feat(turbo): add direct LLM chat mode with runtime toggle

Add 极速 mode that streams via /api/direct-chat without the lead agent stack,
plus VITE_TURBO_MODE_VISIBLE in runtime-config.js (default off). Also disable
global memory, tighten Qwen/vLLM thinking control for flash mode, and force
thinking off on the direct-chat endpoint.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
董宏凯andCursor committed 2026-09-09 21:58:36 +08:00
1 parent 4f7f07504a
commit d765afe2ac
23 files changed
+1295 -394

No files matched your search

@@ -40,77 +40,77 @@ logger = logging.getLogger(__name__)
WRITING_SETUP_MAX_ROUNDS = 8
REPORT_STRUCTURE_SETUP_MAX_ROUNDS = 8
_ES_QUERY_ROUTING_OPEN_TAG = '<skill_routing skill="es_query">'
_SKILLS_BLOCKED_BY_EXCLUDED_TOOL: dict[str, frozenset[str]] = {
# 内置 tool 与历史 custom skill 目录存在两种拼写;禁用编排工具时两者都必须从系统提示和
# runtime skill allowlist 中消失,防止误配的席位仍看到「加载编排技能」。
"agent_orchestration": frozenset({"agent_orchestration", "agent-orchestration"}),
}
_ROUNDTABLE_COORDINATOR_AGENT_ID = "roundtable-coordinator"
_NON_COORDINATOR_ORCHESTRATION_BOUNDARY = """
<orchestration_capability_boundary>
你不是多智能体圆桌总控,因此没有编排、派活或调度其它智能体的权限。
不得加载、读取、检索或调用 agent_orchestration / agent-orchestration,
不得声称准备使用编排技能;请直接完成当前问答或你的本职任务。
</orchestration_capability_boundary>
""".strip()
def _enforce_coordinator_only_tool_exclusions(
agent_id: str | None,
excluded_tools: list[str] | None,
) -> list[str] | None:
"""Make orchestration unavailable to every agent except the fixed roundtable coordinator."""
if agent_id == _ROUNDTABLE_COORDINATOR_AGENT_ID:
return excluded_tools
effective = list(excluded_tools or [])
if "agent_orchestration" not in effective:
effective.append("agent_orchestration")
return effective
def _append_orchestration_capability_boundary(system_prompt: str, agent_id: str | None) -> str:
"""Reinforce the physical tool/skill gate in every non-coordinator system prompt."""
if agent_id == _ROUNDTABLE_COORDINATOR_AGENT_ID:
return system_prompt
return f"{system_prompt}\n\n{_NON_COORDINATOR_ORCHESTRATION_BOUNDARY}"
def _filter_available_skills_for_excluded_tools(
available_skills: list[str] | None,
excluded_tools: list[str] | None,
) -> list[str] | None:
"""让工具级禁用同步收紧 skill allowlist,必要时把「全部」展开为安全列表。"""
if not excluded_tools:
return available_skills
excluded = {name.strip().lower() for name in excluded_tools if name}
blocked = {
skill_name
for tool_name, skill_names in _SKILLS_BLOCKED_BY_EXCLUDED_TOOL.items()
if tool_name in excluded
for skill_name in skill_names
}
if not blocked:
return available_skills
if available_skills is None:
# ``None`` 原本表示「全部启用技能可见」,无法表达「除了总控专属技能之外的全部」。
# 此时显式展开存储中的启用技能;扫描失败则 fail closed,避免角色隔离失效。
try:
from deerflow.skills.storage import get_or_new_skill_storage
available_skills = [
skill.name for skill in get_or_new_skill_storage().load_skills(enabled_only=True)
]
except Exception: # noqa: BLE001
logger.warning("Failed to expand skills while applying excluded-tool boundary", exc_info=True)
available_skills = []
return [name for name in available_skills if name.strip().lower() not in blocked]
_ES_QUERY_ROUTING_OPEN_TAG = '<skill_routing skill="es_query">'
def _log_es_query_routing_registration(
_SKILLS_BLOCKED_BY_EXCLUDED_TOOL: dict[str, frozenset[str]] = {
# 内置 tool 与历史 custom skill 目录存在两种拼写;禁用编排工具时两者都必须从系统提示和
# runtime skill allowlist 中消失,防止误配的席位仍看到「加载编排技能」。
"agent_orchestration": frozenset({"agent_orchestration", "agent-orchestration"}),
}
_ROUNDTABLE_COORDINATOR_AGENT_ID = "roundtable-coordinator"
_NON_COORDINATOR_ORCHESTRATION_BOUNDARY = """
<orchestration_capability_boundary>
你不是多智能体圆桌总控,因此没有编排、派活或调度其它智能体的权限。
不得加载、读取、检索或调用 agent_orchestration / agent-orchestration,
不得声称准备使用编排技能;请直接完成当前问答或你的本职任务。
</orchestration_capability_boundary>
""".strip()
def _enforce_coordinator_only_tool_exclusions(
agent_id: str | None,
excluded_tools: list[str] | None,
) -> list[str] | None:
"""Make orchestration unavailable to every agent except the fixed roundtable coordinator."""
