@lark-apaas/coding-steering 0.1.31 → 0.1.32-beta.0

package/steering/design-html/skills/pptx-style-extract/scripts/query.py

@@ -0,0 +1,562 @@
+#!/usr/bin/env python3
+"""L 层查询入口：按需从阶段一产物里捞数，避免整读 extract.json / shapes.json。
+
+    python3 query.py <stage1-outdir> <子命令> [选项]
+
+      shapes   --part slide7.xml [--kind text|pic|shape|table|group] [--ph] [--limit N]
+      colors   [--top N] [--class design|editor|aux]
+      fonts    [--all]                     默认只列 rendered_n>0 的族
+      text-scale [--top N]
+      images   [--fullscreen] [--repeat] [--top N]
+      clusters [--multi]                   --multi 只看跨文件同素材簇
+      media    [--candidate]
+      summary                               阶段一体检：form/themes/UNRESOLVED/降级信号
+      get <点路径>                          取 extract.json 任意字段（如 get canvas）
+      slides                                每页背景 + 版式
+      layouts                               版式清单
+
+输出是给 agent 读的紧凑表格：定宽列 + 表头，数值直接可抄进产物。
+`shapes` 读 ref/shapes.json（体量大，务必带 --part 过滤），其余读 extract.json。
+"""
+import argparse
+import json
+import math
+import os
+import sys
+from collections import Counter
+
+
+def load(outdir, name):
+    p = os.path.join(outdir, name)
+    if not os.path.exists(p):
+        raise SystemExit('query.py: 找不到 %s' % p)
+    with open(p, encoding='utf-8') as f:
+        return json.load(f)
+
+
+def short(part):
+    return part.rsplit('/', 1)[-1] if part else '-'
+
+
+def trunc(s, n):
+    s = '' if s is None else str(s).replace('\n', ' ')
+    return s if len(s) <= n else s[:n - 1] + '…'
+
+
+def table(rows, headers, aligns=None):
+    if not rows:
+        print('（无匹配）')
+        return
+    cols = len(headers)
+    w = [len(str(h)) for h in headers]
+    for r in rows:
+        for i in range(cols):
+            w[i] = max(w[i], len(str(r[i])))
+    aligns = aligns or ['<'] * cols
+    fmt = '  '.join('{:%s%d}' % (aligns[i], w[i]) for i in range(cols))
+    print(fmt.format(*headers))
+    print('  '.join('-' * w[i] for i in range(cols)))
+    for r in rows:
+        print(fmt.format(*[('' if x is None else x) for x in r]))
+
+
+# ------------------------------------------------------------------- 子命令
+KIND_GROUPS = {'text': None, 'pic': ('pic',), 'shape': ('sp', 'cxnSp'),
+               'table': ('table',), 'group': ('grpSp',)}
+
+
+def cmd_shapes(a, outdir):
+    data = load(outdir, os.path.join('ref', 'shapes.json'))
+    shapes = data['shapes'] if isinstance(data, dict) else data
+    if a.part:
+        want = a.part
+        shapes = [s for s in shapes if short(s['part']) == want or s['part'] == want]
+        if not shapes:
+            parts = sorted({short(s['part']) for s in (data['shapes'] if isinstance(data, dict) else data)})
+            raise SystemExit('query.py: 没有 part=%s。可选: %s' % (want, ', '.join(parts[:12]) + ' …'))
+    if a.kind:
+        if a.kind == 'text':
+            shapes = [s for s in shapes if (s.get('text') or {}).get('paragraphs')]
+        else:
+            keep = KIND_GROUPS[a.kind]
+            shapes = [s for s in shapes if s.get('kind') in keep]
+    if a.ph:
+        shapes = [s for s in shapes if s.get('ph')]
+    rows = []
+    for s in shapes[:a.limit]:
+        b = s.get('box') or {}
+        txt = ''
+        for p in (s.get('text') or {}).get('paragraphs', []):
+            txt += ''.join(r.get('text') or '' for r in p.get('runs', []))
+            if len(txt) > 40:
+                break
