@jsonstudio/llms 0.6.230 → 0.6.375

package/dist/conversion/hub/semantic-mappers/gemini-mapper.js

@@ -171,6 +171,117 @@ function collectParameters(payload) {
     }
     return Object.keys(params).length ? params : undefined;
 }
+function appendChatContentToGeminiParts(message, targetParts) {
+    const content = message.content;
+    if (typeof content === 'string') {
+        const text = content.trim();
+        if (text.length) {
+            targetParts.push({ text });
+        }
+        return;
+    }
+    if (!Array.isArray(content)) {
+        return;
+    }
+    const items = content;
+    for (const block of items) {
+        if (block == null)
+            continue;
+        if (typeof block === 'string') {
+            const text = block.trim();
+            if (text.length) {
+                targetParts.push({ text });
+            }
+            continue;
+        }
+        if (typeof block !== 'object') {
+            const text = String(block);
+            if (text.trim().length) {
+                targetParts.push({ text: text.trim() });
+            }
+            continue;
+        }
+        const record = block;
+        const rawType = record.type;
+        const type = typeof rawType === 'string' ? rawType.toLowerCase() : '';
+        // Text-style blocks
+        if (!type || type === 'text') {
+            const textValue = typeof record.text === 'string'
+                ? record.text
+                : typeof record.content === 'string'
+                    ? record.content
+                    : '';
+            const text = textValue.trim();
+            if (text.length) {
+                targetParts.push({ text });
+            }
+            continue;
+        }
+        // Image-style blocks -> Gemini inlineData
+        if (type === 'image' || type === 'image_url') {
+            // Prefer OpenAI-style image_url.url, but also accept uri/url/data.
+            let url;
+            const imageUrlRaw = record.image_url;
+            if (typeof imageUrlRaw === 'string') {
+                url = imageUrlRaw;
+            }
+            else if (imageUrlRaw && typeof imageUrlRaw === 'object' && typeof imageUrlRaw.url === 'string') {
+                url = imageUrlRaw.url;
+            }
+            else if (typeof record.uri === 'string') {
+                url = record.uri;
+            }
+            else if (typeof record.url === 'string') {
+                url = record.url;
+            }
+            else if (typeof record.data === 'string') {
+                url = record.data;
+            }
+            const trimmed = (url ?? '').trim();
+            if (!trimmed.length) {
+                // Fallback: at least emit a textual marker so内容不会完全丢失
+                targetParts.push({ text: '[image]' });
+                continue;
+            }
+            let mimeType;
+            let data;
+            // data:URL → inlineData { mimeType, data }
+            if (trimmed.startsWith('data:')) {
+                const match = /^data:([^;,]+)?(?:;base64)?,(.*)$/s.exec(trimmed);
+                if (match) {
+                    mimeType = (match[1] || '').trim() || undefined;
+                    data = match[2] || '';
+                }
+            }
+            if (data && data.trim().length) {
+                const inline = {
+                    inlineData: {
+                        data: data.trim()
+                    }
+                };
+                if (mimeType && mimeType.length) {
+                    inline.inlineData.mimeType = mimeType;
+                }
+                targetParts.push(inline);
+            }
+            else {
+                // 非 data: URL 暂时作为文本 URL 传递，保持语义可见
+                targetParts.push({ text: trimmed });
+            }
+            continue;
+        }
+        // 默认：回退为文本 JSON 表示，避免静默丢失内容
+        try {
+            const jsonText = JSON.stringify(record);
+            if (jsonText.trim().length) {
+                targetParts.push({ text: jsonText });
+            }
+        }
+        catch {
+            // ignore malformed block
+        }
+    }
+}
 function buildGeminiRequestFromChat(chat, metadata) {
     const contents = [];
     const emittedToolOutputs = new Set();
@@ -191,9 +302,7 @@ function buildGeminiRequestFromChat(chat, metadata) {
             role: mapChatRoleToGemini(message.role),
             parts: []
         };
-        if (typeof message.content === 'string' && message.content.trim().length) {
-            entry.parts.push({ text: message.content });
-        }
+        appendChatContentToGeminiParts(message, entry.parts);
         const toolCalls = Array.isArray(message.tool_calls) ? message.tool_calls : [];
         for (const tc of toolCalls) {
             if (!tc || typeof tc !== 'object')
@@ -335,7 +444,18 @@ function safeParseJson(value) {
     }
 }
 function ensureFunctionResponsePayload(value) {
+    // Gemini function_response.response 字段在 CloudCode/Gemini CLI 协议里对应的是
+    // protobuf Struct（JSON object），而不是顶层数组。
+    // 这里做一层规范化：
+    // - 对象：直接透传；
+    // - 数组：包一层 { result: [...] } 避免把数组作为 Struct 根节点；
+    // - 原始值：包一层 { result: value }，并把 undefined 映射为 null。
     if (value && typeof value === 'object') {
+        if (Array.isArray(value)) {
+            return {
+                result: value
+            };
+        }
         return value;
     }
     return {

