背景:v0.5.2 工具调用失效修复后,对模型全能力矩阵(工具/思考/多模态/流式/ 压缩/摘要/记忆/故障转移/余额)做契约级核查,发现并修复两处同类时序/边界缺陷。 P1 MCP 工具运行中增删不同步引擎: - 症状:运行中添加/启用 MCP server 后,已打开会话拿不到新工具;断开 server 后已有引擎仍持有失效工具定义(模型发起调用才报 Unknown tool) - 根因:setToolsAll 仅在启动期与工具开关时调用,MCP 连接/断开路径缺失 (与 v0.5.2 修复的懒创建陷阱同类——"变更点 × 同步路径"未全覆盖) - 修复:MCPManager 新增 setOnToolsChanged 回调,connectServer 注册完成 / disconnectServer 注销完成后触发;main.ts 注入回调同步全部已存在引擎。 懒创建引擎由 createEngine 实时拉取(v0.5.2),三条路径(启动/懒创建/ 运行中变更)全覆盖。README"无需重启动态发现"的宣称至此真实成立。 P1 maxTokens 超模型上限直接 400: - 症状:引擎默认 maxTokens=63488,OpenAI gpt-4o(16384)/gpt-4.1(32768)、 Anthropic opus/haiku(32000)、MiMo standard(32768) 每次请求 400,等于不可用 - 修复:五个 adapter(DeepSeek/Agnes/MiMo/OpenAI/Anthropic)统一按 MODEL_INFO.maxOutputTokens 钳制;MiMo 保留 thinking 兜底 32768 语义; Anthropic thinking budget 在钳制后的 max_tokens 内二分,自动跟随 测试(224 → 236 用例): - 新增 maxTokens 钳制契约测试 ×9(max-tokens-clamp.test.ts):mock fetch 记录真实请求体断言——超限钳制(MiMo standard/OpenAI gpt-4o/Anthropic opus)/ 未超限原样传递(DeepSeek/Agnes/MiMo pro/o3-mini/sonnet)/ 推理模型字段名 / 未配置默认值安全性 - 新增 MCP 动态同步端到端测试 ×3(mcp-tools-sync.test.ts):mock MCP SDK + 真实 MCPManager/ToolRegistry/AgentEngineManager——先建引擎再连 server,断言同一会话请求的 tools 动态更新 / 断开后移除失效定义 / 回调异常不阻断 MCP 主流程 能力矩阵核查结论(无回归确认): 工具调用主链路 ✓(v0.5.2)/ SubAgent 工具 ✓(delegate 实时 resolveTools)/ thinking 热更新 ✓(baseConfig 合并 路径无懒创建陷阱)/ 多模态当轮 ✓ / 压缩与孤立 tool 消息配对 ✓ / 摘要分层 ✓ / 记忆注入 ✓ / 故障转移 ✓ / 余额 ✓(v0.5.2)。已知设计限制:历史轮图片不 回传(attachments 仅存缩略图,图片只在发送当轮注入上下文)。 验证: lint 0 / typecheck 双工程 0 / test:electron 236 全过 / build 成功
224 lines
7.9 KiB
TypeScript
224 lines
7.9 KiB
TypeScript
/**
|
||
* Agnes AI Provider Adapter
|
||
*
|
||
* OpenAI 兼容 API。支持 Tool Calling、Thinking 模式、多模态(图片 — URL + Base64)。
|
||
*
|
||
* 独立继承 BaseAdapter,通过 shared/openai-format 和 shared/sse-stream 复用
|
||
* OpenAI 兼容格式构建和 SSE 流式解析逻辑。不与其他 Provider Adapter 耦合。
|
||
*
|
||
* 与 DeepSeek 的差异:
|
||
* - Thinking 模式使用 chat_template_kwargs(非 thinking 字段)
|
||
* - 默认 max_tokens 更大(65536 vs 8192)
|
||
*
|
||
* @see apis/agnes-ai-api-docs-20260625.html
|
||
*/
|
||
|
||
import { BaseAdapter } from './base-adapter';
|
||
import type { MetonaRequest, MetonaResponse, MetonaStreamEvent } from '../types';
|
||
import { MetonaFinishReason } from '../types';
|
||
import type { MetonaModelInfo } from '../types/metona-adapter';
|
||
import { buildOpenAICompatibleMessages, buildOpenAICompatibleTools } from './shared/openai-format';
|
||
import { parseSSEStream, parseOpenAICompatibleResponse } from './shared/sse-stream';
|
||
import log from 'electron-log';
|
||
|
||
export class AgnesAdapter extends BaseAdapter {
|
||
// H-2 修复: provider → providerId(规范要求)
|
||
override readonly providerId: string = 'agnes';
|
||
readonly supportedModels = ['agnes-2.0-flash'];
|
||
readonly supportsToolCalling = true;
|
||
readonly supportsThinking = true;
|
||
|
||
// H-2 修复: Agnes 模型元信息(1M 上下文,65.5K 最大输出)
|
||
private static readonly MODEL_INFO: Record<string, MetonaModelInfo> = {
|
||
'agnes-2.0-flash': {
|
||
id: 'agnes-2.0-flash',
|
||
name: 'Agnes 2.0 Flash',
|
||
