功能迁移,页面整体样式修改

This commit is contained in:
董宏凯 committed 2026-05-21 16:30:11 +08:00
1 parent 04664c8d42
commit bfd75da308
41 files changed
+2232 -1920

No files matched your search

@@ -1,15 +1,3 @@
from .graph import (
AI_WRITING_GRAPH,
build_ai_writing_graph,
get_ai_writing_checkpointer,
get_ai_writing_graph,
shutdown_ai_writing_graph,
)
from .graph import make_ai_writing_graph
__all__ = [
"AI_WRITING_GRAPH",
"build_ai_writing_graph",
"get_ai_writing_graph",
"get_ai_writing_checkpointer",
"shutdown_ai_writing_graph",
]
__all__ = ["make_ai_writing_graph"]
@@ -3,6 +3,7 @@ from __future__ import annotations
import logging
from langchain_core.runnables import RunnableConfig
from langgraph.graph import END, StateGraph
from langgraph.types import interrupt
@@ -15,6 +16,30 @@ from .nodes.editor import editor_node
logger = logging.getLogger(__name__)
async def _mark_session_done(config: RunnableConfig | None) -> None:
"""把业务表行 status 写成 ``done``。
pause_draft_confirm action=finalize / pause_review_confirm force_finalize 时调,
让 LangGraph Server 路径下新会话有正确终态(老路径靠 ``_stream_graph_segment``
finally 块写终态,新路径完全不走它)。失败不影响 graph 流程。
"""
if not config:
return
thread_id = (config.get("configurable") or {}).get("thread_id")
if not thread_id:
return
from deerflow.persistence.ai_writing_sessions import get_default_store
from deerflow.runtime.user_context import AUTO
repo = get_default_store()
if repo is None:
return
try:
await repo.update(thread_id, user_id=AUTO, status="done")
except Exception:
logger.exception("ai_writing: mark_session_done failed (non-fatal) thread_id=%s", thread_id)
# ── 干预节点 ──────────────────────────────────────────────────────────────────
async def pause_material_confirm(state: AIWritingState) -> dict:
@@ -127,7 +152,7 @@ async def pause_outline_confirm(state: AIWritingState) -> dict:
return updates
async def pause_draft_confirm(state: AIWritingState) -> dict:
async def pause_draft_confirm(state: AIWritingState, config: RunnableConfig) -> dict:
"""PAUSE-3:草稿确认"""
draft = state["current_draft"]
user_input: dict = interrupt({
@@ -141,6 +166,7 @@ async def pause_draft_confirm(state: AIWritingState) -> dict:
action = user_input.get("action", "to_editor")
if action == "finalize":
await _mark_session_done(config)
return {
"last_user_intervention": user_input,
"status": "done",
@@ -223,7 +249,7 @@ async def pause_section_help(state: AIWritingState) -> dict:
}
async def pause_review_confirm(state: AIWritingState) -> dict:
async def pause_review_confirm(state: AIWritingState, config: RunnableConfig) -> dict:
"""PAUSE-4:编辑打回确认"""
user_input: dict = interrupt({
"pause_point": "review_confirm",
@@ -234,6 +260,7 @@ async def pause_review_confirm(state: AIWritingState) -> dict:
})
if user_input.get("overrideVerdict") or user_input.get("action") == "force_finalize":
await _mark_session_done(config)
return {
"last_user_intervention": user_input,
"status": "done",
@@ -299,25 +326,26 @@ def after_review_pause(state: AIWritingState) -> str:
return "end" if state["status"] == "done" else "writer_draft"
# ── 组装 Graph ────────────────────────────────────────────────────────────────
def build_ai_writing_graph(checkpointer=None):
"""构建并 compile AI 写作 graph。
# ── LangGraph Server 入口 ────────────────────────────────────────────────────
#
# compile 出裸 graph(不带 checkpointer),由 LangGraph Server 注入主 checkpointer
# (langgraph.json 里 ``make_checkpointer`` 配的那个)。AI 写作的 thread 状态与
# 主聊天共享同一份 saver —— LangGraph schema 用 ``(thread_id, checkpoint_ns)`` 做
# 主键,UUID 不会撞。
def make_ai_writing_graph(config=None):
"""LangGraph Server graph factory:不接收 checkpointer,由 Server 注入。
Args:
checkpointer: 可选的 LangGraph checkpointer。
- ``None`` → 退回 ``MemorySaver`` —— 仅供测试 / 开发使用。
进程或容器重启后会话状态全部丢失,前端拿着旧 session_id 调
``/resume`` 会得到 404。
- 其他 → 传入的 (Async)BaseCheckpointSaver 实例,用于持久化。
config: LangGraph Server 调用工厂时注入的 RunnableConfig(暂未消费)。
"""
g = StateGraph(AIWritingState)
g.add_node("intent_parser", intent_parser_node)
g.add_node("researcher", researcher_node)
g.add_node("pause_material", pause_material_confirm)
g.add_node("writer_outline", writer_outline_node)
g.add_node("pause_outline", pause_outline_confirm)
g.add_node("intent_parser", intent_parser_node)
g.add_node("researcher", researcher_node)
g.add_node("pause_material", pause_material_confirm)
g.add_node("writer_outline", writer_outline_node)
g.add_node("pause_outline", pause_outline_confirm)
g.add_node("writer_draft", writer_draft_node)
g.add_node("pause_section_help", pause_section_help)
g.add_node("pause_draft", pause_draft_confirm)
