feat: 支持 thinking/reasoning 提取与传递

- respond.js: buildChatCompletion 和 buildChatCompletionChunk 新增 reasoningContent 参数，输出 OpenAI o1 兼容的 reasoning_content 字段 - openai/routes.js: 解析请求体中的 reasoning 布尔标志，prompt 预览扩展到 100 字 - queue.js: 将 reasoning 标志传递给 generate - lmarena_text.js: 解析 SSE 中的 ag: 行提取思考过程 - deepseek_text.js: 识别 THINK fragment 类型，收集 thinking 内容 - gemini_text.js: 从 protobuf 响应的 rc[37] 路径提取 thinking 内容
2026-06-16 21:03:59 +08:00 · 2026-04-05 00:08:38 +08:00
parent bf553005ef
commit 7f5f0816ec
8 changed files with 243 additions and 36 deletions
@@ -79,7 +79,7 @@ async function configureModel(page, modelConfig, meta = {}) {
 * @param {string[]} imgPaths - 图片路径数组 (此适配器不支持)
 * @param {string} [modelId] - 模型 ID
 * @param {object} [meta={}] - 日志元数据
- * @returns {Promise<{text?: string, error?: string}>}
+ * @returns {Promise<{text?: string, reasoning?: string, error?: string}>}
 */
 async function generate(context, prompt, imgPaths, modelId, meta = {}) {
    const { page, config } = context;
@@ -108,8 +108,10 @@ async function generate(context, prompt, imgPaths, modelId, meta = {}) {
        logger.debug('适配器', '启动 API 监听...', meta);

        let textContent = '';
+        let thinkingContent = '';  // thinking 内容
        let isComplete = false;
        let isCollecting = false;  // 当前最后一个 fragment 是否为 RESPONSE 类型
+        let isCollectingThinking = false;  // 是否正在收集 thinking

        const responsePromise = page.waitForResponse(async (response) => {
            const url = response.url();
@@ -133,14 +135,21 @@ async function generate(context, prompt, imgPaths, modelId, meta = {}) {

                        // --- 处理 fragment 列表变更，更新 isCollecting 状态 ---

-                        // 初始响应中可能已有 fragments (如 SEARCH / RESPONSE)
+                        // 初始响应中可能已有 fragments (如 THINK / SEARCH / RESPONSE)
                        if (data.v?.response?.fragments && Array.isArray(data.v.response.fragments)) {
                            for (const fragment of data.v.response.fragments) {
                                if (fragment.type === 'RESPONSE') {
                                    isCollecting = true;
+                                    isCollectingThinking = false;
                                    if (fragment.content) textContent += fragment.content;
+                                } else if (fragment.type === 'THINK') {
+                                    // DeepSeek 使用 THINK (不是 THINKING)
+                                    isCollectingThinking = true;
+                                    isCollecting = false;
+                                    if (fragment.content) thinkingContent += fragment.content;
                                } else {
                                    isCollecting = false;
+                                    isCollectingThinking = false;
                                }
                            }
                        }
@@ -150,9 +159,15 @@ async function generate(context, prompt, imgPaths, modelId, meta = {}) {
                            for (const fragment of data.v) {
                                if (fragment.type === 'RESPONSE') {
                                    isCollecting = true;
+                                    isCollectingThinking = false;
                                    if (fragment.content) textContent += fragment.content;
+                                } else if (fragment.type === 'THINK') {
+                                    isCollectingThinking = true;
+                                    isCollecting = false;
+                                    if (fragment.content) thinkingContent += fragment.content;
                                } else {
                                    isCollecting = false;
+                                    isCollectingThinking = false;
                                }
                            }
                        }
@@ -164,9 +179,15 @@ async function generate(context, prompt, imgPaths, modelId, meta = {}) {
                                    for (const fragment of item.v) {
                                        if (fragment.type === 'RESPONSE') {
                                            isCollecting = true;
+                                            isCollectingThinking = false;
                                            if (fragment.content) textContent += fragment.content;
+                                        } else if (fragment.type === 'THINK') {
+                                            isCollectingThinking = true;
+                                            isCollecting = false;
+                                            if (fragment.content) thinkingContent += fragment.content;
                                        } else {
                                            isCollecting = false;
+                                            isCollectingThinking = false;
                                        }
                                    }
                                }
@@ -182,8 +203,12 @@ async function generate(context, prompt, imgPaths, modelId, meta = {}) {
                        // 带路径的 content 操作 (如 response/fragments/-1/content)
                        if (data.p && typeof data.v === 'string') {
                            const match = data.p.match(/response\/fragments\/(-?\d+)\/content/);
-                            if (match && isCollecting) {
-                                textContent += data.v;
+                            if (match) {
+                                if (isCollecting) {
+                                    textContent += data.v;
+                                } else if (isCollectingThinking) {
+                                    thinkingContent += data.v;
+                                }
                            }
                        }

