diff --git a/.gitignore b/.gitignore
index 661b8f0..3fc8ecb 100644
--- a/.gitignore
+++ b/.gitignore
@@ -36,3 +36,9 @@ tests/extension_cases_18/
 *.bak-*
 bioflow_service/bioflow.db
 *.before-*
+
+# paper_writer_service 运行时产物（不入库）
+paper_writer_service/paper.db
+paper_writer_service/paper.db-shm
+paper_writer_service/paper.db-wal
+paper_writer_service/paper.log
diff --git a/deploy/paper_restart.sh b/deploy/paper_restart.sh
new file mode 100755
index 0000000..23737b3
--- /dev/null
+++ b/deploy/paper_restart.sh
@@ -0,0 +1,10 @@
+#!/bin/bash
+# deploy/paper_restart.sh —— 论文撰写服务重启（8097，dev 树明文；不触碰 8090/8091/8098）。
+# 用法：bash deploy/paper_restart.sh
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
+
+bash "$REPO_ROOT/paper_writer_service/restart.sh"
+echo "paper_writer_service 重启已触发（端口 8097，日志 paper_writer_service/paper.log）"
diff --git a/deploy/toggle_paper.sh b/deploy/toggle_paper.sh
new file mode 100755
index 0000000..ba0e96b
--- /dev/null
+++ b/deploy/toggle_paper.sh
@@ -0,0 +1,51 @@
+#!/bin/bash
+# deploy/toggle_paper.sh —— 论文撰写快速开关（dev 树 175；不触碰生产 176 /srv/plantai）。
+# 用法：bash deploy/toggle_paper.sh off|on
+# 原理：BFF main.py 按 PAPER_WRITER_ENABLED 决定是否注册 /api/paper 路由；
+#       不注册时 /api/paper/health 返回 404，前端探活落 offline，触发词落 docgen 单发降级。
+set -uo pipefail
+
+ACTION="${1:-}"
+SDK_DIR="/home/xzq/5090/teamWorkspace/zh/PlantAI-rag-new/sdk"
+ENV_FILE="$SDK_DIR/.env"
+PY="/home/xzq/5090/teamWorkspace/zh/plantai-rag-new-venv/bin/python"
+BIND="${PAPER_BFF_BIND:-0.0.0.0}"
+
+if [ "$ACTION" != "off" ] && [ "$ACTION" != "on" ]; then
+  echo "用法: $0 <off|on>" >&2
+  exit 2
+fi
+
+VAL="true"
+[ "$ACTION" = "off" ] && VAL="false"
+
+# 1. 改 env（追加或覆盖）
+if grep -q "^PAPER_WRITER_ENABLED=" "$ENV_FILE"; then
+  sed -i "s/^PAPER_WRITER_ENABLED=.*/PAPER_WRITER_ENABLED=$VAL/" "$ENV_FILE"
+else
+  echo "PAPER_WRITER_ENABLED=$VAL" >> "$ENV_FILE"
+fi
+echo "PAPER_WRITER_ENABLED -> $VAL"
+
+# 2. 重启 BFF 8090（dev 树；仅取 LISTEN 者；venv 必须绝对路径，cd sdk 后才能 import app.main）
+kill $(lsof -ti tcp:8090 -sTCP:LISTEN) 2>/dev/null || true
+sleep 2
+cd "$SDK_DIR" && set -a && . .env && set +a
+( cd "$SDK_DIR" && setsid "$PY" -m uvicorn app.main:app --host "$BIND" --port 8090 \
+  > /tmp/plantai-bff.log 2>&1 < /dev/null & ) && disown -a
+
+# 3. 自检
+for i in $(seq 1 20); do
+  sleep 2
+  curl -s -m 3 http://127.0.0.1:8090/api/health >/dev/null 2>&1 && break
+done
+curl -s -m 10 http://127.0.0.1:8090/api/health | head -c 120
+echo
+if [ "$ACTION" = "off" ]; then
+  CODE=$(curl -s -o /dev/null -w "%{http_code}" -m 8 http://127.0.0.1:8090/api/paper/health)
+  echo "paper/health HTTP $CODE (off 期望 404)"
+  [ "$CODE" = "404" ] && echo "OFF_OK" || echo "OFF_CHECK_FAIL"
+else
+  curl -s -m 8 http://127.0.0.1:8090/api/paper/health | head -c 150
+  echo
+fi
diff --git a/paper_writer_service/__init__.py b/paper_writer_service/__init__.py
new file mode 100644
index 0000000..bafbf10
--- /dev/null
+++ b/paper_writer_service/__init__.py
@@ -0,0 +1 @@
+"""paper_writer_service —— 论文撰写独立 MCP 服务（8097）。"""
diff --git a/paper_writer_service/app/__init__.py b/paper_writer_service/app/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/paper_writer_service/app/config.py b/paper_writer_service/app/config.py
new file mode 100644
index 0000000..2f3fe8e
--- /dev/null
+++ b/paper_writer_service/app/config.py
@@ -0,0 +1,53 @@
+"""config.py —— paper 服务（8097）环境变量驱动配置（pydantic-settings）。
+
+全部键可用环境变量覆盖（前缀 PAPER_；LLM 沿用 docgen 的 DOCGEN_LLM_* 直连模式，
+由 services/paper/llm.py 直接读 os.environ）。零硬编码宿主路径：相对路径按 __file__ 推导。
+"""
+from __future__ import annotations
+
+from functools import lru_cache
+from pathlib import Path
+
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+_SERVICE_ROOT = Path(__file__).resolve().parent.parent  # paper_writer_service/
+
+
+def _resolve(v: str) -> str:
+    """仓库内相对路径按 __file__ 推导服务根解析为绝对路径。"""
+    p = Path(v).expanduser()
+    if not p.is_absolute():
+        p = _SERVICE_ROOT / p
+    return str(p)
+
+
+class Settings(BaseSettings):
+    model_config = SettingsConfigDict(
+        env_file=".env",
+        env_file_encoding="utf-8",
+        extra="ignore",
+        env_prefix="PAPER_",
+    )
+
+    app_name: str = "plantai-paper-writer-service"
+    version: str = "0.1.0"
+
+    db_path: str = "paper.db"                  # 独立 SQLite（与 BFF 完全隔离）
+    vstore_base_url: str = "http://127.0.0.1:8094"
+    read_timeout: float = 360.0                # write_paper_section 同步长调用读超时
+
+    @property
+    def db_path_abs(self) -> str:
+        return _resolve(self.db_path)
+
+    @property
+    def vstore_base(self) -> str:
+        return self.vstore_base_url.rstrip("/")
+
+
+@lru_cache(maxsize=1)
+def get_settings() -> Settings:
+    return Settings()
+
+
+settings = get_settings()
diff --git a/paper_writer_service/app/db.py b/paper_writer_service/app/db.py
new file mode 100644
index 0000000..3c00d53
--- /dev/null
+++ b/paper_writer_service/app/db.py
@@ -0,0 +1,168 @@
+"""db.py —— paper 服务独立 SQLite（aiosqlite + WAL，无 ORM），三张表。
+
+- papers：论文主表（brief/outline/final_markdown/references/disclaimer）
+- paper_sections：章节表（逐节 markdown + 状态）
+- citations：结构化引用（temp_n 生成期全局递增，final_n complete 按出现序重排写回）
+
+模式沿用 bioflow/docgen 同款：CREATE TABLE IF NOT EXISTS + aiosqlite 单连接 + WAL + 写锁。
+"""
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import time
+from collections.abc import AsyncIterator, Iterable
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any
+
+import aiosqlite
+
+from .config import get_settings
+
+logger = logging.getLogger(__name__)
+
+_SCHEMA = """
+CREATE TABLE IF NOT EXISTS papers (
+    id              TEXT PRIMARY KEY,
+    session_id      TEXT NOT NULL DEFAULT '',
+    title           TEXT NOT NULL DEFAULT '',
+    paper_type      TEXT NOT NULL DEFAULT '',
+    language        TEXT NOT NULL DEFAULT '',
+    format_req      TEXT NOT NULL DEFAULT '',
+    style           TEXT NOT NULL DEFAULT '',
+    status          TEXT NOT NULL DEFAULT 'outline',
+    brief_json      TEXT NOT NULL DEFAULT '{}',
+    outline_json    TEXT NOT NULL DEFAULT '{}',
+    final_markdown  TEXT NOT NULL DEFAULT '',
+    references_json TEXT NOT NULL DEFAULT '[]',
+    disclaimer      TEXT NOT NULL DEFAULT '',
+    created_at      TEXT NOT NULL,
+    updated_at      TEXT NOT NULL
+);
+CREATE INDEX IF NOT EXISTS idx_papers_session ON papers(session_id);
+CREATE INDEX IF NOT EXISTS idx_papers_status ON papers(status);
+
+CREATE TABLE IF NOT EXISTS paper_sections (
+    id          TEXT PRIMARY KEY,
+    paper_id    TEXT NOT NULL REFERENCES papers(id) ON DELETE CASCADE,
+    section_id  TEXT NOT NULL,
+    title       TEXT NOT NULL DEFAULT '',
+    key_points  TEXT NOT NULL DEFAULT '[]',
+    status      TEXT NOT NULL DEFAULT 'pending',
+    markdown    TEXT NOT NULL DEFAULT '',
+    word_count  INTEGER NOT NULL DEFAULT 0,
+    updated_at  TEXT NOT NULL,
+    UNIQUE(paper_id, section_id)
+);
+CREATE INDEX IF NOT EXISTS idx_paper_sections_paper ON paper_sections(paper_id);
+
+CREATE TABLE IF NOT EXISTS citations (
+    id             INTEGER PRIMARY KEY AUTOINCREMENT,
+    paper_id       TEXT NOT NULL REFERENCES papers(id) ON DELETE CASCADE,
+    section_id     TEXT NOT NULL DEFAULT '',
+    temp_n         INTEGER,
+    final_n        INTEGER,
+    kind           TEXT NOT NULL DEFAULT 'title',
+    ref_text       TEXT NOT NULL DEFAULT '',
+    source_doc_id  TEXT NOT NULL DEFAULT '',
+    snippet        TEXT NOT NULL DEFAULT '',
+    created_at     TEXT NOT NULL
+);
+CREATE INDEX IF NOT EXISTS idx_citations_paper ON citations(paper_id);
+"""
+
+
+def now_iso() -> str:
+    return datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
+
+
+def dumps_json(value: Any, default: Any = None) -> str:
+    try:
+        return json.dumps(value, ensure_ascii=False)
+    except (TypeError, ValueError):
+        return json.dumps(default, ensure_ascii=False)
+
+
+def loads_json(text: str | None, default: Any = None) -> Any:
+    if text is None:
+        return default
+    try:
+        return json.loads(text)
+    except (TypeError, ValueError):
+        return default
+
+
+class Database:
+    """aiosqlite 单连接封装（WAL；写串行化，同 bioflow 模式）。"""
+
+    def __init__(self, path: str | Path | None = None) -> None:
+        self._path: Path | None = Path(path) if path is not None else None
+        self._conn: aiosqlite.Connection | None = None
+        self._lock = asyncio.Lock()
+
+    @property
+    def path(self) -> Path:
+        if self._path is None:
+            self._path = Path(get_settings().db_path_abs)
+        return self._path
+
+    async def connect(self) -> None:
+        self.path.parent.mkdir(parents=True, exist_ok=True)
+        self._conn = await aiosqlite.connect(self.path)
+        self._conn.row_factory = aiosqlite.Row
+        await self._conn.execute("PRAGMA journal_mode=WAL")
+        await self._conn.execute("PRAGMA foreign_keys=ON")
+        await self._conn.execute("PRAGMA busy_timeout=5000")
+        await self._conn.executescript(_SCHEMA)
+        await self._conn.commit()
+
+    async def close(self) -> None:
+        if self._conn is not None:
+            await self._conn.close()
+            self._conn = None
+
+    @property
+    def conn(self) -> aiosqlite.Connection:
+        if self._conn is None:
+            raise RuntimeError("Database 未初始化：先 await init_db()")
+        return self._conn
+
+    async def fetchall(self, sql: str, params: Iterable[Any] = ()) -> list[aiosqlite.Row]:
+        async with self._conn.execute(sql, tuple(params)) as cursor:
+            return list(await cursor.fetchall())
+
+    async def fetchone(self, sql: str, params: Iterable[Any] = ()) -> aiosqlite.Row | None:
+        async with self._conn.execute(sql, tuple(params)) as cursor:
+            return await cursor.fetchone()
+
+    async def execute(self, sql: str, params: Iterable[Any] = ()) -> int:
+        async with self._lock:
+            cur = await self._conn.execute(sql, tuple(params))
+            await self._conn.commit()
+            return cur.lastrowid
+
+
+db = Database()
+
+
+async def init_db(path: str | Path | None = None) -> Database:
+    # 原地改单例路径后连接（不做 db = Database(path) 重赋值）：
+    # 服务模块以 `from ...db import db` 持有了该单例引用，重赋值会让它们拿到旧对象。
+    if path is not None:
+        db._path = Path(path)
+    await db.connect()
+    return db
+
+
+async def get_db() -> AsyncIterator[Database]:
+    yield db
+
+
+async def close_db() -> None:
+    await db.close()
+
+
+def monotonic() -> float:
+    return time.monotonic()
diff --git a/paper_writer_service/app/main.py b/paper_writer_service/app/main.py
new file mode 100644
index 0000000..efc2bae
--- /dev/null
+++ b/paper_writer_service/app/main.py
@@ -0,0 +1,84 @@
+"""main.py —— paper 服务应用入口（uvicorn 入口，明文白名单文件）。
+
+- 独立部署树 paper_writer_service/（FR：独立进程 8097 + 独立 SQLite paper.db）。
+- lifespan：init_db（幂等建表）→ 显式进入 MCP session_manager.run()（Starlette 不自动
+  运行子 Mount 的 lifespan，streamable-http 处理请求时要求 task group 已初始化）→ 清理。
+- 端点：/health（GET，BFF 探活反代）+ /mcp（MCP streamable-http）。
+- 不依赖 BFF/生产服务（8090 等绝不探测/连接）。
+"""
+from __future__ import annotations
+
+import logging
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI
+from fastapi.middleware.cors import CORSMiddleware
+
+from .config import get_settings
+from .db import close_db, db, init_db
+from .mcp import paper_mcp
+from .services.paper import search
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+# MCP streamable-http app（调用即创建 session_manager，供 lifespan 进入 run()）。
+# 端点地址：/mcp（streamable_http_path="/" + 挂载前缀 /mcp）。
+_paper_mcp_app = paper_mcp.streamable_http_app(streamable_http_path="/")
+
+
+def create_app() -> FastAPI:
+    settings = get_settings()
+    app = FastAPI(title=settings.app_name, version=settings.version)
+
+    app.add_middleware(
+        CORSMiddleware,
+        allow_origins=["*"],
+        allow_credentials=False,
+        allow_methods=["*"],
+        allow_headers=["*"],
+    )
+
+    @asynccontextmanager
+    async def _lifespan(_app: FastAPI):
+        # 测试注入：db 已由测试预置（init_db(tmp_path)），跳过启动初始化。
+        try:
+            await init_db()
+        except Exception:
+            logger.exception("main: 数据库初始化失败（health 将报 db 降级）")
+
+        # MCP session manager 的 task group 必须在服务运行期内保持活跃。
+        async with paper_mcp.session_manager.run():
+            try:
+                yield
+            finally:
+                try:
+                    await close_db()
+                except Exception:
+                    pass
+                await search.close_client()
+
+    app.router.lifespan_context = _lifespan
+
+    @app.get("/health")
+    async def health() -> dict[str, object]:
+        db_ok = True
+        db_error: str | None = None
+        try:
+            await db.fetchone("SELECT 1 AS ok")
+        except Exception as exc:  # noqa: BLE001
+            db_ok = False
+            db_error = str(exc)
+        return {
+            "ok": True,
+            "status": "ok" if db_ok else "degraded",
+            "db": {"connected": db_ok, "error": db_error},
+            "version": settings.version,
+        }
+
+    # paper MCP 能力外挂（mcp 2.x：MCPServer.streamable_http_app）。
+    app.mount("/mcp", _paper_mcp_app)
+    return app
+
+
+app = create_app()
diff --git a/paper_writer_service/app/mcp.py b/paper_writer_service/app/mcp.py
new file mode 100644
index 0000000..a754fe9
--- /dev/null
+++ b/paper_writer_service/app/mcp.py
@@ -0,0 +1,48 @@
+"""app/mcp.py —— paper_writer_service MCP server（单 streamable-http 端点 /mcp）。
+
+9 个工具：propose_paper_outline / write_paper_section / get_paper_progress / search_refs /
+revise_paper_section / revise_paper / get_paper / complete_paper / cancel_paper。
+
+mcp 2.x 注意：MCPServer（非 FastMCP，来自 mcp.server.mcpserver）；工具返回 dict，由 MCP
+序列化为 CallToolResult（structuredContent + content:[{type:"text",text:"<json>"}]）。
+本文件位于 app/ 层（非 services/），可用 pydantic（mcp 自身依赖 pydantic 构造 schema）。
+"""
+from __future__ import annotations
+
+from typing import Any
+
+from .services.paper import complete, outline, search, write
+
+
+def build_paper_mcp() -> Any:
+    from mcp.server.mcpserver import MCPServer
+
+    mcp = MCPServer(
+        "paper-writer",
+        title="PlantAI 论文撰写",
+        description="论文撰写能力单元：大纲生成/逐节撰写/文献引用/全文合并重排/修订。",
+    )
+
+    mcp.tool(name="propose_paper_outline",
+             description="根据论文主题与写作要求生成大纲并建论文（内部检索知识库注入选题线索）。")(outline.propose_paper_outline)
+    mcp.tool(name="write_paper_section",
+             description="撰写论文某一章节（同步长调用，90-180s）。按全局引用编号表生成正文与引用。")(write.write_paper_section)
+    mcp.tool(name="get_paper_progress",
+             description="查询论文与各章节撰写进度/字数/引用数。")(complete.get_paper_progress)
+    mcp.tool(name="search_refs",
+             description="在知识库(vstore)中检索与查询相关的文献片段。")(search.search_refs)
+    mcp.tool(name="revise_paper_section",
+             description="按用户意见重写论文某一章节，新引用追加到引用表末尾。")(write.revise_paper_section)
+    mcp.tool(name="revise_paper",
+             description="按用户意见判定需要修改哪些章节，返回受影响节清单与修订计划。")(write.revise_paper)
+    mcp.tool(name="get_paper",
+             description="拉取论文全文 markdown、参考文献列表与免责声明。")(complete.get_paper)
+    mcp.tool(name="complete_paper",
+             description="合并全文、按首次出现顺序重排引用编号、渲染 GB/T 7714 文献列表。")(complete.complete_paper)
+    mcp.tool(name="cancel_paper",
+             description="取消论文（置 cancelled）。")(complete.cancel_paper)
+
+    return mcp
+
+
+paper_mcp = build_paper_mcp()
diff --git a/paper_writer_service/app/schemas.py b/paper_writer_service/app/schemas.py
new file mode 100644
index 0000000..8f3ca22
--- /dev/null
+++ b/paper_writer_service/app/schemas.py
@@ -0,0 +1,10 @@
+"""schemas.py —— paper 服务共享常量（明文白名单文件）。
+
+MCP 工具入参/出参均为 dict（由 mcp 层序列化为 content:[{type:text,text:<json>}]），
+本文件只放状态/引用类型常量，避免服务层散落魔法字面量。
+"""
+from __future__ import annotations
+
+PAPER_STATUSES = ("outline", "writing", "complete", "cancelled", "error")
+SECTION_STATUSES = ("pending", "writing", "done", "error")
+CITATION_KINDS = ("doi", "pmid", "title")
diff --git a/paper_writer_service/app/services/__init__.py b/paper_writer_service/app/services/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/paper_writer_service/app/services/paper/__init__.py b/paper_writer_service/app/services/paper/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/paper_writer_service/app/services/paper/cite.py b/paper_writer_service/app/services/paper/cite.py
new file mode 100644
index 0000000..ec205c5
--- /dev/null
+++ b/paper_writer_service/app/services/paper/cite.py
@@ -0,0 +1,65 @@
+"""cite.py —— 引用协议：GB/T 7714 渲染、[n] 校验、complete 按出现序重排。"""
+from __future__ import annotations
+
+import re
+from typing import Any
+
+_CITE_RE = re.compile(r"\[(\d+)\]")
+
+
+def find_cite_numbers(markdown: str | None) -> list[int]:
+    """正文中出现的所有 [n] 编号（含重复，按出现顺序）。"""
+    return [int(m) for m in _CITE_RE.findall(markdown or "")]
+
+
+def render_gbt7714(meta: dict[str, Any], *, fallback_year: str = "2026") -> str:
+    """按可用元数据渲染 GB/T 7714 近似条目。
+
+    - 有 DOI → [J/OL] + DOI
+    - 有 PMID → [J] + PMID
+    - 无 DOI/PMID（知识库资料）→ 降级格式 + [EB/OL]
+    """
+    title = (meta.get("title") or "").strip()
+    journal = (meta.get("journal") or "").strip()
+    authors = (meta.get("authors") or "").strip()
+    year = (meta.get("year") or fallback_year).strip()
+    doi = (meta.get("doi") or "").strip()
+    pmid = (meta.get("pmid") or "").strip()
+    head = f"{authors} " if authors else ""
+    if doi:
+        jpart = f"{journal}, {year}." if journal else f"{year}."
+        return f"{head}{title}[J/OL]. {jpart} DOI:{doi}."
+    if pmid:
+        jpart = f"{journal}, {year}." if journal else f"{year}."
+        return f"{head}{title}[J]. {jpart} PMID:{pmid}."
+    return f"{head}{title}. 知识库资料, {year}. [EB/OL]."
+
+
+def render_disclaimer() -> str:
+    return "参考文献依据知识库资料生成，格式为 GB/T 7714 近似，建议人工核对。"
+
+
+def first_appearance_order(full_markdown: str, temp_numbers: list[int]) -> list[int]:
+    """按全文首次出现顺序返回 temp_n；未出现的排最后（保持原序）。"""
+    seen: dict[int, int] = {}
+    for m in _CITE_RE.finditer(full_markdown or ""):
+        n = int(m.group(1))
+        if n not in seen and n in temp_numbers:
+            seen[n] = m.start()
+    ordered = sorted((n for n in temp_numbers if n in seen), key=lambda n: seen[n])
+    ordered += [n for n in temp_numbers if n not in seen]
+    return ordered
+
+
+def renumber_map(temp_numbers: list[int], full_markdown: str) -> dict[int, int]:
+    """temp_n → final_n（按全文首次出现顺序 1..N）。"""
+    ordered = first_appearance_order(full_markdown, temp_numbers)
+    return {temp: final for final, temp in enumerate(ordered, start=1)}
+
+
+def replace_numbers(markdown: str | None, mapping: dict[int, int]) -> str:
+    def _sub(m: re.Match[str]) -> str:
+        n = int(m.group(1))
+        return f"[{mapping.get(n, n)}]"
+
+    return _CITE_RE.sub(_sub, markdown or "")
diff --git a/paper_writer_service/app/services/paper/complete.py b/paper_writer_service/app/services/paper/complete.py
new file mode 100644
index 0000000..f88c24d
--- /dev/null
+++ b/paper_writer_service/app/services/paper/complete.py
@@ -0,0 +1,106 @@
+"""complete.py —— get_paper / get_paper_progress / complete_paper / cancel_paper。"""
+from __future__ import annotations
+
+from typing import Any
+
+from ...db import db, dumps_json, loads_json, now_iso
+from .cite import render_disclaimer, renumber_map, replace_numbers
+
+
+async def _load(paper_id: str) -> dict[str, Any]:
+    paper = await db.fetchone("SELECT * FROM papers WHERE id = ?", (paper_id,))
+    if not paper:
+        return {"ok": False, "code": "NOT_FOUND", "message": "论文不存在"}
+    sections = [dict(r) for r in await db.fetchall(
+        "SELECT * FROM paper_sections WHERE paper_id = ? ORDER BY rowid", (paper_id,))]
+    citations = [dict(r) for r in await db.fetchall(
+        "SELECT * FROM citations WHERE paper_id = ? ORDER BY temp_n", (paper_id,))]
+    return {"ok": True, "paper": dict(paper), "sections": sections, "citations": citations}
+
+
+def _merge_sections(paper: dict[str, Any], sections: list[dict[str, Any]]) -> str:
+    parts = [f"# {paper.get('title', '')}"]
+    for s in sections:
+        parts.append(f"\n## {s['title']}\n\n{s.get('markdown', '')}")
+    return "\n".join(parts)
+
+
+async def get_paper(paper_id: str, format: str = "md") -> dict[str, Any]:
+    p = await _load(paper_id)
+    if not p.get("ok"):
+        return p
+    paper, sections, citations = p["paper"], p["sections"], p["citations"]
+    full_markdown = paper.get("final_markdown") or _merge_sections(paper, sections)
+    references = loads_json(paper.get("references_json"), []) or [
+        {"n": c.get("final_n") or c.get("temp_n"), "text": c["ref_text"], "kind": c["kind"],
+         "source_doc_id": c["source_doc_id"]} for c in citations
+    ]
+    return {
+        "ok": True, "paper_id": paper_id, "title": paper["title"], "status": paper["status"],
+        "sections": [
+            {"section_id": s["section_id"], "title": s["title"], "status": s["status"],
+             "word_count": s["word_count"]} for s in sections
+        ],
+        "full_markdown": full_markdown, "references": references,
+        "disclaimer": paper.get("disclaimer", ""),
+    }
+
+
+async def get_paper_progress(paper_id: str) -> dict[str, Any]:
+    paper = await db.fetchone("SELECT * FROM papers WHERE id = ?", (paper_id,))
+    if not paper:
+        return {"ok": False, "code": "NOT_FOUND", "message": "论文不存在"}
+    paper = dict(paper)
+    sections = [dict(r) for r in await db.fetchall(
+        "SELECT * FROM paper_sections WHERE paper_id = ? ORDER BY rowid", (paper_id,))]
+    cit_rows = [dict(r) for r in await db.fetchall(
+        "SELECT section_id, COUNT(*) AS c FROM citations WHERE paper_id = ? GROUP BY section_id", (paper_id,))]
+    cit_by_sec = {r["section_id"]: r["c"] for r in cit_rows}
+    current = next((s["section_id"] for s in sections if s["status"] != "done"), None)
+    return {
+        "ok": True, "paper_id": paper_id, "status": paper["status"], "current_section_id": current,
+        "sections": [
+            {"section_id": s["section_id"], "title": s["title"], "status": s["status"],
+             "word_count": s["word_count"], "citation_count": cit_by_sec.get(s["section_id"], 0)}
+            for s in sections
+        ],
+    }
+
+
+async def complete_paper(paper_id: str) -> dict[str, Any]:
+    p = await _load(paper_id)
+    if not p.get("ok"):
+        return p
+    paper, sections, citations = p["paper"], p["sections"], p["citations"]
+    full = _merge_sections(paper, sections)
+    temp_numbers = [c.get("temp_n") for c in citations if c.get("temp_n")]
+    mapping = renumber_map(temp_numbers, full)
+
+    final_full = replace_numbers(full, mapping)
+    for c in citations:
+        await db.execute("UPDATE citations SET final_n = ? WHERE id = ?", (mapping.get(c.get("temp_n")), c["id"]))
+    for s in sections:
+        await db.execute("UPDATE paper_sections SET markdown = ? WHERE id = ?",
+                         (replace_numbers(s.get("markdown"), mapping), s["id"]))
+
+    ordered = sorted(citations, key=lambda c: mapping.get(c.get("temp_n"), 0))
+    references = [
+        {"n": mapping.get(c.get("temp_n"), c.get("temp_n")), "text": c["ref_text"],
+         "kind": c["kind"], "source_doc_id": c["source_doc_id"]} for c in ordered
+    ]
+    has_degraded = any(c["kind"] == "title" for c in citations)
+    disclaimer = render_disclaimer() if has_degraded else ""
+    await db.execute(
+        "UPDATE papers SET status = 'complete', final_markdown = ?, references_json = ?, disclaimer = ?, updated_at = ? WHERE id = ?",
+        (final_full, dumps_json(references), disclaimer, now_iso(), paper_id),
+    )
+    return {"ok": True, "full_markdown": final_full, "references": references,
+            "renumber_map": mapping, "disclaimer": disclaimer}
+
+
+async def cancel_paper(paper_id: str) -> dict[str, Any]:
+    paper = await db.fetchone("SELECT id FROM papers WHERE id = ?", (paper_id,))
+    if not paper:
+        return {"ok": False, "code": "NOT_FOUND", "message": "论文不存在"}
+    await db.execute("UPDATE papers SET status = 'cancelled', updated_at = ? WHERE id = ?", (now_iso(), paper_id))
+    return {"ok": True}
diff --git a/paper_writer_service/app/services/paper/llm.py b/paper_writer_service/app/services/paper/llm.py
new file mode 100644
index 0000000..b5d75c6
--- /dev/null
+++ b/paper_writer_service/app/services/paper/llm.py
@@ -0,0 +1,187 @@
+"""services/paper/llm.py —— LLM 调用（httpx 裸调，拷贝自 sdk/app/services/docgen/llm.py）。
+
+- complete_text(system, user, *, temperature=0.3, max_tokens=8000) -> str
+- complete_json(...)（response_format json_object + 一次 JSON 修复重试）
+- 4 次尝试（1.5s / 3s / 6s 指数退避，429 / 5xx / 超时）
+- env 键：DOCGEN_LLM_BASE_URL / DOCGEN_LLM_API_KEY / DOCGEN_LLM_MODEL / DOCGEN_LLM_TIMEOUT
+"""
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import os
+from typing import Any
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_LLM_BASE_URL = "https://api.deepseek.com/v1"
+DEFAULT_LLM_MODEL = "deepseek-chat"
+DEFAULT_LLM_TIMEOUT = 300.0
+BACKOFF_SECONDS = (1.5, 3.0, 6.0)
+MAX_ATTEMPTS = 4
+
+
+class LLMError(Exception):
+    """LLM 调用失败（重试尽/响应异常）。"""
+
+
+def _llm_config() -> dict[str, Any]:
+    return {
+        "base_url": (os.environ.get("DOCGEN_LLM_BASE_URL") or DEFAULT_LLM_BASE_URL).rstrip("/"),
+        "api_key": os.environ.get("DOCGEN_LLM_API_KEY", ""),
+        "model": os.environ.get("DOCGEN_LLM_MODEL") or DEFAULT_LLM_MODEL,
+        "timeout": float(os.environ.get("DOCGEN_LLM_TIMEOUT") or DEFAULT_LLM_TIMEOUT),
+    }
+
+
+def _extract_text(payload: dict[str, Any]) -> str:
+    try:
+        choices = payload.get("choices") or []
+        if choices:
+            content = choices[0].get("message", {}).get("content")
+            if isinstance(content, str):
+                return content
+            if isinstance(content, list):
+                return "".join(
+                    str(c.get("text", "")) if isinstance(c, dict) else str(c) for c in content
+                )
+    except Exception:
+        pass
+    return ""
+
+
+async def _chat_once(
+    client: httpx.AsyncClient,
+    cfg: dict[str, Any],
+    system: str,
+    user: str,
+    *,
+    temperature: float,
+    max_tokens: int,
+    json_mode: bool,
+    timeout: float | None = None,
+) -> str:
+    headers = {"Content-Type": "application/json"}
+    if cfg["api_key"]:
+        headers["Authorization"] = f"Bearer {cfg['api_key']}"
+    body: dict[str, Any] = {
+        "model": cfg["model"],
+        "messages": [
+            {"role": "system", "content": system},
+            {"role": "user", "content": user},
+        ],
+        "temperature": temperature,
+        "max_tokens": max_tokens,
+    }
+    if json_mode:
+        body["response_format"] = {"type": "json_object"}
+    resp = await client.post(
+        f"{cfg['base_url']}/chat/completions",
+        headers=headers,
+        json=body,
+        timeout=timeout if timeout is not None else cfg["timeout"],
+    )
+    resp.raise_for_status()
+    payload = resp.json()
+    if not isinstance(payload, dict):
+        raise LLMError("LLM 响应形态异常（非 JSON 对象）")
+    text = _extract_text(payload)
+    if not text:
+        raise LLMError("LLM 响应缺少 choices[0].message.content")
+    return text
+
+
+async def _complete(
+    system: str,
+    user: str,
+    *,
+    temperature: float,
+    max_tokens: int,
+    json_mode: bool,
+    timeout: float | None = None,
+) -> str:
+    cfg = _llm_config()
+    last_exc: BaseException | None = None
+    async with httpx.AsyncClient() as client:
+        for attempt in range(MAX_ATTEMPTS):
