Skip to content

CHANGELOG.md

Latest commit

 

History

History
125 lines (98 loc) · 5.73 KB

CHANGELOG.md

File metadata and controls

125 lines (98 loc) · 5.73 KB
Raw

Changelog

v5.5 [v55-full-upgrade] — 2026-02-28

Full upgrade: new MCP servers, deprecation fixes, BossManager integration, slayer chat parsing.

New Features

  • OSRS Wiki MCP Server: 5 tools — search, get_page, get_item, get_monster, item_price
  • DreamBot Docs MCP Server: 5 tools — search_class, get_class_methods, search_method, get_enum, list_classes
  • BossManager: Wired into AIAgent lifecycle (onStart/onLoop/onExit)
  • ChatFeedback → SlayerTaskTracker: Slayer chat parsing integration for task assignment detection
  • PaintOverlay: Added BOSS section for boss tracking stats

Fixes

  • All 15 getID() deprecation warnings replaced with getId() across codebase

Config

  • .mcp.json: Added wiki-mcp-server and dreambot-docs-mcp server entries
  • Version bumped to v5.5, BUILD_TAG updated to v55-full-upgrade

v5.4 [audit-fix-Feb28] — 2026-02-28

Full codebase audit (6 agents) + fix pass (7 agents). Build tag: audit-fix-Feb28

Critical Fixes

  • ActionExecutor: Added missing drinkCombatPotion(), drinkPrayerPotion(), drinkAntifire() for actions 89-91 (was compile-blocking)
  • AIAgent: Wired ScriptedHelperManager (with SlayerHelper + QuestHelper), SlayerTaskTracker, and DeathRecovery lifecycle — actions 82-88 were permanently dead before this
  • TrainingMonitor: NUM_ACTIONS 88 → 92 (was causing ArrayIndexOutOfBoundsException for actions 88-91)
  • DeathRecovery: Fixed EquipmentSlot.BOOTS → FEET, GLOVES → HANDS (DreamBot API)

Reward System Fixes

  • Valuable loot reward: Now tracks prevInventoryIds snapshot — only rewards NEW items, not existing inventory
  • Food eaten reward: Gated on lastAction == 12 — no longer fires on drops/deposits
  • XP delta cap: Changed < to <= (exactly 1000 XP gains no longer discarded)
  • CurriculumManager: All 27 reward component keys now explicitly mapped per phase (was 12/27, 15 fell to default)
  • Cooldown reward path: Now uses curriculum scaling (was bypassing it)

Neural Net / Training Fixes

  • PPOTrainer: NaN guard now breaks both inner and outer loops (was only breaking minibatch loop)
  • ModelLoader: Reinitializes network on dimension mismatch instead of throwing exception
  • Tensor softmax: Added div-by-zero guard when all actions masked
  • MetaTrainer: Removed dead bufContexts allocation
  • ModelLoader: readIntLE now checks for EOF
  • PPOTrainer comments: Fixed stale minibatch size (32→64) and learning rate (3e-4→2e-4)

Security Fixes

  • RuneLite CORS: Restricted from wildcard * to http://127.0.0.1
  • RuneLite 404: Escaped URI in JSON error response (was JSON injection vector)
  • RuneLite threads: NanoHTTPD now uses daemon threads (was preventing clean JVM shutdown)

Game Logic Fixes

  • GameRules Rule 6: Now actually blocks attacking NPCs 15+ levels above player (was only logging warning)
  • GameRules: Prayer deactivation guarded — only calls API when a protection prayer is active
  • GameRules: Rule numbering fixed to sequential 1-9
  • StateEncoder: Zeroed duplicate combat style dims [200-202] and food dims [251-252] (checkpoint compatible)
  • TickCombatTracker: resetCombat() now called on death (was retaining stale timing data)

Overlay / Monitoring

  • PaintOverlay: Header now shows version + build tag + dimension summary
  • Version: Bumped to v5.4 with [audit-fix-Feb28] build tag

Documentation

  • CLAUDE.md: Fully rewritten — state 278, hidden 192, actions 92, NanoHTTPD, all endpoints documented
  • Comments: Fixed all stale action count references (60/80/85/88 → 92) across AIAgent + ActionExecutor

Dead Code Identified (kept for future use)

  • BossManager / BossHelper / BossProfile — built but not wired into AIAgent
  • QuestUnlockTree — data file, never referenced from other classes

v5.3 — 2026-02-27

Added

  • ShopManager: F2P shop prices (24 items, 4 shops), gear deficit calculation
  • Shop actions 60-69: open/buy/close/walk to 4 shops, withdraw coins
  • Magic actions 70-79: 4 strikes, 4 bolts, set/cancel autocast
  • Sell actions 80-81: sell_to_shop, sell_all_junk
  • Scripted helper actions 82-84: start_slayer, start_quest, abort_helper
  • CurriculumManager: 4-phase training (BASICS→COMBAT→GEAR→AUTONOMY)
  • QuestManager: 13 tracked quests, priority scoring
  • ScriptedHelper framework: SlayerHelper + QuestHelper
  • BossProfile/BossManager: 9 bosses in 3 tiers
  • State expanded to 278 dims, actions to 92, hidden to 192

v5.2 — 2026-02-27

Added

  • ChatFeedback: Game chat widget parser for cooldown/too far/blocked/success/loot/level up
  • TrainingZone: 9 F2P zones with combat level ranges
  • Smart NPC targeting with blacklist and level cap
  • Death recovery: 30-tick combat block after death, naked penalty, redirect to bank
  • Run energy masking when energy < 10%
  • Zone-based rewards (+0.05/tick in zone, +0.5 arrival)

v5.1 — 2026-02-27

Added

  • Kill detection fix
  • State 239→255, network hidden 128→192
  • Damage dealt/taken rewards
  • Food supply chain tracking
  • Home teleport tracking
  • Expanded MCP monitoring endpoints

v5.0 — 2026-02-27

Added

  • GearScorer: Tier-based equipment scoring with DPS sub-tiers
  • TickCombatTracker: NPC attack timing for prayer flicking
  • BankGearCache: Bank gear scanning + upgrade detection
  • PaintOverlay: In-game stats rendering (11 sections)
  • XP delta cap, per-tick reward clamp, gear/combat/toggle rewards
  • GameRules: Once-per-episode thresholds, auto prayer flicking
  • Temperature schedule: 1.5→1.0→0.7

v4.0 — 2026-02-27

Initial RL System

  • Actor-critic LSTM with PPO training
  • MetaNN for hierarchical goal conditioning
  • 40 actions, 227-dim state
  • Basic combat, movement, prayer, banking