if agent_id == _ROUNDTABLE_COORDINATOR_AGENT_ID:
return excluded_tools
effective = list(excluded_tools or [])
if "agent_orchestration" not in effective:
effective.append("agent_orchestration")
return effective
def _append_orchestration_capability_boundary(system_prompt: str, agent_id: str | None) -> str:
"""Reinforce the physical tool/skill gate in every non-coordinator system prompt."""
if agent_id == _ROUNDTABLE_COORDINATOR_AGENT_ID:
return system_prompt
return f"{system_prompt}\n\n{_NON_COORDINATOR_ORCHESTRATION_BOUNDARY}"
def _filter_available_skills_for_excluded_tools(
available_skills: list[str] | None,
excluded_tools: list[str] | None,
) -> list[str] | None:
"""让工具级禁用同步收紧 skill allowlist,必要时把「全部」展开为安全列表。"""
if not excluded_tools:
return available_skills
excluded = {name.strip().lower() for name in excluded_tools if name}
blocked = {
skill_name
for tool_name, skill_names in _SKILLS_BLOCKED_BY_EXCLUDED_TOOL.items()
if tool_name in excluded
for skill_name in skill_names
}
if not blocked:
return available_skills
if available_skills is None:
# ``None`` 原本表示「全部启用技能可见」,无法表达「除了总控专属技能之外的全部」。
# 此时显式展开存储中的启用技能;扫描失败则 fail closed,避免角色隔离失效。
try:
from deerflow.skills.storage import get_or_new_skill_storage
available_skills = [
skill.name for skill in get_or_new_skill_storage().load_skills(enabled_only=True)
]
except Exception: # noqa: BLE001
logger.warning("Failed to expand skills while applying excluded-tool boundary", exc_info=True)
available_skills = []
return [name for name in available_skills if name.strip().lower() not in blocked]
def _log_es_query_routing_registration(
*,
system_prompt: str,
app_config: AppConfig,
@@ -401,9 +401,9 @@ def _build_middlewares(
if summarization_middleware is not None:
middlewares.append(summarization_middleware)
# Add TodoList middleware if plan mode is enabled
is_plan_mode = cfg.get("is_plan_mode", False)
is_bootstrap = cfg.get("is_bootstrap", False)
# Add TodoList middleware if plan mode is enabled
is_plan_mode = cfg.get("is_plan_mode", False)
is_bootstrap = cfg.get("is_bootstrap", False)
is_writing_setup = cfg.get("is_writing_setup", False)
is_report_structure_setup = cfg.get("is_report_structure_setup", False)
todo_list_middleware = _create_todo_list_middleware(is_plan_mode)
@@ -433,7 +433,9 @@ def _build_middlewares(
# Add MemoryMiddleware (after TitleMiddleware) — skip in bootstrap mode (agent creation flow) and ephemeral runs
memory_injection_enabled = cfg.get("memory_injection_enabled", True)
if not is_bootstrap and not is_writing_setup and not is_report_structure_setup and not is_ephemeral:
memory_cfg = getattr(resolved_app_config, "memory", None)
memory_globally_enabled = memory_cfg is not None and getattr(memory_cfg, "enabled", False)
if memory_globally_enabled and not is_bootstrap and not is_writing_setup and not is_report_structure_setup and not is_ephemeral:
middlewares.append(MemoryMiddleware(agent_name=agent_name, memory_config=resolved_app_config.memory, injection_enabled=memory_injection_enabled))
