14c79dbfdd
两段式第二段的 regenerate-module 类:改玩法 = 据 intent 有界重写 game-logic.js,不动其余。
复用机制(参数化、不复制 resume/熔断/收口):
- cheap_studio.run_studio 加 4 个默认 None 可选参(system_prompt/initial_kick/write_whitelist/prepare)
切到 modify 态,create 路零行为变化(params=None 等价原逻辑)
- cheap_toolkit.build_toolkit 加 write_whitelist(非白名单 basename 写直接拒)
- cheap_roles.build_modify_system_prompt(复用 create 红线契约块)
cheap_modify.execute_regenerate_modify:
- 取 intent(空→failed)→ materialize base 源 → 重写前后对非目标文件算 hash
- 有界重写(写边界收窄到只 game-logic.js)→ 九门
- status=succeeded ⟺ 九门过 ∧ 非目标稳;manifest{file,kind:behavior,intent,changed,untouchedStable}(断言②地基)
worker_service._process_regenerate_job(regen_fn):把 M3 的 regenerate-module 显式 failed 占位换成真执行;
deterministic/create/生成路一字未动。
测试:test_a11_m4_regenerate 10/10(全注入桩零 LLM)+ 全回归独立复跑 73 测试全绿。
一次真 LLM smoke(单跑):intent=连击递增 → status=succeeded、untouchedStable=True、九门 pass、
attempts=1、¥0.31、137s;独立 diff 佐证 5 非目标文件字节相同、仅 game-logic.js 变。
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
276 lines
16 KiB
Python
276 lines
16 KiB
Python
"""
|
||
cheap_studio.py — 便宜档 ReAct 主编排（spike 核心；对照源：Node gen.mjs 主循环 + tier2 studio.py:run_studio）。
|
||
|
||
组装：内置 Agent + ReActConfig + 外层有界 resume 循环（照 studio.py:404-458）。done 由 cheap_toolkit 的 finish
|
||
工具触发（工具内复核 check+build 绿）；finish 收敛后收口 stage → smoke → ensure_play_spec → 循环外九门 play。
|
||
熔断（CircuitBreakerMiddleware）/ trace（Tier2TraceMiddleware）/ 历史压缩（ContextConfig）经 import 复用 tier2，零重写。
|
||
|
||
为什么要外层 resume：AgentScope 2.0.2 内置 ReAct 在"模型产出无 tool_call 的纯文本回合"即退出（看一次门没绿就停）。
|
||
外层在 agent 停下后，若【没真 finish】且【门没绿】且【还有预算】就带反馈再 reply()——跨 reply 保留 memory = 原地续修。
|
||
这一圈正对应 Node gen.mjs 的 while 主循环。
|
||
|
||
CLI：cheap-worker/.venv/bin/python cheap-worker/cheap_studio.py --id cheap-run1 --brief "一个简单的点击得分小游戏"
|
||
"""
|
||
|
||
import json
|
||
import sys
|
||
import time
|
||
from pathlib import Path
|
||
|
||
# 跨包 import 兜底 + key 注入（_bootstrap 模块级把 tier2/gen-worker 加进 sys.path）。
|
||
sys.path.insert(0, str(Path(__file__).resolve().parent)) # → cheap-worker/（CLI 直跑兼容）
|
||
import _bootstrap # noqa: E402,F401
|
||
from cheap_roles import build_system_prompt # noqa: E402
|
||
from cheap_toolkit import CheapSession, build_toolkit # noqa: E402
|
||
import cheap_run # noqa: E402
|
||
|
||
# tier2 框架层（import config 触发代理旁路 + 加载 agentscope，必须在裸 import agentscope/openai 之前）。
|
||
from worker import config # noqa: E402
|
||
from worker.middleware import ( # noqa: E402
|
||
CircuitBreakerMiddleware, Tier2TraceMiddleware, Tier2CircuitBreak,
|
||
)
|
||
# AgentScope（此时 agentscope 已由 config import 链加载、代理旁路已装）。
|
||
from agentscope.agent import Agent, ReActConfig # noqa: E402
|
||
from agentscope.message import UserMsg # noqa: E402
|
||
from agentscope.state import AgentState # noqa: E402
|
||
from agentscope.permission import PermissionContext, PermissionMode # noqa: E402
|
||
|
||
|
