9bd428040f
n=5 收敛环 iteration 3 坐实:上一版 fix①-B(on_acting 拦截畸形参数)日志 0 触发=无效。 源码核实根因:AgentScope _agent.py:1376 jsonschema.validate 失败 → _handle_error_tool_call → return(Step1),在 Step3 工具执行/on_acting 钩子之前,故畸形调用【根本不进 on_acting】, 上一版是死代码(单测 mock next_handler 吐错误、但线上错误不经它)。已 revert。 重实挂到真能触达 agent 的层:cheap_studio resume 续修循环 wrap writer.reply,catch Tier2CircuitBreak——【仅 missing-param 类 stuck】(reason 含 'is a required property'=MiniMax 漏 必填 path 撞满阈值)不当终态杀,还有 resume 预算 + 未越 ¥ 软停就带贴脸补参反馈续修(与 gate-fail resume 同机制、跨 reply memory 保留=原地续修)。只清 stuck 连击计数(_fail_repeat/_last_fail_sig, 不动 spent_rmb/tool_calls 防误清预算);非漏参 stuck(反复 check 红线/build 编译错=真死圈)/budget/ 其它熔断照旧向上抛终态、不放行。整环验证随 n=5 re-run(iteration 4)。 Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
477 lines
32 KiB
Python
477 lines
32 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 2.0.3 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 hashlib
|
||
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_budget # noqa: E402 预算两段式同源工厂(CLI 与生产 Service 读同一配置源)
|
||
import cheap_run # noqa: E402
|
||
import cheap_verify # noqa: E402 U-A2：便宜档「丰富度」LLM 验证 agent（非阻塞·只报告）
|
||
|
||
# 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.gate_judge import judge_cheap_verdict # noqa: E402 CLI 回喂:九门判据与生产 RepairMiddleware 同源
|
||
from worker.middleware import ( # noqa: E402 breaker 本体经 cheap_budget 同源工厂构造,此处只用 trace + 熔断异常
|
||
Tier2TraceMiddleware, Tier2CircuitBreak,
|
||
)
|
||
# C1b trace 落库：复用 tier2 observability 的 JSONL sink（cheap-worker 不另造 sink）。
|
||
# 非阻塞铁律：导入失败只告警，回落到无 sink 模式，绝不阻断生成主链。
|
||
try:
|
||
from observability.trace import make_jsonl_sink as _make_jsonl_sink # noqa: E402
|
||
_TRACE_SINK_AVAILABLE = True
|
||
except Exception as _import_err:
|
||
_TRACE_SINK_AVAILABLE = False
|
||
print(f"[cheap_studio] trace sink import 失败（回落无 sink 模式，不影响生成）：{_import_err}",
|
||
flush=True)
|
||
# T5 压缩崩封口(CLI 路同样吃到:压缩失败 fail-open 跳过、不杀生成主链;钉 2.0.3,纪律同 m3_stream_patch)。
|
||
from ctx_compress_patch import apply_ctx_compress_patch # noqa: E402
|
||
apply_ctx_compress_patch()
|
||
# 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)
|
||
|
||
|
||
# fix③(死圈修复 2026-07-08):brief 落 evidence/brief.json 供根因取证。
|
||
# 为什么单独落而不只靠 run-summary.brief:① 早落(scaffold 后即写)——run 硬失败/挂死时 run-summary 收不了口、
|
||
# brief.json 已在盘上,取证不断链;② 带 sha256 完整性戳(核 worker 收到的 brief 与后端下发字节一致,查"是不是 brief
|
||
# 被截断/篡改害的生成翻车");③ 限长脱敏(超 PROMPT_MAX 截断 + truncated 标记),防超长 brief 撑爆落盘。
|
||
# 落 evidence/(forensics 目录,不随 src/ 进素材市场交易),绝不落工程根——源工程会被交易,brief 是用户私有输入、不得泄进可交易源。
|
||
_BRIEF_PROMPT_MAX = 1000 # brief 落盘限长(字符);超出截断 + truncated=True,sha256 仍按【完整原文】算(截断不掩盖完整性核对)
|
||
|
||
|
||
def _persist_brief_forensics(game_id: str, brief: str, source: str = "job.brief") -> None:
|
||
"""把本局 brief 落 evidence/brief.json(取证用,见上方注释)。非阻塞:任何异常只告警、绝不咬生成主链。"""
|
||
try:
|
||
text = brief or ""
|
||
raw = text.encode("utf-8")
|
||
rec = {
|
||
"brief": text[:_BRIEF_PROMPT_MAX], # 限长:超长只留前 N 字符(脱敏 + 防撑爆)
