Skip to content

configuration.md

Latest commit

 

History

History
164 lines (116 loc) · 6.08 KB

configuration.md

File metadata and controls

164 lines (116 loc) · 6.08 KB

Configuration Guide

Overview

All games in scripts/play_game.py are configured using OmegaConf. Each game has its configuration file at ./src/agent_client/configs/{game}/config.yaml. Configurations are categorized into three main sections: runner, env, and agent.

Core Configuration Sections

Runner Configuration

Controls the evaluation loop behavior.

Parameter Description Type Required Default
max_steps Maximum number of game steps for evaluation int Yes -

Example:

runner:
  max_steps: 50

Environment Configuration

Game-specific settings and task definitions.

Common Parameters

Parameter Description Type Required Default
task Description of the task to be completed string Yes -
action_mode How agent interacts with the game:"semantic" or "gui" string Yes "semantic"
log_path Path where logs are saved (auto-generated from top-level log_path) string Auto -

GUI Mode Parameters

Only needed for games using GUI interaction (e.g., Angry Birds, PVZ, Metro).

Parameter Description Type When Needed Default
window_title Game window title for capture (supports regex) string GUI mode Game-specific
coor_trans Enable coordinate transformation (1000×1000 → actual resolution) bool GUI with VL models false

Common GUI timing parameters:

  • screenshot_interval: Time between screenshots (seconds)
  • wait_after_action: Wait time after executing actions (seconds)
  • timeout: Maximum time before giving up (seconds)

Game-Specific Parameters

Each game may have unique parameters:

  • Slay the Spire: player_class, ascension_level, seed, mod_input_path, mod_output_path
  • Snake: board_size, num_obstacles, game_mode ("discrete" or "realtime")
  • TwentyFourtyEight: target_tile, show_graphic
  • Angry Birds: slingshot_pos_x, slingshot_pos_y, slingshot_pull_ratio
  • PVZ: grid_rows, grid_cols, plant_slot_width

Example:

env:
  task: "Slay the Spire"
  action_mode: "semantic"
  player_class: "IRONCLAD"
  ascension_level: 0
  seed: 0

Agent Configuration

LLM and agent behavior settings.

Required Parameters

Parameter Description Type Required Default
llm_name LLM model name string Yes -
agent_type Agent type to use string Yes "zeroshot_agent"
prompt_path Python module path to prompt string Yes -

Supported LLM models:

  • OpenAI: gpt-4o, gpt-4o-mini
  • Google: gemini-2.5-flash, gemini-2.5-pro
  • Qwen: qwen3-vl-8b-instruct, qwen3-vl-32b-instruct
  • Seed: doubao-seed-1-8-251228

Supported agent types:

  • zeroshot_agent: Direct task completion without memory
  • memory_agent: Uses long-term memory and skill management

LLM API Configuration

For commercial APIs or self-hosted models:

Parameter Description Type Required Default
api_key API key for authentication string API-based ""
api_base_url Base URL for API endpoint string Self-hosted ""
temperature Sampling temperature float No 0.0
repetition_penalty Repetition penalty float No 1.0

Example:

agent:
  llm_name: gpt-4o
  temperature: 0.0
  agent_type: zeroshot_agent
  prompt_path: agent_servers.slay_the_spire.prompts.semantic.zeroshot_agent

Advanced Parameters

Long-term Memory (for memory_agent)
Parameter Description Type Default
long_term_memory_len Number of recent memories to retrieve int 10
embedding_model Embedding model:"openai" or "qwen" string "openai"
embedding_config Additional embedding configuration dict {}

Example:

agent:
  agent_type: memory_agent
  long_term_memory_len: 10
  embedding_model: "qwen"
  embedding_config:
    model: "text-embedding-v4"
    dimensions: 1024
Video Reflection
Parameter Description Type Default
use_reflection Enable video reflection experience bool false
reflection_format Reflection file format:"json" or "txt" string "json"
reflection_file_path Path to reflection file string Auto-generated

Example:

agent:
  use_reflection: true
  reflection_format: "json"
  reflection_file_path: "data/reflections/Snake_reflections.json"
Debug Mode
Parameter Description Type Default
debug_mode Enable debug logging bool false