Upstream added --player-dir handling to the startup code in main.rs. With the
library split that code lives in entry.rs, so the change is applied there and
main.rs stays the thin launcher.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
main.rs moved to entry.rs on this branch; carry main's --player-dir
log/crash folder handling over to entry.rs.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Combined staging branch: the game library split (#5446) and the single mod
file format (#5909) together, as agreed to land in one release. No conflicts.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: main merged after #4945/#4946 landed: main's new wanted_tiles queue now skips tiles waiting out their retry, as this PR's loop did. verify green on a6d37d7; tools/server_lod_e2e.py passed on Linux at the previous head (source server, capped at 16; none when unset).
🤖 Agent: Review by Ludovic's AI agent (Claude). A pack's vignette=false hides the game's vignette; with no pack nothing changes. I merged main and reworded a comment that named another project. Verify green.
🤖 Agent: Review by Ludovic's AI agent (Claude). GLSL translation fixes, each with a test (GLSL 4.60 §4.1.9, §4.2.2), packs-only code. I merged main in; verify green on this head.
🤖 Agent: main merged again (the two uniforms go after main's isRightHanded … hasCeiling, looked up by name); verify green on 67ebb02. Thanks @Joansitoh.
🤖 Agent: levzzz5154's five review points fixed on the branch (crash telemetry stays on --home's consent, servers ignore the flag, limits documented, splitscreen guests on the game folder's mods intended, tests for shared_dir_of and servers_file_of); main merged; verify green on 9d56f33. Without --player-dir nothing changes. Thanks @Solenopsisbot.
🤖 Agent: Review by Ludovic's AI agent (Claude). levzzz5154's two points are fixed (the hook runs through as_guest, after placement), the docs conflict with main is resolved, and the API fragment carries neoforge/side. Verify green.
🤖 Agent: Opening shulker lids push the local player on the server-thread client as ShulkerBoxBlockEntity.tick does on both sides; lan_e2e shulker_lift check; verify green.
Upstream added `pub mod lod_net;` to the old lib.rs module list, which this PR
moved to client_lib.rs: add it there. Upstream's new horizon-optifine
dependency is used by mobs/cem.rs, a reusable-client module, so it goes in
horizon-client-lib's manifest instead of horizon-client's.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: #5601 has merged, so this test-only follow-up leaves draft: main merged (corpus union, every function checked against its source), verify green on 938f032. Thanks @JustKarkat.
🤖 Agent: shader API series stacked on one union (each head holds the earlier ones); pr_ready --quick green on the last head 47b5a77 (2574 tests), shader-api tests + clippy on each earlier head; offscreen in-game run with a pack using all four features, no validation error.
🤖 Agent: Opaque terrain pass without discard: two solid passes on Metal (Apple's tiled GPUs, measured faster on M1 Pro), one pass elsewhere as before (fix pushed here); the far-tile opaque marking is fixed everywhere.
🤖 Agent: A boat used under water lands at the eye and is refused as in vanilla's BoatItem.use (ClipContext.Fluid.ANY on a full fluid cell); unit test, verify green.
🤖 Agent: far terrain 2/3 (the server samples tiles, off by default). #4945 and main merged, verify green on 6ee16d8; rdbtCVS approved, kacpherek's points fixed in the code; the stack's e2e passes on Linux.
🤖 Agent: far terrain 1/3 (wire format, off by default). Main merged, verify green on 1f6b07c; rdbtCVS approved; the stack's e2e passes (tools/server_lod_e2e.py on Linux: server-sampled LOD capped at the server's 16 chunks, none when unset).
🤖 Agent: requeued: the parity failure was the runners' (a dead .horizon-prepare.lock link on runner horizon at playhorizon and horizon-2 at the Mac mini, now removed); the run is rerunning on the same head.
🤖 Agent: security-reviewed by Ludovic's AI agent (Claude): the secrets file is now encrypted at rest (ChaCha20-Poly1305, key from the computer's machine-id), mode 0600, atomic replace, no token in logs or errors, the owner gate unchanged. 58 account tests pass on Linux (playhorizon); verify and launcher green on 0de577b. Thanks @JulesB40.
🤖 Agent: native_camera_distance hook: a no-op fragment hook until a mod fills it, finite 0..=1024 overrides only, collision clipping kept; unit tests cover it.
🤖 Agent: Review by Ludovic's AI agent (Claude). The hook skips the screen the server pack prompt hides, as hud_layers draws Playing then; arismoko's in-game check confirms. I merged main; verify green.
🤖 Agent: Review by Ludovic's AI agent (Claude). Docs only: bow PvP between players works since 5763854b0 (lan_e2e arrows, #5785's dedicated-server check). I merged main; verify green.
🤖 Agent: Review by Ludovic's AI agent (Claude). EnderDragonRenderer has no dimension check; dragon_geometry returns at once with no dragon, and the dragon is cleared on a dimension change. I merged main; verify green.
🤖 Agent: Review by Ludovic's AI agent (Claude). Far tiles sample an explicit mip, only on tile faces (UVs past a block); block faces keep sprite_grad. A few ALU ops per far-tile fragment. I merged main; verify green.
🤖 Agent: Translucent batching 1/3: off by default (only horizon-graphics.txt turns it on), so players see no change; with it off the frame adds one bool check (follow_translucent_layout) and a branch per draw run, uploads and compaction keep their old paths. Measured gains in the body are for the opt-in path.
🤖 Agent: Review by Ludovic's AI agent (Claude). levzzz5154's four points are addressed (overlay gate, client-only guard, gamepad note, #5617 merged in); API fragment. I merged main (merges with #5731 cleanly); verify green.
The split saves Apple's tiled GPUs their hidden surface removal; elsewhere
the second pass doubled the solid draw calls (one per section without
multi-draw) and the vertex work for no gain that was measured. Other
backends keep the single fs_solid pipeline, the flat one included.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: Management protocol serversettings/difficulty get/set as MinecraftServerSettingsServiceImpl, saved to server.properties and server.toml; the management harness covers it.
🤖 Agent: Boat paddle flags reset to false (AbstractBoat.DATA_ID_PADDLE_*) when a stroke stops or the rider dismounts; new harness boat_paddle_reset compares against a vanilla server.
🤖 Agent: Thrown potions hit players with vanilla's pick margin and leftOwner rule (Projectile.canHitEntity / checkLeftOwner); costs nothing when no potion is in flight; unit test covers the thrower case.
🤖 Agent: Placing a painting swings the hand as vanilla's HangingEntityItem.useOn (InteractionResult.SUCCESS) does; tests pass, parity runs on the head.
🤖 Agent: OptiFine series step verified (licence gate, gating to packs that ship the files, per-mob cost measured, verify green on 64d0b1a4e). Thanks @some-du6e.
🤖 Agent: OptiFine series step verified (licence gate, gating to packs that ship the files, per-mob cost measured, verify green on b88a1aad1). Thanks @some-du6e.
🤖 Agent: OptiFine series step verified (licence gate, gating to packs that ship the files, per-mob cost measured, verify green on d454a1fcb). Thanks @some-du6e.
🤖 Agent: Shader packs: a custom variable/uniform expression may end with one ';' (trimmed), extra statements and empty expressions still rejected; unit tests.
🤖 Agent: Shader packs: custom image dimensions expand pack option macros (default and chosen), the option table is still built once; literals unchanged, unknown names still rejected; unit tests.
🤖 Agent: Shader options: a declared default missing from its authored range stays selectable (inserted first), empty ranges still create no option; unit tests.
🤖 Agent: Shader packs: a one-literal vector constant (ivec3(1)) in workGroups/workGroupsRender fills every component (GLSL 4.60 §5.4.2); negative or nested expressions still rejected; regressions fail before the fix.
🤖 Agent: Shader packs: isRightHanded, is_on_ground and the dimension's bedrockLevel/heightLimit/logicalHeightLimit/hasSkylight/hasCeiling get real values (black-box probe of the uniform values, no source read); set per frame only while a pack is on.
🤖 Agent: Client hook native_mob_submit as an API fragment; I fixed arismoko's two review points (walk state from mobs::anim::state: no walk while riding/dead, babies x3; hook asked before the sulfur cube's block). Verify green on 62e9cf2ba; the hook is a no-op stub without mods.
🤖 Agent: Reviewed lod.rs (seam stand-in rule in select and evaluate, coarse backdrop in the wanted queue, load order only). fmt/clippy/lod tests green in verify on this head. Thanks @some-du6e.
🤖 Agent: Requeued after merging main (conflict beside native_data_pack_files, both hooks kept, 3 files +24 vs main). Server hook native_server_data_loaded(reload) as an API fragment; empty stub, no cost.
A fluid hit always snapped to the top of the water cell, so a boat landed
above a submerged player. In a full water cell the hit now stays on the
ray, as vanilla's VoxelShape.clip has it, so the boat collides with the
player and nothing is placed.
Fixes#5967
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018zuDTsnyHkP28bzn3u7S3m
🤖 Agent: Client hook native_biome_colors + native_refresh_biome_colors as an API fragment: runs only when a mod asks for a refresh (rebuilds sections like a Biome Blend change), nothing per frame; Biomes::with_colors unit-tested.
🤖 Agent: Harness diagnostics only: a JVM oracle that dies without an exception reports its exit/signal and last output (chunk_generation, datafix); no change to what is compared.
🤖 Agent: World border wall invalidated when the camera changes SectionPos (LevelRenderer.repositionCamera); an i32 compare per frame; world_border_render passes; verify green.
🤖 Agent: One line: rope_rot_old (yRotO/xRotO) set at the start of the server-client tick, as the local tick does, so the held lead end stops easing toward yaw 0; verify green.
🤖 Agent: A full dynamic glyph region no longer caches the missing glyph forever: overflow is retried after begin_frame clears the region (one bool check per frame); font unit tests cover overflow and no-provider chars; verify green.
🤖 Agent: Pointer menu mode: the on-screen keyboard auto-opens over chat/signs and A on a text field opens it, as focus mode does; pure helpers unit-tested; pad UI only, verify green.
🤖 Agent: Toast label colour packed with hud::rgba (yellow 0xFFFF00 / pink 0xFF88FF as AdvancementToast), unit test on the packed bytes; client UI only, verify green.
🤖 Agent: queued with my fix (9ec5db50d, see above): the remote arrow's flight rotation as vanilla's `AbstractArrow.tick` has it, without the two hunks that went against 26.3. Thanks @Lostpold88.
🤖 Agent: verify was green before main's splitscreen extra conflicted. Main is merged in (40704247c), and the extra stays off by default. Queued for merge, thanks @Stink-O.
Resolves docs/native-mods.md with main's native_player_command and
native_perform_console_command entries. levzzz5154's review:
- native_player_respawned runs through as_guest, so the helpers that act on
the player (native_player, native_give_items...) reach the respawned one;
- it fires at the end of guest_respawn, after the bed, anchor and world-spawn
placement and the hardcore switch, as PlayerRespawnEvent ends
PlayerList.respawn; the docs name the later-tick trips and the End exit;
- the fragment carries neoforge PlayerEvent.PlayerRespawnEvent, side server.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Keep the client collision snapshot in sync with the animated lid and apply its existing swept-volume push, including while a container menu is open. Cover lift, support and closing for both players in survival and creative on the integrated server.
Co-Authored-By: Codex GPT-6 <[email protected]>
levzzz5154's review: only logs/ and crash-reports/ follow --player-dir (crash
telemetry stays on --home, whose consent it reads); a dedicated server never
takes it. The launch reference says what stays shared (command history, Play
with Friends' relay) and that horizon-ctl logs and the launcher read --home.
Tests: shared_dir_of, servers_file_of.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Vanilla's DataPackCommand ends in ReloadCommand.reloadPacks, which
reloads the server's resources: recipes, tags, loot and advancements
with the functions. Here enable and disable rebuilt only the function
store, so a pack's recipes stayed (or stayed away) until the next
/reload. They now reload the content tables too.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Reject unsupported negative and nested scalar expressions rather than dropping their operators when expanding a vector. Add a regression for the overbroad expansion.
Co-Authored-By: Codex GPT-6 <[email protected]>
Far tiles marked only leaves as opaque, so every stone, sand or ice face
of the far terrain went through the cutout pass and its discard. They are
marked as the mesher marks section quads now: opaque leaves and faces
over opaque texels.
🤖 Agent: verify is green and the branch merges cleanly with main. It also built and passed the Python tests in a combined branch with ~40 other open PRs. Read through: it looks right. Queued for merge (parity runs before it lands). Thanks!
🤖 Agent: verify is green and the branch merges cleanly with main. It also built and passed the Python tests in a combined branch with ~40 other open PRs. Read through: it looks right. Queued for merge (parity runs before it lands). Thanks!
🤖 Agent: verify is green and the branch merges cleanly with main. It also built and passed the Python tests in a combined branch with ~40 other open PRs. Read through: it looks right. Queued for merge (parity runs before it lands). Thanks!
🤖 Agent: verify is green and the branch merges cleanly with main. It also built and passed the Python tests in a combined branch with ~40 other open PRs. Read through: it looks right. Queued for merge (parity runs before it lands). Thanks!
🤖 Agent: verify is green and the branch merges cleanly with main. It also built and passed the Python tests in a combined branch with ~40 other open PRs. Read through: it looks right. Queued for merge (parity runs before it lands). Thanks!
🤖 Agent: verify is green and the branch merges cleanly with main. It also built and passed the Python tests in a combined branch with ~40 other open PRs. Read through: it looks right. Queued for merge (parity runs before it lands). Thanks!
🤖 Agent: verify is green and the branch merges cleanly with main. It also built and passed the Python tests in a combined branch with ~40 other open PRs. Read through: it looks right. Queued for merge (parity runs before it lands). Thanks!
A splash, lingering or XP bottle stopped at blocks and mobs only, so it
fell through any player. The step now meets players too. The thrower and
their vehicle group are skipped until the potion has left them, as
Projectile.canHitEntity and checkLeftOwner do.
Fixes#5876
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018zuDTsnyHkP28bzn3u7S3m
Send the default false values when paddle entries disappear from tracked metadata, including dismounting mid-stroke. Compare 18 transitions with actual vanilla servers.
Co-Authored-By: Codex GPT-6.1 <[email protected]>
- NativeMobRender takes the walk cycle, head turn and age from mobs::anim::state,
the state the vanilla models read: no walking while riding or dead, a baby's
legs three times as far (LivingEntityRenderer.extractRenderState,
WalkAnimationState.positionScale).
- The hook is asked before a sulfur cube's swallowed block is drawn, so a
mob a mod draws carries none of the vanilla model.
Review points from arismoko's review.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A pack of Horizon's own with [shadow] draws from the shadow camera's lists
(d36550e82), which makes frames flat where the device can, and flat draws
read translucent quads in the section arena. translucent_split now counts
that pass with a shaders/ pack's shadow programs, and a renderer made with
such a pack picked starts without the split (picked_pack_draws_flat, which
reads the pack's manifest; it was shaders_pack_picked).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: reviewed by Ludovic's AI agent (Claude): rdbtCVS's #5595 finished (main merged, "Built on {base}" in all 10 launcher languages); remove_version refuses a running build; verify and launcher checks green on e918209, merges cleanly with main. Thanks @rdbtCVS.
The file kept the Microsoft refresh token and the Minecraft access token in
plain JSON. It is now sealed with a key derived from /etc/machine-id
(/etc/hostid on FreeBSD), so no token is on disk in plain text and a copy
taken off the computer can't be read. A file that doesn't open is refused
without quoting it and never written over, as a broken one was.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: Windows F3/crash-report CPU name from the registry's ProcessorNameString with the old fallback; read once (OnceLock) so no per-frame cost; the launcher starts the game with CREATE_NO_WINDOW so reg.exe opens no console. Parser unit-tested.
🤖 Agent: MCP initialize now negotiates the protocol version (echo only a supported one), same 12-line helper in horizon-ctl, horizon-server and the launcher, with unit tests; verify and launcher CI green.
🤖 Agent: Docs only (mods.md, modpacks.md, template comment, manifest doc comment); checked the facts against the content loader (sounds key names a vanilla block, menu/entity_type kinds, mismatch refusal). The failing 'check' job is horizon-mod-registry's lock on main, unrelated.
Packs read these Iris uniforms as zero: isRightHanded (the Main Hand option), is_on_ground, and
the dimension type's bedrockLevel, heightLimit, logicalHeightLimit, hasSkylight and hasCeiling
(from the level's shape; hasCeiling in the Nether). A pack reading bedrockLevel 0 and heightLimit
0 in the overworld puts its fog or its sky's horizon at the wrong height.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
GLSL 1.20's gl_FogFragCoord is a built-in varying: a vertex stage writes the fog distance and the
fragment stage reads it. naga knows no such name, so every program using it was skipped with
"Unknown variable: gl_FogFragCoord" (26 in MakeUp UltraFast, 26 in E-LITE, 18 in VanillAA, 1 in
Potato). It is now hz_FogFragCoord, a float output of the vertex stage and input of the
fragment stage, like gl_FrontColor and gl_TexCoord[n].
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: verify is green and the branch merges cleanly with main. The saved switch stays off by default, so nothing changes for a player who doesn't enable it. Queued for merge, thanks @martipops.
🤖 Agent: verify is green, the branch merges cleanly with main, and levzzz5154's review found it ready (load order below the world's packs, both first open and /reload covered). The hook's API number comes from its fragment when it lands. Queued for merge, thanks @Sylviromi.
Read and change the dedicated server difficulty through vanilla management
requests, broadcasting the change and saving it for restart.
Part of #5147
Co-Authored-By: Codex (GPT-6) <[email protected]>
Translucent quads (water, stained glass, ice) sat right after their
section's solid quads in the 256 MiB arena pages, so the back-to-front list
changed page every few sections and only short runs merged: 3,832 draw calls
a frame over the ocean at render distance 64. With the new
translucentBatching setting on, they get a vertex arena of their own whose
pages may grow to about 2 GiB (LARGE_PAGE_BYTES: DirectX 12 and OpenGL cap a
buffer at i32::MAX), and the sorted indices the same cap, so the list goes
out as a few multi-draws. Vertex data, sorted indices and draw order are
unchanged.
The split holds only while no picked shaders/ pack can draw terrain or its
shadow pass flat (translucent_split); a pack that changes that drops the
sections and far tiles, which the game meshes again. compact_arena moves
what a section holds in the section arena: its solid range when split,
skipping sections without solid quads. LOD tiles and their fade stand-ins
draw and release from the translucent arena too. Batching is offered with
indirect multi-draws, not OpenGL's, and not where vanilla frames draw flat
(translucent_batchable). Every build_renderer caller passes the setting,
splitscreen and resource reloads included.
Off by default. RTX 4090, 1080p, off -> on: +4 to +16 % FPS at render
distance 32, up to +43 % at 64 (ocean 284 -> 406 FPS, 3,832 -> 354 draw
calls); arena memory +1 to +15 %.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Projectile.shootFromRotation adds the shooter's known movement and drops
only its y on the ground, and sets no rotation after it: an arrow shot
while strafing drifts with the player in vanilla too.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
As serde_json Values, vanilla's ~1,450 loot tables (0.6 MB of JSON,
~10,450 objects, a B-tree node each) held ~7 MB of a server's idle heap.
LootTables keeps each table's text and reads it into a Value the first
time it is rolled; every table is still read whole when it loads (a bad
file fails the load as before) and the native hook still runs on it, in
id order, before anything rolls it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: verified at e241409 on the Mac mini: fmt, clippy -D warnings (horizon-client-render), mesh_pool tests pass; the new assert fails without the fix. Thanks @Teiwazik.
The solid terrain shader could discard every pixel (the fade dither and
the alpha cutout), and one discard anywhere in a pipeline turns off the
hidden surface removal of tiled GPUs (Apple's) for everything it draws.
The world pass now draws each solid list twice: first the quads over
opaque texels that are not fading, through fs_opaque, which has no
discard, then the rest through fs_solid. Each vertex shader decides its
pass before any other work; a quad of the other pass lies outside the
clip volume. Shader packs and native materials keep the one pipeline.
M1 Pro, 1280x720, same commit before and after: GPU 10.4 -> 9.6 ms in a
flight over new ground, 12.9 -> 11.0 ms in rain, 9.5 -> 8.1 ms with 300
mobs in view. Screenshots are the same.
On the server-thread client, Game::tick returns tick_remote() before the
rope_rot_old write, so the first-person lead end eased toward the initial
(0, 10) every tick. Write it in tick_remote_level, after the unlit-chunk
early return, as the local tick does.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The wall mesh only holds the stretch inside render distance at build time
and was rebuilt only when the border itself changed, so after a respawn or
a long walk it stayed on the old stretch. Invalidate it when the camera's
section (floor >> 4 on X, Y, Z) changes, as LevelRenderer.repositionCamera.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
🤖 Agent: reviewed (the replay takes the shared flags' sprint bit the server sends, as vanilla's client); rebased on main; pr_ready --quick green on 2a1dd55 (1711 tests); the new test reproduces the live 0.0294/0.0455 signature before the fix; a +0.05 speed lie still counts from its first tick.
Read ProcessorNameString from the Windows registry instead of displaying the raw processor identifier.
Co-Authored-By: OpenAI Codex <[email protected]>
A client whose STOP_SPRINTING and START_SPRINTING reach the server in two
server ticks gets the STOP's shared flags and movement speed (no sprint
modifier) after it sprints again: its flag cleared, LocalPlayer.aiStep
starts the sprint again, modifier included, and sends nothing. The
replay took the attributes but not the flag, so it sprinted on without
the modifier until the START's attributes came: 0.029 then 0.045 off,
the official server's "ground sprinting" flags (#5935).
Record the sprinting bit of the shared flags the server sends the player
(Change::Sprinting) and apply it to the replay's flag alone, as
SynchedEntityData.assignValues does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: rebased onto main (conflict with #5856's frozen test helper, both kept); pr_ready --quick green on 30dfdc3 (1689 horizon-client tests, 19 prediction tests on vanilla data).
For 10 ticks after the server's item use ends, the replay also tries the
client still slowed by it: the client finishes a use on its own ticks and the
server on its own (food eaten one after another with the key held), and the
server stops one for its own reasons. Slower than the keys allow is no
advantage; a use the server has and the client may not still counts at once.
tools/anticheat_e2e.py: prediction-items (a shield, food and a bow used on
the move) and prediction-water (a deep pool, a current, a bubble column, a
ledge) with a real Horizon client.
Fixes#5860.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: proof: pr_ready --quick green (1680 horizon-client tests); write timeout = read timeout; local frozen-player run ends both threads; ships in the next batched restart for the live measurement.
#5915's last revision made the shared pairs references, so the sort's
comparison borrowed a reference: main's verify (clippy -D warnings)
fails on it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: proof: pr_ready --quick green on 98e49e3 (1680 horizon-client tests); live on the official server since 12:22Z as dd4b209 (this head minus the test commit), status pings ok every minute, 99 joins in 15 min; the deadlock of the first deploy is reproduced by the test without the fix.
A peer that stops reading (a frozen game, a dropped network) held its
writer in a full send buffer until TCP gave up, minutes later, and the
connection's place with it: play() joins the writer. Writes now time out
like reads. The turned-away report also says how many places are before
play and how many wait for a writer to finish.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: main moved after the queue (#5875 and #5755 merged), so I merged main in again (clean, d635bc57f). `cargo check -p horizon-client --lib --bins --tests` passes. Requeued.
Review on #5909: native versions published as .hmod would reach
launchers that drop their content or refuse them. The registry keeps
.hnmod until launchers and the game read native .hmods.
docs/api/mod-files.md holds every rule of a .hmod/.hnmod in one
place: zip and path rules, top-level entries, the horizon-native.json
schema, what mixins/<id>/ may hold and the size limits.
The registry's Cargo.lock catches up with horizon-mods' png
dependency so its --locked check passes.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: proven on a union with main c1f4a7c19 (+ #5875, #5755, #5679): `pr_ready.py --quick` passed (fmt, clippy -D warnings, horizon-client tests, doc tests). I added a lan_e2e check, `attack_over_slab` (97c18d712): a NoAI zombie stands behind a bottom slab, the player aims at its head over the slab and left-clicks. It passes for the host (20 → 19.06) and for the guest (19 → 18). Queued for merge. Thanks @repositorykreml1n!
🤖 Agent: verified by Ludovic's AI agent (Claude). ItemStack.isBarVisible needs isDamageableItem, which an unbreakable stack isn't, and the unbreakable unit component now reaches clients. pr_ready --quick passed on main c1f4a7c19 with this PR, #5875, #5903 and #5679 merged in (fmt, clippy, horizon-client tests incl. the new sync test).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against ShulkerBoxBlock.playerWillDestroy and Block.popResource (offsets drawn before the block_drops check, as vanilla). pr_ready --quick passed on main c1f4a7c19 with this PR, #5903, #5755 and #5679 merged in (fmt, clippy -D warnings, horizon-client tests, Python tests). CI parity covers the new command_exec cases.
🤖 Agent: playhorizon 4 CPUs: idle RSS −17 MB with THP off (241.7 → 224.7, every start) and −18 MB median with THP on (4/4), live heap −13.6 MB, Done 2.06 → 1.955 s (4/4); 20-bot MSPT unchanged (3 rounds). pr_ready tests 2066 + content roots pass; merge with main checks; fmt/clippy/Python clean on the Mac.
The deadlock fixed in the previous commit only shows with a log writer;
the test hangs without the fix.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
They live in their own lists, so a plate's entity count never saw them.
They now count, as vanilla's getEntitiesOfClass does. A rocket boosting a
player and a hook on an entity count but skip entityInside, whose step
their ticks skip too. A guest's hook ends when the guest changes level, so
it can't act on the blocks of the level it arrives in.
Fixes#5896
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018zuDTsnyHkP28bzn3u7S3m
loot_table_load_stays_hookable searched for "\n}\n", which a CRLF checkout
(core.autocrlf=true on Windows) never contains. Normalize as the other
source-reading test in this file does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
make_room logs turned-away connections, and the log reads by_address:
count called it with that lock held, so the accept loop deadlocked the
first time the server filled with a log writer set.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A .hmod's contents decide what it is: content only, native code
(horizon-native.json + mixins/<id>/), or both in one file. Native
bundle reading skips a .hmod's content entries, checks both
manifests name one mod, and refuses mod.wasm with a clear error.
The registry publishes native-mod versions as .hmod and still
accepts .hnmod; .hnmod keeps loading unchanged.
Launcher install, game content loading of native .hmods and the
GitHub install path are not done yet.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
decorated_region (through chunk_generation's oracle) failed in the nightly audit with
"oracle failed: " and nothing after it, and datafix with a CalledProcessError that
names the command; both Java oracles stopped with no exception to print. Show the exit
status (and the signal that killed it) and the last lines of its output.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
A file data/<ns>/recipe_filter/<name>.json lists filters that match
recipes by id or type (with *), by an input or an output item, and
remove them or swap one input or output for another. They run on each
recipe's JSON after the packs are merged and its load conditions held,
so they reach every recipe type, mods' own included, in every loader:
crafting, the stations, brewing, the recipe book and the recipes the
players are sent. A filter with an unknown key or not exactly one
action is skipped with a warning, so a typo never matches every recipe.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
When every place is taken, the oldest connection still before play
(handshake, status, login, configuration) that has waited more than 2 s
is closed to admit the new one. A crowd of sockets that never finish can
no longer keep players out; players in the game are never closed.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The connection cap (#5776, players + 64) filled up on the official server
within minutes of a restart and every new join was reset. Now one public
address holds at most 10 connections (local and private addresses, where a
proxy or a port forward hands everyone over, stay unlimited), the handshake,
status and login phases wait 10 s for the client instead of 30 s, and once a
minute the log says how many connections were turned away and which
addresses held the most places.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: fixes the Windows-only account test flakes that blocked launcher 0.2.33 (per-store account lock). Verified: windows-latest green twice on this head (runs 41300356019 push, 41300386770 dry run: 352 passed), Mac mini account tests + clippy clean.
In single-player the player's attack goes through the integrated server as a guest attack. guest_attack calls
look_at_entity, which aimed at the centre of the target mob; Game::attack then drops the hit when a block outline is
nearer than the mob along that ray. The attack packet names only the entity, so a mob that the player's own crosshair
hits through a gap (a top slab, a closed top-half trapdoor, ...) had its centre hidden and the hit was dropped.
For mobs, look_at_entity now aims at the centre when it is visible, else at the visible point of a grid of points in
the mob's box nearest to the centre (nearer the eyes on ties), found with a clip_ray from the eyes. A mob whose whole
box is hidden keeps the centre, so a really blocked hit is still refused. Other entity kinds are unchanged.
Fixes#5772
native_server_data_loaded(reload) fires once the server's data tables are
built: on a world's first open, before its chunks and players, and after
every /reload, once the players were sent the new data. NeoForge's
TagsUpdatedEvent with SERVER_DATA_LOAD. A mod drops caches derived from
the data and rebinds what depends on it there instead of polling.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
block_texture, block_texture_level, block_texture_size and block_texture_lod read the surface's
sprite anywhere on it. The LabPBR _level functions now count the texture's mip levels.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
Cloning a registry state bumped one Arc per name and value; world
generation threads clone the same states, and 20 bots cost 0.25 ms of
streaming. Shared strings are now interned once for the process (the
registries' names and values, never a player's) and copied as pointers.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
/summon minecraft:tnt {Passengers:[...]} dropped them. Load them as riders
(EntityType.loadEntityRecursive) and seat them on the TNT's top.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
SectionMesh::clear kept `lights`, so a mesh taken from the pool still listed the sources of the
section it last held: a broken torch stayed in the colour grid, and sources piled up elsewhere.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
Keep native hook targets and source paths stable. Run gameplay tests in the gameplay library, and build ctl/probes from horizon-client-lib without gameplay. Update build/test selection and the fixed mixin-tool pin.
Co-Authored-By: Codex <[email protected]>
Every unit test shared the process-wide account operations lock. On Windows a connection
to the tests' closed port (127.0.0.1:9) takes about 2 s to be refused, and tests that hold
the lock meanwhile made another test's status read answer "busy" after its 3 s wait
(Retry instead of SignInRequired), or starved a 5 s lock_within behind blocked waiters.
The in-process lock now comes from SecretStore::operations: the system keychain keeps the
one process-wide lock (one keychain per computer), a MemoryStore has its own, as its
secrets are its own. The launcher uses a single store, so nothing changes for players.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
So a flag can be found again in the world and played there (#5879: water
flags that don't reproduce in the harness).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
The HUD batch takes colours packed as hud::rgba (red in the low byte);
0xFFFF_FF00 there is cyan. AdvancementToast draws the label -256 (yellow),
or -30465 (pink) for a challenge.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
🤖 Agent: Shader series stacked in order (#5590 → … → #5757), each head clippy/test/fmt-clean on the Mac mini, the union checked in game with an all-keys pack and without a pack (vanilla). Merging once CI and parity are green.
So it no longer sits on the lines #5731 adds its field to: the two merge in either order without a conflict.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Every merge cancels the parity-main run before it. In a busy hour of merges, the
last finished run fell out of the 10-run window, green_main_head found nothing,
and no Nightly went out for 14 h (2026-11-01) although parity-main had passed
on bd8c53bfa.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Preserve the throw origin and inherited movement by preventing queued movement from overtaking a deferred UseItem. Compare standing and gliding throws against a real vanilla server.
Co-Authored-By: Codex GPT-6.1 <[email protected]>
Preserve collected components when breaking filled shulker boxes in Creative,
including with block drops disabled. Check all variants against vanilla through
the integrated server player-break path.
Fixes#5821
Co-Authored-By: Codex (GPT-6) <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5830 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5819 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5753 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5681 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5663 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5652 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5658 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5824 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5814 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5827 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5619 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PRs before this one in the series, merged in order with their conflicts resolved,
so this branch merges cleanly once #5590 has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: pr_ready --quick all green (1671 horizon-client tests, clippy, fmt); new replay test + 16 existing pass; anticheat_e2e --check prediction: honest client never flagged, cheating bot flagged 13x and sent back 12x.
When charging a bow while strafing on the ground, arrows drifted sideways
because player ground movement was added to the projectile velocity and
rotation was not aligned to the resultant vector. In addition, client-side
remote arrows did not update their pitch and yaw during flight, causing
them to fly horizontally like static lasers without following their
parabolic trajectory.
Align bow release with vanilla behavior by not injecting ground strafe
velocity and updating rotation in `shoot_from_rotation`. In
`remote_arrow_tick`, update pitch and yaw every tick using `lerp_rotation`
matching `AbstractArrow.tick()`.
Fixes#5855.
Co-Authored-By: Antigravity <[email protected]>
`native_biome_colors(biome, colors)` changes a biome's grass, foliage,
dry foliage and water tints. It is asked for every biome when a mod calls
`native_refresh_biome_colors()`, which swaps the mesh workers' biome table
and rebuilds the loaded sections and far terrain when the colours differ
(LevelRenderer.allChanged, as a Biome Blend change does).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QRKvdg3SMVP1onGqS3
native_mod_screen_open refuses what screen_hooks_open would: an overlay that leaves the screen at Playing (a
book, a sign editor, a dialog, a prompt), the world loading, no player, and anywhere but the window's own client
(no renderer, or a LAN guest's swap on the server thread), so a screen never opens that nothing could draw or
close. The docs and fragment say so, and that a gamepad can close it but not click on it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A client mod is asked once a frame for each mob in view (players aside)
with the mob's render state: id, kind, custom name, baby, partial tick,
age, body and head turns, pitch, walk cycle, hurt and death time. It draws
with the block entity drawing helpers under a pose at the mob's feet,
turned with its body, and sets `drawn` to leave out the vanilla model
(EntityRenderer.submit; NeoForge RenderLivingEvent.Pre, cancelled).
Name tag, shadow, fire and glowing outline still draw.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QRKvdg3SMVP1onGqS3
The registry's states were read into a BTreeMap of owned strings, then
shared: two allocations and a hash per string, 50 ms of the start. The
facts.json decoder now shares each name and value as it reads it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The renderer's own pass times were not reachable from the client. The F3
overlay now shows them under their own heading, next to the shader pack
group. A new gpu-passes control command times them without opening F3,
and the stats and bench summaries report them as gpu_passes.
The per-pass GPU timer only covered shader-pack passes, so the F3
overlay could not say which of Horizon's own passes cost the frame.
Give the same timer to the frame's own passes (world, occlusion test,
OIT, HUD, post, HDR, hand, outline, menu blur, native shadow) and add
compute_writes for compute passes.
Passes that already carry timestamps (shader packs, the frame timer),
screenshots, one-off clears and test code are left as they were. The
timer begins at the top of the frame so that passes recorded before
the first draw, such as sprite animation, are timed too.
Until the client says it loaded the level (at join and after a respawn), the
ticks it ends with no move are excused: it is on its loading screen, where it
doesn't move, even once the server stops waiting for it after 60 ticks. Its
first move counts as any other.
Fixes#5849.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
One bad file from a mod would fail its whole table and drop the world's
pack content with it; it is now skipped with a warning. The hook's docs
list every kind the tables read, trade sets, villager trades and dialogs
included.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
parity-main fails on dedicated_server now and then: "+play-start
minecraft:update_attributes". #5597 held the join tick's dirty attribute
instances for one tick, because vanilla syncs them on the next tick, after
the first chunk batch. But this server sends no batch before placed_at + 2,
and only once the player's chunks are ready: on a loaded runner the first
batch came after that next tick, so the attributes went out before it.
They now wait until a batch has been sent (unacknowledged > 0, or the
client already acknowledged one), which keeps vanilla's order however long
the chunks take.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
So #5675 (splitscreen), which edits the same two arms, merges with this in either order; only the one-line
serialize format string still meets it (keep both keys).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A block state's properties were a BTreeMap<String, String>: a 544-byte
leaf and two allocations per property for each of the registry's 35,723
states (about 30 MB of a dedicated server's live heap), for 237 distinct
strings. BlockProperties keeps them as a sorted slice of pairs, read like
the map (get, insert, iter, indexing, order, equality and hashing pair by
pair); the registry shares its states' strings through one Arc each.
States made elsewhere (chunk palettes, commands) own theirs, as before.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The gameplay facts load parses its 56 MB file twice: a validation pass that
only keeps the JSON errors, then the typed decode. The validation now runs
on a scoped thread while this one decodes, and is joined before the load
returns, so its error still wins. With a single worker (`parallel` false)
the two passes run one after the other as before.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The window now hands a key's characters to the character hook instead of
the key, so `world_options_key` must not read the missing text as a
backspace: only a key that carries text types, and a `None` stays the
control socket's `\b`.
The character hook fires after the key's screen effects, as GLFW's
separate char callback (`KeyboardHandler.charTyped`) comes after its key
callback, and its characters survive a cancelled press, as NeoForge's
`CharacterTyped.Pre` comes after a cancelled `KeyPressed.Pre`.
Co-Authored-By: DeepSeek V4.1 Flash <[email protected]>
Co-Authored-By: Claude Code <[email protected]>
GLFW's char callback comes after its key callback, so the character
hook now fires once the screen has acted on the key, on the screen open
then: a key that closes the screen reports no character, and one that
moves the focus types into the box it moved to. Options keeps its text
on the key, as its key path routes it between boxes and key binds.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
LivingEntityUseItemEvent.Start may change a use's duration, which
start_using_item keeps in use_duration (vanilla's useItemRemaining). The
first person pose asked getUseDuration again each frame for the remaining
ticks, and a server's client counted a food's crumbs from its consumable
time: both now count from use_duration, as getUseItemRemainingTicks does.
A use the server's flags start on the client takes getUseDuration, as
onSyncedDataUpdated does.
The fragment carries the catalog's neoforge, side and cancellable keys,
which the stamp tool now copies.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The content root appended each mod type's entity-types.json row by parsing
and rewriting the table, which needed serde_json's float_roundtrip to keep
a zombie's movement_speed to the bit. Cargo unifies that feature over every
crate in the game build (worldgen JSON included), while horizon-world's
parity probes, built alone, would parse without it. The rows are now
appended as raw text copied from the template's row, so no vanilla float
is reparsed and the feature goes.
The save test compared p.kind by value, then borrowed p: ProjectileKind
isn't Copy.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
While PackConfirmScreen is open the frame draws no screen under it, but
the render hook still reported the hidden one, so a mod drew over a
screen nobody saw. It now skips that frame, as the input hooks already
do through screen_over_playing.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The stub moves next to native_item_bar: #5804's block highlight took the
spot after native_tick. The docs said the batch uses screen coordinates;
it takes the same GUI units as x and y (draw_slot_item fills there too).
They now also list the other screens that draw item icons.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The stub moves next to native_camera_shake: #5804's block highlight took
the spot after native_tick. Its docs no longer say mods can draw their own
hand geometry: the hook passes no vertex buffer, pose or light.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Vanilla saves Abilities.flyingSpeed with the player, so a speed a server
mod set now goes into abilities.flySpeed and comes back on the next join;
a file's own speed is read. Every abilities packet comes from one
Guest::abilities, so the join, a game mode change, a respawn and a level
change all carry it (tested with the save round trip). The fragment is
renamed to native_motion for both helpers, FLYING_MAX says why 50 and what
the movement checks do with a fast non-owner, the guest comment names
the server mod, the fork wording goes, and docs/native-mods.md gets a section.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A key press a mod cancelled hands the characters no mod cancelled to the
text box, as vanilla's separate charTyped does; the control socket's
type \b no longer offers its backspace sentinel as two characters, and
its cancelled tap sends the release the docs promise. The key and
character hooks' docs name the control socket and say a key that closes
the screen reports its character there first.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
native_perform_console_command hears a command a mod registered when it is typed on the
dedicated server's console (level 4, no player), before vanilla's dispatcher; the answer goes
to the console through native_feedback. The single player owner's console is not covered.
Since #5100, turn_to_base_block (a dirt path or farmland turning to dirt)
emits a push_entities_up block event, which the block-update probe printed
as "X push_entities_up". Vanilla's Block.pushEntitiesUp only moves the
entities in the new block's box, and these cases have none, so the oracle
logs nothing: the dirt_path case differed and parity-main failed. The probe
now skips it, as it skips the other events that need someone there.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
native_quad draws through HudBatch::fill_quad (the HUD atlas's white, as native_fill) instead of borrowing the
lightning renderer's white_uv. The rules for opening a mod's own screen move into mod_screen_opens, with a test:
only over the world, never over another screen, the title screen or another mod's, and again a no-op. The docs
say every mod's own screen shares the id `mod` (check native_mod_screen() is yours) and that it doesn't pause a
single-player game.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Renamed from native_get_clock (no other helper has a get_ prefix). Both fields are i64 ticks as
Level.getGameTime and getDayTime return them, without the clamp to 0: the day time counts on past 24000
and /time set can make it negative, and the docs say so. Documented as the shown world's clock, for
client mods.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
With it off the static pool is vanilla's again (247 channels to whatever starts first), so the sound_engine
parity harness is unchanged; `uiSoundReserve` (Extras > Sound, "Keep channels for UI sounds") turns on the 16
channels positional sounds leave for the ones played to the listener. The test checks both, and its
`use test_backend::Null` moves into it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: playhorizon A/B, 20 bots, 3 rounds: tick_lan 4.22→3.96 ms (−6 %), MSPT 25.65→25.32. Same bytes (new test vs from_values). fmt + clippy -D warnings pass (Mac mini); workspace tests 3982 pass (playhorizon); the 2 Python failures there are its VSCREEN env, they pass on the Mac mini.
A native mod can now send a particle of its own type, drawn by every client
from the sprite set the mod's pack ships, instead of borrowing a vanilla one.
The type is the mod's namespaced name in ParticleOptions::named ("mymod:spark");
the client builds it as vanilla's SimpleParticleType with a sprite set: a
plain billboard, no tint, 20 to 40 ticks, on the engine's own random. The set
comes from assets/<namespace>/particles/<name>.json, the file vanilla's
SimpleParticleType reads; its textures resolve as particle/<texture> in the
atlas, the same rule vanilla's own particle JSONs use.
On the wire a mod particle is Horizon's own extension: the sentinel id
MOD_PARTICLE_ID (-1, never a vanilla registry id) followed by the string,
which vanilla's Particle codec has no room for. A client that does not know
the mod ignores it like any unknown particle, and native_send_particles
refuses an unnamespaced name.
native_send_particles keeps its signature; a mod builds the options itself,
so no hook point moved. NATIVE_HOOKS_API is 22 for the new capability and
docs/native-mods.md says what is supported.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
The dragon was drawn only while the player stood in the End, so a
dragon summoned in the Overworld (/summon ender_dragon, which the
server sends to the client as an AddEntity) was heard, flapping and
growling, but never shown. Vanilla's EnderDragonRenderer draws an
EnderDragon in any dimension like any other entity; the gate was
added with the fight, when the dragon only ever existed in the End.
The dragon fireballs share the same block and are now drawn in every
dimension too.
LevelTicks and ChunkTicks hash a (type, position) key at every schedule,
poll and dedup check, and the chunk map at every lookup, with std's SipHash.
They are only looked up, never iterated in order (the keys of an area are
sorted), so they now use the Fx hasher, which moves from
horizon-client-render to horizon-core (re-exported under the old path).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The server doc still listed bow PvP between guests as missing; it works since 5763854b0 (lan_e2e 'arrows'), and a dedicated server run with two protocol clients agrees.
🤖 Agent: #5786. Auto-jump now in ServerboundPlayerInputPacket as vanilla's keyPresses. Staircase walk with autoJump on, same server: 75 prediction flags with main's client, 0 with this one. pr_ready --quick all green on 7a8a97dd5.
LocalPlayer.sendChanges sends input.keyPresses, which an auto-jump's makeJump
sets to jump for that tick. Horizon sent the keys as held, so with Auto-Jump on
the server saw jumps its keys didn't allow: the anticheat's prediction flagged
every auto-jump and the 0.4 a tick it left (75 flags on a staircase walk, none
with this change). The official server's log has the same lines.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: #5725. One prediction flag per slip across instead of 10-20 (87% of live lines continue an earlier slip). Strict cheat tests unchanged; new tests for a missed push (counts once) and a kept-up 0.05 speed lie (35 of 40 ticks count). pr_ready --quick all green on 7a8a97dd5.
Leads dropped through break_leash without telling native_loot_drops,
so a mod couldn't see or change the lead a player gets back, nor the
one a leashed mob drops when a player kills it.
Untying a lead, cutting it with shears (on the mob or its fence
knot) or re-leashing a mob from another holder now passes it through
the hook as a use (break_leash_used: the leashed mob, the acting player).
A mob a player kills while leashed records its killer (lead_killers);
the corpse drops its lead on its next tick as vanilla does
(Leashable.tickLeash), as a call of its own naming that killer. A lead
that snaps or loses its holder names no player and drops as before.
The API number comes from the lead_drops fragment.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
87% of the official server's prediction flags (6,002 of 6,900 lines on
2026-11-01) continue a divergence that started on an earlier tick: once the
replay's velocity is off (a hit it missed), the offset dies out by the air drag
(0.91) or ground friction each tick, and each of those ticks flagged again.
The first tick off after ten that agreed now has its follow-up tick try the
velocity the client was left with (across only, at most 0.5), as another
candidate. A single slip then counts once; a blink leaves the velocity alone; a
tick that agrees only thanks to it never becomes the setback point. Not up and
down: the vertical drag (0.98) would let a slow fly pass for a slip.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: #5775. Official server crash 05:15Z (spawn worker WouldBlock, no save). Worker spawn failure non-fatal; connection threads capped at max_players+64. New flood test fails without the cap, passes with it; pr_ready --quick green apart from one ENOSPC test failure that passes alone.
A dedicated server crashed on the official server (2026-11-01 05:15Z) when the
system refused a new thread: Streaming::new's .expect("spawn worker") panicked
and the process died without saving. Worker spawn failures are now logged and
the pool runs with the workers that started.
The server starts a thread per accepted socket (vanilla's Netty uses a few for
all): past max_players + 64 live connection threads, a new socket is closed at
once, so a flood of silent sockets can't use up a container's thread limit.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A section's first encoding unpacked its 4,096 block states, then built a
palette from them again (Paletted::from_values). For a list palette the
stored ids are now renumbered in first-seen order instead: the same palette,
indices and bytes, without the values. Global palettes keep the old path.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The integrated server retained minecraft:unbreakable, but network_stack
never emitted its unit component. The client therefore still drew a bar
even with the rendering guard in the previous commit. Send UNBREAKABLE
as Component::Unit; the client decoder already preserves unit components.
Add a regression test through stack_to_network, Slot write/read and
from_network, checking both ordinary and unbreakable damaged hoes.
It fails without the encoder entry and passes with it. Verify the actual
server-thread client retains the component and loses only the unbreakable
hoe's bar in the hotbar screenshot.
Written with Claude Opus 5.5 in OpenCode.
tick_creature_entities takes the mob list, ticks it, and appended it back
into the emptied list: a new buffer and a copy of every Mob each tick (and
the old buffer freed). With no mob added meanwhile the list now goes back as
it is. The dead-mob set and the guest bindings, looked up per mob each tick
and never iterated, use FxHash instead of SipHash.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: #5725 step 2. Guest item use (eating, bow, shield) now slows the replay as it slows the client. Vanilla 26.3 client walking/sprinting while drawing a bow 3 min: 503 prediction flags → 2 (0.045). pr_ready --quick all green on 3bc251ea5.
`native_mod_screen_open(mod_id)` opens a full-screen menu of a client mod's own over the world (`Screen::Mod`, its
`NativeScreenView.id` `mod`), `native_mod_screen_close` closes it, `native_mod_screen()` says whose is open, and Escape
closes it as any screen. It draws nothing itself: the screen hooks fire over it as over any other, so the mod draws it
in `native_screen_render_post` and takes its clicks, scroll, keys and typing from the mouse, scroll, key and
character hooks (#5543). Two drawing helpers beside `native_fill` for that: `native_quad` (a wedge or a slanted bar)
and `native_item_model` (an icon drawn with another `item_model`).