package/dist/conversion/hub/semantic-mappers/responses-mapper.js

@@ -148,6 +148,20 @@ function serializeSystemContent(message) {
     }
     return undefined;
 }
+function mergeMetadata(a, b) {
+    if (!a && !b) {
+        return undefined;
+    }
+    if (!a && b) {
+        return jsonClone(b);
+    }
+    if (a && !b) {
+        return jsonClone(a);
+    }
+    const left = jsonClone(a);
+    const right = jsonClone(b);
+    return { ...left, ...right };
+}
 export class ResponsesSemanticMapper {
     async toChat(format, ctx) {
         const payload = format.payload || {};
@@ -209,13 +223,15 @@ export class ResponsesSemanticMapper {
             .map(message => serializeSystemContent(message))
             .filter((content) => typeof content === 'string' && content.length > 0);
         const capturedContext = chat.metadata?.responsesContext;
+        const envelopeMetadata = chat.metadata && isJsonObject(chat.metadata) ? chat.metadata : undefined;
         const responsesContext = isJsonObject(capturedContext)
             ? {
                 ...capturedContext,
+                metadata: mergeMetadata(capturedContext.metadata, envelopeMetadata),
                 originalSystemMessages
             }
             : {
-                metadata: (chat.metadata && isJsonObject(chat.metadata) ? chat.metadata : undefined),
+                metadata: envelopeMetadata,
                 originalSystemMessages
             };
         const responsesResult = buildResponsesRequestFromChat(requestShape, responsesContext);

package/dist/conversion/hub/standardized-bridge.js

@@ -60,6 +60,20 @@ export function standardizedToChatEnvelope(request, options) {
     const metadata = {
         context: adapterContext
     };
+    const sourceMeta = (request.metadata && typeof request.metadata === 'object'
+        ? request.metadata
+        : undefined);
+    if (sourceMeta) {
+        if (sourceMeta.webSearch && typeof sourceMeta.webSearch === 'object') {
+            metadata.webSearch = sourceMeta.webSearch;
+        }
+        if (sourceMeta.forceWebSearch === true) {
+            metadata.forceWebSearch = true;
+        }
+        if (sourceMeta.forceVision === true) {
+            metadata.forceVision = true;
+        }
+    }
     if (typeof adapterContext.toolCallIdStyle === 'string' && adapterContext.toolCallIdStyle.length) {
         metadata.toolCallIdStyle = adapterContext.toolCallIdStyle;
     }

