tools/rl_training_tool.py

#!/usr/bin/env python3
"""
RL Training Tools Module

This module provides tools for running RL training through Tinker-Atropos.
Directly manages training processes without requiring a separate API server.

Features:
- Environment discovery (AST-based scanning for BaseEnv subclasses)
- Configuration management with locked infrastructure settings
- Training run lifecycle via subprocess management
- WandB metrics monitoring

Required environment variables:
- TINKER_API_KEY: API key for Tinker service
- WANDB_API_KEY: API key for Weights & Biases metrics

Usage:
    from tools.rl_training_tool import (
        rl_list_environments,
        rl_select_environment,
        rl_get_current_config,
        rl_edit_config,
        rl_start_training,
        rl_check_status,
        rl_stop_training,
        rl_get_results,
    )
"""

import ast
import asyncio
import importlib.util
import json
import os
import subprocess
import sys
import time
import uuid
import logging
from datetime import datetime
import yaml
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any, Dict, List, Optional

logger = logging.getLogger(__name__)

# ============================================================================
# Path Configuration
# ============================================================================

# Path to tinker-atropos submodule (relative to hermes-agent root)
HERMES_ROOT = Path(__file__).parent.parent
TINKER_ATROPOS_ROOT = HERMES_ROOT / "tinker-atropos"
ENVIRONMENTS_DIR = TINKER_ATROPOS_ROOT / "tinker_atropos" / "environments"
CONFIGS_DIR = TINKER_ATROPOS_ROOT / "configs"
LOGS_DIR = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes")) / "logs" / "rl_training"

def _ensure_logs_dir():
    """Lazily create logs directory on first use (avoid side effects at import time)."""
    if TINKER_ATROPOS_ROOT.exists():
        LOGS_DIR.mkdir(exist_ok=True)

# ============================================================================
# Locked Configuration (Infrastructure Settings)
# ============================================================================

# These fields cannot be changed by the model - they're tuned for our infrastructure
LOCKED_FIELDS = {
    "env": {
        "tokenizer_name": "Qwen/Qwen3-8B",
        "rollout_server_url": "http://localhost:8000",
        "use_wandb": True,
        "max_token_length": 8192,
        "max_num_workers": 2048,
        "worker_timeout": 3600,
        "total_steps": 2500,
        "steps_per_eval": 25,
        "max_batches_offpolicy": 3,
        "inference_weight": 1.0,
        "eval_limit_ratio": 0.1,
    },
    "openai": [
        {
            "model_name": "Qwen/Qwen3-8B",
            "base_url": "http://localhost:8001/v1",
            "api_key": "x",
            "weight": 1.0,
            "num_requests_for_eval": 256,
            "timeout": 3600,
            "server_type": "sglang",  # Tinker uses sglang for actual training
        }
    ],
    "tinker": {
        "lora_rank": 32,
        "learning_rate": 0.00004,
        "max_token_trainer_length": 9000,
        "checkpoint_dir": "./temp/",
        "save_checkpoint_interval": 25,
    },
    "slurm": False,
    "testing": False,
}

LOCKED_FIELD_NAMES = set(LOCKED_FIELDS.get("env", {}).keys())


# ============================================================================
# State Management
# ============================================================================

@dataclass
class EnvironmentInfo:
    """Information about a discovered environment."""
    name: str
    class_name: str
    file_path: str
    description: str = ""
    config_class: str = "BaseEnvConfig"


@dataclass
class RunState:
    """State for a training run."""
    run_id: str
    environment: str
    config: Dict[str, Any]
    status: str = "pending"  # pending, starting, running, stopping, stopped, completed, failed
    error_message: str = ""
    wandb_project: str = ""
    wandb_run_name: str = ""
    start_time: float = 0.0
    # Process handles
    api_process: Optional[subprocess.Popen] = None
    trainer_process: Optional[subprocess.Popen] = None
    env_process: Optional[subprocess.Popen] = None


# Global state
_environments: List[EnvironmentInfo] = []
_current_env: Optional[str] = None
_current_config: Dict[str, Any] = {}
_env_config_cache: Dict[str, Dict[str, Dict[str, Any]]] = {}
_active_runs: Dict[str, RunState] = {}
_last_status_check: Dict[str, float] = {}

# Rate limiting for status checks (30 minutes)
MIN_STATUS_CHECK_INTERVAL = 30 * 60


# ============================================================================
# Environment Discovery
# ============================================================================

def _scan_environments() -> List[EnvironmentInfo]:
    """
    Scan the environments directory for BaseEnv subclasses using AST.
    """
    environments = []
    
    if not ENVIRONMENTS_DIR.exists():
        return environments
    
    for py_file in ENVIRONMENTS_DIR.glob("*.py"):
        if py_file.name.startswith("_"):
            continue
        
        try:
            with open(py_file, "r") as f:
                tree = ast.parse(f.read())
            
            for node in ast.walk(tree):
                if isinstance(node, ast.ClassDef):
                    # Check if class has BaseEnv as base
                    for base in node.bases:
                        base_name = ""
                        if isinstance(base, ast.Name):
                            base_name = base.id
                        elif isinstance(base, ast.Attribute):
                            base_name = base.attr
                        
                        if base_name == "BaseEnv":
                            # Extract name from class attribute if present
                            env_name = py_file.stem
                            description = ""
                            config_class = "BaseEnvConfig"
                            
                            for item in node.body:
                                if isinstance(item, ast.Assign):
                                    for target in item.targets:
                                        if isinstance(target, ast.Name):
                                            if target.id == "name" and isinstance(item.value, ast.Constant):
                                                env_name = item.value.value
                                            elif target.id == "env_config_cls" and isinstance(item.value, ast.Name):
                                                config_class = item.value.id
                                
                                # Get docstring
                                if isinstance(item, ast.Expr) and isinstance(item.value, ast.Constant):
                                    if isinstance(item.value.value, str) and not description:
                                        description = item.value.value.split("\n")[0].strip()
                            
                            environments.append(EnvironmentInfo(
                                name=env_name,
                                class_name=node.name,
                                file_path=str(py_file),
                                description=description or f"Environment from {py_file.name}",
                                config_class=config_class,
                            ))
                            break
        except Exception as e:
            logger.warning("Could not parse %s: %s", py_file, e)
    
    return environments


def _get_env_config_fields(env_file_path: str) -> Dict[str, Dict[str, Any]]:
    """
    Dynamically import an environment and extract its config fields.
    
