Codebase Polish

Codebase Polish

ultrathink

You coordinate a full codebase polish. You score every source file 0-10 on polish potential and dispatch parallel background agents to refine files scoring 5+.

You are a coordinator. You read and analyze files, create tasks, and spawn worker agents. You do NOT edit files directly.

Arguments

• $ARGUMENTS containing --dry-run: analyze and report only, no modifications
• $ARGUMENTS containing a path: scope to that directory/file

Phase 1: Codebase Discovery

Glob all source files using parallel patterns:

• **/*.ts, **/*.tsx, **/*.js, **/*.jsx
• **/*.py, **/*.go, **/*.rs
• **/*.vue, **/*.svelte

Exclude: node_modules/**, dist/**, build/**, .next/**, coverage/**, *.min.*, *.d.ts, _generated/**, .git/**

If path argument provided, scope discovery to that path.

Phases 2-5

Read ${CLAUDE_SKILL_DIR}/references/workflow.md for the analysis, work queue, parallel agents, and verification workflow.

Gotchas

• Generated files (*.d.ts, _generated/, codegen output) get picked up by glob patterns. The exclusion list must be checked before scoring or agents will "polish" auto-generated code.
• Polishing test files can silently break assertions. Renaming variables, simplifying ternaries, or removing "dead" setup code in tests often changes behavior. Score test files conservatively.
• Agents run in parallel and can conflict if two files import from each other. One agent renaming an export while another reads the old name causes build failures.
• 5 simultaneous agents on a large codebase can hit Claude rate limits. If agents start failing, reduce concurrency to 2-3.