diff --git a/src/cmdforge/cli/__init__.py b/src/cmdforge/cli/__init__.py
index 67a0ff2..5cda11c 100644
--- a/src/cmdforge/cli/__init__.py
+++ b/src/cmdforge/cli/__init__.py
@@ -62,12 +62,14 @@ def main():
     p_test.set_defaults(func=cmd_test)
 
     # 'run' command
-    p_run = subparsers.add_parser("run", help="Run a tool")
+    # NOTE: Options like --provider must come BEFORE the tool name due to argparse.REMAINDER
+    # Example: cmdforge run --provider mock my-tool (not: cmdforge run my-tool --provider mock)
+    p_run = subparsers.add_parser("run", help="Run a tool (options must come before tool name)")
     p_run.add_argument("name", help="Tool name")
     p_run.add_argument("-i", "--input", help="Input file (reads from stdin if piped)")
     p_run.add_argument("-o", "--output", help="Output file (writes to stdout if omitted)")
     p_run.add_argument("--stdin", action="store_true", help="Read input interactively (type then Ctrl+D)")
-    p_run.add_argument("-p", "--provider", help="Override provider")
+    p_run.add_argument("-p", "--provider", help="Override provider (must come before tool name)")
     p_run.add_argument("--dry-run", action="store_true", help="Show what would happen without executing")
     p_run.add_argument("--show-prompt", action="store_true", help="Show prompts in addition to output")
     p_run.add_argument("-v", "--verbose", action="store_true", help="Show debug information")
@@ -126,6 +128,12 @@ def main():
     p_prov_test.add_argument("name", help="Provider name")
     p_prov_test.set_defaults(func=cmd_providers)
 
+    # providers for-tools
+    p_prov_for_tools = providers_sub.add_parser("for-tools", help="List providers used by specified tools")
+    p_prov_for_tools.add_argument("tools", nargs="+", help="Tool names to check")
+    p_prov_for_tools.add_argument("--warm", action="store_true", help="Warm up the providers (send minimal prompt)")
+    p_prov_for_tools.set_defaults(func=cmd_providers)
+
     # Default for providers with no subcommand
     p_providers.set_defaults(func=lambda args: cmd_providers(args) if args.providers_cmd else (setattr(args, 'providers_cmd', 'list') or cmd_providers(args)))
 
diff --git a/src/cmdforge/cli/provider_commands.py b/src/cmdforge/cli/provider_commands.py
index 588da33..e40be32 100644
--- a/src/cmdforge/cli/provider_commands.py
+++ b/src/cmdforge/cli/provider_commands.py
@@ -67,6 +67,8 @@ def cmd_providers(args):
         return _cmd_providers_test(args)
     elif args.providers_cmd == "check":
         return _cmd_providers_check(args)
+    elif args.providers_cmd == "for-tools":
+        return _cmd_providers_for_tools(args)
     return 0
 
 
@@ -269,6 +271,117 @@ def _cmd_providers_test(args):
     return 0 if result.success else 1
 