package/dist/conversion/responses/responses-openai-bridge.js

@@ -190,17 +190,50 @@ function mergeResponsesTools(originalTools, fromChat) {
 export function buildResponsesRequestFromChat(payload, ctx, extras) {
     const chat = unwrapData(payload);
     const out = {};
+    const forceWebSearch = !!ctx &&
+        isObject(ctx.metadata) &&
+        isObject(ctx.metadata.webSearch) &&
+        ctx.metadata.webSearch.force === true;
     // 基本字段
     out.model = chat.model;
     // tools: 反向映射为 ResponsesToolDefinition 形状
-    const responsesToolsFromChat = mapChatToolsToBridge(chat.tools, {
+    const chatTools = Array.isArray(chat.tools) ? chat.tools : [];
+    // 对于 openai-responses upstream，内建 web_search 由官方服务器处理。
+    // Chat 侧注入的 server-side web_search 函数（带 engine/query/recency/count）
+    // 仅用于非 Responses provider 的 server-tool 回环；在这里构造真正的
+    // `/v1/responses` 请求时，需要：
+    //   1) 不再把函数版 web_search 透传上游；
+    //   2) 若检测到 Chat 侧启用了 web_search 且原始请求中没有 builtin web_search，
+    //      则补一个 `{ type: "web_search" }` 内建工具给 OpenAI Responses。
+    const hasServerSideWebSearch = !forceWebSearch && chatTools.some((tool) => {
+        const fn = tool && typeof tool === 'object' ? tool.function : undefined;
+        const name = typeof fn?.name === 'string' ? fn.name.trim().toLowerCase() : '';
+        return name === 'web_search';
+    });
+    const toolsForBridge = hasServerSideWebSearch
+        ? chatTools.filter((tool) => {
+            const fn = tool && typeof tool === 'object' ? tool.function : undefined;
+            const name = typeof fn?.name === 'string' ? fn.name.trim().toLowerCase() : '';
+            return name !== 'web_search';
+        })
+        : chatTools;
+    const responsesToolsFromChat = mapChatToolsToBridge(toolsForBridge, {
         sanitizeName: sanitizeResponsesFunctionName
     });
     // Prefer Chat‑normalized tools, but if the original Responses payload carried
     // non‑function tools (such as builtin `web_search`), merge them back so that
     // upstream `/v1/responses` providers see their original tool definitions.
     const originalTools = Array.isArray(ctx?.toolsRaw) ? ctx.toolsRaw : undefined;
-    const mergedTools = mergeResponsesTools(originalTools, responsesToolsFromChat);
+    let mergedTools = mergeResponsesTools(originalTools, responsesToolsFromChat);
+    if (hasServerSideWebSearch) {
+        const normalizeType = (value) => typeof value === 'string' ? value.trim().toLowerCase() : '';
+        const hasBuiltinWebSearch = (mergedTools && mergedTools.some((tool) => normalizeType(tool.type) === 'web_search')) ||
+            (originalTools && originalTools.some((tool) => normalizeType(tool.type) === 'web_search'));
+        if (!hasBuiltinWebSearch) {
+            const injected = { type: 'web_search' };
+            mergedTools = mergedTools ? [...mergedTools, injected] : [injected];
+        }
+    }
     if (mergedTools?.length) {
         out.tools = mergedTools;
     }

