Frontend-Backend Plugin
A cross-stack guard rail for projects that use Supabase + Vercel + Next.js + React + Node. Its single job: stop the frontend and the backend from drifting apart.
Tiered investigation pipeline. Every workflow (analyze, build, debug, verify, schema-sync, cleanup, db-setup) starts with a triage step that picks the smallest pipeline that proves the change correct: T0 inline fix (no agents), T1 primary specialist alone, T2 primary + mistake-monitor watchdog, T3 full INVESTIGATION QUARTET per independent sub-domain — primary + Deep Explorer + Deep Validator + mistake-monitor — fanned out concurrently in one message. Auto-escalates one tier on migration / RLS / auth / payments / enum / public-export / generated-column / multi-page-UI / production-risk wording. Cross-cutting gates (type-check, lint, build, browser smoke when UI changed) still run at every tier — the tier only sizes the agent fleet, never the verification surface.
Streaming N-round re-dispatch. Helpers (Deep Explorer, Deep Validator, Mistake Monitor) dispatch with run_in_background: true while the primary writer runs foreground. Findings stream into per-source-agent queues; the orchestrator fires re-dispatches per finding the moment each source agent is free. N-round cap (default N=3) with overlapping findings merging into the next round; T2 = 90s / T3 = 180s wall budget per source agent across all rounds. Beyond cap or budget, findings surface as deferred. Every Agent() call passes model: inherit so parallel branches never downgrade.
What it does
- DB schema parity — every code field that maps to a Supabase column must use the same name. The plugin actively hunts for mismatches and fixes them in code (it never silently renames a DB column).
- Careful feature analysis before implementation — for any new feature, runs an analysis pass (DB → API → UI → permissions → realtime/cache) before writing code.
- Root-cause debugging — when fixing a bug, refuses to stop at the first symptom; walks back to the originating cause across the full stack.
- Browser verification — tiered. T1 = legacy single-observer pass. T2/T3 =
comprehensive-verify federation: verify-coordinator fans out 6 (T2) or 15 (T3) departments via the shared dept-verifier agent (each dept gets its procedure passed in the prompt). Departments cover Visual / Functional / Console / Network / DOM / Security / State+Storage / Realtime / SEO / i18n / Mobile-Touch / PWA / Analytics / Cross-Page / Performance. All use mcp__Claude_in_Chrome__* only — never substitutes.
- Mixes the official skills — delegates to
supabase:supabase, supabase:supabase-postgres-best-practices, vercel:nextjs, vercel:vercel-functions, vercel:env-vars, vercel:vercel-cli, vercel:react-best-practices, vercel:knowledge-update, etc.
- Uses MCP connectors — Supabase MCP for schema/migrations, Vercel MCP for deployments/logs, Sentry MCP for production errors, Context7 MCP for library docs.
- Standing cleanup mandate — after every change, removes anything unnecessary on FE AND BE: dead code, unused deps, unused env vars, unused assets, orphan DB columns/tables/indexes/policies/buckets/edge-functions/cron-jobs/extensions.
- Production-safety / non-regression — verifies in Claude in Chrome and smoke-tests adjacent features. Never breaks something that was working.
Skills
Jeffallan
affaan-m
f
Randroids-Dojo
wshobson
JuliusBrussee