Repository navigation
v0.3.2
New in v0.3.2 — deterministic quality gates and authoring accelerators, distilled
from 58 automated production runs (65 human correction rounds):
- Three new strict gates:
text_render_audit.py(raw_/^, math labels, wrap,
container overflow),raster_policy_audit.py(tile registration, redrawable-class
paste, area budget),arrow_micro_geometry_gate.py(endpoint overshoot, head-only
fragments, corner spurs). review_checklist_gate.py --require-sweep: every review finding must carry a
defect class (A-G, seereferences/defect-taxonomy.md) and a page-wide
same-class sweep count.- Reviewer prompt: up to six orchestrator-selected crop pairs; raster-policy class
added; sweep fields required in findings. - SKILL.md rules 91-98: raster policy, subscript/wrap policy, arrow micro-geometry
contract, class sweep, post-repair verification, review resolution, findings
persistence, strict definition of done. scripts/schema_hints.py+references/scene-grammar-pitfalls.md: advisory
authoring-time grammar pre-check with exact correct spellings.references/component-cheatsheet.md: source-morphology to candidate-component
decision aid with discriminating features and renderer gotchas.tests/test_gates_v04.py: 22 unit tests covering the new gates.
Verified on two production-grade replay benchmarks: review recall on human
correction items improved 17%->83% / 67%->92% / 100% (no false-positive increase);
blind A/B preferred the gated pipeline; end-to-end time on the same figure dropped
143 -> 52 minutes across internal iterations with zero checklist regressions.
Visiomaster v0.3.1 更新说明
本次更新聚焦“箭头拓扑审查与重建闭环”。它不是新增某一类固定图形模板,而是让局部拓扑复杂、连接语义强的图在复刻、审查、修复和下一轮重建之间更可追踪。
核心更新
- 强化
metadata.arrow_plan:严格复刻前先逐条记录原图可见箭头的来源、目标、端点、路径形态、线型、箭头头和语义 intent。 - 强化 scene 绑定规则:可见 edge 需要通过
arrow_plan_id绑定原图箭头事实;一个 arrow plan 默认只能对应一条 scene/motif edge。 - 增加 motif 内部连线可审计能力:
nodes[].motif_edges[]可以声明内部 connector,并绑定arrow_plan_id,避免内部线条游离在审查体系之外。 - 增强 review 模板:
make_review_assets.py会从metadata.arrow_plan自动生成 topology checklist,让 reviewer 按箭头 id 逐项检查。 - 增强 repair brief:
review_findings_to_repair_plan.py保留checklist_refs,并生成arrow_plan_repair_targets,把问题映射到具体 edge、motif 和可编辑字段。 - 增强 regeneration packet:
prepare_regeneration_packet.py会带上 topology checklist、visual checklist、checklist validation 和 arrow repair targets,并展开到 Markdown prompt。 - 强化 validator/audit:新增 arrow-plan 覆盖率、多 edge 误用、端点/路径不匹配、local motif 规则和 motif edge 覆盖检查。
适用场景
本次优化适用于局部拓扑复杂、连接语义强、审查修复链路要求高的图形复刻任务,包括论文模块图、神经网络结构图、系统架构图、流程控制图和多分支数据流图等。
它重点解决几类通用问题:局部连接关系密集但整体相似度掩盖错误、箭头端点和路径形态不稳定、内部组件连线无法被审计、视觉审查结果难以映射回 scene 修改目标。
相比 v0.3.0
v0.3.0 建立了“原图/复刻图审查 -> rebuild brief -> regeneration packet”的完整复刻闭环;v0.3.1 把其中最容易失真的箭头拓扑进一步结构化:
source arrow inventory
-> metadata.arrow_plan
-> scene edge / motif edge binding
-> topology checklist
-> review findings
-> arrow_plan_repair_targets
-> regeneration prompt
这样下一轮 LLM 不只知道“箭头不对”,还会看到应该修改哪个 edge_id、哪个 motif_edges[] 绑定,以及需要保持的端点、路径和语义约束。
注意事项
motif_edges目前主要用于审计、映射和重建提示,不等同于所有 renderer 内部连线都已完全声明式渲染。- Python gate 负责发现结构和证据链问题,不能替代原图/复刻图的视觉审查。
- 本工具仍优先支持 Windows + Microsoft Visio 桌面版 +
pywin32。
验证情况
python -m compileall -q scripts tests sync_to_skill.py:通过python -m pytest -q:通过- 当前示例 scene 验证通过