🤖 Agent: read the diff at dddcc9972 ("Fix the crash on every start with HDR on in macOS, and start with HDR ") against the 26.3 source it cites; verify green, merges cleanly with main. Proof: cg25 union a051cc285 (local branch cg25/union in t3-ef805fdb: main 6d3489d30 + 38 contributor heads merged cleanly, no fixups): clippy -p horizon-client --all-targets -D warnings clean, horizon-client lib 1105 + bin 133 + integration tests pass, parity --changed 355/355 harnesses PASS on the Mac mini (Cargo.lock changed by #3210, so every harness ran; tab_list passed on its one retry).
🤖 Agent: read the diff at 513aed64c ("Name the projectile in the death message when it has no owner") against the 26.3 source it cites; verify green, merges cleanly with main. Proof: cg25 union a051cc285 (local branch cg25/union in t3-ef805fdb: main 6d3489d30 + 38 contributor heads merged cleanly, no fixups): clippy -p horizon-client --all-targets -D warnings clean, horizon-client lib 1105 + bin 133 + integration tests pass, parity --changed 355/355 harnesses PASS on the Mac mini (Cargo.lock changed by #3210, so every harness ran; tab_list passed on its one retry).
🤖 Agent: read the diff at 22cab3eeb ("Open windowed after the window system kills the game in fullscreen") against the 26.3 source it cites; verify green, merges cleanly with main. Proof: cg25 union a051cc285 (local branch cg25/union in t3-ef805fdb: main 6d3489d30 + 38 contributor heads merged cleanly, no fixups): clippy -p horizon-client --all-targets -D warnings clean, horizon-client lib 1105 + bin 133 + integration tests pass, parity --changed 355/355 harnesses PASS on the Mac mini (Cargo.lock changed by #3210, so every harness ran; tab_list passed on its one retry).
🤖 Agent: read the diff at 8057d2751 ("Worn dyed leather armour shows its dye colour") against the 26.3 source it cites; verify green, merges cleanly with main. Proof: cg25 union a051cc285 (local branch cg25/union in t3-ef805fdb: main 6d3489d30 + 38 contributor heads merged cleanly, no fixups): clippy -p horizon-client --all-targets -D warnings clean, horizon-client lib 1105 + bin 133 + integration tests pass, parity --changed 355/355 harnesses PASS on the Mac mini (Cargo.lock changed by #3210, so every harness ran; tab_list passed on its one retry).
🤖 Agent: read the diff at f507a067c ("The breeze's wind and the wind charge scroll their textures") against the 26.3 source it cites; verify green, merges cleanly with main. Proof: cg25 union a051cc285 (local branch cg25/union in t3-ef805fdb: main 6d3489d30 + 38 contributor heads merged cleanly, no fixups): clippy -p horizon-client --all-targets -D warnings clean, horizon-client lib 1105 + bin 133 + integration tests pass, parity --changed 355/355 harnesses PASS on the Mac mini (Cargo.lock changed by #3210, so every harness ran; tab_list passed on its one retry).
🤖 Agent: read the diff at 28f906c37 ("A summoned or egg-spawned mob is not persistent") against the 26.3 source it cites; verify green, merges cleanly with main. Proof: cg25 union a051cc285 (local branch cg25/union in t3-ef805fdb: main 6d3489d30 + 38 contributor heads merged cleanly, no fixups): clippy -p horizon-client --all-targets -D warnings clean, horizon-client lib 1105 + bin 133 + integration tests pass, parity --changed 355/355 harnesses PASS on the Mac mini (Cargo.lock changed by #3210, so every harness ran; tab_list passed on its one retry).
🤖 Agent: read the diff at 93e61ef93: port_regen.python() probes python3/python/py and keeps the first that runs Python 3 (cached), with a test. verify green. Queued for the merger.
