Every harness gave the same result on the Mac mini as on Linux, except
options and chunk_generation, where Rust's float functions use the platform
libm and round differently on macOS: those stay off the Mac. gate.py and
supervise.py no longer force 8 local harnesses, so parity.py keeps the
desktop to 3. The ThinkPad is listed (2 harnesses, nice 19) but not used by
default. Oracle runs get a 15-minute hang guard instead of 5, since one oracle
on a fully loaded machine can take more than 5 minutes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The oracle spent 58 of every world's 62 seconds waiting for the idle
threads of the game's IO pool to time out; it now exits once its outputs are
written (same outputs on all 22 worlds). The harness runs in 6 shards, keeps
the vanilla outputs so the oracle's results are reused, and declares its Rust
scope (datafix, NBT, region files, compression and horizon-core positions):
other horizon-world changes no longer rerun it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
parity.py:
- A harness declaring PARITY_SHARDS = N runs as N entries, NAME#1..NAME#N,
each told its share through PARITY_SHARD; NAME selects them all.
- PARITY_RUST_SCOPE lists the Rust sources a harness's probes can reach.
It applies only if closed (nothing in it names a module, crate-root
re-export or macro outside it, and nothing outside implements anything for
its types); crate roots then count without their other module declarations.
- Bytecode caches no longer count as dependencies.
- With other hosts, this machine runs 3 harnesses at once and only those
under 800 MB, admitted against the memory free at each start; hosts can
exclude harnesses and run at a nice level.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A worker started over ssh (parity.py --worker) runs harnesses on a mirror of
the worktree on another machine, here the Mac mini. One longest-first queue
feeds every machine, remote ones first; passes come back with the content
hash of each dependency, so they are keyed on exactly what ran. Harnesses
that need this machine are listed in tools/parity_hosts.json.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The old server jars and the 22 worlds made from them are inputs, but they
lived in each worktree's .local and were lost when its outputs were
cleared. They now sit beside the shared vanilla data, written atomically,
and the harness generates any that are missing (under a lock, since
several worktrees may run it at once).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Drowned register Zombie's goals and then Drowned.addBehaviourGoals (go to
water, trident and melee attacks with okTarget, go to beach, swim up,
RandomStrollGoal) with DrownedMoveControl, Drowned.travelInWater and
AmphibiousPathNavigation. Skeletons get their bow from finalizeSpawn and
pick the bow goal or the AbstractSkeleton$1 melee goal from the held item,
as reassessWeaponGoal does. Monster.aiStep adds 2 to noActionTime in bright
light (raiders always), which stops RandomStrollGoal; LookAtPlayerGoal
supports Mob.class targets and InteractGoal, used by the wandering trader
(with vanilla's AvoidEntityGoal order), illagers, ravagers and vexes.
Entities below the world take 4 damage a tick instead of vanishing, water
travel uses the 0.8F/0.6F/0.3F float constants, and MobWorld carries the
level's sea level.
The oracle refreshes the player's water state after placing it (it is never
ticked, so isInWater kept the previous scenario's value) and keeps its world
under the harness build directory, deleting it before halting instead of
leaving /tmp/horizon-mob-ai* behind. The probe reports KILLED removals.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tools/gate.py took a crashed tools/parity.py for a pass with no new
failures. The runner now records each completed run, and the gate fails
unless its own run finished. A missing or empty trace (the disk was full)
no longer crashes the runner.
After a harness passes, the runner deletes its large regenerated outputs
(64 MB or more), keeping the vanilla oracle outputs that reusing an oracle
run needs. Full suites left 2-4 GB files in every worktree and filled the
disk.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Zombies (and husks, drowned, zombie villagers, zombified piglins) now
register ZombieAttackTurtleEggGoal first, as vanilla does, so their goal
selector draws the same random numbers. LivingEntity.pushEntities runs in
aiStep against the player and other mobs, with pushes to other mobs applied
at once as Entity.push does (replacing the post-tick push_apart). Spider
target goals check the light before their random interval, LookAtPlayerGoal
uses the non-combat targeting conditions, raiders register the
PatrollingMonster/Raider goals before their own, and the witch runs its
aiStep (heal cooldown, drinking, particle roll) before super.aiStep with
NearestHealableRaiderTargetGoal's nextBoolean.
