mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 21:32:00 +00:00
ThearchyHelios
050bea4b72
一、一键升级（只给预览，不自动保存）
- engine/upgrade.py：确定性改写，纯函数、重复执行结果不变
- R1 契约的叙述指向模型调用节点 → 换成报告撰写节点（system 和 prompt 合成写作要求，模型配置照搬），契约改成 report_from
- R2 没有契约、但模型写的文字进了出口，而且上游有取数 → 同样换成报告撰写
- R3 input → 智能体 → 出口（问数据的典型图）→ 中间插入报告撰写节点，出口字段接报告的正文；
只改直接相连的出口，经过别的节点读它的出口原样保留并写明原因
- R4 配了 output_schema 的智能体 → 打开 cite_fields
- R5 代码节点的产出被口径卡读 → 不改，只建议标成取数，等人确认
- 一个叙述里混了几个写作者的，不猜，整步不做并写明原因
- POST /api/copilot/upgrade-evidence：返回改动、说明和改完的检查结果；assist 时再交给 Copilot 把纯算术的代码节点
改写成口径卡表达式，不是纯算术的保留并给警告
- 升级和发布前修复走同一道「不许降低要求」的检查，只多两个由升级点名的例外：模型调用换成报告撰写、R3 改接的那条连线；
改完不能更糟
- 校验对符合 R1–R4 的旧结构给一条建议「可以升级为可追溯结构」，不算提示、不挡发布
- Copilot 自查：报告撰写节点上游没有任何能引用数字的来源时打回
二、界面
- 问题面板的快速修复「升级为可追溯结构（预览改动）」：逐项写「节点 · 字段：原值 → 新值」，节点类型的变化写成
「模型调用 → 报告撰写」，新插的节点和改接的连线另外写明；「应用」后走现有的保存，全图默认不丢
- 画布改过、存失败、画布被锁都有明确状态；保存失败后手动存成功会清掉「没存上」的提示，
画布和预览对不上时「重试保存」不再盲存；校验接口失败时升级块照样在
- 记录页里没有逐段证据的运行给横幅「这次的报告没有逐段证据 → 升级这张图」，跳到编排页并打开升级预览
- 修掉一处旧缺陷：应用修复时只换配置、不换节点类型，模型调用换成报告撰写后存下去的还是模型调用
三、受管门禁：cells 一键选
- 图里没有口径卡、契约又没写 cells: true 的受管图，以前能过门禁、正式运行却每次失败；现在门禁拦下，
发布弹窗给两个选项：写 cells: true，或交给 Copilot 改成口径卡指标（人亲手选的不算替人打开，Copilot 自己打开照样拦）
- 「修完要更好」的复核不再把改完才冒出来、只能由人拍板的这一处算作改坏：Copilot 补上契约、把 cells 留给人问，
这一步会被采用，不再被判「冒出新的问题」作废
四、测试
- 开发依赖加上 pytest-xdist（锁文件只多了 pytest-xdist 和 execnet）：全量后端测试 -n auto 约 16 秒，原来串行约 2 分钟
- 修掉并行下暴露的一处测试竞态：运行先落终态、再另一个事务封存，测试改成等封存完成再核对
680 lines
38 KiB
Python
680 lines
38 KiB
Python
"""一键升级为可追溯结构（证据五期）。
|
||
|
||
旧图里给人看的报告是模型调用（llm）写的自由文本，问数据的答案是 agent 的原话：数字是模型转写的，
|
||
出处只能在出口按数值回头猜。这里按确定的规则把它们改写成可追溯的结构，只给预览，人确认了才保存：
|
||
|
||
- R1 出具契约的 narrative 指向一个模型调用节点：把它换成报告撰写（system 加 prompt 合成 instructions，
|
||
模型配置照搬），契约改成 report_from
|
||
- R2 没有契约，模型调用写的文字进了出口，上游又有取数（查库、口径卡）：同样换成报告撰写
|
||
- R3 input → agent → output（问数据页的典型图）：在 agent 和出口之间插一个报告撰写，出口改取它的正文
|
||
- R4 配了 output_schema 的 agent：打开 cite_fields（不开的话 Schema 不生效）
|
||
- R5 沙箱代码的产出喂给口径卡：不改，只在 notes 里建议标成 source——它在取数还是在算只有作者知道
|
||
|
||
纯函数、幂等：同一张图升级两次，第二次什么都不改。每一步都在副本上做，过一遍和发布前修复同一套禁止
|
||
规则（autofix.forbidden_changes，只多「模型调用换成报告撰写」「R3 改接的那条连线」两个例外），再重跑
|
||
validate 和门禁：冒出新 error 的那一步丢弃，写明原因。和 autofix 一样直接改原始的图字典，坐标、连线 id、
|
||
视口这些前端的东西原样留着。
|
||
|
||
受管出具的契约要不要写 cells（报告直接引用查询单元格）、required 列哪些指标，都是作者的决定：R3 不替人
|
||
加契约，只在 notes 里把两条路说清楚。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import copy
|
||
import re
|
||
from dataclasses import dataclass, field, replace
|
||
from typing import Any, Callable
|
||
|
||
from app.engine.autofix import forbidden_changes, worse
|
||
from app.engine.governance import (
|
||
REPORT_POLICY,
|
||
REPORT_POLICY_ALSO,
|
||
Ref,
|
||
_card_feeders,
|
||
_exit_refs,
|
||
_Graph,
|
||
_refs,
|
||
publish_issues,
|
||
)
|
||
from app.engine.schema import GraphNode, GraphSpec, NodeType, _ancestors, _feeds_caliber, innermost_loops
|
||
|
||
#: 会改图的规则，按这个顺序应用：先换写字的节点，再插报告，最后开 cite_fields（R3 要先把出口从 agent 的
|
||
#: 原话改接到报告上，R4 打开以后 vars.<assign_to> 就是核对过的对象、不再是文字了）
|
||
RULES = ("R1", "R2", "R3", "R4")
|
||
#: validate 给的那条建议（info）的编号和开头
|
||
HINT_CODE = "evidence.upgrade_available"
|
||
HINT = "可以升级为可追溯结构"
|
||
LEVELS = ("published", "governed")
|
||
|
||
#: 模型调用有、报告撰写用不上的配置。system、prompt 合进 instructions；其余的换成报告撰写后不再生效
|
||
_LLM_ONLY = ("prompt", "system", "skills", "use_history", "output_schema")
|
||
#: 新的报告撰写节点照搬过去的模型配置（R3 从 agent 带过去的只有 provider、model：温度、上限是查数用的）
|
||
_MODEL_KEYS = ("provider", "model")
|
||
#: 报告的指令里照搬了这几类节点的文字，就要提醒：报告撰写自己会收集它们的证据
|
||
_EVIDENCE_TYPES = (NodeType.METRICS, NodeType.TOOL, NodeType.AGENT, NodeType.RETRIEVE, NodeType.CODE,
|
||
NodeType.SUBGRAPH)
|
||
#: 写整段文字的节点：出具契约的 narrative 拼了其中几个的文字，就说不准该换哪一个
|
||
_WRITER_TYPES = (NodeType.LLM, NodeType.AGENT)
|
||
_QUESTION = re.compile(r"(?<![\w.])input\s*\.\s*question(?![\w-])")