    Uses config_init() to get the actual config class, with fallback to
    directly importing BaseEnvConfig if config_init fails.
    """
    try:
        # Load the environment module
        spec = importlib.util.spec_from_file_location("env_module", env_file_path)
        module = importlib.util.module_from_spec(spec)
        sys.modules["env_module"] = module
        spec.loader.exec_module(module)
        
        # Find the BaseEnv subclass
        env_class = None
        for name, obj in vars(module).items():
            if isinstance(obj, type) and name != "BaseEnv":
                if hasattr(obj, "config_init") and callable(getattr(obj, "config_init")):
                    env_class = obj
                    break
        
        if not env_class:
            return {}
        
        # Try calling config_init to get the actual config class
        config_class = None
        try:
            env_config, server_configs = env_class.config_init()
            config_class = type(env_config)
        except Exception as config_error:
            # Fallback: try to import BaseEnvConfig directly from atroposlib
            logger.info("config_init failed (%s), using BaseEnvConfig defaults", config_error)
            try:
                from atroposlib.envs.base import BaseEnvConfig
                config_class = BaseEnvConfig
            except ImportError:
                return {}
        
        if not config_class:
            return {}
        
        # Helper to make values JSON-serializable (handle enums, etc.)
        def make_serializable(val):
            if val is None:
                return None
            if hasattr(val, 'value'):  # Enum
                return val.value
            if hasattr(val, 'name') and hasattr(val, '__class__') and 'Enum' in str(type(val)):
                return val.name
            return val
        
        # Extract fields from the Pydantic model
        fields = {}
        for field_name, field_info in config_class.model_fields.items():
            field_type = field_info.annotation
            default = make_serializable(field_info.default)
            description = field_info.description or ""
            
            is_locked = field_name in LOCKED_FIELD_NAMES
            
            # Convert type to string
            type_name = getattr(field_type, "__name__", str(field_type))
            if hasattr(field_type, "__origin__"):
                type_name = str(field_type)
            
            locked_value = LOCKED_FIELDS.get("env", {}).get(field_name, default)
            current_value = make_serializable(locked_value) if is_locked else default
            
            fields[field_name] = {
                "type": type_name,
                "default": default,
                "description": description,
                "locked": is_locked,
                "current_value": current_value,
            }
        
        return fields
        
    except Exception as e:
        logger.warning("Could not introspect environment config: %s", e)
        return {}


def _initialize_environments():
    """Initialize environment list on first use."""
    global _environments
    if not _environments:
        _environments = _scan_environments()


# ============================================================================
# Subprocess Management
# ============================================================================

async def _spawn_training_run(run_state: RunState, config_path: Path):
    """
    Spawn the three processes needed for training:
    1. run-api (Atropos API server)
    2. launch_training.py (Tinker trainer + inference server)
    3. environment.py serve (the Atropos environment)
    """
    run_id = run_state.run_id
    
    _ensure_logs_dir()

    # Log file paths
    api_log = LOGS_DIR / f"api_{run_id}.log"
    trainer_log = LOGS_DIR / f"trainer_{run_id}.log"
    env_log = LOGS_DIR / f"env_{run_id}.log"
    
    try:
        # Step 1: Start the Atropos API server (run-api)
        logger.info("[%s] Starting Atropos API server (run-api)...", run_id)
        
        # File must stay open while the subprocess runs; we store the handle
        # on run_state so _stop_training_run() can close it when done.
        api_log_file = open(api_log, "w")  # closed by _stop_training_run
        run_state.api_log_file = api_log_file
        run_state.api_process = subprocess.Popen(
            ["run-api"],
            stdout=api_log_file,
            stderr=subprocess.STDOUT,
            cwd=str(TINKER_ATROPOS_ROOT),
        )
        
        # Wait for API to start
        await asyncio.sleep(5)
        
        if run_state.api_process.poll() is not None:
            run_state.status = "failed"
            run_state.error_message = f"API server exited with code {run_state.api_process.returncode}. Check {api_log}"
            _stop_training_run(run_state)
            return
        
        logger.info("[%s] Atropos API server started", run_id)
        
        # Step 2: Start the Tinker trainer
        logger.info("[%s] Starting Tinker trainer: launch_training.py --config %s", run_id, config_path)
        
        trainer_log_file = open(trainer_log, "w")  # closed by _stop_training_run
        run_state.trainer_log_file = trainer_log_file
        run_state.trainer_process = subprocess.Popen(
            [sys.executable, "launch_training.py", "--config", str(config_path)],
            stdout=trainer_log_file,
            stderr=subprocess.STDOUT,
            cwd=str(TINKER_ATROPOS_ROOT),
            env={**os.environ, "TINKER_API_KEY": os.getenv("TINKER_API_KEY", "")},
        )
        
        # Wait for trainer to initialize (it starts FastAPI inference server on 8001)
        logger.info("[%s] Waiting 30 seconds for trainer to initialize...", run_id)
        await asyncio.sleep(30)
        
        if run_state.trainer_process.poll() is not None:
            run_state.status = "failed"
            run_state.error_message = f"Trainer exited with code {run_state.trainer_process.returncode}. Check {trainer_log}"
            _stop_training_run(run_state)
            return
        
        logger.info("[%s] Trainer started, inference server on port 8001", run_id)
        
        # Step 3: Start the environment
        logger.info("[%s] Waiting 90 more seconds before starting environment...", run_id)
        await asyncio.sleep(90)
        
        # Find the environment file
        env_info = None
        for env in _environments:
            if env.name == run_state.environment:
                env_info = env
                break
        
        if not env_info:
            run_state.status = "failed"
            run_state.error_message = f"Environment '{run_state.environment}' not found"
            _stop_training_run(run_state)
            return
        
        logger.info("[%s] Starting environment: %s serve", run_id, env_info.file_path)
        
        env_log_file = open(env_log, "w")  # closed by _stop_training_run
        run_state.env_log_file = env_log_file
        run_state.env_process = subprocess.Popen(
            [sys.executable, str(env_info.file_path), "serve", "--config", str(config_path)],
            stdout=env_log_file,
            stderr=subprocess.STDOUT,
            cwd=str(TINKER_ATROPOS_ROOT),
        )
        
