WTStarMark/dsh-plugin-mesh

dsh-plugin-mesh is an independent visualization site built around the DSH ecosystem rather than a native DSH plugin or client. It crawls GitHub via the public search API for repositories carrying tags such as dsh, dsh-plugin, dsh-plugin-desktop, dsh-plugin-market, and dsh-plugins, then arranges them by function into an interactive concentric mesh of 19 sectors including agent skills, model integrations, desktop clients, interface panels, skins, plugin marketplaces, conversation interaction, memory and context, voice reminders, tool commands, development debugging, web frontends, usage quotas, security permissions, remote access, integration bridges, file management, reading docs, and desktop pets. The official repository deepseek-ai/deepseek-harness is pinned to the center, and every sector subdivides into 2–5 fine branches governed by 61 classification rules shared verbatim between the front-end layout and the Python classifier.

The typical workflow begins by visiting the live deployment at <http://104.129.51.126/> or by cloning the repository and running npm run serve:lan (which requires Node 20 or newer) to launch the local preview at port 8788. To refresh the underlying data, users copy .env.example to .env, supply a GITHUB_TOKEN, run python3 backend/check_token.py to confirm the quota has risen from 10 to 30 requests per minute, and then invoke python3 backend/collect.py --once for a single pass or --loop --interval 3600 for continuous hourly refreshes. The site lets visitors click any sector to expand it into a full circle with its fine branches re-fanned, filter by tag-AND, language, archived status, and keyword, and reuse pre-generated 420×168 self-contained SVG repository cards in other READMEs or web pages through /api/card/<owner>/<name>.svg and the read-only /api/repos JSON endpoint.

Dependencies are deliberately minimal: the front-end requires only Node 20 with zero npm packages, while the back-end is written entirely against the Python standard library with zero pip requirements and zero model calls. The recommended token is a fine-grained GitHub personal access token whose permissions can all stay blank since the collector only reads public search results; the token mainly raises the search quota from 10 to 30 requests per minute and should be rotated roughly every 90 days. The project is released under the MIT license, exposes only a single HTTP port (port 80 in production), rejects any non-GET/HEAD write with HTTP 405, and applies IP-based sliding-window rate limiting that returns 429 when exceeded. Operators using pm2 must invoke --loop rather than --watch, otherwise the manager restarts the collector via SIGINT on every snapshot write and silently drops captured repositories.