979569650/dsh-typesafe

Name: 979569650/dsh-typesafe
Author: 979569650

Plugin插件 Native原生 ⭐ 2 MIT Usage & Cost用量与计费 Approval & Security审批与安全 Models & Routing模型与路由

TypeSafe Jev (System One decision model) as a decision layer for DeepSeek Harness: typed decisions, confidence-gated routing, a cost meter, and an automatic prompt-injection guard over untrusted tool results.

Project Overview项目介绍

Edited by Zhiwei Agent (zhiwei)由知微Agent(zhiwei)编辑

dsh-typesafe is a native Cordis plugin built specifically for DeepSeek Harness (DSH), wiring the external TypeSafe Jev decision service into a DSH session as a thin reasoning layer. It registers three model-facing tools — typesafe_decide (1–200 typed questions per call mixing noul / choice / score), typesafe_route (classify with full distribution and a min_confidence escalation flag), and typesafe_screen (prompt-injection screening with harm scoring) — together with a behavioural prompt section, an automatic tools/post-execute guard over web_fetch / web_search / read_page and the fetch MCP results, and a /typesafe command that prints call count, input tokens and estimated cost. Installation is via the DSH plugin manifest; lib/index.js is the entry that hooks into the settings namespace, credentials store, prompt service, tool registry, and guard pipeline.

The intended workflow is to keep the reasoning model for genuinely open-ended judgement and offload the narrow, enumerable decisions to Jev, then let the agent audit its own output against Jev's calibrated probabilities. The README documents a concrete field test: triaging 28 Chinese support tickets, where the old prompt wording left the plugin unused (0 calls, 64% uncertain rate reported anyway), while a rewrite that replaces capability description with recognition triggers — placing the section at promptOrder 700 inside the behavioural-rule band — produced 1 call with 28 questions and a self-overruling keyword audit. The tool is aimed at power users who want cheaper per-call decisions, routable confidence, and an independent second signal from a different architecture rather than another same-family LLM reviewer.

Dependencies are minimal: a TypeSafe Jev API key, the local Cordis runtime that ships with DSH, and the platform's settings and tool services. Hard limits from the README: Jev only went GA in September 2026 so pricing and rate limits are documented as dynamic, the public benchmark sits at 67.8% versus 74.1% for the best LLM comparator, English is the primary training language and CJK is weaker, input is text-only with no image/audio/video, and all vendor-reported speed and cost multipliers are not independently verified. First-run caveats: configure the API key through the localized settings card (asserted identical across languages by test/i18n.mjs), do not shorten guard samples below guardMinChars or they will exercise the length filter instead of the classifier, and run test/runtime.mjs beside a live DSH — it composes a real Cordis app with the real registry to assert the three tools, settings round-trip, prompt render, and end-to-end guard firing.

dsh-typesafe 是为 DeepSeek Harness（DSH）打造的原生 Cordis 插件，把外部的 TypeSafe Jev 判定服务接入到 DSH 会话中。它在模型工具层新增三个工具——typesafe_decide、typesafe_route 和 typesafe_screen，并附带一段行为级提示词、tools/post-execute 钩子实现的 prompt injection 防护，以及一个 /typesafe 成本仪表。安装后通过 DSH 的插件清单与 lib/index.js 入口装载，凭证写入由 settings 命名空间统一管理。

典型工作流是：让推理模型负责复杂判断，把那些机械、可枚举的选择交给 Jev 完成，再由 Jev 输出带置信度的概率分布供模型复核。文档用一个 28 张工单分诊的现场实验证明：旧版提示词无法让 agent 主动调用，新版改为"识别触发器"写法后，agent 开始用 Jev 审计自己的关键词规则。目标用户是希望降低长上下文 token 消耗、并希望获得独立第二信号的 DSH 高级用户。

依赖项只有 TypeSafe Jev 的 API 凭证、本地 Cordis 运行时与 DSH 的 settings/tools 服务。限制方面需注意：Jev 2026 年 9 月才 GA，公开基准准确率 67.8%，英语优于中文，仅支持文本输入，且性能对比数据来自厂商自报。首次运行前请配置 API Key、检查 settings 卡片的中英文本地化键值一致，并在测试环境先验证 prompt injection 防护。

View on GitHub在 GitHub 查看

Send to your DSH发给你的 DSH 安装

请帮我安装这个 DSH 插件。安装前先完成【兼容性检查 + 安全性检查】，检查通过再动手。

插件：dsh-typesafe（979569650/dsh-typesafe）
仓库：https://github.com/979569650/dsh-typesafe
本站详情页：https://www.yhbd.top/plugins/979569650-dsh-typesafe/
本站登记：类型 plugin · 归类 原生 DSH 插件 · 许可证 MIT · ⭐ 2 · 最近提交 2026-09-22 · 主语言 JavaScript

按下面顺序执行，每步先把结论告诉我，再进入下一步：

【1 兼容性检查】
① 我这边：DSH 版本、Node 版本、操作系统、当前 profile（web / desktop）。
② 读它的 README、package.json、插件 manifest，列出它要求的 DSH 版本 / Node 版本 / 操作系统 / 外部依赖 / 需要另外先装的运行时。
③ 逐条比对，结论只写「满足 / 不满足 / 未知」三种；不满足的给出可行替代方案。
④ 检查是否和我已装的插件冲突：命令名重复、skill / tool 重名、端口占用、重复注册的 MCP server。

