Detects prompt injection by its effect on a sacrificial canary model, not just pattern matching: untrusted input hits a powerless model first, a behavioral check reads the residue, and it returns block, flag, or pass before your primary model acts. Inbound preflight sensor, not a guarantee.

DSH Web 输入体验增强插件：可配置发送/换行键位 + 自定义输入框右键菜单 + 面板尺寸记忆 + 快捷指令（分类 / 跨分类移动 / 关·每次·仅首次三种插入模式 / 存 ~/.dsh/quick-prompts.json）+ 提示词优化（0.11.1 起工具行有独立「优化」按钮、带秒表与斜杠命令保护、写回前比对草稿） + OpenCode 会话请求头 + Windows 默认终端切换（把模型的 PowerShell 工具换成 Git Bash）+ 官方持久终端六件套（terminal_open/r

Framework-aware code intelligence MCP server — 88 framework integrations, 81 languages, 72.7% fewer input tokens to review a pull request, comprehension at parity

Git status badges for DeepSeek Harness: an input-row chip on every install, plus sidebar session-row badges (git status, PR/CI action tokens) after one manual step: `npx dsh-git-badge apply`, then restart dsh web. Event-driven freshness over SSE.

Local dashboard for Claude Code and Codex CLI sessions: which Claude Code session needs input, every tool call and Claude Code subagent live, token cost and quota, several Claude accounts. Run npx ccdeck.

Edit images with Google Nano Banana 2 (image-to-image edit endpoint) on RunComfy. Documents Nano Banana Edit's strengths (preserve subject identity, swap background, localize edits with spatial language, multi-image batch edits up to 20 inputs), the schema, and when to route to GPT Image 2 edit / Fl