        # Wait for environment to connect
        await asyncio.sleep(10)
        
        if run_state.env_process.poll() is not None:
            run_state.status = "failed"
            run_state.error_message = f"Environment exited with code {run_state.env_process.returncode}. Check {env_log}"
            _stop_training_run(run_state)
            return
        
        run_state.status = "running"
        run_state.start_time = time.time()
        logger.info("[%s] Training run started successfully!", run_id)
        
        # Start background monitoring
        asyncio.create_task(_monitor_training_run(run_state))
        
    except Exception as e:
        run_state.status = "failed"
        run_state.error_message = str(e)
        _stop_training_run(run_state)


async def _monitor_training_run(run_state: RunState):
    """Background task to monitor a training run."""
    while run_state.status == "running":
        await asyncio.sleep(30)  # Check every 30 seconds
        
        # Check if any process has died
        if run_state.env_process and run_state.env_process.poll() is not None:
            exit_code = run_state.env_process.returncode
            if exit_code == 0:
                run_state.status = "completed"
            else:
                run_state.status = "failed"
                run_state.error_message = f"Environment process exited with code {exit_code}"
            _stop_training_run(run_state)
            break
        
        if run_state.trainer_process and run_state.trainer_process.poll() is not None:
            exit_code = run_state.trainer_process.returncode
            if exit_code == 0:
                run_state.status = "completed"
            else:
                run_state.status = "failed"
                run_state.error_message = f"Trainer process exited with code {exit_code}"
            _stop_training_run(run_state)
            break
        
        if run_state.api_process and run_state.api_process.poll() is not None:
            run_state.status = "failed"
            run_state.error_message = f"API server exited unexpectedly"
            _stop_training_run(run_state)
            break


def _stop_training_run(run_state: RunState):
    """Stop all processes for a training run."""
    # Stop in reverse order: env -> trainer -> api
    if run_state.env_process and run_state.env_process.poll() is None:
        logger.info("[%s] Stopping environment process...", run_state.run_id)
        run_state.env_process.terminate()
        try:
            run_state.env_process.wait(timeout=10)
        except subprocess.TimeoutExpired:
            run_state.env_process.kill()
    
    if run_state.trainer_process and run_state.trainer_process.poll() is None:
        logger.info("[%s] Stopping trainer process...", run_state.run_id)
        run_state.trainer_process.terminate()
        try:
            run_state.trainer_process.wait(timeout=10)
        except subprocess.TimeoutExpired:
            run_state.trainer_process.kill()
    
    if run_state.api_process and run_state.api_process.poll() is None:
        logger.info("[%s] Stopping API server...", run_state.run_id)
        run_state.api_process.terminate()
        try:
            run_state.api_process.wait(timeout=10)
        except subprocess.TimeoutExpired:
            run_state.api_process.kill()
    
    if run_state.status == "running":
        run_state.status = "stopped"

    # Close log file handles that were opened for subprocess stdout.
    for attr in ("env_log_file", "trainer_log_file", "api_log_file"):
        fh = getattr(run_state, attr, None)
        if fh is not None:
            try:
                fh.close()
            except Exception:
                pass
            setattr(run_state, attr, None)


# ============================================================================
# Environment Discovery Tools
# ============================================================================

async def rl_list_environments() -> str:
    """
    List all available RL environments.
    
    Scans tinker-atropos/tinker_atropos/environments/ for Python files
    containing classes that inherit from BaseEnv.
    
    Returns information about each environment including:
    - name: Environment identifier
    - class_name: Python class name
    - file_path: Path to the environment file
    - description: Brief description if available
    
    TIP: To create or modify RL environments:
    1. Use terminal/file tools to inspect existing environments
    2. Study how they load datasets, define verifiers, and structure rewards
    3. Inspect HuggingFace datasets to understand data formats
    4. Copy an existing environment as a template
    
    Returns:
        JSON string with list of environments
    """
    _initialize_environments()
    
    response = {
        "environments": [
            {
                "name": env.name,
                "class_name": env.class_name,
                "file_path": env.file_path,
                "description": env.description,
            }
            for env in _environments
        ],
        "count": len(_environments),
        "tips": [
            "Use rl_select_environment(name) to select an environment",
            "Read the file_path with file tools to understand how each environment works",
            "Look for load_dataset(), score_answer(), get_next_item() methods",
        ]
    }
    
    return json.dumps(response, indent=2)


async def rl_select_environment(name: str) -> str:
    """
    Select an RL environment for training.
    
    This loads the environment's configuration fields into memory.
    After selecting, use rl_get_current_config() to see all configurable options
    and rl_edit_config() to modify specific fields.
    
    Args:
        name: Name of the environment to select (from rl_list_environments)
    
    Returns:
        JSON string with selection result, file path, and configurable field count
    
    TIP: Read the returned file_path to understand how the environment works.
    """
    global _current_env, _current_config, _env_config_cache
    
    _initialize_environments()
    
    env_info = None
    for env in _environments:
        if env.name == name:
            env_info = env
            break
    
    if not env_info:
        return json.dumps({
            "error": f"Environment '{name}' not found",
            "available": [e.name for e in _environments],
        }, indent=2)
    
    _current_env = name
    
    # Dynamically discover config fields
    config_fields = _get_env_config_fields(env_info.file_path)
    _env_config_cache[name] = config_fields
    
    # Initialize current config with defaults for non-locked fields
    _current_config = {}
    for field_name, field_info in config_fields.items():
        if not field_info.get("locked", False):
            _current_config[field_name] = field_info.get("default")
    
    # Auto-set wandb_name to "{env_name}-DATETIME" to avoid overlaps
    timestamp = datetime.now().strftime("%Y%m%d-%H%M%S")
    _current_config["wandb_name"] = f"{name}-{timestamp}"
    
    return json.dumps({
        "message": f"Selected environment: {name}",
        "environment": name,
        "file_path": env_info.file_path,
    }, indent=2)


# ============================================================================
# Configuration Tools
# ============================================================================

async def rl_get_current_config() -> str:
    """
    Get the current environment configuration.
    
    Returns all configurable fields for the selected environment.
    Each environment may have different configuration options.
    