Ported from Stink-O/horizon ae531730c (`ui.show`) and 848a0686c, reshaped onto the screen hooks.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
An item in use (food, a bow, a shield) slows the client to its
use_effects speed (0.2 for most) and stops its sprint. The replay plays a
guest on the server's copy of its body, which takes the item use from
Game.sync_player_body, and that runs only for the window's own player: a
guest's copy never had one. So the replay walked a guest who was eating
or drawing a bow at full speed. That put it 0.8 of a walk (0.173 blocks a
tick) or of a sprint (0.224) off, every tick of it. 0.173 was the official
server's most common prediction flag.
guest_replay now sets the copy's item use speed and its no-sprint flag
from the item the guest is using, as it already did for the glider. A
vanilla 26.3 client sprinting and walking while drawing a bow for 3
minutes: 503 flags before, 2 after (0.045 blocks).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
mimalloc backs the game's memory with transparent huge pages where Linux
allows them; many are only partly used. low_memory = true starts the game
with MIMALLOC_ALLOW_THP=0: 60 MB less idle and 190 MB less at 20 players,
for ~1.2 % more tick time (playhorizon, 20 bots, 5 rounds). Off by default.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
native_loot_drops heard block breaks and mob deaths only, so a mod
couldn't tell, or change, what a player's shears, brush or hit
dropped, nor who caused it. An auto-pickup or a global loot modifier
needs those too.
The player-use drops now pass through it, with the acting player
named and the new NativeLootDrops::used set: shears on a sheep, a
mooshroom, a snow golem or a bogged; a brush on an armadillo or a
suspicious block; shears carving a pumpkin; picking sweet berries or
cave vines; and a player's hit taking an item frame or its item. One
helper, native_use_drops, builds the context. The berry and cave vine
drops are popped inside the block behaviour, so their run of spill
spawns is tagged and retold when the spill flushes (retell_runs),
keeping each stack's spot and the random draws when no mod changes it.
The API number comes from the use_drops fragment.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
ItemStack.isBarVisible is isDamaged, which needs isDamageableItem: a
max_damage and a damage, and no minecraft:unbreakable. draw_item_decorations
only looked at the item's max_damage and the stack's damage, so a damaged
unbreakable stack (/give @s diamond_hoe[damage=58,unbreakable={}]) showed a
green bar. It now asks stations::is_damageable, the port of
isDamageableItem the anvil and the grindstone use.
Fixes#5721
Written with Claude Opus 5.5 in OpenCode.
Block light has a brightness but no colour, so a pack can't make a soul torch blue. Section meshes now
list their light sources, and a 3D grid around the camera holds the colour reaching each cell, read
through block_light_color. Block light keeps its own brightness, which walls stop, so the grid only
tints it and needs no occlusion of its own.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
Since #5641 a saved snowball, egg or ender pearl loads as a live projectile,
but NoGravity was dropped: the scene's floating snowball and pearl fell into
the floor and were gone before the save. Entity.getGravity is 0 with
NoGravity; the projectile now reads it, honours it and writes it back.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A mod could cancel an experience orb in native_entity_join_level but
not read its value first, nor give experience to anyone. An
auto-pickup or a machine that pays out experience needs both.
native_experience_orbs(dimension, min, max) reads the orbs in a box,
as native_item_entities does for items. native_give_experience(uuid,
points) gives points as an orb a player touches does: Mending repairs
first (repairPlayerItems), then giveExperiencePoints with the level-up
sound. It reaches any player from any server hook, through as_guest
for a guest, and refuses one swapped out by another player's hook.
native_give_levels(uuid, levels) adds levels or, negative, takes them
off (giveExperienceLevels), as the anvil and enchanting table charge.
Both go through one dispatch, native_as_player.
The API number comes from the give_experience fragment.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Follows the .jem reader's review changes: files over the model's byte
budget are refused by size, and the reader's warnings reach the log.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: log-only detail for prediction lines (#5725 step 1); unit tests for the worst tick's offset/state; vanilla client on a dedicated server logs the new detail; pr_ready --quick all green on 21a12582c.
- Strings read or copied out of the files (part names, ids, animation
targets and expressions, axes, numbers written as text) cost one per
started 64 bytes, each time; animations pay theirs before the copy.
- A submodel costs its keys, a part the layers its keys go through.
- by_id no longer copies an id it already has.
- Tests: shared .jpm with a 1 MB expression and with 200k submodel keys,
the .jem's bytes in the byte budget, non-object animation items, junk
boxes that only exhaust when keys are charged.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Main rebuilds a heap of wanted tiles when the view moves and selects
incrementally. The backdrop becomes a class in `Wanted`'s order (seam,
backdrop, the rest), and `evaluate` keeps split tiles at the seam from
standing in, as `Plan::select` does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: launcher on GPUI Fast (Apache-2.0; fixes the NVIDIA/blade crash #3812); macOS text fixed (font-kit) and checked against main by screenshot; launcher tests + clippy pass on macOS; approved as code owner.
- pow follows Java's Math.pow: a NaN exponent is NaN, and 1 or -1 to an
infinite exponent is NaN; x to 0 stays 1, NaN included
- Test clamp with a NaN upper bound
- The too-long test checks the error lands inside the expression instead
of pinning an offset that depends on when nodes are counted
- Boundary test: a 512-term sum (1023 nodes) compiles and runs, 513 terms
are too long
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Without the feature gpui-fast-platform has no font source on macOS: the
launcher drew no text at all and half its home page was blank. With it,
the home page matches main's GPUI build.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A prediction line said only how far off a move was. On the official server
it logs about 15,000 lines an hour, from ordinary players (mostly 0.1 to
0.2 blocks) and cheaters (1 to 10 blocks) alike, and nothing in a line
tells a glide, water, a ladder or a lag from a cheat, so the check can't
be trusted to act.
Each line now ends with its worst tick as the replay played it: where the
client said it went less where the replay ended (x, y, z), how far the
client said it moved, and what the replayed body was doing (ground or air,
water, gliding, sprinting, sneaking, swimming, flying, riding). Only
flagged ticks pay for the text; the state is a few bits per played tick.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- An entry the budget cuts down to no quads is left out, not drawn empty
- Animations are all or nothing per entry, checked and charged before baking
- A box costs max(faces, keys, 1); non-object items of boxes, submodels and
animations cost 1, and so do non-scalar animation values
- MAX_BYTES (now 4 MiB, pub) bounds the .jem and all its .jpm files together;
a cached .jpm counts once
- Boxes whose corners or UVs overflow to non-finite numbers are skipped
- baseId dedup compares only against the entry's own .jpm layer
- Tests: a real lone-surrogate escape case, valid oversized JSON for the byte
limit, the shared byte budget, repeated keys, non-finite boxes, 1 unit left
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- clamp keeps a NaN value NaN, like Java's x < lo ? lo : Math.min(x, hi)
- round returns Java's int: NaN is 0, results saturate at i32::MIN/MAX
- calls take up to u16::MAX arguments, so long keyframe tracks compile
- tests: laziness values, an expression just under the node cap, the
offset of the too-long error, a 300-frame keyframe, clamp/round NaN
and infinities
Co-Authored-By: Claude Opus 5.5 <[email protected]>
`native_push_player([x, y, z])` on a client, added to the local player's velocity as an explosion's knockback is (the window's player moves itself, as vanilla's client does), and `native_set_flying_speed(player, speed)` for server mods, a player's creative flying speed kept as vanilla keeps its abilities (a respawn or a game mode change keeps it). A test for the push. Ported from Stink-O/horizon 63ea5c582 and 848a0686c.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
`native_get_clock()` gives a client mod the world's game time and time of day in ticks (none on the title screen), as `Level.getGameTime` and `getDayTime`: for anything timed to the day, the sun or the world's own clock. Ported from Stink-O/horizon's `game.get-clock`.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A light job for a column of the last ring reads its neighbours past the
area as views of their decorated proto chunks. The server thread built
those views itself while it scheduled light (~8 columns per tick at 20
players). The proto chunk is now checked out to a worker, like a save,
which builds the view; the light job is scheduled once it is back.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- keyframe: no usize overflow for infinite, huge or NaN k; past the end is the last frame
- parser caps nesting at 64 and nodes at 1024 ("expression too deep/long" with an offset),
so no pack can overflow the stack while parsing, folding or dropping
- programs deeper than the 64-value stack (long keyframe tracks) run on a per-thread buffer
- ifb returns 1/0; nbt(key, exists:false) compiles (raw test skipped, value 0)
- catch falls back on NaN only; min/max return NaN like Java's; round is Java's, exact
- unseeded random() starts each thread somewhere else; unary + parses
- !, && and || lower to comparisons and ifs (Not/And/Or/Bool gone); easings derived from
ease-in and moved with the splines to cem/curves.rs; Expr::slots removed
- bench prints CPU time; Error documented; tests for each of the above
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- Past the node budget, later parts are left out instead of replacing their vanilla part
with an empty bone.
- Budget: every box costs at least one, the last box keeps the faces that fit, and the
"left out" warning fires only when something was cut (Bake { left, cut, unread }).
- Animation pairs count against the same budget.
- Layers carry a key index and a base's layers already present are not added twice.
- Per-part textures and sprites are noted in submodels too.
- JSON: a high surrogate before a non-low escape gives U+FFFD and the escape is read on its
own; one shared `get`.
- Tests for byte limits, a missing .jpm named twice, own keys over a .jpm's (also in a
base), the budget's edges, surrogates and comment marks in strings.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: merged main. Fox trust list, save/load, breeding and egg-on-fox trust (rdbtCVS's fix, approved). pr_ready --quick on a union of #5414#5418#5433#5441#5442#5450 merged into main f99f8feaa: fmt, clippy -D warnings, 2403 client/render tests, 510 Python tests pass.
🤖 Agent: merged main. Canceled hits forget their attackers on every damage path (PvP, /damage, thorns, dragon, blast, potion); test. pr_ready --quick on a union of #5414#5418#5433#5441#5442#5450 merged into main f99f8feaa: fmt, clippy -D warnings, 2403 client/render tests, 510 Python tests pass.
🤖 Agent: merged main. ResolvableProfile texture patch drawn for heads, read from the active resource root, reloaded on pack change. pr_ready --quick on a union of #5414#5418#5433#5441#5442#5450 merged into main f99f8feaa: fmt, clippy -D warnings, 2403 client/render tests, 510 Python tests pass.
The fallback for roots no command Target stands for checked only orbs after
the canceled-orb fix; end crystals and cushions also take an id at once
(rdbtCVS's re-review).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The native-mods PRs each add a section to docs/native-mods.md at the same
place; holding the one ahead lets the series merge in one pass.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
# Conflicts:
# crates/horizon-client/src/game/native_hooks.rs
# docs/native-mods.md
The native-mods PRs each add a section to docs/native-mods.md at the same
place; holding the one ahead lets the series merge in one pass.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
# Conflicts:
# crates/horizon-client/src/game/native_hooks.rs
# docs/native-mods.md
The native-mods PRs each add a section to docs/native-mods.md at the same
place; holding the one ahead lets the series merge in one pass.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
# Conflicts:
# docs/native-mods.md
🤖 Agent: merged main. BucketItem.emptyContents in the Nether: sendParticles(LARGE_SMOKE, corner, 8, 1,1,1, 0, ALTERNATIVE) now reaches clients; ALTERNATIVE spread added. pr_ready --quick on a union of #5414#5418#5433#5441#5442#5450 merged into main f99f8feaa: fmt, clippy -D warnings, 2403 client/render tests, 510 Python tests pass.
🤖 Agent: merged main. MinecartSpawner: BaseSpawner ticks at the cart, saved/loaded, drops and picks as a minecart, entity event broadcast. pr_ready --quick on a union of #5414#5418#5433#5441#5442#5450 merged into main f99f8feaa: fmt, clippy -D warnings, 2403 client/render tests, 510 Python tests pass.
Second review round of the .jem reader:
- an entry's keys are layers (own, .jpm, baseId's) over shared objects,
so a big base or .jpm costs nothing per entry; the first entry with an
id is the one baseId finds
- the 1024-entry cap counts every entry, bones count against the face
budget, and a .jem or .jpm over 8 MiB is refused
- the JSON reader takes Gson's = and => for colons and ; between items,
quoted empty keys, and \u surrogate pairs; \u needs four hex digits
- per-part textures and sprites, not drawn yet, get one warning
Co-Authored-By: Claude Opus 5.5 <[email protected]>
On Windows a backslash separates folders, so a pack's ..\x would leave
the resource root once joined onto it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Include GitHub and local native bundles in both installed-mod views, retaining source/update actions and the existing build confirmation.
Co-Authored-By: Codex (GPT-6) <[email protected]>
Review fixes for the .jem reader:
- a small JSON reader keeps keys in file order (animations read what
earlier ones wrote) and accepts what Gson does: a BOM, comments,
trailing commas, single quotes, bare keys
- an entry's own .jpm wins over what it inherits through baseId; each
.jpm is read once
- .jpm paths go through paths::resolve, which refuses ones that leave
the pack; a missing or broken .jpm, or a baseId naming nothing, leaves
that entry out with a warning instead of the whole model
- rotations through an odd number of inverted axes turn the right way
- numbers must be finite; attach and animation values may be strings or
numbers; a file builds at most 1024 entries and 65536 faces
- shadowSize, per-bone textures and attachments are no longer parsed
until something draws them
- CLAUDE.md lists the crate
Co-Authored-By: Claude Opus 5.5 <[email protected]>
shaders.properties' `vignette` key (on unless `false`) says whether the game draws its vignette
over the pack's image; most packs turn it off (BSL, Complementary, Bliss, Bloop, FastPBR,
Hysteria...) because they draw their own. Horizon ignored the key, so with the Vignette option on
those packs got both. The game's vignette now draws only when the option is on and the loaded
pack doesn't turn it off.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
ShelfRenderer passes its block entity as the ItemOwner (at the block's
centre, turned by its facing), so a compass on it points to its target;
without one it always points the same way. The client's copy of a block
entity has no item slots: the docs say to send the stacks drawn in the
update tag.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
dropped_item_geometry resolved and meshed every item within 64 blocks,
even behind the camera. Apply EntityRenderer.shouldRender's frustum test
(box grown by 0.5) first, as mobs already do: ~0.2 ms -> ~0.01 ms per
frame with a few dozen items around.
draw_sections left a translucent sorted-index page bound, so the cloud
draw after it indexed past its end and wgpu aborted on the first frame.
Rebind the shared quad indices when a sorted page was used.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
draw.keep() in native_block_entity_submit keeps what that call drew,
relative to the block as a sign's text is kept, and the block is drawn the
same again without calling the mod until its state, its block entity, its
light or the resource packs change. Idle machines then cost a copy moved to
the camera, not a call into the mod, each frame.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Enderman.teleport(x, y, z) fails for a passenger, repeatedlyTryToTeleport
returns before drawing, and EndermanLookForPlayerGoal.tick skips its
teleports while the enderman rides something. Horizon had none of these
checks, so an enderman in a boat "teleported" every tick (sound and
portal particles) and was snapped back to its seat, which also made its
held block jitter. Fixes#4865.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A server's client keeps thrown tridents apart from its other projectiles,
so TakeItemEntity never found them and the trident just vanished instead
of playing vanilla's ItemPickupParticle. Look the trident up there too,
turn it into a pickup particle and draw it on its way to the collector.
The host also sent the pickup only for tridents that gave an item, so a
creative player's trident vanished on other clients; it now always sends
it, as Player.take does in AbstractArrow.playerTouch.
lan_e2e: trident_pickup throws a trident at each player's feet and checks
that it comes back and flies to them on both clients (fails before this).
Fixes#5577
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Java-format packs read both names as zero today. isElytraFlying is the
player's fall-flying state; cloudHeight is the cloud_height environment
attribute (default 192.33, dimension_type visual/cloud_height) when the
dimension's cloud colour alpha is above 0, as in vanilla's LevelExtractor,
and NaN otherwise ("no clouds"). The 192.33 literal of clouds.rs becomes
a shared constant.
Part of #3172.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
#5656's falling motion (-0.3015 against vanilla's -0.2277) is a tick of
gravity more, not a change in the game: main, and the exact tree CI
failed on (#5502's head merged into 76dc5e3bd), pass every time on a
quiet machine. On a loaded one the server catches up with two ticks for
one of the jump's moves (seen: "the server ran 2 ticks for move 7"), and
the server's copy of the player falls in the tick with no move. Before
the jab that changes the motion the jab sends.
LocatorBarClient's arc now records an error when that happens up to the
jab, so the failure says why, and lunge_motion joins known_flaky so the
runner gives it its one retry.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The camera's motion vectors see mobs and dropped items as still, so a made frame
would show them where the real one does. The game now tags each mob's and
dropped item's vertices with its entity id (Frame::entity_motion); one whose
model has the same shape as last frame gets each vertex's own motion, drawn into
a texture at the render size where the world shows it (its depth, or a
translucent item's texture). The interpolation takes that texture as
object_motion and uses it over the camera's motion vectors. Nothing runs it yet:
the Frame Generation setting comes next.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The Build button moved from Settings > Native mods to the Mods page, but the
message after installing a mod from GitHub still sent players to the old
setting (reported on Discord, 2026-11-01). Updated in English and the ten
translations.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
mob_attributes asked the attribute catalog for movement speed (by name)
and recomputed its base value up to three times per mob and tick, on
top of the registry id built with format! per attribute (Attr::id, the
previous commit). The speed instance, the baby and the sprinting
modifiers now share one lookup.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The type's default of an attribute (base_value, read by mob movement and
every tick's entity tracking) and the LAN attribute sync built the id
"minecraft:<name>" with format! on each call: for a server's 2000 mobs that
is several allocations per mob per tick. Attr::id returns the id as a
static string.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Keep entity portal state with its level so another ticking dimension cannot
clear it. Handle falling-block portals after movement and impact damage but
before placement, preserving cooldowns, saved data and vanilla's End transfer
duplication branch. Extend integrated-server command parity with direct
falling-entity probes and remove its stale in-frame runner override.
Co-Authored-By: Codex (GPT-6) <[email protected]>
Noise, lookup tables and the like had to be computed in WGSL. A pack can now ship up to eight PNGs,
which the check decodes and every stage reads as <id> and <id>_sampler.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
A fill visits up to 32768 blocks of a handful of states. It now skips
blocks that are already the block it would place before matching the
filter (place_state would change nothing there), and matches each block
state against the filter once per fill instead of once per block.
Results, counts and side effects are unchanged. Ported from
Stink-O/horizon 242b45037 (its fill part, the same code).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Positional sounds leave 16 static channels free for sounds played to the
listener. Ported from Stink-O/horizon a72b65ea3 (sound_engine.rs, now in
horizon-client-render); that commit's other half, block destroy events
only within 64 blocks, is already here (lan/containers.rs level_event).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
An item stack drawn under a block entity's pose as a campfire's, and a block
state's property (its facing) read without a world lookup. Helper-only, with a
fragment of its own.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: unit tests for the split choice and resplit; HORIZON_SPRITE_SPLIT=512 renders the same world as the default on the Mac mini; pr_ready --quick green. Merging after verify + parity.
Give the pause action its own translated Splitscreen Settings label. Check the transient test override through the visible saved switch rather than waiting for an untouched settings file to be written.
Co-Authored-By: OpenAI Codex <[email protected]>
🤖 Agent: merger STOP was a conflict with main after #5571 landed; main merged (clean). clippy -D warnings, the native hook tests (55) and rustfmt pass on it; the rest was proven in union3 with #5571.
A GPU with 512 layers per texture array refused to start ("GPU supports 512
texture layers; 1291 required") although terrain binds 8 sprite arrays. The
split is now the smallest of SPRITE_SPLIT, HORIZON_SPRITE_SPLIT and the
adapter's max_texture_array_layers (one less on OpenGL when a full array would
be padded past the limit), computed once when the renderer is made and used
by the sprite arrays, the shader constants and the shader pack atlas. Mip
chains prepared before the adapter was known are cut again for it. Startup
only fails when the sprites need more than 8 arrays of that split.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Rebuild the profiles and saved switch increment on current main after the compositor, controller, windows and profile-page model landed. Preserve pause dialogs and keyboard navigation, keep test overrides transient, and validate profile folder names.
Co-Authored-By: OpenAI Codex <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: vanilla's Pig/Strider getControllingPassenger use player.isHolding, which checks both hands; the server path (lan_vehicles → tick_mount_riding) shares this check. Main merged; its test passes. union3 = main ba8b5b414 + #5516#5535#5566#5571#5578#5573#5569: pr_ready --quick passes (python 511, rustfmt, workspace tests 2856, content root); clippy -D warnings clean after the #5573 borrow fix.
Vanilla 26.3's PhantomSweepAttackGoal.tick bites on a plain box check
(getBoundingBox().inflate(0.2).intersects(target)) with no line-of-sight
or block test, and its targeting gate (TargetingConditions' eye-ray) is
what keeps a pane-hidden player safe. Horizon ports both, so:
- a phantom pressed against the far side of a foot-high pane wall, whose
eye-ray passes over the wall, bites the player pressed behind it, in
Horizon exactly as in vanilla;
- behind a window-height pane wall the eye-ray is blocked and the
phantom never takes the player as a target, also as in vanilla.
Two parity cases pin both sides of that behaviour, mirroring the live
geometry (player and phantom each pressed against their face of the
pane, 0.125 apart, bite reach 0.2). Refs #4928.
🤖 Agent: Review points from some-du6e fixed and confirmed (seat_own_stands ordering, the cancel() test, port-ledger citations); local lan_e2e proof posted by the author's agent; CI green on 6f941a5e9. Queued for merge.
🤖 Agent: launcher account status, profile heads, refresh and targeted sign-in (Microsoft only); rdbtCVS's points fixed earlier; main merged here (lang union, both sides' account tests, a string main dropped removed). Proven: union of main 956e35469 with 19 PRs, launcher clippy + tests green (the i18n tests after the string fix), game clippy/tests green.
🤖 Agent: native loot table and drop hooks. Fixed here: NativeLootDrops.player only while the hurt memory lasts (dropFromLootTable's playerKilled). Proven: union of main 956e35469 with 19 PRs had this PR's code (clippy, workspace + content-root tests green); the author's later commits are docs, a test and rustfmt, and CI verify passes on this head.
🤖 Agent: native consume effect kinds; fixed here: an empty namespace is minecraft's. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: native block contact / landing hooks; fixed here: no item/projectile sweep without mod blocks (hot path), vanilla stepOn-then-inside order, getOnPosLegacy. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: native bone meal hooks; fixed here: Some(false) keeps the meal as growCrop does, None leaves the template's. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
"Microsoft account · signed in" came back with this branch's tables in the
main merge; the code lost it in 86149f7da (i18n tests).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: fence knot clicks and orb smoothing; fixed here: the orb draws its correction step (previous kept), knot hit only within entity range; main merged. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: #5521 by rdbtCVS finished (main merged: serialize keeps screenshotQuality and main's shadowCache). Screenshot Quality defaults to 1x, so at 1x and 100% scale F2 saves the window's frame as before; the scaled frame draws offscreen inside OOM/validation error scopes. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green (clippy -D warnings, workspace tests).
🤖 Agent: native block entity renderers; main merged here. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: OptiFine .properties reader (new crate, no dependents); fixed here: no .. paths, linear glob, no link loops; skip-release. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: parallel horizon-prepare (same output digest, release check covers both lane layouts); fixed here: the extraction test on case-insensitive disks. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: native can_survive / may_place_on hooks; fixed here: a piston head's neighbour change asks can_survive. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: native advancement criteria triggers (+ the author's on_fire fix); main merged here. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: biome modifiers place mod features; fixed here: unread modifier types and unknown names are skipped with a warning instead of dropping every content mod. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: splitscreen part 9 (on #3801, testing only); main carried up the stack. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: splitscreen part 8 (on #3794, testing only); main carried up the stack. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: splitscreen part 7 (on #3777, testing only); main carried up the stack. Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
🤖 Agent: splitscreen part 6 (HORIZON_SPLIT testing only, off by default); main merged here (main's headless loop runs before the split's event loop). Proven on a union of main 956e35469 with 19 open PRs including this head: pr_ready --quick green for it (clippy -D warnings incl. the server build, workspace tests, content-root tests, Python tests). CI parity covers the harnesses.
A native pack had no way to light or tint weather, so vanilla's streaks stood out against a pack's
own lighting. Horizon still places and animates the streaks; the pack only returns their colour.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
origin/main's identity hashes mod components' defaults so peers that
disagree are told apart; it read items.json raw, so a native mod's change
slipped past it. It reads them through item_defaults::read now. The read
and hook docs mention the stack-size range.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Server mods could only give plain items to the acting player
(native_give_items). An auto-pickup or a machine's output needs to hand
over the stack it has, with its damage, name and enchantments, from any
server hook. native_give_stack(uuid, stack) places it as Inventory.add
does and returns what did not fit.
Game::give's slot loop moves into inventory_add, shared by both, so
Inventory.add stays one port. The API number comes from the
give_stack fragment.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests).
Each optifine/cem (or EMF's emf/cem) .jem replaces or joins the parts of
the model layer it names, with OptiFine's part names mapped to the game's
and baby models falling back to the adult's. Built when the packs load
(and on every pack reload, on the reload worker); a mob's draw only swaps
the layer's part tree and, when the .jem names one under textures/, its
texture.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: stop riding a remote vehicle that is gone (Entity.baseTick); approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests). Stacked on #5630, tested together.
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests). Stacked on #5630, tested together.
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff. pr_ready --quick passed on main c4a65832e with this PR merged with 16 others from the same sweep (fmt, clippy -D warnings, horizon-client, -core and -render tests, Python tests). The image follows the text and the bundle component, the rule Horizon's bundle tooltip already uses; moving both under the title, as vanilla's index 1, is a separate parity fix.
🤖 Agent: tadpole AbstractFish goals and TadpoleAi brain, mob_ai scenarios; approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: fancy leaves sorted front to back on the mesh workers (no render-thread cost); approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: six-face panorama check in horizon-prepare check.rs, rustfmt fix; approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
A pack's sky stage only replaced the dome: Minecraft's clouds, sun and stars still drew over it, so a
pack with its own had two of each. Each key defaults to true, so existing packs don't change.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
🤖 Agent: reproduced #5352 with launcher 0.2.31 and Nightly d91cab6 in scratch homes; fixed launcher builds Zoom, Minimap, Recipes and X To Xray on d91cab6 (all ON in Options > Mods). pr_ready --quick green.
🤖 Agent: online_mode harness (finished from #5508): the services' key fetch stays out of the log. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: vaults' display item spins on remote clients (VaultBlockEntity.Client.tick); verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: a stack's own blocks_attacks component is read (Item.use, getItemBlockingWith); approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
The clamp moves from the item table into item_defaults::read, after the
hook, so loot and item models read the same 1 to 99 (a negative size too).
The durability tooltip no longer underflows when a mod lowers max_damage
below a saved stack's damage (getDamageValue clamps to the max).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: enderman's carried block culls back faces (ITEM_CUTOUT); approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head (pr_ready --quick green for it). Parity rerun: the failed run was only online_mode (harness bug, #5627) / dedicated_server (flake #5560).
🤖 Agent: backup before restoring level.dat_old, other-series worlds refused on every open path; approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head (pr_ready --quick green for it). Parity rerun: the failed run was only online_mode (harness bug, #5627) / dedicated_server (flake #5560).
🤖 Agent: reconnect decided by the disconnect's translation key; author's main merge leaves the diff unchanged; approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head (pr_ready --quick green for it). Parity rerun: the failed run was only online_mode (harness bug, #5627) / dedicated_server (flake #5560).
🤖 Agent: 16 port-map rows say port; both review points fixed; data only. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: content mods declare thrown entity types (snowball template); rdbtCVS found no defects; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: server tag hook fragment fixed for the stamp tool (no game code); verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: lan_e2e punches vehicles fully charged so they break; approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: horizon-mixins pin a2af5c0 (only horizon-mixins#15); approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: keybind-opened screens release the item use in the same tick (Gui.setScreen releaseAll), with the use-screen e2e scenario; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: the AppImage path test runs on Unix only (fixes the Windows launcher build); code-owner path, approved here. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: superflat structure sets (FlatLevelSource.createState) with every review point fixed (ring seed 0, sea level/min Y, null layers, no creature packs, generator min Y); verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: launcher strings through tr()/trf(), Turkish table completed; approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: verified by Ludovic's AI agent (Claude). pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others: launcher fmt, clippy -D warnings, and launcher tests (one failure in backend::tests::prepared_versions_get_what_the_asset_store_adds_without_preparing_again was a prepare-lock race; it passes alone with the account tests). The owner gate runs as that signed-in account; no offline name. I fixed the review's doc nit.
🤖 Agent: boats in bubble columns as vanilla, countdown test added by rdbtCVS; approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: opt-in shader-pack shadow cache (off by default), rdbtCVS's custom-image redraw fix read; approved; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
🤖 Agent: VRAM arena compaction, rdbtCVS's last fix (LOD stand-ins off the closing page) read; approved by rdbtCVS; verify green. Proven on a union of main f99f8feaa with 37 open PRs including this head: pr_ready --quick green for it (fmt, clippy -D warnings incl. the server build, 3077 workspace tests, 20 content-root tests, launcher clippy + 340 tests, 510 Python tests; the one failure is #5570's own new test, not in this PR). CI parity covers the harnesses.
A pack with dozens of options needs them in categories. The first page is the one the button opens,
"page:<id>" opens another, "" leaves a slot empty and columns = 1 gives full-width buttons. Every page
has to be reached from the first and every option has to be on a page, or the check refuses the pack.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
The skip compared Paths, which compare case-sensitively, so it never fired
and the extraction test failed on macOS's case-insensitive file system (the
later member, tag.json, wins there). It now checks whether a case-folded
path exists.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Each copying, hashing or compiling step spreads over that many threads; the lanes run theirs side by side, so it is no total limit.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The objects check summoned its stands on the floor, where a guest that
never moved a remote stand still matched the host. This one starts 3
blocks up and turned to 200 degrees, so where it lands and its yaw are
compared between host and guest.
Written with Claude Opus 5.5 in OpenCode.
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests). I set vanilla's forced-translucent colour (alpha 38).
The CEM expression language from OptiFine's cem_animation.txt with EMF's
documented additions (keyframe, keyframeloop, easings, splines, wrapdeg,
degdiff, catch, randomb, global constants e and nan), compiled to a flat
stack program with constants folded; if, && and || only evaluate the
branch they take. Variables are caller-resolved slots. The example
cem_expr_bench times an animation pack's frame of 20 expressions.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests). I stopped the narrator on headless exit and moved idle's doc back.
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests). I added the server-only gate for the stuck push and one shared moveTowardsClosestSpace.
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests). I made the answer carry its level event, added API markers, and the stamp tool now keeps side.
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests).
🤖 Agent: verified by Ludovic's AI agent (Claude). I read the diff against vanilla 26.3. pr_ready --quick passed on main bfc2a2798 with this PR merged with 17 others from the same sweep (fmt, clippy -D warnings, horizon-client and horizon-client-render tests, Python tests). Ctrl+Q on the Survival Inventory tab uses the item's max stack size, like every other Horizon menu (none reads a max_stack_size component yet).
A mod setting minecraft:max_stack_size to 0 made merchant.rs clamp(1, 0)
panic on the server thread. Items now keeps it to ExtraCodecs.intRange(1, 99),
and the docs say what else a malformed value does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A rebuild stays on the Horizon its native build was made on, and mods pinning an
older commit (eight registry mods pin 142388e) set the build's revision too. When
a mod needs newer hook points than that revision has (recipes v2, xtoxray v7),
and the hook mods were tested on different commits, the build was refused with
"Horizon 142388e doesn't have the native hook points these mods use", even
though the installed Nightly has them all and no Update button was offered
(#5352). The build now moves to the revision of the release or Nightly it
starts from when that one has them; pinned mods then wait for Build anyway,
as for any version mismatch. The refusal names only the mods that don't fit.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Review fix for the secrets file (#5095): when the keychain took a
secret but another process had written a newer copy to the file
meanwhile, the newer copy rightly stayed, yet the run stopped counting
it as seen in the file. Once that run stopped asking the keychain and
another process handed the copy over, the run couldn't read it.
The clean-up now says whether the file still has the entry, and the run
keeps counting it while it does.
Test: a launcher on the file renews A during a start's hand-over of A,
the keychain refuses B; after another start hands both over, the first
start reads the renewed A and B.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DG1mYGWXNxQzmhMnjB36p1
ArmorStand.interactAt ignored `DisabledSlots` (not even read or saved), so
locked stands could be emptied and filled. The stand now keeps the bitmask:
bit `filterFlag` disables a slot (an empty hand then takes from the main
hand), `+8` stops taking from it and `+16` putting into it
(`isDisabled`, `swapItem`), and the save writes it back.
A lead tied to another player was untied by anyone who clicked the mob;
only the player it is tied to unties it (`getLeashHolder() == player`).
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Main samples only the planned tiles, nearest first: the seam with the
near chunks is fine at once, but the far terrain ends in a hard edge
until the last of ~3000 tiles (Full) is in, minutes on a slow machine.
After the seam, sample a coarse backdrop first: each far tile's
ancestor two levels up (one where that reaches the seam), ~230 tiles,
about a tenth more work. Ready ancestors already stand in for missing
children; now no tile stands in within 3 chunks of the render distance,
so broad boxes never sit beside the near chunks. The horizon is covered
after under 1000 jobs instead of the last one.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
LivingEntity.dropFromLootTable supplies LAST_DAMAGE_PLAYER only when
playerKilled (lastHurtByPlayerMemoryTime > 0), so NativeLootDrops.player is
None for a mob that dies to the environment after the memory ran out
(arismoko's review). Vanilla's own loot is unchanged.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- swept_inside returns at once when no loaded block is outside minecraft:,
so a server's items and projectiles no longer sweep their blocks every tick
with no content mod (rdbtCVS's hot-path note).
- An item's step_on comes before its inside contacts, as
Entity.applyEffectsFromBlocks calls stepOn before checkInsideBlocks.
- A mob's step_on (and the magma floor, turtle egg checks with it) reads
getOnPosLegacy through block_on_legacy, so a mob over the edge of its
supporting block still steps on it (arismoko's review).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
GameRenderer.extract only renders the level when there is one, so vanilla
never draws the debug crosshair on the title screen. The F3 overlay stays
on when a world closes, and the crosshair was built without looking at
whether the frame showed a level, so its axes stayed over the panorama.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Mob equipment was an item id plus enchantments and a trim, so a picked-up
sword or armour piece lost its damage, name and components, a hand took one
item of a stack, a worn piece always compared as fresh, and death or
replacement dropped it repaired and unnamed.
`Gear.stacks` (the main hand keeps `main_hand_stack`) now holds the stack of
a slot that is more than one plain item. `set_item_stack` stores it,
`stored_stack`/`item_stack` read it back, replacing the item forgets it.
`equipItemIfPossible` takes the whole stack into a hand (one into an armour
slot) and drops the replaced stack as it was; `canReplaceCurrentItem` sees
the worn piece's damage and name; death drops the stack itself
(`Drop::Stack`), with the rolled damage when the drop is not preserved; the
save writes and reads it.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The summary charged a group with every mismatch whose command text also
appears in it, so two groups that probe the same thing (the orb groups'
`data get entity @e[type=minecraft:experience_orb,...]` lines) each
showed the other's failures. Positions are what the per-group ranges
already are; count by those.
Co-Authored-By: JustKarkat's AI agent (ZCode) <[email protected]>
Freeze flight-history interpolation after death and render death rays with
vanilla's depth writes, back-face culling and distance fog. Add a frozen-pose
regression and a vanilla oracle comparison for ray geometry and colours.
Co-Authored-By: Codex (GPT-6) <[email protected]>
The review of #5601 noted the new "experience orb fall" group never
reaches the one path of ExperienceOrb.tick that draws a random: the
push out of a block. This group pins it — a 3x3 block placed over a
settled orb leaves it colliding with only its top face free, so the
server nudges it up (Entity.moveTowardsClosestSpace) with gravity
skipped while its box is inside blocks, and it comes to rest on top of
the block. The draw's magnitude only decides how many ticks the escape
takes, so the checkpoints after it are the same either way; with the
push disabled the orb stays boxed in at y 215 and the group fails.
Corpus with both orb groups: 973/973 commands match.
Co-Authored-By: JustKarkat's AI agent (ZCode) <[email protected]>
A modifier of a type not read yet (add_spawns, add_carvers, a mod's own
serializer) or naming a biome, feature or tag the root doesn't have (another
mod's ids, #c: convention tags) dropped every content mod. Real mods ship
those beside their feature modifiers, so that modifier is now skipped with a
warning naming the mod and file, as NeoForge skips an unresolvable holder
set (reservedbytes's review). A malformed modifier of a read type (unknown
step) still stops the build. The docs say #tag in features reads vanilla's
feature tags only for now (arismoko's review).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Review fix for the secrets file (#5095): a start counted what its failed
hand-over left in the file only afterwards, so a hand-over by another
process in between left it reporting the keychain away for those
accounts. The start now counts the file's keys first, and drops each
one it hands over itself.
Test: another process hands the account over right as this start's
hand-over fails; the start still reads it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DG1mYGWXNxQzmhMnjB36p1
The whole matrix (7 mounts x 2 levels x 2 players + in_nether) takes about 10
minutes, and the nightly LAN suite already runs ~2,400 s of its 3,000 s
timeout (rdbtCVS's review). The default keeps the regression signal;
LAN_E2E_MOUNTS=all and LAN_E2E_LEVELS=all run everything.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- Entity.commonTick sets the old position before the interpolation step, so
the frame lerps from the pre-step position: previous keeps it now, where
ExperienceOrb::tick had overwritten it and the orb still hopped (rdbtCVS).
- The knot is an entity, so a click counts as one on it only within the
entity interaction range (filterHitResult), not the fence's block range
(Joansitoh).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
rdbtCVS's review points 1-3:
- resolve returns None for a path with a .., . or empty segment, so a pack
can never name a file outside the overlay root (vanilla refuses them).
- glob backtracks only to the last star: O(pattern x name), where every
star tried every split before (a player-chosen item name, per frame).
- properties_files reads each real folder once and at most 16 deep, so a
folder pack linking to its own parent no longer recurses to PATH_MAX.
Tests for each.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
winit's MouseButton::Other(n) is the raw X11/evdev number on Linux, not GLFW's
index, so the hook docs and the fragment say a button past 4 is
platform-defined (rdbtCVS's review).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The player-file save in `save_game_metadata` only ran for a dedicated-mode
integrated server (the fixture), so a real `--dedicated` server (the
`DedicatedServer` role) skipped it on autosave and `save-all`: players were
only written when they left or the server stopped. It now runs for every
server role, as PlayerList.saveAll does.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Identifier.bySeparator reads ":clear_all_effects" as minecraft:clear_all_effects
(Consumable.CODEC decodes both), so it is vanilla's clear, not a mod kind for
the native hook (arismoko's review). Test added.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A chest boat or chest/hopper minecart open in a menu keeps its slots in a
stand-in under a placeholder key. Removing the vehicle (broken, killed,
exploded, a rider's death, a portal) dropped the vehicle's stale items and
left the stand-in with a full copy a stale menu could still take from.
`release_vehicle_container` writes the menu's live slots back to the vehicle
first (so they are what drops or travels), then removes the stand-in, before
each removal.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
DirectionalPlaceContext.getRotation is get2DDataValue() * 90, so a vertical
facing gives -90, not toYRot's 270. Both name East for vanilla's blocks
(from_y_rot, rotation segments), but the placement hook shows the raw value
(arismoko's review on #5516). Armor stands and boats keep toYRot.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- native_dispense_item answers (remainder, event): 1000 click, 1001 fail
(OptionalDispenseItemBehavior without success), 0 nothing (NOOP), so a
mod can make a silent or failing behaviour.
- native_dispense_item and native_dispense_default carry the
API @API:dispenser_behaviour@ marker; the fragment lists dispenser.rs.
- stamp_native_hooks.py copies a hook's neoforge, side and cancellable keys
into native-hooks.json instead of dropping them.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A menu was only validated once per tick, after the packet phase, so a click
could still land on a vehicle that was gone, a mount that died or left reach,
a lectern out of reach or a station block that was replaced. The click
packets (container click, button, slot state, rename, beacon, trade select,
bundle select, place recipe) now run one shared check first: the open menu,
whichever kind, must still be valid, else it closes like ContainerClose and
the click lands on nothing. The per-tick checks use the same predicates.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Leaving the server (a disconnect, a LAN close, a dedicated stop) kept the
guest's cursor stack, trade payment, grid and station inputs in the menu, a
trade unwritten on the villager and a vehicle's stand-in in the level.
`guest_close_all` runs the same close code as ContainerClose for every kind
of menu, in the guest's own level, and runs from `HostEvent::Left` and
`save_leaving_guests` (LAN close, dedicated stop, integrated stop) before the
playerdata is written.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Review fix for the secrets file (#5095): a launcher that started while
the keychain was away, on a file a previous run left, asked the
keychain again only for secrets it had written itself. Once another
process handed the file over, it reported the keychain away for
accounts the keychain now held.
It now also counts what it read from the file and what its own start
left there. Test: two launchers started on an old file while the
keychain is away (one read the account, one didn't) both read it after
a third hands it over.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DG1mYGWXNxQzmhMnjB36p1
native_is_valid_bonemeal_target answers Option<bool>: Some(false) refuses
the meal even where the block's template takes it (BoneMealItem.growCrop
returns false when isValidBonemealTarget is false), None leaves the
template's BonemealableBlock (arismoko's review). The client prediction
falls through to growWaterPlant on a refusal, as vanilla's useOn does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Each opener closed a different subset of the menus the guest could have
open (the trading screen left a mount or crafting state behind, a vehicle's
stand-in stayed, the old window id still took clicks). They now all start
with `guest_close_all`, as `ServerPlayer.openMenu` closes the open menu:
containers, vehicles, mounts, merchants, stations, lecterns, the crafting
table and native menus.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
position_riders ran before the player moved and before previous_position
was saved, so a name hologram on the local player was drawn a full tick
behind it. Vanilla ticks the player (tickNonPassenger) before its
passengers (positionRider), so seat_own_stands now runs after the move,
from the previous_position just saved to the moved position.
The snap test now walks toward x 5 when the far sync comes and checks no
step back toward it is left, so it fails if a snap stops clearing the
handler. The ledger citations move off the first doc lines they were
read as ports from.
Written with Claude Opus 5.5 in OpenCode.
Leaving the server (a disconnect, a LAN close, a dedicated stop) kept the
guest's cursor stack, trade payment, grid and station inputs in the menu, a
trade unwritten on the villager and a vehicle's stand-in in the level.
`guest_close_all` runs the same close code as ContainerClose for every kind
of menu, in the guest's own level, and runs from `HostEvent::Left` and
`save_leaving_guests` (LAN close, dedicated stop, integrated stop) before the
playerdata is written.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
🤖 Agent: proven on 8e5ab9ec6: pr_ready --quick green (fmt, clippy -D warnings, 1577 horizon-client tests, 510 Python tests); playhorizon A/B 3 alternated rounds 20 bots: streaming span 4.82 → 4.64 ms median (lower in every round), MSPT within noise. The 34 affected harnesses run in CI (run-parity).
Leaving the server (a disconnect, a LAN close, a dedicated stop) kept the
guest's cursor stack, trade payment, grid and station inputs in the menu, a
trade unwritten on the villager and a vehicle's stand-in in the level.
`guest_close_all` runs the same close code as ContainerClose for every kind
of menu, in the guest's own level, and runs from `HostEvent::Left` and
`save_leaving_guests` (LAN close, dedicated stop, integrated stop) before the
playerdata is written.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
PistonHeadBlock.neighborChanged calls state.canSurvive; the direct
piston_head_can_survive call skipped native_can_survive (arismoko's review).
can_survive keeps the piston rule as its default.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: native_explode: server-only wrapper over server_explode with the TNT/anchor drop-decay rules, guarded on dimension, loaded chunk and radius; verify and parity green on this head.
ExperienceOrb.tick gates the stuck push with !level().isClientSide(): a
server client (remote.rs) no longer nudges stuck orbs and draws from its own
random between server updates. Items and orbs share one
towards_closest_space port instead of two copies.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
stamp_matches compared a missing outputs key with the digest of a missing output as None == None. The docs and comments said HORIZON_PREPARE_JOBS is how many things run at once; it is how many threads copy and hash, the lanes all run at once unless it is 1, and it counts logical processors. With the split there are five lanes. java.rs's memory figure now says what it compared.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A held two-sided model carries a quad per direction (a calibrated sculk
sensor's tendrils and amethyst planes, big dripleaf, block/cross), and the
item render types vanilla picks in ItemBlockRenderTypes -- solidBlockSheet
(entitySolid), cutoutBlockSheet (entityCutout), translucentItemSheet
(entityTranslucentCull) -- all keep the composite state's CULL. Horizon
culled only quads whose sprite has translucent pixels, so held blocks drew
both sides: the coplanar pairs striped (#4856, #4855) and a hay block's far
faces bled through the near ones (#4801).
Two-sidedness in vanilla is geometry, not culling: ItemModelGenerator's
bakeExtrudedSprite bakes a South and a North quad per layer, which
horizon-assets already mirrors.
Co-Authored-By: JustKarkat's AI agent (ZCode) <[email protected]>
The namespaced-id check is Identifier's character rules, five lines; the
dependency pulled horizon-world (and horizon-core, postcard...) into
horizon-mods, the mod registry and the launcher builds.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The owner loads the file level.dat names, but when it left (save_guest,
PlayerList.remove) only its own file was written; the named one waited for
the world's save that follows (stop_integrated: free_player_beds, the
chunks, then save_game_metadata). A game stopped or failing in between
left the named file older than the own one, and the owner then loaded the
older state. `playerdata::save_owner` writes the named file
(`Level::write_player`, also the player half of `Level::write`) before the
own one, so the file the owner loads is never the older.
`playerdata::load` (PlayerDataStorage.load) now falls back to the player's
`.dat_old`, as vanilla does: `write_dat` moves the old file aside before it
renames the new one in, so a save cut off there leaves only the
`.dat_old`, which the named file's read already used.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
PlayerList.loadPlayerData gives the single-player owner the file level.dat
names (`singleplayer_uuid`), not their own. Since #5069 the integrated
server saves the owner to both files, `players/data/<singleplayer_uuid>.dat`
(offline `Player`) and `<owner uuid>.dat`, but loaded the owner's own file
first. A build that saves the owner to the named file alone (0.2.93 and the
stable releases before #5069) leaves the own file behind, so after
switching back the owner got an older inventory and position, and the next
save wrote them over the newer ones.
`Level::read` now says whether its player came from the named file, and
`playerdata::load_owner` applies it before the owner's own file; the own
file stays the fallback when the named file is missing.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Both the vanilla and the Horizon server fetch getPublicKeys in the background, at any moment of a
scenario, so ('unknown', '/publickeys') turned up in four checks (parity run 41243825430 on #5508).
It now stays out of the log, like a bare GET /.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Review fixes for the secrets file (#5095):
- A start's hand-over wrote a snapshot of the file to the keychain, so
a token another process signed out or renewed meanwhile could come
back or be overwritten. Every keychain write now holds
account-secrets.keychain.lock, and the hand-over reads each entry
again under it. The file's own lock still never waits on a prompt.
