baozaotumao
df48c36422
把规则从原 ppt-gen 项目泛化为通用脚手架，覆盖 19 个文件： - ddd/architecture/design-discipline/error-codes/communication-* 等： 领域术语、错误码、消息示例、扩展点、状态流统一为「订单」示范域 - agent-dag/agent.md/agent-concurrency/agent-llm-*：DAG 节点、配置、并发、 LLM 适配从「对话→预览→生成→截图→合成」改为「intake→process→validate→finalize」 - backend-db：Session 实体/状态机/ORM/仓储/checkpoint 对齐新流水线 - 移除幻灯片专属依赖与产物：agent 去 playwright/python-pptx， Dockerfile 去 Chromium 系统库，.gitignore 去 *.pptx - 保留合法通用项：ANSI（PTY 处理）、Playwright（E2E 测试工具）、Gemini CLI（真实 provider） 验证：gemini-cli/openai-api 两变体生成退出 0，生成项目零 PPT 残留， backend/agent 骨架均可 import 运行。
160 lines
7.1 KiB
Markdown
160 lines
7.1 KiB
Markdown
---
|
||
paths:
|
||
- "agent/**"
|
||
---
|
||
|
||
# Agent DAG 编排与异常体系
|
||
|
||
SessionState 聚合根、DAG 节点表、runner 协调器（含 checkpoint 恢复与写入点）以及异常体系与传播规则。
|
||
|
||
> 下文以一条「会话→LLM 处理→结果」的通用流水线作示范，仅占位；节点与状态按你的业务替换，DAG 结构、checkpoint 与异常规则不变。
|
||
|
||
## DAG 编排（dag/）
|
||
|
||
### SessionState（domain/state.py，聚合根状态）
|
||
|
||
> 属领域层：是会话聚合根的状态载体，封装状态机不变量（如合法的 status 迁移），纯逻辑零 IO。
|
||
|
||
```python
|
||
@dataclass
|
||
class SessionState:
|
||
session_id: str
|
||
work_dir: Path
|
||
input_payload: dict | None = None
|
||
result_path: Path | None = None
|
||
status: Literal[
|
||
"idle", "intake", "processing", "validating",
|
||
"finalizing", "done", "revising", "error"
|
||
] = "idle"
|
||
error_message: str | None = None
|
||
```
|
||
|
||
### DAG 节点（dag/nodes.py）
|
||
|
||
节点签名统一：
|
||
```python
|
||
async def node_xxx(state: SessionState, send: Callable, recv_queue: asyncio.Queue) -> SessionState
|
||
```
|
||
|
||
| 节点 | 触发条件 | 行为 |
|
||
|------|---------|------|
|
||
| `session_start` | 收到 `start_session` | 创建 work_dir，启动 LLM 会话（PTY 或 API），就绪检查 |
|
||
| `intake` | 会话就绪后 | 收集/校验输入；用户 `user_input` → 转交 LLM |
|
||
| `process` | 输入就绪 | 调 LLM 产出结果，流式推 `log` 进度 |
|
||
| `validate` | `process` 完成 | 业务规则校验产出，不合规则回错误或重试 |
|
||
| `finalize` | `validate` 通过 | 落地最终结果文件，推 `result_ready` |
|
||
| `revise_loop` | `done` 状态下收到 `user_input` | 转交 LLM → 重进 `process` |
|
||
|
||
### DAG 协调器（dag/runner.py）
|
||
|
||
`runner.py` 有两个职责：**恢复入口**（连接建立时检查 checkpoint）和**正常流程**（串联所有节点）。
|
||
|
||
```python
|
||
async def run_session(state: SessionState, send, recv_queue: asyncio.Queue,
|
||
checkpoint: dict | None = None):
|
||
try:
|
||
# ── 断点恢复入口 ──────────────────────────────────────────────
|
||
if checkpoint:
|
||
resume_status = checkpoint.get("status")
|
||
|
||
if resume_status == "done":
|
||
# 完全恢复：直接推送已有结果，无需任何 LLM 调用
|
||
state.result_path = Path(checkpoint["result_path"])
|
||
await send({"type": "result_ready", "session_id": state.session_id})
|
||
await send({"type": "done"})
|
||
# 进入 revise 循环等待用户继续
|
||
await _revise_loop(state, send, recv_queue)
|
||
return
|
||
|
||
if resume_status in ("validating", "finalizing"):
|
||
# 中间产物已在磁盘，跳过 LLM，直接从校验节点恢复
|
||
state.result_path = Path(checkpoint["result_path"])
|
||
state.status = "validating"
|
||
logger.info("Resuming from checkpoint | status={}", resume_status)
|
||
state = await validate(state, send) # ← checkpoint 写入点 A
|
||
state = await finalize(state, send) # ← checkpoint 写入点 B
|
||
await send({"type": "done"})
|
||
await _revise_loop(state, send, recv_queue)
