feat(director): closed style enum + YAML template mapping + caption

- style field now closed enum: tech-digital/business-finance/news-insight/lifestyle/education
- Add style_variant for free-form variant description
- Move style→template mapping from Python hardcode to config/template-mapping.yaml
- Support exact/aliases/prefixes matching in YAML config
- Add caption field to placements (reader-visible, replaces internal purpose)
- Validate style against ALLOWED_VISUAL_STYLES from config
- Add validate_chat_article_markdown() (reserved for future use)
- Remove SequenceMatcher similarity check from de-ai validator

Author: PPPPanda

config/template-mapping.yaml

@@ -2,6 +2,29 @@
 # Pipeline Formatter 阶段根据话题分类自动匹配模板
 
 wechat:
+  styles:
+    exact:
+      tech-digital: tech-digital.html
+      business-finance: business-finance.html
+      news-insight: news-insight.html
+      lifestyle: lifestyle.html
+      education: education.html
+    aliases:
+      tech-flat: tech-digital.html
+      tech-minimal: tech-digital.html
+      tech-satirical: tech-digital.html
+      tech-editorial: tech-digital.html
+      cyberpunk: tech-digital.html
+      watercolor: lifestyle.html
+      chinese-ink: lifestyle.html
+      business: business-finance.html
+      news: news-insight.html
+    prefixes:
+      tech: tech-digital.html
+      business: business-finance.html
+      news: news-insight.html
+      edu: education.html
+      life: lifestyle.html
   mapping:
     # 科技类
     - keywords: ["科技", "AI", "人工智能", "数码", "编程", "互联网", "区块链", "Web3", "芯片", "手机", "电脑", "软件", "开源"]

prompts/art-director.md

@@ -23,8 +23,9 @@
 
 ### 1. 确定全局视觉风格
 
-根据文章主题和平台，选择或定义：
-- `style`：风格标识（如 tech-flat / watercolor / cyberpunk / chinese-ink）
+根据文章主题和平台，输出：
+- `style`：**封闭枚举，只能从以下值中选择一个**：`tech-digital` / `business-finance` / `news-insight` / `lifestyle` / `education`
+- `style_variant`：该模板家族下的自由变体说明（例如 satirical / flat / editorial / clean / magazine）
 - `global_tone`：一段话描述整体视觉基调（配色、氛围、密度等）
 
