feat(tier2): 统一 trace 落库 sink + SAA schema + 生成控制面只读后端(切片一 C1/C3a)

JsonlFileSink 把五字段 trace 落 workdir/trace.jsonl(非阻塞·写失败不阻断生成);studio.py sink 接线;SAA 扩展段 schema(接口对称五核心+内容不对称 ext);管理面 3 只读端点 roles/traces/cost(惰性 import 保住「仅 import service.app 不牵 fastapi」红线)。测试 test_jsonl_sink 3 / test_admin_routes 6 全绿。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
lili 2026-06-30 01:01:57 -07:00
parent e6525aaa3d
commit f4cea4677a
8 changed files with 769 additions and 4 deletions

View File

@ -18,6 +18,19 @@ SAA 那一轨的扩展段由 Java 线另一个 adapter 产,字段与本契约的
消费链路按 H 族图说 H2 是「Event System → TracingMiddleware → OpenTelemetry span → AgentScope Studio」。trace 记录最终落进已部署的 MySQL 加对象存储,不上重型可观测中间件;观测要早建,是为了 spike 调试和成本对账当下就用得上,而不是要先铺一套独立基建。`traceId` 贯穿整条生成任务链路,既是反查键,也是成本关联键——它对接 `tier2-verdict.schema.json``evidence.traceId`,让一次验收终判能反查到它对应的全过程轨迹。成本侧由 `cost.py` 读 new-api 的 `logs.quota` 权威口径,按 `traceId` 关联后把人民币金额 best-effort 回填进 `cost.cost_rmb`(trace 阶段只记 token,不携金额估算)。
## 两份 schema 的关系
本目录维护两份 trace schema,对应两条异构生成线:
| schema 文件 | 对应生成线 | 特有扩展段 |
|---|---|---|
| `tier2-trace-event.schema.json` | tier2 富游戏自治线(Python / AgentScope ReAct) | `ext`:reasoning / action / observation 三段(ReAct 推理→动作→观察) |
| `saa-trace-event.schema.json` | SAA 16 节点廉价生成线(Java / Spring AI Alibaba) | `ext`:nodeIndex / nodeName / nodeStatus(节点位置与执行状态) |
核心五字段(`traceId` / `step` / `cost` / `verdict` / `timestamp`)在两份 schema 里**完全同名同义**——这是 H1「接口对称」子集,让两条线写进同一张 trace 表时无需 join 也能按 `traceId` 对账。`ext` 扩展段各写各的、字段不要求对齐(H1「内容不对称」):tier2 的推理/动作/观察三段 SAA 轨不存在,SAA 的节点编号/名称/状态 tier2 轨也不存在,**没有的字段绝不编造**。
谁消费哪份:目前 tier2 侧 adapter(`tier2/gen-worker/observability/trace.py``TraceAdapter`)已产出并落盘(`JsonlFileSink` 写进 `_tier2-gen/<game_id>/trace.jsonl`);SAA 侧 adapter 在 Java 后端待产线接线(控制面 phase-1)。
## 与 H 族图说的对应
设计权威是 `docs/architecture/架构/生成引擎/tier2细节图说-H-观测与成本.md`