+            try:
+                return await _chat_once(
+                    client, cfg, system, user,
+                    temperature=temperature, max_tokens=max_tokens, json_mode=json_mode,
+                    timeout=timeout,
+                )
+            except LLMError:
+                raise
+            except (httpx.HTTPStatusError, httpx.HTTPError, httpx.TimeoutException) as exc:
+                last_exc = exc
+                status = getattr(getattr(exc, "response", None), "status_code", None)
+                retryable = status is None or status == 429 or status >= 500
+                if not retryable or attempt == MAX_ATTEMPTS - 1:
+                    raise LLMError(f"LLM 调用失败（HTTP {status or '网络'}）：{exc}") from exc
+                await asyncio.sleep(BACKOFF_SECONDS[min(attempt, len(BACKOFF_SECONDS) - 1)])
+    raise LLMError(f"LLM 调用失败：{last_exc}")
+
+
+async def complete_text(
+    system: str,
+    user: str,
+    *,
+    temperature: float = 0.3,
+    max_tokens: int = 8000,
+    timeout: float | None = None,
+) -> str:
+    """文本补全。失败抛 LLMError。"""
+    return await _complete(
+        system, user, temperature=temperature, max_tokens=max_tokens, json_mode=False,
+        timeout=timeout,
+    )
+
+
+def _repair_json(text: str) -> str:
+    text = text.strip()
+    if text.startswith("```"):
+        text = text.strip("`")
+        if text.lower().startswith("json"):
+            text = text[4:]
+        text = text.strip()
+    if not text:
+        raise LLMError("LLM 返回空内容")
+    if text.startswith("{"):
+        return text
+    start = text.find("{")
+    end = text.rfind("}")
+    if start >= 0 and end > start:
+        return text[start:end + 1]
+    raise LLMError("LLM 输出中未找到 JSON 对象")
+
+
+async def complete_json(
+    system: str,
+    user: str,
+    *,
+    temperature: float = 0.3,
+    max_tokens: int = 8000,
+) -> dict[str, Any]:
+    """JSON 补全（json_object + 一次修复重试）。"""
+    text = await _complete(
+        system, user, temperature=temperature, max_tokens=max_tokens, json_mode=True
+    )
+    try:
+        return json.loads(text)
+    except ValueError:
+        pass
+    text2 = await _complete(
+        system + "\n（重要：你的回答必须是单个合法 JSON 对象，不要输出其他任何内容）",
+        user,
+        temperature=temperature,
+        max_tokens=max_tokens,
+        json_mode=True,
+    )
+    try:
+        return json.loads(text2)
+    except ValueError:
+        repaired = _repair_json(text2)
+        return json.loads(repaired)
diff --git a/paper_writer_service/app/services/paper/outline.py b/paper_writer_service/app/services/paper/outline.py
new file mode 100644
index 0000000..2bac884
--- /dev/null
+++ b/paper_writer_service/app/services/paper/outline.py
@@ -0,0 +1,112 @@
+"""outline.py —— propose_paper_outline：建 papers 行 → search_refs 注入线索 → LLM 出大纲 → 预填 sections。"""
+from __future__ import annotations
+
+import uuid
+from typing import Any
+
+from ...db import db, dumps_json, now_iso
+from . import llm, search
+
+_OUTLINE_SYSTEM = (
+    "你是植物学论文撰写助手。根据用户给出的论文主题与写作要求，生成论文大纲。"
+    "只输出一个合法 JSON 对象，结构为："
+    '{"title": 论文标题, "abstract_expect": 摘要预期内容(一句话), '
+    '"sections": [{"section_id": "s1", "title": "节标题", "key_points": ["要点1", "要点2"], "target_words": 1200}, ...]}。'
+    "章节 5-6 节，覆盖 引言/材料与方法/结果/讨论/结论 等综述常见结构；target_words 每节 1200-2500。"
+)
+
+_DEFAULT_SECTIONS = [
+    ("s1", "引言"),
+    ("s2", "材料与方法"),
+    ("s3", "结果"),
+    ("s4", "讨论"),
+    ("s5", "结论与展望"),
+]
+
+
+async def propose_paper_outline(
+    topic: str,
+    requirements: str = "",
+    paper_type: str = "",
+    language: str = "",
+    length: str = "",
+    format_req: str = "",
+    style: str = "",
+    kb_hints: list[dict[str, Any]] | None = None,
+) -> dict[str, Any]:
+    # 1. 知识库线索：优先用调用方传入的 kb_hints，否则现场 search_refs（失败静默降级）
+    hints = list(kb_hints or [])
+    if not hints:
+        try:
+            r = await search.search_refs(query=topic, top_k=8)
+            hints = list(r.get("items") or [])
+        except Exception:
+            hints = []
+
+    # 2. LLM 出大纲
+    user = f"论文主题：{topic}\n"
+    if paper_type:
+        user += f"论文类型：{paper_type}\n"
+    if requirements:
+        user += f"写作要求：{requirements}\n"
+    if language:
+        user += f"语言：{language}\n"
+    if length:
+        user += f"篇幅：{length}\n"
+    if format_req:
+        user += f"格式：{format_req}\n"
+    if style:
+        user += f"风格：{style}\n"
+    if hints:
+        user += "可参考知识库文献线索（仅选题依据，非硬性引用）：\n" + "\n".join(
+            f"- {h.get('title', '')}: {str(h.get('snippet', ''))[:120]}" for h in hints[:5]
+        )
+    try:
+        outline = await llm.complete_json(_OUTLINE_SYSTEM, user)
+    except llm.LLMError as exc:
+        return {"ok": False, "code": "LLM_ERROR", "message": f"大纲生成失败：{exc}"}
+
+    title = str(outline.get("title") or topic).strip() or topic
+    sections = outline.get("sections") or []
+    if not sections:
+        sections = [
+            {"section_id": sid, "title": t, "key_points": [], "target_words": 1500}
+            for sid, t in _DEFAULT_SECTIONS
+        ]
+
+    # 3. 落库 papers + paper_sections
+    paper_id = uuid.uuid4().hex[:16]
+    await db.execute(
+        "INSERT INTO papers (id, title, paper_type, language, format_req, style, status, brief_json, outline_json, created_at, updated_at) "
+        "VALUES (?, ?, ?, ?, ?, ?, 'outline', ?, ?, ?, ?)",
+        (
+            paper_id, title, paper_type, language, format_req, style,
+            dumps_json({
+                "topic": topic, "requirements": requirements, "length": length,
+                "kb_hints": hints,
+            }),
+            dumps_json({
+                "title": title,
+                "abstract_expect": outline.get("abstract_expect", ""),
+                "sections": sections,
+            }),
+            now_iso(), now_iso(),
+        ),
+    )
+    for s in sections:
+        await db.execute(
+            "INSERT INTO paper_sections (id, paper_id, section_id, title, key_points, status, updated_at) "
+            "VALUES (?, ?, ?, ?, ?, 'pending', ?)",
+            (
+                uuid.uuid4().hex[:16], paper_id,
+                s.get("section_id") or "s1", s.get("title", ""),
+                dumps_json(s.get("key_points") or []), now_iso(),
+            ),
+        )
+    return {
+        "ok": True,
+        "paper_id": paper_id,
+        "title": title,
+        "abstract_expect": outline.get("abstract_expect", ""),
+        "sections": sections,
+    }
diff --git a/paper_writer_service/app/services/paper/search.py b/paper_writer_service/app/services/paper/search.py
new file mode 100644
index 0000000..3eb7f08
--- /dev/null
+++ b/paper_writer_service/app/services/paper/search.py
@@ -0,0 +1,101 @@
+"""search.py —— vstore 8094 知识库检索客户端（search_refs 后端）。
+
+- httpx 裸调 GET /api/vstore/search-chunks
+- 失败不抛异常：返回 {"ok": False, "error": {...}}，工具层降级为空结果
+- vstore chunk 条目无 doi/pmid/journal/year（仅 doc_id/doc_name/snippet/score），
+  故 search_refs 的 meta 为尽力而为（通常为空）→ 引用多为 title kind → 降级格式 + disclaimer
+"""
+from __future__ import annotations
+
+import asyncio
+import os
+from typing import Any
+
+import httpx
+
+DEFAULT_VSTORE_BASE_URL = "http://127.0.0.1:8094"
+RETRY_BACKOFF = (0.5, 1.0, 2.0)
+MAX_RETRIES = 2
+DEFAULT_TIMEOUT = 30.0
+
+_client: httpx.AsyncClient | None = None
+
+
+def _base_url() -> str:
+    return (os.environ.get("VSTORE_BASE_URL") or DEFAULT_VSTORE_BASE_URL).rstrip("/")
+
+
+def get_client() -> httpx.AsyncClient:
+    global _client
+    if _client is None:
+        _client = httpx.AsyncClient(
+            base_url=_base_url(), timeout=httpx.Timeout(DEFAULT_TIMEOUT, connect=5.0)
+        )
+    return _client
+
+
+def set_client(client: httpx.AsyncClient) -> None:
+    """注入 client（测试用 MockTransport）。"""
+    global _client
+    _client = client
+
+
+async def close_client() -> None:
+    global _client
+    if _client is not None:
+        await _client.aclose()
+        _client = None
+
+
+async def search_chunks(
+    query: str,
+    *,
+    top_k: int = 8,
+    min_score: float = 0.30,
+    timeout: float = 12.0,
+) -> dict[str, Any]:
+    """GET /api/vstore/search-chunks。失败返回 {ok:False, error:{...}}（不抛）。"""
+    client = get_client()
+    last_exc: BaseException | None = None
+    for attempt in range(MAX_RETRIES + 1):
+        try:
+            resp = await client.get(
+                "/api/vstore/search-chunks",
+                params={"q": query, "top_k": top_k, "min_score": min_score},
+                timeout=timeout,
+            )
+            if resp.status_code < 500:
+                try:
+                    payload = resp.json()
+                except ValueError:
+                    return {"ok": False, "error": {"code": "BAD_RESPONSE", "message": "vstore 响应非 JSON"}}
+                if not isinstance(payload, dict):
+                    return {"ok": False, "error": {"code": "BAD_RESPONSE", "message": "vstore 响应形态异常"}}
+                return payload
+            last_exc = RuntimeError(f"vstore {resp.status_code}")
+        except (httpx.HTTPError, httpx.TimeoutException) as exc:
+            last_exc = exc
+        if attempt < MAX_RETRIES:
+            await asyncio.sleep(RETRY_BACKOFF[min(attempt, len(RETRY_BACKOFF) - 1)])
+    return {"ok": False, "error": {"code": "UPSTREAM_DOWN", "message": f"vstore 不可达：{last_exc}"}}
+
+
+async def search_refs(paper_id: str = "", query: str = "", top_k: int = 8) -> dict[str, Any]:
+    """search_refs 工具：透传 vstore 检索，映射为 {doc_id,title,snippet,score,meta}。"""
+    if not query:
+        return {"ok": False, "code": "BAD_REQUEST", "message": "缺少检索词 query"}
+    r = await search_chunks(query, top_k=max(1, min(int(top_k), 20)))
+    if not r.get("ok"):
+        return {"ok": True, "items": [], "warning": (r.get("error") or {}).get("message", "")}
+    items: list[dict[str, Any]] = []
+    for it in r.get("items") or []:
+        if not isinstance(it, dict):
+            continue
+        items.append({
+            "doc_id": it.get("doc_id") or "",
+            "title": it.get("doc_name") or it.get("doc_id") or "",
+            "snippet": it.get("snippet") or "",
+            "score": it.get("score"),
+            "meta": {},  # vstore chunk 无 doi/pmid/journal/year，尽力为空
+        })
+    return {"ok": True, "items": items}
diff --git a/paper_writer_service/app/services/paper/write.py b/paper_writer_service/app/services/paper/write.py
new file mode 100644
index 0000000..4787749
--- /dev/null
+++ b/paper_writer_service/app/services/paper/write.py
@@ -0,0 +1,232 @@
+"""write.py —— 章节撰写与修订（write_paper_section / revise_paper_section / revise_paper）。"""
+from __future__ import annotations
+
+from typing import Any
+
+from ...db import db, dumps_json, loads_json, now_iso
+from . import llm
+from .cite import find_cite_numbers, render_gbt7714
+
+_WRITE_SYSTEM = (
+    "你是植物学论文撰写助手，正在撰写论文的某一章节。"
+    "引用文献一律用 [n] 指代，n 必须取自下方提供的全局编号列表；如需引用新文献，编号按当前最大编号依次 +1。"
+    "禁止杜撰 DOI/PMID/作者/期刊——只能使用编号列表或知识库线索中真实给出的元数据。"
+    "只输出一个合法 JSON 对象：{\"markdown\": 章节正文(不含章标题), \"citations_used\": "
+    "[{\"temp_n\": 编号, \"title\": \"...\", \"authors\": \"...\", \"journal\": \"...\", "
+    "\"year\": \"...\", \"doi\": \"...\", \"pmid\": \"...\", \"source_doc_id\": \"...\", \"snippet\": \"...\"}]}。"
+    "citations_used 必须列出本节正文中出现的每一个 [n] 引用（既有编号也列出其已有元数据）。"
+)
+
+_REVISE_SYSTEM = (
+    "你是植物学论文撰写助手，正在按用户意见修订论文的某一章节。"
+    "保持引用协议不变：[n] 指代，n 只能取自全局编号列表或最大编号 +1；禁止杜撰元数据。"
+    "只输出一个合法 JSON 对象：{\"markdown\": 修订后的章节正文, \"citations_used\": [...]}。"
+)
+
+
+# ---------------------------------------------------------------------------
+# 读取辅助
+# ---------------------------------------------------------------------------
+
+async def _get_paper(paper_id: str) -> dict[str, Any] | None:
+    row = await db.fetchone("SELECT * FROM papers WHERE id = ?", (paper_id,))
+    return dict(row) if row else None
+
+
+async def _get_sections(paper_id: str) -> list[dict[str, Any]]:
+    rows = await db.fetchall(
+        "SELECT * FROM paper_sections WHERE paper_id = ? ORDER BY rowid", (paper_id,)
+    )
+    return [dict(r) for r in rows]
+
+
+async def _get_citations(paper_id: str) -> list[dict[str, Any]]:
+    rows = await db.fetchall(
+        "SELECT * FROM citations WHERE paper_id = ? ORDER BY temp_n", (paper_id,)
+    )
+    return [dict(r) for r in rows]
+
+
+def _global_citations_text(citations: list[dict[str, Any]]) -> str:
+    if not citations:
+        return "（当前无已编号文献；新引用从 [1] 开始）"
+    return "\n".join(f"[{c['temp_n']}] {c['ref_text']}" for c in citations)
+
+
+def _max_temp_n(citations: list[dict[str, Any]]) -> int:
+    return max((c.get("temp_n") or 0 for c in citations), default=0)
+
+
+def _build_write_user(paper: dict, outline: dict, sec: dict, cit_text: str, style: str, cite_mode: str) -> str:
+    parts: list[str] = []
+    parts.append(f"论文标题：{paper.get('title', '')}")
+    parts.append(f"论文类型：{paper.get('paper_type') or '综述'}")
+    parts.append(f"语言：{paper.get('language') or '中文'}")
+    parts.append(f"格式要求：{paper.get('format_req') or '无'}")
+    parts.append(f"写作风格：{style or paper.get('style') or '学术严谨'}")
+    parts.append("")
+    parts.append("论文大纲：")
+    for s in outline.get("sections") or []:
+        mark = "（当前撰写节）" if s.get("section_id") == sec.get("section_id") else ""
+        parts.append(f"- {s.get('section_id')}: {s.get('title')}{mark} 要点：{'；'.join(s.get('key_points') or [])}")
+    parts.append("")
+    parts.append(f"现在撰写章节「{sec.get('title', '')}」（section_id={sec.get('section_id')}）。")
+    if sec.get("key_points"):
+        parts.append("本节要点：" + "；".join(loads_json(sec.get("key_points"), [])))
+    parts.append("")
+    parts.append("全局已编号文献列表（[n] 引用只能从这里取值，或从最大编号 +1 起新增）：")
+    parts.append(cit_text)
+    return "\n".join(parts)
+
+
+def _validate_numbers(markdown: str, valid: set[int]) -> list[int]:
+    return [n for n in find_cite_numbers(markdown) if n not in valid]
+
+
+def _kind_of(meta: dict[str, Any]) -> str:
+    if meta.get("doi"):
+        return "doi"
+    if meta.get("pmid"):
+        return "pmid"
+    return "title"
+
+
+async def _persist_new_citations(paper_id: str, section_id: str, used: list[Any], existing_nums: set[int]) -> list[dict[str, Any]]:
+    """把 citations_used 中的新引用（temp_n 不在既有集合）落库，返回新增摘要。"""
+    new_citations: list[dict[str, Any]] = []
+    seen_new: set[int] = set()
+    for c in used:
+        if not isinstance(c, dict):
+            continue
+        tn = int(c.get("temp_n") or 0)
+        if not tn or tn in existing_nums or tn in seen_new:
+            continue
+        seen_new.add(tn)
+        ref_text = render_gbt7714(c)
+        await db.execute(
+            "INSERT INTO citations (paper_id, section_id, temp_n, kind, ref_text, source_doc_id, snippet, created_at) "
+            "VALUES (?, ?, ?, ?, ?, ?, ?, ?)",
+            (
+                paper_id, section_id, tn, _kind_of(c), ref_text,
+                c.get("source_doc_id") or "", c.get("snippet") or "", now_iso(),
+            ),
+        )
+        new_citations.append({"temp_n": tn, "ref_text": ref_text})
+    return new_citations
+
+
+# ---------------------------------------------------------------------------
+# 写节
+# ---------------------------------------------------------------------------
+
+async def write_paper_section(paper_id: str, section_id: str, style: str = "", cite_mode: str = "") -> dict[str, Any]:
+    paper = await _get_paper(paper_id)
+    if not paper:
+        return {"ok": False, "code": "NOT_FOUND", "message": "论文不存在"}
+    sections = await _get_sections(paper_id)
+    sec = next((s for s in sections if s["section_id"] == section_id), None)
+    if not sec:
+        return {"ok": False, "code": "NOT_FOUND", "message": "章节不存在"}
+    citations = await _get_citations(paper_id)
+    outline = loads_json(paper.get("outline_json"), {})
+    cit_text = _global_citations_text(citations)
+    existing_nums = {c.get("temp_n") for c in citations}
+
+    user = _build_write_user(paper, outline, sec, cit_text, style, cite_mode)
+    markdown = ""
+    used: list[Any] = []
+    bad: list[int] = []
+    for attempt in range(2):
+        try:
+            res = await llm.complete_json(_WRITE_SYSTEM, user, max_tokens=8000)
+        except llm.LLMError as exc:
+            return {"ok": False, "code": "LLM_ERROR", "message": f"写节 LLM 失败：{exc}"}
+        markdown = res.get("markdown") or ""
+        used = res.get("citations_used") or []
+        used_nums = {int(c.get("temp_n") or 0) for c in used if isinstance(c, dict)}
+        valid = existing_nums | used_nums
+        bad = _validate_numbers(markdown, valid)
+        if not bad:
+            break
+        user += f"\n（上次正文出现的引用编号 {sorted(bad)} 非法；只能使用已有编号或从 {_max_temp_n(citations) + 1} 起的新编号，重试。）"
+    else:
+        return {"ok": False, "code": "CITE_VALIDATION_FAILED", "message": f"引用编号非法：{sorted(bad)}"}
+
+    new_citations = await _persist_new_citations(paper_id, section_id, used, existing_nums)
+    await db.execute(
+        "UPDATE paper_sections SET markdown = ?, word_count = ?, status = 'done', updated_at = ? WHERE paper_id = ? AND section_id = ?",
+        (markdown, len(markdown), now_iso(), paper_id, section_id),
+    )
+    await db.execute("UPDATE papers SET status = 'writing', updated_at = ? WHERE id = ?", (now_iso(), paper_id))
+    return {"ok": True, "section_id": section_id, "status": "done", "markdown": markdown,
+            "citations_used": used, "new_citations": new_citations}
+
+
+# ---------------------------------------------------------------------------
+# 修订
+# ---------------------------------------------------------------------------
+
+async def revise_paper_section(paper_id: str, section_id: str, instruction: str) -> dict[str, Any]:
+    paper = await _get_paper(paper_id)
+    if not paper:
+        return {"ok": False, "code": "NOT_FOUND", "message": "论文不存在"}
+    sections = await _get_sections(paper_id)
+    sec = next((s for s in sections if s["section_id"] == section_id), None)
+    if not sec:
+        return {"ok": False, "code": "NOT_FOUND", "message": "章节不存在"}
+    citations = await _get_citations(paper_id)
+    outline = loads_json(paper.get("outline_json"), {})
+    existing_nums = {c.get("temp_n") for c in citations}
+
+    user = _build_write_user(paper, outline, sec, _global_citations_text(citations), "", "")
+    user += "\n\n原章节正文：\n" + (sec.get("markdown") or "")
+    user += f"\n\n修订要求：{instruction}"
+
+    markdown = ""
+    used: list[Any] = []
+    bad: list[int] = []
+    for attempt in range(2):
+        try:
+            res = await llm.complete_json(_REVISE_SYSTEM, user, max_tokens=8000)
+        except llm.LLMError as exc:
+            return {"ok": False, "code": "LLM_ERROR", "message": f"修订 LLM 失败：{exc}"}
+        markdown = res.get("markdown") or ""
+        used = res.get("citations_used") or []
+        used_nums = {int(c.get("temp_n") or 0) for c in used if isinstance(c, dict)}
+        valid = existing_nums | used_nums
+        bad = _validate_numbers(markdown, valid)
+        if not bad:
+            break
+        user += f"\n（引用编号 {sorted(bad)} 非法，重试。）"
+    else:
+        return {"ok": False, "code": "CITE_VALIDATION_FAILED", "message": f"引用编号非法：{sorted(bad)}"}
+
+    new_citations = await _persist_new_citations(paper_id, section_id, used, existing_nums)
+    await db.execute(
+        "UPDATE paper_sections SET markdown = ?, word_count = ?, status = 'done', updated_at = ? WHERE paper_id = ? AND section_id = ?",
+        (markdown, len(markdown), now_iso(), paper_id, section_id),
+    )
+    return {"ok": True, "section_id": section_id, "markdown": markdown,
+            "citations_used": used, "citation_delta": len(new_citations), "new_citations": new_citations}
+
+
+async def revise_paper(paper_id: str, instruction: str) -> dict[str, Any]:
+    """LLM 判定受影响章节，返回受影响节清单 + 修订计划（由调用方逐节 revise_paper_section）。"""
+    paper = await _get_paper(paper_id)
+    if not paper:
+        return {"ok": False, "code": "NOT_FOUND", "message": "论文不存在"}
+    sections = await _get_sections(paper_id)
+    outline = loads_json(paper.get("outline_json"), {})
+
+    system = (
+        "你是论文修订计划助手。根据用户修订意见与论文大纲，判定需要修改哪些章节。"
+        "只输出一个合法 JSON 对象：{\"affected_section_ids\": [\"s1\", ...], \"plan\": 修订计划(一句话)}。"
+    )
+    sec_list = "\n".join(f"{s['section_id']}: {s['title']}" for s in sections)
+    user = f"论文标题：{paper.get('title')}\n大纲章节：\n{sec_list}\n\n修订意见：{instruction}"
+    try:
+        res = await llm.complete_json(system, user, max_tokens=2000)
+    except llm.LLMError as exc:
+        return {"ok": False, "code": "LLM_ERROR", "message": f"修订计划 LLM 失败：{exc}"}
+    affected = [s for s in (res.get("affected_section_ids") or []) if any(x["section_id"] == s for x in sections)]
+    return {"ok": True, "affected_section_ids": affected, "plan": res.get("plan", "")}
diff --git a/paper_writer_service/build.sh b/paper_writer_service/build.sh
new file mode 100755
index 0000000..663d69b
--- /dev/null
+++ b/paper_writer_service/build.sh
@@ -0,0 +1,119 @@
+#!/usr/bin/env bash
+# build.sh —— PAPER 服务编译硬化（复制自 bioflow_service/build.sh，白名单改造）。
+# 白名单(明文)：app/main.py(uvicorn 入口)、app/config.py(运维可读面)、
+# app/schemas.py(常量面)、app/db.py(aiosqlite 协程取消热区——Cython 协程断流断言
+# abort 是既有教训，明文保留)、app/mcp.py(MCP 工具层，mcp 依赖 pydantic + async 工具)、
+# services/paper/{llm,search,write,outline,complete}.py(长 LLM/httpx/aiosqlite 调用，
+# 断流崩溃高危点，明文走 CPython 原生协程)+ 全部 __init__.py。
+# 仅纯同步逻辑(cite.py 等)可编译。
+# 用法:build.sh <source-root> <output-dir> [python]
+set -euo pipefail
+umask 027
+
+SOURCE_ROOT="${1:?Usage: $0 <source-root> <output-dir> [python]}"
+OUTPUT_DIR="${2:?Usage: $0 <source-root> <output-dir> [python]}"
+PYTHON="${3:-python3}"
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+VISIBILITY_TOOL="${ELF_VISIBILITY_TOOL:-$SCRIPT_DIR/../tooling/compile/tools/elf_visibility.py}"
+
+SOURCE_ROOT="$(realpath "$SOURCE_ROOT")"
+OUTPUT_DIR="$(realpath -m "$OUTPUT_DIR")"
+
+[[ -d "$SOURCE_ROOT" ]] || { echo "Paper service source missing: $SOURCE_ROOT" >&2; exit 2; }
+[[ -x "$PYTHON" ]] || { echo "Python is not executable: $PYTHON" >&2; exit 2; }
+[[ -f "$VISIBILITY_TOOL" ]] || { echo "ELF visibility tool missing: $VISIBILITY_TOOL" >&2; exit 2; }
+for command_name in gcc strip rsync; do command -v "$command_name" >/dev/null; done
+
+case "$OUTPUT_DIR" in
+  "/"|"$(realpath -m "$HOME")"|"$SOURCE_ROOT")
+    echo "Unsafe output directory: $OUTPUT_DIR" >&2; exit 2 ;;
+esac
+[[ ! -e "$OUTPUT_DIR" ]] || { echo "Output exists; refusing to overwrite: $OUTPUT_DIR" >&2; exit 2; }
+
+echo "[PAPER BUILD 1/5] Copying source (excluding tests)"
+mkdir -p "$OUTPUT_DIR"
+rsync -a --exclude '__pycache__' --exclude 'test_*.py' --exclude 'data' --exclude '.env' \
+  --exclude 'build.sh' --exclude 'restart.sh' --exclude 'pytest.ini' \
+  "$SOURCE_ROOT/" "$OUTPUT_DIR/"
+
+WHITELIST_PY=(
+  "app/main.py"
+  "app/config.py"
+  "app/schemas.py"
+  "app/db.py"
+  "app/mcp.py"
+  "app/services/paper/llm.py"
+  "app/services/paper/search.py"
+  "app/services/paper/write.py"
+  "app/services/paper/outline.py"
+  "app/services/paper/complete.py"
+)
+_is_whitelisted() {
+  local rel="$1"
+  [[ "$(basename "$rel")" == "__init__.py" ]] && return 0
+  for w in "${WHITELIST_PY[@]}"; do
+    case "$rel" in
+      $w) return 0 ;;
+    esac
+  done
+  return 1
+}
+
+echo "[PAPER BUILD 2/5] Discovering Python modules"
+PY_FILES=()
+while IFS= read -r -d '' path; do PY_FILES+=("$path"); done \
+  < <(find "$OUTPUT_DIR" -type f -name '*.py' -print0)
+FILTERED=()
+for py_file in "${PY_FILES[@]}"; do
+  rel="${py_file#"$OUTPUT_DIR"/}"
+  if _is_whitelisted "$rel"; then
+    echo "  [whitelist] $rel (明文保留)"
+    continue
+  fi
+  FILTERED+=("$py_file")
+done
+PY_FILES=("${FILTERED[@]}")
+TOTAL="${#PY_FILES[@]}"
+(( TOTAL > 0 )) || { echo "No paper service modules selected" >&2; exit 3; }
+
+INCLUDE_PY="$("$PYTHON" -c 'import sysconfig; print(sysconfig.get_config_var("INCLUDEPY") or "")')"
+EXT_SUFFIX="$("$PYTHON" -c 'import sysconfig; print(sysconfig.get_config_var("EXT_SUFFIX") or ".so")')"
+
+echo "[PAPER BUILD 3/5] Cython preflight ($TOTAL modules)"
+COUNT=0; FAILURES=0
+for py_file in "${PY_FILES[@]}"; do
+  COUNT=$((COUNT+1))
+  directory="$(dirname "$py_file")"
+  module="$(basename "$py_file" .py)"
+  c_file="$directory/$module.c"
+  echo "  [$COUNT/$TOTAL] ${py_file#"$OUTPUT_DIR"/}"
+  if ! "$PYTHON" -m cython -3 \
+      --directive always_allow_keywords=true,annotation_typing=False \
+      --no-docstrings "$py_file" -o "$c_file" 2>/dev/null; then
+    FAILURES=$((FAILURES+1))
+  fi
+done
+(( FAILURES == 0 )) || { echo "Paper service Cython failed: $FAILURES" >&2; exit 3; }
+
+echo "[PAPER BUILD 4/5] GCC hidden visibility + strip"
+COUNT=0
+for py_file in "${PY_FILES[@]}"; do
+  COUNT=$((COUNT+1))
+  directory="$(dirname "$py_file")"
+  module="$(basename "$py_file" .py)"
+  c_file="$directory/$module.c"
+  so_file="$directory/$module$EXT_SUFFIX"
+  gcc -shared -pthread -fPIC -O2 -fvisibility=hidden \
+    -I"$INCLUDE_PY" -o "$so_file" "$c_file"
+  strip --strip-unneeded "$so_file"
+  "$PYTHON" -B "$VISIBILITY_TOOL" --hide-non-pyinit --check "$so_file" >/dev/null
+  rm -f -- "$c_file" "$py_file"
+done
+
+echo "[PAPER BUILD 5/5] Final checks"
+SO_COUNT="$(find "$OUTPUT_DIR" -type f -name '*.so' | wc -l)"
+[[ "$SO_COUNT" -eq "$TOTAL" ]] || { echo "count mismatch: $TOTAL vs $SO_COUNT" >&2; exit 4; }
+if find "$OUTPUT_DIR" -type f -name '*.c' -print -quit | grep -q .; then
+  echo "Generated C source remains" >&2; exit 4
+fi
+echo "Paper service compile+harden complete: modules=$SO_COUNT output=$OUTPUT_DIR"
diff --git a/paper_writer_service/pytest.ini b/paper_writer_service/pytest.ini
new file mode 100644
index 0000000..4087298
--- /dev/null
+++ b/paper_writer_service/pytest.ini
@@ -0,0 +1,6 @@
+[pytest]
+python_files = test_*.py
+markers =
+    paper_unit: 论文服务单元测试（引用协议/大纲/写节/重排/complete）
+    paper_mcp: MCP 协议互操作测试
+asyncio_mode = auto
diff --git a/paper_writer_service/restart.sh b/paper_writer_service/restart.sh
new file mode 100755
index 0000000..07f74c4
--- /dev/null
+++ b/paper_writer_service/restart.sh
@@ -0,0 +1,34 @@
+#!/usr/bin/env bash
+# restart.sh —— PAPER 服务重启脚本（8097；不触碰生产服务 8090/8091/8092/8093/8094/8098）。
+# 零硬编码：venv/python/仓库路径全部从脚本位置推导或环境变量覆盖。
+# 用法: restart.sh [python] [port]
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
+PYTHON="${1:-$REPO_ROOT/../plantai-rag-new-venv/bin/python}"
+PORT="${2:-8097}"
+HOST="${PAPER_BIND_HOST:-127.0.0.1}"
+
+[[ -x "$PYTHON" ]] || { echo "Python not found: $PYTHON" >&2; exit 2; }
+
+APP_MODULE="app.main:app"
+if command -v lsof >/dev/null 2>&1; then
+  PID="$(lsof -ti tcp:"$PORT" -sTCP:LISTEN 2>/dev/null || true)"
+  if [[ -n "$PID" ]]; then
+    echo "[restart] 停止 $PORT 旧进程 pid=$PID"
+    kill $PID 2>/dev/null || true
+    sleep 2
+  fi
+fi
+
+cd "$SCRIPT_DIR"
+# 加载运行时配置（.env 不入 git）：services/paper/llm.py 直接读 DOCGEN_LLM_* 裸键，
+# search.py 读 VSTORE_BASE_URL——需 set -a 显式导出（同 bioflow_service/restart.sh 先例）。
+set -a
+[ -f .env ] && source .env
+set +a
+echo "[restart] 启动论文撰写服务 :$PORT（$PYTHON）"
+nohup "$PYTHON" -m uvicorn "$APP_MODULE" --host "$HOST" --port "$PORT" \
+  > "$SCRIPT_DIR/paper.log" 2>&1 &
+echo "[restart] 完成；日志 paper_writer_service/paper.log"
diff --git a/sdk/app/config.py b/sdk/app/config.py
index 6aa7222..caa9685 100644
--- a/sdk/app/config.py
+++ b/sdk/app/config.py
@@ -76,6 +76,11 @@ class Settings(BaseSettings):
     bioflow_enabled: bool = False            # 三形态部署开关（默认关：只部署 rag 无生信痕迹）
     bioflow_service_url: str = "http://127.0.0.1:8098"  # 生信服务 8098（BFF 流式反代目标）
 
