fusion

name	fusion
description	Hybrid smart model fusion with auto-shifting between consensus, sequential, specialist, and fallback modes. Use whenever the user wants multiple models to collaborate on a task - automatically selects the optimal strategy based on task analysis.

Fusion - Hybrid Smart Model Collaboration

Use this skill when the user wants to leverage multiple AI models on a single task. The skill automatically analyzes the request and selects the optimal fusion strategy.

Available Fusion Modes

The skill operates in one of four modes, auto-selected based on task analysis:

1. Consensus Mode

• When: Complex reasoning, fact-finding, strategic decisions
• How: Query 2-3 models in parallel, synthesize agreement/disagreement
• Models: Analysis-focused models (o1, o3, Claude-3-Opus, GPT-4)
• Use case: "What are the trade-offs of microservices vs monolith?" or "Should we adopt technology X?"

2. Sequential Mode

• When: Multi-step tasks requiring iterative refinement
• How: Pipeline models in sequence (think → create → review → refine)
• Models: Specialized by step (reasoning → coding → review)
• Use case: "Build a React component with tests" or "Design and implement an API"

3. Specialist Mode

• When: Task type is clearly defined (coding, creative, analysis)
• How: Route to models optimized for that domain
• Models: Task-specific experts (DeepSeek-Coder, Claude-3.5-Sonnet, GPT-4V)
• Use case: "Write a Python script for data analysis" or "Create a marketing tagline"

4. Fallback Mode

• When: Reliability is critical, primary model might fail
• How: Try primary, automatically failover to alternatives
• Models: Primary + 2-3 backups
• Use case: Production code generation, critical analysis

Task Analysis & Routing

When this skill triggers, analyze the user's request for:

1. Complexity signals: "compare", "evaluate", "pros/cons", "multiple perspectives" → Consensus mode
2. Step signals: "build", "implement", "create", "then" → Sequential mode
3. Domain signals:
   • Code: "write", "implement", "debug", "function" → Specialist (coding models)
   • Creative: "creative", "write", "design", "brainstorm" → Specialist (creative models)
   • Analysis: "analyze", "explain", "research", "compare" → Specialist (analysis models)
4. Reliability signals: "critical", "production", "important" → Fallback mode

Model Pool Configuration

Default models organized by capability (configurable via environment):

• Coding: deepseek-coder, gpt-4, claude-3-5-sonnet
• Analysis: o1, o3, claude-3-opus, gpt-4
• Creative: gpt-4, claude-3-5-sonnet, command-r
• Vision: gpt-4-vision, claude-3-opus-vision

Who picks the models?

Fusion supports two model-selection strategies, controlled by the selection_strategy parameter on execute_fusion:

• auto (default) — Fusion's classifier picks the best models for the task. Hands-off.
• user — Use exactly the models the user specified in the models array. Fusion will not override or add models. The classifier still suggests a mode so the agent knows how to run them, but the model list is the user's final choice.

When to ask the user vs. decide for them

Default to auto. Only offer user selection when the user signals they care about which model runs. Signals:

• They name a model ("use Claude", "can GPT-4 handle this?")
• They ask to choose ("which models?", "can I pick?")
• They have a strong preference ("I don't want o1", "stick to open-source models")
• They're iterating and unhappy with a previous model's output

When you detect these signals:

1. Call mcp__fusion__get_models to fetch the available pools.
2. Present the pools to the user with a short description of each.
3. Ask which model(s) they want (use the AskUserQuestion tool if available, or ask inline).
4. Call mcp__fusion__execute_fusion with selection_strategy: "user" and their chosen models.

If the user does not care, use selection_strategy: "auto" and proceed silently.

MCP Integration

If the Fusion MCP server is configured, use these tools:

• mcp__fusion__analyze_task - Get task classification (mode + recommended models + confidence)
• mcp__fusion__execute_fusion - Run multi-model execution. Supports selection_strategy ("auto" | "user"), mode, and models.
• mcp__fusion__get_models - List available models grouped by capability, with a description of each pool's strengths.

If MCP is not available, use skill-based delegation to existing agents with model hints.