Elephant-Goldfish Model (EGM) for Mistral Vibe

Design with an elephant. Review with a goldfish.

This is a standalone Agent Plugins 1.0 package that ports the Elephant-Goldfish Model (Google, Apache-2.0) to Mistral Vibe. It adds spec-driven development skills to Vibe: the Elephant is your main session (full project context) which explores ideas and drafts specifications; Goldfish are fresh-context subagents that review documents without the earlier conversation, exposing unstated assumptions, hallucinations, and blind spots. You make the trade-offs and approve the spec before implementing.

Install

Vibe discovers plugins in ~/.vibe/plugins/<name>/ (user scope — every project) and .vibe/plugins/<name>/ (project scope — trusted folders only).

One-liner (user scope):

git clone https://github.com/Nosenzor/elephant-goldfish-vibe.git ~/.vibe/plugins/egm

Then run /reload in Vibe (or start a new session). The skills appear as /egm:* slash commands and the reviewer agents as egm:<persona> subagents. Check what loaded with /plugins.

Alternatives:

# Project scope: share the plugin with everyone who clones and trusts the repo
git clone https://github.com/Nosenzor/elephant-goldfish-vibe.git .vibe/plugins/egm

# Manual: download this repo as a zip and unpack it into either plugins dir
# (the unpacked folder must be named "egm" and contain plugin.json at its root)

Update: git -C ~/.vibe/plugins/egm pull then /reload-plugins in Vibe.

Commands

You can use each skill on its own; you do not have to follow the full sequence.

How Goldfish run on Vibe

Each reviewer persona is a plugin subagent in ai.mistral.vibe/agents/*.toml (namespaced egm:<name>, e.g. egm:zoom-out, egm:design-critic, egm:comprehension, egm:structural, egm:style, egm:implementation-readiness, egm:codebase-details, egm:inventive-catalyst, egm:synthesizer). Skills spawn them with the task tool (agent: "egm:zoom-out"), so every review runs in a fresh context window — that isolation is what makes it a Goldfish review. Reviewers are read-only (safety = "safe", only read_file / grep / web_fetch); they return their report as the task result. The skills dispatch reviewers in parallel (flat swarm) and aggregate the findings in the main session.

Reviewing with a different model (the duck)

By default reviewers run on the session's active model. To have a different model review — the same one you use for /peer rubber-duck review — set active_model in the reviewer TOMLs to your duck's model alias:

# ai.mistral.vibe/agents/zoom-out.toml (and siblings)
active_model = "my-duck"    # the same alias you configured under [utility_models] peer

Then /reload. The task tool has no per-spawn model override, so the agent profile is the place to pin it.

Conventions

• Artifact-first: drafts and reports are written to files in the working directory (concepts_brief.md, design_draft.md, *_report.md, docs/designs/<feature>.md); chat only gets the path. Never dump long documents into chat.
• Gates: interactive checkpoints use ask_user_question — strictly one question per call, as the sole tool call in the turn, then yield. Open-ended input (file paths) is asked in plain chat.
• Rubrics: the document standards reviewers audit against live in resources/standards/ (feature-spec, implementation-readiness, zoom-out, heilmeier, re-implementation); shared behavior rules in resources/execution-rules.md.
• Severities are text: Blocker | Major | Nit.

Attribution & license

Ported from google/elephant-goldfish-model (Apache-2.0). Original essay: Dave Rensin, Elephants, Goldfish, and the New Golden Age of Software Engineering. See NOTICE and LICENSE.