merge(工单k): 收敛环台账 infra_flags 归因穿线(observe-only)——设计降级/轮墙/熔断/软停入台账+退路树 reasons 打印,治 F-2 R1 误判分散盲区 · 102/102 · fable 终审过(双墙共旗以 degraded 共现消歧,记录在案)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
lili
commit
bf00638c8b
@ -256,6 +256,9 @@ def result_to_record(result: dict, *, run_id: str, brief_variant: str,
|
||||
decision=decision, # type: ignore[arg-type]
|
||||
finished=bool(result.get("finished")),
|
||||
timestamp=timestamp,
|
||||
# 编排层 infra 归因(observe-only):从 studio result 抽 infra_flags(缺则空列表,向后兼容旧桩)。
|
||||
# 防御式 list(...) 复制,避免与 result 内引用共享;取值 ∈ run_record.INFRA_FLAGS 受控值集。
|
||||
infra_flags=list(result.get("infra_flags") or []),
|
||||
)
|
||||
|
||||
|
||||
|
||||
245
tier2/gen-worker/tests/test_infra_flags.py
Normal file
245
tier2/gen-worker/tests/test_infra_flags.py
Normal file
@ -0,0 +1,245 @@
|
||||
"""test_infra_flags.py —— 收敛环台账 infra 归因穿线单测(工单 k,observe-only 第一迭代)。
|
||||
|
||||
跑:cd tier2/gen-worker && cheap-worker/.venv/bin/python -m pytest tests/test_infra_flags.py -q
|
||||
|
||||
验收门(≥4):
|
||||
① studio 态 flag 采集(构造 fake 运行态)—— test_studio_collects_* / test_infra_flags_from_runtime_*;
|
||||
② batch_run 落账含 infra_flags —— test_result_to_record_carries_infra_flags;
|
||||
③ 旧行(无字段)解析兼容 —— test_from_jsonl_line_backward_compat;
|
||||
④ decide_n5 reasons 含分布行且出口不变(含 --no-strong-baseline 语义原样)——
|
||||
test_decide_n5_reasons_infra_dist_and_exit_unchanged。
|
||||
|
||||
穿线红线复核:不塞 fail_system(游戏系统桶语义不污染)、退路树出口枚举与判定不动、台账向后兼容。
|
||||
"""
|
||||
import json
|
||||
|
||||
from worker.run_record import (
|
||||
RunRecord,
|
||||
infra_flags_from_runtime,
|
||||
INFRA_FLAGS,
|
||||
INFRA_FLAG_DESIGN_TEAM_DEGRADED,
|
||||
INFRA_FLAG_WRITER_ITER_WALL,
|
||||
INFRA_FLAG_STEP_CAP_TRIPPED,
|
||||
INFRA_FLAG_SOFT_BUDGET_TRIPPED,
|
||||
INFRA_FLAG_WALL_TIMEOUT,
|
||||
)
|
||||
from worker import fallback_tree as ft
|
||||
|
||||
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
# 构造小工具
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
def _mk(ok: bool, *, fail_system=None, infra=None) -> RunRecord:
|
||||
"""造一条 RunRecord:ok=True → 过门款;ok=False → 失败款(可带 fail_system / infra_flags)。"""
|
||||
if ok:
|
||||
return RunRecord(run_id="cv-r1-M3-面包-0", model="MiniMax-M3", stage="play",
|
||||
brief_variant="面包", pass_gate=True, repairs=0,
|
||||
decision="accept", finished=True, infra_flags=list(infra or []))
|
||||
return RunRecord(run_id="cv-r1-M3-面包-0", model="MiniMax-M3", stage="build",
|
||||
brief_variant="面包", pass_gate=False, repairs=1,
|
||||
decision="fix", finished=False, fail_system=fail_system,
|
||||
infra_flags=list(infra or []))
|
||||
|
||||
|
||||
def _round(n_ok: int, n_fail: int, *, fail_system=None, infra=None) -> list:
|
||||
"""造一轮:n_ok 过门 + n_fail 失败(失败款统一带 fail_system / infra_flags)。"""
|
||||
return ([_mk(True) for _ in range(n_ok)]
|
||||
+ [_mk(False, fail_system=fail_system, infra=infra) for _ in range(n_fail)])
|
||||
|
||||
|
||||
class _FakeTimeoutErr(Exception):
|
||||
"""fake 设计团队超时异常(带 .kind='timeout',镜像 design_team.DesignTeamError 的 kind 约定)。"""
|
||||
kind = "timeout"
|
||||
|
||||
|
||||
class _FakeDegradeErr(Exception):
|
||||
"""fake 设计团队非超时降级异常(无 kind,如预算耗尽 / leader 空产出 / 团队异常)。"""
|
||||
|
||||
|
||||
class _FakeBreaker:
|
||||
"""fake CircuitBreakerMiddleware:只暴露 infra 采集读的 budget_soft_tripped 标记。"""
|
||||
def __init__(self, budget_soft_tripped: bool = False) -> None:
|
||||
self.budget_soft_tripped = budget_soft_tripped
|
||||
|
||||
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
# ① 纯映射:infra_flags_from_runtime(编排层信号 → 受控值集)
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
