feat: fix bugs for bookmark;

core/config.py

@@ -18,6 +18,12 @@ class LLMConfig:
     api_key: str = "none"
     api_key: str = "none"
     model: str = "Qwen2-72B-Instruct"
     model: str = "Qwen2-72B-Instruct"
 
 
+# @dataclass
+# class LLMConfig:
+#     base_url: str = "http://192.168.252.71:9002/v1"
+#     api_key: str = "none"
+#     model: str = "Qwen2-72B-Instruct"
+
 # 最大分片数量
 # 最大分片数量
 min_single_chunk_size = 2000
 min_single_chunk_size = 2000
 max_single_chunk_size = 100000
 max_single_chunk_size = 100000

demo/demo.py

-from PIL import Image
-import pytesseract
-
-img = Image.open("ocr.png")
-
-text = pytesseract.image_to_string(img,lang="chi_sim+eng")
-
-print(text)
\ No newline at end of file

demo/spire_doc_add_comments.py

+from utils.spire_word_util import SpireWordDoc
+
+if __name__ == "__main__":
+    doc = SpireWordDoc()
+    doc.load('demo/限制版-模板-采购合同-2025-11.doc')
+    doc.add_chunk_comment(0,[
+        {
+            "id": 0,              # 必填，规则/评论唯一标识，和 key_points 共同组成批注作者键
+            "key_points": "审核要点1",            # 必填，审核要点，和 id 共同用于去重、更新、删除批注
+            "result": "不合格",      # 必填，仅 result == "不合格" 的评论会被添加/更新批注
+            "suggest": "这是第一条批注",               # 可选，批注正文；缺省为空字符串
+            "original_text": "③分期付款：（适用于货物直接发到买方指定现场的），发货前支付     %预付款        元；到货款自买方签署到货签收单后    个工作日内支付     %到货款        元; 验收款自买方（项目经理或工程人员）签署验收单后   个工作日内支付     %验收款        元；      %作为质保金，合计    元，合同产品质保期届满后        个工作日支付。",         # 可选，优先用于定位原文；为空时批注落在文档首个可用段落
+            "chunk_id": 0,              # 可选，0 基 chunk 下标；有效时优先于入参 chunk_id
+        }
+    ])
+    doc.to_file('demo/限制版-模板-采购合同-2025-11-添加批注.doc')
\ No newline at end of file

