The vignette, drawn over every frame, went from one texture sample to up to
four texel loads and per-pixel address arithmetic: +0.85 ms of GPU a frame on
an HD 620 (integer modulo, removed in the commit before), still +0.4 ms after
it.
- A blurred sprite gets a one-texel border in the atlas, copied from its
opposite edges: the hardware's bilinear sample of the atlas with the
entity group's linear sampler is then GL_LINEAR with repeat, edges
included. Clamp-to-edge keeps the sample point half a texel inside the
sprite, which is what the hardware's clamp does for both filters.
- Each sampling has its own pipeline, the shader specialised by override
constants: a repeat sprite runs main's arithmetic and one sample, nearest
or linear; only clamp adds a clamp. The sampling moves from the vertices
to the quad, so the vertex layout is main's again.
- A black vignette changes no pixel (it darkens by texture x colour), and is
no longer drawn: in daylight its brightness fades to 0 within half a
minute, so most frames outdoors skip it, which main never could.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify passed and a proof run on the Mac mini: the self-check passed (parity 22/22), and six TNT minecarts now chain (6 → 0; 5 survive on main). Thanks @levzzz5154!
🤖 Agent: merged after verify passed and a proof run on the Mac mini: the self-check passed (parity 22/22), and lan_e2e minecarts passes in both single-player modes. Thanks @levzzz5154!
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
The server sends a knocked-back, pushed or jumping mob's move at once
(Entity.needsSync), and the remote client walks every living entity along
the server's positions with vanilla's SteppedInterpolationHandler: each
position takes the type's update interval to reach, sped up while steps
pile up. Before, the client waited for the next due update and then
lerped there in fewer ticks, so a hit pig jumped away fast.
tools/test_stepped_interpolation_parity.py drives vanilla's unmodified
InterpolationData and the Rust port with the same packets and ticks
(random cases and a punched pig's knockback); 135236 positions and
rotations match bit for bit.
Fixes#1374
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ItemEntity.tick: an item stuck inside a full block (one set on it) has no
physics on the server and moves toward the closest free side
(Entity.moveTowardsClosestSpace: north, south, west, east, up; 0.1 to 0.3 a
tick), instead of staying in the block (#1252).
A server's client (LAN guests, and the window on the server thread) digs at
the same speed as the server: getDestroyProgress with the submerged mining
speed, haste, fatigue and efficiency, not the plain tool speed, so it no
longer finishes early under water and gets the block put back (#1254).
lan_e2e: `drops` (a diamond under a block set on it comes out, in both
views) and `underwater` (dirt by hand at eye height under water: about
halfway at 35 ticks on client and server, broken by 105).
Fixes#1252, #1254.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SetWorldSpawnCommand sets the level's spawn and its angle; a player without
a bed or anchor respawns there (it used to set the player's own spawn
point, so other players and later deaths kept the old one). The angle is
saved in level.dat's spawn (yaw, pitch), and a guest whose saved respawn
point is the world spawn follows it when it moves.
lan_e2e `worldspawn`: /setworldspawn 24 blocks east, then the host and the
guest die and respawn there.
Fixes#1310.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ItemCooldowns ticks once a tick for each player: a server's players count
their shield cooldown down in tick_guests, and the host's world in tick_world
(it was in tick_using, which ran for the host and again for every guest's
swap, so a guest's shield came back early on the server thread).
A guest that dies stops digging (its mining, the delay and the crack go), so
the block it was breaking doesn't break once it respawns.
lan_e2e: `shield` (a vindicator's axe disables each player's shield; the
server's cooldown then falls 40 in 40 ticks) and `deathmining` (die while
breaking obsidian, respawn, the obsidian stays). `lan status` and the agent
state report the shield cooldown.
Fixes#1315, #1316.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Entity.applyEffectsFromBlocks: MagmaBlock.stepOn burns a player standing on
it who is not sneaking (hot_floor, 1), CactusBlock and CampfireBlock.entityInside
hurt what touches them (cactus 1, campfire 1, soul campfire 2, lit only).
A server's players (LAN guests, and the host on the server thread) now get
these and the grown sweet berry bush prick along the moves their client sent
that tick: the server rebuilds the tick's movement history and runs the
damage half of the block effects.
lan_e2e `contacts`: magma, sneaking on magma, a lit campfire, a cactus, grown
and young berry bushes, for the host and the guest.
Fixes#1305, #1311, #1312, #1313.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The server sends a player's own air supply and absorption amount
(Entity.DATA_AIR_SUPPLY_ID, Player.DATA_PLAYER_ABSORPTION_ID) when they
change, and the client applies them: on the server thread and as a LAN
guest the breath meter follows the server's Respiration-extended air
(#1304) and the golden hearts show (#1306). The connected client's tick
sets the helmet's Respiration and the Nether's FAST_LAVA as the window's
own tick does (#1309).
AbsorptionAmount is written from the live body and read back on load,
in the window's world and for a server's players (#1307).
lan_e2e: absorption, breath and lava checks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The camera overlays (the vignette every frame) read up to four texels a
pixel, each wrapped with two integer modulos. Intel GPUs emulate integer
division. Reducing the coordinate to one period first (fract, as the shader
did before) leaves every read at most one texel outside the sprite, so a
select wraps it. Same texels and weights; clamp-to-edge is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the agent-control-2 integration of all six PRs on the Mac mini. agent_control_e2e passes in CLI and MCP (one known survival flake), launcher_mcp_e2e passes, No AI is refused, and the options, screen_layout, menu_click and world_options parity harnesses pass. Ludovic asked for this on 2026-10-23.
