中文 | English
一个自己开发自己的 AI coding agent:桌面端里的 agent 按 backlog 连跑,把 kanzei 的下一个功能写出来、测完、提交、发版。你现在看到的大部分代码,就是它自己写的。
自用 AI coding agent,Rust 重写自 opencode(参考其 V2 架构),核心是自研 harness 系统。设计原则:好用压倒一切。
kanzei 深受它们启发,但走了几条它们没走的路:
| kanzei 的做法 | |
|---|---|
| 规则在代码,不在提示词 | 权限、文档格式、状态机、引用完整性全是硬门禁(注册表 + 拦截器):模型格式写错会被引擎拒绝并教它重试,而不是靠 system prompt 恳求 |
| 工单即文件 | 需求/缺陷/目标是纯 markdown,你手改、agent 用工具改,同一份文件;拖拽排序 = 改文件顺序,agent 按你排的顺序取活 |
| 订阅复用 + 多通道混跑 | Codex 订阅、Claude Code 订阅、Kimi API、本地 Ollama 在同一界面切换;并行子代理默认跑本地小模型,探索零成本 |
| 透明压倒一切 | 上下文占用进度条、压缩可召回、子代理实时轨迹、权限规则可查——agent 看到什么、花了什么,你都看得到 |
| 双人格 | 结伴开发(对话优先,拿不准就问)与自主推进(backlog 驱动,连跑不问)是两个 agent,不再互相打架 |
| 轻 | Rust + Tauri,安装包 5MB、二进制 19MB,无 Node 运行时 |
| 为一个人服务 | 没有企业功能、没有多租户:交互按主人的习惯生长,这是它敢叫"好用压倒一切"的原因 |
| Crate | 职责 |
|---|---|
kanzei-harness |
统一扩展层:六注册表(agents/tools/commands/skills/context-sources/permissions)+ 拦截器链硬门禁 + 双模式 profile(dev/research) |
kanzei-llm |
LLM 协议层:Anthropic Messages + OpenAI Chat 兼容(Ollama/LM Studio/DeepSeek/Kimi...),统一 LlmEvent 流,代理支持(loopback 豁免) |
kanzei-core |
session 运行时:runner 循环、调度(steer/queue)、SQLite 事件溯源、压缩 |
kanzei-tools |
内置工具:read(反向 seek tail)/ write(写后校验)/ bash(动态 shell 检测)... |
kanzei |
CLI 入口(kz);M3 起为 Tauri 桌面端的后端 |
安装包(推荐):从 Releases 下载 kanzei-setup-*.exe 双击安装(用户级,含开始菜单/卸载器)。应用启动会静默检查新版本,设置页可一键更新。
源码构建(开发机):
.\scripts\release.ps1 # test → release build → 安装 kz + kzapp 到 ~/.cargo/bin(带 pending 自更新)
.\scripts\package.ps1 -Publish # 打 NSIS 安装包并发布到 GitHub Releases(安装版的更新源)
kz --version桌面端(主要入口):kzapp — 类 VSCode 布局:项目管理 / 需求缺陷侧边栏(可展开、状态流转)/ 流式对话(工具块折叠)/ 模型直选 / 权限弹窗(拒绝/一次/总是)/ 运行日志面板 / 设置页。
CLI:kz run "任务",kz req|defect|source|finding list|add|...(人用直通)。
模型通道(~/.kanzei/kanzei.toml,设置页可视化编辑 + 一键测试):
codex:gpt-5.6-sol|terra|luna— 复用 Codex CLI 订阅登录态,自动刷新claude:claude-opus-5|sonnet-5|haiku-4-5— 复用 Claude Code 长效令牌(claude setup-token)kimi:kimi-k3等任意 OpenAI 兼容端点(key 支持环境变量或直填)ollama:<model>— 本地模型,loopback 自动免代理;fast 角色驱动并行 task 子代理
权限:Ruleset last-match-wins,默认 ask;弹窗选"总是允许"会把泛化规则(bash 取命令首词前缀)写进项目 .kanzei/kanzei.toml。
- ✅ M0-M2:三协议 LLM 层、harness 六注册表与双模式、SQLite 事件溯源、steer/queue 调度、会话持久化与回放、自动压缩
- ✅ 多 agent:task 并行子代理(fast/primary 双档)、双状态人格(结伴开发/自主推进)、连跑与目标驱动自举
- ✅ 桌面端:对话为主布局 + 右侧活动面板、历史对话管理、附件(图片/PDF)、todo/question 工具、发行版安装包与应用内更新
- ⏳ 进行中:多进程解耦(R-030)、并行对话线程与 worktree 合并(R-050)、MCP
日常开发由 kanzei 自举完成(dev agent 按 backlog 连跑),需求/缺陷全录在 .kanzei/project/。
设计规格参照:docs/design/harness-m1.md 与 docs/reference/(拷自 opencode 的 CONTEXT.md 与 specs/v2)。
An AI coding agent that develops itself: the agent inside the desktop app works through its own backlog — writing, testing, committing, and releasing kanzei's next feature. Most of the code you're reading was written by it.
