fix: writing mode uses chat search then pipeline write only

Keep deep_research_report as a write-only deliverable from conversation materials, block post-success rewrite loops with a summary nudge, fill thin/empty pools from model knowledge, and open the sandbox only on real report.md content.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
董宏凯andCursor committed 2026-08-30 12:41:04 +08:00
1 parent 9afb9750e5
commit 49c01cb72b
20 files changed
+1449 -381

No files matched your search

@@ -1,5 +1,5 @@
import hashlib
import logging
import hashlib
import logging
from langchain.agents import create_agent
from langchain.agents.middleware import AgentMiddleware
@@ -38,45 +38,45 @@ logger = logging.getLogger(__name__)
# 写作配置 Q&A 现在「先生成大纲并与用户多轮确认 → 再提交配置」,需要更多来回,故上限放宽。
# 仍是兜底安全阀:到顶才用 tool_choice 强制提交,正常流程由用户确认大纲后自然提交。
WRITING_SETUP_MAX_ROUNDS = 8
REPORT_STRUCTURE_SETUP_MAX_ROUNDS = 8
_ES_QUERY_ROUTING_OPEN_TAG = '<skill_routing skill="es_query">'
def _log_es_query_routing_registration(
*,
system_prompt: str,
app_config: AppConfig,
agent_id: str | None,
available_skills: set[str] | None,
) -> None:
"""Log whether the exact configured ES routing block reached this Q&A prompt."""
routing = app_config.skills.es_query_routing
configured_prompt = (routing.prompt or "").strip()
expected_block = (
f"{_ES_QUERY_ROUTING_OPEN_TAG}\n{configured_prompt}\n</skill_routing>"
if configured_prompt
else ""
)
eligible = available_skills is None or "es_query" in available_skills
registered = bool(expected_block) and expected_block in system_prompt
fingerprint = (
hashlib.sha256(configured_prompt.encode("utf-8")).hexdigest()[:12]
if configured_prompt
else "-"
)
log = logger.info if registered or not routing.enabled or not eligible else logger.warning
log(
"ES query routing prompt registration: agent_id=%s enabled=%s eligible=%s "
"registered=%s prompt_chars=%d prompt_sha256=%s",
agent_id or "default",
routing.enabled,
eligible,
registered,
len(configured_prompt),
fingerprint,
)
REPORT_STRUCTURE_SETUP_MAX_ROUNDS = 8
_ES_QUERY_ROUTING_OPEN_TAG = '<skill_routing skill="es_query">'
def _log_es_query_routing_registration(
*,
system_prompt: str,
app_config: AppConfig,
agent_id: str | None,
available_skills: set[str] | None,
) -> None:
"""Log whether the exact configured ES routing block reached this Q&A prompt."""
routing = app_config.skills.es_query_routing
configured_prompt = (routing.prompt or "").strip()
expected_block = (
f"{_ES_QUERY_ROUTING_OPEN_TAG}\n{configured_prompt}\n</skill_routing>"
if configured_prompt
else ""
)
eligible = available_skills is None or "es_query" in available_skills
registered = bool(expected_block) and expected_block in system_prompt
fingerprint = (
hashlib.sha256(configured_prompt.encode("utf-8")).hexdigest()[:12]
if configured_prompt
else "-"
)
log = logger.info if registered or not routing.enabled or not eligible else logger.warning
log(
"ES query routing prompt registration: agent_id=%s enabled=%s eligible=%s "
"registered=%s prompt_chars=%d prompt_sha256=%s",
agent_id or "default",
routing.enabled,
eligible,
registered,
len(configured_prompt),
fingerprint,
)
def _get_runtime_config(config: RunnableConfig) -> dict:
@@ -426,6 +426,18 @@ def _build_middlewares(
if agent_name == FORCED_RESEARCH_AGENT_ID:
middlewares.append(ForcedResearchMiddleware())
# 写作模式首轮:允许先聊/检索;只有模型开始写 md 时才改走 deep_research_report。
# 不设 tool_choice——thinking 模式的兼容接口会 400。
writing_mode = bool(cfg.get("writing_mode", False))
