feat: comprehensive v2 upgrade — streaming, error KB, file upload, layout analysis

Major changes:
- Streaming: LLM统一 _BaseLLM 接口 (invoke + stream), generate/modify/correct
  节点使用 get_stream_writer() 实现逐字输出, UI 节点平铺展开自动折叠
- Prompt外部化: 7个prompt拆分到 prompts/*.md, loader.py 支持热重载
- 错误自增长: backend/error_kb.py — 指纹去重 + ChromaDB持久化,
  correct_jrxml→validate 通过时自动入库, retrieve同时搜索错误KB
- 文件上传: backend/file_parser.py — PDF/DOCX/图片/文本解析,
  侧边栏多文件上传, 文本自动注入下一条消息
- A4模板识别: backend/layout_analyzer.py — 三种模式(完整A4/行片段修改/行片段新建),
  PaddleOCR元素提取 + 行分组 + JRXML section匹配
- 会话历史下载: jrxml_versions版本追踪 + 侧边栏历史版本下载按钮
- 预览修复: route_after_save跳过预览/导出意图的验证循环
- Ctrl+C修复: JS注入拦截Streamlit裸c键清缓存

Docs: CLAUDE.md (完整项目文档), ROADMAP.md (改进路线图)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

app.py

@@ -1,7 +1,15 @@
-"""Streamlit 多轮对话 UI，用于 JRXML 生成代理。"""
+"""Streamlit 多轮对话 UI，用于 JRXML 生成代理。
+
+支持:
+- 流式输出（LLM 逐字展示）
+- 节点平铺展开（每个处理阶段独立展示）
+- 完成后自动折叠节点区
+- 过程总结卡片
+"""
 
 import os
 import sys
+from pathlib import Path
 
 import streamlit as st
 
@@ -23,7 +31,70 @@ st.set_page_config(
     initial_sidebar_state="expanded",
 )
 
-# ---- URL 参数：session_id ----
+# 阻止 Streamlit 裸 'c' 键清除缓存，保留 Ctrl+C 复制行为
+st.components.v1.html("""
+<script>
+(function() {
+    const parent = window.parent.document;
+    parent.addEventListener('keydown', function(e) {
+        // 仅拦截裸 'c' 键（非 Ctrl/Cmd 组合）
+        if (e.key === 'c' && !e.ctrlKey && !e.metaKey && !e.altKey) {
+            const tag = parent.activeElement ? parent.activeElement.tagName : '';
+            if (tag !== 'INPUT' && tag !== 'TEXTAREA' && !parent.activeElement.isContentEditable) {
+                e.stopImmediatePropagation();
+                e.preventDefault();
+            }
+        }
+    }, true);
+})();
+</script>
+""", height=0)
+
+# ---- 节点名称 → 中文标签 ----
+NODE_LABELS = {
+    "load_session":       "📂 加载会话",
+    "process_input":      "📝 记录输入",
+    "manage_context":     "🧠 管理上下文",
+    "save_state_snapshot": "💾 保存快照",
+    "classify_intent":    "🔍 识别意图",
+    "retrieve":           "📚 检索模板",
+    "generate":           "⚙️ 生成 JRXML",
+    "modify_jrxml":       "🔧 修改 JRXML",
+    "validate":           "✅ 验证",
+    "explain_error":      "🔎 分析错误",
+    "correct_jrxml":      "🛠 自动修正",
+    "finalize":           "📋 完成",
+    "handle_consult":     "💬 咨询回答",
+    "handle_undo":        "↩ 撤销操作",
+    "handle_reset":       "🔄 重置会话",
+    "save_session":       "💾 保存会话",
+}
+
+INTENT_LABELS = {
+    "initial_generation": "新建报表",
+    "modify_report":      "修改报表",
+    "preview_report":     "预览报表",
+    "export_pdf":         "导出 PDF",
+    "export_jrxml":       "下载 JRXML",
+    "undo_modification":  "撤销修改",
+    "consult_question":   "咨询问题",
+    "reset_session":      "重置会话",
+}
+
+SKIP_NODES = {"load_session", "process_input", "manage_context",
+              "save_state_snapshot", "save_session"}
+
+
+def _render_jrxml(jrxml: str, max_lines: int = 30):
+    """展示 JRXML 代码（折叠、限行）。"""
+    lines = jrxml.strip().split("\n")
+    preview = "\n".join(lines[:max_lines])
+    if len(lines) > max_lines:
+        preview += f"\n... (共 {len(lines)} 行)"
+    st.code(preview, language="xml")
+
+
+# ---- URL 参数 ----
 query_params = st.query_params
 url_session_id = query_params.get("session_id", "")
 