contextWindow: 1_000_000,
|
||
maxOutputTokens: 65_536,
|
||
supportsToolCalling: true,
|
||
supportsThinking: true,
|
||
description: 'Agnes AI 快速版,1M 上下文,支持多模态图片(URL + Base64)与思考模式',
|
||
},
|
||
};
|
||
|
||
// ===== POST /chat/completions (非流式) =====
|
||
|
||
// H-2 修复: chat → send(规范要求)
|
||
async send(request: MetonaRequest): Promise<MetonaResponse> {
|
||
const body = this.toNativeRequest(request, false);
|
||
|
||
// #24 修复: 使用 fetchWithTimeout 替代 getFetchSignal + fetch,确保 timer 清理
|
||
const response = await this.fetchWithTimeout(
|
||
`${this.config.baseURL}/chat/completions`,
|
||
{
|
||
method: 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
Authorization: `Bearer ${this.config.apiKey}`,
|
||
...this.config.headers,
|
||
},
|
||
body: JSON.stringify(body),
|
||
},
|
||
this.config.timeoutMs ?? 300_000,
|
||
);
|
||
|
||
if (!response.ok) {
|
||
await this.throwHttpError(response, 'Agnes AI API error');
|
||
}
|
||
|
||
const data = (await response.json()) as Record<string, unknown>;
|
||
const parsed = parseOpenAICompatibleResponse(data);
|
||
|
||
return {
|
||
meta: {
|
||
requestId: request.meta.requestId,
|
||
provider: this.providerId,
|
||
model: (data.model as string) ?? this.config.defaultModel,
|
||
latencyMs: 0,
|
||
timestamp: Date.now(),
|
||
},
|
||
content: parsed.content,
|
||
reasoningContent: parsed.reasoningContent,
|
||
toolCalls: parsed.toolCalls,
|
||
usage: parsed.usage,
|
||
finishReason: parsed.finishReason as MetonaFinishReason,
|
||
};
|
||
}
|
||
|
||
// ===== POST /chat/completions (流式) =====
|
||
|
||
// H-2 修复: chatStream → sendStream(规范要求)
|
||
async *sendStream(request: MetonaRequest): AsyncIterable<MetonaStreamEvent> {
|
||
const body = this.toNativeRequest(request, true);
|
||
|
||
// #24 修复: 使用 fetchWithTimeout 替代 getFetchSignal + fetch,确保 timer 清理
|
||
const response = await this.fetchWithTimeout(
|
||
`${this.config.baseURL}/chat/completions`,
|
||
{
|
||
method: 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
Authorization: `Bearer ${this.config.apiKey}`,
|
||
...this.config.headers,
|
||
},
|
||
body: JSON.stringify(body),
|
||
},
|
||
this.config.timeoutMs ?? 300_000,
|
||
);
|
||
|
||
if (!response.ok || !response.body) {
|
||
await this.throwHttpError(response, 'Agnes AI stream error');
|
||
}
|
||
|
||
yield* parseSSEStream(
|
||
// 非空断言:上方 if 已确保 response.body 不为 null
|
||
response.body!,
|
||
request.meta.requestId,
|
||
request.meta.sessionId,
|
||
request.meta.iteration,
|
||
);
|
||
}
|
||
|
||
/**
|
||
* H-2 修复: 获取上下文窗口大小(规范要求)
|
||
*
|
||
* v0.3.1: 优先使用配置注入的 contextWindow,回退到 MODEL_INFO 默认值。
|
||
* Agnes OpenAI 兼容 API 不支持 context_window 参数,此值仅用于
|
||
* Engine 压缩判断和前端 UI 显示。
|
||
* 注意:Agnes API 未提供 /models 端点,listModels 使用基类默认实现。
|
||
*/
|
||
override getContextWindow(): number {
|
||
// v0.3.1: 优先使用配置注入的 contextWindow
|
||
if (typeof this.config.contextWindow === 'number' && this.config.contextWindow > 0) {
|
||
return this.config.contextWindow;
|
||
}
|
||
