{
  "title": {
    "en": "The same shop, two ways: Codex alone and Codex with Laydyne",
    "ja": "同じ店舗を2通りで：Codexだけと、Codex＋Laydyne"
  },
  "note": {
    "en": "Steps as recorded in the T1 runs of 2026-10-06 (docs/verification/mcp-efficiency). Token figures are means of two runs; see facts.json.",
    "ja": "2026-10-06のT1の実行記録どおりの手順（docs/verification/mcp-efficiency）。トークン数は2回の平均で、出典はfacts.json。"
  },
  "steps": [
    {
      "id": "brief",
      "actor": "user",
      "title": {
        "en": "One brief for both",
        "ja": "同じ依頼を両方へ"
      },
      "detail": {
        "en": "A 12 × 9 m shop with an entrance in the south wall, a counter and a display table near it, two shelf rows with a 1.2 m aisle, every passage at least 0.9 m. Deliver a dimensioned plan.svg, a 3D view and a short report.",
        "ja": "南の壁に入口がある12 × 9 mの店舗。入口近くにカウンターと展示台、通路1.2 mで2列の棚、すべての通路0.9 m以上。寸法入りのplan.svg、3Dの表示、短い報告を渡す。"
      },
      "tools": [],
      "output": {
        "en": "The same task text, plus how to deliver in each setup",
        "ja": "同じ課題文と、それぞれの環境での渡し方"
      }
    },
    {
      "id": "a-write",
      "actor": "agent",
      "title": {
        "en": "A · Codex alone writes a generator",
        "ja": "A・Codexだけ：生成プログラムを書く"
      },
      "detail": {
        "en": "Codex writes one Python script that holds the layout, checks overlaps and passage widths itself and writes model.json, plan.svg and a three.js page.",
        "ja": "Codexが配置を持つPythonスクリプトを1本書き、重なりと通路幅を自分で確かめ、model.json・plan.svg・three.jsのページを書き出す。"
      },
      "tools": [
        "shell",
        "apply_patch"
      ],
      "output": {
        "en": "model.json, plan.svg, view3d.html (about 4 model requests)",
        "ja": "model.json・plan.svg・view3d.html（モデルへの要求は約4回）"
      }
    },
    {
      "id": "b-discover",
      "actor": "agent",
      "title": {
        "en": "B · Codex looks up the Laydyne tools",
        "ja": "B・CodexがLaydyneの道具を調べる"
      },
      "detail": {
        "en": "Codex lists the Laydyne tools and reads the definitions of the ones it needs. It puts the server instructions in front of each of the 67 tool texts (161,345 tokens in all, cut to about 10,000 when printed).",
        "ja": "CodexがLaydyneの道具を一覧し、使う道具の定義を読む。サーバーの指示文は67個すべての道具の説明の前に繰り返される（合計161,345トークン、表示時は約1万に切り詰め）。"
      },
      "tools": [],
      "output": {
        "en": "The calls it will make",
        "ja": "これから呼ぶ道具の使い方"
      }
    },
    {
      "id": "b-build",
      "actor": "laydyne",
      "title": {
        "en": "B · Build in the open Studio tab",
        "ja": "B・開いているStudioで作る"
      },
      "detail": {
        "en": "get_space for the revision, create_space or edit_space with parts by kind; the user sees each change live.",
        "ja": "get_spaceでrevisionを読み、create_spaceかedit_spaceで部品を種類ごとに置く。利用者は変更をその場で見る。"
      },
      "tools": [
        "get_space",
        "get_part_catalog",
        "create_space",
        "edit_space"
      ],
      "output": {
        "en": "The shop as an editable model",
        "ja": "編集できるモデルとしての店舗"
      }
    },
    {
      "id": "b-check",
      "actor": "laydyne",
      "title": {
        "en": "B · Check and look",
        "ja": "B・確かめて見る"
      },
      "detail": {
        "en": "check_space finds overlaps and narrow passages; get_view_image draws the scaled 2D plan (SVG with dimensions and aisle widths) and a 3D picture.",
        "ja": "check_spaceで重なりと狭い通路を探し、get_view_imageで縮尺付き2D平面図（寸法と通路幅入りのSVG）と3Dの画像を描く。"
      },
      "tools": [
        "check_space",
        "get_view_image"
      ],
      "output": {
        "en": "plan.svg saved by the agent, a 3D picture (about 11 model requests in all)",
        "ja": "エージェントが保存したplan.svgと3Dの画像（モデルへの要求は合計約11回）"
      }
    },
    {
      "id": "compare",
      "actor": "user",
      "title": {
        "en": "Same quality, different cost",
        "ja": "品質は同じ、費用は別"
      },
      "detail": {
        "en": "Both passed the same scripted checks in every run. Codex alone used about 95,000 input tokens; with Laydyne about 481,000, mostly because each of its extra model requests re-sends Codex's own context of about 21,000 tokens.",
        "ja": "どの実行も同じ採点スクリプトに合格。入力トークンはCodexだけで約9.5万、Laydyneありで約48.1万。主な理由は、Laydyneありの方がモデルへの要求が多く、その都度Codex自身の約2.1万トークンの文脈を送り直すため。"
      },
      "tools": [],
      "output": {
        "en": "Choose by what you need besides the files: a shared live model, checks, drawing sheets, simulation",
        "ja": "ファイル以外に何が要るかで選ぶ：共有できる生きたモデル、検査、図面シート、シミュレーション"
      }
    }
  ]
}