Loop Engineering 落地 pi：从六块积木到可跑驱动器#

1
// pi-loop.ts — 在 pi 上跑出 Loop Engineering 的最小骨架
2
import { createAgentSession, SessionManager, AuthStorage, ModelRegistry } from "@earendil-works/pi-coding-agent";
3
import { readFileSync, writeFileSync, existsSync } from "node:fs";
4
import { execSync } from "node:child_process";
5

6
const GOAL   = process.argv[2] ?? "fix: single quote in company name crashes bill save";
7
const STATE  = "STATE.md";
8
const RUBRIC = "uv run pytest passes (all green); no syntax errors; diff <= 200 lines";
9
const MAX_ITERS = 20;
10

11
// 反馈信号——loop 的命脉（客观二元判定，不是 agent 自评）
12
function feedback() {
13
  try {
14
    const out = execSync("uv run pytest", { encoding: "utf-8", cwd: process.cwd() });
15
    return { pass: /passed/.test(out) && !/failed/.test(out), out };
16
  } catch (e: any) { return { pass: false, out: String(e) }; }
17
}
18
const loadMemory = () => existsSync(STATE) ? readFileSync(STATE, "utf-8") : "(no prior)";
19
const saveMemory = (i: number, v: string) =>
20
  writeFileSync(STATE, `\n## iter ${i}\nverdict: ${v}\n`, { flag: "a" });
21

22
// 认证与模型注册表——maker 和 checker 共用一份
23
const authStorage   = AuthStorage.create();
24
const modelRegistry  = ModelRegistry.create(authStorage);
25
// 用 registry.find 拿模型（接受任意字符串 id；getModel 要字面量联合类型，不实用）
26
const resolveModel = (id: string) => modelRegistry.find("anthropic", id)!;
27

28
// 生成器——一个落盘的 pi session（session.jsonl 就是它的记忆外存）
29
const { session: maker } = await createAgentSession({
30
  cwd: process.cwd(),
31
  sessionManager: SessionManager.create(process.cwd()),
32
  tools: ["read", "bash", "edit", "write"],
33
  model: resolveModel(process.env.LOOP_MAKER_MODEL ?? "claude-opus-4-8"),
34
  authStorage,
35
  modelRegistry,
36
});
37

38
for (let i = 0; i < MAX_ITERS; i++) {
39
  const { pass, out } = feedback();
40
  if (pass) { console.log(`done in ${i} iterations`); break; }
41

42
  // maker：读 STATE.md 接力上一轮，做最小修复
43
  await maker.prompt(`Goal: ${GOAL}
44

45
Prior attempts and rejected approaches (from STATE.md):
46
${loadMemory()}
47

48
Latest test feedback:
49
${out}
50

51
Diagnose the ROOT cause (not symptom), make the smallest fix,
52
do not refactor unrelated code.`);
53

54
  // 验证器——独立 session + 换模型 + 只读工具，绝不让 maker 批自己的作业
55
  const verdict = await verifier(GOAL, RUBRIC, out);
56
  saveMemory(i, verdict);                  // 落盘——agent 会忘，仓库不会
57
  if (verdict.startsWith("PASS")) { console.log(`passed at iter ${i}`); break; }
58
}
59
maker.dispose();
60

61
async function verifier(goal: string, rubric: string, testOut: string) {
62
  const { session: checker } = await createAgentSession({
63
    sessionManager: SessionManager.inMemory(),   // 干净上下文，不继承 maker 的对话
64
    tools: ["read", "grep"],                       // 只读——checker 不改代码，只评分
65
    model: resolveModel(process.env.LOOP_CHECKER_MODEL ?? "claude-haiku-4-5"), // 故意换一个模型
66
    authStorage,
67
    modelRegistry,
68
  });
69
  const chunks: string[] = [];
70
  const unsub = checker.subscribe(e => {
71
    if (e.type === "message_update" && e.assistantMessageEvent.type === "text_delta")
72
      chunks.push(e.assistantMessageEvent.delta);
73
  });
74
  await checker.prompt(`Goal: ${goal}
75
Rubric: ${rubric}
76
Test output:
77
${testOut.slice(0, 4000)}
78

79
Does the current repo meet the rubric? Reply ONLY 'PASS' or 'FAIL: <reason>'.`);
80
  unsub();
81
  checker.dispose();
82
  return chunks.join("");
83
}

1
pi-loop/
2
├── pi-loop.ts          # 上面那段驱动器
3
├── extensions/
4
│   ├── cron-loop.ts    # Automation 扩展
5
│   └── worktree-per-turn.ts  # Worktree 扩展
6
├── skill/
7
│   └── SKILL.md        # 项目约定固化（"最小修复、不重构无关代码"等）
8
├── app/
9
│   ├── bill.py         # 故意写坏的：单引号公司名崩 save
10
│   ├── test_bill.py    # 会被 loop 反复跑的客观信号
11
│   └── STATE.md        # 跨轮进度（loop 自己读写）
12
└── package.json