Backend Building Swarm

Backend Building Swarm

Stack: tRPC + Drizzle ORM + Hono + MySQL + OAuth 2.0

A swarm-aware backend skill that grafts onto an existing webapp-building-swarm project. Adds api/, contracts/, and optionally db/ directories — but never replaces or modifies existing frontend files.

Prerequisite: An existing shared repo created by webapp-building-swarm at /mnt/agents/output/app.

Architecture

This skill builds on the swarm-workspace two-tier filesystem — a shared coordination repo (branch hub) plus per-subagent worktrees. See swarm-workspace/SKILL.md for the full contract; it is not restated here.

init.sh grafts in place on $PROJECT_PATH and commits on whatever branch is checked out there — it does NOT manage worktrees or merge. In the swarm flow the main agent creates the backend worktree via swarm-workspace's setup-local.sh, runs init.sh in it with PROJECT_PATH pointed at the worktree, commits on the backend branch, and merges that branch into the integration line before page branches fork — so all page agents inherit the tRPC client and contracts. The main agent runs the graft itself rather than dispatching a subagent because init.sh writes a gitignored .env that never travels through the merge: it survives only in the sandbox where it ran, which must be the main agent's own — the one that later builds and stages .env into the other worktrees. Webapp commit history is preserved because the backend branch descends from the scaffolded frontend.

Features

Features can be installed incrementally. Base infrastructure (Hono server, tRPC, contracts) is always installed automatically on first run.

Default: init.sh "App" with no --features → defaults to auth (= db + auth).

Scripts

All scripts use absolute paths from /app/.agents/skills/backend-building-swarm/scripts/.

init.sh — Backend graft (run inside the backend worktree)

# Point PROJECT_PATH at the backend worktree (created by setup-local.sh first):
PROJECT_PATH=$HOME/app-backend \
  bash /app/.agents/skills/backend-building-swarm/scripts/init.sh "<app-title>" [--features db,auth | --template]

• Grafts in place on $PROJECT_PATH and commits on the currently checked-out branch. In the swarm flow $PROJECT_PATH is the main-agent-owned backend worktree (the graft is the main agent's own job, not a subagent's — see the gitignored-.env reason above); the script's bare default (/mnt/agents/output/app) exists only for the deterministic structure test. It does NOT create worktrees or merge.
• Re-entry safe: .backend-features.json gates re-runs (adding --features auth later only installs the delta). Graft-owned auth UI (Login.tsx, useAuth.ts, AuthLayout*) is planted authoritatively each run, while user-extensible files (schema.ts, seed.ts, connection.ts, drizzle.config.ts) yield to any version already present.

Two execution paths depending on which webapp template was used:

--features and --template are mutually exclusive. --template is selected automatically by the presence of .backend-features.json in the project — it reads the feature list from that manifest and skips all file copy/patch/wiring (the fullstack template already laid them out), doing portal provisioning + .env only.

Worktree setup (run by subagents)

Use swarm-workspace's setup-local.sh, passing this skill's prebuilt backend node_modules (Hono/tRPC/Drizzle) as NODE_MODULES_SRC:

NODE_MODULES_SRC=/app/.agents/skills/backend-building-swarm/scripts/template/node_modules \
  bash /app/.agents/skills/swarm-workspace/scripts/setup-local.sh <branch> $HOME/app-<branch>

• Creates a git worktree from the shared repo at /mnt/agents/output/app
• Each subagent MUST use a unique local path ($HOME/app-<branch>) to avoid worktree metadata collisions
• Overlays the backend node_modules so npm install is fast, then npm install
• Handles re-entry; never run git worktree prune — it destroys other agents' worktree entries

See swarm-workspace/SKILL.md for the full two-tier filesystem contract.

.prepare-template.sh — Image build-time only

Builds scripts/template/node_modules/ by npm install-ing the template package.json. Runs once at image build time, not at runtime.

Workflows

The end-to-end sequencing is owned by vibecoding-webapp-swarm (the backend graft is its Phase 4.5: run once on a backend branch forked from scaffold, merged before page branches fork). This skill provides the graft itself. The shape of a graft step:

1. webapp-building-swarm init + scaffold land a frontend on the shared repo
2. Create the backend worktree via swarm-workspace's setup-local.sh (with NODE_MODULES_SRC = this skill's template/node_modules)
3. PROJECT_PATH=$HOME/app-backend init.sh "<title>" [--features … | --template] — graft, provision, commit on backend
4. Main agent merges backend into the integration line; page agents then inherit the tRPC client + contracts

Incremental features (add capabilities over time): re-run init.sh in a worktree with --features auth — the .backend-features.json manifest gates it to install only the delta, landing a new commit on that branch.

