Get it for macOS

Runs a fresh-context review on every Markdown spec and applies only the changes you explicitly approve. Spec files stay as plain Markdown in your project checkout, listed on a dedicated spec workspace page and in a tab beside each conversation. Seven agent skills cover the workflow from brainstorm and draft through staged reviews, findings and proposals to resolution and drift reports. Reviews run as BB workers on the provider configured for the reviewed project, or as a tool-less OpenAI-compatible endpoint that needs a base URL, model id and API key.


Overview

View source

What you get

A spec workspace for BB that adapts Google's Elephant-Goldfish Model. Your design documents stay as Markdown files in your project checkout. A Spec workspace page in the sidebar lists the registered specs of each project, and an Elephant Goldfish tab beside each conversation shows the same reviews, findings and proposals inside the thread where the work happens. Seven agent skills (brainstorm, draft, staged review, review loop, resolve, drift report, bootstrap) cover the workflow from idea to reviewed document.

How it works

The model separates three roles:

  • Elephant — your normal BB conversation. It knows the project and the decisions behind the spec. It can contest a finding with a rationale or propose a revised document, but it never edits the spec directly.
  • Goldfish — the fresh-context reviewer. Each review stage is prompted with only the reviewer persona, the document standard, the numbered spec and the context files you selected, so it carries none of the conversation's history. In the Advanced endpoint mode those prompt inputs are all it ever gets; a BB worker reviewer receives the same prompt but also keeps BB's normal tool permissions and can read the workspace within them.
  • You — the human who decides. You resolve findings (request fix, accept trade-off with a reason, or dismiss), inspect every proposed diff and explicitly confirm before anything is applied.

Reviews run in rounds: full pipeline, feature spec, implementation readiness, chosen brainstorm lenses, or custom reviewers. A successfully applied change can start the next round automatically once no finding remains unresolved.

Reviewer engines

BB workers are the default. Each stage launches a new BB worker thread in the reviewed project's environment. Workers can use tools and write according to BB's permission mode, so the first launch requires explicit installation-wide consent, revocable at any time. BB worker threads are not a sandbox; revoking BB's own permissions is what changes what workers can do.

Advanced endpoint is the optional tool-less mode: a fresh OpenAI-compatible model call that receives only the persona, standard, spec and chosen context files. It cannot read other files, run commands or change anything, and tool-call responses are rejected. It needs a base URL ending in /v1, a model id and an API key stored in BB's plugin secret store. If the selected engine is unavailable, the launch is refused with its reason; the plugin never substitutes another engine.

What is never automatic

Drafting an Elephant request only fills the composer draft — you review and send it yourself. The Elephant proposes changes; it does not apply them. Applying a proposal always needs your explicit confirmation against its diff, with a version check and a backup, and the plugin never blindly re-applies a change after a restart. No commits, pushes or issue sync are performed. Reviewer context is an explicit reading list; linked files and URLs are never fetched automatically.

Requirements

BB 0.47 or newer, Plugin SDK 0.6.79, and a local project checkout visible to the BB server. Markdown specs live in your repository. BB workers use the provider and model configured for the reviewed project's BB environment; the Advanced endpoint mode needs an OpenAI-compatible endpoint and key instead. The plugin works with local sources and does not operate on specs stored on another enrolled machine.

More from Marco Michely

More in Tasks & Workflows 55

Plan, track, route, schedule, or automate work.

View all