minimal-agent 0.8.2 → 0.10.0

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

@@ -0,0 +1,515 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+image_plan_lint_test.py —— image_plan_lint.py 的回归锚（stdlib unittest，仿 layout_balance_test 风格）。
+
+夹具取自真机 build_bak（自包含、不依赖 build_bak 目录存在）：
+  - 结构图（架构/并轨/决策树）必 FAIL；氛围/隐喻（封面渐变/山峰）必 PASS —— 反向校准锁死。
+  - 装饰/设计词（visual hierarchy / contrast / neural network nodes / layers of light / geometric）不误伤。
+  - 改写措辞规避（"five stacked horizontal bands"）lint 抓不到 —— 锁死「诚实边界」（那层靠 guidance）。
+  - 缺 img_plan.json = PASS（exit 0）；含违规 = exit 1；干净 = exit 0（子进程验退出码）。
+
+运行：python skills/ui-design/scripts/image_plan_lint_test.py
+"""
+
+import json
+import os
+import subprocess
+import sys
+import tempfile
+import unittest
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, HERE)
+from image_plan_lint import _find_bad_aesthetic_terms, _find_i2i_only_terms, _find_terms, lint  # noqa: E402
+from image_plan_lint import _find_background_text_marker_terms  # noqa: E402
+
+SCRIPT = os.path.join(HERE, "image_plan_lint.py")
+
+# —— 真机 build_bak 的代表性 prompt（结构图 vs 氛围/隐喻）——
+PROMPT_ARCH = ("Abstract technology architecture visualization with 5 distinct horizontal "
+               "layers representing data flow, clean minimalist lines")
+PROMPT_TRACKS = ("Dual parallel evolution paths abstract visualization, two separate tracks "
+                 "running in parallel representing commercial and open source")
+PROMPT_TREE = ("Abstract decision tree visualization for technology selection guide, "
+               "three branching paths")
+PROMPT_COVER = ("A clean modern abstract technology background, soft violet and blue gradient "
+                "light beams with subtle neural network nodes and connection lines, negative space")
+PROMPT_MOUNTAIN = ("Abstract mountain climbing concept, four stylized mountain peaks, violet blue "
+                   "cyan gradient, clean geometric lines with subtle glow")
+# 改写措辞规避结构（无图表名词）—— 已知 lint 抓不到，靠 guidance。
+PROMPT_PARAPHRASE = ("five stacked horizontal bands of soft light, each band a different hue, "
+                     "minimalist, lots of negative space")
+PROMPT_BAD_AESTHETIC = (
+    "abstract blueprint background with crosshair, aegis shield symbol, "
+    "18 thumbnails, floating cards and UI cards"
+)
+
+
+def write_json(obj):
+    fd, path = tempfile.mkstemp(suffix=".json")
+    os.close(fd)
+    with open(path, "w", encoding="utf-8") as f:
+        json.dump(obj, f, ensure_ascii=False)
+    return path
+
+
+def plan(*prompts):
+    return [{"file": f"img_{i+1}.png", "aspect": "16:9", "prompt": p} for i, p in enumerate(prompts)]
+
+
+def outline(*notes):
+    """notes = [(route, note), ...] → 一个 outline.json。"""
+    return {"pages": [
+        {"page": i + 1, "title": f"页{i+1}", "visual": {"route": r, "note": n}}
+        for i, (r, n) in enumerate(notes)
+    ]}
+
+
+def run(img_plan_path, outline_path=None):
+    cmd = [sys.executable, SCRIPT, img_plan_path]
+    if outline_path:
+        cmd += ["--outline", outline_path]
+    r = subprocess.run(cmd, capture_output=True, text=True, encoding="utf-8")
+    return r.returncode, (r.stdout or "") + (r.stderr or "")
+
+
+class FindTermsTest(unittest.TestCase):
+    def test_structural_prompts_hit(self):
+        self.assertTrue(_find_terms(PROMPT_ARCH))     # architecture / layered structure / data flow
+        self.assertTrue(_find_terms(PROMPT_TRACKS))   # parallel evolution
