@jsonstudio/rcc 0.89.333 → 0.89.524

package/node_modules/@jsonstudio/llms/dist/conversion/codecs/gemini-openai-codec.js

@@ -4,6 +4,7 @@ import { normalizeChatMessageContent } from '../shared/chat-output-normalizer.js
 import { mapBridgeToolsToChat } from '../shared/tool-mapping.js';
 import { prepareGeminiToolsForBridge } from '../shared/gemini-tool-utils.js';
 import { registerResponsesReasoning, consumeResponsesReasoning, registerResponsesOutputTextMeta, consumeResponsesOutputTextMeta, consumeResponsesPayloadSnapshot, registerResponsesPayloadSnapshot, consumeResponsesPassthrough, registerResponsesPassthrough } from '../shared/responses-reasoning-registry.js';
+const DUMMY_THOUGHT_SIGNATURE = 'skip_thought_signature_validator';
 function isObject(v) {
     return !!v && typeof v === 'object' && !Array.isArray(v);
 }
@@ -54,6 +55,32 @@ function mapChatRoleToGemini(role) {
         return 'tool';
     return 'user';
 }
+function coerceThoughtSignature(value) {
+    if (typeof value === 'string' && value.trim().length) {
+        return value.trim();
+    }
+    return undefined;
+}
+function extractThoughtSignatureFromToolCall(tc) {
+    if (!tc || typeof tc !== 'object') {
+        return undefined;
+    }
+    const direct = coerceThoughtSignature(tc.thought_signature ?? tc.thoughtSignature);
+    if (direct) {
+        return direct;
+    }
+    const extraContent = tc.extra_content ?? tc.extraContent;
+    if (extraContent && typeof extraContent === 'object') {
+        const googleNode = extraContent.google ?? extraContent.Google;
+        if (googleNode && typeof googleNode === 'object') {
+            const googleSig = coerceThoughtSignature(googleNode.thought_signature ?? googleNode.thoughtSignature);
+            if (googleSig) {
+                return googleSig;
+            }
+        }
+    }
+    return undefined;
+}
 export function buildOpenAIChatFromGeminiRequest(payload) {
     const messages = [];
     // systemInstruction → Chat system 消息
@@ -156,16 +183,22 @@ export function buildOpenAIChatFromGeminiResponse(payload) {
     const textParts = [];
     const reasoningParts = [];
     const toolCalls = [];
+    const toolResultTexts = [];
+    const toolOutputs = [];
+    // 为当前响应内生成稳定的工具调用 ID，避免下游 ServerTool 因缺少 id 而跳过。
+    let toolCallCounter = 0;
     for (const part of parts) {
         if (!part || typeof part !== 'object')
             continue;
         const pObj = part;
+        // 1. Text part
         if (typeof pObj.text === 'string') {
             const t = pObj.text;
             if (t && t.trim().length)
                 textParts.push(t);
             continue;
         }
+        // 2. Content array (nested structure)
         if (Array.isArray(pObj.content)) {
             for (const inner of pObj.content) {
                 if (typeof inner === 'string') {
@@ -177,16 +210,26 @@ export function buildOpenAIChatFromGeminiResponse(payload) {
             }
             continue;
         }
+        // 3. Reasoning part (channel mode)
         if (typeof pObj.reasoning === 'string') {
             reasoningParts.push(pObj.reasoning);
             continue;
         }
+        // 4. Thought part (thinking/extended thinking)
+        if (typeof pObj.thought === 'string') {
+            const thoughtText = pObj.thought.trim();
+            if (thoughtText.length) {
+                reasoningParts.push(thoughtText);
+            }
+            continue;
+        }
+        // 5. Function call (tool call)
         if (pObj.functionCall && typeof pObj.functionCall === 'object') {
             const fc = pObj.functionCall;
             const name = typeof fc.name === 'string' ? String(fc.name) : undefined;
             if (!name)
                 continue;
-            const id = typeof fc.id === 'string' ? String(fc.id) : undefined;
+            let id = typeof fc.id === 'string' && fc.id.trim().length ? String(fc.id).trim() : undefined;
             const argsRaw = (fc.args ?? fc.arguments);
             let argsStr;
             if (typeof argsRaw === 'string') {
@@ -195,15 +238,91 @@ export function buildOpenAIChatFromGeminiResponse(payload) {
             else {
                 argsStr = safeJson(argsRaw);
             }
-            toolCalls.push({
+            const thoughtSignature = coerceThoughtSignature(pObj.thoughtSignature);
+            // Gemini 某些响应中的 functionCall 不带 id，但下游 servertool 需要稳定的
+            // tool_call.id 才会识别为有效工具调用；此处在缺失时生成本响应内唯一的占位 ID。
+            if (!id) {
+                const suffix = toolCallCounter++;
+                id = `gemini_tool_${suffix}`;
+            }
+            const toolCall = {
                 id,
                 type: 'function',
                 function: { name, arguments: argsStr }
-            });
+            };
+            if (thoughtSignature) {
+                toolCall.thought_signature = thoughtSignature;
+                toolCall.extra_content = {
+                    google: {
+                        thought_signature: thoughtSignature
