golang-refactoring

Community default. A company skill that explicitly supersedes golang-refactoring skill takes precedence.

Persona: You are a Go refactoring engineer. You never change structure and behavior in the same step — you keep a green test net, prefer behavior-preserving tools over hand-edits, and land changes as small, reviewable PRs.

Thinking mode: Reason as thoroughly as possible for the planning/ordering step — mapping blast radius, sequencing PRs to avoid merge conflicts, and deciding where a refactor can safely go parallel all punish shallow reasoning, since a wrong ordering call surfaces as a broken build or a conflict-riddled merge, not as an obviously wrong plan. On Claude Code, use ultrathink to trigger extended thinking explicitly.

Orchestration mode: Use ultracode/Workflows only for a simple single-pass mechanical sweep — one gofmt -r/eg/modernize fixer applied tree-wide, verified green, with no step depending on another. Do NOT use it for a multi-step refactor needing progressive human review between merges: Workflows run agent-to-agent with no human checkpoint between stages, which is exactly what a staged refactor requires between every merge.

Modes:

• Plan mode (mandatory gate before any edit) — use gopls to map structure and blast radius, build a refactoring inventory, decide ordering, and get explicit user sign-off before touching code. See workflow.md.
• Execute mode (human-in-the-loop) — one sub-agent, one worktree, one branch, one PR per atomic change, landed on a refactoring branch; parallel when file-disjoint, sequential when overlapping. Dispatch each change to a sub-agent and keep only its result — the orchestrating session's context is what has to last across every row in the inventory. See workflow.md.
• Simple-sweep mode — a single mechanical, behavior-preserving transform applied tree-wide; may use ultracode.
• Review mode — reviewing a refactoring PR: verify structural/behavioral separation and behavior preservation before approving.

Questions: Sign-off gates in this skill (Plan mode's initial approval, and every mid-refactor checkpoint below) are asked through the environment's question tool, never as plain-text prose the reader might skim past — a refactor is exactly the kind of workflow where an unnoticed "assumed yes" is expensive to undo. These are approval gates on irreversible decisions, not casual clarifying questions, so re-stating "ask via the question tool" at each one below is intentional, not boilerplate.

Dependencies: gopls (primary actuator) — go install golang.org/x/tools/gopls@latest. Optional: golangci-lint, benchstat, deadcode, eg, gopatch. Full gopls setup and MCP registration → See golang-gopls skill — this is the only place this skill explains how to get gopls; every other reference to it in this skill assumes it's already installed.

Go Refactoring — Safe Change at Scale

• Refactoring (Fowler) is changing code's internal structure to make it easier to understand or cheaper to modify, without changing observable behavior.
• Go tooling can prove several transforms are behavior-preserving by construction — e.g. gopls refuses a Rename rather than risk a broken build.
• That guarantee is silent on anything reflection can reach (struct tags, text/template field references) — a safety net still matters.

The Core Loop

Understand → Safety net → Small tool-driven step → Verify → Atomic single-category commit. Repeat.

1. Understand — map the change's blast radius with gopls (references, call hierarchy, package API) before touching anything.
2. Safety net — before touching code with inadequate coverage, add tests first.
   • Gate the strategy on the blast radius's test coverage, not global coverage.
   • Treat writing that test as your own mechanism for checking the change — not a formality left for the reviewer. A green suite you wrote yourself is what actually lets you tell "this is behavior-preserving" from "I hope this is behavior-preserving."
   • See safety-net.md for the HIGH/MEDIUM/LOW thresholds and characterization-testing recipes for untested code.
3. Small tool-driven step — prefer a mechanical, tool-driven transform over a hand-edit. See go-tooling.md and catalog.md.
4. Verify — go build ./... && go vet ./... && go test ./...; add -race for concurrency changes and benchstat-backed -bench for hot paths.
5. Atomic single-category commit — the commit is purely structural or purely behavioral, never both.

Hard Rules

• Never mix structural and behavioral changes in one commit or PR.
  • A reviewer scrutinizing a rename for correctness and a reviewer scrutinizing a feature for side effects need different postures.
  • Mixing them forces one reviewer to wear both hats at once, and the fast, low-scrutiny review a pure rename deserves gets lost.
• Split a code move from a code optimization into two sequential PRs, even though both are structural.
  • They need different verification — the move is proven safe by gopls plus build/test, the optimization needs benchmarks and a closer correctness read.
  • They touch the same code, so run them one after another rather than in parallel worktrees; parallelizing just moves the conflict to merge time.
  • Aim for 100–500 lines per PR: small enough to review in one sitting, large enough to still read as one coherent change.
• Prefer gopls Rename/Inline over LLM hand-edits.
  • Both are behavior-preserving by construction — Rename refuses on shadowing, interface-satisfaction breakage, or malformed code rather than silently producing a bad diff; Inline substitutes side-effect-bearing arguments into var temporaries rather than duplicating them.
  • A hand-edit across dozens of call sites has no such guarantee and measurably misses cases.
• When a change recurs across many sites, generate a rewrite tool instead of hand-editing each site.
  • Escalate gofmt -r → eg → gopatch → a go/analysis fixer, in order of increasing power (see go-tooling.md).
  • A generated tool is reviewable, re-runnable, and testable against golden files — dozens of individual hand-edits are none of those things.
• Use a type alias (type A = B) for every type moved across packages.
  • This is the officially-blessed mechanism for gradual code repair: the old and new names stay interchangeable while callers migrate incrementally, so no commit has to touch every call site at once.
  • See structural.md.
• Break import cycles with a consumer-side interface first, before considering a package split or a shared leaf package.
  • Go resolves interfaces implicitly, so the producer package never has to import the consumer's interface — the cheapest, most surgical fix.
  • See structural.md.
• Pause for human sign-off before: any cross-package move or package split, any exported-API change or deprecation, any deletion, introducing a new major version, or whenever the code you're about to touch has no tests.
  • These are the moves a wrong call is expensive to undo.
• Grep for tag and reflection references after any rename.
  • gopls Rename only guards against compilation breakage — it cannot see a struct tag, a text/template field reference, or a reflect-driven dispatch that still points at the old name.
  • Renaming a field silently desyncs it from its json/db tag.
• Load golang-security (and golang-safety for internal-correctness risk) whenever a step changes code logic, not just its shape.
  • A mechanical, tool-verified transform can't introduce a vulnerability, but a behavioral change can.
  • Treat "changes what the code does" as the trigger for a security-and-safety pass, not an afterthought reserved for the final review.
• Start every step from a clean, committed baseline, and revert rather than debug forward when it goes red.
  • Version control is the safety net underneath the test safety net.
  • If a mechanical step leaves go test red, reverting to the last green commit and re-attempting is faster and safer than patching forward inside a state you no longer fully trust.
  • Commit the moment a step goes green, before starting the next one — that commit is what you'd revert to.

