feat: 同sender消息合并、数字人格精炼与系统治理

.github/workflows/release.yml

@@ -38,46 +38,10 @@ jobs:
       - name: Install Python dependencies
         run: uv sync --group ci -p ${{ env.PYTHON_VERSION }}
 
-      - name: Validate release tag matches package version
+      - name: Validate release tag, build versions, and changelog
         env:
           TAG_NAME: ${{ github.ref_name }}
-        run: |
-          uv run python - <<'PY'
-          import os
-          import pathlib
-          import re
-          import sys
-          import tomllib
-
-          tag_name = os.environ["TAG_NAME"]
-          expected_version = tag_name.removeprefix("v")
-
-          pyproject = tomllib.loads(pathlib.Path("pyproject.toml").read_text(encoding="utf-8"))
-          project_version = str(pyproject["project"]["version"]).strip()
-
-          init_text = pathlib.Path("src/Undefined/__init__.py").read_text(encoding="utf-8")
-          match = re.search(r'__version__\s*=\s*"([^"]+)"', init_text)
-          if match is None:
-              raise SystemExit("Could not find __version__ in src/Undefined/__init__.py")
-          init_version = match.group(1).strip()
-
-          errors: list[str] = []
-          if project_version != init_version:
-              errors.append(
-                  f"Version mismatch: pyproject.toml={project_version}, src/Undefined/__init__.py={init_version}"
-              )
-          if project_version != expected_version:
-              errors.append(
-                  f"Tag/version mismatch: tag={tag_name}, expected package version={expected_version}, actual={project_version}"
-              )
-
-          if errors:
-              raise SystemExit("\n".join(errors))
-
-          print(
-              f"Validated release version {project_version} from tag {tag_name}, pyproject.toml, and src/Undefined/__init__.py"
-          )
-          PY
+        run: uv run python scripts/release_notes.py validate --tag "$TAG_NAME"
 
       - name: Cache Ruff
         uses: actions/cache@v4
@@ -572,9 +536,6 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Fetch tags
-        run: git fetch --tags --force
-
       - name: Download build artifacts
         uses: actions/download-artifact@v4
         with:
@@ -583,47 +544,15 @@ jobs:
 
       - name: Build release notes
         shell: bash
-        run: |
-          TAG_NAME="${{ github.ref_name }}"
-          TAG_TYPE=$(git cat-file -t "$TAG_NAME")
-          TAG_CONTENT=""
-          if [ "$TAG_TYPE" = "tag" ]; then
-            TAG_CONTENT=$(git tag -l --format='%(contents)' "$TAG_NAME")
-            TAG_CONTENT=$(echo "$TAG_CONTENT" | sed '/-----BEGIN PGP SIGNATURE-----/,/-----END PGP SIGNATURE-----/d')
-          fi
-          PREV_TAG=$(git describe --tags --abbrev=0 "${TAG_NAME}^" 2>/dev/null || git rev-list --max-parents=0 HEAD)
-          if [ -n "$TAG_CONTENT" ]; then
-            echo "$TAG_CONTENT" > tag_message.txt
-            printf '\n---\n\n' >> tag_message.txt
-          else
-            : > tag_message.txt
-          fi
-          cat >> tag_message.txt <<'NOTES'
-          ## 📝 Detailed Changes
-          NOTES
-          FEAT_COMMITS=$(git log ${PREV_TAG}..${TAG_NAME} --grep='^feat' --pretty=format:'* %s (%h)' 2>/dev/null || true)
-          if [ -n "$FEAT_COMMITS" ]; then
-            echo -e "\n### 🚀 Features" >> tag_message.txt
-            echo "$FEAT_COMMITS" >> tag_message.txt
-          fi
-          FIX_COMMITS=$(git log ${PREV_TAG}..${TAG_NAME} --grep='^fix' --pretty=format:'* %s (%h)' 2>/dev/null || true)
-          if [ -n "$FIX_COMMITS" ]; then
-            echo -e "\n### 🐛 Bug Fixes" >> tag_message.txt
-            echo "$FIX_COMMITS" >> tag_message.txt
-          fi
-          OTHER_COMMITS=$(git log ${PREV_TAG}..${TAG_NAME} --grep='^feat\|^fix' --invert-grep --pretty=format:'* %s (%h)' 2>/dev/null || true)
-          if [ -n "$OTHER_COMMITS" ]; then
-            echo -e "\n### 🛠 Maintenance & Others" >> tag_message.txt
-            echo "$OTHER_COMMITS" >> tag_message.txt
-          fi
+        run: python3 scripts/release_notes.py notes --tag "${{ github.ref_name }}" --output release_notes.md
 
