feat(mcp): 申购单/出库单生成与抽取结果局部修正(2a)

- src/infra/com_thread.py:Office COM 调用的单线程串行 actor。Word/Excel 并发驱动会互相抢
  ActiveDocument 与模态框,因此所有 COM 操作排队在同一线程执行(与 P3 的 Playwright actor 分开,
  COM 套间与 greenlet 约束不同)
- mcp__arb__doc_requisition:Excel 申购单,产物写 cases/<id>/artifacts/,支持 exclude 关键词与
  project 覆盖。实机验证:2 张普通发票 + 1 笔支付 → 排除「万用表」后生成 19456 字节 .xls,5.0s
- mcp__arb__doc_consumable:Word 出库单;纯差旅返回 skipped 而非报错。当前被缺模板挡住(见下)
- mcp__arb__case_patch_extraction:只改传入的键,回显每个字段的 from/to,未知键告警但仍写入,
  与现值相同时不写(避免无意义重抽),金额/日期仍走 normalize
- server.py:工具数 15 → 18,playbook 增加第 6 步(文档产物)与核对修正说明

发现(未处理,需产品决定):Word 模板「易耗品、出库单.doc」在 4fbb4c3(2026-08-31,同一提交新增
requisition.py)被删除,至今不在仓库里。因此 consumable.py 的出库单路径自那以后一直不可用;
Web 端 _try_fill_consumable_doc 捕获异常后只记日志,所以普通发票提交一直在静默跳过出库单生成。
已确认历史版本是空白校用表单(无姓名、无卡号、无长数字串),是否恢复或改为配置模板路径待定。

门禁:ruff / format --check / mypy / deptry 全绿,pytest 132 passed。
This commit is contained in:
wandering
2026-10-19 20:07:38 +08:00
parent 2f08bbd6e8
commit b2d271c5ed
4 changed files with 219 additions and 2 deletions
+44
View File
@@ -0,0 +1,44 @@
"""Office COM 调用的单线程串行化
Word/Excel 自动化只能在同一线程里稳定工作,而且两个 MCP 工具并发驱动
同一个 Office 应用会互相干扰(模态框、剪贴板、ActiveDocument 抢占)。
因此所有 COM 调用统一提交到一个专属工作线程串行执行。
与 Playwright 的关系:P3 的浏览器用同样的模式,但那是另一个 actor,
不要复用同一个线程——COM 套间与 greenlet 的约束不同。
"""
from __future__ import annotations
import threading
from collections.abc import Callable
from concurrent.futures import ThreadPoolExecutor
from typing import Any
_executor: ThreadPoolExecutor | None = None
_lock = threading.Lock()
def _pool() -> ThreadPoolExecutor:
global _executor
if _executor is None:
_executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="com")
return _executor
def run_com[T](fn: Callable[..., T], *args: Any, **kwargs: Any) -> T:
"""在专用 COM 线程上串行执行 fn,并等待结果。
异常原样抛出(包装在 Future.result 里会保留 traceback)。
"""
with _lock: # 同一时刻只允许一个 Office 操作
future = _pool().submit(fn, *args, **kwargs)
return future.result()
def shutdown() -> None:
"""释放线程(进程退出或测试收尾用)。"""
global _executor
if _executor is not None:
_executor.shutdown(wait=False)
_executor = None
+8 -2
View File
@@ -13,7 +13,7 @@ from typing import Any
from mcp.server import MCPServer
from .tools import analysis, extraction, materials
from .tools import analysis, documents, extraction, materials
SERVER_NAME = "auto-reimbursement"
@@ -33,7 +33,10 @@ INSTRUCTIONS = """财务报销自动化工具集(发票抽取交接 → 匹配
或用 mcp__arb__validate_info(draft=...) 不落盘反复试。不完整就向用户要补充材料,禁止自己宣布可提交。
拼 JSON 前先 mcp__arb__info_skeleton(info_type) 取骨架:required 为 true 的键必须存在(含值为空串的备注类
字段),骨架与 validation_rules.json 同源,别凭记忆造结构。
6. 每步之后用 mcp__arb__case_status 复查态势;阶段由产物推导,没有内部状态机。
6. 普通发票用 mcp__arb__doc_consumable 生成 Word 出库单、mcp__arb__doc_requisition 生成 Excel 申购单
(COM 自动化,需本机装有 Word/Excel;纯差旅会返回 skipped,不是错误)。
人工核对后要改单张票据的某个字段,用 mcp__arb__case_patch_extraction(只传要改的键,会回显 from/to)。
7. 每步之后用 mcp__arb__case_status 复查态势;阶段由产物推导,没有内部状态机。
mcp__arb__case_reset(scope) 可清派生产物重跑;mcp__arb__case_list 看全部会话。
尚未实现(**现在不要调用,也不要在回答里承诺**):Word 出库单与 Excel 申购单、浏览器填报、
@@ -53,11 +56,14 @@ TOOLS: list[Callable[..., dict[str, Any]]] = [
extraction.case_read_text,
extraction.case_save_extraction,
extraction.extraction_spec,
extraction.case_patch_extraction,
analysis.match_payments,
analysis.save_info,
analysis.validate_info,
analysis.rules_get,
analysis.info_skeleton,
documents.doc_consumable,
documents.doc_requisition,
]
+125
View File
@@ -0,0 +1,125 @@
"""Word 出库单与 Excel 申购单生成工具
都是 COM 自动化,依赖 Windows + Microsoft Office;统一走 src.infra.com_thread
的单线程 actor,避免两个工具调用同时驱动 Word/Excel。
"""
from __future__ import annotations
from pathlib import Path
from typing import Any
from ...cases import ARTIFACTS_DIR_NAME, CACHE_DIR_NAME, PROJECT_ROOT, CaseStore
from ...config import SESSION_CONFIG_KEYS, load_config
from ...infra.com_thread import run_com
from ...infra.documents import (
CONSUMABLE_DOC_FILENAME,
REQUISITION_FILENAME,
fill_consumable_from_template,
generate_requisition,
load_requisition_items,
)
from ..envelope import fail, ok
_store = CaseStore()
def _case_config(d: Path) -> dict[str, Any]:
"""项目配置 + 会话目录里 config.json 的用户字段覆盖。"""
config = load_config()
cfg_path = d / "config.json"
if cfg_path.exists():
import json
try:
raw: dict[str, Any] = json.loads(cfg_path.read_text(encoding="utf-8"))
except Exception:
raw = {}
for key in SESSION_CONFIG_KEYS:
if key in raw:
config[key] = raw[key]
return config
def doc_consumable(case_id: str, output_name: str | None = None) -> dict[str, Any]:
"""生成易耗品出库单(Word,仅普通发票需要)。
取会话里的 payment_records.csv 填入项目根目录模板的副本,输出到
cases/<id>/artifacts/。纯差旅发票会返回 skipped,不是错误。
需要 Windows + 已安装 Microsoft Word。
"""
d = _store.resolve(case_id)
if d is None:
return fail(f"会话不存在: {case_id}")
csv_path = d / "payment_records.csv"
if not csv_path.exists():
return fail("缺少 payment_records.csv", next_action="先 mcp__arb__match_payments 生成汇总")
groups = _groups(d)
if groups is not None and not groups.get("general_count", 0):
return ok(skipped=True, reason="本会话没有普通发票(纯差旅),不需要易耗品出库单")
template = PROJECT_ROOT / CONSUMABLE_DOC_FILENAME
if not template.exists():
return fail(f"出库单模板不存在: {template}", next_action="把模板放回项目根目录")
out = d / ARTIFACTS_DIR_NAME / (output_name or CONSUMABLE_DOC_FILENAME)
out.parent.mkdir(parents=True, exist_ok=True)
config = _case_config(d)
try:
run_com(fill_consumable_from_template, csv_path, template, out, config=config)
except ImportError:
return fail("缺少 pywin32,无法驱动 Word", next_action="uv sync 安装项目依赖", needs="pywin32")
except Exception as e:
return fail(f"生成出库单失败: {e}", next_action="检查 Word 是否被其他实例占用或有弹窗未关闭", doc=str(out))
return ok(doc=str(out), filename=out.name, bytes=out.stat().st_size, skipped=False)
def doc_requisition(case_id: str, exclude: list[str] | None = None, project: str | None = None) -> dict[str, Any]:
"""生成低值易耗品申购单(Excel,仅普通发票)。
exclude 传物品名关键词可剔除不参与申购的发票;project 覆盖「对应实验实训项目」,
默认取 normal_info 的报销事由。需要 Windows + 已安装 Microsoft Excel。
"""
d = _store.resolve(case_id)
if d is None:
return fail(f"会话不存在: {case_id}")
keywords = {k.strip() for k in (exclude or []) if k.strip()}
items = load_requisition_items(d / CACHE_DIR_NAME, exclude_keywords=keywords or None)
if not items:
return fail("没有可申购的普通发票", next_action="确认已抽取普通发票,或本会话是纯差旅无需申购单")
ts_template = _find_template()
if ts_template is None:
return fail(f"未找到申购单模板 {REQUISITION_FILENAME}*.xls", next_action="把模板放到项目根目录")
out = d / ARTIFACTS_DIR_NAME / f"低值易耗品申购_{d.name}.xls"
out.parent.mkdir(parents=True, exist_ok=True)
try:
produced = run_com(
generate_requisition,
d,
ts_template,
out,
project,
keywords or None,
)
except ImportError:
return fail("缺少 pywin32,无法驱动 Excel", next_action="uv sync 安装项目依赖", needs="pywin32")
except Exception as e:
return fail(f"生成申购单失败: {e}", next_action="检查 Excel 是否被其他实例占用或有弹窗未关闭", doc=str(out))
return ok(doc=str(produced), filename=Path(produced).name, items=len(items), bytes=Path(produced).stat().st_size)
def _groups(d: Path) -> dict[str, Any] | None:
import json
p = d / "invoice_groups.json"
if not p.exists():
return None
try:
data: dict[str, Any] = json.loads(p.read_text(encoding="utf-8"))
return data
except Exception:
return None
def _find_template() -> Path | None:
stem = REQUISITION_FILENAME.replace(".xls", "")
candidates = sorted(PROJECT_ROOT.glob(f"{stem}*.xls"))
return candidates[-1] if candidates else None
+42
View File
@@ -90,6 +90,48 @@ def case_save_extraction(case_id: str, source_file: str, data: dict[str, Any]) -
return ok(cached=str(path), invoice_type=normalized.get("invoice_type"), warnings=warnings)
def case_patch_extraction(case_id: str, source_file: str, patch: dict[str, Any]) -> dict[str, Any]:
"""局部修正已落盘的抽取结果(人工核对后改字段用)。
只合并传入的键,返回每个被改字段的 from/to,避免静默覆盖;
未在 schema 中出现的键会告警但仍写入(有些字段是给下游自定义用的)。
"""
d = _store.resolve(case_id)
if d is None:
return fail(f"会话不存在: {case_id}")
if not isinstance(patch, dict) or not patch:
return fail("patch 必须是非空对象")
from ...core.extraction.cache import load_cache as _load
from ...core.extraction.cache import save_extraction
current = _load(d).get(source_file)
if current is None:
return fail(f"该材料还没有抽取结果: {source_file}", next_action="先 case_save_extraction 落一份完整结果")
from ..schemas import DOCUMENT_SCHEMAS, INVOICE_TYPES
doc_type = str(current.get("invoice_type", ""))
schema = DOCUMENT_SCHEMAS.get(doc_type, {})
known = set(schema.get("required", [])) | set(schema.get("optional", [])) | {"invoice_type", "_source_file"}
warnings: list[str] = []
changed: dict[str, dict[str, Any]] = {}
merged = dict(current)
for key, value in patch.items():
if key.startswith("_"):
warnings.append(f"忽略内部字段 {key} 的修改")
continue
if key not in changed and merged.get(key) != value:
changed[key] = {"from": merged.get(key), "to": value}
if doc_type in INVOICE_TYPES and known and key not in known:
warnings.append(f"{key} 不在 {doc_type} 的字段定义里,确认拼写无误")
merged[key] = value
if not changed:
return ok(source_file=source_file, changed={}, warnings=warnings, note="patch 与现值相同,未做写入")
normalized, normalize_warnings = normalize_extraction(merged)
normalized["_source_file"] = source_file
path = save_extraction(d, source_file, normalized)
return ok(source_file=source_file, changed=changed, warnings=warnings + normalize_warnings, cached=str(path))
def extraction_spec(kind: str | None = None) -> dict[str, Any]:
"""返回抽取字段定义。kind 省略给全部文档类型;'info' 给 travel_info/normal_info 骨架。"""
return ok(**schema_doc(kind))