sync update from LAN repo (source dhk-1 @ 3b1ec5b4)

This commit is contained in:
gozeros committed 2026-11-05 10:56:34 +08:00
1 parent da6521b6cb
commit 264e2200d6
995 files changed
+148105 -13399

No files matched your search

@@ -21,9 +21,14 @@ from deerflow.agents.middlewares.memory_middleware import MemoryMiddleware
from deerflow.agents.middlewares.multi_page_html_middleware import MultiPageHtmlMiddleware
from deerflow.agents.middlewares.prompt_prefix_middleware import PromptPrefixMiddleware
from deerflow.agents.middlewares.rag_citation_middleware import RagCitationMiddleware
from deerflow.agents.middlewares.recent_turns_middleware import RecentTurnsMiddleware
from deerflow.agents.middlewares.setup_agent_approval_middleware import SetupAgentApprovalMiddleware
from deerflow.agents.middlewares.setup_report_structure_approval_middleware import SetupReportStructureApprovalMiddleware
from deerflow.agents.middlewares.setup_writing_approval_middleware import SetupWritingApprovalMiddleware
from deerflow.agents.middlewares.smart_reasoning_retrieval_middleware import (
SMART_REASONING_TOOL_NAMES,
SmartReasoningRetrievalMiddleware,
)
from deerflow.agents.middlewares.subagent_limit_middleware import SubagentLimitMiddleware
from deerflow.agents.middlewares.summarization_middleware import BeforeSummarizationHook, DeerFlowSummarizationMiddleware
from deerflow.agents.middlewares.task_external_context_middleware import TaskExternalContextMiddleware
@@ -35,8 +40,11 @@ from deerflow.agents.middlewares.view_image_middleware import ViewImageMiddlewar
from deerflow.agents.thread_state import ThreadState
from deerflow.config.agents_config import load_agent_config, load_agent_soul, validate_agent_id
from deerflow.config.app_config import AppConfig, get_app_config
from deerflow.config.system_settings import get_memory_runtime_settings
from deerflow.config.system_settings import get_conversation_context_settings, get_memory_runtime_settings
from deerflow.models import create_chat_model
from deerflow.tools.builtins.llmwiki_reasoning_tools import (
CLASSIC_LLMWIKI_TOOL_NAMES,
)
logger = logging.getLogger(__name__)
@@ -133,9 +141,55 @@ _KNOWLEDGE_SPACE_RETRIEVAL_EXCLUSIONS = {
"invoke_acp_agent",
"agent_orchestration",
"task",
"deep_research_report",
}
_SMART_REASONING_TOOL_NAMES = SMART_REASONING_TOOL_NAMES
_CLASSIC_LLMWIKI_TOOL_NAMES = CLASSIC_LLMWIKI_TOOL_NAMES
_SMART_REASONING_KNOWLEDGE_SCOPE_PROMPT = """
<knowledge_space_scope>
本轮已启用普通知识库“智能推理检索”。这表示所选知识库作为普通问答中的按需资料工具可用,不表示每个问题都必须检索,也不把当前对话切换成独立的检索流程。正常问答、技能、Web、MCP、子智能体及其它工具仍按原有权限和意图规则使用。
先判断当前问题是否真的需要知识库中的事实:
