synth-ai 0.2.0__py3-none-any.whl → 0.2.1.dev0__py3-none-any.whl

synth_ai/environments/examples/crafter_classic/agent_demos/test_crafter_react_agent.py

@@ -0,0 +1,1110 @@
+#!/usr/bin/env python3
+"""
+Test script to run ReAct agents against Crafter environment on synth service (port 8901)
+Tests on multiple easy Crafter instances with enhanced debugging
+"""
+
+import asyncio
+import json
+import uuid
+import math
+import argparse
+import toml
+import logging
+from datetime import datetime
+from typing import Dict, Any, Optional, List, Set
+from pydantic import BaseModel, Field
+from httpx import AsyncClient
+import sys
+import os
+from pathlib import Path
+from tqdm.asyncio import tqdm_asyncio
+
+# Add the src directory to the path
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..", "..", "..", "src"))
+
+from synth_ai.zyk import LM
+from synth_ai.zyk.lms.tools.base import BaseTool
+import numpy as np
+
+
+# --- Configuration Class ---
+class CrafterConfig:
+    """Configuration for Crafter evaluation."""
+
+    def __init__(self, config_path: Optional[str] = None):
+        # Default values
+        self.model_name: Optional[str] = None  # Must be provided via config or CLI
+        self.num_instances = 3
+        self.max_turns = 20
+        self.difficulty = "easy"
+        self.service_base_url = "http://localhost:8901"
+        self.service_timeout = 30.0
+        self.seed = 42
+        self.save_traces = True
+        self.save_detailed_results = True
+        self.verbose = False
+
+        # Custom OpenAI endpoint support
+        self.custom_openai_base_url = None  # e.g., "https://lora-inference-service-xyz.modal.run"
+        self.custom_openai_api_key = "dummy"  # Default dummy key for custom endpoints
+
+        # Load from TOML if provided
+        if config_path and os.path.exists(config_path):
+            self.load_from_toml(config_path)
+
+        # Fail fast if no model name provided
+        # Configure custom OpenAI endpoint if specified
+        self._configure_custom_openai()
+
+    def load_from_toml(self, config_path: str):
+        """Load configuration from TOML file."""
+        config = toml.load(config_path)
+
+        # Extract eval settings
+        eval_config = config.get("eval", {})
+        self.model_name = eval_config.get("model_name", self.model_name)
+        self.num_instances = eval_config.get("episodes", self.num_instances)
+        self.max_turns = eval_config.get("max_steps", self.max_turns)
+        self.difficulty = eval_config.get("difficulty", self.difficulty)
+        self.seed = eval_config.get("seed", self.seed)
+
+        # Extract service settings
+        service_config = config.get("service", {})
+        self.service_base_url = service_config.get("base_url", self.service_base_url)
+        self.service_timeout = service_config.get("timeout", self.service_timeout)
+
+        # Extract output settings
+        output_config = config.get("output", {})
+        self.save_traces = output_config.get("save_traces", self.save_traces)
+        self.save_detailed_results = output_config.get(
+            "save_detailed_results", self.save_detailed_results
+        )
+
+        # Extract custom OpenAI endpoint settings
+        openai_config = config.get("openai", {})
+        self.custom_openai_base_url = openai_config.get("base_url", self.custom_openai_base_url)
+        self.custom_openai_api_key = openai_config.get("api_key", self.custom_openai_api_key)
+
+    def _configure_custom_openai(self):
+        """Configure environment variables for custom OpenAI endpoint if specified."""
+        if self.custom_openai_base_url:
+            # Ensure the base URL ends with /v1 for OpenAI compatibility
+            base_url = self.custom_openai_base_url.rstrip("/")
+            if not base_url.endswith("/v1"):
+                base_url += "/v1"
+
+            # Set environment variables for OpenAI SDK
+            os.environ["OPENAI_BASE_URL"] = base_url
+            os.environ["OPENAI_API_KEY"] = self.custom_openai_api_key
+
+            print(f"🔧 Configured custom OpenAI endpoint: {base_url}")
+            print(f"   API Key: {self.custom_openai_api_key}")
+
+            # Auto-detect if this looks like a fine-tuned model and add ft: regex support
+            if self.model_name and (
+                self.model_name.startswith("ft:") or "lora" in self.model_name.lower()
+            ):
+                self._add_ft_regex_support()
+
+    def _add_ft_regex_support(self):
+        """Add ft: regex pattern to OpenAI naming regexes if not already present."""
+        try:
+            import re
+            from synth_ai.zyk.lms.core import vendor_clients
+
+            # Check if ft: pattern already exists
+            ft_pattern = re.compile(r"^ft:.*$")
+            if not any(
+                pattern.pattern == ft_pattern.pattern
+                for pattern in vendor_clients.openai_naming_regexes
+            ):
+                # Add ft: pattern at the beginning to catch all fine-tuned models
+                vendor_clients.openai_naming_regexes.insert(0, ft_pattern)
+                print(f"✅ Added ft:* regex pattern for fine-tuned model support")
+        except Exception as e:
+            print(f"⚠️  Warning: Could not add ft: regex pattern: {e}")
+
+    def set_custom_endpoint(self, base_url: str, api_key: str = "dummy"):
+        """Programmatically set custom OpenAI endpoint."""
+        self.custom_openai_base_url = base_url
+        self.custom_openai_api_key = api_key
+        self._configure_custom_openai()
+
+
+# --- Global Config ---
+config = CrafterConfig()
+
+
+# --- Helper to build crafter semantic mapping ---
+def get_crafter_semantic_mapping():
+    """Build the crafter semantic ID to item name mapping."""
