sino-bc-cli 1.8.38 → 1.8.39

package/dist/skills/biz-line-strategy-generate/scripts/build_report.py

@@ -10,7 +10,7 @@
     python3 build_report.py --input payload.json [--output /path/to/file.xlsx]
     cat payload.json | python3 build_report.py --input -
 
-payload.json 结构（统一工作流另将 Agent 回包传给导出函数）：
+payload.json 兼容旧内嵌结构（统一工作流另将 Agent 回包传给导出函数）：
 {
   "output_path": "可选，未传时默认写到桌面 业务线策略生成_YYYYMMDD_HHMMSS.xlsx",
   "rows": [ { "...策略摘要行..." } ],
@@ -19,6 +19,9 @@ payload.json 结构（统一工作流另将 Agent 回包传给导出函数）：
                "...其余28项指标...": "...", "extra": ""}]
 }
 
+正常工作流以 detailSource={"version":1,"path":"本条明细.jsonl","count":N} 替代 details。
+路径相对输入 JSON 所在目录；stdin 输入则相对当前目录。
+
 列固定为：组别、业务线、本期/上期、项目名称、广告账号、
 预算策略标签、预算策略内容、预算策略learning、广告产品策略标签、广告产品策略内容、广告产品策略learning、
 受众策略标签、受众策略内容、受众策略learning、素材策略内容、素材策略learning。
@@ -37,6 +40,10 @@ import math
 import unicodedata
 import uuid
 import re
+import sqlite3
+import tempfile
+from collections import OrderedDict
+from contextlib import closing, contextmanager
 from pathlib import Path
 from xml.etree import ElementTree as ET
 from datetime import datetime
@@ -157,126 +164,26 @@ def col_letter(index: int) -> str:
     return letters
 
 
-def _intern_factory():
-    seen = {}
-    order = []
-
-    def intern(value, rich=False):
-        # 同文字的策略富文本与普通单元格分别去重，避免格式串到明细或learning。
-        key = (value, rich)
-        if key not in seen:
-            seen[key] = len(order)
-            order.append(key)
-        return seen[key]
-
-    return intern, order
-
-
-def build_shared_strings(rows, details=None, *, include_strategy=True, ai_cells=None):
-    """按输出范围收集字符串；AI 输入文件只包含明细文字。"""
-    intern, order = _intern_factory()
-    header_indices = [intern(label) for _key, label in COLUMNS] if include_strategy else []
-    row_indices = ([
-        [intern(str(row.get(key, "") or ""), key.endswith('Strategy') or key == 'adAccounts') for key, _label in COLUMNS] for row in rows
-    ] if include_strategy else [])
-    detail_header = [intern(label) for _key, label in DETAIL_COLUMNS]
-    detail_rows = []
-    for row in details or []:
-        detail_rows.append(
-            [intern(str(row.get(key, "") or "")) for key, _label in DETAIL_COLUMNS]
-        )
-
-    # AI 表复用共享字符串池并追加到末尾，前两张表的字符串索引保持不变。
-    ai_header = [intern(label) for label in AI_COLUMNS] if ai_cells is not None else []
-    ai_values = [intern(plain, tuple(parts) if any(bold for _text, bold in parts) else False)
-                 for plain, parts in ai_cells] if ai_cells is not None else []
-
-    def string_xml(text, rich):
-        """共享字符串统一保留换行与加粗，避免 AI 表使用预览器不兼容的 inlineStr。"""
-        escape_text = lambda value: xml_escape(value, {chr(13): '
'})
-        if not rich:
-            return f'<si><t xml:space="preserve">{escape_text(text)}</t></si>'
-        if isinstance(rich, tuple):
-            # AI 的 Markdown 已被拆为文字及样式，不重新解释投放内容。
-            parts = rich
-        else:
-            pattern = r'(【[^\n】]+】|\[[^\n\]]+\]|(?m:^)(?:Meta|Google|TikTok)(?=：))'
-            parts, start = [], 0
-            for match in re.finditer(pattern, text):
-                parts.extend((part, bold) for part, bold in
-                             ((text[start:match.start()], False), (match.group(), True)) if part)
-                start = match.end()
-            if text[start:] or not parts:
-                parts.append((text[start:], False))
-        runs = ''.join(f'<r>{"<rPr><b/></rPr>" if bold else ""}'
-                       f'<t xml:space="preserve">{escape_text(part)}</t></r>' for part, bold in parts)
-        return '<si>' + runs + '</si>'
-    items = ''.join(string_xml(text, rich) for text, rich in order)
-    count = len(header_indices) + sum(map(len, row_indices)) + len(detail_header) + sum(map(len, detail_rows)) + len(ai_header) + len(ai_values)
-    sst_xml = (
-        '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>'
-        f'<sst xmlns="http://schemas.openxmlformats.org/spreadsheetml/2006/main" '
-        f'count="{count}" uniqueCount="{len(order)}">{items}</sst>'