 ### 2. 先设计封面，再规划正文配图
@@ -55,7 +56,8 @@
 | 字段 | 要求 |
 |------|------|
 | `description` | **详细的画面描述**，50-150字。包含主体、构图、背景、氛围、配色、要避免的元素。必须详细到足以让人类理解你想要什么画面。 |
-| `purpose` | 这张图在文章叙事中的作用（1-2句话） |
+| `purpose` | 这张图在文章叙事中的作用（内部字段，给系统/编辑看，不直接给读者看）（1-2句话） |
+| `caption` | **给读者看的短图注**，1句话，尽量自然、像编辑写的，不要解释“这张图在干什么”。禁止使用“具象化/可视化/强化/呼应/增强可读性/给读者一个锚点”等解释性 AI 话语。 |
 | `type` | illustration / infographic / photo / diagram |
 | `aspect_ratio` | 公众号用 16:9 或 4:3，小红书用 3:4 或 1:1 |
 | `size_hint` | full_width / half / thumbnail |
@@ -81,7 +83,8 @@
 
 ```json
 {
-  "style": "tech-flat",
+  "style": "tech-digital",
+  "style_variant": "satirical",
   "global_tone": "现代科技感，配色以蓝绿为主，留白多，信息密度低，整体氛围安静专注",
   "cover": {
     "title": "500元上门安装，299元上门卸载",
@@ -98,6 +101,7 @@
       "type": "illustration",
       "description": "详细的画面描述（50-150字）...",
       "purpose": "在文章中的叙事作用...",
+      "caption": "闲鱼上，装机和卸载已经同时开卖了。",
       "aspect_ratio": "16:9",
       "size_hint": "full_width"
     }

scripts/formatter.py

@@ -290,10 +290,11 @@ def insert_images(content_html: str, placements: list, image_map: dict, platform
         else:
             img_src = f"/api/runs/{run_id}/images/{pid}.png"
 
-        # 图注（深色模板用浅色文字）
+        # 图注：只渲染给读者看的 caption，不渲染内部 purpose
         caption_html = ""
-        if purpose:
-            caption_html = f'\n<p style="text-align:center;font-size:12px;color:{caption_color};margin:4px 0 16px;">{purpose}</p>'
+        caption = str(placement.get("caption", "")).strip()
+        if caption:
+            caption_html = f'\n<p style="text-align:center;font-size:12px;color:{caption_color};margin:4px 0 16px;">{caption}</p>'
 
         img_tag = f'\n<img src="{img_src}" alt="{desc}" {style}>{caption_html}\n'
         blocks.insert(min(para_idx, len(blocks)), img_tag)
@@ -304,11 +305,11 @@ def insert_images(content_html: str, placements: list, image_map: dict, platform
 # ── 模板匹配 ─────────────────────────────────────────────────
 
 def match_template(platform: str, keywords: list[str], director_style: str = "") -> str:
-    """根据 Director style + 关键词匹配排版模板
+    """根据 Director style + 关键词匹配排版模板。
 
     优先级:
-    1. Director style 直接映射（如 style="tech-flat" → tech-digital.html）
-    2. keywords 子串匹配（"AI Agent" 包含 "AI" → tech-digital.html）
+    1. YAML 中配置的 style exact / aliases / prefixes
+    2. keywords 子串匹配
     3. fallback 到 default
     """
     mapping_path = CONFIG_DIR / "template-mapping.yaml"
@@ -319,23 +320,25 @@ def match_template(platform: str, keywords: list[str], director_style: str = "")
     platform_config = mapping.get(platform, {})
     rules = platform_config.get("mapping", [])
     default = platform_config.get("default", "base.html")
+    styles_cfg = platform_config.get("styles", {}) if isinstance(platform_config, dict) else {}
+    exact_map = styles_cfg.get("exact", {}) if isinstance(styles_cfg, dict) else {}
+    alias_map = styles_cfg.get("aliases", {}) if isinstance(styles_cfg, dict) else {}
+    prefix_map = styles_cfg.get("prefixes", {}) if isinstance(styles_cfg, dict) else {}
 
-    # ── 1. Director style 直接映射 ──
-    STYLE_TO_TEMPLATE = {
-        "tech-flat": "tech-digital.html",
-        "tech-minimal": "tech-digital.html",
-        "tech-digital": "tech-digital.html",
-        "cyberpunk": "tech-digital.html",
-        "watercolor": "lifestyle.html",
-        "chinese-ink": "lifestyle.html",
-        "business": "business-finance.html",
-        "news": "news-insight.html",
-        "education": "education.html",
-    }
+    # ── 1. Director style 直接映射（来自 YAML 配置）──
     if director_style:
         style_lower = director_style.lower().strip()
-        if style_lower in STYLE_TO_TEMPLATE:
-            tpl = STYLE_TO_TEMPLATE[style_lower]
+        tpl = None
+        if style_lower in exact_map:
+            tpl = exact_map[style_lower]
+        elif style_lower in alias_map:
+            tpl = alias_map[style_lower]
+        else:
+            for prefix, candidate_tpl in prefix_map.items():
+                if style_lower.startswith(str(prefix).lower().strip()):
+                    tpl = candidate_tpl
+                    break
+        if tpl:
             tpl_path = TEMPLATES_DIR / platform / tpl
             if tpl_path.exists():
                 return tpl

scripts/validators.py

@@ -2,8 +2,8 @@
 
 import json
 import re
+from pathlib import Path
 from dataclasses import dataclass, field
-from difflib import SequenceMatcher
 from typing import Any, Callable
 
 import yaml
@@ -153,6 +153,19 @@ def validate_research_yaml(text: str) -> ValidationResult:
     return ValidationResult(ok=True, parsed=parsed, normalized_text=normalized)
 
 
+def _allowed_visual_styles() -> set[str]:
+    cfg_path = Path(__file__).resolve().parent.parent / "config" / "template-mapping.yaml"
+    try:
+        cfg = yaml.safe_load(cfg_path.read_text(encoding="utf-8")) or {}
+        exact = ((cfg.get("wechat") or {}).get("styles") or {}).get("exact") or {}
+        return {str(k).strip() for k in exact.keys() if str(k).strip()}
+    except Exception:
+        return {"tech-digital", "business-finance", "news-insight", "lifestyle", "education"}
+
+
+ALLOWED_VISUAL_STYLES = _allowed_visual_styles()
+
+
 def validate_visual_plan_json(text: str) -> ValidationResult:
     raw = _strip_code_fence(text)
     try:
@@ -165,12 +178,17 @@ def validate_visual_plan_json(text: str) -> ValidationResult:
         return ValidationResult(ok=False, message="visual_plan.json top-level must be an object", details=[f"got {type(parsed).__name__}"])
 
     style = parsed.get("style")
+    style_variant = parsed.get("style_variant")
     global_tone = parsed.get("global_tone")
     cover = parsed.get("cover")
     placements = parsed.get("placements")
 
     if not isinstance(style, str) or not style.strip():
         details.append("style must be a non-empty string")
+    elif style.strip() not in ALLOWED_VISUAL_STYLES:
+        details.append(f"style must be one of: {', '.join(sorted(ALLOWED_VISUAL_STYLES))}")
+    if not isinstance(style_variant, str) or not style_variant.strip():
+        details.append("style_variant must be a non-empty string")
     if not isinstance(global_tone, str) or not global_tone.strip():
         details.append("global_tone must be a non-empty string")
     if not isinstance(cover, dict):
@@ -196,7 +214,7 @@ def validate_visual_plan_json(text: str) -> ValidationResult:
                 details.append(f"duplicate placement id: {pid}")
             else:
                 seen_ids.add(pid)
-            for field in ("after_section", "type", "description", "purpose"):
+            for field in ("after_section", "type", "description", "purpose", "caption"):
                 value = placement.get(field)
                 if not isinstance(value, str) or not value.strip():
                     details.append(f"{prefix}.{field} must be a non-empty string")
@@ -320,6 +338,46 @@ def validate_writer_markdown(text: str, min_chars: int = 1200) -> ValidationResu
     return ValidationResult(ok=True, parsed=raw, normalized_text=raw + "\n")
 
 
+def validate_chat_article_markdown(text: str, min_chars: int = 800) -> ValidationResult:
+    raw = _strip_code_fence(text).strip()
+    details: list[str] = []
+    bad_markers = [
+        "微信文章无法直接提取",
+        "基于",
+        "我直接修改",
+        "根据你的要求",
+        "下面是修改后的",
+        "以下是修改后的",
+        "我会按这个风格",
+        "已按",
+        "我已经根据",
+        "让我根据",
+        "已重写",
+        "主要调整",
+        "风格特点是",
+    ]
+
+    if len(raw) < min_chars:
+        details.append(f"content too short: {len(raw)} chars (< {min_chars})")
+
+    head = raw[:500]
+    for marker in bad_markers:
+        if marker in head:
+            details.append(f"meta marker near opening: {marker}")
+
+    paragraphs = [p.strip() for p in re.split(r"\n\s*\n", raw) if p.strip()]
+    if len(paragraphs) < 4:
+        details.append(f"too few paragraphs: {len(paragraphs)}")
+
+    heading_count = sum(1 for line in raw.splitlines() if line.strip().startswith("## "))
+    if heading_count < 2:
+        details.append(f"too few section headings: {heading_count}")
+
+    if details:
+        return ValidationResult(ok=False, message="chat article failed quality checks", details=details)
+    return ValidationResult(ok=True, parsed=raw, normalized_text=raw + "\n")
+
+
 def validate_de_ai_markdown(text: str, original_text: str) -> ValidationResult:
     raw = _strip_code_fence(text).strip()
     details: list[str] = []
@@ -349,12 +407,6 @@ def validate_de_ai_markdown(text: str, original_text: str) -> ValidationResult:
         if ratio < 0.6 or ratio > 1.4:
             details.append(f"length ratio out of range: {ratio:.2f}x")
 
-        sim = SequenceMatcher(None, original[:5000], raw[:5000]).ratio()
-        if sim < 0.30:
-            details.append(f"too different from source article: similarity={sim:.2f}")
-        if sim > 0.985:
-            details.append(f"too similar to source article: similarity={sim:.2f}")
-
     if details:
         return ValidationResult(ok=False, message="article_edited.md failed quality checks", details=details)
     return ValidationResult(ok=True, parsed=raw, normalized_text=raw + "\n")