- 问候、闲聊、自我介绍、“你有哪些技能/会做什么/有什么工具”、界面或系统设置说明、对用户已提供文本的改写,以及不依赖库内专属事实的一般推理,直接回答,不得为了展示流程而调用知识库工具。
- 用户询问所选资料、库内人物/事件/制度/数据,或回答需要从库内核验具体事实时,才调用下面的知识库工具。用户明确要求只依据知识库时,执行更严格的仅库内回答。
- 知识库工具与其它正常工具并列。按问题选择最合适的来源;不要仅因为用户选中了知识库就屏蔽、忽略或替代普通技能和工具。
需要检索时按证据缺口选择工具,不存在强制顺序,也不要求同时覆盖原文和 Wiki:
1. ordinary_knowledge_search 适合用 1 至 5 个聚焦改写检索原始文档分片;ordinary_document_chunks 用于读取命中分片的邻近上下文或文档后续分页。
2. ordinary_wiki_search 适合用安全 POSIX 正则发现相关 Wiki 页面、概念关系和时间线;搜索结果只是页面入口,确实需要正文时再调用 ordinary_wiki_read_page,并按 next_cursor 续读长页面。
若返回 native_regex_supported=false,本轮不得再次尝试 Wiki 正则,也不得改用目录枚举或逐页读取模拟搜索;继续使用已取得的原文证据、其它正常工具和模型原有能力完成回答。
3. ordinary_chunk_regex_search 仅在需要核验精确数字、编号、日期、原句或固定词形时使用。没有实际证据缺口就停止检索并回答。
4. 同一用户回合不得重复相同 query、相同正则、相同 page_ref+cursor 或相同文档分页。某个表达式无结果时最多做一次有实质变化的改写。新问题若需要库内事实,应重新取得相关证据,不把旧轮工具正文当作当前证据。
5. 检索后用连贯自然语言直接回答。Wiki 与原始文档冲突时,以能定位到原始文档切片的证据为准并说明差异。本模式不套用经典知识库编号参考文献模式:
- 专名、作品/文件/活动标题、日期、时间、数字、单位和数量关系必须逐项回到原始文档切片核对,并按证据原样保留;不得近义改写专名或混用相似名称。证据未明确给出的细节不要补写。
- 最终回答不要求使用 [1]、[2] 等编号,也不要求生成文末参考文献列表。
- 除非用户本轮明确要求引用格式,否则不要为知识库检索结果额外编排编号或文献列表;若实际使用 Web 等外部来源,仍遵守对应工具的引用规则。
- 不要向用户暴露工具名、内部 ID 或检索协议。具体事实证据不足时明确说明缺少什么;通用背景补充应使用“一般而言”等清晰措辞,不能伪装成库内结论。
</knowledge_space_scope>"""
def _without_classic_citation_contracts(system_prompt: str) -> str:
"""Remove only the classic numbered-RAG citation contract.
Smart knowledge results have their own structured trace, but normal Web
tools remain available and must retain the base prompt's URL citation
contract.
"""
cleaned = system_prompt
for tag in ("rag_citation_mode",):
opening = f"<{tag}>"
closing = f"</{tag}>"
while opening in cleaned:
start = cleaned.find(opening)
end = cleaned.find(closing, start + len(opening))
if end < 0:
break
cleaned = cleaned[:start] + cleaned[end + len(closing) :]
return cleaned
def _knowledge_space_strict_exclusions(excluded_tools: list[str] | None) -> list[str]:
return list(dict.fromkeys([*(excluded_tools or []), *sorted(_KNOWLEDGE_SPACE_RETRIEVAL_EXCLUSIONS)]))
@@ -186,6 +240,51 @@ def _get_runtime_config(config: RunnableConfig) -> dict:
return cfg
def _runtime_tool_is_enabled(cfg: dict, tool_name: str) -> bool:
"""Return whether a middleware may rely on a runtime-filtered tool.
Tool-driving middleware must follow the same allow/exclude contract as the
tool loader. Otherwise a middleware can synthesize a call to a tool that
was deliberately removed from the graph. That happened in roundtable
synthesis: seat deliverables containing ``.md`` made the Markdown delivery
middleware inject ``write_file`` into the text-only coordinator, replacing
its final summary with an impossible tool call.