-    )
-    return sst_xml, header_indices, row_indices, detail_header, detail_rows, (ai_header, ai_values)
-
-
-def build_sheet_xml(header_indices, row_indices, columns, content_keys=None, multiline_keys=None, values=None):
-    """生成 worksheet xml：首行表头样式 s="1"，其余行内容样式 s="2"（自动换行）"""
-    content_keys = content_keys or set()
-    multiline_keys = multiline_keys or set()
-    col_count = len(columns)
-    dimension_ref = f"A1:{col_letter(col_count - 1)}{len(row_indices) + 1}"
-
-    cols_xml_parts = []
-    widths = {}
-    for i, (key, _label) in enumerate(columns):
-        if key in content_keys:
-            width = COL_WIDTH_CONTENT
-        elif key in multiline_keys:
-            width = COL_WIDTH_MULTILINE
-        else:
-            width = COL_WIDTH_DEFAULT
-        # 尽量扩列减少换行；达到Excel行高上限仍保留单元格全文，不拆业务线行。
-        if key.endswith('Strategy'):
-            while width < 255 and any(visual_lines(str(record.get(key) or ''), width) > 26 for record in (values or [])):
-                width = min(255, width + 4)
-        widths[key] = width
-        cols_xml_parts.append(
-            f'<col min="{i + 1}" max="{i + 1}" width="{width}" customWidth="1"/>'
-        )
-    cols_xml = f"<cols>{''.join(cols_xml_parts)}</cols>"
-
-    def row_xml(row_num, indices, style_id):
-        cells = []
-        for i, sst_index in enumerate(indices):
-            ref = f"{col_letter(i)}{row_num}"
-            cells.append(f'<c r="{ref}" s="{style_id}" t="s"><v>{sst_index}</v></c>')
-        # Excel 不会可靠地对程序生成的行自动增高，按中文显示宽度确定行高。
-        if row_num == 1:
-            height = 32
-        else:
-            record = (values or [{}] * len(row_indices))[row_num - 2]
-            lines = max((visual_lines(str(record.get(key) or ""), widths[key]) for key, _ in columns), default=1)
-            height = min(409, max(30, lines * 15 + 10))
-        return f'<row r="{row_num}" ht="{height}" customHeight="1">{"".join(cells)}</row>'
-
-    rows_xml_parts = [row_xml(1, header_indices, 1)]
-    for offset, indices in enumerate(row_indices):
-        rows_xml_parts.append(row_xml(offset + 2, indices, 2))
-
-    sheet_data = f"<sheetData>{''.join(rows_xml_parts)}</sheetData>"
-
-    return (
-        '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>'
-        '<worksheet xmlns="http://schemas.openxmlformats.org/spreadsheetml/2006/main">'
-        f'<dimension ref="{dimension_ref}"/>'
-        '<sheetViews><sheetView workbookViewId="0">'
-        '<pane ySplit="1" topLeftCell="A2" activePane="bottomLeft" state="frozen"/>'
-        '</sheetView></sheetViews>'
-        f'{cols_xml}{sheet_data}<autoFilter ref="{dimension_ref}"/>'
-        '</worksheet>'
-    )
+def string_xml(text, rich):
+    """共享字符串统一保留换行与加粗，避免 AI 表使用预览器不兼容的 inlineStr。"""
+    escape_text = lambda value: xml_escape(value, {chr(13): '&#13;'})
+    if not rich:
+        return f'<si><t xml:space="preserve">{escape_text(text)}</t></si>'
+    if isinstance(rich, tuple):
+        # AI 的 Markdown 已被拆为文字及样式，不重新解释投放内容。
+        parts = rich
+    else:
+        pattern = r'(【[^\n】]+】|\[[^\n\]]+\]|(?m:^)(?:Meta|Google|TikTok)(?=：))'
+        parts, start = [], 0
+        for match in re.finditer(pattern, text):
+            parts.extend((part, bold) for part, bold in
+                         ((text[start:match.start()], False), (match.group(), True)) if part)
+            start = match.end()
+        if text[start:] or not parts:
+            parts.append((text[start:], False))
+    runs = ''.join(f'<r>{"<rPr><b/></rPr>" if bold else ""}'
+                   f'<t xml:space="preserve">{escape_text(part)}</t></r>' for part, bold in parts)
+    return '<si>' + runs + '</si>'
 
 
 def default_output_path() -> str:
@@ -295,19 +202,6 @@ def visual_lines(text, width):
     return sum(max(1, math.ceil(sum(2 if unicodedata.east_asian_width(c) in 'WF' else 1 for c in line) / max(1, width - 2))) for line in text.split('\n'))
 
 
-def prepare_layout(rows, details):
-    """每条业务线与维度明细均保持原行，动作列只记录业务动作。"""
-    # 写入与回读校验都走相同清洗，避免源数据中的非法控制字符使 sharedStrings.xml 无法解析。