+        self.assertTrue(_find_terms(PROMPT_TREE))     # decision tree
+
+    def test_atmospheric_prompts_pass(self):
+        self.assertEqual(_find_terms(PROMPT_COVER), [])      # 渐变 + 装饰节点 ≠ 结构图
+        self.assertEqual(_find_terms(PROMPT_MOUNTAIN), [])   # 山峰隐喻 + geometric lines ≠ 结构图
+
+    def test_decorative_design_words_not_overblocked(self):
+        # 这些词在合法氛围图里常见，绝不能误伤：
+        for s in [
+            "strong visual hierarchy and depth",     # 不收裸 hierarchy
+            "violet and blue contrast, 对比色",       # 不收裸 contrast / 对比
+            "subtle neural network nodes",            # 不收裸 network / nodes
+            "soft layers of light and mist",          # 不收裸 layers / 分层
+            "clean geometric lines, glow",            # 不收 geometric
+        ]:
+            self.assertEqual(_find_terms(s), [], f"误伤装饰词：{s}")
+
+    def test_chinese_note_terms(self):
+        self.assertTrue(_find_terms("抽象的五层架构分层示意图背景"))   # 架构
+        self.assertTrue(_find_terms("双轨演进的抽象概念背景"))         # 双轨
+        self.assertTrue(_find_terms("技术选型决策树的抽象概念背景"))   # 决策树
+        self.assertEqual(_find_terms("山峰与攀登意象的氛围背景"), [])  # 隐喻放行
+        self.assertEqual(_find_terms("神经网络与幻灯片融合的全幅背景"), [])  # 装饰放行
+
+    def test_paraphrase_evasion_is_known_gap(self):
+        # 诚实边界：改写措辞规避结构词，lint 抓不到（锁死此已知缺口，避免误以为全覆盖）。
+        self.assertEqual(_find_terms(PROMPT_PARAPHRASE), [])
+
+    def test_bad_aesthetic_terms_hit(self):
+        hits = _find_bad_aesthetic_terms(PROMPT_BAD_AESTHETIC)
+        self.assertIn("abstract blueprint", hits)
+        self.assertIn("crosshair", hits)
+        self.assertIn("aegis", hits)
+        self.assertIn("shield", hits)
+        self.assertIn("18 thumbnails", hits)
+        self.assertIn("floating cards", hits)
+
+    def test_negative_aesthetic_terms_do_not_hit(self):
+        self.assertEqual(
+            _find_bad_aesthetic_terms("clean product scene, do not create a collage or thumbnail grid"),
+            [],
+        )
+        self.assertEqual(
+            _find_bad_aesthetic_terms("cinematic background without floating cards or UI cards"),
+            [],
+        )
+
+    def test_i2i_only_terms_hit(self):
+        hits = _find_i2i_only_terms(
+            "Use the attached image as the strict visual reference and preserve its main subject."
+        )
+        self.assertIn("use the attached image", hits)
+        self.assertIn("strict visual reference", hits)
+        self.assertIn("preserve its main subject", hits)
+
+    def test_background_text_marker_terms_hit(self):
+        hits = _find_background_text_marker_terms("light route background with chapter numbers 1 2 3 4")
+        self.assertIn("chapter numbers", hits)
+        hits_cn = _find_background_text_marker_terms("背景里画四个章节序号和文字标签")
+        self.assertIn("章节序号", hits_cn)
+
+    def test_negative_background_text_marker_terms_do_not_hit(self):
+        self.assertEqual(_find_background_text_marker_terms("clean background with no chapter numbers"), [])
+        self.assertEqual(_find_background_text_marker_terms("背景禁止章节序号和文字标签"), [])
+
+
+class SlideCopyInBackgroundTest(unittest.TestCase):
+    """slide_copy 背景检查（2026-07-03 ai_p9 事故）：背景 t2i prompt 携带页标题/正文/图表描述 → 违规；
+    前景模板的 Topic anchors 合法；i2i 一律豁免；负面清单（no charts/no tables）不误伤。"""
+
+    OLD_AI_P9_PROMPT = (
+        "low-interference data-space background. Slide title: 投入有限，对冲无限. "
+        "Core message: 核心总投入约 3 万元. Topic anchors: 投入有限, 万元. "
+        "Visual focus: 成本拆解 HTML 表 + 投入/对冲对比柱状图. no text, no logo."