||
def _rec(msg: str) -> None:
|
||
"""进度日志走 stderr（保 CLI stdout 末行可作机读 JSON）。"""
|
||
print(f"[cheap_studio] {msg}", file=sys.stderr, flush=True)
|
||
|
||
|
||
def _bypass_state() -> AgentState:
|
||
"""无人值守跳过工具 ASK（否则 FunctionTool 默认 check_permissions 返回 ASK 卡死；对 studio.py:_bypass_state）。"""
|
||
return AgentState(permission_context=PermissionContext(mode=PermissionMode.BYPASS))
|
||
|
||
|
||
def _user_msg(text: str) -> UserMsg:
|
||
"""2.0.1+ UserMsg 必须带 name（对 studio.py:user_msg）。"""
|
||
return UserMsg(name="user", content=text)
|
||
|
||
|
||
def _verdict_brief(v) -> dict:
|
||
"""从九门 verdict.json 抽 {pass, failedGates}。"""
|
||
if not isinstance(v, dict):
|
||
return {"pass": None, "failedGates": None}
|
||
guards = v.get("guards") or {}
|
||
failed = [k for k, g in guards.items() if isinstance(g, dict) and g.get("pass") is False]
|
||
return {"pass": v.get("pass"), "failedGates": failed}
|
||
|
||
|
||
def _furthest_stage(*, finished, staged, smoke_ran, played) -> str:
|
||
"""据收口流程实际到达的步，判本 run 走到的最远阶段（scaffold/code/stage/smoke/play）。
|
||
|
||
取最远到达者：scaffold 恒成功（失败提前返回、不到此）；finished=code 已收敛；staged=esbuild 打包成功；
|
||
smoke_ran=跑过冒烟（无论过否）；played=跑过九门 play（无论过否）。供 9d trace.stage（后端审核台读）。
|
||
"""
|
||
if played:
|
||
return "play"
|
||
if smoke_ran:
|
||
return "smoke"
|
||
if staged:
|
||
return "stage"
|
||
if finished:
|
||
return "code"
|
||
return "scaffold"
|
||
|
||
|
||
def build_trace_source(verdict, driver_type, attempts, stage, model) -> dict:
|
||
"""组 run-summary 的 9d-trace 源维度（additive 纯函数·可单测；run_studio 收口段调）。
|
||
|
||
把收口段已算出但没存进 summary 的维度结构化产出，供 result_out 组 payload["trace"]、
|
||
后端 D11 ReadinessScorer 据此算就绪分。纯字典构造、零外部依赖、同输入恒同输出。
|
||
|
||
Args:
|
||
verdict: 完整九门 verdict（含 guards，其中 guards.H_progress 是嵌套对象 {pass,checks,latch,...}）；
|
||
play 没跑（finished=False / run_gates=False / smoke 失败）时传 None。
|
||
driver_type: ensure_play_spec 产的 driver 类型（ps.get("driverType")）；没产 play-spec 时 None。
|
||
attempts: resume 轮数（int，=summary 既有字段）。本函数接收以构成完整 trace-源契约，但 repairs 由
|
||
result_out 据 summary.attempts 派生（max(0,attempts-1)），故此处**不回写 attempts**
|
||
（守「不动既有键」）。
|
||
stage: 本 run 走到的最远阶段（scaffold/code/stage/smoke/play 之一，_furthest_stage 产）。
|
||
model: 终态模型名（便宜档单模型 = _bootstrap.SPIKE_MODEL）。
|
||
|
||
Returns:
|
||
dict（仅新增键，供 summary.update）：verdictFull / driverType / models / stage。
|
||
"""
|
||
return {
|
||
# 完整 verdict（保留 guards 嵌套，含 H_progress.pass）；play 没跑 → None。U2 在 result_out 内削成 {pass,guards}。
|
||
"verdictFull": verdict if isinstance(verdict, dict) else None,
|
||
# driver 类型（ensure_play_spec 产）；没产 play-spec（对照路 run_gates=False）→ None。
|
||
"driverType": driver_type,
|
||
# 便宜档单模型 → {code: SPIKE_MODEL} map（镜像 wg1 models 多模型 map 形态）。
|
||
"models": {"code": model},
|
||
# 最远阶段。
|
||
"stage": stage,
|
||
}
|
||
|
||
|
||