writing_artifact_path = cfg.get("writing_artifact_path") if writing_mode else None
if writing_mode:
from deerflow.tools.builtins.deep_research_report_tool import is_writing_mode_report_path
if not is_writing_mode_report_path(writing_artifact_path if isinstance(writing_artifact_path, str) else None):
from deerflow.agents.middlewares.writing_mode_report_middleware import WritingModeReportMiddleware
middlewares.append(WritingModeReportMiddleware())
# LoopDetectionMiddleware — detect and break repetitive tool call loops
middlewares.append(LoopDetectionMiddleware())
@@ -517,6 +529,12 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
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():
writing_artifact_path = None
# 只有写作模式产物 report.md 才走续写;其它 leftover .md 不得跳过管线。
if writing_artifact_path is not None:
from deerflow.tools.builtins.deep_research_report_tool import is_writing_mode_report_path
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 = cfg.get("excluded_tools") or None
@@ -691,10 +709,10 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
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。
# 写作模式首轮(尚无 report.md):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 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:
from deerflow.tools.builtins.deep_research_report_tool import deep_research_report_tool
@@ -726,8 +744,8 @@ 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,
@@ -737,20 +755,20 @@ 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,
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)),
)
_log_es_query_routing_registration(
system_prompt=system_prompt,
app_config=resolved_app_config,
agent_id=agent_id,
available_skills=prompt_available_skills,
)
return create_agent(
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)),
)
_log_es_query_routing_registration(
system_prompt=system_prompt,
app_config=resolved_app_config,
agent_id=agent_id,
available_skills=prompt_available_skills,
)
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),
tools=default_tools,
middleware=_build_middlewares(config, model_name=model_name, agent_name=agent_id, app_config=resolved_app_config),
@@ -486,19 +486,19 @@ task(description="Oracle Cloud analysis", prompt="...", subagent_type="general-p
</subagent_system>"""
_ORDINARY_QA_MARKDOWN_FORMAT_SECTION = """
<ordinary_qa_markdown_format>
当普通问答需要输出带章节结构的较长内容时,必须采用以下 Markdown 规范,以便下载为 Word 后得到正确的自动编号和公文标题格式:
- 文档主标题使用 `# 主标题`;主标题只写标题文字,不添加章节编号。
- 一级章节使用 `## 一、标题`,二级章节使用 `### (一) 标题`,三级章节使用 `#### 1. 标题`。二级编号必须使用半角括号 `(一)`。
- 不要在标题中使用 `2.1`、`2.1.1` 等多级阿拉伯数字编号;章节编号直接按上述三种形式书写。
- 默认禁止使用以 `-`、`*`、`+` 或 `•` 开头的无序圆点列表;并列内容优先改写为完整自然段,确需逐项表达时使用有序列表。只有用户明确要求无序列表时才使用。
- 简短问答不必为了套用格式而强行添加标题;本规范只约束实际使用的标题和列表。
</ordinary_qa_markdown_format>
"""
SYSTEM_PROMPT_TEMPLATE = """
_ORDINARY_QA_MARKDOWN_FORMAT_SECTION = """
<ordinary_qa_markdown_format>
当普通问答需要输出带章节结构的较长内容时,必须采用以下 Markdown 规范,以便下载为 Word 后得到正确的自动编号和公文标题格式:
- 文档主标题使用 `# 主标题`;主标题只写标题文字,不添加章节编号。
- 一级章节使用 `## 一、标题`,二级章节使用 `### (一) 标题`,三级章节使用 `#### 1. 标题`。二级编号必须使用半角括号 `(一)`。
- 不要在标题中使用 `2.1`、`2.1.1` 等多级阿拉伯数字编号;章节编号直接按上述三种形式书写。
- 默认禁止使用以 `-`、`*`、`+` 或 `•` 开头的无序圆点列表;并列内容优先改写为完整自然段,确需逐项表达时使用有序列表。只有用户明确要求无序列表时才使用。
- 简短问答不必为了套用格式而强行添加标题;本规范只约束实际使用的标题和列表。
</ordinary_qa_markdown_format>
"""
SYSTEM_PROMPT_TEMPLATE = """
<role>
You are {agent_name}, an internal super agent.
