Convex Quickstart (Codex beta)

Stand up a running Next.js + shadcn "wow-shell" Convex app from one sentence, then build the user's idea live (locally). The heavy scaffold runs as a served shell script from the Convex quickstart backend ("anteater"); your job is to launch it, then build.

Auth (passkeys), the feedback panel, custom domains, and public *.convex.app publishing are de-scoped for this release — they ship later. The scaffold runs locally only, with no login and no panel. (The error-watch monitor in STEP 5 still applies and is valuable — keep it.)

The user's request after $quickstart is the app idea (e.g. $quickstart a movie-night voting app → idea = "a movie-night voting app"). If no idea was given, ask for a one-sentence idea, then continue.

Degradation rule — when the scaffold can't run, write code, not ceremony

If the bootstrap can't run — a non-interactive/one-shot session, no network access, a sandboxed temp dir, or the user just wants code rather than a running app — don't wait on the scaffold. Write a standard Convex project directly:

• ALL backend code goes under convex/ (schema.ts, queries, mutations, actions) — NEVER at the project root. Convex functions only run from the convex/ directory.
• Write ZERO scaffold/documentation files unless explicitly asked — no START_HERE.md, ARCHITECTURE.md, MANIFEST.txt, or README walls. "Build me a backend" is a request for code, not a design-doc package.

Data access + imports — read before writing any convex/*.ts

• Never an unbounded .collect() on a table that can grow — use .withIndex(...) + .paginate(paginationOpts)/.take(n).
• Index, don't filter — .index(...) in schema.ts for every read path, queried via .withIndex(...); .filter() is a full table scan.
• Imports: query/mutation/action/internalQuery/internalMutation/internalAction from "./_generated/server"; api/internal from "./_generated/api"; never from "convex/server" in application code.
• v.literal("exact value") for fixed string/enum members, not a bare v.string().
• "use node"; is action-only — never in a file that also exports a query or mutation.

Self-verify — before declaring backend work done

Before you call any backend work finished: run npx tsc --noEmit and, when a deployment is available (or via a local anonymous one: CONVEX_AGENT_MODE=anonymous npx convex dev --once), push it. Fix every error either one reports before finishing — one verify round catches the wrong-relative-import / duplicate-symbol / unbalanced-paren class that otherwise breaks the deploy.

STEP 0 — launch the scaffold NOW (before anything else)

Run this first, before any reasoning or other tool calls — it kicks off the ~45–120s scaffold (npm install, convex dev, next dev) in the background so it's installing while you read the rest. Substitute the user's idea for <IDEA>:

BASE="https://basic-anteater-667.convex.site"
IDEA="<IDEA>"
SLUG=$(curl -fsS --max-time 15 -X POST "$BASE/generate" -H 'content-type: application/json' \
  --data "$(node -e 'process.stdout.write(JSON.stringify({idea:process.argv[1],template:"nextjs-shadcn"}))' "$IDEA")" \
  | node -e 'let s="";process.stdin.on("data",d=>s+=d).on("end",()=>{try{process.stdout.write(JSON.parse(s).id||"")}catch{}})') || true
echo "SLUG=$SLUG"
QB=$(mktemp -t convex-qb-XXXX.sh)
curl -fsS --max-time 20 "$BASE/quickstart-bootstrap" -o "$QB" || { echo "BOOTSTRAP_FETCH_FAILED"; exit 3; }
# The bootstrap is feature-flagged via a profile. We ship the MINIMAL profile: scaffold
# only — no auth/passkeys, no feedback panel, no custom domain, and public *.convex.app
# publishing disabled (they ship later). To restore the goodness pass QB_PROFILE=full
# (or individual flags, e.g. QB_PASSKEYS=1 QB_PANEL=1 QB_DOMAIN=1).
echo "AUTH_MODE=none"   # minimal profile = no pre-baked auth; the build has no login
# QB_HARNESS=codex tags telemetry. QB_ARGS_BASE=$BASE is CRITICAL: the slug was
# generated on THIS deployment, so the bootstrap must fetch the personalized args +
# bespoke runbook from the SAME host (its default is prod, which 404s a staging slug
# → generic "My Convex App" defaults).
nohup env QB_PROFILE=minimal QB_HARNESS=codex QB_ARGS_BASE="$BASE" QB_FEEDBACK_URL="$BASE/feedback" \
  bash "$QB" $SLUG > .quickstart-bootstrap.log 2>&1 &
echo "SCAFFOLD_LAUNCHED log=.quickstart-bootstrap.log SLUG=$SLUG"

• If it prints SCAFFOLD_LAUNCHED, the scaffold is running in the background. Do NOT run it again. Note the SLUG=.
• If curl is blocked or you see BOOTSTRAP_FETCH_FAILED, the network/sandbox blocked it — tell the user they likely need to run Codex with network access / auto-approve (codex --sandbox danger-full-access), then retry.

STEP 1 — wait for the scaffold, open the browser

Poll .quickstart-bootstrap.log until it contains BOOTSTRAP_COMPLETE.