|
||
"bytes": len(raw), # 完整 brief 的 UTF-8 字节数
|
||
"sha256": hashlib.sha256(raw).hexdigest(), # 完整 brief 摘要:核 worker↔后端字节一致
|
||
"truncated": len(text) > _BRIEF_PROMPT_MAX, # 是否因限长被截
|
||
"promptMax": _BRIEF_PROMPT_MAX,
|
||
"source": source, # 来源溯源(默认 job.brief)
|
||
}
|
||
ev_dir = cheap_run.game_dir(game_id) / "evidence"
|
||
ev_dir.mkdir(parents=True, exist_ok=True)
|
||
(ev_dir / "brief.json").write_text(json.dumps(rec, ensure_ascii=False, indent=2), encoding="utf-8")
|
||
_rec(f"brief 落盘取证 → {ev_dir / 'brief.json'} bytes={rec['bytes']} truncated={rec['truncated']}")
|
||
except Exception as e: # noqa: BLE001 非阻塞铁律:取证落盘绝不影响生成主链
|
||
_rec(f"brief 落盘异常(已忽略,不影响生成):{type(e).__name__}: {e}")
|
||
|
||
|
||
# fix①-B(重实):missing-param 类 stuck 续修时给 agent 的贴脸补参反馈(比默认 schema 校验错更明确,并催改用 edit_file 短载荷)。
|
||
_MALFORMED_RESUME_FEEDBACK = (
|
||
"【重要·纠偏】你刚才反复漏了工具调用的必填参数(通常是 path),调用被拒并触发了熔断。"
|
||
"重发工具调用时【必须一次带全参数】:write_file 要 {path, content}、edit_file 要 {path, old, new}——path 绝不能省。"
|
||
"小改强烈建议用 edit_file(只发变化的几行、参数短、最不易漏 path),别对整个 game-logic.js 用 write_file"
|
||
"(内容越长越容易把 path 漏掉)。现在:先 read_file 看清你要改的文件路径,再带【完整参数】重发修改,然后 check → build → finish。"
|
||
)
|
||
|
||
|
||
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 _closeout_gates(game_id, *, port, cdp_port) -> dict:
|
||
"""CLI 收口门流水线 stage → smoke → ensure_play_spec → 九门 play(一次跑齐;回喂循环与收口段共用)。
|
||
|
||
与 Service 路 cheap_gates.run_cheap_gates 同序,多回传 staged/smoke_ok/driver_type 供 run-summary 组装
|
||
(Service 路这些由 collector/driver 另采)。起 play 前清残留 verdict(红线③,镜像 cheap_gates.py:28):
|
||
serve-and-play.sh 早退/超时不写新 verdict 也不删旧的,不清则读回上一轮绿 verdict → 假绿。
|
||
"""
|
||
out = {"staged": False, "smoke_ok": None, "driver_type": None, "verdict": None}
|
||
st = cheap_run.stage(game_id)
|
||
out["staged"] = st["ok"]
|
||
if not st["ok"]:
|
||
_rec(f"stage FAIL: {st['output'][:300]}")
|
||
return out
|
||
sm = cheap_run.smoke(game_id, port=port, cdp_port=cdp_port)
|
||
out["smoke_ok"] = sm["ok"]
|
||
_rec(f"smoke {'PASS' if sm['ok'] else 'FAIL'}(抓 state 供 play-spec)")
|
||
if not sm["ok"]:
|
||
# smoke 失败真因必须可见(2026-07-04 bake_off 实证:并发争用致 100% FAIL→全场 key-cycle 回退考卷,
|
||
# 而失败原因被吞在 raw 里,烧掉一轮诊断)——打印 raw 尾供归因。
|
||
_rec(f"smoke raw 尾:{(sm.get('raw') or '')[-300:]}")
|
||
ps = cheap_run.ensure_play_spec(game_id, sm.get("state"))
|
||
# driver 类型:本轮新产的直接带;复用旧 spec(hash 一致)时读盘补(spec 是判卷单一来源)。
|
||
dt = ps.get("driverType")
|
||
if not dt:
|
||
try:
|
||
spec = json.loads((cheap_run.wg1_game_dir(game_id) / "play-spec.json").read_text(encoding="utf-8"))
|
||
dt = (spec.get("driver") or {}).get("type")
|
||
except Exception: # noqa: BLE001 —— 读不出只降级 None
|
||
dt = None
|
||
out["driver_type"] = dt
|
||
_rec(f"play-spec {'已产 driver=' + str(dt) if ps.get('wrote') else ps.get('reason', '?')}")
|
||
(cheap_run.wg1_game_dir(game_id) / "evidence" / "verdict.json").unlink(missing_ok=True) # 红线③:清陈旧 verdict
|
||
