wafer-cli 0.2.7__py3-none-any.whl → 0.2.9__py3-none-any.whl

wafer/cli.py

@@ -30,6 +30,14 @@ import typer
 
 from .config import WaferConfig, WaferEnvironment
 from .inference import infer_upload_files, resolve_environment
+from .problems import (
+    download_problems,
+    get_problem_path,
+    get_problems_path,
+)
+from .problems import (
+    list_problems as list_problems_fn,
+)
 
 app = typer.Typer(
     help="GPU development toolkit for LLM coding agents",
@@ -91,11 +99,15 @@ def main_callback(ctx: typer.Context) -> None:
     # Install exception hook to catch SystemExit and mark failures
     original_excepthook = sys.excepthook
 
-    def custom_excepthook(exc_type, exc_value, exc_traceback):
+    def custom_excepthook(
+        exc_type: type[BaseException],
+        exc_value: BaseException,
+        exc_traceback: object,
+    ) -> None:
         global _command_outcome
         # Mark as failure if SystemExit with non-zero code, or any other exception
         if exc_type is SystemExit:
-            exit_code = exc_value.code if hasattr(exc_value, 'code') else 1
+            exit_code = exc_value.code if hasattr(exc_value, "code") else 1
             if exit_code != 0 and exit_code is not None:
                 _command_outcome = "failure"
         else:
@@ -200,6 +212,13 @@ kernelbench_app = typer.Typer(
 )
 evaluate_app.add_typer(kernelbench_app, name="kernelbench")
 
+# Nested subcommand for gpumode format
+gpumode_app = typer.Typer(
+    help="Evaluate kernels in GPUMode format (custom_kernel/ref_kernel functions)",
+    invoke_without_command=True,
+)
+evaluate_app.add_typer(gpumode_app, name="gpumode")
+
 # =============================================================================
 # Dev commands (internal, used by web app proxy)
 # =============================================================================
@@ -242,6 +261,10 @@ app.add_typer(amd_app, name="amd")
 isa_app = typer.Typer(help="ISA analysis for AMD GPU code objects (.co files)")
 amd_app.add_typer(isa_app, name="isa")
 
+# Kernel Scope - static ISA analysis for Triton kernels
+kernel_scope_app = typer.Typer(help="Static ISA analysis for Triton compilation artifacts")
+amd_app.add_typer(kernel_scope_app, name="kernel-scope")
+
 # =============================================================================
 # Skill management (wafer skill ...)
 # =============================================================================
@@ -396,6 +419,122 @@ def skill_status() -> None:
             typer.echo(f"{tool_name}: Not installed")
 
 
+# =============================================================================
+# Provider auth management (wafer auth ...)
+# =============================================================================
+
+provider_auth_app = typer.Typer(help="Manage API keys for cloud GPU providers")
+app.add_typer(provider_auth_app, name="auth")
+
+
+@provider_auth_app.command("login")
+def provider_auth_login(
+    provider: str = typer.Argument(
+        ...,
+        help="Provider name: runpod, digitalocean, or modal",
+    ),
+    api_key: str | None = typer.Option(
+        None,
+        "--api-key",
+        "-k",
+        help="API key (if not provided, reads from stdin)",
+    ),
+) -> None:
+    """Save API key for a cloud GPU provider.
+
+    Stores the key in ~/.wafer/auth.json. Environment variables
+    (e.g., WAFER_RUNPOD_API_KEY) take precedence over stored keys.
+
+    Examples:
+        wafer auth login runpod --api-key rp_xxx
+        wafer auth login digitalocean --api-key dop_v1_xxx
+        echo $API_KEY | wafer auth login runpod
+    """
+    import sys
+
+    from wafer_core.auth import PROVIDERS, save_api_key
+
+    # Validate provider
+    if provider not in PROVIDERS:
+        typer.echo(f"Error: Unknown provider '{provider}'", err=True)
+        typer.echo(f"Valid providers: {', '.join(PROVIDERS.keys())}", err=True)
+        raise typer.Exit(1)
+
+    # Get API key from option or stdin
+    if api_key is None:
+        if sys.stdin.isatty():
+            typer.echo(f"Enter API key for {PROVIDERS[provider]['display_name']}:")
+            api_key = typer.prompt("API key", hide_input=True)
+        else:
+            api_key = sys.stdin.read().strip()
+
+    if not api_key:
+        typer.echo("Error: No API key provided", err=True)
+        raise typer.Exit(1)
+
+    # Save the key
+    save_api_key(provider, api_key)
+    typer.echo(f"API key saved for {PROVIDERS[provider]['display_name']}")
+    typer.echo("Stored in: ~/.wafer/auth.json")
+
+
+@provider_auth_app.command("logout")
+def provider_auth_logout(
+    provider: str = typer.Argument(
+        ...,
+        help="Provider name: runpod, digitalocean, or modal",
+    ),
+) -> None:
+    """Remove stored API key for a cloud GPU provider.
+
+    Examples:
+        wafer auth logout runpod
+        wafer auth logout digitalocean
+    """
+    from wafer_core.auth import PROVIDERS, remove_api_key
+
+    # Validate provider
+    if provider not in PROVIDERS:
+        typer.echo(f"Error: Unknown provider '{provider}'", err=True)
+        typer.echo(f"Valid providers: {', '.join(PROVIDERS.keys())}", err=True)
+        raise typer.Exit(1)
+
+    if remove_api_key(provider):
+        typer.echo(f"API key removed for {PROVIDERS[provider]['display_name']}")
+    else:
+        typer.echo(f"No stored API key found for {PROVIDERS[provider]['display_name']}")
+
+
+@provider_auth_app.command("status")
+def provider_auth_status() -> None:
+    """Show authentication status for all cloud GPU providers.
+
+    Displays which providers have API keys configured and where
+    the keys are coming from (environment variable or auth.json).
+
+    Example:
+        wafer auth status
+    """
+    from wafer_core.auth import get_all_auth_status
+
+    statuses = get_all_auth_status()
+
+    typer.echo("Cloud GPU Provider Authentication Status")
+    typer.echo("=" * 45)
+
+    for status in statuses:
+        if status.is_authenticated:
+            source_str = f"({status.source})" if status.source else ""
+            typer.echo(f"  {status.display_name}: ✓ {status.key_preview} {source_str}")
+        else:
+            typer.echo(f"  {status.display_name}: ✗ Not configured")
+            typer.echo(f"      Run: wafer auth login {status.provider}")
+            typer.echo(f"      Or set: {status.key_url}")
+
+    typer.echo("")
+    typer.echo("Note: Environment variables take precedence over stored keys.")
+
+
 @app.command(hidden=True)
 def run(
     command: str = typer.Argument(..., help="Command to run in Docker container"),
@@ -1289,86 +1428,37 @@ def evaluate(  # noqa: PLR0913
             --benchmark --defensive
 
     Subcommands:
-        make-template  Generate template files for this format
+        gpumode        Use GPUMode format (functional) - RECOMMENDED
         kernelbench    Use KernelBench format (ModelNew class)
+        make-template  Generate template files for this format (deprecated)
     """
     # If a subcommand is being invoked, skip the main evaluation logic
     if ctx.invoked_subcommand is not None:
         return
 
