dsh-plugin-dreaming 1.0.0

Dreaming: measurable, correctable "consciousness strength" for skills

dsh-plugin-dreaming is a global DeepSeek Harness plugin (Host + Web panel) for a concrete problem:
a skill description is a claim, and claims routinely disagree with delivery. The model loads the whole
body, then discovers the skill cannot do the job — tokens wasted.

It is not a logging plugin; it is a capability self-model:

• every skill load is concurrently probed with a cheap call that predicts what the skill can really
  deliver for this task, where it will mislead, and whether a cheaper path exists;
• it observes what actually happened (followed the skill, abandoned it, redid the work, token cost);
• at turn end a judge compares expectation with reality and extracts reusable rules;
• each skill keeps a capability belief plus a consciousness score (fit accuracy / boundary accuracy /
  precision / waste rate);
• the measured truth is injected before use (system-prompt section + skill-result annotation);
• a nightly dream revisits mismatches and safely edits user-level skill descriptions and boundary
  sections — backup first, journal, hash verification, conflict refusal, single-step revert.

No edits by default: dream.autoApply = false records suggestions only; enable it explicitly to let
dreams apply changes. The panel lives in Settings → Dreaming and shows consciousness, per-skill measured
history, recent calls, the dream journal and revert controls.

Install (DSH ≥ 0.2.1, from GitHub):

dsh plugin --profile <profile> add git+https://github.com/puppetcat-fire/dsh-plugin-dreaming.git

• Source and docs: https://github.com/puppetcat-fire/dsh-plugin-dreaming
• MIT; 261 unit tests; no test-framework dependency.
• Not published to npm yet; install straight from Git (npm publishing is planned).

Try it and tell us if the measured consciousness disagrees with your intuition — that is exactly the
signal we want.

---

Dreaming：让 skill 的“意识强度”可测量、可修正

dsh-plugin-dreaming 是一个 DSH 全局插件（Host + Web 面板），解决一个具体问题：skill 的描述是
“声明能力”，但声明和实际交付经常不一致——模型加载了整段正文，才发现它做不了这件事，token 白花。

它不是日志插件，而是能力自模型：

• 每次 skill 加载时并发做一次便宜的 probe，预测它对这个任务真正能交付什么、哪里会踩坑、
  有没有更便宜的替代路径；
• 观察加载之后真实发生的事（是否按 skill 执行、是否放弃/返工、token 代价）；
• 回合结束后 judge 预期 vs 现实，产出可复用规则；
• 每个 skill 维护能力信念与意识强度指标（fit 准确率 / 边界准确率 / 精确率 / 浪费率）；
• 在模型使用 skill 前，把实测真相注入上下文（系统提示段落 + skill 结果注解）；
• 夜间做梦：回到不匹配案例，安全地修正用户级 skill 的描述与能力边界——先备份、写日志、
  哈希校验、冲突拒绝、单次回滚。

默认不改文件：dream.autoApply = false，做梦只记录建议；要自动修正需显式打开。面板在
Settings → Dreaming，可查看意识强度、每个 skill 的实测历史、最近调用与梦境日志，并手动回滚。

安装（DSH ≥ 0.2.1，从 GitHub 安装）：

dsh plugin --profile <profile> add git+https://github.com/puppetcat-fire/dsh-plugin-dreaming.git

• 源码与文档：https://github.com/puppetcat-fire/dsh-plugin-dreaming
• MIT；261 个单元测试；无测试框架依赖。
• 目前未发布到 npm，直接从 Git 安装（npm 发布计划中）。

欢迎试用并反馈：如果它测出的“意识强度”和你的直觉不一致，那正是我们想知道的。