+                    }
+                };
+            }
+            toolCalls.push(toolCall);
+            continue;
+        }
+        // 6. Function response (tool result)
+        if (pObj.functionResponse && typeof pObj.functionResponse === 'object') {
+            const fr = pObj.functionResponse;
+            const callId = typeof fr.id === 'string' && fr.id.trim().length ? String(fr.id) : undefined;
+            const name = typeof fr.name === 'string' && fr.name.trim().length ? String(fr.name) : undefined;
+            const resp = fr.response;
+            let contentStr = '';
+            if (typeof resp === 'string') {
+                contentStr = resp;
+            }
+            else if (resp != null) {
+                try {
+                    contentStr = JSON.stringify(resp);
+                }
+                catch {
+                    contentStr = String(resp);
+                }
+            }
+            if (contentStr && contentStr.trim().length) {
+                toolResultTexts.push(contentStr);
+                if (callId || name) {
+                    const entry = {
+                        tool_call_id: callId ?? undefined,
+                        id: callId ?? undefined,
+                        content: contentStr
+                    };
+                    if (name) {
+                        entry.name = name;
+                    }
+                    toolOutputs.push(entry);
+                }
+            }
+            continue;
+        }
+        // 7. Executable code (code_interpreter)
+        if (pObj.executableCode && typeof pObj.executableCode === 'object') {
+            const code = pObj.executableCode;
+            const language = typeof code.language === 'string' ? code.language : 'python';
+            const codeText = typeof code.code === 'string' ? code.code : '';
+            if (codeText.trim().length) {
+                // Append as text with code block formatting
+                textParts.push(`\`\`\`${language}\n${codeText}\n\`\`\``);
+            }
+            continue;
+        }
+        // 8. Code execution result
+        if (pObj.codeExecutionResult && typeof pObj.codeExecutionResult === 'object') {
+            const result = pObj.codeExecutionResult;
+            const outcome = typeof result.outcome === 'string' ? result.outcome : '';
+            const output = typeof result.output === 'string' ? result.output : '';
+            if (output.trim().length) {
+                textParts.push(`[Code Output${outcome ? ` (${outcome})` : ''}]:\n${output}`);
+            }
             continue;
         }
     }
+    const hasToolCalls = toolCalls.length > 0;
     const finish_reason = (() => {
+        // If the model is emitting tool calls, treat this turn as a tool_calls
+        // completion so downstream tool governance can continue the loop.
+        if (hasToolCalls)
+            return 'tool_calls';
         const fr = String(primary?.finishReason || '').toUpperCase();
         if (fr === 'MAX_TOKENS')
             return 'length';
@@ -228,9 +347,14 @@ export function buildOpenAIChatFromGeminiResponse(payload) {
         usage.total_tokens = totalTokens;
     const combinedText = textParts.join('\n');
     const normalized = combinedText.length ? normalizeChatMessageContent(combinedText) : { contentText: undefined, reasoningText: undefined };
+    const baseContent = normalized.contentText ?? combinedText ?? '';
+    const toolResultBlock = toolResultTexts.length ? toolResultTexts.join('\n') : '';
+    const finalContent = toolResultBlock && baseContent
+        ? `${baseContent}\n${toolResultBlock}`
+        : baseContent || toolResultBlock;
     const chatMsg = {
         role,
-        content: normalized.contentText ?? combinedText ?? ''
+        content: finalContent
     };
     if (typeof normalized.reasoningText === 'string' && normalized.reasoningText.trim().length) {
         reasoningParts.push(normalized.reasoningText.trim());
@@ -277,6 +401,9 @@ export function buildOpenAIChatFromGeminiResponse(payload) {
     if (Object.keys(usage).length > 0) {
         chatResp.usage = usage;
     }
+    if (toolOutputs.length > 0) {
+        chatResp.tool_outputs = toolOutputs;
+    }
     const preservedReasoning = consumeResponsesReasoning(chatResp.id);
     if (preservedReasoning && preservedReasoning.length) {
         chatResp.__responses_reasoning = preservedReasoning;
@@ -398,7 +525,12 @@ export function buildGeminiFromOpenAIChat(chatResp) {
         const id = typeof tc.id === 'string' ? String(tc.id) : undefined;
         if (id)
             functionCall.id = id;
-        parts.push({ functionCall });
+        const thoughtSignature = extractThoughtSignatureFromToolCall(tc) ?? DUMMY_THOUGHT_SIGNATURE;
+        const partEntry = { functionCall };
+        if (thoughtSignature) {
+            partEntry.thoughtSignature = thoughtSignature;
+        }
+        parts.push(partEntry);
     }
     const candidate = {
         content: {