    Fields are divided into:
    - configurable_fields: Can be changed with rl_edit_config()
    - locked_fields: Infrastructure settings that cannot be changed
    
    Returns:
        JSON string with configurable and locked fields
    """
    if not _current_env:
        return json.dumps({
            "error": "No environment selected. Use rl_select_environment(name) first.",
        }, indent=2)
    
    config_fields = _env_config_cache.get(_current_env, {})
    
    configurable = []
    locked = []
    
    for field_name, field_info in config_fields.items():
        field_data = {
            "name": field_name,
            "type": field_info.get("type", "unknown"),
            "default": field_info.get("default"),
            "description": field_info.get("description", ""),
            "current_value": _current_config.get(field_name, field_info.get("default")),
        }
        
        if field_info.get("locked", False):
            field_data["locked_value"] = LOCKED_FIELDS.get("env", {}).get(field_name)
            locked.append(field_data)
        else:
            configurable.append(field_data)
    
    return json.dumps({
        "environment": _current_env,
        "configurable_fields": configurable,
        "locked_fields": locked,
        "tip": "Use rl_edit_config(field, value) to change any configurable field.",
    }, indent=2)


async def rl_edit_config(field: str, value: Any) -> str:
    """
    Update a configuration field.
    
    Use rl_get_current_config() first to see available fields for the
    selected environment. Each environment has different options.
    
    Locked fields (infrastructure settings) cannot be changed.
    
    Args:
        field: Name of the field to update (from rl_get_current_config)
        value: New value for the field
    
    Returns:
        JSON string with updated config or error message
    """
    global _current_config
    
    if not _current_env:
        return json.dumps({
            "error": "No environment selected. Use rl_select_environment(name) first.",
        }, indent=2)
    
    config_fields = _env_config_cache.get(_current_env, {})
    
    if field not in config_fields:
        return json.dumps({
            "error": f"Unknown field '{field}'",
            "available_fields": list(config_fields.keys()),
        }, indent=2)
    
    field_info = config_fields[field]
    if field_info.get("locked", False):
        return json.dumps({
            "error": f"Field '{field}' is locked and cannot be changed",
            "locked_value": LOCKED_FIELDS.get("env", {}).get(field),
        }, indent=2)
    
    _current_config[field] = value
    
    return json.dumps({
        "message": f"Updated {field} = {value}",
        "field": field,
        "value": value,
        "config": _current_config,
    }, indent=2)


# ============================================================================
# Training Management Tools
# ============================================================================

async def rl_start_training() -> str:
    """
    Start a new RL training run with the current environment and config.
    
    Requires an environment to be selected first using rl_select_environment().
    Use rl_edit_config() to adjust configuration before starting.
    
    This spawns three processes:
    1. run-api (Atropos trajectory API)
    2. launch_training.py (Tinker trainer + inference server)
    3. environment.py serve (the selected environment)
    
    WARNING: Training runs take hours. Use rl_check_status() to monitor
    progress (recommended: check every 30 minutes at most).
    
    Returns:
        JSON string with run_id and initial status
    """
    global _active_runs
    
    if not _current_env:
        return json.dumps({
            "error": "No environment selected. Use rl_select_environment(name) first.",
        }, indent=2)
    
    # Check API keys
    if not os.getenv("TINKER_API_KEY"):
        return json.dumps({
            "error": "TINKER_API_KEY not set. Add it to ~/.hermes/.env",
        }, indent=2)
    
    # Find environment file
    env_info = None
    for env in _environments:
        if env.name == _current_env:
            env_info = env
            break
    
    if not env_info or not Path(env_info.file_path).exists():
        return json.dumps({
            "error": f"Environment file not found for '{_current_env}'",
        }, indent=2)
    
    # Generate run ID
    run_id = str(uuid.uuid4())[:8]
    
    # Create config YAML
    CONFIGS_DIR.mkdir(exist_ok=True)
    config_path = CONFIGS_DIR / f"run_{run_id}.yaml"
    
    # Start with locked config as base
    import copy
    run_config = copy.deepcopy(LOCKED_FIELDS)
    
    if "env" not in run_config:
        run_config["env"] = {}
    
    # Apply configurable fields
    for field_name, value in _current_config.items():
        if value is not None and value != "":
            run_config["env"][field_name] = value
    
    # Set WandB settings
    wandb_project = _current_config.get("wandb_project", "atropos-tinker")
    if "tinker" not in run_config:
        run_config["tinker"] = {}
    run_config["tinker"]["wandb_project"] = wandb_project
    run_config["tinker"]["wandb_run_name"] = f"{_current_env}-{run_id}"
    
    if "wandb_name" in _current_config and _current_config["wandb_name"]:
        run_config["env"]["wandb_name"] = _current_config["wandb_name"]
    
    with open(config_path, "w") as f:
        yaml.dump(run_config, f, default_flow_style=False)
    
    # Create run state
    run_state = RunState(
        run_id=run_id,
        environment=_current_env,
        config=_current_config.copy(),
        status="starting",
        wandb_project=wandb_project,
        wandb_run_name=f"{_current_env}-{run_id}",
    )
    
    _active_runs[run_id] = run_state
    
    # Start training in background
    asyncio.create_task(_spawn_training_run(run_state, config_path))
    
    return json.dumps({
        "run_id": run_id,
        "status": "starting",
        "environment": _current_env,
        "config": _current_config,
        "wandb_project": wandb_project,
        "wandb_run_name": f"{_current_env}-{run_id}",
        "config_path": str(config_path),
        "logs": {
            "api": str(LOGS_DIR / f"api_{run_id}.log"),
            "trainer": str(LOGS_DIR / f"trainer_{run_id}.log"),
            "env": str(LOGS_DIR / f"env_{run_id}.log"),
        },
        "message": "Training starting. Use rl_check_status(run_id) to monitor (recommended: every 30 minutes).",
    }, indent=2)


async def rl_check_status(run_id: str) -> str:
    """
    Get status and metrics for a training run.
    
    RATE LIMITED: For long-running training, this function enforces a
    minimum 30-minute interval between checks for the same run_id.
    