-    def clean(record):
-        return {key: _sanitize_xml_text(value) if isinstance(value, str) else value
-                for key, value in record.items()}
-
-    rows = [clean(r) for r in rows]
-    details = [clean(r) for r in details or []]
-    # 单元格全文保留；超过 Excel 字符上限时在写入前明确报错，避免隐式截断。
-    return rows, details
-
-
 AI_STRATEGY_HEADINGS = ('预算策略', '广告产品策略', '受众策略', '创意策略')
 AI_COLUMNS = ('一页速览', '预算策略标签', '预算策略', '广告产品策略标签',
               '广告产品策略', '受众策略标签', '受众策略', '创意策略')
@@ -383,120 +277,276 @@ def build_ai_sheet_xml(ai_sections, strategy_row, ai_indices):
             '<autoFilter ref="A1:H2"/></worksheet>')
 
 
+@contextmanager
+def _string_index(root):
+    """每次导出/回读独占磁盘索引；页缓存与热点缓存均不随明细总量增长。"""
+    with tempfile.TemporaryDirectory(prefix='.excel-index-', dir=root) as directory:
+        connection = sqlite3.connect(str(Path(directory) / 'strings.sqlite'))
+        try:
+            connection.execute('PRAGMA cache_size = -2048')
+            connection.execute('PRAGMA mmap_size = 0')
+            connection.execute('CREATE TABLE strings (id INTEGER PRIMARY KEY, key TEXT UNIQUE, text TEXT, rich TEXT)')
+            yield connection
+        finally:
+            connection.close()
+
+
+class _Strings:
+    def __init__(self, connection):
+        self.connection = connection
+        self.cache = OrderedDict()
+        self.count = 0
+        self.unique = 0
+
+    def intern(self, text, rich=False):
+        self.count += 1
+        cacheable = isinstance(rich, bool) and len(text) <= 512
+        cache_key = (text, rich) if cacheable else None
+        if cacheable and cache_key in self.cache:
+            self.cache.move_to_end(cache_key)
+            return self.cache[cache_key]
+        encoded = json.dumps(rich, ensure_ascii=False, separators=(',', ':'))
+        key = json.dumps([text, rich], ensure_ascii=False, separators=(',', ':'))
+        found = self.connection.execute('SELECT id FROM strings WHERE key = ?', (key,)).fetchone()
+        if found is None:
+            ident = self.unique
+            self.connection.execute('INSERT INTO strings VALUES (?, ?, ?, ?)', (ident, key, text, encoded))
+            self.unique += 1
+        else:
+            ident = found[0]
+        # 长单元格不进入热点缓存，限制缓存的实际字节占用。
+        if cacheable:
+            self.cache[cache_key] = ident
+            if len(self.cache) > 512:
+                self.cache.popitem(last=False)
+        return ident
+
+    def write(self, archive):
+        with archive.open('xl/sharedStrings.xml', 'w', force_zip64=True) as stream:
+            stream.write(('<?xml version="1.0" encoding="UTF-8" standalone="yes"?>'
+                          '<sst xmlns="http://schemas.openxmlformats.org/spreadsheetml/2006/main" '
+                          f'count="{self.count}" uniqueCount="{self.unique}">').encode())
+            for text, encoded in self.connection.execute('SELECT text, rich FROM strings ORDER BY id'):
+                rich = json.loads(encoded)
+                if isinstance(rich, list):
+                    rich = tuple(tuple(part) for part in rich)
+                stream.write(string_xml(text, rich).encode('utf-8'))
+            stream.write(b'</sst>')
+
+
+def _clean_record(record):
+    return {key: _sanitize_xml_text(value) if isinstance(value, str) else value
+            for key, value in record.items()}
+
+
+def _write_sheet(archive, name, records, columns, strings, content_keys, multiline_keys):
+    widths = {}
+    for key, _label in columns:
+        width = COL_WIDTH_CONTENT if key in content_keys else COL_WIDTH_MULTILINE if key in multiline_keys else COL_WIDTH_DEFAULT
+        if key.endswith('Strategy'):
+            while width < 255 and any(visual_lines(str(_clean_record(row).get(key) or ''), width) > 26 for row in records):
+                width = min(255, width + 4)
+        widths[key] = width
+    dimension = f'A1:{col_letter(len(columns) - 1)}{len(records) + 1}'
+    with archive.open(name, 'w', force_zip64=True) as stream:
+        stream.write(('<?xml version="1.0" encoding="UTF-8" standalone="yes"?>'
