@jsonstudio/llms 0.6.567 → 0.6.586

package/dist/conversion/codecs/gemini-openai-codec.js

@@ -360,7 +360,9 @@ export function buildOpenAIChatFromGeminiResponse(payload) {
     if (Number.isFinite(totalTokens))
         usage.total_tokens = totalTokens;
     const combinedText = textParts.join('\n');
-    const normalized = combinedText.length ? normalizeChatMessageContent(combinedText) : { contentText: undefined, reasoningText: undefined };
+    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
@@ -525,16 +527,43 @@ export function buildGeminiFromOpenAIChat(chatResp) {
         let argsStruct;
         const rawArgs = fn.arguments;
         if (typeof rawArgs === 'string') {
+            const trimmed = rawArgs.trim();
+            if (trimmed.startsWith('{')) {
+                try {
+                    const parsed = JSON.parse(rawArgs);
+                    if (isObject(parsed)) {
+                        argsStruct = parsed;
+                    }
+                    else {
+                        argsStruct = { _raw: rawArgs };
+                    }
+                }
+                catch {
+                    argsStruct = { _raw: rawArgs };
+                }
+            }
+            else {
+                argsStruct = { _raw: rawArgs };
+            }
+        }
+        else if (isObject(rawArgs)) {
+            argsStruct = rawArgs;
+        }
+        else if (Array.isArray(rawArgs)) {
             try {
-                argsStruct = JSON.parse(rawArgs);
+                argsStruct = { _raw: JSON.stringify(rawArgs) };
             }
             catch {
-                argsStruct = { _raw: rawArgs };
+                argsStruct = { _raw: String(rawArgs) };
             }
         }
+        else if (rawArgs != null) {
+            argsStruct = { _raw: String(rawArgs) };
+        }
         else {
-            argsStruct = rawArgs ?? {};
+            argsStruct = {};
         }
+        // Gemini request/response wire uses `args` for functionCall payload.
         const functionCall = { name, args: argsStruct };
         const id = typeof tc.id === 'string' ? String(tc.id) : undefined;
         if (id)

package/dist/conversion/codecs/openai-openai-codec.js

@@ -89,8 +89,9 @@ export class OpenAIOpenAIConversionCodec {
         // Response-side filters (idempotent w.r.t existing logic)
         engine.registerFilter(new ResponseToolTextCanonicalizeFilter()); // response_pre
         try {
-            const { ResponseToolArgumentsToonDecodeFilter } = await import('../../filters/index.js');
+            const { ResponseToolArgumentsToonDecodeFilter, ResponseApplyPatchToonDecodeFilter } = await import('../../filters/index.js');
             engine.registerFilter(new ResponseToolArgumentsToonDecodeFilter()); // response_pre, runs before stringify
+            engine.registerFilter(new ResponseApplyPatchToonDecodeFilter()); // response_pre, runs before stringify
         }
         catch { /* optional */ }
         engine.registerFilter(new ResponseToolArgumentsStringifyFilter()); // response_post

package/dist/conversion/codecs/responses-openai-codec.js

@@ -147,11 +147,12 @@ export class ResponsesOpenAIConversionCodec {
             debug: { emit: () => { } }
         };
         const engine = new FilterEngine();
-        // Response-side filters：文本标准化 → TOON decode（可选）→ shell/basics → finish_reason 不变式
+        // Response-side filters：文本标准化 → TOON decode（可选）→ apply_patch 结构化补丁规范化 → shell/basics → finish_reason 不变式
         engine.registerFilter(new ResponseToolTextCanonicalizeFilter()); // response_pre
         try {
-            const { ResponseToolArgumentsToonDecodeFilter } = await import('../../filters/index.js');
+            const { ResponseToolArgumentsToonDecodeFilter, ResponseApplyPatchToonDecodeFilter } = await import('../../filters/index.js');
             engine.registerFilter(new ResponseToolArgumentsToonDecodeFilter()); // response_pre, runs before stringify
+            engine.registerFilter(new ResponseApplyPatchToonDecodeFilter()); // response_pre, runs before stringify
         }
         catch { /* optional */ }
         engine.registerFilter(new ResponseToolArgumentsStringifyFilter()); // response_post

