LisonEvf/dsh-stock-panel

@lisonevf/dsh-stock-panel is a web bundle plugin built specifically for the DeepSeek Harness dsh runtime, registered via dsh plugin --profile web add and mounted into the conversation page's right-hand full-height view slot as a persistent "A股工作台" tab. Installation is a single command against the prebuilt tarball at the GitHub Release URL (lisonevf-dsh-stock-panel-1.6.0.tgz), and because pnpm skips lifecycle build scripts for that asset no extra flags or credentials are required; users can alternatively install straight from main by appending --config.dangerouslyAllowAllBuilds=true, which forwards the flag to pnpm and triggers an on-the-fly build that takes roughly 50 seconds, after which dsh web must be restarted and the browser hard-refreshed (Ctrl+Shift+R) so the new bundle is picked up by the in-page __ModuleLoader__.

The plugin is organised around the A-share trading day: after-close review walks a seven-step checklist and persists a next-day expectation list (capped at five items across seven fields), the pre-open auction compares each item line by line, intraday the UI deliberately exposes only three questions, and the closing auction helps finalise positioning, while seven views (Market, Review, Battle Plan, Self-Mined Sectors, Workbench, Screener, Alert Rules) follow the clock automatically until the user clicks a nav entry, after which manual navigation takes over. Data flows through the host-side bridge which registers 19 built-in tools — 15 market tools plus 4 self-mined-concept tools, of which 8 are also surfaced to the conversation agent — and all upstream endpoints are configured by the user, because the plugin ships and redistributes no market data of its own; connection failures surface as a visible red banner rather than silent empty states.

The most distinctive feature is the "Self-Mined Sectors" view, which derives sector groupings from the market's own money flow rather than vendor labels: it computes 60-day residual returns over the A-share universe (with optional PCA and market-mean removal), builds a Pearson correlation matrix, applies hierarchical clustering with edge-cutting, filters weak and transitive links plus isolated names, then batches the surviving clusters to the LLM for naming in roughly 35 seconds, displaying each cluster with a strength bar and source pill (model theme, official industry, or official concept) alongside a diff against official classifications. Strict non-goals are enforced in the README: no stock picks, no price predictions, no buy/sell signals, no win-rate claims, and AI output is always read-only — the "一键研判" action routes through the host's ctx.llm for strict JSON, and nothing the model produces is ever written into the user's list unless the user clicks "+" on each item individually, so personal drafts are never overwritten. Prerequisites are Node.js 18+ and a global @deepseek-ai/dsh install, the tag #v1.6.0 will not install (it predates a publish script rename), and engineering quality is held by a CI suite including a 26-rule ux-contract gate, density and contrast checks, a 665 KB client bundle budget, and 93 naming-related assertions across 28 offline node:test files, all under the MIT license.