Personal AI coding agent, rewritten in Rust from opencode (modeled on its V2 architecture). The core differentiator is a custom harness system: six registries (agents / tools / commands / skills / context sources / permissions) resolved into immutable snapshots, with hard gates enforced in code rather than prompts.
Design principle: usability above all — this is a personal daily-driver tool, optimized for transparent context management, multi-agent collaboration, speed, minimal interruptions, and clear information.
Deeply inspired by both — but it takes a few roads they don't:
- Rules live in code, not prompts — permissions, doc formats, state machines, and citation integrity are hard gates (registries + interceptors); a malformed tool call gets rejected with a repair hint instead of being begged away in the system prompt
- Tickets are files — requirements/defects/goals are plain markdown that you and the agent co-edit through the same engine; drag-to-reorder the list and the agent works top-down in your order
- Subscription reuse, mixed fleets — Codex login, Claude Code token, Kimi API, and local Ollama side by side; parallel subagents default to the free local model
- Transparency above all — context meter, recallable compaction, live subagent traces, inspectable permission rules: you always see what the agent sees and spends
- Two personas — pair-programming (conversation-first, asks when unsure) vs. autonomous (backlog-driven, never asks) are separate agents that no longer fight inside one prompt
- Light — Rust + Tauri: 5 MB installer, 19 MB binary, no Node runtime
- Built for one person — no enterprise features, no multi-tenant; the UX grows around its owner's habits
Download kanzei-setup-*.exe from Releases (per-user NSIS installer with uninstaller and Start Menu entry). The app silently checks for updates on startup; one-click update from Settings.
From source (dev machine):
.\scripts\release.ps1 # test → build → install kz + kzapp to ~/.cargo/bin (with pending self-update)
.\scripts\package.ps1 -Publish # build NSIS installer and publish to GitHub Releases- Three-protocol LLM layer: Anthropic Messages / OpenAI Chat / OpenAI Responses, streaming state machines with proxy support (loopback exempt)
- Subscription reuse: Codex CLI login (
codex:gpt-5.6-*), Claude Code long-lived token (claude:*), plus any OpenAI-compatible endpoint (Kimi, Ollama, ...) - Parallel read-only subagents (
tasktool, fast/primary tiers) with live progress streaming to an activity panel - Dual agent personas: pair-programming (conversation-first, default) vs. autonomous (backlog-driven with auto-continue), plus a research profile with enforced source citations
- Markdown-native project tracking: requirements / defects / goals as plain markdown with engine-enforced IDs, state machines, and archives — the agent develops kanzei itself off this backlog
- Event-sourced sessions (SQLite): steer/queue scheduling, restart recovery, conversation replay, automatic context compaction
- Tauri desktop app: conversation-first layout, activity sidebar, permission dialogs with persistent allow-rules, in-app doc viewer, priority/complexity-coded backlog
Chinese sections above are the source of truth; this summary tracks them loosely.
PolyForm Noncommercial 1.0.0 — 可自由使用、修改、分发,禁止商业用途。 Free to use, modify, and redistribute for noncommercial purposes only.