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

@@ -43,6 +43,41 @@ if TYPE_CHECKING:
logger = logging.getLogger(__name__)
def _parent_isolated_run_config(run_name: str) -> dict[str, Any]:
"""Keep LangGraph runtime, drop only inherited LLM streaming callbacks.
``StreamWriter`` reads ``get_config()[configurable]`` at emit time. Replacing
the child config with a blank dict makes ``report_delta`` / ``report_thinking``
raise and get swallowed, so the sandbox only appears after the file lands.
"""
parent: dict[str, Any] = {}
try:
from langchain_core.runnables.config import var_child_runnable_config
current = var_child_runnable_config.get()
if isinstance(current, dict):
parent = dict(current)
except Exception: # noqa: BLE001 — isolation must never fail the completion
parent = {}
tags = [tag for tag in (parent.get("tags") or []) if tag != "nostream"]
tags.append("nostream")
parent["callbacks"] = []
parent["tags"] = tags
parent["run_name"] = run_name
return parent
def _clear_inherited_runnable_config(isolated: dict[str, Any]) -> Callable[[], None]:
"""Install the isolated config for the duration of ``astream``."""
try:
from langchain_core.runnables.config import var_child_runnable_config
except ImportError:
return lambda: None
token = var_child_runnable_config.set(isolated)
return lambda: var_child_runnable_config.reset(token)
# ── message conversion ──────────────────────────────────────────────────────
@@ -347,26 +382,39 @@ class DeerFlowCompletionBackend:
text_parts: list[str] = []
usage: dict[str, Any] = {}
inline_think_filter = InlineThinkTagFilter()
async for chunk in model.astream(to_langchain_messages(messages), **kwargs):
if self._cancel is not None:
self._cancel.raise_if_cancelled()
reasoning = _reasoning_delta(chunk)
if reasoning and on_reasoning is not None:
await on_reasoning(reasoning)
# A few OpenAI-compatible models put their reasoning in the
# ordinary content channel instead of ``reasoning_content``.
# Split it from Markdown even when tags span tokens: callers may
# display a live thought trace, but it never joins ``text``.
delta, inline_reasoning = inline_think_filter.push_parts(_content_to_text(chunk))
if inline_reasoning and on_reasoning is not None:
await on_reasoning(inline_reasoning)
if delta:
text_parts.append(delta)
if on_delta is not None:
await on_delta(delta)
chunk_usage = _extract_usage(chunk)
if any(chunk_usage.values()):
usage = chunk_usage
# Isolate from the parent LangGraph run: inherited callbacks would
# replay every report token into the chat ``messages`` stream, so the
# left-hand bubble and the sandbox would write the same prose at once.
# Custom ``report_delta`` events remain the only live preview channel.
isolated_config = _parent_isolated_run_config(f"deep_research:{operation}")
reset_parent_config = _clear_inherited_runnable_config(isolated_config)
try:
async for chunk in model.astream(
to_langchain_messages(messages),
config=isolated_config,
**kwargs,
):
if self._cancel is not None:
self._cancel.raise_if_cancelled()
reasoning = _reasoning_delta(chunk)
if reasoning and on_reasoning is not None:
await on_reasoning(reasoning)