The mob AI oracle's player is now a fully loaded, hurtable client: its
connection starts with a 60-tick load timer and 26.3 moved the hurt cooldown
to damageCooldownTime, both of which only ServerPlayer.tick counts down, so
whether mobs could hurt the player depended on the scenario order. The probe
reports the player's hurt sound, and an rng-trace feature records the call
sites of Horizon's mob random draws next to the oracle's trace.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A dispensed spawn egg's mob gets the riders finalizeSpawn adds (chicken
jockeys, spider jockeys, strider jockeys). The harness leaves which mob
comes out a baby and rides to the spawning port's finalization draws, as
it already does for the level random of spawning cases. Cases no longer
grow plants out of a dispenser or build a door from mismatched halves,
and the oracle runs every case three times so an unseeded value is
marked even when two runs happen to agree.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A pass is keyed by the content of the harness's dependencies, so it is
also valid in any worktree with the same files. Passes are published to
parity-shared/ beside the shared vanilla data, and a worktree adopts one
whose key matches its own files instead of running the harness: an
agent's gate only runs what its own changes affect, instead of the whole
suite from scratch. Every runner on the machine also shares
PARITY_MACHINE_JOBS slots, so several agents' gates can't overload it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fixture's chunks are plains (empty chunks read as badlands, which has
no bone meal flowers) and carry the case's sky and block light, as the
oracle level does, so grass bone meal places flowers and crops check
their light like vanilla's. The oracle's client sounds include the lava
fizz (1501). Cases build two-block plants whole: a lone lower half is not
a world vanilla makes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Desert, snowy, swamp and jungle nights, lush, sulfur and dripstone caves,
warm and frozen oceans, crimson forests and slime chunks. A cube mob grown
by finalizeSpawn takes its constructor's head rotation from the entity's
unseeded random (AbstractCubeMob.onSyncedDataUpdated), so that yaw alone is
not compared.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tools/test_natural_spawn_parity.py compares whole chunk spawn passes,
spawnCategoryForPosition from chosen sites and chunk generation packs
against vanilla's NaturalSpawner on generated box worlds.
Divergences fixed on the way: the reduced water ambient tag is
reduce_water_ambient_spawns, and the room check uses each type's spawn
dimensions scale (4 for slimes and magma cubes, 2 for sulfur cubes).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Villagers, piglins, piglin brutes, allays, dolphins, foxes and pandas can
pick up loot from construction (their constructors call
setCanPickUpLoot), so they take dispensed equipment without spawn
finalization. Nautilus saddles play item.nautilus_saddle_equip. The
oracle level is the overworld, so fire in an obsidian frame lights a
portal there too, and it answers getEntity(id) and level-wide entity
queries, which leash cutting needs. Unseeded item motion is compared
within its full spread (item motion plus a shearing sheep's push).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fixes from 3076 to 3459 (game event, cat, frog, painting, goat, goat
horn, filtered text, POI, blending, spawn egg, effect duration, archaeology,
double-sided signs, the 1.20 experiment and chunk statuses), the block
entities their schemas add, and block entities' remainders (a decorated
pot's shards stay next to its sherds). The sample world generator logs in
to 1.18.2 to 1.19.4 servers and gates the scene by version.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Opening a world saved by an older version runs the file fixes, level.dat and
the DataFixers upgrades of every file instead of refusing it. The save
round-trip harness treats a lit campfire's cooking progress as a timer in the
vanilla views: Horizon saves it as loaded (stage 3 checks it exactly), and it
moves on by however many ticks each vanilla server ran before data get.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tools/parity_shim.py stands in for java and javac in harness runs: it
records each oracle run's outputs and the content of every file under the
paths its arguments name, and replays the run while those are unchanged.
After a Rust-only change a harness only pays for its Rust side
(block_update 53 s -> 5 s). A harness that fails while reusing oracle
results is rerun fresh, and stops reusing them if that passes; --all and
--fresh never reuse them.
