🔥 Reasonix：2.9k Stars 的 DeepSeek 原生 AI 编码 Agent，99.8% Cache 命中率一天省 $50 > 🚀 [esengine/DeepSeek-Reasonix](https://github.com/esengine/DeepSeek-Reasonix) | ⭐ 2,972 | 🛠 TypeScript | 📅 2026-04-21 --- 老实说，我试过在 DeepSeek 上跑 Claude Code 的替代方案，效果一言难尽——要么 API 绕着圈打水漂，要么会话一长 token 成本就起飞。Reasonix 这个项目有意思的地方在于：它**从第一行代码就是为 DeepSeek 的 prefix-cache 写的**，不是哪个通用框架魔改出来的。 项目才发布两周就冲到 2.9k stars，oosmetrics 上在 Agent 类别增速 Top 2、LLM 类别 Top 3。 ⚡ **核心卖点：Cache 稳定性是架构级的** Reasonix 不做多 provider 支持——它就绑定 DeepSeek，因为 prefix-cache 的字节级稳定性需要 loop 层的精心适配。实测数据：**435M 输入 token，99.82% cache hit，全天花费 ~$12**。换没有 cache 优化的客户端跑同样负载，v4-flash 上至少要 ~$61。五个小时的 LLM 活，省了 $50。 🛠 **一条命令上手** ```bash npm install -g reasonix reasonix code my-project ``` 首次运行粘贴 DeepSeek API key，key 持久化后可以用很久。不想全局装就用 npx： ```bash cd my-project npx reasonix code ``` 三种运行模式： | 命令 | 场景 | |------|------| | `reasonix code [dir]` | 编码 Agent（默认，有文件系统 + shell 工具） | | `reasonix chat` | 纯聊天模式（无文件写入权限） | | `reasonix run "task"` | 一次性任务，输出到 stdout，适合管道 | 🔧 **三大架构支柱** 1. **Cache 优先的推理循环**——每个工具调用都复用同一个 prefix，不产生新的 cache miss。系统提示、技能、记忆全部嵌入 prefix，不走二次加载 2. **工具调用修复**——当 DeepSeek 返回格式有偏差的 tool call 时，Reasonix 内部自动修正，不把异常抛给用户 3. **成本控制体系**——`/effort` 旋钮调节推理深度，`reasonix stats` 实时看 token 消耗和 cache 命中率 💡 **技能系统（不用远程注册中心）** ```bash # 创建项目级技能 /skill new my-skill # 创建全局技能（跨项目可用） /skill new my-skill --global ``` 技能文件就是 Markdown + frontmatter，支持 `runAs: subagent` 让技能跑在独立的子 Agent loop 里，不干扰主会话的上下文。 📡 **QQ 远程通道** Reasonix 支持把 QQ 作为 chat/code 会话的远程通信通道。先启动会话，然后在终端里输入： ``` /qq connect ``` 之后助手回复和确认提示都能从 QQ 继续，不需要一直守在终端前。 **要点总结** - DeepSeek 专用编码 Agent，99.8% prefix-cache 命中率，成本是通用方案的 1/5 - 一条 `npm install -g reasonix && reasonix code ./project` 就能开始用 - 内置 MCP、技能、记忆、Web 搜索、权限控制——不需要额外工具链 - 桌面客户端（Tauri）已出 prerelease，CLI 仍然是主力接口 - MIT 开源，社区驱动，`good first issue` 标签的工单对新贡献者友好 --- 🔥 Reasonix: 2.9k Stars DeepSeek-Native AI Coding Agent — 99.8% Cache Hit Rate Saves $50/Day > 🚀 [esengine/DeepSeek-Reasonix](https://github.com/esengine/DeepSeek-Reasonix) | ⭐ 2,972 | 🛠 TypeScript | 📅 2026-04-21 --- Honestly, I've tried running DeepSeek on generic agent harnesses—token costs balloon in long sessions and cache hits are hit-or-miss. Reasonix is different: **it was built from day one around DeepSeek's prefix-cache mechanic**, not retrofitted. Two weeks old, 2.9k stars. Ranked Top 2 in Agents and Top 3 in LLMs by growth velocity on oosmetrics. ⚡ **Cache Stability Is an Architectural Invariant** Reasonix is DeepSeek-only on purpose. Real-world data: **435M input tokens in one day, 99.82% cache hit rate, ~$12 total**. Same workload without cache optimization on v4-flash: ~$61. That's **$50/day saved** on a single heavy session. 🛠 **Start in One Command** ```bash npm install -g reasonix reasonix code my-project ``` Paste your DeepSeek API key on first run—it persists. Or use `npx` to try without installing: ```bash cd my-project npx reasonix code ``` Three modes: | Command | Use Case | |---------|----------| | `reasonix code [dir]` | Coding agent (filesystem + shell tools) | | `reasonix chat` | Plain chat (no file write access) | | `reasonix run "task"` | One-shot task, streams to stdout | 🔧 **Three Architecture Pillars** 1. **Cache-first loop** — every tool call reuses the same prefix. System prompts, skills, memory all embedded into the prefix, zero reloads. 2. **Tool-call repair** — auto-fixes malformed tool calls from DeepSeek instead of crashing. 3. **Cost control** — `/effort` knob adjusts reasoning depth, `reasonix stats` shows live token/cache metrics. 💡 **Skill System (No Remote Registry)** ```bash /skill new my-skill # project-level /skill new my-skill --global # cross-project ``` Skills are Markdown with frontmatter. `runAs: subagent` spawns an isolated sub-agent loop so heavy skills don't pollute the main context. 📡 **QQ Remote Channel** Start a session first, then connect from inside: ``` /qq connect ``` After that, follow-ups and confirmations continue through QQ—no need to stay glued to the terminal. **Key Takeaways** - DeepSeek-native coding agent with 99.8% prefix-cache hit rate—costs 1/5 of generic solutions - `npm install -g reasonix && reasonix code ./project` to get started - Built-in MCP, skills, memory, web search, permissions—no extra toolchain needed - Tauri desktop client available as prerelease; CLI remains the primary interface - MIT licensed, community-driven, `good first issue` tickets are beginner-friendly