# Minecraft server — itzg NeoForge

## Summary

`itzg/minecraft-server:java21`, NeoForge 1.21.1, offline-mode + authlib-injector
pointing at drasl. Mods come from a local `./server-mods` volume (mounted at
`/mods`), NOT packwiz — see `04-mods.md`. itzg installs MC + NeoForge + libraries
on first boot and auto-syncs `/mods` → `/data/mods` every boot; the stack assumes
the internet is present (for the NeoForge installer), so `TYPE=NEOFORGE` boots
normally every time.

## Server config

```
TYPE=NEOFORGE, VERSION=1.21.1, NEOFORGE_VERSION=21.1.x
REMOVE_OLD_MODS=TRUE          # ./server-mods is the authoritative mod set
volume  ./server-mods:/mods:ro
```

`./server-mods` is populated by `tooling/sync-server-mods.sh` (the `both`/`server`
subset of `mods-sides.json`, copied from `$DISTRIBUTION_WEB_ROOT`). First boot
installs MC + NeoForge + libraries into the `mc_data` volume; every boot itzg
mirrors `/mods` into `/data/mods` and prunes removed ones (`REMOVE_OLD_MODS`).

## Auth / config

```
ONLINE_MODE=TRUE
ENFORCE_SECURE_PROFILE=FALSE                 # pairs with drasl SignPublicKeys=false
JVM_OPTS=-javaagent:/extras/authlib-injector.jar=http://auth.${BASE_DOMAIN}/authlib-injector
```
authlib URL uses the `auth` subdomain; resolves internally via the caddy alias.
(Server still needs the agent even though Fjord clients have it built in.)

**`ONLINE_MODE` MUST be true.** authlib-injector works by redirecting the
server's normal online-auth handshake from Mojang to Drasl. With
`ONLINE_MODE=FALSE` the server skips that handshake entirely → players get
offline-hash UUIDs, no session validation, and **no skins** (the server never
fetches textures from Drasl). The thing that must stay *off* for 1.21+ is the
**secure profile** (`ENFORCE_SECURE_PROFILE=FALSE` + Drasl `SignPublicKeys=false`),
not online-mode. Symptom of getting this wrong: everyone can join but everyone is
Steve, and `usercache.json` shows offline-hash UUIDs.

Memory: `INIT_MEMORY 4G`, `MAX_MEMORY 10G`, `USE_AIKAR_FLAGS TRUE`.
RCON enabled for mc-backup.

## Server operators (OP)

**Do NOT use the itzg `OPS` env var.** It resolves usernames to UUIDs via a
Mojang lookup, but this server is `ONLINE_MODE=TRUE` validated against Drasl —
Drasl assigns its own UUIDs (not Mojang, not the offline-hash). A Mojang-resolved UUID
would not match the player's real session UUID, so the op entry would be inert
(and could clobber a correct `ops.json` on reboot).

**Op via RCON instead.** The player must have connected at least once so their
real Drasl UUID is cached in `/data/usercache.json`; then `op <name>` writes that
cached UUID to `ops.json`. On the host:

```
ssh cochi
docker exec minecraft rcon-cli list                 # confirm cache / who's on
docker exec minecraft sh -c 'cat /data/usercache.json'   # verify name→UUID cached
docker exec minecraft rcon-cli op <name>
docker exec minecraft sh -c 'cat /data/ops.json'    # verify level 4 + right UUID
```

`ops.json` lives in the `mc_data` volume → persists across restarts; no reboot
needed (takes effect live). Current op: `oier`
(`7fbc9adb-bcf4-3913-9a99-46d46acd1009`), level 4, set 2026-06-21.

## Simple Voice Chat — `voice_host` (NOT repo-tracked)

Voice chat is UDP `24454/udp`, published in compose. The mod auto-generates
`/data/config/voicechat/voicechat-server.properties` in the `mc_data` volume.

**Gotcha — `voice_host` must be set or remote clients get a red/crossed icon.**
Behind any NAT/Docker the server announces an address for clients' UDP
connection; if `voice_host=` is empty it auto-detects the internal Docker IP →
unreachable → the secret (TCP) is sent but the UDP handshake never completes
("Sent secret" with no following "successfully connected to voice chat").

Set on prod:

```
docker exec minecraft sed -i 's/^voice_host=.*/voice_host=ulicraft.net/' \
  /data/config/voicechat/voicechat-server.properties
docker restart minecraft   # mod reads props only at startup
```

**⚠️ This value lives in the `mc_data` volume, NOT auto-managed from this repo.**
It survives restarts and `compose down/up`, but a **world wipe / fresh `mc_data`**
regenerates the file with `voice_host=` empty → voice silently breaks again.

A known-good prod snapshot is committed at
`server-configs/voicechat-server.properties` (reference copy, `voice_host=ulicraft.net`).
After a world reset, either re-run the `sed` + `docker restart` above, or restore
the snapshot:

```
docker cp server-configs/voicechat-server.properties \
  minecraft:/data/config/voicechat/voicechat-server.properties
docker restart minecraft
```

(Snapshot is not auto-mounted — to make it durable, add a bind mount of
`server-configs/voicechat-server.properties` into `/data/config/voicechat/` in
compose.)

Verify after a client joins:
`docker logs minecraft 2>&1 | grep -i "connected to voice"` → expect
`Player <name> ... successfully connected to voice chat`.

Note: a red voice icon can ALSO be client-side — a VPN/zero-trust client
(Cloudflare One/WARP, FortiClient) on the player's machine silently drops the
voice UDP. Rule out VPNs before touching the server.