+    # ---- 论文撰写服务（paper_writer_service 8097；PAPER_WRITER_ENABLED 为假时路由不注册）----
+    paper_writer_enabled: bool = False
+    paper_writer_url: str = "http://127.0.0.1:8097"
+    paper_writer_read_timeout: float = 360.0  # write_paper_section 同步长调用读超时（秒）
+
     # ---- 生产形态：BFF 静态托管 ui/dist（ARCHITECTURE.md §3，M6 挂载）----
     frontend_dist_dir: str = "ui/dist"  # 生产形态：静态托管 ui/dist
 
diff --git a/sdk/app/main.py b/sdk/app/main.py
index c494a7f..d15fac9 100644
--- a/sdk/app/main.py
+++ b/sdk/app/main.py
@@ -24,6 +24,7 @@ from .config import get_settings
 from .db import close_db, init_db
 from .routers import capabilities, docgen, extraction, figures, health, kg, legacy, qa, sessions
 from .routers import bioflow_proxy
+from .routers import paper
 from .routers import private_docs
 from .upstream import close_client
 
@@ -59,6 +60,8 @@ def create_app() -> FastAPI:
     app.include_router(private_docs.router)  # 私有资料问答（FR-PRIV-004）
     if get_settings().bioflow_enabled:
         app.include_router(bioflow_proxy.router)  # 生信反代（BIOFLOW_ENABLED 开关，FR-BIO-028）
+    if get_settings().paper_writer_enabled:
+        app.include_router(paper.router)  # 论文撰写（PAPER_WRITER_ENABLED 开关）
 
     @app.exception_handler(RequestValidationError)
     async def _validation_handler(_: Request, exc: RequestValidationError) -> JSONResponse:
diff --git a/sdk/app/routers/docgen.py b/sdk/app/routers/docgen.py
index 9e8c236..bceea50 100644
--- a/sdk/app/routers/docgen.py
+++ b/sdk/app/routers/docgen.py
@@ -184,6 +184,60 @@ def _read_artifact_file(settings: Any, rel: str) -> bytes | None:
         return None
 
 
+async def insert_artifact(
+    db: Database,
+    *,
+    session_id: str,
+    title: str,
+    markdown: str,
+    md_rel: str | None = None,
+    template_name: str = "",
+    template_key: str = "",
+    template_css: str = "",
+    html: str = "",
+    letterhead: str = "",
+    file_name: str = "",
+) -> str:
+    """写制品 md（+可选 html）落盘 + INSERT doc_artifacts，返回 artifact_id。
+
+    docgen 与论文撰写（paper）共用：paper 仅传 markdown/title/template_name。
+    """
+    artifact_id = str(uuid.uuid4())
+    artifact_dir = _artifact_dir(get_settings())
+    if not md_rel:
+        md_rel = f"artifacts/{artifact_id}.md"
+    if markdown:
+        safe_rel = _safe_rel(md_rel)
+        if safe_rel:
+            path = artifact_dir / safe_rel
+            try:
+                path.parent.mkdir(parents=True, exist_ok=True)
+                path.write_text(markdown, encoding="utf-8")
+            except OSError as exc:
+                logger.warning("docgen: 制品落盘失败（不影响返回）: %s", exc)
+    html_rel = ""
+    if html:
+        html_rel = f"artifacts/{artifact_id}.html"
+        safe_rel = _safe_rel(html_rel)
+        if safe_rel:
+            path = artifact_dir / safe_rel
+            try:
+                path.parent.mkdir(parents=True, exist_ok=True)
+                path.write_text(html, encoding="utf-8")
+            except OSError as exc:
+                logger.warning("docgen: 模版 HTML 落盘失败（不影响返回）: %s", exc)
+                html_rel = ""
+    await db.execute(
+        "INSERT INTO doc_artifacts (id, session_id, status, title, template_name, "
+        "template_key, template_css, md_path, docx_path, pdf_path, html_path, letterhead, file_name, error, created_at) "
+        "VALUES (?, ?, 'done', ?, ?, ?, ?, ?, NULL, NULL, ?, ?, ?, NULL, ?)",
+        (artifact_id, session_id, title,
+         template_name, template_key, template_css, md_rel,
+         html_rel or None, letterhead, file_name, _db_now()),
+    )
+    return artifact_id
+
+
 # ---------------------------------------------------------------------------
 # 制品目录配额 / TTL 惰性清理（TL 裁决 2026-09-03：V1 由 BFF 实现）
 # ---------------------------------------------------------------------------
@@ -516,13 +570,7 @@ async def _run_docgen_task(
             on_progress=on_progress,
         )
 
-        # ---- 以下与修复前同步路径一致：落盘 + artifact 落库 + 配额复查 ----
-        artifact_id = str(uuid.uuid4())
-        markdown = str(result.get("markdown") or "")
-        # pipeline 的 md_path 在 DOCGEN_ROOT 为绝对路径时也是绝对路径——
-        # 落库要求相对路径（_safe_rel 拒绝绝对路径，否则导出报「制品内容缺失」，
-        # 2026-09-10 dev 自测实测）。绝对且在制品根内 → 相对化；否则回退
-        # artifacts/<artifact_id>.md
+        # ---- 落盘 + artifact 落库（抽公共 insert_artifact，论文撰写共用）----
         artifact_dir = _artifact_dir(get_settings())
         md_rel = ""
         raw_md_path = str(result.get("md_path") or "")
@@ -535,17 +583,6 @@ async def _run_docgen_task(
                     md_rel = raw_md_path
             except ValueError:
                 md_rel = ""
-        if not md_rel:
-            md_rel = f"artifacts/{artifact_id}.md"
-        if markdown:
-            safe_rel = _safe_rel(md_rel)
-            if safe_rel:
-                path = artifact_dir / safe_rel
-                try:
-                    path.parent.mkdir(parents=True, exist_ok=True)
-                    path.write_text(markdown, encoding="utf-8")
-                except OSError as exc:
-                    logger.warning("docgen: 制品落盘失败（不影响返回）: %s", exc)
 
         template_key = str(result.get("template_key") or "")
         if not template_key and template is not None and template.get("source") == "builtin":
@@ -557,34 +594,26 @@ async def _run_docgen_task(
             (template or {}).get("name", "") or (template or {}).get("title", "")
         )
         html = str(result.get("html") or "")
-        html_rel = ""
-        if html:
-            html_rel = f"artifacts/{artifact_id}.html"
-            safe_rel = _safe_rel(html_rel)
-            if safe_rel:
-                path = _artifact_dir(get_settings()) / safe_rel
-                try:
-                    path.parent.mkdir(parents=True, exist_ok=True)
-                    path.write_text(html, encoding="utf-8")
-                except OSError as exc:
-                    logger.warning("docgen: 模版 HTML 落盘失败（不影响返回）: %s", exc)
-                    html_rel = ""
+
         try:
             quota_after = await _check_quota(db, get_settings())
             if quota_after is not None:
                 logger.warning("docgen: 生成后制品目录已超配额（下一次 generate/export 将被拒绝）")
         except Exception as exc:
             logger.warning("docgen: 生成后配额复查失败（忽略）: %s", exc)
-        await db.execute(
-            "INSERT INTO doc_artifacts (id, session_id, status, title, template_name, "
-            "template_key, template_css, md_path, docx_path, pdf_path, html_path, letterhead, file_name, error, created_at) "
-            "VALUES (?, ?, 'done', ?, ?, ?, ?, ?, NULL, NULL, ?, ?, ?, NULL, ?)",
-            (artifact_id, session_id, title,
-             tpl_display, template_key, template_css, md_rel,
-             html_rel or None,
-             str(result.get("letterhead") or "")[:60],
-             str(result.get("file_name") or name_for_prompt(description) or title or "")[:60],
-             _db_now()),
+
+        artifact_id = await insert_artifact(
+            db,
+            session_id=session_id,
+            title=title,
+            markdown=str(result.get("markdown") or ""),
+            md_rel=md_rel or None,
+            template_name=tpl_display,
+            template_key=template_key,
+            template_css=template_css,
+            html=html,
+            letterhead=str(result.get("letterhead") or "")[:60],
+            file_name=str(result.get("file_name") or name_for_prompt(description) or title or "")[:60],
         )
         # 终态 progress 快照：全部节 done（限频写库可能截断最后一帧）
         final_progress = [
diff --git a/sdk/app/routers/paper.py b/sdk/app/routers/paper.py
new file mode 100644
index 0000000..b95adf3
--- /dev/null
+++ b/sdk/app/routers/paper.py
@@ -0,0 +1,90 @@
+"""routers/paper.py —— 论文撰写路由（对话大脑）。
+
+条件注册（PAPER_WRITER_ENABLED 为假不注册，同 bioflow_proxy 模式）。
+端点：health（探活反代）/ turns（一轮状态机）/ confirm / resume / progress / citations。
+"""
+from __future__ import annotations
+
+import asyncio
+from typing import Any
+
+import httpx
+from fastapi import APIRouter, Depends
+from pydantic import BaseModel
+
+from ..config import get_settings
+from ..db import Database, get_db
+from ..services.paper import agent
+from ..services.paper.mcp_client import PaperMcpClient
+
+router = APIRouter(prefix="/api/paper", tags=["paper"])
+
+
+class PaperTurnRequest(BaseModel):
+    session_id: str
+    text: str
+    paper_id: str | None = None
+
+
+class PaperConfirmRequest(BaseModel):
+    session_id: str
+
+
+def _client() -> PaperMcpClient:
+    s = get_settings()
+    return PaperMcpClient(s.paper_writer_url, s.paper_writer_read_timeout)
+
+
+@router.get("/health")
+async def paper_health() -> dict[str, Any]:
+    """探活反代 8097 /health；离线返回 offline 标记（前端据此隐藏 paper 入口）。"""
+    s = get_settings()
+    try:
+        async with httpx.AsyncClient(timeout=10.0) as client:
+            resp = await client.get(f"{s.paper_writer_url.rstrip('/')}/health")
+            resp.raise_for_status()
+            return resp.json()
+    except Exception:  # noqa: BLE001
+        return {"ok": False, "status": "offline"}
+
+
+@router.post("/turns")
+async def paper_turn(payload: PaperTurnRequest, db: Database = Depends(get_db)) -> dict[str, Any]:
+    blocks = await agent.handle_turn(db, payload.session_id, payload.text, payload.paper_id)
+    return {"ok": True, "blocks": blocks}
+
+
+@router.post("/papers/{paper_id}/confirm")
+async def paper_confirm(paper_id: str, payload: PaperConfirmRequest, db: Database = Depends(get_db)) -> dict[str, Any]:
+    asyncio.create_task(agent.run_writing_loop(db, paper_id, payload.session_id))
+    return {"ok": True, "paper_id": paper_id}
+
+
+@router.post("/papers/{paper_id}/resume")
+async def paper_resume(paper_id: str, payload: PaperConfirmRequest, db: Database = Depends(get_db)) -> dict[str, Any]:
+    asyncio.create_task(agent.run_writing_loop(db, paper_id, payload.session_id))
+    return {"ok": True, "paper_id": paper_id}
+
+
+@router.get("/papers/{paper_id}")
+async def paper_progress(paper_id: str) -> dict[str, Any]:
+    client = _client()
+    try:
+        return await client.call_tool("get_paper_progress", {"paper_id": paper_id})
+    except Exception as exc:  # noqa: BLE001
+        return {"ok": False, "code": "UPSTREAM_DOWN", "message": str(exc)}
+
+
+@router.get("/papers/{paper_id}/citations/{n}")
+async def paper_citation(paper_id: str, n: int) -> dict[str, Any]:
+    client = _client()
+    try:
+        paper = await client.call_tool("get_paper", {"paper_id": paper_id})
+    except Exception as exc:  # noqa: BLE001
+        return {"ok": False, "code": "UPSTREAM_DOWN", "message": str(exc)}
+    if not paper.get("ok"):
+        return paper
+    for ref in paper.get("references") or []:
+        if ref.get("n") == n:
+            return {"ok": True, "citation": ref}
+    return {"ok": False, "code": "NOT_FOUND", "message": f"引用 [{n}] 不存在"}
diff --git a/sdk/app/schemas.py b/sdk/app/schemas.py
index 2b3a3ad..f71a157 100644
--- a/sdk/app/schemas.py
+++ b/sdk/app/schemas.py
@@ -99,6 +99,58 @@ class ArtifactBlock(BaseModel):
     title: str = ""
     template_name: str = ""
     status: str = "done"  # running|done|error（生成中/已就绪/失败）
+    paper_id: str = ""       # 论文撰写功能：关联论文 id（DocPane 据此拉 citation popover）
+    citation_count: int = 0  # 论文撰写功能：引用条数
+
+
+# ---------------------------------------------------------------------------
+# §2c 论文块（paper-writer；BFF 仅持久化校验，字段契约归前端 Paper*Card）
+# ---------------------------------------------------------------------------
+
+class PaperBriefBlock(BaseModel):
+    """论文引导表单卡：预填已抽取 brief 字段，缺失字段待补。"""
+
+    model_config = ConfigDict(extra="allow")
+
+    type: Literal["paper_brief"] = "paper_brief"
+    paper_id: str = ""
+    status: str = "pending"  # pending|submitted
+    fields: list[dict[str, Any]] = Field(default_factory=list)
+
+
+class PaperOutlineBlock(BaseModel):
+    """论文大纲确认卡：章节勾选/编辑要点。"""
+
+    model_config = ConfigDict(extra="allow")
+
+    type: Literal["paper_outline"] = "paper_outline"
+    paper_id: str = ""
+    title: str = ""
+    status: str = "pending"  # pending|confirmed
+    sections: list[dict[str, Any]] = Field(default_factory=list)
+
+
+class PaperProgressBlock(BaseModel):
+    """论文节级进度卡：逐节撰写状态。"""
+
+    model_config = ConfigDict(extra="allow")
+
+    type: Literal["paper_progress"] = "paper_progress"
+    paper_id: str = ""
+    status: str = "writing"  # writing|done|error|cancelled
+    sections: list[dict[str, Any]] = Field(default_factory=list)
+    current_section_id: str = ""
+
+
+class PaperRefsBlock(BaseModel):
+    """论文参考文献卡：文末 GB/T 7714 列表 + disclaimer。"""
+
+    model_config = ConfigDict(extra="allow")
+
+    type: Literal["paper_refs"] = "paper_refs"
+    paper_id: str = ""
+    references: list[dict[str, Any]] = Field(default_factory=list)
+    disclaimer: str = ""
 