🤖 Agent: read the diff at ce848af06: zip hardening: the raw buffer's capacity is capped at the archive length and inflate at 1 GiB with saturating maths, so a zip64 entry claiming u64::MAX/2 errs instead of aborting; real packs are unchanged (the limit stays size*2+1024 below 1 GiB). Test included. verify green; no harness reaches it. Queued for the merger, which waits for this head's parity run.
🤖 Agent: read the diff at a8903cba0: test only: the shulker's body stays at 0 while its head turns (`ShulkerBodyRotationControl`, `ShulkerLookControl`). verify green, parity green on this head (merged into main), merges cleanly.
🤖 Agent: read the diff at 1d693cb3c: `MouseHandler.grabMouse`/`releaseMouse`: the cursor is centred only when the grab state changes; fixture test updated. verify green, parity green on this head (merged into main), merges cleanly.
🤖 Agent: read the diff at f10558380: `WorldSelectionList.handleNewLevels`: an emptied list opens Create New World, whose Cancel returns to the title. verify green, parity green on this head (merged into main), merges cleanly.
🤖 Agent: read the diff at fa940d5b6: `entity_view_scale` extracted from `Entity.setViewScale` (clamp(renderDistance / 8, 1, 2.5) * entityDistanceScaling) with tests for the 48-block chicken range. verify green, parity green on this head (merged into main), merges cleanly.
🤖 Agent: read the diff at fcc7a818b: the remote-mining fixture now records that the hit sound plays once at dig start (`ClientLevel.addBreakingBlockEffects`), plus the `dig_sounds` LAN check for #2277. verify green, parity green on this head (merged into main), merges cleanly.
🤖 Agent: verify green; pr_ready on the Mac mini passes (94 affected harnesses, 93 run and pass incl. the worldgen ones; lan_integration by name only; new test decoration_is_shared_by_biome_source). Through the Singleplayer screen, 2 workers, medians of 5: client joined 0.45 → 0.31 s (create) and 0.46 → 0.31 s (reload); reload first controllable frame 0.60 → 0.42 s.
🤖 Agent: read the diff at d3be8dbf0: horizon-server HTTP API: request and header lines over 8 KiB are a 400 instead of growing without bound; test with a 100 KB line. verify green; no harness reaches it. Queued for the merger, which waits for this head's parity run.
Mojang republished the 26.3 version JSON again. The live copy (5df1795e)
points at asset index 34 917ff6d7; neither was in research-lock.json, so
`horizon-prepare --prepare` from a fresh official-launcher install stops
with "is not a 26.3 version file Horizon knows", and the sounds step warns
about an unknown asset index. Jars, libraries and javaVersion are
unchanged; the index differs from 0c8f43b2 only in 142 minecraft/lang/*
objects and pack.mcmeta (uz_uz moved in the language list).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UE3PMKbnTcxfBSeT5uQWCK
🤖 Agent: read the diff at a77639022: test only: every graphics preset keeps animated textures on (#2064 stays fixed). verify green; parity selects no harness for it. Queued for the merger, which waits for this head's parity run.
Compare managed compiler and PATH entries as platform paths, including Windows executable suffixes. Fixes the Windows assertion that blocked the native-capable launcher release.
Validation: pr_ready passed (401 Python and 219 launcher tests), and the release dry run passed tests and packaged smoke checks on Linux, macOS and Windows.
Co-Authored-By: Codex <noreply@openai.com>
A wall of 3458 text signs ran at 128 fps against 1642 without them: each
frame laid every sign's text out again glyph by glyph (Font::glyph_of was a
quarter of the main thread), then regrouped every entity quad by material
through a per-quad index list.
- block_entity_render: a sign's text vertices are kept relative to its
block and moved to the camera each frame. They are laid out again when
the sign data, block state, light, glowing outline distance or font
(Font::epoch: a new font, a glyph cache reset, the Unicode and Japanese
options) changes; signs out of view for 100 ticks are dropped.
- gfx/entity_batch: quads are grouped in runs of one material and whole
runs copied (2-4x faster on model-like input, the same when every quad
changes material).
