@yuhan1124/draw-prompt 0.4.12 → 0.4.13

package/README.md

@@ -174,7 +174,7 @@ status                                            # 数据 + 下游通道健康
 4. 只有当用户明确要求“文字必须绝对准确/可后处理”或出图反馈属于 `text_error` 时，才切到 `--strict-text` + `overlay` 两段式兜底。
 5. 用 `visual-check` 验证成品图尺寸、画幅、亮度、对比度和基础细节。
 6. 参考图改图用 `edit-check` 验证“主体保留 + 背景/目标确实变化”。
-7. 模板或策略变动后，在开发仓库本地跑 `visual-regress references/visual-cases.jsonl`，确认多场景回归通过；它会真实编译 `convert`、`variants`、`series`、`adapt`、`compose` 等单图/多输出入口。用 `expect_count`、`expect_asset_type(s)`、`expect_aspect(s)`、`expect_required_text(_all)`、`forbid_required_text`、`expect_prompt_contains`、`forbid_prompt_contains` 把真实场景的产品意图固化成门禁。
+7. 模板或策略变动后，在开发仓库本地跑 `visual-regress references/visual-cases.jsonl`，确认多场景回归通过；它会真实编译 `convert`、`rewrite`、`edit`、`variants`、`series`、`adapt`、`compose`、`brand`、`character`、`data-viz` 等单图/多输出入口。用 `expect_count`、`expect_asset_type(s)`、`expect_aspect(s)`、`expect_required_text(_all)`、`forbid_required_text`、`expect_safety_rewrite`、`expect_prompt_contains`、`forbid_prompt_contains` 把真实场景的产品意图固化成门禁。
 
 这条链路的默认目标不是替 gpt-image-2 重做排版引擎，而是减少跑偏、遗漏和廉价风格；
 两段式 overlay 只是文字极端稳定性兜底，不作为普通用户的默认体验。

package/SKILL.md

@@ -8,7 +8,7 @@ description: >-
   画图的指令"、"优化我的出图 prompt"、"按我的风格生成 prompt"，或在用 GPT Image 2 /
   gpt-image-2 出图前需要一段精准提示词时，使用本 skill。
 metadata:
-  version: 0.4.12
+  version: 0.4.13
   openclaw:
     anyBins: ["uv", "python3"]
 ---

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yuhan1124/draw-prompt",
-  "version": "0.4.12",
+  "version": "0.4.13",
   "description": "Convert natural-language image requests into high-quality gpt-image-2 prompts and Codex handoff blocks.",
   "type": "commonjs",
   "bin": {

package/references/conversion-skill-plan.md

@@ -369,7 +369,7 @@ prompt，再交给下游出图。
 | `ui_dashboard` | ui | 侧边栏、顶部栏、KPI、图表、表格 |
 | `slide_corporate_report` | slide | 保留用户显式区域、栏目/行列结构、图标位置、页脚和留白 |
 | `diagram_rag` | diagram | User -> Retriever -> Vector DB -> LLM -> Answer |
-| `diagram_system` | diagram | 分层盒子、有向箭头、图例 |
+| `diagram_system` | diagram | 分层盒子、有向箭头、只保留用户要求的标签 |
 
 ## 质量门
 

package/scripts/prompt_cli.py

@@ -138,7 +138,7 @@ def ensure_home() -> None:
 
 
 SCHEMA_VERSION = 1
-COMPILER_VERSION = "0.4.12"
+COMPILER_VERSION = "0.4.13"
 