@@ -191,6 +216,8 @@ async function generate(context, prompt, imgPaths, modelId, meta = {}) {
                        if (data.v && typeof data.v === 'string' && !data.p && !data.o) {
                            if (isCollecting) {
                                textContent += data.v;
+                            } else if (isCollectingThinking) {
+                                thinkingContent += data.v;
                            }
                        }

@@ -233,7 +260,16 @@ async function generate(context, prompt, imgPaths, modelId, meta = {}) {

        logger.info('适配器', `已获取文本内容 (${textContent.length} 字符)`, meta);
        logger.info('适配器', '文本生成完成，任务完成', meta);
-        return { text: textContent.trim() };
+
+        const trimmedThinking = thinkingContent.trim();
+        const result = { text: textContent.trim() };
+
+        // 返回结果（如果有 thinking 则包含 reasoning）
+        if (trimmedThinking) {
+            logger.info('适配器', `已获取思考过程 (${trimmedThinking.length} 字符)`, meta);
+            result.reasoning = trimmedThinking;
+        }
+        return result;

    } catch (err) {
        // 顶层错误处理
@@ -179,6 +179,12 @@ async function generate(context, prompt, imgPaths, modelId, meta = {}) {
            const imageUrl = imageUrls[0] + '=s1024-rj';
            logger.info('适配器', `找到 ${imageUrls.length} 张图片，开始下载...`, meta);

+            // 提取图片生成的详细描述（thinking）
+            const thinking = extractImageThinking(bodyBuffer);
+            if (thinking) {
+                logger.info('适配器', `提取到详细描述，长度: ${thinking.length}`, meta);
+            }
+
            // 使用封装的下载函数
            const imgDlCfg = config?.backend?.pool?.failover || {};
            const result = await useContextDownload(imageUrl, page, {
@@ -190,7 +196,8 @@ async function generate(context, prompt, imgPaths, modelId, meta = {}) {
            }

            logger.info('适配器', '已获取图片，任务完成', meta);
-            return result;
+            // 返回图片和 thinking（如果有）
+            return thinking ? { ...result, reasoning: thinking } : result;
        }

    } catch (err) {
@@ -428,4 +435,57 @@ function extractAiTextFromResponse(bodyBuffer) {
        }
    }
    return best;
+}
+
+/**
+ * 深度遍历，查找长文本描述（图片生成的 thinking/详细描述）
+ * 排除 URL、base64、分类器名称等非描述性长字符串
+ * @param {any} root - 要遍历的对象
+ * @returns {string} 最长的描述文本，未找到则返回空字符串
+ */
+function findLongDescriptionDeep(root) {
+    const candidates = [];
+    const stack = [root];
+
+    while (stack.length) {
+        const cur = stack.pop();
+        if (!cur) continue;
+
+        if (typeof cur === 'string') {
+            if (cur.length > 200 &&
+                !cur.startsWith('http') &&
+                !cur.startsWith('data:') &&
+                !cur.includes('googleapis.com') &&
+                !cur.includes('googleusercontent.com') &&
+                !/^[A-Za-z0-9+/=]{100,}$/.test(cur)) {
+                candidates.push(cur);
+            }
+        } else if (Array.isArray(cur)) {
+            for (const v of cur) stack.push(v);
+        } else if (typeof cur === 'object') {
+            for (const v of Object.values(cur)) stack.push(v);
+        }
+    }
+
+    if (candidates.length === 0) return '';
+    return candidates.reduce((a, b) => a.length >= b.length ? a : b, '');
+}
+
+/**
+ * 从响应体 Buffer 中提取图片生成的详细描述（thinking）
+ * @param {Buffer} bodyBuffer - 响应体 Buffer
+ * @returns {string} 详细描述文本，未找到则返回空字符串
+ */
+function extractImageThinking(bodyBuffer) {
+    const frames = parseLenFramedResponse(bodyBuffer);
+    const payloads = extractPayloads(frames);
+
+    let best = '';
+    for (const payload of payloads) {
+        const text = findLongDescriptionDeep(payload);
+        if (text.length > best.length) {
+            best = text;
+        }
+    }
+    return best;
 }
@@ -27,7 +27,7 @@ const TARGET_URL = 'https://gemini.google.com/app?hl=en';
 * @param {string[]} imgPaths - 图片路径数组
 * @param {string} [modelId] - 模型 ID (此适配器未使用)
 * @param {object} [meta={}] - 日志元数据
- * @returns {Promise<{text?: string, error?: string}>}
+ * @returns {Promise<{text?: string, reasoning?: string, error?: string}>}
 */
 async function generate(context, prompt, imgPaths, modelId, meta = {}) {
    const { page, config } = context;
@@ -172,11 +172,11 @@ async function generate(context, prompt, imgPaths, modelId, meta = {}) {
        const bodyBuffer = await apiResponse.body();
        logger.debug('适配器', `收到响应体，字节数: ${bodyBuffer.length}`, meta);

