Categorize Phase Executor

Context Tier: MINIMAL

Input

Receives via prompt:

• plugin_path: Path to plugin directory
• improvement_ids: List of improvement IDs to categorize

NOT PROVIDED (context isolation):

• Full analysis data (not needed for categorization)
• Plugin contents (not needed)
• Grounding results (done in next phase)

Execution

1. Read Improvements from State

   python scripts/state_manager.py read "$plugin_path" --field mutable
   

   Extract improvements list.

2. Categorize by Priority

   Apply severity rules:

   • HIGH: estimated_reduction > 0.30 OR fixes pattern violation
   • MEDIUM: 0.10 <= estimated_reduction <= 0.30
   • LOW: estimated_reduction < 0.10

   Group improvements into priority batches:

   improvements_by_priority:
     HIGH: [improvement_ids with HIGH priority]
     MEDIUM: [improvement_ids with MEDIUM priority]
     LOW: [improvement_ids with LOW priority]
   

3. Store Categorization in State

   python scripts/state_manager.py update "$plugin_path" categorized_improvements "$CATEGORIZED_JSON"
   

Output

categorization_complete:
  by_priority:
    HIGH:
      count: [count]
      ids: [list of IDs]
    MEDIUM:
      count: [count]
      ids: [list of IDs]
    LOW:
      count: [count]
      ids: [list of IDs]
  total_improvements: [count]

State Integration

• Reads: mutable.improvements (list with estimated_reduction, priority fields)
• Writes: mutable.categorized_improvements (grouped by HIGH/MEDIUM/LOW)

Priority Logic

This phase implements the severity-based batching strategy:

1. HIGH priority improvements get full grounding (all 4 agents)
2. MEDIUM priority get partial grounding (2 agents)
3. LOW priority get minimal grounding (1 agent)

This reduces grounding overhead by ~60% for typical improvement sets.