 
 # ---------------------------------------------------------------------------
@@ -245,6 +297,7 @@ class PrivateQARequest(BaseModel):
 Block = Union[
     TextBlock, TableBlock, CodeBlock, StreamBlock,
     PipelineBlock, ToolBlock, SourcesBlock, GraphBlock, ArtifactBlock,
+    PaperBriefBlock, PaperOutlineBlock, PaperProgressBlock, PaperRefsBlock,
     BioflowPathBlock, BioflowProgressBlock, BioflowFormBlock, BioflowFileBlock,
     BioflowArchiveBlock, BioflowQaBlock, BioflowReuseBlock,
     PrivateDocBlock,
diff --git a/sdk/app/services/paper/__init__.py b/sdk/app/services/paper/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/sdk/app/services/paper/agent.py b/sdk/app/services/paper/agent.py
new file mode 100644
index 0000000..eea4cc8
--- /dev/null
+++ b/sdk/app/services/paper/agent.py
@@ -0,0 +1,267 @@
+"""services/paper/agent.py —— 论文对话规则状态机（brief→outline→writing→complete/revise）。
+
+LLM 仅做 brief 字段抽取（complete_json）；大纲/正文/引用/合并全部由 paper 服务完成。
+不引入 function calling（BFF 无工具循环先例）。助手消息由本模块进程内直写 messages 表。
+"""
+from __future__ import annotations
+
+import json
+import logging
+import uuid
+from datetime import datetime, timezone
+from typing import Any
+
+from ...config import get_settings
+from ..docgen.llm import complete_json as llm_complete_json
+from .mcp_client import PaperMcpClient, PaperMcpError
+
+logger = logging.getLogger(__name__)
+
+BRIEF_FIELDS = [
+    {"key": "topic", "label": "论文主题", "options": None},
+    {"key": "paper_type", "label": "论文类型", "options": ["综述", "研究论文", "开题报告", "实验报告"]},
+    {"key": "language", "label": "语言", "options": ["中文", "英文"]},
+    {"key": "length", "label": "篇幅（字数）", "options": ["3000字", "5000字", "8000字"]},
+    {"key": "format_req", "label": "格式要求", "options": ["GB/T 7714", "APA", "MLA"]},
+    {"key": "style", "label": "写作风格", "options": ["学术严谨", "通俗易懂"]},
+    {"key": "requirements", "label": "其他要求", "options": None},
+]
+
+REQUIRED_BRIEF_KEYS = ("topic", "paper_type", "language", "length")
+
+_BRIEF_SYSTEM = (
+    "你是论文撰写需求抽取助手。从用户输入中抽取论文写作 brief。"
+    "只输出一个合法 JSON 对象：{\"topic\":\"\",\"paper_type\":\"\",\"language\":\"\","
+    "\"length\":\"\",\"format_req\":\"\",\"style\":\"\",\"requirements\":\"\"}。"
+    "抽取不到的字段填空字符串。"
+)
+
+
+def _db_now() -> str:
+    return datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
+
+
+def _text_block(html: str) -> dict[str, Any]:
+    return {"type": "text", "html": html}
+
+
+def _new_client() -> PaperMcpClient:
+    s = get_settings()
+    return PaperMcpClient(s.paper_writer_url, s.paper_writer_read_timeout)
+
+
+async def _write_message(
+    db,
+    session_id: str,
+    *,
+    kind: str = "paper",
+    status: str = "done",
+    blocks: list[dict[str, Any]],
+    text: str | None = None,
+) -> str:
+    """直写助手消息（kind=paper），返回 message_id。"""
+    mid = str(uuid.uuid4())
+    now = _db_now()
+    await db.execute(
+        "INSERT INTO messages (id, session_id, role, kind, status, blocks, raw, text, created_at, updated_at) "
+        "VALUES (?, ?, 'assistant', ?, ?, ?, NULL, ?, ?, ?)",
+        (mid, session_id, kind, status, json.dumps(blocks, ensure_ascii=False), text, now, now),
+    )
+    await db.execute(
+        "UPDATE sessions SET status = CASE WHEN status = 'idle' THEN 'running' ELSE status END, updated_at = ? WHERE id = ?",
+        (now, session_id),
+    )
+    return mid
+
+
+async def extract_brief(text: str) -> dict[str, str]:
+    """LLM 抽取 brief 字段（失败返回全空，走 paper_brief 卡兜底追问）。"""
+    try:
+        brief = await llm_complete_json(_BRIEF_SYSTEM, text, max_tokens=1000)
+    except Exception as exc:  # noqa: BLE001
+        logger.warning("paper: brief 抽取失败：%s", exc)
+        brief = {}
+    return {f["key"]: str((brief or {}).get(f["key"]) or "").strip() for f in BRIEF_FIELDS}
+
+
+def _missing_brief_keys(brief: dict[str, str]) -> list[str]:
+    return [k for k in REQUIRED_BRIEF_KEYS if not brief.get(k)]
+
+
+# ---------------------------------------------------------------------------
+# 状态机入口
+# ---------------------------------------------------------------------------
+
+async def handle_turn(db, session_id: str, text: str, paper_id: str | None = None) -> list[dict[str, Any]]:
+    """一轮对话：推进状态机并落库助手消息，返回消息的 blocks 列表。"""
+    client = _new_client()
+    if not paper_id:
+        return await _trigger(db, session_id, text, client)
+    return await _followup(db, session_id, text, paper_id, client)
+
+
+async def _trigger(db, session_id: str, text: str, client: PaperMcpClient) -> list[dict[str, Any]]:
+    brief = await extract_brief(text)
+    missing = _missing_brief_keys(brief)
+    if missing:
+        fields = [{**f, "value": brief.get(f["key"], "")} for f in BRIEF_FIELDS]
+        blocks = [
+            _text_block("好的，我来帮你写这篇论文。请补充以下信息（也可以直接打字告诉我）："),
+            {"type": "paper_brief", "paper_id": "", "status": "pending", "fields": fields},
+        ]
+        await _write_message(db, session_id, blocks=blocks, text="请补充论文信息")
+        return blocks
+
+    outline = await client.call_tool("propose_paper_outline", brief)
+    if not outline.get("ok"):
+        blocks = [_text_block(f"论文大纲生成失败：{outline.get('message', '服务未就绪')}，请重试。")]
+        await _write_message(db, session_id, blocks=blocks, text=blocks[0]["html"])
+        return blocks
+    return await _outline_message(db, session_id, outline)
+
+
+async def _outline_message(db, session_id: str, outline: dict[str, Any]) -> list[dict[str, Any]]:
+    paper_id = outline["paper_id"]
+    sections = [
+        {"section_id": s.get("section_id"), "title": s.get("title", ""),
+         "key_points": s.get("key_points") or [], "selected": True}
+        for s in (outline.get("sections") or [])
+    ]
+    blocks = [
+        _text_block(
+            f"已为你生成论文大纲《{outline.get('title', '')}》。"
+            "你可以勾选/编辑章节，或直接说修改意见；确认后我开始逐节撰写。"
+        ),
+        {"type": "paper_outline", "paper_id": paper_id, "title": outline.get("title", ""),
+         "status": "pending", "sections": sections},
+    ]
+    await _write_message(db, session_id, blocks=blocks, text=blocks[0]["html"])
+    return blocks
+
+
+async def _followup(db, session_id: str, text: str, paper_id: str, client: PaperMcpClient) -> list[dict[str, Any]]:
+    prog = await client.call_tool("get_paper_progress", {"paper_id": paper_id})
+    if not prog.get("ok"):
+        blocks = [_text_block(f"未找到论文（{paper_id}），请重新发起撰写。")]
+        await _write_message(db, session_id, blocks=blocks, text=blocks[0]["html"])
+        return blocks
+
+    status = prog.get("status")
+    if status == "complete":
+        return await _revise(db, session_id, paper_id, text, client)
+    if status == "cancelled":
+        blocks = [_text_block("论文已取消。如需重新撰写，请告诉我新主题。")]
+        await _write_message(db, session_id, blocks=blocks, text=blocks[0]["html"])
+        return blocks
+    if status in ("outline", "writing"):
+        # 大纲/撰写阶段：把 text 视为补充要求，重新生成大纲（原大纲作废）
+        brief = await extract_brief(text)
+        outline = await client.call_tool("propose_paper_outline", {**brief, "topic": brief.get("topic") or text})
+        if outline.get("ok"):
+            await client.call_tool("cancel_paper", {"paper_id": paper_id})
+            return await _outline_message(db, session_id, outline)
+        blocks = [_text_block("大纲修改失败，请重试。")]
+        await _write_message(db, session_id, blocks=blocks, text=blocks[0]["html"])
+        return blocks
+    blocks = [_text_block("论文撰写中，请稍候。")]
+    await _write_message(db, session_id, blocks=blocks, text=blocks[0]["html"])
+    return blocks
+
+
+async def _revise(db, session_id: str, paper_id: str, text: str, client: PaperMcpClient) -> list[dict[str, Any]]:
+    plan = await client.call_tool("revise_paper", {"paper_id": paper_id, "instruction": text})
+    if not plan.get("ok"):
+        blocks = [_text_block(f"修订计划生成失败：{plan.get('message', '')}")]
+        await _write_message(db, session_id, blocks=blocks, text=blocks[0]["html"])
+        return blocks
+    affected = plan.get("affected_section_ids") or []
+    if not affected:
+        blocks = [_text_block("没有需要修改的章节。")]
+        await _write_message(db, session_id, blocks=blocks, text=blocks[0]["html"])
+        return blocks
+
+    for sid in affected:
+        res = await client.call_tool(
+            "revise_paper_section", {"paper_id": paper_id, "section_id": sid, "instruction": text}
+        )
+        if not res.get("ok"):
+            blocks = [_text_block(f"章节修订失败：{res.get('message', '')}")]
+            await _write_message(db, session_id, blocks=blocks, text="修订失败")
+            return blocks
+
+    comp = await client.call_tool("complete_paper", {"paper_id": paper_id})
+    paper = await client.call_tool("get_paper", {"paper_id": paper_id})
+    if comp.get("ok") and paper.get("ok"):
+        return await _artifact_message(db, session_id, paper, "已按你的意见修改，文献编号已重排。")
+    blocks = [_text_block("修改后重新生成失败，请重试。")]
+    await _write_message(db, session_id, blocks=blocks, text="修改失败")
+    return blocks
+
+
+# ---------------------------------------------------------------------------
+# 写节循环（confirm/resume 后台任务）
+# ---------------------------------------------------------------------------
+
+async def run_writing_loop(db, paper_id: str, session_id: str) -> None:
+    """后台逐节写节循环：写完→complete→get_paper→写 artifact→artifact 块。"""
+    from ...routers.docgen import insert_artifact  # 惰性导入，避免服务层→路由层常驻耦合
+
+    client = _new_client()
+    await _write_message(db, session_id, blocks=[
+        {"type": "paper_progress", "paper_id": paper_id, "status": "writing", "sections": [], "current_section_id": ""},
+    ], text="开始撰写")
+    try:
+        while True:
+            prog = await client.call_tool("get_paper_progress", {"paper_id": paper_id})
+            if not prog.get("ok"):
+                await _write_message(db, session_id, blocks=[_text_block("撰写进度查询失败。")], text="撰写失败")
+                return
+            sections = prog.get("sections") or []
+            pending = [s for s in sections if s.get("status") != "done"]
+            if not pending:
+                break
+            sec = pending[0]
+            res = await client.call_tool("write_paper_section", {"paper_id": paper_id, "section_id": sec["section_id"]})
+            if not res.get("ok"):
+                await _write_message(db, session_id, blocks=[_text_block(
+                    f"章节「{sec.get('title', '')}」撰写失败：{res.get('message', '')}"
+                )], text="撰写失败")
+                return
+            wc = len(res.get("markdown") or "")
+            await _write_message(db, session_id, blocks=[_text_block(
+                f"「{sec.get('title', '')}」已完成，约 {wc} 字。"
+            )], text=f"「{sec.get('title', '')}」已完成")
+
+        comp = await client.call_tool("complete_paper", {"paper_id": paper_id})
+        if not comp.get("ok"):
+            await _write_message(db, session_id, blocks=[_text_block(f"论文合并失败：{comp.get('message', '')}")], text="合并失败")
+            return
+        paper = await client.call_tool("get_paper", {"paper_id": paper_id})
+        if not paper.get("ok"):
+            await _write_message(db, session_id, blocks=[_text_block("论文拉取失败。")], text="拉取失败")
+            return
+        await _artifact_message(db, session_id, paper, "论文撰写完成！你可以在右侧文档页查看全文，或用自然语言提出修改意见。")
+    except PaperMcpError as exc:
+        logger.warning("paper: 写节循环失败：%s", exc)
+        await _write_message(db, session_id, blocks=[_text_block(f"论文服务异常：{exc}")], text="论文服务异常")
+    except Exception as exc:  # noqa: BLE001
+        logger.exception("paper: 写节循环异常")
+        await _write_message(db, session_id, blocks=[_text_block(f"撰写失败：{exc}")], text="撰写失败")
+
+
+async def _artifact_message(db, session_id: str, paper: dict[str, Any], nl: str) -> list[dict[str, Any]]:
+    from ...routers.docgen import insert_artifact  # 惰性导入
+
+    refs = paper.get("references") or []
+    artifact_id = await insert_artifact(
+        db, session_id=session_id, title=paper.get("title", ""),
+        markdown=paper.get("full_markdown", ""), template_name="论文",
+    )
+    blocks = [
+        _text_block(nl),
+        {"type": "artifact", "artifact_id": artifact_id, "title": paper.get("title", ""),
+         "template_name": "论文", "status": "done", "paper_id": paper.get("paper_id", ""),
+         "citation_count": len(refs)},
+    ]
+    await _write_message(db, session_id, blocks=blocks, text=nl)
+    return blocks
diff --git a/sdk/app/services/paper/mcp_client.py b/sdk/app/services/paper/mcp_client.py
new file mode 100644
index 0000000..ed71e14
--- /dev/null
+++ b/sdk/app/services/paper/mcp_client.py
@@ -0,0 +1,104 @@
+"""services/paper/mcp_client.py —— 裸 JSON-RPC over httpx 客户端（BFF pyproject 无 mcp 依赖，零新增）。
+
+与 paper_writer_service /mcp/ 端点通信：initialize → tools/call。
+响应为 SSE（text/event-stream）：`event: message\ndata: <json>\n\n`；会话由 mcp-session-id 头维系。
+端点注意：/mcp 无斜杠会 307 重定向，必须请求 /mcp/。
+"""
+from __future__ import annotations
+
+import json
+import logging
+from typing import Any
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+PROTOCOL_VERSION = "2025-03-26"
+
+
+class PaperMcpError(Exception):
+    """paper 服务 RPC/工具调用失败。"""
+
+    def __init__(self, message: str, *, cause: BaseException | None = None) -> None:
+        super().__init__(message)
+        self.cause = cause
+
+
+def _parse_sse_body(text: str) -> dict[str, Any] | None:
+    """从 SSE 响应抽取 `data:` 行的 JSON；非 SSE 则整段按 JSON 解析。"""
+    for line in text.splitlines():
+        line = line.strip()
+        if line.startswith("data:"):
+            try:
+                return json.loads(line[5:].strip())
+            except ValueError:
+                return None
+    try:
+        return json.loads(text)
+    except ValueError:
+        return None
+
+
+class PaperMcpClient:
+    """paper 服务 MCP 客户端：维护 mcp-session-id，逐个 JSON-RPC 调用。"""
+
+    def __init__(self, base_url: str, timeout: float = 360.0) -> None:
+        self.base_url = (base_url or "http://127.0.0.1:8097").rstrip("/")
+        self.timeout = timeout
+        self._session_id: str | None = None
+        self._rpc_id = 0
+        self._initialized = False
+
+    async def _post(self, method: str, params: dict[str, Any]) -> dict[str, Any]:
+        self._rpc_id += 1
+        body = {"jsonrpc": "2.0", "id": self._rpc_id, "method": method, "params": params}
+        headers = {
+            "Content-Type": "application/json",
+            "Accept": "application/json, text/event-stream",
+        }
+        if self._session_id:
+            headers["mcp-session-id"] = self._session_id
+        async with httpx.AsyncClient(timeout=self.timeout) as client:
+            resp = await client.post(f"{self.base_url}/mcp/", headers=headers, json=body)
+            resp.raise_for_status()
+            sid = resp.headers.get("mcp-session-id")
+            if sid:
+                self._session_id = sid
+            text = resp.text
+        payload = _parse_sse_body(text)
+        if payload is None:
+            raise PaperMcpError(f"paper 服务响应无法解析：{text[:200]}")
+        if payload.get("error"):
+            raise PaperMcpError(f"paper 服务 RPC 错误：{payload['error']}")
+        return payload.get("result") or {}
+
+    async def initialize(self) -> dict[str, Any]:
+        return await self._post("initialize", {
+            "protocolVersion": PROTOCOL_VERSION,
+            "capabilities": {},
+            "clientInfo": {"name": "plantai-bff-paper", "version": "0.1.0"},
+        })
+
+    async def _ensure_initialized(self) -> None:
+        """MCP streamable-http 要求先 initialize（建立会话 + 取 mcp-session-id），
+        否则 tools/call 会被服务端 400 拒绝。"""
+        if not self._initialized:
+            await self.initialize()
+            self._initialized = True
+
+    async def call_tool(self, name: str, arguments: dict[str, Any]) -> dict[str, Any]:
+        await self._ensure_initialized()
+        result = await self._post("tools/call", {"name": name, "arguments": arguments})
+        if result.get("isError"):
+            raise PaperMcpError(f"工具 {name} 返回错误：{result}")
+        sc = result.get("structuredContent")
+        if sc is not None:
+            return sc
+        for block in result.get("content") or []:
+            if isinstance(block, dict) and block.get("type") == "text" and block.get("text"):
+                try:
+                    return json.loads(block["text"])
+                except ValueError:
+                    return {"raw": block["text"]}
+        return result
diff --git a/tests/paper/conftest.py b/tests/paper/conftest.py
new file mode 100644
index 0000000..139a5f3
--- /dev/null
+++ b/tests/paper/conftest.py
@@ -0,0 +1,80 @@
+"""tests/paper/conftest.py —— 论文服务测试共享夹具（tmp db + 假 LLM + 假检索）。
+
+全部 mock LLM/vstore，不依赖生产服务（8090/8094 等绝不探测）。
+镜像 bioflow_service/tests/conftest.py 的 _run（新建事件循环）+ 临时库模式。
+"""
+from __future__ import annotations
+
+import asyncio
+import sys
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).resolve().parent.parent.parent
+for _p in (str(REPO_ROOT), str(REPO_ROOT / "sdk")):
+    if _p not in sys.path:
+        sys.path.insert(0, _p)
+
+from paper_writer_service.app import db as dbmod  # noqa: E402
+from paper_writer_service.app.services.paper import llm as llmmod  # noqa: E402
+from paper_writer_service.app.services.paper import search as searchmod  # noqa: E402
+
+
+def _run(coro):
+    """同步跑一个协程（新建 event loop；bioflow/docgen 测试同款可靠模式）。"""
+    return asyncio.new_event_loop().run_until_complete(coro)
+
+
+class FakeLLM:
+    """可编程假 LLM：按顺序返回预设 dict；或抛 LLMError。记录全部调用。"""
+
+    def __init__(self, responses=None, error=None):
+        self.responses = list(responses or [])
+        self.error = error
+        self.calls: list[dict] = []
+
+    async def complete_json(self, system, user, **kwargs):
+        self.calls.append({"system": system, "user": user})
+        if self.error:
+            raise llmmod.LLMError(self.error)
+        return self.responses.pop(0) if self.responses else {}
+
+    async def complete_text(self, system, user, **kwargs):
+        self.calls.append({"system": system, "user": user})
+        if self.error:
+            raise llmmod.LLMError(self.error)
+        return self.responses.pop(0) if self.responses else ""
+
+
+# ---------------------------------------------------------------------------
+# DB：单例改路径重连（服务模块以 from ...db import db 持有单例引用）
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def paper_db(tmp_path):
+    dbmod.db._path = Path(tmp_path) / "paper.db"
+    _run(dbmod.db.connect())
+    yield dbmod.db
+    _run(dbmod.db.close())
+
+
+@pytest.fixture
+def fake_llm(monkeypatch):
+    fake = FakeLLM()
+    monkeypatch.setattr(llmmod, "complete_json", fake.complete_json)
+    monkeypatch.setattr(llmmod, "complete_text", fake.complete_text)
+    return fake
+
+
+@pytest.fixture
+def fake_search(monkeypatch):
+    async def _search_refs(paper_id="", query="", top_k=8):
+        return {"ok": True, "items": [
+            {"doc_id": "d1", "title": "水稻SPL基因家族研究进展", "snippet": "SPL 是植物特有转录因子家族。", "score": 0.9, "meta": {}},
+            {"doc_id": "d2", "title": "植物SPL转录因子功能", "snippet": "SPL 调控开花与株型。", "score": 0.8, "meta": {}},
+        ]}
+
+    monkeypatch.setattr(searchmod, "search_refs", _search_refs)
+    return _search_refs
diff --git a/tests/paper/manual_acceptance.md b/tests/paper/manual_acceptance.md
new file mode 100644
index 0000000..f6848de
--- /dev/null
+++ b/tests/paper/manual_acceptance.md
@@ -0,0 +1,33 @@
+# 论文撰写功能 手工验收手册
+
+## 前置（175 服务器，dev 树明文）
+
+- `paper_writer_service` 8097 已拉起（`bash deploy/paper_restart.sh`）。
+- BFF 8090 已启用论文路由（`bash deploy/toggle_paper.sh on`，即 `.env` 里 `PAPER_WRITER_ENABLED=true`）。
+- 前端 Vite dev（5180）或 BFF 静态托管 dist 就绪，Mac 侧走 SSH 隧道访问。
+
+探活自检：
+
+```bash
+curl -s http://127.0.0.1:8097/health                     # {"ok":true,...,"db":{"connected":true}}
+curl -s http://127.0.0.1:8090/api/paper/health            # 应透传 8097 /health
+curl -s http://127.0.0.1:8090/api/paper/papers/nope       # {"ok":false,...,"code":"NOT_FOUND"}（路由已注册）
+```
+
+## 验收剧本（10 步）
+
+1. **触发**：输入「帮我写一篇水稻 SPL 基因家族的综述论文」→ 出现 NL 引导 + `paper_brief` 表单卡（预填已抽取字段，缺失项待补）。
+2. **补全 brief**：类型=综述 / 篇幅=5000字 / 语言=中文 / 格式=GB/T 7714 → 点「生成大纲」→ 出现 `paper_outline` 大纲卡（5-6 节）。
+3. **NL 改大纲**：输入「把展望合并到讨论」→ 系统复述确认并更新大纲卡。
+4. **确认**：点「确认并开始撰写」→ 出现 `paper_progress` 进度卡，逐节推进，每节完成有 NL 小结。
+5. **完成**：全部节完成后自动挂 `artifact` 块并切右侧「文档」TAB；全文含文末 GB/T 7714 参考文献列表 + 降级 disclaimer（无 DOI/PMID 的知识库资料）。
+6. **引用点击**：点击正文 `[3]` → popover 显示标题 / GB 文本 / 来源片段。
+7. **修改**：输入「把讨论部分第二段改得更严谨」→ 修改确认，文献编号已重排。
+8. **导出**：文档 TAB 点「导出 PDF」/「导出 DOCX」→ 成功下载。
+9. **关服务降级**：`bash deploy/paper_restart.sh` 停掉后（或 kill 8097）→ 前端触发词落 docgen 单发（无 paper 入口）。
+10. **恢复续写**：重启 paper 服务 + BFF 后 resume → 跳过已完成节继续。
+
+## 备注
+
+- 引用为知识库（vstore 8094）自动检索注入；知识库 chunk 无 DOI/PMID 元数据，故多为降级格式 + disclaimer（见方案 §3.3）。
+- 长文一致性：citations 全局只增不改，complete 按首次出现顺序重排编号；正文与文末列表同源生成。
diff --git a/tests/paper/pytest.ini b/tests/paper/pytest.ini
new file mode 100644
index 0000000..ddc06ec
--- /dev/null
+++ b/tests/paper/pytest.ini
@@ -0,0 +1,5 @@
+[pytest]
+python_files = test_*.py
+markers =
+    paper_unit: 论文服务单元测试（引用协议/大纲/写节/重排/complete）
+    paper_mcp: MCP 协议层测试（list_tools/call_tool）
diff --git a/tests/paper/run_all.sh b/tests/paper/run_all.sh
new file mode 100755
index 0000000..50efff9
--- /dev/null
+++ b/tests/paper/run_all.sh
@@ -0,0 +1,12 @@
+#!/bin/bash
+# tests/paper/run_all.sh —— 论文撰写功能测试总入口（后端 pytest + 前端 intent 单测）
+set -uo pipefail
+ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
+PY="${PAPER_TEST_PY:-/home/xzq/5090/teamWorkspace/zh/plantai-rag-new-venv/bin/python}"
+
+echo "== 论文后端（pytest：工具逻辑 + MCP 协议 + BFF 状态机 + 块 schema） =="
+"$PY" -m pytest "$ROOT/tests/paper/" -c "$ROOT/tests/paper/pytest.ini" -q
+
+echo
+echo "== 论文前端 intent 单测（node） =="
+node "$ROOT/tests/paper/test_paper_intent_frontend.mjs"
diff --git a/tests/paper/test_mcp_protocol.py b/tests/paper/test_mcp_protocol.py
new file mode 100644
index 0000000..eed6dff
--- /dev/null
+++ b/tests/paper/test_mcp_protocol.py
@@ -0,0 +1,69 @@
+"""tests/paper/test_mcp_protocol.py —— MCP 协议层测试（list_tools / call_tool 往返）。
+
+验证：9 个工具正确暴露、call_tool 返回可解析的 JSON（structuredContent 或 content 文本块）。
+镜像 bioflow_service/tests/test_mcp_server.py 的 call_tool 集成模式。
+"""
+from __future__ import annotations
+
+import json
+