def test_infra_flags_from_runtime_pure_mapping():
|
||||
# 无信号 → 空列表。
|
||||
assert infra_flags_from_runtime() == []
|
||||
|
||||
# 设计团队超时降级 → design_team_degraded + wall_timeout(稳定序=值集登记序)。
|
||||
assert infra_flags_from_runtime(design_degraded=True, design_timeout=True) == [
|
||||
INFRA_FLAG_DESIGN_TEAM_DEGRADED, INFRA_FLAG_WALL_TIMEOUT]
|
||||
|
||||
# 设计团队非超时降级 → 只 design_team_degraded。
|
||||
assert infra_flags_from_runtime(design_degraded=True) == [INFRA_FLAG_DESIGN_TEAM_DEGRADED]
|
||||
|
||||
# 单写撞轮数墙 → writer_iter_wall。
|
||||
assert infra_flags_from_runtime(writer_iter_wall=True) == [INFRA_FLAG_WRITER_ITER_WALL]
|
||||
|
||||
# 熔断 kind 映射:step_cap → step_cap_tripped;timeout → wall_timeout。
|
||||
assert infra_flags_from_runtime(breaker_kind="step_cap") == [INFRA_FLAG_STEP_CAP_TRIPPED]
|
||||
assert infra_flags_from_runtime(breaker_kind="timeout") == [INFRA_FLAG_WALL_TIMEOUT]
|
||||
# budget/stuck 不入首批值集(由 circuit_break 字段另记)→ 空。
|
||||
assert infra_flags_from_runtime(breaker_kind="budget") == []
|
||||
assert infra_flags_from_runtime(breaker_kind="stuck") == []
|
||||
|
||||
# ¥ 软预算软停 → soft_budget_tripped。
|
||||
assert infra_flags_from_runtime(budget_soft_tripped=True) == [INFRA_FLAG_SOFT_BUDGET_TRIPPED]
|
||||
|
||||
# 去重 + 稳定序:设计团队超时(wall_timeout)叠加熔断 timeout(wall_timeout)→ wall_timeout 只一份;
|
||||
# 多信号同现按 INFRA_FLAGS 登记序输出(不随触发先后抖动)。
|
||||
got = infra_flags_from_runtime(
|
||||
design_degraded=True, design_timeout=True, writer_iter_wall=True,
|
||||
breaker_kind="timeout", budget_soft_tripped=True)
|
||||
assert got == [INFRA_FLAG_DESIGN_TEAM_DEGRADED, INFRA_FLAG_WRITER_ITER_WALL,
|
||||
INFRA_FLAG_SOFT_BUDGET_TRIPPED, INFRA_FLAG_WALL_TIMEOUT]
|
||||
assert got.count(INFRA_FLAG_WALL_TIMEOUT) == 1 # 去重坐实
|
||||
# 输出只含受控值集里的项。
|
||||
assert all(f in INFRA_FLAGS for f in got)
|
||||
|
||||
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
# ① studio 态采集(构造 fake 运行态,委托纯映射)
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
def test_studio_collects_infra_flags_from_fake_runtime():
|
||||
# 延迟 import(studio 拉 agentscope;与既有 middleware 测试同环境,baseline 已证可 import)。
|
||||
from worker.agent_loop import studio
|
||||
|
||||
# 设计团队墙钟超时降级 → design_team_degraded + wall_timeout(studio 从异常 .kind 抽 design_timeout)。
|
||||
flags = studio._infra_flags_from_runtime(
|
||||
design_error=_FakeTimeoutErr("超时"), breaker=_FakeBreaker(),
|
||||
breaker_tripped=None, hit_writer_iter_wall=False)
|
||||
assert flags == [INFRA_FLAG_DESIGN_TEAM_DEGRADED, INFRA_FLAG_WALL_TIMEOUT]
|
||||
|
||||
# 非超时降级 → 只 design_team_degraded。
|
||||
flags = studio._infra_flags_from_runtime(
|
||||
design_error=_FakeDegradeErr("预算耗尽"), breaker=_FakeBreaker(),
|
||||
breaker_tripped=None, hit_writer_iter_wall=False)
|
||||
assert flags == [INFRA_FLAG_DESIGN_TEAM_DEGRADED]
|
||||
|
||||
# 撞轮数墙 + step_cap 熔断 + 软停:三信号齐 → 稳定序输出。
|
||||
flags = studio._infra_flags_from_runtime(
|
||||
design_error=None, breaker=_FakeBreaker(budget_soft_tripped=True),
|
||||
breaker_tripped={"kind": "step_cap", "reason": "步数超顶"}, hit_writer_iter_wall=True)
|
||||
assert flags == [INFRA_FLAG_WRITER_ITER_WALL, INFRA_FLAG_STEP_CAP_TRIPPED,
|
||||
INFRA_FLAG_SOFT_BUDGET_TRIPPED]
|
||||
|
||||
# 全无编排层事件(正常收敛的 run)→ 空列表。
|
||||
flags = studio._infra_flags_from_runtime(
|
||||
design_error=None, breaker=_FakeBreaker(),
|
||||
breaker_tripped=None, hit_writer_iter_wall=False)
|
||||
assert flags == []
|
||||
|
||||
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
# ② batch_run 落账含 infra_flags(result → RunRecord 接线)
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
def test_result_to_record_carries_infra_flags():
