victor10035445/dsh-v-skill-links

dsh-v-skill-links is a native incremental patch plugin for DeepSeek Harness Web (dsh web); it maps any local directory's first-level subdirectories that contain a SKILL.md file into a new "Mapped Skills" group inside the chat input's / slash menu, with no npm dependencies and no build step. Installation is done through the harness CLI via dsh plugin --profile web add "github:victor10035445/dsh-v-skill-links", or alternatively via a link: path to a local clone, or via npm pack followed by dsh plugin --profile web add <absolute-tgz-path>, and the plugin requires Harness 0.1.2-rc.1 or newer because it relies on snapshot-store and UI primitive seed tokens exposed by the web shell; configuration lives under $DSH_HOME\profiles\web\cordis.patch.yml under the id: dsh-v-skill-links patch entry, accepting absolute, ~, and relative (relative-to-dsh-startup-directory) paths in directories and a buttons array.

The plugin exposes three entry points beyond the core mapper: mapped skills are invoked through / then a name token (the matching SKILL.md body is injected as a <skill_content> block at the agent/pre-step hook without entering the model's <available_skills> catalog), custom instructions appear under a "Custom Instructions" group and fire the prompt as an immediate user message on selection, and quick buttons render in a grid below the composer where clicks append the prompt to the input (optionally with autoSend, and optionally bound to a {provider, model, reasoningEffort?} triple that switches the session model in place, with graceful fallback when the remote catalog is unreachable). A settings page at Settings → Skill Management offers three tabs (Skill Mapping, Custom Instructions, Quick Buttons) sharing a single draft buffer; saving writes through to the patch layer with hot reload, scanning is first-level-subdirectory only, hidden dot-directories are skipped, case-insensitive substring matching with prefix-priority ranking is used, and stable .v-qb-grid, .v-qb-btn, data-v-button CSS hooks respect the official theme tokens.

This is an additive, non-replacing plugin — it does not disable or replace any official plugin and its mapped skills never appear in the model's system-prompt skill list. Limitations documented in the README include: command names take precedence over skill tokens, name collisions with native skills produce parallel injections, custom instructions require an active agent (sub-agents not yet resumed return 404), pre-step injection is not replayed when a session is restored, only first-level subdirectories are indexed, and the button grid is global yet session-scoped (so sub-agent composers also render it, and blocking composers do not prevent append). The repository is MIT-licensed, ships nine smoke tests under pnpm test, and first-run requires restarting dsh web and refreshing the page so the loader entry and dsh.client bundle wiring take effect.