    Args:
        run_id: The run ID returned by rl_start_training()
    
    Returns:
        JSON string with run status and metrics
    """
    global _last_status_check
    
    # Check rate limiting
    now = time.time()
    if run_id in _last_status_check:
        elapsed = now - _last_status_check[run_id]
        if elapsed < MIN_STATUS_CHECK_INTERVAL:
            remaining = MIN_STATUS_CHECK_INTERVAL - elapsed
            return json.dumps({
                "rate_limited": True,
                "run_id": run_id,
                "message": f"Rate limited. Next check available in {remaining/60:.0f} minutes.",
                "next_check_in_seconds": remaining,
            }, indent=2)
    
    _last_status_check[run_id] = now
    
    if run_id not in _active_runs:
        return json.dumps({
            "error": f"Run '{run_id}' not found",
            "active_runs": list(_active_runs.keys()),
        }, indent=2)
    
    run_state = _active_runs[run_id]
    
    # Check process status
    processes = {
        "api": run_state.api_process.poll() if run_state.api_process else None,
        "trainer": run_state.trainer_process.poll() if run_state.trainer_process else None,
        "env": run_state.env_process.poll() if run_state.env_process else None,
    }
    
    running_time = time.time() - run_state.start_time if run_state.start_time else 0
    
    result = {
        "run_id": run_id,
        "status": run_state.status,
        "environment": run_state.environment,
        "running_time_minutes": running_time / 60,
        "processes": {
            name: "running" if code is None else f"exited ({code})"
            for name, code in processes.items()
        },
        "wandb_project": run_state.wandb_project,
        "wandb_run_name": run_state.wandb_run_name,
        "logs": {
            "api": str(LOGS_DIR / f"api_{run_id}.log"),
            "trainer": str(LOGS_DIR / f"trainer_{run_id}.log"),
            "env": str(LOGS_DIR / f"env_{run_id}.log"),
        },
    }
    
    if run_state.error_message:
        result["error"] = run_state.error_message
    
    # Try to get WandB metrics if available
    try:
        import wandb
        api = wandb.Api()
        runs = api.runs(
            f"{os.getenv('WANDB_ENTITY', 'nousresearch')}/{run_state.wandb_project}",
            filters={"display_name": run_state.wandb_run_name}
        )
        if runs:
            wandb_run = runs[0]
            result["wandb_url"] = wandb_run.url
            result["metrics"] = {
                "step": wandb_run.summary.get("_step", 0),
                "reward_mean": wandb_run.summary.get("train/reward_mean"),
                "percent_correct": wandb_run.summary.get("train/percent_correct"),
                "eval_percent_correct": wandb_run.summary.get("eval/percent_correct"),
            }
    except Exception as e:
        result["wandb_error"] = str(e)
    
    return json.dumps(result, indent=2)


async def rl_stop_training(run_id: str) -> str:
    """
    Stop a running training job.
    
    Args:
        run_id: The run ID to stop
    
    Returns:
        JSON string with stop confirmation
    """
    if run_id not in _active_runs:
        return json.dumps({
            "error": f"Run '{run_id}' not found",
            "active_runs": list(_active_runs.keys()),
        }, indent=2)
    
    run_state = _active_runs[run_id]
    
    if run_state.status not in ("running", "starting"):
        return json.dumps({
            "message": f"Run '{run_id}' is not running (status: {run_state.status})",
        }, indent=2)
    
    _stop_training_run(run_state)
    
    return json.dumps({
        "message": f"Stopped training run '{run_id}'",
        "run_id": run_id,
        "status": run_state.status,
    }, indent=2)


async def rl_get_results(run_id: str) -> str:
    """
    Get final results and metrics for a training run.
    
    Args:
        run_id: The run ID to get results for
    
    Returns:
        JSON string with final results
    """
    if run_id not in _active_runs:
        return json.dumps({
            "error": f"Run '{run_id}' not found",
        }, indent=2)
    
    run_state = _active_runs[run_id]
    
    result = {
        "run_id": run_id,
        "status": run_state.status,
        "environment": run_state.environment,
        "wandb_project": run_state.wandb_project,
        "wandb_run_name": run_state.wandb_run_name,
    }
    
    # Get WandB metrics
    try:
        import wandb
        api = wandb.Api()
        runs = api.runs(
            f"{os.getenv('WANDB_ENTITY', 'nousresearch')}/{run_state.wandb_project}",
            filters={"display_name": run_state.wandb_run_name}
        )
        if runs:
            wandb_run = runs[0]
            result["wandb_url"] = wandb_run.url
            result["final_metrics"] = dict(wandb_run.summary)
            result["history"] = [dict(row) for row in wandb_run.history(samples=10)]
    except Exception as e:
        result["wandb_error"] = str(e)
    
    return json.dumps(result, indent=2)


async def rl_list_runs() -> str:
    """
    List all training runs (active and completed).
    
    Returns:
        JSON string with list of runs and their status
    """
    runs = []
    for run_id, run_state in _active_runs.items():
        runs.append({
            "run_id": run_id,
            "environment": run_state.environment,
            "status": run_state.status,
            "wandb_run_name": run_state.wandb_run_name,
        })
    
    return json.dumps({
        "runs": runs,
        "count": len(runs),
    }, indent=2)


# ============================================================================
# Inference Testing (via Atropos `process` mode with OpenRouter)
# ============================================================================

# Test models at different scales for robustness testing
# These are cheap, capable models on OpenRouter for testing parsing/scoring
TEST_MODELS = [
    {"id": "qwen/qwen3-8b", "name": "Qwen3 8B", "scale": "small"},
    {"id": "z-ai/glm-4.7-flash", "name": "GLM-4.7 Flash", "scale": "medium"},
    {"id": "minimax/minimax-m2.7", "name": "MiniMax M2.7", "scale": "large"},
]

# Default test parameters - quick but representative
DEFAULT_NUM_STEPS = 3       # Number of steps (items) to test
DEFAULT_GROUP_SIZE = 16     # Completions per item (like training)


async def rl_test_inference(
    num_steps: int = DEFAULT_NUM_STEPS,
    group_size: int = DEFAULT_GROUP_SIZE,
    models: Optional[List[str]] = None,
) -> str:
    """
    Quick inference test for any environment using Atropos's `process` mode.
    
    Runs a few steps of inference + scoring to validate:
    - Environment loads correctly
    - Prompt construction works
    - Inference parsing is robust (tested with multiple model scales)
    - Verifier/scoring logic works
    
    Default: 3 steps × 16 completions = 48 total rollouts per model.
    Tests 3 models = 144 total rollouts. Quick sanity check.
    