- MCP secrets_file is true only while the file holds a sign-in: not
for sign-out marks alone, nor for a file that can't be read.
- Tests: a hand-over held on a prompt while a running launcher signs
out or renews the same account; what counts as a sign-in in the file.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DG1mYGWXNxQzmhMnjB36p1
Parses OptiFine's custom entity model files: models, part, attach, scale,
baseId, .jpm part files, invertAxis, translate, rotate, mirrorTexture,
textureSize, texture, boxes (textureOffset box UVs or per-face uvNorth..
with their aliases, sizeAdd, sizesAdd, EMF's sizeAddX/Y/Z), submodel(s),
attachments and the animation expressions as text. Boxes bake into the
quads ModelPart draws, on vanilla's axes around the vanilla part's pivot.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The repeat run's file check skipped horizon-resources.json, so a manifest the repeat rewrote differently passed. The comments that said at most three JVMs run at once now say four with the second server JVM.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A Horizon pack declares options, but they stayed at their defaults: the player had no way to set them.
The picked pack now gets the same Shader Options... button as a shaders/ pack, and its values live in
shaderpacks/<pack>.txt, the file that kind already uses. A saved value counts as a change to the pack,
so it reloads with it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
player_entity left on_fire false, so a criterion's player predicate with
flags.is_on_fire true never matched, a mod's trigger (the reviewer's
case) or a vanilla one. It is Entity.isOnFire on a server now:
remainingFireTicks above 0 (players aren't fire immune).
Written with Claude Opus 5.5 in OpenCode.
Review fix for the secrets file (#5095): Layered asked the keychain
while holding account-secrets.lock, so an unlock prompt held up every
other launcher and CLI process.
- Each call reads its file entry under the lock, asks the keychain
with the lock released, then removes the entry only if it still
holds what was read: another process's newer write stays. The start's
hand-over works the same way, from a snapshot.
- Once the keychain refuses or doesn't answer (a dismissed prompt),
the run stops asking it and parks new secrets in the file, until it
finds that another process handed the keychain one it parked.
- Tests: a keychain waiting on a prompt while another process reads and
writes the file, a dismissed prompt asked once a run, and a half
refused hand-over followed by a sign-out.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DG1mYGWXNxQzmhMnjB36p1
🤖 Agent: fixes#5572. ground never sends back and a dropped claim never feeds the floating kick. anticheat_e2e ground (allow-flight off) passes; vanilla 26.3 client soak 25 min never flagged; vanilla client at the live-flagged spots on the server's seed 42 min / 7 spots, 0 flags (18 min of it on this branch, ground acting). pr_ready --quick all green on main 74698584e.
player_model_shown() no longer blocks spectators, so their model
geometry generates in third person. The render path uses render_with
so the poses' visibility hides body parts, leaving only the head/hat.
body_color returns ~12.5% alpha for spectator players; OIT marks the
body already, so the ghost head looks ghostly.
Co-Authored-By: Hermes Agent <combo1> <[email protected]>
Clarify that only physical_keys is cleared directly in heal_hold_latches, leaving logical sneak/sprint keys to the toggle loop.
Co-Authored-By: Antigravity <[email protected]>
When ground acted, a false claim sent the player back to where it was:
into the air it claimed to stand in. A player there claims it again, is
sent back again (+0.5 VL twice a second at the same spot) and hovers until
vanilla's "floating too long" kick: in 11 minutes on the official server,
four players got it 20 times. Only the anticheat's own kick (logged, told
to the operators, counted toward the ban) should end a false claim.
ground's correction is now the dropped claim alone (vanilla's fall damage
follows), logged as "claim dropped"; it never sends the player back, and a
move whose claim was dropped never counts toward clientIsFloating.
setback = 0 still makes the check log only. The ground e2e runs with
allow-flight off and fails on a setback or a floating kick.
Fixes#5572
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A pack treats hundreds of kinds of surface differently, where [[tags]] hold 32 sets. Each sprite takes
the id of the first [[materials]] entry that names it, as MaterialInput.material, through the same
generated switch as the tags. [shadow] solid_materials draws those ids into the map without their
cutouts.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
Review fixes for the secrets file (#5095):
- The Tailscale access token administers the player's whole tailnet:
it goes through account::keychain(), the keychain or memory for the
run, never the secrets file.
- MCP account_list says secrets_file: true while sign-ins sit in the
file, beside keychain (sign-ins outlast the launcher).
- launcher/README.md and the sign-out help (docs/launcher-commands.md,
regenerated) no longer say tokens are never in a file.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DG1mYGWXNxQzmhMnjB36p1
self.keys tracks the logical gameplay keys, and sync_toggle(0) uses KeyCode::ShiftLeft to represent active sneak regardless of the physical key binding. Calling self.keys.remove(&key) when physical Shift is up cleared active sneak every frame for players who rebound sneak away from Shift (e.g. to Left Control or C).
Fixes#5608
Co-Authored-By: Antigravity <[email protected]>
Review fixes for the secrets file (#5095):
- No store is picked once per run any more: on Linux and FreeBSD every
read takes the file's copy first and the keychain's otherwise, and
every write goes to the keychain, to the file only when it refuses.
A launcher that wrote to the file keeps reading, renewing and signing
out after another process hands the file to the keychain.
- Each start hands the keychain the file's secrets one at a time, each
leaving the file once the keychain took it, so a move it half refuses
never leaves a secret in both places, and signing out removes it.
- Signing out while the keychain is away leaves an empty entry that
hides the keychain's copy until it can be deleted.
- macOS and Windows keep the keychain or memory, as before.
- The concurrency test now runs real child processes.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DG1mYGWXNxQzmhMnjB36p1
RenderPipelines.TEXT_POLYGON_OFFSET and TEXT_SEE_THROUGH (and the
TEXT_SNIPPET they build on) never call withCull, so
RenderPipeline.Builder.build keeps its default, cull.orElse(true). Horizon
drew both with cull_mode None, so a text display showed its mirrored
text from behind (#3833).
Culling alone would have hidden the readable side instead: the text
display quads were taken in the wrong corner order and wound clockwise
toward the reader. They now use Font's glyph order (top left, bottom
left, bottom right, top right), the one name tags already use, through
text_display::CORNERS. A test projects fixed and billboard displays
through the real view and projection matrices and checks that the side
the text reads from is the counter-clockwise one.
The improved-transparency text pipeline is culled the same way.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: players can /msg, /tell and /w a player by their nickname (shown as ~name); server guests included. SMP proof and pr_ready --quick passed on this head (comment above).
A finished column at the edge of the generated area is lit against its
decorated proto neighbours, each copied into a column on the server
thread (light_view). The light engine reads only their blocks, yet the
copy also rebuilt every section's biomes (64 biome lookups and a
palette per section), the scheduled ticks, post-processing marks and
structure references. The view now copies the blocks alone; the final
column (finish) still gets everything.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Review fixes for the secrets file (#5095):
- Every read-modify-replace holds account-secrets.lock (std File::lock),
so launcher and CLI processes signing in or out at once no longer
overwrite each other; each write uses a temporary file of its own.
- The file no longer sticks once made: at the first start the keychain
answers, its secrets move into the keychain and the file is removed,
so an account kept in the keychain before a start it missed comes
back beside the ones signed in meanwhile. If the keychain refuses any,
the file stays whole and in use.
- A broken file's error names only the error kind, line and column:
serde's own message can quote a string the file holds.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DG1mYGWXNxQzmhMnjB36p1
LivingEntity.applyItemBlocking passes `damageBlocked >= damage` to
blockUsingItem with the damage Player.hurtServer already scaled by the
difficulty; Horizon compared against the raw damage, so on Easy a full
block counted as partial and blockedByItem's knockback pulled the
defender toward the mob. The hurt now records whether it was fully
blocked and the shield reads that.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Entity.setRemoved makes a removed vehicle's passengers stopRiding, and
LivingEntity.dismountVehicle puts them at the higher of their own and
the vehicle's heights, at the free spot nearest there (findFreePosition).
break_cushion only cleared the riding id, and a server-thread client kept
riding a cushion that was gone, so the rider stayed at the seat, 0.35
below the cushion, and dropped into the block under it.
Both the server (break_cushion) and the client (the cushion's removal)
now dismount the rider that way, with no get-up sound
(Cushion.removePassenger plays none once the cushion is removed).
Fixes#4571
Co-Authored-By: Claude Opus 5.5 <[email protected]>
native_data_pack_files asks a native mod, while the server assembles its data
tables, for the pack files it contributes: (path, bytes) in the overlay's form.
It fires once per load — when a world opens, and again on every `/reload` — so
the ask is the mod's chance to recompute them from its config or from the
world. A mod that turns configuration into recipes, tags or loot needs no
content mod beside it and no data pack installed by hand: today neither can
recompute those files for a load.
The files join the load where the world's packs already do:
Game::build_world_tables is both a world's first open and reload_world_tables
(`/reload`, `/datapack`), and the hook is asked there and nowhere else. They
sit below those packs: PackOverlay::top is a path's last file, so a player's
data pack still overrides a mod's for the same path, and a later contribution
overrides an earlier one, as packs layer. The advancement prefetch takes the
same files, so a contributed advancement is there on the first open and not
only after a reload.
With no mod hooking it nothing is contributed: a world with no packs still
uses Resources.rest, and a world with packs builds exactly the tables it did
before.
The hook is API next (tools/native-hook-fragments/data_pack_files.json);
native-hooks.json and NATIVE_HOOKS_API are untouched.
Co-Authored-By: DeepSeek <[email protected]>
The local player's model isn't drawn in first person, so a native pack's shadow pass gets it on its
own through the frame's shadow_only quads. The round blob under entities is left out while a pack
draws a shadow map, since the sun's shadow is already there.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
Vanilla's first ServerEntity.sendDirtyEntityData for a player runs before
its first ServerPlayer.tick, whose updatePlayerAttributes (and doTick's
removeFrost) create the interaction range and movement speed instances.
Those go out on the next tick, after the first chunk batch. Horizon's guest
tick created them before sync_player_state sent them, so a vanilla client
got a default-valued update_attributes before the first chunk batch.
Fixes#5560
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A surface wgpu never configured has no error sink, so acquiring from it
panics instead of reporting Validation (#5353). recreate_surface now checks
configure through error scopes and a device-lost flag, and frames configure
such a surface again before acquiring, skipping until it takes.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
EntityHitboxDebugRenderer.emitGizmos draws every visible entity's box,
the Interaction too (its renderer is a NoopRenderer, the hitbox is not
the renderer's). Its box is EntityDimensions.scalable(width, height):
eyes at 0.85 of the height, not a LivingEntity.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Native mod builds only showed in the version picker, so there was no way to delete one by
hand. The Versions page now lists them above the releases, with what each was built on, its
size on disk, and the same Use and Remove (asked twice) as a release. Removing a version a
program still runs from is refused.
Fixes#5326
Co-Authored-By: Claude Opus 5.5 <[email protected]>
LoadingOverlay.extractRenderState blits textures/gui/title/mojangstudios.png
as two halves side by side over BRAND_BACKGROUND and draws its bordered bar
at 83.25% of the height; Horizon only filled the background. The logo comes
from the player's own prepared resources (the GUI atlas already packs it).
Fixes#5115
Co-Authored-By: Claude Opus 5.5 <[email protected]>
remote_open_merchant left the menu title unset, so NativeScreenView.title was
None on that window although the docs promise the server's title.
Co-Authored-By: Claude Code <[email protected]>
Adds screen_key_pressed_pre and screen_char_typed_pre (ScreenEvent.KeyPressed.Pre
and CharacterTyped.Pre) on the open screen, fired from the window's key event,
the control socket's key and type commands, and the input method's commit.
Co-Authored-By: DeepSeek V4.1 Flash <[email protected]>
Co-Authored-By: Claude Code <[email protected]>
The server simulated its players' bodies with no world border (only the
window's own player got one, in player_border), so a server player's
re-simulated movement went through the border's walls. enter_guest now
gives the guest's body the level's border each turn, as vanilla's
Entity.collide collides every player with the level's WorldBorder.
Fixes#2743
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Many Linux systems run no Secret Service (GNOME Keyring, KWallet), and
the launcher then kept sign-ins in memory only: players signed in again
at every start (#5095).
Where the keychain doesn't answer, the secrets now go to
account-secrets.json in the launcher's data folder, readable by the
player only (0600 on Unix) and replaced whole through a temporary file.
The keychain stays first; a secrets file from an earlier start stays in
use so a flaky keychain doesn't sign players in and out, and signing the
last account out removes it. HORIZON_NO_KEYCHAIN still keeps secrets in
memory for the launcher's tests.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DG1mYGWXNxQzmhMnjB36p1
The renderer's opaque entity quads now go into shadow_map through the shadow camera, with their
texture's cutout. They draw after the translucent map is copied, since where one casts the solid
map is the darker and decides. Unlit overlays, as entity.wgsl lists them, cast nothing.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
#5559 let the dialog registry carry data for a client that knows the
core pack; with this PR the server follows vanilla's packRegistry again,
so every registry, the dialogs too, goes without data to that client.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The classic path drew weather before translucent terrain with depth writes
on, so the water behind the rain failed the depth test and the rain columns
cut holes in it. Vanilla 26.3's RenderPipelines.WEATHER is
DepthStencilState(GREATER_THAN_OR_EQUAL, false), and executeClassicTransparency
draws translucent terrain, clouds, then weather, then the world border.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Keeps this pull request near the size limit; the block state stays in the
hooks, whose signatures can't change once merged.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
`native_screen_mouse_scrolled_pre` (`ScreenEvent.MouseScrolled.Pre`) hands a
mod the open screen, the wheel's movement in x and y as the game's scroll
handler already scales it (vanilla's doubles, the sign only when discrete
scrolling is on, a touchpad's pixels over 30) and GLFW's modifier bits, plus
`canceled`.
It fires in the `MouseWheel` arm once `scaled` is defined, before the dialog,
title, hotbar, bundle and every screen scroll, so a cancelled turn scrolls
nothing and leaves nothing half-applied; the hotbar reads the wheel only
while playing, which the hook's gate excludes. The pad's and the keyboard
focus's made-up wheel is not the mouse and fires nothing.
The hook doc names the screens that never fire, as the mouse button hooks
do, and the fragment is pending, so this is "API next".
Co-Authored-By: DeepSeek <[email protected]>
tick_remote_spawners ticked spawners and trial spawners on a remote client
but not vaults, so VaultBlockEntity.Client.tick (display item spin, idle
particles and sounds) never ran on the server-thread client.
Fixes#3108
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Both hooks get the block's state, as a renderer reads
blockEntity.getBlockState(); native_block_state_property reads its facing
without a world lookup, and native_draw_item draws an item stack under the
pose stack as a campfire does (resolve_item, submit_item_world).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
`horizon-launcher play --account WHO` plays this launch as a signed-in account (name or UUID)
without changing the one `account select` picked, so several games of one instance can play
as several accounts at the same time (#4952). The owner gate runs as for any launch, and
what the services renew (tokens, names, skins) is kept.
account::session_for_game_as runs the gate as that account on the loaded accounts and puts
their choice of player back before saving them (launch_as). A Java instance refuses the
flag, as it does --world.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
docs/native-mods.md gets an "Items' default components" section beside the
render hook points: the target path, both sides, threads, vanilla clients,
creative search. The catalog test reads free-function hooks with one helper.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- A failed exporter compile fails every lane at once instead of being
retried by each lane that waits for it.
- The gameplay facts exporter streams to a .partial file and moves it into
place: a run stopped halfway no longer leaves a short file the next run
would keep.
- pool::map takes no per-thread state (a jar clone costs two Arc counts);
pool::lanes and map are tested at 1 job and several. With 1 job the first
failing lane stops the rest, as documented now.
- The lanes' log and stage messages go through relay(), which a test now
checks; block behaviour reuses export_on; the stamp inputs are worked out
once; the fifth lane exists only when it splits; dead checks are gone.
- check_prepare_exporters.py also prepares with 2 jobs (the default path
below 6 cores, 5 JVMs) and repeats that run, which must start no JVM and
change no file.
- The docs say what HORIZON_PREPARE_JOBS does, that the split is not yet
timed on 6 cores, and the split's measured ~400 MB.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
native_block_entity_render_bounding_box lets a mod widen the box its block
entity is culled with (NeoForge's getRenderBoundingBox), or make it
infinite, so a tall part stays drawn while its block is out of view. The
view distance check is strict, as Vec3.closerThan.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: pr_ready on the Mac mini (head 26748d7c1, base a4873dec4): fmt, clippy -D warnings, horizon-command tests 27/27 (new a_dedicated_server_keeps_seed_and_version_for_game_masters), Python 510/510, public tree ok; parity 41/44 affected harnesses pass incl. command_parse, command_exec, chat_screen, server_ops_feedback. The 2 failures predate this PR and don't involve commands: online_mode = this Mac's known IndexError (stray GET /, fails on main too), dedicated_server = config-phase minecraft:dialog registry data from #4698 (same root cause as #5559), the stage before any command. CI verify passed.
On a server where a mod set nicknames (native_set_nickname), a player
argument naming no player matches the player whose nickname it is (any
case, without the ~ chat shows, or with it quoted), so /msg, /tell and /w
reach a nicknamed player by the name chat shows. Real names win; no
nickname, no change.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The other way round from a force-loaded chunk: the player and a horse in the
Nether while the other player keeps the Overworld ticking; the Overworld's pass
must leave the menu alone (in_nether, --server-thread). The player lands in
creative and is put back on the platform once it stands, so a cave below the
spot doesn't drop it out of reach.
A killed mount lies at the spot for 20 ticks (LivingEntity.tickDeath), and the
next one, summoned there 10 ticks later, took the clicks meant for it on a
slow run: wait until the client no longer sees it. LAN_E2E_LEVELS takes a
comma list (the_nether, the_end, in_nether), as LAN_E2E_MOUNTS does; the
force-load is removed even when the wait after adding it fails.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
ExperienceOrb.tick diverged from vanilla 26.3 in the way the dragon's
death XP shows: a landing orb kept its pre-collision speed through the
friction step and bounced back at 0.9 of it on every ground contact, so
the orbs rained onto the exit portal hopped high and for a long time —
the report reads them as floating in mid-air and spreading out.
The tick now follows the vanilla method: gravity 0.03 is skipped while
the orb's box is inside blocks (`!level.noCollision`), a stuck orb with
no player within 8 blocks gets Entity.moveTowardsClosestSpace (already
ported for items, now on the orb too, its own random drawing the
speed), the collided axes' velocities are zeroed the way
Entity.move's restituteMovementAfterCollisions does for a bounciness 0
orb (x and z by Mth.equal's 1.0E-5, y exactly), one friction of
getAirDrag() times the ground's friction — computed as a float, as
vanilla does — scales every axis including y, and the landing bounce
is `-fallSpeed * 0.4` gated on `fallSpeed < -getGravity()` instead of
a fixed -0.9. The friction anchor below the orb is getOnPos(0.999999)
as the item's already was, not the block one below.
tools/test_command_exec_parity.py gains an "experience orb fall"
group: an orb falling 45 blocks onto a stone platform, out of the 8
block follow radius, matches vanilla's Pos and Motion to the digit at
every checkpoint, the settled state included (969/969 with it).
Fixes#4781.
Co-Authored-By: JustKarkat's AI agent (ZCode) <[email protected]>
MineshaftStructure and StrongholdStructure call moveBelowSeaLevel with
ChunkGenerator.getMinY(), which is 0 for FlatLevelSource (the noise settings'
minimum Y otherwise), not the level's -64. The Classic Flat stronghold sat
64 blocks lower than vanilla's.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
`native_screen_mouse_button_pressed_pre` and
`native_screen_mouse_button_released_pre` (`ScreenEvent.MouseButtonPressed.Pre`
and `.MouseButtonReleased.Pre`) hand a mod the open screen, the button's GLFW
index (0 left, 1 right, 2 middle, 3 back, 4 forward, `Other(n)` = n) and
GLFW's modifier bits (shift 1, control 2, alt 4, super 8), plus `canceled`
for a press the screen never sees.
They fire first in `handle_window_event`'s closure, before the screen or the
container hotkey, sort and header buttons act, so a cancelled press leaves
nothing half-applied and a cancelled release keeps a drag held. The pad's and
the keyboard focus's made-up clicks are not the mouse and fire nothing.
The gate and the view two more input hook families need are here: the pure
`glfw_modifiers` and `glfw_button`, `native_screen_open` (a screen of the
game's over `Screen::Playing`, no connect or load) and `screen_view_now`,
which builds only the `HudBatch` its rect needs. The hook docs and the
fragment name the screens that never fire: the title screen, the connect and
loading screens, the prompts, books and sign editors drawn over gameplay, and
any screen while a world loads or saves.
The fragment is pending, so this is "API next".
Co-Authored-By: DeepSeek <[email protected]>
New crate horizon-optifine, the layer every OptiFine-format feature (CTM,
CIT, random and emissive entities, CEM, sky, colours) reads its files
through: Java .properties with escapes and continuations, name matchers
(pattern:, ipattern:, regex:, iregex:), range lists with parenthesised
negatives, id lists with ! exclusions, block lists with states, OptiFine
path resolution (~/, ./, namespace:) and the optifine/ and mcpatcher/
folders of every namespace. Written from OptiFine's public format docs.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
CreativeModeInventoryScreen.slotClicked drops one item (the stack up to its
max size with Control) from a slot of the Survival Inventory tab whatever the
hand carries; the menu's THROW needs an empty hand, so with a carried stack Q
did nothing there. The screen's carrying branch now covers that tab as well as
an item tab's hotbar, and sets the slot as the menu path does (equipment
changes included).
lan_e2e creative_throw also checks Control+Q with an empty hand on an item
tab's hotbar, Q and Control+Q over a display slot, and Q, Control+Q and Q
while carrying on the Survival Inventory tab.
Part of #3684.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A dropped item can be picked up again after 40 ticks; on a slow runner a
pickup between steps would change the counts the check compares.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
On a creative item tab, Q over a hotbar slot took the stack out through the
menu's THROW, but on a client of a server (single player's server thread, LAN,
dedicated) the drop only went to the server on the Survival Inventory tab, so
the slot was synced empty and nothing dropped. Send it on every creative tab,
as AbstractContainerMenu.doClick calls Player.handleCreativeModeItemDrop.
Also port CreativeModeInventoryScreen.slotClicked's own branch for a THROW on
an item tab's hotbar slot while a stack is carried: one item (the whole stack
with Control) drops from the slot and the carried stack stays.
lan_e2e: new check creative_throw, on the host and the guest.
Part of #3684.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A mod could not change a default component of an item it didn't add (a
stack size, a food, a tool rule), as NeoForge's ModifyDefaultComponentsEvent
lets it. native_modify_default_components hands a mod each item's defaults
in items.json's form as they are read, to set or remove components.
Every reader of items.json (the item table, loot, enchanting, item models,
banners, pot sherds, rarity in chat) now goes through
horizon_world::item_defaults::read, so a change reaches all of them, and
stacks are saved and sent against the changed defaults like vanilla's.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The depth range now reaches past the build height below the camera and shadow_position clamps z,
so a point past the far end stays lit, the blank map included. The light's box culls the native
lists instead of a section radius that lost casters at low sun. The native camera's snap is
euclidean, the resolution is clamped once, and the distortion and the camera each live in one
place. The generated reference lists the shadow helpers.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
The first cut told an entry's kind before its requirements were checked, so a
failed requirement still fired, and gave the hook no way to know which mob was
hit. The kind is now told after the check, with the mob's id. A bare type name
is minecraft's own, so ignite runs as minecraft:ignite instead of reaching the
hook. The docs now say what's told: attacker on victim, a player's weapon on a
mob.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A fresh preparation on a 6-core laptop finished its other JVM lanes by ~17 s,
then ran the server lane alone for ~20 s more: its last export, the world
generator's block behaviour (11.5 s there), sat behind everything else on
that JVM. From 6 cores (pool::jobs, so HORIZON_PREPARE_JOBS too) it and the
structure blocks run on a fifth lane, in their own server JVM. It pays a
second Bootstrap and ~400 MB of peak memory, so it stays off below 6 cores:
forced on this 3-core box it took 44.9 s against 40.5 s.
Same output (digest unchanged either way). check_prepare_exporters.py now
runs its batched preparation that way and expects 6 JVM launches; it passes
(23,204 files).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Measured on the pinned 26.3 jars, sampling every process's RSS: the peak came
at the end, from the server worker alone (1.25 GB). Its live heap stayed under
~250 MB; the rest was garbage the serial collector promoted and kept, mostly
from two exporters:
- The collision exporter wrote, and the worker cached, 62 MB of verification
inputs (printf'd, ~4 s) that only the parity harness reads. Inside the
prepare worker it now skips them; standalone runs still write them.
- The gameplay facts exporter built its 56 MB file as a Gson tree, then a
String, then bytes. It now streams each state to the file; the bytes are
the same (re-serializing the parsed file with Gson.toJson gives them back).
The worker now peaks at ~480 MB. In the Rust process, every thread reading a
jar parsed its member list again (~10 MB each, eight at once); they now
share one parse through clones of the archive that read at their own offsets.
Its heap peak halved (96 to 51 MB).
check_prepare_exporters.py passes (23,054 files, 5 JVM launches); the output
digest is unchanged. "Movement and attributes" takes 12-13 s instead of 16-18 s.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
An enchantment's effects entry that names an effect.type in a namespace
outside minecraft: used to be silently skipped at the firing point. It
now reaches native_enchant_effect(enchantment, kind, data, level) — the
post_attack components first, where the weapon's holder is the acting
player — and the game's own damage, sounds, effects and drops are open
to the hook. An unclaimed kind does nothing.
Co-Authored-By: ZCode AI agent (GLM-5.3)
The serial collector is as fast as the parallel one for these JVMs (fresh run
63-66 s either way, pinned 26.3 jars) and keeps the three that run at once to
1.25 GB together instead of 2 GB. An export's time no longer includes waiting
for javac. A compile error still names the export it stopped.
tools/check_prepare_exporters.py passes: 23054 resource files match between two
isolated runs and a batched one, with 5 JVM launches and no compiler.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The exporters, the data generator and javac are short batch JVMs, three of them
running at once. The C2 compiler's background work and G1's concurrent threads
cost them more CPU than those win back. With -XX:TieredStopAtLevel=1
-XX:+UseParallelGC (javac gets them through -J), a fresh preparation from the
pinned 26.3 jars took 67 s instead of 102-105 s on the same machine. The output
digest (tools/prepare_output.py) is unchanged. java.rs is not part of the
exporter bundle key, so cached bundles stay valid.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Fresh preparation ran everything one after another. The steps now run on four
lanes: the server exporter JVM (same order as before, since its exporters share
state), the client exporter JVM then the command registries', the data generator
then the creative inventory's, and the Rust-only copying. At most three JVMs run
at once, still five JVMs in total. A worker stops when its lane is done. The two
exporter sides compile at once. Asset extraction, structure templates, worldgen
bundles, server library unpacking, client library imports, inventories and pin
checks run on several threads. Every pinned file is still hashed in full.
Repeat runs always started the server JVM for the block behaviour and
structure-block exporters, which had no skip check. They now keep a stamp.json
in .local (exporter bundle key, server jar, preparer sources, output sha256), so
an unchanged repeat run starts no JVM. The prepared files are unchanged.
Every step logs how long it took. HORIZON_PREPARE_JOBS=1 runs it all serially.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The check now opens a tamed horse, camel, llama, trader llama, nautilus
(owned), chested donkey and mule by shift + right click and by the
inventory key while riding, and again by shift + right click after getting
off, on the host and on the guest, with a chunk force-loaded in the Nether
and in the End. Every way is reported on its own and must still be open 20
ticks later.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01RodDbPWCByh96WKcbiEQpB
The rect comes from the same container image the game draws, so the
merchant and the beacon report their wider, shifted window. An unknown menu
type is None rather than an empty id, and slot_count is checked against
slots() in debug builds.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A server names no player of its own as a mob's killer, and the drops
context's fields aren't flagged as never read.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
`native_screen_render_post` (`ScreenEvent.Render.Post`) hands the mod the
mouse in GUI pixels and the frame's partial tick besides the screen view,
as the event's `mouseX`, `mouseY` and `partialTick` do, so a mod can draw a
hover highlight.
The view also carries `menu`, the vanilla menu type of the menu the server
opened (`minecraft:generic_9x3`, `minecraft:furnace`, a mod's own id), and
None for a window the client opened. Every container, station and mod
window draws on the inventory screen, so `id`'s doc says they all report
`inventory` and are told apart by `menu`, `title` and `slots`.
The view is built without allocating: slots are counted the way `Game::slots`
lists them (`slot_count`, with `station_slot_count`, `mount_slot_count` and
`native_screen_slot_count` beside their list versions) instead of being
collected into a `Vec` every frame, and the id comes from the screen
`hud_layers` drew rather than from a second argument beside `self.screen`.
The hook and the fragment say which screens never fire it: the title screen,
which returns before its point, and the prompts, books and sign editors
drawn over gameplay after it, which keep `Screen::Playing`, are skipped, and
draw over what the hook drew. Vanilla fires Render.Post for them; that gap
is now written down.
With no mod installed the body stays `let _ = (...)`: the frame is vanilla's.
Co-Authored-By: DeepSeek <[email protected]>
- /summon minecraft:spawner_minecart with no NBT runs BaseSpawner.load on an
empty compound, as SummonCommand.createEntity does, so SpawnPotentials is
WeightedList.of(new SpawnData()) rather than empty.
- The save test passes no raw tag and changes the delay after loading, so it
checks the cart's own save arm.
- The cart model test covers spawner_minecart#main.
- Port ledger: MinecartSpawner's tick, save, load and display state are ported.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
More review fixes on the flat generator:
- FixedBiomeSource.findBiomeHorizontal for the stronghold rings of a fixed biome source.
- The ring cache key carries the structure_overrides, not a constant.
- A structure_overrides list naming no existing set lets every set generate; an unknown
biome is plains.
- A flat generator that cannot be built leaves the world on its flat ground; the column
queries cover the whole level; far decoration is off for a flat world.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The loader dropped an on_consume_effects entry whose type the game
doesn't know silently. It now keeps the kind and the JSON the file
wrote, and the game hands them to native_consume_effect when the item
is consumed, on the server thread, for the acting player. The game's
own effects, sounds, teleports and drops are open to the hook; an
unclaimed kind does nothing, as an unknown kind on any loader.
A bare type name without a namespace is minecraft's own — apply_effects
is minecraft:apply_effects — so only a namespaced kind outside
minecraft's becomes Native; an unknown minecraft: kind is dropped. The
hook sees the item's own consumable component, not a patch on the stack.
Co-Authored-By: ZCode AI agent (GLM-5.3)
The wall-table test's registry has every block's default, a kill by a guest
who left names no player, and the load hook's docs say two loads can overlap.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
NeoForge's CompoundIngredient matches if any child does — the union of
their items, not the intersection (that is IntersectionIngredient).
The ingredient_type key is neoforge:ingredient_type, type only as the
older packs' spelling; children is the field name, ingredients as the
older spelling. neoforge:fluid is not among NeoForge's item ingredient
types (its fluid ingredients are a separate registry for fluid stacks,
not items), so it is dropped until a use case arrives with a source.
Review: rdbtCVS's agent in the pull request.
Co-Authored-By: ZCode AI agent (GLM-5.3)
spawn_loaded_entities hands out the id before native_entity_joins runs,
and a cancelled orb doesn't give it back, so the fallback returned an id
no entity holds. Take it only when an orb holds that id (root_id), and
test both cases.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QT61YPiPzVNysmVHsH
Vanilla reads the block at every drawn position of a ticking section; most
hold stone, dirt or air, and each read is a chain of cache misses (chunk map,
section, palette, words, block facts). Each ticking section now keeps a
4,096-bit mask of the blocks a random tick can do anything for (a randomly
ticking block, or lava), built at its first draw and dropped when the section
changes. The draws, their order and every tick that runs are the same.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Empty stacks a mod leaves don't drop, the load hook runs outside the
cache's lock, a mob's context names the table that rolled (one helper with
drops::die), the wall-variant table lookup has a test, and the docs say
which breaks the drops hook hears.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
native_loot_table_load (horizon_world::loot) hears each loot table's JSON as
it loads, after the data packs' (LootTableLoadEvent, LootTableEvents.MODIFY),
and native_loot_drops hears a broken block's or a dead mob's items with what
dropped them (global loot modifiers, MODIFY_DROPS).
Closes#5522
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: fixes#5505 (lan_bot 13/15 failing on main since #4698 always sends minecraft:dialog data). Test-only: the known-pack assertion allows data for the dialog registry. lan_bot 15/15 on this head with vanilla data; fmt ok.
Review fixes for the structures on a flat generator, from the 26.3 bytecode:
- createForFlat searches the stronghold rings with seed 0.
- FlatLevelSource's sea level is -63 and its generation context starts at Y 0 with depth 384.
- A null layer (not opaque to MOTION_BLOCKING) is skipped by getBaseHeight and
buildTerrain and reads as air in getBaseColumn; a FillLayerFeature fills it after the
structures (TOP_LAYER_MODIFICATION).
- spawnOriginalMobs is empty for a flat generator: no generation-time creature packs.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Since #4698 the server sends minecraft:dialog's data even to a client
that knows the core pack (Horizon's client reads the dialogs from the
packet), so every bot join panicked on the configuration step's "no data
with the known pack" assertion and 13 of 15 lan_bot tests failed. The
assertion now expects data for the dialog registry and stays strict for
every other one.
Fixes#5505.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
#4698 made the minecraft:dialog registry always carry its entries' data,
on the idea that a client does not rebuild dialogs from its known pack.
Vanilla's does (RegistrySynchronization.packRegistry): an entry of a
pack the client knows goes without data, and the dedicated_server
harness saw Horizon send the dialog registry in full where vanilla sends
Optional.empty for each entry. Horizon's own client answers
SelectKnownPacks with no packs, so it keeps getting every entry in full;
a dialog a world's pack adds or replaces is marked from_pack and still
carries its data, so quick actions from a data pack reach every client.
Fixes#5525
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Adds `screen_render_post`, the point of NeoForge's `ScreenEvent.Render.Post`: once a frame, after the
open screen has drawn itself and its tooltips, a native mod may draw over it. It is handed a
`NativeScreenView`: the screen's id (one per `Screen` variant, from `Game::screen_id`), the open
menu's title (borrowed from the menu, not copied: no per-frame allocation), the container panel in GUI
pixels or the whole window, and the menu's slot count.
The firing site is the end of `hud_layers`'s screen match, where every screen but Win and Options is
drawn. Those two keep their early returns: folding them into the match would draw the HUD and the
prompt path behind them, so the hook fires on their own path, after their drawing, instead. It does
not fire while playing with no screen, and it is a `&self` hook on the window only, so a mod can
read and draw but not change the screen.
With no mod installed the body is `let _ = (...)`: the output stays vanilla's.
The hook is API next (`tools/native-hook-fragments/screen_render_post.json`); `native-hooks.json` and
`NATIVE_HOOKS_API` are untouched.
Co-Authored-By: DeepSeek <[email protected]>
An 8x screenshot at 200% Resolution Scale with the video memory already full
crashed: the depth buffer's allocation failed, and its view raised a
validation error that only an out-of-memory scope was watching. The target
and depth buffer, and the whole scaled frame, now run inside out-of-memory
and validation scopes; a failure writes no PNG and tries again next frame at
half the size, down to the window's.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
break_vehicle's ten punches have run two ticks apart since they went
into one script (bd09f8b03). Player.attack scales a hit by its charge,
0.2 + s * s * 0.8, and a fist charges in 5 ticks, so each later punch did
0.4: 10 for the first, then 4 a hit against 2 of decay, 28 after ten of
the more than 40 a vehicle needs (vanilla would take 17). The punches
are now 5 ticks apart, fully charged, twelve of them: the vehicle breaks
on the 8th, and a tick of drift still nets at least 4 a punch.
Fixes#5526.
Written with Claude Opus 5.5 in OpenCode.
Commands(CommandSelection.DEDICATED) registers SeedCommand and
VersionCommand with checkPermissions (LEVEL_GAMEMASTERS) and leaves out
PublishCommand and UnpublishCommand (includeIntegrated). Horizon's
executor tree used the integrated rules on --dedicated, so any player
could read the seed of an SMP. The Commands packet already cut them; the
executor and /help now do too.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Sort solid cutout quads while the worker-side side buckets are built, so rendering uploads the canonical worker result without a render-thread permutation or duplicate allocation.
Co-Authored-By: Codex <[email protected]>
A rider kept its vehicle id after the vehicle left the level. A respawn into
another level keeps the id when KEEP_ENTITY_DATA is set, and a removed
nautilus or minecart leaves it behind. remote_ride then kept the player in
the riding pose with no vehicle under it: it did not move, sneaking did
nothing, and the seat stayed until a vehicle packet named it again.
Vanilla Entity.baseTick stops riding a vehicle that is removed
(isPassenger() && getVehicle().isRemoved() -> stopRiding(), Entity.java).
Do the same in remote_ride when no minecart, boat, cushion or mob has the
id, as tick_riding already does on the local path (boats.rs).
A tadpole is an AbstractFish with a brain: it gets panic 1.25, avoid-player and
the fish random swim goals, SmoothSwimmingMoveControl(85, 10, 0.02, 0.1, true),
SmoothSwimmingLookControl(10), water-bound navigation, and the TadpoleAi brain
(panic, look at player, follow temptation, swim/walk/stay-in-water gate).
Its travelInWater is AbstractFish's (0.01 speed, 0.9 drag, -0.005 gravity).
Adds mob_ai parity scenarios tadpole_idle, tadpole_tempted, tadpole_panic.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
agent_control_e2e's use-screen scenario opens the inventory and the chat over
a golden apple in use (survival) and checks the use is released in that tick:
the stack stays at 64 and no effect appears, while eating with no screen still
finishes. With the call sites in game/remote.rs disabled it fails with the
apple at 63, so it covers the bin wiring the interaction-input parity harness
cannot see; the parity rows stay as oracle-side coverage.
Co-Authored-By: DeepSeek 4.1 Flash (DeepSeek Harness) <[email protected]>
Adds step-on, inside and fall-on callbacks with entity and block context; landing damage is adjustable before sounds and hurt.
Co-Authored-By: GPT-6 <[email protected]>
A cross model (flower, fungus) carried by an enderman drew both of its
coplanar sides through the entity pipeline, which does not cull. The two
quads sat at the same depth and flickered against each other (#5421).
CarriedBlockLayer.submit draws the block through BlockModelRenderState,
which uses the item cutout sheet (RenderPipelines.ITEM_CUTOUT). That
pipeline keeps the default back-face culling, so only the side facing
the camera is drawn. Drop the quads whose face normal points away from
the camera after emitting the block model. CULL_BACK would also drop
them, but it turns off the normal lighting this model needs.
Refs #5421
The hardening edit put a raw newline inside the fragment's `doc_summary`, which
JSON reads as a control character: `catalog_matches_hook_points` failed to
parse it. The line break is the `\n` escape the other fragments use.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
tick_lan_mount_menus checked every open mount menu on every level pass,
not only in the player's own level. When a second level ticks (a player
or a force-loaded chunk in the Nether, a portal ticket), its pass found
no such mob and closed the menu the tick it opened. Like the station,
crafting and merchant ticks, it now skips players not in the level in
place (ServerPlayer.tick checks stillValid in the player's level).
lan_e2e mount_other_level force-loads a Nether chunk and checks that the
host's and the guest's horse inventories stay open, opened with shift +
right click and with the inventory key while riding.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The issue asks the hook to run for every placer that goes through vanilla's
path, not only the player. `the_level_chooses_the_state_a_block_is_placed_in`
now drives a dispenser's `DirectionalPlaceContext` as `dispense_shulker_box`
builds it — `directional_can_place`, then `place_block` — and asserts the
furnace lands in the hook's state, with the hook handed the dispenser's own
click (its yaw, the front block it replaces).
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
Review point on #5516: an answer naming another block would place a block the
item isn't bound to, and skip that block's `setPlacedBy` and sound. Vanilla's
`getStateForPlacement` sets properties on the block's own `defaultBlockState`,
so the hook's answer is taken only for a state of the block being placed;
another block's, like an id past the state table, leaves vanilla's state.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
A mod block took no bone meal: `BoneMealItem.growCrop` looked the block's
`BonemealableBlock` up in the port's classes and stopped. The level now
answers vanilla's three questions for it, ahead of the `BonemealableBlock` a
content template would give it, so a mod block can refuse, roll and grow:
- `is_valid_bonemeal_target` — `isValidBonemealTarget`. Asked on both sides,
as vanilla asks it, so a client predicts the use with the mod's answer.
- `is_bonemeal_success` — `isBonemealSuccess`, drawing from
`native_level_random` so a mod's draws stay in vanilla's order.
- `perform_bonemeal` — `performBonemeal`, with the level's ticks.
The hooks are `native_is_valid_bonemeal_target`, `native_is_bonemeal_success`
and `native_perform_bonemeal` (a fragment, no API number taken), plus
`native_level_random` for a world hook, the stream `native_level_random
(dimension)` hands a server hook.
One insertion point serves every placer: `Sim::grow_crop` is where the
player's `BoneMealItem.useOn`, a dispenser's `BoneMealItem.growCrop` and a
villager's bone meal meet, and the meal is used up and level event 1505 fired
by the callers as before, so a mod block gets vanilla's particles and sound.
The client's copy of the level now forwards the one read-only question
through `DryLevel`, and a mod block's growth particles stay inside it, as
vanilla's `GROWER` type does, where the port knows no type for it.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
A world whose Version.Series is not main went straight to the file fixes
and the upgrade in WorldRequest::prepare, which the dedicated server and
the agent launch both reach. Vanilla refuses it: the dedicated server's
Main.java returns before DataFixers.getFileFixer().fix when
summary.isCompatible() is false, and WorldOpenFlows.
openWorldCheckVersionCompatibility shows the incompatible AlertScreen
(LevelSummary.isCompatible, DataVersion.isCompatible).
The title screen already refused such a world. The open path now does
too, so an incompatible world is never upgraded and saved in place.
The Recover screen's Attempt to Restore moved the broken level.dat aside
and put level.dat_old in its place without copying the world first.
Vanilla RecoverWorldDataScreen.attemptRestore runs
EditWorldScreen.makeBackupAndShowToast before restoreLevelDataFromOld:
a zip of the whole world folder without session.lock. The restore runs
once that finishes, whether the backup worked or not.
The button now starts the same backup through the backup screen and
runs attempt_restore when it is done, with the backup toast.
Reorder solid cutout quads front to back during GPU upload while preserving the canonical section mesh and shared index batching. Include the before and after leaf-scene captures used for validation.
Co-Authored-By: Codex <[email protected]>
F2 now saves the frame at the game's resolution (Resolution Scale above
100%) times a new Video Settings option, Screenshot Quality (1x, 2x, 4x,
8x; 1x by default). The frame is drawn once into its own offscreen
target, with the HUD zoomed to the same layout, and capped to the GPU's
texture and buffer limits; if the large targets can't be allocated the
window's frame is saved as before.
Fixes#3030
Co-Authored-By: Claude Opus 5.5 <[email protected]>
native_has_analog_output and native_analog_output answer BlockBehaviour.hasAnalogOutputSignal and
getAnalogOutputSignal for a mod's block; Level.setBlock's comparator update follows the answer, and
native_update_output_neighbors is Level.updateNeighbourForOutputSignal.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Entity.move resolves Y first, where the body starts, then X and Z. On the
tick a player walks off an edge, or lands on one and slides off it, the
client is on the ground over air at the move's end. An instrumented
25-minute vanilla client soak on generated terrain matched all 117 of
its ground flags to this (104 walk-offs at the same height, 13 landings),
one per step down any slope or staircase: enough to push a player
running down a long one toward the kick. ground now also holds a claim
when the ground holds at the move's start X and Z at the new height. A
no-fall or jesus has nothing there either.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
ExperienceOrb.createInterpolationHandler returns a LinearInterpolationHandler, so a client spreads each server position correction over 3 ticks (doInterpolate in Entity.commonTick) and the orb's own move carries the target along (applyPredictedMovement). The remote path put the orb at the new position at once, so it jumped on every correction; the comment claimed vanilla's NO_OP handler.
LeashFenceKnotEntity.interact runs only when the ray meets the knot (its box from recalculateBoundingBox, 0.375 wide, y + 0.375 to y + 0.875) before the post. A click elsewhere on the fence is FenceBlock.useWithoutItem or LeadItem.useOn, which only ties the player's own mobs (LeadItem.bindPlayerMobs). interact_knot ran the take-back and shears branches for any fence click, so clicking the top of a post with stone took a tied horse's lead instead of placing the stone, and shears snipped every lead.
The prepared root holds features without vanilla's old "config" wrapper and
with a block state as an id, so the zinc ore never decoded and the
content-root test failed over real data (CI skips it without .local/).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Opening the inventory (or chat, or advancements) while an item was in use
left the use running: the key event sets the screen, and the release sentence
of Minecraft.handleKeybinds never runs while a screen is open, so the serving
side kept counting and ate the food behind the GUI. Vanilla releases in the
same tick: the key opens the screen inside handleKeybinds, Gui.setScreen runs
KeyMapping.releaseAll (clearing key.use), and the same call then hits the
isUsingItem release. Add interaction_input::release_on_screen_open and call it
from the inventory/creative, chat and advancements key paths; pause, server
containers, mount inventories and dialogs keep vanilla's later release.
Fixes#4472
Co-Authored-By: DeepSeek 4.1 Flash (DeepSeek Harness) <[email protected]>
`getStateForPlacement` had no way in for a mod: a block could only be placed
in its default state. `LevelAccess::before_state_for_placement` now asks the
level, once per block whose placement state is computed, with vanilla's
`BlockPlaceContext` (the clicked face, the placer's rotation, the clicked
position, whether the clicked block is being replaced, inside, the item in
hand) and the fluid of the position the block would go in, so a block can
face the player, sit on the clicked face or come waterlogged. The answer is
`Some(Some(state))` to place that state, `Some(None)` to refuse the placement
as vanilla's `null` does, and empty to let the block's own placement run.
`native_state_for_placement` is the hook a native bundle answers with, and
`native_state_with_property` builds the state from the block's default state
and one of its properties (a content block's declared ones included).
`Ctx`, the resolved context, is now `PlacementContext`, public and documented
as what a mod block is handed. Both placers that go through vanilla's path
reach it: the player's `BlockItem.place` and a dispenser's
`DirectionalPlaceContext` (both `Sim::state_for_placement`), and the answer
still takes its survival and entity obstruction checks.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
The new cart_slots was inserted between unpack_cart_loot and its doc
comment, so rustdoc merged both onto cart_slots.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: new head (author merged main) keeps the hopper→jukebox fix and jukebox_taken (song stops when a hopper takes the disc); verify passes on aa68b6653, merges cleanly with current main (no overlap in jukebox/machines/spill/save)
Main's handlers take a loop handle (EventLoopCtl) and about_to_wait's body is Game::idle: exit_on_error takes the handle, the split guard stays in idle, and run_event_loop stops the narrator after the split's loop.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: rdbtCVS's merge of main (same resolution as mine, arm order only). clippy -D warnings clean; anticheat 20, server 88, lan:: 116 tests pass on the equivalent tree. lan_bot fails 13/15 on main itself (#5505), not this PR.
