google/elephant-goldfish-model (EGM) · Google, from an approach described by Dave Rensin · GitHub, v0.1.1, November 6, 2026
Google open-sources EGM, a plugin that reviews specs with fresh-context agents
EGM is an Apache-2.0 plugin for Claude Code, Codex CLI and Antigravity, with experimental Cursor, OpenCode and GitHub Copilot adapters. An "elephant" agent holds the project, requirements and earlier decisions and drafts the specification. "Goldfish" reviewers examine the draft and its referenced source "without the earlier conversation." The human resolves trade-offs and approves the design before anything is implemented. The plugin ships seven skills: brainstorm, draft, resolve, bootstrap, review, review loop and drift report, which compares checked-in documents with the current source.
The README says reviews require native fresh-context subagents and that EGM stops the review if the host cannot create them. It also says "A review does not replace tests or human judgment." The repository showed about 14.1k stars and 1k forks on the evening of November 7. The first tag, v0.1.0, was published October 27. Rensin's original essay is dated April 27, 2026, and carries an October 21 note that Google allowed his team to open-source the plugin.
The release notes are candid about the state of it. v0.1.0 is described as "not ready for stable-release claims," and the exact-artifact install, workflow and isolation checks for the release ZIPs remain open in issue #60. The test record, dated October 22, lists reviewer isolation as passed on Claude Code 2.1.308 and Codex CLI 0.163.0, and as unverified on Claude Cowork cloud (issue #105). Claude Code isolation requires version 2.1.271 or later.
We read the README and release notes. We did not run the plugin, and the project publishes no measurement showing that fresh-context review finds more defects than ordinary review.
Why it matters: The design bet is that a reviewer with no memory of the drafting conversation will catch assumptions the drafter can no longer see. That is testable, and nothing published here tests it yet, so treat EGM as a workflow to trial on one real spec, not as a measured gain. If you adopt it, confirm that your host gives reviewers a clean context, since that is the property the maintainers themselves have not verified everywhere.