## Writable bind mounts (mods / config / kubejs)

These three used to be `:ro`, which meant edits could only flow host → container
and the server could never write back. They are now mounted **rw**:

| Mount | Host source | Notes |
|---|---|---|
| `/mods` | `./server-mods` | still overwritten by `tooling/sync-server-mods.sh` |
| `/data/config` | `./server-config` | NEW bind — was inside `mc_data` |
| `/data/kubejs/{server_scripts,startup_scripts,data,assets}` | `$DISTRIBUTION_WEB_ROOT/servers/ulicraft-1.21.1/files/kubejs/…` | writes land in the distribution git checkout |

Still `:ro` on purpose: `./runtime` → `/extras` (vendored jars) and
`CustomSkinLoader` (static json).

**Unchanged rule — the kubejs ROOT must stay writable.** Only the four *source*
subdirs are bind-mounted; kubejs writes `config/`, `cache/`, `generated/`,
`logs/` under `/data/kubejs` at boot, and mounting the whole dir makes
`BaseProperties.save()` NPE so kubejs never initialises. Dropping `:ro` does not
change that — the root is what matters, not the mode.

### Seed `./server-config` before first boot

A bind mount **shadows** what is under it. An empty `./server-config` hides the
existing `/data/config` and every mod regenerates its defaults — tuning silently
reverted. The old files are not destroyed (they stay in `mc_data`, under the
mount) but the server stops seeing them. So, once per host:

```
docker compose stop minecraft
tooling/export-server-config.sh          # oxps: MC_UID=1002 MC_GID=1002 tooling/…
docker compose up -d --force-recreate minecraft
```

The script halts rather than guessing: refuses if `./server-config` is non-empty,
if the container is still running, or if the `mc_data` volume name is ambiguous.

### Ownership — rw mounts need the server's uid

itzg runs the server as **uid 1000**. A writable bind mount is only writable if
the host dir permits uid 1000; otherwise the server logs permission errors (or
kubejs/config writes fail) even though the mount is rw. On `cochi` the host user
is uid 1000, so this is free. On `oxps` the host user `oier` is **uid 1002** —
`./server-config` gets chowned to `MC_UID`/`MC_GID` by the export script, so pass
the right values there, and expect the dir not to be editable as your own user
without sudo or a shared group.

`$DISTRIBUTION_WEB_ROOT` (kubejs) and `./server-mods` must satisfy the same rule
now that they are rw.

### Consequence — kubejs edits mutate the distribution repo

`/data/kubejs/{server_scripts,startup_scripts,data,assets}` are the live
distribution checkout. Anything the server writes there (or an in-game
`/kubejs reload` regenerating a file) shows up as an **uncommitted change in the
distribution repo**, on that host only. `cochi` and `oxps` hold separate
checkouts that already drift — check `git status` in `$DISTRIBUTION_WEB_ROOT`
before pulling or publishing from either.

## Accessing the `mc_data` volume

`mc_data` is a **named docker volume**, not a bind mount — there is no path in
this repo that mirrors the world. Same on both hosts (`cochi`, `oxps`); on `oxps`
the stack lives at `~/mc/ulicraft-server` and the host is LAN-only, so `ssh oxps`
works from the LAN only.

Preferred — go through the container (`container_name: minecraft`, world at
`/data`):

```
ssh oxps
docker exec -it minecraft bash          # interactive poke around /data
docker exec minecraft ls /data/config   # one-shot read
```

Copy files in/out without root:

```
docker cp minecraft:/data/logs/latest.log ./latest.log
docker cp ./foo.json minecraft:/data/config/
```

Host path (root-owned, needs sudo). The volume name is prefixed with the compose
project name, which defaults to the **directory name** — `ulicraft-server_mc_data`
unless the checkout dir differs, so resolve it rather than hardcoding:

```
docker volume ls | grep mc_data
docker volume inspect ulicraft-server_mc_data -f '{{.Mountpoint}}'
#   → /var/lib/docker/volumes/ulicraft-server_mc_data/_data
sudo ls /var/lib/docker/volumes/ulicraft-server_mc_data/_data
```

Root-free equivalent when the container is down (also the shape used by the
`chown` fix in `plan/22-migration.md`):

```
docker run --rm -v ulicraft-server_mc_data:/data alpine ls /data
```

**⚠️ Do not write world files while the server is running** — the running server
holds region files open and will overwrite or corrupt concurrent edits. Stop it
first (`docker compose stop minecraft`), edit, then start. Reads are fine.

For a consistent snapshot of the world, prefer `backups/` (mc-backup mounts
`mc_data:/data:ro` and quiesces saves over RCON — see `plan/07-mc-backup.md`)
over copying `/data/world` out from under a live server.

## Tasks

- [ ] Compose: authlib URL `http://auth.${BASE_DOMAIN}/authlib-injector`
- [ ] Compose: mount `./server-mods:/mods:ro` + `REMOVE_OLD_MODS=TRUE`
- [ ] Run `tooling/sync-server-mods.sh` before bring-up (populates `./server-mods`)
- [ ] Mount `./runtime` for authlib-injector.jar at `/extras`
- [ ] `depends_on`: drasl (authlib must resolve at boot)
- [ ] Confirm `ENFORCE_SECURE_PROFILE=FALSE` ↔ drasl `SignPublicKeys=false`
- [ ] Verify a client joins with their drasl skin
- [ ] After any world wipe: re-set `voice_host` (see Simple Voice Chat section)