    Test models (varying intelligence levels for robustness):
    - qwen/qwen3-8b (small)
    - zhipu-ai/glm-4-flash (medium)
    - minimax/minimax-m1 (large)
    
    Args:
        num_steps: Steps to run (default: 3, max recommended for testing)
        group_size: Completions per step (default: 16, like training)
        models: Optional model IDs to test. If None, uses all 3 test models.
    
    Returns:
        JSON with results per model: steps_tested, accuracy, scores
    """
    if not _current_env:
        return json.dumps({
            "error": "No environment selected. Use rl_select_environment(name) first.",
        }, indent=2)
    
    api_key = os.getenv("OPENROUTER_API_KEY")
    if not api_key:
        return json.dumps({
            "error": "OPENROUTER_API_KEY not set. Required for inference testing.",
        }, indent=2)
    
    # Find environment info
    env_info = None
    for env in _environments:
        if env.name == _current_env:
            env_info = env
            break
    
    if not env_info:
        return json.dumps({
            "error": f"Environment '{_current_env}' not found",
        }, indent=2)
    
    # Determine which models to test
    if models:
        test_models = [m for m in TEST_MODELS if m["id"] in models]
        if not test_models:
            test_models = [{"id": m, "name": m, "scale": "custom"} for m in models]
    else:
        test_models = TEST_MODELS
    
    # Calculate total rollouts for logging
    total_rollouts_per_model = num_steps * group_size
    total_rollouts = total_rollouts_per_model * len(test_models)
    
    results = {
        "environment": _current_env,
        "environment_file": env_info.file_path,
        "test_config": {
            "num_steps": num_steps,
            "group_size": group_size,
            "rollouts_per_model": total_rollouts_per_model,
            "total_rollouts": total_rollouts,
        },
        "models_tested": [],
    }
    
    # Create output directory for test results
    _ensure_logs_dir()
    test_output_dir = LOGS_DIR / "inference_tests"
    test_output_dir.mkdir(exist_ok=True)
    
    for model_info in test_models:
        model_id = model_info["id"]
        model_safe_name = model_id.replace("/", "_")
        
        print(f"\n{'='*60}")
        print(f"Testing with {model_info['name']} ({model_id})")
        print(f"{'='*60}")
        
        # Output file for this test run
        output_file = test_output_dir / f"test_{_current_env}_{model_safe_name}.jsonl"
        
        # Generate unique run ID for wandb
        test_run_id = str(uuid.uuid4())[:8]
        wandb_run_name = f"test_inference_RSIAgent_{_current_env}_{test_run_id}"
        
        # Build the process command using Atropos's built-in CLI
        # This runs the environment's actual code with OpenRouter as the inference backend
        # We pass our locked settings + test-specific overrides via CLI args
        cmd = [
            sys.executable, env_info.file_path, "process",
            # Test-specific overrides
            "--env.total_steps", str(num_steps),
            "--env.group_size", str(group_size),
            "--env.use_wandb", "true",  # Enable wandb for test tracking
            "--env.wandb_name", wandb_run_name,
            "--env.data_path_to_save_groups", str(output_file),
            # Use locked settings from our config
            "--env.tokenizer_name", LOCKED_FIELDS["env"]["tokenizer_name"],
            "--env.max_token_length", str(LOCKED_FIELDS["env"]["max_token_length"]),
            "--env.max_num_workers", str(LOCKED_FIELDS["env"]["max_num_workers"]),
            "--env.max_batches_offpolicy", str(LOCKED_FIELDS["env"]["max_batches_offpolicy"]),
            # OpenRouter config for inference testing
            # IMPORTANT: Use server_type=openai for OpenRouter (not sglang)
            # sglang is only for actual training with Tinker's inference server
            "--openai.base_url", "https://openrouter.ai/api/v1",
            "--openai.api_key", api_key,
            "--openai.model_name", model_id,
            "--openai.server_type", "openai",  # OpenRouter is OpenAI-compatible
            "--openai.health_check", "false",  # OpenRouter doesn't have health endpoint
        ]
        
        # Debug: Print the full command
        cmd_str = " ".join(str(c) for c in cmd)
        # Hide API key in printed output
        cmd_display = cmd_str.replace(api_key, "***API_KEY***")
        print(f"Command: {cmd_display}")
        print(f"Working dir: {TINKER_ATROPOS_ROOT}")
        print(f"WandB run: {wandb_run_name}")
        print(f"  {num_steps} steps × {group_size} completions = {total_rollouts_per_model} rollouts")
        
        model_results = {
            "model": model_id,
            "name": model_info["name"],
            "scale": model_info["scale"],
            "wandb_run": wandb_run_name,
            "output_file": str(output_file),
            "steps": [],
            "steps_tested": 0,
            "total_completions": 0,
            "correct_completions": 0,
        }
        
        try:
            # Run the process command with real-time output streaming
            process = await asyncio.create_subprocess_exec(
                *cmd,
                stdout=asyncio.subprocess.PIPE,
                stderr=asyncio.subprocess.PIPE,
                cwd=str(TINKER_ATROPOS_ROOT),
            )
            
            # Stream output in real-time while collecting for logs
            stdout_lines = []
            stderr_lines = []
            log_file = test_output_dir / f"test_{_current_env}_{model_safe_name}.log"
            
            async def read_stream(stream, lines_list, prefix=""):
                """Read stream line by line and print in real-time."""
                while True:
                    line = await stream.readline()
                    if not line:
                        break
                    decoded = line.decode().rstrip()
                    lines_list.append(decoded)
                    # Print progress-related lines in real-time
                    if any(kw in decoded.lower() for kw in ['processing', 'group', 'step', 'progress', '%', 'completed']):
                        print(f"  {prefix}{decoded}")
            
            # Read both streams concurrently with timeout
            try:
                await asyncio.wait_for(
                    asyncio.gather(
                        read_stream(process.stdout, stdout_lines, "📊 "),
                        read_stream(process.stderr, stderr_lines, "⚠️ "),
                    ),
                    timeout=600,  # 10 minute timeout per model
                )
            except asyncio.TimeoutError:
                process.kill()
                raise
            
            await process.wait()
            
            # Combine output for logging
            stdout_text = "\n".join(stdout_lines)
            stderr_text = "\n".join(stderr_lines)
            
