Перейти к содержимому
Mineforgian

Steve AI

Steve, a personalized AI agent model built to assist you

Загрузки
5K
Подписчики
13
Обновлён
18 августа 2026 г.
Лицензия
MIT

Опубликован 7 ноября 2025 г.

Steve functions as an AI Agent — or an entire team of Agents if you enable multiple instances. You simply describe what you want, and Steve understands the context, forms a plan, and executes it. The concept is the same as natural-language code assistants, except here the “output” is not code: it's embodied Steves operating directly inside your Minecraft world.

The interface is simple: press K to open the command panel and type your request. The agents handle everything — interpretation, planning, and in-game execution. Ask “mine some iron”, and the agent reasons about spawn depths, navigates to the correct layer, locates ore veins, and extracts resources. Request a house, and it evaluates available materials, generates an appropriate layout, and builds it block by block.

Where things get truly interesting is multi-agent coordination. When multiple Steves work on the same task, they don’t just perform actions independently — they actively collaborate, preventing conflicts and optimizing workload distribution. Tell three agents to build a castle, and they’ll automatically partition the blueprint, assign sections among themselves, and construct it in parallel.

These agents aren’t following scripted behaviors. They operate entirely from natural-language instructions, enabling:

  • Resource extraction with optimal mining strategies
  • Autonomous building, including layout planning and material management
  • Combat and defense, with threat assessment and team coordination
  • Exploration and gathering using dynamic pathfinding
  • Collaborative execution, with automatic load-balancing and conflict prevention

How It Works

Each Steve runs an agent loop:

  1. Your command is sent to an LLM (powered by Groq for high-speed inference).
  2. The LLM decomposes your request into structured game instructions.
  3. The instructions execute using Minecraft’s actual mechanics.
  4. If a step fails, the agent queries the LLM again to generate a new plan.

Multi-Agent Coordination

When multiple Steves work together, a dedicated coordination system ensures seamless team behavior:

  • Tasks are automatically divided into sections.
  • Each agent claims a portion of the work.
  • Agents avoid placing blocks in the same location.
  • If one finishes early, responsibilities are rebalanced.

All coordination is handled server-side through a deterministic manager that tracks builds, assigns tasks, and prevents race conditions or inconsistent states.

Ченджлог

1.5.1Релиз1.20.4, 1.20.5, 1.20.6 · 14 июля 2026 г.

Steve AI 1.5.1 Packaging Hotfix

This hotfix corrects the production artifact that caused Minecraft to crash while loading Steve AI.

Fixed

  • Fixed the startup crash caused by the Caffeine runtime library being absent from the installed mod JAR.
  • The complete Forge Jar-in-Jar artifact is now produced as steve-ai-mod-1.5.1.jar.
  • The dependency-free development artifact is now clearly named steve-ai-mod-1.5.1-slim.jar and must not be installed in Minecraft.
  • Maven publishing now uses the complete Jar-in-Jar artifact instead of the slim JAR.
  • The GitHub release workflow now selects the exact installable JAR instead of relying on a filename glob.

Build safeguards

  • Added a verifyReleaseJar Gradle task that checks the installable filename and required runtime contents.
  • Release verification now fails if Caffeine, Resilience4j, or Jar-in-Jar metadata is missing.
  • Updated the build and release documentation to distinguish the installable and development artifacts.

Verification

  • Clean Gradle build completed successfully.
  • 22 JUnit tests passed with no failures, errors, or skipped tests.
  • The dedicated Forge GameTest server passed its required test.
  • The LLM cache initialized successfully during the Forge runtime test.
  • The final installable JAR contains Caffeine 3.1.8 and all five required Resilience4j 2.1.0 modules.
1.5.0Релиз1.20.1 · 14 июля 2026 г.

Steve AI 1.5.0

This release focuses on reliability, safer AI actions, multiplayer behavior, and Forge packaging.

Highlights

  • AI planning no longer blocks the Minecraft server thread.
  • New commands replace outdated plans, and stopped plans cannot resume later.
  • LLM responses are parsed and validated before any Minecraft action runs.
  • Procedural structure building works again when no NBT template is available.
  • Collaborative builds now distribute work correctly, avoid duplicate block claims, and stay isolated by dimension.
  • The release JAR now includes the Caffeine and Resilience4j libraries required at runtime.

