mcpbr 0.4.16__py3-none-any.whl → 0.5.0__py3-none-any.whl

mcpbr/dry_run.py

@@ -0,0 +1,532 @@
+"""Dry-run mode for previewing evaluations without executing them.
+
+This module provides functionality to preview what an evaluation would do,
+validate configurations, estimate costs and time, and check infrastructure
+readiness -- all without making actual API calls or running tasks.
+
+Useful for debugging configuration issues and estimating costs before
+committing to a full evaluation run.
+"""
+
+import logging
+import os
+import shutil
+from dataclasses import dataclass, field
+
+import docker
+from rich.console import Console
+from rich.panel import Panel
+from rich.table import Table
+
+from .benchmarks import create_benchmark
+from .config import HarnessConfig
+from .config_validator import ConfigValidator, ValidationResult
+from .pricing import format_cost, get_model_pricing
+
+logger = logging.getLogger(__name__)
+
+# Historical estimates for average tokens per task by benchmark type.
+# These are rough estimates based on typical evaluation runs and are used
+# when no better data is available. Values represent (input_tokens, output_tokens).
+_ESTIMATED_TOKENS_PER_TASK: dict[str, tuple[int, int]] = {
+    "swe-bench-lite": (50_000, 10_000),
+    "swe-bench-verified": (50_000, 10_000),
+    "swe-bench-full": (50_000, 10_000),
+    "humaneval": (5_000, 2_000),
+    "mbpp": (5_000, 2_000),
+    "gsm8k": (3_000, 1_000),
+    "math": (5_000, 2_000),
+    "truthfulqa": (2_000, 500),
+    "bigbench-hard": (5_000, 2_000),
+    "hellaswag": (2_000, 500),
+    "arc": (3_000, 1_000),
+    "apps": (10_000, 5_000),
+    "codecontests": (15_000, 8_000),
+    "bigcodebench": (10_000, 5_000),
+    "leetcode": (10_000, 5_000),
+    "codereval": (20_000, 8_000),
+    "repoqa": (30_000, 5_000),
+    "toolbench": (10_000, 5_000),
+    "aider-polyglot": (30_000, 10_000),
+    "terminalbench": (20_000, 8_000),
+    "gaia": (15_000, 5_000),
+    "agentbench": (20_000, 8_000),
+    "webarena": (20_000, 8_000),
+    "mlagentbench": (25_000, 10_000),
+    "intercode": (15_000, 5_000),
+    "cybergym": (30_000, 10_000),
+    "mcptoolbench": (10_000, 5_000),
+    "custom": (10_000, 5_000),
+    "mmmu": (5_000, 2_000),
+    "longbench": (30_000, 5_000),
+    "adversarial": (10_000, 5_000),
+}
+
+# Historical average minutes per task (wall-clock time).
+# Accounts for Docker setup, agent execution, and evaluation.
+_ESTIMATED_MINUTES_PER_TASK: dict[str, float] = {
+    "swe-bench-lite": 8.0,
+    "swe-bench-verified": 8.0,
+    "swe-bench-full": 8.0,
+    "humaneval": 2.0,
+    "mbpp": 2.0,
+    "gsm8k": 1.0,
+    "math": 2.0,
+    "truthfulqa": 0.5,
+    "bigbench-hard": 1.5,
+    "hellaswag": 0.5,
+    "arc": 1.0,
+    "apps": 4.0,
+    "codecontests": 6.0,
+    "bigcodebench": 4.0,
+    "leetcode": 4.0,
+    "codereval": 6.0,
+    "repoqa": 5.0,
+    "toolbench": 3.0,
+    "aider-polyglot": 7.0,
+    "terminalbench": 5.0,
+    "gaia": 5.0,
+    "agentbench": 6.0,
+    "webarena": 6.0,
+    "mlagentbench": 7.0,
+    "intercode": 4.0,
+    "cybergym": 8.0,
+    "mcptoolbench": 3.0,
+    "custom": 3.0,
+    "mmmu": 2.0,
+    "longbench": 5.0,
+    "adversarial": 3.0,
+}
+
+
+@dataclass
+class DryRunResult:
+    """Result of a dry-run evaluation preview.
