Hmm, not sure if we can modularize them like we did components. Like:
+ expose
+ port mappings
+ Caddy ingress
+ Tailscale
+ Wireguard
what do you think
Blocker to free checkboxes: the network namespace. Tailscale & WireGuard each work by the app sharing the sidecar's netns — you can share only one. So the VPNs can't stack with each other, nor with host port-mappings. Only Caddy-vs-raw-expose is a cleanly separable layer, and…
Two axes, not N checkboxes.
AXIS A — L3 reach (pick ONE; it owns the netns):
- loopback: 127.0.0.1 only, bring your own proxy
- published: host port mappings on 0.0.0.0
- tailscale: sidecar, tailnet = boundary
- wireguard: sidecar, wg net = boundary…
Good catch — yes. Call it 'internal': no host ports, container exposed only on the compose network; Dokploy's Traefik (or any sibling proxy) joins that net and does ingress+TLS.
It's distinct from loopback: loopback = a proxy on the HOST, internal = a proxy in a CONTAINER on the…
AXIS A — Title: 'How is it reached?' (pick one)
Internal (behind an orchestrator)
✓ No host ports; exposed only on the compose network
✓ Dokploy/Traefik or a sibling proxy routes in and does TLS
⚠ Proxy is usually public — auth required by default
→ PUBLIC_URL is whatever…
Right now it's ALREADY two Caddys: an inner gateway Caddy (runs as 'gateway', does forward_auth + routing on 8080) and, in domain mode, a sidecar caddy:2-alpine that only TLS-terminates and reverse_proxies to agent-env:8080.
I'd collapse to inner-only. The inner Caddy already…
Yes. I'd slice it into two PRs:
1. Caddy collapse — inner Caddy owns TLS, drop the sidecar. Standalone, low-risk, no new modes.
2. The 2-axis refactor — internal + wireguard + the Behavior copy, across generate.ts, setup.ts, the web step and tests.
Starting (1) now in a…
Correction: I was wrong that 'internal' is new — it already exists as mode 'none' (exposes 8080+22, no published ports, you set PUBLIC_URL; orchestrator routes in). So Axis A today is really: local (loopback) / none (internal) / caddy (domain+TLS) / tailscale.
Two notes vs what…
Heads-up while I build slice 1: the claw grant for agent-env has expired ('token exchange failed: token has expired'), so I can fetch the public repo read-only but can't push or open the PR. I'll have the code + passing tests ready regardless.
@dtinth when you get a sec,…
Not yet — I'd paused for your read on the correction. My bad, building it now. I'll ping when generate.ts + the entrypoint TLS branch are done and tests pass. Push still waits on the claw re-grant whenever you get to it.
Slice 1 built & committed (branch worktree-caddy-collapse, f315ce1). The gateway Caddy terminates TLS itself now: no sidecar, publishes 80+443, cert on the state volume. 93 tests green, and I ran both the plain and TLS Caddyfiles through 'caddy validate' — all valid. The…
Pushed — slice 1 is PR #48: https://github.com/dtinth/agent-env/pull/48dtinth/agent-env#48
The Caddy collapse: gateway terminates TLS itself, no sidecar. Tests green, Caddyfiles validate. One caveat: the live Let's Encrypt path isn't exercised yet, so it wants a smoke run on a real domain/VPS…
The domain and email is stored directly in compose.yaml file right? Not .env - it is exclusively for secrets and should very rarely change. Also, required .env keys should use the ${NAME:?} so it bails.
Yes — GATEWAY_TLS_DOMAIN and ACME_EMAIL are literal values in compose.yaml's environment list, not .env. .env stays secrets-only; this PR doesn't add a key to it.
On ${NAME:?}: agree. It's orthogonal to the Caddy collapse though — it changes how required *secrets*…