-    # Validate required args when running evaluation (not subcommands)
-    missing_args = []
-    if implementation is None:
-        missing_args.append("--impl/-i")
-    if reference is None:
-        missing_args.append("--reference")
-    if test_cases is None:
-        missing_args.append("--test-cases")
-
-    if missing_args:
-        typer.echo("Error: Missing required arguments", err=True)
-        typer.echo(f"  Required: {', '.join(missing_args)}", err=True)
-        typer.echo("", err=True)
-        typer.echo(
-            "Usage: wafer evaluate --impl KERNEL.py --reference REF.py --test-cases TESTS.json",
-            err=True,
-        )
-        typer.echo("", err=True)
-        typer.echo("Run 'wafer evaluate --help' for full options.", err=True)
-        typer.echo("Run 'wafer evaluate make-template DIR' to generate starter files.", err=True)
-        raise typer.Exit(1)
-
-    from .evaluate import EvaluateArgs, run_evaluate
-
-    args = EvaluateArgs(
-        implementation=implementation,
-        reference=reference,
-        test_cases=test_cases,
-        target_name=target or "",
-        benchmark=benchmark,
-        profile=profile,
-        defensive=defensive,
-        sync_artifacts=sync_artifacts,
-        gpu_id=gpu_id,
+    # Bare 'wafer evaluate' is no longer supported - must use subcommand
+    typer.echo("Error: 'wafer evaluate' requires a subcommand.", err=True)
+    typer.echo("", err=True)
+    typer.echo("Available subcommands:", err=True)
+    typer.echo(
+        "  gpumode      Evaluate GPUMode format (custom_kernel/ref_kernel functions)", err=True
     )
-
-    try:
-        # Use trio_asyncio to run async code that uses both trio and asyncio
-        # (AsyncSSHClient uses asyncssh which is asyncio-based, bridged via trio_asyncio)
-        import trio_asyncio
-
-        result = trio_asyncio.run(run_evaluate, args)
-    except KeyboardInterrupt:
-        typer.echo("\nInterrupted by user", err=True)
-        raise typer.Exit(130) from None
-    except Exception as e:
-        # Unwrap ExceptionGroup (from Trio nurseries) to show actual error
-        if hasattr(e, "exceptions") and e.exceptions:
-            for exc in e.exceptions:
-                typer.echo(f"Error: {type(exc).__name__}: {exc}", err=True)
-        else:
-            typer.echo(f"Error: {e}", err=True)
-        raise typer.Exit(1) from None
-
-    # Print results
-    if result.success:
-        typer.echo("")
-        typer.echo("=" * 60)
-        status = "PASS" if result.all_correct else "FAIL"
-        typer.echo(f"Result: {status}")
-        score_pct = f"{result.correctness_score:.1%}"
-        typer.echo(f"Correctness: {result.passed_tests}/{result.total_tests} ({score_pct})")
-        if result.geomean_speedup > 0:
-            typer.echo(f"Speedup: {result.geomean_speedup:.2f}x")
-        if result.artifact_path:
-            typer.echo(f"Artifacts: {result.artifact_path}")
-        typer.echo("=" * 60)
-
-        if not result.all_correct:
-            raise typer.Exit(1)
-    else:
-        typer.echo(f"Error: {result.error_message}", err=True)
-        raise typer.Exit(1)
+    typer.echo("  kernelbench  Evaluate KernelBench format (ModelNew class)", err=True)
+    typer.echo("", err=True)
+    typer.echo("Examples:", err=True)
+    typer.echo(
+        "  wafer evaluate gpumode --impl kernel.py --reference ref.py --test-cases tests.json",
+        err=True,
+    )
+    typer.echo(
+        "  wafer evaluate kernelbench --impl impl.py --reference ref.py --benchmark", err=True
+    )
+    typer.echo("", err=True)
+    typer.echo(
+        "Run 'wafer evaluate gpumode --help' or 'wafer evaluate kernelbench --help' for options.",
+        err=True,
+    )
+    raise typer.Exit(1)
 
 
 TEMPLATE_KERNEL = '''\
@@ -1503,8 +1593,59 @@ def evaluate_make_template(
 # KernelBench format evaluation
 # =============================================================================
 
-# Path to KernelBench problems (relative to wafer root)
-KERNELBENCH_ROOT = Path(__file__).parent.parent.parent.parent / "research" / "KernelBench"
+
+def _get_kernelbench_root() -> Path | None:
+    """Get KernelBench problems root, preferring downloaded location."""
+    # First check downloaded location
+    downloaded = get_problems_path("kernelbench")
+    if downloaded is not None:
+        kb_root = downloaded / "KernelBench"
+        if kb_root.exists():
+            return kb_root
+        return downloaded
+
+    # Fall back to legacy location (for development)
+    legacy = Path(__file__).parent.parent.parent.parent / "research" / "KernelBench" / "KernelBench"
+    if legacy.exists():
+        return legacy
+
+    return None
+
+
+@kernelbench_app.command("download")
+def kernelbench_download(
+    force: bool = typer.Option(False, "--force", "-f", help="Re-download even if exists"),
+) -> None:
+    """Download KernelBench problems from GitHub.
+
+    Downloads the problem set to ~/.cache/wafer/problems/kernelbench/
+
+    Examples:
+        wafer evaluate kernelbench download
+        wafer evaluate kernelbench download --force  # Re-download
+    """
+    try:
+        path = download_problems("kernelbench", force=force, verbose=True)
+        typer.echo("")
+        typer.echo(f"Problems available at: {path}")
+        typer.echo("Run 'wafer evaluate kernelbench list-problems' to see available problems.")
+    except Exception as e:
+        typer.echo(f"Error downloading problems: {e}", err=True)
+        raise typer.Exit(1) from None
+
+
+@kernelbench_app.command("list-problems")
+def kernelbench_list_problems() -> None:
+    """List available KernelBench problems.
+
+    Examples:
+        wafer evaluate kernelbench list-problems
+    """
+    try:
+        list_problems_fn("kernelbench", verbose=True)
+    except ValueError as e:
+        typer.echo(str(e), err=True)
+        raise typer.Exit(1) from None
 
 
 @kernelbench_app.callback(invoke_without_command=True)
@@ -1528,9 +1669,18 @@ def kernelbench_evaluate(  # noqa: PLR0913
         help="GPU target name. See 'wafer config targets list' for available targets.",
         autocompletion=complete_target_name,
     ),
+    pool: str | None = typer.Option(
+        None,
+        "--pool",
+        "-p",
+        help="Target pool name. Acquires first available target from the pool. "
+        "Define pools in ~/.wafer/config.toml under [pools.<name>].",
+    ),
     benchmark: bool = typer.Option(False, "--benchmark", help="Run performance benchmarks"),
     profile: bool = typer.Option(False, "--profile", help="Enable profiling"),
-    inputs: Path | None = typer.Option(None, "--inputs", help="Custom inputs file to override get_inputs()"),
+    inputs: Path | None = typer.Option(
+        None, "--inputs", help="Custom inputs file to override get_inputs()"
+    ),
     seed: int = typer.Option(42, "--seed", help="Random seed for weight initialization"),
     defensive: bool = typer.Option(
         False, "--defensive", help="Enable defensive timing to detect evaluation hacking"
@@ -1588,12 +1738,54 @@ def kernelbench_evaluate(  # noqa: PLR0913
         )
         raise typer.Exit(1)
 
+    # Validate --target and --pool are mutually exclusive
+    if target and pool:
+        typer.echo("Error: Cannot specify both --target and --pool", err=True)
+        raise typer.Exit(1)
+
     from .evaluate import KernelBenchEvaluateArgs, run_evaluate_kernelbench
 
+    # If pool specified, acquire a target from the pool
+    resolved_target = target or ""
+    pool_lock_context = None
+
+    if pool:
+        from .target_lock import acquire_from_pool
+        from .targets import filter_pool_by_auth, get_pool
+
+        try:
+            pool_targets = get_pool(pool)
+        except FileNotFoundError as e:
+            typer.echo(f"Error: {e}", err=True)
+            raise typer.Exit(1) from None
+
+        # Filter to only targets with valid auth
+        usable_targets, skipped = filter_pool_by_auth(pool_targets)
+        if skipped:
+            typer.echo(f"Skipping targets without auth: {', '.join(skipped)}", err=True)
+
+        if not usable_targets:
+            typer.echo(f"Error: No usable targets in pool '{pool}'", err=True)
+            typer.echo("  All targets require authentication that is not configured.", err=True)
+            typer.echo("  Run 'wafer auth status' to see which providers need setup.", err=True)
+            raise typer.Exit(1) from None
+
+        typer.echo(f"Acquiring target from pool '{pool}' ({len(usable_targets)} targets)...")
+        pool_lock_context = acquire_from_pool(usable_targets)
+        acquired_target = pool_lock_context.__enter__()
+
+        if acquired_target is None:
+            typer.echo(f"Error: All targets in pool '{pool}' are busy", err=True)
+            typer.echo(f"  Targets: {', '.join(usable_targets)}", err=True)
+            raise typer.Exit(1)
+
+        typer.echo(f"Acquired target: {acquired_target}")
+        resolved_target = acquired_target
+
     args = KernelBenchEvaluateArgs(
         implementation=implementation,
         reference=reference,
-        target_name=target or "",
+        target_name=resolved_target,
         benchmark=benchmark,
         profile=profile,
         inputs=inputs,
@@ -1613,6 +1805,10 @@ def kernelbench_evaluate(  # noqa: PLR0913
     except Exception as e:
         typer.echo(f"Error: {e}", err=True)
         raise typer.Exit(1) from None
+    finally:
+        # Release pool lock if we acquired one
+        if pool_lock_context is not None:
+            pool_lock_context.__exit__(None, None, None)
 