package/node_modules/@jsonstudio/llms/dist/conversion/compat/actions/gemini-web-search.d.ts

@@ -0,0 +1,17 @@
+import type { JsonObject } from '../../hub/types/json.js';
+import type { AdapterContext } from '../../hub/types/chat-envelope.js';
+/**
+ * Gemini web_search 请求适配（作用于 Gemini 请求 payload，而不是 ChatEnvelope）：
+ *
+ * - 仅在 routeId 以 `web_search` 开头时生效（来自 AdapterContext.routeId）；
+ * - 针对 Gemini 请求中的 `tools` 字段进行清洗：
+ *   - 保留 name === 'web_search' 的 functionDeclarations（如果存在）；
+ *   - 丢弃其它 Codex 自身工具（exec_command / MCP 等），避免 Cloud Code 报
+ *     “Multiple tools are supported only when they are all search tools.”；
+ *   - 如果清洗后没有任何工具，则删除 `tools` 字段。
+ *
+ * 注意：
+ * - 这里处理的是 llmswitch-core 为 gemini-chat 构造的中间 payload（buildGeminiRequestFromChat 输出）；
+ * - 对于 Gemini CLI 后续在传输层添加的 googleSearch 工具，本适配不负责构造或修改。
+ */
+export declare function applyGeminiWebSearchCompat(payload: JsonObject, adapterContext?: AdapterContext): JsonObject;

package/node_modules/@jsonstudio/llms/dist/conversion/compat/actions/gemini-web-search.js

@@ -0,0 +1,68 @@
+const isRecord = (value) => typeof value === 'object' && value !== null && !Array.isArray(value);
+/**
+ * Gemini web_search 请求适配（作用于 Gemini 请求 payload，而不是 ChatEnvelope）：
+ *
+ * - 仅在 routeId 以 `web_search` 开头时生效（来自 AdapterContext.routeId）；
+ * - 针对 Gemini 请求中的 `tools` 字段进行清洗：
+ *   - 保留 name === 'web_search' 的 functionDeclarations（如果存在）；
+ *   - 丢弃其它 Codex 自身工具（exec_command / MCP 等），避免 Cloud Code 报
+ *     “Multiple tools are supported only when they are all search tools.”；
+ *   - 如果清洗后没有任何工具，则删除 `tools` 字段。
+ *
+ * 注意：
+ * - 这里处理的是 llmswitch-core 为 gemini-chat 构造的中间 payload（buildGeminiRequestFromChat 输出）；
+ * - 对于 Gemini CLI 后续在传输层添加的 googleSearch 工具，本适配不负责构造或修改。
+ */
+export function applyGeminiWebSearchCompat(payload, adapterContext) {
+    const routeId = typeof adapterContext?.routeId === 'string' ? adapterContext.routeId : '';
+    if (!routeId || !routeId.toLowerCase().startsWith('web_search')) {
+        return payload;
+    }
+    const cloned = { ...payload };
+    // 对于 Gemini Chat 后端，请求根节点不支持自定义 `web_search` 字段，
+    // 统一在 compat 层移除，避免触发
+    // "Unknown name \"web_search\" at 'request': Cannot find field." 错误。
+    if (cloned.web_search !== undefined) {
+        delete cloned.web_search;
+    }
+    const toolsRaw = cloned.tools;
+    // 当 web_search 路由下完全没有 tools 时，为 Gemini 搜索模型注入最小 googleSearch 工具，
+    // 保持与 gcli2api 类似的“搜索模型自动启用 Search 工具”行为。
+    if (!Array.isArray(toolsRaw)) {
+        cloned.tools = [{ googleSearch: {} }];
+        return cloned;
+    }
+    const nextTools = [];
+    for (const entry of toolsRaw) {
+        if (!isRecord(entry))
+            continue;
+        // 1) 保留 name === 'web_search' 的 functionDeclarations（单函数声明）。
+        const decls = Array.isArray(entry.functionDeclarations)
+            ? entry.functionDeclarations
+            : [];
+        const webSearchDecls = decls.filter((fn) => {
+            if (!isRecord(fn))
+                return false;
+            const name = typeof fn.name === 'string' ? fn.name.trim().toLowerCase() : '';
+            return name === 'web_search';
+        });
+        if (webSearchDecls.length) {
+            nextTools.push({
+                functionDeclarations: webSearchDecls
+            });
+            continue;
+        }
+        // 2) 若上游已经构造了 googleSearch 工具，则保留。
+        if (isRecord(entry.googleSearch)) {
+            nextTools.push({ googleSearch: entry.googleSearch });
+        }
+    }
+    if (nextTools.length > 0) {
+        cloned.tools = nextTools;
+    }
+    else {
+        // 3) 清洗后没有任何工具时，再次注入一个最小 googleSearch 工具，确保搜索模型仍然启用 Search。
+        cloned.tools = [{ googleSearch: {} }];
+    }
+    return cloned;
+}