@@ -35,7 +106,6 @@ if "graph" not in st.session_state:
 if "pending_action" not in st.session_state:
     st.session_state.pending_action = None
 
-# 确定活跃的 session_id
 if "agent_state" not in st.session_state:
     if url_session_id:
         data = load_session(url_session_id)
@@ -58,8 +128,8 @@ if "agent_state" not in st.session_state:
 current_session_id = st.session_state.agent_state.get("session_id", "")
 
 
-def run_agent(user_input: str) -> dict:
-    """运行代理图，返回最终状态。"""
+def run_agent(user_input: str):
+    """运行代理图：流式渲染节点进度 + LLM 文本。"""
     agent_state = st.session_state.agent_state
 
     if agent_state.get("current_jrxml") and agent_state.get("status") == "pass":
@@ -68,120 +138,155 @@ def run_agent(user_input: str) -> dict:
     agent_state["user_input"] = user_input
     agent_state["retry_count"] = 0
 
+    # ---- UI 容器 ----
+    streaming_placeholder = st.empty()   # 流式文本
+    nodes_container = st.container()      # 节点进度区
+    summary_placeholder = st.empty()      # 总结卡片
+
+    # 节点追踪
+    executed_nodes: list[dict] = []  # {name, label, status, detail}
+    stream_text = ""
+    stream_active = False
+    current_stream_node = ""
     final_state = None
-    with st.chat_message("assistant"):
-        status_placeholder = st.empty()
-        jrxml_placeholder = st.empty()
 
-        for event in st.session_state.graph.stream(agent_state):
-            for node_name, node_state in event.items():
-                final_state = node_state
-                if node_name == "classify_intent":
-                    intent = node_state.get("intent", "")
-                    intent_labels = {
-                        "initial_generation": "🆕 识别为新建报表请求",
-                        "modify_report": "✏️ 识别为修改报表请求",
-                        "preview_report": "👁 识别为预览请求",
-                        "export_pdf": "📄 识别为导出PDF请求",
-                        "export_jrxml": "📥 识别为导出JRXML请求",
-                        "undo_modification": "↩ 识别为撤销请求",
-                        "consult_question": "💬 识别为咨询问题",
-                        "reset_session": "🔄 识别为重置会话请求",
-                    }
-                    label = intent_labels.get(intent, f"🔍 意图: {intent}")
-                    status_placeholder.info(label)
-                elif node_name == "generate":
-                    status_placeholder.info("🔧 正在生成 JRXML...")
-                elif node_name == "modify_jrxml":
-                    status_placeholder.info("🔧 正在根据您的请求修改 JRXML...")
-                elif node_name == "validate":
-                    if node_state.get("status") == "pass":
-                        status_placeholder.success("✅ 验证通过！")
-                    else:
-                        status_placeholder.warning("⚠ 验证失败，正在分析错误...")
-                elif node_name == "explain_error":
-                    explanation = node_state.get("natural_explanation", "")
-                    status_placeholder.warning(f"🔍 {explanation}")
-                elif node_name == "correct_jrxml":
-                    status_placeholder.info(f"🛠 正在自动修正（尝试 {node_state.get('retry_count', 1)}）...")
-                elif node_name == "handle_consult":
-                    pass
-                elif node_name == "handle_undo":
-                    status_placeholder.info("↩ 已撤销上一步修改")
-                elif node_name == "handle_reset":
-                    status_placeholder.info("🔄 会话已重置")
-                elif node_name == "manage_context":
-                    pass
-                elif node_name == "save_state_snapshot":
-                    pass
-                elif node_name == "save_session":
-                    pass
-                elif node_name == "finalize":
-                    pass
+    try:
+        for event in st.session_state.graph.stream(
+            agent_state, stream_mode=["updates", "custom"]
+        ):
+            mode, data = event
 
-        if final_state:
-            st.session_state.agent_state = final_state
-            intent = final_state.get("intent", "")
+            if mode == "updates":
+                for node_name, node_state in data.items():
+                    label = NODE_LABELS.get(node_name, node_name)
+                    if node_name not in SKIP_NODES:
+                        executed_nodes.append({
+                            "name": node_name,
+                            "label": label,
+                        })
 
