Refactoring Checklist

The tech lead names the modules, packages, or services in scope and the concrete signal that signals done — cyclomatic complexity below a threshold, a SonarQube quality gate passing, or a specific class extracted. Vague scopes ("clean up the auth code") drift into multi-sprint rewrites.

Run SonarQube, Code Climate, or Semgrep against the target paths. Flag long parameter lists (typically 4+), god classes, duplicated blocks, and high cyclomatic complexity. Capture the baseline metrics so post-refactor improvement is measurable.

Use jscpd, PMD CPD, or SonarQube duplication detection to surface copy-paste blocks. Resist the urge to DRY accidental duplicates — if two services happen to validate emails the same way today, extracting a shared helper couples them forever.

Run the suite with coverage (Istanbul, JaCoCo, coverage.py, SimpleCov) and capture line and branch coverage for the files in scope. Below 70% branch coverage is a red flag — write characterization tests before touching anything.

For undertested code, pin the current behavior with characterization tests before changing anything — even bugs. The point is to detect behavior change during refactoring, not to validate that the existing behavior is correct.

The full suite must pass on main before the refactor branch is cut. Quarantine flaky tests explicitly — don't start a refactor with a yellow CI status, because regressions will hide in the noise.

Branch from a known-green main sha. Plan to keep the branch short-lived (under a week) and rebase daily — long-lived refactor branches collide with concurrent feature work and become a separate merge project.

Break methods longer than ~30 lines or with multiple levels of abstraction into named helpers. Use the IDE's refactoring tool (IntelliJ, Rider, VS Code) rather than hand-editing — automated extraction preserves call sites and signatures correctly.

Replace `data`, `tmp`, `mgr`, `helper` with names that say what the value represents in the domain. Use the IDE rename refactor so references update across the repo. Commit renames as their own small commits to keep diffs reviewable.

Switch statements over a type tag that recur in three or more places are good candidates. Don't reach for polymorphism for a single 3-branch conditional — the indirection cost outweighs the win. See Fowler's "Replace Conditional with Polymorphism" for the canonical mechanics.

Methods with 4+ parameters that move together (date range, address fields, pagination) get bundled into a value object or DTO. This often surfaces a missing domain concept that has been hiding in plain sight.

For hot paths, run JMH, BenchmarkDotNet, pytest-benchmark, or k6 load tests on both shas and compare. Watch for the classic refactor regression: a clean Extract Method that re-allocates inside a tight loop, or a polymorphic dispatch that defeats inlining.

The PR description names the smell, the technique applied, and the metric improved (cyclomatic complexity from 24 to 8, duplication from 12% to 2%). Link the baseline SonarQube run and the new one. Reviewers shouldn't have to reverse-engineer the intent.

GitHub auto-requests CODEOWNERS, but page the relevant tech lead in Slack for refactors that span ownership boundaries. Avoid LGTM-without-reading on large refactors — split the PR if a reviewer can't realistically read it in one sitting.

Push fixup commits during review; squash on merge so the history reads as a coherent change. Confirm the GitHub Actions or CircleCI pipeline goes green on the final sha — required status checks should make a red merge impossible, but re-verify.

Pure refactors merge directly. Anything that changes observable behavior — even subtly, like a new code path — goes behind a LaunchDarkly, Unleash, or Statsig flag with a named owner and a cleanup ticket.

Set a 30-day expiry on the flag and assign a named owner. Stale flags are a chronic source of dead code paths and exploding test matrices — the quarterly flag review should already be on the calendar.

Argo Rollouts, Spinnaker, or a manual ALB weighted target group. Watch error rate and p99 latency on the Datadog or Grafana dashboard for at least 30 minutes before promoting. Roll back on any divergence from baseline.