- Integrated GPUs other than Apple's draw far terrain to 128 chunks at
most, whatever LOD Distance says; its description says so, and the
status reports the distance in use (`lod.effective_distance`).
- On machines with four threads or fewer, far terrain starts no detail and
one sample job per 500 ms while the near terrain loads, so the near
chunks and the frame get the cores first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Upload far meshes within about 2 ms a frame (at least one); the rest
wait, with their removals and the cutout, and apply in order.
- Gather mesh inputs across frames within about 2 ms, from a snapshot of
the active tiles' samples, instead of hundreds at once.
- Select tiles only when the plan, samples, covered columns or dirty tiles
changed; add the covered stand-ins only after a replan.
- Find the tiles touching an invalidated chunk per level, not by scanning
every active tile; share tile samples with mesh jobs (Arc) instead of
cloning them.
- Recompute the full-detail columns only when their completeness, the
center or the radius changed.
- Observe chunks for the far summaries within about 1 ms a frame (at least
one), unloaded ones first, with each block state's kind read once instead
of matching block names per block.
LODs off is unchanged: none of this runs without far terrain.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Far tiles mesh every gap in their columns: caves and aquifers in visited and
refined chunks, and the gaps kelp and seagrass leave in water. Most are closed
on every side, so their floors, ceilings and walls cost vertices and never
show.
Before meshing, a tile fills each gap below a column's top that no
neighbouring column's open air looks into (every neighbour's top is at or
above the gap's top) with the material below it, or with the fluid above it
(kelp in water). Cave mouths stay open; unknown neighbours and cells meeting
full-detail or finer terrain count as open.
Seed 1234, RD12, LOD 64, overview: 4.93 M to 3.17 M far-terrain quads.
Follow-up to Online5277's terrain LODs (#771).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Full-detail sections leaving the render distance stopped being drawn at once,
while the distant tiles still left their columns out until the next mesh job
finished: for a few seconds after moving (longer on slow machines) a hole
opened there, lined with the tiles' seal walls. Those took the light recorded
for buried blocks, none, so the hole showed as a dark band.
Lod::take_cutout hands the renderer the columns the finished meshes leave
out. Their resident sections stay drawn, even past the render distance, and
sections unloaded while still left out are released only once the tiles
cover them. With LODs off the cutout is empty and nothing changes.
Side faces above the top of the column they face open onto air. Reduced
columns and the seals expose sides that were buried, so those parts take at
least the sky light over that top (black cliffs in far tiles). Faces into a
cave keep their recorded light.
Follow-up to Online5277's terrain LODs (#771).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify and CI parity passed. horizon-server update (SHA-256 check, world backup, switch, health wait, automatic rollback) and rollback; 6-hourly update notice in log, status and panel. Proven end to end on the Mac mini and the ThinkPad (healthy update, crash rollback, timeout rollback, manual rollback, real GitHub download); 37 unit tests.
🤖 Agent: merged after verify and CI parity passed, and a combined check with rdbt's #796 on current main (fmt, clippy, 359 tests, 14 world-gen harnesses). YUNG/Moog terrain adaptation around structures (Better Fortresses netherrack inside bridges 39–41 % → 2.5–7 %), its carry, and YUNG modifiers.
🤖 Agent: merged after verify and CI parity passed, and a combined check with rdbt's #796 on current main (fmt, clippy, 359 tests, 14 world-gen harnesses). YUNG/Moog terrain adaptation around structures (Better Fortresses netherrack inside bridges 39–41 % → 2.5–7 %), its carry, and YUNG modifiers.
🤖 Agent: merged after verify and CI parity passed, and a combined check with rdbt's #796 on current main (fmt, clippy, 359 tests, 14 world-gen harnesses). YUNG/Moog terrain adaptation around structures (Better Fortresses netherrack inside bridges 39–41 % → 2.5–7 %), its carry, and YUNG modifiers.
🤖 Agent: merged, thanks rdbt! Bulk density arithmetic, the shared Beardifier kernel and Perlin column reuse: 12.685 → 10.069 ms/chunk Overworld terrain (−20.6 %, one worker). We merged current main in (density/tests.rs: kept both test sets) and checked it together with #955/#956/#959: fmt, clippy -D warnings, 359 core/world/port-pack tests, and 14 world-gen harnesses (density, noise, structure_density, server_world_gen, world_spawn, decorated_region, …) all pass.
🤖 Agent: merged after verify passed. The renderer logs each start-up step once per device (driver, requested features and limits, surface, pipelines, first acquire/submit/present), so a native crash's last log line names the step.