+        # 有效字号按 OOXML 优先级找第一个有声明的层：run > 段落 defRPr > 本形状 lstStyle
+        text = s.get('text') or {}
+        sz = next((r.get('sz_px') for p in text.get('paragraphs', [])
+                   for r in p.get('runs', []) if r.get('sz_px')), None)
+        if sz is None:
+            sz = next((p['defRPr']['sz_px'] for p in text.get('paragraphs', [])
+                       if (p.get('defRPr') or {}).get('sz_px')), None)
+        if sz is None:
+            sz = next((lvl.get('sz_px') for lvl in (text.get('lstStyle') or {}).values()
+                       if lvl.get('sz_px')), None)
+        fill = (s.get('fill') or {}).get('type')
+        col = ((s.get('fill') or {}).get('color') or {}).get('resolved')
+        rows.append([s.get('kind'), s.get('id'), trunc(s.get('name'), 22),
+                     '%s/%s' % (s['ph']['type'], s['ph'].get('idx')) if s.get('ph') else '-',
+                     b.get('x'), b.get('y'), b.get('w'), b.get('h'),
+                     sz or '-', (fill or '-') + (('=' + col) if col else ''),
+                     trunc(txt, 34)])
+    table(rows, ['kind', 'id', 'name', 'ph', 'x', 'y', 'w', 'h', 'sz', 'fill', 'text'],
+          ['<', '>', '<', '<', '>', '>', '>', '>', '>', '<', '<'])
+    print('\n%d 个形状%s' % (len(shapes), '（截断到 %d）' % a.limit if len(shapes) > a.limit else ''))
+
+
+def cmd_colors(a, outdir):
+    d = load(outdir, 'extract.json')
+    freq = d['color_freq']
+    if a.klass:
+        freq = [c for c in freq if c['class'] == a.klass]
+    rows = [[c.get('resolved'), c['n'], c['class'],
+             trunc(','.join(c.get('raw') or []), 40),
+             trunc(','.join('%s:%s' % kv for kv in (c.get('layers') or {}).items()), 26)]
+            for c in freq[:a.top]]
+    table(rows, ['resolved', 'n', 'class', 'raw tokens', 'layers'],
+          ['<', '>', '<', '<', '<'])
+    print('\n共 %d 支色（design %d / editor %d / aux %d）'
+          % (len(d['color_freq']),
+             sum(1 for c in d['color_freq'] if c['class'] == 'design'),
+             sum(1 for c in d['color_freq'] if c['class'] == 'editor'),
+             sum(1 for c in d['color_freq'] if c['class'] == 'aux')))
+
+
+def cmd_fonts(a, outdir):
+    d = load(outdir, 'extract.json')
+    fams = d['font_families']
+    if not a.all:
+        fams = [f for f in fams if f.get('rendered_n', f['n'])]
+    rows = [[f['family'], f['n'], f.get('rendered_n', f['n']),
+             ','.join(str(w) for w in f.get('weights') or []) or '-',
+             f.get('bold_runs', 0),
+             trunc(','.join('%s:%s' % kv for kv in (f.get('sources') or {}).items()), 34),
+             'Y' if f.get('in_theme') else '',
+             'renders_no_text' if f.get('renders_no_text') else '']
+            for f in fams]
+    table(rows, ['family', 'n', 'rendered', 'weights', 'bold', 'sources', 'theme', 'flag'],
+          ['<', '>', '>', '<', '>', '<', '<', '<'])
+    print('\n提示：判「在用」看 rendered —— n 含空段落声明（不渲染文字）。'
+          ' 全部族用 --all。')
+
+
+def cmd_text_scale(a, outdir):
+    d = load(outdir, 'extract.json')
+    rows = [[e['sz_px'], e['n'],
+             trunc(','.join('%s:%s' % kv for kv in (e.get('sources') or {}).items()), 34),
+             trunc(','.join('%s:%s' % kv for kv in (e.get('layers') or {}).items()), 24),
+             ','.join(str(w) for w in (e.get('weights') or {})) or '-',
+             trunc(','.join(str(k) for k in (e.get('line_height_mult') or {})), 18)]
+            for e in d['text_scale'][:a.top]]
