llm-adapter

name	llm-adapter
description	Add a new LLM provider to siyuan-agent - create ChatXxx class extending ChatOpenAI, reasoning support, modelKwargs mapping

LLM Provider Adapter

为 siyuan-agent 添加新的 LLM 供应商适配。

架构

src/llms/
  reasoning.ts       # 共享: injectReasoningContent, DEEPSEEK_PROFILES, ModelProfile
  deepseek.ts        # ChatDeepSeek extends ChatOpenAI
  <provider>.ts      # 新供应商: ChatXxx extends ChatOpenAI
src/core/
  chat-model.ts      # 编排层: createChatModel() 按 providerType 分发

新增供应商步骤

1. 创建 src/llms/<provider>.ts

import { ChatOpenAI } from "@langchain/openai";
import type { ReasoningEffort } from "../types";

export class ChatXxx extends ChatOpenAI {
    // (a) kwargs 映射 — 将统一的 ReasoningEffort 翻译为供应商 API 参数
    static getModelKwargs(effort: ReasoningEffort = "default"): Record<string, any> {
        if (effort === "off") return { /* 关闭思考的参数 */ };
        if (effort === "high") return { /* 高深度思考的参数 */ };
        return {};
    }

    constructor(fields: ChatXxxInput = {}) {
        super({
            ...fields,
            apiKey: fields.apiKey || process.env.XXX_API_KEY,
            configuration: {
                baseURL: "https://api.xxx.com",
                dangerouslyAllowBrowser: true,
                ...fields.configuration,
            },
        });
        // (b) 如需 reasoning 提取/回传，patch completions
        this.patchCompletions();
    }

    // (c) 如模型返回 reasoning_content 字段，patch 响应转换
    private patchCompletions() { /* 参考 deepseek.ts */ }

    // (d) 如模型用 <think> 标签包裹思考内容，override 流式处理
    override async *_streamResponseChunks(...) { /* 参考 deepseek.ts */ }
}

2. 注册到 src/types/model-config.ts

在 ModelProviderType 联合类型中添加新值:

export type ModelProviderType = "openai-compatible" | "deepseek" | "xxx";

更新 inferProviderType() 的启发式判断。

3. 接入 src/core/chat-model.ts

在 createChatModel 中添加分支:

if (config.providerType === "xxx") {
    return new ChatXxx({
        ...commonFields,
        modelKwargs: ChatXxx.getModelKwargs(options.reasoningEffort),
    });
}

统一用户设置: ReasoningEffort

用户只需设置一个枚举值，每个供应商自行翻译:

ReasoningEffort	含义	DeepSeek 映射
"default"	供应商默认	{}
"off"	关闭思考	{ thinking: { type: "disabled" } }
"low"	低深度思考	{ reasoning_effort: "high", thinking: { type: "enabled" } }
"high"	高深度思考	{ reasoning_effort: "max", thinking: { type: "enabled" } }

Reasoning 数据流

响应侧: API delta.reasoning_content → patch → additional_kwargs.reasoning_content → stream-runtime.ts → UI
请求侧: additional_kwargs.reasoning_content → injectReasoningContent → API request messages (回传上下文)

关键: convertMessagesToCompletionsMessageParams 不处理 reasoning_content，必须在 completionWithRetry 中注入。

关键文件

• src/llms/deepseek.ts — 完整参考实现
• src/llms/reasoning.ts — 共享工具函数
• src/core/chat-model.ts — 编排层
• src/types/model-config.ts — 类型定义
• src/core/stream-runtime.ts — reasoning 渲染消费端

Task

$ARGUMENTS