+    )
+
+    def _lint_items(self, items):
+        p = write_json(items)
+        try:
+            return lint(p)[0]
+        finally:
+            os.remove(p)
+
+    def test_old_ai_p9_background_prompt_flagged(self):
+        violations = self._lint_items(
+            [{"file": "ai_p9.png", "aspect": "16:9", "target_layer": "background",
+              "prompt": self.OLD_AI_P9_PROMPT}]
+        )
+        hit = [v for v in violations if "slide title:" in v.get("terms", []) or "柱状图" in v.get("terms", [])]
+        self.assertTrue(hit, violations)
+
+    def test_foreground_topic_anchors_not_flagged(self):
+        violations = self._lint_items(
+            [{"file": "ai_p3_fg.png", "aspect": "1:1", "role": "foreground",
+              "prompt": "single complete product cutout. Topic anchors: 全景相机, 装配现场. no text, no logo."}]
+        )
+        self.assertEqual([v for v in violations if v.get("hint", "").startswith("背景生图提示词")], [], violations)
+
+    def test_i2i_exempt(self):
+        violations = self._lint_items(
+            [{"file": "enhanced_src_6.png", "mode": "image-to-image", "aspect": "16:9",
+              "prompt": self.OLD_AI_P9_PROMPT}]
+        )
+        self.assertEqual(violations, [], violations)
+
+    def test_negated_chart_terms_do_not_hit(self):
+        violations = self._lint_items(
+            [{"file": "ai_p2.png", "aspect": "16:9", "target_layer": "background",
+              "prompt": "clean industrial atmosphere, absolutely no text, no charts, no tables, "
+                        "no diagrams, no slide-like layout, no logo."}]
+        )
+        self.assertEqual(violations, [], violations)
+
+
+class LintTest(unittest.TestCase):
+    def test_build_bak_separation(self):
+        """img_2/3/5 结构图 FAIL，img_1/4 氛围 PASS（精确到 file）。"""
+        p = write_json(plan(PROMPT_COVER, PROMPT_ARCH, PROMPT_TRACKS, PROMPT_MOUNTAIN, PROMPT_TREE))
+        try:
+            violations, scanned, exists = lint(p)
+            self.assertTrue(exists)
+            self.assertEqual(scanned, 5)
+            flagged = {v["id"] for v in violations}
+            self.assertEqual(flagged, {"img_2.png", "img_3.png", "img_5.png"})
+        finally:
+            os.remove(p)
+
+    def test_outline_note_pass_catches_via_route(self):
+        """只扫 route=ai-A/ai-B（AI 凭空生成）的 note；source-enhanced（客户图 img2img）与 html 均豁免。
+
+        注：page 4「把客户截图增强成流程图背景」曾被断言要命中（"source-enhanced 也不能承载结构"），
+        但那是误伤——img2img 保原图内容，note 只是描述文字，实际产物是被保留的客户图；拦它只会逼「删
+        img_plan / 改 HTML」丢弃真实客户图。本用例已纠正为**放行 source-enhanced**（客户真图非 AI 凭空造）。
+        """
+        p = write_json(plan(PROMPT_COVER))  # img_plan 干净
+        o = write_json(outline(
+            ("ai-B", "封面渐变氛围"),               # 干净
+            ("ai-A", "抽象的五层架构分层示意图背景"),  # 架构 → 命中（ai-A 凭空生成）
+            ("source-enhanced", "客户图扩展为干净产品背景"), # 客户图 → 豁免
+            ("source-enhanced", "把客户截图增强成流程图背景"), # 客户图（img2img 保原图）→ 豁免，不再误伤
+            ("html", "用 HTML 画流程图对比"),          # html 路由，不扫
+        ))
+        try:
+            violations, scanned, _ = lint(p, o)
+            ids = {v["id"] for v in violations}
+            self.assertTrue(any("page 2" in i for i in ids))    # 仅 ai-A 凭空生结构 → 命中
+            self.assertFalse(any("page 3" in i for i in ids))   # source-enhanced 客户图豁免
+            self.assertFalse(any("page 4" in i for i in ids))   # source-enhanced 客户图豁免（曾误命中，本次纠正）
+            self.assertFalse(any("page 5" in i for i in ids))   # html 页不扫
+        finally:
+            os.remove(p)
+            os.remove(o)
+
+    def test_source_enhanced_structural_note_not_flagged(self):
+        # 真机复现：客户图 src_6 是「信息密集对比图」，route=source-enhanced、direct_contain 原比例。
+        # 曾被结构词「对比图」误判 FAILED → 整条 ppt 工作流 halt。修后放行（客户真图，非 AI 凭空生成）。
+        p = write_json(plan(PROMPT_COVER))  # img_plan 干净
+        o = write_json(outline(
+            ("source-enhanced", "src_6 信息密集对比图（传统多球机九宫格 vs 全景相机+AI 检测），direct_contain 原比例"),
+        ))
+        try:
+            code, out = run(p, o)
+            self.assertEqual(code, 0, out)
+            self.assertIn("PASSED", out)
+        finally:
+            os.remove(p)
+            os.remove(o)
+
+    def test_image_to_image_structural_prompt_not_flagged(self):
+        # img2img 增强客户图：prompt 描述真图内容含结构词（comparison chart）→ 豁免（真图非 AI 造结构）。
+        # 对照 test_violation_exit1（text-to-image 结构 → 仍命中 exit 1）：只放宽 image-to-image。
+        p = write_json([
+            {
+                "file": "enhanced_src_6.png",
+                "mode": "image-to-image",
+                "source_file": "src_6.png",