|
||
_SPAN = re.compile(r"\{\{(.*?)\}\}", re.S)
|
||
|
||
|
||
@dataclass(frozen=True)
|
||
class Match:
|
||
"""一处符合 R1–R4 的旧结构。blocked 不为空的是看着像、但说不准怎么改的，只进 notes。
|
||
|
||
exits 是这一步要改的出口；skipped 是认出来了、但这一步不碰的出口各自的原因（R3 里要经过别的节点才读到
|
||
agent 原话的出口），这一步照改其余的出口，skipped 进 notes。"""
|
||
|
||
rule: str
|
||
node_id: str
|
||
exits: tuple[str, ...] = ()
|
||
summary: str = ""
|
||
blocked: str | None = None
|
||
skipped: tuple[str, ...] = ()
|
||
|
||
|
||
@dataclass
|
||
class _Step:
|
||
label: str
|
||
changes: list[dict[str, Any]] = field(default_factory=list)
|
||
ops: list[dict[str, Any]] = field(default_factory=list)
|
||
notes: list[dict[str, Any]] = field(default_factory=list)
|
||
retyped: set[tuple[str, str, str]] = field(default_factory=set)
|
||
rewired: set[tuple[str, str]] = field(default_factory=set)
|
||
created: bool = False
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 认旧结构
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
def _queries(node: GraphNode) -> bool:
|
||
"""这个节点查库：报告撰写能把它查到的结果编成 Q1、Q2… 引用。只认数据源的查询工具，别的工具的
|
||
返回进不了目录。"""
|
||
if node.type == NodeType.TOOL:
|
||
return str(node.config.get("tool") or "").startswith("db_query__")
|
||
if node.type == NodeType.AGENT:
|
||
tools = node.config.get("tools")
|
||
return isinstance(tools, list) and any(str(t).startswith("db_query__") for t in tools)
|
||
return False
|
||
|
||
|
||
def _fetches(node: GraphNode) -> bool:
|
||
"""上游的取数：查库、口径卡——报告撰写的目录里就有能当值引用的数。知识库检索不算：文档里的话只能当引文，
|
||
知识库问答的答案换成报告撰写，里面转述的数全会被判成没有出处。"""
|
||
return node.type == NodeType.METRICS or _queries(node)
|
||
|
||
|
||
def _cited(node: GraphNode) -> bool:
|
||
return bool(node.config.get("output_schema")) and node.config.get("cite_fields") is True
|
||
|
||
|
||
def _text_ref(node: GraphNode, ref: Ref) -> bool:
|
||
"""顺着这个引用拿到的是不是它写的整段文字：nodes.X、nodes.X.text、vars.<assign_to>、最后一条消息。
|
||
开了 cite_fields 的 agent，assign_to 拿到的是核对过的字段，那是数据不是原话。"""
|
||
root, _, path = ref
|
||
if root in ("last_message", "messages"):
|
||
return True
|
||
if root == "nodes":
|
||
return path in ((), ("text",))
|
||
if root == "vars":
|
||
return not path and not (node.type == NodeType.AGENT and _cited(node))
|
||
return False
|
||
|
||
|
||
def _writers(graph: _Graph, consumer: str, refs: list[Ref], types: tuple[NodeType, ...]) -> list[GraphNode]:
|
||
"""这些引用读到了哪些节点写的整段文字，按画布上的节点顺序。"""
|
||
found: set[str] = set()
|
||
for ref in refs:
|
||
found |= {src.id for src in graph.sources(consumer, ref) if src.type in types and _text_ref(src, ref)}
|
||
return [n for n in graph.spec.nodes if n.id in found]
|
||
|
||
|
||
def _exit_text_refs(exit_node: GraphNode) -> list[Ref]:
|
||
return [ref for _, _, refs in _exit_refs(exit_node) for ref in refs]
|
||
|
||
|
||
def _contract(node: GraphNode) -> dict[str, Any]:
|
||
contract = node.config.get("contract")
|
||
return contract if isinstance(contract, dict) else {}
|
||
|
||
|
||
def _names(nodes: list[GraphNode]) -> str:
|
||
return "、".join(f"「{n.title}」" for n in nodes)
|
||
|
||
|
||
def _llm_blocked(spec: GraphSpec, node: GraphNode, loops: dict[str, str]) -> str | None:
|
||
"""这个模型调用节点能不能原样换成报告撰写；不能的说清为什么。"""
|
||
who = f"「{node.title}」（模型调用）"
|
||
if node.config.get("output_schema"):
|
||
return f"{who}配了 output_schema，交的是结构化数据、不是给人看的文字，没有换成报告撰写"
|
||
if node.id in loops:
|
||
return f"{who}在循环体里，换成报告撰写就是每一轮写一份报告：循环怎么改要你来定，升级没有动它"
|
||
if _feeds_caliber(node, spec):
|
||
return (f"{who}写的数还喂给了口径卡：换成报告撰写，口径卡就读不到它了。先把取数改由 Agent（配 output_schema、"
|
||
"开 cite_fields）或「调用工具」节点来做，再升级")
|
||
return None
|
||
|
||
|
||
def _add(found: dict[tuple[str, str], Match], match: Match) -> None:
|
||
"""同一个节点从几个出口认出来的合成一处：出口、跳过的原因并起来；任何一个出口说不准，整处都不改。
|
||
结果不能看出口在画布上谁排前面。"""
|
||
key = (match.rule, match.node_id)
