# Contributors 按代码、测试、报告、词库任一贡献列。排名不分先后。 ## 代码与缺陷 - **[@AlaIchhe](https://github.com/AlaIchhe)** PR [#2](https://github.com/hongdougao111/dsh-bagu/pull/2) —— 修 AI 接梗一开就崩。 `bridge()` 从 `memes.js` 换成新实现后返回 `{hits, usage, tags}`, 但 AI 分支还按旧版读 `r.same.length`,AI 接梗开着又命中梗时必炸 `TypeError`。 真机上一天炸了 15 次,因为 `aiBridge` 默认关、条件短路,130 个测试全绿也没暴露。 他复现、定位到行、给了 `aiHintOf()` 的修法,还顺手发现 `bagu_meme` 那条 `ReferenceError: list is not defined`。 ## 词库 - **[梗Wiki](https://cnmeme.wiki)** —— 词库全部来自这里。 6800 多条词条的来源、用法、标签、别名、摘要,是那个站点一条一条攒出来的。 最值钱的 `usage` 字段(这个梗怎么接)也是它整理的。 ## 做法与判据 - **[thawline/dsh-memewiki](https://github.com/thawline/dsh-memewiki)** —— 「不等模型来查,先把答案摆在它面前」这个做法是它先做的,词库内联分发也是跟它学的。 - **[beagle style-criteria](https://github.com/existential-birds/beagle)** —— 代码注释判据(把代码念一遍的那类)。 - **[voku/code-slop](https://github.com/voku/agent-skills)** —— comments must explain why, never what the code already expresses. - **[sloplint](https://github.com/DW-Studio/sloplint)** —— 评分公式。 - **[lieflat-less-ai-tone](https://github.com/larashero3-dotcom/lieflat-less-ai-tone)** —— 两轮对照实验语料,1169 篇 / 520 万字。 ## 语料 回归测试里的真人语料,都是从公开页面摘的:孙笑川吧、v吧、B 站评论区、 git 仓库的 commit message。这些说话的人不知道自己在帮一个插件做校准, 但没有他们那 200 多条原话,误拦率是量不出来的。 --- 想加名字就提 PR。改了代码、补了测试、报了能复现的 bug、或者指出一条误拦, 都算。写清楚做了什么就行,不用客气。