@@ -325,214 +353,23 @@ def build_ai_writing_graph(checkpointer=None):
g.add_node("pause_review", pause_review_confirm)
g.set_entry_point("intent_parser")
g.add_edge("intent_parser", "researcher")
g.add_edge("intent_parser", "researcher")
g.add_conditional_edges("researcher", after_researcher,
{"pause_material": "pause_material", "writer_draft": "writer_draft"})
{"pause_material": "pause_material", "writer_draft": "writer_draft"})
g.add_conditional_edges("pause_material", after_material_pause,
{"researcher": "researcher", "writer_outline": "writer_outline"})
{"researcher": "researcher", "writer_outline": "writer_outline"})
g.add_edge("writer_outline", "pause_outline")
g.add_conditional_edges("pause_outline", after_outline_pause,
{"writer_outline": "writer_outline", "writer_draft": "writer_draft"})
{"writer_outline": "writer_outline", "writer_draft": "writer_draft"})
g.add_conditional_edges("writer_draft", after_writer_draft,
{"pause_section_help": "pause_section_help", "pause_draft": "pause_draft"})
{"pause_section_help": "pause_section_help", "pause_draft": "pause_draft"})
g.add_conditional_edges("pause_section_help", after_section_help_pause,
{"researcher": "researcher", "writer_draft": "writer_draft"})
{"researcher": "researcher", "writer_draft": "writer_draft"})
g.add_conditional_edges("pause_draft", after_draft_pause,
{"end": END, "writer_draft": "writer_draft", "editor": "editor"})
{"end": END, "writer_draft": "writer_draft", "editor": "editor"})
g.add_conditional_edges("editor", after_review,
{"pause_draft": "pause_draft", "pause_review": "pause_review"})
{"pause_draft": "pause_draft", "pause_review": "pause_review"})
g.add_conditional_edges("pause_review", after_review_pause,
{"end": END, "writer_draft": "writer_draft"})
{"end": END, "writer_draft": "writer_draft"})
if checkpointer is None:
# MemorySaver 仅作为 fallback —— 提示一下,正式部署应通过
# get_ai_writing_graph() 拿到带持久化 checkpointer 的 graph。
from langgraph.checkpoint.memory import MemorySaver
checkpointer = MemorySaver()
return g.compile(checkpointer=checkpointer)
# ── 持久化 checkpointer 单例(懒加载)─────────────────────────────────────────
# 模块级 graph 槽位。import 时为 None;首次 ``get_ai_writing_graph()`` 调用时
# 按 config.yaml 的 checkpointer 段懒构建出真正的 graph。
# 测试可通过 ``monkeypatch.setattr(graph_mod, "AI_WRITING_GRAPH", stub)``
# 直接注入桩;本模块所有访问点都走 ``get_ai_writing_graph()`` 这一入口,会优先
# 返回 module attribute 的值,保留测试桩注入路径。
AI_WRITING_GRAPH = None
# 持有 async checkpointer(底层 aiosqlite Connection 或 psycopg pool 的所有者)。
# 进程生命周期内常驻;shutdown 时由 ``shutdown_ai_writing_graph`` 显式关闭。
_async_checkpointer_singleton = None
def _fallback_to_memory_saver(reason: str):
"""退回 InMemorySaver 的统一兜底入口。
把 reason 打到 warning 日志里方便运维定位。InMemorySaver 进程重启会丢
状态,但**当次写作能跑完** —— 我们的首要目标是「写作流程能继续」,
持久化是锦上添花。
"""
from langgraph.checkpoint.memory import InMemorySaver
logger.warning(
"ai_writing: 持久化 checkpointer 不可用,退回 InMemorySaver。原因: %s",
reason,
)
return InMemorySaver()
async def _get_async_checkpointer():
"""按 config.yaml 的 checkpointer 段返回一个 async checkpointer 实例。
**绝不抛异常** —— 任何失败一律退回 InMemorySaver。这是为了保证「写作
流程能继续」这个首要目标:持久化坏了用户至少还能写完当次稿,重启会丢
会话但不影响当下创作。
优先级:
1. ``checkpointer:`` 段(``type=sqlite|postgres|memory``)
2. fallback 到 ``InMemorySaver`` 并打 warning
返回的 saver 与主 agent 公用同一 sqlite 文件 —— LangGraph 的 schema
用 ``(thread_id, checkpoint_ns)`` 做主键,AI 写作的 thread_id 是
会话 uuid,跟主 agent 的 LangGraph thread_id 不会撞。
"""
global _async_checkpointer_singleton
if _async_checkpointer_singleton is not None:
return _async_checkpointer_singleton
try:
from deerflow.config.checkpointer_config import get_checkpointer_config
cp_cfg = get_checkpointer_config()
except Exception:
logger.exception("ai_writing: 加载 checkpointer 配置失败")
_async_checkpointer_singleton = _fallback_to_memory_saver("加载 config 抛异常")
return _async_checkpointer_singleton
if cp_cfg is None:
_async_checkpointer_singleton = _fallback_to_memory_saver(
"config.yaml 未配置 checkpointer 段"
)
return _async_checkpointer_singleton
if cp_cfg.type == "memory":
from langgraph.checkpoint.memory import InMemorySaver
_async_checkpointer_singleton = InMemorySaver()
logger.info("ai_writing: 配置 type=memory,使用 InMemorySaver")
return _async_checkpointer_singleton
if cp_cfg.type == "sqlite":
try:
import aiosqlite
from langgraph.checkpoint.sqlite.aio import AsyncSqliteSaver
from deerflow.runtime.store._sqlite_utils import (
ensure_sqlite_parent_dir,
resolve_sqlite_conn_str,
)
conn_str = resolve_sqlite_conn_str(
cp_cfg.connection_string or "ai_writing_checkpoints.db"
)
ensure_sqlite_parent_dir(conn_str)