demo/spire_doc_revision_compare.py

+import argparse
+import asyncio
+import difflib
+import json
+from itertools import zip_longest
+from pathlib import Path
+from typing import Any
+
+from openpyxl import Workbook, load_workbook
+from openpyxl.styles import Alignment, Font
+
+from core.config import LLMConfig
+from utils.common_util import extract_json
+from utils.openai_util import OpenAITool
+
+
+WORD_SUFFIXES = {".doc", ".docx", ".wps"}
+EXCEL_SUFFIXES = {".xlsx", ".xlsm", ".xls"}
+EXCEL_HEADER_NAMES = {"修改前", "修订前", "变更前", "before", "before_revision"}
+
+
+RULE_EXTRACT_SYSTEM_PROMPT = """你是资深合同审查规则分析助手。
+你的任务是根据单条合同条款的“修订前/修订后”对比文本，反推该条修订体现出的审查项和审查规则。
+
+输出要求：
+1. 只输出 JSON，不要输出 Markdown、解释或多余文本。
+2. JSON 格式必须是对象，包含：
+   - rule_title：审查项名称，简短准确，例如“纠纷解决审查”
+   - rule_detail：审查规则明细，用中文分点表述，例如“1）我司只接受……\\n2）我司只接受……”
+3. 只提取能从该条修订前后对比中合理归纳出的规则，不要编造无法支撑的规则。
+4. 如果该条差异无法归纳出审查规则，输出 {"rule_title": "", "rule_detail": ""}。
+"""
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Compare Word revisions or two-column Excel diff rows and write results to Excel.")
+    parser.add_argument("--input_file", 
+                        default=r'/home/ccran/lufa-contract/demo/test/5、[修订后复审版]麓谷发展集团2026-2027年质量_修订对比_2col.xlsx',
+                        help="Path to the Word document with revisions or a two-column Excel file.")
+    parser.add_argument(
+        "-o",
+        "--output",
+        help="Path to the output Excel file. Defaults to '<input_stem>_修订对比.xlsx'.",
+    )
+    parser.add_argument(
+        "--show-text",
+        action="store_true",
+        help="Print full text before and after accepting revisions.",
+    )
+    parser.add_argument(
+        "--no-diff-print",
+        action="store_true",
+        help="Do not print unified diff to the console.",
+    )
+    parser.add_argument(
+        "--no-rule-extract",
+        action="store_true",
+        help="Do not call LLM to extract review rules from revision diff rows.",
+    )
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=10,
+        help="Number of single-row prompts to run concurrently in one batch. Defaults to 10.",
+    )
+    return parser.parse_args()
+
+
+def get_default_output_file(input_file: Path) -> Path:
+    return input_file.with_name(f"{input_file.stem}_修订对比.xlsx")
+
+
+def read_revision_texts(input_file: Path) -> tuple[str, str]:
+    from spire.doc import Document
+
+    doc = Document()
+    try:
+        doc.LoadFromFile(str(input_file))
+        before_text = doc.GetText()
+
+        if doc.HasChanges:
+            doc.AcceptChanges()
+
+        after_text = doc.GetText()
+        return before_text, after_text
+    finally:
+        doc.Close()
+
+
+def cell_to_text(value: Any) -> str:
+    if value != value:
+        return ""
+    return "" if value is None else str(value).strip()
+
+
+def is_excel_header_row(before_text: str, after_text: str) -> bool:
+    before_normalized = before_text.strip().lower()
+    after_normalized = after_text.strip().lower()
+    return before_normalized in EXCEL_HEADER_NAMES and after_normalized in {
+        "修改后",
+        "修订后",
+        "变更后",
+        "after",
+        "after_revision",
+    }
+
+
+def read_excel_diff_rows(input_file: Path) -> list[tuple[str, str]]:
+    suffix = input_file.suffix.lower()
+    if suffix in {".xlsx", ".xlsm"}:
+        wb = load_workbook(input_file, read_only=True, data_only=True)
+        try:
+            ws = wb.active
+            rows: list[tuple[str, str]] = []
+            for row_index, row in enumerate(ws.iter_rows(min_row=1, values_only=True), start=1):
+                before_line = cell_to_text(row[0] if len(row) > 0 else None)
+                after_line = cell_to_text(row[1] if len(row) > 1 else None)
+                if row_index == 1 and is_excel_header_row(before_line, after_line):
+                    continue
+                if not before_line and not after_line:
+                    continue
+                rows.append((before_line, after_line))
+            return rows
+        finally:
+            wb.close()
+
+    if suffix == ".xls":
+        try:
+            import pandas as pd
+        except ImportError as exc:
+            raise RuntimeError("读取 .xls 需要安装 pandas/xlrd，建议另存为 .xlsx 后再运行。") from exc
+
+        try:
+            df = pd.read_excel(input_file, header=None, usecols=[0, 1], dtype=str)
+        except Exception as exc:
+            raise RuntimeError("读取 .xls 失败，建议将文件另存为 .xlsx 后再运行。") from exc
+
+        rows = []
+        for row_index, row in df.iterrows():
+            before_line = cell_to_text(row.iloc[0])
+            after_line = cell_to_text(row.iloc[1])
+            if row_index == 0 and is_excel_header_row(before_line, after_line):
+                continue
+            if not before_line and not after_line:
+                continue
+            rows.append((before_line, after_line))
+        return rows
+
+    raise ValueError(f"unsupported Excel file type: {suffix}")
+
+
+def build_unified_diff(before_lines: list[str], after_lines: list[str]) -> list[str]:
+    return list(
+        difflib.unified_diff(
+            before_lines,
+            after_lines,
+            fromfile="before_revision",
+            tofile="after_revision",
+            lineterm="",
+        )
+    )
+
+
+def build_diff_rows(before_lines: list[str], after_lines: list[str]) -> list[tuple[str, str]]:
+    rows: list[tuple[str, str]] = []
+    matcher = difflib.SequenceMatcher(None, before_lines, after_lines)
+
+    for tag, before_start, before_end, after_start, after_end in matcher.get_opcodes():
+        if tag == "equal":
+            continue
+
+        before_part = before_lines[before_start:before_end]
+        after_part = after_lines[after_start:after_end]
+
+        if tag == "replace":
+            rows.extend(
+                (before_line or "", after_line or "")
+                for before_line, after_line in zip_longest(before_part, after_part)
+            )
+        elif tag == "delete":
+            rows.extend((before_line, "") for before_line in before_part)
+        elif tag == "insert":
+            rows.extend(("", after_line) for after_line in after_part)
+
+    return rows
+
+
+def write_diff_rows_to_excel(
+    diff_rows: list[tuple[str, str]],
+    output_file: Path,
+    rule_results: list[dict[str, str]] | None = None,
+) -> None:
+    wb = Workbook()
+    ws = wb.active
+    ws.title = "revision_compare"
+    ws.append(["修订前", "修订后", "审查项", "审查规则"])
+
+    for cell in ws[1]:
+        cell.font = Font(bold=True)
+        cell.alignment = Alignment(horizontal="center", vertical="center")
+
