feat: 新增 agentfx 任务深链与研判报告入库流程

支持 goPath=agentfx 的 iframe 嵌入问答、内置研判助手种子与报告导入协助卡,并完善深度研究流式报告能力。

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
董宏凯andCursor committed 2026-08-27 16:32:22 +08:00
1 parent 6e9e89a6cb
commit 52b9cbf707
34 files changed
+4488 -392

No files matched your search

@@ -175,10 +175,14 @@ class DeerFlowCompletionBackend:
config: DeepResearchConfig,
event_sink: EventSink | None = None,
cancellation: CancellationToken | None = None,
force_disable_thinking: bool = False,
) -> None:
self._config = config
self._events = event_sink
self._cancel = cancellation
# 与 lead agent 同语义的「思考硬关」开关(内网 vLLM/Qwen 只声明
# supports_thinking 时,thinking_enabled=False 不会注入 enable_thinking=false)。
self._force_disable_thinking = force_disable_thinking
self._models: dict[str, BaseChatModel] = {}
# Running token totals (keyed by model name) for the final usage report.
self._usage_totals: dict[str, dict[str, int]] = {}
@@ -196,7 +200,11 @@ class DeerFlowCompletionBackend:
return cached
from deerflow.models import create_chat_model
model = create_chat_model(name=name, thinking_enabled=self._config.thinking_enabled)
model = create_chat_model(
name=name,
thinking_enabled=self._config.thinking_enabled,
force_disable_thinking=self._force_disable_thinking,
)
self._models[cache_key] = model
return model
@@ -57,7 +57,7 @@ _TOOL_SOURCE_TYPES: dict[str, str] = {
}
# Tool results that never carry report materials.
_IGNORED_TOOLS = {"ask_clarification", "write_todos"}
_IGNORED_TOOLS = {"ask_clarification", "write_todos", "deep_research_report"}
async def harvest_thread_materials(
@@ -80,7 +80,24 @@ async def harvest_thread_materials(
return [], ""
values = checkpoint_tuple.checkpoint.get("channel_values", {}) or {}
messages = values.get("messages") or []
return harvest_messages_materials(messages, query=query)
def harvest_messages_materials(
messages: list[Any],
*,
query: str,
) -> tuple[list[ResearchMaterial], str]:
"""Convert an in-memory message list into research materials.
Same conversion as :func:`harvest_thread_materials` minus the checkpoint
read — for callers that already hold the messages (e.g. a tool reading
``runtime.state`` inside the same run).
Returns ``(materials, agent_summary)`` — the summary is the agent's last
visible text (its collection小结) and doubles as a fallback material when
the thread produced no harvestable tool results.
"""
materials = _harvest_tool_messages(messages, query=query)
agent_summary = _last_assistant_text(messages)
if not materials and agent_summary.strip():
@@ -235,5 +252,6 @@ class StaticHarvestMaterialProvider:
__all__ = [
"MAX_HARVESTED_MATERIALS",
"StaticHarvestMaterialProvider",
"harvest_messages_materials",
"harvest_thread_materials",
]
@@ -5,12 +5,12 @@ from langchain.agents.middleware import AgentMiddleware
from langchain_core.runnables import RunnableConfig
from deerflow.agents.lead_agent.prompt import apply_prompt_template
from deerflow.agents.middlewares.clarification_middleware import ClarificationMiddleware
from deerflow.agents.middlewares.forced_research_middleware import (
FORCED_RESEARCH_AGENT_ID,
ForcedResearchMiddleware,
filter_forced_research_tools,
)
from deerflow.agents.middlewares.clarification_middleware import ClarificationMiddleware
from deerflow.agents.middlewares.forced_research_middleware import (
FORCED_RESEARCH_AGENT_ID,
ForcedResearchMiddleware,
filter_forced_research_tools,
)
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
@@ -373,21 +373,21 @@ def _build_middlewares(
middlewares.append(SkillReviewMiddleware())