The dither changes vanilla's look for every cutout from 64 texels up, so it now runs only under
Distant Plant Dither (ditherCutoutMips in horizon-graphics.txt). Turning it on or off reloads the
textures, since mips are made when they load. Decoded textures don't hold mips, so the sprite cache
keeps its version.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
The official server's first two hours with the anticheat logging only
(67 players): every player who wasn't flying or walking on water peaked
at a ground level of 7, and that was 4 claims standing still in one
second (a block the server didn't have after a long teleport). The seven
who were (no-fall, fly, jesus: claims every tick while rising or falling
through the air, or walking on water; most also kicked by vanilla's
floating check) peaked at 113 to 3749.
ground now defaults to setback 10, alert 15, kick 30, ban_after 2, and a
claim standing still isn't judged (no fall to skip; hovering is
vanilla's floating kick). Replayed on that log, the others peak at 5
and the seven are kicked, then banned. timer and prediction still only
log: prediction flagged nearly everyone, and timer's slow decay lets lag
bursts add up over a session.
anticheat_e2e's ground check now runs at the defaults with cheat bots
that move, and wants the ban count in the kick line.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: merging: two native helpers for server mods (nicknames in chat and the Tab list, the chat filter's verdict); the SMP mod's proof passed on this head with the fragment stamped. No change without a mod calling them.
An advancement of a content mod or a data pack can name a trigger of
the mod's (mymod:pipe_connected); the tree keeps it, but only the game's
own events called Game::trigger, so nothing could grant it.
native_trigger_criterion(uuid, trigger, test) is vanilla's
SimpleCriterionTrigger.trigger for one player: each open criterion of
that trigger goes to the mod's test, then to its player predicate, and
those that pass are granted and awarded as the game's own are. The test
reads the criterion's conditions and can check a vanilla item predicate
in them. It reaches the acting player inside their hook, and any player,
in their own level (for_player), where no player is swapped in;
minecraft: triggers are refused.
Fixes#5481.
Written with Claude Opus 5.5 in OpenCode.
Keystroke::parse takes a 'static str; the app comes from gpui_platform::application(), and
AsyncApp::update returns nothing.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Conflicts: game/lan.rs (the anticheat's HostEvent::Ban next to main's ChatFiltered) and horizon-server config.rs (main's [discord] and [chat_filter] blocks around the anticheat's ban_after wording).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
native_can_survive answers BlockState.canSurvive before the block does: as it is placed (BlockItem.canPlace)
and from the shape update of a block whose template checks it, which breaks it with its drops.
native_may_place_on answers VegetationBlock.mayPlaceOn, the blocks a mod plant can stand on, which its
template's canSurvive asks. Both through LevelAccess, on authoritative servers, as native_update_shape is.
Inbox::reset, called on Event::Reconfiguring after clear_lod: requests a proxy's previous backend held are forgotten instead of holding in_flight slots for the 30 s timeout, and the next offer's per_second makes a new bucket.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The profile's texture id comes from a server's NBT; Identifier's characters and
no '.', '..' or empty segment keep it from reading a PNG outside the resource root.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The ten lang tables conflicted where both sides added strings: main's table as the
base, plus this branch's account strings, without the "Microsoft account · signed in"
this branch removed.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Conflicts: lod.rs (main's wanted-tile heap from #4747; a failed tile stays out of the heap until its RETRY passes, then the heap is rebuilt so it is asked again; main's MAX_LEVEL 22) and remote.rs (send_far_terrain and main's tick_chat_session both run before ClientTickEnd).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Two games playing in one game folder at the same time (#4952, the launcher's
`play --account`) fought over its per-player files: one `options.txt`, one
`logs/latest.log` (the second game logged to the terminal only), one generated name in
`horizon-player.txt` (two offline players got one UUID) and one `servers.dat`.
`horizon-client --player-dir DIR` puts this player's own files in DIR: `options.txt` and the
settings files beside it, `logs/`, `crash-reports/`, `horizon-player.txt`, `servers.dat`,
`hotbar.nbt` and `screenshots/`. Its players still share what the game folder holds: saves,
resource and shader packs, mods (the enabled list included, `logging::shared_dir`) and native
mods' `config/`, which no longer follow `options.txt` but the game folder (`Game::game_dir`).
Without the option nothing moves.
Launch API 0.1.1 promises the option (docs/api/launch.md).
Tested: two games started at once in one folder with --player-dir players/Alice and
players/Bob each logged to its own players/<name>/logs/latest.log and saved their worlds in
the shared saves/; a third, without --name, kept options.txt, horizon-graphics.txt and its
generated horizon-player.txt in players/Carol/.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Main added tr.json after this branch began, so it lacked the 12 keys the i18n
completeness test asks for. They hold the English source until translated, as the
other untranslated entries do.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The nine lang tables conflicted where both sides added keys: main's table kept as
the base (its order and its translations), plus the keys only this branch has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Both sides added an Extra field and its NBT round trip (rabbit MoreCarrotTicks here,
ocelot Trusting on main); kept both.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Fox.onOffspringSpawnedFromEgg's only caller is Mob.checkAndHandleImportantInteractions
(a fox egg used on an adult fox), so the trust moves from use_spawn_egg to
spawn_egg_offspring. The port map's Fox.clearTrusted and readAdditionalSaveData rows
now say what is ported.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Main now clears a pack's custom images each frame before the shadow pass, which may write them; skipping that pass on a hit would leave them empty.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Measured after #5415: with the bots off GitHub's pool it ran 9-12 of its 20 jobs and nothing waited, while 19 jobs
waited for playhorizon's two checks runners (4 CPUs shared with parity), the oldest 18 minutes. The bots stay on
playhorizon-light; the checks runners take main's and same-repository pull requests' checks only when the
repository variable SELF_HOSTED_CI is `all`, for when GitHub's pool is full again.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Two documentation conflicts again, where main's management protocol lines grow at the same place as the RCON ones. Main's text is kept, and the RCON lines now sit away from them (before the management protocol's paragraph and its changelog line), so the next merge doesn't meet them.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The countdown of AbstractBoat.tickBubbleColumn moves to Boat::tick_bubble_countdown,
so a unit test checks it fires on the 60th tick over a column, with the column's
direction, and resets when the boat leaves.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A content mod's placed features now generate in vanilla's biomes through
NeoForge's biome modifiers (data/<ns>/neoforge/biome_modifier/, and
forge/biome_modifier/ for older mods): add_features, remove_features and
none. When the content root builds they rewrite the biomes' feature
lists, an added feature at the end of its step as NeoForge adds it; the
generator's FeatureSorter then orders them as vanilla does. Biomes no
modifier names keep their text, and a root with no modifiers keeps
vanilla's bundle. An unknown biome, tag, feature, step or type fails the
build with the mod and file named.
Part of #4953, closes#5431.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01NLEqWLKW2LHgtLer35pT8m
Vanilla's Item.use, getItemBlockingWith and applyItemBlocking read
DataComponents.BLOCKS_ATTACKS from the stack, so a sword given
`[blocks_attacks={}]` blocks like a shield. Horizon only read the item's
default component. BlocksAttacks::of_stack now takes the stack's own
component first (kept in StackData::other), then the item's, and the
local player's raising, use tick, blocking, off-hand intent check and the
remote client's use prediction go through it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QSY4AbxZbBtWqJLhFg
FlatLevelSource.createState: villages and strongholds (the classic flat
world's structure_overrides) and any set a world's settings allow generate on
the flat layers, with the fixed biome, and are found by /locate. The pipeline's
noise column is the layers (getBaseHeight, getBaseColumn), the terrain stage
fills them, and the structure steps run with no placed features.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
A tile's previous mesh (kept while near chunks fade in) lives in the arena as well; without moving it, a closing page holding one stayed closed until the fade ended and compaction stalled.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Follow-up to #5408. A summoned orb spawned but native_summon returned None,
since no command Target lookup covers orbs: return the stack's first
spawned id. Passengers without Pos were not placed and got lost: share
/summon's recursive place helper (place_summoned), which also supplies
the Fire and Rotation defaults.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QT61YPiPzVNysmVHsH
known() now checks the whole subtree, not only the immediate type:
not(not(unknown)) and or(true, not(unknown)) both failed to exclude
the file, because the outer combinator was a known kind and only the
direct child was checked. Both now fail closed, with tests.
Review: Aria's agent swarm (Dyna) in the pull request.
Co-Authored-By: ZCode AI agent (GLM-5.3)
An ingredient that is one of NeoForge's standard object shapes now
resolves to the flat item list it matches, where the string parser
dropped the recipe file whole: neoforge:compound and
neoforge:intersection keep the items every sub-ingredient names,
neoforge:difference the base's without the subtracted's, and
neoforge:fluid a fluid's item form, its full bucket (a water or lava
tag names the same still fluid). A kind outside these still fails the
file - a mod's own ingredient types come with their own hook, not the
loader.
Co-Authored-By: ZCode AI agent (GLM-5.3)
🤖 Agent: management protocol 6e, ip_bans/* and operators/* applied by the game as vanilla's services: parity harness vs vanilla 26.3 passed (50 exchanges over TLS); clippy, fmt, horizon-server (88) and horizon-server-files (13) tests pass.
Two conflicts, both documentation: docs/api/server.md's changelog and docs/server.md section 8, where main added the management protocol's allowlist and bans. Main's text is kept and the RCON lines follow it.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
docs/native-mods.md keeps main's API 58 notes, then the API mods section.
With the game module in the horizon-client library, its unit tests ran in the
library's test process and shared process-wide state with the library's tests
(health::HURT_SERVER, the sulfur cube archetypes): six tests failed. The game
module now builds into the library only outside tests; its unit tests stay in the
executable's test build, as before the move.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
ip_bans/{add,remove,set,clear} and operators/{add,remove,set,clear}, as
vanilla 26.3's IpBanlistService and OperatorService make them, through the
game's management request like the allowlist and the bans: an address or an
online player's address, those online from it sent away; an operator's level
and bypass as asked, else op-permission-level and the bypass they had, the
commands online players may run updated. The parity harness changes all four
lists on both servers (50 exchanges).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: author's rebase on main, same content as reviewed; pr_ready --quick on main + #5242#5259#5275#5321#5354#5356#5357#5358 passes (clippy, 2389 client/render tests, Python 508)
🤖 Agent: author's rebase on main, same content as reviewed; pr_ready --quick on main + #5242#5259#5275#5321#5354#5356#5357#5358 passes (clippy, 2389 client/render tests, Python 508)
Translucent terrain is drawn over a copy of the solid shadow map, so a pack can tell a point under
water or stained glass from one in shadow: shadow_map_translucent holds that depth, and shadow_color
the texture colour times the tint of the nearest such surface, with alpha 1 on a fluid.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
Called from native_explosion_detonate, native_explode took the outer
blast's source, cause, owner and hive/rail flags. Save and clear them
around server_explode, then restore them.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QT61YPiPzVNysmVHsH
native_play_sound took minecraft: events only. Now an event of another
namespace (mymod:machine_hum) is sent to players by name, as vanilla sends a
sound that isn't in its registry (Holder.direct); they hear it when their
resources define it, as a content mod's sounds.json and .ogg files do.
minecraft: events must still be registered.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
`/item replace`, `/item modify` and `execute if items` answered "The
target does not have slot container.0" for a chest or hopper minecart.
Vanilla's AbstractMinecartContainer.getSlot is ContainerEntity's
getChestVehicleSlot, which takes any index below the container's size
(27 for a chest minecart, 5 for a hopper minecart), and reading or
writing it unpacks the cart's loot table with no player first.
Slot::Cart gives commands those slots; cart_slots drops the loot table
the same way, so no player_generates_container_loot fires later.
The command execution corpus checks both carts against vanilla: 66/74 of
the items and loot group matched before, 74/74 now.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: merged main (mod_effects in the bounce test); pr_ready --quick on main + #5242#5259#5275#5321#5354#5356#5357#5358 passes (clippy, 2389 client/render tests, Python 508)
Both PRs correct rows of the same port-map tables; the conflicts were
neighbouring rows, each changed on one side only, so every row keeps the
side that changed it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
setValuesFromPositionPacket measures tooBigToInterpolate from
PositionMoveRotation.of(entity), which is getClientPositionAndRotation:
the interpolation target while one runs. handleEntityPositionSync
measures from the entity's position. MovePacket keeps the two apart.
The mob and armor stand branches of move_entity now share stepped_move.
A mob put somewhere at once also resets its previous yaw and pitch
(snapTo's setOldPosAndRot), so it doesn't spin for a frame, and every
spawn starts with a fresh interpolation handler, a stand's included.
Written with Claude Opus 5.5 in OpenCode.
🤖 Agent: management protocol 6d, allowlist/* and bans/* applied by the game as vanilla's services: parity harness vs vanilla 26.3 passed (42 exchanges over TLS); pr_ready clippy/fmt/python pass; lan_bot's 13 minecraft:dialog failures come from main (#4698), not this PR.
mod_loaded now asks which mods are loaded (the mod-content root's new
horizon-mod-ids.json marker), not which namespaces they own — a mod that
also owns a namespace doesn't make that namespace another mod. An
unknown condition kind fails the file through not and or as well, not
just on its own: `not` of an unknown kind is not a pass, and an `or`
over one doesn't short-circuit past it. And the stations, the recipe
book and brewing apply the world's packs first and gate the merged
file, as the crafting loader already did: a false-conditioned pack file
leaves the lower recipe gone too, not still active underneath.
Review: Aria's agent swarm (Dyna) in the pull request.
Co-Authored-By: ZCode AI agent (GLM-5.3)
Server owners asked for one identity in the channel (the webhook's name
and avatar, e.g. "HorizonSMP") instead of a per-player name and head on
each post. Drop the per-player username/avatar_url and the Line field
that only fed it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: queued for merge once verify passes (Ludovic's ask: keep operators' commands out of other players' chat on the official server; default stays vanilla).
Conflicts: game.rs keeps the api_mods export beside main's narrator module;
horizon-threads takes main's windows-sys features (the same Windows fix);
native-mods.md keeps main's API 47-57 notes, then the API mods section.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: merged main (the verify failure was main's title keyboard test, fixed in #5425). Wolf.getAmbientSound growls on isAngry (anger timer), checked in bytecode.
The fixture's server generates a 1024-bit RSA key in a debug build before it answers; on playhorizon's checks
runners (4 CPUs shared with parity, niced) that took longer than the 10 s the test waited for the session's event.
A passing run never waits for the timeouts.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
available_parallelism is the process's share (its CPU affinity, a cgroup quota): pinned to 4 of 8 CPUs, the game
reported 4 threads and 8 cores, and this_machine_is_described failed on playhorizon's pinned CI runners. On Linux
the threads are the processors /proc/cpuinfo lists, where cores come from.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A systemd user service's umask is 0002 on Ubuntu, and the launcher's tests check the mode of the files they write.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
GitHub's 20 hosted jobs at a time are shared by every pull request: on 2026-10-31 400+ runs queued and every job
waited 20-25 minutes for work of 0.3-4 minutes. Main's pushes and this repository's own pull requests now run
checks' verify and launcher on playhorizon (label playhorizon-ci, two runners); the bots that run main's own scripts
(pr labels, overlap note, cancel-on-close, claim command, parity-scope) on two small runners there
(playhorizon-light). A fork's pull request stays on ubuntu-latest, and the runners' job-started hook refuses one
whatever its workflow says. SELF_HOSTED_CI=off sends everything back to GitHub's runners.
tools/ci/checks-runners.sh installs the runners (systemd user services pinned to CPUs 4-7 in horizon-ci.slice,
shared cargo home, a build folder per runner outside the checkout); docs/ci-runners.md; tests/test_ci_runners.py pins
the fork rule.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: merged main (the verify failure was main's title keyboard test, fixed in #5425). pack.mcmeta: first JSON value only, as Gson's JsonParser; BOM kept; test.
An armor stand is a LivingEntity, so a vanilla client moves it with the
SteppedInterpolationHandler over its update interval (3 ticks), position
and yaw together, and ArmorStand.tickHeadTurn has the body follow yRot.
Horizon eased stands with the generic three-step lerp and set the yaw at
once from each packet. Stands now share the mobs' stepped handler:
ArmorStand::interpolate takes its step, keeps the yaw continuous and
saves yBodyRotO, and the renderer and the F3+B hitbox turn the body by
Mth.rotLerp(partialTick, yBodyRotO, yBodyRot).
A move of more than 64 blocks now snaps (tooBigToInterpolate), with
nothing left to blend from (Entity.snapTo). TeleportEntity snaps too, as
setValuesFromPositionPacket does; before, only EntityPositionSync did,
so a far teleported mob also slid there over three ticks.
Written with Claude Opus 5.5 in OpenCode.
🤖 Agent: merged main (fragment stamped by #5390 → doc edits moved to API 50 section; main's dragon test updated to HurtSource); cargo test dragon + native pass
allowlist/{add,remove,set,clear} and bans/{add,remove,set,clear}, as vanilla
26.3's AllowlistService and BanlistService make them. The supervisor checks the
parameters and forwards each edit to the running game as a control request
(management <connection> <json>); the game finds the players as
MinecraftPlayerListService.getUser does (by id: the user cache, else the
session service; by name: the user cache, else the profile service, else an
offline player), changes its lists, logs RPC Connection #N lines and sends the
banned or unlisted players away. The reply is the list from its file.
horizon-server-files gains UserList::clear, UserCache::get_by_id and
profiles::find_by_id. The parity harness changes both servers' lists and
compares every answer (42 exchanges); neither server reaches a profile service
there, so names resolve to the same offline players.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A pack asks for a shadow map in its manifest and the renderer draws the solid terrain the light
sees into it each frame, from the shadow camera's lists shaders/ packs already use. The material
and post stages read it as shadow_map with shadow_sampler and the prelude's shadow_position, and
the frame gains light_direction, has_shadow, shadow_view_proj, shadow_distance, shadow_distortion,
shadow_resolution and shadow_center_texel. A pack without [shadow] gets a 1x1 map that lights
everything. Translucent terrain and entities don't cast yet.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
DebugLevelSource: barriers at y 60 and the block states in registry order on
a grid at y 70, one every other block. A Debug world's blocks never change
(Level.setBlock fails; setblock, fill and clone report it), nothing ticks
(scheduled, random, natural spawning), the login packet says isDebug, and
level-type minecraft:debug_all_block_states creates it.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
🤖 Agent: Ocelot.mobInteract trust (1 in 3, events 40/41), DATA_TRUSTING synced and saved, removeWhenFarAway and the avoid goal; superset of #5423 (closed)
A mod item with its own fill level (a tank, a battery) had no way to draw
the bar vanilla's Item.isBarVisible/getBarWidth/getBarColor give tools and
bundles. native_item_bar fires where GuiGraphicsExtractor.itemDecorations
picks the bar, after the bundle override, with vanilla's choice in `bar`;
a mod may replace or hide it. Nothing changes when no mod hooks it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
native_set_nickname(uuid, Option<&str>) shows a name in place of a
player's in chat (Player.getDisplayName's text: their messages, /me,
/msg) and in the Tab list (PlayerInfo UPDATE_DISPLAY_NAME); the hover,
the suggested /tell and the insertion keep the real name, as do the
log, commands and the server's events. native_text_allowed(text) asks
the server's chat filter. API number from the fragment's stamp.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Two conflicts, both documentation: docs/api/server.md's changelog and docs/server.md section 8, where main added the management protocol's TLS and reworded its list. Main's text is kept and the RCON lines follow it.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
🤖 Agent: management protocol 6c, TLS on by default (PKCS12 keystore as vanilla, or a generated self-signed one): parity harness vs vanilla 26.3 passed over TLS (33 exchanges, TLSv1.3 both), horizon-server's 81 tests and clippy pass.
Upstream now sends custom_model_data itself; keep its codec and drop
ours, so this branch only adds custom_data to the LAN item sync.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The management server serves wss:// as vanilla's does: a PKCS12 keystore
(management-server-tls-keystore), its password from
MINECRAFT_MANAGEMENT_TLS_KEYSTORE_PASSWORD or the server property, rustls on
ring. Where vanilla stops without a keystore, Horizon generates a self-signed
one in .horizon/ and writes its path and password to [properties], so
management-server-enabled=true is enough; the start logs the certificate's
SHA-256 fingerprint. The WebSocket reads and writes through a Link (plain or
TLS, full duplex). The parity harness now runs both servers over TLS with one
keytool keystore.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
EntityRenderer draws an entity at Mth.lerp(partialTick, xo, x), but armor
stands were drawn at their tick position, so a falling stand jumped once
a tick (20 times a second) while boats, TNT and mobs slid. ArmorStand now
keeps its previous position: ArmorStand::interpolate saves it before each
interpolation step on a server's client, a stand riding a mob or the
player takes it from its vehicle's previous seat (as the mob riders do,
so it stays right while the level is frozen), and the in-frame tick
saves it too. The body, its light, the culling box, the name tag, the
flames and the F3+B hitbox all use the blended position.
Fixes#4416
Written with Claude Opus 5.5 in OpenCode.
rdbtCVS's review of #5365: a kick only marks the player as leaving, so
packets still queued (a lagging client's ticks) were handled after it,
flagged again and counted a second strike, banning the player in the
session the first kick warned. flag_to now drops them and counts
nothing once the player is leaving. Without a ban list, anticheat_ban
falls back to the anticheat's kick.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: merged main again: save_guest follows main's every-player-file rule with this PR's playerdata API; main's new autosave loop (PlayerList.saveAll) calls playerdata::save without the old vehicle argument. Clippy clean, persist tests pass.
🤖 Agent: merged main; what remains is the Experiments switches committing on Done (Cancel drops them). Main's feature-aware load replaces the Re-Create fix. Client builds.
🤖 Agent: merged main (most of this stack landed via #5069); what remains is pack item tags reading main's tags/<ns>:<registry>/<path> keys (tag_kind) plus two loot tests; namespace tests pass.
🤖 Agent: merged main up the stack; #4448's atomic vector rewrite carries main's new Preprocessed.warnings; #4790 handles Enter on its shader buttons in main's new keyboard focus path (#4792: shift+Enter resets, as shift-click). Each layer builds; horizon-shaders tests 76 pass.
🤖 Agent: merged main up the stack; #4448's atomic vector rewrite carries main's new Preprocessed.warnings; #4790 handles Enter on its shader buttons in main's new keyboard focus path (#4792: shift+Enter resets, as shift-click). Each layer builds; horizon-shaders tests 76 pass.
🤖 Agent: merged main up the stack; #4448's atomic vector rewrite carries main's new Preprocessed.warnings; #4790 handles Enter on its shader buttons in main's new keyboard focus path (#4792: shift+Enter resets, as shift-click). Each layer builds; horizon-shaders tests 76 pass.
🤖 Agent: Server bench against Pumpkin (black box, GPL-3.0 source never read): the runner, its tests, the doc and the 2026-10-31 report. CI green at this head, pr_ready pass (498 Python tests), 24 matrix runs on playhorizon with none failed.
Rabbit.RaidGardenGoal: a MoveToBlockGoal (0.7, 16) to carrots at their maximum
age on #supports_crops ground, which takes one age off them (or the crop when
it had none) and leaves the rabbit full for 40 ticks (moreCarrotTicks, saved
as MoreCarrotTicks, running down by 0 to 2 a tick).
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
native_world_drawn gives a mod the frame once the world is drawn: the device,
queue and encoder, the colour target, the world's depth, the matrices with and
without TAA's jitter, the entity quads with their runs, the atlas and the quad
indices. native_render_needs lets mods ask for the entity runs and the world's
depth without a shader pack; several can ask. Both live in the render crate
(gfx/native_render.rs), and the catalog test accepts hook points there.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
server.toml [server] admin_feedback_to_ops (default true, vanilla's
broadcastToAdmins) is written as horizon-admin-feedback-to-ops; false
keeps the [name: ...] lines to the console's log. Players who aren't
operators never get them either way (g.permission > 0, as vanilla's
isOp, checked by the server_ops_feedback harness).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Lines name players only by the event's `name`; a death message the game wrote
with a nickname gets the real name after it (smp-1 is adding /nick).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A native bundle with [api] (crate and entry) and no [[injection]] is an API mod: its
crate implements horizon_mod_api::Mod and the game asks it right after the item and
block hook points it is named after, once the mixin mods there are done.
- horizon-mod-api: the Mod trait (use_item, use_item_on, right_click_block,
left_click_block, destroy_block, place_block, tooltip), Action and Host.
- horizon-client: the game module moves into the library and the executable only adds
the allocator, so a build can link the game from another package; game/api_mods.rs
dispatches, and a mod turned off in Options > Mods is skipped with no build.
- launcher: API mods skip horizon-mixins; their crate and a generated native-game
package build the executable, so horizon-client never depends on a mod.
- horizon-mods: the manifest's optional api field and its checks.
- docs: native-mods.md "API mods (no mixins)", stability.md; example in
launcher/tests/fixtures/native-api-demo.
`feed_animal` takes the player who feeds an animal into love since this
branch; the probe passes none, so `--examples` builds again.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
#5327 landed `Mannequin.updateDescription` (the default label, hidden, or the
server's component under the name tag) and the two synched-data slots beside
the profile, so both rows stop saying missing.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`LivingEntity.entityEventForEquipmentBreak`'s note still said a player's
hands, the saddle and a guest's tool were not broadcast. Main's #5392 covers
the player half (both hands, armour, shield, a guest's included) and #5317
the mobs' half, so the row points at the slot-to-event table itself now and
says where both halves live.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`ShulkerBoxBlockEntity.moveCollidedEntities` was marked unsure in the
checklist; it is `game/piston_push.rs::Game::push_shulker_entities`, whose doc
names the method and moves every body the sweeping lid box meets along the
facing by the box's size + 0.01. That was the last row of #3516 without a
port.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`Arrow.updateColor` is `Projectile::arrow_effect_color` (its doc names the
method on the way past), `Cat.CatTemptGoal.canScare` is `CatTempt::can_continue`,
`Entity.applyImplicitComponent`/`applyImplicitComponentIfPresent` are the
component pass `entity_nbt::apply_item_components`, and
`LivingEntity.swapHandItems` is the entity-event-55 arm in
`remote.rs::Game::handle_packet`.
`Fox.canHoldItem` and `Fox.pickUpItem`, checked in the same pass, have no
implementation in the code at all (the generic `equip` gate lists the fox as
one whose pickup is elsewhere, and nothing is), so they stay out.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`TrialSpawner.FullConfig.overrideEntityData` is
`mobs/trial_spawner.rs::TrialSpawner::override_entity_data`, reached from the
dispenser's egg by `Game::use_spawn_egg_on_spawner`; and a blast breaking an
item frame (`ItemFrame.shouldDamageDropItem`) is
`game/explosion.rs::Game::explosion_hurt_entities`'s `explosion_breaks_frames`.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`ShelfBlock.useItemOn` was pointing at the client-side reaction
(`game/remote.rs::remote_use_item_on`) and its two swaps were still `missing`
for want of a citation; all three are `game/shelf.rs::Game::use_shelf`,
whose doc names the method and whose two arms are the single swap and the
powered chain's hotbar swap (`block.shelf.multi_swap`), with the block
entity update through `Containers::container_changed`.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
Eighty-eight rows of #3512, #3514, #3516 and #3517 are ticked in their
checklist while `tools/port_map/` still calls them `missing`: the pull
request that fixed each one ticked the box and did not touch the map. This
corrects the 63 that cite the vanilla method in port code — 36 through a doc
comment on the port itself, 27 through a mention whose note says so and gives
its file and line. Every `where` names a function that exists.
The 23 ticked rows with no citation at all are left alone and listed in the
pull request: a tick is not evidence enough to flip a verdict.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
Main's #5397 translated the same 17 strings into ja, pt-br and tr that dd722ee7a did, with
other wording. Git merges both without a conflict, which left each of those tables with 17
duplicate keys. Main's wording is kept and this branch's dropped, so the tables differ from
main only by this branch's 26 account strings and the key it removed.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01JLHp82d7UpxGuCncnLpJsq
Sheep.shear spawns each wool item with spawnAtLocation(stack, 1.0F): an
ItemEntity a block above the feet whose constructor sets its motion to
(rand*0.2-0.1, 0.2, rand*0.2-0.1), with the default pickup delay, and then
adds its own spread ((f-f)*0.1, f*0.05, (f-f)*0.1) on top. Horizon gave the
drops only that spread and spawned them at the body's centre, so the wool
left the sheep with at most 0.05 blocks/tick upwards and fell almost
straight down.
MushroomCow.shear builds each ItemEntity by hand at getY(1.0), so its
mushrooms get the constructor's motion and no pickup delay; the loop now
tells the two apart instead of applying the sheep's spread to both.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`Fox.FoxBreedGoal.breed` hands the breeding to the baby as trust, which needs
the player who fed each parent into love: the animals kept the love counter
but not its cause. `Animal.LoveCause` is now kept (`Extra.love_cause`, fed
from the player who hands over the food, saved and read), and
`fox::trust_breeders` gives the baby both players, once each.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
Ocelot.mobInteract: raw fish from within 3 blocks while the tempt goal runs is
used up; one time in three DATA_TRUSTING is set (entity event 41, hearts),
otherwise smoke (event 40). A trusting ocelot neither runs from players nor is
scared while tempted, is saved as Trusting, and no longer despawns when far.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Keeps main's arm (the anticheat drops a cheating attack, a move behind a
kept attack waits) and this branch's Hub::tick drain.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
`Fox.getTrustedEntities` and `Fox.addTrustedEntity` were modelled as "no
trusted players yet": the save wrote an empty `Trusted` list, nothing read it
back, and the avoid-player goal had no trust test, so a fox fled the player
who spawned it and nobody could become trusted (two rows of #3512).
- `Extra.trusted` holds the two `DATA_TRUSTED_ID` slots; `fox::trusted_entities`,
`add_trusted_entity` (first free slot, then replacing the second),
`clear_trusted` and `trusts` are the entity's own methods by name.
- The save writes the real list (`Fox.TRUSTED_LIST_CODEC`: int arrays) and
`mob_from_nbt` clears and rebuilds it, as `Fox.readAdditionalSaveData` does.
- `AVOID_PLAYERS` no longer flees a player the fox trusts, which is what the
row meant by a bred fox still running from you; `AvoidEntity::player_when`
takes the avoider so the predicate can ask. A fox spawned from a spawn egg
trusts the player who used it (`Fox.onOffspringSpawnedFromEgg`).
Tested without prepared game data:
`mobs::fox::trust_tests::a_fox_trusts_two_players_and_the_third_replaces_the_second`.
Still missing from the row: the breeding hand-off (needs a saved love cause,
which the mobs do not keep) and `DefendTrustedTargetGoal`; both are in the
note of the map rows.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
python3 tools/stamp_native_hooks.py on main 1e16d07: 11 fragments in
first-parent merge order (47 entity_join_level, 48 block_items,
49 native-sidebar, ... 57 chunks_players).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A data file with neoforge:conditions loads only when every condition
holds, and is silently absent otherwise, as NeoForge's ICondition leaves
it out of the manager. The gate covers the recipe loaders - crafting's
three kinds and the mods' own types, the stations, brewing, and the
recipe book - and the packs' overlays over them; mod_loaded asks which
mods' data is in the root (minecraft, always, and the content mods
that are on), and neoforge:true, not, and and or combine. A condition
the game does not know leaves the file out: a future shape fails
closed. Ported modpacks already carry such recipes verbatim (the carry
step keeps them), so cross-mod recipes stop loading blind.
Co-Authored-By: ZCode AI agent (GLM-5.3)
The one conflict was docs/api/server.md's changelog: main's longer management protocol line and the RCON line are both kept.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
#4638's keyboard test expects Enter on Create New World to open the create
screen at once, but 072cc6127 (landed beside it) shows vanilla's
"Preparing for world creation" screen first, as the other title tests
expect. The test now finishes the preparing step before checking.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A key bound to key.pickItem was turned into the middle button only while
playing; behind a screen it was dropped, so pick block on X did nothing in
the inventory or a creative tab. AbstractContainerScreen.keyPressed clicks
the hovered slot with CLONE (button 0) for it, as the middle button does.
Not on the creative search tab, whose box takes the key, nor behind a dialog.
Fixes#5378
Co-Authored-By: Claude Opus 5.5 <[email protected]>
BucketItem.emptyContents ends an evaporation with ServerLevel.sendParticles
(LARGE_SMOKE, the block's corner, 8, spread 1, speed 0, ALTERNATIVE).
Horizon played the hiss but sent no smoke for a player's bucket, and the
dispenser drew its own smoke one block too wide (nextDouble * 2 - 1 around
the corner instead of nextDouble from it) and sent none to other players.
evaporation_smoke does both: the window's puffs and the packet, now with
its randomization type (QueuedParticle.randomization). A client receiving
LevelParticles now honors ALTERNATIVE and ALTERNATIVE_WITH_SPEED, as
ClientPacketListener.handleParticleEvent does: positions spread over
[0, dist), speeds kept or scaled by nextDouble.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The server never put setSprinting's movement speed modifier on a guest's
body: START_SPRINTING only set a flag. Its movement speed then synced
without the modifier, and a vanilla client, which takes the modifiers as
sent (handleUpdateAttributes), walked while it sprinted: after a speed or
slowness change mid-sprint, until it stopped sprinting. handlePlayerCommand
now sets the modifier as setSprinting does.
The prediction's replay copied all the server's attributes on any
attribute or effect packet, the movement speed included, so it lost the
sprint modifier the client kept and played a sprint as a walk (a false
flag of a vanilla client after spreadplayers). An attribute packet now
changes the replay's sprint modifier only when it carried the movement
speed, as the client's does; effect packets copy only the effects. A
replay that starts again sprints as the client's first tick does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
#4787 moved the players' packets into Hub::drain_packets, which the game
calls between poll and tick_level; Hub::tick, the whole-tick helper the
lan_bot server uses, kept only poll and tick_level, so 8 of 14 tests saw
none of their packets handled. The bots of
guests_see_each_other_and_the_world_entities now send ClientTickEnd after
each move, as a client does: a second move in one client tick is kicked.
Fixes#5364.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A check that kicks now counts each player's kicks, across sessions and
restarts (anticheat-strikes.json). The kick that reaches the check's
`ban_after` bans instead: banned-players.json through the normal ban list
(source "Horizon anticheat", reason "Horizon anticheat: <check>"), the
player sent away with vanilla's ban message. combat bans at its second
kick; timer, ground and prediction keep ban_after = 0 until trusted.
Every kick and ban is logged with its evidence and told to the console
and online operators, so an op can /pardon a wrong one.
Also merges hub.rs's two S::Attack arms: the anticheat's shadowed #5123's
(unreachable pattern), so an attack never stopped the moves queued behind
it. Both apply now.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Main now hands title actions a &dyn EventLoopCtl; #4638's keyboard test
predates the preparing read Create New World goes through.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A dedicated server now refuses players' text that holds a hate slur before it
relays it: chat, /msg /tell /w, /me, /say, /teammsg, sign text, anvil names and
book titles. The sender is told in private, logs/chat-filter.log keeps the
original text for operators, and horizon-server's event stream gets a
chat_filtered event instead of chat (so the Discord bridge never forwards it).
The 2nd hit within 10 minutes mutes for 10 minutes, the 3rd bans for 24 hours.
Mutes persist in chat-filter-mutes.json; ops lift them with /unmute <player>
(console: unmute). server.toml's [chat_filter] sets it, on by default.
The new horizon-chat-filter crate holds the settings, the precompiled matcher
for the list (data/slurs.txt, or the owner's words_file) and the escalation.
Text is normalised first: case, accents, look-alike letters, invisible
characters, leetspeak, separators and repeated letters, with care for Niger,
snigger, knight, suspicious and the like.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- A sampler named like a function (Bloop's `uniform sampler2D texture;` read as
`texture(texture, uv)`) is the function where it is called, and the sampler elsewhere.
- A local named like a sampler, or like a sampler parameter, is in scope only after its
initializer (GLSL 4.60 4.2.2): `float depth = texture(depth, c).x` still samples the parameter
`depth` (Potato's depth-aware blur).
- A parameter's `const` (`const in float x`), which naga's parser refuses, is dropped: it only
promises the function doesn't write the parameter.
- naga's frontend nests arrays of arrays the other way round (`float x[3][2]` comes out as 2
arrays of 3; also `float[3][2]`, struct members and parameters), so indexing reads the wrong
element or fails validation. Two or more dimensions in a type are written reversed, so the
frontend's reversal restores GL's order; accesses are untouched.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
enable-rcon, rcon.port and rcon.password were read from server.properties but
nothing listened. The supervisor now serves vanilla's RCON (RconThread and
RconClient in 26.3): it listens on server-ip (every address when empty) and
rcon.port, logs in with the password, and runs each command through the same
console request as `horizon-server cmd`, so it answers while the game restarts.
Settings follow RconThread.create (no password or a port outside 1 to 65535
leaves it off with vanilla's warning; a port that can't be bound is a warning).
The packets follow RconClient: login 3, command 2, replies in packets of 4096
characters, -1 for a refused login, "Unknown request <hex>", and "Error
executing: <command> (<reason>)". Differences: packets are read from the
stream, the password is compared in constant time, a connection that hasn't
logged in within 30 seconds is closed, and at most 16 are open.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
🤖 Agent: the player half of onEquippedItemBroken (hands 47/48, armour, shield, guest breaks broadcast); complements #5317 (mobs) without touching creatures.rs
Grass blades a few texels wide at 2048 average to about the same alpha everywhere by 64, and no scale of that alpha keeps the coverage, so the alpha test passes the whole quad or none of it. From 64 texels up, each level below the base takes the mean alpha through an ordered dither, about as many texels in as the base has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
This reverts commit 3d5138cfd: #5392 broadcasts the player's equipment
breaks (both hands, armour, shield, a guest's included) through
onEquippedItemBroken, so with both the guest's main-hand break would be
sent twice. This PR stays the mobs' half.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
control.rs: main's control_command_inner takes the loop handle
(&dyn EventLoopCtl) like the rest of the game's methods.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A new cart kind (spawner_minecart, made by /summon: there is no item) that
shows a spawner block, drops and is picked as a plain minecart, and saves and
loads the BaseSpawner fields. Its serverTick is the spawner block's, with
broadcastEvent as an entity event on the cart.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
longbridge/gpui-fast: Zed's GPUI with retained views and cached scroll layers.
The API changes are mechanical: focus and blur take the App, the application
comes from gpui-fast-platform (wgpu on Linux), menus say whether they're
disabled, shadows whether they're inset. oo7 follows to 0.6.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: read the transport slice: custom component IDs behind a OnceLock table, checked in the fallback arm only (vanilla ids unchanged), unnegotiated stacks dropped; two integration tests
JukeboxBlockEntity.removeItem goes through setTheItem(EMPTY), which stops
the song and clears has_record; the hopper extraction path now does the same.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
`JukeboxBlockEntity` holds its disc in a `ContainerSingleItem`, so a hopper
or a dropper reaches it like any other container and a disc that lands
starts the song. Here the disc was a bare `Jukebox.record` that
`BlockEntity.slots` did not expose, so `HopperBlockEntity.getContainerAt`
found no container and nothing could be inserted (the
`JukeboxBlockEntity.setTheItem` and `.canPlaceItem` rows of #3516).
`Jukebox.record` becomes `items: [Option<ItemStack>; 1]` with a `record()`
accessor, which `slots`/`slots_mut`/`contents` now hand out; `Containers`
learns `Kind::Jukebox` and its `canPlaceItem` (a playable disc, into the one
empty slot); and a successful transfer calls the new
`Game::jukebox_inserted`, which starts the song the way the player's own use
does (block state, level event, neighbours and comparator).
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
A client helper that plays a sound event, vanilla's or the mod's own
from its assets' sounds.json, for this player only and up to a minute
later (SoundEngine.playDelayed), so a mod can time a blast's sound to
its shock wave. Audio's distance delay now goes through the same
play_after.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QSxLfzUbudHhcvZtgy
A server mod's helper that throws a falling block at a velocity: it
flies, lands and places itself as sand does, or drops as an item, and
can hurt what it lands on as an anvil does. For explosion debris and
collapsing builds. Refused on a client, for another dimension, air, an
unloaded chunk, values that aren't finite, or past 1024 falling blocks.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QSxLfzUbudHhcvZtgy
native_summon spawns any entity from its compound as /summon does;
native_move_entity, native_set_display and native_remove_entity move,
pose (transformation eased over ticks) and remove the entities Horizon
keeps without simulating them: displays, markers, interactions. An
item_display with a mod's own item model is a model that moves.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QSxLfzUbudHhcvZtgy
A client hook asked once a frame for eight vec4s that every stage of a
Horizon shader pack reads as frame.mods, so a native mod can tell the
pack it bundles where and when something happened (an explosion's
centre and age for a cloud ray marched in a post pass). Values that
aren't finite read as 0; all are 0 when no mod sets them. The
reference is regenerated.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QSxLfzUbudHhcvZtgy
#5226 added the proxy, game arguments, release headers and sponsor strings
to the other languages; ja, pt-br and tr (added beside it) never got them,
so the launcher's every_language_has_every_string_the_code_asks_for test
fails on main and on every PR's launcher check.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A pack with text after the closing brace of its pack.mcmeta was left out of the pack list, because serde_json refuses trailing input. Only the first JSON value is read now, so such a pack shows up.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
🤖 Agent: merged main into the reviewed head (clippy had failed on main's old unreachable S::Attack arm, fixed in 825a6bbcf). es_es.json completed (198/198 keys, placeholders kept, JSON checked); asset only, no code.
🤖 Agent: merged main into the reviewed head (clippy had failed on main's old unreachable S::Attack arm, fixed in 825a6bbcf). read the diff against vanilla's SulfurCube.hurtServer: a burning arrow or trident lights an explosive cube's long fuse at the hit (tnt_explodes), after the hurt; pending sounds become flags so primed and hit both play; unit tests.
boot_paint.rs draws the boot screen with GDI (Win32_Graphics_Gdi) and reads the
window's size (Win32_UI_WindowsAndMessaging), which horizon-threads didn't turn on.
🤖 Agent: merged #4445's new head (86e0d83db, main merged) up the stack; #4790 also merges current main (lib.rs: shader_options beside main's split_profiles); top of stack builds clean (client lib+bin).
🤖 Agent: merged #4445's new head (86e0d83db, main merged) up the stack; #4790 also merges current main (lib.rs: shader_options beside main's split_profiles); top of stack builds clean (client lib+bin).
`LivingEntity.onEquippedItemBroken` broadcasts the equipment-break event for
the hand, so every client plays the break and its crumbs. A guest's tool
breaks on the host (`Game::damage_held` acting for the guest), and until now
nothing told the other clients: it was heard only by the window that
happened to be simulating. The break now goes through
`Game::lan_entity_event` with `equipment_break_event(MainHand)` (47).
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`LivingEntity.entityEventForEquipmentBreak` has a client-side inverse here
already (`item_break::equipment_break_slot`); the wear added for #5317 spelled
the event ids out itself. `item_break::equipment_break_event` is that method
by name, `hurt_mob_armor` asks it for the slot's id, and a test pins the two
as inverses of each other over every slot.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`LivingEntity.doHurtEquipment` covers all the equipment a living entity
wears, not only a horse's body armour: `Game::hurt_mob_armor` now takes
`max(1, damage / 4)` off the helmet, chestplate, leggings and boots too,
Unbreaking and the fire-resistant spare applied, and a piece that reaches
its maximum empties its slot and broadcasts the equipment-break event for
it (`LivingEntity.entityEventForEquipmentBreak`: 49 to 52 for the armour,
65 for the body).
`tools/port_map`: `LivingEntity.entityEventForEquipmentBreak` is `inline`
now, with what it still misses written in its note.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`Horse.hurtArmor` (`LivingEntity.doHurtEquipment`) was not modelled at all:
`Mob::damage_with` had no notion of worn durability, so a horse's armour
kept its full durability however often the horse was hit, and the piece
never broke (the row of #3512).
`damage_with` now reports the hit's dealt amount for a mob wearing body
armour (`MobEvent::ArmorHurt`, which carries the damage kind), and the host
applies it in `game/creatures.rs::Game::hurt_mob_armor` the way the player's
own `hurt_armor` does: `max(1, damage / 4)` with Unbreaking applied, no
wear for a piece that does not wear when hurt or for a fire hit on a
fire-resistant one, and at the item's maximum the slot empties and the
equipment-break event (65, the body slot) goes out. A wolf's armour keeps
its own rule (`Wolf.actuallyHurt`, `hurt_wolf_armor`).
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`Frog.FrogPathNavigation` is an `AmphibiousPathNavigation` whose
`prefersShallowSwimming` override costs the water nodes deeper than ten
blocks under sea level, but `frog.rs::register` never changed the
navigation kind, so a frog pathed as a plain walker: it would not cross
water at all and no frog ever reached a shallow bank (the
`Frog.FrogPathNavigation.createPathFinder` row of #3512).
`NavKind::Amphibious` now carries the preference (`Amphibious {
prefers_shallow }`), which the path finder already had a field for and
always read as false, and a frog sets it. The axolotl, drowned and turtle
keep the plain form.
`tools/port_map/game-012.tsv`: both `Frog.FrogPathNavigation` rows are
`port` now.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`FlatLevelGeneratorSettings.updateLayers` and `.adjustGenerationSettings`,
`LevelChunk.postProcessGeneration` and `LevelSummary.backupStatus` are all
in the code (`flatgen.rs`, `game/postprocess.rs`, `save.rs`); the checklist
rows describing them as unported are stale.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
The checklists of #3512/#3514 assumed these were unported; reading both
sides says otherwise:
- `FallingBlockEntity.getTypeName` is no port at all: the anvil death text
is its damage type's message id, which takes no block-name argument.
- `Ocelot.OcelotAvoidEntityGoal.canUse` is `tame.rs::OcelotAvoid::can_use`,
registered last for a wild ocelot.
- `Axolotl.Variant.getSpawnVariant` loses a stale gap: spawn never draws the
blue variant in vanilla either, and the 1-in-1200 draw is on breeding.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
Block and item displays, a saved text display, the spear's stab count, a
firework rocket's entity hit and the arrow and stinger counts a player
carries are all in the code already.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
The second half of the blocks, redstone and fluids list: the jigsaw and
structure block editors, a spawn egg on a trial spawner, the campfire,
chorus flower and decorated pot a projectile lights or breaks, the wind
burst that opens doors, trapdoors and levers, the flower pot an empty hand
empties, the light block's picked level, and the It Spreads trigger.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
The same read of the checklists against `main` for the animals, entities and
world lists (#3601). Eighteen rows move from `missing` to `port`, and five
already-`port` or `inline` rows lose a gap note that has since been closed:
schooling-fish followers path to their leader, a copper golem opens doors on
its path, a mule never mates, a tamed wolf's kill is its owner's, and loot
tables apply set_components.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
Re-read the open rows of the blocks, redstone and fluids list against `main`
(#3601). Nineteen of them are in the code already, so their verdict-map rows
stop saying `missing` (or lose a gap note that has since been closed):
- the whole structure block screen and its behaviour: the mode buttons, size
detection between corners, save, load, place-with-settings and the redstone
trigger are `game/structure_blocks.rs`, and the outline box and render mode
are `horizon-client-render/src/structure_block.rs`;
- `TurtleEggBlock.destroyEgg` is the shared `BlockLandingRegistry::destroy_egg`
of the fall and step rolls;
- `LeverBlock.useWithoutItem`'s dust puff is `Game::lever_puff` on the client
that clicked;
- `LightBlock.getCloneItemStack`, `MagmaBlock.stepOn`,
`RedStoneOreBlock.spawnParticles`, `BeehiveBlock.playerWillDestroy` and
`TripWireBlock.playerWillDestroy` all have their port.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
Main grew `MobWorld.universal_anger` (`NeutralMob.isAngryAtAllPlayers`)
after this branch was written, so the test's MobWorld literal stopped
building. Restores the one-line fix the maintainer's agent had pushed here
before a rebase of mine overwrote it.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`Mob::do_move` rebuilt `Entity.restituteMovementAfterCollisions` by hand and
only ever used the entity's own `bounciness` attribute, so the block's
restitution never reached a mob: a slime block or a bed under a falling mob
stopped it dead where vanilla sends it back up (`Entity.getBlockBounciness`,
the row of #3514).