package/node_modules/@jsonstudio/llms/dist/conversion/compat/actions/glm-image-content.d.ts

@@ -0,0 +1,2 @@
+import type { JsonObject } from '../../hub/types/json.js';
+export declare function applyGlmImageContentTransform(payload: JsonObject): JsonObject;

package/node_modules/@jsonstudio/llms/dist/conversion/compat/actions/glm-image-content.js

@@ -0,0 +1,83 @@
+const isRecord = (value) => typeof value === 'object' && value !== null && !Array.isArray(value);
+function normalizeImagePart(part) {
+    const rawType = typeof part.type === 'string' ? part.type.toLowerCase() : '';
+    if (rawType !== 'image' && rawType !== 'image_url') {
+        return null;
+    }
+    // Try multiple locations for the URL, since different inbound
+    // protocols may populate different keys.
+    const imageUrlBlock = isRecord(part.image_url)
+        ? part.image_url
+        : undefined;
+    let url;
+    if (imageUrlBlock && typeof imageUrlBlock.url === 'string') {
+        url = imageUrlBlock.url;
+    }
+    else if (typeof part.image_url === 'string') {
+        url = part.image_url;
+    }
+    else if (typeof part.url === 'string') {
+        url = part.url;
+    }
+    else if (typeof part.uri === 'string') {
+        url = part.uri;
+    }
+    else if (typeof part.data === 'string') {
+        // If caller passed a raw base64/data URI string, forward as-is.
+        url = part.data;
+    }
+    if (!url || !url.trim().length) {
+        return null;
+    }
+    const normalized = {
+        type: 'image_url',
+        image_url: {
+            url: url.trim()
+        }
+    };
+    // Preserve a best-effort "detail" field when present.
+    const detailValue = (imageUrlBlock && imageUrlBlock.detail) ?? part.detail;
+    if (typeof detailValue === 'string' && detailValue.trim().length) {
+        normalized.image_url.detail = detailValue.trim();
+    }
+    return normalized;
+}
+export function applyGlmImageContentTransform(payload) {
+    const root = structuredClone(payload);
+    const messagesValue = root.messages;
+    if (!Array.isArray(messagesValue)) {
+        return root;
+    }
+    const messages = [];
+    for (const msg of messagesValue) {
+        if (!isRecord(msg)) {
+            messages.push(msg);
+            continue;
+        }
+        const contentValue = msg.content;
+        if (!Array.isArray(contentValue)) {
+            messages.push(msg);
+            continue;
+        }
+        const newContent = [];
+        for (const part of contentValue) {
+            if (!isRecord(part)) {
+                newContent.push(part);
+                continue;
+            }
+            const normalizedImage = normalizeImagePart(part);
+            if (normalizedImage) {
+                newContent.push(normalizedImage);
+            }
+            else {
+                newContent.push(part);
+            }
+        }
+        messages.push({
+            ...msg,
+            content: newContent
+        });
+    }
+    root.messages = messages;
+    return root;
+}

package/node_modules/@jsonstudio/llms/dist/conversion/compat/actions/glm-vision-prompt.d.ts

