revert: relay 移出 agent 仓库
relay 是 workbuddy2api 的前缀组件,跟着网关走才合理,放到 tools/workbuddy-api。
This commit is contained in:
@@ -1,6 +0,0 @@
|
||||
FROM node:22-alpine
|
||||
WORKDIR /app
|
||||
COPY src/ ./src/
|
||||
# 镜像内只放模板;真身部署时用 volume 挂进来
|
||||
COPY config.example.json ./config.json
|
||||
CMD ["node", "src/server.js"]
|
||||
-105
@@ -1,105 +0,0 @@
|
||||
# Context Relay
|
||||
|
||||
把「每轮重传全量历史」换成「只传增量」。挡在 `workbuddy2api` 前面,对上游透明。
|
||||
|
||||
**为什么不改上游**:`workbuddy2api` 一天几十个提交,fork 维护不起。这层只在两者之间做
|
||||
「查表 + 拼数组 + pipe」,不改协议、不改 SSE。
|
||||
|
||||
---
|
||||
|
||||
## 协议
|
||||
|
||||
客户端多带两个头:
|
||||
|
||||
| header | 说明 |
|
||||
|---|---|
|
||||
| `X-Context-Id` | `<core_id>:<user>:<session_id>`,会话级唯一标识 |
|
||||
| `X-Context-Hash` | 客户端认为本服务已存的那段消息的 sha256;**空字符串 = 全量替换** |
|
||||
|
||||
请求体的 `messages` **只放新增**(首轮/压缩后放全量),其余字段原样。
|
||||
|
||||
### 三种走向
|
||||
|
||||
| 条件 | 行为 |
|
||||
|---|---|
|
||||
| 没带 `X-Context-Id`,或模型没开 store | 完全透明转发,和直连一致 |
|
||||
| `X-Context-Hash` 为空 | 整段替换缓存 |
|
||||
| 缓存存在且 hash 命中 | 缓存 + 新增拼起来转发 |
|
||||
| 其他 | **409** `{"code":"context_mismatch"}`,客户端改发全量 |
|
||||
|
||||
409 是唯一的重试信号:会话被压缩、本服务重启、缓存过期,全都走这一条路,客户端不需要
|
||||
自己判断「该不该发全量」。
|
||||
|
||||
---
|
||||
|
||||
## 配置
|
||||
|
||||
`config.json`(部署时放,不入库;模板见 `config.example.json`):
|
||||
|
||||
```jsonc
|
||||
{
|
||||
"port": 7865,
|
||||
"upstream": "http://workbuddy2api:7863",
|
||||
"max_body_mb": 64,
|
||||
"store": {
|
||||
"ttl_ms": 604800000, // 7 天
|
||||
"max_bytes": 5242880, // 单个会话上限,超了该会话不缓存(不影响别人)
|
||||
"max_total_bytes": 268435456,// 全部会话总量闸,超了 LRU 淘汰
|
||||
"models": { "cn:*": true, "global:*": false }
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
- **单会话上限**:超过就不缓存这个会话,下次请求 409 → 全量,行为仍然正确
|
||||
- **不设条数上限**:条数上限会在「会话最长、最需要增量」的时候把人打回全量,正好反着来
|
||||
- **模型白名单**:精确匹配优先,其次只认尾部 `*` 的前缀通配;不写 = 不开
|
||||
|
||||
---
|
||||
|
||||
## 部署
|
||||
|
||||
```yaml
|
||||
# docker-compose.yml
|
||||
ctx-relay:
|
||||
build: ./relay
|
||||
container_name: ctx-relay
|
||||
restart: unless-stopped
|
||||
volumes:
|
||||
- ./relay/config.json:/app/config.json:ro
|
||||
networks: [agent-mcp-net]
|
||||
```
|
||||
|
||||
nginx 侧只新增一条路径,不影响现有的 `/`:
|
||||
|
||||
```nginx
|
||||
location /v1c/ {
|
||||
proxy_pass http://ctx-relay:7865/;
|
||||
proxy_http_version 1.1;
|
||||
proxy_set_header Connection "";
|
||||
proxy_buffering off;
|
||||
proxy_read_timeout 3600s;
|
||||
}
|
||||
```
|
||||
|
||||
客户端 base URL 改成 `https://<host>/v1c`。
|
||||
|
||||
---
|
||||
|
||||
## 观测
|
||||
|
||||
每个带 context 的请求打一行:
|
||||
|
||||
```
|
||||
19:38:27 [relay] kicer:task-xxx delta=2 total=51 上传 3.1KB 省 196.9KB 缓存 1.2MB/256.00MB