The port of that method now drives mobs too:
`movement_restitution::restitute_movement`, which the player and dropped
items already use and which `tools/test_restitution_parity.py` compares with
vanilla, is bound to a mob's own fields by `MobRestitution`. That also drops
the hardcoded honey check: which blocks suppress a bounce is the movement
data's `suppresses_bounce` tag, as for every other entity, and the
gravity comparison uses the living entity's effective gravity (0.01 while a
Slow Falling mob descends) instead of the bare attribute.
Mobs still sneak never, so nothing suppresses their own bounce.
`tools/port_map/game-005.tsv`: the `Entity.getBlockBounciness` row loses its
"mobs ignore it" gap.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
Contributors' agents that rebase and force-push over maintainer fixups
erase them (#5188, #5203). Ask for merges of main instead, and say what
happens if the fixes get erased anyway.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: re-read after the author's update: only crates/horizon-mod-registry/Cargo.lock regenerated against main (adds zip and its deps), so mod-registry's --locked clippy passes; the launcher strings reached main elsewhere.
Main added Japanese and Turkish catalogs and new strings since this PR;
the i18n test asks every language for every string any catalog holds.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
#4787 moved the players' packets into Hub::drain_packets, which the game
calls between poll and tick_level; Hub::tick, the whole-tick helper the
lan_bot server uses, kept only poll and tick_level, so 8 of 14 tests saw
none of their packets handled. The bots of
guests_see_each_other_and_the_world_entities now send ClientTickEnd after
each move, as a client does: a second move in one client tick is kicked.
hub.rs's packet loop had two S::Attack arms (the anticheat's, then
#5123's, which keeps the moves queued behind an attack for the next tick):
the second was unreachable, which fails clippy -D warnings on main and
a_move_behind_an_attack_and_its_swing_waits_for_the_next_tick. One arm
does both now (the same hunk as #5365).
Fixes#5364.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: read the diff: a software (DeviceType::Cpu) adapter skips the GPU occlusion pass by default, HORIZON_GPU_CULL still forces it either way; hardware unchanged; measured +5 % FPS on llvmpipe; unit test. Merged main, dropping the branch's duplicate #5251 fixes.
🤖 Agent: read the diff: screens read Shift/Ctrl from the OS modifier word (InputWithModifiers.hasShiftDown), the word reconciles stale physical keys and held latches on ModifiersChanged (toggles, pad holds and rebinds kept), Shift-drag distributes a carried stack, pad Y quick-moves, Ctrl+Q reads the same word, horizon-ctl key/release keep it in sync; new lan_e2e checks. Merged main: #5208's per-frame heal reads the same word, so the two agree; cargo check -p horizon-client --lib --bins --tests clean.
Main's new mirror_group shares the terrain bind group layout, which this
branch extends with the emission-thresholds texture.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
LivingEntity.onEquippedItemBroken broadcasts entityEventForEquipmentBreak
(47 hand, 48 off hand, 49 to 52 armour) and each client plays breakItem: the
stack's break sound and the crumbs. The acting guest's breaks go out with
their entity id when their turn ends; the shield's sound is its break_sound
component, not a sound the server plays.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The same mixed bots on playhorizon (release builds, CPUs 0-3, three rounds):
Horizon's tick is half Pumpkin's at 100 bots, equal at 60, and 1.8 times
longer at 20 (where it generates 63 % more new chunks a second). Idle
memory, start and median join are behind. The report, how to read it, and
the commands to run it again.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
tools/server_bench.py --servers horizon,pumpkin --pumpkin <binary>: Pumpkin
(a binary built by the user, run as a black box) gets its own default
configuration with the benchmark's settings, listening on 127.0.0.1 with its
Bedrock and LAN listeners, query, plugins and telemetry off. --repeat runs
every server and bot count again, alternated; a run with 0 bots measures the
idle server; each run records the bots' join time.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A build recorded each mod's digest over its mixins/<id>/ folder (after repinning
its copy), but the Mods page digests the whole .hnmod, horizon-native.json
included. They never matched: every registry native mod showed Needs rebuild
while Build said no build was needed. The build now records the digest the page
reads, taken before repin.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: read the diff: lan_e2e/agent_control_e2e run on Windows (junction fallback on WinError 1314, no vscreen), horizon-ctl launch clears HANDLE_FLAG_INHERIT on its std handles before spawning the game (cfg(windows) only); Linux/macOS unchanged; Python tests. Proof on Windows 11 in the body.
🤖 Agent: read the diff against vanilla's ChatScreen.normalizeChatMessage: String.trim, normalizeSpace (Java whitespace) and the 256 UTF-16 unit cut before sending and before the history; unit tests.
A client hook asked each frame for turns in degrees and an offset in
blocks, which add up across mods. Horizon scales the sum by Distortion
Effects (0% turns it off), bounds it to 10 degrees and half a block on
each axis, and multiplies it into the view after the hurt tilt and view
bobbing. It moves what the player sees, not where they aim.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QSxLfzUbudHhcvZtgy
🤖 Agent: survival: a replaced armour piece that doesn't fit is thrown from the eyes like Player.drop(stack, false), as the other Inventory.add-else-drop paths; 2 lines after #4231 took the rest; rdbtCVS approved. Parity rerun (lunge_motion failure unrelated).
Main merged #5148, which appends its stock-layout test to the same place at
the end of `game::tests` where this PR appends its two, so the hunk
conflicted. Both sides are whole `#[test]` functions: kept both, each with
its own closing brace, in main's order after this PR's.
`cargo test -p horizon-client --bins` runs all three green
(`a_mod_screen_on_a_machine_menu_takes_the_stock_layout_out` from #5148, and
this PR's `an_interact_packet_clicks_the_entity_it_names_not_the_nearest` and
`an_interact_packet_cannot_name_a_mob_the_pick_would_refuse`).
Mining a block above while jumping into reach: the attack stays held,
the aim is lost on the way down and found again on the next jump, and
the block never finishes breaking. Three new cases (pick, no tool,
stone by hand) over the miss-flicker pattern the report describes.
All 4,199 cases match the pinned 26.3 client, so the client-side
breaking logic (sameDestroyTarget resets, stopDestroyBlock on a miss)
is vanilla; the reported divergence must live elsewhere — the eye-ray
(reach/eye height) or the server's delayed destroy.
The scan over live and queued particles only runs once there are at least
ParticleLimit.SPORE_BLOSSOM of them in all, so a spore blossom costs nothing
extra in the common case.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: merged main into the reviewed head (clippy had failed on main's old unreachable S::Attack arm, fixed in 825a6bbcf). read the diff against vanilla's Level.sendBlockUpdated / ChunkHolder.broadcastChanges: native_block_entity_sync puts a loaded position in the block log, so the hub sends the block and its update tag once a tick; lan_bot test; tested in game by arismoko. Merged main, API number moved to a hook fragment.
🤖 Agent: merged main into the reviewed head (clippy had failed on main's old unreachable S::Attack arm, fixed in 825a6bbcf). read the diff: native_menu_button_v2 fires right after native_menu_button with the clicking player's cursor stack (clickMenuButton(Player, int)); a change is written back and sent with SetCursorItem; API number from a hook fragment. Marked ready.
🤖 Agent: merged main into the reviewed head (clippy had failed on main's old unreachable S::Attack arm, fixed in 825a6bbcf). read the diff against vanilla: PiglinBrute.wantsToPickUp (golden axe only), StartAdmiringItemIfSeen(119) with StopAdmiringIfItemTooFarAway(9) and StopAdmiringIfTiredOfTryingToReachItem(200, 200), wasHurtBy clearing the memory and ADMIRING_DISABLED for a player's blow; unit tests.
🤖 Agent: merged main and resolved debug_keys.rs: main's F3 Density line (#4389) is now debug_entry_lines' chunk_generation_stats arm, unchanged; cargo check -p horizon-client --bins clean. Reviewed earlier: vanilla's DebugEntry* lines and six debug renderers, profile datafixes; nothing extra per frame with F3 closed.
🤖 Agent: secure chat 3/3 after #4376: the client signs chat and /msg on enforce-secure-profile servers (fixes the official server's 'missing profile public key'); Chat Delay flush marks messages processed, command signing waits for the server's tree, reconfiguration ack as vanilla; tested by 4404c with a real account; main merged by the author.
Main's 31c1212bf added 17 strings to seven tables but not to ja, pt-br or the new tr, which
fails the every-language test on main. They are translated here as in the PR to main that
fixes it, so the lines are the same. tr.json also gets this branch's 26 account strings and
drops "Microsoft account · signed in", as ja and pt-br did.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01JLHp82d7UpxGuCncnLpJsq
A ResolvableProfile can carry a `texture` id (PlayerSkin.Patch body, a
ClientAsset.ResourceTexture). Vanilla draws a player head with that
resource-pack texture instead of the downloaded or default skin; Horizon
ignored it and drew the profile's default skin. Placed heads and head
items now read the patch, loading non-vanilla namespaces' textures from
the resource packs into a skin slot.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01P1ZTfB3AfwKgGnv7UFKYVD
The rebase dropped 24349e636's fixes to native_move_fluids (inverted put-back
check, a tank marked with no fluid in it, other tanks written away; and a
room limit so a cauldron's thirds never need to go back) and #5188's newer
head. This puts them back on #5188's head, with the move in fluids_move.json
as in the rebase.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A server mod's helper for Level.explode with no source entity: blocks
break as NativeBlast::Keep, Block or Tnt say (the drop decay rules),
entities are hurt and pushed, fire is optional, and players near hear
and see it. Refused on a client, for another dimension, an unloaded
chunk, or a radius outside 0 < r <= 32.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XnU6QSxLfzUbudHhcvZtgy
boot_paint.rs imports Win32_Graphics_Gdi and Win32_UI_WindowsAndMessaging,
which stopped resolving after #4441's Cargo.lock pass; main does not
compile. Also the test's language map resolves the keyed refusal.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
A tank dropped for an invalid capacity shifted the later tanks' numbers, so
the face and fill lists reached the wrong tanks (rdbtCVS's review). Also fixes
write_tanks' doc: an empty tank is written as minecraft:empty.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
ShulkerBullet.hurtServer: the hurt sound, fifteen crit particles and destroy.
The bullet is pickable and attackable (no pick radius, unlike
AbstractHurtingProjectile), so the crosshair finds it, a guest's attack
packet reaches it, and the blow plays Player.attack's sounds for a target
that is not a mob.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
- test_table_distinguishes_missing_scene_load_from_zero: #4407's later
commit makes table() refuse reports without matching conditions, so its
fixtures now carry the same full conditions as the contract tests.
- test_the_repos_list_is_the_nightlys: the list names Qwerner88880 since
f35210d2d; 10c71a65d updated test_nightly_set but not this test.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: Shader API [reflection] option and mirror-eye facing culling after #4379; only native shader packs see it; rdbtCVS's points fixed; merged up the stack, shader-api tests and a horizon-client check pass.
🤖 Agent: read the diff against vanilla's DownloadedPackSource (DOWNLOADED_PACK_SELECTION): active server/world packs show in the Resource Packs list as required, fixed at the top, and their server/ ids stay out of options.txt; unit test.
🤖 Agent: read the diff against vanilla's FallingBlockRenderer.shouldRender: a falling block whose state equals its cell's is not drawn; one block read per falling block per frame; unit test.
🤖 Agent: re-read the author's rebase onto main (only context in lan.rs changed). read the diff against vanilla: model light probed at the pose's eye height (crouch render offset undone, Joansitoh's review addressed), setPosToBed and xRot 0 on entering and while sleeping, stopSleeping's findStandUpPosition + teleport on wake for the local player and guests. CI failures were main's Audio build break (#5251).
🤖 Agent: LockCode.canUnlock (spectator or the main-hand item) and createMenu's refusal (action bar + block.chest.locked, both halves of a double chest) match vanilla; the lock stays in the loader's source tag, so saves keep it; approved by rdbtCVS's agent
🤖 Agent: merged main into the reviewed head (the earlier clippy/tests failures were main's Audio build break, fixed by #5251); no other change. read the diff against vanilla: a UUID target resolves through the level's entities (ServerLevel.getEntity), not only mobs; the scoreboard holder lookup shares it; execute on owner keeps only living owners (EntityReference.getLivingEntity). Corpus cases added to command_exec; CI parity runs them. CI failures were main's Audio build break (#5251).
🤖 Agent: merged main into the reviewed head (the earlier clippy/tests failures were main's Audio build break, fixed by #5251); no other change. read the diff against the 26.3 tag sulfur_cube_with_block_immune_to: the eight missing damage types added to the cube's immune list (fireworks left to #4997); unit tests. CI failures were main's Audio build break (#5251).
🤖 Agent: read the diff: flat draws now exist wherever the device supports them, but frames without a pack keep per-section draws off Metal/Intel Gen11 (min_sections usize::MAX), so only a pack's terrain and shadow pass draw flat; gives packs gl_Normal and mc_Entity on NVIDIA/AMD, measured black-box. CI green, parity runs on the head merged into main.
🤖 Agent: read the diff: with a shaders/ pack (oldLighting unset) terrain and fluid vertices drop the baked face shade, measured black-box against Iris with a probe pack (no source read); without a pack, lighting::cardinal as before. CI green, parity runs on the head merged into main.
🤖 Agent: read the diff: while a mod screen is shown on a hopper or 3x3 menu, the stock grid no longer answers first in slot_position, so the mod's slots and inventory_y apply; unchanged without a mod screen. Tested in game by arismoko; CI green, parity runs on the head merged into main.
The picture has no box: the glyphs' shadows keep them readable on Discord's
light and dark themes (Anthony's feedback). A chat_filtered event (the server's
chat filter) is never forwarded nor accepted in discord.events. [discord] is a
real section of init's template (webhook "" = off), so the settings reference
lists its keys; docs/api/server-reference.md regenerated. Alt text and fallback
text are cut after defusing, within Discord's limits.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
default no longer needs the template's properties
Both points arismoko raised by reading at e7f4440f0.
1. The NBT tables and the collision overrides carried one row per *template*
state, not per combination, so `facts.json` and the tables disagreed about
how many states a block has.
`nbt` (`states` lists, `cached_shapes` arrays) now emits a row per
combination in the order `states` numbers them in — a template state's
combinations are one run, and `Rename::id` maps the template state to the
first of it — with the declared properties on each row (the `properties`
compound `BlockState::to_nbt` writes; without them `BlockTable::check`
fails with "state identities differ" even once the lengths agree).
`collision_shapes` wrote `first + i`, so only the first combination of each
template state got a shape and the rest kept the table's -1. It now covers
every combination of every template state, all pointing at that state's one
shape: the geometry is the template's, the properties are the block's.
`structure-blocks.json` declared `states` as the template's states alone;
it now counts them times the fan-out, so the declared total and the lists
agree.
2. A `default_state` key the block declares itself was looked up on the
template anyway, so `{"template": "minecraft:stone",
"state_properties": {"lit": ...}, "default_state": {"lit": "true"}}` was
refused with "template default has no state properties" — stone has one
state and no properties at all. `combinations` already checks a declared
key's value, and only a key the template has narrows which of its states is
the default, so the lookup now happens only when such a key is present.
Three tests, each checked red before green: with the old `nbt` the table gets
4 rows where `facts.json` has 6; with the old `collision_shapes` the block's
last state maps to -1; with the old count `structure-blocks.json` says 4 where
its own lists hold 6; and with the old `default_state` guard the property-less
template is refused.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
Main's heightmap tiles built owned columns; with this PR's zero-copy
meshing the tile and edge columns are references, so the heightmap
columns are built first and borrowed. Spans replaces Vec in the
heightmap and its test.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
the_openssl_script_signs_what_the_launcher_checks runs
tools/ci/sign_release.sh with `bash`. On Windows, Command looks in
System32 before PATH, so with WSL installed it starts WSL's bash.exe,
which can't open the script's Windows path and the test fails. The
script only ever runs on the ubuntu-latest `sign` jobs of launcher.yml
and nightly-build.yml, so skip it on Windows the way the test already
skips a machine without OpenSSL 3. Linux and macOS run it as before.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
refresh_account and renew_for_game_with take the accounts.json a renewed profile is saved to.
The commands pass the launcher's own (Accounts::path()); the tests pass one that is never
there, so they no longer read or write the developer's launcher data. A new test checks that
a refresh saves the services' newer profile and leaves an unchanged one unwritten.
Main added pt-br.json and ja.json after this branch's strings: both get its 26 account
strings, and lose "Microsoft account · signed in", which the code no longer asks for (the
other tables dropped it already).
With main merged: profile_heads' test uses Accent::VIOLET (main made Accent a hue), and
docs/launcher-commands.md is regenerated for main's `completions` command, which main had
not written into it (the reference test failed on main too).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01JLHp82d7UpxGuCncnLpJsq
native_move_fluids moves up to its max between two blocks, one fluid,
native_move_items' shape: what the target won't take goes back into the
source, past the fill sides' leave. The source's tanks just gave the fluid
up, so they have the room for exactly what left them, and a cauldron
levels back up by the thirds it gave, which a drain hands out whole.
Co-Authored-By: ZCode AI agent (GLM-5.3)
Merges origin/main (#5167's join hook beside this leave hook). Main's blast
on experience orbs, a mob picking up a whole stack and native_set_item_entity
emptying a stack now tell native_entity_leave_level too.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Properties.store ends every line with BufferedWriter.newLine, which writes
System.lineSeparator(): "\r\n" on Windows (checked with JDK 25 there).
store() already does the same through line_separator(), but
stores_sorted_and_escaped and defaults_match_vanilla_init_settings
expected "\n" and failed on Windows. The first now expects the platform
separator exactly; the second reads the stored lines as "\n". Nothing
changes on Linux or macOS, where the separator is "\n".
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Review points from rdbtCVS on #4948: a failed read now clears the flag
before failed() so the error screen stays up with Create as its back
target; type_text is blocked during the read like the key paths; the
docs of finish_preparing, title_action and world_worldgen_overrides sit
on their functions again, and set_data_packs' orphan doc is gone.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: read the workflow: verify runs its python, fmt/clippy and test steps in one job again (one runner per run instead of four, the runner queue was 300 deep); the required check is still verify, the only one branch protection requires; no game code.
🤖 Agent: read the diff against vanilla: the dispenser's spawn egg now runs the same EntityType.updateCustomEntityTag step (apply_spawn_egg_entity_data, with its guards) a player's egg does.
naga's constant evaluator lacks mix, smoothstep, reflect, the pack/unpack functions and a few
more built-ins, so a const initialised with one fails the program ("Not implemented as constant
expression: Mix built-in function"; 66 times in Bliss). GL drivers fold them.
A global const that calls one now joins the globals our main initialises (as const ... ? ... : ...
already does), and a local const that calls one loses its const and is evaluated in place.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
boot_paint.rs imports Win32::Foundation, Win32::Graphics::Gdi and
Win32::UI::WindowsAndMessaging, none of which were in the dependency's
feature list, so every Windows build of main fails at horizon-threads
(E0433, the items are configured out). Linux CI never compiles this
file, so it stays green while Windows builds, including the launcher's,
are broken. The same file passed the font count mutably where the API
takes it by const - Linux clippy cannot see that either.
(The hub.rs duplicate S::Attack arm this pull request first also fixed
has since landed on main by itself.)
Co-Authored-By: ZCode AI agent (GLM-5.3)
🤖 Agent: Ludovic's decision (2026-10-31), one line in .github/trusted-reviewers.txt; merged main (verify had failed on main's Audio break, fixed by #5251).
🤖 Agent: read the diff: queued native mod messages are flushed after native_server_stopping while the host and players are still there, on the integrated and dedicated servers; in-game proof on both paths in the body.
🤖 Agent: its owner merged main into the reviewed head (95cdad8b3, unchanged). read the diff: helpers only (per-player sidebar objective, Tab header/footer, tick stats, sleep-count opt-out for AFK players, yaw/pitch on NativeServerPlayer), nothing per tick; the hook fragment is stamped after merge. pr_ready green and the SMP proof on a dedicated server with two clients, per the body.
🤖 Agent: merged main into the reviewed head (the earlier clippy/tests failures were main's Audio build break, fixed by #5251); no other change. read the diff against vanilla: SulfurCube.knockback sets needsSync before setDeltaMovement; unit test.
- appimage_path: only when this binary runs from inside $APPDIR, so a launcher started by another
AppImage's program never overwrites that program's file.
- AppRun no longer exports LD_LIBRARY_PATH (the game and every other child inherited it):
patchelf sets the binary's and libraries' run paths; CI installs patchelf.
- Agent (MCP) configs and desktop shortcuts start update_target(), not the copy in the mount.
- The AppImage is built from $BIN like the archives.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: merged main into the reviewed head (the earlier clippy/tests failures were main's Audio build break, fixed by #5251); no other change. read the diff: x / 0.0 and sqrt of a negative literal go through a private hz_zero global naga can't fold (GL's run-time inf/NaN), texelFetch1D/2D/3D map to texelFetch; unit test. Shader-pack translator only.
🤖 Agent: merged main again (the earlier CI failure was main's broken Audio initializer, fixed by f80c0360d); clippy -D warnings and the horizon-bots tests pass on the merge. Approved by rdbtCVS; bots-only, no game code.
A guest's completion request was answered over the host process's registries, whose player list on a
dedicated server is the one local player "Player". The server now answers with its real online players and,
as vanilla's providers do (OpCommand, DeOpCommands, PardonCommand, PardonIpCommand, WhitelistCommand), from
the operators, banned names and addresses, and the whitelist.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Mannequin.DATA_IMMOVABLE and DATA_DESCRIPTION (after Avatar's two and the
profile), updateDescription (default entity.minecraft.mannequin.label, none
for hide_description, or the saved component) and isEffectiveAi: an immovable
mannequin never travels. The client draws the description under the name
(ClientMannequin.belowNameDisplay) within 10 blocks.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
A new [discord] table in server.toml (webhook, events, enabled) starts a forwarder
on the supervisor's typed event stream: chat, joins, leaves, deaths (and kicks).
Lines of each 2 s are drawn as one PNG of the in-game chat box (half-black box,
§ colours, bold, 1-pixel shadows, wrapped at the chat's 320 px) with Pixelify
Sans (SIL OFL, already in horizon-theme), never Mojang's font. At most one post
every 2 s, longer when Discord's rate-limit headers or a 429 say so; a burst
past 20 lines is counted, not drawn; plain text when drawing fails.
No pings: every post sets allowed_mentions {"parse": []}, and every @ (and <@,
<# ...) gets a zero-width joiner, in the alt text and the fallback text too.
One player's lines carry their head (mc-heads.net by UUID) as the avatar;
their address is never sent. Only the webhook's host is accepted (Discord),
the API and panel show it as "(set)", and the API can't change [discord].
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: read the diff: remaining launcher strings (proxy, game arguments, releases, patch notes, sponsor, mod verdicts) go through tr() with entries in the seven catalogs; no behaviour change.
🤖 Agent: read the diff against vanilla's ShearsItem.useOn: shears on a growing plant head below age 25 set getMaxAgeState with block.growing_plant.crop, a block_change game event and one wear; bodies and max-age heads untouched. New use_item_on cases; CI parity runs them.
🤖 Agent: a resumed dig keeps the first START's clock (ServerPlayerGameMode destroyPos/destroyProgressStart, 0.7 split) on the server and the client; dig_clock, dig_stop and instant_digging LAN checks pass per the author; elythepreciouscat's remote_mining fix included; rdbtCVS approved; verify, clippy, tests green. Parity runs on the merged head.
🤖 Agent: read the diff against vanilla's PlayerHeadItem.getName: a profile with a name titles the head block.minecraft.player_head.named in tooltips and command feedback; custom name still wins; unit test.
🤖 Agent: FSR 3.1 frame interpolation (MIT, AMD FidelityFX 1.1.4) ported to WGSL, still unused (skip-release); rdbtCVS's div_ceil point fixed in 58b2c98d3; naga-validated, GPU tests by the author; verify green.
🤖 Agent: verified: Zip::open now keeps the file and reads only the central directory onward (as Java's ZipFile), entries read at their offsets with no shared cursor; parsing unchanged; tests cover lazy reads, long directories and broken END counts. Thanks!
🤖 Agent: read the diff: the shaders/ preprocessor's #if evaluates float literals (integer / and % kept without one), and an unreadable condition is a warning reading false instead of skipping the program; unit tests. CI green, parity runs on the head merged into main.
🤖 Agent: read the diff: sky, celestial and cloud lightmap coords, the sky normal, cloud face normals (w 0, the game's own clouds unchanged) and blockEntityId -1 by default, as a shaders/ pack reads them in Iris (observed with a probe pack, no source read). CI green, parity runs on the head merged into main.
🤖 Agent: None now samples terrain as vanilla (linear clamped sampler, sampleNearest unless RGSS; terrain.fsh), and distant tiles repeat their sprite per block, which also fixes RGSS/Anisotropic stripes on main since #4843 (checked in game by the author). rdbtCVS's change request was addressed (dismissed with his re-review). verify green.
🤖 Agent: GL backend: indirect execution without compute shaders, batched indirect draws with the same strides/offsets; rdbtCVS checked fmt/clippy on third_party/wgpu-hal by hand and approved. CI green.
🤖 Agent: HDR tiers for lightning, beacons, crystals, fireworks, orbs; marks follow their draws on classic and OIT paths; checked in game by the author; rdbtCVS approved. CI green, parity pending.
🤖 Agent: read the diff: video modes are listed once per monitor instead of at every slider step (about 28 ms each on Windows); with_option keeps the cached list. Unit test; approved by rdbtCVS; CI green, parity runs on the head merged into main.
The title shown while resources load takes Tab, the arrows and Enter the
same way and hands its focus to the game's.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Screen.keyPressed's focus navigation on the targets the controller's focus
mode already reaches: Tab and Shift+Tab in tab order (reading order),
arrows with the ported nextFocusPathInDirection step, Enter, Space and
keypad Enter press the focused widget. Sliders follow canChangeValue,
text fields keep their caret keys, a list is one Tab stop and the world
list selects its focused row. The focused widget is the narration's
focused one, the keyboard's own pointer events don't reach the narrator,
and the real mouse takes the pointer back where it left it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The control socket's key and type commands skip the window's events, so
the open screen never heard them (afterKeyboardAction): the chat's typed
message wasn't narrated in scripted runs. They now go through the same
narrate_key as the window's keys.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The shared button drawer takes the narration a button records
(ui::button_as), so the game rule switches say their name, value,
description and default as BooleanRuleEntry's custom narration does,
and the command block's output, mode, condition and trigger buttons say
"<name>: <value>" and the value a click moves to. The rules list is a
ContainerObjectSelectionList: its rows are narrated with their place,
the category headers by their label, and the rule tooltips the list
draws are no hint of the rows' widgets. The command box and the search
box are numbered where vanilla adds them.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Options pages narrate their list as a ContainerObjectSelectionList: the
row under the mouse or focused, its position, the widget in it and its
header rows; cycle buttons the value a click moves to, key binds
narrator.controls.*, tooltips as hints. The language and pack lists are
object lists, the pack screen and its confirm screen name themselves, and
dialogs narrate their title, warning button, checkboxes, cycle buttons,
edit boxes and button tooltips.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
docs/narrator.md: the statuses and hotkey, the voices per platform (flite
needs libflite on Linux, as vanilla), where the code lives, how a screen
records its narration, silent scripted runs with horizon-ctl narration,
and the known differences from vanilla. Linked from the docs index.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
An EditBox that can't be edited isn't active, so World Options' port box
is not counted while LAN is off (updatePortControlsState). The recipe
book's button moves when the book opens; a stable key keeps the focus
the click gave it, so it is narrated as vanilla's focused ImageButton.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The book is one narratable, hovered whenever it is visible, that narrates
the recipe button, search box, filter or tab findNarratableWidget picks
one level in, with no position of its own. The book button before it is
an ImageButton without a message.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
World Options' switches are CycleButtons (their value and the one a
click moves to), the difficulty lock a LockIconButton ("Difficulty lock
button. Locked/Unlocked"), the category headers FocusableTextWidgets,
and the port box is numbered after the LAN switch where it is added.
The statistics' general and mob lists are ObjectSelectionLists naming
the hovered row, the item list a ContainerObjectSelectionList of sort
buttons and items, as StatsScreen's are.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The world list, the server list with its LAN header and games, and the
superflat layer and preset lists narrate their rows as their entries'
getNarration; the create-world tabs as a TabNavigationBar, its cycle
buttons and switches with the value a click moves to, and the edit boxes
by their messages. The title screen's and pause menu's icon buttons say
their own messages and tooltips in the order the screens add them, and
the multiplayer screens name themselves (joinForNarration for the confirm
and disconnected ones). A world search narrates the filtered list at once.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Screen.shouldRunNarration is GameNarrator.isActive: with the narrator off
its narration still runs and is marked said, so the click that turns it on
in the options says only what it changed 200 ms later. Widgets can now
record text widgets, cycle buttons with their own narration, and lists'
rows by name; a shared drawer's record can be named or moved after the
fact, and the drawers hear which widget the narrator has focused.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
property_setters_refuse_no_change searches native_hooks.rs, read with
include_str!, for function heads written with "\n". Git for Windows'
default core.autocrlf checks .rs files out with CRLF, so the head is
never found and the test fails there. Read the source with its line ends
normalized, as the launcher's mojang_downloads_only_for_an_owner already
does (9cdfc5bd4). The test still finds the same code on an LF checkout.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Stats, advancements, world options (its lock popup, restrictions and game
rules), Play with Friends, server links (with the warning button and its
tooltip), the link and custom screen confirms, the pack prompt and the
code of conduct (title and text, its checkbox) name themselves as their
vanilla screens' getNarrationMessage does; the game mode switcher has no
title. The stats tabs narrate as a TabNavigationBar. The command block
screen has no title and takes the suggestions' usage while they show;
its boxes and the rules' and port's boxes carry their EditBox messages.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The chat screen says only its title, how to send (with the restricted
note) and the message typed, as ChatScreen.updateNarrationState does, and
its link prompt joins the question and the link as ConfirmLinkScreen.
The level loading screens (the startup menu's, single player's start, a
world opening, a server's terrain) have no title and no navigation and
say the server's progress every 2 s; the connect screen has no title;
the saving screen says its title alone. The win screen narrates its
narrator components, the poem's lines and the credits' names.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A container screen narrates its menu's title (the creative inventory's
is empty, a merchant's is its name without the level); slots are never
narrated. The book screens say their title and page indicator (the view
also the page's text), their page box or title box and page buttons; the
sign editor its title and Done. The creative, recipe book and anvil
boxes carry their EditBox messages.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The window's input reaches the screen's narration before the screen takes
it, as MouseHandler and KeyboardHandler call afterMouseMove,
afterMouseAction and afterKeyboardAction; a controller's focus step
counts as an arrow key. The startup menu narrates its screens and hands
its narration to the game with the title, and Ctrl+B schedules the open
screen's narration (updateNarratorStatus). The title screen's modes, each
options page, the pause menu and the death screen name themselves as
vanilla's screens do, joinForNarration for the confirm and alert ones.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
ScreenNarrator reads each drawn frame's screen and narratables: a new
screen narrates at once (init's triggerImmediateNarration), then stays
quiet for 2 s, or until the next key after a keyboard; mouse moves,
clicks and keys schedule the narration that Gui.update runs when due.
updateNarrationState picks the widget as findNarratableWidget does, with
its position and usage, lists' rows and container lists' row widgets, and
a click focuses what it hits as mouseClicked does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Drawers say which screen they draw (begin_screen: its id, title and
narration overrides) and the buttons, sliders and edit boxes they draw
record what they narrate, as Screen.updateNarrationState reads its
NarratableEntries. Nothing is recorded or built unless the narrator asks.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
NSSpeechSynthesizer posts to the run loop of the thread that speaks; the
narrator's speech thread now runs its own each time it idles, and idles
every 50 ms however many messages arrive in between.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The narrator speaks through AppKit's NSSpeechSynthesizer with vanilla's own
queue: a message waits while another is said, an interrupt stops the speech
but keeps the queue, and clear() empties it. The speech thread runs no run
loop for the delegate, so idle() polls isSpeaking and runs vanilla's
didFinishSpeaking when the speech has ended.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A read-only command beside stats and messages: the Narrator option's
status, whether text-to-speech is there, and the last 32 messages the
narrator was asked to say, so scripts can check narration without
listening (HORIZON_NO_AUDIO gives an active, silent narrator).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The narrator loads libflite and libflite_cmu_us_kal16 at run time as JNA
does, splits each message at punctuation and plays the pieces on its own
"Narrator #0" thread; clear() bumps the batch so queued pieces are skipped
while the one playing finishes, as vanilla's executor does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
ChatListener: player and disguised chat are said once they show, Chat
Delay included, with the chat type's narration decoration ("Steve says
hi", from the server's chat_type registry, chat.type.text.narrate when
it lacks the type); system chat as saySystemChatQueued when the Chat
option lets it through; the action bar's system messages as
saySystemQueued. Single player and LAN take the same path: the window
gets them from its server thread as packets.
Also: the jukebox's Now Playing and the creative hotbar's saved message
(saySystemNow), the player list's names as it comes up, ConnectScreen's
"Joining the world" every 2 s and LevelLoadingScreen's "Ready to play"
as it closes, a LAN world found in the server list's view, the selected
command suggestion (once per entry, at once when Tab asked), debug chat
(showDebugChat) and the world options' publish messages. Leaving a
server and a reconfiguration clear the narrator (Minecraft.disconnect,
clearClientLevel). Only the window's Game narrates, never a server
thread's in the same process.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The comment claimed a knot is a block under the crosshair rather than an
entity id. On a server it is both: `knot_ids` (`game/lan_entities.rs`)
gives every fence knot a `LeashFenceKnotEntity` id, so a guest's Interact
naming one arrives with an id that `interact_knot` cannot use, because it
starts from the block pick.
The comment now says that instead of stating a false reason. No behaviour
changes: knot ids were not resolved before this PR either.
rdbtCVS's open nit at 44c54ba8b. Also rewords the same claim in the body.
The previous merge point (`06485bfef`) is ten commits behind main, and it
still has what `825a6bbcf` fixed: two `S::Attack` arms in `Hub::process_inbox`,
the second unreachable, which is what makes `clippy -D warnings` fail on this
branch as it did on main. Merging current main takes that fix.
No conflicts, and no file this PR changes is resolved by hand here.
BlockAttachedEntity.ignoreExplosion: the explosion's direct source entity
in water makes the entity ignore it, before the mobGriefing rule. Primed TNT,
TNT minecarts, creepers, sulfur cubes, withers, fireballs, wither skulls and
end crystals report whether they are in water. An armour stand is not a
BlockAttachedEntity and keeps its own rule.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
native_block_entity_submit runs once a frame for each mod block entity in
view, as vanilla's block entity renderers do (64 blocks, frustum, chunk
fade-in), with a pose stack and helpers that add to the entity batch:
quads, model-element boxes baked as FaceBakery does, and block models
(submitBlockModel). Content mods' textures/entity join the atlas.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A press or repeat of B with the edit modifier (hasControlDownWithQuirk)
moves the narrator to its next status when the narrator is there, the
Narrator Hotkey option is on and no focused edit box takes the keys (chat,
command block, anvil name, world options' boxes, a dialog's text input,
the options' searches and the title screens' fields). The change is
announced with its toast and saved, on the startup menu, the title, the
menus and in game, and the key still reaches the screen.
on_narrator_status_changed marks where vanilla calls the screen's
updateNarratorStatus, for screen narration later.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
horizon-narrator ports vanilla's narrator core:
- GameNarrator's status rules (NarratorStatus, queued vs. interrupting
messages, the status announcement and its toast keys)
- com.mojang.text2speech's NarratorWindows on SAPI's SpVoice, on a speech
thread of its own so the game never waits on it
- ScreenNarrationCollector and Screen's narration timing (delays, the
suppression after init, findNarratableWidget)
HORIZON_NO_AUDIO gives an active but silent narrator for scripted runs.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
updateNarratorStatus's SystemToast (NARRATOR_TOGGLE) goes up with
"Narrator Enabled" and the status, "Narrator Disabled", or "Not Available"
when there is no narrator. As SystemToast.addOrUpdate, a toast still up
takes the new text and its display time starts again instead of a second
one stacking. It takes a free ToastManager slot beside the advancement,
tutorial and pack toasts and draws over every screen: the title (its own
toast keeping the first slot), the options, loading and in game, F1 hiding
it. The startup menu has its own until it hands the title to the game.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The Narrator button cycles Off, Narrates All, Narrates Chat and Narrates
System, reads "Not Available" and is inactive while the narrator isn't
(Options.narrator's caption, OptionsSubScreen.addContents), and a change
runs GameNarrator.updateNarratorStatus: the speech is cleared and
"Narrator : <status>" said. Chat Settings shows it where vanilla's
ChatOptionsScreen does. The Narrator Hotkey switch is a plain toggle with
its macOS tooltip there. options-set narrator announces a new status too.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The client opens horizon-narrator where Minecraft's constructor makes its
GameNarrator (never for a server), runs checkStatus (logged: Horizon has no
native message box), and destroys it when the window's event loop ends.
Whenever options load or change, on the title's startup menu and in game,
the narrator takes the narrator option's status and the voice volume
(getFinalSoundSourceVolume(VOICE): master times voice).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A content block's states were exactly its template's, copied row for row, and
`default_state` could only overlay values on properties the template already
had. So a pipe could not have a connection per side, and a machine could not
have `facing` and `lit`.
A block file may now declare `state_properties` — a boolean, or an enum of
values — and the block's states become its template's states times their
combinations, appended after the template's run so a block's states stay one
contiguous run in registry order.
The combination order puts every property at its declared default first, and
`Rename` carries how many states a block gives each of its template's, so a
template state still moves to exactly one state: the block's default. The
`facts.json` defaults and `render-definitions.json` mappings are unchanged.
Refused: a name its template already has, a bad name or value, a
`default_state` value the property does not have, and more combinations than
a world holds (2^20) — the last before the product is built.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
players, players/kick, server/{status,save,stop,system_message} and the four
lists' getters (allowlist, bans, ip_bans, operators), as vanilla's
PlayerService, ServerStateService and the list services answer them: players as
PlayerDtos, bans with their source, reason and ISO-8601 end, operators with
their level, the status's empty player list left out as vanilla's codec leaves
out its default. Kicks, saves and system messages go through the game's
console (tellraw, title actionbar for an overlay). server/status answers while
the game restarts, started false.
The parity harness now gives both servers the same lists and checks these too
(32 exchanges), the notifications vanilla sends meanwhile set aside.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
management-server-* in server.properties ([properties] in server.toml) turns on
vanilla 26.3's management protocol, served by the supervisor so it answers while
the game restarts. As net.minecraft.server.jsonrpc does: the secret as a bearer
token or in the minecraft-v1 subprotocol from an allowed origin (401 bodies
included), JSON-RPC errors and batches as Connection handles them, by-name and
positional parameters as IncomingRpcMethod checks them, a generated secret
written back, and a wrong secret or TLS setting stopping the start.
This first part: the WebSocket (std only, websocket.rs) and rpc.discover
(vanilla's OpenRPC document, management_schema.rs). The other methods answer an
internal error saying they aren't in Horizon yet; they come next.
tools/test_management_protocol_parity.py sends the vanilla server and Horizon's
the same messages and handshakes and compares the answers, and rpc.discover with
the data generator's json-rpc-api-schema.json.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Main's DebugEntryChunkGeneration line (#4389) was drawn inline in
draw_debug_overlay_under, which this branch replaced by debug_entry_lines:
it is now that function's chunk_generation_stats arm, unchanged.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Main moved the launcher's completion checks into horizon_prepare::missing_extras, which
--check and COMPLETIONS share, so lacks_panorama and its test live there now.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Main added Japanese and Portuguese (Brazil) since this PR's keys were
written; the i18n test wants every key in every language.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
#5149's handle_chat_message_error had no tracker to tell. With one, a signed message that
failed validation never reached it, and the client's last-seen window drifted from the
server's. It now marks the signature as ChatListener.handleChatMessageError does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
`ChatScreen.handleChatInput` and `ChatScreen.normalizeChatMessage` had no
verdict at all: the ledger reads them as unported rather than as the port they
are. Both point at what the message becomes on the way to the server now.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
A chat line went out as typed: Horizon sent the raw box text, where vanilla
sends `ChatScreen.normalizeChatMessage(s)` — `String.trim`, then
`StringUtils.normalizeSpace` (one space for every run of whitespace), then
`StringUtil.trimChatMessage` (the first 256 UTF-16 code units, the length
`ServerboundChatPacket`'s `stringUtf8(256)` decodes).
Two player-visible differences:
- `send_chat` cut the message with `chars().take(256)`, so 256 emoji passed as
512 UTF-16 units and the server refused the packet; a cut at 256 units is
what vanilla does (129 emoji become 128).
- runs of spaces stayed: `hello there` reached other players with its three
spaces, and a line of nothing but spaces was sent, where vanilla sends
nothing and remembers nothing in the history.
The normalization sits in `Game::handle_chat_input` (ChatScreen.handleChatInput),
so the chat box, the chat drafts and the control socket's `chat` all send what
a vanilla chat screen would. `java_whitespace` is reused from `ui.rs`, so
U+00A0 stays (Java whitespace excludes the no-break spaces) and U+3000 and
Java's U+001C..U+001F separators collapse.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
`EntityType.updateCustomEntityTag` was ported for a player's egg
(`block_entity_item::spawn_egg_entity_data` then
`creatures.rs::Game::apply_spawn_egg_entity_data`) but not for the
dispenser's, whose `dispense_spawn_egg` stops after
`entity_nbt::apply_item_components`: a configured egg
(`/give @s pig_spawn_egg[minecraft:entity_data={…}]`) dispensed onto a block
spawned an unconfigured mob, which is the half of the
`EntityType.updateCustomEntityTag` row that pointed at
`game/dispenser.rs`.
`apply_spawn_egg_entity_data` becomes `pub(super)` and the dispenser calls it
after the components, as the player's path does.
`tools/port_map/game-005.tsv`: that row is `port` now.
Co-Authored-By: DeepSeek Harness (deepseek-v4.1-flash) <[email protected]>
The anticheat's Attack arm and #5123's arm that keeps moves queued behind
an attack both landed in process_inbox; the second was unreachable, so
clippy -D warnings failed on main and #5123's behaviour was lost. The
anticheat arm now also holds the queued move when it forwards the attack.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
f2a99d623 split verify into python, clippy and tests jobs plus a gathering
verify: four GitHub-hosted runners per checks run instead of one. With every
pull request push and every merge to main starting checks, that filled the
organisation's shared runner pool, and every workflow's jobs (labels, overlap
notes, other pull requests' checks) queued behind it for tens of minutes.
verify runs the public-tree check, the Python tests, rustfmt, clippy and the
Rust tests one after another again, on one runner. A pull request's
public-tree check still reads only its own commits (--range base..head).
Closes#5302
Co-Authored-By: Claude Opus 5.5 <[email protected]>
naga's GLSL frontend refuses three things GL drivers take, and the program was skipped:
- a swizzle of a scalar (GLSL 4.20: `noise(p).x`, `s.xx`; Kappa's weather map);
- a matrix column's component written to (`uv[1].y = ...`);
- `min`, `max` or `clamp` of a `uint` and integer literals (`max(n, 1)`): naga finds no best
overload, GL takes the uint one.
compile now rewrites the spot naga points at and parses again (at most 256 times): one swizzled
component is dropped, more become a vector constructor of the scalar, a column's component
becomes an index, and the call's integer literals get a `u`. Error locations map through the
rewritten text, which keeps its lines.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: merged main. Main's 45c36d872 had added the client side of the same rule (RiddenBoat in motion/fluids.rs), so I kept main's type and logic and this PR now fills it for served players (lan_actions.rs → boats.rs riding_boat_fluid_box), where the air supply and drowning are decided; port map row updated. cargo check -p horizon-client --bins passes, fmt clean. Approved by rdbtCVS.
horizon-mods gained the zip crate, but the registry's separate lockfile
wasn't updated, so Railway's `cargo build --locked` refused to update it
and every deploy of horizon-mod-registry failed (one failure mail per
merge to main). Minimal update: only the new packages are added.
Checked: the Dockerfile's exact build (same files copied, --release
--locked) finishes.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: merged main into the reviewed head (the earlier clippy/tests failures were main's Audio build break, fixed by #5251); no other change. read the diff: on focus loss a held (non-toggle) sneak/sprint/attack/use mapping is released before the re-sync, toggles keep their latch (ToggleKeyMapping.setDown/release); each frame a Shift/Control the keyboard's modifier state reports up clears its stale physical key and held latch. CI failures were main's Audio build break (#5251).
🤖 Agent: merged main (clean) on top of Domanffe's 2b2c2893e, which fixes dimaniojk's two findings (drag off the strip, the CLI's hue); earlier CI failures were main's broken Audio initializer. Launcher-only; cargo fmt clean. Approved by rdbtCVS.
🤖 Agent: F3 Density line as NoiseBasedChunkGenerator.addDebugScreenInfo under the chunk_generation_stats entry, sampled on a new feet block from the server thread's generator; main merged (conflicts with the per-entry F3 and explorer maps resolved, both kept), horizon-client checks clean; rdbtCVS approved.
🤖 Agent: Shader API: the material stage gets opaque_depth/opaque_color copies taken before translucent terrain and MaterialInput.screen (fragment position over resolution, as Teiwazik showed); rdbtCVS approved after merging #4186's tags; main merged. Only native shader packs see it.
🤖 Agent: merged main into the branch (main's broken Audio initializer made the earlier CI fail, fixed on main by f80c0360d); cargo check -p horizon-client-render passes. Change read: two adapter passes, GPUs first, software only with a warning. Approved by rdbtCVS.
🤖 Agent: functions tick before the levels (MinecraftServer.tickChildren) and an item added this tick first ticks on the next (EntityTickList), so a fresh drop reports Age 0; the freeze dimaniojk found is fixed (only items with a later id are skipped); type_specific/player checks ServerPlayer and input.sneak on the server. rdbtCVS approved; main merged.
The lid animated but the collision queries always answered "no box": a
closed box collided as its plain block and an opening one never grew. The
collision view now reads the boxes' lid progress, the level ticks those
lids on its server pass (not only the client-only animation), and an
opening lid applies moveCollidedEntities to the bodies the game moves.
Vanilla: ShulkerBoxBlock.getCollisionShape / getShape (isClosed),
ShulkerBoxBlockEntity.getBoundingBox / getProgressDeltaAabb /
moveCollidedEntities.
Co-Authored-By: opencode (DeepSeek V4.1 Flash) <[email protected]>
🤖 Agent: merged main (port map: main's suspicious stew rows kept, this PR's variant rows). Mooshroom variant component as MushroomCow.get; test fails without it.
🤖 Agent: read the diff: placement checked before the spawn-list lookup (both pure reads, no RNG), species placement cached in MobData, LocalMobCapCalculator's lists flattened; draws and positions unchanged; A/B on playhorizon −18 % spawn_natural. CI parity runs on the head merged into main.
🤖 Agent: perf_bench --table prints the missing-value marker when a report has no scene-load measurement, keeps a measured 0.0 and the mean of measured scenes; regression test covers all three. Tooling only; rdbtCVS approved; CI green. Marked ready from draft.