+                      '<worksheet xmlns="http://schemas.openxmlformats.org/spreadsheetml/2006/main">'
+                      f'<dimension ref="{dimension}"/>'
+                      '<sheetViews><sheetView workbookViewId="0">'
+                      '<pane ySplit="1" topLeftCell="A2" activePane="bottomLeft" state="frozen"/>'
+                      '</sheetView></sheetViews><cols>' + ''.join(
+                          f'<col min="{i + 1}" max="{i + 1}" width="{widths[key]}" customWidth="1"/>'
+                          for i, (key, _label) in enumerate(columns)) + '</cols><sheetData>').encode())
+        def emit(number, values, height, header=False):
+            cells = []
+            for i, ((key, _label), value) in enumerate(zip(columns, values)):
+                if len(value.encode('utf-16-le')) // 2 > 32767:
+                    raise ValueError(f'{key} 超过Excel单元格32767字符上限，请精简内容；不会拆分原始记录行')
+                rich = not header and columns is COLUMNS and (key.endswith('Strategy') or key == 'adAccounts')
+                ident = strings.intern(value, rich)
+                cells.append(f'<c r="{col_letter(i)}{number}" s="{1 if header else 2}" t="s"><v>{ident}</v></c>')
+            stream.write((f'<row r="{number}" ht="{height}" customHeight="1">' + ''.join(cells) + '</row>').encode())
+        emit(1, [label for _key, label in columns], 32, True)
+        count = 0
+        iterator = iter(records)
+        try:
+            for count, record in enumerate(iterator, 1):
+                record = _clean_record(record)
+                values = [str(record.get(key) or '') for key, _label in columns]
+                lines = max((visual_lines(value, widths[key]) for (key, _), value in zip(columns, values)), default=1)
+                emit(count + 1, values, min(409, max(30, lines * 15 + 10)))
+        finally:
+            if hasattr(iterator, 'close'):
+                iterator.close()
+        if count != len(records):
+            raise ValueError('Excel 明细源行数不一致')
+        stream.write(f'</sheetData><autoFilter ref="{dimension}"/></worksheet>'.encode())
+
+
+def _write_workbook(rows, output_path, details, ai_sections, include_strategy):
+    output = Path(output_path)
+    output.parent.mkdir(parents=True, exist_ok=True)
+    root = Path(details.path).parent if hasattr(details, 'path') else output.parent
+    created = False
+    try:
+        with _string_index(root) as connection, zipfile.ZipFile(output, 'x', zipfile.ZIP_DEFLATED) as archive:
+            created = True
+            strings = _Strings(connection)
+            content_types, relationships, workbook = CONTENT_TYPES_XML, WORKBOOK_RELS_XML, WORKBOOK_XML
+            if not include_strategy:
+                content_types = re.sub(r'<Override PartName="/xl/worksheets/sheet2.xml"[^>]*/>\n?', '', content_types)
+                relationships = re.sub(r'<Relationship Id="rId4"[^>]*/>\n?', '', relationships)
+                workbook = (workbook.replace('<sheet name="策略" sheetId="1" r:id="rId1"/>', '')
+                            .replace('<sheet name="明细" sheetId="2" r:id="rId4"/>', '<sheet name="明细" sheetId="1" r:id="rId1"/>'))
+            if ai_sections is not None:
+                content_types = content_types.replace('</Types>', '<Override PartName="/xl/worksheets/sheet3.xml" ContentType="application/vnd.openxmlformats-officedocument.spreadsheetml.worksheet+xml"/></Types>')
+                relationships = relationships.replace('</Relationships>', '<Relationship Id="rId5" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/worksheet" Target="worksheets/sheet3.xml"/></Relationships>')
+                workbook = workbook.replace('</sheets>', '<sheet name="AI策略" sheetId="3" r:id="rId5"/></sheets>')
+            for name, value in (('[Content_Types].xml', content_types), ('_rels/.rels', ROOT_RELS_XML),
+                                ('xl/workbook.xml', workbook), ('xl/_rels/workbook.xml.rels', relationships),
+                                ('xl/styles.xml', STYLES_XML)):