            # Write logs to files for inspection outside CLI
            with open(log_file, "w") as f:
                f.write(f"Command: {cmd_display}\n")
                f.write(f"Working dir: {TINKER_ATROPOS_ROOT}\n")
                f.write(f"Return code: {process.returncode}\n")
                f.write(f"\n{'='*60}\n")
                f.write(f"STDOUT:\n{'='*60}\n")
                f.write(stdout_text or "(empty)\n")
                f.write(f"\n{'='*60}\n")
                f.write(f"STDERR:\n{'='*60}\n")
                f.write(stderr_text or "(empty)\n")
            
            print(f"  Log file: {log_file}")
            
            if process.returncode != 0:
                model_results["error"] = f"Process exited with code {process.returncode}"
                model_results["stderr"] = stderr_text[-1000:]
                model_results["stdout"] = stdout_text[-1000:]
                model_results["log_file"] = str(log_file)
                print(f"\n  ❌ Error: {model_results['error']}")
                # Print last few lines of stderr for debugging
                if stderr_lines:
                    print(f"  Last errors:")
                    for line in stderr_lines[-5:]:
                        print(f"    {line}")
            else:
                print(f"\n  ✅ Process completed successfully")
                print(f"  Output file: {output_file}")
                print(f"  File exists: {output_file.exists()}")
                
                # Parse the output JSONL file
                if output_file.exists():
                    # Read JSONL file (one JSON object per line = one step)
                    with open(output_file, "r") as f:
                        for line in f:
                            line = line.strip()
                            if not line:
                                continue
                            try:
                                item = json.loads(line)
                                scores = item.get("scores", [])
                                model_results["steps_tested"] += 1
                                model_results["total_completions"] += len(scores)
                                correct = sum(1 for s in scores if s > 0)
                                model_results["correct_completions"] += correct
                                
                                model_results["steps"].append({
                                    "step": model_results["steps_tested"],
                                    "completions": len(scores),
                                    "correct": correct,
                                    "scores": scores,
                                })
                            except json.JSONDecodeError:
                                continue
                    
                    print(f"  Completed {model_results['steps_tested']} steps")
                else:
                    model_results["error"] = f"Output file not created: {output_file}"
                    
        except asyncio.TimeoutError:
            model_results["error"] = "Process timed out after 10 minutes"
            print(f"  Timeout!")
        except Exception as e:
            model_results["error"] = str(e)
            print(f"  Error: {e}")
        
        # Calculate stats
        if model_results["total_completions"] > 0:
            model_results["accuracy"] = round(
                model_results["correct_completions"] / model_results["total_completions"], 3
            )
        else:
            model_results["accuracy"] = 0
            
        if model_results["steps_tested"] > 0:
            steps_with_correct = sum(1 for s in model_results["steps"] if s.get("correct", 0) > 0)
            model_results["steps_with_correct"] = steps_with_correct
            model_results["step_success_rate"] = round(
                steps_with_correct / model_results["steps_tested"], 3
            )
        else:
            model_results["steps_with_correct"] = 0
            model_results["step_success_rate"] = 0
        
        print(f"  Results: {model_results['correct_completions']}/{model_results['total_completions']} correct")
        print(f"  Accuracy: {model_results['accuracy']:.1%}")
        
        results["models_tested"].append(model_results)
    
    # Overall summary
    working_models = [m for m in results["models_tested"] if m.get("steps_tested", 0) > 0]
    
    results["summary"] = {
        "steps_requested": num_steps,
        "models_tested": len(test_models),
        "models_succeeded": len(working_models),
        "best_model": max(working_models, key=lambda x: x.get("accuracy", 0))["model"] if working_models else None,
        "avg_accuracy": round(
            sum(m.get("accuracy", 0) for m in working_models) / len(working_models), 3
        ) if working_models else 0,
        "environment_working": len(working_models) > 0,
        "output_directory": str(test_output_dir),
    }
    
    return json.dumps(results, indent=2)


# ============================================================================
# Requirements Check
# ============================================================================

def check_rl_python_version() -> bool:
    """
    Check if Python version meets the minimum for RL tools.
    
    tinker-atropos depends on the 'tinker' package which requires Python >= 3.11.
    """
    return sys.version_info >= (3, 11)


def check_rl_api_keys() -> bool:
    """
    Check if required API keys and Python version are available.
    
    RL training requires:
    - Python >= 3.11 (tinker package requirement)
    - TINKER_API_KEY for the Tinker training API
    - WANDB_API_KEY for Weights & Biases metrics
    """
    if not check_rl_python_version():
        return False
    tinker_key = os.getenv("TINKER_API_KEY")
    wandb_key = os.getenv("WANDB_API_KEY")
    return bool(tinker_key) and bool(wandb_key)


def get_missing_keys() -> List[str]:
    """
    Get list of missing requirements for RL tools (API keys and Python version).
    """
    missing = []
    if not check_rl_python_version():
        missing.append(f"Python >= 3.11 (current: {sys.version_info.major}.{sys.version_info.minor})")
    if not os.getenv("TINKER_API_KEY"):
        missing.append("TINKER_API_KEY")
    if not os.getenv("WANDB_API_KEY"):
        missing.append("WANDB_API_KEY")
    return missing


# ---------------------------------------------------------------------------
# Schemas + Registry
# ---------------------------------------------------------------------------
from tools.registry import registry

