Spike: a forge on Cloudflare, spawned with OpenTofu (#339)
Status: the module is written and its front door is tested; nothing has been
applied. This run had no tofu binary, no Cloudflare token and no bucket
credential (the run-secret injection it depends on was not declared for it), and
no bwrap. So tofu plan, apply and destroy are unproven, and each open
question below says whether it is evidence or reading. Nothing here claims
a measurement that was not made.
What exists
infra/forge-cloudflare/: Worker + Durable Object front door (worker.js,
tested in test/forge-cloudflare.test.ts), proxied DNS record, optional tunnel
for a body with no public address, S3 backend with use_lockfile and a
per-forge key. Names must match ^spike-…, so the module cannot address the
production tunnel or zone records. Every record carries a comment naming the
spike. The provider (v5) syntax was written from memory and is unchecked by
tofu validate.
Gap found: the Cloudflare provider has no Containers resource, so a Container
body is deployed with wrangler and passed in as body_url. "Apply produces a
forge" therefore needs two tools for that body, one for Fly (fly deploy).
The five questions
| Question | Answer | Basis |
|---|---|---|
| bubblewrap inside the container | Unknown; probably no. src/crew/sandbox.ts probes bwrap --unshare-user and falls back to unsandboxed. Containers run in a VM without user-namespace privileges we control; expect the probe to fail. Check: run the probe in a Container. | reading |
| SQLite WAL across snapshot/restore | Unknown. A disk snapshot taken mid-write holds a consistent main file plus -wal, which SQLite recovers on open, provided both are captured together. The forge's own snapshot (docs/SNAPSHOT.md) does not depend on disk snapshots, so restore via S3 sidesteps the question. | reading |
| HTTPS git instead of SSH | Yes for the forge's clients. Inbound is HTTP only on Containers, so FORGE_SSH_PORT=0 and clone/push over HTTPS with tokens. The app terminal (FORGE_APP_SSH) is off there. Fly keeps SSH. | reading |
| Does a crew run keep the container awake past 10 min idle | Unknown. Idleness is measured on inbound requests; a run is outbound work inside the container, so assume it does not count. Mitigation to test: the Durable Object pings the body every few minutes while a run is open, or sleepAfter is raised. | reading |
| 30-day snapshot retention | Fine for a cache, wrong for a record. The nightly S3 snapshot is the record; the platform snapshot is a warm-start only. | reading |
Verdict (provisional)
- Container body: right for a forge that is mostly idle, small, and used
from a browser: zero idle cost. Wrong for a forge whose crew runs long, wants bwrap, or wants SSH.
- Fly: the default for a forge people work in daily: a volume, raw TCP for
SSH, a full VM (bwrap likely works), scale-to-zero still available.
- Firecracker on metal: only when many forges share hardware and need
kernel isolation between them (docs/EXTRACT.md, _What is given up_); more
to run than a spike warrants.
Both bodies sit behind the same Worker front door, so choosing later costs a variable, not a rewrite.
celld could run that front door on our own machines; see
docs/research/celld.md.
To finish the spike
- Declare the Cloudflare token and bucket key as run secrets; install
tofu. tofu init/validate/plan/applywithspike-name; fetch/healthz,/mcp.- Run the bwrap probe, hold a run past 10 minutes, kill and restore the volume.
tofu destroy, then confirm the zone and account list nothing taggedspike-.