minimal-agent 0.8.2 → 0.10.0

package/skills/ui-design/scripts/image_prompt_rewrite.py

@@ -0,0 +1,713 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""Rewrite PPT image prompts so generated images stay page-bound.
+
+This script is intentionally deterministic. It does not invent design content;
+it binds each image plan item to the current user request, page takeaway,
+composition role, VI tone, and source-image plan. The goal is to prevent generic
+"sleek device / product photography" prompts from producing random monitors,
+controllers, bottles, or other unrelated objects.
+
+生产者-闸门契约（单一词表真相源，治「机器自产自拦」halt）
+========================================================
+本脚本是 t2i prompt 的**唯一生产/净化点**，下游 `image_plan_lint.py` 只做兜底验证。
+因此本脚本必须保证产出**必过** lint：直接 import lint 的黑名单
+（STRUCTURAL/BAD_AESTHETIC/BACKGROUND_TEXT_MARKER/GENERIC_OBJECT_RISK）作为
+`RISKY_TERMS`——lint 加词，这里自动跟上，永不再出现「rewrite 拼进 design 字段的
+中文词（如页 visual_focus 里的『行业缩略图』、页标题里的『架构/流程』）被 lint
+拦下 → 整条工作流确定性 halt」的死端（2026-07-01 真机事故）。
+落点三层：① anchor/标题/takeaway/visual_focus 等**机器拼接段**先剥 RISKY 词；
+② LLM 原 prompt 命中 RISKY 词 → 强制重组成安全的页面绑定 prompt（= yaml 声明的
+「结构图回退」语义：结构交 HTML 层，AI 图退化为安全底图，确定性完成不必等人修）；
+③ 最终 prompt 再整体剥一遍作闭环兜底。i2i（客户图增强）提示词不经此净化——
+lint 对 image-to-image 本就豁免，保真意图不同。
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import sys
+from pathlib import Path
+from typing import Any
+
+# 同目录直跑（python3 .../image_prompt_rewrite.py）时 sys.path[0] 即脚本目录；
+# 被 importlib/测试等其它方式加载时补一道 __file__ 目录守卫，保证 lint 可 import。
+_SCRIPTS_DIR = str(Path(__file__).resolve().parent)
+if _SCRIPTS_DIR not in sys.path:
+    sys.path.insert(0, _SCRIPTS_DIR)
+import image_plan_lint as _lint  # noqa: E402
+
+# lint 黑名单的并集 = 本脚本禁止在 t2i prompt 里出现的词（单一真相源在 lint）。
+RISKY_TERMS = (
+    list(_lint.STRUCTURAL_TERMS)
+    + list(_lint.BAD_AESTHETIC_TERMS)
+    + list(_lint.BACKGROUND_TEXT_MARKER_TERMS)
+    + list(_lint.GENERIC_OBJECT_RISK_TERMS)
+)
+
+
+def _find_risky(text: str) -> list[str]:
+    """text 里非否定式命中的 lint 黑名单词（复用 lint 的否定式判定）。"""
+    return _lint._find_non_negated_terms(text, RISKY_TERMS)
+
+
+def _strip_risky_terms(text: str) -> str:
+    """从 text 里剥掉所有非否定式命中的 RISKY 词（用于机器拼接段与最终闭环兜底）。"""
+    text = str(text or "")
+    guard = 0
+    hits = _find_risky(text)
+    while hits and guard < 6:
+        for term in hits:
+            text = re.sub(re.escape(term), " ", text, flags=re.IGNORECASE)
+        text = re.sub(r"[ \t]{2,}", " ", text)
+        hits = _find_risky(text)
+        guard += 1
+    return text.strip()
+
+
+PAGE_FILE_RE = re.compile(r"ai_p(\d+)(?:_fg)?\.(?:png|jpe?g|webp)$", re.IGNORECASE)
+
+GENERIC_RISK_TERMS = [
+    "professional product photography",
+    "sleek device",
+    "futuristic screen",
+    "generic monitor",
+    "consumer electronics",
+    "game controller",
+    "cosmetic bottle",
+    "medicine bottle",
+    "shampoo bottle",
+]
+
+BACKGROUND_TEXT_MARKER_RE = re.compile(
+    r"(?:章节|序号|编号|号码|数字|标号|标签|题注|文字|文案|1\s*/\s*2\s*/\s*3\s*/\s*4|"
+    r"\b(?:number|numbers|numbered|chapter|chapters|label|labels|caption|captions|text|words|letters)\b)",
+    re.I,
+)
+
+FOREGROUND_STRUCTURE_RE = re.compile(
+    r"(?:\d+\s*张|\d+\s*个|\d+\s*行|\d+\s*列|\d+\s*组|横排|竖排|并排|中部|顶部|底部|左侧|右侧|"
+    r"大字|主标|文字|文案|takeaway|标签|编号|序号|图表|柱图|柱状|饼图|折线|表格|数据表|架构|流程|对比图|"
+    r"时间线|时间轴|箭头|胶囊|节点|连线|\bSVG\b|"
