mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 19:01:58 +00:00
ThearchyHelios
392 lines
20 KiB
Python
392 lines
20 KiB
Python
"""蓝图和关节的生成（N2、N3、N11）：接到 api/copilot_gen.Generation 上的两对落地与自查。
|
||
|
||
- BlueprintApplier / BlueprintChecker：蓝图助手只收 bp_* 操作，落地走 blueprints/ops.apply_bp_op（来源记 assistant，
|
||
只提议：助手写的项都是「待确认」）。复用口径卡的操作按这一轮读好的口径核对并补上它读的列（look/calibers）；
|
||
越界被拒的照实交回模型一次、交付时列给人看（同 AppApplier 的做法）。
|
||
- TemplateApplier / TemplateChecker：填关节只收 fill_joint {joint, value}，而且只收这一组里模板声明过的生成关节；
|
||
改别的位置、别的组的关节、或者用别的操作，一律拒绝并照实交回（照 apps/ops.polish_problem 的写法）。自查跑 GraphChecker
|
||
的 SQL 检查（只读守卫、猜列、数据源范围、目录检查）、指标式的白名单（只许 cell() 取格加算术）和 engine/lineage.static_stops
|
||
（静态追到原始数据），再和骨架逐节点比对（tpl.skeleton_changed）。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import copy
|
||
import json
|
||
import re
|
||
from typing import Any
|
||
|
||
from app.blueprints.ops import BpRefused, apply_bp_op
|
||
from app.blueprints.spec import BP_OPS
|
||
from app.look.base import Applied
|
||
|
||
#: 越界被拒的编号（交回模型、列给人看）
|
||
BP_REFUSED = "bp.refused"
|
||
JOINT_REFUSED = "tpl.joint_refused"
|
||
JOINT_UNFILLED = "tpl.joint_unfilled"
|
||
METRIC_EXPR = "tpl.metric_expr"
|
||
METRIC_COLUMN = "tpl.metric_column"
|
||
SQL_REJECTED = "tpl.sql_rejected"
|
||
#: 指标式里能用的函数：取格和算术（叙述层不做算术，算术只在口径卡里，P7）
|
||
METRIC_FUNCS = frozenset({"cell", "round", "abs", "min", "max"})
|
||
|
||
BLUEPRINT_REPAIR_PROMPT = (
|
||
"这是你刚改好的蓝图：\n{spec}\n\n检查发现下面这些问题：\n{issues}\n\n"
|
||
"只修这些问题，别的不要动。用 bp_* 操作输出修正；拿不准、要用户拿主意的用 bp_ask 问。最后一行输出 done。{looks}")
|
||
TEMPLATE_REPAIR_PROMPT = (
|
||
"这是你刚填的关节：\n{values}\n\n检查（和发布同一套规则）发现下面这些问题：\n{issues}\n\n"
|
||
"只修这些问题：用 fill_joint 重新填有问题的关节（整段重写），别的不要动。最后一行输出 done。{looks}")
|
||
REPAIR_LOOKS_PROMPT = "\n\n这一轮查看过的内容（和起草时看到的是同一份）：\n{text}"
|
||
ISSUE_LINE_PROMPT = "- {where}{message}"
|
||
|
||
|
||
def _issue_lines(issues: list[dict[str, Any]]) -> str:
|
||
return "\n".join(ISSUE_LINE_PROMPT.format(where=f"{i.get('path') or i.get('joint') or i.get('node_id') or ''} ",
|
||
message=i.get("for_model") or i.get("message") or "").replace(" ", " ")
|
||
for i in issues)
|
||
|
||
|
||
# ==========================================================================
|
||
# 蓝图
|
||
# ==========================================================================
|
||
|
||
|
||
class BlueprintApplier:
|
||
"""蓝图的落地：bp_* 操作按 ops.apply_bp_op 落到 spec 上（每落一个换一份新的），来源记 assistant。
|
||
|
||
plan、done 记下说明；reply 记下（这句话不需要改蓝图）；bp_ask 记进反问（Generation 读到它就结束这一轮）；别的操作一概
|
||
不收。calibers 是这一轮读好的口径（look/calibers.CaliberKind）：复用口径卡的操作按它核对、补上读的列。"""
|
||
|
||
def __init__(self, spec: dict[str, Any], *, calibers: Any = None) -> None:
|
||
self.spec = spec
|
||
self.calibers = calibers
|
||
self.effective: list[dict[str, Any]] = []
|
||
self.plan: dict[str, Any] | None = None
|
||
self.replied: dict[str, Any] | None = None
|
||
self.explanation = ""
|
||
#: 越界被拒的操作：[{op, path, message, code}]
|
||
self.refused: list[dict[str, Any]] = []
|
||
self._reported = 0
|
||
self._landed: list[str] = []
|
||
self._continuing = False
|
||
self._repairing = False
|
||
|
||
def apply(self, op: dict[str, Any]) -> Applied:
|
||
kind = str(op.get("op") or "")
|
||
if kind == "plan":
|
||
if self.plan is not None and self._continuing:
|
||
return Applied(False, False)
|
||
self.plan = self.plan or op
|
||
return Applied(False, True)
