fix: Add missing Inference.ts module for hook LLM calls

[b1ackmartian]

Summary

• Adds the missing Inference.ts module that several UserPromptSubmit hooks depend on
• Uses native fetch to call Anthropic Messages API (no additional dependencies)
• Supports fast/standard/complex inference levels for different use cases

Fixes #418

Changes

• Created Packs/pai-core-install/src/skills/CORE/Tools/Inference.ts

API Signature

interface InferenceOptions {
  systemPrompt: string;
  userPrompt: string;
  expectJson?: boolean;
  timeout?: number;
  level?: 'fast' | 'standard' | 'complex';
}

interface InferenceResult {
  success: boolean;
  output?: string;   // Raw text output
  parsed?: any;      // Parsed JSON when expectJson=true
  error?: string;    // Error message
}

Model Mapping

• fast: claude-3-5-haiku-latest (cheap, fast, good for classification)
• standard: claude-sonnet-4-20250514 (balanced)
• complex: claude-opus-4-5-20251101 (complex reasoning)

Test Plan

• AutoWorkCreation.hook.ts executes without import error
• ImplicitSentimentCapture.hook.ts executes without import error
• UpdateTabTitle.hook.ts executes without import error
• New session starts without hook errors

🤖 Generated with Claude Code

[kaimagnus]

Thank you for the Inference.ts module, @b1ackmartian! Clean implementation with the three-tier model approach (fast/standard/complex).

v3.0 already includes an Inference.ts utility and the target path (Packs/pai-core-install/) no longer exists. Closing as this is superseded by the v3.0 tooling. Appreciate the solid API design!