claude-code/plans/gleaming-routing-mercury-impl.md

External LLM Integration Implementation Plan

For Claude: REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.

Goal: Enable agents to use external LLMs (Copilot, Z.AI, Gemini) via CLI tools with a session toggle.

Architecture: Python router reads model-policy.json, invokes appropriate CLI (opencode/gemini), returns response. State file controls Claude vs external mode. PA exposes toggle commands.

Tech Stack: Python 3, subprocess, JSON state files, opencode CLI, gemini CLI

---

Task 1: Create External Mode State File

Files:

• Create: ~/.claude/state/external-mode.json

Step 1: Create state file

cat > ~/.claude/state/external-mode.json << 'EOF'
{
  "enabled": false,
  "activated_at": null,
  "reason": null
}
EOF

Step 2: Verify file

Run: cat ~/.claude/state/external-mode.json | jq . Expected: Valid JSON with enabled: false

Step 3: Commit

git -C ~/.claude add state/external-mode.json
git -C ~/.claude commit -m "feat(external-llm): add external-mode state file"

---

Task 2: Extend Model Policy

Files:

• Modify: ~/.claude/state/model-policy.json

Step 1: Read current file

Run: cat ~/.claude/state/model-policy.json | jq .

Step 2: Add external_models section

Add to model-policy.json (after skill_delegation section):

"external_models": {
  "copilot/gpt-5.2": {
    "cli": "opencode",
    "cli_args": ["--provider", "copilot", "--model", "gpt-5.2"],
    "use_cases": ["reasoning", "fallback"],
    "tier": "opus-equivalent"
  },
  "copilot/sonnet-4.5": {
    "cli": "opencode",
    "cli_args": ["--provider", "copilot", "--model", "sonnet-4.5"],
    "use_cases": ["general", "fallback"],
    "tier": "sonnet-equivalent"
  },
  "copilot/haiku-4.5": {
    "cli": "opencode",
    "cli_args": ["--provider", "copilot", "--model", "haiku-4.5"],
    "use_cases": ["simple"],
    "tier": "haiku-equivalent"
  },
  "zai/glm-4.7": {
    "cli": "opencode",
    "cli_args": ["--provider", "zai", "--model", "glm-4.7"],
    "use_cases": ["code-generation"],
    "tier": "sonnet-equivalent"
  },
  "gemini/gemini-3-pro": {
    "cli": "gemini",
    "cli_args": ["-m", "gemini-3-pro"],
    "use_cases": ["long-context"],
    "tier": "opus-equivalent"
  }
},
"claude_to_external_map": {
  "opus": "copilot/gpt-5.2",
  "sonnet": "copilot/sonnet-4.5",
  "haiku": "copilot/haiku-4.5"
},
"task_routing": {
  "reasoning": "copilot/gpt-5.2",
  "code-generation": "zai/glm-4.7",
  "long-context": "gemini/gemini-3-pro",
  "default": "copilot/sonnet-4.5"
}

Step 3: Validate JSON

Run: cat ~/.claude/state/model-policy.json | jq . Expected: Valid JSON, no errors

Step 4: Commit

git -C ~/.claude add state/model-policy.json
git -C ~/.claude commit -m "feat(external-llm): add external model definitions to policy"

---

Task 3: Create Router Directory Structure

Files:

• Create: ~/.claude/mcp/llm-router/
• Create: ~/.claude/mcp/llm-router/providers/
• Create: ~/.claude/mcp/llm-router/providers/__init__.py

Step 1: Create directories

mkdir -p ~/.claude/mcp/llm-router/providers

Step 2: Create init.py

touch ~/.claude/mcp/llm-router/providers/__init__.py

Step 3: Verify structure

Run: ls -la ~/.claude/mcp/llm-router/ Expected: providers/ directory exists

Step 4: Commit

git -C ~/.claude add mcp/llm-router/
git -C ~/.claude commit -m "feat(external-llm): create llm-router directory structure"

---

Task 4: Create OpenCode Provider

Files:

• Create: ~/.claude/mcp/llm-router/providers/opencode.py

Step 1: Write provider

#!/usr/bin/env python3
"""OpenCode CLI wrapper for Copilot, Z.AI, and other providers."""

import subprocess
from typing import List


def invoke(cli_args: List[str], prompt: str, timeout: int = 300) -> str:
    """
    Invoke opencode CLI with given args and prompt.