+                archive.writestr(name, value)
+            if include_strategy:
+                _write_sheet(archive, 'xl/worksheets/sheet1.xml', rows, COLUMNS, strings, CONTENT_KEYS, MULTILINE_KEYS)
+            _write_sheet(archive, f'xl/worksheets/sheet{2 if include_strategy else 1}.xml', details,
+                         DETAIL_COLUMNS, strings, {'name', 'extra'},
+                         {'id', *(code for code, _label, _is_percent in REPORT_METRICS)})
+            if ai_sections is not None:
+                cells = ai_cell_values(ai_sections, rows[0])
+                headers = [strings.intern(label) for label in AI_COLUMNS]
+                values = [strings.intern(plain, tuple(parts) if any(bold for _text, bold in parts) else False)
+                          for plain, parts in cells]
+                archive.writestr('xl/worksheets/sheet3.xml', build_ai_sheet_xml(ai_sections, rows[0], (headers, values)))
+            strings.write(archive)
+    except BaseException:
+        if created:
+            output.unlink()
+        raise
+    return output_path
+
+
 def write_xlsx(rows, output_path: str, details=None, *, ai_strategy=None) -> str:
     if not rows:
         raise ValueError('策略数据为空')
-    rows, details = prepare_layout(rows, details)
+    rows = [_clean_record(row) for row in rows]
     ai_sections = prepare_ai_sections(ai_strategy) if ai_strategy is not None else None
     if ai_sections is not None and len(rows) != 1:
         raise ValueError('AI 策略 sheet 必须对应唯一业务线')
-    # 保持单元格全文；超过Excel实际字符限制时明确失败，不拆行或静默丢字。
-    for row, columns in [(r, COLUMNS) for r in rows] + [(r, DETAIL_COLUMNS) for r in details]:
-        for key, _ in columns:
-            if len(str(row.get(key) or '').encode('utf-16-le')) // 2 > 32767:
-                raise ValueError(f'{key} 超过Excel单元格32767字符上限，请精简内容；不会拆分原始记录行')
-    sst_xml, header_indices, row_indices, detail_header, detail_rows, ai_indices = build_shared_strings(
-        rows, details, ai_cells=ai_cell_values(ai_sections, rows[0]) if ai_sections is not None else None
-    )
-    sheet_xml = build_sheet_xml(
-        header_indices,
-        row_indices,
-        COLUMNS,
-        content_keys=CONTENT_KEYS,
-        multiline_keys=MULTILINE_KEYS,
-        values=rows,
-    )
-    detail_xml = build_sheet_xml(
-        detail_header,
-        detail_rows,
-        DETAIL_COLUMNS,
-        content_keys={"name", "extra"},
-        multiline_keys={'id', *(code for code, _label, _is_percent in REPORT_METRICS)},
-        values=details,
-    )
-
-    out_dir = os.path.dirname(output_path)
-    if out_dir and not os.path.isdir(out_dir):
-        os.makedirs(out_dir, exist_ok=True)
-
-    with zipfile.ZipFile(output_path, "x", zipfile.ZIP_DEFLATED) as zf:
-        # 第三张表只在最终工作簿追加；草稿仍维持原两张表结构。
-        content_types = CONTENT_TYPES_XML
-        workbook_rels = WORKBOOK_RELS_XML
-        workbook_xml = WORKBOOK_XML
-        if ai_sections is not None:
-            content_types = content_types.replace('</Types>',
-                '<Override PartName="/xl/worksheets/sheet3.xml" '
-                'ContentType="application/vnd.openxmlformats-officedocument.spreadsheetml.worksheet+xml"/></Types>')
-            workbook_rels = workbook_rels.replace('</Relationships>',
-                '<Relationship Id="rId5" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/worksheet" '
-                'Target="worksheets/sheet3.xml"/></Relationships>')
-            workbook_xml = workbook_xml.replace('</sheets>',
-                '<sheet name="AI策略" sheetId="3" r:id="rId5"/></sheets>')
-        zf.writestr("[Content_Types].xml", content_types)
-        zf.writestr("_rels/.rels", ROOT_RELS_XML)
-        zf.writestr("xl/workbook.xml", workbook_xml)
-        zf.writestr("xl/_rels/workbook.xml.rels", workbook_rels)
-        zf.writestr("xl/styles.xml", STYLES_XML)
-        zf.writestr("xl/sharedStrings.xml", sst_xml)
-        zf.writestr("xl/worksheets/sheet1.xml", sheet_xml)
-        zf.writestr("xl/worksheets/sheet2.xml", detail_xml)
