Google has open-sourced the Elephant-Goldfish Model (EGM) plugin, an Apache 2.0 design-and-review layer for AI coding agents that enforces spec-first development. The workflow separates concerns: a persistent "Elephant" session drafts a specification via three commands—/egm-brainstorm (six-lens swarm covering YAGNI, token cost, UX, migration, observability, tech debt), /egm-draft (section-by-section design doc), and /egm-review—while stateless "Goldfish" subagents with zero conversation history review that spec across five sequential stages: Initial Zoom Out, Comprehension Test (paraphrasing to detect ambiguity), Critic Review (design, structural, style), Intermediate Zoom Out, and Implementation Readiness (file paths, data structures, signatures). Findings are ranked Blocker, Major, or Nit; each stage permits one re-audit. EGM supplies 7 skills, 9 reviewer personas, and 6 brainstorm lenses, but no model weights—it runs on any harness supporting fresh subagents. Verified hosts include Antigravity CLI 1.4.5 (16 native reviewers), Codex 0.163.0 (15 reviewers, 2 live tests in 573.13s), Codex CLI v0.1.0 ZIP (full suite in 482.30s), and Claude Code 2.1.271+ (restricted to Read, Grep, Glob tools with CLAUDE.md excluded). Claude Cowork hit usage limits during stages 5–7. Issue #237 documents AGENTS.md leakage into Codex reviewers, highlighting unverified isolation. Unlike GitHub Spec Kit or Kiro, which center human review, EGM’s distinction is mandatory zero-context spec review before code generation; Superpowers isolates subagents mainly at implementation. The plugin is free; costs derive from host model usage. No accuracy benchmarks are published; evidence rests on adoption and end-to-end harness runs.