# PositionArtifactCapMiddleware — position-roundtable seats (and summary)
# PositionArtifactCapMiddleware — position-roundtable seats (and summary)
# may only emit N outputs write_file calls per user turn. Action-plan skips
# the context flag so it can still write report + JSON.
from deerflow.agents.middlewares.position_artifact_cap_middleware import PositionArtifactCapMiddleware
middlewares.append(PositionArtifactCapMiddleware())
# The built-in forced-research responder has a hard per-turn collection
# gate plus temporary-Python-only sandbox IO. This is runtime enforcement,
# not merely SOUL prompt steering, so weak models cannot answer before a
# successful retrieval/skill-script result exists.
if agent_name == FORCED_RESEARCH_AGENT_ID:
middlewares.append(ForcedResearchMiddleware())
# LoopDetectionMiddleware — detect and break repetitive tool call loops
middlewares.append(PositionArtifactCapMiddleware())
# The built-in forced-research responder has a hard per-turn collection
# gate plus temporary-Python-only sandbox IO. This is runtime enforcement,
# not merely SOUL prompt steering, so weak models cannot answer before a
# successful retrieval/skill-script result exists.
if agent_name == FORCED_RESEARCH_AGENT_ID:
middlewares.append(ForcedResearchMiddleware())
# LoopDetectionMiddleware — detect and break repetitive tool call loops
middlewares.append(LoopDetectionMiddleware())
# Inject custom middlewares before ClarificationMiddleware
@@ -638,20 +638,32 @@ 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 = get_available_tools(
default_tools = 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,
app_config=resolved_app_config,
)
if agent_id == FORCED_RESEARCH_AGENT_ID:
default_tools = filter_forced_research_tools(default_tools)
app_config=resolved_app_config,
)
if agent_id == FORCED_RESEARCH_AGENT_ID:
default_tools = filter_forced_research_tools(default_tools)
if is_notebook_mode:
from deerflow.tools.builtins.notebook_search_tool import notebook_search_tool
default_tools = [notebook_search_tool] + default_tools
# 写作模式首轮(尚无 md 产物):挂深度研究报告工具,报告由管线生成而非
# 模型手工 write_file。续写轮(已有 writing_artifact_path)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:
from deerflow.tools.builtins.deep_research_report_tool import deep_research_report_tool
default_tools = [*default_tools, deep_research_report_tool]
logger.info(
"Writing mode first report turn: deep_research_report tool mounted for agent=%s",
agent_id,
)
# Custom agents are steered ONLY by their own SOUL.md — no generic
# super-agent system template and no personal memory injection. We branch to
@@ -271,42 +271,42 @@ answers — and never create a skill silently without user confirmation.
"""
def _build_es_query_routing_section(
*,
app_config: AppConfig | None = None,
available_skills: set[str] | None = None,
) -> str:
"""Return the configurable ``es_query`` routing rule for the system prompt.
A custom agent with an explicit skill allowlist only receives the rule when
``es_query`` is in that allowlist. The default lead agent has no allowlist,
so the operator-controlled config switch is authoritative there.
"""
if available_skills is not None and "es_query" not in available_skills:
return ""
config = app_config
if config is None:
try:
from deerflow.config import get_app_config
config = get_app_config()
except Exception:
logger.debug("Failed to load es_query routing config", exc_info=True)
return ""
routing = getattr(getattr(config, "skills", None), "es_query_routing", None)
if routing is None or not routing.enabled:
return ""
prompt = (routing.prompt or "").strip()
if not prompt:
return ""
return f'<skill_routing skill="es_query">\n{prompt}\n</skill_routing>'
def _build_available_subagents_description(available_names: list[str], bash_available: bool, *, app_config: AppConfig | None = None) -> str:
def _build_es_query_routing_section(
*,
app_config: AppConfig | None = None,
available_skills: set[str] | None = None,
) -> str:
"""Return the configurable ``es_query`` routing rule for the system prompt.
A custom agent with an explicit skill allowlist only receives the rule when
``es_query`` is in that allowlist. The default lead agent has no allowlist,
so the operator-controlled config switch is authoritative there.
"""
if available_skills is not None and "es_query" not in available_skills:
return ""
config = app_config
if config is None:
try:
from deerflow.config import get_app_config
config = get_app_config()
except Exception:
logger.debug("Failed to load es_query routing config", exc_info=True)
return ""
routing = getattr(getattr(config, "skills", None), "es_query_routing", None)
if routing is None or not routing.enabled:
return ""
prompt = (routing.prompt or "").strip()
if not prompt:
return ""
return f'<skill_routing skill="es_query">\n{prompt}\n</skill_routing>'
def _build_available_subagents_description(available_names: list[str], bash_available: bool, *, app_config: AppConfig | None = None) -> str:
"""Dynamically build subagent type descriptions from registry.