     # Print results
     if result.success:
@@ -1659,6 +1855,13 @@ def kernelbench_make_template(
         # Overwrite existing
         wafer evaluate kernelbench make-template level1/1 --force
     """
+    # Get problems root (downloaded or legacy)
+    kb_root = _get_kernelbench_root()
+    if kb_root is None:
+        typer.echo("Error: KernelBench problems not found.", err=True)
+        typer.echo("Run 'wafer evaluate kernelbench download' to download problems.", err=True)
+        raise typer.Exit(1)
+
     # Parse problem ID
     parts = problem.split("/")
     if len(parts) != 2:
@@ -1670,10 +1873,10 @@ def kernelbench_make_template(
         level_str = f"level{level_str}"
 
     # Find the problem file
-    problem_dir = KERNELBENCH_ROOT / "KernelBench" / level_str
+    problem_dir = kb_root / level_str
     if not problem_dir.exists():
         typer.echo(f"Error: KernelBench level directory not found: {problem_dir}", err=True)
-        typer.echo(f"Make sure KernelBench is at: {KERNELBENCH_ROOT}", err=True)
+        typer.echo("Run 'wafer evaluate kernelbench download' to download problems.", err=True)
         raise typer.Exit(1)
 
     # Find matching problem file
@@ -1708,37 +1911,335 @@ def kernelbench_make_template(
 
     output = output.resolve()
 