Fixes#1461 (the cost per sign; the vanilla text distance of 64 blocks
and the look are unchanged).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`Decoration::new` decodes the placed features of every biome a source can
place: about 140 ms of the ~185 ms `OverworldGenerator::for_world` took on
each world opening (Mac mini). It depends on the worldgen resources and the
biome source alone, not on the seed, so `GeneratorData` now keeps it by
biome source (`decoration_for`, like the ring cache) and every generator
built for that source shares it.
The world screens' preload (#2481) already reads the worldgen bundle while
the player picks or creates a world; it now also builds the Overworld's
decoration (`prepare_decoration`), so opening the world finds it ready.
Reopening a world or returning to a dimension in the same session builds
nothing again either.
Mac mini, --profile fast, offscreen, seed 1234, RD 12, 2 workers, server
thread, through the Singleplayer screen, medians of 5 alternated with main:
server ready to join: create 0.45 -> 0.31 s, reload 0.46 -> 0.31 s
first controllable frame: reload 0.60 -> 0.42 s, create 0.85 -> 0.79 s
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: read the diff at aef4186de: test_parity_reach's fixture records dependencies with `as_posix()`, as the tool does. verify green. Queued for the merger.
🤖 Agent: read the diff at 28e3286e4: port_harness also picks up Windows `C:...x.log` paths from javac's "see ..." lines. verify green. Queued for the merger.
🤖 Agent: read the diff at 5096b4110: launcher offline names: Forget also deletes the name's skin and cape files, async skin picks bail if the name was forgotten meanwhile, account rows load-mutate-save, and "New name" rolls a fresh PlayerNNNNN into every game home. Tests included. verify and launcher green; no harness reaches launcher code. Queued for the merger, which waits for this head's parity run.
🤖 Agent: verify green; pr_ready on the Mac mini passes (34 affected harnesses, 33 run and pass, lan_integration by name: 7 lan_e2e checks incl. spawned/respawn/portals/forced-levels pass). World entry, 2 workers, RD 12, medians of 3: create 8.05 → 3.57 s, reload 3.72 → 1.42 s to all view columns loaded.
`chat_line` logged `[chat]` and then `Chat::add_message` logged `[CHAT]` for
the same message, so every line appeared twice, and the `[chat]` copy slipped
past the `[CHAT]` filter `log_for_report` uses to keep chat out of an attached
log.
Vanilla's `ChatComponent.logChatMessage` writes the one line `add_message`
already writes, so the `chat_line` call is the leftover.
Fixes#3261
Co-Authored-By: Hermes Agent (Nous Research) <noreply@nousresearch.com>
🤖 Agent: read the diff at 557a82738: a spawn egg on a trial spawner follows `TrialSpawner.overrideEntityToSpawn` -> `FullConfig.overrideEntity` in the 26.3 source: data reset, normal and ominous configs replaced by `Holder.direct(withSpawning(type))` (saved inline through the new `TrialConfig::encode`, `optionalFieldOf` defaults left out), then `setState(INACTIVE)`; plain spawners unchanged. Round-trip test included. verify and parity green on this head. Queued for the merger.
A server thread (single player's, a LAN host's, a dedicated server's) ran
`streaming.update` once a tick, and `schedule_generation` keeps at most
three load or generate jobs per worker in flight. Joining a world was
capped at about six columns a tick (120 a second with two workers) while
the workers idled most of each tick. Vanilla's server runs its chunk
tasks while it waits for the next tick (`MinecraftServer.waitUntilNextTick`).
`pump_pregen`, now `pump_chunks`, already did that for a /pregen task and
force-loaded chunks; it now also pumps while the players' chunks are in
flight, except while the server is paused or a respawn search waits for
a chunk (its area is only asked for by the tick's own update).