async def run_studio(game_id, brief, *, max_iters=40, max_resumes=6, max_tokens=16000,
|
||
port=4320, cdp_port=9222, run_gates=True,
|
||
system_prompt=None, initial_kick=None, write_whitelist=None, prepare=None):
|
||
"""跑便宜档一局生成（scaffold → ReAct 写 src/ → done 门 → 收口 stage+smoke+九门）。返回 run-summary dict。
|
||
|
||
run_gates=True（默认，spike 单局）：收口跑全套 stage→smoke→ensure_play_spec→九门 play。
|
||
run_gates=False（generation-only，给 compare_node 对照用）：收口只到 stage+smoke 为止，**不**自动产 play-spec、
|
||
**不**跑内部九门 play —— 由调用方在生成与 play 之间注入金标 spec 再单独 play，把驱动器从对照变量里摘掉
|
||
（Codex C1：run_studio 内部已 play，对照需把生成与 play 拆开）。
|
||
|
||
A11 M4 模块重生成复用本编排（同一 ReAct + resume + 熔断 + 三层校验收口，不另造），靠四个可选参切到 modify 态，
|
||
都默认 None=create 原行为（create 路零改动）：
|
||
· system_prompt：None=create 的 build_system_prompt；modify 传 build_modify_system_prompt（只改玩法）。
|
||
· initial_kick：None=create 的「按 brief 造游戏」首条 kick；modify 传「只重写 game-logic.js 实现 intent」。
|
||
· write_whitelist：None=create 不收窄；modify 传 {"game-logic.js"} 把写边界收窄到只许写玩法文件。
|
||
· prepare：None=create 的 scaffold（clone _template）；modify 传「已由上游 scaffold+materialize base 源」
|
||
的 noop（game_dir 已就绪、不要再 scaffold 覆盖掉 base 源）。签名同 scaffold：(game_id)->{ok,output}。
|
||
"""
|
||
t0 = time.time()
|
||
_rec(f"model={_bootstrap.SPIKE_MODEL} id={game_id} brief=「{brief}」max_iters={max_iters} max_resumes={max_resumes}")
|
||
|
||
prep = prepare or cheap_run.scaffold # create=scaffold clone _template；modify=上游已预备的 noop
|
||
sc = prep(game_id)
|
||
if not sc["ok"]:
|
||
return {"ok": False, "gameId": game_id, "fail": "准备失败：" + sc["output"]}
|
||
_rec(f"准备就绪 amgen-{game_id}（{'scaffold clone _template + SAA 信封' if prepare is None else 'modify: base 源已预备'}）")
|
||
|
||
session = CheapSession(game_id=game_id)
|
||
toolkit = build_toolkit(session, write_whitelist=write_whitelist) # modify 收窄到只许写 game-logic.js
|
||
model = _bootstrap.build_cheap_model(max_tokens=max_tokens)
|
||
# 熔断：便宜档开 ¥ 累进硬闸（M1 U4 / KTD5）+ 设便宜档专属 ¥10 硬上限（区别 tier2 的 ¥3 genconfig 默认）。
|
||
# on_model_call 在每次裸模型调用前按「已花 + 本次预估」判、越线即 fail-closed 抛 Tier2CircuitBreak(budget)；
|
||
# 取价不可达时 _ensure_pricing 降级为次数闸（既有行为，不静默超支、不静默阻断）。步数/超时仍用 tier2 默认。
|
||
breaker = CircuitBreakerMiddleware(enable_rmb_gate=True, rmb_hard_limit=10.0)
|
||
tracer = Tier2TraceMiddleware(trace_id=game_id)
|
||
|
||
writer = Agent(
|
||
name="cheap-writer",
|
||
system_prompt=system_prompt or build_system_prompt(game_id), # modify 传 build_modify_system_prompt
|
||
model=model,
|
||
toolkit=toolkit,
|
||
middlewares=[tracer, breaker], # trace 外层、熔断内层（列表序=洋葱序）
|
||
state=_bypass_state(),
|
||
react_config=ReActConfig(max_iters=max_iters),
|
||
context_config=config.build_context_config(), # 历史压缩
|
||
)
|
||
|
||
kick = initial_kick or (
|
||
f"请按这个 brief 造一款游戏：「{brief}」。先 read_file 读手册（.agents/skills/littlejs-game-dev.md）"
|
||
"和你的起点 game-logic.js 再动手；核心玩法实现完、check 与 build 都绿了就立即 finish。")
|
||