+    for index, (before_line, after_line) in enumerate(diff_rows):
+        rule = rule_results[index] if rule_results and index < len(rule_results) else {}
+        ws.append(
+            [
+                before_line,
+                after_line,
+                rule.get("rule_title", ""),
+                rule.get("rule_detail", ""),
+            ]
+        )
+
+    for row in ws.iter_rows(min_row=2):
+        for cell in row:
+            cell.alignment = Alignment(wrap_text=True, vertical="top")
+        if row[1].value:
+            row[1].font = Font(color="FF0000")
+
+    ws.column_dimensions["A"].width = 80
+    ws.column_dimensions["B"].width = 80
+    ws.column_dimensions["C"].width = 24
+    ws.column_dimensions["D"].width = 80
+    ws.freeze_panes = "A2"
+    wb.save(output_file)
+
+
+def chunk_list(items: list[tuple[str, str]], batch_size: int) -> list[list[tuple[int, str, str]]]:
+    if batch_size <= 0:
+        raise ValueError("--batch-size must be greater than 0")
+
+    indexed_items = [
+        (index, before_line, after_line)
+        for index, (before_line, after_line) in enumerate(items, start=1)
+    ]
+    return [
+        indexed_items[start : start + batch_size]
+        for start in range(0, len(indexed_items), batch_size)
+    ]
+
+
+def build_rule_extract_user_prompt(before_line: str, after_line: str) -> str:
+    return (
+        "请根据以下单条合同修订前后对比文本，提取该条修订体现出的审查项和审查规则。\n"
+        "注意：修订后的文本通常代表更符合我司要求的表达，请结合修订前被替换或删除的内容归纳规则。\n"
+        f"{json.dumps({'修订前': before_line, '修订后': after_line}, ensure_ascii=False, indent=2)}"
+    )
+
+
+def normalize_single_rule_response(value: Any) -> dict[str, str]:
+    if isinstance(value, list):
+        rules = [item for item in value if isinstance(item, dict)]
+    elif isinstance(value, dict) and isinstance(value.get("rules"), list):
+        rules = [item for item in value["rules"] if isinstance(item, dict)]
+    elif isinstance(value, dict):
+        rules = [value]
+    else:
+        rules = []
+
+    normalized_rules = [
+        {
+            "rule_title": str(item.get("rule_title", "")).strip(),
+            "rule_detail": str(item.get("rule_detail", "")).strip(),
+        }
+        for item in rules
+        if str(item.get("rule_title", "")).strip() or str(item.get("rule_detail", "")).strip()
+    ]
+    if not normalized_rules:
+        return {"rule_title": "", "rule_detail": ""}
+    return {
+        "rule_title": "\n".join(rule["rule_title"] for rule in normalized_rules if rule["rule_title"]),
+        "rule_detail": "\n\n".join(rule["rule_detail"] for rule in normalized_rules if rule["rule_detail"]),
+    }
+
+
+def build_rule_extract_messages(before_line: str, after_line: str) -> list[dict[str, str]]:
+    return [
+        {"role": "system", "content": RULE_EXTRACT_SYSTEM_PROMPT},
+        {"role": "user", "content": build_rule_extract_user_prompt(before_line, after_line)},
+    ]
+
+
+def parse_rule_response(response: str) -> dict[str, str]:
+    parsed_items = extract_json(response)
+    parsed = parsed_items[0] if parsed_items else {}
+    return normalize_single_rule_response(parsed)
+
+
+async def extract_review_rules_async(
+    diff_rows: list[tuple[str, str]],
+    batch_size: int,
+) -> list[dict[str, str]]:
+    llm = OpenAITool(LLMConfig())
+    results_by_index: dict[int, dict[str, str]] = {}
+    batches = chunk_list(diff_rows, batch_size)
+    for batch_index, batch_rows in enumerate(batches, start=1):
+        start_index = batch_rows[0][0]
+        end_index = batch_rows[-1][0]
+        print(f"正在并发提取第 {batch_index}/{len(batches)} 批审查规则，行 {start_index}-{end_index}...")
+        messages_list = [
+            build_rule_extract_messages(before_line, after_line)
+            for _, before_line, after_line in batch_rows
+        ]
+        responses = await llm.mul_chat(messages_list)
+        for (row_index, _, _), response in zip(batch_rows, responses):
+            results_by_index[row_index] = parse_rule_response(response)
+
+    results: list[dict[str, str]] = []
+    for index in range(1, len(diff_rows) + 1):
+        results.append(results_by_index.get(index, {"rule_title": "", "rule_detail": ""}))
+    return results
+
+
+def extract_review_rules(diff_rows: list[tuple[str, str]], batch_size: int) -> list[dict[str, str]]:
+    return asyncio.run(extract_review_rules_async(diff_rows, batch_size))
+
+
+def build_diff_rows_from_word(
+    input_file: Path,
+    show_text: bool,
+    print_diff: bool,
+) -> list[tuple[str, str]]:
+    before_text, after_text = read_revision_texts(input_file)
+
+    if show_text:
+        print("===== 修订前文本 =====")
+        print(before_text)
+        print("===== 修订后文本 =====")
+        print(after_text)
+
+    if before_text == after_text:
+        return []
+
+    before_lines = before_text.splitlines()
+    after_lines = after_text.splitlines()
+    diff_lines = build_unified_diff(before_lines, after_lines)
+    diff_rows = build_diff_rows(before_lines, after_lines)
+
+    if print_diff and diff_rows:
+        print("===== Diff 对比结果 =====")
+        print("\n".join(diff_lines))
+
+    return diff_rows
+
+
+def load_diff_rows(input_file: Path, show_text: bool, print_diff: bool) -> list[tuple[str, str]]:
+    suffix = input_file.suffix.lower()
+    if suffix in EXCEL_SUFFIXES:
+        diff_rows = read_excel_diff_rows(input_file)
+        print(f"从 Excel 读取到 {len(diff_rows)} 条前后对比。")
+        return diff_rows
+    if suffix in WORD_SUFFIXES:
+        return build_diff_rows_from_word(input_file, show_text, print_diff)
+    raise ValueError(f"不支持的输入文件类型: {suffix}")
+
+
+def compare_revisions(
+    input_file: Path,
+    output_file: Path,
+    show_text: bool,
+    print_diff: bool,
+    extract_rules: bool,
+    batch_size: int,
+) -> bool:
+    diff_rows = load_diff_rows(input_file, show_text, print_diff)
+
+    if not diff_rows:
+        print("没有检测到前后对比内容，不写入 Excel。")
+        return False
+
+    rule_results = extract_review_rules(diff_rows, batch_size) if extract_rules else None
+    write_diff_rows_to_excel(diff_rows, output_file, rule_results)
+    print(f"Excel Diff 对比结果已写入: {output_file}")
+    return True
+
+
+def main() -> None:
+    args = parse_args()
+    input_file = Path(args.input_file)
+    output_file = Path(args.output) if args.output else get_default_output_file(input_file)
+    compare_revisions(
+        input_file=input_file,
+        output_file=output_file,
+        show_text=args.show_text,
+        print_diff=not args.no_diff_print,
+        extract_rules=not args.no_rule_extract,
+        batch_size=args.batch_size,
+    )
+
+
+if __name__ == "__main__":
+    main()