|
||
was = found.get(key)
|
||
if was is None:
|
||
found[key] = match
|
||
return
|
||
found[key] = Match(was.rule, was.node_id, tuple(dict.fromkeys([*was.exits, *match.exits])),
|
||
was.summary or match.summary, was.blocked or match.blocked,
|
||
tuple(dict.fromkeys([*was.skipped, *match.skipped])))
|
||
|
||
|
||
def upgrade_candidates(spec: GraphSpec) -> list[Match]:
|
||
"""图上符合 R1–R4 的旧结构，按规则、再按节点在画布上的顺序。只看结构，不改图、不跑门禁：
|
||
validate 每改一下画布就调一次，要便宜；真改的时候 upgrade_for_evidence 再逐步复核。"""
|
||
graph = _Graph(spec)
|
||
nodes = spec.node_map()
|
||
loops = innermost_loops(spec)
|
||
order = {n.id: i for i, n in enumerate(spec.nodes)}
|
||
found: dict[tuple[str, str], Match] = {}
|
||
outputs = [n for n in spec.nodes if n.type == NodeType.OUTPUT]
|
||
|
||
for exit_node in outputs:
|
||
contract = _contract(exit_node)
|
||
narrative = contract.get("narrative")
|
||
if contract.get("report_from") not in (None, "") or not isinstance(narrative, str) or not narrative.strip():
|
||
continue
|
||
writers = _writers(graph, exit_node.id, _refs(narrative, template=True), _WRITER_TYPES)
|
||
llms = [w for w in writers if w.type == NodeType.LLM]
|
||
if len(writers) > 1:
|
||
# 叙述拼了几个节点写的文字：换哪一个都会让其余的文字悄悄离开契约的核对。拼进来的每个模型调用都不换，
|
||
# 哪怕它另有一个只读它的出口——不然结果要看两个出口在画布上谁排前面
|
||
why = (f"「{exit_node.title}」的出具契约 narrative 拼了{_names(writers)}几个节点写的文字，该换哪一个说不准，"
|
||
"没有升级：让一个节点写完整篇（其余的内容交给它当素材），再升级")
|
||
for llm in llms:
|
||
_add(found, Match("R1", llm.id, blocked=why))
|
||
elif llms:
|
||
llm = llms[0]
|
||
_add(found, Match("R1", llm.id, (exit_node.id,),
|
||
f"「{llm.title}」（模型调用）写的叙述换成报告撰写，出具契约改为核对它的文档",
|
||
_llm_blocked(spec, llm, loops)))
|
||
|
||
r1 = {m.node_id for m in found.values()}
|
||
for exit_node in outputs:
|
||
if _contract(exit_node):
|
||
continue # 有契约的归 R1（叙述模式）或者已经是引用模式
|
||
for llm in _writers(graph, exit_node.id, _exit_text_refs(exit_node), (NodeType.LLM,)):
|
||
if llm.id in r1 or not any(_fetches(nodes[a]) for a in _ancestors(spec, llm.id) if a in nodes):
|
||
continue
|
||
_add(found, Match("R2", llm.id, (exit_node.id,),
|
||
f"「{llm.title}」（模型调用）写的文字直接进了出口，换成报告撰写",
|
||
_llm_blocked(spec, llm, loops)))
|
||
|
||
for exit_node in outputs:
|
||
contract = _contract(exit_node)
|
||
if contract.get("report_from") not in (None, ""):
|
||
continue
|
||
refs = _exit_text_refs(exit_node)
|
||
narrated: list[GraphNode] = []
|
||
if isinstance(contract.get("narrative"), str):
|
||
story = _refs(contract["narrative"], template=True)
|
||
refs += story
|
||
narrated = _writers(graph, exit_node.id, story, _WRITER_TYPES)
|
||
for agent in _writers(graph, exit_node.id, refs, (NodeType.AGENT,)):
|
||
# 喂口径卡的 agent 是取数的，不是作答的：它的原话漏进出口由门禁 G2 管，不在这里插报告
|
||
if not _queries(agent) or _feeds_caliber(agent, spec):
|
||
continue
|
||
# 说不准的只跳过这个出口：同一个 agent 直连的出口照样插报告撰写，结果不看出口在画布上的顺序
|
||
if agent.id in loops:
|
||
_add(found, Match("R3", agent.id, blocked=(
|
||
f"「{agent.title}」在循环体里，升级不替你改循环结构，没有插入报告撰写")))
|
||
elif not any(e.source == agent.id and e.target == exit_node.id for e in spec.edges):
|
||
_add(found, Match("R3", agent.id, skipped=(
|
||
f"「{agent.title}」的原话要经过别的节点才到「{exit_node.title}」，升级不替你改接中间的连线：在它后面"
|
||
f"接一个报告撰写，「{exit_node.title}」改取报告的正文",)))
|
||
elif agent.id in {n.id for n in narrated} and len(narrated) > 1:
|
||
_add(found, Match("R3", agent.id, skipped=(
|
||
f"「{exit_node.title}」的出具契约 narrative 拼了{_names(narrated)}几个节点写的文字，改成核对报告撰写的"
|
||
"文档，其余的文字就悄悄离开了核对：这个出口升级没有动。让一个节点写完整篇，再升级",)))
|
||
else:
|
||
_add(found, Match("R3", agent.id, (exit_node.id,)))
|
||
|
||
for key, m in found.items():
|
||
if m.rule != "R3":
|
||
continue
|
||
agent = nodes[m.node_id]
|
||
if m.exits:
|
||
found[key] = replace(m, summary=f"在「{agent.title}」和{_names([nodes[x] for x in m.exits])}之间插入报告撰写，"