package/dist/conversion/shared/anthropic-message-utils.js

@@ -266,6 +266,7 @@ export function buildOpenAIChatFromAnthropic(payload) {
             continue;
         }
         const textParts = [];
+        const imageBlocks = [];
         const toolCalls = [];
         const reasoningParts = [];
         const toolResults = [];
@@ -284,6 +285,29 @@ export function buildOpenAIChatFromAnthropic(payload) {
                     reasoningParts.push(thinkingText);
                 }
             }
+            else if (t === 'image') {
+                const source = block.source;
+                if (source && typeof source === 'object') {
+                    const s = source;
+                    const srcType = typeof s.type === 'string' ? s.type.toLowerCase() : '';
+                    let url;
+                    if (srcType === 'url' && typeof s.url === 'string') {
+                        url = s.url;
+                    }
+                    else if (srcType === 'base64' && typeof s.data === 'string') {
+                        const mediaType = typeof s.media_type === 'string' && s.media_type.trim().length
+                            ? s.media_type.trim()
+                            : 'image/png';
+                        url = `data:${mediaType};base64,${s.data}`;
+                    }
+                    if (url && url.trim().length) {
+                        imageBlocks.push({
+                            type: 'image_url',
+                            image_url: { url: url.trim() }
+                        });
+                    }
+                }
+            }
             else if (t === 'tool_use') {
                 const name = requireTrimmedString(block.name, 'tool_use.name');
                 const id = requireTrimmedString(block.id, 'tool_use.id');
@@ -310,10 +334,22 @@ export function buildOpenAIChatFromAnthropic(payload) {
         }
         const hasText = typeof normalized.contentText === 'string' && normalized.contentText.length > 0;
         const hasReasoning = mergedReasoning.length > 0;
-        if (hasText || hasRawText || toolCalls.length > 0 || hasReasoning) {
+        if (hasText || hasRawText || toolCalls.length > 0 || hasReasoning || imageBlocks.length > 0) {
+            let contentNode = (hasText ? normalized.contentText : undefined) ?? combinedText ?? '';
+            if (imageBlocks.length > 0) {
+                const blocks = [];
+                const textPayload = (hasText ? normalized.contentText : undefined) ?? combinedText ?? '';
+                if (typeof textPayload === 'string' && textPayload.trim().length) {
+                    blocks.push({ type: 'text', text: textPayload.trim() });
+                }
+                for (const img of imageBlocks) {
+                    blocks.push(jsonClone(img));
+                }
+                contentNode = blocks;
+            }
             const msg = {
                 role,
-                content: (hasText ? normalized.contentText : undefined) ?? combinedText ?? ''
+                content: contentNode
             };
             if (toolCalls.length)
                 msg.tool_calls = toolCalls;
@@ -690,7 +726,8 @@ export function buildAnthropicRequestFromOpenAIChat(chatReq) {
             targetShape = mirrorShapes[mirrorIndex];
             mirrorIndex += 1;
         }
-        const text = collectText(m.content).trim();
+        const contentNode = m.content;
+        const text = collectText(contentNode).trim();
         if (role === 'system') {
             if (!text) {
                 throw new Error('Anthropic bridge constraint violated: Chat system message must contain text');
@@ -715,6 +752,58 @@ export function buildAnthropicRequestFromOpenAIChat(chatReq) {
             continue;
         }
         const blocks = [];
+        if (Array.isArray(contentNode)) {
+            // Preserve or synthesize image blocks where possible, and fall back to text for the rest.
+            for (const entry of contentNode) {
+                if (!entry || typeof entry !== 'object')
+                    continue;
+                const node = entry;
+                const t = typeof node.type === 'string' ? node.type.toLowerCase() : '';
+                if (t === 'image' && node.source && typeof node.source === 'object') {
+                    // Pass-through Anthropic image block as-is.
+                    blocks.push({
+                        type: 'image',
+                        source: jsonClone(node.source)
+                    });
+                    continue;
+                }
+                if (t === 'image_url') {
+                    let url = '';
+                    const imageUrl = node.image_url;
+                    if (typeof imageUrl === 'string') {
+                        url = imageUrl;
+                    }
+                    else if (imageUrl && typeof imageUrl === 'object' && typeof imageUrl.url === 'string') {
+                        url = imageUrl.url;
+                    }
+                    const trimmed = url.trim();
+                    if (!trimmed.length)
+                        continue;
+                    const source = {};
+                    if (trimmed.startsWith('data:')) {
+                        const match = /^data:([^;,]+)?(?:;base64)?,(.*)$/s.exec(trimmed);
+                        if (match) {
+                            const mediaType = (match[1] || '').trim() || 'image/png';
+                            source.type = 'base64';
+                            source.media_type = mediaType;
+                            source.data = match[2] || '';
+                        }
+                        else {
+                            source.type = 'url';
+                            source.url = trimmed;
+                        }
+                    }
+                    else {
+                        source.type = 'url';
+                        source.url = trimmed;
+                    }
+                    blocks.push({
+                        type: 'image',
+                        source
+                    });
+                }
+            }
+        }
         if (text) {
             blocks.push({ type: 'text', text });
         }