Gameplay and action fixes

  • Added strict validation for action names, coordinates, quantities, block lists, and structure dimensions.
  • Protected regions are now respected when mining, placing blocks, building structures, and placing mining torches.
  • Mining only counts blocks that Minecraft successfully destroyed.
  • Tunnel mining no longer removes the floor below Steve.
  • Resource gathering now delegates to the real mining action instead of returning a placeholder result.
  • The unfinished crafting action is no longer exposed as a supported AI action.
  • Combat now attacks at the intended rate instead of applying damage every tick.
  • Pathfinding, following, and block placement stop navigation cleanly after a timeout.
  • Action failures are contained so one broken action cannot break Steve's entire tick loop.
  • Action interceptors now receive start, completion, and failure events.

Structures

  • Fixed a bug that prevented all procedural structures from being generated.
  • Collaborative construction can now be completed by a single Steve across every build section.
  • Cancelled, failed, and timed-out builds release their shared state correctly.
  • Completed structures are registered only after every block has been processed.
  • Structure templates preserve block-state properties such as door halves and orientation data.
  • External NBT templates can be loaded from the structures directory.
  • Invalid template names are rejected to prevent unsafe file paths.
  • Generated houses now place valid upper and lower door blocks.

Server and multiplayer

  • Removed client-only GUI references from common server code.
  • Client GUI registration now happens exclusively during client setup.
  • Steve entities are reconciled with the manager after world and chunk loading.
  • Steve names are validated, case-insensitive, and limited to command-safe characters.
  • Duplicate names and failed entity spawns no longer leave stale manager entries.
  • Steve action executors are created lazily after persisted names are loaded.
  • Event-bus executors are shut down when their Steve entity is removed.
  • Commands that spawn, remove, stop, or direct a Steve now require operator permission level 2.

AI provider reliability

  • Fixed asynchronous retry handling so failed futures are retried correctly.
  • Cache keys now include request parameters, preventing responses with different settings from colliding.
  • Replaced the external SHA-256 helper with Java's built-in implementation.
  • Improved fallback intent matching and removed unsafe default actions for unknown requests.
  • Gemini API keys are sent through the x-goog-api-key header instead of the request URL.
  • Improved handling of malformed, fenced, nested, or partially invalid JSON responses.

Client fixes

  • Fixed duplicate GUI ticking when the overlay screen was open.
  • The GUI now reads Steve entities from the client world instead of server-only singleton state.
  • GUI commands are sent through the normal client-to-server command connection.

Packaging and compatibility

  • Added ForgeGradle Jar-in-Jar metadata and embedded runtime dependencies.
  • Added the missing Gradle wrapper JAR so ./gradlew works from a fresh clone.
  • Release builds can receive their version from the Git tag.
  • Tightened metadata compatibility to Minecraft 1.20.1 and Forge 47.x.
  • Updated the example token limit to match the runtime default of 8,000 tokens.

Tests

  • Added regression coverage for LLM response parsing, fallback behavior, retries, task validation, collaborative construction, and event-bus inheritance.
  • Current test result: 20 tests passed, 0 failed.

Known limitation

  • Crafting is not implemented in this release and is intentionally rejected instead of reporting false success.
1.4.2Релиз1.20.4, 1.20.5, 1.20.6 · 5 июля 2026 г.
  • Bug fixes.
1.4.1Релиз1.20.4, 1.20.5, 1.20.6 · 14 мая 2026 г.

Bug fixes

1.4.0Релиз1.20.4, 1.20.5, 1.20.6 · 4 апреля 2026 г.

Core Mod

  • Minecraft Forge mod entry point (SteveMod) with full lifecycle management (FML events, server start/stop)
  • In-game command interface: open panel with K, issue natural-language commands via /steve spawn <name>

Agent & Action System

  • ActionExecutor — orchestrates action execution with retry logic and error handling
  • CollaborativeBuildManager — multi-agent workload balancing and conflict resolution for building tasks
  • Task / ActionResult — typed task/result models for the agent pipeline
  • Built-in action implementations:
    • BuildStructureAction — autonomous structure planning and block-by-block placement
    • CombatAction — threat assessment and combat coordination
    • CraftItemAction — recipe lookup and crafting execution
    • FollowPlayerAction — pathfinding to track a target player
    • GatherResourceAction — resource location and gathering
    • IdleFollowAction — passive follow behaviour when idle
    • MineResourceAction — optimal mining location selection and execution
    • NavigateAction — point-to-point navigation