package/dist/conversion/compat/actions/glm-history-image-trim.d.ts

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

package/dist/conversion/compat/actions/glm-history-image-trim.js

@@ -0,0 +1,88 @@
+const isRecord = (value) => typeof value === 'object' && value !== null && !Array.isArray(value);
+function shouldDropInlineImagePart(part) {
+    const rawType = typeof part.type === 'string' ? part.type.toLowerCase() : '';
+    if (rawType !== 'image' && rawType !== 'image_url' && rawType !== 'input_image') {
+        return false;
+    }
+    const imageUrlBlock = isRecord(part.image_url)
+        ? part.image_url
+        : part;
+    const urlRaw = typeof imageUrlBlock.url === 'string'
+        ? imageUrlBlock.url
+        : typeof imageUrlBlock.data === 'string'
+            ? imageUrlBlock.data
+            : '';
+    const url = urlRaw.trim();
+    if (!url) {
+        return false;
+    }
+    // GLM 4.7 在历史消息中携带 data:image/base64 时会返回 1210，
+    // 因此仅在历史中丢弃这类 inline image 片段。
+    return url.startsWith('data:image');
+}
+export function applyGlmHistoryImageTrim(payload) {
+    const root = structuredClone(payload);
+    const modelRaw = root.model;
+    const modelId = typeof modelRaw === 'string' ? modelRaw.trim().toLowerCase() : '';
+    if (!modelId || !modelId.startsWith('glm-4.7')) {
+        return root;
+    }
+    const messagesValue = root.messages;
+    if (!Array.isArray(messagesValue)) {
+        return root;
+    }
+    const messages = messagesValue.filter(msg => isRecord(msg));
+    if (!messages.length) {
+        return root;
+    }
+    // 仅在历史消息中进行裁剪：保留最后一条 user 完整内容。
+    let lastUserIdx = -1;
+    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') {
+            lastUserIdx = i;
+            break;
+        }
+    }
+    if (lastUserIdx === -1) {
+        return root;
+    }
+    const nextMessages = [];
+    for (let i = 0; i < messages.length; i += 1) {
+        const msg = messages[i];
+        const role = typeof msg.role === 'string' ? msg.role.toLowerCase() : '';
+        if (i < lastUserIdx && role === 'user') {
+            const contentValue = msg.content;
+            if (!Array.isArray(contentValue)) {
+                nextMessages.push(msg);
+                continue;
+            }
+            const newContent = [];
+            for (const part of contentValue) {
+                if (!isRecord(part)) {
+                    newContent.push(part);
+                    continue;
+                }
+                if (shouldDropInlineImagePart(part)) {
+                    // 丢弃历史中的 data:image/* 片段
+                    // eslint-disable-next-line no-continue
+                    continue;
+                }
+                newContent.push(part);
+            }
+            if (!newContent.length) {
+                // 历史消息只剩下 inline image 时，直接移除整条消息。
+                // 避免向 GLM 发送纯图片历史导致 1210。
+                // eslint-disable-next-line no-continue
+                continue;
+            }
+            const cloned = { ...msg, content: newContent };
+            nextMessages.push(cloned);
+            continue;
+        }
+        nextMessages.push(msg);
+    }
+    root.messages = nextMessages;
+    return root;
+}