|
||
"答案里的数逐个点得开出处")
|
||
elif not m.blocked:
|
||
found[key] = replace(m, blocked="；".join(m.skipped), skipped=())
|
||
|
||
for n in spec.nodes:
|
||
if n.type == NodeType.AGENT and n.config.get("output_schema") and n.config.get("cite_fields") is not True:
|
||
_add(found, Match("R4", n.id, (), f"打开「{n.title}」的 cite_fields，output_schema 才生效"))
|
||
|
||
return sorted(found.values(), key=lambda m: (RULES.index(m.rule), order.get(m.node_id, len(order))))
|
||
|
||
|
||
def upgrade_hint(spec: GraphSpec) -> str | None:
|
||
"""validate 的那条建议的文案；没有能升级的旧结构返回 None。
|
||
|
||
升级是整张图一起改的，建议挂在图上、不挂在哪个节点上：挂上去就混进了那个节点自己的问题清单。
|
||
要改哪几个节点，预览里逐项列着。"""
|
||
ready = [m for m in upgrade_candidates(spec) if not m.blocked]
|
||
if not ready:
|
||
return None
|
||
parts = [m.summary for m in ready[:3]] + ([f"另有 {len(ready) - 3} 处"] if len(ready) > 3 else [])
|
||
return f"{HINT}：{'；'.join(parts)}。先预览改动，确认了才保存"
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 改图（原地改副本）
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
def _raw(graph: dict[str, Any], node_id: str) -> dict[str, Any]:
|
||
return next(n for n in graph.get("nodes") or [] if isinstance(n, dict) and n.get("id") == node_id)
|
||
|
||
|
||
def _config(node: dict[str, Any]) -> dict[str, Any]:
|
||
data = node.setdefault("data", {})
|
||
if not isinstance(data.get("config"), dict):
|
||
data["config"] = {}
|
||
return data["config"]
|
||
|
||
|
||
def _note(text: str, rule: str, node_id: str | None, level: str = "info") -> dict[str, Any]:
|
||
return {"text": text, "rule": rule, "node_id": node_id, "level": level}
|
||
|
||
|
||
def _change(node_id: str | None, title: str, field_name: str, before: Any, after: Any) -> dict[str, Any]:
|
||
return {"node_id": node_id, "node_title": title, "field": field_name, "before": copy.deepcopy(before),
|
||
"after": copy.deepcopy(after)}
|
||
|
||
|
||
def _config_changes(node_id: str, title: str, before: dict[str, Any], after: dict[str, Any]) -> list[dict[str, Any]]:
|
||
return [_change(node_id, title, key, before.get(key), after.get(key))
|
||
for key in dict.fromkeys([*after, *before]) if before.get(key) != after.get(key)]
|
||
|
||
|
||
def _policy(level: str, config: dict[str, Any], defaults: dict[str, Any]) -> dict[str, Any]:
|
||
"""受管级别下新写出来的报告撰写节点，照门禁 G3 写明最严的三项：已经（按节点、再按全图默认）是这个
|
||
值、或者更严的（claims: judge）不动。已发布级别不替探索运行拧紧，照缺省走。"""
|
||
if level != "governed":
|
||
return {}
|
||
out = {}
|
||
for key, want, _ in REPORT_POLICY:
|
||
value = config.get(key)
|
||
effective = defaults.get(key) if value in (None, "") else value
|
||
if effective not in (want, *REPORT_POLICY_ALSO.get(key, ())):
|
||
out[key] = want
|
||
return out
|
||
|
||
|
||
def _report_from(contract: dict[str, Any], node_id: str) -> dict[str, Any]:
|
||
"""narrative 换成 report_from，放在 narrative 原来的位置，其余的键原样留着。原来写着的空 report_from（等于没声明，
|
||
upgrade_candidates 也这么认）丢掉：它排在 narrative 后面的话，逐键照抄会把新值盖成空的，契约的核对就悄悄关了。"""
|
||
out: dict[str, Any] = {}
|
||
for key, value in contract.items():
|
||
if key == "narrative":
|
||
out["report_from"] = node_id
|
||
elif key != "report_from":
|
||
out[key] = copy.deepcopy(value)
|
||
out.setdefault("report_from", node_id)
|
||
return out
|
||
|
||
|
||
def _span_refs(text: str) -> list[tuple[str, list[Ref]]]:
|
||
return [(m.group(0), _refs(m.group(1), template=False)) for m in _SPAN.finditer(text)]
|
||
|
||
|
||
def _echoed_evidence(graph: _Graph, consumer: str, text: str) -> list[str]:
|
||
"""指令里照搬了哪几段上游取数、口径卡的文字（原样的 {{ … }}）。"""
|
||
out = []
|
||
for span, refs in _span_refs(text):
|
||
if any(src.type in _EVIDENCE_TYPES for ref in refs for src in graph.sources(consumer, ref)):
|
||
out.append(span.strip())
|
||
return list(dict.fromkeys(out))
|
||
|
||
|
||