+from conftest import _run
+from paper_writer_service.app.mcp import paper_mcp
+
+EXPECTED_TOOLS = {
+    "propose_paper_outline", "write_paper_section", "get_paper_progress", "search_refs",
+    "revise_paper_section", "revise_paper", "get_paper", "complete_paper", "cancel_paper",
+}
+
+
+def _content_text(result):
+    """从 CallToolResult 抽取文本内容（无 structuredContent 时）。"""
+    for block in getattr(result, "content", None) or []:
+        text = getattr(block, "text", None)
+        if text:
+            return json.loads(text)
+    return {}
+
+
+def _call(name, args):
+    r = _run(paper_mcp.call_tool(name, args))
+    return getattr(r, "structuredContent", None) or _content_text(r)
+
+
+def test_list_tools_exposes_nine():
+    tools = _run(paper_mcp.list_tools())
+    names = {getattr(t, "name", None) for t in tools}
+    assert EXPECTED_TOOLS == names
+
+
+def test_call_tool_propose_outline_roundtrip(paper_db, fake_llm, fake_search):
+    fake_llm.responses.append({
+        "title": "水稻SPL基因家族研究进展",
+        "abstract_expect": "综述",
+        "sections": [{"section_id": "s1", "title": "引言", "key_points": [], "target_words": 1500}],
+    })
+    r = _call("propose_paper_outline", {"topic": "水稻SPL基因家族", "paper_type": "综述"})
+    assert r.get("ok") is True
+    assert r["paper_id"]
+    assert r["title"] == "水稻SPL基因家族研究进展"
+
+
+def test_call_tool_not_found(paper_db):
+    r = _call("get_paper", {"paper_id": "nope"})
+    assert r.get("ok") is False
+    assert r.get("code") == "NOT_FOUND"
+
+
+def test_call_tool_content_is_json_string(paper_db, fake_llm, fake_search):
+    """验证 content[0].text 是合法 JSON 字符串（BFF 裸 JSON-RPC 客户端解析面）。"""
+    fake_llm.responses.append({
+        "title": "T", "abstract_expect": "A",
+        "sections": [{"section_id": "s1", "title": "引言", "key_points": [], "target_words": 1500}],
+    })
+    raw = _run(paper_mcp.call_tool("propose_paper_outline", {"topic": "水稻SPL"}))
+    blocks = getattr(raw, "content", None) or []
+    assert blocks, "CallToolResult 无 content 块"
+    text = getattr(blocks[0], "text", None)
+    assert text is not None
+    payload = json.loads(text)  # 应为合法 JSON 字符串
+    assert payload.get("ok") is True
diff --git a/tests/paper/test_paper_bff_agent.py b/tests/paper/test_paper_bff_agent.py
new file mode 100644
index 0000000..f567881
--- /dev/null
+++ b/tests/paper/test_paper_bff_agent.py
@@ -0,0 +1,123 @@
+"""tests/paper/test_paper_bff_agent.py —— BFF 论文大脑状态机测试（假 MCP server + 假 LLM）。
+
+跑通 brief→outline→confirm(写节循环)→complete→artifact 全状态机。
+"""
+from __future__ import annotations
+
+import pytest
+
+from conftest import _run
+from app import db as bff_db
+from app.routers import docgen as docgen_mod
+from app.services.paper import agent
+
+
+class FakePaperService:
+    """脚本化 paper 服务：按工具名返回预设结果，模拟写节循环推进。"""
+
+    def __init__(self):
+        self.paper_id = "p1"
+        self.s1_done = False
+        self.s2_done = False
+
+    async def call_tool(self, name, arguments):
+        if name == "propose_paper_outline":
+            return {"ok": True, "paper_id": self.paper_id, "title": "水稻SPL基因家族研究进展",
+                    "abstract_expect": "综述", "sections": [
+                        {"section_id": "s1", "title": "引言", "key_points": [], "target_words": 1500},
+                        {"section_id": "s2", "title": "讨论", "key_points": [], "target_words": 1500},
+                    ]}
+        if name == "get_paper_progress":
+            sections = [
+                {"section_id": "s1", "title": "引言", "status": "done" if self.s1_done else "pending",
+                 "word_count": 0, "citation_count": 0},
+                {"section_id": "s2", "title": "讨论", "status": "done" if self.s2_done else "pending",
+                 "word_count": 0, "citation_count": 0},
+            ]
+            return {"ok": True, "paper_id": self.paper_id, "status": "writing",
+                    "sections": sections, "current_section_id": None}
+        if name == "write_paper_section":
+            sid = arguments["section_id"]
+            if sid == "s1":
+                self.s1_done = True
+            else:
+                self.s2_done = True
+            return {"ok": True, "section_id": sid, "status": "done",
+                    "markdown": "内容" * 50, "citations_used": [], "new_citations": []}
+        if name == "complete_paper":
+            return {"ok": True, "full_markdown": "# 全文\n内容",
+                    "references": [{"n": 1, "text": "X.", "kind": "title", "source_doc_id": "d1"}],
+                    "renumber_map": {}, "disclaimer": "d"}
+        if name == "get_paper":
+            return {"ok": True, "paper_id": self.paper_id, "title": "水稻SPL基因家族研究进展",
+                    "status": "complete", "full_markdown": "# 全文",
+                    "references": [{"n": 1, "text": "X.", "kind": "title", "source_doc_id": "d1"}],
+                    "disclaimer": "d"}
+        if name == "cancel_paper":
+            return {"ok": True}
+        if name == "revise_paper":
+            return {"ok": True, "affected_section_ids": ["s1"], "plan": "修订引言"}
+        if name == "revise_paper_section":
+            return {"ok": True, "section_id": arguments["section_id"], "markdown": "改后", "citations_used": []}
+        raise AssertionError(f"unexpected tool: {name}")
+
+
+@pytest.fixture
+def bff_database(tmp_path):
+    _run(bff_db.init_db(tmp_path / "bff.db"))
+    _run(bff_db.db.execute(
+        "INSERT INTO sessions (id, title, status, meta, created_at, updated_at) "
+        "VALUES ('s1', 't', 'idle', '{}', '2026-01-01T00:00:00Z', '2026-01-01T00:00:00Z')"))
+    yield bff_db.db
+    _run(bff_db.close_db())
+
+
+def _fake_brief(complete=True):
+    async def _complete_json(system, user, **kwargs):
+        if complete:
+            return {"topic": "水稻SPL基因家族", "paper_type": "综述", "language": "中文",
+                    "length": "5000字", "format_req": "GB/T 7714", "style": "学术严谨", "requirements": ""}
+        return {"topic": "水稻SPL", "paper_type": "", "language": "", "length": ""}
+    return _complete_json
+
+
+@pytest.fixture
+def fake_insert_artifact(monkeypatch):
+    async def _fake(db, **kwargs):
+        return "art-fake"
+    monkeypatch.setattr(docgen_mod, "insert_artifact", _fake)
+    return _fake
+
+
+def test_trigger_complete_brief_proposes_outline(bff_database, monkeypatch, fake_insert_artifact):
+    svc = FakePaperService()
+    monkeypatch.setattr(agent, "_new_client", lambda: svc)
+    monkeypatch.setattr(agent, "llm_complete_json", _fake_brief(complete=True))
+
+    blocks = _run(agent.handle_turn(bff_database, "s1", "帮我写一篇水稻SPL基因家族综述"))
+    assert blocks[-1]["type"] == "paper_outline"
+    assert blocks[-1]["paper_id"] == "p1"
+
+    rows = _run(bff_database.fetchall("SELECT * FROM messages WHERE session_id = 's1'"))
+    assert len(rows) == 1
+    assert rows[0]["kind"] == "paper"
+
+
+def test_trigger_incomplete_brief_returns_form(bff_database, monkeypatch):
+    monkeypatch.setattr(agent, "_new_client", lambda: FakePaperService())
+    monkeypatch.setattr(agent, "llm_complete_json", _fake_brief(complete=False))
+    blocks = _run(agent.handle_turn(bff_database, "s1", "帮我写论文"))
+    assert blocks[-1]["type"] == "paper_brief"
+    assert blocks[-1]["status"] == "pending"
+
+
+def test_writing_loop_completes_to_artifact(bff_database, monkeypatch, fake_insert_artifact):
+    svc = FakePaperService()
+    monkeypatch.setattr(agent, "_new_client", lambda: svc)
+    _run(agent.run_writing_loop(bff_database, "p1", "s1"))
+
+    assert svc.s1_done and svc.s2_done
+    rows = _run(bff_database.fetchall("SELECT * FROM messages WHERE session_id = 's1' ORDER BY rowid"))
+    assert "paper_progress" in rows[0]["blocks"]
+    assert "artifact" in rows[-1]["blocks"]
+    assert "art-fake" in rows[-1]["blocks"]
diff --git a/tests/paper/test_paper_blocks_schema.py b/tests/paper/test_paper_blocks_schema.py
new file mode 100644
index 0000000..3abc1b0
--- /dev/null
+++ b/tests/paper/test_paper_blocks_schema.py
@@ -0,0 +1,35 @@
+"""tests/paper/test_paper_blocks_schema.py —— 块 schema 回归（未知块拒绝 + 新块通过）。"""
+from __future__ import annotations
+
+import pytest
+from pydantic import TypeAdapter
+
+from app import schemas
+
+_BLOCK = TypeAdapter(schemas.Block)
+
+
+def test_paper_blocks_validate():
+    cases = [
+        {"type": "paper_brief", "paper_id": "p1", "status": "pending", "fields": []},
+        {"type": "paper_outline", "paper_id": "p1", "title": "t", "status": "pending", "sections": []},
+        {"type": "paper_progress", "paper_id": "p1", "status": "writing", "sections": [], "current_section_id": ""},
+        {"type": "paper_refs", "paper_id": "p1", "references": [], "disclaimer": ""},
+    ]
+    for c in cases:
+        b = _BLOCK.validate_python(c)
+        assert b.type == c["type"]
+
+
+def test_artifact_block_paper_fields():
+    b = _BLOCK.validate_python({
+        "type": "artifact", "artifact_id": "a1", "title": "t",
+        "paper_id": "p1", "citation_count": 3,
+    })
+    assert b.paper_id == "p1"
+    assert b.citation_count == 3
+
+
+def test_unknown_block_rejected():
+    with pytest.raises(Exception):
+        _BLOCK.validate_python({"type": "totally_unknown_block"})
diff --git a/tests/paper/test_paper_intent_frontend.mjs b/tests/paper/test_paper_intent_frontend.mjs
new file mode 100644
index 0000000..d3e263d
--- /dev/null
+++ b/tests/paper/test_paper_intent_frontend.mjs
@@ -0,0 +1,42 @@
+// tests/paper/test_paper_intent_frontend.mjs —— 论文撰写意图前端单测（node 直测）
+// 验证 paperIntent 纯函数 + route() 零变化抽样（30 条全量由 test_intent_frontend.mjs 兜底）。
+import { pathToFileURL } from 'node:url';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const here = path.dirname(fileURLToPath(import.meta.url));
+const repo = path.resolve(here, '../..');
+const { paperIntent, route } = await import(pathToFileURL(path.join(repo, 'ui/src/intent.js')).href);
+
+const cases = [
+  ['帮我写一篇水稻 SPL 基因家族综述论文', true],
+  ['撰写一篇关于小麦育种的论文', true],
+  ['拟一份综述', true],
+  ['帮我写一个关于疫情防控的通知', false],
+  ['写一份项目总结报告', false],
+  ['写一首诗', false],
+  ['水稻基因是什么？', false],
+  ['', false],
+];
+let fail = 0;
+for (const [t, exp] of cases) {
+  const got = paperIntent(t);
+  const ok = got === exp;
+  if (!ok) fail++;
+  console.log(`${ok ? 'PASS' : 'FAIL'} paperIntent(${JSON.stringify(t)}) = ${got}（期望 ${exp}）`);
+}
+
+const routeChecks = [
+  ['给我写一篇关于小麦育种的论文', false, false, 'docgen'],
+  ['写一段水稻基因的介绍', false, false, 'qa'],
+];
+for (const [t, a, b, exp] of routeChecks) {
+  const got = route(t, a, b);
+  const ok = got === exp;
+  if (!ok) fail++;
+  console.log(`${ok ? 'PASS' : 'FAIL'} route(${JSON.stringify(t)}) = ${got}（期望 ${exp}）`);
+}
+
+const total = cases.length + routeChecks.length;
+console.log(`===== paperIntent 前端单测：${total - fail}/${total} PASS =====`);
+process.exit(fail === 0 ? 0 : 1);
diff --git a/tests/paper/test_paper_tools.py b/tests/paper/test_paper_tools.py
new file mode 100644
index 0000000..353a9c5
--- /dev/null
+++ b/tests/paper/test_paper_tools.py
@@ -0,0 +1,256 @@
+"""tests/paper/test_paper_tools.py —— 论文服务工具逻辑单元测试（假 LLM + 假检索）。
+
+覆盖：propose_paper_outline / write_paper_section（含引用校验重试）/ get_paper_progress /
+revise_paper_section / revise_paper / get_paper / complete_paper（重排 + disclaimer）/ cancel_paper +
+引用协议纯函数（GB/T 7714 渲染、[n] 重排）。
+"""
+from __future__ import annotations
+
+import pytest
+
+from conftest import _run
+from paper_writer_service.app import db as dbmod
+from paper_writer_service.app.services.paper import cite, complete, outline, write
+
+OUTLINE_RES = {
+    "title": "水稻SPL基因家族研究进展",
+    "abstract_expect": "综述水稻SPL基因家族的结构、功能与调控机制。",
+    "sections": [
+        {"section_id": "s1", "title": "引言", "key_points": ["SPL家族概述"], "target_words": 1500},
+        {"section_id": "s2", "title": "SPL家族结构与进化", "key_points": ["结构特征"], "target_words": 2000},
+        {"section_id": "s3", "title": "SPL功能与调控", "key_points": ["功能"], "target_words": 2000},
+        {"section_id": "s4", "title": "讨论与展望", "key_points": ["展望"], "target_words": 1500},
+        {"section_id": "s5", "title": "结论", "key_points": ["总结"], "target_words": 1000},
+    ],
+}
+
+
+def _propose(fake_llm, fake_search):
+    fake_llm.responses.append(dict(OUTLINE_RES))
+    return _run(outline.propose_paper_outline("水稻SPL基因家族", paper_type="综述", language="中文"))
+
+
+# ---------------------------------------------------------------------------
+# propose_paper_outline
+# ---------------------------------------------------------------------------
+
+def test_propose_outline_creates_paper_and_sections(paper_db, fake_llm, fake_search):
+    r = _propose(fake_llm, fake_search)
+    assert r["ok"] is True
+    assert r["title"] == "水稻SPL基因家族研究进展"
+    assert len(r["sections"]) == 5
+
+    papers = _run(dbmod.db.fetchall("SELECT * FROM papers"))
+    assert len(papers) == 1
+    secs = _run(dbmod.db.fetchall("SELECT * FROM paper_sections ORDER BY rowid"))
+    assert [s["section_id"] for s in secs] == ["s1", "s2", "s3", "s4", "s5"]
+    assert all(s["status"] == "pending" for s in secs)
+
+
+def test_propose_outline_llm_error(paper_db, fake_llm, fake_search):
+    fake_llm.error = "deepseek 超时"
+    r = _run(outline.propose_paper_outline("水稻SPL"))
+    assert r["ok"] is False
+    assert r["code"] == "LLM_ERROR"
+
+
+# ---------------------------------------------------------------------------
+# write_paper_section（引用校验 + 落库）
+# ---------------------------------------------------------------------------
+
+def test_write_section_creates_citations(paper_db, fake_llm, fake_search):
+    paper_id = _propose(fake_llm, fake_search)["paper_id"]
+    fake_llm.responses.append({
+        "markdown": "SPL 是植物特有的转录因子家族[1]，水稻有 19 个成员[2]。",
+        "citations_used": [
+            {"temp_n": 1, "title": "水稻SPL基因家族研究进展", "journal": "植物学报", "year": "2024",
+             "doi": "", "pmid": "", "source_doc_id": "d1", "snippet": "x"},
+            {"temp_n": 2, "title": "植物SPL转录因子功能", "journal": "", "year": "2023",
+             "doi": "", "pmid": "", "source_doc_id": "d2", "snippet": "y"},
+        ],
+    })
+    r = _run(write.write_paper_section(paper_id, "s1"))
+    assert r["ok"] is True
+    assert r["status"] == "done"
+    assert len(r["new_citations"]) == 2
+
+    cits = _run(dbmod.db.fetchall("SELECT * FROM citations ORDER BY temp_n"))
+    assert [c["temp_n"] for c in cits] == [1, 2]
+    assert cits[0]["kind"] == "title"
+    sec = _run(dbmod.db.fetchone("SELECT * FROM paper_sections WHERE paper_id = ? AND section_id = 's1'", (paper_id,)))
+    assert sec["status"] == "done"
+    assert "[1]" in sec["markdown"]
+
+
+def test_write_section_invalid_citation_retries_once(paper_db, fake_llm, fake_search):
+    paper_id = _propose(fake_llm, fake_search)["paper_id"]
+    fake_llm.calls.clear()  # 重置计数，只统计写节调用（大纲已消费 1 次）
+    # 第一次引用非法编号 [99]，第二次合法
+    fake_llm.responses.append({
+        "markdown": "SPL 家族[99]。",
+        "citations_used": [{"temp_n": 1, "title": "t", "source_doc_id": "d1"}],
+    })
+    fake_llm.responses.append({
+        "markdown": "SPL 家族[1]。",
+        "citations_used": [{"temp_n": 1, "title": "t", "source_doc_id": "d1"}],
+    })
+    r = _run(write.write_paper_section(paper_id, "s1"))
+    assert r["ok"] is True
+    assert len(fake_llm.calls) == 2  # 一次失败 + 一次重试成功
+
+
+def test_write_section_invalid_citation_fails_after_retries(paper_db, fake_llm, fake_search):
+    paper_id = _propose(fake_llm, fake_search)["paper_id"]
+    fake_llm.responses.append({"markdown": "SPL 家族[99]。", "citations_used": []})
+    fake_llm.responses.append({"markdown": "SPL 家族[88]。", "citations_used": []})
+    r = _run(write.write_paper_section(paper_id, "s1"))
+    assert r["ok"] is False
+    assert r["code"] == "CITE_VALIDATION_FAILED"
+
+
+def test_write_section_not_found(paper_db, fake_llm):
+    r = _run(write.write_paper_section("nope", "s1"))
+    assert r["ok"] is False
+    assert r["code"] == "NOT_FOUND"
+
+
+# ---------------------------------------------------------------------------
+# get_paper_progress / get_paper
+# ---------------------------------------------------------------------------
+
+def test_progress_reflects_done_section(paper_db, fake_llm, fake_search):
+    paper_id = _propose(fake_llm, fake_search)["paper_id"]
+    fake_llm.responses.append({
+        "markdown": "正文[1]。",
+        "citations_used": [{"temp_n": 1, "title": "t", "source_doc_id": "d1"}],
+    })
+    _run(write.write_paper_section(paper_id, "s1"))
+    r = _run(complete.get_paper_progress(paper_id))
+    assert r["ok"] is True
+    assert r["status"] == "writing"
+    s1 = next(s for s in r["sections"] if s["section_id"] == "s1")
+    assert s1["status"] == "done"
+    assert s1["citation_count"] == 1
+
+
+# ---------------------------------------------------------------------------
+# revise_paper_section / revise_paper
+# ---------------------------------------------------------------------------
+
+def test_revise_section_appends_new_citation(paper_db, fake_llm, fake_search):
+    paper_id = _propose(fake_llm, fake_search)["paper_id"]
+    fake_llm.responses.append({
+        "markdown": "旧正文[1]。",
+        "citations_used": [{"temp_n": 1, "title": "旧文献", "source_doc_id": "d1"}],
+    })
+    _run(write.write_paper_section(paper_id, "s1"))
+    # 修订：新增 [2]
+    fake_llm.responses.append({
+        "markdown": "新正文[1][2]。",
+        "citations_used": [
+            {"temp_n": 1, "title": "旧文献", "source_doc_id": "d1"},
+            {"temp_n": 2, "title": "新文献", "source_doc_id": "d3"},
+        ],
+    })
+    r = _run(write.revise_paper_section(paper_id, "s1", "更严谨"))
+    assert r["ok"] is True
+    assert r["citation_delta"] == 1
+    cits = _run(dbmod.db.fetchall("SELECT temp_n FROM citations ORDER BY temp_n"))
+    assert [c["temp_n"] for c in cits] == [1, 2]
+
+
+def test_revise_paper_returns_affected(paper_db, fake_llm, fake_search):
+    paper_id = _propose(fake_llm, fake_search)["paper_id"]
+    fake_llm.responses.append({"affected_section_ids": ["s2", "s3"], "plan": "修订讨论与结论"})
+    r = _run(write.revise_paper(paper_id, "把讨论改严谨"))
+    assert r["ok"] is True
+    assert r["affected_section_ids"] == ["s2", "s3"]
+
+
+# ---------------------------------------------------------------------------
+# complete_paper（重排 + GB/T 7714 + disclaimer）
+# ---------------------------------------------------------------------------
+
+def test_complete_renumbers_by_first_appearance(paper_db, fake_llm, fake_search):
+    paper_id = _propose(fake_llm, fake_search)["paper_id"]
+    # s1 先引用 [2]，s2 后引用 [1] → complete 后 [2] 应重排为 [1]
+    fake_llm.responses.append({
+        "markdown": "引言提到 B[2]。",
+        "citations_used": [{"temp_n": 2, "title": "文献B", "source_doc_id": "d2"}],
+    })
+    _run(write.write_paper_section(paper_id, "s1"))
+    fake_llm.responses.append({
+        "markdown": "正文提到 A[1]。",
+        "citations_used": [{"temp_n": 1, "title": "文献A", "source_doc_id": "d1"}],
+    })
+    _run(write.write_paper_section(paper_id, "s2"))
+
+    r = _run(complete.complete_paper(paper_id))
+    assert r["ok"] is True
+    # 首次出现顺序：B(temp 2) 在 s1，A(temp 1) 在 s2 → 2→1, 1→2
+    assert r["renumber_map"] == {2: 1, 1: 2}
+    assert r["references"][0]["n"] == 1
+    assert r["references"][0]["text"].startswith("文献B")
+    assert r["disclaimer"] != ""  # title kind → 降级 disclaimer
+
+    paper = _run(dbmod.db.fetchone("SELECT * FROM papers WHERE id = ?", (paper_id,)))
+    assert paper["status"] == "complete"
+    assert "[1]" in paper["final_markdown"] and "[2]" in paper["final_markdown"]
+
+
+def test_complete_paper_no_citations_no_disclaimer(paper_db, fake_llm, fake_search):
+    paper_id = _propose(fake_llm, fake_search)["paper_id"]
+    fake_llm.responses.append({"markdown": "无引用正文。", "citations_used": []})
+    _run(write.write_paper_section(paper_id, "s1"))
+    r = _run(complete.complete_paper(paper_id))
+    assert r["ok"] is True
+    assert r["references"] == []
+    assert r["disclaimer"] == ""
+
+
+def test_get_paper_returns_full_markdown(paper_db, fake_llm, fake_search):
+    paper_id = _propose(fake_llm, fake_search)["paper_id"]
+    fake_llm.responses.append({"markdown": "正文[1]。", "citations_used": [
+        {"temp_n": 1, "title": "t", "source_doc_id": "d1"}]})
+    _run(write.write_paper_section(paper_id, "s1"))
+    _run(complete.complete_paper(paper_id))
+    r = _run(complete.get_paper(paper_id))
+    assert r["ok"] is True
+    assert r["status"] == "complete"
+    assert "水稻SPL基因家族研究进展" in r["full_markdown"]
+    assert len(r["references"]) == 1
+
+
+# ---------------------------------------------------------------------------
+# cancel_paper
+# ---------------------------------------------------------------------------
+
+def test_cancel_paper(paper_db, fake_llm, fake_search):
+    paper_id = _propose(fake_llm, fake_search)["paper_id"]
+    r = _run(complete.cancel_paper(paper_id))
+    assert r["ok"] is True
+    paper = _run(dbmod.db.fetchone("SELECT status FROM papers WHERE id = ?", (paper_id,)))
+    assert paper["status"] == "cancelled"
+
+
+def test_cancel_paper_not_found(paper_db):
+    r = _run(complete.cancel_paper("nope"))
+    assert r["ok"] is False
+    assert r["code"] == "NOT_FOUND"
+
+
+# ---------------------------------------------------------------------------
+# 引用协议纯函数（cite）
+# ---------------------------------------------------------------------------
+
+def test_gbt7714_doi_pmid_title():
+    assert "DOI:10.1000/x" in cite.render_gbt7714({"title": "T", "journal": "J", "year": "2024", "doi": "10.1000/x"})
+    assert "PMID:123" in cite.render_gbt7714({"title": "T", "journal": "J", "year": "2024", "pmid": "123"})
+    degraded = cite.render_gbt7714({"title": "知识库标题", "year": "2023"})
+    assert degraded.endswith("知识库资料, 2023. [EB/OL].")
+
+
+def test_renumber_map_first_appearance():
+    mapping = cite.renumber_map([1, 2, 3], "见[3]和[1]，以及[2]")
+    assert mapping == {3: 1, 1: 2, 2: 3}
+    assert cite.replace_numbers("见[3]和[1]", {3: 1, 1: 2}) == "见[1]和[2]"
diff --git a/tooling/compile/compile_sdk.sh b/tooling/compile/compile_sdk.sh
index fdd6ad7..e7cc003 100755
--- a/tooling/compile/compile_sdk.sh
+++ b/tooling/compile/compile_sdk.sh
@@ -31,7 +31,7 @@ rsync -a --exclude '__pycache__' --exclude 'test_*.py' --exclude 'data' --exclud
   "$SOURCE_ROOT/" "$OUTPUT_DIR/"
 