|
||||
```
|
||||
|
||||
`/health` 返回当前会话数与占用字节。
|
||||
|
||||
---
|
||||
|
||||
## 边界
|
||||
|
||||
- **内存态**:进程重启缓存即丢 → 一次全量重建,之后恢复增量
|
||||
- **并发**:同一 `X-Context-Id` 串行(core 本来就串行,这把锁只防两个客户端同时打)
|
||||
- **流式**:响应原样 `pipe`,不解析、不改帧、不缓冲
|
||||
- **压缩**:等价于一次全量替换,无需特殊处理
|
||||
@@ -1,15 +0,0 @@
|
||||
{
|
||||
"port": 7865,
|
||||
"upstream": "http://workbuddy2api:7863",
|
||||
"max_body_mb": 64,
|
||||
"store": {
|
||||
"ttl_ms": 604800000,
|
||||
"max_bytes": 5242880,
|
||||
"max_total_bytes": 268435456,
|
||||
"models": {
|
||||
"cn:deepseek-v4.1-flash": true,
|
||||
"cn:*": true,
|
||||
"global:*": false
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -1,280 +0,0 @@
|
||||
#!/usr/bin/env node
|
||||
'use strict';
|
||||
// 上下文缓存中继:把「每轮重传全量历史」换成「只传增量」。
|
||||
//
|
||||
// 客户端带两个头:
|
||||
// X-Context-Id <core_id>:<user>:<session_id>
|
||||
// X-Context-Hash 客户端认为本服务已存的那段消息的 sha256(空 = 全量替换)
|
||||
// 命中就把缓存拼在新增前面转发给上游;不命中回 409,客户端改发全量。
|
||||
// 不带这两个头 → 完全透明转发,行为与直连一致。
|
||||
//
|
||||
// 本服务只做三件事:查表、拼数组、pipe 响应。不解析 SSE、不改帧、不缓冲。
|
||||
|
||||
const http = require('node:http');
|
||||
const https = require('node:https');
|
||||
const fs = require('node:fs');
|
||||
const path = require('node:path');
|
||||
const crypto = require('node:crypto');
|
||||
|
||||
// ---------- 配置 ----------
|
||||
const DEFAULTS = {
|
||||
port: 7865,
|
||||
upstream: 'http://127.0.0.1:7863',
|
||||
max_body_mb: 64,
|
||||
store: {
|
||||
ttl_ms: 7 * 24 * 3600 * 1000,
|
||||
max_bytes: 5 * 1024 * 1024, // 单个会话上限
|
||||
max_total_bytes: 256 * 1024 * 1024, // 全部会话总量闸(安全阀)
|
||||
models: {},
|
||||
},
|
||||
};
|
||||
|
||||
function loadConfig() {
|
||||
const file = process.argv[2] || path.join(__dirname, '..', 'config.json');
|
||||
let raw = {};
|
||||
if (fs.existsSync(file)) {
|
||||
try {
|
||||
raw = JSON.parse(fs.readFileSync(file, 'utf8'));
|
||||
} catch (e) {
|
||||
console.error('[relay] 配置不是合法 JSON:' + file + ' → ' + e.message);
|
||||
process.exit(1);
|
||||
}
|
||||
}
|
||||
const store = Object.assign({}, DEFAULTS.store, raw.store || {});
|
||||
store.models = store.models && typeof store.models === 'object' ? store.models : {};
|
||||
return Object.assign({}, DEFAULTS, raw, { store });
|
||||
}
|
||||
const CFG = loadConfig();
|
||||
|
||||
function log(...a) {
|
||||
console.log(new Date().toISOString().slice(11, 19), '[relay]', ...a);
|
||||
}
|
||||
|
||||
// ---------- 模型白名单 ----------
|
||||
// 精确匹配优先,其次只认尾部带 * 的前缀通配(cn:*)。不写 = 不开。
|
||||
function storeEnabled(model) {
|
||||
const m = CFG.store.models;
|
||||
const name = String(model || '');
|
||||