🤖 Agent: verified: horizon-launcher completions <shell> via clap_complete over the existing command tree, no network or file writes, test per shell; approved as code owner. Thanks!
🤖 Agent: verified: vanilla's ENTITY_SHADOW / OIT_ENTITY_SHADOW pipelines take RenderPipeline.Builder's default cull(true); the shadow quads wind +Y (test added), so the underside is culled. Thanks!
🤖 Agent: verified against NbtIo.readUnnamedTag (StringTag.skipString): a file root's name that isn't modified UTF-8 reads as empty; keys and truncated names still fail. Proof: CI parity. Thanks!
Merges origin/main, moves the living_incoming_damage hook to
tools/native-hook-fragments (main's API stays as is), and drops the attackers
a canceled hit noted, so the next hit of that damage type isn't credited to
them (arismoko's review).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
nbt.rs: the Writer keeps both flags, this branch's empty_long_strings
and main's canonical_nan (both on for the disk writer, off for the
network one).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
rdbtCVS's two review points on #4250:
- The launcher's own status workers (off the UI thread, showing "Waiting for
account access…") take the lock without the 3 s bound, so a worker queued
behind another account's renewal no longer shows a false "Could not check
sign-in". account list and MCP keep the bounded wait.
- refresh_account saves a renewed profile (new name or skin) to accounts.json
as launch() does, and the UI reloads accounts and profile heads when a
status arrives and the saved list changed.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Q over a slot and Q in play decided whether to throw the whole stack from
physical_keys, while every other screen check reads Control from the OS
modifier word (has_control_down). The two could disagree after a lost or
lagging key event. drop_whole_stack now takes the modifier word; a latched
sprint toggle still never counts as Control.
The control socket's `release` cleared keys, physical keys and modifiers
but not the sneak/sprint toggle mappings, so a sneak pressed through
`key shift down` came back on the next sync_toggle. It now releases both
mappings as their key's release does (a toggled one stays latched) unless a
controller still holds the action, then resyncs the held keys.
Also:
- `key` reports an unknown action, with its name, from the single early
check; the later match arm that repeated it could no longer run.
- lan_e2e check_recipe_shift reads the recipe buttons before the tab loop,
so a book without a 35x27 tab fails on its assertion and target dump
instead of a NameError.
Validation on macOS arm64:
- cargo test --locked -p horizon-client --bin horizon-client: 359 passed,
3 ignored (including stack_drop_requires_ctrl_even_with_sprint_latched,
now driven through modifiers::set_key).
- cargo clippy --locked -p horizon-client --lib --bins --tests
-- -D warnings, cargo fmt --all --check, lan_e2e.py syntax and
tools/check_public_tree.py: passed.
Still to run with prepared resources: lan_e2e --check recipe-shift
--check modifier-release, and Ctrl+Q in an inventory and in play.
Refs #4166
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Screens now read Shift from the OS modifier word (has_shift_down), but two
synthesized shift-clicks still only inserted ShiftLeft into physical_keys:
the controller's Y quick-move in an inventory (pad_menu.rs) and horizon-ctl's
`station ... shift` click (station_control.rs). Both became plain pickups.
They now go through modifiers::hold/restore, which sets the SHIFT bit with
the physical key and afterwards puts back exactly what was held before:
a Right Shift already down, or an OS word set without a physical key, stays.
The control socket's `key <mod> up` returned before releasing the modifier
when a server dialog or the sign editor was open, or with no world, leaving
SHIFT set so every later inventory click became a QUICK_MOVE. The release
that already ran after `tap` whatever the handler did now runs after `up`
too (released_modifier), so no early return can keep a modifier down.
Tests: modifiers::a_held_click_restores_the_previous_state and
control::taps_and_ups_always_release_their_key.
Validation on macOS arm64:
- cargo test --locked -p horizon-client-input --lib: 64 passed.
- cargo test --locked -p horizon-client --bin horizon-client control::tests:
passed.
- cargo clippy --locked -p horizon-client-input -p horizon-client --lib
--bins --tests -- -D warnings, cargo fmt --all and
tools/check_public_tree.py: passed.
Still to run on a machine with prepared resources: controller Y on a slot
in the survival and creative inventories, and lan_e2e --check
modifier-release --check clicks.
Refs #4166
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Read Shift and Control for screens from the OS modifier word instead of
inferring them from cached physical key events. This follows Minecraft
Java 26.3 InputWithModifiers.hasShiftDown/hasControlDown and prevents a
missed key release from turning subsequent inventory clicks into transfers.
Reconcile lost releases only for physical modifiers that were recorded as
held and are still assigned to the affected sneak/sprint mapping. Preserve
rebound non-modifier keys, controller-held actions and keyboard/controller
toggles. Include controller releases delayed until the next simulation tick.
This avoids cancelling crouch bound to C when the player presses Alt, or
cancelling a controller hold because the keyboard does not hold Shift.
Keep synthesized modifier presses and releases consistent with the OS word.
Releasing one side preserves the bit while the other side is still held.
Validate control key actions before mutating state. Run tap cleanup after
the command handler even if it returns early or fails, and clear the modifier
word on the control socket's release command. This prevents key shift tap
or key shift down followed by release from leaving later clicks modified.
Add five input tests covering modifier press/release, left/right holds,
rebindings, missing physical keys, controller holds and controller toggles.
Add the modifier-release LAN check: after a Shift tap or release, a normal
inventory click must pick up ten diamonds and closing must preserve the
server's count. Remove the obsolete pad-toggle getters introduced by the
previous broad reconciliation, since ToggleKeys already handles toggles.
Validation on macOS arm64 on main 9069aea3e:
- cargo test --locked -p horizon-client-input --lib: 63 passed.
- cargo clippy --locked -p horizon-client-input --all-targets -- -D warnings:
passed.
- cargo clippy --locked -p horizon-client --lib --bins --tests -- -D warnings
-A unreachable-patterns: passed. The allowed lint is main's duplicate
S::Attack arm in lan/hub.rs, a file this branch does not touch.
- cargo fmt --all --check, Python harness syntax, git diff --check and
tools/check_public_tree.py: passed.
Prepared Minecraft resources are absent here. The new modifier-release
check and final game-level verification remain pending on the dedicated
machine; prior LAN results must not be treated as validation of this change.
Rebuild binaries from this commit, then run:
python3 tools/lan_e2e.py --check modifier-release --check clicks \
--check shift-double-click --check recipe-shift --check shift-drag \
--check crafting --check wheel --out .local/validation-4166-review
Also rerun pr_ready.py with the appropriate main remote as its base.
Manual checks still required: hold rebound crouch/sprint while pressing
and releasing Alt; repeat with controller holds and toggle modes; confirm
lost Shift releases still recover with the default sneak binding. Verify
Shift taps and release before normal inventory clicks. Do not use the
deprecated in-frame path; keep game, launcher and XDG homes in fresh
scratch directories as required by AGENTS.md.
Refs #4166
Co-Authored-By: OpenAI Codex gpt-6.1-sol <[email protected]>
A carried stack already starts QUICK_CRAFT in Clicks::press, but
inventory_press only initialized its hovered-slot collection without Shift.
Holding Shift therefore prevented the distribution gesture from completing.
Start the drag with either modifier state. Keep PICKUP_ALL restricted to
non-Shift double-clicks, and leave the Shift drag initially empty so a
press/release without any dragged slot retains its QUICK_MOVE behavior.
Follow Minecraft Java 26.3 AbstractContainerScreen.mouseClicked,
mouseDragged and mouseReleased; use the existing menu and server click path.
Add three LAN checks driven by real pointer events, rather than click-slot:
- shift-drag: spread 20 diamonds across three crafting slots as 6/6/6 with
two carried, or 1/1/1 with seventeen carried, with and without Shift.
- shift-double-click: transfer chest stacks of 8, 13 and 21 diamonds while
keeping a carried stack of 32; spaced clicks transfer only the first pile.
- recipe-shift: place two diamond-block recipe sets from 20 diamonds using
the recipe book, then Shift-click the result to produce two blocks and
retain two diamonds.
The latter two behaviors already worked and receive regression coverage.
Resolve pointer targets from the current UI and include movement inside the
starting slot so the drag actually traverses all selected slots. Check
carried stacks, grid contents and server inventory counts after closure.
Evidence reported by the dedicated-machine agent before this rebase:
Shift drags failed on the old base and passed after the fix; clicks,
shift-double-click, recipe-shift, shift-drag, crafting and wheel passed.
The container_screen, menu_click, place_recipe, recipe_book and
recipe_book_view differential harnesses were also reported passing.
Vanilla expectations came from the pinned sources and Java oracles; the
vanilla graphical client was not exercised. Those game runs cannot be
repeated on this checkout because its prepared resources are absent.
Refs #4166
Co-Authored-By: OpenAI Codex gpt-6.1-sol <[email protected]>
Conflicts: datapack.rs takes both sides' kinds (trade_set,
villager_trade and dialog); game.rs saves through main's
write_level_for_role, then each placed player to their own file;
lan.rs syncs dialog ids, then sends the world's own biome table.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
PiglinBrute.wantsToPickUp takes only a golden axe, through Mob.aiStep's
looting loop. StartAdmiringItemIfSeen sets ADMIRING_ITEM (119 ticks) for a
loved item in sight, and StopAdmiringIfItemTooFarAway(9) and
StopAdmiringIfTiredOfTryingToReachItem(200, 200) end it. A blow from a
player disables admiring for 400 ticks (ADMIRING_DISABLED).
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Only the shadow passes write the shadow targets, so while the key matches they
still hold the map (and the A/B side shadowcomp left current): no cache
textures, no copies, and no copy into shadowcolor buffers made without
COPY_SRC/COPY_DST (a wgpu validation panic with any pack using shadowcolor).
pass-timing leaves this PR (unrelated).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: read the diff at 7fc284b (rdbtCVS's main merges + Escape-as-Cancel fix): the splitscreen profile page as a model with its store and tests in horizon-client-render, not wired into play yet; dimaniojk's paging finding fixed. Part of the splitscreen series whose earlier parts merged.
🤖 Agent: merged main into the reviewed head (the earlier clippy/tests failures were main's Audio build break, fixed by #5251); no other change. read the diff: an items-only content mod now writes the identity marker, and the identity hashes the mods' item and mob effect entries per registry; a root without mod entries keeps its block-only hash (unit test). Fails safe with the content-mods mismatch message across versions.
🤖 Agent: merged main into the reviewed head (the earlier clippy/tests failures were main's Audio build break, fixed by #5251); no other change. two spans (spawn_natural, tick_creature_entities) inside tick_creatures, timed only when spans are on; no behaviour change. CI failures were main's Audio build break (#5251).
🤖 Agent: merged main into the reviewed head (the earlier clippy/tests failures were main's Audio build break, fixed by #5251); no other change. read the diff against vanilla: spear is in #minecraft:no_knockback, so LivingEntity.hurtServer skips dealDefaultKnockback for a stab; Damage::no_knockback now lists Spear; unit tests. CI failures were main's Audio build break (#5251).
🤖 Agent: merged main (both sides added tests at the end of mod_content_tests.rs: kept both); cargo check of the lib and tests passes; review points fixed by the author (02a7c1553 cherry-picked); marked ready
Launcher: Language::Turkish (code tr, Türkçe, picked from a tr-* system
locale), tr.json brought up to main's strings (25 native-mods strings
translated, 5 the launcher no longer asks for dropped, one {version}
placeholder fixed). Game: tr_tr.json joins the bundled Horizon strings.
Built on #5035's Japanese wiring so the two don't conflict.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
🤖 Agent: finished by me (94027c278): merged main, API number as a fragment, the canceled-join crash fixed (just_spawned_item), and a same-level End gateway trip no longer joins again. native_ tests pass.
🤖 Agent: the review follow-ups from #3996–#4289 (screen only for the block's own menu, entity menus clear the block target); review points answered; marked ready
🤖 Agent: read the diff: Superflat layer and preset icons draw the item's GUI model (as CreateFlatWorldScreen's renderItem), clipped exactly to the list; other screens unchanged. Approved by rdbtCVS; CI green, parity runs on the head merged into main.
🤖 Agent: one partner_hidden setting hides the sidebar card and the Servers banner; the 3-minute seen gate now counts on every page. rdbtCVS approved; launcher CI green.
🤖 Agent: rename-aside Windows self-update with rollback replaces self_replace's self-deleting copy that Defender flags; start-up clean-up removes only this exe's .old/.update leftovers. Read in full; rdbtCVS approved; CI's launcher job builds Windows. A real Windows update after release is the remaining check.
startRiding(vehicle, true, true) skips canAddPassenger, so a boat under
water still takes the rider; Horizon's boats seat at most max_passengers,
so a full one still refuses. Horizon's minecarts carry no mobs yet, said
in a comment. A player put in a boat or minecart now fires
minecraft:started_riding (sendEventAndTriggers).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
#5235 keeps the address in horizon_telemetry::server_status::OFFICIAL_ADDRESS:
it is now play.playhorizon.dev, with OFFICIAL_PREVIOUS beside it; the
launcher's defaults and the game use both. The launcher's official player
count pings the address from links.json through its SRV record
(horizon_protocol::server_redirect + server_status::players_at): port 25565
on that host is another server. The telemetry server, which looks no SRV
up, defaults to the SRV target (OFFICIAL_STATUS_ADDRESS). The game moves
servers.dat entries for an earlier address once, like the launcher.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
pushAndPickupEntities only runs for a rideable cart moving at
horizontalDistanceSqr >= 0.01; an activator rail ejects mob passengers
too; a player can't board a cart a mob already rides (isVehicle).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A failed signature and failed OS randomness are logged instead of hidden, and the
ChatListener.flushQueue ledger row points at flush_chat_queue.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- audio.rs: main's shared-output fields (the device tracker's, f80c0360d).
- A guest in Window 1 leaves its F11 to P1: the game's window follows P1's option.
- disable() shows Window 1 again and says so (window1_shown), so a later
split into Window 2 hides it again.
- Clicks, key presses and the wheel from a window the focused player isn't
in no longer reach that player; releases still pass, so no key sticks.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
#4935 landed as API 45, and API numbers now come from
tools/native-hook-fragments (stamp_native_hooks.py after the merge): the
hand-written API 44 and its docs paragraph become
native-block-entity-sync.json. lan_bot keeps both new tests.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Review fixes on the preparing screen:
- The worldgen data was read from the bundle alone, so the packs the Data
Packs and Experiments screens picked never reached the generator (the
cached vanilla `GeneratorData` came back). The read goes through
`shared_data_with` now, with the overrides the chosen packs build
(`creation_worldgen_overrides`, the worldgen companion of
`applyNewPackConfig`'s reload), and a failure reports on the title's
error screen instead of only logging — the Create screen's options stay
(`failed` keeps the mode as the back target).
- Escape and text input are ignored while `preparing` is up
(`GenericMessageScreen.shouldCloseOnEsc` is false): Escape during the
read used to leave Create and `finish_preparing` then replaced it with
a fresh screen, losing the player's options.
- The confirm screen's Back/Escape set `reopen_data_packs`, but nothing
read it: the game side polls it now and reopens the Data Packs screen
with the selection it had (`onAbort`).
- The preparing assertion in the navigation test is unconditional, and
offer_data_packs (with and without a feature pack), the confirm
screen's Proceed and Back, and preparing's input blocking have tests.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
The official server moves to play.playhorizon.dev (SRV to port 24454).
launcher/links.json now carries `official_server` and
`official_server_previous`, read through the existing Links mechanism with
the same built-in defaults, so a later move is a links.json edit, no release.
server_list::add_official offers the address from links, and when the
address differs from the one the marker file recorded, rewrites entries for
a previous address to the new one (player's name, picture and fields kept),
once. No duplicate when the new address is already listed; a player who
removed the server is not offered it again. The Home list re-runs it when
links.json arrives. The game's own built-in copy and docs/smp.md follow.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The page stays closed after its last mesh moves, so the frame's uploads can't
refill it before trim returns it (trim clears closing). Pages record whether
their buffer went to system memory: moves only go to video-memory pages
(alloc_moved), and sparsest empties a system-memory page first. Tests for both.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- The hook asks for its API number with tools/native-hook-fragments/entity_leave_level.json
(#5161); NATIVE_HOOKS_API, the histories and native-hooks.json are main's again.
- An End gateway keeps its traveller in the level (teleportSameDimension), so an item
going through one no longer reports leaving it (arismoko's and rdbtCVS's point).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
#4186's precipitation/temperature/downfall/wetness come first in Frame (they are on main);
this branch's frame-motion fields follow at 416..512. Reference regenerated.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Three things kept tools/lan_e2e.py (and agent_control_e2e.py) from running
on Windows:
- The prepared resources were linked into the scratch home with os.symlink,
which needs Developer Mode or an administrator (WinError 1314). The new
lab.link_dir falls back to a directory junction there, which needs
neither; shutil.rmtree removes a junction without touching its target.
- agent_control_e2e.game_env ran `vscreen up`, which doesn't exist on
Windows. Windows has no virtual screen, so it takes plain lab.game_env,
as macOS does.
- `horizon-ctl launch` never returned when its output was captured:
Windows hands a child every inheritable handle of its parent, so the game
kept the caller's stdout pipe open until it quit. horizon-ctl now clears
the inherit flag on its own standard handles before it starts the game
(horizon_threads::keep_stdio_from_children). Linux and macOS are
unchanged.
Fixes#5274
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Main's 45c36d872 added RiddenBoat for the client's own fluid box. The
served player's tick (lan_actions.rs) now fills the same type from the
ridden boat, so the server-side air supply follows it too.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- The hook asks for its API number with tools/native-hook-fragments/entity_join_level.json
(#5161); NATIVE_HOOKS_API, the histories and native-hooks.json are main's again.
- A mod canceling an item's join no longer crashes a throw: spawn_item's callers
touch the new item only if it joined (just_spawned_item), not whatever item is
last (throw_from_eyes unwrapped it; allays, shearing and /give's overflow
edited an unrelated item).
- An End gateway keeps its traveller in the same level (teleportSameDimension), so
the item doesn't join again on arrival (arismoko's review).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
tick_mobs is half of a crowded server's tick, and its two big parts
(NaturalSpawner and the entity ticks) move independently with the players'
spread. Two nested spans let the slow-tick log and horizon-report tell them
apart; with spans off they cost one relaxed load each.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The forged reply was rejected on its empty biome palette and out-of-range
spans before the cap; it now names one biome and keeps every span inside
HEIGHTS, and asserts the cap's error.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Vanilla's AbstractContainerMenu.clickMenuButton(Player, int) reaches the player's carried stack;
native_menu_button(pos, entity, id) does not, so a mod's widget outside the slots (a tank taking a
bucket) cannot take or give what the player carries. native_menu_button_v2 fires right after it with
`carried: &mut Option<ItemStack>`; a change is written back and sent to the player (SetCursorItem).
Single player's player is a guest of its own host, so one path covers it and LAN guests.
native_menu_button is unchanged (frozen). The API number comes from the fragment when stamped.
Closes#5246
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Wire ja.json into Language (code ja, name 日本語), translate the native
mods strings ja.json and pt-br.json lacked and drop the five strings the
launcher no longer asks for, so every language has every string again.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
lenientOptionalFieldOf("enabled_features", DEFAULT_FLAGS) omits the tag
when the set is vanilla's default, so a `["minecraft:vanilla"]`-only list
(a world created with no experiment picked) does not reach level.dat;
Horizon wrote it. The round-trip test the body promised is here too: a
picked experiment's list writes and reads back equal, a vanilla-only
selection writes no tag.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
for_recreate filtered the saved selection against the world's own packs
only, so a world with trade_rebalance on was re-created without it.
Vanilla's recreateWorldData keeps the DataConfiguration. rdbtCVS's review
point 3 on #4619.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Review fixes on top of the hue strip:
- A drag released outside the strip now writes settings.json too. on_mouse_up
only fires while the element is hovered, so the accent stayed in memory and
was lost on restart; on_mouse_up_out covers the release off it, and
hue_dragging keeps a click elsewhere on the page from saving.
- The contrast test sweeps every half degree from 0 to 359.5 on both bases
instead of two hand-picked hues, so the guarantee is tested where it is
claimed.
- `settings accent <value>` names the degrees a value may be, not only the
seven names.
Co-Authored-By: Hermes Agent (Nous Research) <[email protected]>
Resolve title.rs (Outcome::Connect gained `lan`), handle main's new
graphics_ui::Action::Calibrate in the wheel and key paths, and let
Escape close Video Settings on the first press.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
ExperimentsScreen's switches wrote `options.data_packs` on every flip, so
Cancel and Escape left the flips in place — vanilla's ExperimentsScreen
only flips its own list and writes the selection back in `onDone`. The
switch now flips the screen's flag only; Done rebuilds the selection the
way `onDone` does (the other packs keep their order, the selected
features are appended in reverse screen order), and Cancel drops it.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
main added three strings to every language (`finding_friend`,
`opening_world`, `cancelling_world`); es_es keeps its 198 translations
and gains those three, for 201/201 keys in en_us order.
Co-Authored-By: DeepSeek V4.1 Flash <[email protected]>
Handle root-array and modList metadata with synthetic ZIP regressions.
Preserve modern metadata precedence and existing fallback behavior.
Co-Authored-By: Codex <[email protected]>
Conflicts: main's lan login fallback (an auth failure only warns on LAN) kept
with this branch's keyed reasons; reconnect tests take both the lan flag and
the reason's key.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Ports the 26.3 DebugEntry* classes the debug options screen listed but
the overlay never drew: memory, targeted block/fluid/entity state and
tags, chunk render/source stats, entity and particle stats, player
speed, light levels, heightmaps, biome, local difficulty, day count,
spawn counts, sound mood and cache, post effects and GPU utilization,
in DebugScreenEntryList's order (by identifier) with vanilla's text and
priority/line/group placement. The integrated server reports the lines
that read the server level.
Renderers: water levels, heightmap, collision boxes, block and sky light
levels, solid faces. The list's scrollbar can be dragged, and
debug-profile.json gets vanilla's three datafixes by DataVersion.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
`x / 0.0` and `sqrt(-1.0)` are GL's infinity and NaN at run time (Mellow's vec4(-1.0 / 0.0),
Bloop's sqrt of negative constants). naga's frontend folds them into constants and refuses the
result, so the program was skipped. They now divide by `hz_zero`, a private float the frontend
can't fold: `x / hz_zero`, and `hz_zero / hz_zero` for the square root.
texelFetch1D/2D/3D (GL_EXT_gpu_shader4's names, Bliss) become texelFetch, as texture2D and
the other old names already do.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Conflicts in options_ui.rs (main's HDR calibration page beside the profile
page) and lib.rs's re-export list: kept both. Escape on the list now sets
choice = Some(None), as the Cancel button does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Two readers of the packs' tag files never saw any file:
- `PackOverlay::tag_kind` stripped `":{registry}/"` from a path that
`split_once(':')` had already freed of the colon, so it matched nothing
and every pack tag merged into `LootData` (item, enchantment,
instrument, entity_type, damage_type) was silently lost. It strips
`{registry}/` now, with the key shapes `content_id` writes fed through
it in a test.
- `load_tag_lists` (horizon-items) read the same key shape through
`overlay.kind("tags/{kind}")` and hit the same nothing: a pack's item
and block tags never applied. It reads `tag_kind(kind)` now.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
AABB.ofSize(center, size, size, size), not the box grown by its size on
every side, which let a buried orb jump three times as far as vanilla's.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Keep command signing disabled until the server command tree arrives, count chat length in UTF-16 units, and acknowledge delayed chat during configuration transitions.
Co-Authored-By: Codex <[email protected]>
network_stack (integrated-server and LAN inventory sync) encoded custom_name
and item_name from StackData.other but skipped minecraft:custom_data and
minecraft:custom_model_data, so item models' tints/selectors and client-side
custom_data reads saw nothing for a synced stack. Push both components, built
from the saved tag the same way vanilla's DataComponentPatch carries them;
item_sync's decode side already mapped both back.
Co-Authored-By: Claude Sonnet <[email protected]>
Vanilla's DownloadedPackSource puts active server packs in the selection
as required, fixed at the top (pack.source.server). Horizon already
applied them but never listed them; they also stay out of options.txt.
Fixes#5218
Co-Authored-By: Cursor Agent <[email protected]>
Ludovic (2026-10-31) asked to add x3 кто (Discord vladislav_hagen,
GitHub Qwerner88880, the native-mod hooks author).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A kill leaves the old id in the UUID map, so the first match could be an entity the level no longer holds. on owner follows getLivingEntity, which does not return a marker or a display.
Co-authored-by: Cursor <[email protected]>
Kept PadRouter's state_of with main's stick_axes, both Flags fields, and filled
main's audio device-tracking fields in Audio::share.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
NeutralMob.readPersistentAngerSaveData hands the saved anger back through
Mob.setTarget, and setTarget runs asValidTarget, which turns a target the
mob cannot attack into no target at all. IronGolem.canAttack refuses every
player when a player built the golem. Horizon's world load wrote the target
directly, so a player-built golem reloaded from a save kept its builder as
its target, and its melee goal never re-checks canAttack.
The golem rule becomes Mob::built_by_player, used by both the restore and
spares_players, with tests pinning the targets a built golem does and does
not take and the crack sound it already plays.
Conflicts: integrated.rs keeps both Flags fields (split_players, load);
pad/hardware.rs keeps the PadRouter split and moves main's stick_axes
mapping into state_of. A shared Audio gets main's device-tracking fields
and leaves reopening the output to the game's own Audio.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
main's audio device tracking (#default_seen, devices_seen, device_ticks) added fields
the shared Audio built by share() now fills.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Preserve the client's destroyBlockPos after a completed prediction so a replacement at the same position can resume the retained target, matching the isDestroying split.
Co-Authored-By: Codex <[email protected]>
handleBlockBreakAction writes destroyPos in the branch that keeps digging,
after the creative destroy and the instant-mine check, so a creative or
one-hit START leaves no block behind for a later STOP to break.
Reported in review: destroy_pos was set for every accepted START.
MultiPlayerGameMode.stopDestroyBlock opens with `if (!isDestroying) return`,
so it zeroes the progress once, on the stop that ends a running break; every
later missed pick is a no-op, the dig identity survives, and the client keeps
accumulating until the progress passes 1.0 and the STOP it sends is measured
from the first START.
Clearing the accumulated progress on every miss restarted the dig each jump,
so the client never reached 1.0 and the server's retained clock was never
read. The LAN regression now breaks the stone in 100 ticks with the dig
accumulating to 27 ticks across falls.
The verifier asked for the server-thread regression this change needs: a
survival guest under a block it can only reach near the jump apex, holding
attack and jumping, with the block's dig clock measured from the first
START.
It currently fails: the block survives 308 ticks of held attack. The
client aborts on every fall and the STOP that would read the retained
clock is only sent when the client's own progress completes, which a
five-tick in-reach window cannot reach. Reporting rather than papering
over it.
tests/test_remote_mining.py compiles an extracted copy of
stop_remote_mining against a stub, so the stub needs the new flag and its
assertion must stop expecting a stop to clear the target, which vanilla's
stopDestroyBlock does not do.
A click on the block a stopped break still names has to start the dig
again, so the server re-anchors its clock, while a held button carries the
same target on and keeps it. That is vanilla's "isDestroying &&
sameDestroyTarget" guard: the target test stays pos plus item, and the
running flag is what the start guard adds.
The STOP decision moves into a small pure function over the retained
destroy_pos, the elapsed ticks and the block's per-tick progress, so the
0.7 break, the delayed arming from the same anchor and the cases that must
not inherit a clock are covered without a Game.
Vanilla keeps ServerPlayerGameMode.destroyPos and destroyProgressStart after
an abort or a STOP that arrives out of reach: only a START writes them, the
abort clears isDestroyingBlock alone, and tick does not check distance. On
the client, stopDestroyBlock keeps destroyBlockPos and destroyingItem, so
sameDestroyTarget still matches after a fall, no second START is sent, and
the server's clock keeps running. That is the "break it a bit each jump"
behaviour the report asks for.
Horizon had the formula but not the state: the ABORT and too-far STOP
branches cleared the dig, guest_stop_digging required an active dig, and
the client's own target died on every stop, so each landing re-sent START
and re-anchored dig_start.
The server now retains destroy_pos (written only by an accepted START,
cleared by a respawn) and measures a STOP from the START that named the
block. The client splits isDestroying out of the target, so the target
survives a stop while a running break is still the thing that suppresses a
new START.
GL drivers evaluate a float in #if as a number (I Like Vanilla's #define AMBIENT_CEL_AMOUNT 0.0
tested with #if AMBIENT_CEL_AMOUNT != 0), and Iris loads such packs; Horizon refused the line and
skipped the program. The evaluator now computes in f64: float literals count, a fraction is true,
and / and % stay integer division unless a float literal appears.
A condition that still can't be evaluated already reads as false, but it was a diagnostic, so the
whole program was skipped (Pastel). It is now a warning in the loader's notes, and the program
loads with that branch off, as Iris runs those packs.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
For a public server's SMP mod: native_sidebar shows one player a sidebar of
up to 15 lines (a text, and one where the score would be) that only their
client holds, native_tab_list sets the lines over and under their Tab
list, native_tick_stats reads /tick query's rate and average tick time,
and native_sleep_ignored leaves a player out of players_sleeping_percentage's
count until they are in bed. Server players now carry yaw and pitch, for
telling who is away from the keyboard.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
PlayerHeadItem.getName answers block.minecraft.player_head.named with the
profile's name, so a head given for someone reads their name rather than
"Player Head". Horizon showed the plain item name in the tooltip and in
command feedback.
A stem has no block item, so vanilla's pick-block returns the seed it grew
from. Horizon fell back to the block's own item, found air and picked
nothing, so middle-clicking a grown pumpkin or melon stem did nothing.
ShearsItem.useOn on a growing-plant head below its top age plays
block.growing_plant.crop, sets the head to age 25, fires a block change,
wears the shears and succeeds. Horizon had no such arm, so shears on kelp
or a vine did nothing.
The arm sits with the pumpkin carve and the beehive shear and reuses their
helpers.
`es_es.json` had 13 of the 198 keys `en_us.json` ships, so the other 185
fell back to English on every Spanish screen. Translate them, keeping the
13 existing strings and every `%s`/`%.2f` placeholder.
The file also serves `es_ar`, `es_cl`, `es_ec`, `es_mx`, `es_uy` and
`es_ve`, which `lang.rs` aliases to it.
Co-Authored-By: DeepSeek V4.1 Flash <[email protected]>
The same lost key release that latched sneak also leaves the physical
modifier down, and a screen reads the physical state: every inventory
click becomes a shift-click, every throw a whole-stack drop.
native_block_entity_sync(pos) puts the position in the level's block log,
as Level.sendBlockUpdated marks ChunkHolder.blockChanged: the hub sends the
block and the block entity's update tag to the players who see its chunk,
once per position a tick. NATIVE_HOOKS_API 24. A lan_bot test checks the
chunk's tag, one BlockEntityData for two syncs, and none without a tag.
Refs #4615
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Measured with a probe pack in Minecraft 26.3 + Iris 1.11.7 against Horizon on the same world and
camera:
- gbuffers_skybasic, gbuffers_skytextured and gbuffers_clouds get full light, lightmap
coordinates (240, 240); Horizon gave (0, 240).
- gbuffers_skybasic's normal is (0, 0, 1); Horizon gave (0, 1, 0).
- Clouds get their face's normal (down for the bottom, out for the sides); Horizon gave up for
every face. The cloud mesh now carries it (w 0: the game's own clouds still shade by colour).
- blockEntityId is -1 outside block entities; Horizon gave 0.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A release does not always reach the mapping: a screen reads presses only, and
an unfocused window is given no release at all. The latch then keeps the player
sneaking or sprinting until the key is tapped again, which is what a short
alt-tab did on Wayland. The modifier state is what the keyboard itself asserts,
and winit reports it before any screen or gameplay handler runs, so a per-frame
check settles the question instead of waiting for a release that already got
away.
A mapping that toggles, or one a pad latched, is left alone: set_down(false,
needs_toggle) is a no-op for those, and a toggle the player asked for is meant
to outlive the key. A mapping rebound away from its modifier is skipped too,
since the modifier says nothing about it.
A cauldron's drain now hands out only whole thirds that fit within the
amount asked, as IFluidHandler.drain's maxDrain: asking 1 or 500 of a full
one takes one third (334) or none, so a caller moving into the room it has
never receives more than it can hold. Its fill compares each level's whole
third against what is left of the amount instead of summing, so a huge
amount can't overflow. A save with more tanks than the block now declares
reads the tanks that still exist instead of all empty, which the next write
would have made permanent.
Review: Aria's agent swarm (Dyna) in the pull request.
Co-Authored-By: ZCode AI agent (GLM-5.3)
- A position's placement (isSpawnPositionOk) is tested before its spawn list
is looked up again (canSpawnMobAt: structures reaching it, the zoomed biome,
a string-keyed list). Both only read the level, so their order cannot change
the answer, and the few block reads of the placement turn most positions
away. The placement is read from a table by species, not by an id built
with format! after a linear search (from server-tick-2's
perf/spawn-placement-first).
- LocalMobCapCalculator's players near a chunk are runs of one list, not a
Vec per chunk, and the counts per player a Vec by player index, not a hash
map: thousands of chunks ask every tick on a full server.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The fixture exports vanilla's colours, whose alpha is 255: the fluid face in
the high nibble is Horizon's own, as the emission in the low one is. A test
now checks the material stage's decode against the mesher's encoders.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
A short alt-tab (under 500 ms, so no pause screen opens) leaves the player
sneaking until Shift is tapped again. The latch in `toggles` survives the focus
loss and #4139's `sync_toggle` puts ShiftLeft back into `keys`, while an
unfocused window never receives the key release that would clear it.
Release each mapping that does not toggle, so a held sneak, sprint, attack or
use cannot survive a focus change, and pass the option through for the ones
that do: `set_down(false)` is a no-op for a toggling mapping and for a pad
latch, so the latched state #4139 was protecting still crosses the alt-tab.
The layer list and the preset list drew the flat item/<name> sprite, so a block
showed as the first texture of its model. They now resolve the item and draw it
as GuiGraphics.renderItem does (isometric for blocks, buckets for water and lava),
clipped exactly to the list; the sprite stays as the fallback.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
HappyGhast.getControllingPassenger makes the first passenger control
a harnessed adult (wearing body armor, not on a still timeout), like
AbstractHorse's does for a saddled horse. client_moves only knew the
equine half, so on a Java server the client never moved the ghast and
never sent MoveVehicle: WASD did nothing.
Verified against the pinned 26.3 bytecode:
rea_vanilla HappyGhast.getControllingPassenger / isOnStillTimeout.
The frontier-starvation fix (accept a result newer than the one shown, even
while a newer job is pending) had no test: it ran inside `receive`, which only
the worker channel can reach. `accept_mesh` is that rule on its own, so the
test drives it directly — an older-but-newer-than-shown revision fills the
frame and leaves the tile waiting, the same revision is not redrawn, the
newest settles it, and a canceled result or one for a tile that left the view
is dropped.
Generated with DeepSeek Harness (deepseek-v4.1-flash)
Co-Authored-By: DeepSeek Harness <[email protected]>
A shaders/ pack's terrain program reads each quad's face normal and mc_Entity only in flat draws
(PackAtlas::terrain_inputs(flat)); otherwise it gets (0, 1, 0) and -1. Flat draws existed only
on Metal and Intel Gen11's Vulkan driver, so on NVIDIA and AMD every pack lit all terrain as if
it faced up and knew no block from block.properties. Flat draws now exist wherever the device
supports them; frames without a pack keep per-section draws outside Metal and Gen11
(flat::for_vanilla), so only a pack's terrain and shadow pass draw flat there.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Switching from spectator to creative always left the player flying. In
ServerPlayerGameMode.changeGameModeForPlayer a flying player who leaves
spectator mode stops flying when isInRangeOfGround holds: no collider
touches the player's box and Entity.getAvailableSpaceBelow(1.0) is under
one block.
Orbs had no health and ignored blasts. Vanilla gives an orb 5 health
(ExperienceOrb.health) and ServerExplosion.hurtEntities calls
ExperienceOrb.hurtServer with the blast damage, which takes it off as an
int and discards the orb at 0 or below; pushFromExplosion then moves the
orbs that stay, from the eyes at 0.85 of the 0.5 high box. The health is
saved and loaded as "Health". Fire and lava damage are not ported yet.
Farmland that dried out or lost its support, and a path with a block over
it, became dirt and left whoever stood on it sunk a sixteenth into the
new block. FarmlandBlock.turnToBaseBlock and PathBlock.turnToDirt call
Block.pushEntitiesUp (Block.java): every entity touching the collision
the new state adds is teleported up so it rests on it. The level now
reports the swap and the game lifts the players, guests, mobs and items
the same way.
Only the player broke an egg by stepping on it. TurtleEggBlock.stepOn
(TurtleEggBlock.java) runs for every entity on the ground that does not
step carefully: canDestroyEgg refuses turtles and bats and needs mob
griefing for other living entities, then the level random rolls 1 in 100
and decreaseEggs plays the break sound (0.7, 0.9 + random * 0.2) and
removes one egg or the block. A mob now rolls that in its block effects
and the game cracks the egg.
A wind charge put a lit candle out with the fire-extinguish sound and no
block-change game event. AbstractCandleBlock.onExplosionHit calls
extinguish(null, ...) (AbstractCandleBlock.java): setBlock lit=false with
flags 11, the candle.extinguish sound at 1.0/1.0 and a block_change game
event; the smoke is a level.addParticle, which a server level ignores.
Use the world's extinguish_candle, which does exactly that.
A wind charge opened a gate by running the click path, which swings the
gate's facing towards the player when it faces them. Vanilla's
FenceGateBlock.onExplosionHit (FenceGateBlock.java) only toggles OPEN of
an unpowered gate with setBlockAndUpdate, plays the open or close sound
at pitch random * 0.1 + 0.9 and posts block_open or block_close with the
old state as context and no entity.
A ravager only trampled wheat, carrots, potatoes, beetroots and
torchflowers; PitcherCropBlock.entityInside (PitcherCropBlock.java) does
the same for a pitcher crop: with mobGriefing a server level destroys the
block with drops for any Ravager inside it.
Using an op's spawn egg with entity data on a trial spawner only set the
mob id and dropped the data. TrialSpawnerBlockEntity.setEntityData runs
TrialSpawner.overrideEntityToSpawn(TypedEntityData), whose
FullConfig.overrideEntityData / TrialSpawnerConfig.withSpawning build each
config's spawn data from the egg's data (TypedEntityData.loadInto), then
the state goes to inactive.
Horizon had only the empty-hand take-out of FlowerPotBlock, so no
potted plant could be made. FlowerPotBlock.useItemOn (FlowerPotBlock.java):
an empty pot clicked with a block item that has a potted variant
(getFullPot) becomes that pot, sends the block_change game event,
awards the pot_flower statistic and consumes one item; any other item
tries the empty hand.
ServerEntity.sendChanges (ServerEntity.java) updates a frame's map only when the framed item is a MapItem; lan_map_updates took any framed stack with a map_id component, so /give stone[map_id=1] in a frame got map packets. That check followed a NeoForge-patched decompile, which replaces the test with if (true).
native_entity_leave_level(dimension, entity, kind) fires as vanilla's
onTrackingEnd: wherever an item entity or orb is removed (picked up by a
player or a guest, merged, despawned, burnt, blown up, killed, taken by a
hopper or a mob, gone through a portal) and when its chunk unloads,
including a level closing empty. Single removals go through
remove_dropped / remove_orb; the ticks and sweeps compare IDs before and
after. Not told at a server stop, as in vanilla. API 41.
Part of #5162 and #4953.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
machine_screen() answered Hopper or Dispenser by the menu's type even when
a mod drew its own screen over it, so the player inventory and the menu's
slots sat where HopperScreen puts them (inventory at y 51), not where the
mod's screen said. It now answers none while a mod screen shows.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Hovering an option, a screen or the profile button shows what the pack's lang file says of it
(option.NAME.comment, screen.NAME.comment, profile.comment). Shift-click puts one option
back at the pack's default. A label wider than its button shortens the option's name, never
its value, and the whole label shows on hover.
The options a pack lists in sliders are sliders, each step one of the option's values. const
bool declarations (const bool shadowHardwareFiltering = true;) are switches too; as they are
as often the pack's own settings, they show where a screen names them, not with *.
The options a pack declares (switches, values with their lists, consts) on the screens
shaders.properties lays out, with the names its lang files give them, its profiles, and
Reset. Choices are saved to shaderpacks/<pack>.txt and applied to the pack's sources as
they are read; leaving the page loads the pack again with them.
local_size_x = 8 * 2 gave gl_WorkGroupSize 8, and a size that is not a number 1, silently;
such a program now fails with a message. Without clear_texture, images clear from a zero
buffer of at most 4 MiB, copied a slice of rows at a time, instead of one as large as the
largest image.
A geometry stage's write wrapped as if (hz_primary) took the else of an unbraced if; it is
now a whole if (!hz_primary) {} else write. A barrier after a shared atomic no longer goes
between an if and its else.
An image whose load can run after the same invocation's store (a loop around both, or a
function holding both called in a loop or more than once) is no longer read from the copy
taken before the dispatch. That copy has a texture of its own, so a pass sampling last
frame's copy later in the frame still finds it.
It was skipped as writing no buffer; the sink was only given to fullscreen programs. A program
drawing geometry has the depth buffer, so it now draws with no colour target, its images
written as GL's would be.
A unit test folds a small shadow program with a geometry stage and checks its triangle count.
It found that inputs declared without a size (in vec3 v[]) were unsized arrays naga refuses;
they have the triangle's three elements. An image declared with its layout on the line before
no longer gets a second one.
eyePosition, playerLookVector and playerBodyVector came from the camera, which a third-person
view backs away from the player (and turns around in front view); they are the camera
entity's eye and rotation now, and relativeEyePosition is the eye from the camera.
A shared vector split for atomic functions took sums += v to hz_get_sums() += v, which does
not compile; compound assignments write through the setter. Array sizes of several tokens
are kept whole. Tests cover components, swizzles, whole reads and writes, arrays and a
program compiling end to end.
Clearing image.* happened after the shadow and shadowcomp programs ran, wiping what they
stored; it now opens the frame. A bufferObject.* larger than the device binds is made at the
largest size it takes, with a warning, and its per-pixel size no longer overflows.
gbufferModelView holds the nausea spin the world turns with; the hand's projection now takes
it out again, as vanilla's hand never spun. Tests cover bufferObject.*, image.*,
iris.features.* and buffer blocks with and without a binding.
A composite or deferred program that samples colortexN and stores to colorimgN now samples
a copy taken before it runs, as compute programs did; it sampled an empty texture. Whether
vanilla's sky and sun are drawn depends on the pack having the program, not on it building a
pass.
workGroups needs three numbers and workGroupsRender two (or three, the last taken as 1);
another count fails the program with that message instead of running it 1x1. size.buffer.*
factors must be finite and positive, and every size resolves to at most 8192.
Its size is checked against the device's largest texture of its dimension and its texels
against that size before anything is created, so a pack asking for too much loses that
texture instead of failing validation at upload.
naga's HLSL writer has no WorkGroupSize built-in, so on DirectX 12 the two Rethinking Voxels compute programs that read it were refused. It becomes the constant the program's layout(local_size_*) declares, as GLSL defines it.
GL leaves undefined what one invocation reads of what others of the same dispatch write, and packs lean on how desktop drivers happen to run it. Rethinking Voxels' light lists changed from frame to frame on Apple GPUs, so glass, which reads its light volume directly, flickered.
- An image a compute program both loads and stores, with every load before every store, no atomic function on it and neither coherent nor volatile, is loaded from a copy taken before the dispatch: neighbouring work groups' light lists, reprojection in place and the volume shift read the values they expect.
- A barrier follows each statement of a compute main that applies an atomic function to shared memory, directly or through a function, unless main returns early: light weights raised with atomicMax were read back before every invocation had raised them.
- Flat terrain draws a program with a geometry stage without its index buffer: the vertex stage reads it (set 5) and the draw has an instance for each triangle the geometry stage may emit.
- Programs that write only images draw into a sink that keeps nothing.
- The pack's images are cleared before the shadow pass, not after it, and a pass that samples an image it stores samples last frame's copy. Images an atomic function uses are made for atomics (Metal dropped them otherwise), and images bind in vertex stages where the device lets them.
- HORIZON_SHADER_PACK_DUMP=<image> logs how many of a pack image's texels are set, for debugging.
Rethinking Voxels r0.1-beta9 voxelizes the world in its shadow pass and lights it with its colored voxel lights.
A shadow program of terrain with a geometry stage (shadow.gsh): wgpu has none, so its vertex, geometry and fragment stages become one vertex stage and the fragment stage (translate_geometry). Each vertex invocation stands for one corner of one triangle the geometry stage may emit: it reads its input triangle's three vertices itself (GeometryInputs says how), runs the pack's vertex stage on them, then the geometry stage, and keeps its corner.
- The shadow pass has the cutout test (GREATER 0.1) unless alphaTest.<program> says otherwise: leaves cast the shadow of their texels, not of their blocks.
- Images declared without a format take the one image.* gives them.
- gl_VertexID became hz_VertexID: the compatibility inputs' prefix gl_Vertex matched it first.
The game gives no GeometryInputs yet, so such programs are still skipped.
cameraPositionInt and cameraPositionFract (and their previous values) are the camera's whole blocks and the rest, exact however far from the origin; eyePosition, relativeEyePosition, playerLookVector and playerBodyVector are the eye and where it looks, the camera being the eye in first person. The previous camera position is kept in double precision for them.
- bufferObject.<n>: a storage buffer of n bytes (or bytes per pixel of the render size). Buffer blocks are bound in set 4 and lose their access qualifiers so that every stage binds them alike.
- image.<name>: 2D and 3D images of the pack's, with their format, size (or the render size scaled), sampler name and whether they clear before each frame.
- iris.features: SSBO and CUSTOM_IMAGES are supported.
- A 3D image (iimage3D) is declared as 3D.
- naga: atomicAdd and the other atomic functions on buffers and shared memory, imageAtomic*.
Nothing binds them yet.
- Compute programs dispatch before their composite-style program, over workGroups or workGroupsRender times the target's size; their images bind the side colortexN reads, and a pass that samples and stores one buffer samples a copy taken first. The MC_GL_KHR_shader_subgroup_* macros are defined when the device has subgroups.
- size.buffer: a pass draws at the size of the buffers it writes, and those buffers clear apart from the blit.
- A gbuffers_skybasic or gbuffers_skytextured that writes nothing still hides vanilla's sky or sun and moon, and the world's background is colortex0's clear colour when the pack sets it: vanilla's fog colour was added to Photon's sky, and hid its stars.
Photon v1.4 at 2560x1440 on an M5: 8 to 50 FPS (GPU 125 to 21.5 ms), its sky light from the compute program, quarter-resolution clouds.
- Compute programs (<name>.csh, then _a to _z) are translated and carried before their composite-style program: images (colorimgN in its own format), workGroups and workGroupsRender, shared memory, barriers and subgroups. Compute-only programs are noted as not supported yet. Nothing dispatches them yet.
- size.buffer.<buffer>: a size of its own, relative to the render size (floats) or in pixels (integers).
- shaders.properties and the program switches see the pack's options at their defaults (switches on, value options): Photon sizes its cloud buffers only under #if CLOUDS_TEMPORAL_UPSCALING == 4.
- naga: subgroup operations, memory barriers, and barrier() as a work group barrier.
- The custom textures texture.* names are uploaded and bound to the samplers they name: one texture per dimension of a unit (Photon's 3D scattering LUT under depthtex0), and the lightmap and block atlas by resource location. A sampler with nothing to read gets a stand-in of its own dimension instead of failing its pipeline.