🤖 Agent: merged, thanks Online5277! Terrain LOD lands off by default. With LODs off it matches main (0 differing pixels in 6 RD16 views, same FPS/GPU ms on the HD 620, same time to full render distance); CI verify and parity pass. We merged main in three times and adapted it to #905/#939 and the packed texture coordinates. The follow-ups from the review (edge band, integrated-GPU cost, lighter far meshes) we're taking on ourselves.
A native crash in the graphics driver leaves no panic or crash report: the
last log line is all a player can send. After the adapter line, log the
driver, the features and limits the device is requested with, the surface
configuration, the pipelines (with multi-draw, occlusion culling and the GPU
timer on or off) and the first frame's acquire, submit and present, once per
device.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify passed and local proof (sculk harness). Sculk drops every unloaded chunk's blocks in one pass: worst tick 583–609 → 38–44 ms with 40 bots.
🤖 Agent: merged after verify passed and local proof (heightmap, chunk_stage, lan, …). Chunk-packet heightmaps are built on the warm-up threads; bytes identical; lan_chunks −11–13 %.
🤖 Agent: merged after verify passed and local proof on the integration commit (8 harnesses, lan_e2e both modes). The server thread runs ahead of the workers on macOS and Windows, like vanilla's Thread.setPriority(8): 40 bots tick p99 101.5/77.7 → 32.8/32.7 ms, ticks over 50 ms 166/146 → 1/2.
🤖 Agent: merged, thanks LookImEcho-Rick! It was still marked draft, but its description was complete and our checks were done, so we marked it ready. /execute in <dimension> now moves the player across dimensions like vanilla (fixes#592). With main merged in and the #947 fix (#962), command_exec, command_parse, server_dimension_players and dedicated_server pass locally, and verify passes. If you meant to add more, a follow-up PR is welcome.
🤖 Agent: merged after verify passed and parity --changed 311/311 locally. Lithostitched apply_random, condition and set_block in the game: CTOV farms grow carrots, potatoes and beetroot, paths and mossy cobblestone appear.
🤖 Agent: merged after verify passed and parity --changed 311/311 locally on the integration commit. YUNG modifiers, piece_in_range and dead-end pools (game side).
🤖 Agent: merged after verify and CI parity passed. The carry reads data files with comments as Minecraft does; six Better Fortresses pools come back (265 → 3,765 pieces over seed 1234's Nether centre).
🤖 Agent: merged after verify passed and local proof on one integration commit of modpacks-9's PRs: parity --changed 311/311, structure-start oracle 3,153/3,153. CI parity's only failure was secure_chat, which this change doesn't touch. Stronghold rings 3/128 off → 0/128 on seeds 1234, 2, 3, 42, -7.
🤖 Agent: merged after verify and CI parity passed (command_exec 764/764 three times locally, 20 affected harnesses). /summon no longer spawns other pending entities and takes their id; the harness settle waits for entities like vanilla's. Fixes#947.
MinecraftServer.spin raises its server thread (Thread.setPriority(8)) when
there are more than four processors. The new horizon-threads crate makes the
same request for the dedicated server's thread and the integrated server's:
user-interactive QoS on macOS, above-normal priority on Windows, nothing on
Linux (an unprivileged thread can't raise itself).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`horizon-server run --dir srv` (what init prints) started the game in srv/ with
--control-socket srv/.horizon/engine.json, which the game resolved inside srv/,
so it crashed at every start. spawn now passes absolute paths.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Vega 11 on Windows (Vulkan) panicked at launch in bytemuck's cast_slice
with TargetAlignmentGreaterAndInputNotAligned, starting with the release
that shipped occlusion culling. The section-test results and the GPU
timer's timestamps were cast in place from the mapped range, which isn't
guaranteed to be 4- or 8-byte aligned. Copy the words out instead.
Fixes#953
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
apply_random becomes horizon:apply_random: its weighted lists, or its set
of lists at weight 1, with the inline lists carried like any other (their
Java processors dropped and counted). condition becomes horizon:condition
when the game runs every condition in it (true, all_of, any_of, not,
random_chance, matching_blocks with plain property values), with its then
and else lists carried. set_block with a simple or weighted state provider
and no block entity change becomes horizon:set_block. Anything else stays
Java and is dropped as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
horizon:apply_random runs one processor list drawn by weight
(WeightedList.getRandom); horizon:condition runs `then` or `else` by a
test of the block (true, all_of, any_of, not, random_chance and
matching_blocks on the placed or the existing block). Both take their
random from Lithostitched's RandomSettings: seeded with the world seed
plus the name's hash, forked positionally at the block, the piece or the
structure's pivot. process_block now gets the piece's position for it.
horizon:set_block is Lithostitched's set_block with a simple or weighted
state provider: the block becomes the state (drawn at the chosen position
from the unsalted seed), keeping the properties it shares with the old one
unless preserve_state is off.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged, thanks rdbt! Terrain sampler scratch buffers are reused between chunks; byte-identical output across the 22 world-gen parity harnesses on the combined commit.