pr = cheap_run.play(game_id, port=port, cdp_port=cdp_port)
|
||
out["verdict"] = pr["verdict"]
|
||
return out
|
||
|
||
|
||
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,
|
||
scaffold_template=None, scaffold_desc=None, verify_richness_enabled=True,
|
||
genre=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。
|
||
|
||
genre（W-GENRE 件④，缺省 None=原行为）：品类键（如 'puzzle'）显式透传给丰富度 LLM 评分——同一次
|
||
judge 调用 additive 追加品类扩展条目（分母独立小计落 richness.genre）；None → 按 scaffold_template
|
||
查 GENRE_BY_TEMPLATE 推断；两路都没有/条目缺 → 自动退化为纯通用底座评分。
|
||
|
||
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}")
|
||
|
||
# create=scaffold clone 模板(扩模板:scaffold_template 传 per-genre 黄金骨架名,如经营=_template-shop)；modify=上游已预备的 noop
|
||
prep = prepare or (lambda gid: cheap_run.scaffold(gid, scaffold_template))
|
||
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'}）")
|
||
_persist_brief_forensics(game_id, brief) # fix③:scaffold 后即落 brief 取证(硬失败/挂死也留得下,不进可交易工程根)
|
||
|
||
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:便宜档预算两段式(W-ARCH②,创始人 2026-07-03 裁决)——软停线 ¥10(越线只许收尾类动作,
|
||
# on_acting 拦新增生成调用)+ RMB 硬地板 ¥15(×1.5,fail-closed 封死上界)。经 cheap_budget 同源工厂,
|
||
# 与生产 Service(cheap_service_app)读同一 genconfig 配置源(取代旧 hard 单段 ¥10,CLI/生产口径统一)。
|
||
# 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 = cheap_budget.build_cheap_breaker()
|
||
# C1b trace 落库：把生成轨迹按五字段（traceId/step/cost/verdict/timestamp + ext）
|
||
# 追加写到产物目录 games/amgen-<id>/trace.jsonl，供成本对账 / replay 用。
|
||
# 非阻塞铁律：接线异常只告警 + 回落无 sink 模式，绝不阻断生成主链。
|
||
# 为什么落盘路径选产物目录：与 run-summary.json / evidence/ 同目录，随产物一起留存，运维按 gameId 定位直观。
|
||
_trace_sink = None
|
||
if _TRACE_SINK_AVAILABLE:
|
||
try:
|
||
_trace_path = cheap_run.game_dir(game_id) / "trace.jsonl"
|
||
_trace_sink = _make_jsonl_sink(_trace_path)
|
||
_rec(f"trace sink 接线 → {_trace_path}")
|
||
except Exception as _sink_err:
|
||
_rec(f"trace sink 构造异常（只告警，回落无 sink，不阻断生成）：{_sink_err}")
|
||
# 阶段四观测 波③ T3-1:CLI 路同款并接 OTLP span sink(发 Collector,service.name=cheap-gen-worker)。
|
||
# 默认关(env CHEAP_OTLP_ENDPOINT 未设)→ 原样返回上面的 jsonl sink,CLI 现有行为字节不变;
|
||
# 开则包 fan-out(jsonl 落盘照旧、OTLP 额外发一份)。best-effort:观测挂掉绝不咬生成(设计 §8)。
|
||
try:
|
||
import cheap_otlp_sink # noqa: PLC0415 函数体内惰性 import,顶层不牵 otel
|
||
_trace_sink = cheap_otlp_sink.build_trace_sink(_trace_sink, trace_id=game_id)
|
||
except Exception as _otlp_err:
|
||
_rec(f"OTLP sink 并接异常（回落原 jsonl sink，不阻断生成）：{_otlp_err}")
|
||
tracer = Tier2TraceMiddleware(trace_id=game_id, sink=_trace_sink)
|
||
|
||
writer = Agent(
|
||
name="cheap-writer",
|
||
system_prompt=system_prompt or build_system_prompt(game_id, scaffold_desc), # modify 传 build_modify_system_prompt；扩模板传 scaffold_desc
|
||
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
|
||
closeout = None # CLI 回喂:最近一次九门收口结果(回喂循环产出,收口段复用、不重复跑门)
|
||
try:
|
||
for attempt in range(max_resumes + 1):
|
||
attempts = attempt + 1
|
||
_rec(f"resume attempt {attempts}/{max_resumes + 1} → writer.reply …")