+    table(rows, ['sz_px', 'n', 'sources', 'layers', 'weights', 'lnSpc'],
+          ['>', '>', '<', '<', '<', '<'])
+    print('\n共 %d 档字号。sources=pPr 是段落级 defRPr（Mac Office 常用）。'
+          % len(d['text_scale']))
+
+
+def cmd_images(a, outdir):
+    d = load(outdir, 'extract.json')
+    imgs = d['images']
+    if a.fullscreen:
+        imgs = [i for i in imgs if i.get('fullscreen')]
+    if a.repeat:
+        imgs = [i for i in imgs if i.get('repeat_fixed')]
+    rows = []
+    for i in imgs[:a.top]:
+        b = (i.get('boxes') or [{}])[0].get('box') or {}
+        dom = ' '.join('%s' % c['hex'] for c in (i.get('dominant_colors') or [])[:3])
+        rows.append([short(i['media']), i['n'], len(i.get('boxes') or []),
+                     b.get('x'), b.get('y'), b.get('w'), b.get('h'),
+                     'Y' if i.get('fullscreen') else '',
+                     len(i.get('repeat_fixed') or []),
+                     ','.join(i.get('variant_group') or []) or '-',
+                     i.get('content_id') or '-',
+                     i.get('luminance') if i.get('luminance') is not None else '-', dom])
+    table(rows, ['media', 'n', 'clus', 'x', 'y', 'w', 'h', 'full', 'rep',
+                 'vgroup', 'cid', 'lum', 'dominant'],
+          ['<', '>', '>', '>', '>', '>', '>', '<', '>', '<', '<', '>', '<'])
+    print('\n共 %d 张被引用素材（满屏 %d / 有固定重复位 %d）'
+          % (len(d['images']), sum(1 for i in d['images'] if i.get('fullscreen')),
+             sum(1 for i in d['images'] if i.get('repeat_fixed'))))
+
+
+# ----------------------------------------------------- L11 配方（现成 CSS）
+def _css_color(c):
+    """{resolved} 已是 #HEX 或 rgba(...)，直接用。"""
+    if not isinstance(c, dict):
+        return None
+    return c.get('resolved') or c.get('hex')
+
+
+def _css_gradient(g):
+    """OOXML gradFill → CSS linear-gradient。ang 自 +x 轴顺时针，CSS 自 12 点，故 +90。"""
+    stops = sorted(g.get('stops') or [], key=lambda s: s.get('pos', 0))
+    if not stops:
+        return None
+    parts = ['%s %g%%' % (_css_color(s.get('color')) or 'transparent', s.get('pos', 0))
+             for s in stops]
+    return 'linear-gradient(%gdeg, %s)' % ((g.get('angle_deg') or 0) + 90, ', '.join(parts))
+
+
+def _fill_css(f):
+    if not isinstance(f, dict):
+        return None, None
+    t = f.get('type')
+    if t == 'solid':
+        c = _css_color(f.get('color'))
+        return ('background: %s' % c, c) if c else (None, None)
+    if t == 'gradient':
+        g = _css_gradient(f)
+        return ('background-image: %s' % g, g) if g else (None, None)
+    if t == 'image':
+        return 'background-image: url(%s)' % (f.get('media') or '').rsplit('/', 1)[-1], None
+    if t == 'none':
+        return 'background: transparent', 'transparent'
+    return None, None
+
+
+def _recipe_css(fill, line, radii, effects):
+    """一组形状 → 可直接粘贴的 CSS 声明块。"""
+    out = []
+    fill_decl, fill_val = _fill_css(fill)
+    line_grad = _css_gradient(line['gradient']) if (line or {}).get('gradient') else None
+    w = int(round((line or {}).get('w_px') or 0)) if line and not (line or {}).get('none') else 0
+    if line_grad and w:
+        # 渐变描边只能用双背景实现：内层填充走 padding-box，描边渐变走 border-box
+        base = fill_val if fill_val and fill_val.startswith('linear-gradient') else \
+            'linear-gradient(%s, %s)' % (fill_val or 'transparent', fill_val or 'transparent')
+        out.append('border: %dpx solid transparent' % w)
+        out.append('background: %s padding-box,\n            %s border-box'
