autodev
Feature auto-development — clarify requirements, generate spec, dispatch dev sub-agent to implement and merge into current branch
Use when asked about harness9 architecture, module design, or how components interact — explains the system design
$ npx -y skills add zhangshenao/harness9 --skill architecture-overview --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/architecture-overviewContext preview
The summary Claude sees to decide when to auto-load this skill.
Use when asked about harness9 architecture, module design, or how components interact — explains the system design
name: architecture-overview description: Use when asked about harness9 architecture, module design, or how components interact — explains the system design trigger: architecture, design, how does, module, component, structure
| 原则 | 说明 | |------|------| | **简洁** | 最小化抽象层,极少的直接依赖 | | **完备** | 覆盖 Agent 运行所需的全部核心模块 | | **生产可用** | 错误恢复、超时控制、路径沙箱、并发安全 |
Turn N: LLM(messages + tools) → 推理 + 工具调用决策 → 并发执行所有工具调用 → Observation 注入 context → Turn N+1 自然终止:模型不再发起工具调用 → 输出最终回复
三重终止保障: 1. **自然终止**:`len(responseMsg.ToolCalls) == 0` 2. **MaxTurns**:默认 50,可通过 `WithMaxTurns` 配置 3. **Context 取消**:外部 `cancel()` 或超时
cmd/harness9 (入口)
├── internal/context (System Prompt 组装)
│ └── internal/skills (Skills 解析 + 索引)
├── internal/engine (ReAct 主循环)
│ ├── internal/provider (LLM 调用)
│ ├── internal/tools (工具注册 + 执行)
│ └── internal/schema (数据类型)
└── internal/env (配置加载)**关键设计决策:接口定义在使用者侧**
用户输入 → RunTUI → eng.RunStream(ctx, prompt)
→ engine.Event stream → 逐 token 追加到对话视图
→ ToolCalls → Spinner 动画 + 耗时计数
→ EventDone → 最终回复渲染到屏幕用户输入 → RunCLI → eng.Run(ctx, prompt)
→ runLoop → LLM Generate
→ ToolCalls → 并发执行 → ToolResults → 继续循环
→ 最终回复打印到 stdout`DefaultPromptBuilder.Build()` 按顺序组装:
1. **基础 Prompt**:角色定义 + workDir 2. **AGENTS.md**:项目级规范(文件不存在时跳过) 3. **Skills 索引**:`- name: description` 列表(为空时跳过)
完整内容示例:
You are harness9, an expert coding assistant...
## Project Guidelines (AGENTS.md)
{AGENTS.md 全文}
## Available Skills
Use the `use_skill` tool to load full content of any skill when needed.
- go-coding-standards: Use when writing or reviewing Go code...
- debugging-guide: Use when debugging Go errors...type LLMProvider interface {
Generate(ctx, messages, tools) (Message, error)
GenerateStream(ctx, messages, tools) (<-chan StreamChunk, error)
}当前实现:
**Anthropic 约束**:user/assistant 消息必须严格交替,禁止连续 assistant 消息。
type BaseTool interface {
Name() string
Definition() schema.ToolDefinition // JSON Schema,传给 LLM
Execute(ctx context.Context, args json.RawMessage) (string, error)
}内置工具:
| 工具 | 说明 | |------|------| | `bash` | Shell 命令执行,workDir 为 CWD | | `read_file` | 文件读取,4096 字节截断 | | `write_file` | 文件写入,自动 mkdir | | `edit_file` | 字符串替换编辑,多级模糊匹配 | | `use_skill` | 按需加载 Skill 全文 |
所有文件工具通过 `safePath()` 校验路径,防止 Path Traversal 攻击。
Local-First · Lightweight · Complete · Production-Ready general-purpose Agent framework, written in Go.
Feature auto-development — clarify requirements, generate spec, dispatch dev sub-agent to implement and merge into current branch
Use when the user invokes /commit or asks to commit changes, after a code review has been completed and the changes are confirmed…
Use when the user invokes /cr, requests a code review, or before committing to verify correctness, security, and quality of new…
Use when debugging Go errors, test failures, or unexpected behavior — step-by-step diagnosis approach
Use when writing or reviewing Go code — explains harness9 project coding conventions and patterns
Use when the user invokes /pr or asks to push changes and open a pull request, after commits are ready to be pushed to a remote…