|
||||
import batch_run
|
||||
|
||||
# 造一个形状对齐 studio result 的假结果:带 infra_flags(设计团队超时降级路)。
|
||||
result = {
|
||||
"model": "MiniMax-M3", "finished": False,
|
||||
"last_verdict": {"decision": "fix", "layerResults": {"L1": {"passed": False}}},
|
||||
"circuit_break": None, "breaker_counters": {"tool_calls": 3, "model_calls": 5},
|
||||
"tokens": {"prompt": 100000, "completion": 30000, "cached": 5000},
|
||||
"cost": {"cost_rmb": 1.2, "tokens_by_model": {}},
|
||||
"wall_s": 90.0, "file_tree": [], "source_project": {},
|
||||
"infra_flags": [INFRA_FLAG_DESIGN_TEAM_DEGRADED, INFRA_FLAG_WALL_TIMEOUT],
|
||||
}
|
||||
rec = batch_run.result_to_record(result, run_id="feie-r1-M3-面包-0",
|
||||
brief_variant="面包", timestamp=1750000000.0)
|
||||
assert rec.infra_flags == [INFRA_FLAG_DESIGN_TEAM_DEGRADED, INFRA_FLAG_WALL_TIMEOUT]
|
||||
# 防御式复制:改 result 里的列表不应回污染 rec(不共享引用)。
|
||||
result["infra_flags"].append("xxx")
|
||||
assert rec.infra_flags == [INFRA_FLAG_DESIGN_TEAM_DEGRADED, INFRA_FLAG_WALL_TIMEOUT]
|
||||
# 落盘 JSONL 往返一致(infra_flags 原样保留)。
|
||||
back = RunRecord.from_jsonl_line(rec.to_jsonl_line())
|
||||
assert back.infra_flags == [INFRA_FLAG_DESIGN_TEAM_DEGRADED, INFRA_FLAG_WALL_TIMEOUT]
|
||||
|
||||
# result 缺 infra_flags 键(旧桩)→ 落空列表,不炸。
|
||||
result.pop("infra_flags")
|
||||
rec2 = batch_run.result_to_record(result, run_id="feie-r1-M3-面包-1",
|
||||
brief_variant="面包", timestamp=1750000000.0)
|
||||
assert rec2.infra_flags == []
|
||||
|
||||
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
# ③ 旧行(无 infra_flags 字段)解析兼容 —— 台账向后兼容红线
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
def test_from_jsonl_line_backward_compat():
|
||||
# 旧台账行:没有 infra_flags 列(模拟本工单前落的行)。
|
||||
old_line = json.dumps({
|
||||
"run_id": "feie-001#1", "model": "MiniMax-M3", "stage": "play",
|
||||
"brief_variant": "面包店", "pass": True, "repairs": 4,
|
||||
"cost_yuan": 2.13, "wall_seconds": 187.4, "decision": "accept", "finished": True,
|
||||
}, ensure_ascii=False)
|
||||
rec = RunRecord.from_jsonl_line(old_line) # 不炸
|
||||
assert rec.infra_flags == [] # 缺字段读为空列表
|
||||
assert rec.pass_gate is True and rec.model == "MiniMax-M3" # 其余字段照常
|
||||
|
||||
# 新行:带 infra_flags,往返一致。
|
||||
new_rec = _mk(False, infra=[INFRA_FLAG_STEP_CAP_TRIPPED])
|
||||
line = new_rec.to_jsonl_line()
|
||||
assert '"infra_flags"' in line and "step_cap_tripped" in line
|
||||
assert RunRecord.from_jsonl_line(line).infra_flags == [INFRA_FLAG_STEP_CAP_TRIPPED]
|
||||
|
||||
# 未知多余列(台账演进新增)仍被过滤,不炸。
|
||||
weird = json.loads(old_line)
|
||||
weird["some_future_col"] = 123
|
||||
weird["infra_flags"] = ["design_team_degraded"]
|
||||
rec3 = RunRecord.from_jsonl_line(json.dumps(weird, ensure_ascii=False))
|
||||
assert rec3.infra_flags == ["design_team_degraded"]
|
||||
|
||||
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
# ④ decide_n5 reasons 含分布行 且 出口不变(含 --no-strong-baseline 语义原样)
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
def test_decide_n5_reasons_infra_dist_and_exit_unchanged():
|
||||
# —— F-2 复现场景:两轮不收敛、失败款 fail_system 全空(退路树按 fail_system 看「分散」)、
|
||||
# 但 infra_flags 全集中「设计团队超时降级路」(design_team_degraded + wall_timeout)。
|
||||
# 这正是被误判分散出 R3_conditional 的生产盲区:穿线后 reasons 应打印出集中的 infra 分布。——
|
||||
infra = [INFRA_FLAG_DESIGN_TEAM_DEGRADED, INFRA_FLAG_WALL_TIMEOUT]
|
||||
rounds_flagged = [_round(3, 2, fail_system=None, infra=infra),
|
||||
_round(3, 2, fail_system=None, infra=infra)]
|
||||
# 同构但不带 infra_flags 的对照(证「加 infra_flags 不改出口」)。
|
||||
rounds_plain = [_round(3, 2, fail_system=None),
|
||||
_round(3, 2, fail_system=None)]
|
||||
|
||||