package/dist/conversion/shared/bridge-message-utils.js

@@ -66,6 +66,59 @@ function collectText(value) {
     }
     return '';
 }
+function extractImageBlocksFromContent(content) {
+    const images = [];
+    const visit = (value) => {
+        if (!value)
+            return;
+        if (Array.isArray(value)) {
+            for (const entry of value)
+                visit(entry);
+            return;
+        }
+        if (typeof value !== 'object') {
+            return;
+        }
+        const record = value;
+        const typeValue = typeof record.type === 'string' ? record.type.toLowerCase() : '';
+        if (typeValue === 'image' || typeValue === 'image_url' || typeValue === 'input_image') {
+            let url = '';
+            const imageUrl = record.image_url;
+            if (typeof imageUrl === 'string') {
+                url = imageUrl;
+            }
+            else if (imageUrl && typeof imageUrl === 'object' && typeof imageUrl.url === 'string') {
+                url = imageUrl.url;
+            }
+            else if (typeof record.url === 'string') {
+                url = record.url;
+            }
+            else if (typeof record.uri === 'string') {
+                url = record.uri;
+            }
+            else if (typeof record.data === 'string') {
+                url = record.data;
+            }
+            const trimmed = url.trim();
+            if (trimmed.length) {
+                let detail;
+                if (imageUrl && typeof imageUrl === 'object' && typeof imageUrl.detail === 'string') {
+                    detail = imageUrl.detail.trim() || undefined;
+                }
+                else if (typeof record.detail === 'string') {
+                    detail = record.detail.trim() || undefined;
+                }
+                images.push({ url: trimmed, detail });
+            }
+            return;
+        }
+        if (Array.isArray(record.content)) {
+            visit(record.content);
+        }
+    };
+    visit(content);
+    return images;
+}
 function extractUserTextFromEntry(entry) {
     if (!entry || typeof entry !== 'object')
         return '';
@@ -94,6 +147,7 @@ export function convertMessagesToBridgeInput(options) {
         const role = coerceBridgeRole(m.role || 'user');
         const content = m.content;
         const collectedText = collectText(content);
+        const imageBlocks = extractImageBlocksFromContent(content);
         const text = role === 'system' ? collectedText : collectedText.trim();
         if (role === 'system') {
             if (collectedText && collectedText.length) {
@@ -170,13 +224,29 @@ export function convertMessagesToBridgeInput(options) {
             }
             continue;
         }
-        if (typeof text === 'string') {
+        if (typeof text === 'string' || imageBlocks.length) {
             const tRole = role === 'assistant' ? 'output_text' : 'input_text';
-            const entry = {
-                role,
-                content: [{ type: tRole, text }]
-            };
-            input.push(entry);
+            const blocks = [];
+            if (typeof text === 'string' && text.length) {
+                blocks.push({ type: tRole, text });
+            }
+            for (const img of imageBlocks) {
+                const block = {
+                    type: 'input_image',
+                    image_url: img.url
+                };
+                if (img.detail) {
+                    block.detail = img.detail;
+                }
+                blocks.push(block);
+            }
+            if (blocks.length) {
+                const entry = {
+                    role,
+                    content: blocks
+                };
+                input.push(entry);
+            }
             if (role === 'user') {
                 const trimmed = typeof text === 'string' ? text.trim() : '';
                 if (trimmed.length) {
@@ -260,6 +330,7 @@ function processMessageBlocks(blocks, normalizeFunctionName, tools, toolNameById
     const toolMessages = [];
     let currentLastCall = lastToolCallId;
     const reasoningSegments = [];
+    const images = [];
     for (const block of blocks) {
         if (!block || typeof block !== 'object')
             continue;
@@ -282,6 +353,18 @@ function processMessageBlocks(blocks, normalizeFunctionName, tools, toolNameById
                 toolMessages.push(tm);
             currentLastCall = nested.lastCallId;
             reasoningSegments.push(...nested.reasoningSegments);