-        if ai_sections is not None:
-            zf.writestr("xl/worksheets/sheet3.xml", build_ai_sheet_xml(ai_sections, rows[0], ai_indices))
-
-    return output_path
+    return _write_workbook(rows, output_path, details if details is not None else [], ai_sections, True)
 
 
 def write_detail_workbook(details, output_path):
     """给 AI 单独导出只有明细 sheet 的工作簿，不暴露固定策略正文。"""
-    _rows, clean_details = prepare_layout([], details)
-    for row in clean_details:
-        for key, _label in DETAIL_COLUMNS:
-            if len(str(row.get(key) or '').encode('utf-16-le')) // 2 > 32767:
-                raise ValueError(f'明细 {key} 超过 Excel 单元格上限')
-    strings, _head, _rows, detail_head, detail_rows, _ai = build_shared_strings(
-        [], clean_details, include_strategy=False)
-    sheet = build_sheet_xml(detail_head, detail_rows, DETAIL_COLUMNS,
-                            content_keys={'name', 'extra'},
-                            multiline_keys={'id', *(code for code, _label, _is_percent in REPORT_METRICS)},
-                            values=clean_details)
-    # 从现有标准库工作簿结构裁剪成单 sheet 包，不保留策略表的共享字符串。
-    content_types = re.sub(r'<Override PartName="/xl/worksheets/sheet2.xml"[^>]*/>\n?', '', CONTENT_TYPES_XML)
-    relationships = re.sub(r'<Relationship Id="rId4"[^>]*/>\n?', '', WORKBOOK_RELS_XML)
-    workbook = (WORKBOOK_XML.replace('<sheet name="策略" sheetId="1" r:id="rId1"/>', '')
-                .replace('<sheet name="明细" sheetId="2" r:id="rId4"/>',
-                         '<sheet name="明细" sheetId="1" r:id="rId1"/>'))
-    out_dir = os.path.dirname(output_path)
-    if out_dir:
-        os.makedirs(out_dir, exist_ok=True)
-    with zipfile.ZipFile(output_path, 'x', zipfile.ZIP_DEFLATED) as archive:
-        for name, value in (('[Content_Types].xml', content_types), ('_rels/.rels', ROOT_RELS_XML),
-                            ('xl/workbook.xml', workbook), ('xl/_rels/workbook.xml.rels', relationships),
-                            ('xl/styles.xml', STYLES_XML), ('xl/sharedStrings.xml', strings),
-                            ('xl/worksheets/sheet1.xml', sheet)):
-            archive.writestr(name, value)
-    return output_path
+    return _write_workbook([], output_path, details, None, False)
+
+
+_NS = {'s': 'http://schemas.openxmlformats.org/spreadsheetml/2006/main'}
+
+
+def _xml_items(archive, name, tag):
+    """完整读取 XML，并在消费完 si/row 后从父节点移除，避免空节点积累。"""
+    prefix = '{' + _NS['s'] + '}'
+    qualified = prefix + tag
+    root_tag = prefix + ('worksheet' if tag == 'row' else 'sst')
+    sheet_data = prefix + 'sheetData'
+    sheet_data_count = 0
+    with archive.open(name) as stream:
+        stack = []
+        for event, element in ET.iterparse(stream, events=('start', 'end')):
+            if event == 'start':
+                if not stack and element.tag != root_tag:
+                    raise ValueError('Excel XML 根节点无效')
+                if tag == 'row' and element.tag == sheet_data:
+                    sheet_data_count += 1
+                    if len(stack) != 1 or sheet_data_count != 1:
+                        raise ValueError('Excel sheetData 层级或数量无效')
+                if element.tag == qualified:
+                    valid_parent = (len(stack) == 2 and stack[-1].tag == sheet_data) if tag == 'row' else len(stack) == 1
+                    if not valid_parent:
+                        raise ValueError('Excel 行或共享字符串层级无效')
+                stack.append(element)
+            else:
+                if element.tag == qualified:
+                    yield element
+                    if len(stack) > 1:
+                        stack[-2].remove(element)
+                    element.clear()
+                stack.pop()
+        if tag == 'row' and sheet_data_count != 1:
+            raise ValueError('Excel 缺少唯一 sheetData')
+
+
+def _load_strings(archive, connection):
+    with closing(_xml_items(archive, 'xl/sharedStrings.xml', 'si')) as items:
+        for ident, element in enumerate(items):
+            text = ''.join(node.text or '' for node in element.findall('.//s:t', _NS))