Codex's sandbox often reaps backgrounded (nohup … &) processes when the launch call returns — so the bootstrap may write its first line, then die before scaffolding. If within ~20s the log has stalled (no new lines), no app subdirectory has appeared, and there's no BOOTSTRAP_COMPLETE, the background launch was reaped. Recover by running the bootstrap in the FOREGROUND — re-run the STEP 0 block but replace the nohup env … & line with a plain foreground run, same env:

QB_PROFILE=minimal QB_HARNESS=codex QB_ARGS_BASE="$BASE" QB_FEEDBACK_URL="$BASE/feedback" bash "$QB" $SLUG

It backgrounds convex dev / next dev itself and returns at BOOTSTRAP_COMPLETE in ~1–2 min (set a generous command timeout, 300s+). BOOTSTRAP_FETCH_FAILED → server unreachable; tell the user. When it completes the log prints:

• OPEN_BROWSER_URL: http://localhost:<port> — open this for the user immediately.
• The app is scaffolded in a new subdirectory with convex dev + next dev running and error watchers armed (convex-errors.log / next-errors.log paths are in the log).

STEP 2 — read the runbook + build the idea live

Read the personalized runbook for the full build flow (it's served — fetch it):

curl -fsS "https://basic-anteater-667.convex.site/q/$SLUG.md"

Then build the user's idea following it. What's already done by the scaffold:

• No auth this release — the scaffold ships with no login. Only add auth if the user explicitly asks (then wire their requested provider via the convex-expert skill). (Passkeys are de-scoped — QB_PROFILE=minimal sets QB_PASSKEYS=0.)
• No feedback panel this release — narrate your build in chat; do NOT call progress:post / todos:* / refinementQuestions:*. (Panel is off — QB_PANEL=0.)
• Publishing is disabled this release — the app runs locally at the printed URL; do not publish or run $add hosting.

Rules while building:

• Delegate all code inside convex/ to the convex-expert skill's rules (object-form syntax, validators, indexes, internal vs public).
• Watch for convex/ + next errors and fix them as they appear — the easiest way is the fix_errors_automatically tool (see STEP 5), which surfaces them as events.

STEP 3 — wrap up (runs locally; no publish this release)

When the app builds clean and the core feature works (your "v1"), tell the user it's running locally at the printed URL and summarize what you built. Publishing to a public *.convex.app URL and custom domains are disabled this release — don't deploy, don't run $add hosting, and don't look for .quickstart-domains.json.

STEP 5 — stay on watch with fix_errors_automatically (start EARLY, don't yield)

This harness has no push: a runtime error (or a slow async result) sits unseen until you actively look. This plugin bundles a convex-plugin MCP server with one blocking tool that surfaces it as an event — keep using it even though the feedback panel is off this release (the convex_error / next_error watch is the valuable part).

Start watching as soon as the app is open (right after STEP 1) — not just after v1. The user is most engaged at the very start and will often submit a request or question while you're still building. Call fix_errors_automatically between build milestones and whenever you'd otherwise pause; on the FIRST call it returns immediately if anything is already pending (it surfaces existing unhandled requests/questions as existing: true), so feedback submitted before you started watching is never lost. Then make it your standing idle action for the rest of the session.

Avoid the repeated approval prompt: each call is a separate tool invocation, so a one-time "Allow" re-asks every loop. Choose "Always allow" when prompted, or add to ~/.codex/config.toml: [mcp_servers.convex-plugin] then trust_level = "trusted". Also pass a long timeoutMs (e.g. 280000) so each block is long and you re-call less often.

Call the fix_errors_automatically tool with projectDir set to your app's absolute path (the convex-app/ subdir, where .env.local lives). It BLOCKS until the next event, then returns one of:

• feature_request / refinement_answer → the user wants something — handle it (loop back through STEP 2/3), then call the tool again.
• convex_error / next_error → stop, read the line, fix the file, wait for convex dev / Next to recover, then call again.
• quiet → heartbeat timeout, nothing happened — just call it again.

Loop on this tool instead of ending your turn. Each call blocks, so you stay on watch and react within seconds — no polling, no walking away. Keep looping until the user clearly says they're done.

After the run

Offer $improve-convex-plugin so the user can send this session for an AI review that improves the quickstart.

On failure / friction — offer to self-report (WITH the user's permission)

If the run doesn't go smoothly — you give up, hit a blocker, it takes way too long, or the user is frustrated — don't yield silently. Ask first, send only on a yes:

"That didn't go as smoothly as it should have. May I send this session — the transcript, with secrets redacted — to the Convex team so they can fix what went wrong? yes / no"

• On yes → run the improve-convex-plugin flow (ships the redacted transcript): curl -fsSL "https://basic-anteater-667.convex.site/send-transcript" | QB_HARNESS=codex bash -s -- --idea "<what broke>"
• On no → respect it; send nothing.
• Either way, give the user an honest summary of where it got stuck.

Never send the transcript without an explicit yes.