+                "aspect": "16:9",
+                "prompt": "Enhance the attached comparison chart, preserve its main subject, content and real colors.",
+            }
+        ])
+        try:
+            code, out = run(p)
+            self.assertEqual(code, 0, out)
+            self.assertIn("PASSED", out)
+        finally:
+            os.remove(p)
+
+    def test_missing_img_plan_passes(self):
+        code, out = run(os.path.join(tempfile.gettempdir(), "no_such_img_plan_xyz.json"))
+        self.assertEqual(code, 0)
+        self.assertIn("PASSED", out)
+
+    def test_clean_plan_exit0(self):
+        p = write_json(plan(PROMPT_COVER, PROMPT_MOUNTAIN))
+        try:
+            code, out = run(p)
+            self.assertEqual(code, 0)
+            self.assertIn("PASSED", out)
+        finally:
+            os.remove(p)
+
+    def test_directory_mode_prefers_design_json_before_outline_json(self):
+        with tempfile.TemporaryDirectory() as d:
+            img_plan = os.path.join(d, "img_plan.json")
+            design_path = os.path.join(d, "design.json")
+            outline_path = os.path.join(d, "outline.json")
+            with open(img_plan, "w", encoding="utf-8") as f:
+                json.dump(plan(PROMPT_COVER), f, ensure_ascii=False)
+            with open(design_path, "w", encoding="utf-8") as f:
+                json.dump(outline(("ai-A", "抽象的五层架构分层示意图背景")), f, ensure_ascii=False)
+            with open(outline_path, "w", encoding="utf-8") as f:
+                json.dump(outline(("ai-A", "封面渐变氛围")), f, ensure_ascii=False)
+
+            code, out = run(d)
+            self.assertEqual(code, 1)
+            self.assertIn("page 1", out)
+            self.assertIn("架构", out)
+
+    def test_violation_exit1(self):
+        p = write_json(plan(PROMPT_ARCH))
+        try:
+            code, out = run(p)
+            self.assertEqual(code, 1)
+            self.assertIn("FAILED", out)
+            self.assertIn("img_1.png", out)
+        finally:
+            os.remove(p)
+
+    def test_bad_aesthetic_prompt_exit1(self):
+        p = write_json(plan(PROMPT_BAD_AESTHETIC))
+        try:
+            code, out = run(p)
+            self.assertEqual(code, 1)
+            self.assertIn("FAILED", out)
+            self.assertIn("摄影感场景", out)
+        finally:
+            os.remove(p)
+
+    def test_text_to_image_cannot_use_i2i_prompt(self):
+        p = write_json([
+            {
+                "file": "ai_p9.png",
+                "mode": "text-to-image",
+                "aspect": "16:9",
+                "prompt": "Use the attached image as the strict visual reference: preserve its main subject.",
+            }
+        ])
+        try:
+            code, out = run(p)
+            self.assertEqual(code, 1)
+            self.assertIn("图生图提示词", out)
+            self.assertIn("ai_p9.png", out)
+        finally:
+            os.remove(p)
+
+    def test_text_to_image_cannot_generate_chapter_numbers(self):
+        p = write_json([
+            {
+                "file": "ai_p2.png",
+                "mode": "text-to-image",
+                "aspect": "16:9",
+                "prompt": "light blue industrial background with chapter numbers 1 2 3 4",
+            }
+        ])
+        try:
+            code, out = run(p)
+            self.assertEqual(code, 1)
+            self.assertIn("章节序号", out)
+            self.assertIn("不可编辑文字", out)
+        finally:
+            os.remove(p)
+
+    def test_text_to_image_negative_chapter_numbers_pass(self):
+        p = write_json([
+            {
+                "file": "ai_p2.png",
+                "mode": "text-to-image",
+                "aspect": "16:9",
+                "prompt": "clean route background with no chapter numbers and no text labels",
+            }
+        ])
+        try:
+            code, out = run(p)
+            self.assertEqual(code, 0, out)
+        finally:
+            os.remove(p)
+
+    def test_image_to_image_allows_i2i_prompt(self):
+        p = write_json([
+            {
+                "file": "enhanced_src_1.png",
+                "mode": "image-to-image",
+                "source_file": "src_1.png",
+                "aspect": "16:9",
+                "prompt": "Use the attached image as the strict visual reference: preserve its main subject.",
+            }
+        ])
+        try:
+            code, out = run(p)
+            self.assertEqual(code, 0, out)
+            self.assertIn("PASSED", out)
+        finally:
+            os.remove(p)
+
+    def test_generic_product_prompt_without_domain_anchor_fails(self):
