Architect: `role-architect`. Date: 2026-05-23. Convoy decomposed into **1 brief** — the surface is two files (one script, one doc) and the fix is a single coherent change (env-var gate + README rewrite); splitting would force the README update to land first or last on its own, which adds review overhead without any parallelization win.
### File plan
| File | Action | Brief | Purpose |
| --- | --- | --- | --- |
| `scripts/setup-neon-db.js` | modified | 1 | Read `ADMIN_INITIAL_PASSWORD` env var at the top of `setupNeonDatabase()`. If unset or empty string, log a clear actionable error (names the env var, points at `.env.local`, suggests `openssl rand -base64 24`) and `process.exit(1)`**before** any DDL or DB connection. Replace `bcrypt.hash('admin123', 12)` with `bcrypt.hash(adminPassword, 12)`. Remove the trailing `console.log(' Admin User: admin@tcgvault.com')` and `console.log(' Admin Password: admin123')` lines so neither the literal nor the env-var value lands in stdout. Keep `ON CONFLICT (email) DO NOTHING` — see R4. Match the existing CJS `require()` style (the file's top is `#!/usr/bin/env node` + `require('dotenv').config(...)`; do **not** convert to ESM in this brief — that's a separate concern flagged in **Out-of-scope** below). |
| `README.md` | modified | 1 | Strip the "🔐 Default Admin Account" section (lines 115-119) entirely. Add `ADMIN_INITIAL_PASSWORD` to the install-step env-example block alongside `POSTGRES_URL` / `JWT_SECRET`, with a one-line note that it is required for `npm run setup-db` and can be set as a CI secret instead of `.env.local` if setup runs from CI. Add a short "First-time admin setup" paragraph under or near "Set up the database" that documents the env-var requirement, the `openssl rand -base64 24` tip, and the fact that operators of envs predating this convoy must rotate the admin password manually (Decision A). |
### API surface
**No new routes, no modified routes.** This convoy is purely operational — the seed script is a CLI tool (`npm run setup-db`) and `README.md` is documentation. The `/api/setup-database.js` route referenced in the original P0 #3 spec was already deleted by `fix-auth-bypass` Brief 3 (commit `fc0dd73`); CI's `forbidden-endpoints` job blocks reintroduction.
### Schema diff
**No schema changes.** No `CREATE TABLE`, no `ALTER`, no `INSERT INTO users` semantic change beyond *which password gets hashed and stored on first run*. Same column shape, same `ON CONFLICT` clause, same `bcrypt` cost factor (12). This is what Decision B turns on — see below.
### Test plan
**No vitest coverage added** (Decision C — C1). The setup script runs once per environment; the fail-loud env-var path is verified by **manual smoke** in the brief's acceptance criteria:
1.**Happy path:** With `ADMIN_INITIAL_PASSWORD=<value>` set in `.env.local`, run `npm run setup-db`. Expect the admin row to be created (or skipped if it already exists) and the script to print neither the literal `'admin123'` nor the chosen password value to stdout. Then `curl -X POST http://localhost:3000/api/auth/login -H 'content-type: application/json' -d '{"email":"admin@tcgvault.com","password":"<value>"}'` returns 200 with a JWT.
2.**Fail-loud path:** Unset `ADMIN_INITIAL_PASSWORD` (`unset ADMIN_INITIAL_PASSWORD` or comment it out in `.env.local`) and run `npm run setup-db`. Expect the script to print the error message and exit with code 1 **before** opening a DB connection.
3.**Idempotency:** With the admin row already present (weak `admin123` hash or otherwise), re-run setup-db with `ADMIN_INITIAL_PASSWORD=<different-value>`. Expect the admin row to be **unchanged** (`ON CONFLICT (email) DO NOTHING` short-circuits the INSERT). This is the documented behavior — rotation is out of scope (Decision A).
4.**README accuracy:** Read README top-to-bottom; confirm no occurrence of `admin123` remains, the new env-example block lists `ADMIN_INITIAL_PASSWORD`, and the "First-time admin setup" copy matches the script's actual behavior.
Implementer pastes the stdout from steps 1 + 2 into the PR description for the reviewer.
**Why not unit tests** (Decision C rationale): the script is run a handful of times per environment lifetime, the failure modes are loud (`process.exit(1)` + clear stderr), and unit-testing the env-var-required path would require either (a) extracting the admin-creation logic into a new `lib/seed-admin.js` module (scope expansion — Decision C option C3) or (b) spawning `node scripts/setup-neon-db.js` from a `vitest` test and asserting on stdout / exit code (slow, brittle, requires mocking `@neondatabase/serverless`). Manual smoke catches the same regressions at a fraction of the LOC cost, and the brief's acceptance criteria forces the implementer to actually run it.