+                   % (base, line_grad))
+    else:
+        if fill_decl:
+            out.append(fill_decl)
+        if w:
+            lc = _css_color((line or {}).get('color')) or 'currentColor'
+            dash = (line or {}).get('dash')
+            style = 'dashed' if dash and 'dash' in dash else 'solid'
+            out.append('border: %dpx %s %s' % (w, style, lc))
+    if radii:
+        lo, hi = min(radii), max(radii)
+        if abs(hi - lo) <= 0.5:
+            out.append('border-radius: %gpx' % round(lo, 1))
+        else:
+            out.append('border-radius: %gpx\x00/* 源内 %g~%gpx 共 %d 档，归一档位由 L11 定 */'
+                       % (round(sum(radii) / len(radii), 1), round(lo, 1), round(hi, 1),
+                          len(set(round(r, 1) for r in radii))))
+    for e in effects or []:
+        if e.get('type') == 'outerShdw':
+            col = _css_color(e.get('color')) or 'rgba(0,0,0,0.25)'
+            blur = e.get('blurRad_px') or 0
+            dist = e.get('dist_px') or 0
+            ang = e.get('dir_deg') or 0
+            dx = dist * math.cos(math.radians(ang))
+            dy = dist * math.sin(math.radians(ang))
+            out.append('box-shadow: %.1fpx %.1fpx %.1fpx %s' % (dx, dy, blur, col))
+    return out
+
+
+def _sig(fill, line, effects):
+    """分组键 = 填充 + 描边 + 效果。**不含圆角**——OOXML 圆角是 min(w,h) 的百分比，
+    同一配方在不同尺寸的卡上绝对 px 必然不同（feishu 三张 glassCard 实测 10.8/12.8/19.5），
+    把它计入键会把一个配方拆成三组；归一到哪一档是 L11 的判断，脚本只报区间。"""
+    f = 'none'
+    if isinstance(fill, dict):
+        if fill.get('type') == 'solid':
+            f = 'solid:%s' % _css_color(fill.get('color'))
+        elif fill.get('type') == 'gradient':
+            f = 'grad:%s' % _css_gradient(fill)
+        else:
+            f = fill.get('type') or 'none'
+    ln = 'none'
+    if isinstance(line, dict) and not line.get('none'):
+        if line.get('gradient'):
+            ln = 'grad:%g:%s' % (line.get('w_px') or 0, _css_gradient(line['gradient']))
+        elif line.get('color'):
+            ln = '%g:%s:%s' % (line.get('w_px') or 0, _css_color(line['color']),
+                               line.get('dash') or '')
+    fx = ','.join(sorted(e.get('type', '') for e in effects or []))
+    return (f, ln, fx)
+
+
+def cmd_recipes(a, outdir):
+    data = load(outdir, os.path.join('ref', 'shapes.json'))
+    shapes = data['shapes'] if isinstance(data, dict) else data
+    groups = {}
+    for s in shapes:
+        fill, line = s.get('fill'), s.get('line')
+        fx = s.get('effects')
+        radius = s.get('radius_px')
+        if not fill and not line and not fx:
+            continue
+        if isinstance(fill, dict) and fill.get('type') == 'image' and not a.all:
+            continue                        # 图片填充是素材不是配方
+        k = _sig(fill, line, fx)
+        if k[0] == 'none' and k[1] == 'none' and not k[2]:
+            continue
+        g = groups.setdefault(k, {'n': 0, 'parts': Counter(), 'sizes': [], 'radii': [],
+                                  'fill': fill, 'line': line, 'fx': fx})
+        g['n'] += 1
+        if radius:
+            g['radii'].append(radius)
+        g['parts'][short(s['part'])] += 1
+        b = s.get('box') or {}
+        if b.get('w'):
+            g['sizes'].append((b['w'], b['h']))
+    rows = sorted(groups.values(), key=lambda g: -g['n'])
+    rows = [g for g in rows if g['n'] >= a.min]
+    if not rows:
+        print('（无满足 --min %d 的配方组）' % a.min)