+            if (nested.images.length)
+                images.push(...nested.images);
+            continue;
+        }
+        if (type === 'input_image') {
+            const url = typeof block.image_url === 'string' ? block.image_url.trim() : '';
+            if (url) {
+                const detail = typeof block.detail === 'string' && block.detail.trim()
+                    ? block.detail.trim()
+                    : undefined;
+                images.push({ url, detail });
+            }
             continue;
         }
         if (type === 'function_call') {
@@ -344,7 +427,7 @@ function processMessageBlocks(blocks, normalizeFunctionName, tools, toolNameById
         }
     }
     const text = textParts.length ? textParts.join('\n').trim() : null;
-    return { text, toolCalls, toolMessages, lastCallId: currentLastCall, reasoningSegments };
+    return { text, images, toolCalls, toolMessages, lastCallId: currentLastCall, reasoningSegments };
 }
 export function convertBridgeInputToChatMessages(options) {
     const { input, tools, normalizeFunctionName, toolResultFallbackText } = options;
@@ -470,7 +553,29 @@ export function convertBridgeInputToChatMessages(options) {
             for (const msg of nested.toolMessages)
                 messages.push(msg);
             const normalizedRole = coerceBridgeRole((explicit.role ?? entry.role) || 'user');
-            if (typeof nested.text === 'string') {
+            if (nested.images.length) {
+                const contentBlocks = [];
+                if (typeof nested.text === 'string' && nested.text.trim().length) {
+                    contentBlocks.push({ type: 'text', text: nested.text });
+                }
+                for (const img of nested.images) {
+                    const imgBlock = { type: 'image_url', image_url: { url: img.url } };
+                    if (img.detail) {
+                        imgBlock.image_url.detail = img.detail;
+                    }
+                    contentBlocks.push(imgBlock);
+                }
+                const msg = {
+                    role: normalizedRole,
+                    content: contentBlocks
+                };
+                const combinedReasoning = combineReasoningSegments(consumeEntryReasoning(), nested.reasoningSegments);
+                if (combinedReasoning.length) {
+                    msg.reasoning_content = combinedReasoning.join('\n');
+                }
+                messages.push(msg);
+            }
+            else if (typeof nested.text === 'string') {
                 pushNormalizedChatMessage(messages, normalizedRole, nested.text, {
                     reasoningSegments: combineReasoningSegments(consumeEntryReasoning(), nested.reasoningSegments)
                 });
@@ -491,9 +596,31 @@ export function convertBridgeInputToChatMessages(options) {
             for (const msg of nested.toolMessages)
                 messages.push(msg);
             const normalizedRole = coerceBridgeRole(entry.role || 'user');
-            if (typeof nested.text === 'string') {
+            if (nested.images.length) {
+                const contentBlocks = [];
+                if (typeof nested.text === 'string' && nested.text.trim().length) {
+                    contentBlocks.push({ type: 'text', text: nested.text });
+                }
+                for (const img of nested.images) {
+                    const imgBlock = { type: 'image_url', image_url: { url: img.url } };
+                    if (img.detail) {
+                        imgBlock.image_url.detail = img.detail;
+                    }
+                    contentBlocks.push(imgBlock);
+                }
+                const msg = {
+                    role: normalizedRole,
+                    content: contentBlocks
+                };
+                const combinedReasoning = combineReasoningSegments(consumeEntryReasoning(), nested.reasoningSegments);
+                if (combinedReasoning.length) {
+                    msg.reasoning_content = combinedReasoning.join('\n');
+                }
+                messages.push(msg);
+            }
+            else if (typeof nested.text === 'string') {
                 pushNormalizedChatMessage(messages, normalizedRole, nested.text, {
-                    reasoningSegments: combineReasoningSegments(consumeEntryReasoning(), nested.reasoningSegments)
+                    reasoningSegments: consumeEntryReasoning()
                 });
             }
             lastToolCallId = nested.lastCallId;

