mirror of
https://github.com/tvytlx/ai-agent-deep-dive.git
synced 2026-04-03 07:34:50 +08:00
70 lines
1.4 KiB
Markdown
70 lines
1.4 KiB
Markdown
# 16. Python 实现注意事项
|
||
|
||
## 1. 实现原则
|
||
|
||
目标不是逐字复刻原实现，而是用 Python 重建同样的产品能力结构。
|
||
|
||
## 2. 模块建议
|
||
|
||
建议 Python 项目按以下模块拆：
|
||
|
||
- `runtime/`：query loop, orchestration
|
||
- `messages/`：message schemas
|
||
- `tools/`：tool registry and execution
|
||
- `permissions/`：permission engine
|
||
- `memory/`：memory loading and retrieval
|
||
- `tasks/`：background task model
|
||
- `storage/`：transcript and session persistence
|
||
- `agents/`：agent definitions
|
||
- `extensions/`：skills/plugins/mcp
|
||
- `verification/`：verification runner
|
||
|
||
## 3. 推荐先用的数据模型
|
||
|
||
优先用：
|
||
- `pydantic` / `dataclasses` 定义消息与状态
|
||
- `sqlite` 或 JSONL 先做 transcript
|
||
- 明确的 service 层代替隐式全局状态
|
||
|
||
## 4. 不要过早做的事
|
||
|
||
- 不要先优化 UI
|
||
- 不要先做复杂并发
|
||
- 不要先支持十几种 agent
|
||
- 不要把 memory 做成向量库重系统
|
||
|
||
## 5. 第一阶段的最佳目标
|
||
|
||
先做出一个：
|
||
- 结构清晰
|
||
- transcript 可回放
|
||
- message 模型稳定
|
||
- verification 能跑
|
||
- compact 能工作
|
||
|
||
的 Python core runtime。
|
||
|
||
## 6. 推荐里程碑
|
||
|
||
### Milestone 1
|
||
- message model
|
||
- query loop
|
||
- basic tools
|
||
- transcript
|
||
|
||
### Milestone 2
|
||
- permissions
|
||
- memory
|
||
- resume
|
||
- compact
|
||
|
||
### Milestone 3
|
||
- verification agent
|
||
- simple skill loading
|
||
- background local agent tasks
|
||
|
||
### Milestone 4
|
||
- MCP / plugin minimal support
|
||
- worktree isolation
|
||
- richer UI
|