    Args:
        cli_args: Provider/model args like ["--provider", "copilot", "--model", "gpt-5.2"]
        prompt: The prompt text
        timeout: Timeout in seconds (default 5 minutes)

    Returns:
        Model response as string

    Raises:
        RuntimeError: If opencode CLI fails
        TimeoutError: If request exceeds timeout
    """
    cmd = ["opencode", "--print"] + cli_args + ["-p", prompt]

    try:
        result = subprocess.run(
            cmd,
            capture_output=True,
            text=True,
            timeout=timeout
        )
    except subprocess.TimeoutExpired:
        raise TimeoutError(f"opencode timed out after {timeout}s")

    if result.returncode != 0:
        raise RuntimeError(f"opencode failed (exit {result.returncode}): {result.stderr}")

    return result.stdout.strip()


if __name__ == "__main__":
    # Quick test
    import sys
    if len(sys.argv) > 1:
        response = invoke(["--provider", "copilot", "--model", "gpt-5.2"], sys.argv[1])
        print(response)
    else:
        print("Usage: opencode.py 'prompt'")

Step 2: Make executable

chmod +x ~/.claude/mcp/llm-router/providers/opencode.py

Step 3: Verify syntax

Run: python3 -m py_compile ~/.claude/mcp/llm-router/providers/opencode.py Expected: No output (success)

Step 4: Commit

git -C ~/.claude add mcp/llm-router/providers/opencode.py
git -C ~/.claude commit -m "feat(external-llm): add opencode provider wrapper"

---

Task 5: Create Gemini Provider

Files:

• Create: ~/.claude/mcp/llm-router/providers/gemini.py

Step 1: Write provider

#!/usr/bin/env python3
"""Gemini CLI wrapper for Google models."""

import subprocess
from typing import List


def invoke(cli_args: List[str], prompt: str, timeout: int = 300) -> str:
    """
    Invoke gemini CLI with given args and prompt.

    Args:
        cli_args: Model args like ["-m", "gemini-3-pro"]
        prompt: The prompt text
        timeout: Timeout in seconds (default 5 minutes)

    Returns:
        Model response as string

    Raises:
        RuntimeError: If gemini CLI fails
        TimeoutError: If request exceeds timeout
    """
    cmd = ["gemini"] + cli_args + ["-p", prompt]

    try:
        result = subprocess.run(
            cmd,
            capture_output=True,
            text=True,
            timeout=timeout
        )
    except subprocess.TimeoutExpired:
        raise TimeoutError(f"gemini timed out after {timeout}s")

    if result.returncode != 0:
        raise RuntimeError(f"gemini failed (exit {result.returncode}): {result.stderr}")

    return result.stdout.strip()


if __name__ == "__main__":
    # Quick test
    import sys
    if len(sys.argv) > 1:
        response = invoke(["-m", "gemini-3-pro"], sys.argv[1])
        print(response)
    else:
        print("Usage: gemini.py 'prompt'")

Step 2: Make executable

chmod +x ~/.claude/mcp/llm-router/providers/gemini.py

Step 3: Verify syntax

Run: python3 -m py_compile ~/.claude/mcp/llm-router/providers/gemini.py Expected: No output (success)

Step 4: Commit

git -C ~/.claude add mcp/llm-router/providers/gemini.py
git -C ~/.claude commit -m "feat(external-llm): add gemini provider wrapper"

---

Task 6: Create Main Router (invoke.py)

Files:

• Create: ~/.claude/mcp/llm-router/invoke.py

Step 1: Write router

#!/usr/bin/env python3
"""
Invoke external LLM via configured provider.

Usage:
    invoke.py --model copilot/gpt-5.2 -p "prompt"
    invoke.py --task reasoning -p "prompt"
    invoke.py --task code-generation -p "prompt" --json