+    try:
+        import crafter
+        import itertools
+
+        # Create a dummy env to get ID mappings
+        dummyenv = None
+        try:
+            dummyenv = crafter.Env()
+            max_id = (
+                max(
+                    max(dummyenv._world._mat_ids.values()),
+                    max(dummyenv._sem_view._obj_ids.values()),
+                )
+                + 1
+            )
+            id_to_item = ["void"] * max_id
+            for name, ind in itertools.chain(
+                dummyenv._world._mat_ids.items(), dummyenv._sem_view._obj_ids.items()
+            ):
+                if name is None:
+                    clean = "none"
+                elif hasattr(name, "__name__"):
+                    clean = name.__name__
+                else:
+                    clean = str(name)
+                id_to_item[ind] = clean.lower()
+            player_idx = id_to_item.index("player")
+            return id_to_item, player_idx
+        finally:
+            if dummyenv:
+                try:
+                    dummyenv.close()
+                except Exception:
+                    pass
+            del dummyenv
+    except ImportError:
+        # Fallback if crafter is not available
+        return None, None
+
+
+def format_semantic_map_view(obs_data: Dict[str, Any], view_size: int = 7) -> str:
+    """Format a semantic map view around the player (ASCII)."""
+    try:
+        # Get mapping list
+        id_to_item, _ = get_crafter_semantic_mapping()
+        if id_to_item is None:
+            return "Map view unavailable (crafter not installed)"
+
+        semantic_map = obs_data.get("semantic_map")
+        player_position = obs_data.get("player_position", [0, 0])
+
+        if semantic_map is None:
+            return "Map view unavailable (no semantic map data)"
+
+        # Ensure numpy array with 2 dimensions
+        sem_arr = np.asarray(semantic_map)
+        if sem_arr.ndim == 1:
+            # Probably flattened; try to infer square size
+            size = int(np.sqrt(sem_arr.size))
+            sem_arr = sem_arr.reshape(size, size)
+        elif sem_arr.ndim != 2:
+            return "Map view unavailable (invalid map dimensionality)"
+
+        px, py = map(int, player_position)
+        half = view_size // 2
+
+        rows = []
+        visible = set()
+        for dy in range(-half, half + 1):
+            row_tokens = []
+            for dx in range(-half, half + 1):
+                x, y = px + dx, py + dy
+                if 0 <= x < sem_arr.shape[0] and 0 <= y < sem_arr.shape[1]:
+                    if dx == 0 and dy == 0:
+                        token = "player"
+                    else:
+                        idx = int(sem_arr[x, y])
+                        token = id_to_item[idx] if idx < len(id_to_item) else "?"
+                else:
+                    token = "void"
+                row_tokens.append(token)
+                if token not in {"void", "player"}:
+                    visible.add(token)
+            rows.append(" ".join(row_tokens))
+
+        map_view = f"\nLocal Map View ({view_size}x{view_size}):\n" + "\n".join(rows)
+        if visible:
+            map_view += "\nVisible items: " + ", ".join(sorted(visible))
+        else:
+            map_view += "\nNo special items visible (mostly grass/empty)"
+        return map_view
+    except Exception as e:
+        return f"Map view error: {e}"
+
+
+# --- Shaped Reward Configuration ---
+# K-values for shaped reward calculation: reward = sum(K * log(count)) for each achievement
+ACHIEVEMENT_K_VALUES = {
+    "collect_coal": 3.0,
+    "collect_diamond": 100.0,
+    "collect_drink": 0.1,
+    "collect_iron": 10.0,
+    "collect_sapling": 0.1,
+    "collect_stone": 1.0,
+    "collect_wood": 1.0,
+    "defeat_skeleton": 1.0,
+    "defeat_zombie": 1.0,
+    "eat_cow": 1.0,
+    "eat_plant": 0.1,
+    "make_iron_pickaxe": 30.0,
+    "make_iron_sword": 30.0,
+    "make_stone_pickaxe": 10.0,
+    "make_stone_sword": 10.0,
+    "make_wood_pickaxe": 3.0,
+    "make_wood_sword": 3.0,
+    "place_furnace": 10.0,
+    "place_plant": 0.1,
+    "place_stone": 1.0,
+    "place_table": 3.0,
+    "wake_up": 0.1,
+}
+
+
+# --- Tool Definitions ---
+class CrafterActionArgs(BaseModel):
+    """Arguments for crafter actions."""
+
+    actions: List[str] = Field(
+        description="List of 1-5 action names to execute in sequence (e.g., ['move_up', 'do', 'mine_down'])"
+    )
+    reasoning: str = Field(description="Brief explanation of why these actions were chosen")
+
+
+class TerminateArgs(BaseModel):
+    """Arguments for termination."""
+
+    reason: str = Field(description="Reason for termination")
+
+
+class CrafterActionTool(BaseTool):
+    """Tool for performing actions in the Crafter environment."""
+
+    name: str = "interact"
+    arguments: type[BaseModel] = CrafterActionArgs
+    description: str = "Perform 1-5 actions in sequence in the Crafter environment."
+
+
+class TerminateTool(BaseTool):
+    """Tool to terminate the episode."""
+
+    name: str = "terminate"
+    arguments: type[BaseModel] = TerminateArgs
+    description: str = "End the episode when finished or no progress can be made."
+
+
+# --- Shaped Reward Helper ---
+def calculate_shaped_reward(achievement_counts: Dict[str, int]) -> Dict[str, Any]:
+    """Calculate shaped reward using K * log(count) for each achievement."""