 
 PACKAGED_SKILL_FILES = [
@@ -2320,7 +2320,7 @@ TEMPLATE_DEFS = {
     "diagram_system": {
         "asset_type": "diagram",
         "label": "系统架构图",
-        "layout": "layered system boxes with directional arrows and a small legend",
+        "layout": "layered system boxes with directional arrows",
         "keywords": ["架构", "系统", "模块", "服务", "流程"],
     },
     "product_hero": {
@@ -2593,6 +2593,9 @@ NON_DISPLAY_TEXTS = {
     "关系",
     "标签",
     "卖点标签",
+    "主视觉",
+    "产品主视觉",
+    "核心",
     "购买转化氛围",
     "但不要过度促销",
 }
@@ -2642,12 +2645,13 @@ def extract_required_texts(request: str, explicit_texts: list[str]) -> list[str]
 
     def clean_text(text: str) -> str:
         text = text.strip(" \t\n\r，,、。；;：:.!?！？")
-        text = re.sub(r"^(?:[^：:]{0,12}(?:指标卡|卡片|列表|标题|模块|节点|步骤|栏目|分支))[：:]", "", text)
+        text = re.sub(r"^(?:[^：:]{0,20}(?:活动路径|指标卡|卡片|列表|标题|模块|节点|步骤|栏目|分支|路径|卡))[：:]", "", text)
         text = re.sub(r"^(?:顶部|底部|中间|左侧|右侧|上方|下方|首页|页面)?(?:主标题|副标题|标题)\s+", "", text)
         text = re.sub(r"^(?:问候语|核心卡片|主要按钮|主按钮|按钮)\s+", "", text)
         text = re.sub(r"^(?:是|为|叫)\s+", "", text)
         text = re.sub(r"^(?:一个|一枚|一项)?(?:明显的|醒目的|主要的|primary\s+)?(.{1,16})按钮$", r"\1", text, flags=re.IGNORECASE)
-        text = re.sub(r"(?:网格|列表|区域|模块)$", "", text)
+        text = re.sub(r"(?:[一二三四五六七八九十\d]+个)?(?:网格|列表|区域|模块|卡片)$", "", text)
+        text = re.sub(r"\s*[一二三四五六七八九十\d]+个$", "", text)
         text = re.sub(r"(?:要)?(?:渲染)?(?:清晰|清楚|可读|明显)$", "", text)
         text = re.sub(r"(?:这些)?元素$", "", text)
         text = re.sub(r"(?:[一二三四五六七八九十\d]+个)?步骤$", "", text)
@@ -2748,12 +2752,13 @@ def merge_texts(primary: list[str], extra: list[str]) -> list[str]:
     seen: set[str] = set()
     for item in primary + extra:
         text = item.strip(" \t\n\r，,、。；;：:.!?！？")
-        text = re.sub(r"^(?:[^：:]{0,12}(?:指标卡|卡片|列表|标题|模块|节点|步骤|栏目|分支))[：:]", "", text)
+        text = re.sub(r"^(?:[^：:]{0,20}(?:活动路径|指标卡|卡片|列表|标题|模块|节点|步骤|栏目|分支|路径|卡))[：:]", "", text)
         text = re.sub(r"^(?:顶部|底部|中间|左侧|右侧|上方|下方|首页|页面)?(?:主标题|副标题|标题)\s+", "", text)
         text = re.sub(r"^(?:问候语|核心卡片|主要按钮|主按钮|按钮)\s+", "", text)
         text = re.sub(r"^(?:是|为|叫)\s+", "", text)
         text = re.sub(r"^(?:一个|一枚|一项)?(?:明显的|醒目的|主要的|primary\s+)?(.{1,16})按钮$", r"\1", text, flags=re.IGNORECASE)
-        text = re.sub(r"(?:网格|列表|区域|模块)$", "", text)
+        text = re.sub(r"(?:[一二三四五六七八九十\d]+个)?(?:网格|列表|区域|模块|卡片)$", "", text)
+        text = re.sub(r"\s*[一二三四五六七八九十\d]+个$", "", text)
         text = re.sub(r"(?:要)?(?:渲染)?(?:清晰|清楚|可读|明显)$", "", text)
         text = re.sub(r"(?:这些)?元素$", "", text)
         text = re.sub(r"(?:[一二三四五六七八九十\d]+个)?步骤$", "", text)
@@ -2768,19 +2773,30 @@ def merge_texts(primary: list[str], extra: list[str]) -> list[str]:
         if text and key not in seen:
             seen.add(key)
             merged.append(text)
-    return merged
+    pruned: list[str] = []
+    compact_items = [(text, re.sub(r"[\s，,、。；;：:.!?！？]+", "", text)) for text in merged]
+    for idx, (text, key) in enumerate(compact_items):
+        if len(key) <= 8 and any(
+            idx != other_idx and key and key in other_key and len(other_key) > len(key)
+            for other_idx, (_, other_key) in enumerate(compact_items)
+        ):
+            continue
+        pruned.append(text)
+    return pruned
 
