gopherstack: Skill for Claude Code

gopherstack-parity-audit

The 5 parity principles

1. Never ship stub methods. No-op handlers, &StubOutput{}, fabricated IDs are all disqualifying. Every routed op must mutate/read real state, return AWS-accurate shapes + error codes, persist when persistence is on. If something genuinely can't be implemented, say so in structural_gaps: — never a half-working stub.
2. Verify wire shapes against the real SDK, not the handler's own output. Real bug classes found this way: wrong XML list wrapper (<member> vs <Item>); wrong root elements; missing response-root nesting when Output has no httpPayload member; ISO8601 strings where the JSON protocol wants epoch-seconds numbers; a hardcoded xml:"StubResponse" XMLName silently overriding every runtime root; stub registrations overwriting real handlers by registration order; missing errCodeLookup entries → 500 InternalFailure. Use gopherstack-sdk-shape to do the verification — don't restate its recipes here.
3. Unit tests are not parity proof. SDK-driven integration tests have caught 8 client-breaking wire bugs that green unit tests missed. Run task build-linux before go test ./test/integration/... — integration tests exercise the real Docker-built binary, not a Go-level shortcut.
4. Grep-based stub hunting has false positives in both directions. An op that returns an empty envelope after real backend logic ran is correct (void-result ops — read the backend method before flagging it). Conversely a "real-looking" op can be a disguised stub, e.g. filtering a map that is never actually populated. Read the backend method, not just the handler.
5. De-stub hygiene, once a stub is found and fixed: remove the stub registration, delete the orphaned stub handler func entirely, drop it from any bare-call stub test manifests, and run golangci-lint before calling it done.

The EC2 stub mechanism (read before touching services/ec2)