+
+    Contains all information about what an evaluation would do, including
+    task details, cost estimates, configuration validation, and infrastructure
+    readiness checks.
+
+    Attributes:
+        benchmark_name: Name of the benchmark to run.
+        total_tasks: Total number of tasks that would be executed.
+        task_ids: List of task IDs that would be executed.
+        estimated_cost_usd: Estimated total cost in USD based on model pricing.
+        estimated_time_minutes: Estimated total wall-clock time in minutes.
+        config_valid: Whether the configuration passed validation.
+        config_errors: List of configuration validation error messages.
+        docker_available: Whether Docker is available and running.
+        mcp_servers_reachable: Mapping of MCP server names to reachability status.
+        warnings: List of warning messages about the evaluation.
+    """
+
+    benchmark_name: str
+    total_tasks: int
+    task_ids: list[str]
+    estimated_cost_usd: float | None
+    estimated_time_minutes: float | None
+    config_valid: bool
+    config_errors: list[str]
+    docker_available: bool
+    mcp_servers_reachable: dict[str, bool]
+    warnings: list[str] = field(default_factory=list)
+
+
+def _check_docker_available() -> bool:
+    """Check whether Docker is available and running.
+
+    Returns:
+        True if Docker daemon is reachable, False otherwise.
+    """
+    try:
+        client = docker.from_env()
+        client.ping()
+        return True
+    except Exception:
+        return False
+
+
+def _check_mcp_server_reachable(command: str) -> bool:
+    """Check whether an MCP server command is available in PATH.
+
+    This checks that the command is installed and executable, which is a
+    necessary prerequisite for the MCP server to be reachable at runtime.
+
+    Args:
+        command: The MCP server command to check (e.g., 'npx', 'uvx').
+
+    Returns:
+        True if the command is found in PATH, False otherwise.
+    """
+    return shutil.which(command) is not None
+
+
+def _validate_config_from_object(config: HarnessConfig) -> ValidationResult:
+    """Validate a HarnessConfig object using the config validator.
+
+    Since ConfigValidator works on files, we perform structural validation
+    directly on the config object instead.
+
+    Args:
+        config: The harness configuration to validate.
+
+    Returns:
+        ValidationResult with errors and warnings.
+    """
+    validator = ConfigValidator()
+
+    # Validate API key (Anthropic provider)
+    if config.provider == "anthropic":
+        validator._validate_api_key()
+
+    # If we have no errors from the validator and the config object was
+    # successfully created (Pydantic validated), it is valid.
+    return ValidationResult(
+        valid=not validator.has_errors,
+        errors=validator.errors,
+        warnings=validator.warnings,
+    )
+
+
+def _estimate_cost(
+    model_id: str,
+    benchmark_name: str,
+    num_tasks: int,
+) -> float | None:
+    """Estimate the cost of running an evaluation based on model pricing.
+
+    Uses historical token usage estimates per benchmark type and the model's
+    pricing to compute an approximate total cost.
+
+    Args:
+        model_id: The model identifier for pricing lookup.
+        benchmark_name: The benchmark name for token estimation.
+        num_tasks: The number of tasks to estimate for.
+
+    Returns:
+        Estimated cost in USD, or None if pricing is unavailable.