package/dist/conversion/hub/pipeline/hub-pipeline.d.ts

@@ -1,10 +1,15 @@
 import { Readable } from 'node:stream';
 import type { StandardizedRequest, ProcessedRequest } from '../types/standardized.js';
 import type { JsonObject } from '../types/json.js';
-import type { VirtualRouterConfig, RoutingDecision, RoutingDiagnostics, TargetMetadata } from '../../../router/virtual-router/types.js';
+import type { VirtualRouterConfig, RoutingDecision, RoutingDiagnostics, TargetMetadata, VirtualRouterHealthStore } from '../../../router/virtual-router/types.js';
 import { type HubProcessNodeResult } from '../process/chat-process.js';
 export interface HubPipelineConfig {
     virtualRouter: VirtualRouterConfig;
+    /**
+     * 可选：供 VirtualRouterEngine 使用的健康状态持久化存储。
+     * 当提供时，VirtualRouterEngine 将在初始化时恢复上一次快照，并在 cooldown/熔断变化时调用 persistSnapshot。
+     */
+    healthStore?: VirtualRouterHealthStore;
 }
 export interface HubPipelineRequestMetadata extends Record<string, unknown> {
     entryEndpoint?: string;
@@ -56,7 +61,6 @@ export declare class HubPipeline {
     private buildAdapterContext;
     private maybeCreateStageRecorder;
     private asJsonObject;
-    private pickRawRequestBody;
     private normalizeRequest;
     private convertProcessNodeResult;
     private materializePayload;