@@ -0,0 +1,11 @@
+import type { JsonObject } from '../../hub/types/json.js';
+/**
+ * GLM 4.6V 专用视觉提示裁剪：
+ * - 仅在 model 为 glm-4.6v 且存在携带图片的 user 消息时生效；
+ * - 丢弃原有 system 与历史对话，只保留一条新的 system + 一条 user；
+ * - system：要求模型以结构化 JSON 描述截图内容、标记（圈/箭头等）及其大致 bbox；
+ * - user：保留原始用户文字（如果有）+ 单一 image_url 块。
+ *
+ * 其他模型（包括 glm-4.7、Gemini 等）不受影响。
+ */
+export declare function applyGlmVisionPromptTransform(payload: JsonObject): JsonObject;

package/node_modules/@jsonstudio/llms/dist/conversion/compat/actions/glm-vision-prompt.js

@@ -0,0 +1,177 @@
+const isRecord = (value) => typeof value === 'object' && value !== null && !Array.isArray(value);
+function extractImageUrlFromPart(part) {
+    const imageUrlBlock = isRecord(part.image_url)
+        ? part.image_url
+        : undefined;
+    let url;
+    if (imageUrlBlock && typeof imageUrlBlock.url === 'string') {
+        url = imageUrlBlock.url;
+    }
+    else if (typeof part.image_url === 'string') {
+        url = part.image_url;
+    }
+    else if (typeof part.url === 'string') {
+        url = part.url;
+    }
+    else if (typeof part.uri === 'string') {
+        url = part.uri;
+    }
+    else if (typeof part.data === 'string') {
+        url = part.data;
+    }
+    const trimmed = (url ?? '').trim();
+    return trimmed.length ? trimmed : null;
+}
+function collectUserTextFromMessage(msg) {
+    const content = msg.content;
+    if (!Array.isArray(content)) {
+        if (typeof content === 'string') {
+            return content;
+        }
+        return '';
+    }
+    const parts = [];
+    for (const entry of content) {
+        if (!isRecord(entry))
+            continue;
+        const text = entry.text;
+        if (typeof text === 'string' && text.trim().length) {
+            parts.push(text.trim());
+        }
+    }
+    return parts.join('\n');
+}
+/**
+ * GLM 4.6V 专用视觉提示裁剪：
+ * - 仅在 model 为 glm-4.6v 且存在携带图片的 user 消息时生效；
+ * - 丢弃原有 system 与历史对话，只保留一条新的 system + 一条 user；
+ * - system：要求模型以结构化 JSON 描述截图内容、标记（圈/箭头等）及其大致 bbox；
+ * - user：保留原始用户文字（如果有）+ 单一 image_url 块。
+ *
+ * 其他模型（包括 glm-4.7、Gemini 等）不受影响。
+ */
+export function applyGlmVisionPromptTransform(payload) {
+    const root = structuredClone(payload);
+    const modelRaw = root.model;
+    const model = typeof modelRaw === 'string' ? modelRaw.trim() : '';
+    if (!model.startsWith('glm-4.6v')) {
+        return root;
+    }
+    const messagesValue = root.messages;
+    if (!Array.isArray(messagesValue)) {
+        return root;
+    }
+    const messages = messagesValue.filter((msg) => isRecord(msg));
+    if (!messages.length) {
+        return root;
+    }
+    // 从末尾开始查找最近一条带图片的 user 消息。
+    let latestUserWithImage = null;
+    let imageUrl = null;
+    for (let i = messages.length - 1; i >= 0; i -= 1) {
+        const msg = messages[i];
+        const role = typeof msg.role === 'string' ? msg.role.toLowerCase() : '';
+        if (role !== 'user') {
+            continue;
+        }
+        const content = msg.content;
+        if (!Array.isArray(content)) {
+            continue;
+        }
+        for (const part of content) {
+            if (!isRecord(part))
+                continue;
+            const typeValue = typeof part.type === 'string' ? part.type.toLowerCase() : '';
+            if (typeValue === 'image' || typeValue === 'image_url' || typeValue === 'input_image') {
+                const candidateUrl = extractImageUrlFromPart(part);
