horizon-client's cpal output and its controllers (gilrs, the DualSense's HID output) move behind the
audio and controllers features, on by default, so the game is unchanged. Built with
--no-default-features it links neither libasound.so.2 nor libudev.so.1. release.yml packages that
build as horizon-engine in horizon-server-linux-x86_64.tar.gz (xtask package-server, checked by
check_imports.py --server); install.sh installs it on Linux (and --archive installs a local one);
horizon-server already prefers horizon-engine. The Dockerfile uses it too, and CI lints it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: reviewed; verify passes (client --help exits 0 without loading the game, so the start check only catches a program the system can't load). Linux installs now name the missing ALSA/udev libraries; services run through the installer's current link.
The first published benchmark for the working version (docs/server.md
section 10, item 4). It ran the v2026.10.18.15 release, as install.sh
installs it, against vanilla 26.3 with 1, 10, 25 and 50 mixed bots. The
servers ran on the ThinkPad (quiet, under nice 19) and the bots on the
Mac mini over the tailnet. Vanilla's default game rules applied on both;
Horizon's server runs the locator bar now.
Both servers kept 20 TPS throughout. Horizon used 3.7 to 5.2 times less
memory. Its average tick was shorter with 1, 10 and 25 bots and 18 %
longer with 50, while it generated about twice the new terrain a second
(the exploring bots fly as fast as each server sends chunks). Its P99
was lower at every count.
benchmarks/server-mixed-2026-10-18/ holds the README, the report and the
machine's load; docs/server.md links it from sections 9 and 10.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On the server thread, lan_e2e --check light failed its border step (#891):
the guest's border chunk still had block light 0 next to glowstone placed
in the chunk past it. The server was right; the check was not.
same_light() waits until the host's light equals the guest's, but on the
server thread the host's game is a client too, and both are dark until
the light update arrives. They agreed at once, same_light returned, and
the next line read the guest's light before the update came.
The check now waits up to 100 ticks for the guest's border light, then
compares the host's with it.
With temporary logging on the server thread: the glowstone lands in a
finished chunk that isn't lit (the ring past what the guest tracks), so
the edit queues a full light job for its neighbours. The border chunk's
job then relights sections 15 to 17, and Hub::broadcast_light_changes
sends them to both clients as a tracked border chunk, as
ChunkMap.isChunkOnTrackedBorder does. Nothing in the engine changes.
Fixes#891
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: Merged. Bit-exact against the per-corner oracle (unit test over a dense grid, 7 seeds, wrapping extremes), CI parity passed, and the surface pass is about 10% faster in chunkgen_bench. Thanks rdbt!
On a Linux without ALSA's library (most minimal server images), the
game exits at once with the loader's "error while loading shared
libraries: libasound.so.2". The supervisor took that for a crash and
restarted it five times over two and a half minutes before giving up
with "see logs/latest.log", which holds nothing about it.
Now:
- `run` starts the game with --help first (it loads no game files) and,
if the system can't load it, stops at once with the loader's line and
the packages to install on Debian/Ubuntu, Fedora, Arch and NixOS;
- `init` prints the same as a warning after "Game:";
- install.sh lists the libraries `ldd` can't find, with the same advice;
- docs/server.md says so in the install section.
Also, a service installed without the installer's launcher on the PATH
ran `<dir>/releases/<tag>/horizon-server`: Linux reports the running
program with links resolved, so the unit pinned that release and an
update never reached it. The fallback now goes through `<dir>/current`
while it points at that release.
Found by the end-to-end install check on the ThinkPad (NixOS, no
nix-ld libraries) with v2026.10.18.15.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: Merged. Thanks rdbt for the consolidation and for merging main yourself! On de6f15e6: fmt, workspace clippy, horizon-client tests and the client build pass, and 19/19 covering harnesses pass (options, world_reopen, lan, dedicated_server, block_removal, section_mesh, resource_pack, menu_click, screen_layout, the pack, world-options, chat and container screens, and the consumable ones). We'll handle the two follow-ups noted in the comment above (the 120 s fatal timeout, and the backdrop terrain built on the event loop).
init and run looked for asset-index.json to recognise a prepared folder,
but horizon-prepare writes that file only with the optional sounds
(`horizon-prepare sounds`). A server prepared the usual way, without
sounds, was never found: init said "no prepared game files" and listed
the folder that held them. They are now recognised by
horizon-resources.json, the manifest every preparation writes and the
one the game itself reads.
Found by installing v2026.10.18.15 with install.sh on the Mac mini and
preparing from the release folder.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: Merged. Matches 26.3's KeyboardHandler.keyPress: the F3 toggle isn't conditional on a world or the absence of a screen, and the title counts as an open screen so it gets no gameplay keys. Fixes#811. Thanks rdbt!
A `fly` benchmark replied with frame times only. With single player on its
server thread, the world's ticks no longer show in the frames, so a flight
could not tell how long the server took.
The summary now has tick_ms and tick_ms_max: the world's smoothed tick time
(the F3 server line) averaged and at its highest over the flight. On the
server thread it is the server's.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: Merged. Fixes#861: a ctl tp/fly move is sent even while the destination chunk isn't loaded, the single-player owner is exempt from the move check as in vanilla, and wait-ready waits for the level around the player. lan, dedicated_server, server_entities and server_dimension_players pass.