package/dist/conversion/hub/pipeline/hub-pipeline.js

@@ -22,13 +22,16 @@ import { runReqOutboundStage1SemanticMap } from './stages/req_outbound/req_outbo
 import { runReqOutboundStage2FormatBuild } from './stages/req_outbound/req_outbound_stage2_format_build/index.js';
 import { runReqOutboundStage3Compat } from './stages/req_outbound/req_outbound_stage3_compat/index.js';
 import { extractSessionIdentifiersFromMetadata } from './session-identifiers.js';
+import { computeRequestTokens } from '../../../router/virtual-router/token-estimator.js';
 export class HubPipeline {
     routerEngine;
     config;
     unsubscribeProviderErrors;
     constructor(config) {
         this.config = config;
-        this.routerEngine = new VirtualRouterEngine();
+        this.routerEngine = new VirtualRouterEngine({
+            healthStore: config.healthStore
+        });
         this.routerEngine.initialize(config.virtualRouter);
         try {
             this.unsubscribeProviderErrors = providerErrorCenter.subscribe((event) => {
@@ -117,6 +120,18 @@ export class HubPipeline {
             }
         }
         const workingRequest = processedRequest ?? standardizedRequest;
+        // 使用与 VirtualRouter 一致的 tiktoken 计数逻辑，对标准化请求进行一次
+        // 上下文 token 估算，供后续 usage 归一化与统计使用。
+        try {
+            const estimatedTokens = computeRequestTokens(workingRequest, '');
+            if (typeof estimatedTokens === 'number' && Number.isFinite(estimatedTokens) && estimatedTokens > 0) {
+                normalized.metadata = normalized.metadata || {};
+                normalized.metadata.estimatedInputTokens = estimatedTokens;
+            }
+        }
+        catch {
+            // 估算失败不应影响主流程
+        }
         const normalizedMeta = normalized.metadata;
         const responsesResume = normalizedMeta && typeof normalizedMeta.responsesResume === 'object'
             ? normalizedMeta.responsesResume
@@ -180,81 +195,60 @@ export class HubPipeline {
         const outboundEndpoint = resolveEndpointForProviderProtocol(outboundAdapterContext.providerProtocol);
         const outboundRecorder = this.maybeCreateStageRecorder(outboundAdapterContext, outboundEndpoint);
         const outboundStart = Date.now();
-        const isResponsesSubmit = normalized.entryEndpoint === '/v1/responses.submit_tool_outputs' &&
-            outboundProtocol === 'openai-responses';
         let providerPayload;
-        if (isResponsesSubmit) {
-            providerPayload =
-                this.pickRawRequestBody(normalized.metadata) ?? normalized.payload;
-            const outboundEnd = Date.now();
-            nodeResults.push({
-                id: 'req_outbound',
-                success: true,
-                metadata: {
-                    node: 'req_outbound',
-                    executionTime: outboundEnd - outboundStart,
-                    startTime: outboundStart,
-                    endTime: outboundEnd,
-                    dataProcessed: {
-                        messages: 0,
-                        tools: 0
-                    },
-                    bypass: 'responses-submit-passthrough'
-                }
-            });
-        }
-        else {
-            const outboundStage1 = await runReqOutboundStage1SemanticMap({
-                request: workingRequest,
-                adapterContext: outboundAdapterContext,
-                semanticMapper: outboundSemanticMapper,
-                contextSnapshot: outboundContextSnapshot,
-                contextMetadataKey: outboundContextMetadataKey,
-                stageRecorder: outboundRecorder
-            });
-            let formattedPayload = await runReqOutboundStage2FormatBuild({
-                formatEnvelope: outboundStage1.formatEnvelope,
-                adapterContext: outboundAdapterContext,
-                formatAdapter: outboundFormatAdapter,
-                stageRecorder: outboundRecorder
-            });
-            formattedPayload = await runReqOutboundStage3Compat({
-                payload: formattedPayload,
-                adapterContext: outboundAdapterContext,
-                stageRecorder: outboundRecorder
-            });
-            providerPayload = formattedPayload;
-            const outboundEnd = Date.now();
-            nodeResults.push({
-                id: 'req_outbound',
-                success: true,
-                metadata: {
-                    node: 'req_outbound',
-                    executionTime: outboundEnd - outboundStart,
-                    startTime: outboundStart,
-                    endTime: outboundEnd,
-                    dataProcessed: {
-                        messages: workingRequest.messages.length,
-                        tools: workingRequest.tools?.length ?? 0
-                    }
+        const outboundStage1 = await runReqOutboundStage1SemanticMap({
+            request: workingRequest,
+            adapterContext: outboundAdapterContext,
+            semanticMapper: outboundSemanticMapper,
+            contextSnapshot: outboundContextSnapshot,
+            contextMetadataKey: outboundContextMetadataKey,
+            stageRecorder: outboundRecorder
+        });
+        let formattedPayload = await runReqOutboundStage2FormatBuild({
+            formatEnvelope: outboundStage1.formatEnvelope,
+            adapterContext: outboundAdapterContext,
+            formatAdapter: outboundFormatAdapter,
+            stageRecorder: outboundRecorder
+        });
+        formattedPayload = await runReqOutboundStage3Compat({
+            payload: formattedPayload,
+            adapterContext: outboundAdapterContext,
+            stageRecorder: outboundRecorder
+        });
+        providerPayload = formattedPayload;
+        const outboundEnd = Date.now();
+        nodeResults.push({
+            id: 'req_outbound',
+            success: true,
+            metadata: {
+                node: 'req_outbound',
+                executionTime: outboundEnd - outboundStart,
+                startTime: outboundStart,
+                endTime: outboundEnd,
+                dataProcessed: {
+                    messages: workingRequest.messages.length,
+                    tools: workingRequest.tools?.length ?? 0
                 }
-            });
-        }
+            }
+        });
         // 为响应侧 servertool/web_search 提供一次性 Chat 请求快照，便于在 Hub 内部实现
         // 第三跳（将工具结果注入消息历史后重新调用主模型）。
+        //
+        // 注意：这里不再根据 processMode(passthrough/chat) 做分支判断——即使某些
+        // route 将 processMode 标记为 passthrough，我们仍然需要保留一次规范化后的
+        // Chat 请求快照，供 stopMessage / gemini_empty_reply_continue 等被动触发型
+        // servertool 在响应阶段使用。
         let capturedChatRequest;
-        if (normalized.processMode !== 'passthrough') {
-            try {
-                capturedChatRequest = JSON.parse(JSON.stringify({
-                    model: workingRequest.model,
-                    messages: workingRequest.messages,
-                    tools: workingRequest.tools,
-                    parameters: workingRequest.parameters
-                }));
-            }
-            catch {
-                capturedChatRequest = undefined;
-            }
+        try {
+            capturedChatRequest = JSON.parse(JSON.stringify({
+                model: workingRequest.model,
+                messages: workingRequest.messages,
+                tools: workingRequest.tools,
+                parameters: workingRequest.parameters
+            }));
+        }
+        catch {
+            capturedChatRequest = undefined;
         }
         const metadata = {
             ...normalized.metadata,
@@ -397,6 +391,13 @@ export class HubPipeline {
         if (conversationId) {
             adapterContext.conversationId = conversationId;
         }
+        const responsesResume = metadata.responsesResume &&
+            typeof metadata.responsesResume === 'object'
+            ? metadata.responsesResume
+            : undefined;
+        if (responsesResume) {
+            adapterContext.responsesResume = responsesResume;
+        }
         if (target?.compatibilityProfile && typeof target.compatibilityProfile === 'string') {
             adapterContext.compatibilityProfile = target.compatibilityProfile;
         }
@@ -420,16 +421,6 @@ export class HubPipeline {
         }
         return value;
     }
-    pickRawRequestBody(metadata) {
-        if (!metadata || typeof metadata !== 'object') {
-            return undefined;
-        }
-        const raw = metadata.__raw_request_body;
-        if (!raw || typeof raw !== 'object') {
-            return undefined;
-        }
-        return raw;
-    }
     async normalizeRequest(request) {
         if (!request || typeof request !== 'object') {
             throw new Error('HubPipeline requires request payload');