|
||
if kind == "done":
|
||
self.explanation = str(op.get("explanation", ""))
|
||
return Applied(False, True)
|
||
if kind == "reply":
|
||
if self.replied is None and not self._repairing:
|
||
self.replied = op
|
||
return Applied(False, False)
|
||
if kind not in BP_OPS:
|
||
return Applied(False, False)
|
||
if self._continuing and op in self.effective:
|
||
return Applied(False, False)
|
||
op = {k: v for k, v in op.items() if k in ("op", "path", "value", "question", "options", "template")}
|
||
try:
|
||
op = self._with_reuse(op)
|
||
self.spec = apply_bp_op(self.spec, op, source="assistant")
|
||
except BpRefused as e:
|
||
self.refused.append({"op": kind, "path": op.get("path"), "message": e.message, "code": e.code})
|
||
return Applied(False, False, e.message)
|
||
self.effective.append(op)
|
||
where = str(op.get("path") or op.get("op"))
|
||
if where not in self._landed:
|
||
self._landed.append(where)
|
||
return Applied(True, True)
|
||
|
||
def _with_reuse(self, op: dict[str, Any]) -> dict[str, Any]:
|
||
"""复用口径卡：按这一轮读好的口径核对，补上那张卡读的取数节点和列。没读口径（calibers 为空）时原样交给存的那一步核。"""
|
||
if self.calibers is None:
|
||
return op
|
||
path = str(op.get("path") or "")
|
||
if op.get("op") == "bp_set" and path.endswith("/reuse") and isinstance(op.get("value"), dict):
|
||
return {**op, "value": self._reuse(op["value"])}
|
||
if op.get("op") == "bp_add" and isinstance(op.get("value"), dict) and isinstance(op["value"].get("reuse"), dict):
|
||
return {**op, "value": {**op["value"], "reuse": self._reuse(op["value"]["reuse"])}}
|
||
return op
|
||
|
||
def _reuse(self, ref: dict[str, Any]) -> dict[str, Any]:
|
||
reuse, reason = self.calibers.reuse_for(ref)
|
||
if reuse is None:
|
||
raise BpRefused(f"这张口径卡不能复用：{reason}", code="bp.data_unmapped")
|
||
return reuse
|
||
|
||
def landed(self) -> list[str]:
|
||
return list(self._landed)
|
||
|
||
def continuing(self) -> None:
|
||
self._continuing = True
|
||
|
||
def repairing(self) -> None:
|
||
self._repairing = True
|
||
|
||
def unreported(self) -> list[dict[str, Any]]:
|
||
out = self.refused[self._reported:]
|
||
self._reported = len(self.refused)
|
||
return out
|
||
|
||
def refusals(self) -> list[dict[str, Any]]:
|
||
"""交付时列给人看的：助手想改什么、为什么没改（警告级）。"""
|
||
return [{"code": r["code"], "level": "warning", "path": r["path"],
|
||
"message": f"助手提的这一处没有采纳：{r['message']}"} for r in self.refused]
|
||
|
||
|
||
class BlueprintChecker:
|
||
"""蓝图的自查：越界被拒、还没交回过的操作，加上卡片内容超出模板能力的地方（bp.template_mismatch）——这些是助手提错了，
|
||
交回它改。缺的项（bp.missing、bp.data_unmapped……）不交回：缺的往往是只有用户知道的事，逼它补只会让它编，由它在
|
||
这一轮里 bp_ask 问，或者留给用户在卡片上填。"""
|
||
|
||
def __init__(self, applier: BlueprintApplier, *, template_id: str | None = None) -> None:
|
||
self.applier = applier
|
||
self.template_id = template_id
|
||
|
||
def lint(self) -> list[dict[str, Any]]:
|
||
from app.blueprints.ops import blueprint_issues
|
||
|
||
return blueprint_issues(self.applier.spec, for_="trial")
|
||
|
||
async def sources(self) -> list[dict[str, Any]]:
|
||
"""对到的数据源和表在不在（要查库，用自己的会话：流式回答时接口的会话可能已经关了）。对到一个不存在的源是助手
|
||
提错了，交回它改，不留给用户到试运行时才碰上。"""
|
||
from app.blueprints.store import data_source_problems
|
||
from app.db.base import SessionLocal
|
||
|
||
spec = self.applier.spec
|
||
template_id = (spec.get("template") or {}).get("id") or self.template_id
|
||
async with SessionLocal() as session:
|
||
return await data_source_problems(session, spec, template_id=template_id)
|
||
|
||
async def issues(self) -> list[dict[str, Any]]:
|
||
refused = [{"code": r["code"], "level": "error", "path": r["path"], "message": r["message"]}
|
||
for r in self.applier.unreported()]
|
||
return [*refused, *(i for i in self.lint() if i["code"] == "bp.template_mismatch" and i["level"] == "error"),
|
||
*await self.sources()]
|
||
|
||
def repair_request(self, issues: list[dict[str, Any]], looks_text: str) -> str:
|
||
return BLUEPRINT_REPAIR_PROMPT.format(spec=json.dumps(self.applier.spec, ensure_ascii=False),
|
||
issues=_issue_lines(issues),
|
||
looks=REPAIR_LOOKS_PROMPT.format(text=looks_text) if looks_text else "")
|
||
|
||
|
||
# ==========================================================================
|
||
# 模板的生成关节
|
||
# ==========================================================================
|
||
|
||
|
||
def joint_problem(joints: dict[str, Any], group: list[Any], op: dict[str, Any]) -> str | None:
|
||
"""填关节时这个操作有没有越界：越界返回交回模型的一句话，没越界返回 None（同 apps/ops.polish_problem 的写法）。
|
||
|
||
joints 是模板声明的全部关节（id → Joint），group 是这一组要填的生成关节。只收 fill_joint、只收这一组的关节、值是一段
|
||
非空的文字。骨架是锁住的：节点、连线、确定性关节（按卡片填的）和别的配置都不归模型管。"""
|
||
from app.blueprints.templates import GENERATED_JOINT_KINDS
|
||
|
||
kind = op.get("op") if isinstance(op, dict) else None
|
||
mine = {j.id for j in group}
|
||
if kind != "fill_joint":
|
||
return (f"填关节只能用 fill_joint 填模板留给你的位置（{'、'.join(sorted(mine))}），不能用 {kind}："
|
||
"模板的节点、连线和别的配置是锁住的")
|
||
joint_id = str(op.get("joint") or "")
|
||
joint = joints.get(joint_id)
|
||
if joint is None:
|
||
return f"模板里没有关节「{joint_id}」：这一轮只能填 {'、'.join(sorted(mine))}"
|
||
if joint.kind not in GENERATED_JOINT_KINDS:
|
||
return f"「{joint.label or joint_id}」按卡片确定地填，不归助手管：要改请回到蓝图卡片"
|
||
if joint_id not in mine:
|
||
return f"「{joint.label or joint_id}」不在这一组里（这一组只填 {'、'.join(sorted(mine))}），由别的那一组填"
|
||
value = op.get("value")
|
||
if not isinstance(value, str) or not value.strip():
|
||
return f"「{joint.label or joint_id}」要写一段非空的文字（{'SQL' if joint.kind == 'sql' else '指标式'}）"
|
||
if len(value) > 20_000:
|
||
return f"「{joint.label or joint_id}」太长了（超过 20000 字）"
|
||
return None
|
||
|
||
|
||
class TemplateApplier:
|
||
"""一组生成关节的落地：fill_joint 把值写到图上那个位置（graph 是这一组自己的那一份，原地改），值记在 values 里。
|
||
|
||
越界的照实交回（refused），自查时交回模型一次，交付时列给人看。"""
|
||
|
||
def __init__(self, graph: dict[str, Any], joints: list[Any], group: list[Any]) -> None:
|
||
self.graph = graph
|
||
self.joints = {j.id: j for j in joints}
|
||
self.group = list(group)
|
||
self.values: dict[str, str] = {}
|
||
self.plan: dict[str, Any] | None = None
|
||
self.replied: dict[str, Any] | None = None
|
||
self.explanation = ""
|
||
self.refused: list[dict[str, Any]] = []
|
||
self.effective: list[dict[str, Any]] = []
|
||
self._reported = 0
|
||
self._continuing = False
|
||
self._repairing = False
|
||
|
||
def apply(self, op: dict[str, Any]) -> Applied:
|
||
from app.blueprints.instantiate import node_config, write_at
|
||
|
||
kind = str(op.get("op") or "")
|
||
if kind == "plan":
|
||
if self.plan is not None and self._continuing:
|
||
return Applied(False, False)
|
||
self.plan = self.plan or op
|
||
return Applied(False, True)
|
||
if kind == "done":
|
||
self.explanation = str(op.get("explanation", ""))
|
||
return Applied(False, True)
