Plugins

Task-to-skill pairing meta-skill with a laziness ladder for minimal loading, a user-editable routing table, and a quarantine, SkillSpector scan, human-approval install flow.

Prompt-polish template injector for DSH: click ✨ in the composer to inject a "polish my prompt" template (current draft auto-embedded), no send, no network, no model — the main agent rewrites with full context. Replaces narrative-prompt-polish (2026-09-14 migration).

Scoped persistent memory for DSH: memory_save/search/replace/remove/list across global, user, project and failure scopes, DSH storage-domain persistence with secret and prompt-injection scanning, session-scope search, retention cleanup, and optional session-start injection; standing context via memory_pin/memory_pins/memory_unpin.

一个可直接链接到 DeepSeek Harness（DSH）Web profile 的独立安全检查插件。在安装或使用不受信任的 Skill 前，先执行浏览器本地粗检，再由用户决定是否调用 DSH 已配置的大模型进行结构化复核。

Task-to-skill pairing meta-skill with a laziness ladder for minimal loading, a user-editable routing table, and a quarantine, SkillSpector scan, human-approval install flow.

Scans skills and MCP configs for prompt injection, homoglyphs, hidden Unicode, dangerous shell, and credential leaks.

On-demand tool loadout for DSH: MCP tools are grouped by server and host plugin tools by prefix; grouped tools hide by default and activate with one call, while tools claimed by no rule stay visible. Ships preset modes (coding / writing / ops / review), a trial zone that promotes new tools after repeated use, gated groups that need user confirmation to activate, an outbound cut while reading the web as prompt-injection defense, and built-in search plus proxy call for long-tail tools.

Deterministic prompt-injection interception: tool results from file reads, repo searches and web fetches are scanned before they enter model context; high-severity findings quarantine-wrap or block, and credential-material files quarantine on the path alone.

基于 DeepSeek Harness 的 GitHub 插件静态安全审查预设，该项目代码由ai生成，主要框架由我搭建，通过“侦察员+审计员”双 Agent 协作、通过实验性ai自测攻防。主要功能有防死机上下文裁剪，有效防御投毒与提示词注入。包含21条安全红线、依赖树连坐排查与自动 YAML 校验，这是一名20岁非科班大学生的AI安全探索，欢迎大佬指导。本项目不能代替人工审计。请勿在生产环境或敏感凭证环境中使用。

AI Agent runtime authorization & evidence verification — tool-call GuardrailProvider, CCS 7-dimension verification standard, MCP/DSH security scanner, SSRF/command-injection/credential-exfil blocking with Ed25519 signed receipts.

Taints the agent when tool results carry untrusted content, gates the privileged calls that follow, and refuses credentials passed to network-capable tools in every mode.