# 无强档对照(--no-strong-baseline 语义):失败分散无集中 → R3_conditional(出口不变)。
|
||||
out_flagged = ft.decide_n5(rounds_flagged, has_strong_baseline=False)
|
||||
out_plain = ft.decide_n5(rounds_plain, has_strong_baseline=False)
|
||||
assert out_flagged["exit"] == ft.EXIT_R3_CONDITIONAL # 误判出口原样(本工单不改判定)
|
||||
assert out_flagged["exit"] == out_plain["exit"] # 加 infra_flags 不改出口
|
||||
# reasons 含各轮 infra 分布行,且点名集中的编排层原因(治盲区的核心穿线证据)。
|
||||
joined = "\n".join(out_flagged["reasons"])
|
||||
assert "infra 分布" in joined
|
||||
assert "design_team_degraded:2" in joined and "wall_timeout:2" in joined
|
||||
assert "r1 infra 分布" in joined and "r2 infra 分布" in joined
|
||||
# 不带 infra_flags 的对照:无分布行(无 flag 不加噪)。
|
||||
assert "infra 分布" not in "\n".join(out_plain["reasons"])
|
||||
|
||||
# 有强档对照:同场景 → R3_天花板(前瞻支);加 infra_flags 亦不改出口。
|
||||
assert ft.decide_n5(rounds_flagged, has_strong_baseline=True)["exit"] == ft.EXIT_R3_CEILING
|
||||
assert (ft.decide_n5(rounds_flagged, has_strong_baseline=True)["exit"]
|
||||
== ft.decide_n5(rounds_plain, has_strong_baseline=True)["exit"])
|
||||
|
||||
# —— 收敛(go)出口也不受 infra_flags 影响:软预算软停发生在最终干净轮上,仍判 go,且打印分布。——
|
||||
go_flagged = [_round(5, 0), [_mk(True, infra=[INFRA_FLAG_SOFT_BUDGET_TRIPPED]) for _ in range(5)]]
|
||||
go_plain = [_round(5, 0), _round(5, 0)]
|
||||
o_go = ft.decide_n5(go_flagged, has_strong_baseline=False)
|
||||
assert o_go["exit"] == ft.EXIT_GO
|
||||
assert o_go["exit"] == ft.decide_n5(go_plain, has_strong_baseline=False)["exit"]
|
||||
assert "soft_budget_tripped:5" in "\n".join(o_go["reasons"])
|
||||
|
||||
# —— 集中表现层仍走 R1(fail_system 分流不被 infra_flags 干扰):证 fail_system 与 infra_flags 各管一摊。——
|
||||
r1_flagged = [_round(1, 4, fail_system="presentation", infra=infra),
|
||||
_round(1, 4, fail_system="presentation", infra=infra)]
|
||||
r1_plain = [_round(1, 4, fail_system="presentation"),
|
||||
_round(1, 4, fail_system="presentation")]
|
||||
assert ft.decide_n5(r1_flagged, has_strong_baseline=False)["exit"] == ft.EXIT_R1_TEMPLATE
|
||||
assert (ft.decide_n5(r1_flagged, has_strong_baseline=False)["exit"]
|
||||
== ft.decide_n5(r1_plain, has_strong_baseline=False)["exit"])
|
||||
@ -80,8 +80,15 @@ class DesignTeamError(Exception):
|
||||
"""工作室设计团队失败信号(模型超时/预算耗尽/leader 未产出有效设计稿)。
|
||||
|
||||
studio 的 _design_stage 接住它 → degrade 回落原单 agent 单轮设计,绝不让设计阶段中断主链。
|
||||
|
||||
kind:失败大类(供 studio 采集编排层 infra_flags 时区分);'timeout' = 整团队墙钟硬超时(降级归
|
||||
wall_timeout),其余(预算耗尽 / leader 空产出 / 团队异常)= None(只归 design_team_degraded)。
|
||||
"""
|
||||
|
||||
def __init__(self, message: str, *, kind: str | None = None) -> None:
|
||||
self.kind = kind
|
||||
super().__init__(message)
|
||||
|
||||
|
||||
# ── 四面专家定义:(工具名, 工具中文描述, system prompt)──────────────────────────
|
||||
# 工具名/描述会被 AgentScope 从 FunctionTool 抽成 leader 可见的工具 schema(描述即 leader 选工具的依据)。
|
||||
@ -257,7 +264,8 @@ async def run_design_team(
|
||||
try:
|
||||
resp = await asyncio.wait_for(leader.reply(_user_msg(kick)), timeout=timeout_s)
|
||||
except asyncio.TimeoutError as e:
|
||||
raise DesignTeamError(f"工作室设计团队超时(> {timeout_s}s):{e}") from e
|
||||
# kind='timeout':整团队墙钟硬超时——studio 采集时据此归 infra_flags 的 wall_timeout(+ design_team_degraded)。
|
||||
raise DesignTeamError(f"工作室设计团队超时(> {timeout_s}s):{e}", kind="timeout") from e
|
||||
except Exception as e: # noqa: BLE001 —— 任何团队异常都转成 DesignTeamError(让 studio 走 degrade)
|
||||
raise DesignTeamError(f"工作室设计团队异常:{type(e).__name__}: {e}") from e
|
||||
|
||||
|
||||
@ -33,7 +33,7 @@ from pathlib import Path
|
||||
|
||||
# 包内/直跑兼容导入(直跑时把 tier2/ 加进 sys.path,使 `gen_worker.*` 可解析)。
|
||||
try:
|
||||
from .. import config, roles, run, genconfig