# 直接持有 aiosqlite.Connection —— 进程生命周期,shutdown 时显式关。
conn = await aiosqlite.connect(conn_str)
saver = AsyncSqliteSaver(conn)
await saver.setup()
_async_checkpointer_singleton = saver
logger.info("ai_writing: 使用 AsyncSqliteSaver (%s)", conn_str)
return saver
except Exception as exc:
# sqlite 失败原因可能是:依赖未装 / 磁盘满 / 文件权限 / 路径不存在 /
# 文件被外部进程独占锁。任何一个都不能让 AI 写作整个不可用。
logger.exception(
"ai_writing: AsyncSqliteSaver 初始化失败,conn_str=%r",
getattr(cp_cfg, "connection_string", None),
)
_async_checkpointer_singleton = _fallback_to_memory_saver(
f"sqlite 初始化失败: {type(exc).__name__}: {exc}"
)
return _async_checkpointer_singleton
if cp_cfg.type == "postgres":
try:
from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver
from psycopg_pool import AsyncConnectionPool
if not cp_cfg.connection_string:
raise ValueError("checkpointer.connection_string 是 postgres 必填项")
pool = AsyncConnectionPool(cp_cfg.connection_string, open=False)
await pool.open()
saver = AsyncPostgresSaver(pool)
await saver.setup()
_async_checkpointer_singleton = saver
logger.info("ai_writing: 使用 AsyncPostgresSaver")
return saver
except Exception as exc:
logger.exception("ai_writing: AsyncPostgresSaver 初始化失败")
_async_checkpointer_singleton = _fallback_to_memory_saver(
f"postgres 初始化失败: {type(exc).__name__}: {exc}"
)
return _async_checkpointer_singleton
# 未知 type —— config 配错也不能让服务挂
_async_checkpointer_singleton = _fallback_to_memory_saver(
f"未知 checkpointer.type={cp_cfg.type!r}"
)
return _async_checkpointer_singleton
async def get_ai_writing_graph():
"""懒加载并返回 AI 写作 graph 单例。
第一次调用按 config 选定的 checkpointer 构建 graph 并缓存;后续调用直接
返回缓存。测试可在调用前通过 ``monkeypatch.setattr`` 直接替换模块级
``AI_WRITING_GRAPH`` —— 本函数会优先返回已设的值。
"""
global AI_WRITING_GRAPH
if AI_WRITING_GRAPH is not None:
return AI_WRITING_GRAPH
checkpointer = await _get_async_checkpointer()
AI_WRITING_GRAPH = build_ai_writing_graph(checkpointer=checkpointer)
return AI_WRITING_GRAPH
def get_ai_writing_checkpointer():
"""返回当前已实例化的 checkpointer(清理任务用它来删 thread)。
若尚未懒构建过则返回 None。
"""
return _async_checkpointer_singleton
async def shutdown_ai_writing_graph() -> None:
"""关闭 saver 持有的底层连接,并清空 graph 单例。Lifespan shutdown 调用。"""
global AI_WRITING_GRAPH, _async_checkpointer_singleton
saver = _async_checkpointer_singleton
_async_checkpointer_singleton = None
AI_WRITING_GRAPH = None
if saver is None:
return
# aiosqlite saver: saver.conn 是 aiosqlite.Connection
conn = getattr(saver, "conn", None)
if conn is not None and hasattr(conn, "close"):
try:
await conn.close()
except Exception:
logger.warning("ai_writing: 关闭 checkpointer connection 失败", exc_info=True)
# postgres saver: saver.pool 是 AsyncConnectionPool
pool = getattr(saver, "pool", None)
if pool is not None and hasattr(pool, "close"):
try:
await pool.close()
except Exception:
logger.warning("ai_writing: 关闭 checkpointer pool 失败", exc_info=True)
return g.compile()
@@ -15,6 +15,29 @@ from langchain_core.runnables import RunnableConfig
logger = logging.getLogger(__name__)
def safe_get_stream_writer():
"""返回一个 stream writer callable;在没有 LangGraph runtime 上下文时退化成 no-op。
AI 写作节点用它把流式 chunk 推到 LangGraph 原生 ``custom`` 通道。
场景:
- LangGraph Server 路径(``client.runs.stream`` 订阅了 ``stream_mode=["custom", ...]``)
→ 真实 writer,事件透传给 SDK 消费方。
- 老 ``/start /stream /resume`` 路径(``graph.astream(stream_mode="updates")``)
→ 真实 writer 仍可调用,只是没人订阅 custom 通道,事件被静默丢弃。
这是阶段 2「硬切到 get_stream_writer」的预期行为:老前端 EventSource 不再收到
token-level chunk,但整体进度卡(draft_ready / review_ready 这些 state 级事件)
仍正常 —— 阶段 3 前端切到 useStream 后立刻补齐。
- 单测里直接调节点函数(无 LangGraph runtime) → 返回 no-op,调用方零分支。
"""
try:
from langgraph.config import get_stream_writer # 延迟 import,便于单测打补丁
return get_stream_writer()
except Exception:
return lambda _payload: None
def get_model(config: RunnableConfig, model_name: str | None = None):
"""创建对话模型。
@@ -1,8 +1,8 @@
"""编辑节点:三项独立 LLM 审查(事实/逻辑/语言),逐 token 流式推送以支持打字机效果。"""
from __future__ import annotations
import asyncio
import logging
from typing import Callable
from langchain_core.messages import HumanMessage, SystemMessage
from langchain_core.runnables import RunnableConfig
@@ -11,7 +11,7 @@ from ..state import AIWritingState, ReviewResult, ReviewIssue
from ..prompts.editor_prompts import (
EDITOR_SYSTEM, FACT_REVIEW_USER, LOGIC_REVIEW_USER, LANGUAGE_REVIEW_USER,
)
from ._utils import get_model, parse_json, _content_to_text, ThinkTagParser
from ._utils import get_model, parse_json, _content_to_text, ThinkTagParser, safe_get_stream_writer
logger = logging.getLogger(__name__)
