mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-18 07:12:00 +00:00
587bf7e27c
起因是一次探索运行:Copilot 改节点时把工具绑定整份抹掉,agent 没有工具只能写「假设调用」, 校验节点的自动修复编出 total_count=0 让 schema 通过,协作成员把工具调用标记当文字输出了四轮, 轮数用完后节点把原话当成功结果交出去——运行标成已完成,数据库一次都没查。 平台可信度(先写失败测试) - Copilot 修改节点改为合并:没写的字段保留,null 才删除;协作成员按名字合并。前后端同一套规则, 工具集合变小时自查报警、改图回执列出变化 - 新校验规则:提示词点名了某个工具而节点没有绑定它,判错误,Copilot 自查会交回去修 - 校验修复只许改格式:冒出原文没有的值就判修复无效;修复这次模型调用计入用量 - 识别模型把工具调用写成文字的情况(DSML、<tool_call> 等标记):纠正一次,仍然如此就判失败并说明原因 - 协作团队用完轮数不再按成功交付:默认失败,可选降档,点出一次都没派到的成员 - 审批恢复后,同一节点里已经执行过的工具调用复用结果,不再重复执行 后端 - 查询时限交给数据库执行(Oracle call_timeout、PG statement_timeout、MySQL max_execution_time、 SQLite progress handler),可按数据源配置;超时报错与实际耗时一致 - 等审批的运行可以真正取消;会话列表带最后一轮状态;测连接结果持久化;探查结构可只看不存; 自定义工具可草稿试跑、写坏的参数定义保存时拒绝;MCP 服务改动后作废旧工具缓存; 检索命中带贡献分解;重建索引在后台跑并报进度;报错统一收进 core/errors 前端 - 公共层:管理页共用件、失败说明(lib/explain)、健康状态进公共库;目录按表各自加载并有超时, 一张表卡住不再拖垮整站;上传显示真实字节进度;离线横幅只播报不变的句子 - 画布:卡片只说游标所在的那一次执行;等审批时可放弃运行;团队收尾判定与未完成说清楚; 回放用量取那一刻的值 - 编排页:Copilot 合并语义与工具变化回执;「用完轮数时」选项;正式运行期间锁住所有改图入口 - 回答流 / 问数据:新失败类型说人话并给下一步;画布与右栏互相定位;数据源范围、回收站、重新核对 - 记录页:「航迹」页签复用画布时间轴,可拖动回放、列出关键时刻;列表行带执行 / 等人 / 失败缩略条 - 外壳:统一的离开前确认(data router),导航、⌘K、快捷键、后退都会先问 验证(沙箱):tsc 干净,pytest 838 passed(两遍),14 个检查脚本中 13 个全绿; check-ui 余 1 项旧断言(按旧日期格式找运行行),3c 更新。
214 lines
8.8 KiB
Python
214 lines
8.8 KiB
Python
from __future__ import annotations
|
|
|
|
import json
|
|
from typing import Any
|
|
|
|
from app.core.events import EventType
|
|
from app.engine.context import NodeContext, NodeError
|
|
from app.engine.expressions import ExpressionError, eval_expression
|
|
from app.engine.state import GraphState, template_context
|
|
|
|
|
|
async def run_input(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
|
"""入口节点。把运行输入按声明的字段整理进变量池,补上默认值。"""
|
|
raw = dict(state.get("input") or {})
|
|
fields = ctx.cfg("fields", []) or []
|
|
resolved: dict[str, Any] = dict(raw)
|
|
|
|
for field in fields:
|
|
name = field.get("name")
|
|
if not name:
|
|
continue
|
|
if name not in resolved or resolved[name] in (None, ""):
|
|
if field.get("required") and field.get("default") in (None, ""):
|
|
raise NodeError(ctx.node.id, f"缺少必填输入:{name}")
|
|
resolved[name] = field.get("default")
|
|
|
|
return {
|
|
"input": resolved,
|
|
"vars": resolved,
|
|
"nodes": {ctx.node.id: resolved},
|
|
}
|
|
|
|
|
|
async def run_output(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
|
"""出口节点。按配置把散落在各节点的结果收成一个结构化成果。
|
|
|
|
配了 contract 就升级成出具契约:核对指标齐不齐、叙述里的数字
|
|
能不能回指到指标集,然后给出三档判定(formal / degraded / withheld)。
|
|
判定和缺口清单一起印在成果上——降档不是悄悄地降。
|
|
"""
|
|
mapping = ctx.cfg("fields", None)
|
|
tctx = template_context(state)
|
|
|
|
if mapping:
|
|
result: dict[str, Any] = {}
|
|
for field in mapping:
|
|
name = field.get("name")
|
|
if not name:
|
|
continue
