@estebanforge/pi-egm

The Elephant-Goldfish Model (EGM) on Pi: Google's open-source spec-driven development workflow — the same plugin thousands of Google engineers use to draft specs with deep-context agents and stress-test them with zero-context reviewers before writing code. Upstream source: google/elephant-goldfish-model.

An Elephant (your Pi session, holding project context) drafts the specification through structured brainstorming and probing questions. Goldfish reviewers — fresh subagents with zero conversation history — read only the draft and the live source, stress-testing the design from conflicting perspectives to expose unstated assumptions and blind spots while changes are still cheap. You make the trade-offs and approve the spec; only then does a coding agent build it.

Vendored verbatim from google/elephant-goldfish-model (Apache-2.0). The only content changes are one harness pointer per skill and the Pi tool name in the reviewer-dispatch enumerations. PI.md carries the Pi dispatch notes; PI.md-related concerns live there, not in the skill text.

Skills

Every skill also loads as a Pi skill (/skill:egm-<name>); the commands exist for parity with the upstream Antigravity and Claude Code UX. The philosophy behind the model is explained at egm.dev and in Dave Rensin's article; the full workflow docs live in the upstream repo.

Requirements

• None beyond Pi. egm_goldfish ships in this package: each reviewer runs in a fresh headless child process with no extensions, no context files, no discovered skills or prompt templates, no persisted session, and the edit tool disabled (bash and write remain, so reviewers can search and deliver their report but cannot modify documents in place), so the zero-context guarantee does not depend on any subagent extension. One known residual exposure: pi still loads APPEND_SYSTEM.md files into the child's system prompt and there is no flag to suppress them. The egm_goldfish tool accepts optional model and thinking_level parameters to run a reviewer on a cheaper pi model; unset means the pi settings default model (provider and keys come from your Pi settings).
• During EGM flows the extension blocks spawn-shaped tools other than egm_goldfish, so a partially compatible subagent extension cannot silently fork context into a reviewer. The guard is a best-effort backstop: once EGM artifacts have been touched in a session (a read of the bundled skills, personas, standards, or PI.md, an egm_goldfish call, or an /egm-* command) it stays armed until the session ends, because EGM flows are multi-turn and plain answers between stages are not flow exits. It is name-based, so a tool with an unusual name could still be called — residual risk we accept rather than chase per-extension compatibility.
• If you run pi-subagent-manager you will notice EGM does not use it: its spawned sessions load your personal package stack, which can inject memory context into a reviewer.

Install

pi install npm:@estebanforge/pi-egm

Start a new session, then type /egm-brainstorm Add webhook delivery to this project. — the command prefills the editor with the skill prompt (press Enter to run, same as the fleet's command pattern). Continue with:

/egm-draft
/egm-review docs/designs/webhooks.md

Commands need an interactive session; headless runs should read the vendored skills/<name>/SKILL.md directly.

Optional: add EGM artifact directory: .dev/docs to your AGENTS.md to keep briefs, drafts, and review reports out of the repo root.

Provenance and license

• Skill, persona, and standards text: google/elephant-goldfish-model (egm.dev), Apache-2.0, Copyright Google LLC. Upstream is the canonical home for content changes.
• Pi integration (PI.md, extensions/, packaging): EstebanForge, Apache-2.0.
• Not an officially supported Google product.

Develop

npm run check          # from the monorepo root: typecheck + tests
npx vitest run packages/pi-egm/tests   # this package's structure tests
pi -e ./packages/pi-egm                # try it without installing