# sish-client

A small, environment-driven connector for a [sish](https://github.com/antoniomika/sish) edge.
It opens SSH reverse tunnels to the edge, claims hostnames there, and relays their traffic to
local targets. In the default mode sish routes by **SNI without terminating TLS**, and each
connection starts with a **PROXY v2 header** carrying the real client address.

```
client ──TLS──▶ edge :443 (sish, SNI routing) ══ssh══▶ sish-client ──▶ target (terminates TLS)
```

Image: `registry.traberph.de/public/sish-client` (Alpine + OpenSSH client + tini, ~20 MB,
runs as `nobody`, works with a read-only root filesystem).

## Quick start

```sh
docker run -d --name sish-client --read-only --tmpfs /tmp \
  -e SISH_HOST=edge.example.com \
  -e SISH_HOST_KEY="ssh-ed25519 AAAA..." \
  -e ROUTES="app.example.com=app:8443" \
  -v ./id_ed25519:/secrets/id_ed25519:ro \
  registry.traberph.de/public/sish-client:latest
```

The connector's **public** key must be authorized on the edge (sish `--authentication-keys-directory`),
and each hostname must be allowed there (sish `--bind-hosts` or `_sish` TXT records).

## Configuration

| Variable | Default | Description |
|---|---|---|
| `SISH_HOST` | *required* | Edge hostname or IP |
| `SISH_PORT` | `2222` | sish SSH port |
| `SISH_USER` | `connector` | SSH user (sish accepts any; shows up in edge logs) |
| `SISH_KEY` | | Private key contents (e.g. from a Kubernetes `secretKeyRef`). Takes precedence over `SISH_KEY_FILE` |
| `SISH_KEY_FILE` | `/secrets/id_ed25519` | Path to a mounted private key |
| `SISH_HOST_KEY` | | Edge host public key, `"<type> <base64>"` (the first two fields of the `.pub` file) |
| `SISH_KNOWN_HOSTS_FILE` | | Alternative to `SISH_HOST_KEY`: a known_hosts file matching `[SISH_HOST]:SISH_PORT` |
| `SISH_STRICT_HOST_KEY` | `true` | `false` disables host key verification (testing only) |
| `ROUTES` | | Routes, separated by commas, spaces or newlines (see below) |
| `ROUTES_FILE` | | File with routes, one per line, `#` comments allowed. Combined with `ROUTES` |
| `TARGET` | | Default `host:port` for routes without an explicit target |
| `DEFAULT_BIND_PORT` | `443` | Edge port for routes without an explicit bind port |
| `SNI_PROXY` | `true` | Ask sish to route by SNI instead of terminating TLS |
| `PROXY_PROTOCOL` | `2` | PROXY header version sent to the target: `1`, `2` or `off` |
| `KEEPALIVE_INTERVAL` | `15` | Seconds between SSH keepalives (3 missed = reconnect) |
| `CONNECT_TIMEOUT` | `10` | SSH connect timeout in seconds |
| `BACKOFF_MIN` / `BACKOFF_MAX` | `2` / `60` | Reconnect backoff in seconds (doubles per failure, resets after a successful session) |
| `SSH_EXTRA_OPTS` | | Extra `ssh` arguments, e.g. `-o Ciphers=aes128-gcm@openssh.com` |
| `STATE_DIR` | `/tmp/sish-client` | Writable directory for the key copy, known_hosts and the ready marker |

### Routes

```
<host>[:<bind-port>][=<target-host>:<target-port>]
```

| Example | Meaning |
|---|---|
| `app.example.com=traefik:8443` | SNI `app.example.com` on edge :443 → `traefik:8443` |
| `app.example.com` | Same, using `TARGET` |
| `*.apps.example.com=traefik:8443` | Wildcard (sish needs `--bind-wildcards`) |
| `app.example.com:80=traefik:8080` | Plain HTTP on edge :80, routed by `Host` header (**needs `SNI_PROXY=false`**, see below) |

**HTTP (:80) and SNI (:443) routes cannot share one connector.** With `SNI_PROXY=true` sish
treats every forward as an SNI listener, which collides with its HTTP listener on :80. Run a
second instance for HTTP routes (e.g. for HTTP→HTTPS redirects):

```sh
SNI_PROXY=false PROXY_PROTOCOL=off ROUTES="app.example.com:80=traefik:8080"
```

If the edge rejects any route, the whole session fails (`ExitOnForwardFailure`) and is retried
with backoff. A misconfigured route is therefore loud (visible in the logs and health status)
instead of silently missing.