@@ -40,9 +40,9 @@ def _parse_issues(raw_issues: list, category: str) -> list[ReviewIssue]:
async def _stream_one_review(
model, user_prompt: str, category: str,
sse_queue: asyncio.Queue | None, session_id: str,
writer: Callable[[dict], None],
) -> str:
"""流式执行单项审查,逐 chunk 推送 review_chunk 事件,返回累积的正文文本(用于 JSON 解析)。"""
"""流式执行单项审查,逐 chunk 经 LangGraph custom 通道推送 review_chunk 事件,返回累积的正文文本(用于 JSON 解析)。"""
full_text = "" # 仅累积正文(非思考),供后续 JSON 解析
parser = ThinkTagParser()
async for chunk in model.astream([
@@ -55,16 +55,42 @@ async def _stream_one_review(
for is_thinking, text in parser.feed(delta):
if not is_thinking:
full_text += text
if sse_queue is not None:
await sse_queue.put(("review_chunk", {
"session_id": session_id,
"category": category,
"text": text,
"is_thinking": is_thinking,
}))
writer({
"type": "review_chunk",
"category": category,
"text": text,
"is_thinking": is_thinking,
})
return full_text
async def _save_review_to_repo(config: RunnableConfig, review: ReviewResult) -> None:
"""把 review_result 异步落到业务表(best-effort,写失败不影响流程)。
``user_id`` 走 ``resolve_user_id(AUTO)`` 读 contextvar(与 intent_parser 建行、
writer_draft 写草稿同套规则)。
"""
cfg = config.get("configurable", {}) or {}
thread_id = cfg.get("thread_id")
if not thread_id:
return
from deerflow.persistence.ai_writing_sessions import get_default_store
from deerflow.runtime.user_context import AUTO
repo = get_default_store()
if repo is None:
return
try:
await repo.update(
thread_id,
user_id=AUTO,
status="review_ready",
review_result=dict(review),
)
except Exception:
logger.exception("editor: save_review failed (non-fatal) thread_id=%s", thread_id)
async def editor_node(state: AIWritingState, config: RunnableConfig) -> dict:
draft = state.get("current_draft")
if not draft:
@@ -78,8 +104,7 @@ async def editor_node(state: AIWritingState, config: RunnableConfig) -> dict:
mats = _materials_summary(pkg)
cfg = config.get("configurable", {})
sse_queue: asyncio.Queue | None = cfg.get("sse_queue")
session_id: str = cfg.get("sse_session_id", "")
writer = safe_get_stream_writer()
ai_writing_config = cfg.get("ai_writing_config", {}) or {}
pass_threshold = int(ai_writing_config.get("pass_threshold", _DEFAULT_PASS_THRESHOLD))
@@ -110,7 +135,7 @@ async def editor_node(state: AIWritingState, config: RunnableConfig) -> dict:
for category, user_prompt in REVIEW_TASKS:
fallback = {"score": 80, "issues": []}
try:
full_text = await _stream_one_review(model, user_prompt, category, sse_queue, session_id)
full_text = await _stream_one_review(model, user_prompt, category, writer)
res = parse_json(full_text, fallback)
except Exception as e:
logger.warning("editor review [%s] failed: %s", category, e)
@@ -120,16 +145,13 @@ async def editor_node(state: AIWritingState, config: RunnableConfig) -> dict:
results[category] = res
# 这一项审查完成 —— 推送结构化结果(前端据此把流式文本替换为评分/意见)
if sse_queue is not None:
issues = _parse_issues(res.get("issues", []), category)
await sse_queue.put(("review_item_done", {
"session_id": session_id,
"category": category,
"score": int(res.get("score", 80)),
"issues": [dict(iss) for iss in issues],
}))
else:
logger.debug("editor [%s] done (no sse_queue)", category)
issues = _parse_issues(res.get("issues", []), category)
writer({
"type": "review_item_done",
"category": category,
"score": int(res.get("score", 80)),
"issues": [dict(iss) for iss in issues],
})
# 汇总
fact_score = int(results["fact"].get("score", 80))
@@ -156,6 +178,9 @@ async def editor_node(state: AIWritingState, config: RunnableConfig) -> dict:
revision_notes=all_issues,
)
# 阶段 2:节点内并存 DB 写入(路由层同样也写一遍,阶段 7 删冗余时移除路由那份)
await _save_review_to_repo(config, review)
# 分项审查出现模型异常时,以非致命 notice 提示用户,审核结论仍照常产出
notices: list[dict] = []
if review_errors:
@@ -1,6 +1,7 @@
"""意图解析节点:将用户自然语言意图结构化为写作配置。"""
from __future__ import annotations
import logging
import re
from langchain_core.runnables import RunnableConfig
@@ -8,6 +9,49 @@ from langchain_core.runnables import RunnableConfig
from ..state import AIWritingState
from ._utils import llm_json
logger = logging.getLogger(__name__)
def _truncate_title(text: str, max_len: int = 50) -> str:
"""与老 /start 路由里 _truncate_title 行为一致:UTF 字符截断 + 省略号。"""
text = (text or "").strip()
if len(text) <= max_len:
return text
return text[:max_len].rstrip() + "…"
async def _upsert_business_row(configurable: dict, state: AIWritingState) -> None:
"""在 intent_parser 入口幂等创建 ai_writing_sessions 业务表行。
LangGraph Server 路径下没有 /start 路由建行,必须节点自建,否则列表页查不到、
transcript / draft / review 后续写入都会因 update 找不到行而静默丢数据。
resume 路径下 thread_id 已经有行,upsert_session 只刷新 updated_at。