|
||||
from .. import config, roles, run, genconfig, run_record
|
||||
from ..toolkit import Tier2Session, build_toolkit
|
||||
from ..middleware import CircuitBreakerMiddleware, Tier2CircuitBreak, Tier2TraceMiddleware
|
||||
from . import design_team # 阶段 1 工作室星形多 agent 设计团队
|
||||
@ -42,7 +42,7 @@ except ImportError: # pragma: no cover —— 直接 python studio.py 兜底
|
||||
# 本文件在 tier2/gen-worker/worker/agent_loop/studio.py;把 gen-worker/ 加进 sys.path,
|
||||
# 使顶层包 `worker` 可解析(包名 worker 合法;目录名 gen-worker 含连字符不可直接 import)。
|
||||
sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
|
||||
from worker import config, roles, run, genconfig # type: ignore
|
||||
from worker import config, roles, run, genconfig, run_record # type: ignore
|
||||
from worker.toolkit import Tier2Session, build_toolkit # type: ignore
|
||||
from worker.middleware import ( # type: ignore
|
||||
CircuitBreakerMiddleware, Tier2CircuitBreak, Tier2TraceMiddleware)
|
||||
@ -109,12 +109,15 @@ async def _design_single_agent(brief: str, model) -> str:
|
||||
return text_of(resp)
|
||||
|
||||
|
||||
async def _design_stage(brief: str, model_factory, *, model_sink: list, use_team: bool = True) -> str:
|
||||
async def _design_stage(brief: str, model_factory, *, model_sink: list,
|
||||
use_team: bool = True) -> tuple[str, Exception | None]:
|
||||
"""阶段 1:产富游戏设计稿。优先工作室星形多 agent 团队(过门头号杠杆),失败 degrade 回单 agent。
|
||||
|
||||
产物形状 = str(连贯设计稿),与原单 agent 完全一致 → 阶段 2 单写消费接口(roles.writer_system 的
|
||||
design_text)零改、爆炸半径最小化。design_team 造的所有 model(leader + 专家)都进 model_sink,
|
||||
供 run_studio 汇总 token 成本(多 agent 烧的 token 必须被成本台账抓到,P0 纪律)。
|
||||
产物形状 = (连贯设计稿 str, 降级异常 or None)。设计稿与原单 agent 完全一致 → 阶段 2 单写消费接口
|
||||
(roles.writer_system 的 design_text)零改、爆炸半径最小化。第二个返回值是**编排层 infra 归因**用:
|
||||
团队降级时把捕获的异常带回(None=未降级),run_studio 据它(及 .kind=='timeout')采集 infra_flags 的
|
||||
design_team_degraded / wall_timeout(治 F-2 R1 生产盲区)。design_team 造的所有 model(leader + 专家)
|
||||
都进 model_sink,供 run_studio 汇总 token 成本(多 agent 烧的 token 必须被成本台账抓到,P0 纪律)。
|
||||
|
||||
Args:
|
||||
brief: 题面。
|
||||
@ -123,20 +126,46 @@ async def _design_stage(brief: str, model_factory, *, model_sink: list, use_team
|
||||
model_sink: 收集本阶段造的所有 model(team 的 + 兜底单 agent 的),供成本汇总。
|
||||
use_team: True → 先试工作室多 agent 团队(默认);False → 直接走单 agent(留旁路/调试用)。
|
||||
"""
|
||||
design_error: Exception | None = None # 团队降级时捕获的异常(供 infra 归因;None=未降级)
|
||||
if use_team:
|
||||
try:
|
||||
# 工作室星形多 agent 设计团队(worker-as-tool;预算约束 + 超限 degrade 在 design_team 内)。
|
||||
return await design_team.run_design_team(brief, model_factory, model_sink=model_sink)
|
||||
return await design_team.run_design_team(brief, model_factory, model_sink=model_sink), None
|
||||
except design_team.DesignTeamError as e:
|
||||
# 团队失败(超时/预算耗尽/leader 无有效产出)→ degrade 回单 agent,绝不中断主链。
|
||||
# 捕获异常带回(其 .kind=='timeout' 标记设计团队墙钟超时),供 infra_flags 采集归因。
|
||||
design_error = e
|
||||
print(f"[tier2-studio] 阶段 1 工作室多 agent 团队失败,degrade 回单 agent 设计:{e}", flush=True)
|
||||
except Exception as e: # noqa: BLE001 —— 任何意外也 degrade(设计阶段绝不能成为新失败点)
|
||||
design_error = e
|
||||
print(f"[tier2-studio] 阶段 1 工作室多 agent 团队异常,degrade 回单 agent 设计:"
|
||||
f"{type(e).__name__}: {e}", flush=True)
|
||||
# degrade / 旁路:单 design agent 单轮。它造的 model 也进 sink 供成本汇总。
|
||||
m = model_factory()
|
||||
model_sink.append(m)
|
||||
return await _design_single_agent(brief, m)
|
||||
return await _design_single_agent(brief, m), design_error
|
||||
|
||||
|
||||
def _infra_flags_from_runtime(*, design_error: Exception | None, breaker,
|
||||
breaker_tripped: dict | None,
|
||||