 
+def _cmd_providers_for_tools(args):
+    """List (and optionally warm) providers used by specified tools."""
+    import sys
+    import re
+    from ..tool import load_tool
+    from ..resolver import resolve_tool, ToolNotFoundError
+
+    # Cloud providers don't need warming - they're always "warm"
+    CLOUD_PREFIXES = ('claude', 'opencode', 'gemini', 'codex', 'gpt', 'openai')
+
+    def is_local_provider(name):
+        """Check if provider is local (needs warming) vs cloud."""
+        name_lower = name.lower()
+        return not any(name_lower.startswith(prefix) for prefix in CLOUD_PREFIXES)
+
+    def estimate_model_size(provider_name):
+        """Estimate model size in billions from provider name/command for sorting.
+
+        Larger models should be warmed first to claim contiguous GPU memory.
+        Returns estimated size in billions (e.g., 70 for 70b model).
+        """
+        # Check provider command for model size hints
+        provider = get_provider(provider_name)
+        text_to_check = f"{provider_name} {provider.command if provider else ''}"
+
+        # Look for patterns like "70b", "14b", "7b", "3b", "1.5b"
+        match = re.search(r'(\d+(?:\.\d+)?)\s*[bB]', text_to_check)
+        if match:
+            return float(match.group(1))
+
+        # Known large model keywords
+        if any(kw in text_to_check.lower() for kw in ['opus', 'large', 'big', '32b', '70b']):
+            return 70.0
+        if any(kw in text_to_check.lower() for kw in ['medium', '14b', '13b']):
+            return 14.0
+        if any(kw in text_to_check.lower() for kw in ['small', 'fast', '3b', '1b', 'tiny', 'mini']):
+            return 3.0
+
+        # Default to medium size
+        return 7.0
+
+    def get_provider(name):
+        """Get provider by name."""
+        providers = load_providers()
+        for p in providers:
+            if p.name == name:
+                return p
+        return None
+
+    tool_names = args.tools
+    providers_used = set()
+
+    for tool_name in tool_names:
+        try:
+            resolved = resolve_tool(tool_name)
+            tool = resolved.tool
+        except ToolNotFoundError:
+            print(f"Warning: Tool '{tool_name}' not found, skipping", file=sys.stderr)
+            continue
+
+        # Extract providers from prompt steps
+        for step in tool.steps:
+            if hasattr(step, 'provider') and step.provider:
+                providers_used.add(step.provider)
+
+        # Also check tool steps for nested tools (recursive would be complex, just note dependency)
+        for step in tool.steps:
+            if hasattr(step, 'tool') and step.tool:
+                # Could recurse here, but for now just note it
+                pass
+
+    # Sort for consistent output
+    providers_list = sorted(providers_used)
+
+    if args.warm:
+        # Only warm local providers (Ollama, etc.)
+        local_providers = [p for p in providers_list if is_local_provider(p)]
+        cloud_providers = [p for p in providers_list if not is_local_provider(p)]
+
+        if cloud_providers:
+            print(f"Skipping {len(cloud_providers)} cloud provider(s): {', '.join(cloud_providers)}", file=sys.stderr)
+
+        if local_providers:
+            # Sort by model size DESCENDING - warm largest models first
+            # This lets large models claim contiguous GPU memory before small models fragment it
+            local_providers_sorted = sorted(
+                local_providers,
+                key=lambda p: estimate_model_size(p),
+                reverse=True
+            )
+
+            print(f"Warming {len(local_providers_sorted)} local provider(s) (largest first)...", file=sys.stderr)
+            for provider_name in local_providers_sorted:
+                size_est = estimate_model_size(provider_name)
+                try:
+                    result = call_provider(provider_name, "Hi", timeout=120)
+                    if result.success:
+                        print(f"  [+] {provider_name} (~{size_est:.0f}B): warm", file=sys.stderr)
+                    else:
+                        print(f"  [-] {provider_name} (~{size_est:.0f}B): failed ({result.error[:50]}...)", file=sys.stderr)
+                except Exception as e:
+                    print(f"  [-] {provider_name} (~{size_est:.0f}B): error ({e})", file=sys.stderr)
+        print(file=sys.stderr)
+
+    # Output provider names (one per line for easy parsing)
+    for p in providers_list:
+        print(p)
+
+    return 0
+
+
 def _cmd_providers_check(args):
     """Check which providers are available."""
     providers = load_providers()
diff --git a/src/cmdforge/providers.py b/src/cmdforge/providers.py
index e251ba1..f3c7a7b 100644
--- a/src/cmdforge/providers.py
+++ b/src/cmdforge/providers.py
@@ -1,6 +1,7 @@
 """Provider abstraction for AI CLI tools."""
 
 import os
+import re
 import shlex
 import subprocess
 import shutil
@@ -11,6 +12,16 @@ from typing import Optional, List
 import yaml
 
 
+# Regex to strip ANSI escape codes (spinners, colors, cursor movements)
+# Handles: CSI sequences (\x1b[...X), including private modes (\x1b[?...X), OSC sequences, and \r
+ANSI_ESCAPE_RE = re.compile(r'\x1b\[[\x20-\x3f]*[\x40-\x7e]|\x1b\].*?\x07|\r')
+
+
+def strip_ansi(text: str) -> str:
+    """Strip ANSI escape codes from text (e.g., Ollama spinner output)."""
+    return ANSI_ESCAPE_RE.sub('', text)
+
+
 # Default providers config location
 PROVIDERS_FILE = Path.home() / ".cmdforge" / "providers.yaml"
 