Mac mini, --profile fast, offscreen, seed 1234, RD 12, server thread,
medians of 3, all 637 view columns in and meshed once:
2 workers: create 8.05 -> 3.57 s, reload 3.72 -> 1.42 s
4 workers: create 4.97 -> 2.16 s, reload 2.23 -> 1.27 s
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A server may put a show_dialog click event on a chat line or a book
page; ClickEvent.Action.SHOW_DIALOG allows it from a server, and vanilla
opens the dialog through Screen.defaultHandleGameClickEvent. Horizon
dropped the click event while parsing the component (no arm for it in
component.rs, no ClickEvent variant to carry it), so clicking the text
did nothing.
The dialog now survives as ClickEvent::ShowDialog (a registry id or the
inline dialog), parses from click_event, round-trips back out, and both
click handlers (chat and BookViewScreen) open it with show_dialog_ref,
the same path a server's show_dialog packet takes.
Co-Authored-By: ZCode <agent@zcode.dev>
Creating a server on Windows failed with a bare "mklink /J failed" (#3200).
servers.rs passed the paths to cmd unquoted, so a user folder with `&`, `^`
or `(` broke cmd's parsing; sources.rs already quoted them. Both now use
one sources::junction, which also puts cmd's own message in the error.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: proven with a real launcher GUI run in isolated dirs (Settings → Native mods row, red warning naming veinminer 0.3.0, trust → cached build reused, status line); launcher clippy and 265 tests green; #3194 merged and this now targets main with only its own 5 files. Verify green at 505de825.
The GitHub Sponsors link was only on the patch notes page. Show it beside
Discord and Website in the sidebar footer (heart icon), from links.json's
sponsor entry, translated in fr and ru.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: proven with a real launcher run in isolated dirs: registry install of Veinminer's .hnmod into an instance, native-build of that instance's enabled native mods, instance pinned to the build; launcher clippy and tests green after merging main (#3140 and #3189 merged). Verify green at 1f6511d0.
The Team tests fixes on the Nightly and hit the cap of 6 builds a day
(Ludovic, 2026-10-28: raise it to 12).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merging levzzz5154's native-mod registry kind; proven with the real Veinminer .hnmod through registry build, the launcher's install check and a native build that ran (see comment).
🤖 Agent: verify passed at 8b5a27ed3. Move only (crate split B7): the furnace's recipe logic leaves containers as the FurnaceCooking extension trait. Full pr_ready green on this head (1261 tests, parity 69 affected / 0 failed); #2959, #2960 and #3131 still merge cleanly. Ready once parity is green.
🤖 Agent: merging: native-build compiles horizon-mixins bundles on their pinned Nightly with the launcher's own Rust; proven by building levzzz's Veinminer .hnmod on 83f15fa (580 s cold, 2m46s warm) and /veinminer answering in the built game.
Settings → Game → Native mods (experimental), off by default. On, Build names the native mods
turned on in the instance Play starts, warns in red that they run with full access to the
computer, and compiles them only from "I trust these mods: build"; the row shows the step while
it runs, the error if it fails, and the build the instance plays. Ported from native-mixins-1.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- native-mod entries (levzzz5154's #3140) are listed as mods and install as
horizon-native-mods/<id>.hnmod (the folder #3112's inventory reads), turned on in
horizon-mods.txt, after the .hnmod is checked as the registry checks it.
- native-build with no folder builds the native mods turned on in the instance, and that
instance then plays the build (the game's own home builds into the Native mods instance).
- Every registry project says who made it; a project not made by Horizon says it is not made
or endorsed by the Horizon project, with the mod policy's link (playhorizon.dev/mods/policy).
A native mod says it runs with full access to the computer and which Horizon it builds on.
- Pruning keeps every native build an instance plays.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merging: nightly publishes horizon-source.tar.gz (signed) and tags its commit nightly-source/<date>-<short>; proven by a full dry run of the workflow (run 40686585965) and tests/test_nightly.py.
Ludovic Marie
degepie
levzzz
rdbt
Alexander Krylov
masa-masa-0314
PH4N7OMx