|
|
result[name] = ctx.render(field.get("value", ""), state)
|
|
else:
|
|
# 没配映射就把最后一条消息当作结果,让新建的图开箱即用
|
|
result = {"result": tctx.get("last_message", "")}
|
|
|
|
contract = ctx.cfg("contract")
|
|
if contract:
|
|
result["_issuance"] = _apply_contract(contract, state, ctx)
|
|
|
|
return {"output": result, "nodes": {ctx.node.id: result}}
|
|
|
|
|
|
def _apply_contract(
|
|
contract: dict[str, Any], state: GraphState, ctx: NodeContext
|
|
) -> dict[str, Any]:
|
|
from datetime import datetime, timezone
|
|
|
|
from app.engine.issuance import decide_tier, trace_numbers
|
|
|
|
nodes = state.get("nodes") or {}
|
|
|
|
# 1) 收指标集:contract.metrics_from 指向一个或多个口径卡节点
|
|
metrics: list[dict[str, Any]] = []
|
|
calibers: list[dict[str, str]] = []
|
|
sources = contract.get("metrics_from") or []
|
|
if isinstance(sources, str):
|
|
sources = [sources]
|
|
# gaps 记的是"校验本身哪里没跑起来",和"校验跑了但没通过"是两回事。
|
|
# 没有它,一个写错的节点名会让所有列表都是空的,看起来和"全部通过"一模一样。
|
|
gaps: list[str] = []
|
|
unresolved: list[str] = []
|
|
#: 指标 id → 它来自哪张口径卡(「销售口径 @ v2」),逐个数字回指时标出处
|
|
caliber_of: dict[str, str] = {}
|
|
for source in sources:
|
|
payload = nodes.get(source)
|
|
if isinstance(payload, dict) and payload.get("kind") == "metric_set":
|
|
metrics.extend(payload.get("metrics") or [])
|
|
calibers.append(
|
|
{
|
|
"node": source,
|
|
"caliber": payload.get("caliber", ""),
|
|
"version": payload.get("caliber_version", ""),
|
|
}
|
|
)
|
|
label = " @ ".join(str(v) for v in (payload.get("caliber"), payload.get("caliber_version")) if v)
|
|
for m in payload.get("metrics") or []:
|
|
caliber_of.setdefault(str(m.get("id")), label)
|
|
else:
|
|
# 节点不存在、还没跑到、或者根本不是口径卡——指标集就是缺的,
|
|
# 不能当作"这次没有指标要查"
|
|
unresolved.append(source)
|
|
if unresolved:
|
|
gaps.append(f"指标源未解析:{', '.join(unresolved[:5])}")
|
|
|
|
present = {m["id"] for m in metrics}
|
|
missing_required = [m for m in (contract.get("required") or []) if m not in present]
|
|
missing_expected = [m for m in (contract.get("expected") or []) if m not in present]
|
|
|
|
# 2) 叙述数字回指
|
|
declared_narrative = str(contract.get("narrative", ""))
|
|
narrative = ctx.render_str(declared_narrative, state)
|
|
# 模板取不到路径时 render 出来是空串,所以"声明了叙述但渲染成空"必须单独识别:
|
|
# 节点改名、vars 写错都会让回指校验静默变成空操作
|
|
if declared_narrative.strip() and not narrative.strip():
|
|
gaps.append("叙述模板渲染为空(路径可能写错了)")
|
|
elif not declared_narrative.strip():
|
|
gaps.append("契约没有声明叙述,数字回指未执行")
|
|
|
|
trace = (
|
|
trace_numbers(narrative, metrics, allow=contract.get("allow_numbers"))
|
|
if narrative.strip()
|
|
else None
|
|
)
|
|
unmatched = trace.unmatched if trace else []
|
|
|
|
# 声明了 metrics_from 却一个指标都没收到,同样是"没查成"
|
|
if sources and not metrics:
|
|
gaps.append("指标集为空,叙述里的数字无从回指")
|
|
|
|
# 上游有协作团队用完轮数、按降档交付的:叙述里的话不是调度者认可的结论,
|