package/dist/conversion/hub/pipeline/stages/resp_outbound/resp_outbound_stage1_client_remap/index.js

@@ -15,7 +15,6 @@ export function runRespOutboundStage1ClientRemap(options) {
         clientPayload = buildResponsesPayloadFromChat(options.payload, {
             requestId: options.requestId
         });
-        mergeOriginalResponsesPayload(clientPayload, options.adapterContext);
     }
     recordStage(options.stageRecorder, 'resp_outbound_stage1_client_remap', clientPayload);
     return clientPayload;
@@ -42,36 +41,3 @@ function resolveAliasMapFromContext(adapterContext) {
     }
     return Object.keys(map).length ? map : undefined;
 }
-function mergeOriginalResponsesPayload(payload, adapterContext) {
-    if (!adapterContext) {
-        return;
-    }
-    const raw = adapterContext.__raw_responses_payload;
-    if (!raw || typeof raw !== 'object' || Array.isArray(raw)) {
-        return;
-    }
-    try {
-        if (payload.required_action == null && raw.required_action != null) {
-            payload.required_action = JSON.parse(JSON.stringify(raw.required_action));
-        }
-    }
-    catch {
-        /* ignore clone errors */
-    }
-    const rawStatus = typeof raw.status === 'string' ? raw.status : undefined;
-    if (rawStatus === 'requires_action') {
-        payload.status = 'requires_action';
-    }
-    // 如果桥接后的 payload 没有 usage，而原始 Responses 载荷带有 usage，则回填原始 usage，
-    // 确保 token usage 不在工具/桥接路径中丢失。
-    const payloadUsage = payload.usage;
-    const rawUsage = raw.usage;
-    if ((payloadUsage == null || typeof payloadUsage !== 'object') && rawUsage && typeof rawUsage === 'object') {
-        try {
-            payload.usage = JSON.parse(JSON.stringify(rawUsage));
-        }
-        catch {
-            payload.usage = rawUsage;
-        }
-    }
-}