package/dist/conversion/shared/responses-output-builder.js

@@ -1,6 +1,41 @@
 import { normalizeFunctionCallId } from './bridge-id-utils.js';
 import { normalizeContentPart } from './output-content-normalizer.js';
 import { expandResponsesMessageItem } from '../../sse/shared/responses-output-normalizer.js';
+function buildToolOutputIndex(response) {
+    const ids = new Set();
+    try {
+        const primary = Array.isArray(response.tool_outputs)
+            ? response.tool_outputs
+            : [];
+        for (const entry of primary) {
+            if (!entry || typeof entry !== 'object')
+                continue;
+            const raw = entry.tool_call_id ||
+                entry.call_id ||
+                entry.id;
+            if (typeof raw === 'string' && raw.trim().length) {
+                const trimmed = raw.trim();
+                // 记录原始 ID（例如 OpenAI 的 toolu_ 前缀），以兼容直接使用
+                // tool_call_id 的客户端；同时记录归一化后的 fc_ 形式，保证与
+                // buildFunctionCallOutput 中 normalizeFunctionCallId 的结果对齐。
+                ids.add(trimmed);
+                try {
+                    const normalized = normalizeFunctionCallId({ callId: trimmed, fallback: trimmed });
+                    if (normalized && normalized !== trimmed) {
+                        ids.add(normalized);
+                    }
+                }
+                catch {
+                    // 归一化失败不应影响主流程
+                }
+            }
+        }
+    }
+    catch {
+        // best-effort: 不因索引构建失败影响主流程
+    }
+    return ids;
+}
 function appendReasoningSegments(target, raw) {
     if (typeof raw !== 'string' || !raw.length) {
         return;
@@ -91,7 +126,13 @@ export function buildResponsesOutputFromChat(options) {
     const usage = normalizeUsage(response.usage);
     const outputTextMeta = response?.__responses_output_text_meta;
     const outputText = resolveOutputText(convertedContent, outputTextMeta);
-    const hasNormalizedToolCalls = normalizedToolCalls.length > 0;
+    // 如果顶层 tool_outputs 已经为所有 tool_calls 提供了结果，说明这些函数调用
+    // 已在服务端（例如 server-side web_search）完成，不应再对客户端暴露
+    // required_action/submit_tool_outputs。此时只需返回 completed 状态即可，避免
+    // 再触发一轮工具回合。
+    const executedIds = buildToolOutputIndex(response);
+    const pendingToolCalls = normalizedToolCalls.filter((entry) => !executedIds.has(entry.id));
+    const hasNormalizedToolCalls = pendingToolCalls.length > 0;
     if (hasNormalizedToolCalls) {
         for (const item of outputItems) {
             if (item.type === 'message') {
@@ -100,7 +141,7 @@ export function buildResponsesOutputFromChat(options) {
         }
     }
     const requiredAction = hasNormalizedToolCalls
-        ? buildRequiredActionFromNormalized(normalizedToolCalls)
+        ? buildRequiredActionFromNormalized(pendingToolCalls)
         : undefined;
     const status = hasNormalizedToolCalls ? 'requires_action' : 'completed';
     return {

package/dist/conversion/shared/tool-filter-pipeline.js

@@ -111,6 +111,7 @@ function applyLocalToolGovernance(chatRequest, rawPayload) {
     }
     const hasImageHint = detectImageHint(messages, rawPayload);
     if (hasImageHint) {
+        // 有图片线索时不干预工具列表，保持由上游（Codex 等）决定 view_image 的暴露与使用。
         return chatRequest;
     }
     const filteredTools = tools.filter((tool) => {