       - name: Create GitHub release
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: |
           gh release create ${{ github.ref_name }} release-downloads/* \
             --title "Undefined ${{ github.ref_name }}" \
-            --notes-file tag_message.txt
+            --notes-file release_notes.md
 
   publish-pypi:
     name: Publish Python package to PyPI

AGENTS.md

@@ -34,7 +34,10 @@ Use `<attachment uid="..."/>` for both images and files. The legacy `<pic uid=".
 Remote attachments are cached only up to `[attachments].remote_download_max_size_mb`; larger items, or all remote items when the value is `0`, are registered as URL references with `source_ref` instead of downloaded file content.
 
 ### Auto processing pipelines
-Automatic extraction pipelines live under `src/Undefined/skills/auto_pipeline/pipelines/<name>/` and use `config.json + handler.py`. Slash commands have higher priority; when a command is dispatched, automatic pipelines and AI auto-reply are skipped. Command inputs and command outputs should be recorded in message history so later AI turns can see the result. For non-command messages, all pipelines detect in parallel and all matches process in parallel before AI auto-reply. Outputs should go through `MessageSender`, which writes history and automatically registers local CQ media or uploaded files as session attachment UIDs.
+Automatic extraction pipelines live under `src/Undefined/skills/pipelines/<name>/` and use `config.json + handler.py`. Slash commands have higher priority; when a command is dispatched, automatic pipelines and AI auto-reply are skipped. Command inputs and command outputs should be recorded in message history so later AI turns can see the result. For non-command messages, all pipelines detect in parallel and all matches process in parallel before AI auto-reply. Outputs should go through `MessageSender`, which writes history and automatically registers local CQ media or uploaded files as session attachment UIDs.
+
+### Same-sender short-window message batching
+Consecutive messages from the same sender within `[message_batcher].window_seconds` are merged into a single AI invocation, so the AI sees the whole batch as `<message>` blocks and decides per-intent (independent request vs. correction/interruption). Pokes always bypass; an at-bot message arriving while a buffer already exists is processed individually so it is not blocked; a first at-bot message that opens the buffer routes the eventual batch through the mention lane. History writes remain unchanged. Configure under `[message_batcher]` (`enabled`, `window_seconds`, `strategy`, `max_window_seconds`, `max_messages_per_batch`, `group_enabled`, `private_enabled`, `pre_send_seconds`, `allow_cancel_after_send`); details in [docs/message-batching.md](docs/message-batching.md). Optional speculative pre-fire (`0 < pre_send_seconds < window_seconds`) dispatches the current batch to the LLM early once the user has been silent for `pre_send_seconds`; new messages can cancel the in-flight call as long as it has not yet sent any reply.
 
 ### User identification in prompts
 The system prompt now includes a rule: **recognize and address users by their QQ ID (`sender_id`)** because nicknames can change. When needing to address a user, use the latest nickname obtained via `group.get_member_info(brief=true)`. Observations recorded in cognitive memory should always include the QQ ID, e.g., “QQ号12345678（昵称张三）做了某事”.

