@yuhan1124/draw-prompt 0.4.7 → 0.4.8

package/README.md

@@ -69,7 +69,8 @@ npx --yes --registry=https://registry.npmjs.org/ @yuhan1124/draw-prompt@latest i
 
 `install-skill` 默认复制 npm 包内的运行文件，避免软链到 npx 缓存导致后续路径失效。它只复制
 `SKILL.md`、CLI、references 和必要元数据；开发仓库里的 `tests/`、`tmp/`、`golden-cases.jsonl`、
-`visual-cases.jsonl` 不会进入 npm 包或安装目录。
+`visual-cases.jsonl` 不会进入 npm 包或安装目录。安装后可直接跑 `doctor`，它会检查
+包文件、版本一致性、核心单图转化和一条真实长输入 `compose` 链路。
 
 开发者本地调试也可以软链 repo：
 

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: 0.4.7
+  version: 0.4.8
   openclaw:
     anyBins: ["uv", "python3"]
 ---

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yuhan1124/draw-prompt",
-  "version": "0.4.7",
+  "version": "0.4.8",
   "description": "Convert natural-language image requests into high-quality gpt-image-2 prompts and Codex handoff blocks.",
   "type": "commonjs",
   "bin": {

package/scripts/prompt_cli.py

@@ -136,7 +136,7 @@ def ensure_home() -> None:
 
 
 SCHEMA_VERSION = 1
-COMPILER_VERSION = "0.4.7"
+COMPILER_VERSION = "0.4.8"
 
 
 PACKAGED_SKILL_FILES = [
@@ -211,6 +211,29 @@ DOCTOR_CASES = [
     },
 ]
 
+DOCTOR_COMPOSE_CASES = [
+    {
+        "id": "long-input-product-plan",
+        "request": (
+            "请把下面这段产品方案整理成 4 张图用于路演：我们要做一个 AI 生图 Prompt 转化 skill。"
+            "目标用户是经常把长文档、产品需求、运营数据、架构说明变成图片的研发和运营同学。"
+            "当前痛点是直接把长输入丢给模型容易跑偏，常见问题包括风格不稳定、中文标签遗漏、架构图被画成 UI、产品图漏卖点、活动战报颜色跑偏。"
+            "方案包括：场景识别、风格预设、意图保持检查、质量门、真实 Image 2 验证、样本反馈学习。"
+            "上线标准是安装后 doctor 通过、常见场景开箱高质量、NPM 公网安装可用。"
+        ),
+        "max_images": 4,
+        "style_preset": "corporate",
+        "expect_count": 4,
+        "expect_no_asset_type": "illustration",
+        "expect_required_text_by_index": {
+            0: ["目标用户"],
+            1: ["当前痛点"],
+            2: ["方案", "场景识别", "风格预设", "质量门"],
+            3: ["上线标准", "NPM 公网安装可用"],
+        },
+    }
+]
+
 
 def package_root() -> Path:
     return Path(__file__).resolve().parent.parent
@@ -4887,6 +4910,30 @@ def split_document_sections(text: str, max_images: int) -> list[str]:
     if buf:
         chunks.append(" ".join(buf).strip())
 