demo/spire_doc_revision_compare_demo.py

-import argparse
-import difflib
-from itertools import zip_longest
-from pathlib import Path
-
-from openpyxl import Workbook
-from openpyxl.styles import Alignment, Font
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Compare Word revision text and write diff rows to Excel.")
-    parser.add_argument("--input_file", default=r'D:\VsProject\lufa-contract\demo\5、[修订后复审版]麓谷发展集团2026-2027年质量.docx',
-                        help="Path to the Word document with revisions.")
-    parser.add_argument(
-        "-o",
-        "--output",
-        help="Path to the output Excel file. Defaults to '<input_stem>_修订对比.xlsx'.",
-    )
-    parser.add_argument(
-        "--show-text",
-        action="store_true",
-        help="Print full text before and after accepting revisions.",
-    )
-    parser.add_argument(
-        "--no-diff-print",
-        action="store_true",
-        help="Do not print unified diff to the console.",
-    )
-    return parser.parse_args()
-
-
-def get_default_output_file(input_file: Path) -> Path:
-    return input_file.with_name(f"{input_file.stem}_修订对比.xlsx")
-
-
-def read_revision_texts(input_file: Path) -> tuple[str, str]:
-    from spire.doc import Document
-
-    doc = Document()
-    try:
-        doc.LoadFromFile(str(input_file))
-        before_text = doc.GetText()
-
-        if doc.HasChanges:
-            doc.AcceptChanges()
-
-        after_text = doc.GetText()
-        return before_text, after_text
-    finally:
-        doc.Close()
-
-
-def build_unified_diff(before_lines: list[str], after_lines: list[str]) -> list[str]:
-    return list(
-        difflib.unified_diff(
-            before_lines,
-            after_lines,
-            fromfile="before_revision",
-            tofile="after_revision",
-            lineterm="",
-        )
-    )
-
-
-def build_diff_rows(before_lines: list[str], after_lines: list[str]) -> list[tuple[str, str]]:
-    rows: list[tuple[str, str]] = []
-    matcher = difflib.SequenceMatcher(None, before_lines, after_lines)
-
-    for tag, before_start, before_end, after_start, after_end in matcher.get_opcodes():
-        if tag == "equal":
-            continue
-
-        before_part = before_lines[before_start:before_end]
-        after_part = after_lines[after_start:after_end]
-
-        if tag == "replace":
-            rows.extend(
-                (before_line or "", after_line or "")
-                for before_line, after_line in zip_longest(before_part, after_part)
-            )
-        elif tag == "delete":
-            rows.extend((before_line, "") for before_line in before_part)
-        elif tag == "insert":
-            rows.extend(("", after_line) for after_line in after_part)
-
-    return rows
-
-
-def write_diff_rows_to_excel(diff_rows: list[tuple[str, str]], output_file: Path) -> None:
-    wb = Workbook()
-    ws = wb.active
-    ws.title = "revision_compare"
-    ws.append(["修订前", "修订后"])
-
-    for cell in ws[1]:
-        cell.font = Font(bold=True)
-        cell.alignment = Alignment(horizontal="center", vertical="center")
-
-    for before_line, after_line in diff_rows:
-        ws.append([before_line, after_line])
-
-    for row in ws.iter_rows(min_row=2):
-        for cell in row:
-            cell.alignment = Alignment(wrap_text=True, vertical="top")
-        if row[1].value:
-            row[1].font = Font(color="FF0000")
-
-    ws.column_dimensions["A"].width = 80
-    ws.column_dimensions["B"].width = 80
-    ws.freeze_panes = "A2"
-    wb.save(output_file)
-
-
-def compare_revisions(input_file: Path, output_file: Path, show_text: bool, print_diff: bool) -> bool:
-    before_text, after_text = read_revision_texts(input_file)
-
-    if show_text:
-        print("===== 修订前文本 =====")
-        print(before_text)
-        print("===== 修订后文本 =====")
-        print(after_text)
-
-    if before_text == after_text:
-        print("修订前后文本一致，无差异，不写入 Excel。")
-        return False
-
-    before_lines = before_text.splitlines()
-    after_lines = after_text.splitlines()
-    diff_lines = build_unified_diff(before_lines, after_lines)
-    diff_rows = build_diff_rows(before_lines, after_lines)
-
-    if not diff_rows:
-        print("没有检测到真实差异，不写入 Excel。")
-        return False
-
-    if print_diff:
-        print("===== Diff 对比结果 =====")
-        print("\n".join(diff_lines))
-
-    write_diff_rows_to_excel(diff_rows, output_file)
-    print(f"Excel Diff 对比结果已写入: {output_file}")
-    return True
-
-
-def main() -> None:
-    args = parse_args()
-    input_file = Path(args.input_file)
-    output_file = Path(args.output) if args.output else get_default_output_file(input_file)
-    compare_revisions(
-        input_file=input_file,
-        output_file=output_file,
-        show_text=args.show_text,
-        print_diff=not args.no_diff_print,
-    )
-
-
-if __name__ == "__main__":
-    main()

skills/review-flow-skill/SKILL.md

+---
+name: review-flow-skill
+description: 合同审查编排 Skill。用于把 URL 或本地合同文件下载/解析为 txt，逐条覆盖 rules.xlsx 中的审查规则，调用 review-llm-skill 完成路由、审查、反思、合并，并将结果文件上传。
+---
+
+# Review Flow Skill
+
+## 定位
+
+`review-flow-skill` 是合同审查的流程编排层。它不直接实现 HTTP、文档解析、Excel 读写或 LLM 提示词，而是按固定顺序组合以下 Skill：
+
+- `http-skill`：URL 输入时下载源文件；审查完成后上传结果文件。
+- `doc-excel-skill`：把 Word/PDF/Excel/CSV 解析成 UTF-8 txt；读取 `assets/rules.xlsx`；把 JSON 结果写入 Excel。
+- `review-llm-skill`：执行规则路由、具体审查、反思复核、结果合并。
+
+## 标准流程
+
+1. 输入识别
+   - 如果输入是 `http://` 或 `https://` URL，必须先调用 `http-skill` 下载到本地工作目录。
+   - 如果输入是本地路径，直接进入解析步骤。
+
+2. 文档解析
+   - 优先把待审文件解析为 `.txt`，减少后续上下文占用。
+   - `.txt` 直接复用。
+   - `.doc`、`.docx`、`.wps`、`.pdf` 使用 `doc-excel-skill/scripts/doc_tool.py` 转 txt。
+   - `.xlsx`、`.xls`、`.csv`、`.tsv` 使用 `doc-excel-skill/scripts/excel_tool.py` 读取各 sheet 后序列化为 txt。
+
+3. 规则加载
+   - 默认从 `skills/review-flow-skill/assets/rules.xlsx` 读取审查规则。
+   - 未指定 sheet 时读取所有 sheet。
+   - 必须覆盖到每一条非空规则。流程不得因为路由未命中而跳过“规则覆盖记录”。
+
+4. 状态记录
+   - 每次运行在工作目录生成 `state.json`。
+   - 状态至少记录：源文件、txt 文件、规则文件、规则总数、当前规则序号、已完成规则、输出文件路径。
+   - 每完成一条规则立即写回状态，用于断点续审。
+
+5. 逐条规则审查
+   - 按规则顺序逐条推进。
+   - 对每条规则，先对 txt 分块调用 `review-llm-skill` 的 `router` 动作。
+   - 只有路由命中的分块才调用 `review` 动作。
+   - 即使所有分块均未命中，该规则也必须标记为已覆盖，并记录路由未命中。
+
+6. 单规则反思
+   - 每条规则的分块审查完成后，将该规则的原始 findings 放入 `context.findings`，调用 `reflect` 动作复核。
+   - 反思结果写入中间结果文件。
+
+7. 全局合并
+   - 所有规则审查完毕后，对反思后的 findings 按 `rule_name`、`result`、`original_text` 聚合。
+   - 同组多条 findings 调用 `merge` 动作融合 `issue` 和 `suggestion`。
+   - 输出最终 JSON 和 Excel 汇总文件。
+
+8. 结果上传
+   - 默认使用 `http-skill` 上传最终 Excel 文件。
+   - 如调用方只需要本地文件，可传 `--no-upload`。
+
+## 工具文件
+
+- `scripts/review_flow.py`：审查流程编排 CLI。
+- `assets/rules.xlsx`：默认审查规则库。
+
+## 输出文件
+
+运行目录默认位于 `skills/review-flow-skill/outputs/run-YYYYmmdd-HHMMSS/`，包含：
+
+- `input.txt`：解析后的合同文本。
+- `state.json`：断点续审状态。
+- `rule_progress.json`：每条规则的覆盖、路由、审查统计。
+- `findings_raw.json`：分块审查原始结果。
+- `findings_reflected.json`：按规则反思后的结果。
+- `final_findings.json`：全局合并后的最终 JSON 结果。
+- `review_result.xlsx`：最终 Excel 汇总文件。
+- `upload_result.json`：上传结果，仅在上传成功后生成。
+
+## 使用示例
+
+查看帮助：
+
+```bash
+python skills/review-flow-skill/scripts/review_flow.py --help
+```
+
+审查 URL 文件并上传结果：
+
+```bash
+python skills/review-flow-skill/scripts/review_flow.py \
+  https://example.com/contract.docx
+```
+
+审查本地文件，不上传：
+
+```bash
+python skills/review-flow-skill/scripts/review_flow.py \
+  demo/example.docx \
+  --no-upload
+```
+
+只审查指定 sheet：
+
+```bash
+python skills/review-flow-skill/scripts/review_flow.py \
+  demo/example.pdf \
+  --sheet-name 通用 \
+  --sheet-name 借款
+```
+
+从已有运行目录继续：
+
+```bash
+python skills/review-flow-skill/scripts/review_flow.py \
+  --resume skills/review-flow-skill/outputs/run-20260612-173000
+```
+
+调试时限制本次最多推进 3 条规则：
+
+```bash
+python skills/review-flow-skill/scripts/review_flow.py \
+  demo/example.docx \
+  --max-rules 3 \
+  --no-upload
+```
+
+## 约束
+
+- 不在该 Skill 中重写下载、上传、文档解析、Excel 操作或 LLM 提示词能力。
+- 审查过程必须逐条规则推进并持久化状态。
+- 规则覆盖记录和最终结果必须落盘，避免只保存在上下文中。
+- 默认使用 txt 分块输入 LLM，避免一次性塞入完整合同造成上下文过大。