ARCHITECTURE.md

@@ -44,8 +44,10 @@ graph TB
             InjectionAgent["InjectionResponseAgent<br/>注入响应生成<br/>[injection_response_agent.py]"]
         end
 
-        CommandDispatcher["CommandDispatcher<br/>命令分发器<br/>• /help /stats /lsadmin<br/>• /addadmin /rmadmin<br/>• /bugfix /faq<br/>[services/command.py]"]
-
+        CommandDispatcher["CommandDispatcher<br/>命令分发器<br/>• /help /stats /admin<br/>• /bugfix /faq<br/>[services/command.py]"]
+
+        MessageBatcher["MessageBatcher<br/>同 sender 短时合并<br/>• 按 (scope, sender_id) 分桶<br/>• T1=window_seconds 结束 batch<br/>• T2=pre_send_seconds 投机预发送<br/>• 拍一拍/buffer 内 @bot 旁路<br/>• 首条 @bot 整批走 mention 队列<br/>[services/message_batcher.py]"]
+
         subgraph QueueSystem["车站-列车 队列系统 (services/)"]
             AICoordinator["AICoordinator<br/>AI 协调器<br/>• Prompt 构建<br/>• 队列管理<br/>• 回复执行<br/>[ai_coordinator.py]"]
             QueueManager["QueueManager<br/>队列管理器<br/>[queue_manager.py]"]
@@ -121,7 +123,7 @@ graph TB
 
         subgraph CommandsLayer["平台指令 (skills/commands/)"]
             Cmd_Core["核心指令<br/>• help • stats"]
-            Cmd_Admin["管理指令<br/>• addadmin<br/>• rmadmin • lsadmin"]
+            Cmd_Admin["管理指令<br/>• admin (ls/add/del)"]
             Cmd_FAQ["FAQ 指令<br/>• faq (ls/view/search/del)"]
             Cmd_Fun["娱乐指令<br/>• bugfix"]
         end
@@ -226,6 +228,8 @@ graph TB
     GitHubSender -->|"发送图片卡片"| OneBotClient
 
     MessageHandler -->|"3. 自动回复"| AICoordinator
+    AICoordinator -->|"3.1 入桶等待合并"| MessageBatcher
+    MessageBatcher -->|"3.2 flush 合并批次"| AICoordinator
     AICoordinator -->|"创建上下文"| RequestContext
     AICoordinator -->|"入队"| QueueManager
     QueueManager -->|"分发"| ModelQueues
@@ -319,7 +323,7 @@ graph TB
 
     class User,Admin,OneBotServer,LLM_API external
     class Main,ConfigLoader,ConfigHotReload,ConfigModels,OneBotClient,Context,WebUI core
-    class MessageHandler,SecurityService,InjectionAgent,CommandDispatcher,AICoordinator message
+    class MessageHandler,SecurityService,InjectionAgent,CommandDispatcher,MessageBatcher,AICoordinator message
     class AIClient,PromptBuilder,ModelRequester,ToolManager,MultimodalAnalyzer,SummaryService,TokenCounter,Parsing ai
     class ToolRegistry,AgentRegistry,AgentToolRegistry,IntroGenerator skills
     class RequestContext,ContextFilter,ResourceRegistry,HistoryManager,MemoryStorage,EndSummaryStorage,CognitiveService,CognitiveJobQueue,CognitiveHistorian,CognitiveVectorStore,CognitiveProfileStorage,FAQStorage,ScheduledTaskStorage,TokenUsageStorage storage
@@ -342,6 +346,7 @@ sequenceDiagram
     participant MH as MessageHandler
     participant SS as SecurityService
     participant CD as CommandDispatcher
+    participant MB as MessageBatcher
     participant AC as AICoordinator
     participant QM as QueueManager
     participant AI as AIClient
@@ -370,7 +375,7 @@ sequenceDiagram
             CD-->>OB: 返回结果
             OB->>U: 发送响应
         else 非命令消息
-            MH->>MH: 并行检测 skills/auto_pipeline 管线
+            MH->>MH: 并行检测 skills/pipelines 管线
             opt 命中自动提取管线
                 MH->>MH: 并行处理全部命中的自动提取
                 MH->>OH: 发送视频/卡片/PDF
@@ -380,7 +385,22 @@ sequenceDiagram
             end
             %% AI处理流程
             MH->>AC: handle_auto_reply()
-
+            alt 拍一拍 / buffer 内 @bot
+                AC->>QM: 立即按优先级入队
+            else 普通消息进合并桶
+                AC->>MB: submit(BufferedMessage)
+                MB-->>MB: TYPING: 重置 T1/T2 静默计时
+                opt T2=pre_send_seconds 到期
+                    MB->>AC: handle_batched_dispatch(投机批次 + BatchDispatchToken)
+                    AC->>QM: 提前入队抢时间
+                end
+                MB-->>MB: T1=window_seconds 到期结束 batch
+                opt 未启用投机或尚未预发送
+                    MB->>AC: handle_batched_dispatch(最终批次)
+                    AC->>QM: 按首条触发的优先级入队
+                end
+            end
+
             %% 上下文创建
             AC->>AC: 创建 RequestContext
             AC->>ST: 保存历史记录
@@ -831,8 +851,8 @@ description: 从 PDF 文件中提取文本和表格，填写表单。当用户
 