|
||
if kind == "reply":
|
||
if self.replied is None and not self._repairing:
|
||
self.replied = op
|
||
return Applied(False, False)
|
||
if problem := joint_problem(self.joints, self.group, op):
|
||
self.refused.append({"op": kind, "joint": op.get("joint"), "message": problem})
|
||
return Applied(False, False, problem)
|
||
joint = self.joints[str(op["joint"])]
|
||
value = str(op["value"]).strip()
|
||
if self.values.get(joint.id) == value:
|
||
return Applied(False, False)
|
||
config = node_config(self.graph, joint.node_id)
|
||
if config is None:
|
||
return Applied(False, False, f"图上没有节点 {joint.node_id}")
|
||
write_at(config, joint.field, value)
|
||
self.values[joint.id] = value
|
||
self.effective.append(op)
|
||
return Applied(True, True)
|
||
|
||
def landed(self) -> list[str]:
|
||
return list(self.values)
|
||
|
||
def continuing(self) -> None:
|
||
self._continuing = True
|
||
|
||
def repairing(self) -> None:
|
||
self._repairing = True
|
||
|
||
def unreported(self) -> list[dict[str, Any]]:
|
||
out = self.refused[self._reported:]
|
||
self._reported = len(self.refused)
|
||
return out
|
||
|
||
def refusals(self) -> list[dict[str, Any]]:
|
||
return [{"code": JOINT_REFUSED, "level": "warning", "joint": r.get("joint"),
|
||
"message": f"助手想改的这一处没有采纳：{r['message']}"} for r in self.refused]
|
||
|
||
|
||
class TemplateChecker:
|
||
"""一组生成关节的自查：越界被拒的、还没填的、SQL 检查（GraphChecker：只读守卫、猜列、目录检查）、指标式白名单和
|
||
静态追到原始数据（engine/lineage.static_stops）、骨架比对。只报这一组的节点上的问题：别的组在并行地填，它们的占位
|
||
不归这一组管。"""
|
||
|
||
def __init__(self, applier: TemplateApplier, *, skeleton: dict[str, Any], sources: list[Any] | None = None,
|
||
checkers: dict[str, Any] | None = None) -> None:
|
||
self.applier = applier
|
||
self.skeleton = skeleton
|
||
self.sources = sources or []
|
||
self.checkers = checkers
|
||
|
||
def _nodes(self) -> dict[str, dict[str, Any]]:
|
||
return {n["id"]: n for n in self.applier.graph.get("nodes") or [] if isinstance(n, dict) and n.get("id")}
|
||
|
||
async def issues(self) -> list[dict[str, Any]]:
|
||
from app.blueprints.instantiate import skeleton_changes
|
||
|
||
out = [{"code": JOINT_REFUSED, "level": "error", "joint": r.get("joint"), "message": r["message"]}
|
||
for r in self.applier.unreported()]
|
||
out += [{"code": JOINT_UNFILLED, "level": "error", "joint": j.id, "node_id": j.node_id,
|
||
"message": f"「{j.label or j.id}」还没填：用 fill_joint 填上"}
|
||
for j in self.applier.group if j.id not in self.applier.values]
|
||
out += self._sql_issues()
|
||
out += self._metric_issues()
|
||
out += skeleton_changes(self.skeleton, self.applier.graph, list(self.applier.joints.values()))
|
||
return out
|
||
|
||
def _sql_issues(self) -> list[dict[str, Any]]:
|
||
"""这一组 SQL 关节所在的取数节点：只读守卫、再按 GraphChecker（和改图助手同一套：猜列、目录检查、画布校验）。"""
|
||
from app.api.copilot_gen import GraphApplier, GraphChecker
|
||
from app.data.guard import SqlRejected, check
|
||
|
||
nodes = self._nodes()
|
||
mine = {j.node_id for j in self.applier.group if j.kind == "sql" and j.id in self.applier.values}
|
||
if not mine:
|
||
return []
|
||
out: list[dict[str, Any]] = []
|
||
kinds = {getattr(s, "name", None): getattr(s, "kind", None) for s in self.sources}
|
||
for node_id in sorted(mine):
|
||
config = (nodes[node_id].get("data") or {}).get("config") or {}
|
||
sql = str((config.get("args") or {}).get("sql") or "")