 # 白名单明文:main.py + config.py + __init__.py(编译前从发现集排除)
-WHITELIST_PY=("app/main.py" "app/config.py" "app/schemas.py" "app/db.py" "app/routers/qa.py" "app/services/sse_relay.py" "app/routers/bioflow_proxy.py")
+WHITELIST_PY=("app/main.py" "app/config.py" "app/schemas.py" "app/db.py" "app/routers/qa.py" "app/services/sse_relay.py" "app/routers/bioflow_proxy.py" "app/routers/paper.py" "app/services/paper/agent.py" "app/services/paper/mcp_client.py")
 # 白名单理由:main.py=uvicorn 入口;config.py=运维可读面;
 # schemas.py=pydantic v2 模型(validator 方法被 Cython 转 cyfunction 触发
 # PydanticUserError,实测编译后无法导入——纯数据契约面,无敏感逻辑,明文保留);
diff --git a/ui/src/api/paper.js b/ui/src/api/paper.js
new file mode 100644
index 0000000..0b52825
--- /dev/null
+++ b/ui/src/api/paper.js
@@ -0,0 +1,34 @@
+// api/paper.js —— 论文撰写端点（paper-writer）
+// 全部经 http.js 封装。turns 短同步；confirm/resume 启动后台写节循环；
+// progress/citations 透传 paper 服务。
+import { get, post } from './http.js';
+
+/** 探活（前端 30s 轮询；离线时触发词落 docgen 单发降级） */
+export function paperHealth() {
+  return get('/api/paper/health');
+}
+
+/** 一轮对话（短同步）：POST /api/paper/turns → {ok, blocks} */
+export function paperTurn(payload) {
+  return post('/api/paper/turns', payload);
+}
+
+/** 确认大纲并启动后台写节循环 */
+export function confirmPaper(paperId, sessionId) {
+  return post(`/api/paper/papers/${encodeURIComponent(paperId)}/confirm`, { session_id: sessionId });
+}
+
+/** 断写后重启循环（跳过已完成节） */
+export function resumePaper(paperId, sessionId) {
+  return post(`/api/paper/papers/${encodeURIComponent(paperId)}/resume`, { session_id: sessionId });
+}
+
+/** 节级进度（写稿期 2s 轮询） */
+export function paperProgress(paperId) {
+  return get(`/api/paper/papers/${encodeURIComponent(paperId)}`);
+}
+
+/** 单条引用来源（正文 [n] 点击 popover） */
+export function paperCitation(paperId, n) {
+  return get(`/api/paper/papers/${encodeURIComponent(paperId)}/citations/${encodeURIComponent(n)}`);
+}
diff --git a/ui/src/components/bench/DocPane.vue b/ui/src/components/bench/DocPane.vue
index 3f3c74c..0c9a849 100644
--- a/ui/src/components/bench/DocPane.vue
+++ b/ui/src/components/bench/DocPane.vue
@@ -15,6 +15,8 @@ import { ref, computed, watch, nextTick, onMounted, onBeforeUnmount } from 'vue'
 import { useDocgenStore } from '../../stores/docgen.js';
 import { DG, BUILTIN_TEMPLATES } from '../../i18n/docgen.js';
 import { renderMarkdownPlain, sanitizeHtml } from '../chat/markdown.js';
+// 论文撰写：正文 [n] → cite-chip，点击拉 citation popover
+import { paperCitation } from '../../api/paper.js';
 // FR-PRIV-006：「我的资料」管理面板（挂空态区顶部，UI_SPEC §9.2）
 import PrivateLibraryPane from './PrivateLibraryPane.vue';
 