+    # 2026-07-02 真机事故（ai_p23）：design visual_focus 里的「三组'一'短标」漏过词表 →
+    # 生图模型照画三条带字药丸横在背景正中、与 HTML 标题叠字。带字元件类词全部拒收。
+    r"短标|角标|徽标|字样|清单|列表|药丸|"
+    r"comparison chart|chart|diagram|timeline|arrow|arrows|capsules?|nodes?|labels?|text|headline|"
+    r"badges?|chips?|pills?|list items?)",
+    re.I,
+)
+
+NEGATIVE_OBJECTS = (
+    "no game controller, no cosmetic bottle, no medicine bottle, no shampoo bottle, "
+    "no generic consumer electronics, no random monitor, no keyboard, no UI mockup, "
+    "no text, no logo, no letters, no watermark"
+)
+
+CHINESE_ANCHOR_RE = re.compile(r"[\u4e00-\u9fff][\u4e00-\u9fffA-Za-z0-9·+\-/]{1,11}")
+ENGLISH_ANCHOR_RE = re.compile(r"\b[A-Za-z][A-Za-z0-9+\-/]{3,}\b")
+ANCHOR_STOPWORDS = {
+    "主题", "指令", "用户", "要求", "生成", "做成", "页面", "背景", "图片", "文字", "内容", "设计",
+    "视觉", "风格", "中文", "文案", "标题", "素材", "客户", "产品", "场景", "重点", "需要",
+    "必须", "用于", "高清", "比例", "信息", "专业", "方案", "系统", "能力", "价值", "结构",
+    "用户原始输入与目标", "主题/指令", "文件作为基础", "生成一套专业中文", "专业中文",
+    "presentation", "visual", "image", "slide", "title", "core", "message", "premium",
+    "product", "detail", "page", "background", "foreground", "style", "clean", "high",
+    "resolution", "text", "safe", "zone",
+}
+
+
+def _load_json(path: Path, default: Any) -> Any:
+    if not path.exists():
+        return default
+    try:
+        return json.loads(path.read_text(encoding="utf-8"))
+    except Exception:
+        return default
+
+
+def _load_text(path: Path) -> str:
+    if not path.exists():
+        return ""
+    return path.read_text(encoding="utf-8", errors="replace")
+
+
+def _text_of(value: Any) -> str:
+    if value is None:
+        return ""
+    if isinstance(value, str):
+        return value
+    if isinstance(value, (int, float, bool)):
+        return str(value)
+    if isinstance(value, list):
+        return " ".join(_text_of(v) for v in value)
+    if isinstance(value, dict):
+        return " ".join(_text_of(v) for v in value.values())
+    return str(value)
+
+
+def _add_anchor(out: list[str], value: str) -> None:
+    value = value.strip(" ，,。.;；:：()（）[]【】<>《》\"'")
+    if len(value) < 2:
+        return
+    low = value.lower()
+    if low in ANCHOR_STOPWORDS:
+        return
+    if FOREGROUND_STRUCTURE_RE.search(value):
+        return
+    # 拒收含 lint 黑名单词的 anchor（如「行业缩略图」）：拼进 prompt 会被下游闸门拦。
+    if _find_risky(value):
+        return
+    if value not in out:
+        out.append(value)
+
+
+def _extract_anchors(*values: Any, limit: int = 12) -> list[str]:
+    text = " ".join(_text_of(v) for v in values)
+    anchors: list[str] = []
+    for match in CHINESE_ANCHOR_RE.findall(text):
+        _add_anchor(anchors, match)
+        if len(anchors) >= limit:
+            return anchors
+    for match in ENGLISH_ANCHOR_RE.findall(text):
+        _add_anchor(anchors, match)
+        if len(anchors) >= limit:
+            return anchors
+    return anchors
+
+
+def _has_anchor(text: str, anchors: list[str]) -> bool:
+    hay = text.lower()
+    return any(anchor.lower() in hay for anchor in anchors)
+
+
+def _has_generic_risk(text: str) -> bool:
+    hay = text.lower()
+    return any(term in hay for term in GENERIC_RISK_TERMS)
+
+
+def _page_map(design: dict[str, Any]) -> dict[int, dict[str, Any]]:
+    pages: dict[int, dict[str, Any]] = {}
+    for page in design.get("pages", []) if isinstance(design, dict) else []:
+        if not isinstance(page, dict):
+            continue
+        try:
+            pages[int(page.get("page"))] = page
+        except Exception:
+            continue
+    return pages
+
+
+def _page_no_from_file(file_name: str) -> int | None:
+    match = PAGE_FILE_RE.search(file_name)
+    if not match:
+        return None
+    try:
+        return int(match.group(1))
+    except ValueError:
+        return None
+
+
+def _page_for_item(item: dict[str, Any], pages: dict[int, dict[str, Any]]) -> dict[str, Any] | None:
+    try:
+        page_no = int(item.get("page"))
+        if page_no in pages:
+            return pages[page_no]
+    except Exception:
+        pass
+    file_name = str(item.get("file") or "")
+    page_no = _page_no_from_file(file_name)
+    if page_no in pages:
+        return pages[page_no]
+    source_file = str(item.get("source_file") or "")
+    needles = [v for v in (file_name, source_file) if v]
+    if needles:
+        for page in pages.values():
+            blob = json.dumps(page, ensure_ascii=False)
+            if any(n in blob for n in needles):
+                return page
+    return None
+
+
+def _deck_meta_anchor(deck_meta: dict[str, Any] | None) -> dict[str, str]:
+    """从 design.json 顶层 deck_meta 提取产品外观/客户身份锚，剥净 RISKY 词后返回。
+
+    2026-07-07 真机 p9 复盘：页面文案省略主语（"拆除 WiFi 模块"不重复"全景相机"）时，
+    纯靠该页 title/takeaway 提取的 anchors 完全抓不到"这是客户自己的产品"这层信息，
+    t2i 生图凭空画出无关设备（通用电路板/工控盒），而非客户真实产品外观。deck_meta
+    是设计阶段读全部 vision.json 后**只写一次**的深度信息，这里读出来供每页 t2i 复用，
+    不依赖任何单页文案是否显式提及产品名。"""
+    deck_meta = deck_meta if isinstance(deck_meta, dict) else {}
+    product = _strip_risky_terms(str(deck_meta.get("product_identity") or "")).strip()
+    customer = _strip_risky_terms(str(deck_meta.get("customer_context") or "")).strip()
+    return {"product_identity": product, "customer_context": customer}
+
+
+def _page_context(
+    page: dict[str, Any] | None,
+    user_request: str,
+    deck_meta: dict[str, Any] | None = None,
+) -> dict[str, str]:
+    page = page or {}
+    title = str(page.get("title") or "")
+    takeaway = str(page.get("one_sentence_takeaway") or page.get("key_point") or title)
+    visual_focus = str(page.get("visual_focus") or page.get("visual_story") or "")
+    grammar = str(page.get("page_grammar") or page.get("composition_style") or "")
+    text_zone = str(page.get("safe_text_zone") or page.get("text_overlay_zone") or page.get("text_zone") or "left")
+    visual_role = str(page.get("visual_role") or page.get("primary_visual_type") or "")
+    vi_style = str(
+        page.get("vi")
+        or page.get("style")
+        or page.get("visual_style")
+        or page.get("theme")
+        or ""
+    )
+    anchors = _extract_anchors(
+        title,
+        takeaway,
+        visual_role,
+        page.get("visual_story"),
+        page.get("semantic_image_prompt"),
+        page.get("visual"),
+        page.get("blocks"),
+        user_request,
+    )
+    anchor_text = ", ".join(anchors[:8]) or "the user-provided topic and page-specific subject"
+    ctx = {
+        "title": title,
+        "takeaway": takeaway,
+        "visual_focus": visual_focus,
+        "grammar": grammar,
+        "text_zone": text_zone,
+        "visual_role": visual_role,
+        "vi_style": vi_style,
+        "user_request": user_request[:800],