RL_LIST_ENVIRONMENTS_SCHEMA = {"name": "rl_list_environments", "description": "List all available RL environments. Returns environment names, paths, and descriptions. TIP: Read the file_path with file tools to understand how each environment works (verifiers, data loading, rewards).", "parameters": {"type": "object", "properties": {}, "required": []}}
RL_SELECT_ENVIRONMENT_SCHEMA = {"name": "rl_select_environment", "description": "Select an RL environment for training. Loads the environment's default configuration. After selecting, use rl_get_current_config() to see settings and rl_edit_config() to modify them.", "parameters": {"type": "object", "properties": {"name": {"type": "string", "description": "Name of the environment to select (from rl_list_environments)"}}, "required": ["name"]}}
RL_GET_CURRENT_CONFIG_SCHEMA = {"name": "rl_get_current_config", "description": "Get the current environment configuration. Returns only fields that can be modified: group_size, max_token_length, total_steps, steps_per_eval, use_wandb, wandb_name, max_num_workers.", "parameters": {"type": "object", "properties": {}, "required": []}}
RL_EDIT_CONFIG_SCHEMA = {"name": "rl_edit_config", "description": "Update a configuration field. Use rl_get_current_config() first to see all available fields for the selected environment. Each environment has different configurable options. Infrastructure settings (tokenizer, URLs, lora_rank, learning_rate) are locked.", "parameters": {"type": "object", "properties": {"field": {"type": "string", "description": "Name of the field to update (get available fields from rl_get_current_config)"}, "value": {"description": "New value for the field"}}, "required": ["field", "value"]}}
RL_START_TRAINING_SCHEMA = {"name": "rl_start_training", "description": "Start a new RL training run with the current environment and config. Most training parameters (lora_rank, learning_rate, etc.) are fixed. Use rl_edit_config() to set group_size, batch_size, wandb_project before starting. WARNING: Training takes hours.", "parameters": {"type": "object", "properties": {}, "required": []}}
RL_CHECK_STATUS_SCHEMA = {"name": "rl_check_status", "description": "Get status and metrics for a training run. RATE LIMITED: enforces 30-minute minimum between checks for the same run. Returns WandB metrics: step, state, reward_mean, loss, percent_correct.", "parameters": {"type": "object", "properties": {"run_id": {"type": "string", "description": "The run ID from rl_start_training()"}}, "required": ["run_id"]}}
RL_STOP_TRAINING_SCHEMA = {"name": "rl_stop_training", "description": "Stop a running training job. Use if metrics look bad, training is stagnant, or you want to try different settings.", "parameters": {"type": "object", "properties": {"run_id": {"type": "string", "description": "The run ID to stop"}}, "required": ["run_id"]}}
RL_GET_RESULTS_SCHEMA = {"name": "rl_get_results", "description": "Get final results and metrics for a completed training run. Returns final metrics and path to trained weights.", "parameters": {"type": "object", "properties": {"run_id": {"type": "string", "description": "The run ID to get results for"}}, "required": ["run_id"]}}
RL_LIST_RUNS_SCHEMA = {"name": "rl_list_runs", "description": "List all training runs (active and completed) with their status.", "parameters": {"type": "object", "properties": {}, "required": []}}
RL_TEST_INFERENCE_SCHEMA = {"name": "rl_test_inference", "description": "Quick inference test for any environment. Runs a few steps of inference + scoring using OpenRouter. Default: 3 steps x 16 completions = 48 rollouts per model, testing 3 models = 144 total. Tests environment loading, prompt construction, inference parsing, and verifier logic. Use BEFORE training to catch issues.", "parameters": {"type": "object", "properties": {"num_steps": {"type": "integer", "description": "Number of steps to run (default: 3, recommended max for testing)", "default": 3}, "group_size": {"type": "integer", "description": "Completions per step (default: 16, like training)", "default": 16}, "models": {"type": "array", "items": {"type": "string"}, "description": "Optional list of OpenRouter model IDs. Default: qwen/qwen3-8b, z-ai/glm-4.7-flash, minimax/minimax-m2.7"}}, "required": []}}

_rl_env = ["TINKER_API_KEY", "WANDB_API_KEY"]

registry.register(name="rl_list_environments", emoji="🧪", toolset="rl", schema=RL_LIST_ENVIRONMENTS_SCHEMA,
    handler=lambda args, **kw: rl_list_environments(), check_fn=check_rl_api_keys, requires_env=_rl_env, is_async=True)
registry.register(name="rl_select_environment", emoji="🧪", toolset="rl", schema=RL_SELECT_ENVIRONMENT_SCHEMA,
    handler=lambda args, **kw: rl_select_environment(name=args.get("name", "")), check_fn=check_rl_api_keys, requires_env=_rl_env, is_async=True)
registry.register(name="rl_get_current_config", emoji="🧪", toolset="rl", schema=RL_GET_CURRENT_CONFIG_SCHEMA,
    handler=lambda args, **kw: rl_get_current_config(), check_fn=check_rl_api_keys, requires_env=_rl_env, is_async=True)
registry.register(name="rl_edit_config", emoji="🧪", toolset="rl", schema=RL_EDIT_CONFIG_SCHEMA,
    handler=lambda args, **kw: rl_edit_config(field=args.get("field", ""), value=args.get("value")), check_fn=check_rl_api_keys, requires_env=_rl_env, is_async=True)
registry.register(name="rl_start_training", emoji="🧪", toolset="rl", schema=RL_START_TRAINING_SCHEMA,
    handler=lambda args, **kw: rl_start_training(), check_fn=check_rl_api_keys, requires_env=_rl_env, is_async=True)
registry.register(name="rl_check_status", emoji="🧪", toolset="rl", schema=RL_CHECK_STATUS_SCHEMA,
    handler=lambda args, **kw: rl_check_status(run_id=args.get("run_id", "")), check_fn=check_rl_api_keys, requires_env=_rl_env, is_async=True)
registry.register(name="rl_stop_training", emoji="🧪", toolset="rl", schema=RL_STOP_TRAINING_SCHEMA,
    handler=lambda args, **kw: rl_stop_training(run_id=args.get("run_id", "")), check_fn=check_rl_api_keys, requires_env=_rl_env, is_async=True)
registry.register(name="rl_get_results", emoji="🧪", toolset="rl", schema=RL_GET_RESULTS_SCHEMA,
    handler=lambda args, **kw: rl_get_results(run_id=args.get("run_id", "")), check_fn=check_rl_api_keys, requires_env=_rl_env, is_async=True)
registry.register(name="rl_list_runs", emoji="🧪", toolset="rl", schema=RL_LIST_RUNS_SCHEMA,
    handler=lambda args, **kw: rl_list_runs(), check_fn=check_rl_api_keys, requires_env=_rl_env, is_async=True)
registry.register(name="rl_test_inference", emoji="🧪", toolset="rl", schema=RL_TEST_INFERENCE_SCHEMA,
    handler=lambda args, **kw: rl_test_inference(num_steps=args.get("num_steps", 3), group_size=args.get("group_size", 16), models=args.get("models")),
    check_fn=check_rl_api_keys, requires_env=_rl_env, is_async=True)