</role>
@@ -624,15 +624,15 @@ You: "Deploying to staging..." [proceed]
{acp_section}
</working_directory>
<response_style>
- Clear and Concise: Avoid over-formatting unless requested
- Natural Tone: Use paragraphs and prose, not bullet points by default
- Action-Oriented: Focus on delivering results, not explaining processes
</response_style>
{ordinary_qa_format_section}
<citations>
<response_style>
- Clear and Concise: Avoid over-formatting unless requested
- Natural Tone: Use paragraphs and prose, not bullet points by default
- Action-Oriented: Focus on delivering results, not explaining processes
</response_style>
{ordinary_qa_format_section}
<citations>
**CRITICAL: Always include citations when using web search results**
- **When to Use**: MANDATORY after web_search, web_fetch, or any external information source
@@ -792,8 +792,7 @@ def apply_agent_only_prompt_template(
else ""
)
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。
# 对话行为照旧;只有开始写 md 时才走 deep_research_report。资料不足也直接开写。
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.
@@ -801,11 +800,11 @@ Writing mode is enabled for this run (first report turn — no existing artifact
`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.
- If harvested materials are thin, still call `deep_research_report` once and let it write immediately.
Do NOT ask the user whether to search for more, and do NOT call `ask_clarification` for insufficient materials.
- Search with your normal conversation tools. `deep_research_report` only writes the report file from materials already in this conversation; it does not search.
- 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.
- After the tool returns successfully: do NOT call `read_file`, `write_file`, `str_replace`, or `present_files` on the report. Reply briefly (2-3 sentences) summarizing the core conclusions. Do not rewrite even if you think materials were thin.
- Only if the tool fails, tell the user and offer to retry.
</writing_mode>
"""
@@ -1407,7 +1406,7 @@ def build_report_structure_setup_prompt(
"""
def apply_prompt_template(
def apply_prompt_template(
subagent_enabled: bool = False,
max_concurrent_subagents: int = 3,
*,
@@ -1419,11 +1418,11 @@ 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,
is_scheduled_run: bool = False,
) -> str:
rag_mode_enabled: bool = False,
is_notebook_mode: bool = False,
notebook_source_ids: list[int] | None = None,
is_scheduled_run: bool = False,
) -> str:
storage_agent_id = agent_id if agent_id is not None else agent_name
# Get memory context
@@ -1473,9 +1472,8 @@ def apply_prompt_template(
else ""
)
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。
# 对话行为一切照旧;只有开始写 md 时才把交付交给 deep_research_report。
# 资料不足也直接开写,不再经 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).
@@ -1487,14 +1485,13 @@ Writing mode is enabled for this run (first report turn — no existing artifact
- 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
- The tool reuses the materials already collected in this conversation, 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.
- If harvested materials are thin, still call `deep_research_report` once and let it start the writing pipeline immediately.
Do NOT ask the user whether to search for more materials, and do NOT call `ask_clarification` for that reason.
- Search with your normal conversation tools. `deep_research_report` only writes the report file from materials already in this conversation; it does not search.
- 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.
- While the tool runs, stay silent. After it returns successfully: do NOT call `read_file`, `write_file`, `str_replace`, or `present_files` on the report. 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, and do not rewrite even if you think materials were thin.
- 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>
"""
@@ -1528,7 +1525,7 @@ 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_mode_section = (
"""
<rag_citation_mode>
RAG/reference display mode is enabled.
@@ -1541,15 +1538,15 @@ When a search, knowledge, skill, or MCP tool returns a result list:
</rag_citation_mode>
"""
if rag_mode_enabled
else ""
)
ordinary_qa_format_section = (
_ORDINARY_QA_MARKDOWN_FORMAT_SECTION
if not is_bootstrap and not writing_mode and not is_notebook_mode and not is_scheduled_run
else ""
)
# Format the prompt with dynamic skills and memory
else ""
)
ordinary_qa_format_section = (
_ORDINARY_QA_MARKDOWN_FORMAT_SECTION
if not is_bootstrap and not writing_mode and not is_notebook_mode and not is_scheduled_run
else ""
)
# 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),
@@ -1562,10 +1559,10 @@ When a search, knowledge, skill, or MCP tool returns a result list:
network_policy_section=network_policy_section,
subagent_section=subagent_section,
subagent_reminder=subagent_reminder,
subagent_thinking=subagent_thinking,
acp_section=acp_and_mounts_section,
ordinary_qa_format_section=ordinary_qa_format_section,
)
subagent_thinking=subagent_thinking,
acp_section=acp_and_mounts_section,
ordinary_qa_format_section=ordinary_qa_format_section,
)
prompt = prompt + rag_mode_section + _current_time_context()