@@ -144,7 +155,7 @@ def delete_provider(name: str) -> bool:
     return False
 
 
-def call_provider(provider_name: str, prompt: str, timeout: int = 300, _tried: Optional[set] = None) -> ProviderResult:
+def call_provider(provider_name: str, prompt: str, timeout: int = 300, max_tokens: Optional[int] = None, _tried: Optional[set] = None) -> ProviderResult:
     """
     Call an AI provider with the given prompt.
 
@@ -152,6 +163,7 @@ def call_provider(provider_name: str, prompt: str, timeout: int = 300, _tried: O
         provider_name: Name of the provider to use
         prompt: The prompt to send
         timeout: Maximum execution time in seconds
+        max_tokens: Optional max output tokens (appends provider-specific flag)
         _tried: Internal set of already-tried providers (prevents infinite loops)
 
     Returns:
@@ -178,6 +190,15 @@ def call_provider(provider_name: str, prompt: str, timeout: int = 300, _tried: O
     # Parse command (expand environment variables)
     cmd = os.path.expandvars(provider.command)
 
+    # Append max_tokens flag for known providers
+    if max_tokens:
+        name_lower = provider_name.lower()
+        if name_lower.startswith("claude") or "claude" in cmd.lower():
+            cmd = f"{cmd} --max-tokens {max_tokens}"
+        elif name_lower.startswith("gemini") or "gemini" in cmd.lower():
+            cmd = f"{cmd} --max-output-tokens {max_tokens}"
+        # opencode and codex use default model limits, no flag available
+
     # Check if base command exists (use shlex for proper quote handling)
     try:
         cmd_parts = shlex.split(cmd)
@@ -191,7 +212,7 @@ def call_provider(provider_name: str, prompt: str, timeout: int = 300, _tried: O
         if provider.fallback and provider.fallback not in _tried:
             import sys
             print(f"[fallback] {provider_name} failed, trying {provider.fallback}...", file=sys.stderr)
-            return call_provider(provider.fallback, prompt, timeout, _tried)
+            return call_provider(provider.fallback, prompt, timeout, max_tokens, _tried)
         return ProviderResult(text="", success=False, error=error_msg)
 
     # Expand ~ for the which check
@@ -212,19 +233,20 @@ def call_provider(provider_name: str, prompt: str, timeout: int = 300, _tried: O
         )
 
         if result.returncode != 0:
-            error_msg = f"Provider exited with code {result.returncode}: {result.stderr}"
-            if "not found" in result.stderr.lower() or "not installed" in result.stderr.lower():
+            stderr_clean = strip_ansi(result.stderr)
+            error_msg = f"Provider exited with code {result.returncode}: {stderr_clean}"
+            if "not found" in stderr_clean.lower() or "not installed" in stderr_clean.lower():
                 error_msg += "\n\nTo install AI providers, run: cmdforge providers install"
             return try_fallback(error_msg)
 
         # Warn if output is empty (provider ran but returned nothing)
-        if not result.stdout.strip():
-            stderr = result.stderr.strip()
+        clean_stdout = strip_ansi(result.stdout)
+        if not clean_stdout.strip():
+            stderr = strip_ansi(result.stderr).strip()
 
             # Check for OpenCode's ProviderModelNotFoundError
             if "ProviderModelNotFoundError" in stderr or "ModelNotFoundError" in stderr:
                 # Extract provider and model info if possible
-                import re
                 provider_match = re.search(r'providerID:\s*"([^"]+)"', stderr)
                 model_match = re.search(r'modelID:\s*"([^"]+)"', stderr)
                 provider_id = provider_match.group(1) if provider_match else "unknown"
@@ -243,7 +265,7 @@ def call_provider(provider_name: str, prompt: str, timeout: int = 300, _tried: O
                 f"Provider returned empty output{stderr_hint}.\n\nThis may mean the model is not available. Try a different provider or run: cmdforge providers install"
             )
 