 
 def extract_structural_labels(request: str, asset_type: str) -> list[str]:
     if asset_type not in {"diagram", "infographic", "slide", "ui", "poster", "product"}:
         return []
     candidates: list[tuple[int, str]] = []
-    list_intro = r"(?:需要)?(?:展示|呈现|列出|包含|包括|含有|分为|覆盖|输出)"
+    list_intro = r"(?:需要)?(?:展示|呈现|列出|包含|包括|含有|分为|覆盖)"
     stop_words = r"(?:\b16\s*:\s*9\b|\b9\s*:\s*16\b|\b3\s*:\s*4\b|\b1\s*:\s*1\b|适合|用于|画幅|aspect|高质量|高清|clean|corporate)"
     patterns = [
         rf"(?:结构(?:从左到右|从右到左|从上到下|从下到上)?|链路|流程)\s*[：:]\s*([^。\n]{{2,240}})",
         rf"(?:数据|指标|核心数据|关键指标|卖点|亮点|功能点|特性|活动路径|路径)\s*(?:包括|包含|有|为)?[：:\s]*([^。；;\n]{{2,180}})",
         rf"{list_intro}\s*(?:这些|以下|对应的)?(?:模块|部分|层|栏目|节点|入口|能力|场景|列表|指标卡|卡片|步骤|分支|数据|指标|卖点|亮点|功能点|特性|路径)?[：:\s]*([^。；;\n]{{2,180}})",
+        rf"(?:顶部|底部|中间|左侧|右侧|上方|下方|最终|末端)?\s*输出\s*(?!可交给|为|成|到|给)([^。；;\n]{{2,180}})",
+        rf"(?:顶部|底部|中间|左侧|右侧|上方|下方|主区域|页面|首页)?\s*(?:有|显示|放置|排列|提供)\s*([^。；;\n]{{2,180}}?(?:[一二三四五六七八九十\d]+个)?(?:模块|卡片|按钮|入口|标签|列表))",
         rf"(?:模块|节点|栏目|部分|层|入口|能力|场景|列表|指标卡|卡片|步骤|分支|数据|指标|卖点|亮点|功能点|特性|路径)\s*(?:包括|包含|有|为)[：:\s]*([^。；;\n]{{2,180}})",
         rf"(?:模块|节点|栏目|部分|层|入口|能力|场景|列表|指标卡|卡片|步骤|分支|数据|指标|卖点|亮点|功能点|特性|路径)[^。；;\n：:]{{0,16}}[：:]([^。；;\n]{{2,180}})",
     ]
@@ -2790,17 +2806,23 @@ def extract_structural_labels(request: str, asset_type: str) -> list[str]:
         )
     for pattern in patterns:
         for match in re.finditer(pattern, request, flags=re.IGNORECASE):
+            if re.search(r"\bicon\s*:|图标", match.group(0), flags=re.IGNORECASE):
+                continue
             value = re.split(stop_words, match.group(1), maxsplit=1, flags=re.IGNORECASE)[0]
             value = re.sub(r"[()（）]", "、", value)
             value = re.sub(r"(?:进入|再到|到|输出|->|→)", "、", value)
             for part_match in re.finditer(r"[^、,，;；|]+", value):
                 part = part_match.group(0).strip(" \t\n\r，,、：:")
-                part = re.sub(r"^(?:[^：:]{0,12}(?:指标卡|卡片|列表|标题|模块|节点|步骤|栏目|分支))[：:]", "", part)
+                part = re.sub(r"^(?:[^：:]{0,20}(?:活动路径|指标卡|卡片|列表|标题|模块|节点|步骤|栏目|分支|路径|卡))[：:]", "", part)
                 part = re.sub(r"^(?:和|与|及|以及|and|再|则)\s*", "", part, flags=re.IGNORECASE).strip()
                 part = re.sub(r"^(?:底部有|顶部有|左侧有|右侧有|上方有|下方有)\s*", "", part)
                 part = re.sub(r"^(?:包含|包括|含有)\s*", "", part)
+                part = re.sub(r"^(?:问候语|核心卡片|主要按钮|主按钮|按钮)\s+", "", part)
+                part = re.sub(r"^(?:一个|一枚|一项)?(?:明显的|醒目的|主要的|primary\s+)?(.{1,16})按钮$", r"\1", part, flags=re.IGNORECASE)
                 part = re.sub(r"\s*(?:和|与|及|以及|and)$", "", part, flags=re.IGNORECASE).strip()
                 part = re.sub(r"(?:这些)?元素$", "", part)
+                part = re.sub(r"(?:[一二三四五六七八九十\d]+个)?(?:网格|列表|区域|模块|卡片)$", "", part)
+                part = re.sub(r"\s*[一二三四五六七八九十\d]+个$", "", part)
                 part = re.sub(r"(?:[一二三四五六七八九十\d]+个)?步骤$", "", part)
                 part = part.strip(" \t\n\r，,、：:")