Model selection priority:
    1. Explicit --model flag
    2. Task-based routing (--task flag)
    3. Default from policy
"""

import argparse
import json
import sys
from pathlib import Path

STATE_DIR = Path.home() / ".claude/state"
ROUTER_DIR = Path(__file__).parent


def load_policy() -> dict:
    """Load model policy from state file."""
    policy_file = STATE_DIR / "model-policy.json"
    with open(policy_file) as f:
        return json.load(f)


def resolve_model(args: argparse.Namespace, policy: dict) -> str:
    """Determine which model to use based on args and policy."""
    if args.model:
        return args.model
    if args.task and args.task in policy.get("task_routing", {}):
        return policy["task_routing"][args.task]
    return policy.get("task_routing", {}).get("default", "copilot/sonnet-4.5")


def invoke(model: str, prompt: str, policy: dict) -> str:
    """Invoke the appropriate provider for the given model."""
    external_models = policy.get("external_models", {})

    if model not in external_models:
        raise ValueError(f"Unknown model: {model}. Available: {list(external_models.keys())}")

    model_config = external_models[model]
    cli = model_config["cli"]
    cli_args = model_config.get("cli_args", [])

    # Import and invoke appropriate provider
    if cli == "opencode":
        sys.path.insert(0, str(ROUTER_DIR))
        from providers.opencode import invoke as opencode_invoke
        return opencode_invoke(cli_args, prompt)
    elif cli == "gemini":
        sys.path.insert(0, str(ROUTER_DIR))
        from providers.gemini import invoke as gemini_invoke
        return gemini_invoke(cli_args, prompt)
    else:
        raise ValueError(f"Unknown CLI: {cli}")


def main():
    parser = argparse.ArgumentParser(
        description="Invoke external LLM via configured provider"
    )
    parser.add_argument(
        "-p", "--prompt",
        required=True,
        help="Prompt text"
    )
    parser.add_argument(
        "--model",
        help="Explicit model (e.g., copilot/gpt-5.2)"
    )
    parser.add_argument(
        "--task",
        choices=["reasoning", "code-generation", "long-context", "general"],
        help="Task type for automatic model routing"
    )
    parser.add_argument(
        "--json",
        action="store_true",
        help="Output as JSON with model info"
    )
    parser.add_argument(
        "--timeout",
        type=int,
        default=300,
        help="Timeout in seconds (default: 300)"
    )

    args = parser.parse_args()

    try:
        policy = load_policy()
        model = resolve_model(args, policy)
        result = invoke(model, args.prompt, policy)

        if args.json:
            output = {
                "model": model,
                "response": result,
                "success": True
            }
            print(json.dumps(output, indent=2))
        else:
            print(result)

    except Exception as e:
        if args.json:
            output = {
                "model": args.model or "unknown",
                "error": str(e),
                "success": False
            }
            print(json.dumps(output, indent=2))
            sys.exit(1)
        else:
            print(f"Error: {e}", file=sys.stderr)
            sys.exit(1)


if __name__ == "__main__":
    main()

Step 2: Make executable

chmod +x ~/.claude/mcp/llm-router/invoke.py

Step 3: Verify syntax

Run: python3 -m py_compile ~/.claude/mcp/llm-router/invoke.py Expected: No output (success)

Step 4: Commit

git -C ~/.claude add mcp/llm-router/invoke.py
git -C ~/.claude commit -m "feat(external-llm): add main router invoke.py"

---

Task 7: Create Delegation Helper

Files:

• Create: ~/.claude/mcp/llm-router/delegate.py

Step 1: Write delegation helper

#!/usr/bin/env python3
"""
Agent delegation helper. Routes to external or Claude based on mode.

Usage:
    delegate.py --tier sonnet -p "prompt"
    delegate.py --tier opus -p "complex reasoning task" --json
"""

import argparse
import json
import subprocess
import sys
from pathlib import Path

STATE_DIR = Path.home() / ".claude/state"
ROUTER_DIR = Path(__file__).parent


def is_external_mode() -> bool:
    """Check if external-only mode is enabled."""
    mode_file = STATE_DIR / "external-mode.json"
    if mode_file.exists():
        with open(mode_file) as f:
            data = json.load(f)
            return data.get("enabled", False)
    return False


def get_external_model(tier: str) -> str:
    """Get the external model equivalent for a Claude tier."""
    policy_file = STATE_DIR / "model-policy.json"
    with open(policy_file) as f:
        policy = json.load(f)
    mapping = policy.get("claude_to_external_map", {})
    if tier not in mapping:
        raise ValueError(f"No external mapping for tier: {tier}")
    return mapping[tier]


def delegate(tier: str, prompt: str, use_json: bool = False) -> str:
    """
    Delegate to appropriate model based on mode.