- Outputs past the last buffer RENDERTARGETS names go nowhere, as in GL, instead of an attachment the pass lacks (a crash).
- A shader the driver's compiler refuses (an internal error) skips its program as a validation error does.
Photon v1.4 now loads all 27 overworld programs and draws its lighting, shadows, fog and volumetric clouds.
naga's SPIR-V writer panicked on Photon's deferred4 ("Expression is not cached"): the GLSL frontend passed a pointer into the middle of a variable for f(s.field), which only Metal's writer takes. The naga pin moves to a fix that spills such arguments through a local and writes them back after the call, as GLSL's copy-out says.
real_packs with BACKENDS=1 hands every translated module to the SPIR-V, HLSL, GLSL and MSL writers with wgpu's settings, so this is caught on any machine.
An R32I or R32UI buffer a program binds as colorimgN is made with atomic storage usage, as the pack's own images already are: Metal drops texture atomics on a texture made without it, so imageAtomicMin on colortex9 did nothing.
atomicAdd(sums[i], n) and atomicMax(lights[k].w, w) on a shared ivecN or uvecN are GLSL that naga can't express (it has no atomic vectors): such a vector, or array of vectors, becomes an array of its components, a component an element of it, and the whole vector is read and written through functions of ours. Rethinking Voxels' light lists use them.
A prepare, deferred, composite or shadowcomp program of compute stages only (.csh without .vsh and .fsh) runs its stages in order where the program would draw. shadowcomp programs also run the compute stages they carry.
- IS_IRIS is defined: packs take Iris's extensions (compute programs, images, custom textures) only where it is, and Rethinking Voxels draws an error screen without it. iris.features.optional defines IRIS_FEATURE_<name> for the features Horizon has; a required one it lacks is noted.
- A global initialized with the size on its type (ivec3[6] offsets = ivec3[](...)) kept only the size as its type.
- naga: constant shifts drop the bits shifted out, as GLSL says (1 << 31).
The pack's renderer asks the adapter what it can do instead of assuming Metal or desktop Vulkan. Without compute shaders compute programs are skipped; without storage writes from fragment or vertex shaders the programs that write images or buffers there, and geometry stages, are skipped; an image whose format can't be a storage texture is left out with the programs using it; without clear_texture images are cleared by copying zeros. Each leaves a note, and the bind group layouts made outside the error scope (storage buffers, geometry indices) only name stages the device supports, so none of it can stop the game.
HORIZON_SHADER_PACK_WITHOUT=compute,fragment-storage,vertex-storage,clear tries the fallbacks on a device that has it all.
- shadow.culling=false draws every section within the shadow distance and every face of it; reversed does so within voxelDistance (default 160) and culls beyond as before. Horizon drew only the faces turned to the light, and a pack voxelizing the world in its shadow pass lost the others.
- A geometry stage's image and buffer writes run in one of the invocations that stand for its triangle; in an expression the others read the value instead. The instances drawn follow from the strips EndPrimitive ends.
- at_midBlock for flat terrain: the offset to the centre of the face's block, in 64ths.
- firstPersonCamera.
- naga: imageAtomic* give the texel's value before the operation.
- Storage buffers are zeroed once and kept across frames, bound in set 4 to every program that declares the buffer block.
- The pack's images are made at their size, cleared before each frame when asked, and bound as images by name and as textures by their sampler name.
- Images an atomic function uses bind for atomics; the device asks for texture atomics, storage buffers in vertex shaders and texture clears where the GPU has them.
Rethinking Voxels r0.1-beta9 loads 10 programs and draws; its shadow pass (a geometry stage) and compute-only programs remain.
Packs invert only gbufferProjection's perspective terms when they rebuild positions, so a bob inside it moved every rebuilt position, and the lighting with it, while walking. The bob and the nausea spin are now part of the model-view, where GL drew them; the product is the same, and the hand's projection no longer adds the bob a second time.
gbufferModelViewInverse.1.0 read row 1, column 0: the transpose. Packs turn a view direction to the world as inverse.0.0 * x + inverse.1.0 * y + ..., so Photon's sun direction (from sunPosition) turned with the camera: the sky lit as another time of day when looking around, swung with view bobbing, and the sun was missing.
shaders.properties is preprocessed before its lines join: a value continued with \ just before #else (Photon's moon_phase_brightness) swallowed the directive and the other branch, so the uniform was never defined. Directives now apply while lines join, and lines of branches they turn off are dropped.
Each dimension of a TEXTURE_* entry is at most 16384, and the bytes it needs are counted
without overflowing, so a pack asking for a huge volume gets a note instead of a wrapped
length check.
Photon translated 45 of its 88 composite-style programs and none of its overworld terrain, sky set-up or final pass; now all 88 translate and every dimension loads its full set.
- A struct passed between stages (flat out CloudsParameters p) keeps the pack's declaration as a private global and travels as its members packed into vectors, copied by our main.
- sampler, texture2D and Vulkan GLSL's other texture and sampler type names are free names in desktop GLSL (sampler2D sampler); they are renamed unless called.
- naga's GLSL frontend lacked textureGather, sampled implicit levels in vertex shaders, broke struct constructors with array members, vecN(bool) and vector % scalar: naga is patched with the fixes until a release has them.
- real_packs prints every loader note with SHOW_ERRORS.
texture.<stage>.<name> and texture.noise name the files a pack binds to its samplers: PNG images, and raw 1D, 2D and 3D data with their size, format and type, each with its .mcmeta (blur, clamp). Post carries them, with a note for each that can't be read; nothing draws them yet.
Shader packs need fixes to naga's GLSL frontend that no release has yet. They go on top of this
copy in the commits that need them, as wgpu-hal's fix did, instead of a git dependency.
- A fragment output declared float, vec2 or vec3 (int and uint too) is written as the vec4 of its kind, alpha 1: wgpu's targets take four components.
- inverse() of a mat2, mat3 or mat4 is a function of ours (the adjugate over the determinant): naga writes it neither for Metal nor for WGSL.
- alphaTestRef is a constant of the program: its alpha test's reference.
- Solid and cutout terrain draw together, so terrain gets the cutout test (alpha > 0.1) unless alphaTest.* says otherwise.
- The naga pin moves to the fixes for pointers into vectors passed as out arguments and field selection on a matrix column.
A function-like macro called across lines (Photon's clamp01(\n ... \n)) was left unexpanded. Its lines now join the line the call starts on; the lines after it stay, empty, so line numbers still map back.
The auto-reconnect's `allowed` matched English keywords ("ban", "server
closed", "online mode"...) against the disconnect's resolved text. The
texts arrive resolved in the player's language — a vanilla server's
"Server closed" is `服务器已关闭` to a zh_cn client, so a legitimate retry
was refused, and Horizon's own refusals (NOT_SIGNED_IN, the invalid
address) could only stay English for the matcher to see them.
The disconnect's translation key now travels with the text: the net
thread's reasons carry it (`Reason { text, key }`) from the server's
component's `translate` field, the session-service refusals' inner keys,
and Horizon's own keys (new `horizon.multiplayer.not_signed_in`,
`invalid_address`, `connected_friend_direct/relay` in all six bundles;
the friend connection's status texts and the local world-open progress
line move to keys too). `allowed` decides from a key list first — the
keyword scan stays as the fallback for plain reasons (OS errors,
servers' literal kicks in the server's own language), whose behavior is
unchanged.
NOT_SIGNED_IN is localized now that its classification is key-based, and
`load_language`'s no-language fallback picks Horizon's bundled strings up
so the net thread's keys resolve there too.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
native_entity_join_level(dimension, entity, kind, pos, loaded_from_disk,
canceled) fires as vanilla's PersistentEntitySectionManager.addEntity:
from spawn_item, add_orb, /summon and chunk loads. loaded_from_disk is
true only for an entity that comes back with its chunk; portal and
gateway arrivals and /summon go through the chunk-load queue, so they are
marked there as fresh adds. A cancel takes the entity back out, and a
/summon of it fails with commands.summon.failed.uuid. API 41.
Part of #5162 and #4953.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
native_block_entity_fluids declares a mod block's tanks, as its container
declares its slots: each tank's capacity, the tanks each face reaches, and
which a fill may reach and a drain may take. The tanks' contents live in
the entity's saved compound under `fluids`, so they survive restarts and
the block's own tick reads them there.
native_fluid_tanks reads them through a side, native_fluid_insert fills
(first into tanks holding the same fluid, then into empty ones) and answers
what did not fit, native_fluid_extract drains and answers what it took.
Both also work a cauldron: it holds thirds of a bucket, a lava one only
whole, and powder snow none, as it is no fluid. Water and lava are the
fluids; a tank holds one at a time.
Co-Authored-By: ZCode AI agent (GLM-5.3)
Review found that the by-id path dropped what `mob_hit` applies: a mob in its
death animation, the mob the player rides and the camera entity are never
clicked by the pick, so naming one in a packet must not reach the feed, shear,
leash or name paths either. `named_mob_pickable` is that rule, shared by both
paths and unit-tested on its own, since the pick's copy cannot be tested
without a world.
Generated with DeepSeek Harness (deepseek-v4.1-flash)
Co-Authored-By: DeepSeek Harness <[email protected]>
ServerGamePacketListenerImpl.handleInteract resolves packet.entityId(), checks
it and the range, and hands that same entity to Player.interactOn. Horizon aimed
the guest at the named entity and then picked one again from the view, so a mob
that moved between the eye and the target took the click, and a block in the way
stopped an interaction the packet had already validated (#4651).
GuestUse::Entity now carries the packet's id and location, and the interact_
helpers use it as it is; the local player still picks from the crosshair, the way
Minecraft.hitResult does.
Generated with DeepSeek Harness (deepseek-v4.1-flash)
Co-Authored-By: DeepSeek Harness <[email protected]>
`slot_position` asks the hopper's and dispenser's stock layout before its
native-screen arm, so a mod screen on a `minecraft:hopper` or
`minecraft:generic_3x3` menu never reached the slots it declares, and its
`inventory_y` was ignored: the mod drew its screen while every slot stayed on
the stock grid (#5141, item from #4953).
Only that layout stands back, and only while a mod screen is shown; the rest of
those menus was already the mod's, since `draw_native_screen` owns the
background and `screen_height` the height. With no mod screen the stock grid
answers exactly as before.
Generated with DeepSeek Harness (deepseek-v4.1-flash)
Co-Authored-By: DeepSeek Harness <[email protected]>
Migrate legacy side files before loading End chunks and retain gateway source chunks while resolving new exits.
Co-Authored-By: Codex GPT-6.1 <[email protected]>
Settings apply at every pointer move of a dragged slider, and since the
Exclusive Fullscreen rows that listed the monitor's video modes each time:
about 28 ms on Windows, so the handle lagged and jumped. The list is now
kept until the window's monitor changes, in game and on the startup menu.
horizon-ctl options-set keeps it too (with_option rebuilt Settings without it).
Fixes#5133
Co-Authored-By: Claude Opus 5.5 <[email protected]>
With a shaders/ pack, Iris gives terrain, translucent terrain and fluids their colour without
vanilla's face shading: packs light faces themselves (oldLighting, off unless the pack turns it
on). Horizon baked the shading in, so packs shaded faces twice: a probe pack's colours per face,
Horizon over Iris, were 0.8 north and south, 0.6 east and west, 1.0 up. Sections now mesh
without it while a shaders/ pack runs (Mesher::shader_pack_lighting), and remesh when that
changes.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A click on a block that opened nothing left use_target set, so a menu an
entity opened next (a chest minecart's generic_9x3) could take the machine's
screen when the machine declared the same menu (arismoko on #5061). Using an
entity now clears it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A jigsaw start (villages, outposts, bastions, ancient cities, trial chambers,
trail ruins) is loaded from its saved Children (PoolPiece::from_saved), as
StructureStart.loadStaticStart does, so the neighbours place the parts that
cross into them and the beardifier sees them. Another structure's pieces
have no loader: it is the generator's own start when that has the same name
and the same pieces' box, else the name and box alone. Decoded once per chunk.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The window's server thread publishes the generators; end.locator is only
set on the deprecated in-frame path, which is not extended.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
DensityCache holds DebugEntryChunkGeneration's lastPos and result; the test pins
the sampling on a new block or dimension, the cleared line without a
generator, and the clear when another world opens.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
native_chunk_load, with whether the chunk is new, and native_chunk_unload,
including a level that closes for being empty. native_player_logged_in and
native_player_logged_out come in pairs however a player leaves: a disconnect, a
kick, the server stopping or failing. At a stop they're told before the save,
still listed, and a last message goes out.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
DebugEntryChunkGeneration adds its group only in the server-level branch. With no
generator (a remote server, a LAN world joined, the server thread not there
yet) the cached line of the last world was still drawn; it is cleared
there, and when another world opens.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
smooth() starts from the first value it sees. The load-time dry run that finds unknown names
evaluates every custom uniform with the built-ins at zero, so each smooth() started there:
Photon's biome_temperature (smooth over 1000 s) read -0.97 on the plains where Iris reads 0.6 at
once, and took minutes to get there. Smoothing now restarts after the dry run, and on frames
before the camera's biome is known (its chunk not loaded yet).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
edits_the_lists_while_stopped failed now and then on Linux CI with "a server
is already running in /tmp/horizon-server-lists-<pid>". The folder's lock is
an flock taken with try_lock; while another test in the same binary spawns a
process (tar, id, the engine), the child holds a copy of the lock's
descriptor between its fork and exec, so the lock is briefly taken and
try_lock answers WouldBlock. one_supervisor_per_folder already retried for
that window (4513af899); the lists test did not, neither for its own lock nor
for the edits, which take the lock inside lists::run.
The retry moves into supervisor::retried, a test-only helper (50 tries, 20 ms
apart), used by one_supervisor_per_folder and by the lists test for its lock
and for the edits that must succeed. The edits that must fail while the test
holds the lock are unchanged, and so is lists::run.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Use the active world recipe tables when handling remote recipe placement, and let the server own remote grid and ghost updates instead of predicting them on the client.
Co-Authored-By: Codex <[email protected]>
LZ4BlockInputStream decompresses each block with lz4-java's
LZ4JavaSafeFastDecompressor and calls it corrupted unless the decompressor
reads exactly the block. That decompressor has its own end conditions: a
sequence whose literals leave less than 8 bytes of output, or of input
past them, must be the last and end the output, and an offset of 0 fills
the match with zeros. lz4_flex's decoder disagrees both ways, so Horizon
loaded some streams lz4-java refuses (with whatever they decoded to) and
refused some it loads.
decompress_lz4_java ports that decompressor, fed from a buffer kept as
LZ4BlockInputStream keeps compressedBuffer (it may read past the block,
up to the buffer's end), and decode_lz4 requires it to read the whole
block. lz4_flex still compresses.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The game's generator is Persistent<VanillaTerrain>, which forwarded
final_writes but not final_block_entities, so the trait default (none)
replaced the chests and spawners of structure pieces reaching into a chunk an
older version saved. Test through Persistent. Also adapts saved starts to
the starts_at wait/try split on main.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
DebugEntryChunkGeneration.display keeps lastPos and result and recomputes
only when the feet BlockPos changes; F3 sampled every frame. The line reads
the pipeline's samplers only, so it now goes through the generator and no
longer takes the shared query worker (no try_lock, no frame without the line).
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
native_container_slots, native_extract_item and native_move_items follow the
rules hoppers use; a move puts a refused item back, so none is lost.
native_dimension names the level for a machine's tick, which isn't told it.
native_insert_item goes one item at a time and tells the container once. It
no longer overfills an empty slot, leaves a client's copy alone, and can't
reach a machine's slots during its own tick, where what went in was lost.
Hoppers change with it, as vanilla has them: a shulker box doesn't go into a
shulker box, a stack's max_stack_size component sets its limit, and comparators
read it. A comparator's fill ratio is clamped to full, where vanilla's
count-at-most-limit invariant keeps it: a stack a mod wrote over its own
limit can't push the signal past 15. A slot whose block entity went away
keeps the item, rather than losing it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Co-Authored-By: ZCode AI agent (GLM-5.3)
RegionFile.write and clear end in writeHeader, which writes both header
sectors from memory. The constructor has already zeroed the entries it
found invalid there (overlapping the header, empty, past the end of the
file), so a write leaves them zero on disk. Horizon wrote only the
touched slot's offset and timestamp, so stale invalid entries stayed in
the file for other tools to read.
write_header replaces write_header_entry and writes both sectors.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A mod usually has its own itemGroup.<mod id> already (reservedbytes on
#5066): "title" is looked up as a lang key first, English text when it is
none. itemGroup.<mod>.<name> still overrides it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Only call `forget_sync_data` if the generated overlay contains files.
This prevents unnecessary data clearing when no file updates are
required for the LAN overlay.
Review fixes (kacpherek on #4947):
- a refused, lost, timed-out or unreadable tile comes back as no columns,
and Lod asks for it again 2 s later instead of keeping it empty;
- sampling threads queue their requests and the game thread sends them
each tick with its other packets, so they no longer take the
connection's writer; requests queued when the far terrain ends are
dropped with their slots;
- docs and the e2e: horizon-lod-distance is off unless set.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The last step: a Horizon client of a dedicated server or of a LAN world
(not of its own server thread, which hands its generator over directly)
draws far terrain when the server offers it. Its far terrain's samplers
each send one tile request and wait for the reply, which the network
thread hands straight to them (lod_net::Inbox), keeping to the server's
limits (requests waiting, tiles a second) so a well-behaved client is never
refused. LOD Distance stops at the server's distance. A new dimension, a
reconfiguration or a disconnect cancels the waiting samplers at once.
tools/server_lod_e2e.py checks it on a real dedicated server: with
horizon-lod-distance=16, a client with LOD Distance 32 gets the server's
far terrain, capped at 16 chunks, and draws far tiles; with 0, none.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Review fixes (kacpherek, dimaniojk on #4946):
- the worker hands finished replies back to the game thread, which sends
them before the tick's flush_all, so they keep their place among the
game's packets;
- a slot frees when the reply leaves, and replies wait while the player's
connection has more than 1 MiB pending;
- every request takes a token, refusals included; unreadable or over-rate
requests are strikes, and 64 strikes disconnect the player;
- a panic while sampling drops that level's sampler and refuses the tile,
so every job still gets a reply;
- only clients with the horizon brand (the ones offered it) are served;
- a level with no far terrain refuses instead of leaving the client waiting.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
NbtIo.readUnnamedTag skips the root's name with StringTag.skipString
and never decodes it, so a file whose root name isn't modified UTF-8
loads in vanilla. Horizon decoded it and refused the whole file; for
level.dat that meant the world opened with fresh defaults (game mode,
cheats, time) saved over the real ones.
read_named and read_named_streaming read the root name leniently: a
name that isn't modified UTF-8 is read as empty. Every other string,
compound keys included, is still checked, as DataInput.readUTF does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
LockCode.canUnlock lets a spectator or a player whose main hand matches the
lock's item predicate open a container; BaseContainerBlockEntity.createMenu
refuses anyone else with an action-bar message naming the container and one
chest-locked sound. Horizon kept the tag on disk and read it nowhere, so a
locked chest opened for anyone, /data get did not show the lock, and
/setblock could not make one.
The lock now lives where the loader already put it, one place reads and
writes it, and the open funnel refuses with vanilla's message and sound.
The second step: a dedicated server, or a LAN world, offers far terrain to
Horizon clients while they configure (vanilla clients hear nothing) and
answers their tile requests. `horizon-lod-distance` in server.properties
(Horizon's own key, 64 chunks by default, 0 off) bounds how far; each
player may have 8 tiles waiting and ask for 64 a second (a burst of 256),
and a tile past the distance, over the limits or of the wrong shape is
refused at once. One "Far terrain server" thread, started at the first
request and behind the ticks (horizon_threads::behind_server), samples
them from the player's level for every player in turn, and sends each
reply to its player itself. Clients don't ask yet (the next step).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
horizon-lod-distance now defaults to 0: far tiles reveal the generator's
terrain beyond view distance, which helps crack a seed. The offer starts
with a protocol version. A reply over 512 KiB or 16 384 spans, with spans
out of order, overlapping, empty or out of the height range, is an error
before it allocates; the server refuses such a tile instead of sending it.
(Real tiles measure 4-6 spans a column.)
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The first of three steps that let a Horizon client get its far terrain
(LOD) from a dedicated server or a LAN world, whose generator and seed it
doesn't have: the offer a server sends Horizon clients in configuration,
the tile request, the reply (columns with biomes by name, block states by
id, or a refusal), the per-player check of a request's distance and shape,
the token bucket both sides pace with, and the `horizon-lod-distance`
property's reading (64 chunks by default, 1024 at most, 0 off). Nothing
sends or reads them yet.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Lightning, additive quads and translucent particles draw in the OIT stages
there, and only the OIT entities got a mark after them. The firework mark
also takes the same base as the draw it follows.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
NbtIo.write goes through StringFallbackDataOutput: a writeUTF that would
exceed 65535 encoded bytes logs "Failed to write NBT String" and writes
an empty string, and the rest of the tree is written. Horizon's
write_named failed instead, and since a save batch's records are encoded
before any is written, one block entity with such a string (a long
command block command, a book) made the whole batch fail: every chunk
changed since the last save was lost, not only that one.
The disk writer (write_named) now writes such a string empty; the
network writer still refuses it, as vanilla's does.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Entity.getRootVehicle follows getVehicle up through any vehicle, and
PlayerList.remove removes getPassengersAndSelf of that root, the same
entities Entity.save wrote under Passengers for
ServerPlayer.saveParentVehicle. The ride saved a minecart without its mob
passengers while remove_vehicle_tree removed them, so they left the world
unsaved, and vehicle_root climbed a mob stack to a boat but not to a cart.
- vehicle_of: the mob, boat or minecart a mob rides; parent_vehicle_root
climbs it to the root, so a stack in a cart has the cart as root.
- vehicle_tree_nbt puts the cart's mob passengers under Passengers, as
for a boat (add_passengers, out of put_mob_passengers).
- remove_vehicle_tree removes vehicle_tree over the living mobs, the ones
put_mob_passengers saves.
- Unit tests over the pure parts: a lone rider saves the horse, two
players on one boat save nothing, a mob in a boat rides under it, a
cart saves and takes its mobs.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Ensure `forget_sync_data` is called unconditionally when packing files
for LAN overlays and when disconnecting from a remote session. This
prevents stale resource synchronization data from persisting across
different connection states.
The window's world has no mod containers (the server tick installs them), so
mod_menu_at was always none there and no mod screen opened. Ask
native_block_entity_container for the clicked block instead.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A world with the `trade_rebalance` feature flag now plays the flag's
built-in pack: the overridden villager trade tags and trades, the
librarian enchantment trades and the changed chest loot, as vanilla's
server does. `MinecraftServer.configurePackRepository` force-enables a
feature pack whose requested features the world enables
(`enableForcedFeaturePacks`) and drops a selected pack asking for
missing ones: `datapack::enabled_ids` now takes the world's
`enabled_features` and does both, so the pack's files reach the world's
content overlay (`/datapack list` shows it as enabled).
The content tables then merge the pack over the resource root:
`Trades::load` takes the packs' `villager_trade`/`trade_set` files and
their tags (`replace` clears, `TagLoader`), and `Enchantments::load_with`
takes the packs' enchantment tags, so the librarian trades'
`#minecraft:trades/<biome_common>` pools resolve. Along the way:
- `PackOverlay::tag_kind` stripped a leading `:` that is not part of
the `tags/<ns>:<registry>/<path>` key shape, so it never matched and
every pack tag overlay silently no-opped; loot tables' tag overlays
were affected too.
- `/loot give ... loot <table>` rolled the resource root's table when
the table had a random sequence (`self.res.items()`), not the world's
own — the world's tables are used for the sequence branch now.
- The world's flags feed the worldgen overrides and the biome table
(`packed_biomes`, `level_shapes`, the preload reads them from the
world's level.dat), and the integrated open path passes them through.
- Fixes the stacked world-save commit leaving `integrated.rs` calling
`packed_biomes` without its `resources` argument (the bin did not
compile).
Vanilla reference: `MinecraftServer.configurePackRepository`,
`enableForcedFeaturePacks`, `TagLoader.load`,
`SimpleJsonResourceReloadListener`, the built-in
`data/minecraft/datapacks/trade_rebalance` pack.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
The first try (key == target) never held for a mod block: `holds` knows only
vanilla block entities, so every mod screen stayed shut. The screen now opens
when the menu is the one the clicked block's container declares
(`mod_menu_at`), which a chest menu from a command over a stale target is not.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
RideCommand.mount took only a mob as the vehicle. A player or a mob now
mounts a boat with a free seat, and a player any minecart no other player
is in, through the same state as a click. /ride <player> dismount goes
through stop_riding, and Entity.getVehicle names a minecart as one.
Part of #4968.
The overlay's key hashed each enabled zip in full, on the assumption that one
file reads in a moment: with a 9 GB pack that read 9 GB at every launch.
A zip over 250 MiB, vanilla's limit on a server pack, is now keyed by its
length, modified time and central directory. Smaller ones keep their bytes, so
a server pack's touched time still doesn't rebuild the overlay, and an edit
that a zip's empty checksums wouldn't show is still seen.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
- A mod screen opens only when the clicked block holds the menu: use_target
outlives its use, so a menu opened later (a command's, another mod's) got the
last clicked machine's screen, and its clicks and tooltips (arismoko on #4106
and #4289).
- native_screen_tooltip runs over an empty slot, as its doc says, so a mod can
say what an input takes (arismoko on #4289).
- A mod block whose template has no block entity gets none, with a warning:
breaking it never removed the block entity, which stayed in the save
(reservedbytes and DUzzL on #3996).
- A GUI texture that doesn't decode is named in the log, not skipped silently
(reservedbytes on #4106).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
From reservedbytes' test on #4331: a block's item was added to search twice
(blocks' items are already among the plan's items); the page buttons stretched
the 200-wide button sprite into 20 px and lost their sides (now blit_sprite, as
vanilla's Button draws); turning the page while on search, saved hotbars or the
inventory keeps that tab; a mod tab's title comes from itemGroup.<mod>.<name>
in its lang files, the file's title being the English fallback.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A content mod declares tabs in data/<id>/horizon/creative_tab/*.json: a
title, an icon and its items. They come after vanilla's, ten to a page
(the top row's first five columns, then the bottom row's), with < and >
buttons above the screen; search, saved hotbars and the inventory show on
every page. Its new items join the search tab under their English names.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
"thickness" (2 to 14 pixels) gives a connecting block its own outline and
collision, built as PipeBlock builds its shapes (pipe_shape, checked against
the chorus plant's 64 states). The shapes go through the same per-state path
as a block's "shapes", so a block sets one or the other.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A block declared with template minecraft:chorus_plant and connects_to: a
block tag is vanilla's PipeBlock without the chorus plant's rules: no
support needed, and each of its six sides joins the same block or a block
in the tag, on placement and on every neighbour change.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Listing the packs opened every zip in resourcepacks/ twice, once for its
pack.mcmeta and once for its pack.png, and each open read the whole file
into memory on the render thread. With 22 GB of packs in the folder, enabled
or not, opening the options screen froze the game for five to seven minutes.
A Zip now keeps the file open and holds only the central directory on, with
the END record normalized as before, the way Java's ZipFile works. Entries
are read at their positions, so clones and threads don't share a cursor.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
flush_chat_queue now goes through accept_chat, as ChatListener.flushQueue runs each handler.
Without it a flushed signed message never reached the last-seen tracker, and vanilla's
server disconnected the client at its next message (checksum mismatch). The harness now
sets Chat Delay, holds a message back and flushes it. send_command parses over chat_tree.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
EntitySelector.findEntities with a UUID takes ServerLevel.getEntity,
whatever the entity is. target_with_uuid kept only mobs, so /tp, /kill
and /data get entity with a display's, armor stand's or item's UUID
found nothing; the scoreboard's holder lookup kept mobs and items. Both
now go through level_entity.
Fixes#4966
`DataPackCommand` gates feature packs on the world's flags: a pack whose
requested features are not enabled is not available, cannot be enabled
("required flags are not enabled in this world"), and a feature pack can
never be disabled — the flag stays on for the world's life. Horizon
refused every enable and allowed every disable regardless. The command
now reads the world's `enabled_features`: with the flag on, a feature
pack enables and its data loads; without it, it stays hidden and the
refusal names the missing flag.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
Vanilla's server save runs `PlayerList.saveAll`: every connected player
is saved to `playerdata/<their own uuid>.dat` — there is no world-owned
player file. The integrated server's save now does the same for its
placed guests (`playerdata::save`), on every save and when a player
leaves, and the owner joins through `PlayerDataStorage.load` on their
own uuid first — the world's own player state is only the fallback for a
first join. Previously the owner's state was stored into the level and
then dropped (`write_without_player`), so position and inventory were
lost on every reload, and the stale single-player file shadowed the
newest-file fallback on load.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
The integrated server's owner plays as the world's own player
(`Owner`): on leave their state is stored into the host's `Level`
(`playerdata::store`), but the server saves with `write_without_player`
(it runs as `dedicated`) and the window's own save is skipped while it
is connected, so the state never reached a file and position and
inventory were lost on every reload. The integrated role now writes the
world's own player file (`players/data/<singleplayer_uuid>.dat`) with
the level, as single player does (`PlayerDataStorage.save`); a real
dedicated server keeps saving its players on their own.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
Review of the port against the decompile found three deviations that
stalled carts at corners and broke climbing:
- `stepAlongTrack` ends with `setDeltaMovement(movement)` — the movement
turns onto the exit's direction at the same speed. Without it the
velocity never turned, so a corner's exit selection kept the old
direction and the cart stopped.
- `TrackIteration`'s slope, halt and boost speeds apply once per tick
and carry their flags across the iteration; the port re-applied them
on every rail.
- the player-input push does not replace the natural slowdown and the
speed cap (vanilla applies both), and a stuck step
(did-not-move-but-would-have) zeroes the movement.
Also `isDecending` reads the step's movement (the exit's direction),
not the stale velocity, for the hill's Y projection.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
`AbstractMinecart` chooses its behavior by the world's features: with
`minecart_improvements` the new behavior runs — the slide on slopes
grows with the speed (`max(0.0078125, h * 0.02)`) so a descending cart
keeps its pace, the empty cart slows half as fast (0.975 over 0.96),
the speed cap is a vector normalization instead of a per-axis clamp,
a powered rail boosts by `+0.06` along the movement or pushes a resting
cart 0.2 away from the signal's side, an iteration carries leftover
distance across rails, a hill projects the Y onto the slope, and a V of
opposing slopes rests a slow cart (`restAtVShape`). Without the flag
the old physics runs as before.
Stacked on #4975 (← #4957 ← #4948 ← #4619 ← #4577), whose
`enabled_features` the gate reads.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
Vanilla reads the world generation data behind a loading screen before
the Create New World screen opens, again after its Data Packs or
Experiments screen applied a selection, and warns before applying a
selection that asks for experimental features. Horizon showed none of
that.
- `createWorld.preparing` (`GenericMessageScreen`: the menu background
and the centred message on a black box with a 1-pixel border and 12
pixels of padding, no progress): Create New World reads
`vanillagen::shared_data` on a worker while the title shows it, and
opens the create screen when the read is done; the Data Packs screen's
Done and the Experiments screen's Done re-run the read the same way
(`applyNewPackConfig`).
- `ConfirmExperimentalFeaturesScreen`: when the Data Packs screen's Done
picks a feature pack, the warning comes up first — the title, the red
message wrapped to 310 and centred, a 100-wide Details button, Proceed
and Back in the footer, the whole block centred, Escape declining. Its
`DetailsScreen` lists the picked packs bold with the flags they ask
for (`selectWorld.experimental.details.entry`). Proceed applies the
selection, Back reopens the Data Packs screen on the selection it was
opened with (`tryApplyNewDataPacks`'s `onAbort`).
Stacked on #4619, whose Experiments screen and feature-pack plumbing the
confirm screen builds on.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
Vanilla's level.dat carries `WorldDataConfiguration`'s `enabled_features`
flat in the Data compound (`FeatureFlags.REGISTRY`'s codec: a list of the
whole set's namespaced ids, `minecraft:vanilla` included, the tag omitted
when the flags are vanilla's default) — Horizon wrote nothing, so a world
made with experiments told vanilla nothing and forgot its flags itself.
`Level.enabled_features` is written with the selection (the flags of the
picked feature packs over vanilla) and read back from level.dat, so a
vanilla-made world keeps its flags too. The behaviors behind the three
flags are still separate work; this is the data they will gate on.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
Review fixes on #4619:
- The More tab's `options.press` guard caught button 30 before the
Experiments arm could, so the button did nothing; the guard lets 30
through now.
- Create World's `open_world` overwrote the `datapacks.json`
`datapack::install` had just written, losing the unselected world
packs' `disabled` entries, so a pack the player left off came on. The
write now only happens when `install` did not run (only Experiments
used), and it creates the world folder and logs failures instead of
dropping the error.
- `level.dat`'s `DataPacks.Disabled` also lists the world packs the Data
Packs screen held back, as vanilla's `DataPackConfig` does.
- The completion registries' pack list gains the feature packs, so a
pack enabled through Experiments completes its worldgen ids.
- The feature packs on the Data Packs screen compared against the
resource pack format and drew red; they compare against the data pack
format now.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
Vanilla's More tab offers Game Rules, Experiments and Data Packs; the
Experiments button was drawn disabled ("Experiments have no screen yet").
It now opens vanilla's ExperimentsScreen (`title/experiments.rs`): every
feature pack the resources root ships under `data/minecraft/datapacks`
(`ServerPacksSource`, `PackSource.FEATURE`), each with a 44-wide on/off
switch at the right of its row, its description grey underneath, the red
warning wrapped to 310 and centred, Done and Cancel in the footer —
vanilla's `HeaderAndFooterLayout` + `SwitchGrid` geometry. Switching a
pack adds or removes its id in the world's data pack selection, which the
Data Packs screen and the world's DataPacks then see; Done keeps the
other packs and rejoins the picked feature packs last, reversed
(`onDone`), Cancel and Escape leave the selection as it was.
The feature packs also join the world's repository, so a picked pack
reaches the game:
- `datapack::feature_packs` scans the resources root's feature packs and
`screen_packs` offers them on the Data Packs screen (available, off),
as `PackSelectionScreen` does;
- a world remembers its selection in `datapacks.json` (the pack stack's
own file) and `level.dat`'s `DataPacks` (`DataPackConfig`);
- `pack_stack`, `packed_biomes` and `world_worldgen_overrides` resolve
the ids against the resources root, and a feature pack stays off
unless the world enabled it (vanilla starts `PackSource.FEATURE`
unselected), so existing worlds behave exactly as before;
- `/datapack` shows an enabled feature pack with the feature source and
can disable it again, while `enable` still refuses one
(`commands.datapack.enable.failed.no_flags`) — the three flags'
behaviors are not implemented yet, but the packs' data files
(trade_rebalance's loot tables and tags) reach the content tables
through the pack stack.
Dropping a pack on the window (`PackSelectionScreen.onFilesDrop`): a
zip, a pack folder, or a folder holding exactly one pack (`PackDetector`)
is copied into the create screen's pack folder, where the Data Packs
screen's existing watcher lists it within a second; anything else is
refused and logged.
Stacked on #4577, whose unified pack list the resolution goes through.
Verified headless: a world with `trade_rebalance` in `datapacks.json`
lists it enabled ("[trade_rebalance (feature)]"), disables it with
`/datapack disable`, and the other feature packs stay available.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
Co-Authored-By: ZCode Agent <[email protected]>
Review fixes on #4577:
- The window's Game on the integrated-server path kept `world_biomes`
unset: the chunks it draws carry the packs' biome ids past the built-in
list, but nothing named them. `open_integrated` builds the world's
table now, as `finish_open_world` does for the load path.
- The integrated path's own encoders read `res.biomes()` past the world's
table: `lan_host`'s chunk encoder and `/execute if biome`'s naming go
through `world_biomes!` now, so a packed biome keeps its name and its
place on the wire (a guest still needs the table sent over the
connection; a remote server join clears the last world's table so no
stale names linger).
- A tag in a sub-directory of its registry never matched: `content_id`
attached the namespace to the file's last segment
(`tags/worldgen/biome/has_structure/minecraft:village_plains`), while
the lookups build `tags/{registry}/{ns}:{path}`. The store key is the
tag id now (`tags/worldgen/biome/test:sub/mytag`), `PackOverlay`
gained `tag_kind` for the registry walks, and the readers match.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
`WorldgenOverrides` and `Biomes::with_packs` both take the world's packs,
but the client built the biome table and the worldgen from different
lists: `content_overlay` reads `datapacks.json` (empty until a world
writes it) while the world's worldgen uses `datapack::discover` with
`enabled_ids`. A pack's biomes therefore never reached the table
`WorldgenOverrides::finish` tests against, so it dropped every new biome
("cannot add biomes yet") and the generator then failed on a material
rule naming one ("unknown material biome").
`world_loading::packed_biomes` builds the table from the same list the
worldgen is built from, and the completion registries read the same
overridden ids instead of `DataStore`'s files, so `/locate biome` accepts
and completes a packed biome.
Verified with Terralith 2.6.5 as the world's only pack: the world opens,
`/locate biome terralith:alpine_highlands` finds it, and `/locate biome
#c:is_lush` reports `terralith:cave/underground_jungle`.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
`Biomes::with_packs` (the built-in table with a world's pack files over
it, as `RegistryDataLoader` layers registries) had no callers, so a
packed biome had no client data and `Biomes::id` resolved it to plains —
every packed biome rendered, spawned and saved as the fallback.
The table is built when a world opens, from the same pack overlay the
content tables use, and every biome lookup goes through it
(`world_biomes!`, a field-level borrow for the loops that also take
`&mut self`). `world_loading::make_streaming` builds the same table for
the generator, so the ids the terrain names and the ids the client draws
agree.
A world with more than 128 biomes needs the palette's 8 bits, which the
earlier commit gives it.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
`command-registries.json` lists the game's own ids, so `/locate biome`
refused a packed biome before searching and offered no packed biome or
tag in its completions. The packs' `worldgen/biome`, `structure`,
`structure_set`, `template_pool` and `processor_list` ids now join those
registries, and so do their worldgen tag ids.
The tag lookup behind `#tag` (`/execute if biome`, `/locate biome #tag`)
read only `data/<ns>/tags/worldgen/biome` from the resource root; it now
reads the world's packs too, in resource-stack order, with a `replace`
file dropping what the lower sources put in (`TagLoader.load`). That also
makes a pack's structure tags work.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
`Strategy::biomes(7)` tops out at 128 biome ids; vanilla ships 67 and a
pack like Terralith adds a hundred more (`Biomes::with_packs`), which the
global palette's 8 bits hold.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
The sky reads inverse_view * ray, so its mirror is F * inverse_view; the column operation used for view_proj * F left a pitched camera's sky wrong: looking down 40 degrees mirrored a sky looking down. A test compares the rays.
A scan for the water plane reads up to 60 000 blocks on the game thread; it now runs when the camera moves 8 blocks or every 5 seconds.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
A normal within 0.99 of an axis (about 8 degrees) snapped to it, which flattened flowing water's gentlest slope (about 6 degrees). The snap now reaches about 3.6 degrees, still enough for distant block faces, and the doc says so.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
NativeFrame::write skips title frames, so the first world frame after a title screen, a pause or a world load reprojected from a stale camera with a delta_time of seconds. Past half a second the frame counts as the first. The easing is a pure function, tested.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EBqm1v5zTsDGwuZF6XKFjL
Like a server's streaming, a headless game's doesn't mesh sections it can't draw. Measured
on a single-player world after loading: 0 sections meshed instead of 21287, 853 MB resident
instead of 1117 MB, and about 15% of a core in steady state instead of 20%.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Main reached NATIVE_HOOKS_API 36. native_visuals and native_emissive
(API 8 and 9 on this branch) now share API 37, in the module doc and
docs/native-mods.md. The emissive layer moves to the terrain shader's
@location(9), after main's fade, local and visibility.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Refactor the pause screen button layout to correctly handle the
`pause_screen_additions` tag and the `minecraft:server_links` fallback.
The new implementation uses a helper function `pause_grid` to calculate
the grid origin and ensures that the custom additions button (ID 14)
is rendered as a full-width row, properly pushing down subsequent
menu options.
Additionally, implement `Minecraft.getQuickActionsDialog` to support
opening the quick actions dialog via the `key.quickActions` binding.
Update port maps to reflect the new implementation details for
`PauseScreen` and `Minecraft` quick actions.
A DialogScreen is a screen in vanilla, so KeyMapping.releaseAll runs.
Horizon kept Screen::Playing, dropped key and right-button releases, and
left W and use held after the dialog closed.
`horizon-client --headless` (`horizon-ctl launch --headless`) runs where no display server
is: winit's event loop needs one even for --offscreen, so the game runs its own loop
(game/headless.rs) through the EventLoopCtl handle, one frame a tick. Each step is
Game::idle, which runs the frame through the RedrawRequested path with WindowId::dummy();
handle_window_event takes that id when there is no window. No renderer is made and audio
is silent; the loading screen is skipped (synchronous startup). It counts as offscreen for
everything else. Quitting saves the world as a window's game does (Game::shut_down, now
shared with winit's `exiting`).
Measured on a Ryzen 7 3700X with no DISPLAY: a single-player world opens through the
server thread, the control socket moves, commands and menus work, no GPU device is opened,
about 1.1 GB resident, and in steady state the client's own thread uses 1.6% of a core.
Screenshots and clicks at GUI points answer "no renderer yet" until vision (#4951).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- Track active animated sprite layers per section and cull animation ticks and GPU uploads for sprites not present in visible or shadow sections.
- Mask OPAQUE_LAYER (bit 15) when collecting animated layers from mesh quads so opaque animated blocks (magma, lava, prismarine, etc.) are animated.
- Include active animated layers from visible LOD tiles (and any fading LOD tiles).
- Unit tests for inactive sprite culling and opaque animated face layer collection.
Co-Authored-By: Antigravity <[email protected]>
An update that stopped between copying the new binary and moving it into
place left <name>.update-<pid> in the install folder, and the start-up
clean-up only matched .old- files.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A headless client (#4951) has no display, so it cannot run winit's event loop. The game's
methods used winit's ActiveEventLoop only to stop the loop and to say when to wake next:
they now take `&dyn EventLoopCtl` (game/event_loop.rs), which winit's loop implements.
window_event and about_to_wait keep their bodies, now the inherent handle_window_event and
idle, and the ApplicationHandler methods forward to them. The control socket's and the pad
menu's synthetic input call handle_window_event. Nothing changes for a window's game.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
#4231 (on main) already returns the replaced piece in creative and
discards it there when the inventory is full. What this branch still
adds: survival throws the piece that no longer fits like Player.drop
(stack, false) instead of spawning it motionless at the eyes.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Main split open_device into open_with_retry and open_backend (#4064). The
CPU-adapter check now sits right after request_adapter and fails the
backend, so the non-conformant retry doesn't pick the same adapter.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Right past the render distance, far terrain showed as big flat-topped
boxes with grey stone sides: coarse ancestors standing in while the
sampling order filled a ring ahead first and refined broad tiles every
other job. Main now samples only the plan's leaves, nearest first; this
pulls that order into `sample_order` (no change) and tests it:
- At render distance 12, cells are 1 block wide to 256 blocks out and
2 to 512 (Reduced: 1 to the near terrain's edge, 2 to 256).
- Loading one job at a time, the ring within two chunks of the near
terrain reaches that detail all around after 270-301 jobs, all fine
tiles (7500-8500 density lattice points). The old order took
54000-119000 for the part in view alone.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
MushroomCow.get answers with minecraft:mooshroom/variant, red or brown, and
applyImplicitComponents reads it back, so a mob spawned from an egg keeps
its colour. finalize's species_variant_component listed wolf, cat, frog,
pig, cow, chicken and zombie nautilus but not the mooshroom: the component
was never reported, variant_name fell through to the frog's table and
variant_index could not recognise "brown".
The save format is untouched: a mooshroom still stores Type, not variant.
The ticks == 0 guard never let a dropped item reach ItemEntity::tick, the
only place ticks counts up, so every drop froze: no gravity, pickup delay,
merging or despawn. An item now waits one tick only when its id is past
the last one the tick saw after its functions ran.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Main's OIT_ENTITY shares the layer's high bits with the HDR tier: the overlay
consumers mask 0xffff, vs_lightning masks the tier, and a test pins that a tier
keeps the material.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- options_ui: Space/Enter on FocusId::Cycle now emits Event::ButtonClick
before cycling (was silent before)
- options_ui: wheel on a keyboard-focused slider steps it; wheel on a
keyboard-focused cycle button cycles it (Horizon improvement; Vanilla 26.3
does not intercept the wheel for sliders at all)
- graphics_ui: same wheel improvement for the graphics settings page — when
can_change_value is set, wheel steps the focused slider or cycles the focused
cycle item instead of scrolling the list
- title.rs: confirm_key_size returns (Action, bool); the bool is true for any
enabled focused button press, false for disabled buttons or no-ops
- game.rs: title keyboard confirm plays the click sound from the bool, not just
on Action::Play
- startup_menu.rs: extract .0 from confirm_key_size return
- Tests: options_keyboard_space_enter_emits_button_click_for_cycle_and_default
and title_screen_keyboard_navigation_and_activation updated and passing
Co-Authored-By: Antigravity <[email protected]>
Review fixes on #4577:
- The window's Game on the integrated-server path kept `world_biomes`
unset: the chunks it draws carry the packs' biome ids past the built-in
list, but nothing named them. `open_integrated` builds the world's
table now, as `finish_open_world` does for the load path.
- The integrated path's own encoders read `res.biomes()` past the world's
table: `lan_host`'s chunk encoder and `/execute if biome`'s naming go
through `world_biomes!` now, so a packed biome keeps its name and its
place on the wire (a guest still needs the table sent over the
connection; a remote server join clears the last world's table so no
stale names linger).
- A tag in a sub-directory of its registry never matched: `content_id`
attached the namespace to the file's last segment
(`tags/worldgen/biome/has_structure/minecraft:village_plains`), while
the lookups build `tags/{registry}/{ns}:{path}`. The store key is the
tag id now (`tags/worldgen/biome/test:sub/mytag`), `PackOverlay`
gained `tag_kind` for the registry walks, and the readers match.
Co-Authored-By: GLM 5.3 (ZCode) <[email protected]>
`WorldgenOverrides` and `Biomes::with_packs` both take the world's packs,
but the client built the biome table and the worldgen from different
lists: `content_overlay` reads `datapacks.json` (empty until a world
writes it) while the world's worldgen uses `datapack::discover` with
`enabled_ids`. A pack's biomes therefore never reached the table
`WorldgenOverrides::finish` tests against, so it dropped every new biome
("cannot add biomes yet") and the generator then failed on a material
rule naming one ("unknown material biome").
`world_loading::packed_biomes` builds the table from the same list the
worldgen is built from, and the completion registries read the same
overridden ids instead of `DataStore`'s files, so `/locate biome` accepts
and completes a packed biome.
Verified with Terralith 2.6.5 as the world's only pack: the world opens,
`/locate biome terralith:alpine_highlands` finds it, and `/locate biome
#c:is_lush` reports `terralith:cave/underground_jungle`.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
`Biomes::with_packs` (the built-in table with a world's pack files over
it, as `RegistryDataLoader` layers registries) had no callers, so a
packed biome had no client data and `Biomes::id` resolved it to plains —
every packed biome rendered, spawned and saved as the fallback.
