Synthesize Phase Executor

Context Tier: METADATA

Input

Receives via prompt:

• plugin_path: Path to plugin directory
• selected_improvement_ids: IDs of improvements user selected to apply

NOT PROVIDED (context isolation):

• Rejected improvements (filtered out by coordinator)
• Full analysis (not needed for synthesis)
• Intermediate reasoning (only final structured data)

Execution

1. Read Selected Improvements from State

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

   Extract:

   • improvements (filter to selected IDs)
   • grounding_results (filter to selected IDs)
   • immutable (for scope metadata)

2. Prepare Synthesis Input Build METADATA-tier input:

   selected_improvements:
     - improvement_id: CTX-001
       description: [what was selected]
       code_change:
         before: [current code]
         after: [improved code]
       grounding_results:
         pattern_compliant: true
         token_estimate: [reduction data]
         risk_level: LOW
   scope_metadata:
     plugin_name: [from analysis]
     files_analyzed: [count]
     improvements_available: [total count from state]
     improvements_selected: [selected count]
     focus_area: [from immutable state]
   

3. Launch Improvement Synthesizer

   Task: Generate final improvement report
   Agent: coordinator-internal/improvement-synthesizer.md
   Prompt:
     selected_improvements: [metadata above]
     scope_metadata: [plugin context]
   

4. Return Report Directly Return the synthesizer's ImprovementReport UNCHANGED to coordinator.

Output

improvement_report:
  executive_summary:
    improvements_applied: [count]
    estimated_token_reduction: [total %]
    patterns_improved: [list]
  changes:
    - file: [path]
      change_type: [type]
      before: [code]
      after: [code]
  next_steps:
    - [actionable step]

State Integration

• Reads:
  • mutable.improvements (selected subset)
  • mutable.grounding_results (selected subset)
  • immutable (plugin_path, focus_area, session_id)
• Writes: Nothing (final phase)

Context Tier Justification

This phase operates at METADATA tier:

• Only selected improvements (rejected ones filtered)
• Only structured grounding results (no intermediate reasoning)
• Only scope metadata (plugin name, counts, focus area)
• Total context: ~200-500 tokens (vs 5-20K tokens in analyze phase)

This represents a 97% context reduction from Phase 1 to Phase 6.