+    """
+    pricing = get_model_pricing(model_id)
+    if pricing is None:
+        return None
+
+    input_tokens, output_tokens = _ESTIMATED_TOKENS_PER_TASK.get(benchmark_name, (10_000, 5_000))
+
+    # Calculate per-task cost
+    input_cost = (input_tokens / 1_000_000) * pricing.input_price_per_mtok
+    output_cost = (output_tokens / 1_000_000) * pricing.output_price_per_mtok
+    per_task_cost = input_cost + output_cost
+
+    return per_task_cost * num_tasks
+
+
+def _estimate_time(
+    benchmark_name: str,
+    num_tasks: int,
+    max_concurrent: int,
+    timeout_seconds: int,
+) -> float:
+    """Estimate the wall-clock time for running an evaluation.
+
+    Uses historical per-task time estimates and accounts for concurrency.
+    The estimate is capped by the configured timeout per task.
+
+    Args:
+        benchmark_name: The benchmark name for time estimation.
+        num_tasks: The number of tasks to run.
+        max_concurrent: Maximum concurrent tasks.
+        timeout_seconds: Configured timeout per task in seconds.
+
+    Returns:
+        Estimated wall-clock time in minutes.
+    """
+    per_task_minutes = _ESTIMATED_MINUTES_PER_TASK.get(benchmark_name, 3.0)
+
+    # Cap per-task time at the configured timeout
+    timeout_minutes = timeout_seconds / 60.0
+    per_task_minutes = min(per_task_minutes, timeout_minutes)
+
+    # Account for concurrency: tasks run in batches of max_concurrent
+    effective_concurrency = max(1, min(max_concurrent, num_tasks)) if num_tasks > 0 else 1
+    total_minutes = (num_tasks / effective_concurrency) * per_task_minutes
+
+    return total_minutes
+
+
+async def dry_run(config: HarnessConfig, verbosity: int = 0) -> DryRunResult:
+    """Preview what an evaluation would do without executing it.
+
+    Loads the benchmark tasks, validates the configuration, checks Docker
+    availability, checks MCP server reachability, and estimates cost and
+    time. Does NOT make any API calls or run any tasks.
+
+    Args:
+        config: The harness configuration to preview.
+        verbosity: Verbosity level (0=minimal, 1=summary, 2=detailed).
+
+    Returns:
+        DryRunResult containing all preview information.
+    """
+    warnings: list[str] = []
+    config_errors: list[str] = []
+    task_ids: list[str] = []
+    total_tasks = 0
+
+    # 1. Validate configuration
+    validation_result = _validate_config_from_object(config)
+    config_valid = validation_result.valid
+    for error in validation_result.errors:
+        config_errors.append(f"{error.field}: {error.error}")
+    for warning in validation_result.warnings:
+        warnings.append(f"Config warning ({warning.field}): {warning.error}")
+
+    # 2. Load benchmark tasks
+    benchmark_name = config.benchmark
+    try:
+        benchmark_kwargs = {}
+        if config.benchmark == "cybergym":
+            benchmark_kwargs["level"] = config.cybergym_level
+
+        benchmark = create_benchmark(config.benchmark, **benchmark_kwargs)
+        tasks = benchmark.load_tasks(
+            sample_size=config.sample_size,
+            filter_difficulty=config.filter_difficulty,
+            filter_category=config.filter_category,
+            filter_tags=config.filter_tags,
+        )
+        total_tasks = len(tasks)
+        task_ids = [t.get("instance_id", f"task_{i}") for i, t in enumerate(tasks)]
+    except Exception as e:
+        warnings.append(f"Failed to load benchmark tasks: {e}")
+        total_tasks = config.sample_size if config.sample_size else 0
+
+    # 3. Check Docker availability
+    docker_available = _check_docker_available()
+    if not docker_available:
+        warnings.append(
+            "Docker is not available. Evaluation requires Docker to create "
+            "isolated task environments."
+        )
+
+    # 4. Check MCP server reachability
+    mcp_servers_reachable: dict[str, bool] = {}
+    if config.comparison_mode:
+        if config.mcp_server_a and config.mcp_server_a.command:
+            name_a = config.mcp_server_a.name or "mcp_server_a"
+            reachable = _check_mcp_server_reachable(config.mcp_server_a.command)
+            mcp_servers_reachable[name_a] = reachable
+            if not reachable:
+                warnings.append(
+                    f"MCP server A command '{config.mcp_server_a.command}' not found in PATH."
