第 12 关
会话
智能体的历史活在内存里,而内存随程序一起结束。会话会一路把它写下来,于是对话能比进程活得更久:你可以明天接着聊,也可以让它分叉。
1/34
手风琴褶数:
- user
- assistant
- tool_result
EventBus
问题
第 2 关的历史是内存里的一个列表。程序一停(一次部署、一次崩溃、有人关掉标签页明天再回来),列表就没了,智能体从零开始,把它早就知道的事情再问一遍。
只在最后保存也不够:一次跑到一半失败的运行,恰恰是你最想查看、或者想从断掉的地方接着跑的那一次。
解决方案
给每段对话一个 id 和一个磁盘上的位置,把历史当成那个文件,而不是一个变量。三个动作就够了:
-
边走边存
每条消息一出现就写进会话文件,一条消息一行。中途就算崩溃,也什么都不会丢。
-
继续
用同一个会话 id 创建的新智能体,会在下一个请求之前把文件读回它的历史里。
-
分叉
把历史复制到一个新会话里去试别的做法。原来的会话保持原样。
JSONL,也就是每行一条 JSON 消息,是常用的格式:加一条消息就是加一行,写了一半的最后一行也很容易发现。Claude Code 和 Codex 也是这样保存自己的会话的,它们就是靠这个来继续之前的对话。
角色
还是那群熟悉的角色,这次踏上了冒险。
- 冒险日志 会话文件
- 每个会话 id 一个存档栏,每条消息一个标记,一出现就写下。
- 手风琴 历史
- 智能体在内存里拥有的东西。游戏一关机它就空了;日志不会。
- CONTINUE 继续
- 用同一个 id 的新智能体,把日志读回它的历史里。
- COPY A QUEST 分叉
- 第二个存档栏,从第一个的副本开始。之后各自增长。
- Astor 循环
- 照常跑循环,每条消息都写进日志。
- 小镇 工具
- 长老、守卫和道具店:
talk_to和buy。
代码
使用 astorlm:传入一个 sessionId 和一个 FileSessionManager。智能体创建时会加载这个会话,之后每条消息都会保存;新的 id 会开一个空会话。fork() 会返回一个基于会话副本的新智能体。
从零手写:第 2 关的循环,先从一个 JSONL 文件读出历史,每条新消息再追加到文件末尾。分叉就是复制文件。
import { FileSessionManager, OpenAIProvider, createLocalAgent, tool } from 'astorlm'
import { z } from 'zod'
const talkTo = tool({
name: 'talk_to',
description: 'Talk to someone in town. Returns what they say.',
schema: z.object({ npc: z.string() }),
execute: async ({ npc }) => town.talk(npc), // your code
})
const buy = tool({
name: 'buy',
description: 'Buy an item at the shop. Returns the price and the gold left.',
schema: z.object({ item: z.string() }),
execute: async ({ item }) => shop.buy(item), // your code
})
// The adventure log: quest-1.jsonl (one message per line) and quest-1.meta.json, in this folder.
const sessionManager = new FileSessionManager({ dir: './sessions' })
const quest = (sessionId: string) =>
createLocalAgent({
// Any OpenAI-compatible endpoint: OpenAI, Ollama, LM Studio, vLLM, a proxy…
provider: new OpenAIProvider({
baseURL: 'http://localhost:11434/v1', // e.g. Ollama's default address
model: 'your-model', // e.g. 'llama3.1', 'gpt-4o-mini'
apiKey: 'YOUR_API_KEY', // local servers usually ignore it
}),
tools: [talkTo, buy],
sessionId, // a new id starts an empty log; a known one loads it
sessionManager,
})
// Day 1. Every message is saved as soon as it's added, so a crash loses nothing.
const day1 = await quest('quest-1')
await day1.run('I need to get into the Cave of Echoes. Find out what it takes, and get what you can.')
// …the process ends. Day 2: a new agent on the same id reads the log back before answering.
const day2 = await quest('quest-1')
console.log(day2.getMessages().length) // 6: yesterday is in the history
await day2.run('I got the Silver Key from the mayor’s daughter. What now?')
// A fork: a new session with a copy of every message. quest-1 stays as it was.
const west = await day2.fork({ newSessionId: 'quest-2' })
await west.run('Suppose the key doesn’t fit. Is there another way in?')