+        "anchors": anchor_text,
+    }
+    ctx.update(_deck_meta_anchor(deck_meta))
+    return ctx
+
+
+# 含数字/金额/百分比的锚点碎片（如「核心总投入约」「3 万元」）：文字渲染型生图模型（Seedream 类）
+# 会把它们照画成假图表标注——背景/增强层锚点一律剔除。
+DIGITY_ANCHOR_RE = re.compile(r"[0-9０-９%％]|万元|亿|千万")
+# 句子碎片判定：CHINESE_ANCHOR_RE 从 takeaway 整句里切出来的片段常含功能词（的/把/被/约/相当于…），
+# 或顶到 12 字上限（截断即非名词）；场景名词锚不该有这些形态。
+SENTENCE_FRAGMENT_RE = re.compile(r"[的了是把被让将约]|相当于|等于|成为|沦为")
+ANCHOR_MAX_NOUN_LEN = 11
+
+
+def _scene_anchor_text(anchor_text: str) -> str:
+    """背景/i2i 层锚点净化：只留场景名词类短语（去数字/金额与句子碎片、限 6 个），空了给通用兜底。
+    页面绑定（anti-generic）靠这些名词，不靠标题/正文原句——原句会被模型当文字画进图里。"""
+    kept = []
+    for a in (anchor_text or "").split(","):
+        a = a.strip()
+        if not a or DIGITY_ANCHOR_RE.search(a) or SENTENCE_FRAGMENT_RE.search(a):
+            continue
+        # 12 字上限截断只发生在 CJK 整句切片上（英文名词天然带词界，不设限）。
+        if re.search(r"[一-鿿]", a) and len(a) > ANCHOR_MAX_NOUN_LEN:
+            continue
+        kept.append(a)
+        if len(kept) >= 6:
+            break
+    return ", ".join(kept) or "the user-provided topic and page-specific scene"
+
+
+def _style_tone(ctx: dict[str, str]) -> str:
+    vi = ctx.get("vi_style") or "the selected VI palette and current user style request"
+    return (
+        f"premium product-detail page visual, cinematic but restrained, clean lighting, "
+        f"high resolution, coherent with {vi}, not decorative clutter"
+    )
+
+
+def _safe_background_focus(text: str) -> str:
+    """Background prompts must not ask the image model to draw labels/numbers.
+
+    Editable PPT text, chapter marks, labels, and numbers belong to HTML/SVG
+    foreground layers. If these words leak into a background image prompt, image
+    models reliably bake non-editable symbols into the bitmap.
+    """
+    text = str(text or "").strip()
+    if not text:
+        return "clean page-relevant atmosphere with clear visual direction"
+    if (
+        BACKGROUND_TEXT_MARKER_RE.search(text)
+        or FOREGROUND_STRUCTURE_RE.search(text)
+        or _find_risky(text)
+    ):
+        return "low-interference complementary background atmosphere with clean negative space, not foreground labels or charts"
+    return text
+
+
+def _is_data_or_html_page(page: dict[str, Any] | None) -> bool:
+    if not isinstance(page, dict):
+        return False
+    text = " ".join(
+        str(page.get(k) or "")
+        for k in (
+            "page_grammar",
+            "composition_style",
+            "visual_role",
+            "primary_visual_type",
+            "semantic_image_prompt",
+        )
+    ).lower()
+    return any(term in text for term in ("data-visual", "html-svg", "chart", "axis", "diagram", "架构", "流程", "图表"))
+
+
+def _product_identity_clause(ctx: dict[str, str]) -> str:
+    """deck 级产品外观/客户身份锚——独立于本页文案是否显式提及产品名（见 _deck_meta_anchor）。
+    空值时返回空字符串，不影响没有真实产品照的示意型 deck。"""
+    product = ctx.get("product_identity") or ""
+    customer = ctx.get("customer_context") or ""
+    if not product and not customer:
+        return ""
+    parts = []
+    if product:
+        parts.append(
+            f"Product visual consistency (this deck's real product — match its actual appearance, "
+            f"never invent a different-looking device): {product}."
+        )
+    if customer:
+        parts.append(f"Customer/industry context: {customer}.")