-    # Check if exists
-    if output.exists() and not force:
-        typer.echo(f"Error: {output} already exists. Use --force to overwrite.", err=True)
+    # Check if exists
+    if output.exists() and not force:
+        typer.echo(f"Error: {output} already exists. Use --force to overwrite.", err=True)
+        raise typer.Exit(1)
+
+    # Copy the file
+    content = problem_file.read_text()
+    output.parent.mkdir(parents=True, exist_ok=True)
+    output.write_text(content)
+
+    typer.echo(f"Created {output}")
+    typer.echo("")
+    typer.echo("Next steps:")
+    typer.echo(f"  1. Read {output} to understand the Model interface")
+    typer.echo("  2. Create an implementation file with your ModelNew class:")
+    typer.echo("")
+    typer.echo("     import torch.nn as nn")
+    typer.echo("")
+    typer.echo("     class ModelNew(nn.Module):")
+    typer.echo("         def __init__(self, ...):")
+    typer.echo("             # Same signature as Model.__init__")
+    typer.echo("             ...")
+    typer.echo("")
+    typer.echo("         def forward(self, ...):")
+    typer.echo("             # Same signature as Model.forward")
+    typer.echo("             # Your optimized implementation here")
+    typer.echo("             ...")
+    typer.echo("")
+    typer.echo("  3. Run evaluation:")
+    typer.echo(f"     wafer evaluate kernelbench --impl my_kernel.py --reference {output}")
+
+
+# =============================================================================
+# GPUMode format evaluation
+# =============================================================================
+
+
+@gpumode_app.command("download")
+def gpumode_download(
+    force: bool = typer.Option(False, "--force", "-f", help="Re-download even if exists"),
+) -> None:
+    """Download GPUMode reference kernels from GitHub.
+
+    Downloads the problem set to ~/.cache/wafer/problems/gpumode/
+
+    Examples:
+        wafer evaluate gpumode download
+        wafer evaluate gpumode download --force  # Re-download
+    """
+    try:
+        path = download_problems("gpumode", force=force, verbose=True)
+        typer.echo("")
+        typer.echo(f"Problems available at: {path}")
+        typer.echo("Run 'wafer evaluate gpumode list-problems' to see available problems.")
+    except Exception as e:
+        typer.echo(f"Error downloading problems: {e}", err=True)
+        raise typer.Exit(1) from None
+
+
+@gpumode_app.command("list-problems")
+def gpumode_list_problems() -> None:
+    """List available GPUMode problems.
+
+    Examples:
+        wafer evaluate gpumode list-problems
+    """
+    try:
+        list_problems_fn("gpumode", verbose=True)
+    except ValueError as e:
+        typer.echo(str(e), err=True)
+        raise typer.Exit(1) from None
+
+
+@gpumode_app.command("make-template")
+def gpumode_make_template(
+    problem: str = typer.Option(
+        ...,
+        "--problem",
+        "-p",
+        help="Problem ID (e.g., 'pmpp/vectoradd_py' or 'amd/fp8-mm')",
+    ),
+    output: Path = typer.Option(
+        None, "--output", "-o", help="Output directory (default: ./<problem_name>/)"
+    ),
+    force: bool = typer.Option(False, "--force", "-f", help="Overwrite existing files"),
+) -> None:
+    """Extract a GPUMode problem as template files.
+
+    Creates a directory with reference.py, task.yml, and other problem files.
+    You then create kernel.py with your custom_kernel implementation.
+
+    Examples:
+        # Extract pmpp vectoradd problem
+        wafer evaluate gpumode make-template --problem pmpp/vectoradd_py
+
+        # Extract to specific directory
+        wafer evaluate gpumode make-template --problem pmpp/vectoradd_py --output ./my-kernel/
+    """
+    import shutil
+
+    # Get problem path
+    problem_path = get_problem_path("gpumode", problem)
+    if problem_path is None:
+        # Check if problems are downloaded
+        if get_problems_path("gpumode") is None:
+            typer.echo("Error: GPUMode problems not downloaded.", err=True)
+            typer.echo("Run 'wafer evaluate gpumode download' first.", err=True)
+        else:
+            typer.echo(f"Error: Problem '{problem}' not found.", err=True)
+            typer.echo(
+                "Run 'wafer evaluate gpumode list-problems' to see available problems.", err=True
+            )
+        raise typer.Exit(1)
+
+    # Determine output path
+    if output is None:
+        output = Path.cwd() / problem.replace("/", "_")
+
+    output = output.resolve()
+
+    # Check if exists
+    if output.exists() and not force:
+        typer.echo(f"Error: {output} already exists. Use --force to overwrite.", err=True)
+        raise typer.Exit(1)
+
+    # Copy the problem directory
+    if output.exists():
+        shutil.rmtree(output)
+    shutil.copytree(problem_path, output)
+
+    typer.echo(f"Created {output}/")
+    typer.echo("")
+    typer.echo("Contents:")
+    for f in sorted(output.iterdir()):
+        if not f.name.startswith("."):
+            typer.echo(f"  {f.name}")
+    typer.echo("")
+    typer.echo("Next steps:")
+    typer.echo("  1. Read reference.py to understand the kernel interface")
+    typer.echo("  2. Create kernel.py with your custom_kernel implementation:")
+    typer.echo("")
+    typer.echo("     def custom_kernel(data):")
+    typer.echo("         # Your optimized implementation")
+    typer.echo("         ...")
+    typer.echo("")
+    typer.echo("  3. Run evaluation:")
+    typer.echo(
+        f"     wafer evaluate gpumode --impl {output}/kernel.py --reference {output}/reference.py \\"
+    )
+    typer.echo(f"         --test-cases {output}/test_cases.json --target <target>")
+
+
+@gpumode_app.callback(invoke_without_command=True)
+def gpumode_evaluate(  # noqa: PLR0913, PLR0915
+    ctx: typer.Context,
+    implementation: Path | None = typer.Option(
+        None, "--impl", "-i", help="Path to implementation kernel file"
+    ),
+    reference: Path | None = typer.Option(
+        None, "--reference", help="Path to reference kernel file"
+    ),
+    test_cases: Path | None = typer.Option(
+        None, "--test-cases", help="Path to test cases JSON file"
+    ),
+    target: str | None = typer.Option(
+        None,
+        "--target",
+        "-t",
+        help="GPU target name. See 'wafer config targets list' for available targets.",
+        autocompletion=complete_target_name,
+    ),
+    pool: str | None = typer.Option(
+        None,
+        "--pool",
+        "-p",
+        help="Target pool name. Acquires first available target from the pool. "
+        "Define pools in ~/.wafer/config.toml under [pools.<name>].",
+    ),
+    benchmark: bool = typer.Option(False, "--benchmark", help="Run performance benchmarks"),
+    profile: bool = typer.Option(False, "--profile", help="Enable profiling"),
+    defensive: bool = typer.Option(
+        False, "--defensive", help="Enable defensive timing to detect evaluation hacking"
+    ),
+    sync_artifacts: bool = typer.Option(
+        True, "--sync-artifacts/--no-sync-artifacts", help="Download artifacts"
+    ),
+    gpu_id: int | None = typer.Option(None, "--gpu-id", help="Override GPU ID"),
+) -> None:
+    """Run kernel evaluation in GPUMode format (functional).
+
+    This format expects:
+    - Implementation: Python file with `custom_kernel(inputs)` function
+    - Reference: Python file with `ref_kernel(inputs)` and `generate_input(**kwargs)` functions
+    - Test cases: JSON file with test parameters
+
+    Examples:
+        # Basic correctness check
+        wafer evaluate gpumode --impl kernel.py --reference ref.py --test-cases tests.json
+
+        # With benchmarking
+        wafer evaluate gpumode --impl kernel.py --reference ref.py --test-cases tests.json \\
+            --target vultr-b200 --benchmark
+
+    Subcommands:
+        download       Download GPUMode problems from GitHub
+        list-problems  List available problems
+        make-template  Extract a problem as template files
+    """
+    # If a subcommand is being invoked, skip the main evaluation logic
+    if ctx.invoked_subcommand is not None:
+        return
+
+    # Validate required args when running evaluation (not subcommands)
+    missing_args = []
+    if implementation is None:
+        missing_args.append("--impl/-i")
+    if reference is None:
+        missing_args.append("--reference")
+    if test_cases is None:
+        missing_args.append("--test-cases")
+
+    if missing_args:
+        typer.echo("Error: Missing required arguments", err=True)
+        typer.echo(f"  Required: {', '.join(missing_args)}", err=True)
+        typer.echo("", err=True)
+        typer.echo(
+            "Usage: wafer evaluate gpumode --impl KERNEL.py --reference REF.py --test-cases TESTS.json",
+            err=True,
+        )
+        typer.echo("", err=True)
+        typer.echo("Run 'wafer evaluate gpumode --help' for full options.", err=True)
+        typer.echo("Run 'wafer evaluate gpumode download' to download problem sets.", err=True)
+        raise typer.Exit(1)
+
+    # Validate --target and --pool are mutually exclusive
+    if target and pool:
+        typer.echo("Error: Cannot specify both --target and --pool", err=True)
+        raise typer.Exit(1)
+
+    from .evaluate import EvaluateArgs, run_evaluate
+
+    # If pool specified, acquire a target from the pool
+    resolved_target = target or ""
+    pool_lock_context = None
+
+    if pool:
+        from .target_lock import acquire_from_pool
+        from .targets import filter_pool_by_auth, get_pool
+
+        try:
+            pool_targets = get_pool(pool)
+        except FileNotFoundError as e:
+            typer.echo(f"Error: {e}", err=True)
+            raise typer.Exit(1) from None
+
+        # Filter to only targets with valid auth
+        usable_targets, skipped = filter_pool_by_auth(pool_targets)
+        if skipped:
+            typer.echo(f"Skipping targets without auth: {', '.join(skipped)}", err=True)
+
+        if not usable_targets:
+            typer.echo(f"Error: No usable targets in pool '{pool}'", err=True)
+            typer.echo("  All targets require authentication that is not configured.", err=True)
+            typer.echo("  Run 'wafer auth status' to see which providers need setup.", err=True)
+            raise typer.Exit(1) from None
+
+        typer.echo(f"Acquiring target from pool '{pool}' ({len(usable_targets)} targets)...")
+        pool_lock_context = acquire_from_pool(usable_targets)
+        acquired_target = pool_lock_context.__enter__()
+
+        if acquired_target is None:
+            typer.echo(f"Error: All targets in pool '{pool}' are busy", err=True)
+            typer.echo(f"  Targets: {', '.join(usable_targets)}", err=True)
+            raise typer.Exit(1)
+
+        typer.echo(f"Acquired target: {acquired_target}")
+        resolved_target = acquired_target
+
+    args = EvaluateArgs(
+        implementation=implementation,
+        reference=reference,
+        test_cases=test_cases,
+        target_name=resolved_target,
+        benchmark=benchmark,
+        profile=profile,
+        defensive=defensive,
+        sync_artifacts=sync_artifacts,
+        gpu_id=gpu_id,
+    )
+
+    try:
+        import trio_asyncio
+
+        result = trio_asyncio.run(run_evaluate, args)
+    except KeyboardInterrupt:
+        typer.echo("\nInterrupted by user", err=True)
+        raise typer.Exit(130) from None
+    except Exception as e:
+        if hasattr(e, "exceptions") and e.exceptions:
+            for exc in e.exceptions:
+                typer.echo(f"Error: {type(exc).__name__}: {exc}", err=True)
+        else:
+            typer.echo(f"Error: {e}", err=True)
+        raise typer.Exit(1) from None
+    finally:
+        # Release pool lock if we acquired one
+        if pool_lock_context is not None:
+            pool_lock_context.__exit__(None, None, None)
+
+    # Print results
+    if result.success:
+        typer.echo("")
+        typer.echo("=" * 60)
+        status = "PASS" if result.all_correct else "FAIL"
+        typer.echo(f"Result: {status}")
+        score_pct = f"{result.correctness_score:.1%}"
+        typer.echo(f"Correctness: {result.passed_tests}/{result.total_tests} ({score_pct})")
+        if result.geomean_speedup > 0:
+            typer.echo(f"Speedup: {result.geomean_speedup:.2f}x")
+        if result.artifact_path:
+            typer.echo(f"Artifacts: {result.artifact_path}")
+        typer.echo("=" * 60)
+
+        if not result.all_correct:
+            raise typer.Exit(1)
+    else:
+        typer.echo(f"Error: {result.error_message}", err=True)
         raise typer.Exit(1)
 
-    # Copy the file
-    content = problem_file.read_text()
-    output.parent.mkdir(parents=True, exist_ok=True)
-    output.write_text(content)
-
-    typer.echo(f"Created {output}")
-    typer.echo("")
-    typer.echo("Next steps:")
-    typer.echo(f"  1. Read {output} to understand the Model interface")
-    typer.echo("  2. Create an implementation file with your ModelNew class:")
-    typer.echo("")
-    typer.echo("     import torch.nn as nn")
-    typer.echo("")
-    typer.echo("     class ModelNew(nn.Module):")
-    typer.echo("         def __init__(self, ...):")
-    typer.echo("             # Same signature as Model.__init__")
-    typer.echo("             ...")
-    typer.echo("")
-    typer.echo("         def forward(self, ...):")
-    typer.echo("             # Same signature as Model.forward")
-    typer.echo("             # Your optimized implementation here")
-    typer.echo("             ...")
-    typer.echo("")
-    typer.echo("  3. Run evaluation:")
-    typer.echo(f"     wafer evaluate kernelbench --impl my_kernel.py --reference {output}")
-
 
 # =============================================================================
 # Push and Remote-Run commands
@@ -1871,7 +2372,7 @@ def _run_direct_mode(
         typer.echo(f"Uploading {upload_dir.name}...")
         try:
             push_result = push_direct(upload_dir, target)
-            workspace_name = push_result.workspace_path
+            workspace_name = push_result.workspace_name
             typer.echo(f"Uploaded {len(push_result.files_uploaded)} files")
         except Exception as e:
             typer.echo(f"Error uploading: {e}", err=True)
@@ -2044,27 +2545,41 @@ def login(
         None, "--token", "-t", help="Access token (skip browser OAuth)"
     ),
     port: int | None = typer.Option(
-        None, "--port", "-p", help="Port for OAuth callback server (default: 8765 for SSH, random for local)"
+        None,
+        "--port",
+        "-p",
+        help="Port for OAuth callback server (local only, ignored for SSH)",
+    ),
+    no_device_code: bool = typer.Option(
+        False,
+        "--no-device-code",
+        help="Force browser OAuth even on SSH (requires port forwarding)",
     ),
 ) -> None:
     """Authenticate CLI with wafer-api via GitHub OAuth.
 