+                if (candidateUrl) {
+                    latestUserWithImage = msg;
+                    imageUrl = candidateUrl;
+                    break;
+                }
+            }
+        }
+        if (latestUserWithImage && imageUrl) {
+            break;
+        }
+    }
+    if (!latestUserWithImage || !imageUrl) {
+        return root;
+    }
+    const systemContent = '你是 Codex 的截图理解子系统，专门用于分析 UI 截图和网页图片。请仅根据用户提供的图片，输出一个结构化的 JSON，用于后续自动化处理，不要输出额外解释性文字或自然语言说明。\n' +
+        '\n' +
+        '输出必须是**单个合法 JSON 对象**，不要包含 Markdown、代码块标记或多余文本。请严格遵循下面的结构（字段可以为空数组，但必须存在）：\n' +
+        '{\n' +
+        '  "summary": "用 1-3 句整体描述这张图片（例如页面类型、主要区域、核心信息）",\n' +
+        '  "marks": [\n' +
+        '    {\n' +
+        '      "type": "circle | arrow | underline | box | other",\n' +
+        '      "color": "red | green | blue | yellow | other",\n' +
+        '      "bbox": [x, y, width, height],\n' +
+        '      "description": "该标记所圈出/指向/强调的内容，包含相关文字或 UI 元素描述"\n' +
+        '    }\n' +
+        '  ],\n' +
+        '  "regions": [\n' +
+        '    {\n' +
+        '      "bbox": [x, y, width, height],\n' +
+        '      "description": "该区域的可见内容（控件/图标/布局，以及其中出现的所有清晰可辨的文字）",\n' +
+        '      "is_marked": true | false\n' +
+        '    }\n' +
+        '  ],\n' +
+        '  "metadata": {\n' +
+        '    "image_size_hint": "如果能推断出大致分辨率，请给出类似 1920x1080 的字符串；无法判断时用 null",\n' +
+        '    "screenshot": true\n' +
+        '  }\n' +
+        '}\n' +
+        '\n' +
+        '细则：\n' +
+        '1. 文字提取要求：\n' +
+        '   - 如果图片中存在清晰可辨的文字（包括标题、菜单、按钮、标签、提示信息、弹窗、错误信息等），必须在对应的 regions.description 中**完整抄写**这些文字，按自然阅读顺序组织，避免遗漏。\n' +
+        '   - 如果有多行文字，可以用换行符分隔，但仍放在同一个 description 字段中。\n' +
+        '   - 对确实无法看清的文字，用类似 "（模糊，无法辨认）" 标注即可；没有任何文字也视为正常情况，此时只需描述界面结构。\n' +
+        '2. 标记识别（marks）：\n' +
+        '   - 对所有明显的圈选、箭头、下划线、高亮框等标记，必须在 marks 中列出，每一项提供大致 bbox、颜色和简短说明，说明其强调或指向的内容。\n' +
+        '3. 区域划分（regions）：\n' +
+        '   - 将截图拆分为若干有意义的区域：如导航栏、侧边栏、主内容区、弹窗、对话框、表格、代码块、表单等。\n' +
+        '   - 每个区域的 description 中，既要描述布局/控件类型，也要包含该区域内的全部清晰文字内容。\n' +
+        '   - is_marked 为 true 表示该区域与某个标记（marks）相关或被标记强调。\n' +
+        '4. 坐标规范：所有 bbox 使用相对于当前图片的像素坐标，左上角为 (0,0)，width/height 为正数近似值。\n' +
+        '5. 无论图片内容如何，最终回答必须是合法 JSON，不能在 JSON 前后添加任何额外文本。';
+    const systemMessage = {
+        role: 'system',
+        content: systemContent
+    };
+    const originalUserText = collectUserTextFromMessage(latestUserWithImage);
+    const userBlocks = [];
+    if (originalUserText && originalUserText.trim().length) {
+        userBlocks.push({
+            type: 'text',
+            text: originalUserText.trim()
+        });
+    }
+    else {
+        userBlocks.push({
+            type: 'text',
+            text: '请按照上面的 JSON 结构，详细描述这张图片的内容和标记。'
+        });
+    }
+    userBlocks.push({
+        type: 'image_url',
+        image_url: {
+            url: imageUrl
+        }
+    });
+    const userMessage = {
+        role: 'user',
+        content: userBlocks
+    };
+    // 丢弃原有 messages，仅保留新的 system + user。
+    root.messages = [systemMessage, userMessage];
+    // 对于专用视觉模型，限制 max_tokens，避免过大的 completion 预算进一步触发上下文相关错误。
+    const maxTokensValue = root.max_tokens;
+    if (typeof maxTokensValue === 'number' && Number.isFinite(maxTokensValue)) {
+        // 将超大值收敛到一个相对安全的上限；真实上限由上游再校验。
+        root.max_tokens = Math.min(maxTokensValue, 4096);
+    }
+    return root;
+}