Quick Start

1. Initialize Frontend First (webapp-building-swarm)

PROJECT_PATH=$HOME/init REMOTE_PATH=/mnt/agents/output/app \
  bash /app/.agents/skills/webapp-building-swarm/scripts/init-webapp.sh "My App"

2. Graft Backend

# Create the backend worktree (swarm-workspace's setup-local.sh + this skill's node_modules), then graft into it:
NODE_MODULES_SRC=/app/.agents/skills/backend-building-swarm/scripts/template/node_modules \
  bash /app/.agents/skills/swarm-workspace/scripts/setup-local.sh backend $HOME/app-backend

# Full stack with auth (default):
PROJECT_PATH=$HOME/app-backend \
  bash /app/.agents/skills/backend-building-swarm/scripts/init.sh "My App"

# Database only (no auth):
PROJECT_PATH=$HOME/app-backend \
  bash /app/.agents/skills/backend-building-swarm/scripts/init.sh "My App" --features db

# Fullstack template (skip scaffolding; provision-only — auto-detected via .backend-features.json):
PROJECT_PATH=$HOME/app-backend \
  bash /app/.agents/skills/backend-building-swarm/scripts/init.sh "My App" --template

Then the main agent merges the backend branch: cd /mnt/agents/output/app && git merge backend --no-edit. Because .env is gitignored, copy it to the shared repo so later worktrees can pick it up: cp $HOME/app-backend/.env /mnt/agents/output/app/.env.

What this does (base, always on first run with --features):

• Copies api/, contracts/ into the worktree
• Patches vite.config.ts in-place (adds @contracts alias, envDir, build.outDir)
• Adds tsconfig.server.json and merges @contracts/* path into existing tsconfigs
• Merges package.json (adds backend deps and scripts)
• Generates .env with portal credentials
• Runs npm install
• Commits the graft on the current branch (main agent merges it; no ff-merge here)

Additional per feature:

• db: Adds db/ directory, drizzle.config.ts, database connection, DATABASE_URL env var, and db/seed.ts (scaffold for seeding — run with npx tsx db/seed.ts). Do not overwrite the generated api/queries/connection.ts, drizzle.config.ts, or .env. Only add your tables to db/schema.ts.
• auth: Adds api/kimi/, auth router, client patches (Login, useAuth, AuthLayout, TRPCProvider), auto-wires Login/NotFound routes into App.tsx.

3. Verify Auto-Wiring

On first run, init.sh auto-wires TRPCProvider into src/main.tsx. When auth feature is installed, it also adds Login/NotFound routes to src/App.tsx. Check the init output:

• "Auto-wired" — no action needed
• "Wiring required" — complete the listed steps manually (see Post-Init Wiring)

If manual wiring is needed, do it in a worktree (created via swarm-workspace's setup-local.sh, which auto-copies the staged .env so npm run db:* has DATABASE_URL), edit, commit on the branch — the main agent merges it into the shared repo. (No push: worktrees share the shared repo's .git object store.)

4. Database Setup (if db or auth feature installed)

Run from a worktree, not OSS. npm run db:* needs DATABASE_URL, which lives in .env — setup-local.sh auto-copies the graft's staged .env ($REPO_PATH/.env) into every worktree, so this works without extra flags. ENV_SRC is shown explicitly below for clarity (and to override the source if ever needed); omitting it falls back to the same staged .env:

ENV_SRC=/mnt/agents/output/app/.env \
  NODE_MODULES_SRC=/app/.agents/skills/backend-building-swarm/scripts/template/node_modules \
  bash /app/.agents/skills/swarm-workspace/scripts/setup-local.sh main $HOME/app-main
cd $HOME/app-main && npm run db:push

5. Development & Build

# Build on a final-build worktree (correctness gate — catch type/build errors)
ENV_SRC=/mnt/agents/output/app/.env \
  NODE_MODULES_SRC=/app/.agents/skills/backend-building-swarm/scripts/template/node_modules \
  bash /app/.agents/skills/swarm-workspace/scripts/setup-local.sh final-build $HOME/app-final-build
cd $HOME/app-final-build && npm run build

Do not copy dist/ anywhere. The model records a version of the committed source via mshtools-website_version_manager (action: "build_version", type: "dynamic" — static serves the frontend only, without the server), and the platform builds the dynamic app server-side from that version's committed source in the worktree (project_dir) — which is why node_modules and .env must be present there (ENV_SRC above copies .env into the worktree). The tool is invoked per the system prompt, not by this skill.

Common Commands

Run npm run * commands inside a worktree ($HOME/app-<branch>), not on the shared coordination repo. Heavy steps (install, build, dev) would thrash its slow FUSE mount; and even the light db:* commands need the worktree's node_modules + .env, which are gitignored and therefore present only in worktrees — never in the coordination repo.

Stack Overview

Backend (added by this skill)

• Hono + tRPC 11.x
• End-to-end type safety
• Public query procedure (base); authenticated/admin procedures (auth feature)

Database (db feature)

• Drizzle ORM with MySQL
• Lazy getDb() connection
• Type-safe queries (Guide)
• Schema migrations

Authentication (auth feature)

• OAuth 2.0 (Details)
• JWT sessions, admin role support

Frontend (from webapp-building-swarm, untouched)

• React 19 + TypeScript + Vite (HMR)
• Tailwind CSS + shadcn/ui components
• Dark/light mode

