dsh-pick: A Plugin Picking Advisor for DeepSeek Harness

What Does It Do?

It is a DSH plugin for DeepSeek Harness that acts as a plugin-picking advisor, turning "which plugin should I install" into a repeatable flow: state your need, find candidates, compare across dimensions, get a verdict. It solves the problem that raw npm download counts are extremely noisy, so the top result by raw numbers is often just the package that ships releases most often. Its core capabilities are denoised metric computation, candidate comparison and ranking, and pushing conclusions into the conversation for scenario-aware analysis.

Core Features

• Denoised metrics: a daily baseline is the median of daily downloads after removing release days and the day after, combined with 7-day growth, release cadence and recency into a weighted health score (baseline 40%, growth 25%, freshness 20%, cadence 15%).
• Need-to-query conversion: Chinese requirement descriptions are turned into search terms via synonyms and bigram splitting, shared by both the CLI and the panel.
• Comparison and ranking: compare 2 to 6 plugins item by item with the best value in each row highlighted, or rank by weekly downloads, baseline, growth, health score or category.
• In-conversation tools: four tools are registered with DeepSeek Harness, so you can ask "which database plugin is reliable?" without opening the panel and the AI queries the catalog and compares on its own.
• Pre-install feasibility warnings: cards flag DSH compatibility, a missing dsh.bundle declaration, tui-client plugins, and overlap with plugins you already have; candidates that cannot be installed get no install button.
• Daily catalog refresh: GitHub Actions rebuilds the catalog every day, and the plugin checks at startup and every 6 hours, falling back to the bundled copy when the fetch fails.

How to Use This Plugin?

After enabling it in DSH, open the full panel from the "pick a plugin" entry on the right sidebar guide page; the top row shows the plugin name, version, update and uninstall buttons, and both update and uninstall are delegated to DSH's own plugin manager, taking effect after a restart. The "+AI" button on candidate cards and conclusion sections packages the need and key numbers into a message sent to the current conversation, sending only conclusions and figures rather than raw data, and never overwriting text you have typed but not sent. You can also skip the panel entirely and describe your need in the conversation, where the AI calls the tools to search the catalog and compare data; every tool result carries a methodology note and the AI is instructed to cite only the numbers the tools return. Clicking "install then confirm" on a candidate card installs a plugin from the catalog, and already-installed ones show an installed marker. The command-line and standalone web mode offer subcommands for building the catalog, backfilling metrics, searching, ranking, exact lookups and local preview, which suits bulk data verification.