Harness probes built in the dev profile are optimised (overflow checks
and debug assertions kept), and the chunk generation harness runs its
vanilla and Horizon sides concurrently.
HORIZON_NO_AUDIO keeps the game off the speakers; tools/parity.py sets it,
since the save round-trip harness launches the game on a virtual screen.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Only the collision shapes overlapping the spawn block stop the fall, so a
wide mob beside the dispenser no longer lands a block high. Spawn egg
cases compare that the new mob speaks and where, not which of its voices
its own unseeded random picked.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
DataFixerUpper reads typed fields without taking them out of a record's
remainder, so a typed field a fix drops comes back on write with its value
from the record's last read: entities keep their old ArmorItems, HandItems,
body_armor_item, saddle, Trident and Particle next to the new fields, as
vanilla 26.3 writes them. Fixes that write and read entities refresh that
map. The area effect cloud particle moves to custom_particle (4420), boats
keep Type (BoatSplitFix's remainder update misses), and region files with a
short header read as empty, as vanilla's RegionFile does. The sample worlds
now use the item formats each version accepts and enable its experiments.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
BoatDispenseItemBehavior spawns the chest boat or chest raft itself: boats
now carry ChestBoat's 27-slot chest, drawn with its model, saved with its
Items, opened by sneaking, spilt when broken. MinecartDispenseItemBehavior
spawns command block minecarts (an empty command). The dispenser harness
now draws fronts per item (fire, portal frames, anchors, crops, copper,
hives, golem and wither builds), every boat wood, more equipment, spawn
eggs, sulfur cube swallowables and leashed targets, over 6000 cases.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Flint and steel uses BaseFireBlock.canBePlacedAt and getState (a portal
frame draws the level random for vertical dispensers), posts block_place
and block_change, and lights only what CampfireBlock, CandleBlock and
CandleCakeBlock.canLight accept. Fire placed inside an empty obsidian
frame now lights the portal (BaseFireBlock.onPlace). Glowstone posts the
anchor's block_change. Shears cut leads (shearOffAllLeashConnections) and
empty a sulfur cube. Dropped leads draw motion from the item's own
random, not the level's. The oracle gives bone meal a chunk generator so
sapling, grass and moss cases are compared instead of erroring.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tools/parity.py traces each harness (its Python, Java oracle sources and
the Rust sources of the probes it runs, from Cargo's dep-info) and reruns
it only when one of those, the vanilla artifacts or the toolchains change.
Harnesses run in parallel within the available memory, and their probes
build without LTO and with 16 codegen units into .local/parity-target: a
one-line change rebuilds every horizon-world probe in 3 seconds instead of
122. The full suite takes about 20 minutes instead of an hour.
tools/gate.py is the check a branch passes before integration, and
tools/supervise.py manages the per-area worktrees: status, new worktrees,
saving unfinished work, and integrating a branch only when the gate passes.
Harnesses find probes through lab.cargo_target() instead of a hard-coded
target/, the save round-trip harness builds the binaries it launches, and
the mob AI harness now fails on divergences like every other harness
(--report-only keeps the old behaviour). Debug builds keep line tables
only, so worktrees stay small.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tools/rea_vanilla.py runs REA's JADX provider over the pinned client and
server JARs: class search, class members, method bodies and callers
(including lambdas). REA and the JADX engine are pinned in
research-lock.json.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Uncommitted changes the agent had in progress when it was stopped; may not
build or pass its harnesses yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Uncommitted changes the agent had in progress when it was stopped; may not
build or pass its harnesses yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Mth.atan2 builds its tables with StrictMath.asin (fdlibm, ported) and
HotSpot's cos; the controls multiply by the folded RAD_TO_DEG float
constant. Mob.checkDespawn runs before the tick and, in 26.3, persistent
mobs only reset their idle time near a player. NearestAttackableTargetGoal
always requires a line of sight and ignores mustReach when searching.
MeleeAttackGoal starts its canUse clock at game time 0. The harness now
records every random draw of the vanilla mobs with its caller, the random
state per tick, and starts each scenario at game time 1000.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>