|
# 数字全都对得上也不能盖「完整出具」
|
|
titles = {n.id: n.title for n in ctx.run.spec.nodes}
|
|
for node_id, payload in nodes.items():
|
|
if isinstance(payload, dict) and payload.get("exhausted"):
|
|
gaps.append(f"协作团队「{titles.get(node_id, node_id)}」用完 {payload.get('rounds', '?')} "
|
|
"轮仍未完成,交来的是成员最后的原话")
|
|
|
|
tier = decide_tier(
|
|
missing_required=missing_required,
|
|
missing_expected=missing_expected,
|
|
unmatched=unmatched,
|
|
strict=bool(contract.get("strict")),
|
|
gaps=gaps,
|
|
)
|
|
|
|
# 逐个数字的出处。以前只有一个计数,界面标不出「这个数来自哪个指标」
|
|
matched = [
|
|
{**hit, **({"caliber": caliber_of[hit["metric"]]} if caliber_of.get(hit.get("metric")) else {})}
|
|
for hit in (trace.matched if trace else [])
|
|
][:50]
|
|
issuance = {
|
|
"tier": tier,
|
|
"calibers": calibers, # 口径版本照常印
|
|
"metrics_checked": len(metrics),
|
|
"missing_required": missing_required,
|
|
"missing_expected": missing_expected, # 缺数据声明照常印
|
|
"unmatched_numbers": unmatched[:20],
|
|
"matched_numbers": len(trace.matched) if trace else 0,
|
|
"matched": matched,
|
|
# 校验没跑全的原因照实印出来——读的人要能分辨"查过都对"和"根本没查"
|
|
"gaps": gaps,
|
|
"declared_at": datetime.now(timezone.utc).isoformat(),
|
|
}
|
|
ctx.emit(
|
|
EventType.ISSUANCE,
|
|
tier=tier,
|
|
missing_required=missing_required,
|
|
missing_expected=missing_expected,
|
|
unmatched=len(unmatched),
|
|
calibers=calibers,
|
|
# 降档的原因要跟着事件走:只有 gaps 的时候,时间线上的出具步骤以前只能写
|
|
# 一个光秃秃的「降档出具」,横幅还说「请对照下方声明」,下方却什么都没有
|
|
gaps=gaps,
|
|
metrics_checked=len(metrics),
|
|
matched_numbers=len(trace.matched) if trace else 0,
|
|
matched=matched,
|
|
)
|
|
return issuance
|
|
|
|
|
|
async def run_transform(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
|
"""数据整形。用受限表达式把上游输出揉成下游要的形状,不用为此专门起一个 LLM。"""
|
|
tctx = template_context(state)
|
|
mode = ctx.cfg("mode", "expression")
|
|
output: Any
|
|
|
|
if mode == "expression":
|
|
expr = ctx.cfg("expression", "")
|
|
if not expr:
|
|
raise NodeError(ctx.node.id, "整形节点没有填表达式")
|
|
try:
|
|
output = eval_expression(expr, tctx)
|
|
except ExpressionError as e:
|
|
raise NodeError(ctx.node.id, f"表达式错误:{e}") from e
|
|
elif mode == "template":
|
|
output = ctx.render_str(ctx.cfg("template", ""), state)
|
|
elif mode == "json":
|
|
rendered = ctx.render_str(ctx.cfg("template", "{}"), state)
|
|
try:
|
|
output = json.loads(rendered)
|
|
except json.JSONDecodeError as e:
|
|
raise NodeError(
|
|
ctx.node.id,
|
|
f"模板渲染出来的不是合法 JSON(第 {e.lineno} 行第 {e.colno} 列附近)。"
|
|
"检查模板里的引号、逗号,字符串值要用 | json 过滤器输出",
|
|
) from e
|
|
else:
|
|
raise NodeError(ctx.node.id, f"未知的整形模式:{mode}")
|
|
|
|
updates: dict[str, Any] = {"nodes": {ctx.node.id: output}}
|
|
var_name = ctx.cfg("assign_to", "")
|
|
if var_name:
|
|
updates["vars"] = {var_name: output}
|
|
return updates
|