// Sessions from scratch: a JSONL file per conversation. Node's standard library only.
import { appendFileSync, copyFileSync, existsSync, mkdirSync, readFileSync } from 'node:fs'
type Message = { role: string; content: unknown; [key: string]: unknown }
const DIR = './sessions'
const pathOf = (id: string) => `${DIR}/${id}.jsonl`
// 1. Load: one JSON message per line. No file yet means a new, empty session.
export function load(id: string): Message[] {
if (!existsSync(pathOf(id))) return []
return readFileSync(pathOf(id), 'utf8')
.split('\n')
.filter(Boolean)
.map((line) => JSON.parse(line))
}
// 2. Save as you go: append each message the moment it exists. A crash mid-run loses nothing.
function append(id: string, message: Message): void {
mkdirSync(DIR, { recursive: true })
appendFileSync(pathOf(id), JSON.stringify(message) + '\n')
}
// 3. Fork: copy the file. From here on, each session only appends to its own.
export function fork(from: string, to: string): void {
copyFileSync(pathOf(from), pathOf(to))
}
// 4. The loop from level 2, with its history loaded first and every new message written down.
export async function runSession(id: string, prompt: string, maxTurns = 10): Promise<string> {
const messages = load(id)
const add = (message: Message) => {
messages.push(message)
append(id, message)
}
add({ role: 'user', content: prompt })
for (let turn = 1; turn <= maxTurns; turn++) {
const choice = await chat(messages) // one model call, as in level 2
add(choice.message)
if (choice.finish_reason !== 'tool_calls') return choice.message.content ?? ''
for (const call of choice.message.tool_calls ?? []) {
add({ role: 'tool', tool_call_id: call.id, content: await runTool(call) })
}
}
throw new Error(`No answer after ${maxTurns} turns`)
}
// Day 1, then day 2 on the same id, then a fork to try another way.
await runSession('quest-1', 'I need to get into the Cave of Echoes. Find out what it takes, and get what you can.')
await runSession('quest-1', 'I got the Silver Key from the mayor’s daughter. What now?')
fork('quest-1', 'quest-2')
await runSession('quest-2', 'Suppose the key doesn’t fit. Is there another way in?')
# Sessions from scratch: a JSONL file per conversation. Standard library only.
import json
import shutil
from pathlib import Path
DIR = Path("sessions")
def path_of(session_id):
return DIR / f"{session_id}.jsonl"
# 1. Load: one JSON message per line. No file yet means a new, empty session.
def load(session_id):
path = path_of(session_id)
if not path.exists():
return []
return [json.loads(line) for line in path.read_text(encoding="utf-8").splitlines() if line]
# 2. Save as you go: append each message the moment it exists. A crash mid-run loses nothing.
def append(session_id, message):
DIR.mkdir(exist_ok=True)
with path_of(session_id).open("a", encoding="utf-8") as log:
log.write(json.dumps(message) + "\n")
# 3. Fork: copy the file. From here on, each session only appends to its own.
def fork(source, target):
shutil.copyfile(path_of(source), path_of(target))
# 4. The loop from level 2, with its history loaded first and every new message written down.
def run_session(session_id, prompt, max_turns=10):
messages = load(session_id)
def add(message):
messages.append(message)
append(session_id, message)
add({"role": "user", "content": prompt})
for _ in range(max_turns):
choice = chat(messages) # one model call, as in level 2
add(choice["message"])
if choice["finish_reason"] != "tool_calls":
return choice["message"].get("content") or ""
for call in choice["message"].get("tool_calls", []):
add({"role": "tool", "tool_call_id": call["id"], "content": run_tool(call)})
raise RuntimeError(f"No answer after {max_turns} turns")
# Day 1, then day 2 on the same id, then a fork to try another way.
run_session("quest-1", "I need to get into the Cave of Echoes. Find out what it takes, and get what you can.")
run_session("quest-1", "I got the Silver Key from the mayor's daughter. What now?")
fork("quest-1", "quest-2")
run_session("quest-2", "Suppose the key doesn't fit. Is there another way in?")
注意事项
- 继续一个会话,就要把一切重发一遍。一段长对话的第 30 天,背着前面 29 天。会话负责保留历史;压缩它(第 9 关)才能让它不超出上下文窗口。
- 分叉复制的是历史,不是世界。火把只买过一次,但两个存档里都有它。工具发出的邮件、下的订单、写的文件,在每个分支里都是真的发生过。
- 继续时检查最后一行。如果程序死在工具调用和它的结果之间,日志会以一个没人回应的调用结尾,而大多数服务商会拒绝这种请求。把它删掉,或者给它补一个错误结果。
- 每个会话只能有一个写入者。两个进程同时往同一个文件追加,行会交错在一起。给每次运行一个自己的会话,或者加一把锁。
- 会话文件是个人数据。里面有用户说过的每一句话和每一个工具结果。决定好它们放在哪里、谁能读、什么时候删除。