Files
agent/app/server/server.js
T
kicer 695abf1f6a 功能:接入 prompt_tokens_details.cached_tokens,消息与模型统计展示缓存命中
- engine: assistant 落库与 SSE 事件带 usage {pt,ct,cached}(上游未报 details 时 cached=null 区分未上报与 0 命中);recordModelUsage 写 cached_tokens;modelUsageStats 聚合 today_cached/total_cached
- db: messages 加 usage TEXT、model_usage 加 cached_tokens 列(建表+迁移双路径,幂等)
- server: /api/pool 统计行透传 today_cached/total_cached
- 前端 msg-actions: 时间戳后新增「In: 1.2k/95%, Out: 456」小字(title 显示精确 tok 数),历史/SSE/摘要卡片三处渲染同参
- 模型卡新增「缓存命中」行:累计缓存 tok · 历史命中率(累计缓存÷累计输入)
2026-09-10 09:17:45 +08:00

130 lines
6.1 KiB
JavaScript
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
// 入口:装配配置、数据库、SSE 总线、路由与静态资源。
const express = require('express');
const http = require('http');
const path = require('path');
const fs = require('fs');
const { loadConfig } = require('./config');
const { initDb } = require('./db');
const { createBus } = require('./sse');
const { createAuth } = require('./auth');
const engine = require('./engine');
const llm = require('./llm');
const mcp = require('./mcp');
const skills = require('./tools/skills');
const log = require('./log');
const cfg = loadConfig(process.env.AGENT_CONFIG);
log.setLevel(cfg.log_level); // 分级日志开关(log.js;config.yml log_level,重启生效)
const db = initDb(cfg.db_path);
const bus = createBus();
engine.setModelPool(cfg, db, bus);
engine.init();
const app = express();
const server = http.createServer(app);
app.use(express.json({ limit: '2mb' }));
// 登录认证(cookie 会话):users 配在 config.yml 的 auth.users;留空则不启用。
// 相比 nginx htpasswd 的 basic auth,cookie 在手机浏览器原生下载时也会自动携带,下载不再失败。
const auth = createAuth(cfg);
app.use(auth.router); // /login、/logout、/api/login 等免认证端点
app.use(auth.middleware); // 其余请求需登录态(cookie 或 Basic 头)
app.use(express.static(path.join(__dirname, 'public')));
app.use('/api/sessions', require('./routes/sessions'));
app.use('/api/sessions', require('./routes/messages'));
app.use('/api/sessions', require('./routes/files'));
app.use('/api/skills', require('./routes/skills'));
app.use('/api/config', require('./routes/config'));
app.use('/api/platforms', require('./routes/platforms'));
app.use('/api/system', require('./routes/system'));
// 模型池状态(前端可展示空闲/占用)+ 每模型 token 用量(今日/累计,输入/输出分开)。
// 从生效配置补齐平台维度字段(platform/mid/vision/type + 平台 key 数 key_count):
// model_pool 表只有 id/display_name/占用态,会话页「选择模型」弹窗需要按平台分组展示。
// 配置里新增但重启前未进池的模型不在此返回(池才是运行时真相)。
app.get('/api/models', (req, res) => {
const rows = db.prepare('SELECT id, display_name, is_busy, current_session_id FROM model_pool ORDER BY id').all();
const stats = engine.modelUsageStats();
// 配置视图:模型唯一键 → 平台/原始 id/vision;平台名 → key 数(cfg.models 含 key 副本,不能直接拿来用)
const cfgNow = engine.getConfig();
const infoById = {};
const keyCountByPlat = {};
for (const p of cfgNow.platforms || []) {
keyCountByPlat[p.name] = (p.keys || []).length;
for (const m of p.models || []) {
infoById[p.name + '/' + m.id] = { platform: p.name, mid: m.id, vision: !!m.vision, type: p.type };
}
}
for (const r of rows) {
const info = infoById[r.id] || {};
r.platform = info.platform || '';
r.mid = info.mid || '';
r.vision = info.vision || false;
r.type = info.type || '';
r.key_count = keyCountByPlat[r.platform] || 0;
const s = stats[r.id] || {};
r.today_prompt = s.today_prompt || 0;
r.today_completion = s.today_completion || 0;
r.today_cached = s.today_cached || 0;
r.total_prompt = s.total_prompt || 0;
r.total_completion = s.total_completion || 0;
r.total_cached = s.total_cached || 0;
}
res.json(rows);
});
app.get('/api/health', (req, res) => res.json({ ok: true, time: Date.now() }));
const PORT = process.env.PORT || cfg.port;
// 启动成功后把模型实际收到的各部分 dump 到文件方便人工核对(data_root 默认 ./data,
// 宿主机对应 <repo>/data/prompt.txt,已被 gitignore)。包含两部分:
// 1) system 消息:模板 + 可用技能段落 + 语言指令(进程内恒定,仅技能增删时变);
// 2) tools 数组:内置 definitions.json + 启动时发现的 MCP 工具,与每次请求的 tools 字段一致。
function dumpPromptBundle() {
const file = path.join(cfg.data_root, 'prompt.txt');
const tools = engine.TOOL_DEFINITIONS.concat(mcp.toolDefinitions());
const parts = [
'# ===== 1. System Prompt(system 消息,进程内恒定)=====\n\n' + llm.systemPromptText((cfg.agent || {}).language),
'\n\n# ===== 2. Tools(' + tools.length + ' 个:内置 + MCP,随每次请求的 tools 字段发送)=====\n\n' + JSON.stringify(tools, null, 2),
];
fs.mkdirSync(path.dirname(file), { recursive: true });
fs.writeFileSync(file, parts.join(''), 'utf8');
log.info('startup', ' prompt_dump=' + file + ' (tools=' + tools.length + ')');
}
// 启动时先发现各 MCP 服务的工具(单个服务失败只警告,不阻止启动),
// 并对技能列表做快照(注入 system prompt,见 llm.setSkillListText),完成后开始监听。
mcp.init(cfg).then(async () => {
// 技能列表快照:模型直接感知已有技能,免先 list_skills 探索;
// 此处失败不阻塞启动,list_skills 仍可做实时兑底。
try {
const list = await skills.listSkills({ cfg });
llm.setSkillListText(skills.formatSkillListForPrompt(list));
log.info('startup', ' skills=' + (list.length ? list.map(s => s.name).join(', ') : '(none)'));
} catch (e) {
log.warn('startup', '技能列表生成失败:' + e.message);
}
server.listen(PORT, () => {
log.info('startup', 'Agent server listening on ' + PORT);
log.info('startup', ' workspace_root=' + cfg.workspace_root);
log.info('startup', ' data_root=' + cfg.data_root);
log.info('startup', ' models=' + cfg.models.map(m => m.id + '@' + m.url).join(', '));
const mcpList = (cfg.mcp_servers || []).map(s => s.name + '@' + s.url + (s.enable === false ? '(禁用)' : '')).join(', ');
log.info('startup', ' mcp_servers=' + (mcpList || '(none)'));
log.info('startup', ' auth=' + (auth.enabled ? 'login: ' + cfg.auth.users.map(u => u.user).join(', ') : '(disabled)'));
log.info('startup', ' log_level=' + cfg.log_level);
// prompt dump 失败不影响服务:只是核对用产物。
try {
dumpPromptBundle();
} catch (e) {
log.warn('startup', 'prompt dump 失败:' + e.message);
}
});
});