mmiani/kotlin-kmp-claude-agent-skills/kotlin-kmp-refactor-safety

KMP Refactor Safety Skill

Use this skill when the task is primarily a refactor, migration, reliability hardening, or architectural cleanup of existing code (not a greenfield feature). This skill is about change discipline and reviewability, not system architecture itself. Pair it with your architecture/style/testing skills as needed.

Primary objective

Deliver refactors that are:

• minimal in surface area
• safe to review and merge
• free of parallel/competing implementations
• regression-resistant (tests + observability)

Non-negotiables

1) Scope control (no opportunistic rewrites)

• Touch only files required to achieve the goal.
• Do not reformat, rename, or “clean up” unrelated code.
• Avoid broad mechanical changes unless explicitly requested.
• Keep diffs small and intention-revealing.

2) No parallel implementations

• Do not introduce a new pattern while leaving the old pattern active.
• End state must be one of:
• old path removed, OR
• old path gated behind a feature flag with a clear removal plan (explicit TODO + tracking issue).
• Avoid duplicate sources of truth. If ownership must move, move it once and rewire callers deterministically.

3) Preserve public contracts unless required

• Keep public APIs stable (interfaces, model shapes, routing/public endpoints, configuration surface).
• If a contract must change:
• update all call sites in the same change-set
• document the breaking change in PR notes/release notes
• add targeted tests proving compatibility and intent

4) Refactor in safe phases

Prefer this order:

1. Foundation: introduce the new core abstraction (e.g., manager/store/service) and wire it without changing behavior.
2. Adoption: migrate existing call sites incrementally to the new abstraction.
3. Lock-in: remove the old path or gate it behind a flag (default off) and make the new path the default.
4. Cleanup: delete dead code, consolidate configuration, reduce complexity.

Never ship an ambiguous half-state where both old and new paths may run without a flag and clear routing.

5) Observability (redacted)

For flows prone to edge cases (auth, payments, retries, background/foreground, webhooks, state restoration):

• add logs/metrics around transitions and decisions
• never log secrets (tokens, passwords, codes, personal data)
• log only:
• booleans/branch decisions
• status codes and operation names
• hashed or truncated IDs if needed (e.g., last 4 chars)
• ensure logs can be disabled or are appropriate for production environments

6) Tests are part of the refactor

Minimum expectations:

• unit tests for pure logic/state transitions/mappers
• regression tests for the bug or failure mode motivating the refactor
• coverage for:
• success path
• failure path
• concurrency/race behavior when applicable (mutex, retry, idempotency)
• avoid brittle tests; prefer deterministic inputs/outputs and stable assertions

7) Idempotency and re-entrancy for callbacks

For events that can repeat (callbacks, deep links, webhooks, background resume):

• handle duplicates safely
• validate inputs before side effects
• operations should be safe to retry without corrupting state

8) Backwards compatibility and migration safety

• Use versioning or feature flags for behavior changes that might break existing users.
• Provide migration steps (data migrations, config updates) in a single place.
• If data shape changes:
• define transitional support window
• ensure rollback strategy (or non-destructive migrations)

9) Performance and resource safety

• Avoid adding disk/network reads on hot paths (e.g., per-request storage reads).
• Prefer caching with explicit invalidation when correctness requires it.
• Ensure retries are bounded (max 1 retry unless specified) and do not create loops.

10) PR hygiene (reviewable output)

• Prefer small, coherent commits with clear messages.
• Avoid reordering code blocks unless necessary.
• Include PR notes:
• What changed
• Why
• How to test
• Flags/migration steps
• Rollback plan (if relevant)

Quick refactor checklist

• [ ] Single source of truth after refactor
• [ ] No stale in-memory state alongside persisted state (unless intentionally synchronized)
• [ ] No competing interceptors/validators/plugins that overwrite each other
• [ ] Duplicate events/callbacks are handled idempotently
• [ ] Failure modes are explicit and do not cause silent data loss/logouts
• [ ] Tests cover the motivating regression and key edge cases
• [ ] Clear migration/flag/rollback notes included