+                    if node_name == "classify_intent":
+                        intent = node_state.get("intent", "")
+                        il = INTENT_LABELS.get(intent, intent)
+                        executed_nodes[-1]["detail"] = f"意图: {il}"
+
+                    elif node_name == "retrieve":
+                        ctx = node_state.get("retrieved_context", "")
+                        executed_nodes[-1]["detail"] = (
+                            f"找到 {len(ctx)} 字符参考模板" if ctx else "未匹配到模板"
+                        )
+
+                    elif node_name in ("generate", "modify_jrxml", "correct_jrxml"):
+                        # 流式文本已在上面的 custom 事件中展示
+                        jrxml = node_state.get("current_jrxml", "")
+                        executed_nodes[-1]["detail"] = f"生成 {len(jrxml)} 字符 JRXML"
+
+                    elif node_name == "validate":
+                        status = node_state.get("status", "")
+                        if status == "pass":
+                            executed_nodes[-1]["detail"] = "验证通过 ✓"
+                        else:
+                            err = node_state.get("error_msg", "")
+                            executed_nodes[-1]["detail"] = f"验证失败: {err[:80]}"
+
+                    elif node_name == "explain_error":
+                        expl = node_state.get("natural_explanation", "")
+                        executed_nodes[-1]["detail"] = expl[:120]
+
+                    elif node_name == "handle_consult":
+                        ans = node_state.get("consult_answer", "")
+                        executed_nodes[-1]["detail"] = ans[:150]
+
+                    final_state = node_state
+
+            elif mode == "custom":
+                cd = data
+                if cd.get("type") == "stream":
+                    stream_text += cd.get("text", "")
+                    stream_active = True
+                    current_stream_node = cd.get("node", "")
+                    streaming_placeholder.code(stream_text, language="xml")
+
+    except Exception as e:
+        st.error(f"工作流异常: {e}")
+        return
+
+    # ---- 渲染节点进度区 ----
+    with nodes_container:
+        with st.expander("处理过程", expanded=False):
+            for i, node in enumerate(executed_nodes):
+                icon = "✓" if i < len(executed_nodes) - 1 else "●"
+                detail_str = f" — {node['detail']}" if node.get("detail") else ""
+                st.caption(f"{icon} {node['label']}{detail_str}")
+
+    # ---- 清除流式占位 ----
+    if stream_active:
+        streaming_placeholder.empty()
+
+    # ---- 总结卡片 ----
+    if final_state:
+        st.session_state.agent_state = final_state
+        intent = final_state.get("intent", "")
+        status = final_state.get("status", "")
+
+        with summary_placeholder.container(border=True):
             if intent == "consult_question":
                 answer = final_state.get("consult_answer", "")
+                st.info(answer)
                 st.session_state.messages.append({
-                    "role": "assistant",
-                    "content": answer,
-                    "type": "consult",
+                    "role": "assistant", "content": answer, "type": "consult",
                 })
-                status_placeholder.empty()
-                st.markdown(answer)
+
             elif intent in ("undo_modification", "reset_session"):
-                # 消息已在节点中添加，不需要额外输出
-                status_placeholder.empty()
-                jrxml_placeholder.empty()
+                st.success("操作已完成")
+                # 消息已在节点中添加
+
             elif intent in ("preview_report", "export_pdf", "export_jrxml"):
                 jrxml = final_state.get("current_jrxml", "")
                 if jrxml:
+                    st.success("✅ 当前报表")
+                    _render_jrxml(jrxml)
                     st.session_state.messages.append({
-                        "role": "assistant",
-                        "content": jrxml,
-                        "type": "jrxml",
+                        "role": "assistant", "content": jrxml, "type": "jrxml",
                     })
-                    status_placeholder.success("✅ 当前报表")
-                    jrxml_placeholder.code(jrxml, language="xml")
                 else:
-                    status_placeholder.warning("⚠ 当前没有报表可以预览或导出。")
-                    jrxml_placeholder.empty()
-            elif final_state.get("status") == "pass":
+                    st.warning("⚠ 当前没有报表可以展示。")
+
+            elif status == "pass":
+                jrxml = final_state.get("current_jrxml", "")
+                st.success("✅ JRXML 生成成功")
+                st.markdown("**生成结果:**")
+                _render_jrxml(jrxml)
+                st.caption("您可以从侧边栏下载文件，或继续对话进行修改。")
                 st.session_state.messages.append({
-                    "role": "assistant",
-                    "content": final_state.get("current_jrxml", ""),
-                    "type": "jrxml",
+                    "role": "assistant", "content": jrxml, "type": "jrxml",
                 })
                 st.session_state.messages.append({
                     "role": "assistant",
                     "content": "✅ JRXML 生成成功！您可以从侧边栏下载文件，或继续修改。",
                     "type": "success",
                 })
-                status_placeholder.success("✅ JRXML 验证通过！")
-                jrxml_placeholder.code(final_state.get("current_jrxml", ""), language="xml")
+
             else:
+                jrxml = final_state.get("current_jrxml", "")
                 error_msg = final_state.get("error_msg", "未知错误")
                 explanation = final_state.get("natural_explanation", "")
                 retries = final_state.get("retry_count", 0)
+                st.error(f"❌ 经过 {retries} 次重试后仍无法生成有效的 JRXML")
+                st.markdown(f"**错误:** {error_msg}")
+                if explanation:
+                    st.markdown(f"**原因:** {explanation}")
+                if jrxml:
+                    with st.expander("查看当前 JRXML"):
+                        _render_jrxml(jrxml, max_lines=80)
+                st.caption("请简化报表结构后重试。")
                 st.session_state.messages.append({
                     "role": "assistant",
-                    "content": final_state.get("current_jrxml", ""),
-                    "type": "jrxml",
-                })
-                st.session_state.messages.append({
-                    "role": "assistant",
-                    "content": f"❌ 经过 {retries} 次重试后仍无法生成有效的 JRXML。\n\n**错误:** {error_msg}\n\n**解释:** {explanation}\n\n请重新描述您的需求或简化报表结构。",
+                    "content": f"❌ 经过 {retries} 次重试后仍无法生成有效的 JRXML。\n\n**错误:** {error_msg}",
                     "type": "error_explanation",
                 })
-                status_placeholder.error(f"❌ 经过 {retries} 次重试后验证失败")
-                jrxml_placeholder.text("")
-        else:
-            st.error("未产生结果，请重试。")
-
-    return final_state
+    else:
+        st.error("未产生结果，请重试。")
 