Mirrors Codex's pattern where agent_type_description is dynamically generated
@@ -583,11 +583,11 @@ User: "staging"
You: "Deploying to staging..." [proceed]
</clarification_system>
{skills_section}
{es_query_routing_section}
{bootstrap_section}
{skills_section}
{es_query_routing_section}
{bootstrap_section}
{deferred_tools_section}
@@ -700,12 +700,12 @@ combined with a FastAPI gateway for REST API access [citation:FastAPI](https://f
# style, clarification system, citation rules, response-style / critical-reminder
# defaults) and no personal memory injection. Only the mechanical sections an
# agent still needs to actually use its configured skills/tools are kept.
AGENT_ONLY_PROMPT_TEMPLATE = """{soul}
{network_policy_section}{skills_section}
{es_query_routing_section}
{deferred_tools_section}
AGENT_ONLY_PROMPT_TEMPLATE = """{soul}
{network_policy_section}{skills_section}
{es_query_routing_section}
{deferred_tools_section}
{writing_mode_section}
@@ -758,16 +758,16 @@ def apply_agent_only_prompt_template(
soul = get_agent_soul(storage_agent_id)
skills_section = get_skills_prompt_section(
available_skills,
app_config=app_config,
display_enabled=rag_mode_enabled,
)
es_query_routing_section = _build_es_query_routing_section(
app_config=app_config,
available_skills=available_skills,
)
deferred_tools_section = get_deferred_tools_prompt_section(app_config=app_config)
skills_section = get_skills_prompt_section(
available_skills,
app_config=app_config,
display_enabled=rag_mode_enabled,
)
es_query_routing_section = _build_es_query_routing_section(
app_config=app_config,
available_skills=available_skills,
)
deferred_tools_section = get_deferred_tools_prompt_section(app_config=app_config)
n = max_concurrent_subagents
subagent_section = _build_subagent_section(n, app_config=app_config) if subagent_enabled else ""
@@ -777,17 +777,36 @@ def apply_agent_only_prompt_template(
if writing_artifact_path
else ""
)
writing_mode_section = (
f"""<writing_mode>
if writing_mode and not writing_artifact_path:
# 与通用模板一致:对话行为照旧,交付时刻的「写 md」动作交给
# deep_research_report 管线;资料不足时先经 ask_clarification 问用户。
writing_mode_section = """<writing_mode>
Writing mode is enabled for this run (first report turn — no existing artifact).
- HARD RULE: the report file is produced ONLY by the `deep_research_report` tool — never write it yourself with `write_file`, never `present_files` it.
- Work normally (chat, search, clarify) until the report is due; then produce it by calling the `deep_research_report` tool:
`topic` = the user's subject copied faithfully, `focus` = the user's extra requirements (if any).
The tool reuses the conversation's collected materials and writes the full report (general-analysis structure with references)
to `/mnt/user-data/outputs/report.md`, presented automatically.
- If the tool reports insufficient materials, call `ask_clarification` to ask whether to search for more; after the user answers,
re-call the tool with the same `topic`/`focus` plus `retrieve=true` (search more) or `retrieve=false` (existing materials only).
NEVER pass `retrieve=true` before the user has answered — the tool rejects it and bounces you back to the question.
- Do NOT write the report body yourself with `write_file`, and do NOT call `present_files` for it.
- After the tool returns, reply briefly (2-3 sentences) summarizing the core conclusions; do not duplicate the report inline.
- Only if the tool fails, tell the user and offer to retry.
</writing_mode>
"""
else:
writing_mode_section = (
f"""<writing_mode>
Writing mode is enabled for this run.
- Your final deliverable MUST be a Markdown file saved under `/mnt/user-data/outputs/` (e.g. `/mnt/user-data/outputs/report.md`).
- If a current Markdown artifact is provided below, revise that file rather than creating a new one unless the user explicitly asks for a separate document.
- After writing the file, call `present_files` with the Markdown file path.
{current_writing_artifact}</writing_mode>
"""
if writing_mode
else ""
)
if writing_mode
else ""
)
acp_section = _build_acp_section(app_config=app_config)
custom_mounts_section = _build_custom_mounts_section(app_config=app_config)
@@ -809,10 +828,10 @@ RAG/reference display mode is enabled.