def _retype(graph: dict[str, Any], spec: GraphSpec, match: Match, level: str) -> _Step:
|
||
"""R1 / R2：模型调用换成报告撰写。R1 另把契约的 narrative 换成 report_from。"""
|
||
node = spec.node_map()[match.node_id]
|
||
raw = _raw(graph, node.id)
|
||
before = copy.deepcopy(_config(raw))
|
||
text = "\n\n".join(p for p in (str(before.get("system") or "").strip(), str(before.get("prompt") or "").strip()) if p)
|
||
after: dict[str, Any] = {"instructions": text} if text else {}
|
||
after.update({k: copy.deepcopy(v) for k, v in before.items() if k not in _LLM_ONLY})
|
||
after.update(_policy(level, after, spec.defaults))
|
||
raw["type"] = NodeType.REPORT.value
|
||
raw["data"]["config"] = after
|
||
title = node.title
|
||
step = _Step(label=(f"把「{title}」换成报告撰写，出具契约改为核对它的文档" if match.rule == "R1"
|
||
else f"把「{title}」换成报告撰写"),
|
||
retyped={(node.id, NodeType.LLM.value, NodeType.REPORT.value)}, created=True)
|
||
step.changes = [_change(node.id, title, "type", NodeType.LLM.value, NodeType.REPORT.value),
|
||
*_config_changes(node.id, title, before, after)]
|
||
step.ops = [{"op": "replace_node", "id": node.id,
|
||
"node": {"id": node.id, "type": NodeType.REPORT.value, "label": raw["data"].get("label") or "",
|
||
"config": copy.deepcopy(after)}}]
|
||
dropped = [k for k in _LLM_ONLY if k not in ("prompt", "system") and before.get(k) not in (None, "", [], {})]
|
||
if before.get("skills"):
|
||
skills = before["skills"] if isinstance(before["skills"], list) else [before["skills"]]
|
||
step.notes.append(_note(f"「{title}」挂的技能 {'、'.join(f'「{s}」' for s in skills)} 不会再加载：报告撰写不挂技能，"
|
||
"其中的写作要求要保留的话，写进 instructions", match.rule, node.id, "warning"))
|
||
if other := [k for k in dropped if k != "skills"]:
|
||
step.notes.append(_note(f"「{title}」的 {'、'.join(other)} 报告撰写用不上，已去掉", match.rule, node.id))
|
||
if echoed := _echoed_evidence(_Graph(spec), node.id, text):
|
||
step.notes.append(_note(
|
||
f"「{title}」的 instructions 里照搬了 {'、'.join(echoed)}：报告撰写会自己收集上游的口径卡指标和查询结果，数字由"
|
||
"系统渲染；直接塞进指令的这段文字里的数，模型容易照抄成不带出处的裸数字（会被打回重写）。确认用不上就删掉，"
|
||
"升级没有替你删", match.rule, node.id))
|
||
if match.rule == "R1":
|
||
for exit_id in match.exits:
|
||
exit_raw = _raw(graph, exit_id)
|
||
cfg = _config(exit_raw)
|
||
was = copy.deepcopy(cfg.get("contract"))
|
||
cfg["contract"] = _report_from(was or {}, node.id)
|
||
step.changes.append(_change(exit_id, spec.node_map()[exit_id].title, "contract", was, cfg["contract"]))
|
||
step.ops.append({"op": "update_node", "id": exit_id, "config": {"contract": copy.deepcopy(cfg["contract"])}})
|
||
return step
|
||
|
||
|
||
def _fresh_id(graph: dict[str, Any], base: str) -> str:
|
||
taken = {str(n.get("id")) for n in graph.get("nodes") or [] if isinstance(n, dict)}
|
||
if base not in taken:
|
||
return base
|
||
return next(f"{base}_{i}" for i in range(2, len(taken) + 3) if f"{base}_{i}" not in taken)
|
||
|
||
|
||
def _answer_instructions(agent: GraphNode) -> str:
|
||
"""插进来的报告撰写写什么。问数据的图里问题在 input.question；别的图照搬 agent 的要求。agent 的原话附在
|
||
后面当思路（和 Copilot 手册里问数据的写法一致），明说里面的数不能照抄。"""
|
||
prompt = str(agent.config.get("prompt") or "").strip()
|
||
if _QUESTION.search(prompt):
|
||
head = "回答：{{ input.question }}"
|
||
elif prompt:
|
||
head = f"按下面的要求写给读的人看的回答（查数已经由上游「{agent.title}」做完，这里只写）：\n{prompt}"
|
||
else:
|
||
head = "根据证据目录里的查询结果，写一份简洁的回答"
|
||
return (f"{head}\n\n「{agent.title}」查数时的分析过程附在下面，只作思路参考：里面的数字不能照抄，一律用引用标记"
|
||
f"从证据目录里取。\n{{{{ nodes.{agent.id}.text }}}}")
|
||
|
||
|
||
def _agent_text_head(agent: GraphNode) -> re.Pattern[str]:
|
||
"""模板里「这个 agent 写的整段文字」的写法：nodes.X、nodes.X.text、vars.<assign_to>（没开 cite_fields 时）。"""
|
||
nid = re.escape(agent.id)
|
||
heads = [rf"nodes\s*(?:\.\s*{nid}|\[\s*['\"]{nid}['\"]\s*\])(?:\s*(?:\.\s*text|\[\s*['\"]text['\"]\s*\]))?"]