-        const text = getFinalAiTextFromResponse(bodyBuffer);
+        const { text, reasoning } = getFinalAiTextFromResponse(bodyBuffer);

        if (text) {
-            logger.info('适配器', `解析成功，文本长度: ${text.length}`, meta);
-            return { text };
+            logger.info('适配器', `解析成功，文本长度: ${text.length}，思考长度: ${reasoning?.length || 0}`, meta);
+            return reasoning ? { text, reasoning } : { text };
        } else {
            return { error: '未能从响应中提取文本' };
        }
@@ -390,19 +390,80 @@ function collectRcTextsDeep(root) {
 }

 /**
- * 从响应体 Buffer 中提取最终 AI 文本
+ * 从单个 payload 中提取文本和 thinking 内容
+ * @param {Array} payload - 解析后的 payload
+ * @returns {{text: string, thinking: string}}
+ */
+function extractTextAndThinking(payload) {
+    let text = '';
+    let thinking = '';
+
+    try {
+        if (!Array.isArray(payload)) return { text, thinking };
+
+        // 找 rc 节点 (通常在 payload[4][0])
+        // 结构: payload[4][0] = ["rc_xxx", ["text..."], ..., [37]: [[thinking]]]
+        let rc = null;
+        if (payload[4] && Array.isArray(payload[4][0]) &&
+            typeof payload[4][0][0] === 'string' && payload[4][0][0].startsWith('rc_')) {
+            rc = payload[4][0];
+        }
+
+        if (!rc) return { text, thinking };
+
+        // 文本在 rc[1][0]
+        if (Array.isArray(rc[1]) && typeof rc[1][0] === 'string') {
+            text = rc[1][0];
+        }
+
+        // thinking 在 rc[37][0][0]
+        // 结构: rc[37] = [["**Thinking Title**\n\nThinking content..."]]
+        try {
+            if (rc[37] && Array.isArray(rc[37]) && rc[37][0] && Array.isArray(rc[37][0])) {
+                if (typeof rc[37][0][0] === 'string') {
+                    thinking = rc[37][0][0];
+                }
+            }
+        } catch {
+            // thinking 提取失败，忽略
+        }
+    } catch {
+        // ignore
+    }
+
+    return { text, thinking };
+}
+
+/**
+ * 从响应体 Buffer 中提取最终 AI 文本和 thinking
 * @param {Buffer} bodyBuffer - 响应体 Buffer
+ * @returns {{text: string, reasoning: string}}
 */
 function getFinalAiTextFromResponse(bodyBuffer) {
    const frames = parseLenFramedResponse(bodyBuffer);
    const payloads = extractPayloads(frames);