+    total_reward = 0.0
+    reward_breakdown = {}
+
+    for achievement, count in achievement_counts.items():
+        if count > 0 and achievement in ACHIEVEMENT_K_VALUES:
+            k_value = ACHIEVEMENT_K_VALUES[achievement]
+            # Use log(count + 1) to handle count=0 case gracefully
+            reward_contribution = k_value * math.log(count + 1)
+            total_reward += reward_contribution
+            reward_breakdown[achievement] = {
+                "count": count,
+                "k_value": k_value,
+                "contribution": reward_contribution,
+            }
+
+    return {"total_shaped_reward": total_reward, "breakdown": reward_breakdown}
+
+
+# --- Base ReAct Agent ---
+class BaseReActAgent:
+    """Base ReAct agent for environment interaction."""
+
+    def __init__(self, llm: LM, max_turns: int = 20, verbose: bool = False):
+        self.llm = llm
+        self.max_turns = max_turns
+        self.verbose = verbose
+        self.history = []
+        self.system_name = "base-react-agent"
+
+        # Define tools in OpenAI format
+        self.tools = [
+            CrafterActionTool(),
+            TerminateTool(),
+        ]
+
+    async def decide(self, obs: str, system_message: str, turn: int) -> Dict[str, Any]:
+        """Get agent decision based on observation."""
+        # Create conversation context
+        context = f"Turn {turn + 1}/{self.max_turns}\n\n{obs}"
+        # Generate response using LLM
+        response_obj = await self.llm.respond_async(
+            system_message=system_message, user_message=context, tools=self.tools
+        )
+
+        tool_calls = response_obj.tool_calls
+
+        # Handle case where tool_calls is None or empty (graceful fallback)
+        if not tool_calls:
+            if self.verbose:
+                print(f"[WARNING] No tool calls returned by LLM, using default action")
+            return {
+                "name": "interact",
+                "parameters": {
+                    "actions": ["do"],
+                    "reasoning": "Default action - no tool call received",
+                },
+            }
+
+        tool_call_data = tool_calls[0]
+
+        # Handle both dict and object formats
+        if isinstance(tool_call_data, dict):
+            tool_name = tool_call_data["function"]["name"]
+            tool_args_str = tool_call_data["function"]["arguments"]
+        else:
+            tool_name = tool_call_data.function.name
+            tool_args_str = tool_call_data.function.arguments
+
+        tool_arguments = json.loads(tool_args_str)
+
+        return {"name": tool_name, "parameters": tool_arguments}
+
+
+# --- Crafter ReAct Agent ---
+class CrafterReActAgent(BaseReActAgent):
+    """ReAct agent for Crafter environment."""
+
+    def __init__(self, llm: LM, max_turns: int = 20, verbose: bool = False):
+        super().__init__(llm, max_turns, verbose)
+        self.system_name = "crafter-react-agent"
+
+    def get_system_message(self) -> str:
+        return """You are CrafterAgent playing Crafter survival environment. Your goal is to unlock as many achievements as possible while staying alive.
+
+You will see a semantic map view showing your surroundings. Use this to navigate toward resources.
+
+Key mechanics:
+• 'do' action: collect wood from trees, stone from deposits, food from cows/plants
+• 'do' does nothing on grass/water - move to find resources first
+• Craft progression: wood → table → wood_pickaxe → stone → stone_pickaxe → iron tools
+• Sleep when energy low to restore and unlock wake_up achievement
+• Use semantic map view to navigate toward resources you can see
+
+Available actions: move_left, move_right, move_up, move_down, do, sleep, place_stone, place_table, place_furnace, place_plant, make_wood_pickaxe, make_stone_pickaxe, make_iron_pickaxe, make_wood_sword, make_stone_sword, make_iron_sword, noop
+
+Strategy:
+1. Look at the semantic map to see what's around you
+2. Move toward trees to collect wood with 'do'
+3. Once you have wood, place a table to enable crafting
+4. Make a wood pickaxe to collect stone more efficiently
+5. Progress to stone pickaxe, then iron tools
+6. Eat food when health is low, sleep when energy is low
+
+You should provide 1-5 actions in sequence for efficient gameplay. Use the semantic map view to navigate toward visible resources.
+
+Example good action sequences:
+- ['move_right', 'move_right', 'do'] (move to tree and collect wood)
+- ['place_table', 'make_wood_pickaxe'] (craft progression)
+- ['move_up', 'do', 'move_down', 'do'] (collect from multiple resources)
+
+Be strategic and use the map view to find resources! Focus on unlocking achievements."""
+
+    def format_observation(self, obs: Dict[str, Any]) -> str:
+        """Format observation for Crafter with rich context."""
