# Conflicts:
#	frontend-web/src/components/workspace/chats/chat-box.tsx
#	frontend-web/src/pages/StrategyRoutes.tsx
#	offline-backend-20260512/backend/app/gateway/routers/agents.py
This commit is contained in:
董宏凯 committed 2026-05-28 15:36:11 +08:00
commit d2f3eba434
81 files changed
+10489 -1653

No files matched your search

@@ -9,6 +9,7 @@ from deerflow.agents.middlewares.clarification_middleware import ClarificationMi
from deerflow.agents.middlewares.loop_detection_middleware import LoopDetectionMiddleware
from deerflow.agents.middlewares.memory_middleware import MemoryMiddleware
from deerflow.agents.middlewares.prompt_prefix_middleware import PromptPrefixMiddleware
from deerflow.agents.middlewares.rag_citation_middleware import RagCitationMiddleware
from deerflow.agents.middlewares.setup_agent_approval_middleware import SetupAgentApprovalMiddleware
from deerflow.agents.middlewares.subagent_limit_middleware import SubagentLimitMiddleware
from deerflow.agents.middlewares.summarization_middleware import BeforeSummarizationHook, DeerFlowSummarizationMiddleware
@@ -254,11 +255,14 @@ def _build_middlewares(
List of middleware instances.
"""
resolved_app_config = app_config or get_app_config()
cfg = _get_runtime_config(config)
middlewares = build_lead_runtime_middlewares(app_config=resolved_app_config, lazy_init=True)
# Inject the admin prompt prefix into the model request (kept out of the
# persisted/displayed message so the user never sees it).
middlewares.append(PromptPrefixMiddleware())
if cfg.get("rag_mode_enabled", False):
middlewares.append(RagCitationMiddleware())
# Add summarization middleware if enabled
summarization_middleware = _create_summarization_middleware(app_config=resolved_app_config)
@@ -266,7 +270,6 @@ def _build_middlewares(
middlewares.append(summarization_middleware)
# Add TodoList middleware if plan mode is enabled
cfg = _get_runtime_config(config)
is_plan_mode = cfg.get("is_plan_mode", False)
is_bootstrap = cfg.get("is_bootstrap", False)
todo_list_middleware = _create_todo_list_middleware(is_plan_mode)
@@ -371,6 +374,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
max_concurrent_subagents = cfg.get("max_concurrent_subagents", 3)
is_bootstrap = cfg.get("is_bootstrap", False)
memory_injection_enabled = bool(cfg.get("memory_injection_enabled", True))
rag_mode_enabled = bool(cfg.get("rag_mode_enabled", False))
writing_mode = bool(cfg.get("writing_mode", False))
writing_artifact_path = cfg.get("writing_artifact_path") if writing_mode else None
if not isinstance(writing_artifact_path, str) or not writing_artifact_path.strip():
@@ -461,6 +465,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
is_bootstrap=True,
app_config=resolved_app_config,
memory_injection_enabled=memory_injection_enabled,
rag_mode_enabled=rag_mode_enabled,
),
state_schema=ThreadState,
)
@@ -495,6 +500,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
available_skills=(set(agent_config.skills or []) if agent_config else None),
app_config=resolved_app_config,
memory_injection_enabled=memory_injection_enabled,
rag_mode_enabled=rag_mode_enabled,
is_notebook_mode=is_notebook_mode,
notebook_source_ids=notebook_source_ids,
),
@@ -667,17 +667,25 @@ def _get_memory_context(agent_name: str | None = None, *, app_config: AppConfig
@lru_cache(maxsize=32)
def _get_cached_skills_prompt_section(
skill_signature: tuple[tuple[str, str, str, str], ...],
display_signature: tuple[tuple[str, str, str], ...],
available_skills_key: tuple[str, ...] | None,
container_base_path: str,
skill_evolution_section: str,
) -> str:
display_by_skill = {name: (mode, custom_instruction) for name, mode, custom_instruction in display_signature}
filtered = [(name, description, category, location) for name, description, category, location in skill_signature if available_skills_key is None or name in available_skills_key]
skills_list = ""
if filtered:
skill_items = "\n".join(
f" <skill>\n <name>{name}</name>\n <description>{description} {_skill_mutability_label(category)}</description>\n <location>{location}</location>\n </skill>"
for name, description, category, location in filtered
)
items = []
for name, description, category, location in filtered:
display_mode, custom_instruction = display_by_skill.get(name, ("", ""))
display_hint = ""
if display_mode:
display_hint = f"\n <display mode=\"{display_mode}\">{custom_instruction}</display>"
items.append(
f" <skill>\n <name>{name}</name>\n <description>{description} {_skill_mutability_label(category)}</description>\n <location>{location}</location>{display_hint}\n </skill>"
)
skill_items = "\n".join(items)
skills_list = f"<available_skills>\n{skill_items}\n</available_skills>"
# Restricted mode: agent has an explicit skill allowlist.
@@ -693,6 +701,19 @@ def _get_cached_skills_prompt_section(
else:
scope_note = ""
display_pattern = (
"""
**Skill Result Display Pattern:**
- If a skill has `<display mode="citation">`, use the skill's returned result order as reference numbers. Cite facts from the first result as `[[1]]`, the second as `[[2]]`, and so on. Use only numbers that correspond to returned results.
