Zoo 3, lean and mean

Releasing Zoo 3, the ultimate lean coding workflow for everyone.

Good old:

1. Spec files so the task does not evaporate halfway through.
2. Automatic spec reviews and code reviews.
3. UI verification with browser use.
4. Pragmatic values of Linus Torvalds and Don Melton.

Awesome new:

1. Co-writes spec with user input. User approves before implementation. (Yes, a step back from a fully automated Zoo 2 approach, but definitely a better balance.)
2. Splits work into small reviewable subtasks and commits.
3. Runs tiered reviews: tests, scripted checks, broad review, visual review.
4. Writes proposals for refactoring and out-of-scope issues found.
5. Presents a readable report at the end, including screenshots of all UI changes.

Compared to Zoo 2, this:

• replaces Zoo Heavy, Zoo Lite and Zoo Zero with a single Zoo workflow that’s enough for everyone
• replaces uber-reviews with tiered reviews, where you’re supposed to bring your own scripted checks (more on this later)
• drops per-step Bureau report files for massive savings in token usage; we track state in a detailed task file instead
• drops heavyweight separation of test writing / implementation / problem solving steps; modern models are good enough to do this ad hoc
• rewrites all skills in a super-compact form for speed, token savings and maintainability

Aside from Zoo skills, the release carries two extra ones:

• terse for a compact matter-of-fact communication style cutting down on tokens and fluff (and it’s a very small skill by size); Zoo Init will offer to install this for you
• linus for the mean part of lean and mean; if you missed this fella from Zoo 1, plug it into your Claude, as Codex is too stuck up to call you a fucking moron with shit for brains

Scripted checks

We’ve learned a lot about “harness engineering” this year, and both OpenAI and Anthropic recommend heavy guardrails.

We call these “scripted checks” — scripts that find and report violations of project-specific coding standards. For example, in Bubblehouse, go run ./cmd/fire-check executes both deterministic checks and highly specific LLM-based checks via codex exec. Turns out, requiring an LLM to reply OK/FAIL to 100 specific questions produces more consistent reviews than feeding it a 100-item coding guide. We also use an AST-pattern-matching pre-analyzer which tells LLMs which lines to look at.

That’s the idea. Building your own checks is an entire project, and is optional, but if you do and teach Zoo how to invoke it (in .zoo/zoo.md or AGENTS.md), Zoo Code Review will run it in parallel with the tests and normal agentic review.

Proposals

Zoo 2.3 introduced proposals for refactorings, and Zoo 3 doubles down on them to better resist out-of-scope findings.

When a reviewer finds an out-of-scope problem, Zoo creates a proposal under .proposals/ instead of derailing the active task. This includes security issues, cleanups, and overall system changes.

Installation

Ask your agent to install all skills from Zoo repo, and run Zoo Init.