View File

@ -0,0 +1,88 @@
{
"$schema": "https://json-schema.org/draft/2020-12/schema",
"$id": "https://wanxiang.ai/contracts/trace/saa-trace-event.schema.json",
"title": "SaaTraceEvent",
"description": "SAA(Spring AI Alibaba)16 节点廉价生成线的统一 trace 事件契约(H 族 · A2)。【这是什么】一条 trace 记录 = 一次生成轨迹里的一步,由 SAA Java adapter 产出。核心五字段(traceId / step / cost / verdict / timestamp)与 tier2-trace-event.schema.json 完全同名同义(H1 接口对称子集);ext 扩展段是 SAA 这一轨独有的,记录当前节点在 16 节点图上的位置与执行状态(H1 内容不对称:各轨各写各的,没有的字段绝不编造)。【为什么独立成约】两条异构生成线写进同一张 trace 表,靠的是「接口对称、内容不对称」。本 schema 定义 SAA 轨写进统一表的记录形状;tier2 那一轨的扩展段在 tier2-trace-event.schema.json 里,字段与本 schema 的 ext 不同名、也不要求对齐。【谁产】SAA Java 线的 trace adapter(game-cloud/game-module-aigc 或 studio 模块,待接线);本 schema 是字段口径约束,不是采集机制。【状态】spec 已立,SAA Java adapter 待产线接线(控制面 phase-1)。详见 contracts/trace/README.md 与 docs/architecture/架构/生成引擎/tier2细节图说-H-观测与成本.md 图 H1。",
"type": "object",
"required": ["traceId", "step", "timestamp", "ext"],
"additionalProperties": false,
"properties": {
"traceId": {
"description": "公共核心子集字段(两条生成线同名同义必填)。一次生成的轨迹主键,贯穿整条生成任务链路;是对账 / 成本关联键,对接 tier2-verdict.schema.json 的 evidence.traceId。SAA 侧由 Java adapter 在构造时钉定,本次轨迹所有 step 共用同一个 traceId。",
"type": "string",
"minLength": 1
},
"step": {
"description": "公共核心子集字段。第几步(SAA 轨 = 第几个执行节点,0 起;对应 ext.nodeIndex。tier2 轨语义为 ReAct 第几轮 reply / 工具调用步,两者语义不同但字段名对称)。",
"type": "integer",
"minimum": 0
},
"cost": {
"description": "公共核心子集字段。这一步折成人民币的成本载体;仅在 SAA 节点真正发起模型调用时非空,其余节点步为 null。cost_rmb 由 cost.py 按 new-api quota 权威口径事后 best-effort 回填(成本权威源是 new-api 计费行,trace 不携金额估算)。",
"type": ["object", "null"],
"$ref": "#/$defs/cost"
},
"verdict": {
"description": "公共核心子集字段。这一步 / 这道门的裁决(可空)。SAA 轨当前在节点执行终态时填入(success / error / skipped),与 ext.nodeStatus 语义一致;门级裁决由裁决引擎 judge 另行回填,不在 adapter 映射阶段产生。",
"type": ["string", "null"]
},
"timestamp": {
"description": "公共核心子集字段。本步节点执行的发生时刻(ISO8601 字符串)。节点缺时间戳时为兜底空串。",
"type": "string"
},
"ext": {
"description": "SAA 扩展段(不对称段,各轨各写各的;tier2 轨的扩展段字段与此不同名也不要求对齐)。记录当前节点在 SAA 16 节点图上的位置(nodeIndex / nodeName)与执行状态(nodeStatus)。additionalProperties:false 锁死段名,防止 adapter 之外注入未约定字段。",
"$ref": "#/$defs/ext"
}
},
"$defs": {
"cost": {
"type": ["object", "null"],
"description": "成本载体(节点发起模型调用时非空)。tokens 由 adapter 实填(从模型调用 response 抓 input_tokens / output_tokens);cost_rmb 由 cost.py 按 new-api quota 事后回填(权威口径,trace 不携金额估算)。",
"additionalProperties": false,
"required": ["tokens"],
"properties": {
"tokens": {
"description": "本次模型调用的 token 用量。in / out 由 adapter 实填;cached 预留缓存命中 token(当前 adapter 未填)。",
"type": "object",
"additionalProperties": false,
"required": ["in", "out"],
"properties": {
"in": { "description": "输入 token 数。", "type": "integer", "minimum": 0 },
"out": { "description": "输出 token 数。", "type": "integer", "minimum": 0 },
"cached": { "description": "缓存命中 token 数(可选,预留;当前 adapter 未实填)。", "type": "integer", "minimum": 0 }
}
},
"cost_rmb": {
"description": "本步折算人民币(可选)。adapter 不填,由 cost.py 读 new-api logs.quota 按 traceId 关联后 best-effort 回填。",
"type": "number",
"minimum": 0
}
}
},
"ext": {
"type": "object",
"description": "SAA 扩展段对象。三个必填字段描述当前节点在 16 节点图上的位置与执行状态。additionalProperties:false 锁死字段集。",
"additionalProperties": false,
"required": ["nodeIndex", "nodeName", "nodeStatus"],
"properties": {
"nodeIndex": {
"description": "当前节点在 SAA 16 节点执行图中的位置(0 起整数;与顶层 step 同义但独立落盘,便于按节点序号切片分析)。",
"type": "integer",
"minimum": 0,
"maximum": 15
},
"nodeName": {
"description": "SAA 节点的可读名称(如 game_concept_generator / code_writer / asset_patcher 等;由 Java adapter 按 SAANode.name 填入,便于在 trace 里直接定位是哪个节点步)。",
"type": "string",
"minLength": 1
},
"nodeStatus": {
"description": "节点执行状态(终态才填;running 表示节点仍在执行中——仅在 adapter 需要记录中间态时使用)。该值同时被填进顶层 verdict 作这一步裁决。",
"type": "string",
"enum": ["pending", "running", "success", "error", "skipped"]
}
}
}
}
}

View File

@ -28,6 +28,8 @@ contracts/trace 现行边界:统一 trace 契约的正式 schema 落处 = contra
from __future__ import annotations
import json
from pathlib import Path
from typing import Any, Callable, Optional
@ -377,6 +379,74 @@ class TraceAdapter:
}
class JsonlFileSink:
"""把统一 trace 事件(TraceStep dict)追加写到 JSONL 文件的落盘 sink。
用法:
sink = JsonlFileSink(workdir / "trace.jsonl")
adapter = TraceAdapter(trace_id="...", sink=sink)
为什么要它:
- 五字段 schema(traceId / step / cost / verdict / timestamp)原样落盘,供成本对账replay ;
- 每条事件是独立 JSON (JSONL),追加写无需整文件加锁,适合长跑生成;
- 落盘路径 = 生成产物 workdir / "trace.jsonl", game_id 一一对应,随产物一起留存
best-effort 铁律( TraceAdapter 一致):
- 目录不存在则自动创建;
- 写失败只 print 告警,**绝不抛异常**绝不阻断主生成流程(trace 是旁路观测)
"""
def __init__(self, output_path: "Path | str") -> None:
"""
Args:
output_path: trace.jsonl 落盘路径(全路径);父目录不存在时在首次写前自动创建
"""
self._path = Path(output_path)
# 预建父目录(best-effort;目录已存在不报错)。
try:
self._path.parent.mkdir(parents=True, exist_ok=True)
except Exception as exc:
# 预建失败(如路径不可写)只告警,每次 __call__ 时再尝试一次。
print(
f"[tier2-trace] 预建 trace.jsonl 目录失败(best-effort,写时再试):"
f"{self._path.parent}{exc}",
flush=True,
)
def __call__(self, trace_step: dict) -> None:
"""把一条 TraceStep dict 序列化为 JSON 追加写一行到 trace.jsonl。
五字段(traceId / step / cost / verdict / timestamp) TraceAdapter.ingest 已填进
trace_step(TraceStep dict 子类),此处原样序列化落盘,保留 ext 扩展段
写失败(磁盘满 / 权限错等) print 告警,绝不抛 一次落库失败不得废掉已成功的生成产物
"""
try:
# 防御性再建一次目录(首次 __init__ 预建可能因时序失败)。
self._path.parent.mkdir(parents=True, exist_ok=True)
line = json.dumps(trace_step, ensure_ascii=False)
with self._path.open("a", encoding="utf-8") as f:
f.write(line + "\n")
except Exception as exc:
print(
f"[tier2-trace] trace.jsonl 写失败(best-effort,不阻断生成,path={self._path}):{exc}",
flush=True,
)
def make_jsonl_sink(output_path: "Path | str") -> "JsonlFileSink":
"""构造 JSONL 落盘 sink,目标路径 output_path;可直接塞进 TraceAdapter(sink=...)。
用法示例(在编排器 run_studio ):
sink = make_jsonl_sink(run._workdir(game_id) / "trace.jsonl")
tracer = Tier2TraceMiddleware(trace_id=game_id, sink=sink)
:param output_path: trace.jsonl 落盘路径(全路径)
:return: JsonlFileSink 实例(Callable[[dict], None])
"""
return JsonlFileSink(output_path)
def with_studio(
trace_id: str,
*,