user_id 走 ``resolve_user_id(AUTO)``:FastAPI 请求线程把它写到 contextvar,
asyncio.create_task 继承父任务的 context,所以这里能拿到当前用户;
contextvar 没设时(测试 / 离线场景)抛 RuntimeError,try/except 兜底跳过。
"""
from deerflow.persistence.ai_writing_sessions import get_default_store
from deerflow.runtime.user_context import AUTO
thread_id = configurable.get("thread_id")
if not thread_id:
return
repo = get_default_store()
if repo is None:
return
user_intent = (state.get("user_intent") or "").strip()
try:
await repo.upsert_session(
id=thread_id,
title=_truncate_title(user_intent) or "AI 写作",
user_intent=user_intent,
user_id=AUTO,
status="in_progress",
)
except Exception:
logger.exception("intent_parser: upsert_session failed (non-fatal) thread_id=%s", thread_id)
# 兜底类型列表,当 config 中未注入 valid_article_types 时使用
_DEFAULT_ARTICLE_TYPES: list[dict] = [
{"key": "news", "label": "新闻资讯"},
@@ -79,6 +123,10 @@ async def intent_parser_node(state: AIWritingState, config: RunnableConfig) -> d
valid_keys: list[str] = [t["key"] for t in valid_article_types]
default_key: str = valid_keys[0] if valid_keys else "analysis"
# 业务表 upsert(LangGraph Server 路径下没有 /start 路由建行,节点入口自建;
# resume 时幂等不重复建。失败不阻断写作流程)。
await _upsert_business_row(configurable, state)
type_options = _build_type_options(valid_article_types)
system = _SYSTEM_TEMPLATE.format(type_options=type_options)
user = _USER_TEMPLATE.format(
@@ -15,7 +15,7 @@ from ..prompts.researcher_prompts import (
EXTRACT_KEYWORDS_SYSTEM, EXTRACT_KEYWORDS_USER,
RANK_MATERIALS_SYSTEM, RANK_MATERIALS_USER,
)
from ._utils import llm_json
from ._utils import llm_json, safe_get_stream_writer
logger = logging.getLogger(__name__)
@@ -103,10 +103,9 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
# 用户指定的检索关键词数量(1~10),决定生成多少个检索词、检索多少个方向
keyword_count = _resolve_keyword_count(state.get("keyword_count"))
# SSE 实时进度推送(与 outline/draft 节点共用 sse_queue 机制)
# 实时进度推送:阶段 2 起统一走 LangGraph 原生 custom 通道。
cfg = config.get("configurable", {})
sse_queue = cfg.get("sse_queue")
sse_session_id: str = cfg.get("sse_session_id", "")
writer = safe_get_stream_writer()
# 素材收集专家专用模型:检索词生成 / 素材排序 可在 config.yaml 的
# ai_writing 段单独配置(keyword_model / rank_model),留空则回退写作主模型。
@@ -116,16 +115,11 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
# 素材总量上限(config.yaml: ai_writing.max_materials),按检索词数均分到每个词
max_materials = _resolve_max_materials(ai_writing_config.get("max_materials"))
async def _emit(message: str) -> None:
if sse_queue is None:
return
await sse_queue.put(("research_progress", {
"session_id": sse_session_id,
"message": message,
}))
def _emit(message: str) -> None:
writer({"type": "research_progress", "message": message})
# Step 1: 提取关键词
await _emit("正在生成检索词…")
_emit("正在生成检索词…")
kw_result = await llm_json(
EXTRACT_KEYWORDS_SYSTEM,
EXTRACT_KEYWORDS_USER.format(
@@ -154,7 +148,7 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
total = len(search_keywords)
# 每个检索词分到的检索条数 = 总量上限 / 检索词数(向上取整)
per_keyword = _per_keyword_limit(max_materials, total)
await _emit(f"检索词已生成,开始检索 {total} 个方向…")
_emit(f"检索词已生成,开始检索 {total} 个方向…")
# 并发搜索,但每完成一个就推送进度(计数在单线程 asyncio 下无需加锁)
completed = 0
@@ -166,9 +160,9 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
finally:
completed += 1
if completed < total:
await _emit(f"已完成 {completed}/{total} 次检索,正在继续检索相关内容…")
_emit(f"已完成 {completed}/{total} 次检索,正在继续检索相关内容…")
else:
await _emit(f"已完成全部 {total} 次检索")
_emit(f"已完成全部 {total} 次检索")
search_tasks = [_search_one(kw) for kw in search_keywords]
all_results_nested: list[list[SearchResult]] = await asyncio.gather(*search_tasks, return_exceptions=True)
@@ -208,7 +202,7 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
search_errors.extend(getattr(provider, "errors", []))
# Step 3: 相关度排序(如素材较多,让 LLM 排序)
await _emit("检索完成,正在筛选排序素材…")
_emit("检索完成,正在筛选排序素材…")
rank_error = ""
if len(raw_materials) > max_materials:
materials_list = "\n".join(
@@ -1,7 +1,6 @@
"""作家节点 Phase 2:按大纲逐段写正文,支持流式 chunk 输出。"""
from __future__ import annotations
import asyncio
import logging
import re
@@ -13,7 +12,7 @@ from ..prompts.writer_prompts import (
DRAFT_SECTION_SYSTEM, DRAFT_SECTION_USER, DRAFT_EXPAND_USER,
WRITER_MODE_LOOSE, WRITER_MODE_STRICT, SECTION_BLOCKED_MARKER,
)
from ._utils import get_model, _content_to_text, count_words, ThinkTagParser
from ._utils import get_model, _content_to_text, count_words, ThinkTagParser, safe_get_stream_writer
logger = logging.getLogger(__name__)
@@ -64,6 +63,35 @@ def _get_section_materials(material_ids: list[str], pkg: dict) -> str:
return "\n".join(lines) if lines else "(无相关素材)"