 1. **外部实体层**：用户、管理员、OneBot 协议端 (NapCat/Lagrange.Core)、大模型 API 服务商
 2. **核心入口层**：main.py 启动入口、配置管理器 (config/loader.py)、热更新应用器 (config/hot_reload.py)、OneBotClient (onebot.py)、RequestContext (context.py)、Runtime API Server (api/app.py → api/routes/ 路由子模块)
-3. **消息处理层**：MessageHandler (handlers.py)、SecurityService (security.py)、CommandDispatcher (services/command.py)、AICoordinator (ai_coordinator.py)、QueueManager (queue_manager.py)、自动处理管线 (skills/auto_pipeline/)、Bilibili/arXiv/GitHub 解析与发送模块
-    自动提取由 `AutoPipelineRegistry` 并行检测、并行处理全部命中的管线；发送结果写入历史后继续进入 AI 自动回复。
+3. **消息处理层**：MessageHandler (handlers.py)、SecurityService (security.py)、CommandDispatcher (services/command.py)、MessageBatcher (services/message_batcher.py)、AICoordinator (ai_coordinator.py)、QueueManager (queue_manager.py)、自动处理管线 (skills/pipelines/)、Bilibili/arXiv/GitHub 解析与发送模块
+    自动提取由 `PipelineRegistry` 并行检测、并行处理全部命中的管线；发送结果写入历史后继续进入 AI 自动回复。
 4. **AI 核心能力层**：AIClient (client.py)、PromptBuilder (prompts.py)、ModelRequester (llm.py)、ToolManager (tooling.py)、MultimodalAnalyzer (multimodal.py)、SummaryService (summaries.py)、TokenCounter (tokens.py)
 5. **存储与上下文层**：MessageHistoryManager (utils/history.py, 10000条限制)、MemoryStorage (memory.py, 置顶备忘录, 500条上限)、EndSummaryStorage、CognitiveService + JobQueue + HistorianWorker + VectorStore + ProfileStorage、MemeService + MemeWorker + MemeStore + MemeVectorStore (表情包库)、FAQStorage、ScheduledTaskStorage、TokenUsageStorage (自动归档)
 6. **技能系统层**：ToolRegistry (registry.py)、AgentRegistry、6个 Agents、11类 Toolsets
@@ -847,6 +867,7 @@ description: 从 PDF 文件中提取文本和表格，填写表单。当用户
 *   **非阻塞发车**：实现了可配置节奏的非阻塞调度循环（默认 **1Hz**）。列车按节奏出发，带走一个请求到后台异步处理。
 *   **高可用性**：即使前一个请求仍在处理（如耗时的网络搜索），新的请求也会按时被分发，不会造成队列堵塞。
 *   **优先级管理**：支持四级优先级（超级管理员 > 私聊 > 群聊@ > 群聊普通），确保重要消息优先响应。
+*   **关停收敛**：`MessageHandler.close()` 会先 flush `MessageBatcher`，再调用 `QueueManager.drain()` 等待已入队请求和在途请求自然完成，最后才停止队列处理器，避免缓冲消息只入队未执行。
 
 ### 6个智能体 Agent