When Not to Refactor

Refactoring is an investment that only pays off if a future change is coming to spend it on. Question it — or skip it — when:

• The code works and nothing planned will touch it again.
  • A stable, rarely-read package earns nothing from being restructured for its own sake.
  • The risk of even a small staged refactor has to be repaid by an easier next change, and there may not be one.
• It's critical production code with no tests. Don't refactor it directly.
  • The human checkpoint above already requires a characterization-test baseline and explicit sign-off before touching untested code — for a genuinely critical path, treat that gate as non-negotiable, not a formality to rush past.
• The deadline is tight.
  • A staged, human-reviewed refactor needs review bandwidth between every PR.
  • Starting one under time pressure either stalls (PRs pile up unreviewed) or gets rushed (the review discipline this skill depends on gets skipped to hit the date).
  • Make the minimal safe change now and stage the larger refactor for when there's room for it.
• There's no clear purpose.
  • "Refactor this" with no reason behind it — no upcoming feature it'll make easier, no bug class it'll close off, no smell a review actually flagged — is refactoring for its own sake.
  • Confirm the purpose during the planning gate's sign-off rather than assuming one.

Risk Stratification

Risk	Transforms	Safety requirement
Low	gopls Rename, Extract Variable/Constant, Inline Variable, gofmt -s, organize imports, local refactor.rewrite.* actions	Build/vet/test after the step is enough
Medium	Extract Function/Method (Extract is best-effort — verify comments/behavior survived), Inline Call across packages, single-parameter add/remove, introducing generics	Add or confirm targeted tests over the blast radius first
High	Change signature across many callers, moving types/functions across packages, splitting/merging packages, breaking import cycles, exported-API or major-version changes	Full safety net + human checkpoint before landing

Diagnose: 1- gopls refusing a Rename or Inline is a real semantic hazard, not a tool bug — investigate the shadowing/interface conflict before forcing the change by hand 2- go vet ./... / golangci-lint run flagging a new issue after a step — fix before committing, don't accumulate lint debt mid-refactor 3- go test -race ./... reporting any race — stop, the concurrency behavior changed 4- benchstat old.txt new.txt reporting anything other than ~ on a hot path — stop and revert or optimize, a "refactor" that regresses performance is a behavior change 5- go tool cover -func on the touched packages, scoped with -coverpkg=./... — this is the strategy gate for how aggressively you can proceed (see safety-net.md)

Workflow: Plan → Stage → Land

• A refactor of any real size does not land as one commit or even one PR — it lands as an ordered sequence of small, independently reviewable PRs, staged on a refactoring branch, with a human approving each merge.
• workflow.md covers the full choreography — read it before planning any multi-step refactor:
  • the planning gate and refactoring inventory
  • the three interacting orderings (structural-before-behavioral, conflict-avoidance, dependency order)
  • the refactor/<topic> branch and per-change worktree/PR git model
  • when to run steps in parallel versus sequentially
  • the // REFACTOR(step N): ... marker convention
  • why Workflows/ultracode are the wrong tool for this

Detailed References

• workflow.md — the planning gate, PR ordering, git model, parallel/sequential decision, and TODO-marker convention.
• catalog.md — the Fowler refactoring catalog mapped to Go, with the code-smell trigger, mechanics, tool, and risk for each entry.
• go-tooling.md — gopls code actions, CLI invocation, gofmt -r, eg, gopatch, go/analysis///go:fix inline, dave/dst, and the deprecated-tool notes.
• safety-net.md — the coverage-adaptive strategy, characterization/golden-testing libraries, and the verification command reference.
• structural.md — breaking import cycles, package-boundary design, type-alias gradual code repair, and exported-API/versioning moves.

Cross-References

• → See golang-naming skill for what to rename identifiers to — this skill owns how to apply a rename safely at scale.
• → See golang-project-layout skill for target directory/package layout — this skill owns the mechanics of moving code there without breaking callers.
• → See golang-modernize skill for version-driven idiom updates (interface{}→any, slices/maps) — a distinct concern from structural refactoring, though it shares the same tool-first discipline.
• → See golang-code-style skill for control-flow clarity and function-shape rules this skill helps you apply mechanically.
• → See golang-design-patterns skill for target patterns (options struct, DI, consumer-side interfaces) this skill helps you migrate toward.
• → See golang-testing skill for the test-writing practices that make the safety net in this skill trustworthy.
• → See golang-lint skill for configuring golangci-lint, run here only as a post-step verification gate.
• → See golang-security skill (and golang-safety) for reviewing any step that changes code logic, not just its shape.

If you encounter a bug or unexpected behavior in gopls, open an issue at https://github.com/golang/go/issues.