async def _save_draft_to_repo(config: RunnableConfig, draft: DraftArticle) -> None:
"""把 current_draft 异步落到业务表(best-effort,写失败不影响流程)。
``user_id`` 走 ``resolve_user_id(AUTO)`` 读 contextvar(FastAPI 请求线程设置,
asyncio.create_task 继承)。intent_parser 在阶段 4 已用同样规则建行;这里 update
时也用 AUTO 才能命中 owner filter。
"""
cfg = config.get("configurable", {}) or {}
thread_id = cfg.get("thread_id")
if not thread_id:
return
from deerflow.persistence.ai_writing_sessions import get_default_store
from deerflow.runtime.user_context import AUTO
repo = get_default_store()
if repo is None:
return
try:
await repo.update(
thread_id,
user_id=AUTO,
status="draft_ready",
draft_title=draft.get("title") or "",
draft_markdown=draft.get("full_markdown") or "",
)
except Exception:
logger.exception("writer_draft: save_draft failed (non-fatal) thread_id=%s", thread_id)
async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> dict:
outline = state.get("current_outline")
if not outline:
@@ -80,10 +108,10 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
writing_mode = state.get("writing_mode", "loose")
loose_titles: set[str] = set(intervention.get("_loose_sections") or [])
# SSE 流式推送所需
# 流式推送:阶段 2 起统一走 LangGraph 原生 custom 通道(get_stream_writer)。
# cfg 仍保留以便阶段 4 取 user_id / thread_id 用。
cfg = config.get("configurable", {})
sse_queue: asyncio.Queue | None = cfg.get("sse_queue")
session_id: str = cfg.get("sse_session_id", "")
writer = safe_get_stream_writer()
# 修改意见(来自用户或编辑)
revision_notes = ""
@@ -174,25 +202,23 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
for is_thinking, text in parser.feed(delta):
if not is_thinking:
section_buf += text
if sse_queue is not None:
await sse_queue.put(("draft_chunk", {
"session_id": session_id,
"section_index": i,
"section_title": sec_title,
"text": text,
"is_thinking": is_thinking,
"is_section_done": False,
}))
writer({
"type": "draft_chunk",
"section_index": i,
"section_title": sec_title,
"text": text,
"is_thinking": is_thinking,
"is_section_done": False,
})
# 章节完成
if sse_queue is not None:
await sse_queue.put(("draft_chunk", {
"session_id": session_id,
"section_index": i,
"section_title": sec_title,
"text": "",
"is_section_done": True,
}))
writer({
"type": "draft_chunk",
"section_index": i,
"section_title": sec_title,
"text": "",
"is_section_done": True,
})
cleaned_buf = section_buf.strip()
blocked_reason = ""
@@ -289,6 +315,11 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
revision_count=revision_count,
)
# 阶段 2:节点内并存 DB 写入(路由层 _stream_graph_segment 同样也会写一遍,
# 阶段 7 删冗余时移除路由层那份)。new path (LangGraph Server) 下路由层
# 不再介入,节点写就是唯一来源。
await _save_draft_to_repo(config, draft)
# 分节写作出现模型异常时,以非致命 notice 提示用户,草稿仍照常产出
notices: list[dict] = []
if section_errors:
@@ -1,7 +1,6 @@
"""作家节点 Phase 1:生成文章大纲,支持流式 chunk 输出。"""
from __future__ import annotations
import asyncio
import json
import logging
@@ -12,7 +11,7 @@ from ..state import AIWritingState, Outline, OutlineSection
from ..prompts.writer_prompts import (
OUTLINE_SYSTEM, OUTLINE_USER, OUTLINE_REVISE_SYSTEM, OUTLINE_REVISE_USER,
)
from ._utils import get_model, parse_json, _content_to_text, ThinkTagParser
from ._utils import get_model, parse_json, _content_to_text, ThinkTagParser, safe_get_stream_writer
logger = logging.getLogger(__name__)
@@ -79,10 +78,8 @@ async def writer_outline_node(state: AIWritingState, config: RunnableConfig) ->
feedback_section=feedback_section,
)
# SSE 流式推送
cfg = config.get("configurable", {})
sse_queue: asyncio.Queue | None = cfg.get("sse_queue")
session_id: str = cfg.get("sse_session_id", "")
# 流式推送:阶段 2 起统一走 LangGraph 原生 custom 通道(get_stream_writer)。
writer = safe_get_stream_writer()
full_text = "" # 只累积正文(非思考)用于 JSON 解析
parser = ThinkTagParser()
@@ -101,12 +98,11 @@ async def writer_outline_node(state: AIWritingState, config: RunnableConfig) ->
for is_thinking, text in parser.feed(delta):
if not is_thinking:
full_text += text
if sse_queue is not None:
await sse_queue.put(("outline_chunk", {
"session_id": session_id,
"text": text,
"is_thinking": is_thinking,
}))
writer({
"type": "outline_chunk",
"text": text,
"is_thinking": is_thinking,
})
except Exception as e:
logger.error("outline streaming failed: %s", e, exc_info=True)
return {
@@ -8,4 +8,39 @@ def make_ai_writing_session_store(session_factory) -> AIWritingSessionRepository
return AIWritingSessionRepository(session_factory)
