Cloudflare deployment

Alchemy declares production resources in alchemy.run.ts. Shared queue and rate-limit settings live in infra/bindings.ts; pnpm run infra:wrangler generates local Wrangler configurations from those settings.

Deploy

Configure the deployment credentials and auth environment described in the deployment runbook (docs/deploying.md in the local checkout), then run:

pnpm run deploy

This invokes Alchemy for the prod stage. The stack contains three Workers, a shared D1 database, a SQLite Durable Object namespace for private Assistant Conversations, background queues, and rate-limit bindings. Configured providers add resources such as the email binding and workspace-export bucket. The web Worker exports the conversation host; API and background bindings address that same host within the stage. D1 stores directory metadata and quotas, while each conversation object stores its transcript and replay.

See optional providers.

Assistant bindings and providers

ASSISTANT_CONVERSATIONS is a SQLite Durable Object binding, declared by Alchemy with its class migration. Keep the named host export in the built web Worker and point API/background cross-Worker bindings at that stage's web script. Generate Wrangler declarations from infrastructure sources; do not edit generated bindings.

Conversation storage works with local persisted D1 and Durable Object state. Remote model credentials are optional: browsing and management work without them, while new generation refuses until a provider is configured. Member REST access also needs the separate assistant OAuth audience. See Private Assistant Conversations.

After deploying an isolated stage, verify disconnect/reconnect, deployment during an answer, explicit Retry and provider cancellation. Local tests and bundle checks are separate from those deployment and provider checks.

Pull-request previews

The Preview workflow deploys same-repository pull requests to pr-<number>, with a separate database, conversation object namespace, queues, and Workers. It migrates and seeds the database and updates a pull-request comment with the URLs. Closing the pull request destroys its stage.

Previews disable optional external providers and set ENVIRONMENT=preview. To manage a preview from a configured deployment shell:

ALCHEMY_STAGE=pr-42 pnpm run deploy:stage
ALCHEMY_STAGE=pr-42 pnpm run db:seed:stage
ALCHEMY_STAGE=pr-42 pnpm run destroy:stage \
  --confirm-target="$CLOUDFLARE_ACCOUNT_ID/pr-42"

Recovery and teardown

Use the recovery runbook (docs/backup-recovery.md in the local checkout) for rollback and restore procedures. Check schema and binding compatibility before rolling back Worker code.

pnpm run destroy --confirm-target="$CLOUDFLARE_ACCOUNT_ID/prod" deletes the production stack. It is a teardown command, not a code rollback.