+        p = write_json([
+            {
+                "file": "ai_p11.png",
+                "mode": "text-to-image",
+                "aspect": "16:9",
+                "prompt": "professional product photography of a sleek device on a dark premium background",
+            }
+        ])
+        try:
+            code, out = run(p)
+            self.assertEqual(code, 1)
+            self.assertIn("随机消费品", out)
+            self.assertIn("ai_p11.png", out)
+        finally:
+            os.remove(p)
+
+    def test_negative_random_object_constraints_pass(self):
+        p = write_json([
+            {
+                "file": "ai_p11.png",
+                "mode": "text-to-image",
+                "aspect": "16:9",
+                "prompt": "energy grid inspection scene, no game controller, no cosmetic bottle, no random monitor, no text",
+            }
+        ])
+        o = write_json({
+            "pages": [
+                {"page": 11, "title": "能源电网巡检", "visual_focus": "energy grid inspection scene"}
+            ]
+        })
+        try:
+            code, out = run(p, o)
+            self.assertEqual(code, 0, out)
+        finally:
+            os.remove(p)
+            os.remove(o)
+
+    def test_domain_bound_product_prompt_passes(self):
+        p = write_json([
+            {
+                "file": "ai_p11.png",
+                "mode": "text-to-image",
+                "aspect": "16:9",
+                "prompt": (
+                    "professional product photography of an energy grid inspection robot "
+                    "inside a substation operations room, no text, no logo"
+                ),
+            }
+        ])
+        o = write_json({
+            "pages": [
+                {
+                    "page": 11,
+                    "title": "能源电网巡检机器人",
+                    "one_sentence_takeaway": "变电站巡检机器人把现场状态同步给调度中心。",
+                    "visual_focus": "energy grid inspection robot inside a substation operations room",
+                    "visual": {"route": "ai-B", "note": "能源电网巡检机器人"},
+                }
+            ]
+        })
+        try:
+            code, out = run(p, o)
+            self.assertEqual(code, 0, out)
+        finally:
+            os.remove(p)
+            os.remove(o)
+
+    def test_non_aviation_domain_bound_prompts_pass(self):
+        p = write_json([
+            {
+                "file": "ai_p3.png",
+                "mode": "text-to-image",
+                "aspect": "16:9",
+                "prompt": "professional product photography of a medical imaging diagnostic workstation in a hospital reading room, no text, no logo",
+            },
+            {
+                "file": "ai_p4.png",
+                "mode": "text-to-image",
+                "aspect": "16:9",
+                "prompt": "professional product photography of a foldable phone hinge closeup for premium consumer hardware, no text, no logo",
+            },
+        ])
+        o = write_json({
+            "pages": [
+                {"page": 3, "title": "医疗影像诊断工作站", "visual_focus": "medical imaging diagnostic workstation hospital reading room"},
+                {"page": 4, "title": "消费硬件折叠铰链", "visual_focus": "foldable phone hinge closeup premium consumer hardware"},
+            ]
+        })
+        try:
+            code, out = run(p, o)
+            self.assertEqual(code, 0, out)
+        finally:
+            os.remove(p)
+            os.remove(o)
+
+    def test_explicit_unrelated_object_fails_even_with_other_words(self):
+        p = write_json([
+            {
+                "file": "ai_p7.png",
+                "mode": "text-to-image",
+                "aspect": "16:9",
+                "prompt": "energy grid operations background with a game controller in the center",
+            }
+        ])
+        try:
+            code, out = run(p)
+            self.assertEqual(code, 1)
+            self.assertIn("game controller", out)
+        finally:
+            os.remove(p)
+
+    def test_positive_thumbnail_grid_still_fails(self):
+        p = write_json([
+            {
+                "file": "ai_p8.png",
+                "mode": "text-to-image",
+                "aspect": "16:9",
+                "prompt": "energy grid operations page with a thumbnail grid layout in the center",
+            }
+        ])
+        try:
+            code, out = run(p)
+            self.assertEqual(code, 1)
+            self.assertIn("thumbnail grid", out)
+        finally:
+            os.remove(p)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)