View File

@ -0,0 +1,272 @@
"""service/admin_routes.py —— tier2 管理面只读 API(phase-1 · U-C3a)。
三个只读 GET 端点,供运营回看生成配置与轨迹:
GET /control/roles 各生成角色当前用的模型( worker/config.py 常量读)
GET /control/traces/{id} run._workdir(traceId)/trace.jsonl 返回全部步骤
GET /control/cost/{id} trace.jsonl new-api quota 口径成本对账
设计决策(附中文注释说明):
全部 GET 只读,不触发任何生成不写文件不起真服务;
worker/config.py 顶层 import agentscope(mini-desktop 上有;6c6g/测试环境无),
本模块对 config 一律惰性 import + try/except,确保 6c6g py_compile / import 不炸;
/control/traces/{id} 文件不存在返 404(REST 语义:资源不存在即 404),响应体带 found=false
方便客户端分支一致处理,优于返回 200+found=false(后者歧义:200 暗示"有了");
/control/cost/{id} 取价不可达 pricingSource="degraded" + 200(不抛 500),
对齐 best-effort 铁律(成本观测不得中断主链);
_workdir_for_trace run._workdir 同口径独立实现(无需 import run),便于轻量测试打桩
"""
from __future__ import annotations
import json
import os
from pathlib import Path
from fastapi import APIRouter
from fastapi.responses import JSONResponse
# 路由前缀 /control;标签 admin 供 OpenAPI docs 分组展示。
router = APIRouter(prefix="/control", tags=["admin"])
# ─────────────────────────────────────────────────────────────────────────────
# 辅助:角色 → 模型映射
# ─────────────────────────────────────────────────────────────────────────────
def _get_roles() -> list[dict]:
"""读各生成角色当前配的模型名。
优先路:import worker.config 取真实编译常量(只在 mini-desktop 上成功,那里装了 agentscope)
回落路:worker.config 因缺 agentscope import 失败时,直接读 worker.genconfig( agentscope 依赖)
两条路产出相同形状:[{"role": ..., "model": ...}, ...]
角色说明:
writer tier2 单写 ReAct agent(默认 MiniMax-M3;env TIER2_MODEL 可临时压)
cheap_flash 便宜档主力(deepseek-v4-flash, new-api OpenAI 兼容路)
cheap_pro 便宜档强档(deepseek-v4-pro,退路树 Q1 )
baseline 跨族对照(当前 deepseek-v4-pro;网关开通 Opus/Fable 后改 generation.yaml 切换)
"""
try:
# 优先路:worker.config 顶层 import agentscope(mini-desktop 真跑时走这条)。
from worker import config as _cfg # noqa: PLC0415 — 惰性 import
return [
{"role": "writer", "model": _cfg.model_name_from_env()},
{"role": "cheap_flash", "model": _cfg.DEEPSEEK_FLASH},
{"role": "cheap_pro", "model": _cfg.DEEPSEEK_PRO},
{"role": "baseline", "model": _cfg.BASELINE_CROSSCHECK},
]
except Exception:
# 回落路:agentscope 未装(6c6g / 测试环境)→ genconfig 直读,无 agentscope 依赖。
from worker import genconfig as _gc # noqa: PLC0415
return [
{"role": "writer",
"model": os.environ.get("TIER2_MODEL",
_gc.get("model", "default_model_name", "MiniMax-M3"))},
{"role": "cheap_flash",
"model": _gc.get("model", "deepseek_flash", "deepseek-v4-flash")},
{"role": "cheap_pro",
"model": _gc.get("model", "deepseek_pro", "deepseek-v4-pro")},
{"role": "baseline",
"model": _gc.get("model", "baseline_crosscheck", "deepseek-v4-pro")},
]
# ─────────────────────────────────────────────────────────────────────────────
# 辅助:workdir 路径计算(与 run._workdir 同口径,无需 import run)
# ─────────────────────────────────────────────────────────────────────────────
def _workdir_for_trace(trace_id: str) -> Path:
"""计算 trace_id 对应的 tier2 工程目录,口径与 worker/run._workdir 完全一致。
为何独立实现而不直接 import worker.run:
run.py 顶层 import archetypes(agentscope ),保持本模块尽量轻量(6c6g / 测试环境友好);
路径计算是纯 pathlib 操作,独立复制最小逻辑更清晰
路径:
本文件在 tier2/gen-worker/service/admin_routes.py 上溯 4 级到仓根:
service/ gen-worker/ tier2/ 仓根(games-development-ai)
GEN_DIR = 仓根/game-runtime/games/_tier2-gen(tier2 专用生成目录)
此函数是测试桩点:测试用 unittest.mock.patch 替换它指向临时目录即可,无需真实工程目录存在
"""
repo_root = Path(__file__).resolve().parents[3]
gen_dir = repo_root / "game-runtime" / "games" / "_tier2-gen"
return gen_dir / trace_id
# ─────────────────────────────────────────────────────────────────────────────
# 辅助:从 trace 步骤提取 tokens_by_model
# ─────────────────────────────────────────────────────────────────────────────
def _extract_tokens_by_model(steps: list[dict]) -> dict[str, dict[str, int]]:
"""从 trace 步骤列表提取 tokens_by_model 聚合(按模型分 in/out/cached token 累加)。
关联规则( to_trace_step 映射规则一致):
- ext.raw.model_name 非空 更新当前模型名(对应 ModelCallStartEvent)
- cost.tokens.in / out 非零 记入当前模型的 bucket(对应 ModelCallEndEvent)
跨多轮模型调用:同一模型的 token 累加;未见模型名时用 "unknown" 兜底
:param steps: trace.jsonl 解析后的步骤列表
:return: {model_name: {"in": Σin, "out": Σout, "cached": Σcached}}
"""
tokens_by_model: dict[str, dict[str, int]] = {}
current_model = "unknown" # 每遇 ModelCallStart 类事件更新
for step in steps:
# 更新当前模型名:从 ext.raw.model_name 读(ModelCallStartEvent 标记)。
ext = step.get("ext") or {}
raw = ext.get("raw") or {}
model_name_in_step = raw.get("model_name")
if model_name_in_step:
current_model = str(model_name_in_step)
# 读 token 计数:cost.tokens.in / out(ModelCallEndEvent / Tier2ModelCallPhaseEvent 填)。
cost = step.get("cost") or {}
tokens = cost.get("tokens") if isinstance(cost, dict) else None
if not isinstance(tokens, dict):
continue
in_tok = int(tokens.get("in") or 0)
out_tok = int(tokens.get("out") or 0)
# cached 字段在 trace 当前格式里一般不出现(由 RecordingChatModel 记录,非 trace 字段),
# 预留:若将来 trace 扩展落 tokens.cached 可直接读到。
cached_tok = int(tokens.get("cached") or 0)
if in_tok == 0 and out_tok == 0:
continue # 零 token 步骤跳过(非 ModelCallEnd)
bucket = tokens_by_model.setdefault(current_model, {"in": 0, "out": 0, "cached": 0})
bucket["in"] += in_tok
bucket["out"] += out_tok
bucket["cached"] += cached_tok
return tokens_by_model
# ─────────────────────────────────────────────────────────────────────────────
# 端点一:GET /control/roles
# ─────────────────────────────────────────────────────────────────────────────
@router.get("/roles")
def get_roles() -> dict:
"""GET /control/roles —— 各生成角色当前配的模型(只读,无副作用)。
数据源:worker/config.py 编译常量(rolemodel 静态映射);agentscope 不可达时回落 genconfig
返回:{"roles": [{"role": "writer", "model": "MiniMax-M3"}, ...]}
"""
return {"roles": _get_roles()}
# ─────────────────────────────────────────────────────────────────────────────
# 端点二:GET /control/traces/{trace_id}
# ─────────────────────────────────────────────────────────────────────────────
@router.get("/traces/{trace_id}")
def get_traces(trace_id: str):
"""GET /control/traces/{traceId} —— 读 trace.jsonl 返回全部步骤。
数据源:run._workdir(traceId)/trace.jsonl(JsonlFileSink 写的 JSONL,五字段:
traceId / step / cost / verdict / timestamp + ext 扩展段)
返回(200):{"traceId":..., "steps":[...], "count":n, "found":true}
文件不存在(404):{"traceId":..., "steps":[], "count":0, "found":false}
REST 选型理由(404 而非 200+found=false):
资源不存在 = 404 HTTP 语义约定,让客户端无需解析 found 字段即可判断;
响应体仍带 found=false + steps 以方便调用方分支一致解析
健壮性:坏行跳过不崩(best-effort,对齐 JsonlFileSink best-effort 铁律)
"""
workdir = _workdir_for_trace(trace_id)
trace_file = workdir / "trace.jsonl"
if not trace_file.exists():
# 资源不存在 → 404;带 found=false 供调用方分支处理。
return JSONResponse(
status_code=404,
content={"traceId": trace_id, "steps": [], "count": 0, "found": False},
)
steps: list[dict] = []
for raw_line in trace_file.read_text(encoding="utf-8").splitlines():
line = raw_line.strip()
if not line:
continue
try:
steps.append(json.loads(line))
except Exception:
# 坏行跳过:不崩、不计入 count(best-effort;trace 写失败不得影响读)。
pass
return {"traceId": trace_id, "steps": steps, "count": len(steps), "found": True}
# ─────────────────────────────────────────────────────────────────────────────
# 端点三:GET /control/cost/{trace_id}
# ─────────────────────────────────────────────────────────────────────────────
@router.get("/cost/{trace_id}")
def get_cost(trace_id: str):
"""GET /control/cost/{traceId} —— 对 trace.jsonl 做 new-api quota 口径成本对账。
数据源:
trace.jsonl 提取 tokens_by_model(按模型分 in/out token 累加)
newapi_pricing.fetch_pricing_params() 活取计费三件套(pricing/qpu/usd_rate)
cost_for_run(tokens, pricing, qpu, usd_rate) new-api quota 公式折¥
返回(200):
{"traceId":..., "costRmb":float, "byModel":{model:{quota,usd,rmb,...}}, "pricingSource":"live"}
降级处理(best-effort 铁律,成本观测不得阻断):
- trace 不存在 404( /traces 同语义)
- token 记录 pricingSource="degraded", costRmb=0.0, byModel={} 返回 200
- 取价不可达(网络/6c6g) pricingSource="degraded", costRmb=0.0, byModel={} 返回 200
- cost 模块 import 失败 同上 degraded 路径(防御)
"""
workdir = _workdir_for_trace(trace_id)
trace_file = workdir / "trace.jsonl"
if not trace_file.exists():
# trace 文件不存在 → 404(与 /traces 端点语义一致)。
return JSONResponse(
status_code=404,
content={"traceId": trace_id, "costRmb": 0.0, "byModel": {},
"pricingSource": "notfound"},
)
# ① 读 trace.jsonl 并解析步骤(坏行跳过,best-effort)。
steps: list[dict] = []
for raw_line in trace_file.read_text(encoding="utf-8").splitlines():
line = raw_line.strip()
if not line:
continue
try:
steps.append(json.loads(line))
except Exception:
pass # 坏行跳过
# ② 提取 tokens_by_model。
tokens_by_model = _extract_tokens_by_model(steps)
if not tokens_by_model:
# 无 token 记录(trace 全是非 ModelCallEnd 步骤,或还没生成过)→ degraded。
print(f"[tier2-admin] GET /cost/{trace_id}: trace 无 token 记录,返回 degraded。",
flush=True)
return {"traceId": trace_id, "costRmb": 0.0, "byModel": {}, "pricingSource": "degraded"}
# ③ 惰性 import cost 模块(observability 不依赖 agentscope,但防御性兜住任何 import 错误)。
try:
from observability.newapi_pricing import fetch_pricing_params # noqa: PLC0415
from observability.cost import cost_for_run # noqa: PLC0415
except Exception as exc:
# 极端环境下 import 失败(如依赖缺失)→ degraded,不抛 500。
print(f"[tier2-admin] GET /cost/{trace_id}: 导入 cost/pricing 模块失败(degraded):{exc}",
flush=True)
return {"traceId": trace_id, "costRmb": 0.0, "byModel": {}, "pricingSource": "degraded"}
# ④ 活取 new-api 计费参数(best-effort:取不到 → degraded,不抛)。
pricing_params = fetch_pricing_params() # 内部 best-effort,失败返 None
if pricing_params is None:
# 取价不可达(6c6g / 网关下线 / httpx 超时)→ degraded 标识,返 200 不抛 500。
print(f"[tier2-admin] GET /cost/{trace_id}: new-api 取价不可达,返回 degraded。", flush=True)
return {"traceId": trace_id, "costRmb": 0.0, "byModel": {}, "pricingSource": "degraded"}
# ⑤ 按 new-api quota 公式折¥(cost_for_run 纯函数,全显式传参,不读全局)。
cost_result = cost_for_run(
tokens_by_model,
pricing_params["pricing"],
pricing_params["qpu"],
pricing_params["usd_rate"],
)
return {
"traceId": trace_id,
"costRmb": cost_result["cost_rmb"],
"byModel": cost_result.get("by_model", {}),
"pricingSource": "live",
}

View File

@ -238,7 +238,7 @@ def build_app(*, title: str = SERVICE_TITLE) -> "FastAPI":
flush=True,
)
return create_app(
app = create_app(
storage=RedisStorage(**storage_params),
message_bus=RedisMessageBus(**bus_params),
workspace_manager=LocalWorkspaceManager(basedir=ws_basedir),
@ -252,6 +252,14 @@ def build_app(*, title: str = SERVICE_TITLE) -> "FastAPI":
title=title,
)
# ── 挂管理面只读路由(phase-1 U-C3a;GET /control/roles · /control/traces/{id} · /control/cost/{id})──
# 惰性 import 红线:admin_routes 顶层 import fastapi。create_app 既已建出 FastAPI app,
# fastapi 在能调 build_app 的环境(mini-desktop)必然在场;在 build_app 体内 import,保证 6c6g
# 仅 `import service.app`(不调 build_app)时不牵 fastapi、不破坏顶层惰性红线。
from .admin_routes import router as admin_router # noqa: PLC0415 —— 惰性 import 红线
app.include_router(admin_router)
return app
def main() -> None:
"""部署入口:起 uvicorn 跑 tier2 Agent Service(只在 mini-desktop 装好 agentscope[full]+redis 后跑)。

View File

@ -0,0 +1,179 @@
"""test_admin_routes.py —— U-C3a 管理面只读 API 单测(TDD:先红后绿)。
FastAPI TestClient(进程内,**不起真 uvicorn 服务**不连 Redis/agentscope不跑 gen)验三个只读端点:
GET /control/roles 200 + 各角色模型(至少含 writer)
GET /control/traces/{id} 临时样本 trace.jsonl 打桩 _workdir 200 + steps 数对 + 五字段;
坏行 best-effort 跳过;不存在 id 404 + found:false
GET /control/cost/{id} 样本 trace + 打桩取价:live costRmb;取价不可达 degraded 标识,不崩
跑法( cheap-worker venv 解释器,绝不 import/读写 cheap-worker 源码):
PYTHONPATH=tier2/gen-worker <venv>/bin/python -m pytest tier2/gen-worker/tests/test_admin_routes.py -v
"""
from __future__ import annotations
import json
import os
import sys
from pathlib import Path
from unittest import mock
import pytest
from fastapi import FastAPI
from fastapi.testclient import TestClient
# 包内/直跑兼容:把 gen-worker/ 加进 sys.path,使 service / observability / worker 可 import
# (PYTHONPATH 已指 gen-worker 时这步是冗余保险;直接 python 跑本文件也能解析包)。
_HERE = os.path.dirname(os.path.abspath(__file__))
_GW = os.path.dirname(_HERE) # gen-worker/
if _GW not in sys.path:
sys.path.insert(0, _GW)
from service.admin_routes import router # noqa: E402 —— 被测路由
# ── 合法样本 trace.jsonl 内容(五字段 + ext)──────────────────────────────────────
# 覆盖:ReplyStart(reply 边界)/ ModelCallStart(带 model_name,供成本按模型归集)/
# ModelCallEnd(带 token,供成本折算)/ ToolResultEnd(verdict=success)。
_SAMPLE_TRACE_LINES = [
{"traceId": "t-sample", "step": 0, "cost": None, "verdict": None,
"timestamp": "2026-06-27T10:00:00",
"ext": {"raw": {"event": "ReplyStartEvent", "agent_name": "writer"}}},
{"traceId": "t-sample", "step": 1, "cost": None, "verdict": None,
"timestamp": "2026-06-27T10:00:01",
"ext": {"raw": {"event": "ModelCallStartEvent", "model_name": "MiniMax-M3"}}},
{"traceId": "t-sample", "step": 2, "cost": {"tokens": {"in": 1200, "out": 800}}, "verdict": None,
"timestamp": "2026-06-27T10:00:02",
"ext": {"raw": {"event": "ModelCallEndEvent", "model_call_end": True}}},
{"traceId": "t-sample", "step": 3, "cost": None, "verdict": "success",
"timestamp": "2026-06-27T10:00:03",
"ext": {"observation": {"phase": "observation", "state": "success"}}},
]
@pytest.fixture
def client() -> TestClient:
"""最小 FastAPI app,只挂被测 admin router(不调 build_app、不连 Redis/agentscope)。"""
app = FastAPI()
app.include_router(router)
return TestClient(app)
def _write_sample_trace(tmp_path: Path) -> Path:
"""在临时目录写样本工程 workdir + trace.jsonl(末尾附一坏行,验 best-effort 跳过)。
返回样本 workdir(tmp_path/'t-sample'),其下 trace.jsonl = 五字段合法行 + 1 坏行
"""
workdir = tmp_path / "t-sample"
workdir.mkdir(parents=True, exist_ok=True)
lines = [json.dumps(o, ensure_ascii=False) for o in _SAMPLE_TRACE_LINES]
lines.append("{ 这是一行坏 JSON —— best-effort 应跳过、不计入 count ") # 坏行
(workdir / "trace.jsonl").write_text("\n".join(lines) + "\n", encoding="utf-8")
return workdir
# ════════════════════════════════════════════════════════════════════════════
# 端点一:GET /control/roles
# ════════════════════════════════════════════════════════════════════════════
def test_roles_returns_writer(client):
"""GET /control/roles → 200 + roles 列表含 writer 角色且其 model 非空。"""
resp = client.get("/control/roles")
assert resp.status_code == 200, resp.text
data = resp.json()
assert "roles" in data, f"响应缺 roles 字段:{data}"
roles = {r["role"]: r["model"] for r in data["roles"]}
assert "writer" in roles, f"roles 应含 writer:{roles}"