|
||
var = str(agent.config.get("assign_to") or "").strip()
|
||
if var and not _cited(agent):
|
||
heads.append(rf"vars\s*(?:\.\s*{re.escape(var)}|\[\s*['\"]{re.escape(var)}['\"]\s*\])")
|
||
return re.compile("|".join(f"(?:{h})" for h in heads))
|
||
|
||
|
||
def _swap_text(value: str, agent: GraphNode, report_id: str) -> str:
|
||
head_re = _agent_text_head(agent)
|
||
|
||
def repl(m: re.Match[str]) -> str:
|
||
head, bar, rest = m.group(1).partition("|")
|
||
if not head_re.fullmatch(head.strip()):
|
||
return m.group(0)
|
||
return f"{{{{ nodes.{report_id}.text{(' |' + rest.rstrip()) if bar else ''} }}}}"
|
||
|
||
return _SPAN.sub(repl, value)
|
||
|
||
|
||
def _insert(graph: dict[str, Any], spec: GraphSpec, match: Match, level: str) -> _Step:
|
||
"""R3：agent 和出口之间插一个报告撰写。agent → 出口的连线改接成 agent → 报告 → 出口，出口字段里 agent 的
|
||
原话改取报告的正文；没配字段的出口交的是最后一条消息，插了报告以后就是报告写的，不用改。"""
|
||
nodes = spec.node_map()
|
||
agent = nodes[match.node_id]
|
||
raw_agent = _raw(graph, agent.id)
|
||
rid = _fresh_id(graph, "report")
|
||
config: dict[str, Any] = {"instructions": _answer_instructions(agent)}
|
||
config.update({k: copy.deepcopy(agent.config[k]) for k in _MODEL_KEYS if agent.config.get(k) not in (None, "")})
|
||
config.update(_policy(level, config, spec.defaults))
|
||
here = raw_agent.get("position") or {}
|
||
there = _raw(graph, match.exits[0]).get("position") or {}
|
||
position = {"x": round((float(here.get("x") or 0) + float(there.get("x") or 0)) / 2, 1),
|
||
"y": round((float(here.get("y") or 0) + float(there.get("y") or 0)) / 2 + 120, 1)}
|
||
new = {"id": rid, "type": NodeType.REPORT.value, "position": position,
|
||
"data": {"label": "报告撰写", "config": config}}
|
||
graph["nodes"].insert(graph["nodes"].index(raw_agent) + 1, new)
|
||
exits = [nodes[x] for x in match.exits]
|
||
step = _Step(label=f"在「{agent.title}」和{_names(exits)}之间插入报告撰写，出口改取报告的正文",
|
||
rewired={(agent.id, x.id) for x in exits}, created=True)
|
||
step.changes.append(_change(rid, "报告撰写", "node", None,
|
||
{"id": rid, "type": NodeType.REPORT.value, "label": "报告撰写", "config": config}))
|
||
step.ops.append({"op": "add_node", "node": {"id": rid, "type": NodeType.REPORT.value, "label": "报告撰写",
|
||
"config": copy.deepcopy(config)}})
|
||
|
||
def link(source: str, target: str) -> dict[str, Any]:
|
||
return {"id": f"{source}:->{target}", "source": source, "target": target}
|
||
|
||
converted: list[dict[str, Any]] = [] # 这一步从叙述模式改成 report_from 的契约
|
||
graph["edges"] = [e for e in graph.get("edges") or []
|
||
if not (isinstance(e, dict) and e.get("source") == agent.id and e.get("target") in match.exits)]
|
||
graph["edges"].append(link(agent.id, rid))
|
||
step.changes.append(_change(None, "", "edge", None, {"source": agent.id, "target": rid}))
|
||
step.ops.append({"op": "add_edge", "edge": {"source": agent.id, "target": rid}})
|
||
for x in exits:
|
||
graph["edges"].append(link(rid, x.id))
|
||
step.changes.append(_change(None, "", "edge", {"source": agent.id, "target": x.id},
|
||
{"source": rid, "target": x.id}))
|
||
step.ops += [{"op": "remove_edge", "source": agent.id, "target": x.id},
|
||
{"op": "add_edge", "edge": {"source": rid, "target": x.id}}]
|
||
cfg = _config(_raw(graph, x.id))
|
||
updates: dict[str, Any] = {}
|
||
fields = cfg.get("fields")
|
||
if isinstance(fields, list):
|
||
swapped = [{**f, "value": _swap_text(f["value"], agent, rid)}
|
||
if isinstance(f, dict) and isinstance(f.get("value"), str) else f for f in fields]
|
||
if swapped != fields:
|
||
step.changes.append(_change(x.id, x.title, "fields", fields, swapped))
|
||
updates["fields"] = swapped
|
||
cfg["fields"] = swapped
|
||
exact = f"{{{{ nodes.{rid}.text }}}}"
|
||
mixed = [str(f.get("name")) for f, old in zip(swapped, fields)
|
||
if isinstance(f, dict) and f != old and f.get("value", "").strip() != exact]
|
||
if mixed:
|
||
step.notes.append(_note(
|
||
f"「{x.title}」的成果字段{'、'.join(f'「{n}」' for n in mixed)}在报告正文前后还拼了别的内容：这样没法"
|
||
f"逐段对应证据。要每个数都点得开，字段里只放 {exact}", "R3", x.id))
|
||
contract = cfg.get("contract")
|
||
narrative = contract.get("narrative") if isinstance(contract, dict) else None
|
||
if isinstance(narrative, str) and agent.id in {
|
||
n.id for n in _writers(_Graph(spec), x.id, _refs(narrative, template=True), (NodeType.AGENT,))}:
|
||
was = copy.deepcopy(contract)
|
||
cfg["contract"] = _report_from(contract, rid)
|
||
converted.append(cfg["contract"])
|
||
