vercel-plugin

Summary Add as the high-priority default for generic agent-building prompts. Route default agent-building guidance to lower-case and its installed docs, while leveraging the Slack Agent Skill for Slack-specific builds. Narrow prompt activation to explicit eve/project signals and refresh generated manifest/catalog. Validation frontmatter validation: passed prompt-routing check: passed for generic, Slack, explicit eve, and negative prompts : passed : not run because is not installed in this environment : blocked because dependencies are not installed and this repo is locked with Supersedes #137 to use the requested branch name.

This PR adds detection logic for which harness the plugin is installed for + an instal ID for first time users who install the Plugin

Summary is a no-op for pure lexical recalls. It rewrites the reason string to say but leaves , so the skill is injected anyway. The result is that any skill landing in the top-5 lexical results is injected regardless of how weak its raw score is — and because then takes the top 2, genuinely relevant skills get displaced by noise. Verified against (0.47.0) — lines 347–362 and 416–422 are byte-identical to the 0.25.0 copy I first hit this on. The mechanism Two constants are deliberately set so lexical alone can match ( on ): Since , top-K membership is the only gate. Raw score never enters the decision — a raw 10.6 and a raw 454.8 both clear identically. () looks like the intended raw-score threshold, but it can't reject: matches . For a pure retrieval recall ( → still , and changes nothing but the message. Observed verbatim in a real session: Suggested fix Make the floor actually reject when there is no exact signal to fall back on: Worth checking the floor separately: with reported for both a raw 10.6 and a raw 454.8, the raw MiniSearch score spans orders of magnitude, so a single absolute cutoff may need to be relative (e.g. a ratio against the top hit) to be meaningful. Note runs immediately after and adds , so any fix to the floor needs to survive that — otherwise a rejected entry is resurrected one step later. Relationship to existing issues #98* — same family of symptom, different mechanism and different fix. That one is exact-keyword matching in non-Vercel projects, remedied by a project-relevance gate. This fires in a project that passes any such gate, via the lexical path (), so #98's fix would not address it. #93* — unrelated to this, and I hit that one too (writing any pulls in ). Not re-reporting; already covered there. Workaround via in . Documented in the plugin's own CLAUDE.md env table, and it cleanly restores exact-match behavior. Environment vercel-plugin 0.25.0 (); scoring code confirmed unchanged on @ 0.47.0 Claude Code 2.1.220, macOS 15 (arm64), Node 24 Project: Next.js 16.3.0, React 19.2.4, deployed on Vercel