- If a skill has `<display mode="list">`, summarize the returned items as a concise list and keep item order stable.
- If a skill has `<display mode="table">`, summarize the result and output `<skill-display tool="SKILL_OR_TOOL_NAME" mode="table" />` where the frontend should render the real table. Do not hand-write the table rows yourself.
- If a skill has `<display mode="card">`, summarize the result and output `<skill-display tool="SKILL_OR_TOOL_NAME" mode="card" />` where the frontend should render the cards.
- If a skill has `<display mode="custom">`, follow the custom display instruction in the `<display>` tag when presenting the skill result.
"""
if display_signature
else ""
)
return f"""<skill_system>
You have access to skills that provide optimized workflows for specific tasks. Each skill contains best practices, frameworks, and references to additional resources.
{scope_note}
@@ -702,13 +723,19 @@ You have access to skills that provide optimized workflows for specific tasks. E
3. The skill file contains references to external resources under the same folder
4. Load referenced resources only when needed during execution
5. Follow the skill's instructions precisely
{display_pattern}
{location_hint}{evolution_block}
{skills_list}
</skill_system>"""
def get_skills_prompt_section(available_skills: set[str] | None = None, *, app_config: AppConfig | None = None) -> str:
def get_skills_prompt_section(
available_skills: set[str] | None = None,
*,
app_config: AppConfig | None = None,
display_enabled: bool = False,
) -> str:
"""Generate the skills prompt section with available skills list."""
skills = _get_enabled_skills_for_config(app_config)
@@ -739,11 +766,30 @@ def get_skills_prompt_section(available_skills: set[str] | None = None, *, app_c
return ""
skill_signature = tuple((skill.name, skill.description, skill.category, skill.get_container_file_path(container_base_path)) for skill in skills)
display_signature: tuple[tuple[str, str, str], ...] = ()
if display_enabled:
try:
from deerflow.config.extensions_config import get_extensions_config
extensions = get_extensions_config()
display_signature = tuple(
(
name,
skill_config.display.mode,
skill_config.display.custom_instruction,
)
for name, skill_config in sorted(extensions.skills.items())
if skill_config.display is not None
and skill_config.display.enabled
and skill_config.display.mode != "none"
)
except Exception:
logger.debug("Failed to load skill display configs for prompt injection", exc_info=True)
available_key = tuple(sorted(available_skills)) if available_skills is not None else None
if not skill_signature and available_key is not None:
return ""
skill_evolution_section = _build_skill_evolution_section(effective_skill_evolution_enabled)
return _get_cached_skills_prompt_section(skill_signature, available_key, container_base_path, skill_evolution_section)
return _get_cached_skills_prompt_section(skill_signature, display_signature, available_key, container_base_path, skill_evolution_section)
def get_agent_soul(agent_id: str | None) -> str:
@@ -868,6 +914,7 @@ def apply_prompt_template(
is_bootstrap: bool = False,
app_config: AppConfig | None = None,
memory_injection_enabled: bool = True,
rag_mode_enabled: bool = False,
is_notebook_mode: bool = False,
notebook_source_ids: list[int] | None = None,
) -> str:
@@ -902,7 +949,11 @@ def apply_prompt_template(
bootstrap_section = _build_bootstrap_section() if is_bootstrap else ""
# Get skills section
skills_section = get_skills_prompt_section(available_skills, app_config=app_config)
skills_section = get_skills_prompt_section(
available_skills,
app_config=app_config,
display_enabled=rag_mode_enabled,
)
# Get deferred tools section (tool_search)
deferred_tools_section = get_deferred_tools_prompt_section(app_config=app_config)
@@ -940,6 +991,21 @@ Writing mode is enabled for this run.
# Build the offline-network policy block (empty unless DEER_FLOW_OFFLINE_MODE is set).
network_policy_section = _build_network_policy_section()
rag_mode_section = (
"""
<rag_citation_mode>
RAG/reference display mode is enabled.
When a search, knowledge, skill, or MCP tool returns a result list:
- If the skill has a display adapter, follow that adapter first.
