Files
deerflow-code/offline-backend-20260512/backend/packages/harness/deerflow/agents/middlewares/title_middleware.py
T
AdministratorandClaude Opus 4.8 acc29c8121 fix(middlewares): 标题模型推理不再泄漏到 UI 流 + open/ephemeral 会话避免无谓 DB 访问
- title: 标题模型用隔离 config(不继承父级 streaming callbacks),force_disable_thinking,<think> 块清洗,避免推理 token 进入实时流/运行日志
- uploads: 无文件的普通对话轮次直接跳过,不再解析线程归属(避免命中 DB)
- thread_data: ephemeral 会话用当前登录用户解析数据目录
- thread_paths: open-/__skill-review__ 等合成线程视为 local-only,直接用当前用户
- frontend: 新增 build:test 构建脚本

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-24 21:20:12 +08:00

271 lines
12 KiB
Python

"""Middleware for automatic thread title generation."""
import logging
import re
from typing import TYPE_CHECKING, Any, NotRequired, override
from langchain.agents import AgentState
from langchain.agents.middleware import AgentMiddleware
from langgraph.runtime import Runtime
from deerflow.config.title_config import get_title_config
from deerflow.models import create_chat_model
if TYPE_CHECKING:
from deerflow.config.app_config import AppConfig
from deerflow.config.title_config import TitleConfig
logger = logging.getLogger(__name__)
class TitleMiddlewareState(AgentState):
"""Compatible with the `ThreadState` schema."""
title: NotRequired[str | None]
title_provisional: NotRequired[bool]
class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
"""Automatically generate a title for the thread after the first user message."""
state_schema = TitleMiddlewareState
def __init__(self, *, app_config: "AppConfig | None" = None, title_config: "TitleConfig | None" = None):
super().__init__()
self._app_config = app_config
self._title_config = title_config
def _get_title_config(self):
if self._title_config is not None:
return self._title_config
if self._app_config is not None:
return self._app_config.title
return get_title_config()
def _normalize_content(self, content: object) -> str:
if isinstance(content, str):
return content
if isinstance(content, list):
parts = [self._normalize_content(item) for item in content]
return "\n".join(part for part in parts if part)
if isinstance(content, dict):
text_value = content.get("text")
if isinstance(text_value, str):
return text_value
nested_content = content.get("content")
if nested_content is not None:
return self._normalize_content(nested_content)
return ""
def _should_set_provisional_title(self, state: TitleMiddlewareState) -> bool:
"""Check if we should set an instant, question-derived placeholder title.
Runs in ``before_model`` so the thread shows a sensible title the moment
the user sends their first message — eliminating the "untitled" window
while the (possibly long) first answer streams. No LLM call, no latency.
"""
config = self._get_title_config()
if not config.enabled:
return False
# Already has a title (provisional or final) — nothing to do.
if state.get("title"):
return False
# Only on the very first turn, as soon as the user message arrives.
messages = state.get("messages", [])
user_messages = [m for m in messages if m.type == "human"]
return len(user_messages) == 1
def _should_generate_title(self, state: TitleMiddlewareState) -> bool:
"""Check if we should generate the final LLM title for this thread."""
config = self._get_title_config()
if not config.enabled:
return False
# Skip only if we already have a *finalized* title. A provisional
# placeholder set in before_model should still be upgraded by the LLM.
if state.get("title") and not state.get("title_provisional"):
return False
# Check if this is the first turn (has at least one user message and one assistant response)
messages = state.get("messages", [])
if len(messages) < 2:
return False
# Count user and assistant messages
user_messages = [m for m in messages if m.type == "human"]
assistant_messages = [m for m in messages if m.type == "ai"]
# Generate title after first complete exchange
return len(user_messages) == 1 and len(assistant_messages) >= 1
def _build_title_prompt(self, state: TitleMiddlewareState) -> tuple[str, str]:
"""Extract user/assistant messages and build the title prompt.
Returns (prompt_string, user_msg) so callers can use user_msg as fallback.
"""
config = self._get_title_config()
messages = state.get("messages", [])
user_msg_content = next((m.content for m in messages if m.type == "human"), "")
assistant_msg_content = next((m.content for m in messages if m.type == "ai"), "")
user_msg = self._normalize_content(user_msg_content)
assistant_msg = self._strip_think_tags(self._normalize_content(assistant_msg_content))
prompt = config.prompt_template.format(
max_words=config.max_words,
user_msg=user_msg[:500],
assistant_msg=assistant_msg[:500],
)
return prompt, user_msg
def _strip_think_tags(self, text: str) -> str:
"""Remove <think>...</think> blocks emitted by reasoning models (e.g. minimax, GLM, DeepSeek-R1).
Also handles malformed cases: an unclosed `<think>` (response truncated mid-reasoning)
or a stray `</think>` (open tag eaten by upstream parsing) — both can otherwise leak
raw reasoning into the title.
"""
text = re.sub(r"<think>[\s\S]*?</think>", "", text, flags=re.IGNORECASE)
text = re.sub(r"<think>[\s\S]*$", "", text, flags=re.IGNORECASE)
text = re.sub(r"^[\s\S]*?</think>", "", text, flags=re.IGNORECASE)
return text.strip()
def _parse_title(self, content: object) -> str:
"""Normalize model output into a clean title string (strips <think> blocks)."""
