feat: 公开定时任务页面、Word 导出、定时任务实时刷新与技能详情重构

- 新增 /public/scheduled-tasks/:taskId 公开只读访问页与对应的
  /api/public/scheduled-tasks 后端路由(无需登录)
- 新增对话整体 / 单条消息 / Markdown 工件 "另存为 Word (.docx)" 能力
- 定时任务列表与运行列表在任务执行中自动轮询刷新,并新增"正在执行"标记
- 定时任务运行消息接口在事件存储为空时回退到 LangGraph checkpoint,
  让正在执行的任务也能立即看到提示词与流式输出
- 技能详情对话框由折叠面板改为 Tab 切换(Detail / SKILL.md),
  并支持 Rendered / Raw 切换、一键复制
- config.yaml 增加 database (MySQL) 与 checkpointer (SQLite) 配置块

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
吉鑫升andClaude Opus 4.7 committed 2026-05-20 17:04:58 +08:00
1 parent 7e75c5b9b9
commit d339db4799
18 files changed
+1617 -125

No files matched your search

@@ -14,7 +14,16 @@ from fastapi import APIRouter, HTTPException, Query, Request
from fastapi.responses import FileResponse, Response, StreamingResponse
from pydantic import BaseModel, Field
from app.gateway.deps import get_current_user, get_local_provider, get_optional_user_from_request, get_run_event_store, get_run_store, get_tag_store
from app.gateway.deps import (
get_checkpointer,
get_current_user,
get_local_provider,
get_optional_user_from_request,
get_run_event_store,
get_run_store,
get_tag_store,
)
from deerflow.runtime import serialize_channel_values
from deerflow.runtime.scheduler import DuplicateTaskNameError, get_scheduled_task_service
router = APIRouter(prefix="/api/scheduled-tasks", tags=["scheduled-tasks"])
@@ -235,31 +244,129 @@ async def get_scheduled_task_run(run_id: str, request: Request) -> dict:
return run
def _event_type_for_checkpoint_message(message: dict) -> str:
"""Map a serialized LangChain message to the same ``event_type`` strings
the run-event store uses, so the frontend's existing role labels keep
working for checkpoint-derived synthetic rows."""
raw_type = message.get("type") or message.get("role") or ""
if raw_type in {"human", "user"}:
return "llm.human.input"
if raw_type in {"ai", "assistant"}:
return "llm.ai.response"
if raw_type == "tool":
return "llm.tool.result"
if raw_type == "system":
return "llm.system"
return f"llm.{raw_type or 'message'}"
async def _checkpoint_messages_for_run(
request: Request, agent_run: dict, *, limit: int
) -> list[dict]:
"""Return a synthetic list of message rows pulled from the LangGraph
checkpoint for an in-flight run.
The event store is updated by the streaming pipeline and may stay empty
until the first message tuple is flushed — meanwhile the checkpoint
already carries the current state (input + accumulating AI deltas). We
project that into the same row shape the frontend uses for the event
store, so the run-detail page can show progress immediately."""
try:
checkpointer = get_checkpointer(request)
except Exception:
return []
thread_id = agent_run.get("thread_id")
if not thread_id:
return []
config = {"configurable": {"thread_id": thread_id, "checkpoint_ns": ""}}
try:
tup = await checkpointer.aget_tuple(config)
except Exception:
return []
if tup is None:
return []
checkpoint = getattr(tup, "checkpoint", {}) or {}
channel_values = checkpoint.get("channel_values") or {}
try:
values = serialize_channel_values(channel_values)
except Exception:
return []
messages = values.get("messages") or []
rows: list[dict] = []
for index, message in enumerate(messages[-limit:]):
if not isinstance(message, dict):
continue
rows.append(
{
"seq": index + 1,
"event_type": _event_type_for_checkpoint_message(message),
"category": "message",
"content": message,
"metadata": {"source": "checkpoint"},
"created_at": None,
}
)
return rows
@router.get("/runs/{run_id}/messages")
async def get_scheduled_task_run_messages(run_id: str, request: Request, limit: int = 500) -> dict:
"""Agent messages produced by a scheduled run's underlying agent run.
Lets the run-detail page show what the agent is doing — useful when a run
is stuck in ``running``. Authorization reuses :func:`_load_run` (owner /
subscriber / admin); the event-store query then passes ``user_id=None``
because access has already been checked.
Resolution order (most authoritative first):
1. Event-store rows for the underlying agent run — these are the per-tuple
streaming events the worker writes as the agent runs.
2. LangGraph checkpoint snapshot — the canonical conversation state. Used
as a fallback when the event store is still empty (early in a run) or
when only partial events have been flushed; we project the snapshot
into the same row shape the frontend already renders so the user can
see the prompt + current AI output while the task is mid-flight.
3. Synthetic human-input row built from ``task.prompt`` — shown even
before the agent run is registered, so a `pending` task at minimum
displays what was queued.
"""
service = get_scheduled_task_service()
user_id = await _user_id(request)
run = await _load_run(service, request, run_id, user_id)
if run is None:
raise HTTPException(status_code=404, detail=f"Run {run_id} not found")
capped = max(1, min(limit, 1000))
# Always available: task prompt as the seeded human input. Lets the
# ``execution process`` tab show *something* even before the agent run is
# registered, which is the user-visible bug we're fixing.
fallback_rows: list[dict] = []
task_id = run.get("task_id")
if task_id:
task = await service.store.get_task_any(task_id)
prompt = (task or {}).get("prompt")
if prompt:
fallback_rows.append(
{
"seq": 0,
"event_type": "llm.human.input",
"category": "message",
# Frontend treats human-input content as a plain string.
"content": str(prompt),
"metadata": {"source": "task_prompt"},
"created_at": run.get("created_at") or run.get("scheduled_for"),
}
)
agent_run_id = run.get("agent_run_id")
if not agent_run_id:
return {"messages": [], "has_more": False}
return {"messages": fallback_rows, "has_more": False}
# Each scheduled run executes on its own throwaway runtime thread — NOT the
# run row's ``scheduler_thread_id`` — so resolve the real thread from the
# run store before querying its message events.
run_store = get_run_store(request)
agent_run = await run_store.get(agent_run_id, user_id=None)
if agent_run is None:
return {"messages": [], "has_more": False}
capped = max(1, min(limit, 1000))
return {"messages": fallback_rows, "has_more": False}
event_store = get_run_event_store(request)
rows = await event_store.list_messages_by_run(
agent_run["thread_id"],
@@ -267,6 +374,16 @@ async def get_scheduled_task_run_messages(run_id: str, request: Request, limit:
limit=capped + 1,
user_id=None,
)
# Event store has no rows yet (run just started, or events still buffering).
# Fall back to the live checkpoint so the user immediately sees the prompt
# + whatever AI output has been streamed so far.
if not rows:
checkpoint_rows = await _checkpoint_messages_for_run(request, agent_run, limit=capped)
if checkpoint_rows:
return {"messages": checkpoint_rows[:capped], "has_more": False}
return {"messages": fallback_rows, "has_more": False}
has_more = len(rows) > capped
return {"messages": rows[:capped] if has_more else rows, "has_more": has_more}