@@ -70,10 +72,66 @@ function mountShadow() {
     // 缺陷 #7b：md 回退预览不注入标题编号——序列化会把标题写回正本，编号不得进源文
     inner = `<div class="doc-body md-text">${renderMarkdownPlain(docgen.md)}</div>`;
   }
+  if (docgen.activeArtifact && docgen.activeArtifact.paper_id) {
+    inner = addCiteChips(inner); // 论文制品：[n] → cite-chip
+  }
   root.innerHTML = inner;
   if (editing.value) activateEditDom();
 }
 
+// ---------- 论文引用 cite-chip ----------
+
+function escapeHtml(s) {
+  return String(s == null ? '' : s)
+    .replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;')
+    .replace(/"/g, '&quot;').replace(/'/g, '&#39;');
+}
+
+function addCiteChips(inner) {
+  return inner.replace(/\[(\d+)\]/g, (m, n) =>
+    `<sup class="cite-chip" data-n="${n}" style="color:#2563eb;cursor:pointer;font-weight:600;margin:0 1px;">[${n}]</sup>`);
+}
+
+function removeCitePopover() {
+  const host = hostEl.value;
+  if (host) {
+    const p = host.querySelector(':scope > .cite-popover');
+    if (p) p.remove();
+  }
+}
+
+function showCitePopover(chip, c) {
+  removeCitePopover();
+  const host = hostEl.value;
+  if (!host || !chip) return;
+  const rect = chip.getBoundingClientRect();
+  const hostRect = host.getBoundingClientRect();
+  const pop = document.createElement('div');
+  pop.className = 'cite-popover';
+  pop.innerHTML = `<div class="cp-title">${escapeHtml(c.title || c.text || '')}</div>` +
+    `<div class="cp-text">${escapeHtml(c.text || '')}</div>` +
+    (c.snippet ? `<div class="cp-snippet">${escapeHtml(c.snippet)}</div>` : '');
+  pop.style.cssText = `position:absolute;left:${rect.left - hostRect.left}px;top:${rect.bottom - hostRect.top + 4}px;max-width:360px;background:#fff;border:1px solid #e5e7eb;border-radius:8px;box-shadow:0 4px 12px rgba(0,0,0,.12);padding:8px 10px;z-index:50;font-size:12px;`;
+  host.appendChild(pop);
+}
+
+function wireCiteChips() {
+  const root = shadowRoot();
+  if (!root || root._citeWired) return;
+  root._citeWired = true;
+  root.addEventListener('click', (e) => {
+    const chip = e.target && e.target.closest ? e.target.closest('.cite-chip') : null;
+    if (!chip) { removeCitePopover(); return; }
+    const n = chip.getAttribute('data-n');
+    const paperId = docgen.activeArtifact && docgen.activeArtifact.paper_id;
+    if (!paperId || !n) return;
+    paperCitation(paperId, n).then((data) => {
+      const c = data && data.citation;
+      if (c) showCitePopover(chip, c);
+    }).catch(() => {});
+  });
+}
+
 watch(() => [docgen.html, docgen.css, docgen.renderNonce], () => {
   // 2026-09-04 问题5：宿主元素可能随 loading 态被 v-if 卸载重建，
   // nextTick 后 mountShadow 才能拿到新宿主；renderNonce 兜底 html 拉取失败
@@ -90,7 +148,7 @@ watch(() => docgen.md, () => {
     if (mode.value === 'preview') mountShadow();
   });
 });
-onMounted(() => nextTick(mountShadow));
+onMounted(() => nextTick(() => { mountShadow(); wireCiteChips(); }));
 
 // ---------- HTML 所见即所得编辑 ----------
 
diff --git a/ui/src/components/chat/AssistantMessage.vue b/ui/src/components/chat/AssistantMessage.vue
index c627bd2..e257186 100644
--- a/ui/src/components/chat/AssistantMessage.vue
+++ b/ui/src/components/chat/AssistantMessage.vue
@@ -28,6 +28,11 @@ import ExtensionConfirmCard from './ExtensionConfirmCard.vue';
 import ExtensionRunCard from './ExtensionRunCard.vue';
 // FR-PRIV-005：私有资料参考块（独立 i18n 文件，不碰 labels.js 红线）
 import { PRIV } from '../../i18n/private.js';
+// 论文撰写：四类块（brief 表单卡 / outline 大纲卡 / progress 进度卡 / refs 内联列表）
+import { PP } from '../../i18n/paper.js';
+import PaperBriefCard from './PaperBriefCard.vue';
+import PaperOutlineCard from './PaperOutlineCard.vue';
+import PaperProgressCard from './PaperProgressCard.vue';
 
 const props = defineProps({
   message: { type: Object, required: true },
@@ -298,6 +303,21 @@ const degradedNote = computed(() => props.message.degradedNote || null);
         <!-- artifact：文档产物卡（P3；「打开预览」→ docgen store + bench 文档 TAB） -->
         <DocCard v-if="b.type === 'artifact'" :block="b" />
 
+        <!-- 论文撰写：brief 引导表单卡 -->
+        <PaperBriefCard v-if="b.type === 'paper_brief'" :block="b" />
+        <!-- 论文撰写：大纲确认卡 -->
+        <PaperOutlineCard v-if="b.type === 'paper_outline'" :block="b" />
+        <!-- 论文撰写：节级进度卡 -->
+        <PaperProgressCard v-if="b.type === 'paper_progress'" :block="b" />
+        <!-- 论文撰写：参考文献内联列表 -->
+        <div v-if="b.type === 'paper_refs'" class="paper-refs">
+          <div class="pr-title">{{ PP.paperRefsTitle }}</div>
+          <ol class="pr-list">
+            <li v-for="r in b.references" :key="r.n">[{{ r.n }}] {{ r.text }}</li>
+          </ol>
+          <div v-if="b.disclaimer" class="pr-disclaimer">{{ b.disclaimer }}</div>
+        </div>
+
         <!-- sources：内联引用 chip（FR-CHAT-015） -->
         <div v-if="b.type === 'sources' && b.items && b.items.length" class="inline-sources">
           <span class="src-label">{{ T.sourcesTitle }}</span>
diff --git a/ui/src/components/chat/Composer.vue b/ui/src/components/chat/Composer.vue
index 47422d2..6ca3741 100644
--- a/ui/src/components/chat/Composer.vue
+++ b/ui/src/components/chat/Composer.vue
@@ -16,13 +16,14 @@ import { ref, computed, onMounted } from 'vue';
 import { MODES as MODE_LABELS, T } from '../../i18n/labels.js';
 import { DG, BUILTIN_TEMPLATES } from '../../i18n/docgen.js';
 import { BF } from '../../i18n/bioflow.js';
-import { route } from '../../intent.js';
+import { route, paperIntent } from '../../intent.js';
 import { route as routeBioflow, isExtensionCreateIntent } from '../../utils/intent.js';
 import { uploadLibrary, listLibrary } from '../../api/docgen.js';
 import { uploadBioflowFile } from '../../api/bioflow.js';
 import { useDocgenStore } from '../../stores/docgen.js';
 import { useChatStore } from '../../stores/chat.js';
 import { useBioflowStore } from '../../stores/bioflow.js';
+import { usePaperStore } from '../../stores/paper.js';
 import { useSettingsStore } from '../../stores/settings.js';
 
 const props = defineProps({
@@ -59,6 +60,9 @@ onMounted(() => {
       settings.load(sid).catch(() => { /* 静默：失败不阻断（设置面板打开时会重试） */ });
     }
   } catch { /* settings store 未就绪（测试/降级）：不阻断 */ }
+  try {
+    usePaperStore().start(); // 论文撰写探活（30s；离线时触发词落 docgen 单发降级）
+  } catch { /* paper store 未就绪：不阻断 */ }
 });
 
 const canSend = computed(() =>
@@ -134,6 +138,16 @@ const bioflowEnabled = computed(() => {
   }
 });
 