|
||
source = str(config.get("tool") or "").removeprefix("db_query__")
|
||
try:
|
||
# 模板写法（{{ input.period_start }}）换成一个字面量再查：守卫按 SQL 本身判，不按渲染出来的值判
|
||
check(re.sub(r"\{\{.*?\}\}", "0", sql, flags=re.S), readonly=True, source_name=source,
|
||
dialect=kinds.get(source))
|
||
except SqlRejected as e:
|
||
out.append({"code": SQL_REJECTED, "level": "error", "node_id": node_id, "field": "args.sql",
|
||
"message": f"取数的 SQL 过不了只读检查：{e}"})
|
||
graph = GraphApplier(copy.deepcopy(nodes), copy.deepcopy(self.applier.graph.get("edges") or []))
|
||
checker = GraphChecker(graph, sources=self.sources, checkers=self.checkers)
|
||
out += [i for i in checker.blocking() or [] if i.get("node_id") in mine]
|
||
return out
|
||
|
||
def _metric_issues(self) -> list[dict[str, Any]]:
|
||
"""指标式：只许 cell(nodes.<这一组的取数节点>, 0, '列名') 取格加算术、比较；静态追得到原始数据；取的列在那条 SQL 里。"""
|
||
import ast
|
||
|
||
from app.engine.expressions import ExpressionError, cell_parts, leaf_refs, parse_expression
|
||
from app.engine.lineage import static_stops
|
||
|
||
nodes = self._nodes()
|
||
fetches = {j.node_id for j in self.applier.group if j.kind == "sql"}
|
||
flat = {nid: {"type": n.get("type"), "title": (n.get("data") or {}).get("label") or nid,
|
||
"config": (n.get("data") or {}).get("config") or {}} for nid, n in nodes.items()}
|
||
out: list[dict[str, Any]] = []
|
||
for j in self.applier.group:
|
||
if j.kind != "metric_expr" or j.id not in self.applier.values:
|
||
continue
|
||
expr = self.applier.values[j.id]
|
||
|
||
def bad(message: str, code: str = METRIC_EXPR) -> None:
|
||
out.append({"code": code, "level": "error", "joint": j.id, "node_id": j.node_id,
|
||
"message": f"「{j.label or j.id}」{message}"})
|
||
|
||
try:
|
||
tree = parse_expression(expr)[0]
|
||
except ExpressionError as e:
|
||
bad(f"写得不对：{e}")
|
||
continue
|
||
calls = {n.func.id for n in ast.walk(tree) if isinstance(n, ast.Call) and isinstance(n.func, ast.Name)}
|
||
if extra := sorted(calls - METRIC_FUNCS):
|
||
bad(f"只能用 cell() 取格再做算术（可以用 {'、'.join(sorted(METRIC_FUNCS))}），不能用 {'、'.join(extra)}")
|
||
continue
|
||
refs = leaf_refs(tree)
|
||
if not refs:
|
||
bad("没有取任何数：要用 cell() 从取数的结果里取格")
|
||
continue
|
||
for path in refs:
|
||
parts = cell_parts(path)
|
||
node_id = parts[0][len("nodes."):] if parts and parts[0].startswith("nodes.") else ""
|
||
if parts is None or node_id not in fetches:
|
||
bad(f"只能用 cell(nodes.{'/'.join(sorted(fetches))}, 0, '列名') 取这一组取数的格，不能写 {path}")
|
||
break
|
||
if parts[1].strip() != "0":
|
||
bad("取数只出一行：行号写 0")
|
||
break
|
||
column = parts[2].strip().strip("'\"")
|
||
sql = str(((flat[node_id]["config"].get("args") or {}).get("sql")) or "")
|
||
if not re.search(rf"(?<![\w]){re.escape(column)}(?![\w])", sql, re.I):
|
||
bad(f"取的列「{column}」在取数的 SQL 里找不到：SQL 要交出这一列（列名照写）", METRIC_COLUMN)
|
||
break
|
||
else:
|
||
if stops := static_stops(flat, expr):
|
||
bad(f"的输入没有追到原始数据（{'、'.join(s.path for s in stops[:3])}）：要用 cell() 取查询结果的格")
|
||
return out
|
||
|
||
def repair_request(self, issues: list[dict[str, Any]], looks_text: str) -> str:
|
||
values = "\n".join(f"{jid}：{value}" for jid, value in self.applier.values.items()) or "（还没填）"
|
||
return TEMPLATE_REPAIR_PROMPT.format(values=values, issues=_issue_lines(issues),
|
||
looks=REPAIR_LOOKS_PROMPT.format(text=looks_text) if looks_text else "")
|