Waza V3.33.0
Updated for stronger models, with less ceremony and sharper evidence.
Changelog
- Frontier Models: Keeps Waza aligned with how Claude, Codex, and other leading models work today by replacing obsolete approval rituals and over-prescribed steps with evidence-first judgment and conditional guidance.
- UI: Infers direction from the current product, screenshots, design tokens, and sibling components before asking, adds a generated-image asset mode, and records target, preserve, and evidence boundaries for repairs.
- Health on Windows: Makes audit launch reliable across Python, Git Bash, shim, and runtime variations while preserving report-only behavior.
- Health Evidence and Safety: Strengthens context evidence, interpreter identity, executable trust, report-only boundaries, and mirror, hotspot, and security signals so audits fail honestly.
- Check, Hunt, and Write: Loads detailed guidance only when relevant, strengthens evidence and deletion review, and preserves author voice and structural assets.
- Install and Packaging: Removes automatic update checks, makes installer guidance fail closed, and verifies packaged references and runtime paths so installation stays explicit and reproducible.
更新日志
- 前沿模型：持续跟进 Claude、Codex 等前沿模型的使用方式，用证据优先的判断和按需加载的指引，替代过时的确认仪式与过度规定的步骤。
- UI：先从当前产品、截图、设计 token 和同类组件推断方向，再决定是否提问；同时新增生成图片资产模式，并用 target、preserve、evidence 明确修复边界。
- Windows Health：兼容 Python、Git Bash、启动器和不同运行时组合，同时保持只报告、不执行修复的安全边界。
- Health 证据与安全：补强上下文证据、解释器身份、可执行文件信任、只报告边界，以及镜像、热点和安全信号，让审计能够诚实失败。
- Check、Hunt 与 Write：只在相关任务中加载详细指引，补强证据与删除审查，并在改写时保留作者口吻和结构资产。
- 安装与打包：移除自动更新检查，让安装指引默认安全失败，并校验打包后的引用和运行时路径，使安装过程保持显式且可复现。
Update: npx skills update -g -y · Claude Desktop · ⭐ tw93/Waza