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
287 lines
10 KiB
TypeScript
287 lines
10 KiB
TypeScript
/**
|
||
* DeepSeek Provider Adapter
|
||
*
|
||
* 基于 OpenAI 兼容 API。支持 Tool Calling、Thinking 模式、流式输出。
|
||
* 模型: deepseek-v4-flash / deepseek-v4-pro(1M 上下文,384K 最大输出)
|
||
*
|
||
* 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。
|
||
*
|
||
* @see apis/deepseek-api-docs-20260518.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';
|
||
|
||
export class DeepSeekAdapter extends BaseAdapter {
|
||
// H-2 修复: provider → providerId(规范要求)
|
||
override readonly providerId: string = 'deepseek';
|
||
readonly supportedModels = ['deepseek-v4-pro', 'deepseek-v4-flash'];
|
||
readonly supportsToolCalling = true;
|
||
readonly supportsThinking = true;
|
||
|
||
// H-2 修复: DeepSeek 模型元信息(1M 上下文,384K 最大输出)
|
||
private static readonly MODEL_INFO: Record<string, MetonaModelInfo> = {
|
||
'deepseek-v4-pro': {
|
||
id: 'deepseek-v4-pro',
|
||
name: 'DeepSeek V4 Pro',
|
||
contextWindow: 1_000_000,
|
||
maxOutputTokens: 384_000,
|
||
supportsToolCalling: true,
|
||
supportsThinking: true,
|
||
description: 'DeepSeek 旗舰模型,1M 上下文,支持深度推理与工具调用',
|
||
},
|
||
'deepseek-v4-flash': {
|
||
id: 'deepseek-v4-flash',
|
||
name: 'DeepSeek V4 Flash',
|
||
contextWindow: 1_000_000,
|
||
maxOutputTokens: 384_000,
|
||
supportsToolCalling: true,
|
||
supportsThinking: true,
|
||
description: 'DeepSeek 快速版,1M 上下文,低延迟推理',
|
||
},
|
||
};
|
||
|
||
// ===== 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 ${this.config.apiKey}`,
|
||
...this.config.headers,
|
||
},
|
||
body: JSON.stringify(body),
|
||
},
|
||
this.config.timeoutMs ?? 120_000,
|
||
);
|
||
|
||
if (!response.ok) {
|
||
await this.throwHttpError(response, 'DeepSeek API error');
|
||
}
|
||
|
||
const data = (await response.json()) as Record<string, unknown>;
|
||
const parsed = parseOpenAICompatibleResponse(data);
|
||
|
||
return {
|
||
meta: {
|
||
requestId: request.meta.requestId,
|
||
provider: this.providerId,
|
||
model: (data.model as string) ?? this.config.defaultModel,
|
||
latencyMs: 0,
|
||
timestamp: Date.now(),
|
||
},
|
||
content: parsed.content,
|
||
reasoningContent: parsed.reasoningContent,
|
||
toolCalls: parsed.toolCalls,
|
||
usage: parsed.usage,
|
||
finishReason: parsed.finishReason as MetonaFinishReason,
|
||
};
|
||
}
|
||
|
||
// ===== POST /chat/completions (流式) =====
|
||
|
||
// H-2 修复: chatStream → sendStream(规范要求)
|
||
async *sendStream(request: MetonaRequest): AsyncIterable<MetonaStreamEvent> {
|
||
const body = this.toNativeRequest(request, true);
|
||
|
||
// #24 修复: 使用 fetchWithTimeout 替代 getFetchSignal + fetch,确保 timer 清理
|
||
const response = await this.fetchWithTimeout(
|
||
`${this.config.baseURL}/chat/completions`,
|
||
{
|
||
method: 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
Authorization: `Bearer ${this.config.apiKey}`,
|
||
...this.config.headers,
|
||
},
|
||
body: JSON.stringify(body),
|
||
},
|
||
this.config.timeoutMs ?? 300_000,
|
||
);
|
||
|
||
if (!response.ok || !response.body) {
|
||
await this.throwHttpError(response, 'DeepSeek stream error');
|
||
}
|
||
|
||
yield* parseSSEStream(
|
||
// 非空断言:上方 if 已确保 response.body 不为 null
|
||
// TypeScript 无法通过 await Promise<never> 正确收窄,需显式断言
|
||
response.body!,
|
||
request.meta.requestId,
|
||
request.meta.sessionId,
|
||
request.meta.iteration,
|
||
);
|
||
}
|
||
|
||
// ===== GET /models =====
|
||
|
||
/**
|
||
* H-2 修复: 返回 MetonaModelInfo[](规范要求)
|
||
*
|
||
* 优先尝试从 API 获取实时模型列表,并合并本地 MODEL_INFO 元数据。
|
||
* API 不可用时回退到 supportedModels。
|
||
*/
|
||
async listModels(): Promise<MetonaModelInfo[]> {
|
||
try {
|
||
const response = await fetch(`${this.config.baseURL}/models`, {
|
||
headers: { Authorization: `Bearer ${this.config.apiKey}` },
|
||
signal: AbortSignal.timeout(10_000),
|
||
});
|
||
if (response.ok) {
|
||
const data = (await response.json()) as { data?: Array<{ id: string }> };
|
||
if (data.data?.length) {
|
||
// 合并 API 返回的模型 ID 与本地元数据
|
||
return data.data.map((m) => DeepSeekAdapter.MODEL_INFO[m.id] ?? { id: m.id });
|
||
}
|
||
}
|
||
} catch {
|
||
// API 不可用时降级
|
||
}
|
||
// 回退到 supportedModels(带本地元数据)
|
||
return this.supportedModels.map((id) => DeepSeekAdapter.MODEL_INFO[id] ?? { id });
|
||
}
|
||
|
||
/**
|
||
* H-2 修复: 获取上下文窗口大小(规范要求)
|
||
*
|
||
* v0.3.1: 优先使用配置注入的 contextWindow,回退到 MODEL_INFO 默认值。
|
||
* DeepSeek OpenAI 兼容 API 不支持 context_window 参数,此值仅用于
|
||
* Engine 压缩判断和前端 UI 显示。
|
||
*/
|
||
override getContextWindow(): number {
|
||
// v0.3.1: 优先使用配置注入的 contextWindow
|
||
if (typeof this.config.contextWindow === 'number' && this.config.contextWindow > 0) {
|
||
return this.config.contextWindow;
|
||
}
|
||
// 回退到 MODEL_INFO
|
||
const modelInfo = DeepSeekAdapter.MODEL_INFO[this.config.defaultModel];
|
||
return modelInfo?.contextWindow ?? 1_000_000;
|
||
}
|
||
|
||
// ===== GET /user/balance =====
|
||
|
||
/**
|
||
* 查询账户余额
|
||
*
|
||
* v0.5.2 修复: DeepSeek 官方 API 实际返回 `balance_infos` 数组格式:
|
||
* { "is_available": true, "balance_infos": [{ "currency": "CNY",
|
||
* "total_balance": "110.00", "granted_balance": "10.00", "topped_up_balance": "100.00" }] }
|
||
* 此前按扁平字段解析(data.total_balance)→ 永远取到 undefined → 恒显示 0。
|
||
* 现优先取 balance_infos[0],回退扁平格式(兼容网关/代理的简化响应)。
|
||
*
|
||
* URL 规范化: 余额端点为 {root}/user/balance(无 /v1 前缀)。用户配置的
|
||
* baseURL 可能带 /v1 或尾斜杠(chat 端点两种写法都合法),此处剥离后拼接。
|
||
*/
|
||
async getBalance(): Promise<{
|
||
currency: string;
|
||
totalBalance: string;
|
||
grantedBalance: string;
|
||
toppedUpBalance: string;
|
||
} | null> {
|
||
try {
|
||
// 规范化 baseURL:去尾斜杠、去尾 /v1(余额端点在根路径下)
|
||
const root = this.config.baseURL.replace(/\/+$/, '').replace(/\/v1$/, '');
|
||
const response = await fetch(`${root}/user/balance`, {
|
||
headers: { Authorization: `Bearer ${this.config.apiKey}` },
|
||
signal: AbortSignal.timeout(10_000),
|
||
});
|
||
if (!response.ok) return null;
|
||
const data = (await response.json()) as {
|
||
is_available?: boolean;
|
||
balance_infos?: Array<{
|
||
currency?: string;
|
||
total_balance?: string;
|
||
granted_balance?: string;
|
||
topped_up_balance?: string;
|
||
}>;
|
||
// 扁平格式字段(网关/代理兼容)
|
||
currency?: string;
|
||
total_balance?: string;
|
||
granted_balance?: string;
|
||
topped_up_balance?: string;
|
||
};
|
||
// 优先官方 balance_infos 数组,回退扁平格式
|
||
const info = data.balance_infos?.[0] ?? data;
|
||
return {
|
||
currency: info.currency ?? 'CNY',
|
||
totalBalance: info.total_balance ?? '0',
|
||
grantedBalance: info.granted_balance ?? '0',
|
||
toppedUpBalance: info.topped_up_balance ?? '0',
|
||
};
|
||
} catch {
|
||
return null;
|
||
}
|
||
}
|
||
|
||
// ========== 私有方法 ==========
|
||
|
||
/**
|
||
* 构建 DeepSeek 原生请求体
|
||
*
|
||
* DeepSeek 特有参数:
|
||
* - thinking: { type: "enabled" } — 启用思考模式
|
||
* - reasoning_effort — 思考强度映射
|
||
* - stream_options: { include_usage: true } — 流式返回 usage
|
||
*/
|
||
private toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
|
||
const messages = buildOpenAICompatibleMessages(request);
|
||
const tools = buildOpenAICompatibleTools(request.tools);
|
||
|
||
// v0.5.3: max_tokens 按模型上限钳制 — 引擎默认 63488 超过部分模型上限时 API 直接 400
|
||
const modelInfo = DeepSeekAdapter.MODEL_INFO[this.config.defaultModel];
|
||
const maxOutput = modelInfo?.maxOutputTokens ?? 384_000;
|
||
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;
|
||
}
|
||
|
||
// Thinking 模式
|
||
// API 默认 thinking.type = "enabled",必须显式发送 disabled 才能关闭
|
||
if (request.params.thinkingEnabled === false) {
|
||
body.thinking = { type: 'disabled' };
|
||
} else if (request.params.thinkingEnabled) {
|
||
body.thinking = { type: 'enabled' };
|
||
const effortMap: Record<string, string> = {
|
||
low: 'high',
|
||
medium: 'high',
|
||
high: 'high',
|
||
max: 'max',
|
||
};
|
||
// DeepSeek API 仅支持 high / max 两档,low/medium 映射为 high
|
||
body.reasoning_effort = effortMap[request.params.thinkingEffort ?? 'high'] ?? 'high';
|
||
}
|
||
|
||
// 停止序列
|
||
if (request.params.stopSequences?.length) {
|
||
body.stop = request.params.stopSequences;
|
||
}
|
||
|
||
return body;
|
||
}
|
||
}
|