.messages)) {
+            const firstMsg = body.messages[0];
+            if (firstMsg && typeof firstMsg.content === 'object' && Array.isArray(firstMsg.content)) {
+                return 'claude';
+            }
+        }
+        // Responses API: has 'input' field (string or array) and optionally 'instructions'
+        if (body.input !== undefined && body.model) {
+            // Has 'input' field → Responses API
+            if (typeof body.input === 'string' || Array.isArray(body.input)) {
+                return 'responses';
+            }
+        }
+        // OpenAI Chat Completions: has 'messages' + 'model'
+        if (body.messages && body.model) {
+            return 'completions';
+        }
+        // Fallback: if has 'input' without 'messages', assume responses
+        if (body.input !== undefined && !body.messages) {
+            return 'responses';
+        }
+    }
+    // Default to completions
+    return 'completions';
+}
+/**
+ * Determine the upstream format from a SourceType string.
+ * Maps the legacy SourceType values to the new Format type.
+ */
+function sourceTypeToFormat(sourceType) {
+    switch (sourceType) {
+        case 'claude':
+        case 'claude-chat':
+            return 'claude';
+        case 'openai':
+            return 'responses';
+        case 'openai-chat':
+            return 'completions';
+        case 'gemini':
+        case 'gemini-chat':
+            return 'gemini';
+        default:
+            return 'completions';
+    }
+}