-    let best = "";
+    let bestText = '';
+    let bestThinking = '';
+
+    // 遍历所有 payload，保留最长的 text 和对应的 thinking
    for (const payload of payloads) {
-        const m = collectRcTextsDeep(payload);
-        for (const text of m.values()) {
-            if (text.length > best.length) best = text;
+        const { text, thinking } = extractTextAndThinking(payload);
+        if (text.length > bestText.length) {
+            bestText = text;
+            bestThinking = thinking;
        }
    }
-    return best;
+
+    // 如果 extractTextAndThinking 没找到（可能结构不匹配），fallback 到 collectRcTextsDeep
+    if (!bestText) {
+        for (const payload of payloads) {
+            const m = collectRcTextsDeep(payload);
+            for (const text of m.values()) {
+                if (text.length > bestText.length) bestText = text;
+            }
+        }
+    }
+
+    return { text: bestText, reasoning: bestThinking };
 }
@@ -134,30 +134,48 @@ async function generate(context, prompt, imgPaths, modelId, meta = {}) {
        }

        // 9. 解析文本流
-        // 格式示例:
+        // SSE 格式说明:
+        // - a0: 回复内容（最终文本）
+        // - ag: 思考过程 (thinking/reasoning) - 仅 thinking 模型有
+        // - a2: 心跳 [{"type":"heartbeat"}]
+        // - ad: 结束标记 {"finishReason":"stop"}
+        // 示例:
+        // ag:"Let me think..."
        // a0:"Hello"
        // a0:" World"
-        // d:{"finishReason":"stop"}
+        // ad:{"finishReason":"stop"}
        let fullText = '';
+        let thinkingText = '';
        const lines = content.split('\n');

        for (const line of lines) {
            if (line.startsWith('a0:')) {
                try {
-                    // 尝试解析 JSON 字符串内容
-                    // line.substring(3) 应该是 JSON 字符串，如 "Hello"
                    const textPart = JSON.parse(line.substring(3));
                    fullText += textPart;
                } catch (e) {
-                    // 如果解析失败，可能是原生文本或其他格式
                    logger.warn('适配器', `解析文本块失败: ${line}`, meta);
                }
+            } else if (line.startsWith('ag:')) {
+                // 思考过程内容
+                try {
+                    const thinkPart = JSON.parse(line.substring(3));
+                    thinkingText += thinkPart;
+                } catch (e) {
+                    logger.warn('适配器', `解析思考块失败: ${line}`, meta);
+                }
            }
        }

        if (fullText) {
            logger.info('适配器', `获取文本成功，长度: ${fullText.length}`, meta);
-            return { text: fullText };
+            const result = { text: fullText };
+            // 如果有思考过程，添加到 reasoning 字段
+            if (thinkingText.trim()) {
+                logger.info('适配器', `获取思考过程，长度: ${thinkingText.length}`, meta);
+                result.reasoning = thinkingText;
+            }
+            return result;
        } else {
            logger.warn('适配器', '未解析到有效文本内容', { ...meta, preview: content.substring(0, 150) });
            // 如果没解析到 a0，尝试直接返回原始内容防空
@@ -123,8 +123,9 @@ export function createOpenAIRouter(context) {
            }

            const { prompt, imagePaths, modelId, modelName } = parseResult.data;
+            const reasoning = data.reasoning === true;

-            logger.info('服务器', `[队列] 请求入队: ${prompt.slice(0, 10)}...`, { id: requestId, images: imagePaths.length });
+            logger.info('服务器', `[队列] 请求入队: ${prompt.slice(0, 100)}...`, { id: requestId, images: imagePaths.length });

            // 加入队列
            queueManager.addTask({
@@ -135,7 +136,8 @@ export function createOpenAIRouter(context) {
                modelId,
                modelName,
                id: requestId,
-                isStreaming
+                isStreaming,
+                reasoning
            });

        } catch (err) {
@@ -94,7 +94,7 @@ export function createQueueManager(queueConfig, callbacks) {
     * @param {TaskContext} task - 任务上下文
     */
    async function processTask(task) {
-        const { res, prompt, imagePaths, modelId, modelName, id, isStreaming } = task;
+        const { res, prompt, imagePaths, modelId, modelName, id, isStreaming, reasoning } = task;
        const startTime = Date.now();

        logger.info('服务器', '[队列] 开始处理任务', { id, remaining: queue.length });
@@ -133,7 +133,7 @@ export function createQueueManager(queueConfig, callbacks) {
            }