+    if len(chunks) <= 1:
+        semantic_patterns = [
+            r"(目标用户(?:是|：|:)[^。；;\n]+)",
+            r"(用户场景(?:是|：|:)[^。；;\n]+)",
+            r"(当前痛点(?:是|：|:)[^。；;\n]+)",
+            r"(主要问题(?:是|：|:)[^。；;\n]+)",
+            r"(方案(?:包括|是|：|:)[^。；;\n]+)",
+            r"(核心能力(?:包括|是|：|:)[^。；;\n]+)",
+            r"(上线标准(?:是|：|:)[^。；;\n]+)",
+            r"(验收标准(?:是|：|:)[^。；;\n]+)",
+        ]
+        semantic_matches: list[tuple[int, str]] = []
+        seen_spans: set[tuple[int, int]] = set()
+        for pattern in semantic_patterns:
+            for match in re.finditer(pattern, text):
+                span = match.span(1)
+                if span in seen_spans:
+                    continue
+                seen_spans.add(span)
+                semantic_matches.append((span[0], match.group(1).strip()))
+        semantic_chunks = [chunk for _, chunk in sorted(semantic_matches, key=lambda item: item[0])]
+        if len(semantic_chunks) >= 2:
+            chunks = semantic_chunks
+
     if len(chunks) <= 1:
         sentences = [s.strip() for s in re.split(r"(?<=[。！？!?；;])\s*", text) if s.strip()]
         chunks = []
@@ -4908,6 +4955,8 @@ def split_document_sections(text: str, max_images: int) -> list[str]:
 
 def choose_compose_asset(chunk: str, index: int) -> str:
     lower = chunk.lower()
+    if re.search(r"^(?:目标用户|用户场景|当前痛点|主要问题|上线标准|验收标准)", chunk):
+        return "infographic"
     if any(k in lower for k in ["架构", "系统", "模块", "链路", "rag", "llm", "retriever", "pipeline"]):
         return "diagram"
     if any(k in lower for k in ["数据", "指标", "报表", "趋势", "占比", "转化率", "漏斗", "图表"]):
@@ -4922,7 +4971,9 @@ def choose_compose_asset(chunk: str, index: int) -> str:
         return "character"
     if index == 0 and any(k in lower for k in ["标题", "主题", "发布", "活动", "封面", "总结"]):
         return "poster"
-    return "illustration"
+    if any(k in lower for k in ["插画", "场景插图", "故事", "氛围图", "场景图", "scene illustration"]):
+        return "illustration"
+    return "infographic"
 
 
 def infer_compose_style(text: str) -> str:
@@ -4946,6 +4997,13 @@ def extract_visual_labels(chunk: str, asset_type: str, limit: int = 5) -> list[s
 
     for match in re.findall(r'"([^"\n]{1,28})"|“([^”\n]{1,28})”|「([^」\n]{1,28})」', chunk):
         add(next((m for m in match if m), ""))
+    for match in re.finditer(r"(目标用户|用户场景|当前痛点|主要问题|方案|核心能力|上线标准|验收标准)\s*(?:是|包括|包含|：|:)\s*([^。；;\n]{2,140})", chunk):
+        add(match.group(1))
+        value = match.group(2)
+        for part in re.split(r"[、,，/|]", value):
+            part = re.sub(r"^(?:和|与|及|以及|and)\s*", "", part.strip(), flags=re.IGNORECASE)
+            if 1 < len(part) <= 18:
+                add(part)
     for match in re.findall(r"\b[A-Z][A-Za-z0-9_-]{1,20}\b", chunk):
         add(match)
     for match in re.findall(r"(?:标题|主题|模块|步骤|节点|页面)[：:\s]*([^，。；;\n]{2,24})", chunk):
@@ -4964,7 +5022,7 @@ def compose_purpose(asset_type: str, index: int) -> str:
         "ui": "界面概念图",
         "product": "产品视觉图",
         "character": "角色设定图",
-        "illustration": "场景说明图",
+        "illustration": "场景插画",
     }
     return f"{index}. {purpose_map.get(asset_type, '配图')}"
 