+    return " " + " ".join(parts)
+
+
+def _background_prompt(ctx: dict[str, str], data_page: bool) -> str:
+    # 🚫 背景层提示词**绝不携带页标题/正文原句/数字/图表描述**（2026-07-03 ai_p9 真机事故：
+    # "Slide title: 投入有限，对冲无限. Core message: 核心总投入约 3 万元… Visual focus: 成本拆解 HTML 表 +
+    # 投入/对冲对比柱状图" 被 Seedream 类文字渲染型模型照着画出一整版烘字假幻灯片，与 HTML 真卡互相遮挡）。
+    # 页面绑定（anti-generic）只靠去数字化的场景名词锚；表格/图表/标题由 HTML 层承载，背景只管氛围与留白。
+    if data_page:
+        subject = (
+            "low-interference data-space background, subtle material depth, "
+            "large clean negative space reserved for the editable HTML/SVG data graphics drawn on top"
+        )
+        visual_focus = "pure atmosphere only; all data visuals are drawn by the HTML layer, not in this image"
+    else:
+        subject = (
+            "realistic page-specific scene or product environment derived from the slide topic, "
+            "visual focus supports the slide message without adding unrelated objects"
+        )
+        visual_focus = _safe_background_focus(ctx["visual_focus"])
+    return (
+        f"{subject}. Scene inspiration keywords (mood reference only — never render these words or any writing): "
+        f"{_scene_anchor_text(ctx['anchors'])}. Visual focus: {visual_focus}."
+        f"{_product_identity_clause(ctx)} "
+        f"Text safe zone: {ctx['text_zone']} — keep it as large clean negative space. "
+        f"{_style_tone(ctx)}. absolutely no text of any kind, no chinese characters, no digits, "
+        f"no charts, no tables, no diagrams, no slide-like layout, {NEGATIVE_OBJECTS}."
+    )
+
+
+def _foreground_prompt(ctx: dict[str, str]) -> str:
+    return (
+        "single complete page-specific product, object, or scene detail as a foreground cutout, "
+        f"matching slide message: {_strip_risky_terms(ctx['takeaway'])}. Topic anchors: {ctx['anchors']}."
+        f"{_product_identity_clause(ctx)} "
+        "transparent or clean dark background, complete subject, no cropping through the main body, "
+        f"{_style_tone(ctx)}. {NEGATIVE_OBJECTS}."
+    )
+
+
+def _rewrite_t2i(
+    item: dict[str, Any],
+    page: dict[str, Any] | None,
+    user_request: str,
+    deck_meta: dict[str, Any] | None = None,
+) -> tuple[str, bool, str]:
+    original = str(item.get("prompt") or "")
+    ctx = _page_context(page, user_request, deck_meta)
+    data_page = _is_data_or_html_page(page)
+    role = str(item.get("role") or item.get("target_layer") or ctx["visual_role"]).lower()
+    file_name = str(item.get("file") or "").lower()
+    is_foreground_file = "_fg." in file_name
+    is_background = not is_foreground_file and role not in ("foreground", "fg", "cutout")
+
+    force_rewrite = False
+    reason = []
+    if "use the attached image" in original.lower() or "preserve its main subject" in original.lower():
+        force_rewrite = True
+        reason.append("i2i_prompt_in_t2i")
+    anchors = _extract_anchors(
+        page,
+        item.get("note"),
+        item.get("role"),
+        item.get("target_layer"),
+        user_request,
+    )
+    if _has_generic_risk(original) and not _has_anchor(original, anchors):
+        force_rewrite = True
+        reason.append("generic_prompt_without_domain_anchor")
+    if not original.strip() or not _has_anchor(original, anchors):
+        force_rewrite = True
+        reason.append("missing_domain_anchor")
+    if is_background and (
+        BACKGROUND_TEXT_MARKER_RE.search(original)
+        or BACKGROUND_TEXT_MARKER_RE.search(ctx["visual_focus"])
+        or FOREGROUND_STRUCTURE_RE.search(original)
+        or FOREGROUND_STRUCTURE_RE.search(ctx["visual_focus"])
+    ):
+        force_rewrite = True
+        reason.append("background_text_marker_conflict")
+    # 背景 prompt 携带页面文案/图表描述（"Slide title:" 等旧模板遗留、或 LLM 自拼）→ 必须重组：
+    # 下游 lint 的 slide_copy 检查会拦（生产者产出必过闸门契约）；旧 img_plan 复跑也借此收敛到新模板。
+    if is_background and _lint._find_slide_copy_terms(original):
+        force_rewrite = True
+        reason.append("slide_copy_in_background_prompt")
+    # 原 prompt 命中 lint 黑名单（结构图/坏审美/背景文字标记/泛消费品）→ 机器重组成
+    # 安全的页面绑定 prompt——「结构图回退」的确定性实现：结构交 HTML 层，AI 图退化为安全底图。
+    risky = _find_risky(original)
+    if risky:
+        force_rewrite = True
+        reason.append("risky_terms:" + "|".join(str(t) for t in risky[:4]))
+
+    if not is_background:
+        prompt = _foreground_prompt(ctx)
+    else:
+        prompt = _background_prompt(ctx, data_page)
+
+    if force_rewrite:
+        # 闭环兜底：模板为静态安全文案、动态段已剥 RISKY 词；万一跨段拼接仍构成命中，整体再剥一遍，
+        # 保证「rewrite 产出必过 lint」的不变量成立（did_change 按真实差异报，二跑幂等 changed=0）。
+        residual = _find_risky(prompt)
+        if residual:
+            prompt = _strip_risky_terms(prompt)
+        return prompt, prompt != original, ",".join(reason)
+
+    # force_rewrite=False：原 prompt 本身已过审（有锚点/无风险词/无背景文字冲突），但"有锚点"只是指
+    # 页面局部词命中，不代表提过客户真实产品外观——该页文案完全可能没提产品名（见本函数头注真机 p9
+    # 复盘），产品身份锚必须补齐，不能因为原 prompt "看起来还行" 就放过。幂等：锚文本已存在则不重复拼。
+    appended = []
+    identity_clause = _product_identity_clause(ctx)
+    if identity_clause and identity_clause.strip() not in original:
+        original = original.rstrip(" .") + "." + identity_clause
+        appended.append("product_identity_anchor")
+    if NEGATIVE_OBJECTS.lower() not in original.lower():
+        original = original.rstrip(" .") + ". " + NEGATIVE_OBJECTS + "."
+        appended.append("append_negative_constraints")
+    if appended:
+        return original, True, ",".join(appended)
+    return original, False, ""
+
+
+def _source_plan_entry(source_plan: dict[str, Any], source_file: str) -> dict[str, Any]:
+    """按 source_file 到 source_layout_plan.json 的 images[] 里找对应条目；找不到返回 {}。"""
+    if not source_file:
+        return {}
+    for im in (source_plan.get("images") or []) if isinstance(source_plan, dict) else []:
+        if isinstance(im, dict) and str(im.get("file") or "") == source_file:
+            return im
+    return {}
+
+
+def _quote_embedded_text(embedded_text: Any, limit: int = 6) -> str:
+    """把 plan_source_layout.py 算好的 embedded_text 列表格式化成可引用的 "'文字'(位置), ..." 短语。
+    与 plan_source_layout.py::_quote_embedded_text 逐字同构（脚本间不耦合，各自拷贝，同一份格式约定）。"""
+    parts = []
+    for it in (embedded_text or [])[:limit]:
+        text = it.get("text") if isinstance(it, dict) else None
+        if not text:
+            continue
+        region = it.get("region") if isinstance(it, dict) else ""
+        parts.append(f"'{text}'" + (f" ({region})" if region else ""))
+    return ", ".join(parts)
+
+
+# "看起来像真正的 i2i 指令"特征词——同时出现在 plan_source_layout.py::_enhance_prompt 的
+# _PRESERVE 模板和本文件 Tier 3 自己的 preserve 模板里，保证幂等：上一轮重建的输出、或设计 LLM
+# 微调过的版本，下一轮都能被正确识别为"genuine"，不会被误判成需要重建的占位符文本。
+_I2I_SIGNATURE_MARKERS = (
+    "attached image",
+    "attached photo",
+    "reference image",
+    "strict visual reference",
+    "preserve its main subject",
+    "preserve the real main subject",
+)
+_I2I_MIN_LEN = 30  # degenerate-match 防护：真正的 i2i 指令（哪怕被精修过）远长于这个下限
+
+
+def _looks_like_i2i_prompt(text: str) -> bool:
+    """原 prompt 是否已经读起来像真正的 image-to-image 指令（非空、非占位符、非误放的 t2i 文本）。"""
+    text = str(text or "")
+    if len(text.strip()) < _I2I_MIN_LEN:
+        return False
+    low = text.lower()
+    return any(m in low for m in _I2I_SIGNATURE_MARKERS)
+
+
+def _has_embedded_text_removal_clause(text: str, embedded_text: Any, limit: int = 6) -> bool:
+    """text 是否已经完整包含"删除这些嵌入文字"的指令——要求含"remov"词干（真的是删除指令，
+    不是提到了这段文字）且**每一条** embedded_text 的原文都以精确子串形式出现（LLM 微调的是
+    措辞，不会改写要求它删除的引用文字本身，故不做模糊匹配）。"""
+    items = [
+        it for it in (embedded_text or [])[:limit]
+        if isinstance(it, dict) and str(it.get("text") or "").strip()
+    ]
+    if not items:
+        return False
+    if "remov" not in text.lower():
+        return False
+    return all(str(it["text"]).strip() in text for it in items)
+
+
+def _append_deltext_clause(original: str, quoted: str) -> str:
+    """Tier 2：就地打补丁而非重建——在已有 prompt（可能带设计 LLM 的页面定制文字）末尾追加缺失
+    的去字指令，复用 _quote_embedded_text 的引用格式，不触碰 original 其余部分。"""
+    clause = (
+        f" Additionally, the original image has this exact baked-in caption/title text that must be "
+        f"completely REMOVED, leaving clean empty space in its place: {quoted}. Never reproduce or invent any text."