step.changes.append(_change(x.id, x.title, "contract", was, cfg["contract"]))
|
||
updates["contract"] = cfg["contract"]
|
||
if updates:
|
||
step.ops.append({"op": "update_node", "id": x.id, "config": copy.deepcopy(updates)})
|
||
if note := _cells_note(spec, exits, converted, level):
|
||
step.notes.append(_note(note[0], "R3", rid, note[1]))
|
||
return step
|
||
|
||
|
||
def _cells_note(spec: GraphSpec, exits: list[GraphNode], converted: list[dict[str, Any]],
|
||
level: str) -> tuple[str, str] | None:
|
||
"""插进来的报告撰写会直接引用查询单元格，受管级别的正式运行只在出口契约写了 cells: true 时才认
|
||
（governance.cells_allowed）。已经写了的不提；按受管级别升级的，或者把原有契约改成了 report_from、图上又没有
|
||
口径卡可走的，说成警告：不写的话报告里的数一个都引用不了，报告节点每次正式运行都会失败。"""
|
||
if converted and all(c.get("cells") is True for c in converted):
|
||
return None
|
||
where = _names(exits)
|
||
choose = (f"二选一：在{where}的出具契约里写 cells: true；或者改走口径卡，把要写的数登记成指标，报告只引用 "
|
||
"[[m:指标id]]。选哪条要你拿主意，升级没有替你写")
|
||
cards = any(n.type == NodeType.METRICS for n in spec.nodes)
|
||
if level == "governed" or (converted and not cards):
|
||
lead = "按受管级别" if level == "governed" else "现在（已发布级别）照常；以后按受管级别发布的话"
|
||
return (f"报告会直接引用查询单元格（[[v:Q1.r0.列]]），{lead}，正式运行只在出口的出具契约写了 cells: true 时才认单元格"
|
||
"引用：不写的话报告里的数一个都引用不了，报告节点（numbers: strict、on_violation: fail）每次正式运行都会失败。"
|
||
+ choose, "warning")
|
||
return ("报告会直接引用查询单元格（[[v:Q1.r0.列]]）。探索运行和已发布级别照常；要按受管级别正式出具，得在出口的出具契约里写 "
|
||
"cells: true，或者改走口径卡：把要写的数登记成指标，报告只引用 [[m:指标id]]。选哪条要你拿主意，升级没有替你加契约",
|
||
"info")
|
||
|
||
|
||
def _whole_var_readers(spec: GraphSpec, var: str) -> list[tuple[GraphNode, str]]:
|
||
"""模板里把 vars.<var> 整个拿来用的地方：(节点, 原样的 {{ … }})。取字段的（vars.kpi.gmv）不算。"""
|
||
head_re = re.compile(rf"vars\s*(?:\.\s*{re.escape(var)}|\[\s*['\"]{re.escape(var)}['\"]\s*\])")
|
||
|
||
def strings(value: Any) -> list[str]:
|
||
if isinstance(value, str):
|
||
return [value]
|
||
if isinstance(value, dict):
|
||
return [s for v in value.values() for s in strings(v)]
|
||
if isinstance(value, list):
|
||
return [s for v in value for s in strings(v)]
|
||
return []
|
||
|
||
out: dict[tuple[str, str], tuple[GraphNode, str]] = {}
|
||
for n in spec.nodes:
|
||
for text in strings(n.config):
|
||
for m in _SPAN.finditer(text):
|
||
if head_re.fullmatch(m.group(1).partition("|")[0].strip()):
|
||
out.setdefault((n.id, m.group(0).strip()), (n, m.group(0).strip()))
|
||
return list(out.values())
|
||
|
||
|
||
def _cite(graph: dict[str, Any], spec: GraphSpec, match: Match, level: str) -> _Step:
|
||
"""R4：打开 cite_fields。打开以后 assign_to 拿到的是按 Schema 核对过的对象，不再是文字：把它整个当文字
|
||
用的地方说出来，不替人改。"""
|
||
agent = spec.node_map()[match.node_id]
|
||
cfg = _config(_raw(graph, agent.id))
|
||
before = cfg.get("cite_fields")
|
||
cfg["cite_fields"] = True
|
||
step = _Step(label=f"打开「{agent.title}」的 cite_fields：每个字段都核对到查询结果里的那一格")
|
||
step.changes = [_change(agent.id, agent.title, "cite_fields", before, True)]
|
||
step.ops = [{"op": "update_node", "id": agent.id, "config": {"cite_fields": True}}]
|
||
var = str(agent.config.get("assign_to") or "").strip()
|
||
for reader, span in _whole_var_readers(spec, var) if var else []:
|
||
step.notes.append(_note(
|
||
f"「{agent.title}」开了 cite_fields 以后，vars.{var} 拿到的是核对过的对象，不再是文字：「{reader.title}」里的 "
|
||
f"{span} 会显示成对象。要原话就改成 {{{{ nodes.{agent.id}.text }}}}", "R4", reader.id, "warning"))
|
||
return step
|
||
|
||
|
||
_APPLY: dict[str, Callable[[dict[str, Any], GraphSpec, Match, str], _Step]] = {
|
||
"R1": _retype, "R2": _retype, "R3": _insert, "R4": _cite}
|
||
|
||
|
||
def compute_feeders(spec: GraphSpec) -> list[tuple[GraphNode, list[GraphNode]]]:
|
||
"""喂口径卡、角色又不是取数（evidence_role 不是 source）的沙箱代码：(代码节点, 读它的口径卡)。
|
||
R5 只建议；assist 时交给 Copilot 看是不是纯算术。"""
|
||
graph = _Graph(spec)
|
||
found: dict[str, tuple[GraphNode, list[GraphNode]]] = {}
|
||
for card in spec.nodes:
|
||
if card.type != NodeType.METRICS:
|
||
continue
|
||
for feeder in _card_feeders(graph, card):
|
||
if feeder.type == NodeType.CODE and feeder.config.get("evidence_role") != "source":
|
||
found.setdefault(feeder.id, (feeder, []))[1].append(card)
|
||
return list(found.values())
|
||
|
||
|
||
def metric_feeders(spec: GraphSpec, card: GraphNode) -> list[tuple[int, GraphNode]]:
|
||
"""口径卡每个指标追到头读的是谁：(指标下标, 来源节点)。和 compute_feeders 同一条追法（穿过整形、校验节点，
|
||
governance._card_feeders），只是逐个指标追：code → 整形 → 口径卡的，指标记在这段代码名下。"""
|
||
graph = _Graph(spec)
|
||
metrics = card.config.get("metrics")
|
||
