DSH Studio: Skill Extraction Pipeline and On-Demand MCP Loading, Two Features in One Plugin

samir2015dzh/dsh-studio is a fork of zhengjy01/dsh-skill-studio, categorized as a workflow in the community directory. It consolidates two functions into a single plugin: a skill workbench and on-demand MCP management. On the skill side, a YES/NO decision is sent three seconds after the end of each round (throttled by 90 seconds for judgment based on the last 10 summaries and 5 minutes for extraction); full extraction is performed only if YES is received. Candidates are written directly to `~/.agents/skills` and only enter the review queue if writing fails. Before merging, it passes through four checkpoints: tombstone scanning, category tagging, local similarity rechecking, and model folding. Merged results shorter than 60% of the original text are rejected. The model utilizes the provider/model configured in DSH, and the plugin does not store API keys. On the MCP side, only tool names are listed at the start of a session (descriptions have been omitted since version 0.8.0), with the full `inputSchema` pulled on-demand by `mcp_load` or `mcp_call`. There are three modes: eager, on-demand (default), and disabled. By default, idle connections are disconnected after 5 minutes without a call, but the cache is retained for transparent reconnection on the next call. The plugin also provides three-tier filtering lists, usage-based grouping, a trash bin, multi-source scanning, and diagnostic markers. It installs into the DSH profile via `link:` and its configuration is centralized in `~/.dsh/dsh-studio/extractor.json` and `mcp.json`.