### Risk list
- **R1 — Existing deployed admin rows are unchanged.** Any environment where `scripts/setup-neon-db.js` has already run with the old `bcrypt.hash('admin123', ...)` keeps the weak hash after this convoy merges, because `ON CONFLICT (email) DO NOTHING` skips the INSERT on re-run. **Mitigation:** Decision A (going-forward only). The brief's "Pre-merge operator checklist" and the README's new copy explicitly state that operators of pre-convoy envs must rotate the admin password manually (via the app's password-change UI or a future `rotate-default-admin` follow-up convoy). The PR description must include this callout for reviewers.
- **R2 — Operator runs `npm run setup-db` without setting the env var and the error is unhelpful.** A bare `throw new Error('missing')` would leave a new operator confused. **Mitigation:** the brief mandates a verbatim-near actionable message that names `ADMIN_INITIAL_PASSWORD`, points at `.env.local`, suggests `openssl rand -base64 24`, and notes the CI-secret alternative. Matches the wording shape Brief 1 of `fix-auth-bypass` used for `JWT_SECRET` in `lib/auth-secret.js`.
- **R3 — Setup-script stdout leaks the new admin password into CI logs.** The current script prints `Admin User: admin@tcgvault.com` and `Admin Password: admin123` after success. If we leave the second line and interpolate the env-var value into it, the chosen password lands in plaintext stdout — visible to anyone with CI log access, including the Vercel deploy log if setup ever runs there. **Mitigation:** the brief deletes both `console.log` lines outright. The success summary becomes `'✅ Admin user ready (email: admin@tcgvault.com)'` with no password echo. The chosen password is set only in the operator's env-var source (`.env.local` or CI secret), where access control already lives.
- **R4 — Changing `ON CONFLICT (email) DO NOTHING` to `DO UPDATE SET password = ...` would silently rotate every existing dev's admin password to whatever they put in `ADMIN_INITIAL_PASSWORD`.** This is **not** what we want; rotation is a separate concern (Decision A). **Mitigation:** the brief explicitly keeps `ON CONFLICT (email) DO NOTHING` unchanged and adds a verbatim acceptance-criterion line forbidding the change. A future `rotate-default-admin` follow-up convoy may add a separate idempotent migration script under `scripts/migrations/YYYY-MM-DD-*.js` if a real audit finds a deploy still carrying the weak hash.
- **R5 — README install-step env block omits the new variable.** A first-time operator following the README copies the env example, runs `npm run setup-db`, and hits the fail-loud error — confusing if the env example didn't mention `ADMIN_INITIAL_PASSWORD`. **Mitigation:** the brief explicitly updates the env block in step 3 of the README's installation section, not just the "Default Admin Account" section.
- **R6 — Three sibling files still hardcode `admin@tcgvault.com` / `admin123`.** Out of scope for this convoy (per the convoy spec's "Out of scope" list — these are historical scripts and a manual-QA doc): `scripts/reset-db.js` (lines ~141-156), `scripts/create-test-users.js` (line ~34), `TESTING_GUIDE.md` (line ~7). **Mitigation:** flag as a follow-up under § Anything flagged but not acted on below. The argument for *not* including them here: `reset-db.js` is a no-go-zone (`scripts/fix-*.js` family — historical / already-run), `create-test-users.js` is a dev seed for non-admin alice / bob / carol accounts where the weak admin reference is informational only, and `TESTING_GUIDE.md` is the test-data table for `create-test-users.js`. Sweeping them together would either widen the convoy's scope to "credential hygiene full sweep" or violate no-go-zones.
### Decomposition
| Brief # | Title | Files | Depends on | Estimated PR size |
| --- | --- | --- | --- | --- |
| 1 | Require `ADMIN_INITIAL_PASSWORD` env var; strip credentials from README | `scripts/setup-neon-db.js`, `README.md` | — | ~25 LOC net (10 added, 15 removed) |
No fan-out — both briefs touch the same script and ship in the same PR on `convoy/drop-public-setup`. `/multitask` is not applicable. Brief 2 must land *after* brief 1 because brief 2's "do not touch the env-var check block" acceptance criterion references brief 1's code shape verbatim.
### Decision A: Existing-admin rotation story → **A1 (going-forward only)**
We ship the env-var gate alone; rotation of any already-deployed weak `admin123` hash is left to manual operator action, documented in the README and the PR description. **Rationale (3 sentences):** this matches the pattern fix-auth-bypass Brief 1 used for `JWT_SECRET` — fix-loud the input path, do not sweep existing data, and queue any data-sweep work as a separate convoy when an audit demands it. The alternative (a one-time idempotent rotation script under `scripts/migrations/`) is sensible but doubles the brief count and introduces a new env var (`ADMIN_RESET_PASSWORD`) for a problem we cannot confirm exists in any specific deploy. If the queued `pick-a-name` / launch-audit convoys surface a real deploy still using the weak hash, the follow-up convoy `rotate-default-admin` is queued in § Anything flagged but not acted on for that purpose.