hit_writer_iter_wall: bool) -> list[str]:
|
||||
"""据 studio 运行态收集编排层 infra_flags(observe-only;委托 run_record 纯映射,绝不碰 fail_system)。
|
||||
|
||||
把四类编排层运行态信号抽成原语后交 run_record.infra_flags_from_runtime 归一(去重、稳定序):
|
||||
- design_error:设计阶段降级时捕获的异常(None=未降级;.kind=='timeout' → 设计团队墙钟超时);
|
||||
- hit_writer_iter_wall:单写外层 resume 轮数墙是否撞到(预算耗尽仍未收敛);
|
||||
- breaker_tripped:熔断返回 dict {'kind':...}(None=未熔断;kind ∈ step_cap/budget/stuck/timeout);
|
||||
- breaker.budget_soft_tripped:¥ 软预算是否软停触发(soft 档越软停线,不断链)。
|
||||
|
||||
单独抽成函数便于用「构造 fake 运行态」单测(不必真跑 agentscope ReAct 全链)。
|
||||
"""
|
||||
return run_record.infra_flags_from_runtime(
|
||||
design_degraded=design_error is not None,
|
||||
design_timeout=(getattr(design_error, "kind", None) == "timeout"),
|
||||
writer_iter_wall=bool(hit_writer_iter_wall),
|
||||
breaker_kind=(breaker_tripped or {}).get("kind") if breaker_tripped else None,
|
||||
budget_soft_tripped=bool(getattr(breaker, "budget_soft_tripped", False)),
|
||||
)
|
||||
|
||||
|
||||
def _bypass_state() -> AgentState:
|
||||
@ -359,11 +388,13 @@ async def run_studio(
|
||||
# 客户端,兜底单 agent 也从它取一个。design_models 收集本阶段造的所有客户端,供下方成本汇总(多 agent
|
||||
# 烧的 token 必须进台账,P0 纪律)。design_text 形状 = str,与原单 agent 一致 → 阶段 2 消费接口零改。
|
||||
design_text = ""
|
||||
design_error: Exception | None = None # 设计团队降级时捕获的异常(供 infra_flags 归因;None=未降级)
|
||||
design_models: list = [] # 阶段 1 造的所有模型客户端(team leader + 专家 + 兜底单 agent),供成本汇总
|
||||
if do_design:
|
||||
design_model_factory = functools.partial(
|
||||
config.build_model, mname, max_tokens=max_tokens, thinking_budget=thinking_budget)
|
||||
design_text = await _design_stage(brief, design_model_factory, model_sink=design_models)
|
||||
design_text, design_error = await _design_stage(
|
||||
brief, design_model_factory, model_sink=design_models)
|
||||
|
||||
# ── 阶段 2:单写 ReAct(放开 max_iters + 九工具 + 四熔断 + trace 接线 + BYPASS 权限)──
|
||||
session = Tier2Session(game_id, play_spec=play_spec)
|
||||
@ -407,6 +438,7 @@ async def run_studio(
|
||||
cumulative_iter = 0
|
||||
|
||||
breaker_tripped = None
|
||||
hit_writer_iter_wall = False # 单写外层 resume 轮数墙:预算耗尽仍未收敛(infra_flags 采集用)
|
||||
final_text = ""
|
||||
# ── 外层有界自纠循环(spike feie-001 根因 1)──
|
||||
# AgentScope 2.0.2 原生 ReAct 在「模型产出无 tool_call 的纯文本回合」即退出(_agent.py:612)——
|
||||
@ -451,6 +483,8 @@ async def run_studio(
|
||||
kick_text = "验收门已全绿。现在直接调 finish 交付源工程,不要再改。"
|
||||
continue
|
||||
if attempt >= max_resumes:
|
||||
# 撞外层 resume 轮数墙:预算耗尽仍未收敛 → 置 infra 标记(writer_iter_wall),供退路树归因。
|
||||
hit_writer_iter_wall = True
|
||||
print(f"[tier2-studio] game={game_id} resume 预算耗尽({max_resumes} 次)仍未收敛,停。",
|
||||
flush=True)
|
||||
break
|
||||
@ -604,6 +638,17 @@ async def run_studio(
|
||||
# ── trace 收口(H1/H2):读 tracer 摘要(traceId / 步数 / 丢弃数),供编排器对账反查 ──
|
||||
trace_summary = tracer.summary()
|
||||
|
||||
# ── 编排层 infra 归因采集(observe-only;治 F-2 R1 生产盲区)──────────────────────────
|
||||
# 把本次 run 的四类编排层运行态信号(设计团队降级 / 单写撞轮数墙 / 熔断返回 kind / ¥ 软预算软停)
|
||||
# 归一成 infra_flags 受控值集,写进 result 供 batch_run 落账、退路树 decide_n5 打印分布归因。
|
||||
# **绝不塞 fail_system**(游戏系统桶),与退路树 R1/R2 分流键语义正交;绝不参与 decision。
|
||||
infra_flags = _infra_flags_from_runtime(
|
||||
design_error=design_error, breaker=breaker,
|
||||
breaker_tripped=breaker_tripped, hit_writer_iter_wall=hit_writer_iter_wall)
|
||||
if infra_flags:
|
||||
print(f"[tier2-studio] game={game_id} 编排层 infra_flags={infra_flags}"
|
||||
"(observe-only,退路树归因用,不改 decision)", flush=True)
|
||||
|
||||
result = {
|
||||
"game_id": game_id,
|
||||
"model": mname,
|
||||
@ -627,6 +672,10 @@ async def run_studio(
|
||||
