Improve Phase Executor

Context Tier: SELECTIVE

Input

Receives via prompt:

• analysis_id: Session ID from analyze phase
• focus_area: context|orchestration|handoff|all

NOT PROVIDED (context isolation):

• Full plugin contents (reads from state)
• Unrelated violations (filtered by focus area)
• Analysis reasoning (only structured data)

Execution

1. Read Analysis from State

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

   Extract analysis_summary for selective context.

2. Filter File References by Focus Area

   Before launching each improver, filter file_refs to pass only relevant files:

   python scripts/file_cache.py get_refs_by_focus "$plugin_path" "$focus_area"
   

   This reduces context duplication by ~2x by passing only files matching the focus area patterns.

3. Launch Improvers Based on Focus Area

   If focus_area = "context":

   Task: Optimize context patterns
   Agent: coordinator-internal/context-optimizer.md
   Prompt:
     analysis_summary: [from state - violations, opportunities for context]
     focus_area: context
     file_refs: [filtered by "context" patterns]
   

   If focus_area = "orchestration":

   Task: Improve orchestration patterns
   Agent: coordinator-internal/orchestration-improver.md
   Prompt:
     analysis_summary: [from state - violations, opportunities for orchestration]
     focus_area: orchestration
     file_refs: [filtered by "orchestration" patterns]
   

   If focus_area = "handoff":

   Task: Improve handoff patterns
   Agent: coordinator-internal/handoff-improver.md
   Prompt:
     analysis_summary: [from state - violations, opportunities for handoff]
     focus_area: handoff
     file_refs: [filtered by "handoff" patterns]
   

   If focus_area = "all": Launch ALL THREE improvers IN PARALLEL, each with focus-filtered file_refs.

4. Collect Improvements Aggregate improvements from all launched improvers.

5. Store Improvements in State

   python scripts/state_manager.py update "$plugin_path" improvements "$IMPROVEMENTS_JSON"
   

Output

improvements_generated:
  improvement_ids:
    - CTX-001
    - ORCH-002
    - HO-003
  count: [total count]
  by_priority:
    HIGH: [count]
    MEDIUM: [count]
    LOW: [count]

State Integration

• Reads: mutable.analysis_summary (PluginAnalysis)
• Writes: mutable.improvements (list of improvements with IDs)

Focus-Based Filtering (CM-001)

Each improver receives ONLY files relevant to its focus area:

• context: agents/.md, coordinator-internal/.md, skills/**/*.md
• orchestration: agents/.md, coordinator-internal/.md, hooks/*.json
• handoff: agents/.md, hooks/.json, coordinator-internal/.md, scripts/.py
• all: All three improvers run in parallel, each with their own filtered refs

This reduces context duplication by ~2x compared to passing all files to all improvers.