"""
if bool(cfg.get("disable_tools", False)):
return False
normalized_name = str(tool_name or "").strip()
raw_allowed = cfg.get("allowed_tools")
if raw_allowed is not None:
allowed = (
{raw_allowed.strip()}
if isinstance(raw_allowed, str) and raw_allowed.strip()
else {
str(name).strip()
for name in raw_allowed
if str(name).strip()
}
if isinstance(raw_allowed, (list, tuple, set))
else set()
)
if normalized_name not in allowed:
return False
raw_excluded = cfg.get("excluded_tools")
excluded = (
{raw_excluded.strip()}
if isinstance(raw_excluded, str) and raw_excluded.strip()
else {
str(name).strip()
for name in raw_excluded
if str(name).strip()
}
if isinstance(raw_excluded, (list, tuple, set))
else set()
)
return normalized_name not in excluded
def _resolve_model_name(requested_model_name: str | None = None, *, app_config: AppConfig | None = None) -> str:
"""Resolve a runtime model name safely, falling back to default if invalid. Returns None if no models are configured."""
app_config = app_config or get_app_config()
@@ -271,6 +370,28 @@ def _create_summarization_middleware(*, app_config: AppConfig | None = None, for
)
def _create_context_reduction_middleware(
*,
app_config: AppConfig | None = None,
summarization_requested: bool = False,
) -> AgentMiddleware | None:
"""Resolve the active context-reduction strategy for this model run.
The administrator-managed recent-turn window has priority over summary
compression and is global once enabled. Its request-only trimming avoids
the summarization LLM call while leaving checkpoint history untouched.
With the setting off, the existing per-request/config summarization
behavior is preserved exactly.
"""
context_settings = get_conversation_context_settings()
if context_settings.recent_turn_window_enabled:
return RecentTurnsMiddleware()
return _create_summarization_middleware(
app_config=app_config,
force_enabled=summarization_requested,
)
def _create_todo_list_middleware(is_plan_mode: bool) -> TodoMiddleware | None:
"""Create and configure the TodoList middleware.
@@ -421,18 +542,31 @@ def _build_middlewares(
# 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):
smart_knowledge_reasoning = (
str(cfg.get("knowledge_retrieval_mode") or "classic").strip().lower()
== "smart_reasoning"
)
if smart_knowledge_reasoning:
# Request-only cleanup keeps checkpoint/UI history intact while each
# new question starts with fresh evidence. It also owns the hard
# forty-call synthesis budget for this mode.
middlewares.append(SmartReasoningRetrievalMiddleware())
if cfg.get("rag_mode_enabled", False) and not smart_knowledge_reasoning:
# Smart reasoning deliberately has no numbered-reference contract.
# Keep this fail-safe in the harness even though the Gateway also
# clears rag_mode_enabled for validated smart runs.
middlewares.append(RagCitationMiddleware())
# Add summarization middleware. The on/off switch now lives in the chat UI
# (上下文压缩 toggle, default off) and arrives as ``summarization_enabled`` in
# the runtime config; the config file only supplies tuning. force_enabled
# lets the per-request toggle turn it on even when config.enabled is false,
# while non-UI paths (channels/scheduled) still fall back to config.enabled.
# Add the active context-reduction middleware. The admin-controlled
# recent-three-turn window is global and takes priority; when it is off,
# the existing chat-level summarization toggle/config fallback is unchanged.
summarization_requested = cfg.get("summarization_enabled", False)
summarization_middleware = _create_summarization_middleware(app_config=resolved_app_config, force_enabled=summarization_requested)
if summarization_middleware is not None:
middlewares.append(summarization_middleware)
context_reduction_middleware = _create_context_reduction_middleware(
app_config=resolved_app_config,
summarization_requested=summarization_requested,
)
if context_reduction_middleware is not None:
middlewares.append(context_reduction_middleware)
# Add TodoList middleware if plan mode is enabled
is_plan_mode = cfg.get("is_plan_mode", False)