if (Object.prototype.hasOwnProperty.call(m, name)) return m[name] === true;
|
||||
for (const k of Object.keys(m)) {
|
||||
if (k.endsWith('*') && name.startsWith(k.slice(0, -1))) return m[k] === true;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
// ---------- 缓存 ----------
|
||||
// key -> { hash, messages, bytes, expireAt, lastUsed }
|
||||
// 两个字节数阈值,语义不同:
|
||||
// max_bytes —— 单个会话上限,超了就这个会话不缓存(下次全量),不影响别人
|
||||
// max_total_bytes —— 全部会话总量闸,超了按 LRU 淘汰最久没用的
|
||||
// 不设条数上限:那会在「会话最长、最需要增量」的时候把人打回全量,正好反着来。
|
||||
const cache = new Map();
|
||||
let cacheBytes = 0;
|
||||
|
||||
function sha256(s) {
|
||||
return crypto.createHash('sha256').update(s).digest('hex');
|
||||
}
|
||||
// 两侧(core 与本服务)都只认 JSON.stringify 的结果:字段顺序取自原始报文,
|
||||
// 客户端拼出来的字符串与服务端一致。
|
||||
function messagesHash(messages) {
|
||||
return sha256(JSON.stringify(messages));
|
||||
}
|
||||
function byteLen(s) {
|
||||
return Buffer.byteLength(s, 'utf8');
|
||||
}
|
||||
|
||||
function touch (key, entry) {
|
||||
entry.lastUsed = Date.now();
|
||||
// Map 插入顺序即 LRU 顺序:删了再插,等价于移到队尾
|
||||
cache.delete(key);
|
||||
cache.set(key, entry);
|
||||
}
|
||||
|
||||
function evictIfNeeded() {
|
||||
const max = CFG.store.max_total_bytes;
|
||||
const now = Date.now();
|
||||
for (const [k, e] of cache) {
|
||||
if (e.expireAt <= now) {
|
||||
cache.delete(k);
|
||||
cacheBytes -= e.bytes;
|
||||
}
|
||||
}
|
||||
while (cacheBytes > max && cache.size) {
|
||||
const oldest = cache.keys().next().value;
|
||||
const e = cache.get(oldest);
|
||||
cache.delete(oldest);
|
||||
cacheBytes -= e.bytes;
|
||||
log('LRU 淘汰 ' + oldest + '(' + fmtBytes(e.bytes) + ')');
|
||||
}
|
||||
}
|
||||
|
||||
function put(key, messages) {
|
||||
const prev = cache.get(key);
|
||||
if (prev) { cacheBytes -= prev.bytes; cache.delete(key); }
|
||||
const bytes = byteLen(JSON.stringify(messages));
|
||||
if (bytes > CFG.store.max_bytes) {
|
||||
// 单会话超上限:直接不缓存。不能截断(截断会破坏上下文),只能退回全量模式。
|
||||
log(key + ' 单会话 ' + fmtBytes(bytes) + ' 超过上限 ' + fmtBytes(CFG.store.max_bytes) + ',不缓存');
|
||||
return null;
|
||||
}
|
||||
const entry = {
|
||||
hash: messagesHash(messages),
|
||||
messages,
|
||||
bytes,
|
||||
expireAt: Date.now() + CFG.store.ttl_ms,
|
||||
lastUsed: Date.now(),
|
||||
};
|
||||
cache.set(key, entry);
|
||||
cacheBytes += bytes;
|
||||
evictIfNeeded();
|
||||
return entry;
|
||||
}
|
||||
|
||||
function get(key) {
|
||||
const e = cache.get(key);
|
||||
if (!e) return null;
|
||||
if (e.expireAt <= Date.now()) {
|
||||
cache.delete(key);
|
||||
cacheBytes -= e.bytes;
|
||||
return null;
|
||||
}
|
||||
touch(key, e);
|
||||