🤖 Agent: merged, thanks rdbt! Spectator scroll-wheel flight speed now follows vanilla (6,791 oracle cases); verified on the combined commit with local_input and the client tests. Fixes#945.
lithostitched:reference runs the processors of other lists in turn.
Carry now writes their processors in its place, in order and references
in them too, from the mod's own processor lists. A reference to a tag or
to a list the mod lacks stays, and is dropped and counted as before.
ChoiceTheorem's Overhauled Village builds most of its lists this way.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Lod::update sorted every unsampled plan node each frame to take at most
four (none while every sample slot is busy), and woke the workers every
frame. It now skips the selection while the slots are full, keeps only
the best 2 * (SAMPLE_JOBS + pending) entries before sorting (a key is
listed at most twice, so the same tiles are taken in the same order),
takes the best fallback with min(), and wakes the workers only when a job
was queued. A finished chunk's detail now goes straight to the tile at
each level 0-4 holding it instead of scanning every ready tile (the others
never overlapped it).
Mac mini, RD12, LOD 256 while far terrain generates, `sample` of the main
thread over 30 s, alternating: Lod::update 10.4 % and 12.3 % of the
samples before, 3.4 % and 3.5 % after; update_world 3497 and 3373
samples before, 668 and 864 after.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
YUNG's API and MoogsStructureLib add their own kernel to the Beardifier:
a size and falloff of its own, carving or burying above and below a piece's
floor, around rigid pieces and junctions. Better Strongholds buries its
stronghold, Better Fortresses carves the netherrack over its bridges and
Moog's mega fortress carves a band around its corridors. A structure or a
pool element holding `horizon:enhanced_terrain_adaptation` now gets that
kernel (horizon_core's TerrainKernel), summed after vanilla's density as
the libraries' mixins do; YUNG's also grows the start's box by its radius.
Vanilla data never has the key, so vanilla worlds are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Better Nether Fortresses relies on three parts of YUNG's API the game did
not have:
- piece_in_range: another piece (of some templates, or any) inside the
piece's box grown sideways, below and above.
- Modifiers, applied once the layout is done, in piece order and with the
assembler's random: a condition, then actions on the piece itself.
transform swaps its template for one picked by setFeatureSeed(x, y, x)
and moves it by an offset turned with the piece; delay_generation moves
it to the end of the piece list.
- Dead-end pools: a piece that got no child though it has other jigsaws
is taken out (its parent's junction, its room and its count given back)
and a piece of its dead-end pool is chosen for the same parent jigsaw.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Minecraft reads data JSON with Gson's lenient reader, which skips `//`,
`#` and `/* */` comments. Six of YUNG's template pools have them (Better
Fortresses' bridges, halls and hall ends; Better Jungle Temples' floors),
so carry refused them as unreadable and the structures lost those pieces:
seed 1234's seven Better Fortresses had 265 pieces in all, now 3,765, and
five Better Jungle Temples 170, now 196. A file that parses once its
comments are gone is carried as plain JSON.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ChunkGeneratorStructureState's ring searches call findBiomeHorizontal,
which samples its 57x57 quarts in order through one resolver; vanilla's
climate index starts each lookup from the previous result and keeps it on
a tie. Horizon built a fresh resolver for every quart, so ties went the
other way: three of seed 1234's 128 strongholds were up to 6 chunks off
(the first one 96 blocks from vanilla's). StartContext::noise_biomes
samples a batch through one resolver and the search hands it the square.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
YUNG's single elements keep their dead-end pool and their modifiers under
horizon:yungsapi, which the game's assembler now reads: the modifier's
condition, and its transform and delay_generation actions on the piece
itself. piece_in_range carries as is. A modifier aimed at another target
than self, or another action, is counted lost.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
YUNG's API's `enhanced_terrain_adaptation` (its presets spelled out, NONE
as size 0) and MoogsStructureLib's (with its band) were counted lost. Carry
now writes them as `horizon:enhanced_terrain_adaptation` on the structure
and on the pool elements that have their own: YUNG's with every piece's
junctions and the start's box grown by the radius, Moog's with only the
terrain-matching pieces' junctions, as each library does.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify and CI parity passed. tick_blocks finds the motion-blocking height from the top of the terrain: 1.98 → 1.36 ms per tick with 10 bots on the Mac mini (−31%).