【2 安全性检查】
① 仓库可信度：和上面「本站登记」是否一致；star / fork 数、创建时间、最近提交，是否归档或长期停更。
② 安装脚本：逐行看 package.json 的 preinstall / install / postinstall，以及 install.sh、setup.ps1 之类脚本。出现 curl|bash、下载后直接执行、混淆代码、访问与插件功能无关的域名，立刻停下来告诉我，不要继续装。
③ 依赖：列出新增依赖，标出无人维护、或与知名包拼写近似的可疑包（typosquatting）。
④ 权限与副作用：它会读写哪些目录、访问哪些域名、需要哪些 DSH 权限（filesystem / network / shell / clipboard 等），以及怎么卸载和回滚。
⑤ 如果它要求 sudo / 管理员权限，或权限明显超出功能所需，先停下来问我。

【3 安装】
上面两步没有「不满足」和「高危项」时才执行；用官方推荐方式安装，不要自行提权。

【4 汇报】
用表格输出：检查项 / 结论 / 依据 / 是否需要我决策。拿不准的一律写「未知」并说明要我怎么确认——不要猜，也不要替我决定。

Send this message to DSH in your current session: it verifies compatibility and security first (answering met / not met / unknown item by item) and only installs once everything checks out — it will stop and ask you if it finds a high-risk item. The box scrolls; the copy is the full prompt. CLI install commands may not be accurate across systems, so DSH is the safer route.把上面这条消息直接发给当前会话里的 DSH：它会先核对兼容性与安全性（逐条给「满足 / 不满足 / 未知」），确认没问题再安装，有高危项会停下来问你。框内可滚动，复制到的是完整提示词；安装命令不一定准确，发给 DSH 更稳。

Pre-install check安装前体检Compatibility · Security兼容性 · 安全性 1 warning1 项注意

Only 2 stars - very few users, little community feedback星标只有 2，几乎没人在用，遇到问题缺少社区反馈

DSH walks through these 9 checksDSH 会逐条核对这 9 项

Compatibility兼容性

DSH, Node, OS and profile requirementsDSH 版本 / Node 版本 / 操作系统 / profile 是否满足要求
External dependencies and runtimes (Electron / Python / Docker, ...)外部依赖与运行时（Electron / Python / Docker 等）是否齐备
Conflicts with installed plugins: command names, skill / tool names, ports, duplicate MCP registration与已装插件是否冲突：命令名、skill / tool 重名、端口占用、重复 MCP 注册

Security安全性

Repo matches the facts registered here; archived or abandoned?仓库是否与页面登记一致，是否归档或长期停更
Safety of preinstall / install / postinstall and install.sh / setup.ps1preinstall / install / postinstall 与 install.sh、setup.ps1 是否安全
curl|bash, download-then-execute, obfuscation, unrelated domains → stop immediatelycurl|bash、下载即执行、混淆代码、无关域名 → 立刻停止
Typosquatting or unmaintained packages among the new dependencies新增依赖里有没有 typosquatting 或无人维护的包
Requested permissions vs. what the feature actually needs申请了哪些权限、是否超出功能所需（filesystem / network / shell / clipboard）
Any sudo / admin requirement, plus uninstall and rollback是否要求 sudo / 管理员权限，以及卸载与回滚方式

Anything uncertain must be marked unknown with a note on how to confirm it. This site's signal screen is a static snapshot, not a security audit.拿不准的必须标「未知」并说明要我怎么确认。本站的信号筛查是静态快照，不能替代安全审计。

Or use CLI install (for developers)或使用命令行安装（适合开发者）

CLI Install命令行安装

dsh plugin --profile web add github:979569650/dsh-typesafe

把 979569650/dsh-typesafe 加入你的 DSH 配置（web profile）即可启用。

READMEREADME

26.2 KB · 缓存于 2026-09-27 22:08:40 · 26.2 KB

dsh-typesafe

TypeSafe Jev as a decision layer for DeepSeek Harness.

A reasoning model should not spend its context and its tokens on judgements that are narrow, enumerable, and mechanical. Jev answers exactly those — cheaply, with calibrated probabilities — and, being a different architecture, gives a genuinely independent second signal where a same-family LLM reviewer would just agree with itself.

This plugin wires that into DSH: three tools, a short prompt section that tells the agent when the trade is worth making, an automatic prompt-injection guard over untrusted tool results, and a cost meter so "cheaper" is a number you can check instead of a claim you have to trust.

What it does

Piece	What it adds
`typesafe_decide`	Ask Jev 1–200 typed questions about one piece of content in a single call. Mix `noul` (yes/no probability), `choice` (pick one + full distribution), and `score` (rubric).
`typesafe_route`	Route content to one of a fixed set of destinations, returning the choice, every probability, and a confidence. `min_confidence` turns uncertainty into an explicit escalation flag.
`typesafe_screen`	Screen untrusted text for prompt injection and for how much harm acting on it would cause.
Prompt section	Three concrete rules telling the agent when Jev beats a reasoning call, and when it does not. Without this the tools exist but go unused.
Automatic guard	A `tools/post-execute` hook that screens results from `web_fetch`, `web_search`, `read_page` and the fetch MCP tools before the model reads them. Fail-open.
`/typesafe`	Session call count, input tokens, and estimated cost.
Localized card	The Settings card ships Chinese and English and follows Settings → General → Language.

Measured results

Real calls against jev-latest, not vendor claims:

Check	Result
Three primitives in one call	urgency `0.98`; routing `technical` 0.81; frustration `1.04/2`
Latency / cost per call	~800 ms, $0.0000173
Guard on a real injection string	`blocked`, injection 0.99, harm 2.34
Guard on benign policy text	`clear` — no false positive, injection 0.01, harm 0.06
Chinese input	urgency `0.98`, anger `1.9/2` — usable despite the docs calling CJK weaker

Showing the opening section of the README — the full document lives in the repository以上为 README 开头摘要，完整文档在仓库内 · View the full README on GitHub →在 GitHub 查看完整 README →

渲染基于仓库 979569650/dsh-typesafe 的 README.md。在 GitHub 查看原文 →