@@ -489,7 +623,16 @@ def _build_middlewares(
# to the global ``knowledge.rag_enabled`` config. No-op when knowledge is
# disabled. Skipped in bootstrap/writing-setup/ephemeral like MemoryMiddleware.
knowledge_cfg = getattr(resolved_app_config, "knowledge", None)
if not is_bootstrap and not is_writing_setup and not is_report_structure_setup and not is_gongwen_template_setup and not is_ephemeral and knowledge_cfg is not None and knowledge_cfg.enabled:
if (
not smart_knowledge_reasoning
and not is_bootstrap
and not is_writing_setup
and not is_report_structure_setup
and not is_gongwen_template_setup
and not is_ephemeral
and knowledge_cfg is not None
and knowledge_cfg.enabled
):
from deerflow.agents.middlewares.knowledge_rag_middleware import KnowledgeRagMiddleware
middlewares.append(KnowledgeRagMiddleware(limit=knowledge_cfg.rag_limit, default_enabled=knowledge_cfg.rag_enabled))
@@ -576,7 +719,16 @@ def _build_middlewares(
# LoopDetectionMiddleware — detect and break repetitive tool call loops
middlewares.append(LoopDetectionMiddleware())
if not writing_mode and not workbench_write_now:
# This middleware creates a synthetic ``write_file`` call after the model
# responds. Mount it only when that tool is actually present under the
# request policy. In particular the Step-2 coordinator is text-only; its
# synthesis input can legitimately quote seat instructions mentioning
# ``.md`` and must still finish with a normal textual summary.
if (
not writing_mode
and not workbench_write_now
and _runtime_tool_is_enabled(cfg, "write_file")
):
middlewares.append(MarkdownFileDeliverableMiddleware())
# Inject custom middlewares before ClarificationMiddleware
@@ -628,18 +780,31 @@ def _build_middlewares(
)
)
# 商业情报助理的确认卡由 DetailConfirmMiddleware 追加在纯正文之后。
# 必须放在 ClarificationMiddleware 之前,这样总结消息本身不带工具调用。
if agent_name == "biz-detail-assistant":
from deerflow.agents.middlewares.detail_confirm_middleware import DetailConfirmMiddleware
middlewares.append(DetailConfirmMiddleware())
# ClarificationMiddleware should always be last
middlewares.append(ClarificationMiddleware())
return middlewares
def _build_setup_middlewares(
config: RunnableConfig,
model_name: str | None,
round_cap_middleware: AgentMiddleware,
*,
app_config: AppConfig,
) -> list[AgentMiddleware]:
"""Build a setup-agent chain while preserving clarification as the final gate."""
middlewares = _build_middlewares(
config,
model_name=model_name,
app_config=app_config,
)
clarification_middleware = middlewares.pop()
if not isinstance(clarification_middleware, ClarificationMiddleware):
raise RuntimeError("ClarificationMiddleware must be the final middleware")
middlewares.extend([round_cap_middleware, clarification_middleware])
return middlewares
def make_lead_agent(config: RunnableConfig):
"""LangGraph graph factory; keep the signature compatible with LangGraph Server."""
runtime_config = _get_runtime_config(config)
@@ -647,6 +812,12 @@ def make_lead_agent(config: RunnableConfig):
return _make_lead_agent(config, app_config=runtime_app_config or get_app_config())
# The Gateway may build this synchronous graph in its dedicated initializer.
# It restores ContextVar changes and keeps MCP discovery on the request loop;
# graph execution/checkpoint operations still run in the original async task.
make_lead_agent._deerflow_background_build = True
def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
# Lazy import to avoid circular dependency
from deerflow.tools import get_available_tools
@@ -661,13 +832,25 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
# create_chat_model(force_disable_thinking=...) so enable_thinking=false is
# actually injected. Default off → no behavior change for normal chats.
thinking_force_disabled = bool(cfg.get("thinking_force_disabled", False))