-    Opens browser for GitHub authentication. Use --token to skip browser.
+    Local: Opens browser for GitHub authentication.
+    SSH: Uses device code flow (no port forwarding needed).
+
     Uses the API environment from config (see 'wafer config show').
 
-    SSH Users:
-    - Automatically uses port 8765 (just set up port forwarding once)
-    - On local machine: ssh -L 8765:localhost:8765 user@host
-    - On remote machine: wafer login
-    - Browser opens locally, redirect works through tunnel
+    SSH Users (Easiest):
+    - Just run: wafer login
+    - Visit the URL and enter the code shown
+    - No port forwarding needed!
+
+    SSH with browser (Advanced):
+    - Use --no-device-code to force browser flow
+    - Requires: ssh -L 8765:localhost:8765 user@host
 
     Manual token option:
     - Visit auth.wafer.ai, authenticate, copy token from URL
     - Run: wafer login --token <paste-token>
 
     Examples:
-        wafer login                    # auto-detects SSH, uses appropriate port
-        wafer login --port 9000        # override port
+        wafer login                    # device code on SSH, browser on local
+        wafer login --no-device-code   # force browser (needs port forwarding on SSH)
+        wafer login --port 9000        # custom port for browser flow
         wafer login --token xyz        # manual token (no browser)
 
         # Change environment:
@@ -2073,7 +2588,7 @@ def login(
     """
     import httpx
 
-    from .auth import browser_login, save_credentials, verify_token
+    from .auth import browser_login, device_code_login, save_credentials, verify_token
     from .global_config import get_api_url, get_supabase_url, load_global_config
 
     # Show which environment we're logging into
@@ -2083,21 +2598,31 @@ def login(
     typer.echo(f"Auth: {get_supabase_url()}")
     typer.echo("")
 
-    # Auto-detect SSH and use fixed port
-    if port is None:
-        is_ssh = bool(os.environ.get("SSH_CONNECTION") or os.environ.get("SSH_CLIENT"))
-        if is_ssh:
-            port = 8765
-            typer.echo("🔒 SSH session detected - using port 8765 for OAuth callback")
-            typer.echo("   Make sure you have port forwarding set up:")
-            typer.echo("   ssh -L 8765:localhost:8765 user@host")
-            typer.echo("")
+    # Auto-detect SSH
+    is_ssh = bool(os.environ.get("SSH_CONNECTION") or os.environ.get("SSH_CLIENT"))
 
-    # Browser OAuth if no token provided
+    # Choose auth method
     refresh_token = None
     if token is None:
         try:
-            token, refresh_token = browser_login(port=port)
+            if is_ssh and not no_device_code:
+                # Use device code flow for SSH (no port forwarding needed)
+                typer.echo("🔒 SSH session detected - using device code authentication")
+                typer.echo("   (No port forwarding required!)")
+                typer.echo("")
+                token, refresh_token = device_code_login()
+            else:
+                # Use browser OAuth for local or if explicitly requested
+                if is_ssh:
+                    typer.echo("🔒 SSH session detected - using browser authentication")
+                    typer.echo("   Make sure you have port forwarding set up:")
+                    if port is None:
+                        port = 8765
+                        typer.echo(f"   ssh -L {port}:localhost:{port} user@host")
+                    else:
+                        typer.echo(f"   ssh -L {port}:localhost:{port} user@host")
+                    typer.echo("")
+                token, refresh_token = browser_login(port=port)
         except TimeoutError as e:
             typer.echo(f"Error: {e}", err=True)
             raise typer.Exit(1) from None
@@ -2146,9 +2671,8 @@ def login(
 @app.command("logout")
 def logout() -> None:
     """Remove stored credentials."""
-    from .auth import clear_credentials
-
     from . import analytics
+    from .auth import clear_credentials
 
     # Track logout event first (while credentials still exist for user identification)
     # Note: track_logout() handles the case where user is not logged in
@@ -2625,6 +3149,7 @@ init_app = typer.Typer(
 
 Choose based on your GPU access:
 
+  local        GPU on current machine (no SSH)
   ssh          Your own hardware via SSH
   runpod       RunPod cloud GPUs (needs WAFER_RUNPOD_API_KEY)
   digitalocean DigitalOcean AMD MI300X (needs WAFER_AMD_DIGITALOCEAN_API_KEY)"""
@@ -2632,6 +3157,92 @@ Choose based on your GPU access:
 targets_app.add_typer(init_app, name="init")
 
 
+@init_app.command("local")
+def init_local(
+    name: str = typer.Option("local", "--name", "-n", help="Target name"),
+    gpu_ids: str = typer.Option("0", "--gpu-ids", "-g", help="Comma-separated GPU IDs"),
+) -> None:
+    """Initialize a local target for GPU on current machine.
+
+    Detects your local GPU and configures a target for direct execution
+    (no SSH). Use this when running wafer on the same machine as the GPU.
+
+    Examples:
+        wafer config targets init local
+        wafer config targets init local --name my-5090 --gpu-ids 0,1
+    """
+    from .targets import save_target
+
+    # Parse GPU IDs
+    try:
+        parsed_gpu_ids = [int(g.strip()) for g in gpu_ids.split(",")]
+    except ValueError:
+        typer.echo(f"Error: Invalid GPU IDs '{gpu_ids}'. Use comma-separated integers.", err=True)
+        raise typer.Exit(1) from None
+
+    typer.echo("Detecting local GPU...")
+
+    try:
+        from wafer_core.gpu_detect import (
+            detect_local_gpu,
+            get_compute_capability,
+            get_torch_requirements,
+        )
+
+        detected_gpu = detect_local_gpu()
+
+        if detected_gpu:
+            typer.echo(f"  Found: {detected_gpu.gpu_name}")
+            if detected_gpu.vendor == "nvidia":
+                typer.echo(f"  CUDA: {detected_gpu.driver_version}")
+            else:
+                typer.echo(f"  ROCm: {detected_gpu.driver_version}")
+            typer.echo(f"  GPU count: {detected_gpu.gpu_count}")
+
+            # Get torch requirements and compute capability
+            torch_reqs = get_torch_requirements(detected_gpu)
+            compute_capability = get_compute_capability(detected_gpu)
+            gpu_type = _extract_gpu_type(detected_gpu.gpu_name)
+
+            typer.echo(f"  PyTorch: {torch_reqs.packages[0]}")
+        else:
+            typer.echo("  No GPU detected (nvidia-smi/rocm-smi not found)", err=True)
+            raise typer.Exit(1)
+
+    except ImportError as e:
+        typer.echo(f"Error: Missing dependency: {e}", err=True)
+        raise typer.Exit(1) from None
+
+    # Build target data
+    target_data = {
+        "name": name,
+        "type": "local",
+        "gpu_ids": parsed_gpu_ids,
+        "gpu_type": gpu_type,
+        "compute_capability": compute_capability,
+        "torch_package": torch_reqs.packages[0],
+        "torch_index_url": torch_reqs.index_url,
+        "vendor": detected_gpu.vendor,
+        "driver_version": detected_gpu.driver_version,
+    }
+
+    try:
+        target = save_target(target_data)
+        typer.echo(f"✓ Created target: {target.name}")
+        typer.echo("  Type: Local (no SSH)")
+        typer.echo(f"  GPU IDs: {parsed_gpu_ids}")
+        typer.echo(f"  GPU Type: {gpu_type}")
+        typer.echo(f"  Compute: {compute_capability}")
+        typer.echo(f"  Torch: {torch_reqs.packages[0]}")
+        typer.echo("")
+        typer.echo(
+            f"Usage: wafer evaluate --target {name} --impl kernel.py --reference ref.py --test-cases tests.json"
+        )
+    except (ValueError, AssertionError) as e:
+        typer.echo(f"Error: {e}", err=True)
+        raise typer.Exit(1) from None
+
+
 @init_app.command("runpod")
 def init_runpod(
     name: str = typer.Option("runpod-mi300x", "--name", "-n", help="Target name"),
@@ -2795,23 +3406,29 @@ def init_ssh(
     host: str = typer.Option(..., "--host", "-H", help="SSH host (user@hostname:port)"),
     ssh_key: str = typer.Option("~/.ssh/id_ed25519", "--ssh-key", "-k", help="Path to SSH key"),
     gpu_ids: str = typer.Option("0", "--gpu-ids", "-g", help="Comma-separated GPU IDs"),
-    gpu_type: str = typer.Option(
-        "H100", "--gpu-type", help="GPU type (H100, A100, B200, MI300X, etc.)"
+    gpu_type: str | None = typer.Option(
+        None, "--gpu-type", help="GPU type (auto-detected if not specified)"
     ),
     docker_image: str | None = typer.Option(
         None, "--docker-image", "-d", help="Docker image (optional)"
     ),
     ncu: bool = typer.Option(False, "--ncu/--no-ncu", help="NCU profiling available"),
+    no_detect: bool = typer.Option(False, "--no-detect", help="Skip GPU auto-detection"),
 ) -> None:
     """Initialize an SSH target for your own GPU hardware.
 