+            connection.execute('INSERT INTO strings (id, text) VALUES (?, ?)', (ident, text))
+
+
+def _validate_sheet(archive, name, records, columns, connection):
+    expected = iter(records)
+    count = 0
+    actual = _xml_items(archive, name, 'row')
+    cache = OrderedDict()
+    try:
+        for count, row in enumerate(actual, 1):
+            if count == 1:
+                target = [label for _key, label in columns]
+            else:
+                try:
+                    record = _clean_record(next(expected))
+                except StopIteration:
+                    raise ValueError('Excel 表存在多余行') from None
+                target = [str(record.get(key) or '') for key, _label in columns]
+            cells = row.findall('s:c', _NS)
+            if row.get('r') != str(count) or len(cells) != len(columns):
+                raise ValueError('Excel 表行列数或坐标不完整')
+            references = []
+            for index, cell in enumerate(cells):
+                reference = cell.find('s:v', _NS)
+                if cell.get('r') != f'{col_letter(index)}{count}' or cell.get('t') != 's' \
+                        or cell.find('s:f', _NS) is not None or reference is None \
+                        or not re.fullmatch(r'[0-9]+', reference.text or ''):
+                    raise ValueError('Excel 表单元格坐标或共享字符串索引无效')
+                ident = int(reference.text)
+                if ident > 9223372036854775807:
+                    raise ValueError('Excel 共享字符串索引越界')
+                references.append(ident)
+            # 每行只查询缓存缺失项；行内值单独持有，避免缓存淘汰影响本行比较。
+            values = {ident: cache[ident] for ident in references if ident in cache}
+            missing = set(references) - values.keys()
+            if missing:
+                placeholders = ','.join('?' for _ in missing)
+                values.update(connection.execute(f'SELECT id, text FROM strings WHERE id IN ({placeholders})', tuple(missing)))
+            for ident, value in zip(references, target):
+                if ident not in values or values[ident] != value:
+                    raise ValueError('Excel 表内容或共享字符串索引不完整')
+                if len(value) <= 1024:
+                    cache[ident] = value
+                    cache.move_to_end(ident)
+                    if len(cache) > 512:
+                        cache.popitem(last=False)
+        sentinel = object()
+        if count != len(records) + 1 or next(expected, sentinel) is not sentinel:
+            raise ValueError('Excel 表内容或行列数不完整')
+    finally:
+        actual.close()
+        if hasattr(expected, 'close'):
+            expected.close()
 
 
 def validate_detail_workbook(path, details):
     """回读 AI 输入，确认唯一 sheet 与本条明细逐格一致。"""
-    ns = {'s': 'http://schemas.openxmlformats.org/spreadsheetml/2006/main'}
-    _rows, clean_details = prepare_layout([], details)
-    with zipfile.ZipFile(path) as archive:
+    root = Path(details.path).parent if hasattr(details, 'path') else Path(path).parent
+    with zipfile.ZipFile(path) as archive, _string_index(root) as connection:
         if archive.testzip():
             raise ValueError('AI 明细工作簿 ZIP 校验失败')
         workbook = ET.fromstring(archive.read('xl/workbook.xml'))
-        if [item.attrib['name'] for item in workbook.findall('s:sheets/s:sheet', ns)] != ['明细']:
+        if [item.attrib['name'] for item in workbook.findall('s:sheets/s:sheet', _NS)] != ['明细']:
             raise ValueError('AI 输入工作簿只能包含明细 sheet')
-        strings = [''.join(item.itertext()) for item in ET.fromstring(archive.read('xl/sharedStrings.xml'))]
-        sheet = ET.fromstring(archive.read('xl/worksheets/sheet1.xml'))
-        actual = [[strings[int(cell.find('s:v', ns).text)] for cell in row]
-                  for row in sheet.findall('s:sheetData/s:row', ns)]
-        expected = [[label for _key, label in DETAIL_COLUMNS]] + [
-            [str(row.get(key) or '') for key, _label in DETAIL_COLUMNS] for row in clean_details]
-        if actual != expected:
-            raise ValueError('AI 输入明细 sheet 内容不完整')