|
||
return
|
||
|
||
# intake / processing / error：无法恢复，LLM 会话已死
|
||
await send({"type": "checkpoint_lost",
|
||
"message": "上次会话在处理过程中中断，需要重新开始"})
|
||
|
||
# ── 正常流程 ──────────────────────────────────────────────────
|
||
state = await session_start(state, send, recv_queue)
|
||
state = await intake(state, send, recv_queue)
|
||
state = await process(state, send, recv_queue)
|
||
state = await validate(state, send) # ← checkpoint 写入点 A
|
||
state = await finalize(state, send) # ← checkpoint 写入点 B
|
||
await send({"type": "done"})
|
||
await _revise_loop(state, send, recv_queue)
|
||
|
||
except AgentException as e:
|
||
logger.error("DAG error | code={} msg={}", e.code, e.message)
|
||
await send({"type": "error", "code": e.code, "message": e.message})
|
||
except Exception as e:
|
||
logger.exception("Unexpected DAG error")
|
||
await send({"type": "error", "code": "INTERNAL_ERROR", "message": str(e)})
|
||
finally:
|
||
cleanup_session(state)
|
||
|
||
|
||
async def _revise_loop(state, send, recv_queue):
|
||
while True:
|
||
msg = await recv_queue.get()
|
||
if msg["type"] == "user_input":
|
||
state = await revise_loop(state, send, recv_queue, msg["text"])
|
||
state = await validate(state, send) # ← checkpoint 写入点 A
|
||
state = await finalize(state, send) # ← checkpoint 写入点 B
|
||
await send({"type": "done"})
|
||
```
|
||
|
||
### Checkpoint 写入点
|
||
|
||
`validate` 和 `finalize` 节点完成后，通过推送特定消息触发 Backend 写入 DB：
|
||
|
||
| 写入点 | 节点 | 推送消息（Backend 监听后写 DB） |
|
||
|--------|------|---------------------------------|
|
||
| A | `validate` 完成 | `{"type": "validated", "result_path": "..."}` |
|
||
| B | `finalize` 完成 | `{"type": "result_ready", "result_path": "..."}` |
|
||
| — | `done` | `{"type": "done"}` → Backend 将 status 更新为 `"done"` |
|
||
|
||
**checkpoint 写入由 Backend 负责（监听 Agent 推送的消息），Agent 只负责推送正确的消息。**
|
||
|
||
## 异常体系（domain/exceptions.py）
|
||
|
||
异常类型是**领域错误词汇**（统一语言的一部分，且 `code` 对应错误码契约 `error-codes.md`），故归领域层 `domain/exceptions.py`，由 domain 拥有。框架级的 `exception_handler` 注册属接口/基础设施关注点，放 `middleware/error_handler.py`，不污染 domain 的纯净（零框架/IO 依赖）。
|
||
|
||
### domain/exceptions.py
|
||
|
||
```python
|
||
class AgentException(Exception):
|
||
def __init__(self, code: str, message: str, status_code: int = 500):
|
||
self.code = code
|
||
self.message = message
|
||
self.status_code = status_code
|
||
|
||
class LLMStartupError(AgentException):
|
||
def __init__(self):
|
||
super().__init__("LLM_STARTUP_ERROR", "LLM 启动/连接失败")
|
||
|
||
class ExternalServiceTimeoutError(AgentException):
|
||
def __init__(self, phase: str):
|
||
super().__init__("EXTERNAL_SERVICE_TIMEOUT", f"外部依赖在 {phase} 阶段超时")
|
||
|
||
class ExternalServiceError(AgentException):
|
||
def __init__(self, detail: str):
|
||
super().__init__("EXTERNAL_SERVICE_ERROR", f"外部依赖返回错误: {detail}")
|
||
|
||
class OutputValidationError(AgentException):
|
||
def __init__(self):
|
||
super().__init__("INVALID_INPUT", "无法从输出中提取有效结果")
|
||
```
|
||
|
||
### 异常传播规则
|
||
|
||
- DAG 节点内捕获异常 → 更新 `state.status = "error"` → 通过 `send({"type": "error", "message": ...})` 推送给前端 → 抛给 runner
|
||
- Runner 捕获 → 记录日志 → 关闭 LLM 会话 → 清理资源
|
||
- **禁止在节点内静默 swallow 异常**
|