Cleanup Dedupe

Find duplicated code and consolidate where it reduces complexity. Auto-extract only token-identical, sizeable blocks. Smaller or divergent duplicates often shouldn't be DRY'd — premature abstraction is worse than 3 similar lines.

Preflight

1. Language detect — jscpd handles 150+ languages, so we run it on whatever is present.
2. Git state: refuse auto-apply on dirty tree.
3. Report dir: ensure .claude/cleanup-reports/ exists.
4. Read existing util conventions: where does the project keep shared utilities? Look for lib/, utils/, shared/, common/, package directories. Extracted code goes there.

Detect

# jscpd - the standard for cross-language clone detection
bunx jscpd --min-tokens 70 --min-lines 30 --threshold 0 --reporters json --output /tmp/jscpd-out . 2>/dev/null \
  || npx jscpd --min-tokens 70 --min-lines 30 --reporters json --output /tmp/jscpd-out .

Parse /tmp/jscpd-out/jscpd-report.json. Each duplication entry has firstFile, secondFile, lines, tokens, and the actual fragment.

Filter the noise before further analysis:

• Drop matches in node_modules/, dist/, build/, .next/, __generated__/, *.generated.*.
• Drop matches inside test files mirroring each other (test setup duplication is often intentional).
• Drop matches that are just type/interface definitions — those go to cleanup-types.
• Drop matches in migration files.

Assess

Write .claude/cleanup-reports/cleanup-dedupe-{YYYY-MM-DD}.md:

# Duplication Assessment — YYYY-MM-DD

## Summary
- Total clones found: N (after filtering noise)
- HIGH confidence (auto-extractable): X
- MEDIUM (similar but divergent): Y
- LOW (structural similarity, intentional): Z
- Estimated LOC saved (HIGH only): ~N

## Clones

### Clone 1 — HIGH (extract to `packages/utils/src/format.ts`)
- Files: `apps/app/features/holdings/format.ts:15-67`, `apps/admin/features/users/format.ts:22-74`
- 52 lines, 380 tokens, identical
- Both implement `formatCompactNumber(n: number): string`
- Extract to: `packages/utils/src/formatCompactNumber.ts`, both sites import.

### Clone 2 — MEDIUM
- Files: A and B
- Similar structure but divergent in 3 spots — the formatting differs by locale, the rounding by precision.
- Recommendation: don't extract yet — the abstraction needs a parameterization design that the human should decide.

## Critical Assessment
[2-3 paragraphs: are duplicates concentrated in one area? Is there a missing shared package? Are the duplicates a sign that an early abstraction would have been wrong, OR that a missing abstraction is hurting?]

Apply

Auto-extract HIGH-confidence only.

Confidence rubric

HIGH (auto-apply):

• ≥30 LOC AND ≥70 tokens (jscpd defaults).
• Token-identical (jscpd's strict mode, after stripping comments and whitespace).
• Same function signature OR reducible to one with no parameter changes.
• Across 2+ files in the same workspace/package boundary OR a clear shared package exists.
• Not in test files, fixtures, or generated code.

MEDIUM (report only):

• Similar but with 1-3 spot divergences — needs parameterization design.
• Cross-package duplication where extraction requires creating a new shared package.
• 15-29 LOC — borderline; sometimes 3 similar functions are clearer than 1 over-parameterized one.

LOW (note, no action):

• Structural similarity (e.g., 5 React components with the same prop-spreading pattern) — usually intentional.
• Test setup boilerplate.

Execution (HIGH only)

1. Determine the destination:

• Same package: existing lib/ or utils/ directory.
• Different packages: shared package if one exists (e.g., packages/utils/), else escalate to MEDIUM (don't auto-create new packages).

1. Create the new util file with the extracted function. Use the canonical name from one of the source files; prefer the more descriptive name.
2. Replace both source occurrences with imports of the new util.
3. Delete the originals.
4. Single commit: chore(cleanup): cleanup-dedupe — extracted N shared utilities.

Verify

bun run check 2>&1 || npx tsc --noEmit && npx eslint .
bun test 2>&1 || pytest 2>&1
# Re-run jscpd to confirm reduction
bunx jscpd --min-tokens 70 --min-lines 30 . 2>&1

If verify fails: revert and downgrade all to MEDIUM.

Output

• "Extracted N shared utilities, saved ~M LOC. K duplicates deferred — see report."
• Report path.
• Verify status.

NEVER

• Extract code from generated/.d.ts/migration files.
• Create a new shared package automatically — escalate that decision.
• Extract React/Vue components on prop-pattern similarity alone — composition vs. abstraction is design judgment.
• Force an abstraction that requires changing function signatures of either source — that's a behavior change, not a refactor.
• DRY two similar test cases — tests benefit from explicitness; let them duplicate.
• Extract code that has different error handling or different side effects in each location — not actually duplicate behavior.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: raintree-technology
• Source: raintree-technology/agent-starter
• License: MIT
• Homepage: https://claude.raintree.technology

Install and usage instructions live in the source repository linked above.