+                )
+        if config.mcp_server_b and config.mcp_server_b.command:
+            name_b = config.mcp_server_b.name or "mcp_server_b"
+            reachable = _check_mcp_server_reachable(config.mcp_server_b.command)
+            mcp_servers_reachable[name_b] = reachable
+            if not reachable:
+                warnings.append(
+                    f"MCP server B command '{config.mcp_server_b.command}' not found in PATH."
+                )
+    elif config.mcp_server and config.mcp_server.command:
+        name = config.mcp_server.name or "mcp_server"
+        reachable = _check_mcp_server_reachable(config.mcp_server.command)
+        mcp_servers_reachable[name] = reachable
+        if not reachable:
+            warnings.append(f"MCP server command '{config.mcp_server.command}' not found in PATH.")
+
+    # 5. Check API key
+    api_key = os.environ.get("ANTHROPIC_API_KEY")
+    if not api_key:
+        warnings.append(
+            "ANTHROPIC_API_KEY environment variable is not set. "
+            "Evaluation requires a valid API key."
+        )
+
+    # 6. Estimate cost
+    estimated_cost = _estimate_cost(config.model, benchmark_name, total_tasks)
+    if estimated_cost is None:
+        warnings.append(
+            f"Could not estimate cost: pricing unavailable for model '{config.model}'. "
+            f"Cost estimation uses historical token usage averages and may vary."
+        )
+    else:
+        # Add a note about estimation accuracy
+        if verbosity >= 1:
+            warnings.append(
+                "Cost estimate is based on historical averages and actual costs may vary "
+                "significantly depending on task complexity and agent behavior."
+            )
+
+    # 7. Estimate time
+    estimated_time = _estimate_time(
+        benchmark_name,
+        total_tasks,
+        config.max_concurrent,
+        config.timeout_seconds,
+    )
+
+    # 8. Budget warning
+    if config.budget is not None and estimated_cost is not None:
+        if estimated_cost > config.budget:
+            warnings.append(
+                f"Estimated cost ({format_cost(estimated_cost)}) exceeds budget "
+                f"({format_cost(config.budget)}). Evaluation may be halted early."
+            )
+
+    return DryRunResult(
+        benchmark_name=benchmark_name,
+        total_tasks=total_tasks,
+        task_ids=task_ids,
+        estimated_cost_usd=estimated_cost,
+        estimated_time_minutes=estimated_time,
+        config_valid=config_valid,
+        config_errors=config_errors,
+        docker_available=docker_available,
+        mcp_servers_reachable=mcp_servers_reachable,
+        warnings=warnings,
+    )
+
+
+def format_dry_run_report(result: DryRunResult) -> None:
+    """Print a rich-formatted dry-run report to the console.
+
+    Displays a comprehensive overview of what the evaluation would do,
+    including task details, cost estimates, infrastructure readiness,
+    and any warnings or errors.
+
+    Args:
+        result: The DryRunResult to format and display.