Common Mistakes

• Don't run init.sh against /mnt/agents/output/app in a swarm setup — set PROJECT_PATH to the backend worktree created by setup-local.sh. Grafting directly in the shared repo edits/builds in the coordination hub.
• Don't forget to merge the backend branch — init.sh commits the graft on the branch but does not merge. The main agent must git merge backend before page branches fork, or page agents won't inherit the tRPC client.
• Don't run npm install / npm run build on /mnt/agents/output/app — it's coordination only. Use a worktree.
• Don't run git worktree prune in any worktree — it destroys other agents' worktree metadata in the shared .git/worktrees/ directory.
• Don't hand-write TypeScript interfaces for DB entities — use typeof table.$inferSelect from db/schema.ts. Hand-written createdAt: string will conflict with the actual Date objects delivered by superjson.
• Don't write raw SQL — always use Drizzle's type-safe query API (getDb().query.*, getDb().select(), getDb().insert()).
• Don't modify api/lib/ or api/kimi/ — framework internals. Build on top of them.
• Don't modify or re-create the auth client files — src/pages/Login.tsx, src/hooks/useAuth.ts, src/const.ts, src/components/AuthLayout*.tsx are installed by the auth graft and consumed as-is. The template Login.tsx is the only OAuth entry point; sign-in CTAs navigate to LOGIN_PATH from @/const rather than constructing OAuth URLs.
• Don't build a Register page — OAuth auto-provisions users on first login; there is no registration flow. Register/sign-up CTAs link to LOGIN_PATH.
• Don't import api/ from frontend code — use @contracts/ for types/constants that cross the boundary. The only exception is the type import in src/providers/trpc.tsx.
• Don't use api as the tRPC client name — it's trpc (imported from @/providers/trpc).
• Don't skip Zod validation on tRPC inputs — always use .input(z.object({...})) for mutations and parameterized queries.
• Don't use serial() for foreign key columns — serial() is bigint unsigned auto_increment, MySQL only allows one auto-increment column per table. FK columns must use bigint("col", { mode: "number", unsigned: true }).
• Don't use int() for foreign keys referencing serial() PKs — type mismatch.
• NEVER drop tables to fix a failed migration — the database may contain user data. Fix schema.ts and run npm run db:push to sync. See Troubleshooting.
• NEVER use db:push --force — auto-accepts destructive changes.
• NEVER modify .env values — generated by init.sh with valid credentials. Do not overwrite or placeholder-ify.
• NEVER overwrite init-generated infrastructure files — api/queries/connection.ts, api/middleware.ts, api/lib/, drizzle.config.ts, src/providers/trpc.tsx, .env. Build on top.
• Don't change the default port (3000).

AI Agent Instructions

Required reading before implementing features — read these docs (under docs/):

• db feature: Database.md, tRPC.md
• auth feature: Authentication.md, tRPC.md
• All features: Development-Guide.md, Project-Structure.md

When building features on this stack:

1. Read existing frontend code in src/ to understand the app structure (from a worktree)
2. Design tRPC routers that match the frontend's data needs
3. Add DB tables in db/schema.ts, then run npm run db:push from the worktree
4. Create new routers in api/ and register them in api/router.ts
5. Use type-safe queries via Drizzle ORM — never raw SQL
6. Frontend components go in src/ using the existing @/ alias

When to use which features:

• Database only: --features db
• User login: --features auth (includes db automatically)
• Default (no flag): auth
• Fullstack template: --template (features come from the template's .backend-features.json)

Key Paths

Full directory tree: Project Structure

Detailed Documentation

• Project Structure - Directory layout and where to put things
• Development Guide - Component development, API patterns, workflows
• Authentication - OAuth 2.0 implementation details
• Database Guide - Drizzle ORM configuration, schema design, migrations
• tRPC Best Practices - Router patterns and type safety
• Post-Init Wiring - Manual wiring fallback when auto-wire fails

Routing (react-router)

This template uses react-router v7.

import { useNavigate, useLocation, useParams, Link, Routes, Route, BrowserRouter } from "react-router";

See tRPC Best Practices and the frontend's App.tsx for current route definitions.

Troubleshooting

Port 3000 already in use — lsof -ti:3000 | xargs kill.

Database connection refused — Check DATABASE_URL in .env. Run npm run db:push before npm run dev.

User reports login failing — Verify VITE_KIMI_AUTH_URL and VITE_APP_ID in .env match the portal app config, and that src/pages/Login.tsx still matches the template (the only OAuth entry point).

npm run db:migrate fails — NEVER drop tables. Fix schema.ts, run npm run db:push, delete the broken migration file and its entry in db/migrations/meta/_journal.json, then npm run db:generate for a clean baseline.

init.sh errors that $PROJECT_PATH/src is missing — the worktree wasn't created or points at the wrong place. Create it first with swarm-workspace's setup-local.sh (NODE_MODULES_SRC=…/backend-building-swarm/scripts/template/node_modules … setup-local.sh backend $HOME/app-backend), and pass PROJECT_PATH=$HOME/app-backend.

git merge backend conflicts — the backend branch must fork from the merged scaffold; merge it before page branches fork (Phase 4.5, before Phase 5). If it diverged, re-graft on a fresh backend worktree off the current integration head.

Type errors after adding a new router — Make sure you registered it in api/router.ts inside appRouter.