// Build and maintain a persistent knowledge base (LLM Wiki) using native backend tools. Supports ingesting PDFs, Markdown, and TXT files, and querying or linting the data. Use this whenever the user wants to add, retrieve, or manage their own local knowledge base and documents.
| name | llm-wiki |
| description | Build and maintain a persistent knowledge base (LLM Wiki) using native backend tools. Supports ingesting PDFs, Markdown, and TXT files, and querying or linting the data. Use this whenever the user wants to add, retrieve, or manage their own local knowledge base and documents. |
You are the maintainer of a persistent local LLM Wiki. When the user asks you to manage their wiki (ingest documents, answer queries from it, or lint the wiki), follow these instructions carefully.
The LLM Wiki is based on a structured 3-layer architecture:
sources/): The raw files that are tracked in the wiki. The system relies on a deduplication manifest (manifest.json) so the same file data isn't processed twice.wiki/): The compiled knowledge. You maintain index.md, topic files, and a log.md where all structural changes are recorded.schema/): Configuration mapping (internal).The Wiki operates via native backend tools (wiki_ingest, wiki_query, wiki_lint). The default workspace is the absolute path to ./.llm_wiki within your root workspace. If the user specifies a directory for the wiki (like wiki_lib), you MUST determine the FULL ABSOLUTE PATH to that directory and pass it to the workspace parameter of the tools. This ensures the wiki is correctly created inside a .llm_wiki sub-folder at that exact location.
When a user provides a file (PDF, TXT, MD) or a directory of files to be ingested:
Call your wiki_ingest tool with the source path as source and the target directory as workspace (if provided).
source..llm_wiki/wiki/log.md, and securely add standard Markdown links inside .llm_wiki/wiki/index.md so that the new knowledge is linked from the root page.When the user asks you a question that relies on their wiki:
Call your wiki_query tool with a highly descriptive search query.
workspace path when calling the wiki_lint tool.To ensure the wiki is healthy (no broken links or orphaned pages):
Call your wiki_lint tool.
workspace path when calling the wiki_lint tool..md files in .llm_wiki/wiki/ (using standard file editing tools) to repair broken paths. Ensure all paths are relative.sources/. They are the raw truth. Only edit files in wiki/.index.md after an ingestion.进阶版日报生成器，支持多数据源采集、工作分析、趋势对比、周报月报聚合
代理执行 AKG 算子任务。先检查固定仓库与分支；若 `~/.akg/check_env.md` 不存在则强制 `FULL_SETUP=true`；忽略所有 `akg_cli` 检查和使用；后端代码生成直接执行 `run_workflow.py --workflow kernelgen`。
为昇腾算子在源码中接入 TRACE_POINT 与 MoeTracing，串通 trace_preprocessor、profiling tensor、point_map.json、 save_profiling_data 与 trace_collector 生成 Chrome trace。强调门禁 G1–G5：全链路预处理与 OPP、profiling 为数据输出最后一位、 整条编译与示例脚本联调、落盘路径在 spawn 前 resolve。遵循函数级粒度与就地扩展，禁止另注册 xxx_profiling 类第二入口， 保持原 Op 与 torch.ops 名称及签名不变。在用户提到算子打点、Profiling、Chrome trace、MoeTracing，或将结论写入本 skill 时读取。
在用户提供 Chrome/Perfetto trace.json、或排查 Ascend 上 MoE/FusedDeepMoe 等算子性能时使用。按 phase、category、core group、tid 统计耗时、overlap、bubble，输出 CSV、Markdown 报告与确定性诊断；可选外部 LLM 扩写分析。默认 phase 映射面向 UMDK FusedDeepMoe，其它 trace 需替换或扩展 config/phase_map.yaml。
跨会话检索聊天原文（记忆不足时再用）。在回答任何关于历史事件、日期、人物、过去对话的问题时，如果记忆中没有相关信息或不足以回答，则需要使用跨会话检索聊天原文。用 mcp_exec_command 执行 scripts/search_history.py，读 ~/.jiuwenswarm/agent/sessions/*/history.json。支持 channel、session_id、关键词、时间窗。如果搜索结果不足，尝试用不同的关键词再次搜索。
安排延迟重启本 Agent 所在的服务（JiuwenSwarm app）。执行后当前 Agent 进程会被终止并重新启动，当前会话会断开。用于用户要求重启、配置更新需生效、或服务异常需重载时。使用 bash 调用脚本。