# Roundtable runs opt in explicitly. Missing/false keeps the global switch.
write_file_compatibility = True if cfg.get("roundtable_write_file_compatibility") else None
reasoning_effort = cfg.get("reasoning_effort", None)
requested_model_name: str | None = cfg.get("model_name") or cfg.get("model")
is_plan_mode = cfg.get("is_plan_mode", False)
disable_tools = bool(cfg.get("disable_tools", False))
subagent_enabled = cfg.get("subagent_enabled", False)
knowledge_space_strict = bool(cfg.get("knowledge_space_strict", False))
if knowledge_space_strict:
knowledge_retrieval_mode = str(
cfg.get("knowledge_retrieval_mode") or "classic"
).strip().lower()
smart_knowledge_reasoning = knowledge_retrieval_mode == "smart_reasoning"
classic_knowledge_space_strict = knowledge_space_strict and not smart_knowledge_reasoning
if smart_knowledge_reasoning:
# Gateway validation owns this mode. Keep a harness-side fail-safe so
# internal callers cannot accidentally construct a smart run without
# its required read-only retrieval tools.
disable_tools = False
if classic_knowledge_space_strict:
subagent_enabled = False
max_concurrent_subagents = cfg.get("max_concurrent_subagents", 3)
is_bootstrap = cfg.get("is_bootstrap", False)
@@ -684,7 +867,10 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
report_structure_setup_skills = cfg.get("valid_retrieval_skills") if is_report_structure_setup else None
report_structure_setup_quick = bool(cfg.get("writing_setup_quick", False)) if (is_report_structure_setup or is_gongwen_template_setup) else False
memory_injection_enabled = bool(cfg.get("memory_injection_enabled", True))
rag_mode_enabled = bool(cfg.get("rag_mode_enabled", False))
# The Gateway owns this flag, but internal harness callers can still pass
# inconsistent config. Smart mode must never inherit the classic prompt
# that requires [1]/[2] references or a reference list.
rag_mode_enabled = bool(cfg.get("rag_mode_enabled", False)) and not smart_knowledge_reasoning
writing_mode = bool(cfg.get("writing_mode", False))
writing_artifact_path = cfg.get("writing_artifact_path") if writing_mode else None
workbench_write_now = bool(cfg.get("workbench_write_now", False))
@@ -706,13 +892,55 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
agent_id,
cfg.get("excluded_tools") or None,
)
if knowledge_space_strict:
raw_allowed_tools = cfg.get("allowed_tools")
if isinstance(raw_allowed_tools, str):
requested_allowed_tools: list[str] | None = (
[raw_allowed_tools] if raw_allowed_tools.strip() else []
)
elif isinstance(raw_allowed_tools, (list, tuple, set)):
requested_allowed_tools = [
str(name) for name in raw_allowed_tools if str(name).strip()
]
else:
requested_allowed_tools = None
if classic_knowledge_space_strict:
requested_excluded_tools = _knowledge_space_strict_exclusions(requested_excluded_tools)
# 确认卡不能和总结写在同一条模型消息里,否则消息列表不显示正文。
# 这个智能体改为先输出表格,再由 DetailConfirmMiddleware 追加协助卡。
if agent_id == "biz-detail-assistant":
requested_excluded_tools = list(dict.fromkeys([*(requested_excluded_tools or []), "ask_clarification"]))
if smart_knowledge_reasoning:
# Smart mode uses the WeKnora-compatible five-surface tool set and
# deliberately skips the legacy all-in-one search tool.
requested_excluded_tools = list(
dict.fromkeys(
[
*(
name
for name in (requested_excluded_tools or [])
if name not in _SMART_REASONING_TOOL_NAMES
),
*sorted(_CLASSIC_LLMWIKI_TOOL_NAMES),
]
)
)
else:
selected_ordinary = cfg.get("llmwiki_knowledge_base_ids")
selected_assistant = cfg.get("assistant_knowledge_base_ids")
classic_enabled_names: set[str] = set()
if isinstance(selected_ordinary, list) and selected_ordinary:
classic_enabled_names.update(_CLASSIC_LLMWIKI_TOOL_NAMES)
elif isinstance(selected_assistant, list) and selected_assistant:
classic_enabled_names.add("llmwiki_search")
requested_excluded_tools = list(
dict.fromkeys(
[
*(
name
for name in (requested_excluded_tools or [])
if name not in classic_enabled_names
),
*sorted(_SMART_REASONING_TOOL_NAMES),
*sorted(_CLASSIC_LLMWIKI_TOOL_NAMES - classic_enabled_names),
]
)
)
try:
agent_config = load_agent_config(agent_id) if not is_bootstrap and not is_writing_setup and not is_report_structure_setup and not is_gongwen_template_setup else None
except FileNotFoundError:
@@ -721,6 +949,12 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
agent_display_name = agent_config.name if agent_config else agent_id
# Custom agent model from agent config (if any), or None to let _resolve_model_name pick the default
agent_model_name = agent_config.model if agent_config and agent_config.model else None