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
On Apple GPUs timestamps go only at pass boundaries, and the frame was
bracketed by two one-thread compute passes. Two things made it read
about 0.01 ms:
- Metal resolves counter samples without waiting for the passes that
write them. Resolved in the frame's own command buffer, the end
timestamp was mostly the one an earlier frame left in the shared
query. Each readback slot now has its own queries, and with markers a
slot is resolved in a later frame, once on_submitted_work_done says the
GPU finished the frame it timed.
- A compute pass depends on nothing, so the end marker ran alongside the
frame's render passes. The end is now a render pass that loads the
frame's target, shading one pixel with every write masked: it runs
after every pass that drew into the target. The start stays a compute
pass (cheaper than a load and store of the target).
M4 Pro, render distance 32, offscreen: GPU p50 0.002 ms before, now
1.95 ms against a 1.95 ms frame (village) and 4.46 ms against 4.51 ms
(mountain), where the frame is GPU-bound. Vulkan and DirectX 12, which
write timestamps between commands, are unchanged but for the per-slot
queries.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the agent-control-2 integration of all six PRs on the Mac mini. agent_control_e2e passes in CLI and MCP (one known survival flake), launcher_mcp_e2e passes, No AI is refused, and the options, screen_layout, menu_click and world_options parity harnesses pass. Ludovic asked for this on 2026-10-23.
🤖 Agent: merged after verify passed and the agent-control-2 integration of all six PRs on the Mac mini. agent_control_e2e passes in CLI and MCP (one known survival flake), launcher_mcp_e2e passes, No AI is refused, and the options, screen_layout, menu_click and world_options parity harnesses pass. Ludovic asked for this on 2026-10-23.
🤖 Agent: merged after verify passed. Ludovic asked on 2026-10-23 for every hardware fact players send (opt-in). The end-to-end run against a local server is in the description: breakdowns need ≥3 installs, with no install ID or exact value. The consent version goes up to 2, so players who opted in are asked again, as the privacy page promises.
Under Wayland the title screen can come before the window is on a screen, so the
report's hardware.display stayed empty (seen end to end in a headless sway).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Horizon Settings gains a row: "AI Agent..." and "No AI: Off". No AI hides the
AI Agent page and closes the control socket at once.
The AI Agent page has an Agent Control switch (horizon-extras.txt
agentControl, off by default): on, the game the player starts opens the
default control endpoint, so horizon-ctl and MCP clients reach it without
--control. Turning it off closes the endpoint at once; dropping a control
server now also stops its listener thread. The page shows whether an agent
is connected, when it last sent a command and the socket, and Copy
Connection Details copies the socket path, a horizon-ctl line to try and the
MCP config for this game.
agent_control_e2e.py --scenario setting covers it: Agent Control on opens the
endpoint without --control, the pages list their widgets, Copy works, and
turning the switch off removes the endpoint while the game keeps running.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Right after a window is made, Wayland doesn't say which screen it is on, so
current_monitor() is None and display was missing (seen end to end in a headless
sway). Fall back to the primary, else the only, monitor, and read it again when a
session ends; agent control's status (the benchmark report's display) uses the same.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The table gains what only the raw socket could do (options pages, connect
and invite codes, disconnect, LAN, Play with Friends, books, signs, mount
inventories, effects, xp, enchant, potion, fill, containers) and new
commands: worlds, world-rename, world-backup, world-delete --confirm (the
Select and Edit World screens, from the title screen only), servers (the
Multiplayer list) and interact-entity (right-click an entity). status,
mouse, pad, release, bench and stats become MCP tools; summon takes a
position. Destructive and network tools carry MCP annotations.
docs/horizon-ctl-commands.md is generated from the table
(horizon-ctl help --markdown) and a test fails when it drifts. The
agent_control_e2e manage scenario exercises the new commands.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
horizon_core::no_ai reads and writes the noAi line of the game's
horizon-extras.txt, so the launcher and the game share one switch. While
it is on the game never opens its control socket, horizon-ctl refuses every
command with exit 4 and kind denied, horizon-ctl mcp and horizon-launcher
mcp refuse to start (and refuse calls if it is turned on mid-session), and
horizon-launcher no-ai [on|off] shows or sets it. The game's extras keep
the key when they rewrite the file. Both e2e harnesses check it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
main (#1271) pins Horizon Settings and Mods as the grid's last buttons;
this branch adds Credits & Attribution after them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
What players send (opt-in, unchanged switches) now describes the whole machine:
CPU cores and base/boost clock, the GPU's maker, integrated or discrete and its
VRAM, the drive under the game folder (NVMe/SSD/HDD), the screen (size, refresh,
scale), on battery or mains, laptop or desktop. Perf, crash and bug items carry
them in `system`; benchmark reports in `hardware`. No serials, host or user
names, or paths. CONSENT_VERSION goes to 2, so players are asked again.
The server keeps the new fields in `systems.extra` (the table is rebuilt once
on old databases, ids kept). /v1/aggregate and /v1/benchmarks/summary add
breakdowns by RAM class, VRAM class, drive, OS and integrated/discrete GPU;
a class is published only with at least 3 different installs, never a raw value.
website/scripts/benchmarks.mjs mirrors the classes on the shared fixtures.
The summary's "not Horizon or plain vanilla" check returned None (accepted)
instead of the rejection; it now matches the site's script.
The launcher's "View what would be sent" and `data show` list every field.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#1271 removed the title's live world backdrop (backdrop_live), so the
locator bar hides on the title alone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>