-        return ProviderResult(text=result.stdout, success=True)
+        return ProviderResult(text=clean_stdout, success=True)
 
     except subprocess.TimeoutExpired:
         return try_fallback(f"Provider timed out after {timeout} seconds")
diff --git a/src/cmdforge/runner.py b/src/cmdforge/runner.py
index 8b70109..b7a9ea0 100644
--- a/src/cmdforge/runner.py
+++ b/src/cmdforge/runner.py
@@ -546,7 +546,7 @@ def execute_prompt_step(
         if provider.lower() == "mock":
             result = mock_provider(prompt)
         else:
-            result = call_provider(provider, prompt)
+            result = call_provider(provider, prompt, max_tokens=step.max_tokens)
 
         if not result.success:
             print(f"Error in prompt step: {result.error}", file=sys.stderr)
@@ -590,7 +590,7 @@ Please try again with valid JSON matching the schema exactly."""
         if provider.lower() == "mock":
             result = mock_provider(current_prompt)
         else:
-            result = call_provider(provider, current_prompt)
+            result = call_provider(provider, current_prompt, max_tokens=step.max_tokens)
 
         if not result.success:
             print(f"Error in prompt step: {result.error}", file=sys.stderr)
diff --git a/src/cmdforge/tool.py b/src/cmdforge/tool.py
index 8f3d4ec..3abfd61 100644
--- a/src/cmdforge/tool.py
+++ b/src/cmdforge/tool.py
@@ -105,6 +105,7 @@ class PromptStep:
     output_schema: Optional[dict] = None  # JSON schema for validation (None = use default)
     max_retries: int = 1  # Retry count on validation failure
     plain_text: bool = False  # Bypass structured output enforcement
+    max_tokens: Optional[int] = None  # Max output tokens (provider-dependent)
 
     def to_dict(self) -> dict:
         d = {
@@ -127,6 +128,8 @@ class PromptStep:
             d["max_retries"] = self.max_retries
         if self.plain_text:
             d["plain_text"] = self.plain_text
+        if self.max_tokens:
+            d["max_tokens"] = self.max_tokens
         return d
 
     @classmethod
@@ -141,7 +144,8 @@ class PromptStep:
             strip_fences=data.get("strip_fences", False),
             output_schema=data.get("output_schema"),
             max_retries=data.get("max_retries", 1),
-            plain_text=data.get("plain_text", False)
+            plain_text=data.get("plain_text", False),
+            max_tokens=data.get("max_tokens")
         )
 
 
diff --git a/tests/test_providers.py b/tests/test_providers.py
index 3db3a5a..5882d3b 100644
--- a/tests/test_providers.py
+++ b/tests/test_providers.py
@@ -12,7 +12,7 @@ from cmdforge.providers import (
     load_providers, save_providers, get_provider,
     add_provider, delete_provider,
     call_provider, mock_provider,
-    DEFAULT_PROVIDERS
+    DEFAULT_PROVIDERS, strip_ansi
 )
 
 
@@ -507,3 +507,122 @@ class TestProviderFallback:
 