// 回退到 MODEL_INFO
|
||
const modelInfo = AgnesAdapter.MODEL_INFO[this.config.defaultModel];
|
||
return modelInfo?.contextWindow ?? 1_000_000;
|
||
}
|
||
|
||
// ========== 私有方法 ==========
|
||
|
||
/**
|
||
* 构建 Agnes AI 原生请求体
|
||
*
|
||
* Agnes AI 特有参数:
|
||
* - 多模态图片:user 消息的 images[] → OpenAI content 数组 [{type:"text"}, {type:"image_url"}]
|
||
* 支持 HTTPS URL 或 base64 Data URI(与 MiMo 一致)
|
||
* - chat_template_kwargs: { enable_thinking: true } — 启用思考模式(非 thinking 字段)
|
||
* - 默认 max_tokens: 65536(1M 上下文,65.5K 最大输出)
|
||
*/
|
||
private toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
|
||
const messages = buildOpenAICompatibleMessages(request);
|
||
const tools = buildOpenAICompatibleTools(request.tools);
|
||
|
||
// === Agnes 多模态:将 images 转为 OpenAI content 数组 ===
|
||
// buildOpenAICompatibleMessages 不处理图片(各 Provider 自行处理)
|
||
const nonSystemMsgs = request.messages.filter((m) => m.role !== 'system');
|
||
let imageCount = 0;
|
||
for (let i = 0; i < messages.length; i++) {
|
||
// messages[0] 是 system,非 system 消息从 messages[1] 开始
|
||
if (i === 0) continue;
|
||
const origMsg = nonSystemMsgs[i - 1];
|
||
if (!origMsg?.images?.length) continue;
|
||
|
||
imageCount += origMsg.images.length;
|
||
|
||
const contentParts: Array<Record<string, unknown>> = [];
|
||
if (origMsg.content) {
|
||
contentParts.push({ type: 'text', text: origMsg.content });
|
||
}
|
||
for (const img of origMsg.images) {
|
||
contentParts.push({
|
||
type: 'image_url',
|
||
image_url: { url: img.url },
|
||
});
|
||
}
|
||
messages[i].content = contentParts;
|
||
}
|
||
|
||
if (imageCount > 0) {
|
||
const firstUrl = request.messages.find((m) => m.images?.length)?.images?.[0]?.url ?? '';
|
||
log.info(
|
||
`[Agnes] Processing ${imageCount} image(s), first URL prefix: ${firstUrl.slice(0, 50)}`,
|
||
);
|
||
}
|
||
|
||
// v0.5.3: max_tokens 按模型上限钳制(agnes-2.0-flash 上限 65536)
|
||
const modelInfo = AgnesAdapter.MODEL_INFO[this.config.defaultModel];
|
||
const maxOutput = modelInfo?.maxOutputTokens ?? 65_536;
|
||
const maxTokens = Math.min(request.params.maxTokens ?? maxOutput, maxOutput);
|
||
|
||
const body: Record<string, unknown> = {
|
||
model: this.config.defaultModel,
|
||
messages,
|
||
temperature: request.params.temperature,
|
||
max_tokens: maxTokens,
|
||
stream,
|
||
};
|
||
|
||
if (stream) {
|
||
body.stream_options = { include_usage: true };
|
||
}
|
||
|
||
if (tools) {
|
||
body.tools = tools;
|
||
}
|
||
|
||
// C-3 修复: Thinking 模式 — Agnes 使用 chat_template_kwargs 而非 thinking
|
||
// Agnes API 仅支持 enable_thinking: true/false,不支持 effort 级别
|
||
// thinkingEffort === 'low' 时映射为 false(不启用深度思考),其他级别映射为 true
|
||
if (request.params.thinkingEnabled) {
|
||
const effort = request.params.thinkingEffort ?? 'high';
|
||
body.chat_template_kwargs = { enable_thinking: effort !== 'low' };
|
||
}
|
||
|
||
// 停止序列
|
||
if (request.params.stopSequences?.length) {
|
||
body.stop = request.params.stopSequences;
|
||
}
|
||
|
||
return body;
|
||
}
|
||
}
|