agentic-plugins

Builds disposable localhost workbenches — Flask + raw sqlite3 + htmx + Server-Sent Events apps — that give an agentic coding/eval/PR/data session a live UI with no build step, npm install, or deploy. The signature move is a two-way human↔agent loop over one SQLite file: the human acts in the browser (htmx POST → SQLite → fragment + SSE invalidation), the agent acts from the terminal (httpx → SQLite → SSE), both seeing the same state update live with no reload. Ships recipes for eval viewers, PR review rooms, document review / redline surfaces, trace replays, refactor cockpits, and more. Use when the user wants a quick local UI for a coding/eval/PR/data task, OR mentions localhost workbench, eval viewer, PR review room, review/redline a document, select text and leave comments, htmx+SSE app, live invalidation, two-way agent loop, or "build me a little UI for this." Do NOT use for production apps or React/Vue/Svelte SPAs (use frontend-design), or static one-shot data pages / HTML reports (use generative-ui).

Score a single skill against the gardener rubric. Reads the skill's SKILL.md, references/, and templates/, then writes a 7-dimension scorecard with per-criterion rationales and a recommended-actions list. Isolated counterpart to /gardener: same rubric, one item at a time. Use when the user asks to audit a skill, score a skill, check if a skill is healthy, or wants to re-score after making changes.

Monthly catalog hygiene audit for Claude Code plugin marketplaces. Inventories every skill and agent, scores each against a 7-dimension rubric via parallel Opus Agents, detects description collisions via Bedrock Cohere v4 embeddings with pairwise cosine, and writes a delta report vs. the prior audit. Proposes resolutions; does not auto-apply. Use when the user asks to audit the catalog, run the gardener, find skill collisions, check plugin hygiene, generate a monthly audit report, or mentions catalog drift.

Propose description rewrites to resolve skill collisions. Given a collision pair from a gardener audit, reads both skills' descriptions, generates before/after rewrites with negative discriminators, and writes a proposal Markdown the user applies manually. Uses current Claude prompting discipline: positive imperatives, motivation clauses, scope explicit. Use when the user asks to fix a collision, rewrite a description, disambiguate two skills, or mentions resolving a gardener-flagged pair.