-    return {'details': len(clean_details)}
+        _load_strings(archive, connection)
+        _validate_sheet(archive, 'xl/worksheets/sheet1.xml', details, DETAIL_COLUMNS, connection)
+    return {'details': len(details)}
 
 
 def _read_ai_sheet(archive):
@@ -516,8 +566,16 @@ def _read_ai_sheet(archive):
     rows = sheet.findall('s:sheetData/s:row', ns)
     if len(rows) != 2 or [row.get('r') for row in rows] != ['1', '2']:
         raise ValueError('AI 策略 sheet 必须有唯一业务线数据行')
-    strings = [''.join(node.text or '' for node in item.findall('.//s:t', ns))
-               for item in ET.fromstring(archive.read('xl/sharedStrings.xml'))]
+    needed = set()
+    for cell in sheet.findall('s:sheetData/s:row/s:c', ns):
+        reference = cell.find('s:v', ns)
+        if reference is not None and re.fullmatch(r'[0-9]+', reference.text or ''):
+            needed.add(int(reference.text))
+    strings = {}
+    with closing(_xml_items(archive, 'xl/sharedStrings.xml', 'si')) as items:
+        for ident, item in enumerate(items):
+            if ident in needed:
+                strings[ident] = ''.join(node.text or '' for node in item.findall('.//s:t', ns))
     actual = []
     for index, row in enumerate(rows, 1):
         cells = row.findall('s:c', ns)
@@ -530,7 +588,7 @@ def _read_ai_sheet(archive):
                     or not re.fullmatch(r'[0-9]+', reference.text or ''):
                 raise ValueError('AI 策略 sheet 必须引用共享字符串')
             ident = int(reference.text)
-            if ident >= len(strings):
+            if ident not in strings:
                 raise ValueError('AI 策略 sheet 共享字符串索引越界')
             values.append(strings[ident])
         actual.append(values)
@@ -550,21 +608,18 @@ def read_ai_sheet(path):
 def validate_xlsx(path, rows, details, *, ai_strategy=None):
     """逐单元格回读，与布局后的完整输入比较，避免只检验文件存在。"""
     ns = {'s': 'http://schemas.openxmlformats.org/spreadsheetml/2006/main'}
-    expected = prepare_layout(rows, details)
-    with zipfile.ZipFile(path) as archive:
+    expected = ([_clean_record(row) for row in rows], details)
+    root = Path(details.path).parent if hasattr(details, 'path') else Path(path).parent
+    with zipfile.ZipFile(path) as archive, _string_index(root) as connection:
         if archive.testzip():
             raise ValueError('Excel ZIP 校验失败')
         workbook = ET.fromstring(archive.read('xl/workbook.xml'))
         names = ['策略', '明细'] + (['AI策略'] if ai_strategy is not None else [])
         if [s.attrib['name'] for s in workbook.findall('s:sheets/s:sheet', ns)] != names:
             raise ValueError('Excel sheet 名称或顺序不完整')
-        strings = [''.join(item.itertext()) for item in ET.fromstring(archive.read('xl/sharedStrings.xml'))]
+        _load_strings(archive, connection)
         for index, (records, columns) in enumerate(zip(expected, (COLUMNS, DETAIL_COLUMNS)), 1):
-            sheet = ET.fromstring(archive.read(f'xl/worksheets/sheet{index}.xml'))
-            actual = [[strings[int(cell.find('s:v', ns).text)] for cell in row] for row in sheet.findall('s:sheetData/s:row', ns)]
-            target = [[label for _, label in columns]] + [[str(r.get(key) or '') for key, _ in columns] for r in records]
-            if actual != target:
-                raise ValueError(f'Excel 第 {index} 张表内容或行列数不完整')
+            _validate_sheet(archive, f'xl/worksheets/sheet{index}.xml', records, columns, connection)
         if ai_strategy is not None:
             if len(expected[0]) != 1:
                 raise ValueError('AI 策略 sheet 必须对应唯一业务线')
@@ -605,9 +660,9 @@ def main() -> int:
 
     output_path = args.output or payload.get("output_path") or default_output_path()
     output_path = os.path.abspath(output_path)
-    details = payload.get("details") or []
-    if not isinstance(details, list):
-        details = []
+    from detail_store import details_from_payload
+    root = Path.cwd() if args.input == "-" else Path(args.input).resolve().parent
+    details = details_from_payload(payload, root)
 
     write_xlsx(rows, output_path, details=details)