config = self._get_title_config()
title_content = self._strip_think_tags(self._normalize_content(content))
title = title_content.strip().strip('"').strip("'")
return title[: config.max_chars] if len(title) > config.max_chars else title
def _isolated_config(self) -> dict[str, Any]:
"""Build a RunnableConfig that does NOT inherit the parent's streaming callbacks.
The leak that pushed title-generation reasoning into the live UI stream came
from inheriting the parent RunnableConfig (and with it the SSE/streaming
callbacks) for the title model call. Here we deliberately start from an empty
config with ``callbacks: []`` so the title model's tokens — including any
``<think>`` reasoning from a reasoning-only deployment — never reach the
frontend stream or the run journal as conversation output. The stored title
is additionally cleaned by ``_parse_title``/``_strip_think_tags``.
"""
return {
"run_name": "title_agent",
"tags": ["middleware:title"],
"callbacks": [],
}
def _fallback_title(self, user_msg: str) -> str:
config = self._get_title_config()
fallback_chars = min(config.max_chars, 50)
if len(user_msg) > fallback_chars:
return user_msg[:fallback_chars].rstrip() + "..."
return user_msg if user_msg else "New Conversation"
def _provisional_title_result(self, state: TitleMiddlewareState) -> dict | None:
"""Set an instant placeholder title from the first user message (no LLM)."""
if not self._should_set_provisional_title(state):
return None
messages = state.get("messages", [])
user_msg_content = next((m.content for m in messages if m.type == "human"), "")
user_msg = self._normalize_content(user_msg_content)
if not user_msg.strip():
return None
title = self._fallback_title(user_msg)
logger.debug("TitleMiddleware: set provisional title %r before first model call", title)
return {"title": title, "title_provisional": True}
def _generate_title_result(self, state: TitleMiddlewareState) -> dict | None:
"""Generate a local fallback title without blocking on an LLM call."""
if not self._should_generate_title(state):
return None
_, user_msg = self._build_title_prompt(state)
return {"title": self._fallback_title(user_msg), "title_provisional": False}
def _force_disable_thinking(self, model: object) -> None:
"""Force-disable thinking on the title model via `extra_body` overrides.
Title generation never needs reasoning. The model factory only emits
disable params when the model config declares `when_thinking_disabled`
or a matching `when_thinking_enabled` shape — many user-added reasoning
models (e.g. GLM-4.5/5) only set `supports_thinking: true` and would
otherwise keep thinking on, which is slow and leaks <think> tags into
the title prompt response. Injecting both common OpenAI-compatible
disable shapes covers Zhipu GLM (`thinking.type`) and Qwen/vLLM
(`chat_template_kwargs.enable_thinking`); other providers ignore
unknown body fields.
"""
if not hasattr(model, "extra_body"):
return
current = getattr(model, "extra_body", None) or {}
merged_chat_template_kwargs = {
**(current.get("chat_template_kwargs") or {}),
"enable_thinking": False,
}
merged = {
**current,
"thinking": {**(current.get("thinking") or {}), "type": "disabled"},
"chat_template_kwargs": merged_chat_template_kwargs,
}
try:
model.extra_body = merged
except Exception:
logger.debug("Could not override extra_body to disable thinking on title model", exc_info=True)
async def _agenerate_title_result(self, state: TitleMiddlewareState) -> dict | None:
"""Polish the title with an LLM, without leaking its reasoning to the UI.
The title model is invoked with an *isolated* config (``_isolated_config``)
that does not inherit the parent's streaming callbacks, so even a
reasoning-only deployment's ``<think>`` tokens never reach the live UI
stream or the run journal. Thinking is also force-disabled on the model,
and the final stored title is stripped of any residual ``<think>`` blocks.
On any failure we fall back to the local question-derived title.
"""
if not self._should_generate_title(state):
return None
config = self._get_title_config()
prompt, user_msg = self._build_title_prompt(state)
try:
model_kwargs: dict[str, Any] = {"thinking_enabled": False, "force_disable_thinking": True}
if self._app_config is not None:
model_kwargs["app_config"] = self._app_config
if config.model_name:
model = create_chat_model(name=config.model_name, **model_kwargs)
else:
model = create_chat_model(**model_kwargs)
self._force_disable_thinking(model)
response = await model.ainvoke(prompt, config=self._isolated_config())
title = self._parse_title(response.content)
if title:
return {"title": title, "title_provisional": False}
except Exception:
logger.debug("Failed to generate async title via LLM; falling back to local title", exc_info=True)
return {"title": self._fallback_title(user_msg), "title_provisional": False}
@override
def before_model(self, state: TitleMiddlewareState, runtime: Runtime) -> dict | None:
return self._provisional_title_result(state)
@override
async def abefore_model(self, state: TitleMiddlewareState, runtime: Runtime) -> dict | None:
return self._provisional_title_result(state)
@override
def after_model(self, state: TitleMiddlewareState, runtime: Runtime) -> dict | None:
return self._generate_title_result(state)
@override
async def aafter_model(self, state: TitleMiddlewareState, runtime: Runtime) -> dict | None:
return await self._agenerate_title_result(state)