Launch

Launch — one command to a running stack

pnpm run bootstrap walks every step end to end: .env collection, Cloudflare connection, build, deploy, and a printed checklist for what only the dashboard can do (Access, Email Routing, Slack/Telegram/Discord app creation). Re-run it any time — answers persist in .env (gitignored, mode 600).

What the script does

1. Preflight — verifies a Docker daemon is reachable (container builds) and connects Alchemy to Cloudflare (browser OAuth once, or set CLOUDFLARE_API_TOKEN + CLOUDFLARE_ACCOUNT_ID in .env).
2. Collects config — dashboard Access emails, worker secrets (GITHUB_TOKEN, AI_GATEWAY_TOKEN, Slack/Telegram/Discord credentials), your paid-CLI subscription credentials (Claude, Codex, Cursor, Devin — blank keeps the lane dark), feature flags (SHIBA_*_SUBSCRIPTION, AGENT_HARNESS, AUTOMATIONS_ENABLED, MEMORY_ENABLED), and the agent mailbox address.
3. Builds and deploys — pnpm build then alchemy deploy. Everything in .env is bound as a secret or var; anything missing there is removed from the Worker on the next deploy, so .env is the source of truth.

After deploy (printed by the script)

• Cloudflare Access — until an Access app covers the hostname, every /api/* answers 401 access_not_configured by design (fail closed). Create it at one.dash.cloudflare.com → Access → Applications on app.tryshiba.dev (or your workers.dev host); ACCESS_EMAILS + WORKERS_SUBDOMAIN in .env pick up the audience automatically on redeploy.
• Custom domain — Workers → Settings → Domains → Add (e.g. app.tryshiba.dev); mirror the hostname into .env as WORKER_HOSTNAME.
• Agent mailbox — Email Routing → Custom addresses → create dev@tryshiba.dev (or your AGENT_MAILBOX) → route to Worker shiba-ai-coworker, then register it via POST /api/mailboxes. Unregistered recipients are rejected by design — register first.
• Sandbox image — the deployed Worker pulls the pushed registry image. On an arm64 machine: docker buildx build --platform linux/amd64 -t registry.cloudflare.com/<account-id>/shiba-ai-coworker-sandbox-dogfood:boot --push apps/backend. (The deploy-worker.yml workflow can build it too.)
• Chat channels — Slack via slack-app-manifest.yaml; Telegram via setWebhook to /api/telegram/webhook with TELEGRAM_WEBHOOK_SECRET; Discord via the Interactions Endpoint at /api/discord/interactions with DISCORD_PUBLIC_KEY.
• Check — https://<host>/api/setup/status after signing in.

When alchemy deploy can’t plan (narrow API token)

The Alchemy plan phase reads D1/Vectorize/KV/R2 state; a token scoped for Workers only gets 401s. Either widen CLOUDFLARE_API_TOKEN or deploy the proven wrangler path: node scripts/assemble-public.mjs → copy apps/backend/wrangler.jsonc to a sibling config with containers[].image swapped to the pushed registry ref → strip absolute-URL and /app*//dashboard* rules from public/_redirects → npx wrangler deploy -c <sibling> → delete the sibling. secrets() can’t be pushed by wrangler without .env values — set them with npx wrangler secret put <NAME>.

Rollback

npx wrangler rollback -c apps/backend/wrangler.jsonc returns the Worker to the previous version. Removing a key from .env and redeploying removes the binding; leave subscription flags blank to take a lane dark.