__all__ = ["AIWritingSessionRow", "AIWritingSessionRepository", "make_ai_writing_session_store"]
# ── 进程级默认 repo 单例 ──────────────────────────────────────────────────────
#
# AI 写作 graph 注册成 LangGraph Server assistant 后,节点跑在 harness 进程里,
# 拿不到 FastAPI 的 ``request.app.state``;harness → app 又是禁止的 import 方向。
# 这里提供一个 app 启动时注册、harness 节点运行时读取的「跨边界单例」通道。
#
# 写入:app/gateway/deps.py 在创建完 session_factory 之后调
# ``register_default_store(app.state.ai_writing_session_store)``
# 读取:节点里 ``get_default_store()``,None 表示未配置 DB(跳过写入即可)。
_default_store: AIWritingSessionRepository | None = None
def register_default_store(store: AIWritingSessionRepository | None) -> None:
"""登记进程级默认 AI 写作 session store。
幂等:可多次调用,覆盖前值。传 ``None`` 用于显式清理(主要给测试)。
"""
global _default_store
_default_store = store
def get_default_store() -> AIWritingSessionRepository | None:
"""返回进程级默认 AI 写作 session store;未配置时为 ``None``。
节点调用方应在拿到 ``None`` 时跳过 DB 写入(写入是「锦上添花」非致命)。
"""
return _default_store
__all__ = [
"AIWritingSessionRow",
"AIWritingSessionRepository",
"make_ai_writing_session_store",
"register_default_store",
"get_default_store",
]
@@ -4,11 +4,11 @@ from __future__ import annotations
from datetime import UTC, datetime
from sqlalchemy import DateTime, String, Text
from sqlalchemy import String, Text
from sqlalchemy.orm import Mapped, mapped_column
from deerflow.persistence.base import Base
from deerflow.persistence.types import PortableLongText
from deerflow.persistence.types import BeijingDateTime, PortableLongText
class AIWritingSessionRow(Base):
@@ -26,5 +26,13 @@ class AIWritingSessionRow(Base):
# 完整对话时间线(前端组装的 progressEvents + completedInterventions),用于历史回显。
# 用 PortableLongText:MySQL 上映射为 LONGTEXT,避免 TEXT 64KB 上限导致保存 500。
transcript: Mapped[str | None] = mapped_column(PortableLongText, nullable=True) # JSON object
created_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), default=lambda: datetime.now(UTC))
updated_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), default=lambda: datetime.now(UTC))
# 时间列必须用 BeijingDateTime(项目规范,与 threads_meta 等其它表对齐)—— MySQL
# DATETIME 不存时区,raw ``DateTime(timezone=True)`` 会把 UTC wall clock 写入
# 列,读出来比真实北京时间晚 8 小时。``BeijingDateTime`` 在 bind 时 UTC→Beijing、
# 在 result 时 attach +08:00,前后端拿到的就是北京时间且 tz-aware。
created_at: Mapped[datetime] = mapped_column(BeijingDateTime(), default=lambda: datetime.now(UTC))
updated_at: Mapped[datetime] = mapped_column(
BeijingDateTime(),
default=lambda: datetime.now(UTC),
onupdate=lambda: datetime.now(UTC),
)
@@ -63,6 +63,52 @@ class AIWritingSessionRepository:
await session.refresh(row)
return self._row_to_dict(row)
async def upsert_session(
self,
*,
id: str,
title: str,
user_intent: str | None = None,
user_id: str | None | _AutoSentinel = AUTO,
status: str = "in_progress",
) -> dict:
"""幂等:行不存在则创建,行存在则保留原数据(不覆盖 draft / review 等已有写入)。
AI 写作 graph 的 ``intent_parser_node`` 在每次进入时调一次(首启 + resume 时
都会跑),所以必须幂等。resume 时若直接 ``create`` 会因主键冲突抛错;用
``update`` 又找不到行而 no-op。这个方法把两条路合一:行不存在新建,行存在
只刷新 ``updated_at`` 以便列表页排序前置。
"""
resolved_user_id = resolve_user_id(user_id, method_name="AIWritingSessionRepository.upsert_session")
now = datetime.now(UTC)
async with self._sf() as session:
row = await session.get(AIWritingSessionRow, id)
if row is None:
row = AIWritingSessionRow(
id=id,
user_id=resolved_user_id,
title=title,
user_intent=user_intent,
status=status,
draft_title=None,
draft_markdown=None,
completed_interventions=None,
review_result=None,
transcript=None,
created_at=now,
updated_at=now,
)
session.add(row)
else:
# 行已存在(resume 路径):仅刷新 updated_at,保留 user_id / title / draft / review 等。
# 若 user_id 之前为空(migration 兼容场景),用解析出来的填上。
if row.user_id is None and resolved_user_id is not None:
row.user_id = resolved_user_id
row.updated_at = now
await session.commit()
await session.refresh(row)
return self._row_to_dict(row)
async def update(
self,
session_id: str,
@@ -56,6 +56,17 @@ class ThreadMetaStore(abc.ABC):
async def update_status(self, thread_id: str, status: str, *, user_id: str | None | _AutoSentinel = AUTO) -> None:
pass
@abc.abstractmethod
async def update_assistant_id(self, thread_id: str, assistant_id: str, *, user_id: str | None | _AutoSentinel = AUTO) -> None:
"""Backfill ``assistant_id`` for a thread row.
Used by ``start_run`` when the row was pre-created via ``POST /threads``
without an assistant_id (the SDK's ``client.threads.create()`` doesn't
pass it), and the first run on the thread reveals the assistant. No-op
if the thread does not exist or the owner check fails.
"""
pass
@abc.abstractmethod
async def update_metadata(self, thread_id: str, metadata: dict, *, user_id: str | None | _AutoSentinel = AUTO) -> None:
"""Merge ``metadata`` into the thread's metadata field.