@@ -5796,6 +5854,68 @@ def doctor_check_case(case: dict) -> dict:
     }
 
 
+def doctor_check_compose_case(case: dict) -> dict:
+    findings: list[dict] = []
+    text = str(case.get("request") or "")
+    max_images = int(case.get("max_images") or 4)
+    chunks = split_document_sections(text, max_images)
+    shared_style = infer_compose_style(text)
+    visual_plan: list[dict] = []
+    for idx, chunk in enumerate(chunks, start=1):
+        asset_type = choose_compose_asset(chunk, idx - 1)
+        labels = extract_visual_labels(chunk, asset_type)
+        purpose = compose_purpose(asset_type, idx)
+        compiled = compile_visual_case(
+            {
+                "id": f"compose-{idx:02d}",
+                "request": f"{purpose}。根据这段内容生成对应画面：{chunk}",
+                "asset_type": asset_type,
+                "style": shared_style,
+                "style_preset": case.get("style_preset"),
+                "text": labels if asset_type in {"diagram", "infographic", "ui"} else [],
+                "target": "raw",
+                "tags": "compose,long-input",
+            },
+            target="raw",
+            include_handoff=False,
+        )
+        visual_plan.append({"chunk": chunk, "compiled": compiled})
+        for item in compiled["lint"]:
+            findings.append({"severity": item["severity"], "rule": f"compose.{idx}.lint.{item['rule']}", "message": item["message"]})
+        for item in compiled["intent_check"]:
+            findings.append({"severity": item["severity"], "rule": f"compose.{idx}.intent.{item['rule']}", "message": item["message"]})
+
+    if case.get("expect_count") and len(visual_plan) != int(case["expect_count"]):
+        doctor_add(findings, "error", "compose.count", f"期望 {case['expect_count']} 张图，实际 {len(visual_plan)}")
+    forbidden_asset = case.get("expect_no_asset_type")
+    if forbidden_asset:
+        for idx, item in enumerate(visual_plan, start=1):
+            asset_type = item["compiled"]["spec"].get("asset_type")
+            if asset_type == forbidden_asset:
+                doctor_add(findings, "error", "compose.asset_type", f"第 {idx} 张图不应路由为 {forbidden_asset}")
+    for raw_index, expected_texts in (case.get("expect_required_text_by_index") or {}).items():
+        index = int(raw_index)
+        if index >= len(visual_plan):
+            doctor_add(findings, "error", "compose.index_missing", f"缺少第 {index + 1} 张图")
+            continue
+        labels = set(visual_plan[index]["compiled"]["spec"].get("required_text") or [])
+        prompt = visual_plan[index]["compiled"]["prompt"]
+        for text_item in expected_texts:
+            if text_item not in labels:
+                doctor_add(findings, "error", "compose.required_text_missing", f"第 {index + 1} 张图缺少必显文字：{text_item}")
+            elif f'"{text_item}"' not in prompt:
+                doctor_add(findings, "error", "compose.required_text_not_quoted", f"第 {index + 1} 张图 Prompt 未逐字引用：{text_item}")
+    return {
+        "name": f"compose-{case['id']}",
+        "pass": not has_lint_error(findings),
+        "count": len(visual_plan),
+        "asset_types": [item["compiled"]["spec"].get("asset_type") for item in visual_plan],
+        "prompt_digests": [item["compiled"]["prompt_digest"] for item in visual_plan],
+        "required_text": [item["compiled"]["spec"].get("required_text") or [] for item in visual_plan],
+        "findings": findings,
+    }
+
+
 def cmd_doctor(args: argparse.Namespace) -> int:
     root = package_root()
     checks = [doctor_check_package(root)]
@@ -5804,6 +5924,11 @@ def cmd_doctor(args: argparse.Namespace) -> int:
             checks.append(doctor_check_case(case))
         except Exception as exc:
             checks.append({"name": case.get("id", "case"), "pass": False, "findings": [{"severity": "error", "rule": "case.exception", "message": str(exc)}]})
+    for case in DOCTOR_COMPOSE_CASES:
+        try:
+            checks.append(doctor_check_compose_case(case))
+        except Exception as exc:
+            checks.append({"name": f"compose-{case.get('id', 'case')}", "pass": False, "findings": [{"severity": "error", "rule": "compose.exception", "message": str(exc)}]})
     passed = all(item["pass"] for item in checks)
     summary = {"pass": passed, "version": COMPILER_VERSION, "checks": len(checks), "failed": len([item for item in checks if not item["pass"]])}
     if args.json: