widescreen-research: Command for Claude Code

Validate State Tracker Pattern

Description

Validates that MCP server operations follow the "state tracker only" principle. The MCP server must NEVER provide intelligence - it only validates, stores state, and returns progress metadata. All intelligence comes from the LLM client.

Core Principle

The LLM is the intelligent client. The server only tracks state.

Rules

The Server Should:

• ✅ Validate parameters (types, required fields)
• ✅ Store state in session (unmodified)
• ✅ Track progress (counts, percentages)
• ✅ Return metadata only
• ✅ Handle errors

The Server Should NEVER:

• ❌ Generate content (ideas, hypotheses, solutions)
• ❌ Analyze content (evaluate, assess, interpret)
• ❌ Make decisions based on content
• ❌ Provide intelligence (infer, deduce, conclude)
• ❌ Transform content (modify, enhance, improve)

Usage

For MCP servers implementing this pattern, run validation before committing:

npm run validate:state-tracker

Example Violations

BAD: Server provides intelligence

// ❌ Generating content
const ideas = this.generateIdeas(prompt);

// ❌ Analyzing content
const quality = this.evaluateQuality(prompt);

// ❌ Making decisions based on content
if (prompt.includes('complex')) {
  return this.useAdvancedMode();
}

GOOD: Server tracks state only

// ✅ Store what LLM provides
const idea = { content: prompt };

// ✅ Use LLM's evaluation (from parameters)
const quality = parameters.quality;

// ✅ Use LLM's decision (from parameters)
const mode = parameters.mode;

Implementation

This pattern ensures:

• Research is reproducible (same state → same behavior)
• Intelligence is auditable (all in LLM prompts/parameters)
• Servers are simple and testable (pure state machines)
• Clear separation of concerns (LLM = intelligence, Server = state)