🤖 Agent: Merged. CI parity and verify passed, and on current main with this branch, horizon-client clippy is clean and all its tests pass (873 lib + integration). Consumables now finish as in vanilla 26.3.
With the world on its server thread, a horizon-ctl `tp` or `fly` moved the
window's player to chunks the server had not sent. The client skips its
tick while its own chunk isn't loaded, so it never sent the move, and the
server kept streaming the old area: "loading terrain..." forever.
- The client sends a move made where nothing is loaded (by horizon-ctl;
a player who walks there has already sent it), so the server streams
the chunks there.
- The integrated server doesn't check its owner's moves, as vanilla's
`shouldCheckPlayerMovement` (`isSingleplayerOwner`); a long `tp` was
"moved too quickly".
- `wait-ready` on a server also waits for the level around the player
(`PlayerLoaded`); the status shows it as `server.loaded`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A client joined to a server (and the window on the server thread) keeps the
server's chat lines in its own list. The event log only counted the local
chat's lines, so `wait-for chat` never saw a line from the server.
The remote client counts the lines it receives. The event log reads that
count and those lines while connected.
lan_e2e gets a `chat` check: a line the host sends reaches the guest's
event log, and on the server thread the host window's too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With single player on its server thread, the window is a client of the
server. The agent's `command` went out as the owner's chat and returned no
output, so harnesses that read a command's feedback (lan_e2e trials) failed.
Control `give` and `time` changed only the window's copy of the world.
The server now runs them as single player's console does without a server
thread (`as_console`): as the world's owner, in the owner's level. The
feedback goes to the owner's chat and comes back to the agent. `give` runs
`give @s`, and `time` sets the server's clock too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: Merged after #855 and the real verify. block_removal 4000/4000 including 319 mining cases; dispenser, crafter, composter, mace and use_item_on pass.
🤖 Agent: Merged after the real verify passed. Prism's metadata renames native jars (same sha1), so prepare now also looks under Prism's name, still checked against Mojang's sha1, with nothing downloaded. Launcher and horizon-prepare tests pass.
With HORIZON_SPIKES=MS, a dedicated or integrated server tick longer than
MS ms is logged with the server thread's CPU time over the tick and the
tick's spans (tick_mobs, streaming, ...). A tick with much more wall time
than CPU time waited for a lock or for a core.
The spans recorded on a server thread were never taken before, so they
grew for as long as the server ran with HORIZON_SPIKES set.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every spawn attempt (`NaturalSpawner.getRandomPosWithin`) asks for the
WORLD_SURFACE height of a column. Horizon scanned the column down for it
each time, a chunk-map lookup and a block-name compare per block. Vanilla
reads the chunk's heightmap.
Chunks now keep WORLD_SURFACE: built on the first read, and dropped by a
`set` that can move a column's surface (`Heightmap.update` leaves writes
two or more below it alone). Natural spawning, the spawn rules' bat check
and beacons read it.
The unused `heights` field (never written after `Chunk::empty`) is gone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ServerLevel.removePlayerImmediately and addRespawnedPlayer: when a player
respawns, every other player tracking it gets RemoveEntities and the next
tick pairs it again as a new entity. The hub only cleared the respawned
player's own tracker, so the others kept the dead body, lying down, wherever
the player went (the host window too, on the server thread).
The in-frame LAN host's death now reaches its guests (LivingEntity.die's
EntityEvent DEATH), and its respawn goes through the same path.
lan_e2e gets a `respawn` check: the host dies and respawns (and on the
server thread the guest too), and the other side's client draws it dead,
then alive. `status` lists the other players a client draws dead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
test_block_removal_parity gains a "mine" removal: VanillaDispenserOracle
calls the real ServerPlayerGameMode.destroyBlock for a survival player holding
a pickaxe, axe, shovel, sword, shears, a mace or nothing, and Horizon's
dispense fixture calls break_block, the path every player's mining takes (the
single player game, the server thread and LAN guests). It compares the drops,
the spilled contents, experience, sounds, events, statistics and the held
stack. On main it shows every mined container losing its contents (#844, fixed
by #855) and three gaps of Horizon's own, fixed here:
- ItemStack.mineBlock: an item with a tool component counts as used, and wears
its damage_per_block (2 for swords, the mace and the trident), also the mace
and trident, whose tool component has no rules. Shears wear on every block
but fire (ShearsItem.mineBlock).
- DecoratedPotBlock.playerWillDestroy: a pot mined with an item in
breaks_decorated_pots, unless an enchantment prevents it, cracks first and
shatters.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The generator's structure-start cache was one Mutex<HashMap>. Each
generated chunk reads the 289 start chunks around it for its references,
and the server thread reads the starts of the chunks its mobs and spawn
attempts are in (`structures_reaching`). While workers generated, the
server thread spent most of its start lookups waiting for that lock.
The cache is now 64 RwLock shards keyed by a hash of the chunk: a lookup
read-locks one shard. Vanilla keeps the starts on each chunk
(`StructureManager.getStartForStructure` -> `chunk.getStartForStructure`),
so its lookups share no lock either. The first starts built for a chunk
are still the ones everyone keeps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>