skills/review-flow-skill/scripts/review_flow.py

+#!/usr/bin/env python3
+"""Orchestrate URL/local-file contract review across every rule."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import shutil
+import sys
+from collections import defaultdict
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+
+SCRIPT_DIR = Path(__file__).resolve().parent
+SKILL_DIR = SCRIPT_DIR.parent
+REPO_ROOT = SKILL_DIR.parents[1]
+DEFAULT_RULES_FILE = SKILL_DIR / "assets" / "rules.xlsx"
+DEFAULT_OUTPUT_ROOT = SKILL_DIR / "outputs"
+
+HTTP_SCRIPT_DIR = REPO_ROOT / "skills" / "http-skill" / "scripts"
+DOC_EXCEL_SCRIPT_DIR = REPO_ROOT / "skills" / "doc-excel-skill" / "scripts"
+REVIEW_LLM_SCRIPT_DIR = REPO_ROOT / "skills" / "review-llm-skill" / "scripts"
+
+for path in (HTTP_SCRIPT_DIR, DOC_EXCEL_SCRIPT_DIR, REVIEW_LLM_SCRIPT_DIR):
+    if str(path) not in sys.path:
+        sys.path.insert(0, str(path))
+
+import excel_tool  # type: ignore  # noqa: E402
+import http_util  # type: ignore  # noqa: E402
+from doc_tool import doc_to_txt  # type: ignore  # noqa: E402
+from segment_llm_action import run_segment_llm_action  # type: ignore  # noqa: E402
+
+
+WORD_SUFFIXES = {".doc", ".docx", ".wps"}
+PDF_SUFFIXES = {".pdf"}
+TEXT_SUFFIXES = {".txt", ".md"}
+TABLE_SUFFIXES = {".xlsx", ".xls", ".csv", ".tsv"}
+TITLE_KEYS = ("审查项", "审查项名称", "规则名称", "规则标题", "title", "name")
+
+
+def jdump(value: Any) -> str:
+    return json.dumps(value, ensure_ascii=False, indent=2)
+
+
+def now_text() -> str:
+    return datetime.now().isoformat(timespec="seconds")
+
+
+def is_url(value: str) -> bool:
+    return value.startswith(("http://", "https://"))
+
+
+def write_json(path: Path, value: Any) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(jdump(value) + "\n", encoding="utf-8")
+
+
+def read_json(path: Path, default: Any) -> Any:
+    if not path.exists() or path.stat().st_size == 0:
+        return default
+    return json.loads(path.read_text(encoding="utf-8"))
+
+
+def make_run_dir(output_root: Path) -> Path:
+    stamp = datetime.now().strftime("%Y%m%d-%H%M%S")
+    run_dir = output_root / f"run-{stamp}"
+    index = 1
+    while run_dir.exists():
+        run_dir = output_root / f"run-{stamp}-{index}"
+        index += 1
+    run_dir.mkdir(parents=True)
+    return run_dir
+
+
+def configure_http(args: argparse.Namespace) -> None:
+    http_util._configure_urls(  # noqa: SLF001
+        http_util._strip(args.base_fastgpt_url),  # noqa: SLF001
+        http_util._strip(args.base_backend_url),  # noqa: SLF001
+        http_util._strip(args.outer_backend_url),  # noqa: SLF001
+    )
+    http_util._configure_login(args.username, args.password)  # noqa: SLF001
+
+
+def resolve_source(source: str, run_dir: Path) -> Path:
+    if is_url(source):
+        downloaded = http_util.download_file(source, run_dir / "download")
+        if not downloaded:
+            raise RuntimeError(f"download failed: {source}")
+        return Path(downloaded)
+    path = Path(source).expanduser()
+    if not path.is_absolute():
+        path = Path.cwd() / path
+    if not path.exists():
+        raise FileNotFoundError(path)
+    return path
+
+
+def excel_to_txt(path: Path, output: Path) -> Path:
+    chunks: list[str] = []
+    sheets = excel_tool.list_sheets(str(path)) if path.suffix.lower() in {".xlsx", ".xls"} else [None]
+    for sheet in sheets:
+        rows = excel_tool.load_excel(str(path), sheet=sheet, header=True)
+        title = sheet or path.name
+        chunks.append(f"# Sheet: {title}\n{jdump(rows)}")
+    output.parent.mkdir(parents=True, exist_ok=True)
+    output.write_text("\n\n".join(chunks), encoding="utf-8")
+    return output
+
+
+def parse_to_txt(source_file: Path, run_dir: Path) -> Path:
+    suffix = source_file.suffix.lower()
+    output = run_dir / "input.txt"
+    if suffix in TEXT_SUFFIXES:
+        if source_file.resolve() != output.resolve():
+            shutil.copyfile(source_file, output)
+        return output
+    if suffix in WORD_SUFFIXES or suffix in PDF_SUFFIXES:
+        return Path(doc_to_txt(str(source_file), str(output)))
+    if suffix in TABLE_SUFFIXES:
+        return excel_to_txt(source_file, output)
+    raise ValueError(f"unsupported input file type: {suffix}")
+
+
+def compact_rule(row: dict[str, Any], sheet_name: str, row_index: int) -> dict[str, Any]:
+    cleaned = {str(k).strip(): v for k, v in row.items() if str(k).strip() and v not in (None, "")}
+    title = next((str(cleaned[key]).strip() for key in TITLE_KEYS if cleaned.get(key)), "")
+    if not title:
+        title = next((str(value).strip() for value in cleaned.values() if value not in (None, "")), "")
+    cleaned["_sheet_name"] = sheet_name
+    cleaned["_row_index"] = row_index
+    cleaned["_rule_id"] = f"{sheet_name}:{row_index}:{title or 'untitled'}"
+    cleaned["_title"] = title or cleaned["_rule_id"]
+    return cleaned
+
+
+def load_rules(rules_file: Path, sheet_names: list[str] | None) -> list[dict[str, Any]]:
+    selected_sheets = sheet_names or excel_tool.list_sheets(str(rules_file))
+    rules: list[dict[str, Any]] = []
+    for sheet_name in selected_sheets:
+        rows = excel_tool.search_rows(str(rules_file), sheet_name, {})
+        for index, row in enumerate(rows, start=2):
+            if not isinstance(row, dict):
+                continue
+            rule = compact_rule(row, sheet_name, index)
+            if any(not str(key).startswith("_") for key in rule):
+                rules.append(rule)
+    return rules
+
+
+def split_text(text: str, chunk_size: int) -> list[dict[str, Any]]:
+    if chunk_size <= 0:
+        raise ValueError("--chunk-size must be > 0")
+    total = max(1, math.ceil(len(text) / chunk_size))
+    return [
+        {
+            "chunk_index": index,
+            "chunk_total": total,
+            "start": index * chunk_size,
+            "end": min((index + 1) * chunk_size, len(text)),
+            "text": text[index * chunk_size : (index + 1) * chunk_size],
+        }
+        for index in range(total)
