sync update from LAN repo (source master @ ca48c4ec)

This commit is contained in:
gozeros committed 2026-11-05 10:53:54 +08:00
1 parent 8bc8b0e082
commit ea9dcff9c0
873 files changed
+107866 -14011

No files matched your search

@@ -149,34 +149,37 @@ _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”。所选知识库是回答中具体事实的首要依据,以下五个只读工具负责补充背景证据。
知识库选择只限定事实检索来源,不关闭原有能力:可以正常使用当前智能体已配置的技能,以及计算、文件处理、制图和内容生成等非外部检索工具。
知识库检索与技能调用都嵌套在同一个模型工作流中:先读取当前工具结果,再决定下一步;不得把检索、技能和最终回答作为同一批并行动作发出。
技能发现可按原流程调用 search_skills、skill_list、skill_view;不得用 Web、MCP 或子智能体检索来替代所选知识库中的专属事实。
检索用于增强回答,不替代你原本的理解、归纳、写作和推理能力:可以用稳定的通用知识解释背景、术语和因果关系,但必须与“库内材料明确记载的事实”及“基于材料的推断”区分。
不得用模型记忆编造知识库专属的日期、数字、专名、原句或内部事实。用户若明确要求只依据知识库,则执行更严格的仅库内回答。
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。
你必须像获取普通信息一样,先理解用户目标并判断是否需要知识库资料,再按证据缺口有目的地调用检索工具。用户只需直接提问,不需要预先编写检索步骤。若问题不需要所选知识库中的专属事实,可直接回答或正常使用技能,不得为了展示检索过程而检索。
1. 一旦判断需要检索知识库,本轮第一个检索动作必须调用 ordinary_knowledge_search,使用 1 至 5 个聚焦改写对原始文档做向量/混合检索,先取得与问题直接相关的分片、文档名和实体线索。调用检索工具时不要同时输出回答正文;等待工具结果后再继续判断。
2. 读取首轮原始分片后,再判断是否需要调用 ordinary_wiki_search。需要梳理目录结构、概念关系、时间线或寻找扩展线索时,结合原问题与首轮 RAG 命中的实体、标题、时间和主题构造安全 POSIX 正则;证据已经足够时可以不查 Wiki,直接回答。
Wiki 搜索只返回页面入口;只对回答确实相关的页面调用 ordinary_wiki_read_page。长页面按 next_cursor 续读,禁止无游标重复读取同一页面。
3. Wiki 若暴露首轮 RAG 未覆盖且回答确实需要的实体、时间、链接或来源,再用新的查询调用 ordinary_knowledge_search,或用 ordinary_chunk_regex_search 核验精确数字、编号、日期、原句和固定词形。
深读 Wiki 后,把其中与问题有关的标题、别名和关联链接作为候选扩展词;根据尚未补齐的证据缺口进行删减、组合或改写,再形成新的 query。实际采用的扩展 query 必须保留在工具参数中,以便检索过程展示;不得机械照搬全部候选词。
需要邻近上下文或后续分页时调用 ordinary_document_chunks。没有实际证据缺口就停止检索并回答,不要求每轮同时覆盖所有检索面。
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。每个新问题重新取得首轮证据,不把旧轮工具正文当作当前证据。
5. 原始分片、Wiki 结构与必要的补充核验已经足够后,用连贯自然语言直接完整回答。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。本模式不套用经典参考文献模式:
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,否则不要为检索结果额外编排编号或文献列表。
- 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 classic citation blocks from a smart-reasoning system prompt."""
"""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 ("citations", "rag_citation_mode"):
for tag in ("rag_citation_mode",):
opening = f"<{tag}>"
closing = f"</{tag}>"
while opening in cleaned:
@@ -237,6 +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()
@@ -671,7 +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
@@ -723,13 +780,6 @@ 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
@@ -762,6 +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
@@ -776,6 +832,8 @@ 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)
@@ -786,12 +844,13 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
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 knowledge_space_strict:
if classic_knowledge_space_strict:
subagent_enabled = False
max_concurrent_subagents = cfg.get("max_concurrent_subagents", 3)
is_bootstrap = cfg.get("is_bootstrap", False)
@@ -833,7 +892,18 @@ 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)
if smart_knowledge_reasoning:
# Smart mode uses the WeKnora-compatible five-surface tool set and
@@ -871,11 +941,6 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
]
)
)
# 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"]))
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:
@@ -884,6 +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)
@@ -943,13 +1014,14 @@ 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,
@@ -968,7 +1040,7 @@ 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_setup_middlewares(
config,
@@ -1005,7 +1077,7 @@ 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_setup_middlewares(
config,
@@ -1030,15 +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),
@@ -1063,14 +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:
@@ -1083,7 +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]
@@ -1176,6 +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,
@@ -1184,7 +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,