# The global prompt suffix remains on unless this particular custom agent
# explicitly opts out. ``getattr`` preserves compatibility with callers
# that supply lightweight agent-config test doubles.
model_prompt_suffix_enabled = bool(
getattr(agent_config, "model_prompt_suffix_enabled", True)
) if agent_config else True
# Final model name resolution: request → agent config → global default, with fallback for unknown names
model_name = _resolve_model_name(requested_model_name or agent_model_name, app_config=resolved_app_config)
@@ -769,10 +1003,6 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
runtime_available_skills = resolve_collector_skill_visibility(
agent_id, runtime_available_skills, runtime_extra_skills(cfg)
)
if knowledge_space_strict:
# Skills can hide their own network/search clients, so selected-space
# runs fail closed by exposing no runtime skill surface.
runtime_available_skills = []
runtime_available_skills = _filter_available_skills_for_excluded_tools(
runtime_available_skills,
requested_excluded_tools,
@@ -784,13 +1014,15 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
"agent_id": agent_id,
"agent_name": agent_display_name or "default",
"model_name": model_name or "default",
"model_prompt_suffix_enabled": model_prompt_suffix_enabled,
"thinking_enabled": thinking_enabled,
"reasoning_effort": reasoning_effort,
"is_plan_mode": is_plan_mode,
"subagent_enabled": subagent_enabled,
"writing_mode": writing_mode,
"writing_artifact_path": writing_artifact_path,
"knowledge_space_strict": knowledge_space_strict,
"knowledge_space_strict": classic_knowledge_space_strict,
"knowledge_retrieval_mode": knowledge_retrieval_mode,
"tool_groups": agent_config.tool_groups if agent_config else None,
"available_skills": runtime_available_skills,
}
@@ -800,7 +1032,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
# 写作配置 Q&A:最小化 agent —— 只挂 setup_writing(提交配置)+ ask_clarification(弹澄清卡)。
# 关键:绝不给全套工具、绝不复用 lead 超级体提示,否则模型会把「新能源汽车市场分析」之类
# 长得像任务的输入直接做掉,而不是转成写作配置提交。
# 末尾追加 WritingSetupRoundCapMiddleware:澄清最多 3 轮,到顶用 tool_choice 强制提交配置卡。
# 在 ClarificationMiddleware 前挂 WritingSetupRoundCapMiddleware:到顶用 tool_choice 强制提交配置卡。
# 快捷模式:用默认选项、不逐项澄清提问(但仍生成并确认大纲)——直接**不挂** ask_clarification,
# 让模型物理上问不了澄清问题,比只靠提示词约束更可靠。
from deerflow.agents.lead_agent.prompt import build_writing_setup_prompt
@@ -808,12 +1040,14 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
setup_tools = [setup_writing] if writing_setup_quick else [setup_writing, ask_clarification_tool]
return create_agent(
model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, app_config=resolved_app_config),
model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, write_file_compatibility=write_file_compatibility, app_config=resolved_app_config),
tools=setup_tools,
middleware=[
*_build_middlewares(config, model_name=model_name, app_config=resolved_app_config),
middleware=_build_setup_middlewares(
config,
model_name,
WritingSetupRoundCapMiddleware(max_rounds=WRITING_SETUP_MAX_ROUNDS),
],
app_config=resolved_app_config,
),
system_prompt=build_writing_setup_prompt(
writing_setup_article_types,
quick_mode=writing_setup_quick,
@@ -843,10 +1077,11 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