            // 调用核心生图逻辑 (通过 Pool 分发)
-            const result = await generate(poolContext, prompt, imagePaths, modelId, { id });
+            const result = await generate(poolContext, prompt, imagePaths, modelId, { id, reasoning });

            // 清除心跳
            if (heartbeatInterval) clearInterval(heartbeatInterval);
@@ -106,9 +106,19 @@ export function sendApiError(res, options) {
 * 构造 OpenAI 格式的聊天完成响应（非流式）
 * @param {string} content - 响应内容
 * @param {string} [modelName] - 模型名称
+ * @param {string} [reasoningContent] - 思考/推理过程内容 (OpenAI o1 格式)
 * @returns {object} OpenAI 格式的响应对象
 */
-export function buildChatCompletion(content, modelName) {
+export function buildChatCompletion(content, modelName, reasoningContent) {
+    const message = {
+        role: 'assistant',
+        content: content
+    };
+
+    if (reasoningContent) {
+        message.reasoning_content = reasoningContent;
+    }
+
    return {
        id: 'chatcmpl-' + Date.now(),
        object: 'chat.completion',
@@ -116,10 +126,7 @@ export function buildChatCompletion(content, modelName) {
        model: modelName || 'default-model',
        choices: [{
            index: 0,
-            message: {
-                role: 'assistant',
-                content: content
-            },
+            message,
            finish_reason: 'stop'
        }]
    };
@@ -130,9 +137,16 @@ export function buildChatCompletion(content, modelName) {
 * @param {string} content - 响应内容
 * @param {string} [modelName] - 模型名称
 * @param {string|null} [finishReason='stop'] - 完成原因
+ * @param {string} [reasoningContent] - 思考/推理过程内容 (OpenAI o1 格式)
 * @returns {object} OpenAI 格式的流式响应块
 */
-export function buildChatCompletionChunk(content, modelName, finishReason = 'stop') {
+export function buildChatCompletionChunk(content, modelName, finishReason = 'stop', reasoningContent) {
+    const delta = { content };
+
+    if (reasoningContent) {
+        delta.reasoning_content = reasoningContent;
+    }
+
    return {
        id: 'chatcmpl-' + Date.now(),
        object: 'chat.completion.chunk',
@@ -140,7 +154,7 @@ export function buildChatCompletionChunk(content, modelName, finishReason = 'sto
        model: modelName || 'default-model',
        choices: [{
            index: 0,
-            delta: { content },
+            delta,
            finish_reason: finishReason
        }]
    };
@@ -112,6 +112,9 @@ function shouldLog(level) {
    return targetIndex >= effectiveEnvIndex;
 }

+// 需要提取到前面用方括号显示的 meta 字段
+const FRONT_META_KEYS = ['id', 'adapter', 'model'];
+
 export function log(level, mod, msg, meta = {}) {
    if (!shouldLog(level)) return;

@@ -121,10 +124,23 @@ export function log(level, mod, msg, meta = {}) {

    // 将消息中的换行符替换为 ↵ 符号，保持日志为单行
    const sanitizedMsg = msg.replace(/\r?\n/g, ' ↵ ');
-    const base = `${ts} [${levelTag}] [${mod}] ${sanitizedMsg}`;

-    const metaStr = Object.keys(meta).length
-        ? ' | ' + Object.entries(meta).map(([k, v]) => {
+    // 提取关键字段放在前面用方括号显示
+    const frontParts = [];
+    const remainingMeta = {};
+    for (const [k, v] of Object.entries(meta)) {
+        if (FRONT_META_KEYS.includes(k) && v !== undefined && v !== null) {
+            frontParts.push(`[${v}]`);
+        } else {
+            remainingMeta[k] = v;
+        }
+    }
+    const frontStr = frontParts.length ? ' ' + frontParts.join(' ') : '';
+
+    const base = `${ts} [${levelTag}] [${mod}]${frontStr} ${sanitizedMsg}`;
+
+    const metaStr = Object.keys(remainingMeta).length
+        ? ' | ' + Object.entries(remainingMeta).map(([k, v]) => {
            if (v instanceof Error) {
                return `${k}=${v.message}`;
            }