     Creates a target config for direct SSH access to a GPU machine.
-    Use for baremetal servers, VMs, or any machine you have SSH access to.
+    Automatically detects GPU type and selects compatible PyTorch version.
 
     Examples:
+        # Auto-detect GPU (recommended)
         wafer config targets init ssh --name my-gpu --host user@192.168.1.100:22
+
+        # Multiple GPUs with NCU profiling
         wafer config targets init ssh --name lab-h100 --host ubuntu@gpu.lab.com:22 --gpu-ids 0,1 --ncu
-        wafer config targets init ssh --name docker-gpu --host user@host:22 --docker-image nvcr.io/nvidia/pytorch:24.01-py3
+
+        # Skip detection, specify manually
+        wafer config targets init ssh --name my-gpu --host user@host:22 --gpu-type H100 --no-detect
     """
     from .targets import save_target
 
@@ -2828,17 +3445,86 @@ def init_ssh(
         typer.echo("Example: user@192.168.1.100:22", err=True)
         raise typer.Exit(1)
 
+    # Auto-detect GPU if not specified
+    detected_gpu = None
+    torch_package = None
+    torch_index_url = None
+
+    if not no_detect:
+        typer.echo(f"Connecting to {host}...")
+        try:
+            import trio
+            import trio_asyncio
+            from wafer_core.async_ssh import AsyncSSHClient
+            from wafer_core.gpu_detect import (
+                detect_remote_gpu,
+                get_compute_capability,
+                get_torch_requirements,
+            )
+
+            expanded_key = str(Path(ssh_key).expanduser())
+
+            async def _detect() -> None:
+                nonlocal detected_gpu, torch_package, torch_index_url
+                # Need trio_asyncio.open_loop() for asyncssh bridge
+                async with trio_asyncio.open_loop():
+                    async with AsyncSSHClient(host, expanded_key) as client:
+                        detected_gpu = await detect_remote_gpu(client)
+
+            trio.run(_detect)
+
+            if detected_gpu:
+                typer.echo(f"  Found: {detected_gpu.gpu_name}")
+                if detected_gpu.vendor == "nvidia":
+                    typer.echo(f"  CUDA: {detected_gpu.driver_version}")
+                else:
+                    typer.echo(f"  ROCm: {detected_gpu.driver_version}")
+
+                # Get torch requirements
+                torch_reqs = get_torch_requirements(detected_gpu)
+                torch_package = torch_reqs.packages[0]  # Just torch, not all packages
+                torch_index_url = torch_reqs.index_url
+                typer.echo(f"  PyTorch: {torch_package}")
+
+                # Use detected GPU type if not specified
+                if not gpu_type:
+                    # Extract GPU name (e.g., "H100" from "NVIDIA H100 80GB HBM3")
+                    gpu_type = _extract_gpu_type(detected_gpu.gpu_name)
+            else:
+                typer.echo("  No GPU detected (nvidia-smi/rocm-smi not found)")
+                if not gpu_type:
+                    gpu_type = "H100"  # Default fallback
+                    typer.echo(f"  Using default: {gpu_type}")
+
+        except Exception as e:
+            typer.echo(f"  Detection failed: {e}", err=True)
+            if not gpu_type:
+                gpu_type = "H100"
+                typer.echo(f"  Using default: {gpu_type}")
+
+    # Fallback if no detection
+    if not gpu_type:
+        gpu_type = "H100"
+
     # Compute capability mappings
-    compute_caps = {
-        "B200": "10.0",
-        "H100": "9.0",
-        "A100": "8.0",
-        "A10": "8.6",
-        "V100": "7.0",
-        "MI300X": "9.4",
-        "MI250X": "9.0",
-    }
-    compute_capability = compute_caps.get(gpu_type, "8.0")
+    if detected_gpu:
+        from wafer_core.gpu_detect import get_compute_capability
+
+        compute_capability = get_compute_capability(detected_gpu)
+    else:
+        compute_caps = {
+            "B200": "10.0",
+            "H100": "9.0",
+            "A100": "8.0",
+            "A10": "8.6",
+            "V100": "7.0",
+            "MI300X": "9.4",
+            "MI250X": "9.0",
+            "RTX 5090": "10.0",
+            "RTX 4090": "8.9",
+            "RTX 3090": "8.6",
+        }
+        compute_capability = compute_caps.get(gpu_type, "8.0")
 
     # Build target data
     target_data = {
@@ -2855,6 +3541,12 @@ def init_ssh(
     if docker_image:
         target_data["docker_image"] = docker_image
 
+    # Add torch requirements if detected
+    if torch_package:
+        target_data["torch_package"] = torch_package
+    if torch_index_url:
+        target_data["torch_index_url"] = torch_index_url
+
     try:
         target = save_target(target_data)
         typer.echo(f"✓ Created target: {target.name}")
@@ -2862,9 +3554,12 @@ def init_ssh(
         typer.echo(f"  Host: {host}")
         typer.echo(f"  GPU IDs: {parsed_gpu_ids}")
         typer.echo(f"  GPU Type: {gpu_type}")
+        typer.echo(f"  Compute: {compute_capability}")
         typer.echo(f"  NCU: {'Yes' if ncu else 'No'}")
         if docker_image:
             typer.echo(f"  Docker: {docker_image}")
+        if torch_package:
+            typer.echo(f"  Torch: {torch_package}")
         typer.echo("")
         typer.echo(
             f"Usage: wafer evaluate --target {name} --impl kernel.py --reference ref.py --test-cases tests.json"
@@ -2874,6 +3569,44 @@ def init_ssh(
         raise typer.Exit(1) from None
 
 
+def _extract_gpu_type(gpu_name: str) -> str:
+    """Extract GPU type from full GPU name.
+
+    Examples:
+        "NVIDIA H100 80GB HBM3" -> "H100"
+        "NVIDIA GeForce RTX 4090" -> "RTX 4090"
+        "AMD Instinct MI300X OAM" -> "MI300X"
+    """
+    gpu_name_upper = gpu_name.upper()
+
+    # Check for known GPU types
+    known_types = [
+        "B200",
+        "B100",
+        "H200",
+        "H100",
+        "A100",
+        "A10",
+        "V100",
+        "RTX 5090",
+        "RTX 5080",
+        "RTX 4090",
+        "RTX 4080",
+        "RTX 3090",
+        "RTX 3080",
+        "MI300X",
+        "MI250X",
+        "MI100",
+    ]
+
+    for gpu_type in known_types:
+        if gpu_type in gpu_name_upper:
+            return gpu_type
+
+    # Fallback: return cleaned name
+    return gpu_name.replace("NVIDIA ", "").replace("AMD ", "").strip()
+
+
 @targets_app.command("add")
 def targets_add(
     file_path: Path = typer.Argument(..., help="Path to target TOML file"),
@@ -2956,6 +3689,93 @@ def targets_show(
         raise typer.Exit(1) from None
 