return e;
|
||||
}
|
||||
|
||||
function fmtBytes(n) {
|
||||
if (n < 1024) return n + 'B';
|
||||
if (n < 1048576) return (n / 1024).toFixed(1) + 'KB';
|
||||
return (n / 1048576).toFixed(2) + 'MB';
|
||||
}
|
||||
|
||||
// ---------- 同 key 串行 ----------
|
||||
// core 本来就是串行发的,这把锁只防「同一会话被两个客户端同时打」导致拼接乱序。
|
||||
const locks = new Map();
|
||||
async function withLock(key, fn) {
|
||||
while (locks.has(key)) await locks.get(key);
|
||||
let release;
|
||||
const p = new Promise((r) => (release = r));
|
||||
locks.set(key, p);
|
||||
try {
|
||||
return await fn();
|
||||
} finally {
|
||||
locks.delete(key);
|
||||
release();
|
||||
}
|
||||
}
|
||||
|
||||
// ---------- 转发 ----------
|
||||
function splitUrl(u) {
|
||||
const m = /^(https?):\/\/([^/]+)(\/.*)?$/.exec(u);
|
||||
if (!m) throw new Error('上游地址非法: ' + u);
|
||||
return { mod: m[1] === 'https' ? https : http, host: m[2], base: m[3] || '' };
|
||||
}
|
||||
const UP = splitUrl(CFG.upstream);
|
||||
|
||||
const HOP = new Set(['connection', 'keep-alive', 'proxy-authenticate', 'proxy-authorization',
|
||||
'te', 'trailer', 'transfer-encoding', 'upgrade', 'host', 'content-length']);
|
||||
|
||||
function forward(bodyBuf, req, res) {
|
||||
const headers = {};
|
||||
for (const [k, v] of Object.entries(req.headers)) {
|
||||
if (!HOP.has(k.toLowerCase())) headers[k] = v;
|
||||
}
|
||||
headers['content-length'] = bodyBuf.length;
|
||||
|
||||
const out = UP.mod.request(
|
||||
{ host: UP.host.split(':')[0], port: UP.host.split(':')[1] || (UP.mod === https ? 443 : 80),
|
||||
path: UP.base + req.url.replace(/^\/v1c/, ''), method: req.method, headers },
|
||||
(upRes) => {
|
||||
res.writeHead(upRes.statusCode, upRes.headers);
|
||||
upRes.pipe(res); // 流式原样回传:SSE 不受影响
|
||||
}
|
||||
);
|
||||
out.on('error', (e) => {
|
||||
if (!res.headersSent) {
|
||||
res.writeHead(502, { 'content-type': 'application/json' });
|
||||
res.end(JSON.stringify({ error: { code: 'upstream_unreachable', message: String(e && e.message || e) } }));
|
||||
} else {
|
||||
res.destroy();
|
||||
}
|
||||
});
|
||||
out.end(bodyBuf);
|
||||
}
|
||||
|
||||
function sendJson(res, status, obj) {
|
||||
const buf = Buffer.from(JSON.stringify(obj));
|
||||
res.writeHead(status, { 'content-type': 'application/json', 'content-length': buf.length });
|
||||
res.end(buf);
|
||||
}
|
||||
|
||||
// ---------- 主流程 ----------
|
||||
const server = http.createServer((req, res) => {
|
||||
if (req.method === 'GET' && req.url === '/health') {
|
||||
return sendJson(res, 200, { status: 'ok', service: 'ctx-relay', conversations: cache.size, bytes: cacheBytes });
|