+        # Extract key information from observation
+        health = obs.get("health", 0)
+        inventory = obs.get("inventory", {})
+
+        # Extract health from inventory if not in main obs
+        if health == 0 and "health" in inventory:
+            health = inventory["health"]
+
+        # Format inventory items (exclude health since we show it separately)
+        inventory_items = []
+        for item, count in inventory.items():
+            if count > 0 and item != "health":
+                inventory_items.append(f"{item}: {count}")
+
+        inventory_str = ", ".join(inventory_items) if inventory_items else "empty"
+
+        # Get achievements
+        achievements = obs.get("achievements") or obs.get("achievements_status", {})
+        unlocked_achievements = [name for name, unlocked in achievements.items() if unlocked]
+        achievements_str = ", ".join(unlocked_achievements) if unlocked_achievements else "none"
+
+        # Get position and other state
+        position = obs.get("position", [0, 0])
+        player_position = obs.get("player_position", position)
+        player_direction = obs.get("player_direction", [0, 1])
+        num_steps = obs.get("num_steps_taken", 0)
+
+        # Check termination status
+        terminated = obs.get("terminated", False)
+
+        # Get semantic map view
+        map_view = format_semantic_map_view(obs, view_size=5)
+
+        return (
+            f"Crafter Game State:\n"
+            f"Step: {num_steps}\n"
+            f"Health: {health}\n"
+            f"Position: {player_position}\n"
+            f"Direction: {player_direction}\n"
+            f"Inventory: {inventory_str}\n"
+            f"Achievements: {achievements_str}\n"
+            f"Terminated: {terminated}\n"
+            f"{map_view}\n\n"
+            f"Available actions: move_left, move_right, move_up, move_down, do, sleep, place_stone, place_table, place_furnace, place_plant, make_wood_pickaxe, make_stone_pickaxe, make_iron_pickaxe, make_wood_sword, make_stone_sword, make_iron_sword, noop\n\n"
+            f"Key mechanics:\n"
+            f"• 'do' action: collect wood from trees, stone from deposits, food from cows/plants\n"
+            f"• 'do' does nothing on grass/water - move to find resources\n"
+            f"• Craft progression: wood → table → wood_pickaxe → stone → stone_pickaxe → iron tools\n"
+            f"• Sleep when energy low to restore and unlock wake_up achievement\n\n"
+            f"Choose 1-5 actions to execute in sequence. Focus on exploring to find resources and crafting tools to unlock achievements."
+        )
+
+
+# --- Episode Runner ---
+async def run_single_episode(
+    client: AsyncClient, agent: CrafterReActAgent, task_instance, instance_num: int
+) -> Dict[str, Any]:
+    """Run a single Crafter episode and return episode metrics."""
+    try:
+        # Create environment using the task instance
+        create_resp = await client.post(
+            f"/env/CrafterClassic/initialize",
+            json={"task_instance": await task_instance.serialize()},
+        )
+
+        if create_resp.status_code != 200:
+            print(
+                f"  Instance {instance_num}: Failed to create environment - {create_resp.status_code}: {create_resp.text}"
+            )
+            return {
+                "eval_metric": 0.0,
+                "rubric": {},
+                "total_reward": 0.0,
+                "num_achievements": 0,
+                "terminated": False,
+                "error": True,
+            }
+
+        env_id = create_resp.json()["env_id"]
+
+        # Get initial observation
+        obs = create_resp.json()["observation"]
+        formatted_obs = agent.format_observation(obs)
+
+        # DEBUG: Print initial state (minimal)
+        print(
+            f"\n  Instance {instance_num}: Starting Crafter survival ({task_instance.metadata.difficulty}, {agent.max_turns} turns max)"
+        )
+
+        # Track episode metrics
+        total_reward = 0.0
+        final_achievements = {}
+        num_achievements = 0
+        terminated = False
+        rollout_length = 0
+
+        # Run episode
+        for turn in range(agent.max_turns):
+            try:
+                # Get agent decision
+                action = await agent.decide(formatted_obs, agent.get_system_message(), turn)
+                # print(f"  ✅ Agent decision received: {action}")
+
+                # # DEBUG: Print agent decision with safer access
+                # try:
+                #     actions = action.get('parameters', {}).get('actions', action.get('arguments', {}).get('actions', []))
+                #     reasoning = action.get('parameters', {}).get('reasoning', action.get('arguments', {}).get('reasoning', 'no reasoning'))
+                #     #print(f"  Turn {turn+1}: Agent chose {actions} - {reasoning}")
+                # except Exception as e:
+                #     print(f"  Turn {turn+1}: Agent action structure: {action}")
+                #     print(f"  Error parsing action: {e}")
+
+                # Check for termination
+                if action["name"] == "terminate":
+                    reason = action.get("parameters", {}).get(
+                        "reason", action.get("arguments", {}).get("reason", "no reason given")
+                    )
+                    print(f"  Agent terminated: {reason}")
+                    break
+
+                # Execute actions in environment with safer access
+                action_sequence = action.get("parameters", {}).get(
+                    "actions", action.get("arguments", {}).get("actions", [])
+                )
+                if not action_sequence:
+                    print(f"  ⚠️  No actions found in: {action}")
+                    continue
+
+                # Convert action names to integers using the proper action map
+                # Define the proper Crafter action mapping
+                CRAFTER_ACTION_MAP = {
+                    "noop": 0,
+                    "move_left": 1,
+                    "move_right": 2,
+                    "move_up": 3,
+                    "move_down": 4,
+                    "do": 5,
+                    "sleep": 6,
+                    "place_stone": 7,
+                    "place_table": 8,
+                    "place_furnace": 9,
+                    "place_plant": 10,
+                    "make_wood_pickaxe": 11,
+                    "make_stone_pickaxe": 12,
+                    "make_iron_pickaxe": 13,
+                    "make_wood_sword": 14,
+                    "make_stone_sword": 15,
+                    "make_iron_sword": 16,
+                }
+
+                action_ints = []
+                for action_name in action_sequence:
+                    if action_name in CRAFTER_ACTION_MAP:
+                        action_int = CRAFTER_ACTION_MAP[action_name]
+                    else:
+                        action_int = 0  # Default to noop
+                    action_ints.append(action_int)
+
+                # Execute each action individually (Crafter expects single actions)
+                for i, action_int in enumerate(action_ints):