)
prompt = AGENT_ONLY_PROMPT_TEMPLATE.format(
soul=soul,
skills_section=skills_section,
es_query_routing_section=es_query_routing_section,
deferred_tools_section=deferred_tools_section,
soul=soul,
skills_section=skills_section,
es_query_routing_section=es_query_routing_section,
deferred_tools_section=deferred_tools_section,
writing_mode_section=writing_mode_section,
subagent_section=subagent_section,
network_policy_section=network_policy_section,
@@ -1421,15 +1440,15 @@ 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,
display_enabled=rag_mode_enabled,
)
es_query_routing_section = _build_es_query_routing_section(
app_config=app_config,
available_skills=available_skills,
)
skills_section = get_skills_prompt_section(
available_skills,
app_config=app_config,
display_enabled=rag_mode_enabled,
)
es_query_routing_section = _build_es_query_routing_section(
app_config=app_config,
available_skills=available_skills,
)
# Get deferred tools section (tool_search)
deferred_tools_section = get_deferred_tools_prompt_section(app_config=app_config)
@@ -1438,8 +1457,35 @@ def apply_prompt_template(
if writing_artifact_path
else ""
)
writing_mode_section = (
f"""
if writing_mode and not writing_artifact_path:
# 首轮:对话行为一切照旧,仅在交付时刻把「写 md」动作交给
# deep_research_report 管线(收割对话资料 + 通用分析结构)。资料不足时
# 工具会返回指令,让模型经 ask_clarification 征询用户是否补充检索。
writing_mode_section = """
<writing_mode>
Writing mode is enabled for this run (first report turn — no existing artifact).
**Mandatory Output Contract**
- HARD RULE: the report file is produced ONLY by the `deep_research_report` tool. You are FORBIDDEN from writing the report yourself — do not call `write_file`
(or any file-writing tool) to create the report, and do not call `present_files` for it.
- Treat the user's request as a research-report writing task. Work normally until delivery time: discuss, search, and clarify exactly as you always would.
- When the report file is due — the user asks to start writing, or you have finished gathering what the conversation needs — call the `deep_research_report` tool:
- `topic`: the user's subject/thesis, copied faithfully.
- `focus`: the user's extra requirements (length, tone, must-cover points), if any.
- The tool reuses the materials already collected in this conversation (and searches more only if the user agrees), then writes the full
report (general-analysis structure with a numbered reference section) to `/mnt/user-data/outputs/report.md`, presented automatically.
- If the tool replies that the harvested materials are insufficient, it returns instructions: immediately call `ask_clarification` asking
whether to search for more materials, and after the user answers, re-call `deep_research_report` with the same `topic`/`focus` plus
`retrieve=true` (user agreed) or `retrieve=false` (user declined — write with what exists).
NEVER pass `retrieve=true` before the user has answered — the tool rejects it and bounces you back to the question.
- Do NOT write the report body yourself with `write_file`, and do NOT call `present_files` for the report — the tool already handles both.
- While the tool runs, stay silent; after it returns, reply briefly (2-3 sentences) summarizing the report's core conclusions and point the user to the Markdown artifact. Do not duplicate the report inline.
- Only if the tool fails, tell the user what went wrong and offer to retry; do not silently fall back to writing the report by hand unless the user explicitly asks.
</writing_mode>
"""
else:
writing_mode_section = (
f"""
<writing_mode>
Writing mode is enabled for this run.
@@ -1456,9 +1502,9 @@ Writing mode is enabled for this run.
{current_writing_artifact}
</writing_mode>
"""
if writing_mode
else ""
)
if writing_mode
else ""
)
# Build ACP agent section only if ACP agents are configured
acp_section = _build_acp_section(app_config=app_config)
@@ -1486,10 +1532,10 @@ When a search, knowledge, skill, or MCP tool returns a result list:
# Format the prompt with dynamic skills and memory
prompt = SYSTEM_PROMPT_TEMPLATE.format(
agent_name=agent_name or storage_agent_id or "CM 助手",
soul=get_agent_soul(storage_agent_id),
skills_section=skills_section,
es_query_routing_section=es_query_routing_section,
bootstrap_section=bootstrap_section,
soul=get_agent_soul(storage_agent_id),
skills_section=skills_section,
es_query_routing_section=es_query_routing_section,
bootstrap_section=bootstrap_section,
deferred_tools_section=deferred_tools_section,
writing_mode_section=writing_mode_section,
memory_context=memory_context,