+        return
+    for i, g in enumerate(rows[:a.top], 1):
+        css = _recipe_css(g['fill'], g['line'], g['radii'], g['fx'])
+        pages = ', '.join('%s×%d' % (p, n) if n > 1 else p
+                          for p, n in g['parts'].most_common(6))
+        if len(g['parts']) > 6:
+            pages += ' …共 %d 处' % len(g['parts'])
+        sz = ''
+        if g['sizes']:
+            ws = sorted(set(round(w) for w, _ in g['sizes']))
+            hs = sorted(set(round(h) for _, h in g['sizes']))
+            sz = '  尺寸 w%s h%s' % (ws[0] if len(ws) == 1 else '%d~%d' % (ws[0], ws[-1]),
+                                    hs[0] if len(hs) == 1 else '%d~%d' % (hs[0], hs[-1]))
+        print('[r%d] 出现 %d 次%s' % (i, g['n'], sz))
+        print('     页: %s' % pages)
+        for decl in css:
+            sub = decl.split('\n')
+            for j, ln in enumerate(sub):
+                if j == len(sub) - 1:
+                    head, _, note = ln.partition('\x00')
+                    print('     %s;%s' % (head, ('  ' + note) if note else ''))
+                else:
+                    print('     %s' % ln)
+        print()
+    print('共 %d 组配方（出现 ≥%d 次）。命名与取舍由 L11 判断，CSS 已可直接粘贴。'
+          % (len(rows), a.min))
+
+
+def cmd_grids(a, outdir):
+    d = load(outdir, 'extract.json')
+    grids = (d.get('spacing_candidates') or {}).get('grids') or []
+    rows = []
+    for g in grids:
+        c, r = g.get('cols'), g.get('rows')
+        fmt = lambda x: ('%d @%gpx' % (x['n'], x['pitch']) if x['regular']
+                         else '(%d @%gpx 不规整)' % (x['n'], x['pitch'])) if x else '-'
+        rows.append([short(g['part']), g['kind'], g['n'],
+                     fmt(c), c['pitch_stdev'] if c else '-',
+                     fmt(r), r['pitch_stdev'] if r else '-',
+                     '%s/%s' % (g.get('filled', '-'), g.get('cells', '-')),
+                     '%g~%g' % tuple(g['item_w']), '%g~%g' % tuple(g['item_h'])])
+    table(rows, ['part', 'kind', 'n', 'cols', 'c-sd', 'rows', 'r-sd',
+                 'filled/cells', 'item w', 'item h'],
+          ['<', '<', '>', '<', '>', '<', '>', '<', '<', '<'])
+    if a.part:
+        for g in grids:
+            if short(g['part']) != a.part:
+                continue
+            print('\n%s %s 详情:' % (short(g['part']), g['kind']))
+            for ax in ('cols', 'rows'):
+                if ax in g:
+                    x = g[ax]
+                    print('  %s n=%d pitch=%g stdev=%g regular=%s'
+                          % (ax, x['n'], x['pitch'], x['pitch_stdev'], x['regular']))
+                    print('     中心 %s' % x['centers'])
+                    print('     起点 %s' % x['starts'])
+    print('\n%d 组栅格。中心分档（非左上角）——同格内元素尺寸可不同仍算同列。'
+          '带「不规整」的轴步距方差超阈值，只作参考不要照抄。' % len(grids))
+
+
+def cmd_clusters(a, outdir):
+    d = load(outdir, 'extract.json')
+    cl = d.get('media_clusters') or []
+    if a.multi:
+        cl = [c for c in cl if c['member_n'] > 1]
+    rows = []
+    for c in cl:
+        b = (c.get('boxes') or [{}])[0]
+        rows.append([c['content_id'], c['member_n'], c['n'],
+                     'Y' if c.get('sha256_identical') else '',
+                     len(c.get('repeat_fixed') or []),
+                     len(c.get('repeat_fixed_cross_media') or []),
+                     trunc(','.join(short(m) for m in c['members']), 52)])
+    table(rows, ['cid', 'files', 'n', 'sha=', 'rep', 'xrep', 'members'],