 
 # ---- 侧边栏 ----
@@ -192,7 +297,6 @@ with st.sidebar:
 
     # 会话管理
     st.markdown("### 会话管理")
-
     sessions = list_all_sessions()
     session_options = {}
     for s in sessions:
@@ -270,6 +374,83 @@ with st.sidebar:
             run_agent("重新来，清空当前报表")
         st.rerun()
 
+    st.divider()
+    st.markdown("### 上传文件")
+    st.caption("支持图片 (OCR)、PDF、Word、文本文件。内容将附加到您的下一条消息中。")
+
+    if "uploaded_files" not in st.session_state:
+        st.session_state.uploaded_files = []  # [{name, text, type}]
+
+    uploaded = st.file_uploader(
+        "选择文件",
+        type=["png", "jpg", "jpeg", "bmp", "webp", "pdf", "docx", "txt", "csv", "json", "xml"],
+        accept_multiple_files=True,
+        key="file_uploader",
+        label_visibility="collapsed",
+    )
+
+    if uploaded:
+        for uf in uploaded:
+            # 去重
+            if any(f["name"] == uf.name for f in st.session_state.uploaded_files):
+                continue
+            import tempfile
+            from backend.file_parser import parse_file
+            from backend.layout_analyzer import analyze_layout
+
+            suffix = Path(uf.name).suffix.lower()
+            with tempfile.NamedTemporaryFile(suffix=suffix, delete=False) as tmp:
+                tmp.write(uf.getvalue())
+                tmp_path = tmp.name
+
+            result = parse_file(tmp_path, suffix)
+
+            # 对图片/PDF 进行 A4 模板布局分析
+            parsed_text = result["text"]
+            parsed_type = result["file_type"]
+            if suffix in (".png", ".jpg", ".jpeg", ".bmp", ".webp", ".pdf"):
+                layout = analyze_layout(tmp_path)
+                tt = layout.get("template_type", "unknown")
+                current_jrxml = st.session_state.agent_state.get("current_jrxml", "")
+
+                if tt == "full_a4":
+                    parsed_text = layout["description"]
+                    parsed_type = "a4_template"
+                elif tt == "partial_rows":
+                    parsed_type = "a4_partial"
+                    if current_jrxml.strip():
+                        # 修改模式：尝试行匹配
+                        from backend.layout_analyzer import match_rows_to_jrxml
+                        match = match_rows_to_jrxml(layout, current_jrxml)
+                        parsed_text = (
+                            f"[行片段修改] 上传图片包含 {layout['total_rows']} 行，"
+                            f"视为 A4 报表的一部分。\n\n"
+                            f"{match['description']}\n\n"
+                            f"--- 行结构 ---\n{layout['description']}"
+                        )
+                    else:
+                        # 新建模式：按 A4 模板处理
+                        parsed_text = layout["description"]
+
+            Path(tmp_path).unlink(missing_ok=True)
+
+            if parsed_text:
+                st.session_state.uploaded_files.append({
+                    "name": uf.name,
+                    "text": parsed_text,
+                    "type": parsed_type,
+                })
+
+    if st.session_state.uploaded_files:
+        for i, f in enumerate(st.session_state.uploaded_files):
+            cols = st.columns([5, 1])
+            with cols[0]:
+                st.caption(f"📎 {f['name']} ({f['type']}, {len(f['text'])} 字符)")
+            with cols[1]:
+                if st.button("✕", key=f"rm_uf_{i}", help="移除"):
+                    st.session_state.uploaded_files.pop(i)
+                    st.rerun()
+
     st.divider()
     st.markdown("### 配置")
     llm_backend = os.getenv("LLM_BACKEND", "cloud")
@@ -280,16 +461,36 @@ with st.sidebar:
 
