Roadmap
These are design directions, not supported commands or installed templates. The docs index and CLI/template help describe shipped behavior. Each addition must fit the existing named App workflow and extension boundaries.
Portable object storage
A small backend-neutral library could provide logical object references over GCS and S3-compatible storage, including R2. Integrate one real app and run the same conformance suite across private provider buckets before generalizing it.
Keep app data access outside the CLI. Existing stored provider URLs need a bounded compatibility reader with explicitly owned host/bucket mappings. Changing an infrastructure setting cannot rewrite shipped app code. Migration needs retained location metadata, resumable copying, checksums, ownership and deletion rules; no claim of provider portability removes the cost of moving bytes.
Edge lifecycle
The url-shortener Worker + D1 template and email-routing already ship.
Extend external runtimes through template-owned Extension.source adapters;
core must not branch on provider names. General Worker deployment history,
versioned rollback and portable object-storage bindings remain future work.
Acceptance needs exact account/name/environment ownership, readiness gates,
foreign-route rejection and recovery after partial provider failure.
Optional Soot operations
A named Soot template could add observation and human handoff. It would consume a reproducible versioned artifact and own its service lifecycle, private state, secrets, backup and probes; agent behavior belongs in Soot packs.
Begin with observation and bounded reporting through an explicitly selected monitoring receiver. Require durable case/outbox state, restart/dedup checks, model-failure handling and recovery updates. Do not grant a daemon the Docker socket or root control snapshot. Unattended repair needs operation identity, resource fencing and reconciliation after SSH loss before authority is expanded.
Security triage is the first concrete use case. The monitoring App already
emits the host security audit and routes
its alerts. A Soot on the same VM could take each alert group, explain it from
the read-only metrics, and hand the decision to a human: reboot, runtime
upgrade, unpublishing a port, or server harden --apply. The Soot itself
would execute none of these.
That design is blocked on two Soot-side pieces:
- An Alertmanager adapter. Soot accepts only its own event shape at
POST /v1/events. - An event-only credential. Today that endpoint takes the full operator token, and Alertmanager must not hold the token that manages the Soot.
Until both exist, Discord receivers remain the delivery path.
Authenticated origin pulls
The origin trusts Cloudflare by IP range in the provider firewall. Any Cloudflare zone, including one an attacker controls, can reach those ranges.
Per-zone authenticated origin pulls would close that gap. It needs three changes:
- a zone-scoped client certificate uploaded and enabled through the Cloudflare API;
- Caddy
client_auththat requires and verifies the certificate per site; - an ordered rollout: enable Cloudflare first, then require the certificate at the origin, with rollback in the reverse order.
Acceptance needs the following evidence:
- the token permission is documented;
- a direct-to-origin request without the certificate is refused;
- foreign zones are rejected;
- recovery works after a partial provider failure.
Read replicas
PostgreSQL remains single-VM without automatic failover. A future read-offload mode could use asynchronous physical replication to an explicitly supplied VM. Same-VM replicas add neither host capacity nor host-failure protection.
Prerequisites include private TLS connectivity, dedicated replication credentials, owned slots and volumes, bounded retained WAL, lag/health reporting, write rejection and an isolated bootstrap/reseed path. Reads must tolerate staleness; read-after-write stays on the primary. A replica is not a backup.
Provider-neutral hosts
Any Debian or Ubuntu VM reachable by SSH can run 2server Apps today. Host
hardening, the security audit and the metadata firewall assume no provider.
Provider behavior sits behind ssh.kind: gcp, the AWS vm: block and the gcp/aws
secret providers. Four gaps still favor GCP or AWS:
- Backups to S3-compatible storage. Postgres backups accept
gs://and AWSs3://only, with instance credentials. R2, B2, Spaces, Hetzner Object Storage and MinIO need an endpoint, static keys held as VM secrets, and a region pattern beyond AWS names. Until then, a Hetzner, DigitalOcean or bare-metal host has no supported backup path. - Managed backup storage (
backupStorage) creates GCS buckets only. - VM identity follows the SSH transport. Lifecycle, disks, origin IP and
backup storage recognise a GCP VM only through
ssh.kind: gcp. Avm: {kind: gcp}block besidevm: {kind: aws}would let a GCP VM reached by plain SSH keep them, withssh.kind: gcpas the fallback. - Metadata addresses beyond 169.254.169.254. The firewall blocks that IPv4 address on 80/443, which covers GCP, AWS, Azure IMDS, DigitalOcean, Hetzner and OpenStack. Azure WireServer (168.63.129.16), Alibaba (100.100.100.200) and AWS IPv6 (fd00:ec2::254) need a per-provider address list and ip6tables.
Acceptance needs a fresh-VM bootstrap, backup and restore drill on at least one non-GCP, non-AWS host.
Release gates
Prioritize current users: artifact installation, fresh-VM bootstrap, repeated release, failure preservation and isolated recovery. New capabilities need adverse-case tests and provider-specific staging evidence. Multi-cloud scheduling, a general plugin loader and autonomous database repair remain outside the current scope. Operator evidence and deployment-specific decisions belong in the consuming repository.