package/dist/conversion/hub/process/chat-process.js

@@ -1,6 +1,6 @@
 import { runChatRequestToolFilters } from '../../shared/tool-filter-pipeline.js';
 import { ToolGovernanceEngine } from '../tool-governance/index.js';
-import { detectLastAssistantToolCategory } from '../../../router/virtual-router/tool-signals.js';
+import { ensureApplyPatchSchema } from '../../shared/tool-mapping.js';
 const toolGovernanceEngine = new ToolGovernanceEngine();
 export async function runHubChatProcess(options) {
     const startTime = Date.now();
@@ -72,6 +72,12 @@ async function applyRequestToolGovernance(request, context) {
             governanceTimestamp: Date.now()
         }
     };
+    // 清理历史图片：仅保留「最新一条 user 消息」中的图片分段，
+    // 避免历史对话中的图片在后续多轮工具 / 普通对话中继续作为多模态负载发给不支持图片的模型。
+    merged = {
+        ...merged,
+        messages: stripHistoricalImageAttachments(merged.messages)
+    };
     if (containsImageAttachment(merged.messages)) {
         if (!merged.metadata) {
             merged.metadata = {
@@ -205,6 +211,63 @@ function castSingleTool(tool) {
         }
     };
 }
+function stripHistoricalImageAttachments(messages) {
+    if (!Array.isArray(messages) || !messages.length) {
+        return messages;
+    }
+    // 找到最新一条 user 消息，仅允许该消息保留图片分段；
+    // 更早的 user 消息中若存在图片，则移除其 image* 分段，保留纯文本与非图片内容。
+    let latestUserIndex = -1;
+    for (let idx = messages.length - 1; idx >= 0; idx -= 1) {
+        const candidate = messages[idx];
+        if (candidate && typeof candidate === 'object' && candidate.role === 'user') {
+            latestUserIndex = idx;
+            break;
+        }
+    }
+    if (latestUserIndex < 0) {
+        return messages;
+    }
+    let changed = false;
+    const next = messages.slice();
+    for (let idx = 0; idx < messages.length; idx += 1) {
+        if (idx === latestUserIndex) {
+            continue;
+        }
+        const message = messages[idx];
+        if (!message || typeof message !== 'object') {
+            continue;
+        }
+        if (message.role !== 'user') {
+            continue;
+        }
+        const content = message.content;
+        if (!Array.isArray(content) || !content.length) {
+            continue;
+        }
+        const filtered = [];
+        let removed = false;
+        for (const part of content) {
+            if (part && typeof part === 'object' && !Array.isArray(part)) {
+                const typeValue = part.type;
+                if (typeof typeValue === 'string' && typeValue.toLowerCase().includes('image')) {
+                    removed = true;
+                    continue;
+                }
+            }
+            filtered.push(part);
+        }
+        if (removed) {
+            const cloned = {
+                ...message,
+                content: filtered
+            };
+            next[idx] = cloned;
+            changed = true;
+        }
+    }
+    return changed ? next : messages;
+}
 function containsImageAttachment(messages) {
     if (!Array.isArray(messages) || !messages.length) {
         return false;
@@ -278,17 +341,7 @@ function castCustomTool(tool) {
         function: {
             name: 'apply_patch',
             description,
-            parameters: {
-                type: 'object',
-                properties: {
-                    patch: {
-                        type: 'string',
-                        description: 'Unified diff patch content (FREEFORM, not JSON)'
-                    }
-                },
-                required: ['patch'],
-                additionalProperties: false
-            },
+            parameters: ensureApplyPatchSchema(),
             strict: true
         }
     };
@@ -355,19 +408,10 @@ function maybeInjectWebSearchTool(request, metadata) {
         : 'selective';
     const intent = detectWebSearchIntent(request);
     if (injectPolicy === 'selective') {
+        // 仅当当前这一轮用户输入明确表达“联网搜索”意图时才注入 web_search。
+        // 不再依赖上一轮工具分类（read/search/websearch），避免形成隐式续写语义。
         if (!intent.hasIntent) {
-            // 当最近一条用户消息没有明显的“联网搜索”关键词时，
-            // 如果上一轮 assistant 的工具调用已经属于搜索类（如 web_search），
-            // 则仍然视为 web_search 续写场景，强制注入 web_search 工具，
-            // 以便在后续路由中按 servertool 逻辑跳过不适配的 Provider（例如 serverToolsDisabled 的 crs）。
-            const assistantMessages = Array.isArray(request.messages)
-                ? request.messages.filter((msg) => msg && msg.role === 'assistant')
-                : [];
-            const lastTool = detectLastAssistantToolCategory(assistantMessages);
-            const hasSearchToolContext = lastTool?.category === 'search';
-            if (!hasSearchToolContext) {
-                return request;
-            }
+            return request;
         }
     }
     const existingTools = Array.isArray(request.tools) ? request.tools : [];