    Args:
        tier: Claude tier (opus, sonnet, haiku)
        prompt: The prompt text
        use_json: Return JSON output

    Returns:
        Model response as string
    """
    if is_external_mode():
        # Use external model
        model = get_external_model(tier)
        invoke_script = ROUTER_DIR / "invoke.py"

        cmd = [sys.executable, str(invoke_script), "--model", model, "-p", prompt]
        if use_json:
            cmd.append("--json")

        result = subprocess.run(cmd, capture_output=True, text=True)

        if result.returncode != 0:
            raise RuntimeError(f"External invoke failed: {result.stderr}")

        return result.stdout.strip()
    else:
        # Use Claude
        cmd = ["claude", "--print", "--model", tier, prompt]

        result = subprocess.run(cmd, capture_output=True, text=True)

        if result.returncode != 0:
            raise RuntimeError(f"Claude failed: {result.stderr}")

        response = result.stdout.strip()

        if use_json:
            return json.dumps({
                "model": f"claude/{tier}",
                "response": response,
                "success": True
            }, indent=2)

        return response


def main():
    parser = argparse.ArgumentParser(
        description="Delegate to Claude or external model based on mode"
    )
    parser.add_argument(
        "--tier",
        required=True,
        choices=["opus", "sonnet", "haiku"],
        help="Claude tier (maps to external equivalent when in external mode)"
    )
    parser.add_argument(
        "-p", "--prompt",
        required=True,
        help="Prompt text"
    )
    parser.add_argument(
        "--json",
        action="store_true",
        help="Output as JSON"
    )

    args = parser.parse_args()

    try:
        result = delegate(args.tier, args.prompt, args.json)
        print(result)
    except Exception as e:
        if args.json:
            print(json.dumps({"error": str(e), "success": False}, indent=2))
            sys.exit(1)
        else:
            print(f"Error: {e}", file=sys.stderr)
            sys.exit(1)


if __name__ == "__main__":
    main()

Step 2: Make executable

chmod +x ~/.claude/mcp/llm-router/delegate.py

Step 3: Verify syntax

Run: python3 -m py_compile ~/.claude/mcp/llm-router/delegate.py Expected: No output (success)

Step 4: Commit

git -C ~/.claude add mcp/llm-router/delegate.py
git -C ~/.claude commit -m "feat(external-llm): add delegation helper"

---

Task 8: Create Toggle Script

Files:

• Create: ~/.claude/mcp/llm-router/toggle.py

Step 1: Write toggle script

#!/usr/bin/env python3
"""
Toggle external-only mode.

Usage:
    toggle.py on [--reason "user requested"]
    toggle.py off
    toggle.py status
"""

import argparse
import json
import sys
from datetime import datetime
from pathlib import Path

STATE_FILE = Path.home() / ".claude/state/external-mode.json"


def load_state() -> dict:
    """Load current state."""
    if STATE_FILE.exists():
        with open(STATE_FILE) as f:
            return json.load(f)
    return {"enabled": False, "activated_at": None, "reason": None}


def save_state(state: dict):
    """Save state to file."""
    with open(STATE_FILE, "w") as f:
        json.dump(state, f, indent=2)


def enable(reason: str = None):
    """Enable external-only mode."""
    state = {
        "enabled": True,
        "activated_at": datetime.now().isoformat(),
        "reason": reason or "user-requested"
    }
    save_state(state)
    print("External-only mode ENABLED")
    print(f"  Activated: {state['activated_at']}")
    print(f"  Reason: {state['reason']}")
    print("\nAll agent requests will now use external LLMs.")
    print("Run 'toggle.py off' or '/pa --external off' to disable.")


def disable():
    """Disable external-only mode."""
    state = {
        "enabled": False,
        "activated_at": None,
        "reason": None
    }
    save_state(state)
    print("External-only mode DISABLED")
    print("\nAll agent requests will now use Claude.")


def status():
    """Show current mode status."""
    state = load_state()
    if state.get("enabled"):
        print("External-only mode: ENABLED")
        print(f"  Activated: {state.get('activated_at', 'unknown')}")
        print(f"  Reason: {state.get('reason', 'unknown')}")
    else:
        print("External-only mode: DISABLED")
        print("  Using Claude for all requests.")


def main():
    parser = argparse.ArgumentParser(description="Toggle external-only mode")
    subparsers = parser.add_subparsers(dest="command", required=True)

    # on command
    on_parser = subparsers.add_parser("on", help="Enable external-only mode")
    on_parser.add_argument("--reason", help="Reason for enabling")