+    """
+    console = Console()
+
+    # Header
+    console.print()
+    console.print(
+        Panel(
+            "[bold]Dry Run Report[/bold]\n[dim]Preview of evaluation without executing tasks[/dim]",
+            border_style="cyan",
+        )
+    )
+
+    # Benchmark & Tasks table
+    task_table = Table(
+        title="Evaluation Overview",
+        show_header=True,
+        header_style="bold cyan",
+    )
+    task_table.add_column("Property", style="bold")
+    task_table.add_column("Value")
+
+    task_table.add_row("Benchmark", result.benchmark_name)
+    task_table.add_row("Total Tasks", str(result.total_tasks))
+    task_table.add_row(
+        "Estimated Cost",
+        format_cost(result.estimated_cost_usd) if result.estimated_cost_usd is not None else "N/A",
+    )
+
+    if result.estimated_time_minutes is not None:
+        hours = int(result.estimated_time_minutes // 60)
+        minutes = int(result.estimated_time_minutes % 60)
+        if hours > 0:
+            time_str = f"{hours}h {minutes}m"
+        else:
+            time_str = f"{minutes}m"
+        task_table.add_row("Estimated Time", time_str)
+    else:
+        task_table.add_row("Estimated Time", "N/A")
+
+    console.print()
+    console.print(task_table)
+
+    # Task IDs (show first 10, then summarize)
+    if result.task_ids:
+        console.print()
+        console.print("[bold]Task IDs:[/bold]")
+        display_count = min(10, len(result.task_ids))
+        for task_id in result.task_ids[:display_count]:
+            console.print(f"  [dim]-[/dim] {task_id}")
+        if len(result.task_ids) > display_count:
+            console.print(f"  [dim]... and {len(result.task_ids) - display_count} more[/dim]")
+
+    # Infrastructure Readiness table
+    infra_table = Table(
+        title="Infrastructure Readiness",
+        show_header=True,
+        header_style="bold cyan",
+    )
+    infra_table.add_column("Check", style="bold")
+    infra_table.add_column("Status", justify="center")
+    infra_table.add_column("Details")
+
+    # Config validation
+    if result.config_valid:
+        infra_table.add_row("Configuration", "[green]PASS[/green]", "Valid")
+    else:
+        error_summary = "; ".join(result.config_errors[:3])
+        if len(result.config_errors) > 3:
+            error_summary += f" (+{len(result.config_errors) - 3} more)"
+        infra_table.add_row("Configuration", "[red]FAIL[/red]", error_summary)
+
+    # Docker
+    if result.docker_available:
+        infra_table.add_row("Docker", "[green]PASS[/green]", "Running")
+    else:
+        infra_table.add_row("Docker", "[red]FAIL[/red]", "Not available")
+
+    # MCP servers
+    for server_name, reachable in result.mcp_servers_reachable.items():
+        if reachable:
+            infra_table.add_row(f"MCP: {server_name}", "[green]PASS[/green]", "Command found")
+        else:
+            infra_table.add_row(f"MCP: {server_name}", "[red]FAIL[/red]", "Command not found")
+
+    # API key
+    api_key = os.environ.get("ANTHROPIC_API_KEY")
+    if api_key:
+        masked = f"{api_key[:8]}...{api_key[-4:]}" if len(api_key) > 12 else "***"
+        infra_table.add_row("API Key", "[green]PASS[/green]", f"Set ({masked})")
+    else:
+        infra_table.add_row("API Key", "[red]FAIL[/red]", "Not set")
+
+    console.print()
+    console.print(infra_table)
+
+    # Warnings
+    if result.warnings:
+        console.print()
+        console.print("[yellow bold]Warnings:[/yellow bold]")
+        for warning in result.warnings:
+            console.print(f"  [yellow]-[/yellow] {warning}")
+
+    # Config errors
+    if result.config_errors:
+        console.print()
+        console.print("[red bold]Configuration Errors:[/red bold]")
+        for error in result.config_errors:
+            console.print(f"  [red]-[/red] {error}")
+
+    # Summary
+    console.print()
+    all_clear = (
+        result.config_valid
+        and result.docker_available
+        and all(result.mcp_servers_reachable.values())
+        and api_key is not None
+    )
+    if all_clear:
+        console.print(
+            Panel(
+                "[green bold]All checks passed.[/green bold]\nThe evaluation is ready to run.",
+                border_style="green",
+            )
+        )
+    else:
+        console.print(
+            Panel(
+                "[red bold]Some checks failed.[/red bold]\n"
+                "Please resolve the issues above before running the evaluation.",
+                border_style="red",
+            )
+        )
+
+    console.print()