+                    step_resp = await client.post(
+                        f"/env/CrafterClassic/step",
+                        json={
+                            "env_id": env_id,
+                            "request_id": str(uuid.uuid4()),
+                            "action": {
+                                "tool_calls": [{"tool": "interact", "args": {"action": action_int}}]
+                            },
+                        },
+                    )
+
+                    if step_resp.status_code != 200:
+                        print(
+                            f"    ❌ Action {i + 1} failed: {step_resp.status_code}: {step_resp.text}"
+                        )
+                        break
+
+                    # Update observation after each action
+                    obs = step_resp.json()["observation"]
+
+                # Check final response status
+                if step_resp.status_code != 200:
+                    break
+
+                # Show final state after all actions
+                formatted_obs = agent.format_observation(obs)
+                step_count = obs.get("num_steps_taken", 0)
+                rollout_length = step_count
+                position = obs.get("player_position", [0, 0])
+                # print(f"  Turn {turn+1}: Actions completed - Step: {step_count}, Position: {position}")
+
+                # Update history with safer access
+                reasoning = action.get("parameters", {}).get(
+                    "reasoning", action.get("arguments", {}).get("reasoning", "")
+                )
+                agent.history.append(f"{', '.join(action_sequence)}: {reasoning[:50]}")
+
+                # Track episode progress - Use the FINAL observation from the last action
+                terminated = obs.get("terminated", False)
+                step_reward = obs.get("reward", 0.0)
+                total_reward += step_reward
+                achievements = obs.get("achievements") or obs.get("achievements_status", {})
+
+                # ALWAYS update final_achievements with the latest observation
+                final_achievements = achievements
+
+                num_achievements = sum(1 for v in achievements.values() if v) if achievements else 0
+
+                if terminated:
+                    print(
+                        f"  ✅ Instance {instance_num}: Episode completed! Achievements: {num_achievements}, Total reward: {total_reward:.3f}"
+                    )
+                    break
+
+            except Exception as e:
+                print(f"  ❌ Error in turn {turn + 1}: {e}")
+                import traceback
+
+                traceback.print_exc()
+                break
+
+        # Cleanup
+        await client.post(f"/env/CrafterClassic/terminate", json={"env_id": env_id})
+
+        # Calculate K-weighted achievement reward
+        achievement_reward = 0.0
+        if final_achievements:
+            for achievement, unlocked in final_achievements.items():
+                if unlocked and achievement in ACHIEVEMENT_K_VALUES:
+                    k_value = ACHIEVEMENT_K_VALUES[achievement]
+                    achievement_reward += k_value * math.log(2)  # log(1+1) for single achievement
+
+        # Use achievement reward as the total reward
+        total_reward = achievement_reward
+
+        # Calculate eval metric and rubric
+        eval_metric = float(num_achievements)  # Simple metric: number of achievements
+
+        # Create rubric with specific achievement checks
+        rubric = {}
+        if final_achievements:
+            rubric = {
+                "collect_wood": 1.0 if final_achievements.get("collect_wood", False) else 0.0,
+                "collect_stone": 1.0 if final_achievements.get("collect_stone", False) else 0.0,
+                "collect_coal": 1.0 if final_achievements.get("collect_coal", False) else 0.0,
+                "collect_iron": 1.0 if final_achievements.get("collect_iron", False) else 0.0,
+                "collect_diamond": 1.0 if final_achievements.get("collect_diamond", False) else 0.0,
+                "place_table": 1.0 if final_achievements.get("place_table", False) else 0.0,
+                "place_furnace": 1.0 if final_achievements.get("place_furnace", False) else 0.0,
+                "make_wood_pickaxe": 1.0
+                if final_achievements.get("make_wood_pickaxe", False)
+                else 0.0,
+                "make_stone_pickaxe": 1.0
+                if final_achievements.get("make_stone_pickaxe", False)
+                else 0.0,
+                "make_iron_pickaxe": 1.0
+                if final_achievements.get("make_iron_pickaxe", False)
+                else 0.0,
+                "make_wood_sword": 1.0 if final_achievements.get("make_wood_sword", False) else 0.0,
+                "make_stone_sword": 1.0
+                if final_achievements.get("make_stone_sword", False)
+                else 0.0,
+                "make_iron_sword": 1.0 if final_achievements.get("make_iron_sword", False) else 0.0,
+                "defeat_skeleton": 1.0 if final_achievements.get("defeat_skeleton", False) else 0.0,
+                "defeat_zombie": 1.0 if final_achievements.get("defeat_zombie", False) else 0.0,
+                "wake_up": 1.0 if final_achievements.get("wake_up", False) else 0.0,
+                "eat_cow": 1.0 if final_achievements.get("eat_cow", False) else 0.0,
+                "eat_plant": 1.0 if final_achievements.get("eat_plant", False) else 0.0,
+            }
+        else:
+            # Default rubric with all zeros
+            rubric = {
+                "collect_wood": 0.0,
+                "collect_stone": 0.0,
+                "collect_coal": 0.0,
+                "collect_iron": 0.0,
+                "collect_diamond": 0.0,
+                "place_table": 0.0,
+                "place_furnace": 0.0,
+                "make_wood_pickaxe": 0.0,
+                "make_stone_pickaxe": 0.0,
+                "make_iron_pickaxe": 0.0,
+                "make_wood_sword": 0.0,
+                "make_stone_sword": 0.0,
+                "make_iron_sword": 0.0,
+                "defeat_skeleton": 0.0,
+                "defeat_zombie": 0.0,
+                "wake_up": 0.0,
+                "eat_cow": 0.0,
+                "eat_plant": 0.0,
+            }
+
+        return {
+            "eval_metric": eval_metric,
+            "rubric": rubric,
+            "total_reward": total_reward,
+            "num_achievements": num_achievements,
+            "achievements": final_achievements,
+            "rollout_length": rollout_length,
+            "terminated": terminated,
+            "error": False,
+        }
+
+    except Exception as e:
+        print(f"  Instance {instance_num}: Error - {e}")
+        import traceback
+
+        traceback.print_exc()
+        return {
+            "eval_metric": 0.0,
+            "rubric": {},
+            "total_reward": 0.0,
+            "num_achievements": 0,
+            "terminated": False,
+            "error": True,
+        }
+
+
+# --- Batch Evaluation ---
+async def evaluate_crafter_batch() -> Dict[str, Any]:
+    """Evaluate Crafter agent on multiple easy instances."""