                 if not part:
@@ -2854,7 +2876,7 @@ def infer_palette_constraints(request: str) -> list[str]:
     ]
     constraints: list[str] = []
     seen: set[str] = set()
-    for sentence in re.split(r"[。；;\n]", request):
+    for sentence in re.split(r"[。；;\n]|(?<=[A-Za-z)])\.\s+", request):
         for fragment in re.split(r"[，,]", sentence):
             fragment = fragment.strip(" \t\r\n。.；;，,")
             if not fragment:
@@ -3112,7 +3134,7 @@ def infer_must_include(asset_type: str, template_id: str, texts: list[str], stri
         "ui": ["requested screen type", "requested UI sections", "clear component hierarchy"],
         "infographic": ["requested information units", "requested relationships", "clear visual hierarchy"],
         "slide": ["widescreen slide canvas", "all requested sections", "requested visual motifs", "crisp readable text hierarchy"],
-        "diagram": ["labeled components", "directional arrows", "legend or flow semantics"],
+        "diagram": ["labeled components", "directional arrows", "clear flow semantics"],
         "product": ["single hero product", "visible material texture", "controlled studio lighting"],
         "photography": ["realistic subject", "specific scene details", "natural imperfections"],
         "character": ["consistent character identity", "turnaround views", "expression close-ups"],
@@ -3451,7 +3473,7 @@ def render_prompt(spec: dict) -> str:
                 brief_block,
                 f"Style / quality envelope for diagram: {diagram_style}. This only controls presentation polish; the user brief wins.",
                 f"Quality controls: crisp typography, clear node grouping, balanced white space, precise alignment, high contrast, restrained palette ({diagram_palette}).",
-                f"Recommended structure: {layout}. Use boxes, groups, arrows, and a feedback loop only where they match the brief.",
+                f"Recommended structure: {layout}. Use boxes, groups, arrows, and a feedback loop only where they match the brief; do not add a legend or extra explanatory labels unless the brief asks for them.",
                 "Do not over-template the diagram. Preserve the user's named modules, reading order, and product-review context.",
                 f"Must include: {must_include}.",
                 text_block,
@@ -4870,6 +4892,8 @@ def visual_case_compile(case: dict) -> dict:
     if tool in {"convert", "rewrite"}:
         compiled = compile_visual_case(case, target=case.get("target") or "codex-image")
         return with_compiled_items({**compiled, "tool": tool}, [{"id": case.get("id") or tool, **compiled}])
+    if tool == "edit":
+        return compile_visual_edit_case(case)
     if tool == "variants":
         return compile_visual_variants_case(case)
     if tool == "series":
@@ -5128,6 +5152,51 @@ def compile_visual_variants_case(case: dict) -> dict:
     return with_compiled_items({"tool": "variants", "count": len(items)}, items)
 