 
+@targets_app.command("probe")
+def targets_probe(
+    name: str = typer.Argument(..., help="Target name"),
+) -> None:
+    """Probe a target to discover available compilation backends.
+
+    Connects to the target and checks what's available:
+    - Triton
+    - torch.compile/inductor
+    - HIP/hipcc or CUDA/nvcc
+    - ROCm or CUDA version
+    - Python packages (torch, triton, etc.)
+
+    Example:
+        wafer config targets probe runpod-mi300x
+    """
+    import trio
+
+    from .targets import ProbeError, load_target, probe_target_capabilities
+
+    try:
+        target = load_target(name)
+    except FileNotFoundError as e:
+        typer.echo(f"Error: {e}", err=True)
+        raise typer.Exit(1) from None
+
+    typer.echo(f"Probing target: {name}...")
+
+    try:
+        capabilities = trio.run(probe_target_capabilities, target)
+    except ProbeError as e:
+        # ProbeError already has actionable context
+        typer.echo(f"\nError: {e}", err=True)
+        raise typer.Exit(1) from None
+    except Exception as e:
+        # Unexpected errors - include type for debugging
+        typer.echo(f"\nUnexpected error probing target: {type(e).__name__}: {e}", err=True)
+        raise typer.Exit(1) from None
+
+    # Display results
+    typer.echo(f"\nTarget: {name}")
+
+    if capabilities.get("gpu_name"):
+        typer.echo(f"  GPU: {capabilities['gpu_name']}")
+    if capabilities.get("compute_capability"):
+        typer.echo(f"  Compute: {capabilities['compute_capability']}")
+
+    typer.echo("\n  Compilation Backends:")
+    backends = capabilities.get("backends", {})
+
+    # Triton
+    triton_ver = backends.get("triton")
+    if triton_ver:
+        typer.echo(f"    ✓ Triton: {triton_ver}")
+    else:
+        typer.echo("    ✗ Triton: not installed")
+
+    # torch.compile
+    if triton_ver and backends.get("torch"):
+        typer.echo("    ✓ torch.compile/inductor: available")
+    else:
+        typer.echo("    ✗ torch.compile/inductor: requires Triton")
+
+    # HIP/CUDA compiler
+    if backends.get("hipcc"):
+        typer.echo(f"    ✓ HIP/hipcc: {backends['hipcc']}")
+    elif backends.get("nvcc"):
+        typer.echo(f"    ✓ CUDA/nvcc: {backends['nvcc']}")
+    else:
+        typer.echo("    ✗ No GPU compiler found")
+
+    # ROCm/CUDA version
+    if capabilities.get("rocm_version"):
+        typer.echo(f"    ROCm: {capabilities['rocm_version']}")
+    if capabilities.get("cuda_version"):
+        typer.echo(f"    CUDA: {capabilities['cuda_version']}")
+
+    typer.echo("\n  Python Environment:")
+    typer.echo(f"    Python: {capabilities.get('python_version', 'unknown')}")
+
+    packages = capabilities.get("packages", {})
+    if packages.get("torch"):
+        typer.echo(f"    PyTorch: {packages['torch']}")
+    if triton_ver:
+        typer.echo(f"    Triton: {triton_ver}")
+
+
 @targets_app.command("remove")
 def targets_remove(
     name: str = typer.Argument(..., help="Target name"),
@@ -3086,6 +3906,92 @@ def targets_pods() -> None:
         typer.echo()
 
 
+# ── Pool commands ───────────────────────────────────────────────────────────
+
+
+@targets_app.command("pool-list")
+def targets_pool_list() -> None:
+    """List all configured target pools.
+
+    Example:
+        wafer config targets pool-list
+    """
+    from .targets import get_pool, list_pools
+
+    pools = list_pools()
+
+    if not pools:
+        typer.echo("No pools configured")
+        typer.echo("")
+        typer.echo("Define pools in ~/.wafer/config.toml:")
+        typer.echo("  [pools.my-pool]")
+        typer.echo('  targets = ["target-1", "target-2"]')
+        return
+
+    typer.echo("Configured pools:\n")
+    for pool_name in pools:
+        try:
+            targets = get_pool(pool_name)
+            typer.echo(f"  {pool_name}: {', '.join(targets)}")
+        except Exception as e:
+            typer.echo(f"  {pool_name}: (error: {e})")
+
+
+@targets_app.command("pool-create")
+def targets_pool_create(
+    name: str = typer.Argument(..., help="Pool name"),
+    targets: list[str] = typer.Argument(..., help="Target names to include in pool"),
+) -> None:
+    """Create or update a target pool.
+
+    Example:
+        wafer config targets pool-create mi300x-pool mi300x-1 mi300x-2 mi300x-3
+    """
+    from .targets import save_pool
+
+    try:
+        save_pool(name, targets)
+        typer.echo(f"Pool '{name}' created with {len(targets)} targets")
+    except FileNotFoundError as e:
+        typer.echo(f"Error: {e}", err=True)
+        raise typer.Exit(1) from None
+
+
+@targets_app.command("pool-status")
+def targets_pool_status(
+    name: str = typer.Argument(..., help="Pool name"),
+) -> None:
+    """Show status of targets in a pool (locked/available).
+
+    Example:
+        wafer config targets pool-status mi300x-pool
+    """
+    from .target_lock import get_lock_holder, is_target_locked
+    from .targets import get_pool
+
+    try:
+        targets = get_pool(name)
+    except FileNotFoundError as e:
+        typer.echo(f"Error: {e}", err=True)
+        raise typer.Exit(1) from None
+
+    typer.echo(f"Pool '{name}' ({len(targets)} targets):\n")
+
+    available = 0
+    for target_name in targets:
+        locked = is_target_locked(target_name)
+        if locked:
+            pid = get_lock_holder(target_name)
+            pid_str = f" (pid {pid})" if pid else ""
+            typer.echo(f"  [busy]  {target_name}{pid_str}")
+        else:
+            typer.echo(f"  [free]  {target_name}")
+            available += 1
+
+    typer.echo("")
+    typer.echo(f"Available: {available}/{len(targets)}")
+
+
 # =============================================================================
 # Billing commands
 # =============================================================================
@@ -3119,7 +4025,9 @@ def billing_usage(
 @billing_app.command("topup")
 def billing_topup(
     amount: int = typer.Argument(25, help="Amount in dollars ($10-$500)"),
-    no_browser: bool = typer.Option(False, "--no-browser", help="Print URL instead of opening browser"),
+    no_browser: bool = typer.Option(
+        False, "--no-browser", help="Print URL instead of opening browser"
+    ),
 ) -> None:
     """Add credits to your account.
 
@@ -3165,7 +4073,9 @@ def billing_topup(
 
 @billing_app.command("portal")
 def billing_portal(
-    no_browser: bool = typer.Option(False, "--no-browser", help="Print URL instead of opening browser"),
+    no_browser: bool = typer.Option(
+        False, "--no-browser", help="Print URL instead of opening browser"
+    ),
 ) -> None:
     """Open Stripe billing portal.
 