LLM Integration

  • TaskPlanner — converts natural-language input into a structured action plan using LLMs
  • PromptBuilder / ResponseParser — prompt construction and response parsing utilities
  • Synchronous clients: OpenAIClient, GroqClient, GeminiClient
  • Asynchronous clients: AsyncOpenAIClient, AsyncGroqClient, AsyncGeminiClient (non-blocking LLM calls)
  • LLMCache — response caching to reduce API calls and latency
  • LLMExecutorService — managed thread pool for async LLM requests
  • LLMFallbackHandler / ResilientLLMClient — automatic provider fallback and resilience policies
  • ResilienceConfig — configurable retry, timeout, and circuit-breaker settings
  • LLMException / LLMResponse — typed error and response models

Execution Engine

  • SteveAPI — public API surface for issuing commands to agents
  • CodeExecutionEngine — safe execution sandbox for generated code snippets
  • MetricsInterceptor — records execution metrics (latency, success rate, token usage)

Memory & World Knowledge

  • SteveMemory — per-agent short-term and long-term memory store
  • WorldKnowledge — shared world state cache (blocks, entities, resources)
  • StructureRegistry — persistent registry of built and planned structures

Plugin System

  • ActionPlugin / ActionFactory / ActionRegistry — extensible plugin API for registering custom actions
  • CoreActionsPlugin — default plugin bundling all built-in actions
  • PluginManager — runtime plugin discovery and lifecycle management via Java ServiceLoader

Structure Generation

  • StructureGenerators — procedural generators for common structures (houses, walls, platforms, etc.)
  • StructureTemplateLoader — loads structure templates from resource files
  • BlockPlacement — value type representing a single block placement operation

Utilities & Events

  • ActionUtils — shared helper methods for action implementations
  • SteveCommands — registers and handles all Forge commands
  • ServerEventHandler — listens to server-side Minecraft events (player join/leave, world tick, etc.)

Configuration

  • TOML config file (config/steve-common.toml) supporting OpenAI, Groq, and Gemini API credentials and model settings
  • Config example file (config/steve-common.toml.example)

Testing

  • Unit test stubs for ActionExecutor, TaskPlanner, WorldKnowledge, and StructureGenerators

Infrastructure

  • Gradle build system with Minecraft Forge (1.20.1-47.3.0) setup
  • gradlew / gradlew.bat wrapper scripts
  • scripts/run_steve.sh helper script for local development
  • Resource files: mods.toml, en_us.json language file, pack.mcmeta
  • .gitignore covering Gradle, IDE, and OS-specific files
1.2.0Релиз1.20.4, 1.20.5, 1.20.6 · 21 февраля 2026 г.
  • Bug fixes
1.1.0Релиз1.20.4, 1.20.5, 1.20.6 · 5 января 2026 г.

🛠 Fixes & Improvements

⛏ Iron Mining

  • Adjusted the optimal iron mining depth to Y=16.
  • Forced Steve to descend until solid ground is found at the correct Y level before starting to mine.
  • Fixed the issue where Steve would mine straight forward at surface level, “flying” and placing torches in mid-air.
  • Improved torch placement logic to prevent placement in invalid (air) positions.

💎 Ore Drops

  • Removed incorrect inventory handling from Steve.
  • Kept destroyBlock(..., true) so blocks correctly drop items on the ground for the player to collect.

🏠 Structure Building (House)

  • Fixed a bug where only half of the house was being built.
  • Removed doors from the build plan (doors require special two-block placement logic and were being dropped).
  • Increased the number of blocks placed per tick to speed up construction.
  • Fixed compilation and initialization errors related to BlockState.

🖥 GUI (SteveGUI)

  • Removed Minecraft formatting codes (§l, §7, §8) that caused garbled characters.
  • Removed special characters (such as ) that caused encoding issues (e.g. “ât”).
  • Cleaned up chat text rendering for better readability.

🧪 Stability & Debugging

  • Fixed multiple potential StringIndexOutOfBoundsException and IndexOutOfBoundsException crashes.
  • Added structured debug logging (NDJSON) for better runtime diagnostics.
  • Stabilized the build process and ensured the mod compiles successfully.

✅ Current Status

  • Mining: Steve now correctly descends to the proper depth before mining.
  • Building: Structures are built fully (without doors).
  • Follow command: Working as expected.
  • UI: No more corrupted or confusing characters.
  • Model responses: Still echo commands (prompt / API limitation, not a code issue).

Комментарии

Загружаем…