Local security audit for AI API relays and LLM proxies: detects prompt injection, model substitution, tool-call rewriting, SSE anomalies, error leakage, and Web3 wallet risks.
Per-round reasoning_effort optimizer for DeepSeek Harness (dsh): auto-downgrades tool-call reasoning for simple tool chains, lifting back for heavy work. Cuts thinking time between tool calls.
把 WebUI 会话流中连续相邻的同工具调用合并为「主卡片+紧凑子行」树状展示,默认覆盖所有内置通用行工具 | Merges consecutive same-tool calls in the WebUI chat flow into a main-card + compact-child-rows tree view, all built-in generic tools by default
Results-only toggle for the DeepSeek Harness Web GUI: hides thinking and tool-call process, keeps one live status line while the agent works, with a per-turn trace line and a settings row. zh-CN/en.
Renders edit and write tool calls as always-expanded side-by-side git-style diffs in a diff viewer, with optional syntax coloring and word-level highlighting.
Animated desktop companion for DeepSeek Harness — live agent deck, mini-chat, streaming replies, model routing, context pressure and native commands, in a widget that docks to your screen edge.
Don't hand over the whole machine, just the decision blocking it: an unattended DeepSeek Harness (DSH) run's tool-call approvals and ask_user_question prompts reach your phone as Feishu cards — desktop first, outbound only.