Doraemon Code

本项目是一个本地运行的 coding agent，主入口是 CLI，运行模式包括：

• plan
• review
• build

当前主链已经支持：

• OpenAI / Anthropic / Google 官方 SDK 直连
• OpenAI-compatible 上游
• built-in tools
• skills
• 本地扩展组和远端 MCP servers

System Overview

                    +------------------+
                    |   User / CLI     |
                    +---------+--------+
                              |
                              v
                    +------------------+
                    | Mode: plan/review|
                    |       /build     |
                    +---------+--------+
                              |
          +-------------------+-------------------+
          |                                       |
          v                                       v
 +--------------------+                 +--------------------+
 | Built-in Tools     |                 | Skills            |
 | read/edit/memory   |                 | prompt activation |
 | research/task      |                 +--------------------+
 +----------+---------+
            |
            v
 +--------------------+        +-----------------------------+
 | Agent Runtime      |<------>| MCP Extensions             |
 | state/trace/loops  |        | local groups + remote MCP  |
 +----------+---------+        +-----------------------------+
            |
            v
 +--------------------+
 | Model Providers    |
 | OpenAI/Anthropic   |
 | Google             |
 +--------------------+

Quick Start

git clone https://github.com/ifnodoraemon/dora-code.git
cd dora-code
pip install -e .
cp .env.example .env
mkdir -p .agent

最小配置示例：

{
  "model": "gpt-5.4",
  "openai_api_key": "YOUR_KEY",
  "openai_api_base": "https://www.packyapi.com/v1"
}

启动：

doraemon

Runtime Model

产品层 mode：

• plan
• review
• build

默认 built-in capability groups：

• plan

  • read
  • memory
  • research
  • task

• review

  • read
  • research
  • task

• build

  • read
  • edit
  • memory
  • research
  • task

对应工具名：

• read
  • read
  • search
  • ask_user
• edit
  • write
  • run
• memory
  • memory_get
  • memory_put
  • memory_search
  • memory_list
• research
  • web_search
  • web_fetch
• task
  • task

Browser / database 不在默认主链里，只能通过 MCP 扩展接入。

CLI Review And Policy Commands

• /review [--base REF] [paths...]：临时进入只读 review 模式，对工作区变更或指定路径做代码审查。
• /tools [all]：查看当前 mode 下工具可见性、审批要求和 sandbox 分类。
• /audit [limit]：查看最近工具策略检查和执行审计。

Provider Config

OpenAI-compatible：

{
  "model": "gpt-5.4",
  "openai_api_key": "YOUR_KEY",
  "openai_api_base": "https://provider.example.com/v1"
}

Anthropic-compatible：

{
  "model": "claude-sonnet-4-6",
  "anthropic_api_key": "YOUR_KEY",
  "anthropic_api_base": "https://provider.example.com/v1"
}

Anthropic 的 base URL 用户仍然填到 /v1，运行时会自动归一化，避免 SDK 拼成重复路径。

Google：

{
  "model": "gemini-3.1-pro-preview",
  "google_api_key": "YOUR_KEY"
}

MCP Config

远端 MCP server 配在 .agent/config.json 的 mcp_servers。

Context7：

{
  "mcp_servers": [
    {
      "name": "context7",
      "transport": "streamable_http",
      "url": "https://mcp.context7.com/mcp",
      "headers": {
        "CONTEXT7_API_KEY": "YOUR_CONTEXT7_API_KEY"
      },
      "tool_prefix": "context7",
      "require_approval": true,
      "enabled": true
    }
  ]
}

GitHub：

{
  "mcp_servers": [
    {
      "name": "github",
      "transport": "streamable_http",
      "url": "https://api.githubcopilot.com/mcp/",
      "headers": {
        "Authorization": "Bearer YOUR_GITHUB_TOKEN"
      },
      "tool_prefix": "github",
      "require_approval": true,
      "enabled": true
    }
  ]
}

建议总是设置 tool_prefix。远端 MCP 工具默认按敏感工具处理，需要审批；只有确认该 server 足够可信且工具无写入风险时，才设置 "require_approval": false。

当前远端 MCP runtime 支持：

• streamable_http
• stdio

并具备：

• fail-open
• 重试
• 会话复用
• 失败冷却
• 默认审批
• JSON-RPC 错误显式返回
• trace source 标记

Computer-Use Demo

受控浏览器 agent 示例在 examples/computer_use_agent：

python -m examples.computer_use_agent run \
  --goal "Mark this month's pending customers as followed up and export the result"

该示例会启动本地 CRM demo，浏览器导航限制在配置的 start URL origin，执行过程会写入 step trace、截图、下载文件和报告。

Real Validation

当前已真实验证过：

• Packy gpt-5.4
• GitHub MCP
• Context7 MCP

最小真实回归命令：

REAL_API_BASE='https://www.packyapi.com/v1' \
REAL_API_KEY='YOUR_KEY' \
REAL_MODEL='gpt-5.4' \
python3 scripts/run_evals.py --category basic

真实 benchmark：

REAL_API_BASE='https://www.packyapi.com/v1' \
REAL_API_KEY='YOUR_KEY' \
REAL_MODEL='gpt-5.4' \
python3 scripts/run_benchmark.py --suite real_repo --limit 5

Runtime Flow

user input
   |
   v
select mode
   |
   v
resolve built-in capability groups
   |
   +--> activate relevant skills
   |
   +--> attach enabled MCP extensions / servers
   |
   v
build model-visible tool list
   |
   v
LLM call
   |
   +--> text response
   |
   +--> tool call
           |
           v
      execute built-in or MCP tool
           |
           v
      append tool result
           |
           v
        next LLM turn

Documentation

• Architecture
• API
• Development
• Contributing