@@ -117,6 +117,14 @@ class MemoryThreadMetaStore(ThreadMetaStore):
record["updated_at"] = now_iso()
await self._store.aput(THREADS_NS, thread_id, record)
async def update_assistant_id(self, thread_id: str, assistant_id: str, *, user_id: str | None | _AutoSentinel = AUTO) -> None:
record = await self._get_owned_record(thread_id, user_id, "MemoryThreadMetaStore.update_assistant_id")
if record is None:
return
record["assistant_id"] = assistant_id
record["updated_at"] = now_iso()
await self._store.aput(THREADS_NS, thread_id, record)
async def update_metadata(self, thread_id: str, metadata: dict, *, user_id: str | None | _AutoSentinel = AUTO) -> None:
record = await self._get_owned_record(thread_id, user_id, "MemoryThreadMetaStore.update_metadata")
if record is None:
@@ -174,6 +174,20 @@ class ThreadMetaRepository(ThreadMetaStore):
await session.execute(update(ThreadMetaRow).where(ThreadMetaRow.thread_id == thread_id).values(status=status, updated_at=datetime.now(UTC)))
await session.commit()
async def update_assistant_id(
self,
thread_id: str,
assistant_id: str,
*,
user_id: str | None | _AutoSentinel = AUTO,
) -> None:
resolved_user_id = resolve_user_id(user_id, method_name="ThreadMetaRepository.update_assistant_id")
async with self._sf() as session:
if not await self._check_ownership(session, thread_id, resolved_user_id):
return
await session.execute(update(ThreadMetaRow).where(ThreadMetaRow.thread_id == thread_id).values(assistant_id=assistant_id, updated_at=datetime.now(UTC)))
await session.commit()
async def update_metadata(
self,
thread_id: str,
@@ -130,8 +130,16 @@ async def run_agent(
stream_subgraphs: bool = False,
interrupt_before: list[str] | Literal["*"] | None = None,
interrupt_after: list[str] | Literal["*"] | None = None,
command: dict[str, Any] | None = None,
) -> None:
"""Execute an agent in the background, publishing events to *bridge*."""
"""Execute an agent in the background, publishing events to *bridge*.
``command`` 来自 LangGraph SDK 的 ``Command`` 协议(``{resume, update, goto}``)。
存在时把它翻译成 ``langgraph.types.Command`` 并作为 ``astream`` 的入参 ——
Pregel 检查到 Command 输入会从最近 checkpoint 处恢复(典型场景:interrupt 之后
的 ``Command(resume=user_input)``);不存在时仍按 ``graph_input`` dict 起新 run。
AI 写作的 4 个干预卡就靠这个把 user_input 传回 ``pause_*_confirm`` 节点。
"""
# Unpack infrastructure dependencies from RunContext.
checkpointer = ctx.checkpointer
@@ -269,10 +277,23 @@ async def run_agent(
logger.info("Run %s: streaming with modes %s (requested: %s)", run_id, lg_modes, requested_modes)
# 7. Stream using graph.astream
# 如果客户端传了 Command(典型:interrupt 后续提的 ``Command(resume=...)``),
# 把它翻译成 langgraph.types.Command 作为 astream 输入;否则按 graph_input dict
# 起新一轮。Pregel 检查到 Command 会从最近 checkpoint 处恢复,而不是重头跑。
stream_input: Any = graph_input
if command is not None:
from langgraph.types import Command as _LGCommand
stream_input = _LGCommand(
resume=command.get("resume"),
update=command.get("update"),
goto=command.get("goto"),
)
if len(lg_modes) == 1 and not stream_subgraphs:
# Single mode, no subgraphs: astream yields raw chunks
single_mode = lg_modes[0]
async for chunk in agent.astream(graph_input, config=runnable_config, stream_mode=single_mode):
async for chunk in agent.astream(stream_input, config=runnable_config, stream_mode=single_mode):
if record.abort_event.is_set():
logger.info("Run %s abort requested — stopping", run_id)
break
@@ -281,7 +302,7 @@ async def run_agent(
else:
# Multiple modes or subgraphs: astream yields tuples
async for item in agent.astream(
graph_input,
stream_input,
config=runnable_config,
stream_mode=lg_modes,
subgraphs=stream_subgraphs,
@@ -10,6 +10,7 @@ Consumers: ``deerflow.runtime.runs.worker`` (SSE publishing) and
from __future__ import annotations
import dataclasses
from typing import Any
@@ -35,6 +36,16 @@ def serialize_lc_object(obj: Any) -> Any:
return obj.dict()
except Exception:
pass
# Dataclass instances —— LangGraph 的 ``Interrupt`` 就是 dataclass 没有
# model_dump 也没有 dict 方法;之前 fallback 到 ``str(obj)`` 会把
# ``Interrupt(value=..., id='...')`` 序列化成字符串,下游 SDK
# (``@langchain/langgraph-sdk``) 期待 dict 结构 → 直接丢弃 → AI 写作
# 的 4 个干预卡永远不弹(参 ai-writing-优化.md §阶段 3 排查记录)。
if dataclasses.is_dataclass(obj) and not isinstance(obj, type):
try:
return {k: serialize_lc_object(v) for k, v in dataclasses.asdict(obj).items()}
except Exception:
pass
# Last resort
try:
return str(obj)
@@ -45,12 +56,21 @@ def serialize_lc_object(obj: Any) -> Any:
def serialize_channel_values(channel_values: dict[str, Any]) -> dict[str, Any]:
"""Serialize channel values, stripping internal LangGraph keys.
Internal keys like ``__pregel_*`` and ``__interrupt__`` are removed
to match what the LangGraph Platform API returns.
``__pregel_*`` 键始终剥离。``__interrupt__`` 仅在**为空时**剥离 —— 因为
``@langchain/langgraph-sdk`` 的 ``useStream`` 只在 ``event === "values"``
通道里读 ``data.__interrupt__`` 并通过 ``stream.interrupt`` 暴露给上层。
AI 写作的 4 个干预卡(material/outline/draft/review)就是靠它触发的;
继续过滤会让前端永远拿不到 interrupt。
主聊天不消费 ``stream.interrupt``,但若我们让一个空数组溜过去,SDK 会
fallback 成 ``{when: "breakpoint"}`` 的伪 interrupt —— 所以仅在数组非空
时保留,两侧都不破坏。
"""
result: dict[str, Any] = {}
for key, value in channel_values.items():
if key.startswith("__pregel_") or key == "__interrupt__":
if key.startswith("__pregel_"):
continue
if key == "__interrupt__" and not value:
continue
result[key] = serialize_lc_object(value)
return result