|
||
try:
|
||
await writer.reply(_user_msg(kick))
|
||
except Tier2CircuitBreak as _cb:
|
||
# fix①-B(重实·resume 层,2026-07-08 n=5 收敛环坐实 heritage 死圈复现)——
|
||
# 根因:MiniMax 长参数下漏 write_file/edit_file 必填 path,schema 校验拒 → 连撞 stuck 阈值 → 终态杀、九门没跑。
|
||
# 先前 on_acting 版无效(_agent.py:1376 校验先于 Step3 工具执行/on_acting、畸形调用直接 return 不进钩子);
|
||
# 改挂此处:missing-param 类 stuck 不当终态,还有 resume 预算就带【贴脸补参反馈】续修——与 gate-fail resume
|
||
# 同机制、跨 reply memory 保留=原地续修、能真正触达 agent。只清 stuck 连击计数(不动 spent_rmb/tool_calls,防误清预算);
|
||
# 非 missing-param 的 stuck(反复 check 同红线 / build 同编译错=真死圈)/ budget / 其它熔断照旧向上抛(终态、不放行)。
|
||
_missing_param = _cb.kind == "stuck" and "is a required property" in (_cb.reason or "")
|
||
if _missing_param and attempt < max_resumes and not getattr(breaker, "budget_soft_tripped", False):
|
||
_rec(f"熔断 stuck(漏必填参 path)→ fix①-B 带补参反馈续修(attempt {attempts}):{(_cb.reason or '')[:70]}")
|
||
breaker._fail_repeat = 0 # 只清 stuck 连击计数(下一 attempt 干净起,不动 spent_rmb/tool_calls)
|
||
breaker._last_fail_sig = None
|
||
kick = _MALFORMED_RESUME_FEEDBACK
|
||
continue
|
||
raise # 其它熔断 / 预算耗尽 / 非漏参 stuck → 终态,交外层 except 记 breaker_tripped
|
||
# ① 真 finish:CLI 回喂对齐(W-S1 单③,对账发现六④)——收敛判据从「check+build 绿」升为
|
||
# 「九门绿」:finish 后跑九门收口,门未全绿且轮数/预算有余时带 C6 反馈 resume 续修,
|
||
# 使 lab 复验口径与生产 RepairMiddleware(finish 点拦截→跑门→judge→回喂)行为同型。
|
||
if session.finished is not None:
|
||
if not run_gates:
|
||
_rec("finish 已接受(check+build 绿)→ 收敛(generation-only,不跑九门回喂)")
|
||
break
|
||
_rec("finish 已接受(check+build 绿)→ 九门收口判(CLI 回喂对齐)")
|
||
closeout = _closeout_gates(game_id, port=port, cdp_port=cdp_port)
|
||
j = judge_cheap_verdict(closeout["verdict"], game_id=game_id,
|
||
staged_dir=cheap_run.wg1_game_dir(game_id))
|
||
vb = _verdict_brief(closeout["verdict"])
|
||
_rec(f"九门 play pass={vb['pass']} failedGates={vb['failedGates']}")
|
||
if j.passed:
|
||
break
|
||
if attempt >= max_resumes:
|
||
_rec("九门未绿,但 resume 轮数耗尽 → 交尽力产物收口")
|
||
break
|
||
# 预算判定与生产同口径:软停已触发 / 实测已花越软停线 → 只收尾、不再回喂(裁决:越线只许收尾)。
|
||
budget_out = bool(getattr(breaker, "budget_soft_tripped", False)) or (
|
||
breaker._rmb_gate_active and breaker.spent_rmb >= breaker.rmb_hard_limit)
|
||
if budget_out:
|
||
_rec(f"九门未绿,但 ¥ 软停线已到(已花 ¥{breaker.spent_rmb:.4f})→ 只收尾、不再回喂")
|
||
break
|
||
session.finished = None # 重置收敛标记,带门反馈续修(跨 reply memory 保留=原地续修)
|
||
_rec(f"九门未绿(failedGates={j.failed_gates})→ 带 C6 反馈回喂续修")
|
||
kick = j.feedback # C6 结构渲染全文(含 phaseNow/driverType/game-log 信号;与 RepairMiddleware 注入同源)
|
||
continue
|
||
# 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 ──
|
||
# CLI 回喂对齐后,run_gates 路的九门收口已在循环内跑过(closeout),此处直接复用最后一轮结果,
|
||
# 不重复跑门;对照路(run_gates=False)保持原「只 stage+smoke」行为零改动。
|
||
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:
|
||
if run_gates:
|
||
cg = closeout or _closeout_gates(game_id, port=port, cdp_port=cdp_port) # 防御:正常路循环内已跑
|
||
staged = cg["staged"]
|
||
smoke_ok = cg["smoke_ok"]
|
||
