Ground Phase Executor

Context Tier: FILTERED

Input

Receives via prompt:

• plugin_path: Path to plugin directory
• categorized_improvements: Priority batches (HIGH/MEDIUM/LOW)

NOT PROVIDED (context isolation):

• Full analysis (not needed for grounding)
• Rejected improvements (filtered out)
• Plugin contents (grounding agents get only improvement text)

Execution

1. Read Categorized Improvements from State

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

   Extract categorized_improvements.

2. Apply Severity-Based Batching

   For HIGH priority improvements: Launch ALL 4 grounding agents IN PARALLEL:

   Task: Check pattern compliance
   Agent: coordinator-internal/grounding/pattern-checker.md
   Prompt:
     improvements_to_check: [HIGH improvements]
     focus_area: [from state]
   

   Task: Estimate token impact
   Agent: coordinator-internal/grounding/token-estimator.md
   Prompt:
     improvements_to_check: [HIGH improvements]
     focus_area: [from state]
   

   Task: Check consistency
   Agent: coordinator-internal/grounding/consistency-checker.md
   Prompt:
     improvements_to_check: [HIGH improvements]
     focus_area: [from state]
   

   Task: Assess risks
   Agent: coordinator-internal/grounding/risk-assessor.md
   Prompt:
     improvements_to_check: [HIGH improvements]
     focus_area: [from state]
   

   For MEDIUM priority improvements: Launch 2 grounding agents IN PARALLEL:

   • pattern-checker.md
   • token-estimator.md

   For LOW priority improvements: Launch 1 grounding agent:

   • pattern-checker.md only

3. Challenge HIGH Priority Improvements

   For HIGH priority improvements ONLY: Launch the challenger agent to validate claims have supporting evidence:

   Task: Validate improvement claims
   Agent: coordinator-internal/grounding/challenger.md
   Prompt:
     improvements: [HIGH improvements with claims]
     relevant_files: [file references if needed]
   

   This agent runs ONLY for HIGH priority items, not MEDIUM or LOW. Maximum 1 round per improvement - no retry loops.

4. Collect Grounding Results Aggregate assessments from all grounding agents, including challenge assessments for HIGH priority improvements.

5. Store Grounding Results in State

   python scripts/state_manager.py update "$plugin_path" grounding_results "$GROUNDING_JSON"
   

Output

grounding_complete:
  grounded_improvements:
    - improvement_id: CTX-001
      priority: HIGH
      grounding:
        pattern_compliant: true
        token_estimate:
          before: 5000
          after: 3500
          reduction_percent: 30
        consistency_check: PASS
        risk_level: LOW
    - improvement_id: ORCH-002
      priority: MEDIUM
      grounding:
        pattern_compliant: true
        token_estimate:
          before: 2000
          after: 1600
          reduction_percent: 20
  total_grounded: [count]

State Integration

• Reads: mutable.categorized_improvements (priority batches)
• Writes: mutable.grounding_results (assessments for each improvement)

Efficiency Gains

Severity-based batching reduces grounding operations:

• Typical improvement set: 10 improvements (3 HIGH, 4 MEDIUM, 3 LOW)
• Without batching: 40 grounding operations (10 × 4 agents)
• With batching: 19 grounding operations (3×4 + 4×2 + 3×1)
• Reduction: 52% fewer grounding operations