)
)
return create_agent(
model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, app_config=resolved_app_config),
model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, write_file_compatibility=write_file_compatibility, app_config=resolved_app_config),
tools=setup_tools,
middleware=[
*_build_middlewares(config, model_name=model_name, app_config=resolved_app_config),
middleware=_build_setup_middlewares(
config,
model_name,
WritingSetupRoundCapMiddleware(
max_rounds=REPORT_STRUCTURE_SETUP_MAX_ROUNDS,
tool_name="setup_report_structure",
@@ -858,7 +1093,8 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
"retrieval_directions / retrieval_skills 不确定就留空、content 必须给出一份 Markdown 大纲。"
),
),
],
app_config=resolved_app_config,
),
system_prompt=setup_prompt,
state_schema=ThreadState,
)
@@ -866,15 +1102,17 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
if is_bootstrap:
# Special bootstrap agent with minimal prompt for initial custom agent creation flow
return create_agent(
model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, app_config=resolved_app_config),
model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, write_file_compatibility=write_file_compatibility, app_config=resolved_app_config),
tools=get_available_tools(
include_mcp=not knowledge_space_strict,
include_mcp=not classic_knowledge_space_strict,
model_name=model_name,
subagent_enabled=subagent_enabled,
is_scheduled_run=cfg.get("is_scheduled_run", False),
excluded_tools=requested_excluded_tools,
allowed_tools=requested_allowed_tools,
allow_agent_orchestration=agent_id in _COORDINATOR_AGENT_IDS,
app_config=resolved_app_config,
write_file_compatibility=write_file_compatibility,
)
+ [setup_agent],
middleware=_build_middlewares(config, model_name=model_name, app_config=resolved_app_config),
@@ -899,14 +1137,16 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
[]
if disable_tools
else get_available_tools(
include_mcp=not knowledge_space_strict,
include_mcp=not classic_knowledge_space_strict,
model_name=model_name,
groups=agent_config.tool_groups if agent_config else None,
subagent_enabled=subagent_enabled,
is_scheduled_run=cfg.get("is_scheduled_run", False),
excluded_tools=requested_excluded_tools,
allowed_tools=requested_allowed_tools,
allow_agent_orchestration=agent_id in _COORDINATOR_AGENT_IDS,
app_config=resolved_app_config,
write_file_compatibility=write_file_compatibility,
)
)
if agent_id == FORCED_RESEARCH_AGENT_ID:
@@ -919,7 +1159,12 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
# 模型手工 write_file。续写轮(已有 report.md)不挂,保持原有「直接编辑
# 该文件」行为。不受 tool_groups/excluded_tools 过滤——它是写作模式合约
# 的核心交付手段。
if writing_mode and not writing_artifact_path and not knowledge_space_strict:
# Workbench delivery writes only from collected materials; it does not
# expand the selected knowledge scope or enable external search tools.
workbench_report_delivery = workbench_write_now and bool(cfg.get("workbench_session_id"))
if writing_mode and not writing_artifact_path and (
not classic_knowledge_space_strict or workbench_report_delivery
):
from deerflow.tools.builtins.deep_research_report_tool import deep_research_report_tool
default_tools = [*default_tools, deep_research_report_tool]
