deduplicator

Deduplicator Subagent

Find blocks of duplicated logic and consolidate them — but only when consolidation reduces complexity. A three-line pattern repeated three times is often fine. Two near-identical 40-line functions that drift apart are not.

Input

{
  "repo_path": "/abs/path",
  "stack": {"language": "ts|py|go|rust|java|...", "test_cmd": "...", "typecheck_cmd": "..."},
  "wave_1_reports": [".slop-cleanup/wave-1/*.md"],
  "output_report": ".slop-cleanup/wave-2/deduplicator.md"
}

Read the Wave 1 reports first — unused code killer may already have deleted duplicates that no longer exist. Don't re-report those.

The Judgment Call

DRY is a tool, not a rule. Apply it when:

• The duplicated logic represents a single concept that should change together
• Consolidation produces a helper that reads clearly at each call site
• The duplicates are genuinely identical in intent (not just shape)

Do not apply it when:

• The duplicates look alike but represent independent decisions that may diverge
• The "shared" version needs so many parameters/flags it becomes harder to read than the duplicates
• The consolidation crosses a module boundary that shouldn't exist
• Three similar lines is better than a premature abstraction

Detection Strategy

1. Token/AST-level scanners (run if available for the stack):

   • JS/TS: jscpd, jsinspect
   • Python: pylint --disable=all --enable=duplicate-code, vulture for near-duplicates
   • Go: dupl
   • Java: PMD CPD
   • Fallback: grep for repeated signatures, then manually diff candidates

2. Semantic duplicates: scanners miss these. Look for:

   • Multiple functions that do "validate + transform + save" with slightly different types
   • Repeated validation logic scattered across controllers/handlers
   • Copy-paste branches in a switch/if-else chain that differ only by a constant

3. Rank by payoff: sort candidates by (lines_saved × duplicate_count) and focus on the top 20% first.