package/dist/conversion/hub/response/provider-response.js

@@ -130,14 +130,6 @@ export async function convertProviderResponse(options) {
         catch {
             // ignore conversation capture errors
         }
-        if (formatEnvelope.payload && typeof formatEnvelope.payload === 'object') {
-            try {
-                options.context.__raw_responses_payload = JSON.parse(JSON.stringify(formatEnvelope.payload));
-            }
-            catch {
-                /* best-effort clone */
-            }
-        }
     }
     formatEnvelope.payload = runRespInboundStageCompatResponse({
         payload: formatEnvelope.payload,

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

@@ -286,6 +286,14 @@ function appendChatContentToGeminiParts(message, targetParts) {
 function buildGeminiRequestFromChat(chat, metadata) {
     const contents = [];
     const emittedToolOutputs = new Set();
+    const adapterContext = metadata?.context;
+    const rawProviderId = adapterContext?.providerId;
+    const normalizedProviderId = typeof rawProviderId === 'string' ? rawProviderId.toLowerCase() : '';
+    const providerIdPrefix = normalizedProviderId.split('.')[0];
+    // 保持对通用 gemini-cli 的保护（避免上游直接执行 functionCall），
+    // 但对于 antigravity.* 明确允许通过 Gemini functionCall 协议执行工具，
+    // 以便完整打通 tools → functionCall → functionResponse 链路。
+    const omitFunctionCallPartsForCli = providerIdPrefix === 'gemini-cli';
     for (const message of chat.messages) {
         if (!message || typeof message !== 'object')
             continue;
@@ -304,10 +312,15 @@ function buildGeminiRequestFromChat(chat, metadata) {
             parts: []
         };
         appendChatContentToGeminiParts(message, entry.parts);
-        const toolCalls = Array.isArray(message.tool_calls) ? message.tool_calls : [];
+        const toolCalls = Array.isArray(message.tool_calls)
+            ? message.tool_calls
+            : [];
         for (const tc of toolCalls) {
             if (!tc || typeof tc !== 'object')
                 continue;
+            if (omitFunctionCallPartsForCli) {
+                continue;
+            }
             const fn = tc.function || {};
             const name = typeof fn.name === 'string' ? fn.name : undefined;
             if (!name)
@@ -324,7 +337,14 @@ function buildGeminiRequestFromChat(chat, metadata) {
             else {
                 argsStruct = fn.arguments ?? {};
             }
-            const part = { functionCall: { name, args: cloneAsJsonValue(argsStruct) } };
+            let argsJson = cloneAsJsonValue(argsStruct);
+            // Gemini / Antigravity 期望 functionCall.args 为对象（Struct），
+            // 若顶层为数组或原始类型，则包装到 value 字段下，避免产生非法的 list 形状。
+            if (!argsJson || typeof argsJson !== 'object' || Array.isArray(argsJson)) {
+                argsJson = { value: argsJson };
+            }
+            const functionCall = { name, args: argsJson };
+            const part = { functionCall };
             if (typeof tc.id === 'string') {
                 part.functionCall.id = tc.id;
             }
@@ -409,7 +429,6 @@ function buildGeminiRequestFromChat(chat, metadata) {
     }
     // Apply claude-thinking compat directly at Gemini mapping time to ensure it is always active
     // for antigravity.*.claude-sonnet-4-5-thinking, regardless of compatibilityProfile wiring.
-    const adapterContext = metadata?.context;
     const compatRequest = applyClaudeThinkingToolSchemaCompat(request, adapterContext);
     return compatRequest;
 }