+          ['<', '>', '>', '<', '>', '>', '<'])
+    print('\n%d 簇（跨文件同素材 %d 簇）。xrep>0 = 同素材散成多文件却占同一位置。'
+          % (len(d.get('media_clusters') or []),
+             sum(1 for c in (d.get('media_clusters') or []) if c['member_n'] > 1)))
+
+
+def cmd_media(a, outdir):
+    d = load(outdir, 'extract.json')
+    rows = []
+    for m in d['media']:
+        if a.candidate and not m.get('candidate'):
+            continue
+        rows.append([short(m['media']), m['ext'], m['bytes'],
+                     'Y' if m.get('candidate') else '', m.get('used_n', 0),
+                     short(m.get('out')) if m.get('exported') else '-',
+                     short(m.get('compressed_out')) or '-',
+                     m.get('compressed_bytes') or '-',
+                     trunc(','.join(m.get('reasons') or []), 34)])
+    table(rows, ['media', 'ext', 'bytes', 'cand', 'used', 'out', 'compressed',
+                 'comp_bytes', 'reasons'],
+          ['<', '<', '>', '<', '>', '<', '<', '>', '<'])
+    print('\n%d 个 media，导出 %d，转码 %d。压缩图 = compressed 列那份，原图 = out 列。'
+          % (len(d['media']), d['counts'].get('media_exported', 0),
+             d['counts'].get('media_transcoded', 0)))
+
+
+def cmd_slides(a, outdir):
+    d = load(outdir, 'extract.json')
+    rows = []
+    for s in d.get('slides') or []:
+        bg = s.get('background') or {}
+        col = (bg.get('color') or {}).get('resolved')
+        rows.append([short(s['part']), short(s.get('layout')),
+                     bg.get('type') or (bg.get('source') or '-'),
+                     col or '-',
+                     len(bg.get('stops') or []) or '-'])
+    table(rows, ['slide', 'layout', 'bg type', 'bg color', 'stops'])
+    print('\n%d 页实例页' % len(d.get('slides') or []))
+
+
+def cmd_layouts(a, outdir):
+    d = load(outdir, 'extract.json')
+    rows = []
+    for l in d['layouts']:
+        bg = l.get('background') or {}
+        rows.append([short(l['part']), trunc(l.get('name'), 26), l.get('type_attr'),
+                     short(l.get('master')), l.get('used_by_slides'), l.get('shape_n'),
+                     trunc(','.join('%s:%s' % kv for kv in (l.get('placeholders') or {}).items()), 30),
+                     (bg.get('color') or {}).get('resolved') or bg.get('type') or '-'])
+    table(rows, ['layout', 'name', 'type', 'master', 'used', 'shapes', 'placeholders', 'bg'],
+          ['<', '<', '<', '<', '>', '>', '<', '<'])
+    print('\n%d 个版式；form_hint=%s（%s）'
+          % (len(d['layouts']), d['form_hint']['form'], d['form_hint']['note']))
+
+
+
+def cmd_summary(a, outdir):
+    d = load(outdir, 'extract.json')
+    fh = d.get('form_hint') or {}
+    tt = d.get('theme_topology') or {}
+    counts = d.get('counts') or {}
+    cv = d.get('canvas') or {}
+    print('canvas    : %s px   源 %s EMU' % (cv.get('px'), (cv.get('source') or {}).get('cx')))
+    print('form      : %s  (%s)' % (fh.get('form'), trunc(fh.get('note'), 72)))
+    print('themes    : %s  mechanism=%s' % (tt.get('themes'), tt.get('mechanism')))
+    print('slides    : %s   layouts %s / masters %s' % (
+        counts.get('slides'), counts.get('layouts'), counts.get('masters')))
+    print('colors    : %s 条   fonts %s 族   images %s' % (
+        len(d.get('color_freq') or []), len(d.get('font_families') or []),
+        len(d.get('images') or [])))
+    multi = [c for c in d.get('media_clusters') or [] if len(c.get('members') or []) > 1]
+    if multi:
+        print('clusters  : %d 个跨文件同素材簇（query.py clusters --multi 细看）' % len(multi))
+    warn = 0
+    unresolved = sum(c.get('n', 0) for c in d.get('color_freq') or []
+                     if 'UNRESOLVED' in str(c.get('resolved', '')))
+    if unresolved:
+        warn += 1
+        print('⚠ UNRESOLVED 色 %d 次——schemeClr 解析有残留，查 ref/ 定位，不能带进语义层' % unresolved)
+    if not d.get('pillow_available', True):
+        warn += 1
+        print('⚠ Pillow 不可用——聚类/取色/转码已降级（%s），产物记 gaps' % d.get('content_cluster_mode'))
+    if d.get('scheme_fallback'):
+        print('note: scheme_fallback 触发 %d 次（rels 断链回退，详见 extract.json）' % len(d['scheme_fallback']))
+    if not warn:
+        print('OK: 无阻塞信号，可进语义层')
+    return 0
+
+
+
+def cmd_get(a, outdir):
+    """按点路径取 extract.json 任意字段，如 get canvas / get theme_topology.mechanism"""
+    d = load(outdir, 'extract.json')
+    cur = d
+    for seg in a.path.split('.'):
+        if isinstance(cur, list):
+            cur = cur[int(seg)]
+        elif isinstance(cur, dict):
+            if seg not in cur:
+                raise SystemExit('get: 无字段 %s（同级键: %s）' % (seg, ', '.join(sorted(cur)[:20])))
+            cur = cur[seg]
+        else:
+            raise SystemExit('get: %s 已是标量' % seg)
+    print(json.dumps(cur, ensure_ascii=False, indent=1)[:20000])
+
+
+def main(argv=None):
+    ap = argparse.ArgumentParser(description=__doc__,
+                                 formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument('outdir')
+    sub = ap.add_subparsers(dest='cmd', required=True)
+
+    p = sub.add_parser('shapes')
+    p.add_argument('--part')
+    p.add_argument('--kind', choices=sorted(KIND_GROUPS))
+    p.add_argument('--ph', action='store_true')
+    p.add_argument('--limit', type=int, default=200)
+    p.set_defaults(fn=cmd_shapes)
+
+    p = sub.add_parser('colors')
+    p.add_argument('--top', type=int, default=40)
+    p.add_argument('--class', dest='klass', choices=('design', 'editor', 'aux'))
+    p.set_defaults(fn=cmd_colors)
+
+    p = sub.add_parser('fonts')
+    p.add_argument('--all', action='store_true')
+    p.set_defaults(fn=cmd_fonts)
+
+    p = sub.add_parser('text-scale')
+    p.add_argument('--top', type=int, default=60)
+    p.set_defaults(fn=cmd_text_scale)
+
+    p = sub.add_parser('images')
+    p.add_argument('--fullscreen', action='store_true')
+    p.add_argument('--repeat', action='store_true')
+    p.add_argument('--top', type=int, default=60)
+    p.set_defaults(fn=cmd_images)
+
+    p = sub.add_parser('clusters')
+    p.add_argument('--multi', action='store_true')
+    p.set_defaults(fn=cmd_clusters)
+
+    p = sub.add_parser('media')
+    p.add_argument('--candidate', action='store_true')
+    p.set_defaults(fn=cmd_media)
+
+    sub.add_parser('summary').set_defaults(fn=cmd_summary)
+
+    p = sub.add_parser('get')
+    p.add_argument('path')
+    p.set_defaults(fn=cmd_get)
+    p = sub.add_parser('recipes')
+    p.add_argument('--min', type=int, default=2)
+    p.add_argument('--top', type=int, default=20)
+    p.add_argument('--all', action='store_true')
+    p.set_defaults(fn=cmd_recipes)
+
+    p = sub.add_parser('grids')
+    p.add_argument('--part')
+    p.set_defaults(fn=cmd_grids)
+
+    sub.add_parser('slides').set_defaults(fn=cmd_slides)
+    sub.add_parser('layouts').set_defaults(fn=cmd_layouts)
+
+    a = ap.parse_args(argv)
+    a.fn(a, os.path.abspath(a.outdir))
+    return 0
+
+
+if __name__ == '__main__':
+    sys.exit(main())