- For `citation`, cite returned items with `[[1]]`, `[[2]]`, etc. Only cite real returned items.
- For `table`, do not hand-write the table. Summarize and place `<skill-display tool="SKILL_OR_TOOL_NAME" mode="table" />` where the table should appear. The frontend will render the real tool result as a stable table.
- For `list` or `card`, summarize and place `<skill-display tool="SKILL_OR_TOOL_NAME" mode="list" />` or `<skill-display tool="SKILL_OR_TOOL_NAME" mode="card" />` where the structured result should appear.
- If no display adapter exists, default to citation mode with numbered citations.
</rag_citation_mode>
"""
if rag_mode_enabled
else ""
)
# Format the prompt with dynamic skills and memory
prompt = SYSTEM_PROMPT_TEMPLATE.format(
@@ -957,7 +1023,7 @@ Writing mode is enabled for this run.
acp_section=acp_and_mounts_section,
)
prompt = prompt + f"\n<current_date>{datetime.now().strftime('%Y-%m-%d, %A')}</current_date>"
prompt = prompt + rag_mode_section + f"\n<current_date>{datetime.now().strftime('%Y-%m-%d, %A')}</current_date>"
if is_notebook_mode:
source_ids = notebook_source_ids or []
@@ -61,6 +61,21 @@ _AUTH_PATTERNS = (
"无权",
"未授权",
)
# Upstream model gateway returned a malformed payload (typically a JSON decode
# error wrapped as `bad_response_status_code`, or a truncated/garbled inference
# response). These are almost always transient inference-side issues and safe
# to retry.
_UPSTREAM_GARBAGE_PATTERNS = (
"bad_response_status_code",
"bad response status",
"expecting ',' delimiter",
"expecting value",
"expecting property name",
"unexpected character",
"unterminated string",
"extra data",
"jsondecodeerror",
)
class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
@@ -139,9 +154,10 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
detail = _extract_error_detail(exc)
lowered = detail.lower()
error_code = _extract_error_code(exc)
error_code_lower = str(error_code).lower() if error_code is not None else ""
status_code = _extract_status_code(exc)
if _matches_any(lowered, _QUOTA_PATTERNS) or _matches_any(str(error_code).lower(), _QUOTA_PATTERNS):
if _matches_any(lowered, _QUOTA_PATTERNS) or _matches_any(error_code_lower, _QUOTA_PATTERNS):
return False, "quota"
if _matches_any(lowered, _AUTH_PATTERNS):
return False, "auth"
@@ -159,6 +175,11 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
return True, "transient"
if _matches_any(lowered, _BUSY_PATTERNS):
return True, "busy"
# Treat upstream gateway garbage (malformed JSON, truncated responses)
# as a retriable transient — the deterministic char-position errors we
# see from vLLM/proxy layers usually clear on a re-roll.
if _matches_any(lowered, _UPSTREAM_GARBAGE_PATTERNS) or _matches_any(error_code_lower, _UPSTREAM_GARBAGE_PATTERNS):
return True, "upstream"
return False, "generic"
@@ -171,21 +192,34 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
def _build_retry_message(self, attempt: int, wait_ms: int, reason: str) -> str:
seconds = max(1, round(wait_ms / 1000))
reason_text = "服务繁忙" if reason == "busy" else "网络/连接异常"
if reason == "busy":
reason_text = "服务繁忙"
elif reason == "upstream":
reason_text = "上游响应异常"
else:
reason_text = "网络/连接异常"
return f"大模型因{reason_text}调用失败,正在第 {attempt}/{self.retry_max_attempts} 次重试,{seconds} 秒后继续…"
def _build_circuit_breaker_message(self) -> str:
return "大模型因连续多次调用失败已被熔断保护,请稍候再试。"
def _build_user_message(self, exc: BaseException, reason: str) -> str:
detail = _extract_error_detail(exc)
def _build_user_message(self, reason: str) -> str:
if reason == "quota":
return "大模型因账户额度不足、账单异常或使用受限而拒绝请求,请检查模型账户后再试。"
if reason == "auth":
return "大模型因鉴权失败或无访问权限而拒绝请求,请检查模型的 API Key 与权限配置后再试。"
if reason in {"busy", "transient"}:
return "大模型因服务繁忙或网络异常,多次重试后仍未恢复,请稍候再继续。"
return f"大模型因「{detail}」原因调用失败。"
if reason == "upstream":
return "大模型当前响应异常(上游模型或网关临时故障),多次重试后仍未恢复,请稍后再试或更换模型。"