### Decision B: `no-go-zones` reading → **operational change is allowed**
The no-go-zones rule prohibits editing `scripts/setup-neon-db.js` for **schema changes** (`CREATE TABLE`, `ALTER`, new columns, idempotency-of-DDL concerns). **Rationale (3 sentences):** this convoy's edit is purely operational — it gates one bcrypt input on an env var and adjusts two `console.log` lines, with zero changes to DDL, table definitions, or the `ON CONFLICT` clause. That's structurally equivalent to Brief 1 of fix-auth-bypass, which added env-var fail-loud to `lib/auth-secret.js` without anyone treating it as a schema-rule violation. The brief documents the distinction explicitly in its "Conventions to follow" section so future agents reading either file see the reasoning.
### Decision C: vitest coverage → **C1 (no tests)**
The setup script runs a handful of times per environment lifetime; its failure modes are loud (`process.exit(1)` + clear stderr) and immediate; and the manual-smoke acceptance criteria in the brief cover both the happy and fail-loud paths. **Rationale (3 sentences):** unit-testing the env-var path would require either extracting a `lib/seed-admin.js` module (Decision C3 — scope expansion, ~2× brief LOC, plus a new file to maintain) or spawning `node scripts/setup-neon-db.js` from vitest and mocking `@neondatabase/serverless` (slow, fragile, low signal). Following the same heuristic the existing vitest suite uses — "tests cover the runtime auth surface that 30+ handlers depend on, not the operational scripts that run once per env" — leaves the test inventory focused on the highest-blast-radius surfaces. If a future convoy extracts seed logic for any other reason (e.g. multi-env seed templates), it can add tests at that point for free.
### Decision D: scope expansion to include CJS → ESM conversion → **Option B (expand this convoy)**
**Discovered mid-convoy:** the implementer for brief 1 confirmed that `scripts/setup-neon-db.js` does not actually run via `npm run setup-db` on Node 22.x. The `bump-next-js` convoy added `"type": "module"` to `package.json` (required for ESLint v9 flat config); the seed script still uses CJS `require()` calls and throws `ReferenceError: require is not defined in ES module scope` on first invocation. The architect's note in § "Anything flagged but not acted on" #1 — "it runs successfully today under Node 22" — was incorrect for Node 22.14.0.
**Decision:** expand this convoy to include brief 2 (`convert-setup-db-to-esm`) rather than queue a separate `convert-setup-db-to-esm` follow-up convoy. **Rationale (3 sentences):** brief 1's env-var gate is theatrical security on a script no operator can actually execute on Node 22.x, so the two fixes are logically coupled and shipping them in one PR creates a single coherent "setup-db is now both safe and functional" change. The CJS→ESM conversion is mechanical (~6 LOC, no functional changes) and touches the same file as brief 1, so review and audit overhead is near-zero. Splitting into two convoys would mean operators on Node 22.x cannot bootstrap a database between PRs — an unnecessary regression window for a fix that fits cleanly in the same surface area. User ratified the expansion (Option B) on 2026-05-23 after the implementer's gate-1 report surfaced the breakage.
These are real findings surfaced during architecture but **deliberately out of scope** for this convoy. Each should be tracked separately so the audit trail survives.
1.~~**CommonJS `require()` in an ESM package (`scripts/setup-neon-db.js`).**~~**Resolved by brief 2 (added 2026-05-23 mid-convoy per Decision D).** Original architect's claim that "it runs successfully today under Node 22" was incorrect — the implementer for brief 1 confirmed the script throws `ReferenceError: require is not defined in ES module scope` on Node 22.14.0. Scope was expanded to include the CJS→ESM conversion in this convoy rather than queue it as a separate follow-up. See § Decisions D for the ratification.
2.**Sibling weak-credential references in `scripts/reset-db.js`, `scripts/create-test-users.js`, and `TESTING_GUIDE.md`.**`reset-db.js` mirrors `setup-neon-db.js`'s admin-INSERT and password echo (lines ~141-156); `create-test-users.js` prints `admin@tcgvault.com / admin123 (Admin)` as a usage hint (line ~34); `TESTING_GUIDE.md` has the same row in its test-account table (line ~7). **Why not fixed here:** the convoy spec's "Out of scope" list explicitly excludes "any other `scripts/seed-*.js` or `scripts/add-*.js` files (those are one-off historical jobs per no-go-zones)", and `reset-db.js` is in the `scripts/fix-*` / historical family. `TESTING_GUIDE.md` is the manual-QA doc that doc-writer is expected to rename to `docs/MANUAL_QA.md` in the launch-polish convoy (P3, ship-readiness § Role-doc-writer findings) — folding it in here adds review overhead. **Follow-up suggestion:** queue a small `purge-weak-creds-from-helpers` convoy as part of the launch-polish phase, or roll it into `pick-a-name` (since the email itself is also changing).
3.**The success-summary `console.log(' Admin User: admin@tcgvault.com')` line.** This brief deletes it along with the password line (R3), but the admin email itself is still hardcoded in the SQL `INSERT` (line 135) and is going to be renamed under the queued `pick-a-name` convoy (P1 #12). Not a credentials issue, but worth noting that the email is still a known constant. **Why not fixed here:** branding decision belongs to `pick-a-name`; this convoy is credentials-only.