@@ -961,12 +1206,23 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
available_skills=prompt_available_skills,
app_config=resolved_app_config,
memory_injection_enabled=memory_injection_enabled,
# Smart ordinary-KB mode has its own authorized database scope.
# The legacy /mnt/knowledge/routing.md provider points at an
# unrelated mounted file tree and can misresolve “这个库”. Suppress
# only that static block; personal memory and Hindsight remain on.
include_legacy_knowledge_routing=not smart_knowledge_reasoning,
rag_mode_enabled=rag_mode_enabled,
is_notebook_mode=is_notebook_mode,
notebook_source_ids=notebook_source_ids,
is_scheduled_run=bool(cfg.get("is_scheduled_run", False)),
)
if smart_knowledge_reasoning:
# The base lead prompt contains a web-research citation/Sources
# contract and an optional numbered RAG block. Neither contract
# applies to the WeKnora-style smart knowledge loop.
system_prompt = _without_classic_citation_contracts(system_prompt)
# 前端可提供当前视觉明暗模式给页面生成类任务。把值放入系统提示而不是
# 指望模型从 ToolRuntime.context 猜测,确保自定义 SOUL 也能稳定使用。
presentation_theme = str(cfg.get("presentation_theme") or "").strip().lower()
@@ -979,11 +1235,21 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
system_prompt = _append_orchestration_capability_boundary(system_prompt, agent_id)
if knowledge_space_strict:
if knowledge_space_strict and smart_knowledge_reasoning:
system_prompt += _SMART_REASONING_KNOWLEDGE_SCOPE_PROMPT
elif knowledge_space_strict:
system_prompt += (
"\n\n<knowledge_space_scope>\n"
"本轮用户已选择知识空间。只能依据所选知识空间的预检索上下文与 llmwiki_search 结果回答;"
"禁止使用 Web、MCP、外部检索技能、子智能体或模型常识补齐。若所选空间证据不足,必须明确说明。\n"
"本轮用户已选择知识空间。知识库选择只限定具体事实的检索来源,不关闭当前智能体已配置的技能,"
"也不关闭计算、文件处理、制图和内容生成等非外部检索能力。先检查系统已注入的预检索证据;证据不足时"
"再根据缺口调用 llmwiki_search。选择多个普通知识库时优先按 knowledge_base_id 分库检索,避免一个库的结果掩盖另一个库。用户询问库的范围、"
"目录或主题概览时,可调用 llmwiki_browse_wiki_index;搜索命中只给出片段而回答需要全文、关系或上下文时,"
"再调用 llmwiki_read_wiki_page 读取相关页面。只读取能填补当前证据缺口的不同页面,不得重复相同页面或"
"游标。原始分片与 Wiki 冲突时以可定位的原始材料为准。\n"
"技能与知识库应协同使用:知识库先提供专属事实和背景,技能负责分析、计算、转换、制图或生成交付物;"
"写文件和展示成品应在必要证据收集完成后进行。可以使用稳定的通用知识解释背景、术语和因果关系,但要"
"明确区分库内事实、基于材料的推断和一般性补充,不得用模型记忆编造知识库专属日期、数字、专名或内部事实。"
"禁止使用 Web、MCP、外部检索技能或子智能体替代所选知识库检索。若所选空间证据不足,必须明确说明。\n"
"</knowledge_space_scope>"
)
@@ -991,6 +1257,14 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
system_prompt += project_system_prompt(cfg)
if cfg.get("knowledge_research_session_id") and disable_tools:
from deerflow.agents.lead_agent.knowledge_research_prompt import KNOWLEDGE_RESEARCH_PROMPT
# Research is an inline document editor. Generic file/tool contracts
# contradict this server-enforced tool-free mode.
system_prompt = KNOWLEDGE_RESEARCH_PROMPT
_log_es_query_routing_registration(
system_prompt=system_prompt,
app_config=resolved_app_config,
@@ -999,7 +1273,15 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
)
return create_agent(
model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, reasoning_effort=reasoning_effort, force_disable_thinking=thinking_force_disabled, app_config=resolved_app_config),
model=create_chat_model(
name=model_name,
thinking_enabled=thinking_enabled,
reasoning_effort=reasoning_effort,
force_disable_thinking=thinking_force_disabled,
write_file_compatibility=write_file_compatibility,
prompt_suffix_enabled=model_prompt_suffix_enabled,
app_config=resolved_app_config,
),
tools=default_tools,
middleware=_build_middlewares(config, model_name=model_name, agent_name=agent_id, app_config=resolved_app_config),
system_prompt=system_prompt,