+    print(f"🎯 Evaluating Crafter on {config.num_instances} {config.difficulty} instances...")
+
+    llm = LM(model_name=config.model_name, formatting_model_name=config.model_name, temperature=0.0)
+
+    # Get easy task instances using the taskset system
+    from synth_ai.environments.examples.crafter_classic.taskset import (
+        CrafterTaskInstance,
+        CrafterTaskInstanceMetadata,
+    )
+    from synth_ai.environments.tasks.core import Impetus, Intent
+
+    easy_task_instances = []
+    for seed in range(config.num_instances):
+        try:
+            metadata = CrafterTaskInstanceMetadata(
+                difficulty=config.difficulty,
+                seed=seed,
+                num_trees_radius=5,  # Good for easy difficulty
+                num_cows_radius=2,
+                num_hostiles_radius=0,  # No hostiles for easy
+            )
+            task_instance = CrafterTaskInstance(
+                id=uuid.uuid4(),
+                impetus=Impetus(
+                    instructions=f"Survive and unlock achievements in an {config.difficulty} environment."
+                ),
+                intent=Intent(rubric={}, gold_trajectories=None, gold_state_diff={}),
+                metadata=metadata,
+                is_reproducible=True,
+                initial_engine_snapshot=None,
+            )
+            easy_task_instances.append(task_instance)
+        except Exception as e:
+            print(f"  ⚠️  Failed to create task instance for seed {seed}: {e}")
+            continue
+
+    print(f"  📝 Generated {len(easy_task_instances)} {config.difficulty} task instances")
+
+    async with AsyncClient(
+        base_url=config.service_base_url, timeout=config.service_timeout
+    ) as client:
+        # Run trajectories in parallel batches of 4
+        batch_size = 4
+        all_results = []
+
+        for batch_start in range(0, len(easy_task_instances), batch_size):
+            batch_end = min(batch_start + batch_size, len(easy_task_instances))
+            batch_instances = easy_task_instances[batch_start:batch_end]
+
+            print(
+                f"  🚀 Running batch {batch_start // batch_size + 1} ({len(batch_instances)} episodes)..."
+            )
+
+            # Create tasks for this batch
+            batch_tasks = []
+            for i, task_instance in enumerate(batch_instances):
+                agent = CrafterReActAgent(llm, max_turns=config.max_turns, verbose=False)
+                batch_tasks.append(
+                    run_single_episode(client, agent, task_instance, batch_start + i + 1)
+                )
+
+            # Run this batch in parallel with progress bar
+            batch_results = await tqdm_asyncio.gather(
+                *batch_tasks, desc=f"Batch {batch_start // batch_size + 1}", unit="episode"
+            )
+            all_results.extend(batch_results)
+
+            print(f"  ✅ Batch {batch_start // batch_size + 1} completed")
+
+        results = all_results
+
+        # Filter out error results
+        valid_results = [r for r in results if not r.get("error", False)]
+
+        if not valid_results:
+            return {
+                "eval_metrics": [],
+                "mean_eval_metric": 0.0,
+                "mean_rubric": {},
+                "num_episodes": 0,
+            }
+
+        # Extract eval metrics and rubrics
+        eval_metrics = [r["eval_metric"] for r in valid_results]
+        mean_eval_metric = sum(eval_metrics) / len(eval_metrics)
+
+        # --- Rollout length statistics ---
+        rollout_lengths = [r["rollout_length"] for r in valid_results]
+        sorted_lengths = sorted(rollout_lengths)
+        n_lengths = len(sorted_lengths)
+        # Median (Q2)
+        if n_lengths % 2 == 1:
+            q2_rollout = sorted_lengths[n_lengths // 2]
+        else:
+            q2_rollout = (sorted_lengths[n_lengths // 2 - 1] + sorted_lengths[n_lengths // 2]) / 2
+        # 90th percentile (P90)
+        p90_index = int(0.9 * (n_lengths - 1))
+        p90_rollout = sorted_lengths[p90_index]
+        max_rollout = sorted_lengths[-1]
+
+        # Calculate mean rubric values
+        all_rubric_keys = set()
+        for r in valid_results:
+            all_rubric_keys.update(r["rubric"].keys())
+
+        mean_rubric = {}
+        for key in all_rubric_keys:
+            values = [r["rubric"].get(key, 0.0) for r in valid_results]
+            mean_rubric[key] = sum(values) / len(values)
+
+        # Calculate shaped reward (training rubric)
+        # Count total achievements across all episodes
+        achievement_counts = {}
+        unique_achievements_per_trajectory = []
+        all_unique_achievements = set()
+
+        for result in valid_results:
+            achievements = result.get("achievements", {})
+            trajectory_achievements = set()
+            for achievement, unlocked in achievements.items():
+                if unlocked:
+                    achievement_counts[achievement] = achievement_counts.get(achievement, 0) + 1
+                    trajectory_achievements.add(achievement)
+                    all_unique_achievements.add(achievement)
+            unique_achievements_per_trajectory.append(trajectory_achievements)
+
+        # Calculate shaped reward using the counts
+        shaped_reward_data = calculate_shaped_reward(achievement_counts)
+
+        # Calculate unique achievements by N trajectories
+        unique_achievements_by_n = {}
+        for n in range(1, len(valid_results) + 1):
+            unique_at_n = set()
+            for i in range(n):
+                unique_at_n.update(unique_achievements_per_trajectory[i])
+            unique_achievements_by_n[n] = len(unique_at_n)
+
+        # Create training rubric (normalized shaped reward components)
+        training_rubric = {}
+        total_episodes = len(valid_results)
+        if shaped_reward_data["breakdown"]:
+            for achievement, data in shaped_reward_data["breakdown"].items():