# 四熔断触发记录(对接 verdict.breakerKind:step_cap/budget/stuck/timeout)。
|
||||
"circuit_break": breaker_tripped,
|
||||
"breaker_counters": {"tool_calls": breaker.tool_calls, "model_calls": breaker.model_calls},
|
||||
# 编排层 infra 归因(observe-only;取值 ∈ run_record.INFRA_FLAGS 受控值集,可空)。
|
||||
# RunRecord 接线(batch_run.result_to_record)从此抽 infra_flags 落台账,供退路树 decide_n5
|
||||
# 打印各轮分布归因(治 F-2 R1「设计团队超时降级路」被误判分散的生产盲区);绝不参与 decision。
|
||||
"infra_flags": infra_flags,
|
||||
# 工作记忆 checkpoint 摘要(U2,A2:跨 resume 累计轮次 + 是否从 checkpoint 续跑)。
|
||||
# 累计轮次由本编排层维护(2.0.2 state.cur_iter 每 reply 归零);checkpoint 详情落 workdir/.tier2-work-memory.json。
|
||||
"work_memory": {"cumulativeIter": cumulative_iter, "resumedFromCheckpoint": resume_ckpt is not None},
|
||||
|
||||
@ -317,6 +317,19 @@ def _fail_system_dist(rounds: list[list[Any]]) -> dict[str, float]:
|
||||
return dist
|
||||
|
||||
|
||||
def _infra_flag_dist(round_records: list[Any]) -> dict[str, int]:
|
||||
"""一轮内各 infra_flag 出现次数(编排层归因;鸭子类型读 rec.infra_flags,decide_n5 保持零依赖)。
|
||||
|
||||
observe-only:只喂 decide_n5 的 reasons 打印(治 F-2 R1 生产盲区),**绝不参与分流/改出口**。
|
||||
统计的是全款(不限失败款):软预算软停等可能发生在最终过门的款上,也是有价值的编排层归因信号。
|
||||
"""
|
||||
dist: dict[str, int] = {}
|
||||
for r in round_records:
|
||||
for f in (getattr(r, "infra_flags", None) or []):
|
||||
dist[f] = dist.get(f, 0) + 1
|
||||
return dist
|
||||
|
||||
|
||||
def decide_n5(rounds: list[list[Any]], *, has_strong_baseline: bool) -> dict[str, Any]:
|
||||
"""n=5 收敛环判定。rounds = 各轮的 RunRecord 列表(有序,Round1 在前)。
|
||||
|
||||
@ -337,6 +350,17 @@ def decide_n5(rounds: list[list[Any]], *, has_strong_baseline: bool) -> dict[str
|
||||
last = fails[-1]
|
||||
reasons.append(f"各轮失败数={fails};收敛判据 KTD1 = 末轮≤1 且 ≥2 轮(压 n=5 首轮裸运气)。")
|
||||
|
||||
# ── 编排层 infra 归因(observe-only;治 F-2 R1 生产盲区)────────────────────────────────
|
||||
# 把各轮的 infra_flags 分布打进 reasons(如「r1 infra 分布={design_team_degraded:2}」),让主持人/
|
||||
# 机器一眼看出「两败局是否集中在设计团队超时降级路」这类编排层原因——过去 fail_system 空时退路树
|
||||
# 误判「分散」出 R3_conditional,归因只能人工翻日志。**只打印、不分支、不改任何出口**(纯观测穿线)。
|
||||
# 放在收敛/未收敛分流之前 → 无论走哪个出口,reasons 都带上分布行(有 flag 的轮才打,无则不加噪)。
|
||||
for _idx, _rd in enumerate(rounds, start=1):
|
||||
_idist = _infra_flag_dist(_rd)
|
||||
if _idist:
|
||||
_shown = ", ".join(f"{k}:{v}" for k, v in _idist.items())
|
||||
reasons.append(f"r{_idx} infra 分布={{{_shown}}}(编排层归因,observe-only,不参与分流)。")
|
||||
|
||||
# ── 收敛分支:末轮 ≤1 ──
|
||||
if last <= 1:
|
||||
if len(rounds) < 2:
|
||||
|
||||
@ -63,6 +63,28 @@ FailSystem = Literal["resource", "merge", "order", "presentation"]
|
||||
# decision:终判三值,取自 verdict.decision(tier2-verdict 契约)。accept/fix/kill。
|
||||
Decision = Literal["accept", "fix", "kill"]
|
||||
|
||||
# infra_flags:编排层失控/降级归因标记(G4 采集字段表之外的加性观测列;**与 fail_system 严格分离**)。
|
||||
# 为什么单独一列、绝不塞进 fail_system:fail_system 是「游戏系统桶」(resource/merge/order/presentation),
|
||||
# 语义是"哪个游戏系统的骨架缺口",退路树 decide_n5 的 R1/R2 分流直接读它;而「设计团队超时降级 /
|
||||
# writer 撞轮数墙 / step_cap 熔断 / 软预算软停 / 墙钟超时」是**编排层**事件(与具体游戏系统无关),塞进
|
||||
# fail_system 会污染系统桶语义、误导 R1/R2 分流。故另立 infra_flags 一列专收编排层归因,治 F-2 R1
|
||||
# 「两败局全在设计团队超时降级路,但 fail_system 空 → 退路树误判分散出 R3_conditional」的生产盲区。
|
||||
# 本列 observe-only:只进 decide_n5 的 reasons 打印供主持人/机器归因,退路树绝不据它分支、不改出口。
|
||||
INFRA_FLAG_DESIGN_TEAM_DEGRADED = "design_team_degraded" # 设计团队(超时/预算/leader 空产出)→ degrade 回单 agent
|
||||
INFRA_FLAG_WRITER_ITER_WALL = "writer_iter_wall" # 单写外层 resume 轮数墙:预算耗尽仍未收敛
|
||||
INFRA_FLAG_STEP_CAP_TRIPPED = "step_cap_tripped" # step_cap 熔断:工具调用步数超硬顶
|
||||
INFRA_FLAG_SOFT_BUDGET_TRIPPED = "soft_budget_tripped" # ¥ 软预算软停触发(soft 档越软停线,不断链)
|
||||