breaker_tripped = None
|
||
attempts = 0
|
||
try:
|
||
for attempt in range(max_resumes + 1):
|
||
attempts = attempt + 1
|
||
_rec(f"resume attempt {attempts}/{max_resumes + 1} → writer.reply …")
|
||
await writer.reply(_user_msg(kick))
|
||
# ① 真 finish → 收敛退出
|
||
if session.finished is not None:
|
||
_rec("finish 已接受（check+build 绿）→ 收敛")
|
||
break
|
||
# ② 门已绿但没 finish → 踹一脚让它 finish
|
||
c_ok = bool(session.last_check and session.last_check.get("ok"))
|
||
b_ok = bool(session.last_build and session.last_build.get("ok"))
|
||
if c_ok and b_ok:
|
||
_rec("门绿但未 finish → 踹 finish")
|
||
kick = "check 与 build 都已 PASS。现在直接调 finish 交付（summary 一句话），不要再改。"
|
||
continue
|
||
# ③ resume 预算耗尽 → 停
|
||
if attempt >= max_resumes:
|
||
_rec("resume 预算耗尽，停")
|
||
break
|
||
# ④ 门没绿 + agent 自停 → 带失败反馈再 reply（跨 reply memory 保留=原地续修）
|
||
fb = "尚未跑出绿 check/build。"
|
||
if session.last_check and not session.last_check.get("ok"):
|
||
fb = "上次 check 失败：\n" + (session.last_check.get("output") or "")[:1500]
|
||
elif session.last_build and not session.last_build.get("ok"):
|
||
fb = "上次 build 失败：\n" + (session.last_build.get("output") or "")[:1500]
|
||
_rec(f"门没绿 + agent 自停 → 带反馈再 reply（attempt {attempts}）")
|
||
kick = (f"你刚才停下了，但 check/build 还没全绿——不要放弃。{fb}\n"
|
||
"请据失败 write_file 针对性修，再 check → build，直到都绿再 finish。先修最关键的错。")
|
||
except Tier2CircuitBreak as e:
|
||
breaker_tripped = {"kind": e.kind, "reason": e.reason}
|
||
_rec(f"熔断 CircuitBreak：kind={e.kind} reason={e.reason}")
|
||
except Exception as e: # noqa: BLE001 顶层兜底，落 breaker 信息便于诊断
|
||
breaker_tripped = {"kind": None, "reason": f"{type(e).__name__}: {e}"}
|
||
_rec(f"未捕获异常：{type(e).__name__}: {e}")
|
||
|
||
# ── 收口：finish 后 stage → smoke → ensure_play_spec → 循环外九门 play ──
|
||
finished = session.finished is not None
|
||
staged = False
|
||
smoke_ok = None
|
||
verdict = None
|
||
driver_type = None # M3b U1：ensure_play_spec 产的 driver 类型（对照路 run_gates=False 不产 → 保持 None）
|
||
if finished:
|
||
st = cheap_run.stage(game_id)
|
||
staged = st["ok"]
|
||
_rec(f"stage {'OK' if staged else 'FAIL: ' + st['output']}")
|
||
if staged:
|
||
sm = cheap_run.smoke(game_id, port=port, cdp_port=cdp_port)
|
||
smoke_ok = sm["ok"]
|
||
_rec(f"smoke {'PASS' if smoke_ok else 'FAIL'}（抓 state 供 play-spec）")
|
||
if run_gates:
|
||
ps = cheap_run.ensure_play_spec(game_id, sm.get("state"))
|
||
driver_type = ps.get("driverType") # M3b U1：driver 类型存进 trace.gatespec.driver（后端 D11 firstPlay 维）
|
||
_rec(f"play-spec {'已产 driver=' + ps.get('driverType', '?') if ps.get('wrote') else ps.get('reason', '?')}")
|
||
pr = cheap_run.play(game_id, port=port, cdp_port=cdp_port)
|
||
verdict = pr["verdict"]
|
||
vb = _verdict_brief(verdict)
|
||
_rec(f"九门 play pass={vb['pass']} failedGates={vb['failedGates']}")
|
||
else:
|
||
_rec("generation-only（run_gates=False）：跳过 ensure_play_spec + 九门 play，交对照方注入金标 spec 后单独 play")
