HakiCat ScaffoldHakiCat Scaffold选择一个项目，用自然语言说明任务。Agent 可以查找代码、修改文件、执行命令和测试，也能委派子代理处理子任务。你可以查看工具执行记录，并按需开启计划审批、自测与改动审阅。
Choose a project and describe a task. The agent can inspect code, edit files, run commands and tests, and delegate subtasks. Inspect tool results and optionally enable plan approval, test feedback and change review.
17 项基础工具，加上 4 项由 Agent 集成的协作工具。当前可用工具会随权限、模式和 MCP 连接变化。
17 base tools plus 4 agent-integrated collaboration tools. Availability depends on permissions, modes and MCP connections.
以下是流程示意，并非实时运行日志。Agent 先读取上下文，再调用工具；工具结果回传模型，继续处理直到完成或达到循环上限。
Illustrative workflow, not a live execution log. The agent reads context, invokes tools and feeds results back to the model until the task finishes or the loop limit is reached.
计划、自测、审阅为可选模式，可按需要切换。
Planning, test feedback and review are optional modes.
开启 Plan 模式后，多步任务可以先提交计划；你批准后再执行，拒绝理由会回传模型。
Enable Plan mode to review a proposed plan before execution. Rejection feedback returns to the model.
开启自测模式后，文件改动可触发测试；失败输出回传模型进行修复。循环有上限，且需要本机安装项目所需工具。
When enabled, edits can trigger tests and failures feed back into repair attempts. The loop is bounded and requires the local project toolchain.
开启审阅模式后，可查看文件 diff，接受改动或恢复记录的文件快照。文件回滚不等于撤销命令的所有副作用。
Review file diffs and accept changes or restore recorded file snapshots. Restoring files does not undo every command side effect.
把可独立完成的子任务交给子代理；同步等待结果，或通过任务工具查看后台状态。效果与模型、任务和工作区上下文有关。
Delegate independent subtasks and wait for results or inspect background status. Outcomes depend on the model, task and available context.
在应用中配置 MCP 服务器的启动命令与参数，连接成功后将其工具加入可用工具表。
Configure an MCP server command and arguments in the app; its tools become available after connection.
支持 stdio 服务器。服务器需要的运行环境（例如 Node.js、Python）由本机提供；连接后工具会在后续工具查询中生效，无需重启应用。
Supports stdio servers. Install the server runtime, such as Node.js or Python, locally. Connected tools become available on subsequent tool queries without restarting the app.
这里展示的是接入能力，不代表已连接文件系统或数据库服务器。服务器故障会被记录并跳过，其他可用工具仍可使用；MCP 工具能访问哪些资源还取决于服务器自身配置。
This describes integration support, not an active filesystem or database connection. Failed providers are recorded and skipped. External tool access also depends on the server configuration.
下方为格式示例，耗时不是性能承诺。一次工具调用记成一条 Step：调用哪个工具、传了什么、返回什么、工具自身耗时多久、结果是否被截断。
The table below is illustrative, not a performance guarantee. A tool call becomes a Step: which tool, what it was given, what came back, how long the tool itself took, and whether the result was clipped.
| # | 工具 | Tool | 输入 / 结果 | Input / result | 耗时 | Duration |
|---|---|---|---|---|---|---|
| 1 | shell | echo NAVIGATION_OK | 34 ms | |||
| 2 | read | nav-a.md → 1.2 KB | 3 ms | |||
| 3 | read | nav-b.md → 1.4 KB | 2 ms | |||
| 4 | grep | NAVIGATION · . → 2 matches | 6 ms | |||
| 5 | edit | label.ts +2 −1 | 8 ms |
文件访问策略与命令执行的边界需要分别理解。
File access policies and command execution have different boundaries.
只读：适合阅读与分析代码；写入工具与命令执行工具不可用。
Read-only: Inspect and analyze code without write or command execution tools.
工作区写入：文件工具按工作区路径限制访问，命令从选定工作目录启动。
Workspace write: File tools enforce workspace paths; commands start in the selected working directory.
完全访问：允许更广的本机文件与命令访问，适合你明确需要的操作。
Full access: Allows broader local file and command access when your task requires it.
命令执行不是操作系统级隔离：shell 命令仍可能访问工作区之外的路径、网络或其他进程。审批策略控制是否等待确认；不会把命令变成隔离容器。
Command execution is not OS-level isolation: Shell commands may reach paths outside the workspace, networks or other processes. Approval controls confirmation, not container isolation.
本地桌面应用不等于离线模型：模型推理请求发送到你配置的 API 服务；对话、选中的文件内容和工具结果可能随请求发送。
A local desktop app does not imply offline inference: Requests go to your configured API. Conversations, selected file content and tool results may be sent with those requests.
下载安装程序，双击运行即可。装好后打开 HakiCat Scaffold，在「设置」里填 Base URL、API Key 和模型名，或直接登录哈基猫 API 账号——登录后会自动拉取可用密钥与模型列表。
Download the installer and run it. When HakiCat Scaffold opens, fill in the Base URL, API key and model name under Settings — or sign in with a HakiCat API account, which fetches your keys and model list automatically.
使用 Windows 64 位安装程序。已有下载链接保留；链接上的安装包版本以实际下载文件为准。
Use the Windows 64-bit installer. The existing download URL is retained; the package version is determined by the downloaded file.
在设置中填写 Base URL、API Key 和模型名称，或使用哈基猫 API 账号登录。建议填写服务商提供的完整 API 基础路径（例如 https://example.com/v1）；裸域名会补 /v1，特殊路径仍应按服务商文档填写。
Set the Base URL, API key and model name, or sign in with a HakiCat API account. Prefer the full API base path supplied by your provider, such as https://example.com/v1. Bare hosts default to /v1; use provider documentation for custom paths.
将工作区设为项目目录。第一次接触陌生项目可先用只读模式分析，再按任务开启写入与命令权限。
Set the workspace to your project directory. Start in read-only mode for unfamiliar projects, then enable editing and commands as needed.
例如：“找到设置保存逻辑，修复空模型名的问题，运行相关测试，并列出改动文件。”从一个明确的小任务开始，再根据结果调整模式。
For example: “Find settings persistence, fix empty model names, run the relevant tests and list changed files.” Start with a clear, scoped task and adjust modes based on results.
产品名 HakiCat Scaffold · 版本 1.0.0 · Copyright © 2026 HakiCat Product HakiCat Scaffold · v1.0.0 · Copyright © 2026 HakiCat
2026 年 10 月 12 日的测试记录，区分已验证与尚未覆盖。
Test records from October 12, 2026, with explicit coverage limits.
本地测试：全项目测试 384 项通过、0 项失败、21 项条件跳过；Windows 应用和 NSIS 安装程序构建成功。
Local tests: 384 passed, 0 failed and 21 conditionally skipped. Windows application and NSIS installer builds succeeded.
真实模型：使用 glm-5.3-flash 验证流式工具调用与子代理委派；子代理在测试目录中正确统计出 2 个 Go 文件。这是一个具体用例的结果，不代表所有任务都能成功。
Live model: Streaming tool calls and delegation were checked with glm-5.3-flash. A delegated task correctly counted 2 Go files in a fixture directory. This verifies that case, not every possible task.
未覆盖：需要账号或充值凭据的条件测试未执行；Windows 环境缺少 C 编译器，未运行 Go 竞态检测；真实外部 MCP 服务器的兼容性需要逐一联调。
Not covered: Credential-dependent account and top-up checks were skipped. Go race detection was not run because the Windows host lacked a C compiler. Real MCP servers require individual compatibility checks.