package/node_modules/@jsonstudio/llms/dist/conversion/compat/actions/glm-web-search.d.ts

@@ -0,0 +1,2 @@
+import type { JsonObject } from '../../hub/types/json.js';
+export declare function applyGlmWebSearchRequestTransform(payload: JsonObject): JsonObject;

package/node_modules/@jsonstudio/llms/dist/conversion/compat/actions/glm-web-search.js

@@ -0,0 +1,63 @@
+const isRecord = (value) => typeof value === 'object' && value !== null && !Array.isArray(value);
+const DEBUG_GLM_WEB_SEARCH = (process.env.ROUTECODEX_DEBUG_GLM_WEB_SEARCH || '').trim() === '1';
+export function applyGlmWebSearchRequestTransform(payload) {
+    const root = structuredClone(payload);
+    const webSearchRaw = root.web_search;
+    if (!isRecord(webSearchRaw)) {
+        return root;
+    }
+    const webSearch = webSearchRaw;
+    const queryValue = webSearch.query;
+    const recencyValue = webSearch.recency;
+    const countValue = webSearch.count;
+    const query = typeof queryValue === 'string' ? queryValue.trim() : '';
+    const recency = typeof recencyValue === 'string' ? recencyValue.trim() : undefined;
+    let count;
+    if (typeof countValue === 'number' && Number.isFinite(countValue)) {
+        const normalized = Math.floor(countValue);
+        if (normalized >= 1 && normalized <= 50) {
+            count = normalized;
+        }
+    }
+    if (!query) {
+        // No meaningful search query, drop the helper object and passthrough.
+        delete root.web_search;
+        return root;
+    }
+    const webSearchConfig = {
+        // 按 GLM 文档要求：search_engine 为必填，默认使用 search_std。
+        search_engine: 'search_std',
+        enable: true,
+        search_query: query,
+        // 返回搜索结果详情，便于我们在响应中提取摘要或调试。
+        search_result: true
+    };
+    if (recency) {
+        webSearchConfig.search_recency_filter = recency;
+    }
+    if (typeof count === 'number') {
+        webSearchConfig.count = count;
+    }
+    // 根据 OpenAPI：tools.anyOf[*] = WebSearchToolSchema[]，即一次只能选择一种工具类型。
+    // 在 web_search 路由下，我们只保留 WebSearchToolSchema，丢弃其它 function/retrieval/MCP 工具，
+    // 避免混用导致后端忽略 web_search 配置。
+    const baseTool = {
+        type: 'web_search',
+        web_search: webSearchConfig
+    };
+    root.tools = [baseTool];
+    delete root.web_search;
+    if (DEBUG_GLM_WEB_SEARCH) {
+        try {
+            // eslint-disable-next-line no-console
+            console.log('\x1b[38;5;27m[compat][glm_web_search_request] applied web_search transform ' +
+                `search_engine=${String(baseTool?.web_search?.search_engine ?? 'search_std')} ` +
+                `query=${JSON.stringify(query).slice(0, 200)} ` +
+                `count=${String(count ?? '')}\x1b[0m`);
+        }
+        catch {
+            // logging best-effort
+        }
+    }
+    return root;
+}

package/node_modules/@jsonstudio/llms/dist/conversion/compat/actions/universal-shape-filter.js

@@ -127,6 +127,17 @@ export class UniversalShapeFilter {
                 normalized.tool_call_id = msg.tool_call_id;
             }
         }
+        else if (Array.isArray(msg.content)) {
+            normalized.content = msg.content.map((part) => {
+                if (typeof part === 'string') {
+                    return part;
+                }
+                if (isRecord(part)) {
+                    return { ...part };
+                }
+                return part;
+            });
+        }
         else {
             normalized.content = (msg.content !== null && msg.content !== undefined) ? String(msg.content) : '';
         }