assert roles["writer"], f"writer 的 model 不应为空:{roles}"
print(f" [ok] roles={roles}")
# ════════════════════════════════════════════════════════════════════════════
# 端点二:GET /control/traces/{id}
# ════════════════════════════════════════════════════════════════════════════
def test_traces_found_with_sample(client, tmp_path):
"""样本 trace.jsonl 打桩 _workdir → 200 + steps 数对(坏行跳过)+ 五字段在。"""
_write_sample_trace(tmp_path)
# 打桩:_workdir_for_trace(tid) → tmp_path/tid(端点据此读 <workdir>/trace.jsonl)。
with mock.patch("service.admin_routes._workdir_for_trace",
side_effect=lambda tid: tmp_path / tid):
resp = client.get("/control/traces/t-sample")
assert resp.status_code == 200, resp.text
data = resp.json()
assert data["found"] is True, f"found 应为 True:{data}"
# 坏行被跳过 → count == 合法样本行数。
assert data["count"] == len(_SAMPLE_TRACE_LINES), \
f"steps 数应为 {len(_SAMPLE_TRACE_LINES)}(坏行跳过),实际 {data['count']}"
assert len(data["steps"]) == data["count"]
# 五字段齐(traceId/step/cost/verdict/timestamp)。
core = {"traceId", "step", "cost", "verdict", "timestamp"}
for s in data["steps"]:
assert core <= set(s.keys()), f"步骤缺五字段,缺 {core - set(s.keys())}:{s}"
# 抽验内容:第 2 步 token、第 3 步 verdict。
assert data["steps"][2]["cost"]["tokens"]["in"] == 1200
assert data["steps"][3]["verdict"] == "success"
print(f" [ok] count={data['count']}(坏行已跳过),五字段齐,token/verdict 对")
def test_traces_not_found_returns_404(client, tmp_path):
"""不存在的 traceId → 404 + found:false + 空 steps。"""
with mock.patch("service.admin_routes._workdir_for_trace",
side_effect=lambda tid: tmp_path / tid):
resp = client.get("/control/traces/does-not-exist")
assert resp.status_code == 404, resp.text
data = resp.json()
assert data["found"] is False, f"found 应为 False:{data}"
assert data["steps"] == [], f"steps 应为空:{data}"
assert data["count"] == 0
print(f" [ok] 不存在 traceId → 404 found=false")
# ════════════════════════════════════════════════════════════════════════════
# 端点三:GET /control/cost/{id}
# ════════════════════════════════════════════════════════════════════════════
def test_cost_live_with_stub_pricing(client, tmp_path):
"""样本 trace + 打桩取价(live)→ 200 + costRmb>0 + pricingSource=live + byModel 含模型。"""
_write_sample_trace(tmp_path)
# 桩 new-api 计费三件套(模拟网关返回,避免真网络;cost_for_run 是纯函数,折算确定)。
stub_params = {
"pricing": {"MiniMax-M3": {"model_ratio": 2.0, "completion_ratio": 4.0, "cache_ratio": 0.1}},
"qpu": 500000.0,
"usd_rate": 7.3,
}
with mock.patch("service.admin_routes._workdir_for_trace",
side_effect=lambda tid: tmp_path / tid), \
mock.patch("observability.newapi_pricing.fetch_pricing_params",
return_value=stub_params):
resp = client.get("/control/cost/t-sample")
assert resp.status_code == 200, resp.text
data = resp.json()
assert "costRmb" in data, f"响应应含 costRmb:{data}"
assert data["pricingSource"] == "live", f"pricingSource 应为 live:{data}"
assert data["costRmb"] > 0, f"样本有 token,costRmb 应 > 0:{data}"
assert "MiniMax-M3" in data["byModel"], f"byModel 应含 MiniMax-M3:{data}"
print(f" [ok] live costRmb={data['costRmb']} byModel={list(data['byModel'])}")
def test_cost_degraded_when_pricing_unreachable(client, tmp_path):
"""取价不可达(fetch_pricing_params→None)→ 200 + pricingSource=degraded + 含 costRmb,不崩。"""
_write_sample_trace(tmp_path)
with mock.patch("service.admin_routes._workdir_for_trace",
side_effect=lambda tid: tmp_path / tid), \
mock.patch("observability.newapi_pricing.fetch_pricing_params",
return_value=None):
resp = client.get("/control/cost/t-sample")
assert resp.status_code == 200, resp.text # 取价失败不抛 500
data = resp.json()
assert data["pricingSource"] == "degraded", f"pricingSource 应为 degraded:{data}"
assert "costRmb" in data, f"degraded 也应带 costRmb 字段:{data}"
print(f" [ok] degraded(取价不可达)costRmb={data['costRmb']} 不崩")
def test_cost_trace_not_found_returns_404(client, tmp_path):
"""cost 对不存在的 traceId → 404(与 /traces 同语义,不崩)。"""
with mock.patch("service.admin_routes._workdir_for_trace",
side_effect=lambda tid: tmp_path / tid):
resp = client.get("/control/cost/no-such-trace")
assert resp.status_code == 404, resp.text
data = resp.json()
assert data["pricingSource"] == "notfound", f"应标 notfound:{data}"
print(f" [ok] cost 不存在 traceId → 404 notfound")