driver_type = cg["driver_type"] # M3b U1:driver 类型存进 trace.gatespec.driver(后端 D11 firstPlay 维)
|
||
verdict = cg["verdict"]
|
||
else:
|
||
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 not smoke_ok:
|
||
_rec(f"smoke raw 尾:{(sm.get('raw') or '')[-300:]}") # 真因可见(同 _closeout_gates 注)
|
||
_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),
|
||
# C1b trace 收口摘要：{traceId, steps, dropped}，供编排器对账（trace.jsonl 落盘路径已在接线处 _rec）。
|
||
"trace": tracer.summary(),
|
||
}
|
||
# ── 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))
|
||
|
||
# ── U-A2：便宜档「丰富度」LLM 评分（非阻塞·只报告·不进 verdict / 不改 ok 达标）──
|
||
# 架构红线：玩法「丰富不丰富」的校验需大模型能力，绝不写成 code-presence/正则/断言（违反创始人红线）。
|
||
# 故纯走 LLM judge（cheap_verify.verify_richness），结果 additive 写 summary["richness"]：
|
||
# · 不进 verdict、不改 summary["ok"]/达标判定、不动任何既有键（result_out._build_trace 只读 named 字段，不读 richness）；
|
||
# · judge 用独立 model 实例，token 不污染生成成本台账（costRmb 仍只算生成）；
|
||
# · 非阻塞双保险：verify_richness 内部已 try/except 兜底 degraded，这里再包一层，verify 失败绝不影响 run。
|
||
# 只在 run_gates（完整 spike 收口）且 finished（有真产物）时跑；对照路（run_gates=False）零改动、不触发 LLM。
|
||
if verify_richness_enabled and run_gates and finished:
|
||
try:
|
||
# 品类扩展 rubric(W-GENRE 件④):显式 genre 参数优先,否则按 scaffold_template 查
|
||
# GENRE_BY_TEMPLATE 推断;命中时同一次评分 additive 追加品类条目(分母独立小计落
|
||
# richness.genre);两路都没有/未登记 → 行为与既有完全一致。
|
||
_genre = genre or (cheap_verify.GENRE_BY_TEMPLATE.get(scaffold_template) if scaffold_template else None)
|
||
summary["richness"] = await cheap_verify.verify_richness(game_id, brief=brief, genre=_genre)
|
||
rv = summary["richness"]
|
||
_rec(f"丰富度 LLM 评分 score={rv.get('score')}/{rv.get('max')} groups={rv.get('groups')} "
|
||
f"genre={(rv.get('genre') or {}).get('key')}:{(rv.get('genre') or {}).get('score')} degraded={rv.get('degraded')}")
|
||
except Exception as e: # noqa: BLE001 非阻塞铁律：richness 绝不影响生成主链
|
||
summary["richness"] = {"score": None, "degraded": True,
|
||
"reason": f"richness 接线异常：{type(e).__name__}: {e}"}
|
||
_rec(f"丰富度评分接线异常（已降级，不影响 run）：{type(e).__name__}: {e}")
|
||
|
||
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']}")
|
||
|
||
# ── 面二观测:发 llm_tokens{kind} 计数 + llm_cache_hit_rate(默认关 / best-effort;绝不咬生成)──
|
||
# 本路(CLI/批跑)model 是 RecordingOpenAIChatModel,records 每条 (in,out,cached) 三元组,cached 由
|
||
# usage.cache_input_tokens 归一 → 三项全可得(生产 Service 路框架默认 model 无 records,cached 拿不到、
|
||
# 见 cheap_service_app collector)。取不到 CHEAP_OTLP_ENDPOINT 即 no-op,与 trace sink 默认关同源;
|
||
# 面五(game-cloud GenMetrics)不发 token 计数,故不双计。发送异常只报告、绝不阻断生成。
|
||
try:
|
||
import cheap_otlp_sink # noqa: PLC0415 惰性 import(顶层不牵 otel)
|
||
_cached_tok = sum(r[2] for r in model.records if len(r) > 2)
|
||
cheap_otlp_sink.record_llm_token_metrics(tok_in, tok_out, _cached_tok)
|
||
except Exception as _me: # noqa: BLE001 面二 metric best-effort,绝不影响生成主链
|
||
_rec(f"面二 metric 发送异常(已忽略,不影响生成):{type(_me).__name__}: {_me}")
|
||
|
||
# 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)
|