|
||
|
||
tok_in, tok_out = model.usage_sum() # 成本落盘(R4)：token 台账始终可取(RecordingOpenAIChatModel.records 累计)。
|
||
summary = {
|
||
"ok": finished,
|
||
"gameId": game_id,
|
||
"brief": brief,
|
||
"model": _bootstrap.SPIKE_MODEL,
|
||
"finished": finished,
|
||
"attempts": attempts,
|
||
"check": {"ok": bool(session.last_check and session.last_check.get("ok"))} if session.last_check else None,
|
||
"build": {"ok": bool(session.last_build and session.last_build.get("ok"))} if session.last_build else None,
|
||
"staged": staged,
|
||
"smoke": {"ok": smoke_ok} if smoke_ok is not None else None,
|
||
"verdict": _verdict_brief(verdict) if verdict is not None else None,
|
||
"breaker": breaker_tripped,
|
||
# 成本落盘(R4)：¥ 走 breaker 同口径(new-api quota 折价 cost.compute)；取价不可达时 spent_rmb=0 且
|
||
# rmbGate=degraded 标识(_ensure_pricing 已打降级日志可查)——不静默超支、不静默阻断。
|
||
"costRmb": round(breaker.spent_rmb, 4),
|
||
"rmbGate": "active" if breaker._rmb_gate_active else "degraded",
|
||
"tokens": {"in": tok_in, "out": tok_out, "total": tok_in + tok_out},
|
||
"doneSummary": session.finished.get("summary") if session.finished else None,
|
||
"wallSec": round(time.time() - t0, 1),
|
||
}
|
||
# ── M3b U1：additive 富化 9d trace 源维度（verdictFull/driverType/models/stage）──
|
||
# 收口段已算出但没存的维度结构化进 summary，供 result_out 组 payload["trace"]、后端 D11 算分；
|
||
# 全 additive（新增键、不动既有键），对照路（run_gates=False）verdict/driver_type 为 None 时各键自然降级。
|
||
stage = _furthest_stage(finished=finished, staged=staged,
|
||
smoke_ran=(smoke_ok is not None), played=(verdict is not None))
|
||
summary.update(build_trace_source(verdict, driver_type, attempts, stage, _bootstrap.SPIKE_MODEL))
|
||
ev_dir = cheap_run.game_dir(game_id) / "evidence"
|
||
ev_dir.mkdir(parents=True, exist_ok=True)
|
||
(ev_dir / "run-summary.json").write_text(json.dumps(summary, ensure_ascii=False, indent=2), encoding="utf-8")
|
||
_rec(f"run-summary 落盘 → {ev_dir / 'run-summary.json'} wallSec={summary['wallSec']}")
|
||
# 注：n=5 收敛环批跑（RunRecord/batch_run）是 plan deferred；spike 单局 run-summary 已含等价字段。
|
||
return summary
|
||
|
||
|
||
if __name__ == "__main__":
|
||
import argparse
|
||
import asyncio
|
||
|
||
ap = argparse.ArgumentParser(description="便宜档 ReAct 主编排（spike 单局）")
|
||
ap.add_argument("--id", default="cheap-run1")
|
||
ap.add_argument("--brief", default="一个简单的点击得分小游戏")
|
||
ap.add_argument("--max-iters", type=int, default=40)
|
||
ap.add_argument("--max-resumes", type=int, default=6)
|
||
ap.add_argument("--max-tokens", type=int, default=16000)
|
||
a = ap.parse_args()
|
||
result = asyncio.run(run_studio(a.id, a.brief, max_iters=a.max_iters,
|
||
max_resumes=a.max_resumes, max_tokens=a.max_tokens))
|
||
# stdout 末行单行 JSON（U6 对照 shell-out 据此解析）。
|
||
print(json.dumps(result, ensure_ascii=False))
|
||
sys.exit(0 if result.get("ok") else 1)
|