     st.divider()
     st.markdown("### 下载")
+
     final = st.session_state.agent_state.get("final_jrxml", "")
+    versions = st.session_state.agent_state.get("jrxml_versions", [])
+
     if final:
         st.download_button(
-            label="📥 下载 JRXML",
+            label="📥 下载最新 JRXML",
             data=final,
             file_name="report.jrxml",
             mime="application/xml",
             use_container_width=True,
         )
 
+    if versions:
+        with st.expander("📋 历史版本", expanded=False):
+            for i, v in enumerate(reversed(versions)):
+                ts = v.get("ts", "")[:16]
+                label = v.get("label", "版本")
+                status = v.get("status", "")
+                icon = "✅" if status == "pass" else "❌"
+                dl_label = f"{icon} v{len(versions)-i} — {label} ({ts})"
+                st.download_button(
+                    label=dl_label,
+                    data=v.get("jrxml", ""),
+                    file_name=f"report_v{len(versions)-i}.jrxml",
+                    mime="application/xml",
+                    use_container_width=True,
+                    key=f"dl_v{i}",
+                )
+
 # ---- 标题 ----
 st.title("📝 JRXML 报表生成器")
 st.caption("用自然语言描述您的报表需求，我将逐步生成可用的 JRXML 模板。")
@@ -297,22 +498,33 @@ st.caption("用自然语言描述您的报表需求，我将逐步生成可用
 # ---- 聊天历史 ----
 for msg in st.session_state.messages:
     with st.chat_message(msg["role"]):
-        if msg["role"] == "assistant" and msg.get("type") == "jrxml":
+        if msg.get("type") == "jrxml":
             with st.expander("查看生成的 JRXML", expanded=False):
                 st.code(msg["content"], language="xml")
-        elif msg["role"] == "assistant" and msg.get("type") == "error_explanation":
+        elif msg.get("type") == "error_explanation":
             st.warning(msg["content"])
-        elif msg["role"] == "assistant" and msg.get("type") == "success":
+        elif msg.get("type") == "success":
             st.success(msg["content"])
-        elif msg["role"] == "assistant" and msg.get("type") == "consult":
+        elif msg.get("type") == "consult":
             st.info(msg["content"])
         else:
             st.markdown(msg["content"])
 
 # ---- 聊天输入 ----
 if prompt := st.chat_input("描述您的报表需求..."):
+    # 拼接上传文件的文本
+    uploaded_texts = []
+    if st.session_state.get("uploaded_files"):
+        for f in st.session_state.uploaded_files:
+            uploaded_texts.append(f"[上传文件: {f['name']}]\n{f['text']}")
+    if uploaded_texts:
+        full_prompt = "\n\n".join(uploaded_texts) + "\n\n---\n用户需求:\n" + prompt
+        st.session_state.uploaded_files = []  # 用后即清
+    else:
+        full_prompt = prompt
+
     st.session_state.messages.append({"role": "user", "content": prompt})
     with st.chat_message("user"):
         st.markdown(prompt)
-    run_agent(prompt)
+    run_agent(full_prompt)
     st.rerun()