+                # Normalize by number of episodes for comparison
+                training_rubric[achievement] = data["contribution"] / total_episodes
+
+        return {
+            "eval_metrics": eval_metrics,
+            "mean_eval_metric": mean_eval_metric,
+            "mean_rubric": mean_rubric,
+            "achievement_counts": achievement_counts,
+            "shaped_reward_data": shaped_reward_data,
+            "training_rubric": training_rubric,
+            "unique_achievements_per_trajectory": unique_achievements_per_trajectory,
+            "all_unique_achievements": all_unique_achievements,
+            "unique_achievements_by_n": unique_achievements_by_n,
+            "num_episodes": len(valid_results),
+            "q2_rollout": q2_rollout,
+            "p90_rollout": p90_rollout,
+            "max_rollout": max_rollout,
+        }
+
+
+async def main():
+    """Run Crafter evaluation."""
+    # Configure logging to reduce verbosity
+    logging.getLogger("httpx").setLevel(logging.WARNING)
+    logging.getLogger("google_genai").setLevel(logging.WARNING)
+    logging.getLogger("google.generativeai").setLevel(logging.WARNING)
+    logging.getLogger("google_genai.models").setLevel(logging.WARNING)
+    logging.getLogger("google_genai.types").setLevel(logging.WARNING)
+
+    print(f"🎮 Crafter ReAct Agent Evaluation")
+    print(f"Model: {config.model_name}")
+    print(f"Service: {config.service_base_url}")
+    print(f"Instances: {config.num_instances}")
+    print(f"Max Turns: {config.max_turns}")
+    print(f"Difficulty: {config.difficulty}")
+    print(f"Time: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}")
+    print("=" * 50)
+
+    # Test service health
+    async with AsyncClient(base_url=config.service_base_url, timeout=10.0) as client:
+        try:
+            health_resp = await client.get("/health")
+            health_data = health_resp.json()
+
+            if "CrafterClassic" not in health_data.get("supported_environments", []):
+                print("❌ CrafterClassic not available on service")
+                return
+
+            print("✅ Service health check passed")
+
+        except Exception as e:
+            print(f"❌ Service health check failed: {e}")
+            return
+
+    # Run evaluation
+    try:
+        results = await evaluate_crafter_batch()
+
+        print("\n" + "=" * 80)
+        print("🏆 FINAL CRAFTER EVALUATION RESULTS")
+        print("=" * 80)
+
+        # Print eval metrics
+        print(f"📊 EVAL METRICS:")
+        print(f"  Episodes: {results['num_episodes']}")
+        print(f"  Individual Scores: {[f'{x:.1f}' for x in results['eval_metrics']]}")
+        print(f"  Mean Eval Metric: {results['mean_eval_metric']:.2f}")
+
+        # Print standard rubric results
+        print(f"\n🎯 STANDARD RUBRIC RESULTS:")
+        if results["mean_rubric"]:
+            for achievement, score in sorted(results["mean_rubric"].items()):
+                print(f"  {achievement}: {score:.2f}")
+        else:
+            print("  No rubric data available")
+
+        # Print shaped reward results
+        print(f"\n🏋️  TRAINING EVAL SCORE (SHAPED REWARD):")
+        shaped_data = results.get("shaped_reward_data", {})
+        print(f"  Total Shaped Reward: {shaped_data.get('total_shaped_reward', 0.0):.3f}")
+
+        # Print achievement counts and contributions
+        achievement_counts = results.get("achievement_counts", {})
+        if achievement_counts:
+            print(f"\n  Achievement Counts Across All Episodes:")
+            for achievement, count in sorted(achievement_counts.items()):
+                k_value = ACHIEVEMENT_K_VALUES.get(achievement, 0.0)
+                contribution = k_value * math.log(count + 1) if count > 0 else 0.0
+                print(
+                    f"    {achievement}: {count} times (K={k_value:.1f}, contribution={contribution:.3f})"
+                )
+        else:
+            print("  No achievements unlocked")
+
+        # Print training rubric (normalized contributions)
+        print(f"\n🎖️  TRAINING RUBRIC (PER EPISODE):")
+        if results.get("training_rubric"):
+            for achievement, score in sorted(results["training_rubric"].items()):
+                print(f"  {achievement}: {score:.3f}")
+        else:
+            print("  No training rubric data available")
+
+        # Print unique achievements analysis
+        print(f"\n🏆 UNIQUE ACHIEVEMENTS ANALYSIS:")
+        all_unique = results.get("all_unique_achievements", set())
+        print(f"  Total Unique Achievements Unlocked: {len(all_unique)}")
+        if all_unique:
+            print(f"  Unique Achievements: {', '.join(sorted(all_unique))}")
+
+        # Print unique achievements by N trajectories
+        unique_by_n = results.get("unique_achievements_by_n", {})
+        if unique_by_n:
+            print(f"\n📊 UNIQUE ACHIEVEMENTS BY N TRAJECTORIES:")
+            for n in sorted(unique_by_n.keys()):
+                print(f"  After {n} trajectories: {unique_by_n[n]} unique achievements")
+
+        # Calculate average achievements per trajectory
+        achievements_per_trajectory = [
+            len(achievements)
+            for achievements in results.get("unique_achievements_per_trajectory", [])
+        ]
+        if achievements_per_trajectory:
+            avg_achievements = sum(achievements_per_trajectory) / len(achievements_per_trajectory)
+            print(f"\n📈 TRAJECTORY ANALYSIS:")
+            print(f"  Average Achievements per Trajectory: {avg_achievements:.2f}")
+            print(f"  Achievements per Trajectory: {achievements_per_trajectory}")
+            print(f"  Best Trajectory: {max(achievements_per_trajectory)} achievements")
+            print(f"  Worst Trajectory: {min(achievements_per_trajectory)} achievements")
+
+        # Overall assessment
+        print(f"\n🔍 ASSESSMENT:")
+        if results["mean_eval_metric"] >= 3.0:
+            print("🎉 Excellent performance - achieving multiple objectives!")