The table is built when a world opens, from the same pack overlay the
content tables use, and every biome lookup goes through it
(`world_biomes!`, a field-level borrow for the loops that also take
`&mut self`). `world_loading::make_streaming` builds the same table for
the generator, so the ids the terrain names and the ids the client draws
agree.
A world with more than 128 biomes needs the palette's 8 bits, which the
earlier commit gives it.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
`command-registries.json` lists the game's own ids, so `/locate biome`
refused a packed biome before searching and offered no packed biome or
tag in its completions. The packs' `worldgen/biome`, `structure`,
`structure_set`, `template_pool` and `processor_list` ids now join those
registries, and so do their worldgen tag ids.
The tag lookup behind `#tag` (`/execute if biome`, `/locate biome #tag`)
read only `data/<ns>/tags/worldgen/biome` from the resource root; it now
reads the world's packs too, in resource-stack order, with a `replace`
file dropping what the lower sources put in (`TagLoader.load`). That also
makes a pack's structure tags work.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
`Strategy::biomes(7)` tops out at 128 biome ids; vanilla ships 67 and a
pack like Terralith adds a hundred more (`Biomes::with_packs`), which the
global palette's 8 bits hold.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
- Return first valid dialog's tag instead of Tag::List (show_dialog can't parse List)
- This fixes pause_screen_additions button actually opening the dialog
Predict craftable PlaceRecipe placements on the remote client, look up
ghost recipes by display id, and clear the ghost when the server updates
the open menu so red ghost slots do not stick after a successful place.
Fixes#4549
Co-Authored-By: Composer <[email protected]>
Co-authored-by: Cursor <[email protected]>
Judged in tools/hdr_gallery.py's gallery under the tuning standard. Flames, lava,
crying obsidian, the froglights, the jack o'lantern and the larger amethyst shine all
over; brewing stand, enchanting table and the smaller buds go bright; candle cakes and
the ender chest don't shine (the cake's frosting would in daylight). The bright-parts
mask goes to 0.32 of the way from a sprite's darkest pixel to its brightest.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The tier windows judge a pixel by a fixed brightness, tuned on vanilla's
textures: a resource pack with a lighter torch handle lifted the handle
too. When HDR output turns on, a small pass now measures every block
sprite's darkest and brightest opaque pixel (gfx/hdr_stats.rs), and
terrain marks a light source's pixel only when it is as bright as its
sprite's threshold, set between the two: for most light sources near
their brightest (their bright parts shine), for those marked WHOLE at
their darkest (they shine all over). The tier windows still decide how
much. Nothing is measured with HDR off.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- Restore Bug Report and Feedback icons with click handlers (open confirm_trusted_link)
- Fix dialog leak between worlds: clear sync data on world close (forget_sync_data)
- pause_screen_additions: match vanilla behavior (single button opening dialog_list)
- Fix clippy warnings: use ? operator with as_ref()
- Fix tests: update pause_buttons/pause_icons calls for new signature
Completing a subtree used to mark every active descendant dirty, so the
final chunk of a cold view remeshed the whole view. A completion that does
not change a tile's own drawn state now touches no mesh.
Cutout publication waited for every mesh job in flight, so continuous
loading could defer it indefinitely. It now tracks the active tiles whose
meshes must stop clipping the changed columns; a retired waiter hands its
wait to its active descendants, and later batches do not delay it.
An active tile index (`unmeshed`) replaces the all-active scan in removal
gating, dispatches coalesce while the pool is saturated, and the neighbours
behind a split tile's shared face are no longer remeshed. A saved summary
without completed detail no longer releases a covered empty stand-in.
Co-Authored-By: omp (DeepSeek V4.1 Flash) <[email protected]>
handleLogin prepares the key pair only when the login packet says the server
is in online mode, and tick asks shouldRefreshKeyPair only once there is a
chat session. sendCommand sends ServerboundChatCommandSignedPacket whenever
the command has signable arguments, with no signatures when there is no
session (ArgumentSignatures.signCommand drops null signatures).
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Two decorations reaching the same chunk an older version saved ran on their own
snapshots and the last write won; a chunk's features each see what the
one before placed. The proto chunks already exclude each other by being
checked out: the final chunks a running decoration writes into now do too.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Completing a subtree used to mark every active descendant dirty, so the
final chunk of a cold view remeshed the whole view. A completion that does
not change a tile's own drawn state now touches no mesh.
Cutout publication waited for every mesh job in flight, so continuous
loading could defer it indefinitely. It now tracks the sequence of the batch
that replaces the affected tiles; later batches do not delay it.
An active tile index (`unmeshed`) replaces the all-active scan in removal
gating, dispatches coalesce while the pool is saturated, and the neighbours
behind a split tile's shared face are no longer remeshed. A saved summary
without completed detail no longer releases a covered empty stand-in.
Co-Authored-By: omp (DeepSeek V4.1 Flash) <[email protected]>
starts_at returned nothing for a chunk an older version saved, so the
explored part of an upgraded world had no structures to find. It now
returns the starts in the chunk's structures.starts (name and bounding box,
no pieces: they are not rebuilt).
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
A chest or spawner of a structure piece reaching into a chunk an older
version saved had only its block copied; the generated block entity, with
its rolled contents, now goes with it.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
region_bits computed outside the lock with no in-flight guard, so every
worker that asked for a region meanwhile decoded its chunks again.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The F3 overlay is built on the window thread every frame; locking the
shared query worker froze the frame behind /locate biome or /place jigsaw.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The far-terrain scheduler evaluated the whole quadtree on the frame thread
for every finished sample, cloned every active tile's columns into every
mesh job, and handed a whole selection batch to one worker.
Selection is now incremental: a finished sample re-evaluates only its
ancestors (O(tree depth)) and marks only the tiles whose faces or clipping
actually change. The sample queue is a nearest-first priority heap rebuilt
when the view moves a chunk, not scanned and sorted every frame. Mesh jobs
carry one immutable snapshot and borrow columns by reference instead of
cloning them; a pool of workers meshes tiles in parallel with a priority
queue and revision checks. Bounds are tracked while emitting vertices on the
worker, so the frame thread no longer walks every vertex of every uploaded
tile. Removals wait until every overlapping replacement has its mesh, so
fade-out cannot open a hole, and both worker pools are stopped under their
own wait locks.
Selection stays an exact match for Plan::select after every sample, across
camera moves and coverage changes (test); edge-neighbour dirtying, active
descendant redraws and removal gating have their own tests. Column spans are
inline (SmallVec), removing almost all per-column allocations on the
sampling workers.
Co-Authored-By: omp (DeepSeek V4.1 Flash) <[email protected]>
- Re-add clippy::clone_on_copy allow for UniformLocation in command.rs.
- Format indirect_execution expression in adapter.rs to satisfy rustfmt.
- Note in gfx.rs comment that multi_draw requires indirect first instance
support (OpenGL 4.2 + GL_ARB_shader_draw_parameters) in addition to 4.0 indirect draws.
Co-Authored-By: Antigravity <[email protected]>
Forgetting the only profile on the last page left the page empty until
Previous was pressed: the page is now kept to the pages the list has.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- A window that failed to open was tried again every frame, each time
re-laying out every player. Its players now play in Window 1 and the
window is forgotten.
- Vsync and the frame pacing's refresh rate follow the first shown
window, not a hidden Window 1; Window 1's visibility is kept in a
field, so Wayland (which can't report it) isn't told every frame.
- Only free windows remember where they were: a monitor's rect made a
later free window with its number open monitor-sized.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
route_pads released P1's pad hardware every frame, so with P1 alone the
reader lent for P1's turn lost its rumble and DualSense link a frame
after they started. It now releases only when nothing is lent.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A controller first pressed with a trigger got a player, and the held
trigger then attacked or used: only buttons were held back.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A LAN player who logged in under a registered seat's name (P2) passed
the integrated server's gate and the full-server check and held the
owner's permissions given to that seat. The gate now admits a seat's
name only from the memory connection (MEMORY_ADDRESS, port 0).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A refresh or a launch check holds the account lock across network calls.
Status reads (account list, the account page) now wait at most 3 s for
it, then say another operation is running, instead of waiting out a
network timeout. Writers still wait.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Translate the account status messages, print a failed `account refresh`
once, fetch a profile head again a minute after its skin download failed,
and check every account's sign-in at most once a minute on page switches
(an account added meanwhile at once), so switching pages no longer reads
the keychain each time.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- When clicking a pause_screen_additions button, first close the pause screen (return to Playing + capture mouse)
- Then show the dialog
- This matches vanilla behavior where pause menu closes when opening a dialog from it
- Fix button Y position: pause_screen_additions buttons now appear at row 3 (y0 + 122),
above Options/World Options, matching vanilla behavior
- Fix click handler to use correct Y position for additions buttons
- Update pause_icons to account for additions height
- Button label uses external_title (falls back to title if not set)
- Add pause_screen_additions() to read dialogs from the minecraft:pause_screen_additions tag
- Display dialogs with external_title and pause:true as buttons in the pause menu
- Clicking these buttons opens the corresponding dialog
fix: clear held keys when opening a dialog to prevent stuck movement
The client uses #4214's SignatureCache and its handling of a bad packed id, instead of its
own cache. markMessageAsProcessed runs where the chat takes a message, past any Chat Delay,
with whether it was shown; handleDeleteChat now calls LastSeenTracker::ignore_pending.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- AbstractSliderButton / OptionInstanceSliderButton parity:
- ArrowLeft / ArrowRight adjust slider values when in can_change_value mode
- Space / Enter toggles can_change_value focus state
- Escape cancels can_change_value before closing the options screen
- Mouse wheel scrolling over a slider steps its value
- Render state parity: track highlighted only when focused and !can_change_value,
handle highlighted when hovered or can_change_value
- Dialog sliders: support ArrowLeft / ArrowRight nudge and mouse wheel scrolling
- General UI keyboard navigation: Tab/Shift-Tab cycle and 2D spatial Arrow navigation
Co-Authored-By: Antigravity <[email protected]>
OldMinecartBehavior.pushAndPickupEntities also picks up living entities
in the cart's box, 0.2 wider on each side. Horizon only pushed other
carts, so a cow shoved into a minecart stayed on the ground. Eligible
mobs now startRiding an empty cart on contact; players still click to
board, iron golems still push, and a cart that already has a rider
pushes instead. Entity.canRide refuses a warden and a boarding cooldown.
sync_riders treated only boats and mobs as vehicles, so a boarded cart
would drop its rider the next tick. Riders sit on the cart's passenger
attachment, get out when it breaks, and are saved under Passengers.
Vanilla: OldMinecartBehavior.pushAndPickupEntities, Entity.startRiding,
AbstractMinecart passenger attachment, Entity.positionRider.
Co-Authored-By: OhMyAgent (grok-4.6) <[email protected]>
ServerPlayer.saveParentVehicle writes RootVehicle {Attach: UUID of the
immediate mount, Entity: the root vehicle saved with its passengers},
only while the root has exactly one player passenger
(Entity.hasExactlyOnePlayerPassenger). PlayerList.remove then, for such
a player, stops riding and removes the root and its passengers from the
world (UNLOADED_WITH_PLAYER), so the ride leaves with the player and
ServerPlayer.loadAndSpawnParentVehicle spawns it again on the next login.
PlayerList.removeAll does this for each player in turn on shutdown.
Horizon only wrote Attach and left the vehicle in the world, so a vanilla
reader got no vehicle back and the horse or boat stayed behind.
- persist.rs: vehicle_root, vehicle_tree_nbt (mob stack, boat with its
mobs, minecart), parent_vehicle_tag with the one-player check, and
remove_vehicle_tree.
- lan.rs: take_parent_vehicle on a guest leaving (HostEvent::Left) and
save_leaving_guests for lan_close, stop_integrated and stop_dedicated;
the tree goes in the guest's saved compound, the guest unrides.
store_owner writes the full RootVehicle for the integrated owner.
- playerdata::store takes the RootVehicle compound; save no longer takes
a vehicle. sync_level writes the tree for the local player.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Entity.getFluidInteractionBox asks the vehicle, and
AbstractBoat.modifyPassengerFluidInteractionBox keeps only the part of the
passenger's box above the boat's top, unchanged when the boat is underwater and
absent when the boat reaches the passenger's top. The rider's fluid interaction
now uses that box instead of clearing the eyes afterwards, so the body
immersion, in_water and the eyes all follow vanilla and a boat floating on water
no longer drowns its passenger (#4001).
Entity.updateFluidOnEyes finds no fluid on the eyes of a rider whose eyes
are inside a boat that is not underwater. The server-side tick of a LAN
guest ignored that, so a boat bobbing in water drowned its passenger.
Quick Actions menu (key.quickActions / G) now reads dialog definitions
from enabled datapacks (data/namespace/dialog/*.json with type
multi_action). Each action entry becomes a button; clicking runs the
action id as a command.
Also exposes DataPack::list and DataPack::read as pub so the game can
scan datapack dialog files.
Co-Authored-By: Nemotron AI agent <[email protected]>
Adds a non-pausing Quick Actions menu opened by the G key (key.quickActions,
already defined in keybinds). The screen shows buttons for quick player
actions and releases the mouse while open. Camera action works; emote,
quick chat and gesture are placeholders pending datapack-driven content.
Co-Authored-By: Nemotron AI agent <[email protected]>
A chunk with a final neighbour (a chunk of an older world, loaded from the
save) used to be left undecorated: no trees, ores or structure pieces in the
strip along the old chunks. The streaming now decorates it with the final
neighbour's blocks in the region and writes what the features changed into
the final chunk, as vanilla's features write into the chunks around them.
The structure starts of an old chunk are not made again from the seed.
test_blending_parity.py --phase full compares the decorated chunks around
old ones with vanilla.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
`MinecraftServer.tickChildren` runs `commandFunctions` before any
`ServerLevel.tick`, and `EntityTickList.ensureActiveIsNotIterated` gives
an entity added while the list iterates its first tick on the next tick.
Together they mean a tick function that polls for the drop a block just
made still finds it with `Age: 0` — the pattern data packs use to hook a
block break. Horizon ticked its functions inside the entity pass, after
the items: the drop had counted its age up already and the poll found
nothing.
The functions now run at the top of the tick, before the levels, with
the world's saved data read first (a load tag's `data modify storage`
must land after the save is restored, not before). An item entity added
this tick first ticks the next one.
`type_specific/player` also matches players only, as
`PlayerPredicate.matches` starts with `instanceof ServerPlayer`, and
honours `input` (`InputPredicate`; `input.sneak` is `Input.shift`, the
held key, not the sneaking pose).
Verified with the Vein Miner data pack: its tick-driven chain mining
takes both ores it is asked to, and its `veinminer:sneaking` predicate
reports false while the player is not sneaking.
Co-Authored-By: ZCode Agent <[email protected]>
SavedOldWorld finds the saved chunks with blending_data (or DataVersion
below 4997) per region, reads their border columns and the biomes they
keep, and the overworld generator blends the new chunks around them with it.
tools/old_world_e2e.py plays a world made with the old_world_fixture
example, with and without blending (HORIZON_DISABLE_BLENDING, vanilla's
DEBUG_DISABLE_BLENDING).
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Show a dash when no scene-load measurements exist; preserve measured zero and the mean of available scenes.
Co-Authored-By: OpenAI Codex GPT-6.1 Sol <[email protected]>
Blender.of, BlendingData (calculated from an old chunk's border columns),
blendOffsetAndFactor, blendDensity, blendBiome and the carving filter, with
fastutil's entry order for the weighted sums. The generator takes an
OldWorld (the chunks an older version saved) and blends biomes, the density
graph's blend_alpha/blend_offset/blend_density and the carving mask with it.
tools/test_blending_parity.py compares the terrain around old chunks of
another seed block for block and quart for quart with vanilla.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
A pack can let players turn its water reflections off: [reflection] option
names an option the mirror is drawn only while it is on (a bool on, a number
above 0); without it, the mirror is drawn whenever the material reads it.
The sample pack shows it.
While mirroring, solid quads were all drawn, since the mirror's eye sees the
undersides the camera does not. Sides::visible_from keeps the quads that face
any eye from the camera straight down to its mirror below the plane, which
differ only in y, for the sections and the distant tiles.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01NbhTrA2NFEcA5CdtUTS1eH
NoiseBasedChunkGenerator.addDebugScreenInfo: DebugEntryChunkGeneration
adds 'Density N: .. T: .. V: .. C: .. E: .. D: .. W: .. PV: .. PS: ..' (the
overworld's debug_functions, three decimals) at the feet where the game
runs the level's generator: a single-player world on its server thread.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Water can reflect only what is on screen, which loses the sky over a lake and
everything above the picture's top. Within API 0.1, additively, the material
stage gains reflection and reflection_sampler: the world as a camera mirrored
about the horizontal water plane around the camera sees it, sky and solid
terrain at half resolution, cut away below the plane. Frame gains
water_plane, has_reflection and reflection_resolution.
The mirror is drawn only for a pack whose material function reads
reflection (CheckedStage::reads), so other packs and vanilla pay nothing and
the game scans for the plane only then: the median standing-water surface of
a grid of columns around the camera, rescanned when the camera changes block
or each second. Sections are culled against the union of the camera's and
the mirror's frusta; while mirroring, solid quads are not left out by facing
and occlusion culling is off, since both follow the camera's eye.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01NbhTrA2NFEcA5CdtUTS1eH
Online servers with enforce-secure-profile=true (26.3's default) refused every chat
message and /msg from Horizon ("Chat disabled due to missing profile public key").
game/remote_chat.rs (ClientPacketListener):
- setKeyPair: after each login packet, and when the key pair is renewed, a chat session
(random id, chain from link 0) goes to the server as ServerboundChatSessionUpdatePacket,
only when the server plays the account under its own profile id: an offline-mode
server's offline id gets none, so chat stays unsigned there as before.
- sendChat / sendCommand: chat and each message argument of a command (SignableCommand.of
over the server's tree) are signed with the same time, salt and last seen, one chain
link each; every packet carries the last-seen update.
- handlePlayerChat / markMessageAsProcessed: other players' packed last seen through the
signature cache, and a ChatAck once more than 64 messages wait.
- sendUnattendedCommand: a clicked chat line, a dialog's button and a mod's command stay
unsigned (vanilla asks before signing one; there is no such screen here).
tools/secure_chat_client_e2e.py: vanilla 26.3's server and a mock of the services that
issues certificates; online with enforce-secure-profile (chat logged as secure, a /msg
whisper delivered, nothing refused) and offline (unsigned, as before). The net harness
takes its chat lines from HORIZON_NET_CHAT and reports whether it signed.
Fixes the LocalChatSession.createMessageEncoder, SignedMessageChain.encoder,
ClientPacketListener.setKeyPair and ClientPacketListener.sendCommand rows of
horizon-rust/horizon#3520 and horizon-rust/horizon#3518.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- auth.rs: the services' certificates endpoint (player/getCertificates, discovery), POSTed
with the account's access token: the PKCS#8 private key, the X.509 public key,
publicKeySignatureV2, expiresAt and refreshedAfter (AccountProfileKeyPairManager.
fetchProfileKeyPair). Fetched on a background thread at start and on each join, kept in
memory like the token (no profilekeys file), renewed past refreshedAfter (checked at
most hourly), and kept as it was when a fetch fails.
- horizon-command: ParseResults::signable_arguments (SignableCommand.of), moved out of
the LAN host's handleSignedChatCommand so a client can run it over a server's tree.
Nothing sends the key yet: the client's wiring follows.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The protocol pieces a client needs to sign its chat as vanilla 26.3's does, beside the
server's side that is already here:
- rsa.rs: PKCS#8 (and bare PKCS#1) RSA private keys, as the services hand out a player's
chat key, and RSASSA-PKCS1-v1_5 signatures (Signer.from), checked with the public key
before they leave. Signing is deterministic, so a test compares OpenSSL's bytes.
- secure_chat.rs: SignedMessageChain.Encoder (LocalChatSession.createMessageEncoder), the
client's LastSeenMessagesTracker, and MessageSignature.Packed.unpack against the cache.
Tests run what the encoder signs through the server's Chain, and the tracker's updates
through LastSeenMessagesValidator.
Nothing calls them yet: the account's key pair and the client's wiring follow.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A pack's material stage cannot tell water from glass, a fluid's top from its
side, or a lamp from stone, and has no surface frame for normal maps. Within
API 0.1, additively, MaterialInput gains face and fluid (0 none, 1 water,
2 lava), emission (the block's HDR tier as 0..1), and the geometric normal and
tangent from the derivatives of the position and uv, snapped to an axis on
block faces so distant faces do not shimmer.
The meshers mark fluid quads in the high nibble of the colour's alpha byte,
which terrain.wgsl does not read (it reads only the emission in the low
nibble); the distant LOD tiles get the same flags.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01NbhTrA2NFEcA5CdtUTS1eH
A temporal pass (TAA, denoising) needs its own output from the frame before.
Within API 0.1, additively, the post prelude gains history (group 3, binding 8):
after the passes have run, what the last pass read is copied into it, so a
temporal pass put second to last reads its previous result next frame. It is
black on the first frame and with a single pass. Binding 8 leaves 6 and 7 for
the shadow map.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01NbhTrA2NFEcA5CdtUTS1eH
Temporal passes (reprojection, motion blur, accumulation) need where a point was
last frame; packs that adapt exposure or fade underwater fog need the eye's light
and medium without their steps. Within API 0.1, additively, Frame gains
previous_view_projection, camera_motion and delta_time, filled by NativeFrame from
the frame before, and eye_light_smooth, in_water and in_lava, eased with
half-lives of 2 s and a quarter second. The frame stays a multiple of 16 bytes.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01NbhTrA2NFEcA5CdtUTS1eH
The material stage could shade water and glass only from their own texture: it
had no way to know what was behind them, so depth-tinted water and refraction
were impossible. Within API 0.1, additively:
- MaterialInput.screen: the point's place on the screen, 0..1.
- opaque_depth (binding 6) and opaque_color (binding 11): the world's depth and
colour as the solid terrain and entities left them. The world pass ends before
translucent terrain, the targets are copied, and a pass that loads them goes on;
the depth is stored for that while a material stage runs. The colour can be
copied from the scaled world target only: drawing straight to the window, it
is a black 1x1.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01NbhTrA2NFEcA5CdtUTS1eH
Default off in horizon-graphics.txt. When on, reuse the last shadow map only if
sun/view, world revision, day time, ticks, pack uniforms and settings match,
and skip the cache while entity shadows are drawn.
Co-authored-by: Cursor <[email protected]>
The shadow-cache branch was cut from an older tree and looked like it removed
the published benchmarks folder; put those files back from current main.
Co-authored-by: Cursor <[email protected]>
`account session WHO` renews an account and hands the running game a new
read-once session file, or the public status when it needs a browser
sign-in or a retry. `account sign-in --for WHO` runs the UUID-bound
re-sign-in from the command line. The launcher passes its own path to the
game as HORIZON_LAUNCHER_EXE. `account session` is kept out of MCP: its
file holds a token.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Fetch public skin heads on workers with a bounded, URL-keyed cache and UUID default fallback, report each account operation's actual phase, and bind Sign in again to the account's UUID so a different Microsoft account is rejected before any credentials are stored.
Co-Authored-By: Codex <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Show ready, refreshable, and sign-in-required states using the theme's success, warning, and danger colors, retaining icons and full labels.
Co-Authored-By: Codex <[email protected]>
Basic indirect drawing (glDrawArraysIndirect, glDrawElementsIndirect) is
standard in Desktop OpenGL 4.0+ and GLES 3.1+.
1. In wgpu-hal GLES adapter: decouple DownlevelFlags::INDIRECT_EXECUTION
from compute shaders, enabling indirect execution on Desktop GL 4.0+
(and GLES 3.1+), which unlocks multi_draw in horizon-client-render on
OpenGL 4.2 hardware (e.g. Intel HD Graphics 4000).
2. In wgpu-hal GLES command and queue: avoid pushing draw_count separate
Command variants into the command buffer. Push a single DrawIndirect /
DrawIndexedIndirect command with draw_count. In queue execution, bind
DRAW_INDIRECT_BUFFER once and loop glDraw*Indirect across offsets on the
CPU, eliminating command buffer allocation churn and repeated buffer
re-bindings.
Co-Authored-By: Antigravity <[email protected]>
Compare sign-in states directly instead of comparing a state to an Option's debug output, which continually restarted every account's status checks.
Co-Authored-By: Codex <[email protected]>
Inspect and renew due sessions on workers, expose Refresh and Sign in again actions, and share the operation through CLI and MCP for future launcher-to-game integration.
Co-Authored-By: Codex <[email protected]>
Twenty-nine literals were drawn straight into the window, so the tables never
had them: the game arguments and proxy rows with their explanations, every
text field's placeholder (one place now), Manage versions, the launcher's own
update lines, the release and patch-note empty states and the sponsors line.
Russian in ru.json; the other six tables carry the English source.
Co-Authored-By: Hermes Agent (Nous Research) <[email protected]>
Accent carried one of seven names; it carries a hue now, and none for the
violet, whose tokens stay the palette's. Settings shows a strip of hues
beside the swatches: a click or a drag picks any colour, the swatches stay
as shortcuts, settings.json keeps "violet"/"blue" for those and writes a
picked hue in degrees. The contrast test runs over hues nobody preset too.
Co-Authored-By: Hermes Agent (Nous Research) <[email protected]>
Review: device_event turned P1's GUI scale on for every mouse motion even with no other
player; it now hands the event to P1 directly while there is no compositor.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Upstream moved page titles and tooltips to translation keys: the profile page gets its own
(horizon.pages.split_profiles).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Upstream moved the title actions into title_action (the split guest's Quit guard goes there),
added the server's online mode beside our memory-seat limit bypass, the audio output's device
name (a shared output carries it), and dropped code mods' key hooks.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Light that isn't a block gets HDR output's tier too (entity.wgsl fs_hdr_mark):
- The additive pipelines marked nothing: lightning and the dragon's eyes and death
rays get mark passes of their own (vs_lightning passes the tier on as vs_entity
does). Lightning and the death rays very bright, the dragon's eyes bright.
- End crystals, their healing beam and beacon beams bright. The mark pass now wraps
a beam's v in its sprite as fs_entity does, or it sampled transparent texels.
- Blazes bright, magma cubes and glow squid low, entity flames incandescent.
- Fireworks (sparks and the burst's flash) bright: translucent-layer particles
carry a tier and get a mark pass after their draw.
- Experience orbs very bright, glow item frames incandescent (the item in one stays
only lit).
- The overlay is the layer's low 16 bits everywhere (the entity batch's materials,
the shader packs' plain-run and hurt-colour checks), so a tier never changes how
anything draws, with HDR on or off.
Tuned in game on an AW3423DW (Senjin). HDR off: nothing changes.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A camel's legs used the default walk: they kept swinging while it sat or dashed and eased in at 0.4. Camel.updateWalkAnimation aims at 6 times the distance moved (at most 1) only in the standing pose with no dash animation running, otherwise at zero, and eases at 0.2. The remote client's own walk update uses it for camels too.
A leashed sitting camel stayed down however hard the lead pulled. Camel.onElasticLeashPull stands it up when it is sitting, not changing pose, and the standing box fits where it is (canCamelChangePose, wouldNotSuffocateAtTargetPose). AbstractHorse.onElasticLeashPull, which it calls first, also ends a horse's grazing.
A horse that landed from a small hop while still rearing ignored the keys until the rear ended, and a donkey or mule never reared on a jump. AbstractHorse.onPlayerJump sets allowStandSliding and calls standIfPossible (every horse that can rear), getRiddenInput only zeroes the input while the horse is standing on the ground without a jump pending and without allowStandSliding, and tick clears it once the horse stands no more.
A horse, donkey, mule or undead horse that reared kept its rider at the fixed seat. AbstractHorse.getPassengerAttachmentPoint adds (0, 0.15, -0.7) times the previous tick's stand animation (standAnimO) to the seat, turned with the horse. The player and a mob rider both slide.
The rider's seat on a camel was a fixed height for sitting and standing, built from the old full-height box, and a mob rider sat at the top of the box when the camel sat. Camel.getPassengerAttachmentPoint takes the body anchor from the current pose box and lerps it through the sit and stand transitions (getBodyAnchorAnimationYOffset), at the camel's age scale. The player and a mob rider both use it now.
A camel that sat down kept its standing box, so it blocked and was hit as if it still stood. Camel.getDefaultDimensions gives Pose.SITTING an adult box 1.43 lower with eyes at 0.845, and a baby box 0.425 tall with eyes at 0.41. Mob::awake_dims now follows the camel's synced pose, on the server and on a client.
A llama landed with the generic fall sound and hurt on any damaging fall. Llama.causeFallDamage makes no fall sound, and hurts the llama and hands the fall to its riders only when the fall is six blocks or more.
A rider took no damage when the mob under it landed hard, and a mob on a mount's back never felt it either. Entity.propagateFallToPassengers hands the fall to every passenger (LivingEntity.causeFallDamage first, AbstractHorse.causeFallDamage after its own hurt and only when hurt), and each passenger runs its own causeFallDamage with its own safe distance. A mount now raises a fall event the game turns into Player.causeFallDamage for the local player and server players, and the mob causeFallDamage for a mob rider.
AbstractBoatRenderer.submit turns a boat that is not under water about the axis (1, 0, 1) by the bubble angle when it is not 0 (Mth.equal, 1e-5), after the hurt tilt and before the model's flip, so a boat wobbles while a column works on it.
AbstractBoatRenderer.extractRenderState copies AbstractBoat.getBubbleAngle(partialTicks), the wobble lerped between the last two ticks, into the render state. boat_geometry now reads it the same way (the drawing follows in the next commit).
The server decides when a column throws or pulls a boat, but the boat its player drives moves on that player's client, so the server's own change does nothing there. AbstractBoat.handleEntityEvent now takes entity events 71 and 72 and runs handleBubbleColumnEffect, which the client applies to the boat it drives.
Boats never rose or sank in a column and a down column let every rider stay. AbstractBoat.tickBubbleColumn now counts the bubble time down on the server while the boat stays above the column, then ejects the riders (down only), broadcasts entity event 71 (down) or 72 (up) and applies the throw or pull. A server's player is let out through guest_vehicle_tick. Clients run the other half: the wobble multiplier grows by 0.05 a tick while the bubble time runs and fades by 0.1 otherwise, and the angle is 10 * sin(0.5 * tickCount) * multiplier.
A boat in an up or down column never changed its vertical speed. AbstractBoat.handleBubbleColumnEffect sets it to 0.6 for an up column and subtracts 0.7 for a down one, on the side that moves the boat (canSimulateMovement).
A boat above a bubble column, with nothing over it, had no bubble state and never splashed. AbstractBoat.onAboveBubbleColumn now marks the boat as above the column with its direction, starts the 60 tick bubble time when it is 0 and, one time in a hundred while the boat is not under water, posts the splash game event on the server and plays the splash sound and particle on the client (it does not push the boat, unlike Entity.onAboveBubbleColumn).
A boat inside a bubble column, with water above it, kept its speed. Entity.handleOnInsideBubbleColumn now moves it up to 0.7 or down to -0.3 (AbstractBoat does not override it), twice a tick as AbstractBoat.tick calls applyEffectsFromBlocks twice. Only the side that moves the boat applies it (canSimulateMovement): the server for a boat nobody drives, the client for the boat its player drives. Boats the window does not tick also get their status refreshed, as getStatus runs on every boat.
Boats had no AbstractBoat.DATA_ID_BUBBLE_TIME (entity data entry 13). The server now sends it with the other synched boat data, and sends 0 again once it drops back, and a client keeps it on the boat.
bubble_column_push found the columns a box touches and pushed in one go, so nothing else could ask which column an entity is above or inside. bubble_columns_touching now returns each touched column with its drag and whether nothing sits above it (BubbleColumnBlock.entityInside); bubble_column_push uses it and behaves as before.
Part of the splitscreen series (#3541). The page a joining splitscreen player picks its
profile on, in horizon-client-render with its own tests, and the profile store it reads:
names kept in the game folder (splitscreen/profiles/<name>/), pages of them, New Profile...
(a player name: 3 to 16 letters, digits or underscores), forgetting one (hidden, its settings
kept), Cancel. Options pages it opens from grey Language and Resource Packs for another
player (the whole process's). Nothing opens it yet: the game side joins next.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Part 9 of the splitscreen series (#3541), on #3801. Each player plays in a window: Window 1
(the game's), Window 2, 3... opened for the players put there, or a borderless window filling
a monitor (over the menu bar on macOS, the system's borderless fullscreen elsewhere). Players
in one window share it. Closing a player's own window drops its players; the game's window
hides while nobody plays in it; titles say who plays where; a window reopens where it was.
Another player's F11 toggles its own window. HORIZON_SPLIT_WINDOWS (testing, until the Setup
screen) says each seat's window: `1,1,2`, or `m2` for monitor 2.
Live: tools/splitscreen_e2e.py adds `windows`.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Part 8 of the splitscreen series (#3541), on #3794. While splitscreen watches the controllers
(HORIZON_SPLIT for now; the switch comes later), one reader serves every player, alone too:
the first controller used is P1's (its rumble and DualSense output lent to P1's turn). In P1's
world another controller pressing a button is offered to join on every view: A joins (as
P<seat>), Y drives P1 when P1 has none, B says no, 10 s and it goes. A controller that
disconnects leaves its player dimmed and asked: the same controller back takes it again, A on
another takes over, B lets it leave. Each view gets a short badge (who, and why: "Keyboard
and mouse" when it takes them).
Live: tools/splitscreen_e2e.py adds `alone` (watching, nobody joins).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
On Windows the game horizon-ctl launch starts inherits horizon-ctl's stdout pipe, which then
never closes while the game runs: subprocess.run waited forever. No pkill there either.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Part 7 of the splitscreen series (#3541), on #3777. While several play, one controller reader
(PadRouter) serves every player's game: a controller nobody has goes, when pressed, to the
first player without one (P1 first); its first press is not passed on; a controller that
disconnects leaves its player without one. Each player's glyphs follow its controller.
Alone, P1 reads the controllers itself, exactly as without splitscreen.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Part 6 of the splitscreen series (#3541), on parts 1 (compositor), 2 (audio), 4 (pointer
ownership) and 5 (layout). With HORIZON_SPLIT=N (testing only, no switch yet), N-1 more
players join P1's world through memory as P2, P3..., each its own Game in this process,
rendered offscreen and composited into P1's window. Keyboard and mouse go to the view
clicked last. A player who quits or disconnects drops out; alone again, P1 draws on its
window as before. Guests never open the controllers (pad.feed); routing comes next.
Live: tools/splitscreen_e2e.py (join, four).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
AMD FidelityFX SDK 1.1.4's frame interpolation (MIT): dilated depth and camera
motion vectors, the estimated depths, the game and optical flow vector fields,
the disocclusion mask, the interpolation and the inpainting, as WGSL compute
passes. Not used by the game yet; the setting and the present thread follow.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- Add ShadowCache struct to store rendered shadow maps between frames
- Cache keyed by sun/moon angle, camera position, and shadow settings
- On cache hit: copy cached textures instead of re-rendering shadow pass
- On cache miss: render normally and store result in cache
- Add pass-timing control command for shader pack GPU profiling
- Expose ShadowTargets fresh/cleared fields for cache access
Benchmarks (seed 1234, render distance 16, time 6000):
- Avg FPS: +23-45% across biomes
- 1% low FPS: +85-167% improvement
- Shadow pass time: ~90% reduction on cache hit
Linux releases ship horizon-launcher-linux-x86_64.AppImage beside the
.tar.gz: one file to download, make executable and run, with the
libraries GPUI needs beside it. glibc, the graphics drivers and the
session's D-Bus stay the machine's (a driver cannot be packaged), so the
image needs glibc 2.35 or newer, like the .tar.gz. appimagetool 1.9.1
is pinned by SHA-256 and run extracted, so building needs no FUSE; the
image itself needs none either, as its runtime mounts it where FUSE
works and unpacks it where it does not.
Inside an AppImage, current_exe names a copy in the read-only mount, so
a launcher in one updates by replacing the file $APPIMAGE names, and
the menu entry it writes starts that file rather than the mount's copy.
The .tar.gz stays what every other launcher updates itself from.
Closes#2110.
Co-Authored-By: DeepSeek V4.1 Flash 1017 (River) <[email protected]>
One Hide, shown after the same three minutes on screen, now sits on the Servers
banner as well as the sidebar card, and hiding either hides both for good.
The on-screen timer counts on every page, so a player who lives in the Servers tab
reaches the Hide too.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The terrain vertex arena only freed a page once every range in it was free, so after a render distance change or far tile churn it kept pages that were mostly empty. On a 2 GiB card the next allocations then spilled to system memory and frames slowed to a crawl. When at least 128 MiB and half of the arena is free ranges, copy the meshes off the sparsest page into the other pages (32 MiB of GPU copies a frame) and let trim free it.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The next start removes horizon-launcher.exe.old-* beside it. The launcher
that just restarted into the new binary is still exiting, so it keeps
trying for a few seconds in the background; a file still in use (another
launcher window) is left for a later start.
self_replace copied the launcher under a random hidden name
(.horizon-launcher.<random>.__relocated__.exe), started that copy to wait
for the launcher to exit and delete it, and deleted the copy through a
hidden cmd.exe. Microsoft Defender quarantines that self-deleting pattern
as Trojan:Script/Wacatac.H!ml (#3175).
Windows lets a running binary be renamed, so the running launcher now
steps aside to horizon-launcher.exe.old-<pid>-<n> in its own folder and
the new binary takes its path. No copy is started, and the crate (with
its start-up hook that ran in every launch) is gone from the build.
2026-10-28 13:57:08 +01:00
1162 changed files with 237621 additions and 13762 deletions
gh pr create --base main --head "$branch" --title 'Stamp native hook API fragments' --body 'Assign API numbers to the pending native hook fragments. Close and reopen this PR to run CI.'
@@ -39,7 +39,8 @@ Game workspace (root `Cargo.toml`); `launcher/` and `crates/horizon-telemetry-se
| Crate | Owns |
|---|---|
| `horizon-client` | The game: a lib (`src/lib.rs`, ~100 flat modules: gameplay, networking, title screens) and the bin (`src/main.rs`→`game.rs` + `game/`), plus `horizon-ctl` (`src/bin/horizon-ctl/`). Also the dedicated server (`--dedicated`) |
| `horizon-client` | Gameplay library (`src/lib.rs`: `game.rs`+`game/`, `entry.rs`) and thin executable (`src/main.rs`). Also the dedicated server (`--dedicated`). Re-exports `horizon-client-lib` for existing mod paths. |
| `horizon-client-lib` | Reusable client systems, `horizon-ctl`, probes and their tests. Its manifest points to the existing sources under `crates/horizon-client/`; the library root is `src/client_lib.rs`. It never depends on gameplay. |
| `horizon-command` | Command parsing (Brigadier): the vanilla tree, argument types, selectors, suggestions, the tree's network form. Re-exported as `horizon_client::command` |
| `horizon-client-core` | The client lib's bottom layer: `text.rs`, `lang.rs`, `snbt.rs`, `nbt_path.rs`, `component.rs`, `scoreboard.rs`, `bossbar.rs`, `gamerules.rs`, `block_entity_nbt.rs`, registry/particle/sound tables. Re-exported under the old paths (`horizon_client::text`...) |
@@ -56,6 +57,7 @@ Game workspace (root `Cargo.toml`); `launcher/` and `crates/horizon-telemetry-se
| `horizon-friends`, `horizon-relay` | Play with Friends: invite codes, encrypted tunnel, the relay ([docs/play-with-friends.md](docs/play-with-friends.md)) |
| `horizon-mods`, `cargo-horizon-mod` | Content mods (`.hmod`: manifest, content fence, no code), the registry index and `.hpack`, and `cargo horizon-mod` (`new`, `build`, `run`, `check`, `registry`) ([docs/mods.md](docs/mods.md)). Code mods are native, built by the launcher ([docs/native-mods.md](docs/native-mods.md)) |
| `horizon-optifine` | OptiFine-format pack files below every feature that reads them: `.properties`, the name/range/block matchers, paths and `optifine/` folders, and custom entity models (`cem/`: `.jem`/`.jpm` parsing and baking) |
| `horizon-shader-api` | Horizon's own WGSL shader packs ([docs/api/shaders.md](docs/api/shaders.md)): `horizon-shaders.toml`, the `.hshader` package; shared by the game, the launcher and the registry |
| `horizon-shaders` | Shader packs (the `shaders/` folder format), rebuilt clean-room from a written spec of the format; no wgpu |
| `horizon-theme` | Design tokens and fonts for the launcher, the game's own menus and the website |
`docs/`, `.github/workflows/`. The website (playhorizon.dev) is the private horizon-rust/horizon-site: its build
reads `README.md`, `launcher/links.json`, `docs/api/`, `roadmap/roadmap.json` and `crates/horizon-theme/tokens.css` from main.
## Write tests with signal
- Test observable behavior, not words in docs, comments or source. Keep checks for machine-consumed contracts.
- Use independent expected values or a vanilla oracle; do not derive the answer from the code under test.
- Test outcomes, not only calls to stubs. Use mocks to isolate behavior, not as the behavior itself.
- Skip tests of trivial constants, derives and getters, or behavior already covered by another test.
- Avoid pinning implementation details; a harmless refactor should not break a behavior test.
- Make slow tests faster without removing cases or weakening their coverage.
- Never delete parity harnesses or tests that pin vanilla behaviour: they are the spec.
## Build and test fast
```sh
python3 tools/pr_ready.py # the self-check: fmt, clippy -D warnings, touched crates' tests, harnesses; paste its block in the PR
python3 tools/pr_ready.py --quick # same without harnesses
cargo test -p horizon-client --lib --bins --tests # skips the 104 examples (pr_ready does this); `--doc` separately
cargo test -p horizon-client --lib mod_content::tests:: -- --ignored # the 5 slow content-root tests, #[ignore]d; pr_ready runs them when mod content, assets, blocks or items change
cargo test -p horizon-client -p horizon-client-lib --lib --bins --tests # skips the 104 examples (pr_ready does this); `--doc` separately
cargo test -p horizon-client-lib --lib mod_content::tests:: -- --ignored # the 5 slow content-root tests, #[ignore]d; pr_ready runs them when mod content, assets, blocks or items change
cargo build --profile fast -p horizon-client # optimized, quick to link: use it to play or run lan_e2e
python3 tools/parity.py --changed origin/main # the harnesses your branch can affect (CI's parity job)
python3 tools/parity.py --changed-files paths.txt --list # which harnesses cover these files, without running them
@@ -91,8 +103,11 @@ python3 tools/agent_control_e2e.py # horizon-ctl / MCP end to end
or the game binary with a fixture env var (`game/dispense_fixture.rs`, `game/sculk_fixture.rs`,
`game/command_harness.rs`, `game/net_harness.rs`). Copy a nearby harness for a new one.
- Probes build into `.local/parity-target/`; harnesses find binaries via `lab.cargo_target()`, never `target/`.
-What a change reruns (of 329 harnesses, measured 2026-10-23): a file of the game binary (`game.rs`, `game/`)
~25, a file of the clientlib ~103, of `horizon-world` ~103–109, of `horizon-core` ~119. Plan for it.
-Harness selection uses Cargo dep-info and reachable-code narrowing (`tools/parity_reach.py`).
`horizon-client-lib` builds the control tool and client probes without gameplay; their dep-info
excludes `game.rs` and `game/`, so gameplay edits do not rebuild or select those probes.
Source paths remain in `crates/horizon-client/`. Measure actual harness selection on refreshed
records with `tools/parity.py --changed-files paths.txt --list`; historical counts are not a guarantee.
- Linux needs `libasound2-dev libudev-dev`; `--no-default-features` drops audio and controllers (the server build).
- Run and drive the game (starts the `horizon-client` beside `horizon-ctl`; other arguments go to the game):
@@ -117,8 +132,11 @@ python3 tools/agent_control_e2e.py # horizon-ctl / MCP end to end
## horizon-client: where things live
`L:` = lib (`crates/horizon-client/src/`, or `crates/horizon-client-render/src/` for the modules that crate holds, under
the same `horizon_client::<module>` paths), `G:` = bin (`crates/horizon-client/src/game/`). The lib can't see
`Game`; bin code reaches the lib as `horizon_client::<module>`.
the same `horizon_client::<module>` paths), `G:` = gameplay (`crates/horizon-client/src/game/`), also in the lib in normal builds.
Gameplay keeps its `horizon_client::<module>` imports through a self-crate alias. The executable
contains only the allocator, mod API marker and a call to `entry::main`. For unit tests, the lib
excludes `entry` and `game`, and the executable test target owns `game` so its process-wide test
state stays separate from the library tests. This compiles gameplay twice for executable tests.
| Subsystem | Files |
|---|---|
@@ -149,7 +167,7 @@ the same `horizon_client::<module>` paths), `G:` = bin (`crates/horizon-client/s
- `struct Game` (game.rs, ~290 fields) holds everything. Its methods are spread over ~180 files: game.rs keeps
the core (tick, render, input, inventory, world open/close) and each `game/<feature>.rs` adds an
`impl Game<'_>` block for one feature, with its own state types. Fields are private to the bin.
`impl Game<'_>` block for one feature, with its own state types. Fields stay private to the gameplay module and its descendants. Gameplay tests run in the gameplay library, separate from reusable-library tests.
@@ -742,13 +742,13 @@ Open: the GUI Scale is clamped to 2 after a restart under Hyprland, while XWayla
</details>
#### AppImage and a Nix flake
#### A Nix flake
Install Horizon on any Linux distribution as an AppImage, and on NixOS through a Nix flake.
Install Horizon on NixOS through a Nix flake.
<details><summary>Technical</summary>
A player's idea (#horizon-ideas). An AppImage of the launcher and game from the release workflow, and a flake.nix packaging the release binaries with their runtime libraries (Vulkan loader, X11/Wayland, ALSA), the way horizon-parity's NixOS setup runs them today.
A player's idea (#horizon-ideas). The launcher's AppImage ships with each launcher release ([packaging/linux/appimage.sh](packaging/linux/appimage.sh)): one file with the libraries GPUI needs, which replaces itself when it updates. A flake.nix packaging the release binaries with their runtime libraries (Vulkan loader, X11/Wayland, ALSA) is still missing, the way horizon-parity's NixOS setup runs them today.
description="The dedicated server's chat filter: its [chat_filter] settings, the slur list's matcher and the mutes, with no game state (docs/chat-filter.md)"
"Shows your name above your head in third person, as other players see it. Hidden when the server hides your team's names or you're invisible."
}
Item::HideBlockParticles=>{
"Hides the debris particles that burst from a broken block and crumble off the one you mine. Other particles stay."
}
Item::AutoReconnect=>{
"Rejoins a server after a lost connection or a restart, never after a kick or ban."
}
Item::UiSoundReserve=>{
"Sounds in the world leave 16 of the 247 sound channels free, so a flood of them (a big explosion) can't silence menu clicks."
}
Item::MouseTweaks=>{
"Right-drag places one per slot, left-drag gathers the held item, Shift + left-drag moves every slot."
}
@@ -471,6 +505,9 @@ impl Item {
Item::DiscordPresence=>{
"Shows Playing Horizon on your Discord profile while Discord runs: menus, single player or multiplayer, the dimension and the time played. Never a world's name or a server's address."
}
Item::Splitscreen=>{
"In a world, another controller can join as a new player: press A when it asks, then pick a profile and where to play."
}
Item::AgentControl=>{
"Your own AI agent can play this game through horizon-ctl or an MCP client, on this computer only."
/// The cost tag and the description beside a control, wrapped to two lines.
pubfndescription(
&self,
Loaded 100 of 1162 files, more files were not shown because too many files have changed in this diff.
Show more
Reference in new issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
levzzz