Several connectors may claim the same hostname. sish then load-balances between them, which is
how you get redundancy (e.g. one connector per ingress replica).

## Health

The container writes `$STATE_DIR/ready` once sish has confirmed **every** route, and removes it
when the tunnel drops. The image's `HEALTHCHECK` tests this file; in Kubernetes use it as a
readiness probe:

```yaml
readinessProbe:
  exec:
    command: ["test", "-f", "/tmp/sish-client/ready"]
  periodSeconds: 5
```

## Examples

### Docker Compose

```yaml
services:
  app:
    image: nginx   # must terminate TLS and accept PROXY protocol on 8443
  connector:
    image: registry.traberph.de/public/sish-client:latest
    read_only: true
    tmpfs: [/tmp]
    cap_drop: [ALL]
    security_opt: [no-new-privileges:true]
    environment:
      SISH_HOST: edge.example.com
      SISH_HOST_KEY: ssh-ed25519 AAAA...
      ROUTES: app.example.com=app:8443
    volumes:
      - ./id_ed25519:/secrets/id_ed25519:ro
    restart: unless-stopped
```

The key file must be readable by uid `65534` inside the container. Compose ignores `uid`/`mode`
on file secrets, so either:
- make the key readable by a group and run with `user: "65534:<gid>"`. With **rootless** Docker
  your host files appear as owned by root inside the container, so the gid is `0`.
- or pass the key via `SISH_KEY` instead of a file.

### Kubernetes sidecar (next to an ingress controller)

```yaml
- name: sish-client
  image: registry.traberph.de/public/sish-client:20260926154233-1a2b3c4 # {"$imagepolicy": "flux-system:sish-client"}
  env:
    - { name: SISH_HOST, value: edge.example.com }
    - { name: SISH_HOST_KEY, value: "ssh-ed25519 AAAA..." }
    - { name: ROUTES_FILE, value: /etc/sish-client/routes.txt }
    - { name: TARGET, value: "127.0.0.1:9443" } # ingress entrypoint with PROXY protocol
    - name: SISH_KEY
      valueFrom: { secretKeyRef: { name: sish-client, key: id_ed25519 } }
  securityContext:
    runAsNonRoot: true
    readOnlyRootFilesystem: true
    allowPrivilegeEscalation: false
    capabilities: { drop: [ALL] }
  volumeMounts:
    - { name: tmp, mountPath: /tmp }
    - { name: routes, mountPath: /etc/sish-client, readOnly: true }
  readinessProbe:
    exec: { command: ["test", "-f", "/tmp/sish-client/ready"] }
```

Routes are only read at connect time. Roll the pods after changing them (e.g. with a config
hash annotation).

## Image tags

The Gitea workflow (`.gitea/workflows/build.yaml`) builds `linux/amd64` and `linux/arm64` and pushes:

| Trigger | Tags |
|---|---|
| Push to `main` | `<yyyymmddHHMMSS>-<sha7>` (UTC) and `latest` |
| Git tag `v*` | `<yyyymmddHHMMSS>-<sha7>` and the git tag, e.g. `v1.0.0` |

Repository secrets required: `REGISTRY_USERNAME`, `REGISTRY_TOKEN` (a token with push access
to `registry.traberph.de/public`). The runner needs Docker (Buildx + QEMU for arm64).

### Flux image automation

```yaml
apiVersion: image.toolkit.fluxcd.io/v1
kind: ImageRepository
metadata: { name: sish-client, namespace: flux-system }
spec:
  image: registry.traberph.de/public/sish-client
  interval: 10m
---
apiVersion: image.toolkit.fluxcd.io/v1
kind: ImagePolicy
metadata: { name: sish-client, namespace: flux-system }
spec:
  imageRepositoryRef: { name: sish-client }
  filterTags:
    pattern: '^(?P<ts>[0-9]{14})-[0-9a-f]{7}$'
    extract: '$ts'
  policy:
    numerical: { order: asc }
```

## Build locally

```sh
docker build -t registry.traberph.de/public/sish-client:dev .
```