Cristian Calabro ran Google's open-source EGM plugin in Codex on a docs-only change, designing and reviewing it before coding, and it caught three things he would have shipped
EGM (Elephant-Goldfish Model) is a design-review plugin Google open-sourced in November: the "elephant" agent holds project context and drafts the spec, "goldfish" reviewer subagents read only the document and referenced code with no earlier conversation, and you decide. Cristian Calabro (@c_calabro) tried it on a tiny request in a repo-checking CLI he's building: make scan --targets show JSON and YAML samples in --help and CLI.md. In Codex CLI he went through brainstorm, draft, review, implementation and a drift report. Results: the brainstorm's seven review lenses warned the four samples could drift apart, so a parity check went into the acceptance criteria; the implementation-plan review noticed the scanner check would write reports into the repo being scanned, against his AGENTS.md, so it moved to a temp folder outside the repo; the closing drift report found the docs never said commit values are compared exactly. On Linux we cloned the EGM repo (0.1.1, Apache-2.0) and ran its offline unit tests: 381 passed, 2 skipped. We did not install the plugin in Codex or Claude Code and run the workflow.
Visit site Elephant-Goldfish Model (EGM) + Codex CLI ↗- Pricing
- Free
- Region
- Global
- Where it's from
- Overseas
- Added
- Nov 13, 2026
- Tested on
- Nov 13, 2026
- Version we used
- google/elephant-goldfish-model git 96e9d6f, plugin 0.1.1; author ran it in Codex CLI with GPT-6 Luna
- Categories
- CodingProductivity
- Original author
- Cristian Calabro @c_calabro ↗
- Our test
- Partly tested
- Cost / time
- The EGM plugin is free under Apache-2.0. Each review spawns several subagents, which uses your model quota in Codex, Claude Code or similar.
The task
When an agent jumps straight to code, unstated assumptions and risks in the spec only show up after shipping, even on a docs-only change.
Steps
- Install the whole plugin (not just the skills folder; the skills read agents/ and resources/ at run time). Codex CLI: codex plugin marketplace add google/elephant-goldfish-model, then codex plugin add egm@egm. Claude Code: claude plugin marketplace add google/elephant-goldfish-model, then claude plugin install egm@egm. Start a new session.
- Describe the feature with egm-brainstorm. It asks a few clarifying questions (he was asked whether samples belong in --help or in an error, and whether to show only required fields), then runs seven review lenses in parallel and writes concepts_brief.md. Have it address the risks.
- Use egm-draft to write design_draft.md section by section: overview, problem, goals and non-goals, technical plan, alternatives, detailed implementation plan, approving each before the next. The final design lands in docs/designs/.
- Review it with egm-review. He picked implementation plan only; it reported two major findings, and he let egm-resolve fix them, producing goldfish_review_report.md to accept.
- Implement (he used superpowers skills in Codex), then run egm-drift-report to compare the design with the code and fill the gaps.
- In Codex you pick EGM skills from the skill picker; in Claude Code and Antigravity use the /egm: commands.
Prompt
/egm:egm-review docs/designs/your-design-doc.md
What the author says
Cristian Calabro @c_calabro (X article, 11 Nov 2026): run in Codex CLI with GPT-6 Luna because the task was simple. On this small task EGM caught three things he would have shipped: the four samples could drift (a parity check was added), the scanner check would write reports into the scanned repo (the reviewer picked up "keep private review material out of the repo" from his AGENTS.md), and the docs never said commit values are compared exactly. During design, typecheck, lint and all tests passed and the scan had no critical or high findings; he did not run the build because egm-draft forbids building outputs during design. Versus superpowers: EGM's goldfish reviews run in parallel with no conversation history and the design is approved section by section; superpowers asks one question at a time and offers two or three approaches.
Our test
On 13 Nov 2026 on Linux we cloned google/elephant-goldfish-model (96e9d6f, plugin.json version 0.1.1, Apache-2.0), set up a Python venv with pytest 9.0.3 and pyyaml 6.0.3 per TESTING.md, and ran pytest: 381 passed, 2 skipped in about 4 s (the skips need local Copilot and Cursor bundles). The tests check skill file format, links, plugin manifests and per-tool install paths; no model is called. The repo has 7 skills (egm-bootstrap, brainstorm, draft, drift-report, resolve, review, review-loop) and 10 reviewer persona files. We did not install the plugin in Codex or Claude Code or run a review.
What we found when we tried it:The plugin package is complete and passes its offline tests, and the install paths match the repo layout. How useful the reviews are needs a run on your own project.
How to get in
The original is an English X article. EGM's docs are at https://github.com/google/elephant-goldfish-model and https://egm.dev/ .
Who it suits
Developers using Codex, Claude Code or Antigravity on real projects who will spend a bit more time on design first; it works on small changes too, at the cost of a few extra documents.
Pitfalls
- Goldfish reviews need subagents with fresh contexts; if your tool can't create them, EGM stops.
- Copying only the skills folder breaks it; install the complete plugin.
- Even a docs-only change produced a concepts brief, design draft, review report and drift report; he kept drafts outside the repo.
- Reviews don't replace tests or your judgment; he approved each stage himself.
- We only ran the repo's offline unit tests, not the workflow in Codex or Claude Code; the three catches are his results.
More in this category
- A demo video per docs page that flags itself when the UI changesVideoScreenplayClaude CodeTool free; uses your agent's quota◐ Partly testedOverseas
- AI backend code that guards against races, leaks and bad migrationsCodingbackendskillsClaude CodePostgreSQLFree; about 2× tokens per task◐ Partly testedOverseas
- An afternoon of AI work, a 3D action game you play in the browserCodingClaude CodeTripoJimengSkill free; Claude quota plus asset tools◐ Partly testedChina
- Keep a long-running AI-built project from turning into a messCodingClaudeCodexoil-delegateNeeds Claude and Codex plans◐ Partly testedChina
- Use a web app once and get an AI write-up of how it worksResearchREAClaude CodeChromeFree tools; needs Claude or Codex◐ Partly testedOverseas