    # off command
    subparsers.add_parser("off", help="Disable external-only mode")

    # status command
    subparsers.add_parser("status", help="Show current mode")

    args = parser.parse_args()

    if args.command == "on":
        enable(args.reason)
    elif args.command == "off":
        disable()
    elif args.command == "status":
        status()


if __name__ == "__main__":
    main()

Step 2: Make executable

chmod +x ~/.claude/mcp/llm-router/toggle.py

Step 3: Test toggle

Run: ~/.claude/mcp/llm-router/toggle.py status Expected: Shows "External-only mode: DISABLED"

Step 4: Commit

git -C ~/.claude add mcp/llm-router/toggle.py
git -C ~/.claude commit -m "feat(external-llm): add toggle script"

---

Task 9: Update Session Start Hook

Files:

• Modify: ~/.claude/hooks/scripts/session-start.sh

Step 1: Read current hook

Run: cat ~/.claude/hooks/scripts/session-start.sh

Step 2: Add external mode check

Add before the final output section:

# Check external mode
if [ -f ~/.claude/state/external-mode.json ]; then
  EXTERNAL_ENABLED=$(jq -r '.enabled // false' ~/.claude/state/external-mode.json)
  if [ "$EXTERNAL_ENABLED" = "true" ]; then
    echo "external-mode:enabled"
  fi
fi

Step 3: Verify hook

Run: bash -n ~/.claude/hooks/scripts/session-start.sh Expected: No output (success)

Step 4: Commit

git -C ~/.claude add hooks/scripts/session-start.sh
git -C ~/.claude commit -m "feat(external-llm): announce external mode in session-start"

---

Task 10: Add Component Registry Triggers

Files:

• Modify: ~/.claude/state/component-registry.json

Step 1: Read current registry

Run: cat ~/.claude/state/component-registry.json | jq '.skills'

Step 2: Add external-mode skill entry

Add to skills array:

{
  "id": "external-mode-toggle",
  "name": "External Mode Toggle",
  "description": "Toggle between Claude and external LLMs",
  "path": "mcp/llm-router/toggle.py",
  "triggers": [
    "use external",
    "switch to external",
    "external models",
    "external only",
    "use copilot",
    "use opencode",
    "back to claude",
    "use claude again",
    "disable external",
    "external mode"
  ]
}

Step 3: Validate JSON

Run: cat ~/.claude/state/component-registry.json | jq . Expected: Valid JSON

Step 4: Commit

git -C ~/.claude add state/component-registry.json
git -C ~/.claude commit -m "feat(external-llm): add external-mode triggers to registry"

---

Task 11: Integration Test

Step 1: Test toggle

~/.claude/mcp/llm-router/toggle.py status
~/.claude/mcp/llm-router/toggle.py on --reason "testing"
~/.claude/mcp/llm-router/toggle.py status
~/.claude/mcp/llm-router/toggle.py off

Expected: Status changes correctly

Step 2: Test router (mock - will fail without actual CLIs)

~/.claude/mcp/llm-router/invoke.py --model copilot/gpt-5.2 -p "Say hello" --json 2>&1 || echo "Expected: CLI not found (normal if opencode not installed)"

Step 3: Test delegation helper status check

python3 -c "
import sys
sys.path.insert(0, '$HOME/.claude/mcp/llm-router')
from delegate import is_external_mode
print(f'External mode: {is_external_mode()}')
"

Expected: "External mode: False"

Step 4: Final commit

git -C ~/.claude commit --allow-empty -m "feat(external-llm): integration complete - gleaming-routing-mercury"

---

Summary

After completing all tasks:

Component	Status
state/external-mode.json	Created
state/model-policy.json	Extended
mcp/llm-router/invoke.py	Created
mcp/llm-router/delegate.py	Created
mcp/llm-router/toggle.py	Created
mcp/llm-router/providers/opencode.py	Created
mcp/llm-router/providers/gemini.py	Created
hooks/scripts/session-start.sh	Updated
state/component-registry.json	Updated

To use:

# Enable external mode
~/.claude/mcp/llm-router/toggle.py on

# Or via PA
/pa --external on
/pa switch to external models

# Invoke directly
~/.claude/mcp/llm-router/invoke.py --task reasoning -p "Explain quantum computing"

# Delegate (respects mode)
~/.claude/mcp/llm-router/delegate.py --tier sonnet -p "Check disk space"