View File

@ -0,0 +1,130 @@
"""test_jsonl_sink.py —— U-C1a JSONL sink 落盘单测(TDD:先红后绿)。
:python3 tests/test_jsonl_sink.py( pytest,__main__ 跑全部断言)
JsonlFileSink:
1. emit 三条 trace step trace.jsonl 落盘;
2. 每行是合法 JSON;
3. 五字段(traceId / step / cost / verdict / timestamp)齐全;
4. 目录不存在 自动创建;
5. 写失败(不可写路径) 只告警不抛,不中断调用方
"""
import sys
import os
import json
import tempfile
from pathlib import Path
_HERE = os.path.dirname(os.path.abspath(__file__))
_GW = os.path.dirname(_HERE) # gen-worker/
sys.path.insert(0, _GW)
from observability.trace import JsonlFileSink, make_jsonl_sink, TraceAdapter # noqa: E402
# ── 测试一:核心落盘 + 五字段完整性 ──────────────────────────────────────────────
def test_jsonl_sink_creates_file_and_lines():
"""emit 三条 trace step → trace.jsonl 落盘、每行合法 JSON、五字段齐、内容对。"""
with tempfile.TemporaryDirectory() as tmpdir:
out = Path(tmpdir) / "trace.jsonl"
sink = JsonlFileSink(out)
# 用 TraceAdapter + dict 桩事件(无需 AgentScope 真环境)驱动 sink。
adapter = TraceAdapter(trace_id="test-trace-001", sink=sink)
fake_events = [
# ReplyStartEvent → 落 raw.reply_boundary / agent_name
{"type": "ReplyStartEvent", "created_at": "2026-06-29T10:00:00",
"reply_id": "r1", "name": "writer"},
# ModelCallEndEvent → cost.tokens 非空
{"type": "ModelCallEndEvent", "created_at": "2026-06-29T10:00:01",
"input_tokens": 100, "output_tokens": 50},
# ToolResultEndEvent → verdict=success
{"type": "ToolResultEndEvent", "created_at": "2026-06-29T10:00:02",
"state": "success"},
]
for e in fake_events:
adapter.ingest(e)
# 断言①:文件已落盘
assert out.exists(), f"trace.jsonl 应存在,但未落盘:{out}"
lines = out.read_text(encoding="utf-8").strip().split("\n")
# 断言②:三条事件 → 三行
assert len(lines) == 3, f"应有 3 行,实际 {len(lines)} 行:\n" + "\n".join(lines)
# 断言③:每行合法 JSON + 五字段齐
CORE_FIELDS = {"traceId", "step", "cost", "verdict", "timestamp"}
for i, line in enumerate(lines):
obj = json.loads(line) # 若非合法 JSON 此处抛
missing = CORE_FIELDS - set(obj.keys())
assert not missing, f"{i} 行缺五字段 {missing}:{line}"
assert obj["traceId"] == "test-trace-001", (
f"{i} 行 traceId 错(期望 test-trace-001,实际 {obj['traceId']})")
assert obj["step"] == i, f"{i} 行 step 错(期望 {i},实际 {obj['step']})"
# 断言④:ModelCallEnd 步 cost 非空且 token 字段正确
second = json.loads(lines[1])
assert second["cost"] is not None, "ModelCallEnd 步 cost 应非空"
assert second["cost"]["tokens"]["in"] == 100, "input_tokens 应为 100"
assert second["cost"]["tokens"]["out"] == 50, "output_tokens 应为 50"
# 断言⑤:ToolResultEnd 步 verdict=success
third = json.loads(lines[2])
assert third["verdict"] == "success", (
f"ToolResultEnd 步 verdict 应为 success,实际 {third['verdict']}")
print(f" [ok] lines={len(lines)}, traceId={json.loads(lines[0])['traceId']}, "
f"cost.tokens.in={second['cost']['tokens']['in']}, verdict={third['verdict']}")
# ── 测试二:嵌套目录自动创建 ─────────────────────────────────────────────────────
def test_jsonl_sink_auto_creates_dir():
"""目录不存在 → sink 自动创建嵌套目录;emit 成功落盘。"""
with tempfile.TemporaryDirectory() as tmpdir:
out = Path(tmpdir) / "subdir" / "nested" / "trace.jsonl"
sink = make_jsonl_sink(out)
step = {"traceId": "t2", "step": 0, "cost": None, "verdict": None,
"timestamp": "2026-06-29T00:00:00",
"ext": {"raw": {"event": "ReplyStartEvent"}}}
sink(step)
assert out.exists(), f"嵌套目录下 trace.jsonl 应自动创建:{out}"
obj = json.loads(out.read_text(encoding="utf-8").strip())
assert obj["traceId"] == "t2", f"traceId 错:{obj['traceId']}"
print(f" [ok] auto-mkdir:{out}")
# ── 测试三:写失败只告警、不抛 ────────────────────────────────────────────────────
def test_jsonl_sink_write_failure_no_raise():
"""写失败(不可写路径) → 只 print 告警,绝不抛,不中断调用方。"""
# /proc/nonexistent/ 在 macOS 下不存在且不可写,触发写失败场景。
bad_path = Path("/proc/nonexistent_for_test_sink/trace.jsonl")
sink = JsonlFileSink(bad_path)
try:
sink({"traceId": "t3", "step": 0, "cost": None, "verdict": None,
"timestamp": "", "ext": {"raw": {"event": "X"}}})
print(" [ok] write-fail → 只告警,不抛")
except Exception as e:
raise AssertionError(f"sink 写失败时不应抛异常,但抛了:{e}") from e
# ── 入口 ────────────────────────────────────────────────────────────────────────
if __name__ == "__main__":
tests = [
("test_jsonl_sink_creates_file_and_lines", test_jsonl_sink_creates_file_and_lines),
("test_jsonl_sink_auto_creates_dir", test_jsonl_sink_auto_creates_dir),
("test_jsonl_sink_write_failure_no_raise", test_jsonl_sink_write_failure_no_raise),
]
passed = failed = 0
for name, fn in tests:
try:
print(f"[RUN] {name}")
fn()
print(f"[PASS] {name}")
passed += 1
except Exception as exc:
import traceback
print(f"[FAIL] {name}: {exc}")
traceback.print_exc()
failed += 1
print(f"\n结果:{passed} 通过 / {failed} 失败")
sys.exit(0 if failed == 0 else 1)