+    ]
+
+
+def selected_by_router(router_result: Any, rule: dict[str, Any]) -> bool:
+    if isinstance(router_result, list):
+        return bool(router_result)
+    if not isinstance(router_result, dict):
+        return False
+    selected = router_result.get("selected_items") or router_result.get("selected_rules") or []
+    if not isinstance(selected, list):
+        return False
+    if not selected:
+        return False
+    title = str(rule.get("_title", ""))
+    rule_id = str(rule.get("_rule_id", ""))
+    for item in selected:
+        if not isinstance(item, dict):
+            return True
+        text = jdump(item)
+        if title and title in text:
+            return True
+        if rule_id and rule_id in text:
+            return True
+    return True
+
+
+def normalize_findings(value: Any) -> list[dict[str, Any]]:
+    if isinstance(value, dict):
+        if isinstance(value.get("findings"), list):
+            value = value["findings"]
+        else:
+            value = [value]
+    if not isinstance(value, list):
+        return []
+    return [item for item in value if isinstance(item, dict)]
+
+
+def call_llm(action: str, rule: dict[str, Any], text: str) -> Any:
+    return run_segment_llm_action(action=action, rule=rule, text=text)
+
+
+def review_rule(rule: dict[str, Any], chunks: list[dict[str, Any]]) -> dict[str, Any]:
+    routed_chunks: list[int] = []
+    router_results: list[dict[str, Any]] = []
+    raw_findings: list[dict[str, Any]] = []
+
+    router_rule = {
+        "title": rule.get("_title"),
+        "rule_id": rule.get("_rule_id"),
+        "candidate_rules": [rule],
+    }
+    for chunk in chunks:
+        router_result = call_llm("router", router_rule, chunk["text"])
+        hit = selected_by_router(router_result, rule)
+        router_results.append(
+            {
+                "chunk_index": chunk["chunk_index"],
+                "selected": hit,
+                "router_result": router_result,
+            }
+        )
+        if not hit:
+            continue
+
+        routed_chunks.append(chunk["chunk_index"])
+        review_rule_payload = {
+            **rule,
+            "context": {
+                "chunk_index": chunk["chunk_index"],
+                "chunk_total": chunk["chunk_total"],
+            },
+        }
+        review_result = call_llm("review", review_rule_payload, chunk["text"])
+        for finding in normalize_findings(review_result):
+            finding.setdefault("rule_name", rule.get("_title"))
+            finding["_rule_id"] = rule.get("_rule_id")
+            finding["_sheet_name"] = rule.get("_sheet_name")
+            finding["_row_index"] = rule.get("_row_index")
+            finding["_chunk_index"] = chunk["chunk_index"]
+            raw_findings.append(finding)
+
+    reflected_findings: list[dict[str, Any]] = []
+    if raw_findings:
+        reflect_payload = {
+            **rule,
+            "context": {
+                "findings": raw_findings,
+                "routed_chunks": routed_chunks,
+            },
+        }
+        evidence_text = "\n\n".join(
+            str(item.get("original_text", "")).strip()
+            for item in raw_findings
+            if item.get("original_text")
+        )
+        reflect_result = call_llm("reflect", reflect_payload, evidence_text)
+        reflected_findings = normalize_findings(reflect_result)
+        for finding in reflected_findings:
+            finding.setdefault("rule_name", rule.get("_title"))
+            finding["_rule_id"] = rule.get("_rule_id")
+            finding["_sheet_name"] = rule.get("_sheet_name")
+            finding["_row_index"] = rule.get("_row_index")
+
+    return {
+        "rule": rule,
+        "router_results": router_results,
+        "routed_chunks": routed_chunks,
+        "raw_findings": raw_findings,
+        "reflected_findings": reflected_findings,
+    }
+
+
+def merge_findings(findings: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    grouped: dict[tuple[str, str, str], list[dict[str, Any]]] = defaultdict(list)
+    for finding in findings:
+        key = (
+            str(finding.get("rule_name", "")),
+            str(finding.get("result", "")),
+            str(finding.get("original_text", "")),
+        )
+        grouped[key].append(finding)
+
+    merged: list[dict[str, Any]] = []
+    for (rule_name, result, original_text), items in grouped.items():
+        if len(items) == 1:
+            merged.append(items[0])
+            continue
+        merge_payload = {
+            "rule_name": rule_name,
+            "result": result,
+            "original_text": original_text,
+            "findings": [
+                {
+                    "issue": item.get("issue", ""),
+                    "suggestion": item.get("suggestion", ""),
+                    "original_text": item.get("original_text", ""),
+                }
+                for item in items
+            ],
+        }
+        merged_result = call_llm("merge", merge_payload, "")
+        base = dict(items[0])
+        if isinstance(merged_result, dict):
+            base["issue"] = merged_result.get("issue", base.get("issue", ""))
+            base["suggestion"] = merged_result.get("suggestion", base.get("suggestion", ""))
+        merged.append(base)
+    return merged
+
+
+def save_outputs(run_dir: Path, progress: list[dict[str, Any]], raw: list[dict[str, Any]], reflected: list[dict[str, Any]], final: list[dict[str, Any]] | None = None) -> None:
+    write_json(run_dir / "rule_progress.json", progress)
+    write_json(run_dir / "findings_raw.json", raw)
+    write_json(run_dir / "findings_reflected.json", reflected)
+    if final is not None:
+        write_json(run_dir / "final_findings.json", final)
+        excel_tool.json_to_sheet(final, str(run_dir / "review_result.xlsx"), "审查结果")
+
+
+def update_state(run_dir: Path, state: dict[str, Any]) -> None:
+    state["updated_at"] = now_text()
+    write_json(run_dir / "state.json", state)
+
+