+        elif results["mean_eval_metric"] >= 1.0:
+            print("✅ Good performance - consistently achieving objectives!")
+        elif results["mean_eval_metric"] >= 0.5:
+            print("⚠️  Moderate performance - some achievements unlocked")
+        else:
+            print("📈 Learning phase - focus on basic survival and resource gathering")
+
+        # Output markdown table row for README collation
+        print(f"\n📋 MARKDOWN TABLE ROW:")
+        print(
+            "| Model            | Episodes | Mean Score | Avg Achievements | Unique Achievements | Shaped Reward | Mean K-Score | Q2 Len | P90 Len | Max Len |"
+        )
+        print(
+            "|------------------|----------|------------|------------------|---------------------|---------------|--------------|--------|---------|---------|"
+        )
+        achievements_per_trajectory = [
+            len(achievements)
+            for achievements in results.get("unique_achievements_per_trajectory", [])
+        ]
+        avg_achievements = (
+            sum(achievements_per_trajectory) / len(achievements_per_trajectory)
+            if achievements_per_trajectory
+            else 0.0
+        )
+        total_unique = len(results.get("all_unique_achievements", set()))
+        shaped_reward = results.get("shaped_reward_data", {}).get("total_shaped_reward", 0.0)
+        mean_k_score = (
+            shaped_reward / results["num_episodes"] if results["num_episodes"] > 0 else 0.0
+        )
+        q2_rollout = results.get("q2_rollout", 0)
+        p90_rollout = results.get("p90_rollout", 0)
+        max_rollout = results.get("max_rollout", 0)
+
+        print(
+            f"| {config.model_name:<16} | {results['num_episodes']:>8} | {results['mean_eval_metric']:>10.2f} | {avg_achievements:>16.2f} | {total_unique:>19} | {shaped_reward:>13.3f} | {mean_k_score:>12.3f} | {q2_rollout:>6} | {p90_rollout:>7} | {max_rollout:>7} |"
+        )
+
+    except Exception as e:
+        print(f"❌ Evaluation failed: {e}")
+
+
+if __name__ == "__main__":
+    # Parse command line arguments
+    parser = argparse.ArgumentParser(description="Run Crafter ReAct Agent Evaluation")
+    parser.add_argument("--config", "-c", type=str, help="Path to TOML configuration file")
+    parser.add_argument("--model", "-m", type=str, help="Model name (overrides config)")
+    parser.add_argument("--episodes", "-e", type=int, help="Number of episodes (overrides config)")
+    parser.add_argument(
+        "--max-turns", "-t", type=int, help="Maximum turns per episode (overrides config)"
+    )
+    parser.add_argument("--difficulty", "-d", type=str, help="Difficulty level (overrides config)")
+
+    # Custom OpenAI endpoint support
+    parser.add_argument(
+        "--openai-base-url",
+        type=str,
+        help="Custom OpenAI-compatible base URL (e.g., https://lora-service.modal.run)",
+    )
+    parser.add_argument(
+        "--openai-api-key",
+        type=str,
+        default="dummy",
+        help="API key for custom endpoint (default: 'dummy')",
+    )
+
+    args = parser.parse_args()
+
+    # Load configuration
+    if args.config:
+        config = CrafterConfig(args.config)
+    else:
+        # Try to load default config
+        default_config_path = (
+            Path(__file__).parent.parent.parent.parent / "evals" / "configs" / "crafter.toml"
+        )
+        if default_config_path.exists():
+            config = CrafterConfig(str(default_config_path))
+        else:
+            config = CrafterConfig()
+
+    # Override with command line arguments
+    if args.model:
+        config.model_name = args.model
+    if args.episodes:
+        config.num_instances = args.episodes
+    if args.max_turns:
+        config.max_turns = args.max_turns
+    if args.difficulty:
+        config.difficulty = args.difficulty
+
+    # Configure custom OpenAI endpoint if provided
+    if args.openai_base_url:
+        config.set_custom_endpoint(args.openai_base_url, args.openai_api_key)
+
+    # Fail fast if model_name still missing
+    if not config.model_name:
+        raise ValueError(
+            "CrafterConfig: 'model_name' must be specified in the TOML config or via --model CLI argument; no fallback default."
+        )
+
+    asyncio.run(main())