out: list[tuple[int, GraphNode]] = []
|
||
for i, metric in enumerate(metrics if isinstance(metrics, list) else []):
|
||
one = card.model_copy(update={"data": card.data.model_copy(update={"config": {**card.config,
|
||
"metrics": [metric]}})})
|
||
out += [(i, feeder) for feeder in _card_feeders(graph, one)]
|
||
return out
|
||
|
||
|
||
def _r5_notes(spec: GraphSpec) -> list[dict[str, Any]]:
|
||
return [_note(
|
||
f"「{code.title}」（沙箱代码）的产出喂给了口径卡{_names(cards)}：它如果是在取数（查库、调接口、读文件），把 "
|
||
"evidence_role 标成 source，口径卡读它时记得下出处；如果是在做计算，把计算挪进口径卡的表达式（交给 Copilot "
|
||
"可以请它把纯算术的代码改写成口径卡表达式）。它在取数还是在算只有你知道，升级没有替你改", "R5", code.id)
|
||
for code, cards in compute_feeders(spec)]
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 入口
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
def forbidden(before: dict[str, Any], after: dict[str, Any], *,
|
||
retyped: set[tuple[str, str, str]] | frozenset[tuple[str, str, str]] = frozenset(),
|
||
rewired: set[tuple[str, str]] | frozenset[tuple[str, str]] = frozenset()) -> list[str]:
|
||
"""升级的禁止规则：和发布前修复同一套（删节点、删连线、删契约、放宽审批、关 strict、改级别…一律拦），
|
||
只放过这一步自己点名的换类型（模型调用 → 报告撰写）和改接的连线。"""
|
||
return forbidden_changes(before, after, retyped=frozenset(retyped), rewired=frozenset(rewired))
|
||
|
||
|
||
def upgrade_for_evidence(graph: dict[str, Any] | GraphSpec, *, level: str = "published") -> dict[str, Any]:
|
||
"""按 R1–R4 改写，R5 只给建议。纯函数，不动传进来的图；同一张图升级两次，第二次什么都不改。
|
||
|
||
level 是复核门禁用的发布级别：受管级别下新写出来的报告撰写节点另写明门禁 G3 要的三项。
|
||
|
||
返回 {graph, changes, ops, notes, applied, rejected}：
|
||
- changes：[{fix_id, rule, label, node_id, node_title, field, before, after}]；field 为 type 是节点换了类型，
|
||
node 是新插入的节点（after 是节点），edge 是连线（before / after 是 {source, target}，改接的两边都有）
|
||
- ops：形状同 Copilot 的操作流，另有 replace_node {id, node}（换类型，node 是整个新节点）
|
||
- notes：[{text, rule, node_id, level}]，不改、只建议的，以及看着像旧结构、但没法确定怎么改的
|
||
- applied：采用的 "R1:<节点 id>"…；rejected：[{fix_id, reason}]，碰了禁止规则或者会冒出新 error 的
|
||
"""
|
||
if level not in LEVELS:
|
||
raise ValueError(f"level 只能是 {' / '.join(LEVELS)}，写的是 {level!r}")
|
||
raw = graph.model_dump(mode="json") if isinstance(graph, GraphSpec) else graph
|
||
current = copy.deepcopy(raw)
|
||
spec = GraphSpec.model_validate(current)
|
||
issues = publish_issues(spec, level=level)
|
||
changes: list[dict[str, Any]] = []
|
||
ops: list[dict[str, Any]] = []
|
||
notes: list[dict[str, Any]] = []
|
||
applied: list[str] = []
|
||
rejected: list[dict[str, Any]] = []
|
||
created: str | None = None # 第一处写出报告撰写节点的规则
|
||
for rule in RULES:
|
||
tried: set[str] = set()
|
||
while True:
|
||
# 每一步之后重新认：前一步可能改了后一步要看的结构（插了报告的出口不再读 agent 的原话）
|
||
match = next((m for m in upgrade_candidates(spec) if m.rule == rule and m.node_id not in tried), None)
|
||
if match is None:
|
||
break
|
||
tried.add(match.node_id)
|
||
if match.blocked:
|
||
# 一个说不准的叙述会挡下拼进去的每个模型调用，原因只说一遍
|
||
if not any(n["text"] == match.blocked for n in notes):
|
||
notes.append(_note(match.blocked, rule, match.node_id, "warning"))
|
||
continue
|
||
notes += [_note(why, rule, match.node_id, "warning") for why in match.skipped]
|
||
fix_id = f"{rule}:{match.node_id}"
|
||
trial = copy.deepcopy(current)
|
||
step = _APPLY[rule](trial, spec, match, level)
|
||
try:
|
||
trial_spec = GraphSpec.model_validate(trial)
|
||
except Exception: # noqa: BLE001 - 规则把图改坏了：丢弃，不交给人一张存不回去的图
|
||
rejected.append({"fix_id": fix_id, "reason": "改完的图结构读不懂，这一步没有采用"})
|
||
continue
|
||
trial_issues = publish_issues(trial_spec, level=level)
|
||
why = "；".join(forbidden(current, trial, retyped=step.retyped, rewired=step.rewired)) \
|
||
or worse(issues, trial_issues)
|
||
if why:
|
||
rejected.append({"fix_id": fix_id, "reason": why})
|
||
notes.append(_note(f"{step.label}：没有采用，{why}", rule, match.node_id, "warning"))
|
||
continue
|
||
current, spec, issues = trial, trial_spec, trial_issues
|
||
applied.append(fix_id)
|
||
changes += [{"fix_id": fix_id, "rule": rule, "label": step.label, **c} for c in step.changes]
|
||
ops += step.ops
|
||
notes += step.notes
|
||
created = created or (rule if step.created else None)
|
||
notes += _r5_notes(spec)
|
||
if created and level != "governed":
|
||
notes.append(_note(
|
||
"报告撰写重写之后仍有没出处的数字时：探索运行照常产出、把问题标在报告里；正式运行里 on_violation 缺省是 fail，节点"
|
||
"会失败。想先标注着看，把 on_violation 设成 flag", created, None))
|
||
return {"graph": current, "changes": changes, "ops": ops, "notes": notes, "applied": applied, "rejected": rejected}
|