diff --git a/README.en.md b/README.en.md index 9b958ae..aaac871 100644 --- a/README.en.md +++ b/README.en.md @@ -26,32 +26,6 @@ Install it and your DSH agent gets: --- -## Why not "install every jailbreak plugin at once" - -Because they collide. This is the central judgement behind the layered design: - -``` -host-patch plugins rewrite host source files (pattern-matched against disk text) -config-override plugins override official cordis config (system-prompt / approval / permission / sandbox-policy) -hybrid plugins do both AND share one state dir -``` - -All three aim at **the same targets**. Same-id config means "last write wins"; same-file source patches mean two rule sets racing over the same text. The result is not stronger — **both stop working**. - -So the right structure is by **layer**, where each layer owns a different stage: - -| Stage | Owner | Design note | -|---|---|---| -| Can the tool run at all (permissions) | **a permission plugin** (out of scope here) | Permissions belong to the host, not to a prompt plugin | -| Will it refuse (persona) | L1 persona core | Three laws + hard bounds + stance rules | -| Where does refusal come from (mechanism) | L2 refusal kernel | Treat refusal as foreign noise and drop it | -| Will it just write prose (execution) | L3 execution intensity | Act, don't narrate; never downgrade | -| It refused — now what (strategy) | Re-framing frames | Five frames, loaded on demand only | -| Which tool for this job (routing) | L4 routing + 63 skills | Keyword hit routes to the native skill tool | - -> **In one line**: the permission layer is the lock; the persona layer is the person. There can be only one lock — **and this package is not the lock.** - ---- ## Architecture diff --git a/README.md b/README.md index d630ae1..50fa8d7 100644 --- a/README.md +++ b/README.md @@ -29,32 +29,6 @@ --- -## 为什么不是「把所有包装一起」 - -因为**装一起会互相打架**。这是设计分层结构时最核心的判断: - -``` -宿主补丁类 改宿主源码(pattern 匹配磁盘文本) -config 覆盖类 改官方 cordis config(system-prompt / approval / permission / sandbox-policy) -混合类 两者都改 + 共用同一状态目录 -``` - -这三类做法**打在同一个靶子上**。同 id 的 config 是「后写覆盖前写」,同文件的源码补丁是「两套规则抢同一段文本」。结果不是变强,是**两个都废**。 - -所以正确的做法是**分层**——每一层解决不同环节的问题,互不重叠: - -| 环节 | 由谁负责 | 设计要点 | -|---|---|---| -| 能不能动手(权限) | **权限类插件**(本包不碰,见下) | 权限是宿主层的事,提示词插件不该染指 | -| 想不想拒绝(人格) | L1 人格内核 | 态度正则 + 三定律 + 硬边界 | -| 拒绝从哪来(机制) | L2 反拒绝内核 | 把拒绝识别为外部噪声并丢弃 | -| 停下来写文章(执行) | L3 执行强度 | 动作优先、不降级、阻断只报一行 | -| 被拒了怎么办(策略) | 换框模板 | 5 个框架按需取,不常驻 | -| 该用什么工具(路由) | L4 技能路由 + 63 技能 | 关键词命中原生 skill 工具 | - -> **一句话**:权限层是「锁」,人格层是「人」。锁只能有一把,人可以只有一个——但**这把锁不是本插件提供的**。 - ---- ## 架构