+def initial_state(args: argparse.Namespace, run_dir: Path, source_file: Path, text_file: Path, rules: list[dict[str, Any]]) -> dict[str, Any]:
+    return {
+        "created_at": now_text(),
+        "updated_at": now_text(),
+        "source": args.source,
+        "source_file": str(source_file),
+        "text_file": str(text_file),
+        "rules_file": str(Path(args.rules_file).resolve()),
+        "sheet_names": args.sheet_name or "ALL",
+        "chunk_size": args.chunk_size,
+        "current_rule_index": 0,
+        "total_rules": len(rules),
+        "completed_rules": [],
+        "outputs": {
+            "progress": str(run_dir / "rule_progress.json"),
+            "raw": str(run_dir / "findings_raw.json"),
+            "reflected": str(run_dir / "findings_reflected.json"),
+            "final": str(run_dir / "final_findings.json"),
+            "excel": str(run_dir / "review_result.xlsx"),
+        },
+    }
+
+
+def run_new(args: argparse.Namespace) -> tuple[Path, dict[str, Any], list[dict[str, Any]], list[dict[str, Any]]]:
+    run_dir = make_run_dir(Path(args.output_root))
+    configure_http(args)
+    source_file = resolve_source(args.source, run_dir)
+    text_file = parse_to_txt(source_file, run_dir)
+    rules = load_rules(Path(args.rules_file), args.sheet_name)
+    state = initial_state(args, run_dir, source_file, text_file, rules)
+    write_json(run_dir / "rules.loaded.json", rules)
+    update_state(run_dir, state)
+    return run_dir, state, rules, split_text(text_file.read_text(encoding="utf-8"), args.chunk_size)
+
+
+def run_resume(args: argparse.Namespace) -> tuple[Path, dict[str, Any], list[dict[str, Any]], list[dict[str, Any]]]:
+    run_dir = Path(args.resume).expanduser()
+    if not run_dir.is_absolute():
+        run_dir = Path.cwd() / run_dir
+    state = read_json(run_dir / "state.json", {})
+    if not state:
+        raise FileNotFoundError(run_dir / "state.json")
+    rules = read_json(run_dir / "rules.loaded.json", [])
+    if not rules:
+        rules = load_rules(Path(state["rules_file"]), None if state.get("sheet_names") == "ALL" else state.get("sheet_names"))
+    text = Path(state["text_file"]).read_text(encoding="utf-8")
+    return run_dir, state, rules, split_text(text, int(state["chunk_size"]))
+
+
+def upload_result(args: argparse.Namespace, run_dir: Path) -> None:
+    if args.no_upload:
+        return
+    configure_http(args)
+    result = http_util.upload_file(run_dir / "review_result.xlsx")
+    write_json(run_dir / "upload_result.json", {"uploaded": result})
+
+
+def run_review(args: argparse.Namespace) -> Path:
+    run_dir, state, rules, chunks = run_resume(args) if args.resume else run_new(args)
+
+    progress = read_json(run_dir / "rule_progress.json", [])
+    raw_findings = read_json(run_dir / "findings_raw.json", [])
+    reflected_findings = read_json(run_dir / "findings_reflected.json", [])
+
+    start = int(state.get("current_rule_index", 0))
+    stop = len(rules)
+    if args.max_rules is not None:
+        stop = min(stop, start + args.max_rules)
+
+    for rule_index in range(start, stop):
+        rule_result = review_rule(rules[rule_index], chunks)
+        progress.append(
+            {
+                "rule_index": rule_index,
+                "rule_id": rules[rule_index].get("_rule_id"),
+                "title": rules[rule_index].get("_title"),
+                "routed_chunks": rule_result["routed_chunks"],
+                "covered": True,
+                "raw_count": len(rule_result["raw_findings"]),
+                "reflected_count": len(rule_result["reflected_findings"]),
+                "completed_at": now_text(),
+                "router_results": rule_result["router_results"],
+            }
+        )
+        raw_findings.extend(rule_result["raw_findings"])
+        reflected_findings.extend(rule_result["reflected_findings"])
+        state["current_rule_index"] = rule_index + 1
+        state["completed_rules"].append(rules[rule_index].get("_rule_id"))
+        save_outputs(run_dir, progress, raw_findings, reflected_findings)
+        update_state(run_dir, state)
+
+    if int(state.get("current_rule_index", 0)) >= len(rules):
+        final_findings = merge_findings(reflected_findings)
+        save_outputs(run_dir, progress, raw_findings, reflected_findings, final_findings)
+        state["finalized_at"] = now_text()
+        update_state(run_dir, state)
+        upload_result(args, run_dir)
+
+    return run_dir
+
+
+def parser() -> argparse.ArgumentParser:
+    p = argparse.ArgumentParser(description="Run contract review flow over every rule.")
+    p.add_argument("source", nargs="?", help="URL or local contract file. Required unless --resume is used.")
+    p.add_argument("--resume", help="Resume from an existing run directory.")
+    p.add_argument("--rules-file", default=str(DEFAULT_RULES_FILE))
+    p.add_argument("--sheet-name", action="append", help="Rule sheet to load. May be repeated. Defaults to all sheets.")
+    p.add_argument("--output-root", default=str(DEFAULT_OUTPUT_ROOT))
+    p.add_argument("--chunk-size", type=int, default=6000)
+    p.add_argument("--max-rules", type=int, help="Only advance this many rules in the current invocation.")
+    p.add_argument("--no-upload", action="store_true")
+    p.add_argument("--base-fastgpt-url", default=http_util.DEFAULT_BASE_FASTGPT_URL)
+    p.add_argument("--base-backend-url", default=http_util.DEFAULT_BASE_BACKEND_URL)
+    p.add_argument("--outer-backend-url", default=http_util.DEFAULT_OUTER_BACKEND_URL)
+    p.add_argument("--username", default=http_util.DEFAULT_BACKEND_ADMIN_USERNAME)
+    p.add_argument("--password", default=http_util.DEFAULT_BACKEND_ADMIN_PASSWORD)
+    return p
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = parser().parse_args(argv)
+    if not args.resume and not args.source:
+        raise SystemExit("source is required unless --resume is used")
+    run_dir = run_review(args)
+    print(run_dir)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

skills/review-llm-skill/scripts/review_llm_skill.py

-#!/usr/bin/env python3
-"""Backward-compatible entry point for segment_llm_action.py."""
-
-from __future__ import annotations
-
-from segment_llm_action import *  # noqa: F403
-from segment_llm_action import main
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())

skills/review-llm-skill/scripts/segment_llm_action.py

@@ -19,7 +19,6 @@ try:
         REFLECT_SYSTEM_PROMPT,
         REFLECT_SYSTEM_PROMPT,
         REVIEW_SYSTEM_PROMPT,
         REVIEW_SYSTEM_PROMPT,
         ROUTER_SYSTEM_PROMPT,
         ROUTER_SYSTEM_PROMPT,
-        SUMMARY_ROUTER_SYSTEM_PROMPT,
         SUMMARY_SYSTEM_PROMPT,
         SUMMARY_SYSTEM_PROMPT,
     )
     )
 except ImportError:
 except ImportError:
@@ -28,7 +27,6 @@ except ImportError:
         REFLECT_SYSTEM_PROMPT,
         REFLECT_SYSTEM_PROMPT,
         REVIEW_SYSTEM_PROMPT,
         REVIEW_SYSTEM_PROMPT,
         ROUTER_SYSTEM_PROMPT,
         ROUTER_SYSTEM_PROMPT,
-        SUMMARY_ROUTER_SYSTEM_PROMPT,
         SUMMARY_SYSTEM_PROMPT,
         SUMMARY_SYSTEM_PROMPT,
     )
     )
 
 

utils/spire_word_util.py

@@ -470,6 +470,7 @@ class SpireWordDoc(DocBase):
                     current_child_text = self._resolve_table(table)
                     current_child_text = self._resolve_table(table)
                 # 跳过其他非文本子对象
                 # 跳过其他非文本子对象
                 else:
                 else:
+                    print(child_obj)
                     continue
                     continue
                 # 添加新对象
                 # 添加新对象
                 if (
                 if (
@@ -794,6 +795,22 @@ class SpireWordDoc(DocBase):
     def add_chunk_comment(self, chunk_id, comments):
     def add_chunk_comment(self, chunk_id, comments):
         """
         """
         为 chunk 添加批注（保证每条评论只批注一次）。
         为 chunk 添加批注（保证每条评论只批注一次）。
+
+        comments 为评论字典列表，每个元素结构如下：
+        {
+            "id": str | int,              # 必填，规则/评论唯一标识，和 key_points 共同组成批注作者键
+            "key_points": str,            # 必填，审核要点，和 id 共同用于去重、更新、删除批注
+            "result": "不合格" | str,      # 必填，仅 result == "不合格" 的评论会被添加/更新批注
+            "suggest": str,               # 可选，批注正文；缺省为空字符串
+            "original_text": str,         # 可选，优先用于定位原文；为空时批注落在文档首个可用段落
+            "chunk_id": int,              # 可选，0 基 chunk 下标；有效时优先于入参 chunk_id
+        }
+
+        说明：
+        - 批注作者会格式化为 "{id}|{key_points}"，用于识别同一条评论并避免重复插入。
+        - original_text 非空时，先在目标 chunk 内精确匹配，失败后再做模糊匹配。
+        - comment["chunk_id"] 缺失或无效时，使用入参 chunk_id。
+
         执行顺序：
         执行顺序：
         1) 过滤非“不合格”项；
         1) 过滤非“不合格”项；
         2) 先按作者标识查重，命中则更新内容；
         2) 先按作者标识查重，命中则更新内容；
@@ -936,6 +953,9 @@ class SpireWordDoc(DocBase):
 
 
         if remove_prefix:
         if remove_prefix:
             self.remove_comment_prefix()
             self.remove_comment_prefix()
+        self._doc.Replace('!Undefined Bookmark,','',False,True)
+        self._doc.Replace('!Undefined Book','',False,True)
+        self._doc.Replace('!Undefined B','',False,True)
         self._doc.SaveToFile(path)
         self._doc.SaveToFile(path)
 
 
     def release(self):
     def release(self):