Changelog

Changelog

5.2.5 - Reinforcement Round Limit Config

Added

• ai.training.max-round-minutes, defaulting to 1, controls the per-generation /ai reinforcement round cap. The command's [round-minutes] argument still overrides the config value for one session.

5.2.4 - Mace Fall Velocity Fix

Fixed

• Bot walking now caps only horizontal movement and preserves vertical velocity, so mace dive bots no longer slow-fall while the movement controller is steering them midair.

5.2.3 - Training Sample Guard

Fixed

• Mixed movement training now only ranks and saves a specialist family brain from bots that actually produced route samples for that same family, preventing unrelated aggregate survival fitness from replacing a specialist brain.

5.2.2 - Mixed Family Training

Changed

• Mixed /ai reinforcement now samples the configured loadout mix, seeds each candidate from its assigned loadout family, ranks results per family, and updates every eligible specialist brain from one round.
• Movement training autosave is enabled by default, while save-only-improved-brain still prevents worse brains from replacing better ones.

5.2.1 - Movement Training Default

Changed

• /ai reinforcement <population-size> <name> [skin] now defaults to the movement-controller trainer.
• legacy remains available as an explicit mode for the old full-replacement training pipeline.
• Movement-training option strings such as family=mace:mix=mace_curriculum are parsed as movement options even without a leading movement: prefix.

5.2.0 - Movement Brain Bank Evaluation Support

Added

• Movement brain bank routing by MovementBranchFamily.
• Manifest plus per-brain persistence under ai/movement/.
• Legacy ai/brain.json import as general_fallback.
• Automatic weighted movement-training loadout assignment.
• Per-family reward profiles and rollout metrics.
• /ai evaluate report export for fixed-seed movement-bank ablations.
• checkMovementOnlyContract Gradle guardrail for movement-only source code.

Notes

• Initial mixed movement training recorded per-family telemetry but updated general_fallback; this was superseded by 5.2.2, which trains eligible specialist families from the mixed loadout population.
• Curriculum mode updates the configured ai.training.curriculum-family brain.
• Live evaluation metrics require an arena runner; the current evaluation export initializes and reports route/fallback/schema state.

5.1.1 - Combat Reliability + Movement Neural Network

See Release Notes 5.1.1 for full details.

Added

Combat reliability:

• Vanilla attack ordering fix: Bot.attack() calls vanilla attack before swing/punch to prevent charge reset.
• Charge-aware planning: CombatDirector waits for full attack charge before committing to attacks.
• Mace recharge planning with gravity-aware airtime tracking.
• Mace airborne tracking with ground clip tolerance and velocity-aware horizontal damping.
• Sweep instrumentation and telemetry.
• Combat telemetry fields: critPred, sweepPred, chargeAtVanillaAttack, chargeAfterVanillaAttack, targetHp, targetHpDelta.
• /bot combatdebug command for per-bot combat trace logging.

Movement neural network:

• Movement-only neural network controlling footwork. It does not control combat.
• CombatDirector.plan() / execute() split with CombatIntent / MovementState coupling.
• 37-value MovementInput schema and 8-value MovementOutput schema.
• MovementOutputApplier with threshold-based movement decisions.
• In-JVM genetic algorithm with tournament selection, uniform crossover, adaptive Gaussian mutation, and elite preservation.
• Brain persistence with schema validation, safe fallback, and backups.
• /ai brain <status|load|save|reset> commands.
• /ai movement command for spawning movement-controller bots.
• /ai reinforcement accepts movement-controller or legacy mode.

Loadouts:

• New loadouts: vanilla, axe, smp, pot, spear.
• /bot loadoutmix command for distributing loadouts across live bots.
• Loadout mix presets: alltypes, core, problem.

Fixed

• Bots no longer waste swings on uncharged attacks.
• Vanilla attack damage calculation now runs before swing/punch animation.
• Mace smash no longer commits when airtime is insufficient for recharge.
• Mace airborne tracking no longer overshoots on moving targets.

Known limitations

• Changing the movement layer shape after training invalidates saved brains.
• Training with large populations impacts server TPS.
• Movement-bank bots must be spawned with /ai movement; normal /bot create bots use the legacy movement path.

Pre-5.1.1 - Combat + Inventory + Presets Overhaul

Added

• Weapon-aware CombatDirector that picks behavior by inventory, distance, cooldown, and dimension.
• Melee, Mace Smash, Trident Momentum Throw, Wind Charge, Ender Pearl, Crystal PvP, Anchor Bomb, Cobweb, Elytra Glide, Totem Swap, and Heal behaviors.
• Full per-bot inventory: 9 hotbar + 27 storage + 4 armor + 1 offhand.
• /bot inventory <name> GUI editor.
• /bot give, /bot armor, /bot weapons commands.
• Built-in loadouts: sword, mace, trident, windcharge, skydiver, hybrid, crystalpvp, anchorbomb, pvp, clear.
• YAML preset system with full NBT round-trip.
• Permissions: terminatorplus.admin, terminatorplus.manage, terminatorplus.*.
• API calls: Terminator.combatTick(LivingEntity) and Terminator.getDimension().

Changed

• Normal bots use vanilla damage, including real crits, shields, and enchantments.
• /bot loadout is per-bot when a name is passed.
• /bot give accepts [bot] [slot] for targeted placement.

Fixed

• Hotbar slot sync when CombatDirector switches weapons.
• Fall damage stacking on mace smash.

Previous versions

Pre-overhaul. See Git history.