+    )
+    return original.rstrip(" .") + "." + clause
+
+
+def _rewrite_i2i(item: dict[str, Any], source_plan: dict[str, Any], page: dict[str, Any] | None, user_request: str) -> tuple[str, bool, str]:
+    original = str(item.get("prompt") or item.get("enhance_prompt") or "")
+    intent = str(item.get("enhance_intent") or item.get("enhance_style") or "")
+    source_file = str(item.get("source_file") or "")
+    plan_entry = _source_plan_entry(source_plan, source_file)
+    text_policy = str(plan_entry.get("text_policy") or "")
+    embedded_text = plan_entry.get("embedded_text")
+    quoted = _quote_embedded_text(embedded_text)
+
+    # 2026-07-07 架构修复（治"规划算对了、执行没服从"这一大类问题里最隐蔽的一种）：这个函数此前
+    # 对**每一条** i2i 都无条件从写死的三分支模板重新拼一遍 prompt，从不把 original（设计 LLM
+    # 抄自 source_plan.enhance_prompt、且 ppt.yaml 明确授权"按 key_point 微调"过的版本）当作
+    # 返回值的基础——只在最后拿 original 跟新 prompt 做字符串比较判断 changed。这意味着设计 LLM
+    # 做的任何页面定制、以及 plan_source_layout.py 未来对 enhance_prompt 措辞的任何改进，全部
+    # 每轮被无条件推翻重建。i2i 天然对 image_plan_lint 豁免（lint 对 mode=="image-to-image" 处处
+    # 门控跳过），信任 original 的自由文本内容没有 lint 失败风险——这解除了 t2i 那种"必须重建成
+    # 安全模板"的唯一理由。改为三层信任分级，统一到 _rewrite_t2i 同款"已过审就不动，只在命中
+    # 具体检查失败时才动手"的哲学：
+    #   Tier 1（原样保留）：original 已经像真 i2i 指令，且该有的去字引用（若 text_policy 要求）
+    #     已完整存在 → 一个字都不改。
+    #   Tier 2（就地打补丁）：original 可用但缺去字指令（= src_6 事故的确切场景：设计 LLM 抄来的
+    #     版本本身没问题，只是漏了这一条）→ 追加缺失子句，保留其余全部定制文字。
+    #   Tier 3（兜底重建，下面的逻辑原样保留）：original 缺失/不像 i2i 指令 → 走原有模板重建。
+    requires_deltext = bool(quoted) and (text_policy == "remove_all_text" if ("dense" in intent or "faithful" in intent) else True)
+    if original.strip() and _looks_like_i2i_prompt(original):
+        if not requires_deltext or _has_embedded_text_removal_clause(original, embedded_text):
+            return original, False, "kept_as_is"
+        return _append_deltext_clause(original, quoted), True, "patched_missing_deltext"
+
+    # ---- Tier 3：兜底重建（original 缺失/不可信时的安全网，逻辑与此前一致）----
+    ctx = _page_context(page, user_request)
+    preserve = (
+        "Use the attached image as the strict visual reference. Preserve the real main subject and key factual content. "
+    )
+    if "dense" in intent or "faithful" in intent:
+        if text_policy == "remove_all_text" and quoted:
+            # 装饰性 chrome 文字（整版截图/海报自带标题/标签，非图表数据）：保结构去字，引用实抄文字。
+            policy = (
+                "Faithfully keep the exact same layout, sub-images, comparison structure and every non-text visual "
+                "detail unchanged, and keep the original framing and aspect ratio (do not crop, stretch or outpaint). "
+                f"However, the original image also has this exact baked-in caption/title text that must be completely "
+                f"REMOVED, leaving clean empty space in its place: {quoted}. Never reproduce or invent any text. "
+                "Use contain-friendly composition."