# NOTE: We intentionally do NOT embed the raw exception detail into the
# user-facing message. Raw `Error code: 400 - {...}` strings — especially
# Python repr-style dict literals — get persisted as AI messages and then
# re-fed to the model on the next turn, where their unescaped quotes /
# braces can corrupt the upstream JSON payload and cause deterministic
# JSONDecodeError loops. Keep the user message clean; full detail is
# already in the backend log via `logger.warning(..., exc_info=exc)`.
return "大模型本次调用未能成功完成,请稍后再试或更换模型。"
def _emit_retry_event(self, attempt: int, wait_ms: int, reason: str) -> None:
try:
@@ -249,7 +283,7 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
)
if retriable:
self._record_failure()
return AIMessage(content=self._build_user_message(exc, reason))
return AIMessage(content=self._build_user_message(reason))
@override
async def awrap_model_call(
@@ -295,7 +329,7 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
)
if retriable:
self._record_failure()
return AIMessage(content=self._build_user_message(exc, reason))
return AIMessage(content=self._build_user_message(reason))
def _matches_any(detail: str, patterns: tuple[str, ...]) -> bool:
@@ -0,0 +1,109 @@
"""Inject stable citation numbering before RAG final answers."""
from __future__ import annotations
import logging
import os
from pathlib import Path
from collections.abc import Awaitable, Callable
from typing import Any, override
from langchain.agents import AgentState
from langchain.agents.middleware import AgentMiddleware
from langchain.agents.middleware.types import ModelCallResult, ModelRequest, ModelResponse
from langchain_core.messages import SystemMessage
from deerflow.runtime.references import build_latest_reference_prompt, build_reference_debug_summary
logger = logging.getLogger(__name__)
_RAG_DEBUG_LOGGER_NAME = "rag_citation_debug"
def _get_rag_debug_logger() -> logging.Logger:
debug_logger = logging.getLogger(_RAG_DEBUG_LOGGER_NAME)
if getattr(debug_logger, "_rag_file_configured", False):
return debug_logger
log_path = Path(os.environ.get("RAG_CITATION_DEBUG_LOG", "logs/rag_citation_debug.log"))
if not log_path.is_absolute():
log_path = Path.cwd() / log_path
log_path.parent.mkdir(parents=True, exist_ok=True)
handler = logging.FileHandler(log_path, encoding="utf-8")
handler.setFormatter(logging.Formatter("%(asctime)s %(levelname)s %(message)s"))
debug_logger.addHandler(handler)
debug_logger.setLevel(logging.INFO)
debug_logger.propagate = False
setattr(debug_logger, "_rag_file_configured", True)
setattr(debug_logger, "_rag_file_path", str(log_path))
return debug_logger
def _log_rag_debug(message: str, *args: Any) -> None:
logger.info(message, *args)
_get_rag_debug_logger().info(message, *args)
class RagCitationMiddleware(AgentMiddleware[AgentState]):
"""Give the model the same numbered references the frontend will render."""
@staticmethod
def _enabled(request: ModelRequest) -> bool:
runtime = getattr(request, "runtime", None)
context = getattr(runtime, "context", None) if runtime is not None else None
if isinstance(context, dict) and context.get("rag_mode_enabled"):
return True
config = getattr(runtime, "config", None) if runtime is not None else None
if isinstance(config, dict):
configurable = config.get("configurable")
if isinstance(configurable, dict) and configurable.get("rag_mode_enabled"):
return True
runtime_context = config.get("context")
if isinstance(runtime_context, dict) and runtime_context.get("rag_mode_enabled"):
return True
return False
@staticmethod
def _patch(request: ModelRequest) -> ModelRequest | None:
if not RagCitationMiddleware._enabled(request):
return None
messages = list(request.messages)
prompt, batch = build_latest_reference_prompt(messages)
if not prompt:
_log_rag_debug(
"middleware reference-prompt enabled=true prompt=false messages=%s",
len(messages),
)
return None
summary = build_reference_debug_summary(messages, [batch] if batch else None)
_log_rag_debug(
"middleware reference-prompt enabled=true prompt=true messages=%s tool_messages=%s "
"raw_items=%s normalized_items=%s kept_items=%s latest_batch_sources=%s",
summary["messages"],
summary["tool_messages"],
summary["raw_items"],
summary["normalized_items"],
summary["kept_items"],
len((batch or {}).get("sources") or []),
)
patched_messages = [
*messages,
SystemMessage(content=prompt),
]
return request.override(messages=patched_messages)
@override
def wrap_model_call(
self,
request: ModelRequest,
handler: Callable[[ModelRequest], ModelResponse],
) -> ModelCallResult:
patched = self._patch(request)
return handler(patched or request)
@override
async def awrap_model_call(
self,
request: ModelRequest,
handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
) -> ModelCallResult:
patched = self._patch(request)
return await handler(patched or request)