         assert result.success is True
         assert "[MOCK]" in result.text
+
+
+class TestStripAnsi:
+    """Tests for ANSI escape code stripping."""
+
+    def test_strip_basic_color_codes(self):
+        """Strip basic color codes like \\x1b[0m."""
+        text = "\x1b[31mred text\x1b[0m"
+        assert strip_ansi(text) == "red text"
+
+    def test_strip_cursor_movements(self):
+        """Strip cursor movement codes."""
+        text = "\x1b[2Jhello\x1b[H"
+        assert strip_ansi(text) == "hello"
+
+    def test_strip_private_mode_sequences(self):
+        """Strip private mode sequences like \\x1b[?25l (hide cursor)."""
+        text = "\x1b[?2026h\x1b[?25lspinner\x1b[?25h\x1b[?2026l"
+        assert strip_ansi(text) == "spinner"
+
+    def test_strip_ollama_spinner(self):
+        """Strip Ollama spinner output."""
+        text = "\x1b[?2026h\x1b[?25l⠋ \x1b[?25h\x1b[?2026l\x1b[?2026h\x1b[?25l⠙ \x1b[?25h"
+        result = strip_ansi(text)
+        # Should only have spinner characters left, not escape codes
+        assert "\x1b" not in result
+        assert "⠋" in result or "⠙" in result
+
+    def test_strip_carriage_return(self):
+        """Strip carriage returns."""
+        text = "line1\rline2"
+        assert strip_ansi(text) == "line1line2"
+
+    def test_preserve_normal_text(self):
+        """Normal text should be preserved."""
+        text = "Hello, World! 123 [test] {data}"
+        assert strip_ansi(text) == text
+
+    def test_strip_osc_sequences(self):
+        """Strip OSC (Operating System Command) sequences."""
+        text = "\x1b]0;Window Title\x07content"
+        assert strip_ansi(text) == "content"
+
+
+class TestMaxTokens:
+    """Tests for max_tokens parameter."""
+
+    @pytest.fixture
+    def temp_providers_file(self, tmp_path):
+        providers_file = tmp_path / ".cmdforge" / "providers.yaml"
+        with patch('cmdforge.providers.PROVIDERS_FILE', providers_file):
+            yield providers_file
+
+    @patch('subprocess.run')
+    @patch('shutil.which')
+    def test_max_tokens_appended_for_claude(self, mock_which, mock_run, temp_providers_file):
+        """max_tokens should append --max-tokens for claude providers."""
+        mock_which.return_value = "/usr/bin/claude"
+        mock_run.return_value = MagicMock(returncode=0, stdout="response", stderr="")
+        save_providers([Provider("claude-haiku", "claude -p --model haiku")])
+
+        call_provider("claude-haiku", "Test prompt", max_tokens=4096)
+
+        # Verify the command included --max-tokens
+        call_args = mock_run.call_args
+        cmd = call_args[0][0]
+        assert "--max-tokens 4096" in cmd
+
+    @patch('subprocess.run')
+    @patch('shutil.which')
+    def test_max_tokens_appended_for_gemini(self, mock_which, mock_run, temp_providers_file):
+        """max_tokens should append --max-output-tokens for gemini providers."""
+        mock_which.return_value = "/usr/bin/gemini"
+        mock_run.return_value = MagicMock(returncode=0, stdout="response", stderr="")
+        save_providers([Provider("gemini", "gemini --model gemini-2.5-pro")])
+
+        call_provider("gemini", "Test prompt", max_tokens=8192)
+
+        call_args = mock_run.call_args
+        cmd = call_args[0][0]
+        assert "--max-output-tokens 8192" in cmd
+
+    @patch('subprocess.run')
+    @patch('shutil.which')
+    def test_max_tokens_not_appended_when_none(self, mock_which, mock_run, temp_providers_file):
+        """No flag should be added when max_tokens is None."""
+        mock_which.return_value = "/usr/bin/claude"
+        mock_run.return_value = MagicMock(returncode=0, stdout="response", stderr="")
+        save_providers([Provider("claude", "claude -p")])
+
+        call_provider("claude", "Test prompt", max_tokens=None)
+
+        call_args = mock_run.call_args
+        cmd = call_args[0][0]
+        assert "--max-tokens" not in cmd
+
+    @patch('subprocess.run')
+    @patch('shutil.which')
+    def test_max_tokens_passed_through_fallback(self, mock_which, mock_run, temp_providers_file):
+        """max_tokens should be passed to fallback provider."""
+        mock_which.return_value = "/usr/bin/claude"
+        # First call fails, second succeeds
+        mock_run.side_effect = [
+            MagicMock(returncode=1, stdout="", stderr="Error"),
+            MagicMock(returncode=0, stdout="fallback response", stderr="")
+        ]
+        save_providers([
+            Provider("primary", "claude -p --model opus", fallback="fallback"),
+            Provider("fallback", "claude -p --model haiku")
+        ])
+
+        result = call_provider("primary", "Test prompt", max_tokens=4096)
+
+        assert result.success is True
+        # Both calls should have had max_tokens
+        calls = mock_run.call_args_list
+        assert "--max-tokens 4096" in calls[0][0][0]
+        assert "--max-tokens 4096" in calls[1][0][0]
+