mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 07:02:00 +00:00
ThearchyHelios
88 lines
3.9 KiB
Python
88 lines
3.9 KiB
Python
"""口径调用审计（next-round 02 的后续）：谁在什么时候、用哪一版口径、带什么参数。
|
||
|
||
`caliber_match` 是封存后的等价证明，不是调用日志；一次 `use_caliber` 组合或一次按口径正式出具，在这里各记一条
|
||
`AuditEvent`（action=caliber.use）。和账本其他部分一样不另建表：audit_events 已经按 subject 建了索引，按
|
||
subject_kind=caliber、subject_id=「工作流@版本」查得到。
|
||
|
||
**只记参数指纹，不记参数值。** 参数可能是遮罩值、手机号这类不该再落一份的东西；记 hash 和字段名，排查「同一次调用
|
||
用了哪组参数」够用，要复核参数去运行输入和比对工件里看（那里本来就有遮罩处理）。指标只记 id，不记算出来的值。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import logging
|
||
from typing import Any, Iterable, Mapping
|
||
|
||
from sqlalchemy.ext.asyncio import AsyncSession
|
||
|
||
ACTION = "caliber.use"
|
||
SUBJECT_KIND = "caliber"
|
||
HOWS = ("composed", "issued")
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
def params_fingerprint(values: Mapping[str, Any] | None) -> tuple[str, list[str]]:
|
||
"""(参数指纹, 字段名)。空值不参与指纹：没填和填空串是同一组参数；字段名排序，保证同一组参数写法不同也相同。"""
|
||
from app.core.artifact_store import canonical_json, content_hash
|
||
|
||
clean = {str(k): v for k, v in (values or {}).items() if v is not None and v != ""}
|
||
return content_hash(canonical_json(clean)), sorted(clean)
|
||
|
||
|
||
async def record_use(session: AsyncSession, *, workflow_id: str, version: int, how: str,
|
||
card: str | None = None, metrics: Iterable[str] = (), params: Mapping[str, Any] | None = None,
|
||
run_id: str | None = None, conversation_id: str | None = None,
|
||
actor: str | None = None, auto: bool | None = None) -> None:
|
||
"""记一条调用。不提交：和调用方自己的事务一起提交。how 只能是 composed / issued。"""
|
||
from app.apps.audit import record
|
||
|
||
if how not in HOWS:
|
||
raise ValueError(f"不认识的调用方式：{how}")
|
||
digest, fields = params_fingerprint(params)
|
||
data: dict[str, Any] = {
|
||
"workflow_id": workflow_id,
|
||
"version": int(version),
|
||
"how": how,
|
||
"params_hash": digest,
|
||
"param_fields": fields,
|
||
"metrics": [str(m) for m in dict.fromkeys(metrics)],
|
||
}
|
||
if card:
|
||
data["card"] = card
|
||
if run_id:
|
||
data["run_id"] = run_id
|
||
if conversation_id:
|
||
data["conversation_id"] = conversation_id
|
||
if auto is not None:
|
||
data["auto"] = bool(auto)
|
||
await record(session, action=ACTION, subject_kind=SUBJECT_KIND,
|
||
subject_id=f"{workflow_id}@{int(version)}", actor=actor, data=data)
|
||
|
||
|
||
async def audit_composed_uses(uses: Iterable[Mapping[str, Any]], *, conversation_id: str | None = None,
|
||
actor: str | None = None) -> int:
|
||
"""一轮生成里所有 use_caliber 落成审计，返回记了几条。取数组合发生在建图阶段，没有运行可挂，只记对话。"""
|
||
uses = list(uses)
|
||
if not uses:
|
||
return 0
|
||
from app.db.base import SessionLocal
|
||
|
||
async with SessionLocal() as session:
|
||
for use in uses:
|
||
await record_use(
|
||
session,
|
||
workflow_id=str(use.get("workflow_id") or ""),
|
||
version=int(use.get("version") or 0),
|
||
how="composed",
|
||
card=str(use.get("card") or "") or None,
|
||
metrics=use.get("metrics") or (),
|
||
params=use.get("inputs") if isinstance(use.get("inputs"), Mapping) else None,
|
||
conversation_id=conversation_id,
|
||
actor=actor,
|
||
)
|
||
await session.commit()
|
||
return len(uses)
|
||
|
||
|
||
__all__ = ["ACTION", "HOWS", "SUBJECT_KIND", "audit_composed_uses", "params_fingerprint", "record_use"]
|