View File

@ -51,11 +51,13 @@ except ImportError: # pragma: no cover —— 直接 python studio.py 兜底
try:
from observability.cost import tokens_by_model_from_records, cost_for_run
from observability.newapi_pricing import fetch_pricing_params, DEFAULT_QUOTA_PER_UNIT, DEFAULT_USD_RATE
from observability.trace import make_jsonl_sink
except Exception: # pragma: no cover —— 直跑/路径未就位兜底
sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
from observability.cost import tokens_by_model_from_records, cost_for_run # type: ignore
from observability.newapi_pricing import ( # type: ignore
fetch_pricing_params, DEFAULT_QUOTA_PER_UNIT, DEFAULT_USD_RATE)
from observability.trace import make_jsonl_sink # type: ignore
# 框架接缝(2.0.2):Agent + ReActConfig + UserMsg + AgentState + BYPASS 权限。
from agentscope.agent import Agent, ReActConfig
@ -365,9 +367,11 @@ async def run_studio(
toolkit = build_toolkit(session)
breaker = CircuitBreakerMiddleware() # 四道熔断 + 软刹(on_reply / on_system_prompt)
# trace 接线(H1/H2):traceId 用 game_id(贯穿本次生成,对接 verdict.evidence.traceId / 成本关联键)。
# sink=None → 只留内存 steps(spike 期收口由编排器读 adapter.summary;真落库 sink 随控制面 phase-1 接)。
# JSONL sink → 把每条 trace step(五字段 schema:traceId/step/cost/verdict/timestamp + ext)
# 追加写到产物 workdir/trace.jsonl,供成本对账 / replay 用;sink 失败 best-effort 只告警不阻断主链。
# 单实例对应单 agent;它与 breaker 都挂 on_reply,由框架按 middlewares 列表序串成洋葱链(trace 在外、纯旁路)。
tracer = Tier2TraceMiddleware(trace_id=game_id)
_trace_jsonl_path = run._workdir(game_id) / "trace.jsonl"
tracer = Tier2TraceMiddleware(trace_id=game_id, sink=make_jsonl_sink(_trace_jsonl_path))
writer_model = config.build_model(
mname, max_tokens=max_tokens, thinking_budget=thinking_budget)
@ -629,7 +633,8 @@ async def run_studio(
# RunRecord 接线(B2)从此抽 cost_yuan / tokens_by_model 填 G4 采集字段。
"cost": cost,
# trace 收口摘要(H1/H2,新增):{traceId, steps, dropped, middlewareDropped}。
# spike 期 trace step 留在 tracer.adapter.steps 内存(sink=None);真落库 sink 随控制面 phase-1 接。
# trace step 既留 tracer.adapter.steps 内存(供本摘要 / replay),又经 JsonlFileSink 落盘
# 到 workdir/trace.jsonl(控制面 phase-1 已接,见上文 make_jsonl_sink 接线点)。
"trace": trace_summary,
"wall_s": wall_s,
"file_tree": session.file_tree(),