@@ -3319,7 +4229,7 @@ def workspaces_exec(
     workspace: str | None = typer.Argument(
         None, help="Workspace name or ID (optional if default set)"
     ),
-    command: list[str] = typer.Argument(..., help="Command to execute on GPU"),
+    command: list[str] = typer.Argument(..., help="Command to execute"),
     timeout: int | None = typer.Option(
         None,
         "--timeout",
@@ -3332,13 +4242,23 @@ def workspaces_exec(
         "-s",
         help="Sync local directory to workspace before executing",
     ),
+    gpu: bool = typer.Option(False, "--gpu", help="Force GPU routing (default behavior)"),
+    cpu: bool = typer.Option(False, "--cpu", help="Run in workspace container (no GPU)"),
+    baremetal: bool = typer.Option(
+        False, "--baremetal", help="Force baremetal target (for hardware counters like ncu/nsys)"
+    ),
     verbose: bool = typer.Option(False, "--verbose", "-v", help="Show [wafer] status messages"),
     quiet: bool = typer.Option(False, "--quiet", "-q", help="Suppress [wafer] status messages"),
 ) -> None:
-    """Execute a command in workspace with GPU routing.
+    """Execute a command in workspace.
+
+    By default, auto-detects whether to route to GPU based on the command.
+    Use --gpu, --cpu, or --baremetal to override.
 
-    Runs the command on the workspace's configured GPU target (Modal, baremetal, etc.)
-    and streams output back. No SSH or zsh plugin required.
+    Routing options:
+      --gpu       Force GPU container (Modal or baremetal with GPU)
+      --cpu       Run in workspace container directly (no GPU)
+      --baremetal Force baremetal target (for ncu, nsys, hardware counters)
 
     If workspace is not specified, uses the default workspace from config,
     or the only workspace if you have exactly one.
@@ -3353,6 +4273,21 @@ def workspaces_exec(
     from .global_config import get_defaults, get_preferences
     from .workspaces import exec_command, resolve_workspace, sync_files
 
+    # Validate mutually exclusive routing flags
+    routing_flags = sum([gpu, cpu, baremetal])
+    if routing_flags > 1:
+        typer.echo("Error: --gpu, --cpu, and --baremetal are mutually exclusive", err=True)
+        raise typer.Exit(1)
+
+    # Determine routing (None = auto-detect)
+    routing: str | None = None
+    if gpu:
+        routing = "gpu"
+    elif cpu:
+        routing = "cpu"
+    elif baremetal:
+        routing = "baremetal"
+
     # Resolve workspace (specified, config default, or single workspace)
     try:
         resolved_workspace = resolve_workspace(workspace)
@@ -3377,7 +4312,8 @@ def workspaces_exec(
         show_status = prefs.mode == "explicit"
 
     if show_status:
-        typer.echo(f"[wafer] Workspace: {resolved_workspace}", err=True)
+        routing_label = routing or "auto"
+        typer.echo(f"[wafer] Workspace: {resolved_workspace} (routing: {routing_label})", err=True)
 
     # Sync files if requested
     if sync is not None:
@@ -3413,8 +4349,15 @@ def workspaces_exec(
         # Remove leading "--" if present (typer passes it through with allow_interspersed_args=False)
         if command and command[0] == "--":
             command = command[1:]
-        # Use shlex.join to properly quote args containing spaces/special chars
-        command_str = shlex.join(command)
+        # Handle two cases:
+        # 1. Single element: user quoted the whole command (e.g., "echo hello world")
+        #    -> use directly, don't re-quote
+        # 2. Multiple elements: user passed separate args (e.g., -- python -c "print(1)")
+        #    -> use shlex.join to properly quote args with spaces
+        if len(command) == 1:
+            command_str = command[0]
+        else:
+            command_str = shlex.join(command)
     else:
         command_str = command
 
@@ -3423,6 +4366,7 @@ def workspaces_exec(
             workspace_id=resolved_workspace,
             command=command_str,
             timeout_seconds=effective_timeout,
+            routing=routing,
         )
     except RuntimeError as e:
         typer.echo(f"Error: {e}", err=True)
@@ -4441,8 +5385,8 @@ def _setup_wafer_core_env() -> None:
     - WAFER_API_URL: If already set, uses that instead of config
     - WAFER_AUTH_TOKEN: If already set, uses that instead of cached token
     """
-    from .global_config import get_api_url
     from .auth import get_valid_token
+    from .global_config import get_api_url
 
     # Set API URL (get_api_url already respects WAFER_API_URL env var)
     os.environ["WAFER_API_URL"] = get_api_url()
@@ -4746,8 +5690,8 @@ def capture_command(  # noqa: PLR0915
     import os
     import tomllib
 
-    from .global_config import get_api_url
     from .auth import get_valid_token
+    from .global_config import get_api_url
 
     # Set environment variables for wafer-core BEFORE importing it
     # wafer-core backend.py reads WAFER_API_URL and WAFER_AUTH_TOKEN from env
@@ -4951,8 +5895,8 @@ def capture_list_command(
     """
     import os
 
-    from .global_config import get_api_url
     from .auth import get_valid_token
+    from .global_config import get_api_url
 
     # Set environment variables for wafer-core BEFORE importing it
     os.environ["WAFER_API_URL"] = get_api_url()
@@ -5301,6 +6245,98 @@ def isa_analyze(
         raise typer.Exit(1) from None
 
 
+# =============================================================================
+# Kernel Scope Commands (wafer amd kernel-scope ...)
+# =============================================================================
+
+
+@kernel_scope_app.command("analyze")
+def kernel_scope_analyze(
+    path: Path = typer.Argument(..., help="Path to file or directory to analyze"),
+    json_output: bool = typer.Option(False, "--json", "-j", help="Output as JSON"),
+    csv_output: bool = typer.Option(False, "--csv", help="Output as CSV"),
+    recursive: bool = typer.Option(
+        True, "--recursive/--no-recursive", "-r", help="Scan directories recursively"
+    ),
+    filter_expr: str | None = typer.Option(
+        None, "--filter", "-f", help="Filter results (e.g., 'spills > 0')"
+    ),
+    output_file: Path | None = typer.Option(None, "--output", "-o", help="Write output to file"),
+    kernel_index: int = typer.Option(0, "--kernel", "-k", help="Kernel index if multiple in file"),
+) -> None:
+    """Analyze Triton compilation artifacts (ISA, LLVM-IR, TTGIR).
+
+    Performs static analysis to extract performance metrics like register
+    pressure, spills, MFMA density, and occupancy limits.
+
+    Supports:
+      - AMDGCN ISA files (.s, .gcn, .asm)
+      - LLVM-IR files (.ll)
+      - TTGIR files (.ttgir, .ttir, .mlir)
+
+    Examples:
+        wafer amd kernel-scope analyze kernel.s
+        wafer amd kernel-scope analyze kernel.s --json
+        wafer amd kernel-scope analyze ~/.triton/cache/ --filter 'spills > 0'
+        wafer amd kernel-scope analyze . -r --csv -o metrics.csv
+    """
+    from .kernel_scope import analyze_command
+
+    try:
+        output = analyze_command(
+            path=str(path),
+            json_output=json_output,
+            csv_output=csv_output,
+            recursive=recursive,
+            filter_expr=filter_expr,
+            output_file=str(output_file) if output_file else None,
+            kernel_index=kernel_index,
+        )
+        typer.echo(output)
+
+    except FileNotFoundError as e:
+        typer.echo(f"Error: {e}", err=True)
+        raise typer.Exit(1) from None
+    except RuntimeError as e:
+        typer.echo(f"Error: {e}", err=True)
+        raise typer.Exit(1) from None
+    except Exception as e:
+        typer.echo(f"Error: {e}", err=True)
+        raise typer.Exit(1) from None
+
+
+@kernel_scope_app.command("metrics")
+def kernel_scope_metrics() -> None:
+    """List available metrics for kernel scope analysis.
+
+    Shows all metrics that can be extracted from Triton compilation
+    artifacts, along with their derivation.
+
+    Examples:
+        wafer amd kernel-scope metrics
+    """
+    from .kernel_scope import metrics_command
+
+    output = metrics_command()
+    typer.echo(output)
+
+
+@kernel_scope_app.command("targets")
+def kernel_scope_targets() -> None:
+    """List supported GPU targets and their specifications.
+
+    Shows hardware specs (VGPRs, SGPRs, LDS, etc.) for each supported
+    AMD GPU architecture.
+
+    Examples:
+        wafer amd kernel-scope targets
+    """
+    from .kernel_scope import targets_command
+
+    output = targets_command()
+    typer.echo(output)
+
+
 def main() -> None:
     """Entry point for wafer CLI."""
     app()