# PLAN: S4 — SSE 流式回复

## Goal
新增 SSE endpoint 实现 AI 回复的逐字流式传输。

## Current State
- S3: call_hermes() 使用 subprocess.run（阻塞，一次性返回全部输出）
- app.py: 已有 POST /api/conversations/<id>/messages（一次性回复）

## What to Build
1. 新增函数 `call_hermes_stream(user_input, timeout=120)` → generator yielding text chunks
   - 使用 subprocess.Popen 启动 `hermes chat -m <input>`
   - 用 stdout.readline() 逐行读取输出
   - 每读到一行就 yield 出去
   - 超时或结束时清理进程

2. 新增 SSE endpoint: `GET /api/conversations/<int:conv_id>/stream?message=<text>`
   - 使用 Flask Response + text/event-stream content type
   - 先保存用户消息到 DB
   - 构建历史上下文
   - 调用 call_hermes_stream() 生成流式输出
   - 每个 chunk 通过 SSE 发送: `data: <chunk>\n\n`
   - 流结束后发送 `data: [DONE]\n\n`
   - 在生成器中累积全文，流结束后保存 AI 回复到 DB
   - 需要设置 threading 让 DB 写入不阻塞主线程

## Files
- `app.py`: 新增 call_hermes_stream() + GET /api/conversations/<id>/stream
- `test_s4_sse.py`: 集成测试

## Data Flow
```
Browser (EventSource) → GET /stream?message=hello
  → Flask saves user msg to DB
  → call_hermes_stream(history + message)
  → Popen(["hermes", "chat", "-m", input])
  → for line in stdout: yield chunk → SSE to client
  → after loop: save full AI reply to DB
```

## Risks
- Flask's default WSGI server (Werkzeug) is single-threaded; SSE streaming blocks one worker
- Need to ensure DB writes don't conflict with streaming
- Popen process cleanup on client disconnect

## Verification
- Syntax check
- Unit test: mock Popen to verify SSE format
- Smoke test: all 8 pass