||||
}
|
||||
|
||||
const chunks = [];
|
||||
let size = 0;
|
||||
const limit = CFG.max_body_mb * 1024 * 1024;
|
||||
req.on('data', (c) => {
|
||||
size += c.length;
|
||||
if (size > limit) {
|
||||
sendJson(res, 413, { error: { code: 'body_too_large', message: 'body > ' + CFG.max_body_mb + 'MB' } });
|
||||
req.destroy();
|
||||
return;
|
||||
}
|
||||
chunks.push(c);
|
||||
});
|
||||
req.on('end', () => {
|
||||
const raw = Buffer.concat(chunks);
|
||||
const ctxId = String(req.headers['x-context-id'] || '').trim();
|
||||
const ctxHash = String(req.headers['x-context-hash'] || '').trim();
|
||||
|
||||
let body;
|
||||
try {
|
||||
body = JSON.parse(raw.toString('utf8'));
|
||||
} catch (_) {
|
||||
return forward(raw, req, res); // 非 JSON 一律透明转发
|
||||
}
|
||||
|
||||
// 没带 context id,或模型没开 store → 原样转发
|
||||
if (!ctxId || !storeEnabled(body.model)) {
|
||||
return forward(raw, req, res);
|
||||
}
|
||||
|
||||
withLock(ctxId, () => {
|
||||
const delta = Array.isArray(body.messages) ? body.messages : [];
|
||||
const deltaBytes = byteLen(JSON.stringify(delta));
|
||||
const entry = get(ctxId);
|
||||
let messages;
|
||||
|
||||
if (ctxHash === '') {
|
||||
messages = delta; // 全量替换(首轮 / 压缩后 / 回退重发)
|
||||
} else if (entry && entry.hash === ctxHash) {
|
||||
messages = entry.messages.concat(delta);
|
||||
} else {
|
||||
log(ctxId + ' hash 不匹配,回 409(delta=' + delta.length + ')');
|
||||
return sendJson(res, 409, {
|
||||
error: { code: 'context_mismatch', message: 'context hash mismatch, resend full messages with empty X-Context-Hash' },
|
||||
});
|
||||
}
|
||||
|
||||
const saved = byteLen(JSON.stringify(messages)) - deltaBytes;
|
||||
put(ctxId, messages);
|
||||
log(ctxId + ' delta=' + delta.length + ' total=' + messages.length +
|
||||
' 上传 ' + fmtBytes(deltaBytes) + ' 省 ' + fmtBytes(saved > 0 ? saved : 0) +
|
||||
' 缓存 ' + fmtBytes(cacheBytes) + '/' + fmtBytes(CFG.store.max_total_bytes));
|
||||
|
||||
const outBody = Buffer.from(JSON.stringify(Object.assign({}, body, { messages })));
|
||||
forward(outBody, req, res);
|
||||
}).catch((e) => {
|
||||
if (!res.headersSent) sendJson(res, 500, { error: { code: 'relay_error', message: String(e && e.message || e) } });
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
server.listen(CFG.port, '0.0.0.0', () => {
|
||||
log('监听 0.0.0.0:' + CFG.port + ' → ' + CFG.upstream);
|
||||
log('store: ttl=' + (CFG.store.ttl_ms / 3600000) + 'h 单会话≤' + fmtBytes(CFG.store.max_bytes) +
|
||||
' 总量≤' + fmtBytes(CFG.store.max_total_bytes) + ' models=' + JSON.stringify(CFG.store.models));
|
||||
});
|
||||
Reference in New Issue
Block a user