# A few OpenAI-compatible models put their reasoning in the
# ordinary content channel instead of ``reasoning_content``.
# Split it from Markdown even when tags span tokens: callers may
# display a live thought trace, but it never joins ``text``.
delta, inline_reasoning = inline_think_filter.push_parts(_content_to_text(chunk))
if inline_reasoning and on_reasoning is not None:
await on_reasoning(inline_reasoning)
if delta:
text_parts.append(delta)
if on_delta is not None:
await on_delta(delta)
chunk_usage = _extract_usage(chunk)
if any(chunk_usage.values()):
usage = chunk_usage
finally:
reset_parent_config()
tail, inline_reasoning_tail = inline_think_filter.finish_parts()
if inline_reasoning_tail and on_reasoning is not None:
@@ -198,35 +198,63 @@ class BasicRunner:
cancellation.raise_if_cancelled()
if not all_materials:
knowledge_fill = _allows_knowledge_fill(request.config.report_instruction)
await events.emit(
"warning",
phase="curating",
payload={"code": "no_materials", "message": "未收集到任何研究材料", "recoverable": False},
payload={
"code": "no_materials",
"message": (
"未收集到外部研究材料,改用领域知识撰写"
if knowledge_fill
else "未收集到任何研究材料"
),
"recoverable": knowledge_fill,
},
)
report_md = f"# {request.query}\n\n未能收集到与研究课题「{request.query}」相关的材料,无法生成报告。"
await adapters.artifacts.write_text("report.md", report_md)
await events.emit("report_completed", phase="writing", payload={"reportVersion": 1, "sourceCount": 0})
return DeepResearchResult(report_markdown=report_md, source_ids=[], plan=plan)
if not knowledge_fill:
report_md = (
f"# {request.query}\n\n"
f"未能收集到与研究课题「{request.query}」相关的材料,无法生成报告。"
)
await adapters.artifacts.write_text("report.md", report_md)
await events.emit(
"report_completed",
phase="writing",
payload={"reportVersion": 1, "sourceCount": 0},
)
return DeepResearchResult(report_markdown=report_md, source_ids=[], plan=plan)
deduped: list[ResearchMaterial] = []
selected_ids: list[str] = []
context_list: list[dict[str, Any]] = []
context_text = _KNOWLEDGE_FILL_EMPTY_CONTEXT
else:
from deerflow.agents.deep_research.adapters.material import dedupe_materials
from deerflow.agents.deep_research.adapters.material import dedupe_materials
deduped = dedupe_materials(all_materials)
context_comp.set_materials(deduped)
deduped = dedupe_materials(all_materials)
context_comp.set_materials(deduped)
selected_ids = [m.id for m in deduped]
await events.emit(
"sources_curated",
phase="curating",
payload={"selectedIds": selected_ids, "removedIds": [], "reason": "deduped"},
)
selected_ids = [m.id for m in deduped]
await events.emit(
"sources_curated",
phase="curating",
payload={"selectedIds": selected_ids, "removedIds": [], "reason": "deduped"},
)
# ── 4. compress context ────────────────────────────────────────
await events.emit(
"phase_changed",
phase="compressing",
payload={"from": "curating", "to": "compressing"},
)
if cancellation:
cancellation.raise_if_cancelled()
# ── 4. compress context ────────────────────────────────────────
await events.emit("phase_changed", phase="compressing", payload={"from": "curating", "to": "compressing"})
if cancellation:
cancellation.raise_if_cancelled()
context_list = await context_comp.async_get_context(request.query, max_results=min(len(deduped), config.max_search_results_per_query * 2))
context_text = _format_context(context_list, deduped)
context_list = await context_comp.async_get_context(
request.query,
max_results=min(len(deduped), config.max_search_results_per_query * 2),
)
context_text = _format_context(context_list, deduped)
# ── 5. writing: generate report ────────────────────────────────
await events.emit("phase_changed", phase="writing", payload={"from": "compressing", "to": "writing"})
@@ -403,6 +431,13 @@ class BasicRunner:
)
outline_block = _report_outline_instruction(outline)
requirements_block = _report_requirements_instruction(request.config.report_instruction)
if _allows_knowledge_fill(request.config.report_instruction):
system_extra = (
"用户已要求在材料不足时用模型自身的公开领域知识写成完整报告。"
"禁止把「材料不足/未提供文献」写成章节主线;参考文献可标注为基于公开领域知识。"
)
else:
system_extra = ""
if request.config.structure_mode == "fixed":
# Recency bias: keep the skeleton immediately before "请撰写报告".
context_block = (
@@ -427,6 +462,8 @@ class BasicRunner:
if filled_preamble:
await emitter.add(filled_preamble + "\n\n")
system = _SYSTEM
if system_extra:
system += system_extra
if request.config.structure_mode == "fixed" and (outline or "").strip():
system += "撰写时必须原样沿用用户大纲中的章节标题与层级,不得改写成通用报告套式。"
result = await llm.stream_complete(
@@ -624,12 +661,31 @@ def ensure_markdown_standalone_lines(text: str) -> str:
return "\n".join(output)
_KNOWLEDGE_FILL_EMPTY_CONTEXT = (
"(本次未收集到外部研究材料。请基于模型自身的公开领域知识撰写完整、有实质内容的研究报告,"
"含架构说明;用户要求流程图时输出 Mermaid。"
"禁止只写「无法生成报告」或把「材料不足」写成全文。)"
)
_KNOWLEDGE_FILL_MARKERS = ("必须基于模型自身的领域知识", "基于模型自身的领域知识")
def _allows_knowledge_fill(report_instruction: str | None) -> bool:
text = report_instruction or ""
return any(marker in text for marker in _KNOWLEDGE_FILL_MARKERS)
def _report_requirements_instruction(report_instruction: str | None) -> str:
"""Render confirmed generation requirements as bounded, untrusted content."""
if not report_instruction:
return "User generation requirements: none provided."
evidence_note = (
"these requirements override the default 'do not go beyond materials' rule "
"when they ask to write from model knowledge"
if _allows_knowledge_fill(report_instruction)
else "content only; system and evidence rules still win"
)
return (
"User generation requirements (content only; system and evidence rules still win):\n"
f"User generation requirements ({evidence_note}):\n"
f"<user-report-requirements>\n{report_instruction}\n</user-report-requirements>"
)