 
+def compile_visual_edit_case(case: dict) -> dict:
+    goal = str(case.get("goal") or visual_case_request(case))
+    if not goal:
+        raise ValueError("edit case 缺少 goal/request 字段")
+    references = [parse_reference(value) for value in case_list(case.get("reference") or case.get("references"))]
+    preserve = case_list(case.get("preserve")) or ["main subject identity, silhouette, material cues, and composition anchors"]
+    changes = case_list(case.get("change")) or [goal]
+    required_text = case_list(case.get("text") or case.get("required_text"))
+    aspect = str(case.get("aspect") or "3:4")
+    asset_type = str(case.get("asset_type") or "product")
+    quality = str(case.get("quality") or "high")
+    reference_block = "; ".join(f"{item['role']}={item['ref']}" for item in references) or "provided reference image(s)"
+    prompt = "\n".join(
+        [
+            f"Edit the provided reference image(s) into a {aspect} {asset_type} result for: {goal}.",
+            f"References: {reference_block}.",
+            "Preserve exactly: " + "; ".join(preserve) + ".",
+            "Change only: " + "; ".join(changes) + ".",
+            f"Visual target: {case.get('style') or 'production-quality realistic edit, consistent lighting, no visible seams'}; quality={quality}.",
+            exact_text_block(required_text),
+            "Avoid: identity drift; unwanted background changes; mismatched perspective; fake logos; garbled text; low-resolution artifacts.",
+        ]
+    )
+    lint = lint_prompt(prompt, asset_type, quality, required_text)
+    spec = {
+        "asset_type": asset_type,
+        "aspect": aspect,
+        "template_id": "edit",
+        "required_text": required_text,
+        "strict_text": False,
+        "quality": quality,
+    }
+    compiled = {
+        "spec": spec,
+        "prompt": prompt,
+        "prompt_digest": prompt_digest(prompt),
+        "lint": lint,
+        "intent_check": [],
+        "handoff": None,
+        "text_overlay_spec": None,
+        "acceptance_criteria": [],
+    }
+    return with_compiled_items({"tool": "edit", **compiled}, [{"id": "edit-01", **compiled}])
+
+
 def compile_visual_series_case(case: dict) -> dict:
     raw_items = case.get("briefs") or case.get("items") or case.get("series") or []
     if isinstance(raw_items, str):
@@ -5370,6 +5439,27 @@ def visual_case_expectation_findings(case: dict, compiled: dict) -> list[dict]:
                 "message": f"期望 template_id={expected_template}，实际={spec.get('template_id')}",
             }
         )
+    expected_templates = case_list(case.get("expect_template_ids"))
+    if expected_templates:
+        actual = [str(item.get("template_id") or "") for item in specs]
+        if actual != expected_templates:
+            findings.append(
+                {
+                    "severity": "error",
+                    "rule": "case.template_ids_mismatch",
+                    "message": f"期望 template_ids={expected_templates}，实际={actual}",
+                }
+            )
+
+    if case.get("expect_safety_rewrite"):
+        if not any(item.get("safety_rewrite") for item in specs):
+            findings.append(
+                {
+                    "severity": "error",
+                    "rule": "case.safety_rewrite_missing",
+                    "message": "期望触发安全改写，但 spec.safety_rewrite 为空",
+                }
+            )
 
     labels_by_item = [set(item.get("required_text") or []) for item in specs]
     labels = set().union(*labels_by_item) if labels_by_item else set()