+/** 论文撰写可用（探活显隐；离线时触发词落 docgen 单发降级——后端不注册时由请求失败兜底） */
+const paperEnabled = computed(() => {
+  try {
+    const st = usePaperStore();
+    return st.availability !== 'offline';
+  } catch {
+    return true;
+  }
+});
+
 const hintText = computed(() =>
   attachments.value.length ? T.composerHint : `${T.composerHint} · ${DG.composerDropHint}`,
 );
@@ -291,17 +305,34 @@ const submit = () => {
       emit('send', text); // 生信 store 未就绪（测试/降级）：漏判安全退化为 QA 路径
     }
   } else if (intent === 'docgen') {
-    // 文档生成分支：chat.sendDocgen（用户消息落库 + 骨架 + 异步生成 + 自动切文档 TAB）
-    // 缺陷 #42：选中模板经 { templateId } 传入（空串/未选 = 自动推荐）
+    // 文档生成分支：论文撰写 sub-route（paperIntent && paperEnabled → sendPaper），
+    // 否则 sendDocgen（降级路径零改动——paper 未启用/未命中时触发词落 docgen 单发）。
+    let paperRoute = false;
     try {
-      useChatStore().sendDocgen(text, readyAttachments.value.map((a) => a.libraryId), {
-        templateId: selectedTemplateId.value || '',
-      }).catch(() => {
-        /* sendDocgen 内部已按失败可见原则标注错误，此处兜底吞掉重复抛出 */
-      });
+      paperRoute = paperEnabled.value && paperIntent(text);
     } catch {
-      /* chat store 未就绪（测试/降级）：漏判安全退化为 QA 路径 */
-      emit('send', text);
+      paperRoute = false;
+    }
+    if (paperRoute) {
+      try {
+        useChatStore().sendPaper(text).catch(() => {
+          /* sendPaper 内部已按失败可见原则标注错误 */
+        });
+      } catch {
+        emit('send', text); // chat store 未就绪（测试/降级）：漏判安全退化为 QA 路径
+      }
+    } else {
+      // 缺陷 #42：选中模板经 { templateId } 传入（空串/未选 = 自动推荐）
+      try {
+        useChatStore().sendDocgen(text, readyAttachments.value.map((a) => a.libraryId), {
+          templateId: selectedTemplateId.value || '',
+        }).catch(() => {
+          /* sendDocgen 内部已按失败可见原则标注错误，此处兜底吞掉重复抛出 */
+        });
+      } catch {
+        /* chat store 未就绪（测试/降级）：漏判安全退化为 QA 路径 */
+        emit('send', text);
+      }
     }
   } else if (intent === 'edit') {
     // 编辑分支：预填 DocPane「用 prompt 修改」并切文档 TAB
diff --git a/ui/src/components/chat/PaperBriefCard.vue b/ui/src/components/chat/PaperBriefCard.vue
new file mode 100644
index 0000000..7bbf872
--- /dev/null
+++ b/ui/src/components/chat/PaperBriefCard.vue
@@ -0,0 +1,64 @@
+<script setup>
+// PaperBriefCard.vue —— 论文引导表单卡（预填已抽取 brief，缺失字段待补）
+// 卡片辅助、NL 主导：提交时把已填字段拼成一句 NL，回 Composer 走 sendPaper 提交。
+import { ref, computed } from 'vue';
+import { PP } from '../../i18n/paper.js';
+import { useChatStore } from '../../stores/chat.js';
+
+const props = defineProps({ block: { type: Object, required: true } });
+
+const fields = computed(() => (Array.isArray(props.block.fields) ? props.block.fields : []));
+const submitting = ref(false);
+const values = ref({});
+
+for (const f of fields.value) {
+  if (!(f.key in values.value)) values.value[f.key] = f.value || '';
+}
+
+function submit() {
+  if (submitting.value || props.block.submitted) return;
+  submitting.value = true;
+  const parts = fields.value
+    .filter((f) => values.value[f.key] && String(values.value[f.key]).trim())
+    .map((f) => `${f.label}：${String(values.value[f.key]).trim()}`);
+  const text = `帮我写一篇论文：${parts.join('；')}`;
+  props.block.submitted = true;
+  try {
+    useChatStore().sendPaper(text).catch(() => { /* 内部已标注错误 */ });
+  } catch { /* chat store 未就绪：降级 */ }
+  submitting.value = false;
+}
+</script>
+
+<template>
+  <div class="paper-brief-card">
+    <div class="pbc-title">{{ PP.paperBriefTitle }}</div>
+    <div class="pbc-hint">{{ PP.paperBriefHint }}</div>
+    <div v-for="f in fields" :key="f.key" class="pbc-field">
+      <label class="pbc-label">{{ f.label }}</label>
+      <select v-if="Array.isArray(f.options) && f.options.length" v-model="values[f.key]" class="pbc-input">
+        <option value=""></option>
+        <option v-for="o in f.options" :key="o" :value="o">{{ o }}</option>
+      </select>
+      <input v-else v-model="values[f.key]" class="pbc-input" type="text" />
+    </div>
+    <div class="pbc-footer">
+      <button v-if="!props.block.submitted" class="pbc-submit" type="button" :disabled="submitting" @click="submit">
+        {{ submitting ? PP.paperBriefSubmitting : PP.paperBriefSubmit }}
+      </button>
+      <span v-else class="pbc-done">✓</span>
+    </div>
+  </div>
+</template>
+
+<style scoped>
+.paper-brief-card { border: 1px solid var(--border, #e5e7eb); border-radius: 8px; padding: 12px; margin: 8px 0; }
+.pbc-title { font-weight: 600; margin-bottom: 4px; }
+.pbc-hint { color: #6b7280; font-size: 12px; margin-bottom: 8px; }
+.pbc-field { display: flex; align-items: center; gap: 8px; margin-bottom: 6px; }
+.pbc-label { min-width: 72px; font-size: 13px; color: #374151; }
+.pbc-input { flex: 1; padding: 4px 8px; border: 1px solid #d1d5db; border-radius: 4px; }
+.pbc-footer { margin-top: 8px; }
+.pbc-submit { padding: 6px 14px; border: none; border-radius: 6px; background: #2563eb; color: #fff; cursor: pointer; }
+.pbc-done { color: #16a34a; }
+</style>
diff --git a/ui/src/components/chat/PaperOutlineCard.vue b/ui/src/components/chat/PaperOutlineCard.vue
new file mode 100644
index 0000000..fea02c2
--- /dev/null
+++ b/ui/src/components/chat/PaperOutlineCard.vue
@@ -0,0 +1,51 @@
+<script setup>
+// PaperOutlineCard.vue —— 论文大纲确认卡（章节勾选/编辑要点 + 确认开写）
+import { ref, computed } from 'vue';
+import { PP } from '../../i18n/paper.js';
+import { useChatStore } from '../../stores/chat.js';
+
+const props = defineProps({ block: { type: Object, required: true } });
+
+const sections = computed(() => (Array.isArray(props.block.sections) ? props.block.sections : []));
+const confirming = ref(false);
+
+function confirm() {
+  if (confirming.value || props.block.status === 'confirmed') return;
+  confirming.value = true;
+  props.block.status = 'confirmed';
+  try {
+    useChatStore().confirmPaper(props.block.paper_id);
+  } catch { /* chat store 未就绪：降级 */ }
+  confirming.value = false;
+}
+</script>
+
+<template>
+  <div class="paper-outline-card">
+    <div class="poc-title">{{ PP.paperOutlineTitle }}：{{ props.block.title }}</div>
+    <div v-for="s in sections" :key="s.section_id" class="poc-section">
+      <input type="checkbox" :checked="s.selected !== false" @change="s.selected = $event.target.checked" />
+      <span class="poc-sec-title">{{ s.title }}</span>
+      <span v-if="s.key_points && s.key_points.length" class="poc-sec-pts">{{ s.key_points.join('；') }}</span>
+    </div>
+    <div class="poc-footer">
+      <button v-if="props.block.status !== 'confirmed'" class="poc-confirm" type="button" :disabled="confirming" @click="confirm">
+        {{ confirming ? PP.paperOutlineConfirming : PP.paperOutlineConfirm }}
+      </button>
+      <span v-else class="poc-confirmed">✓ 已启动撰写</span>
+    </div>
+    <div class="poc-modify-hint">{{ PP.paperOutlineModify }}</div>
+  </div>
+</template>
+
+<style scoped>
+.paper-outline-card { border: 1px solid var(--border, #e5e7eb); border-radius: 8px; padding: 12px; margin: 8px 0; }
+.poc-title { font-weight: 600; margin-bottom: 8px; }
+.poc-section { display: flex; align-items: baseline; gap: 6px; margin-bottom: 4px; }
+.poc-sec-title { font-weight: 500; }
+.poc-sec-pts { color: #6b7280; font-size: 12px; }
+.poc-footer { margin-top: 10px; }
+.poc-confirm { padding: 6px 14px; border: none; border-radius: 6px; background: #16a34a; color: #fff; cursor: pointer; }
+.poc-confirmed { color: #16a34a; }
+.poc-modify-hint { margin-top: 6px; color: #9ca3af; font-size: 12px; }
+</style>
diff --git a/ui/src/components/chat/PaperProgressCard.vue b/ui/src/components/chat/PaperProgressCard.vue
new file mode 100644
index 0000000..f4c1c08
--- /dev/null
+++ b/ui/src/components/chat/PaperProgressCard.vue
@@ -0,0 +1,34 @@
+<script setup>
+// PaperProgressCard.vue —— 论文节级进度卡（逐节撰写状态，写稿期 2s 轮询驱动更新）
+import { computed } from 'vue';
+import { PP } from '../../i18n/paper.js';
+
+const props = defineProps({ block: { type: Object, required: true } });
+
+const sections = computed(() => (Array.isArray(props.block.sections) ? props.block.sections : []));
+const statusLabel = computed(() => {
+  if (props.block.status === 'done') return PP.paperProgressDone;
+  if (props.block.status === 'error') return '撰写失败';
+  if (props.block.status === 'cancelled') return '已取消';
+  return PP.paperProgressWriting;
+});
+</script>
+
+<template>
+  <div class="paper-progress-card">
+    <div class="ppc-title">{{ PP.paperProgressTitle }} · {{ statusLabel }}</div>
+    <div v-for="s in sections" :key="s.section_id" class="ppc-section" :class="'is-' + s.status">
+      <span class="ppc-dot">{{ s.status === 'done' ? '✓' : s.status === 'writing' ? '…' : '○' }}</span>
+      <span class="ppc-sec-title">{{ s.title }}</span>
+    </div>
+  </div>
+</template>
+
+<style scoped>
+.paper-progress-card { border: 1px solid var(--border, #e5e7eb); border-radius: 8px; padding: 12px; margin: 8px 0; }
+.ppc-title { font-weight: 600; margin-bottom: 8px; }
+.ppc-section { display: flex; align-items: center; gap: 6px; margin-bottom: 4px; }
+.ppc-dot { width: 16px; text-align: center; color: #9ca3af; }
+.ppc-section.is-done .ppc-dot { color: #16a34a; }
+.ppc-section.is-writing .ppc-dot { color: #2563eb; }
+</style>
diff --git a/ui/src/i18n/paper.js b/ui/src/i18n/paper.js
new file mode 100644
index 0000000..f5b2489
--- /dev/null
+++ b/ui/src/i18n/paper.js
@@ -0,0 +1,28 @@
+// i18n/paper.js —— 论文撰写（paper-writer）全部新文案（独立文件不碰 labels.js 红线）
+
+export const PAPER_STEPS = [
+  { name: '抽取需求', desc: '解析论文主题与写作要求', phase: 'planning' },
+  { name: '生成大纲', desc: '检索知识库并规划章节', phase: 'retrieval' },
+  { name: '逐节撰写', desc: '按大纲逐节生成正文', phase: 'generation' },
+  { name: '文献引用', desc: '自动引用知识库文献并编号', phase: 'evidence' },
+  { name: '合成全文', desc: '合并全文并渲染参考文献', phase: 'done' },
+];
+
+export const PP = {
+  paperBriefTitle: '论文撰写引导',
+  paperBriefHint: '请补充以下信息（也可直接打字告诉我）',
+  paperBriefSubmit: '生成大纲',
+  paperBriefSubmitting: '生成中…',
+  paperOutlineTitle: '论文大纲',
+  paperOutlineConfirm: '确认并开始撰写',
+  paperOutlineConfirming: '启动中…',
+  paperOutlineModify: '说一句修改意见，如「把展望合并到讨论」',
+  paperProgressTitle: '撰写进度',
+  paperProgressWriting: '撰写中…',
+  paperProgressDone: '撰写完成',
+  paperRefsTitle: '参考文献',
+  paperRefsDisclaimer: '参考文献依据知识库资料生成，格式为 GB/T 7714 近似，建议人工核对。',
+  paperCiteTitle: '引用来源',
+};
+
+export default PP;
diff --git a/ui/src/intent.js b/ui/src/intent.js
index e51b6f6..153a5de 100644
--- a/ui/src/intent.js
+++ b/ui/src/intent.js
@@ -107,4 +107,15 @@ export function route(text, hasAttachments, hasActiveArtifact) {
   return 'qa';
 }
 
+// —— 论文撰写意图（paper-writer）：在 docgen 分支内做 sub-route ——
+// 「论文/综述」主题词 + 写类动词 → 论文撰写（结构化引用体系），
+// 其余 docgen（报告/通知等）仍走原 sendDocgen。纯函数，node 直测。
+const PAPER_TYPE_WORDS = ['论文', '综述'];
+
+export function paperIntent(text) {
+  const s = String(text == null ? '' : text).trim();
+  if (!s) return false;
+  return PAPER_TYPE_WORDS.some((w) => s.includes(w)) && hasAnyWord(s, WRITE_VERB);
+}
+
 export default route;
diff --git a/ui/src/stores/chat.js b/ui/src/stores/chat.js
index 438a468..2140670 100644
--- a/ui/src/stores/chat.js
+++ b/ui/src/stores/chat.js
@@ -26,6 +26,10 @@ import { generate, getTask } from '../api/docgen.js';
 import { useSettingsStore } from './settings.js';
 import { useDocgenStore } from './docgen.js';
 
+// ---- 论文撰写（paper-writer）：独立分支（send/_runStream/sendDocgen 零改动） ----
+import { paperTurn, confirmPaper as confirmPaperApi, paperProgress as paperProgressApi } from '../api/paper.js';
+import { PP as PP_LABELS } from '../i18n/paper.js';
+
 // ---- P2 生信流程编排：独立分支（send/_runStream/sendDocgen 零改动） ----
 import { BF } from '../i18n/bioflow.js';
 import { route as routeBioflow } from '../utils/intent.js';
@@ -901,6 +905,10 @@ function newStreamCtx(sessionId, msgList) {
 const DOCGEN_TASK_POLL_MS = 2000;
 const docgenTaskPolls = new Map(); // task_id → { timer, sessionId, aMsg }
 
+// 论文撰写（paper-writer）写稿期轮询注册表：paper_id → {timer, sessionId}
+const PAPER_POLL_MS = 2000;
+const paperPolls = new Map();
+
 export const useChatStore = defineStore('chat', {
   state: () => ({
     sessionId: null,   // 当前会话 id（与 sessions.activeId 对应）
@@ -1566,6 +1574,172 @@ export const useChatStore = defineStore('chat', {
       entry.timer = setTimeout(tick, DOCGEN_TASK_POLL_MS);
     },
 
+    // ===============================================================
+    // 论文撰写（paper-writer）：sendPaper —— 用户消息落库 → POST /api/paper/turns
+    // （短同步，BFF agent 直写助手消息并返回 blocks）→ 本地回填助手消息。
+    // 写稿期由 confirmPaper 触发后台循环，前端 2s 轮询 paperProgress 更新进度卡，
+    // complete 后重载会话取 agent 直写的 artifact 块并接管到文档 TAB。
+    // ===============================================================
+    async sendPaper(description, { userMsg } = {}) {
+      const q = String(description == null ? '' : description).trim();
+      if (!q || this.busy) return;
+      const sessions = useSessionsStore();
+      if (!this.sessionId) {
+        try {
+          const sess = await sessions.create();
+          this.sessionId = sess.id;
+        } catch {
+          return; // 创建失败：列表页可见错误
+        }
+      }
+      if (this.sessionId && !sessions.items.some((s) => s.id === this.sessionId)) {
+        sessions.items.unshift({ id: this.sessionId, title: q.slice(0, 30), status: 'running', created_at: new Date().toISOString(), updated_at: new Date().toISOString(), message_count: 0, meta: {} });
+      }
+      const sessionId = this.sessionId;
+      const msgList = this.messages;
+      this.sessionMsgs[sessionId] = msgList;
+      const ctx = newStreamCtx(sessionId, msgList);
+      this.streamCtx[sessionId] = ctx;
+      this.lastResult = null;
+
+      // 1. 用户消息入列 + 落库
+      let userId = null;
+      let userMsgRef = null;
+      if (userMsg) {
+        userMsgRef = userMsg;
+        userId = userMsg.id;
+      } else {
+        const raw = { id: null, role: 'user', kind: 'paper', status: 'done', text: q, mode: 'ask', meta: {}, blocks: [{ type: 'text', html: q }], raw: null, created_at: new Date().toISOString(), updated_at: new Date().toISOString() };
+        msgList.push(raw);
+        userMsgRef = msgList[msgList.length - 1];
+        try {
+          const saved = await createMessage(sessionId, { role: 'user', text: q });
+          userId = saved.id;
+          userMsgRef.id = userId;
+        } catch (err) {
+          userMsgRef.persistError = err;
+          this._finishCtx(ctx, { ok: false, error: err, message: userMsgRef, silentAbort: false });
+          return;
+        }
+      }
+      sessions.patchLocal(sessionId, { status: 'running' });
+      try { await updateSessionQuiet(sessionId, { status: 'running' }); } catch { /* 状态 PATCH 失败不阻断 */ }
+
+      // 2. POST turns（短同步）
+      let payload;
+      try {
+        payload = await paperTurn({ session_id: sessionId, text: q });
+      } catch (err) {
+        const marker = this._errorMarker(err.code);
+        const text = err.message || (marker && marker.text) || T.upstreamError;
+        const raw = { id: null, role: 'assistant', kind: 'paper', status: 'error', streamErr: (marker && marker.ann) || 'message', errText: text, blocks: [{ type: 'text', html: text }], raw: null, created_at: new Date().toISOString(), updated_at: new Date().toISOString() };
+        msgList.push(raw);
+        const am = msgList[msgList.length - 1];
+        this._persistFinal(am, sessionId, 'error', {}).catch(() => {});
+        this._finishCtx(ctx, { ok: false, error: { code: err.code || 'NETWORK', message: text, retryable: true }, message: am, sessionId, userMessageId: userId, assistantMessageId: am.id });
+        return;
+      }
+
+      // 3. 助手消息（agent 已落库；本地回填 blocks）
+      const blocks = Array.isArray(payload && payload.blocks) ? payload.blocks : [];
+      const raw = { id: null, role: 'assistant', kind: 'paper', status: 'done', text: null, blocks: blocks.length ? blocks : [{ type: 'text', html: '（论文服务未返回内容）' }], raw: payload || null, created_at: new Date().toISOString(), updated_at: new Date().toISOString() };
+      msgList.push(raw);
+      const aMsg = msgList[msgList.length - 1];
+      ctx.aMsg = aMsg;
+      this.runningMsgId = null;
+      try {
+        const saved = await createMessage(sessionId, { role: 'assistant', kind: 'paper', status: 'done', blocks: raw.blocks });
+        aMsg.id = saved.id;
+      } catch {
+        aMsg.id = null;
+      }
+      delete this.streamCtx[sessionId]; // 释放 busy（无流）
+      this.lastResult = { ok: true, payload, message: aMsg, sessionId, userMessageId: userId, assistantMessageId: aMsg.id };
+      this.scrollSeq++;
+    },
+
+    /** 确认大纲 → POST confirm 启动后台写节循环 → 2s 轮询进度 */
+    async confirmPaper(paperId) {
+      const sessionId = this.sessionId;
+      if (!paperId || !sessionId) return;
+      try {
+        await confirmPaperApi(paperId, sessionId);
+      } catch { /* 启动失败：卡片可见错误 */ }
+      this._startPaperPoll(paperId, sessionId);
+    },
+
+    _startPaperPoll(paperId, sessionId) {
+      if (paperPolls.has(paperId)) return;
+      const entry = { timer: null, sessionId };
+      paperPolls.set(paperId, entry);
+      const clear = () => {
+        if (entry.timer != null) clearTimeout(entry.timer);
+        if (paperPolls.get(paperId) === entry) paperPolls.delete(paperId);
+      };
+      const tick = async () => {
+        if (!paperPolls.has(paperId) || paperPolls.get(paperId) !== entry) return;
+        let prog = null;
+        try {
+          prog = await paperProgressApi(paperId);
+        } catch {
+          prog = null; // 瞬时失败：继续轮询
+        }
+        if (!paperPolls.has(paperId) || paperPolls.get(paperId) !== entry) return;
+        if (prog && prog.ok) {
+          this._patchPaperProgress(sessionId, prog);
+          if (prog.status === 'complete' || prog.status === 'cancelled' || prog.status === 'error') {
+            clear();
+            if (prog.status === 'complete') {
+              try {
+                await this.loadSession(sessionId); // 重载取 agent 直写的 artifact 块
+                await this._adoptPaperArtifact(sessionId);
+              } catch { /* 重载失败不阻断 */ }
+            }
+            return;
+          }
+        }
+        entry.timer = setTimeout(tick, PAPER_POLL_MS);
+      };
+      entry.timer = setTimeout(tick, PAPER_POLL_MS);
+    },
+
+    /** 把进度同步进消息流里的 paper_progress 块 */
+    _patchPaperProgress(sessionId, prog) {
+      const list = this.sessionMsgs[sessionId] || [];
+      for (const m of list) {
+        if (!m || m.role !== 'assistant' || !Array.isArray(m.blocks)) continue;
+        const pb = m.blocks.find((b) => b && b.type === 'paper_progress');
+        if (pb) {
+          pb.status = prog.status;
+          pb.current_section_id = prog.current_section_id || '';
+          pb.sections = (prog.sections || []).map((s) => ({ section_id: s.section_id, title: s.title, status: s.status }));
+        }
+      }
+      this.scrollSeq++;
+    },
+
+    /** 找到最新含 paper_id 的 artifact 块并接管到文档 TAB */
+    async _adoptPaperArtifact(sessionId) {
+      const list = this.sessionMsgs[sessionId] || [];
+      for (const m of list) {
+        if (!m || m.role !== 'assistant' || !Array.isArray(m.blocks)) continue;
+        const art = m.blocks.find((b) => b && b.type === 'artifact' && b.paper_id);
+        if (art) {
+          try {
+            const { useDocgenStore } = await import('../stores/docgen.js');
+            await useDocgenStore().adoptArtifact(art, { initialMd: '' });
+          } catch { /* docgen store 未就绪不阻断 */ }
+          if (this.sessionId === sessionId) {
+            try {
+              const { useBenchStore } = await import('../stores/bench.js');
+              useBenchStore().setActiveTab('doc');
+            } catch { /* bench 未就绪不阻断 */ }
+          }
+          return;
+        }
+      }
+    },
+
     // ---------------------------------------------------------------
     // sendBioflow（P2）：生信对话式编排独立分支——用户消息落库（既有机制）→
     // 生信骨架（path 块）→ POST /api/bioflow/chat/stream（SSE 经 sse/parser.js
diff --git a/ui/src/stores/paper.js b/ui/src/stores/paper.js
new file mode 100644
index 0000000..edc82f8
--- /dev/null
+++ b/ui/src/stores/paper.js
@@ -0,0 +1,36 @@
+// stores/paper.js —— 论文撰写探活 + 开关（镜像 health.js 的 30s 周期探活）
+// availability: 'online' | 'offline' | 'unknown'。offline 时 Composer 触发词落 docgen 单发降级。
+import { defineStore } from 'pinia';
+import { paperHealth } from '../api/paper.js';
+
+const HEALTH_INTERVAL = 30000;
+
+export const usePaperStore = defineStore('paper', {
+  state: () => ({
+    availability: 'unknown',
+    polling: false,
+  }),
+  actions: {
+    async probe() {
+      try {
+        const data = await paperHealth();
+        this.availability = data && data.ok ? 'online' : 'offline';
+      } catch {
+        this.availability = 'offline';
+      }
+    },
+    start() {
+      if (this.polling) return;
+      this.polling = true;
+      this.probe();
+      this._timer = setInterval(() => {
+        if (document.hidden) return;
+        this.probe();
+      }, HEALTH_INTERVAL);
+    },
+    stop() {
+      clearInterval(this._timer);
+      this.polling = false;
+    },
+  },
+});