# KnowledgeRagMiddleware — inject knowledge-base recall (phase 2). Gated
@@ -574,10 +576,10 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
# actually injected. Default off → no behavior change for normal chats.
thinking_force_disabled = bool(cfg.get("thinking_force_disabled", False))
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)
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)
max_concurrent_subagents = cfg.get("max_concurrent_subagents", 3)
is_bootstrap = cfg.get("is_bootstrap", False)
# 写作配置 Q&A:与 is_bootstrap 同构的另一种「最小化 bootstrap」——挂 setup_writing 工具、
@@ -603,11 +605,11 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
if not is_writing_mode_report_path(writing_artifact_path):
writing_artifact_path = None
agent_id = validate_agent_id(cfg.get("agent_id") or cfg.get("agent_name"))
requested_excluded_tools = _enforce_coordinator_only_tool_exclusions(
agent_id,
cfg.get("excluded_tools") or None,
)
agent_id = validate_agent_id(cfg.get("agent_id") or cfg.get("agent_name"))
requested_excluded_tools = _enforce_coordinator_only_tool_exclusions(
agent_id,
cfg.get("excluded_tools") or None,
)
try:
agent_config = load_agent_config(agent_id) if not is_bootstrap and not is_writing_setup and not is_report_structure_setup else None
@@ -656,14 +658,14 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
from deerflow.agents.deep_research.retrieval import merge_collector_skill_allowlist, runtime_extra_skills
runtime_available_skills = merge_collector_skill_allowlist(
agent_id, runtime_available_skills, runtime_extra_skills(cfg)
)
runtime_available_skills = _filter_available_skills_for_excluded_tools(
runtime_available_skills,
requested_excluded_tools,
)
prompt_available_skills = set(runtime_available_skills) if runtime_available_skills is not None else None
runtime_available_skills = merge_collector_skill_allowlist(
agent_id, runtime_available_skills, runtime_extra_skills(cfg)
)
runtime_available_skills = _filter_available_skills_for_excluded_tools(
runtime_available_skills,
requested_excluded_tools,
)
prompt_available_skills = set(runtime_available_skills) if runtime_available_skills is not None else None
config["metadata"].update(
{
@@ -744,14 +746,14 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
# 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),
tools=get_available_tools(
model_name=model_name,
subagent_enabled=subagent_enabled,
is_scheduled_run=cfg.get("is_scheduled_run", False),
excluded_tools=requested_excluded_tools,
allow_agent_orchestration=agent_id == _ROUNDTABLE_COORDINATOR_AGENT_ID,
app_config=resolved_app_config,
)
tools=get_available_tools(
model_name=model_name,
subagent_enabled=subagent_enabled,
is_scheduled_run=cfg.get("is_scheduled_run", False),
excluded_tools=requested_excluded_tools,
allow_agent_orchestration=agent_id == _ROUNDTABLE_COORDINATOR_AGENT_ID,
app_config=resolved_app_config,
)
+ [setup_agent],
middleware=_build_middlewares(config, model_name=model_name, app_config=resolved_app_config),
system_prompt=apply_prompt_template(
@@ -771,19 +773,19 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
# Default lead agent (unchanged behavior)
is_notebook_mode = bool(cfg.get("notebook_search_space_id"))
notebook_source_ids: list[int] = cfg.get("notebook_source_ids") or []
default_tools = (
[]
if disable_tools
else get_available_tools(
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,
allow_agent_orchestration=agent_id == _ROUNDTABLE_COORDINATOR_AGENT_ID,
app_config=resolved_app_config,
)
)
default_tools = (
[]
if disable_tools
else get_available_tools(
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,
allow_agent_orchestration=agent_id == _ROUNDTABLE_COORDINATOR_AGENT_ID,
app_config=resolved_app_config,
)
)
if agent_id == FORCED_RESEARCH_AGENT_ID:
default_tools = filter_forced_research_tools(default_tools)
if is_notebook_mode:
@@ -825,8 +827,8 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
writing_mode=writing_mode,
writing_artifact_path=writing_artifact_path,
)
else:
system_prompt = apply_prompt_template(
else:
system_prompt = apply_prompt_template(
subagent_enabled=subagent_enabled,
max_concurrent_subagents=max_concurrent_subagents,
agent_id=agent_id,
@@ -839,22 +841,22 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
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)),
)
# 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()
if presentation_theme in {"dark", "light"}:
system_prompt += (
"\n\n【当前界面主题】\n"
f"当前页面呈现主题为 `{presentation_theme}`。若本轮生成汇报页面,请优先选择同色系 HTML 模板;"
"除非用户明确要求覆盖该主题。"
)
system_prompt = _append_orchestration_capability_boundary(system_prompt, agent_id)
_log_es_query_routing_registration(
is_scheduled_run=bool(cfg.get("is_scheduled_run", False)),
)
# 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()
if presentation_theme in {"dark", "light"}:
system_prompt += (
"\n\n【当前界面主题】\n"
f"当前页面呈现主题为 `{presentation_theme}`。若本轮生成汇报页面,请优先选择同色系 HTML 模板;"
"除非用户明确要求覆盖该主题。"
)
system_prompt = _append_orchestration_capability_boundary(system_prompt, agent_id)
_log_es_query_routing_registration(
system_prompt=system_prompt,
app_config=resolved_app_config,
agent_id=agent_id,