+            )
+        else:
+            policy = (
+                "Faithfully restore clarity only; do not rewrite charts, screenshots, numbers, labels, UI structure or data. "
+                "Use contain-friendly composition, no invented text."
+            )
+    elif "cutout" in intent or "remove_background" in intent:
+        policy = (
+            "Create a clean foreground cutout with transparent or plain background; remove original colored/gradient background, "
+            "keep the complete subject, no cropping through devices, no new text or logo."
+        )
+        if quoted:
+            policy += f" The original image has this exact baked-in text that must be completely removed: {quoted}."
+    else:
+        policy = (
+            "Recompose into the target slide ratio with clean negative space for editable PPT text; remove distracting original text, "
+            "do not create a collage or thumbnail grid."
+        )
+        if quoted:
+            policy += f" Specifically, this exact baked-in text must be completely removed: {quoted}."
+    # i2i 同样不拼标题/正文原句（增强图常被用作 hero/背景层，模型会把原句画成新文字；
+    # 内容锚定靠 attached image 本身 + 去数字化场景名词锚即可）。
+    prompt = (
+        f"{preserve}{policy} Scene inspiration keywords (mood reference only — never paint any new words): "
+        f"{_scene_anchor_text(ctx['anchors'])}. Source file: {source_file}. Target visual role: {intent or ctx['visual_role']}. "
+        f"Text safe zone: {ctx['text_zone']}. {_style_tone(ctx)}. no new text, no logo, no letters, no watermark."
+    )
+    if prompt != original:
+        return prompt, True, "rebuilt_fallback"
+    return original, False, ""
+
+
+def rewrite_plan(build: Path) -> dict[str, Any]:
+    plan_path = build / "img_plan.json"
+    plan = _load_json(plan_path, [])
+    if not isinstance(plan, list):
+        raise ValueError("img_plan.json must be an array")
+    design = _load_json(build / "design.json", {})
+    source_plan = _load_json(build / "source_layout_plan.json", {})
+    user_request = _load_text(build / "user_request.md")
+    pages = _page_map(design if isinstance(design, dict) else {})
+    deck_meta = design.get("deck_meta") if isinstance(design, dict) else {}
+
+    changed = []
+    i2i_tier_counts = {"kept_as_is": 0, "patched_missing_deltext": 0, "rebuilt_fallback": 0}
+    for item in plan:
+        if not isinstance(item, dict):
+            continue
+        file_name = str(item.get("file") or "")
+        page = _page_for_item(item, pages)
+        mode = str(item.get("mode") or "text-to-image").lower()
+        before = str(item.get("prompt") or "")
+        if mode == "image-to-image":
+            prompt, did_change, reason = _rewrite_i2i(item, source_plan if isinstance(source_plan, dict) else {}, page, user_request)
+            if reason in i2i_tier_counts:
+                i2i_tier_counts[reason] += 1
+        else:
+            prompt, did_change, reason = _rewrite_t2i(item, page, user_request, deck_meta)
+            item["mode"] = "text-to-image"
+        item["prompt"] = prompt
+        if did_change:
+            changed.append(
+                {
+                    "file": file_name,
+                    "mode": item.get("mode"),
+                    "reason": reason,
+                    "before": before[:240],
+                    "after": prompt[:360],
+                }
+            )
+
+    plan_path.write_text(json.dumps(plan, ensure_ascii=False, indent=2), encoding="utf-8")
+    report = {
+        "changed_count": len(changed),
+        "changed": changed,
+        "i2i_tier_counts": i2i_tier_counts,
+    }
+    (build / "image_prompt_rewrite_report.json").write_text(
+        json.dumps(report, ensure_ascii=False, indent=2),
+        encoding="utf-8",
+    )
+    lines = ["# 生图提示词重写报告", "", f"- changed_count: {len(changed)}", ""]
+    for entry in changed:
+        lines.append(f"## {entry['file']}")
+        lines.append(f"- mode: {entry['mode']}")
+        lines.append(f"- reason: {entry['reason']}")
+        lines.append(f"- after: {entry['after']}")
+        lines.append("")
+    (build / "image_prompt_rewrite_report.md").write_text("\n".join(lines).rstrip() + "\n", encoding="utf-8")
+    return report
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Rewrite PPT image prompts with page/domain/style anchors.")
+    parser.add_argument("--build", default="./build", help="Build directory")
+    args = parser.parse_args()
+    report = rewrite_plan(Path(args.build))
+    print(f"IMAGE_PROMPT_REWRITE_OK changed={report['changed_count']}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())