INFRA_FLAG_WALL_TIMEOUT = "wall_timeout" # 墙钟超时(设计团队墙钟超时 / 单写 reply 墙钟或单步静默熔断)
|
||||
# 受控值集(首批五项;**可扩**——新增编排层失控信号只需在此登记一个常量,studio 采集点与本集同源)。
|
||||
INFRA_FLAGS: tuple[str, ...] = (
|
||||
INFRA_FLAG_DESIGN_TEAM_DEGRADED, INFRA_FLAG_WRITER_ITER_WALL,
|
||||
INFRA_FLAG_STEP_CAP_TRIPPED, INFRA_FLAG_SOFT_BUDGET_TRIPPED, INFRA_FLAG_WALL_TIMEOUT,
|
||||
)
|
||||
InfraFlag = Literal[
|
||||
"design_team_degraded", "writer_iter_wall", "step_cap_tripped",
|
||||
"soft_budget_tripped", "wall_timeout",
|
||||
]
|
||||
|
||||
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
# RunRecord —— G4 采集字段表的数据类(每款 run 一条,落 JSONL 台账)
|
||||
@ -195,6 +217,16 @@ class RunRecord:
|
||||
"""agent 对 driver 的编辑次数(仅第二段)。来源:第二段编排观测;
|
||||
用途:防 agent 用改 driver 逃避卡死探测。"""
|
||||
|
||||
# ── 编排层 infra 归因组(加性观测;**与上面 fail_system 游戏系统桶严格分离**,见 INFRA_FLAGS 注释)──
|
||||
infra_flags: list[str] = field(default_factory=list)
|
||||
"""编排层失控/降级归因标记列表(取值 ∈ INFRA_FLAGS 受控值集,可空、可扩)。来源:studio 运行态在
|
||||
设计团队降级点 / 单写外层 resume 轮数墙 / 熔断返回点 / ¥ 软预算软停点采集(见 agent_loop/studio.py
|
||||
的 _infra_flags_from_runtime 收集口),经 batch_run 落账带入;用途:治 F-2 R1 生产盲区——退路树
|
||||
decide_n5 把各轮 infra_flags 分布打进 reasons 供主持人/机器归因,但**绝不据它分支或改出口**
|
||||
(observe-only)。默认空列表(field(default_factory=list) 防可变默认在实例间共享);旧台账行无此列时
|
||||
from_jsonl_line 读为空列表(向后兼容,r{N} 轮分组不受影响)。**绝不与 fail_system 混用**:fail_system
|
||||
是游戏系统桶(退路树 R1/R2 分流键),本列是编排层事件,语义正交、各管一摊。"""
|
||||
|
||||
def to_jsonl_line(self) -> str:
|
||||
"""序列化成一行 JSONL(末尾不带换行;append_record 落盘时补 '\\n')。
|
||||
|
||||
@ -225,6 +257,47 @@ class RunRecord:
|
||||
return cls(**filtered)
|
||||
|
||||
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
# infra_flags_from_runtime —— 编排层运行态信号 → 受控 infra_flags 值集(纯映射,零副作用)
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
def infra_flags_from_runtime(
|
||||
*,
|
||||
design_degraded: bool = False,
|
||||
design_timeout: bool = False,
|
||||
writer_iter_wall: bool = False,
|
||||
breaker_kind: Optional[str] = None,
|
||||
budget_soft_tripped: bool = False,
|
||||
) -> list[str]:
|
||||
"""把 studio 运行态的编排层信号归一成 infra_flags(去重、稳定序;纯函数,供 studio 采集点委托)。
|
||||
|
||||
信号 → 标记映射(与 INFRA_FLAGS 受控值集同源;新增信号在此加一条分支 + 上面登记一个常量即可):
|
||||
- design_degraded → design_team_degraded(设计团队降级回单 agent);
|
||||
- design_timeout → wall_timeout(设计团队墙钟超时;通常与 design_degraded 同现,但独立成标记);
|
||||
- writer_iter_wall → writer_iter_wall(单写外层 resume 轮数墙);
|
||||
- breaker_kind=='step_cap' → step_cap_tripped;=='timeout' → wall_timeout
|
||||
(熔断 kind 见 middleware.Tier2CircuitBreak;budget/stuck 暂不入首批值集,由 circuit_break 字段另记);
|
||||
- budget_soft_tripped → soft_budget_tripped(¥ 软预算软停,不断链)。
|
||||
|
||||
稳定序:按 INFRA_FLAGS 登记序去重排列,便于台账/断言比对(不随触发先后抖动);无信号 → 空列表。
|
||||
纯映射不碰 fail_system(游戏系统桶),严守编排层归因与游戏系统语义的正交分界。
|
||||
"""
|
||||
hit: set[str] = set()
|
||||
if design_degraded:
|
||||
hit.add(INFRA_FLAG_DESIGN_TEAM_DEGRADED)
|
||||
if design_timeout:
|
||||
hit.add(INFRA_FLAG_WALL_TIMEOUT) # 设计团队墙钟超时也归 wall_timeout
|
||||
if writer_iter_wall:
|
||||
hit.add(INFRA_FLAG_WRITER_ITER_WALL)
|
||||
if breaker_kind == "step_cap":
|
||||
hit.add(INFRA_FLAG_STEP_CAP_TRIPPED)
|
||||
elif breaker_kind == "timeout":
|
||||
hit.add(INFRA_FLAG_WALL_TIMEOUT) # 单写 reply 墙钟/单步静默熔断
|
||||
if budget_soft_tripped:
|
||||
hit.add(INFRA_FLAG_SOFT_BUDGET_TRIPPED)
|
||||
# 按受控值集登记序输出(稳定、去重),便于台账比对与断言。
|
||||
return [f for f in INFRA_FLAGS if f in hit]
|
||||
|
||||
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
# append_record —— 唯一的显式落盘出口(JSONL 台账追加;本模块唯一副作用)
|
||||
# ──────────────────────────────────────────────────────────────────────────
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user