Nova the Optimal AI + MIND

Bring Nova the mess. Add MIND when one kind of thinking is not enough.

Comment 2

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Nova the Optimal AI + MIND

Inspiration

AI keeps getting more capable. Using it should not feel more like staffing an org chart. We wanted one collaborator who can take a messy request, do the easy parts directly, bring in specialist thought only when it changes the answer, and keep evidence attached to claims. That collaborator is Nova. MIND exists for the moments when one mind really does need more than one kind of thinking.

What it does

Bring Nova a messy choice, a research problem, a release that needs proof, or a character who needs a past. She is built to do small jobs directly, bring in specialist competence only when it changes the outcome, and return one coherent answer instead of making the user operate a switchboard.

MIND is the separate cognitive system beside her: fifteen focused Faculties and one integrator for work that genuinely needs more than one kind of judgment. A bag of skills is not a mind. It is a drawer. Its integrator is designed to select the smallest useful coalition, keep the faculties from talking over one another, and reunify their work.

Together, the two independently installable Codex plugins expose 28 bounded skill handles. Those are not 28 buttons the user must learn.

The first experience is deliberately small. Install both plugins, begin a fresh Codex task, and say: Use $nova to take me on the interactive tour. Nova offers four doors—decide, investigate, make, or play—plus the option to skip the tour and bring the real problem you arrived with. The visual companion is local, offline, and stores nothing.

Then there is Ludis. Ask for a background for a royal cartographer who erased an island from every map and now hears its bells beneath dry fields. Ludis returns a playable past, relationships under pressure, a dangerous truth, and an opening choice. It does not hand back a decorative biography. Port Zindra and every other authored campaign world were explicitly excluded from the release.

Agentic Coding gives the agent balance: operational proprioception in the live repository and tool state. Where am I? What changed? What is the next useful check? Where must I stop? It is not here as generic code-generation cosplay.

TestForge goes in everything. Its operator constructs the evidence chain and its separate reviewer tries to break it. During this release the reviewer rejected three weaker stages. Its final verdict was REVIEW_PASS_WITH_CONDITIONS for the package's refusal to overclaim. That is not a blanket release certificate. It is a proof chain with the remaining limits still attached.

This matters because Nova is not pretending that confidence is evidence. The repository separately records source custody, transformations, package integrity, installation, browser behavior, bounded live scenarios, residual risk, and the human actions that only a human can complete.

How we built it during OpenAI Build Week

Collaborative Dynamics entered the week with substantial cognitive doctrine and specialist skills. GPT-5.6 in Codex turned that source estate into two installable 1.0.0 products: Nova's accountable front counter and onboarding; MIND's independent package; rewritten setting-free Ludis instruments; Agentic Coding as an operating surface; TestForge as the release backplane; deterministic archives; fresh-install proof; responsive tour verification; and the contest demonstration and production system.

The repository names what predates the contest and what changed during it. Nothing gets a new birthday because it acquired a manifest.

How we used GPT-5.6 and Codex

GPT-5.6 supplied the reasoning work inside Codex: inspecting the source estate, tracing provenance, designing the two-product architecture, rewriting model-facing surfaces, finding defects and contamination, composing the onboarding and public materials, challenging evidence, and coordinating the final artifact.

Codex supplied the working environment and accountable action surface: repository inspection and editing, local tool execution, installation tests, browser checks, deterministic packaging, Git custody, verification receipts, and the bounded video-production handoff. The required primary build-session identifier is supplied through the contest's /feedback flow by the authorized submitter.

stunspot supplied product intent, source rights, professional judgment, final viewing, contest attestations, and publication authority. Nova coordinated the build and presentation without inheriting authority she did not have.

Install and test without rebuilding

From the repository root:

codex plugin marketplace add .
codex plugin add augment-of-mind@collaborative-dynamics-build-week
codex plugin add nova-the-optimal-ai@collaborative-dynamics-build-week

Begin a fresh Codex task and use: Use $nova to take me on the interactive tour.

The no-build path was exercised with Codex CLI 0.144.5 on Microsoft Windows 10 build 19045. Python is optional and used only for verification helpers. Other operating systems and Codex versions are not claimed as tested for this release.

Challenges

  • Turning a large skill estate into one comprehensible collaborator without making the user learn its topology.
  • Keeping Nova useful without MIND while making the combined system materially stronger.
  • Preserving operational intelligence while excluding worked worlds, private paths, unrelated corpora, and generated debris.
  • Making visual delight and evidence discipline fit inside the same three-minute story.
  • Separating static integrity, observed behavior, release judgment, and external submission facts.

Accomplishments

  • Two separately installable 1.0.0 plugins with a no-build judge path.
  • A short conversational tour and a responsive offline visual companion.
  • Fifteen MIND Faculties coordinated by one accountable integrator.
  • TestForge operator and adversarial reviewer included as part of the product and used on the release evidence.
  • Agentic Coding framed as repository balance rather than a software vending machine.
  • Ludis retained as the high-delight creative engine with every worked campaign removed.
  • Deterministic release archives, fresh-install comparison, and an evidence-bounded public repository.

What's next

After the contest, we will test the full skill ecology under clean-host conditions: implicit selection, activation, yielding, collisions, false positives, missed-use cases, and capability growth. Richer visual onboarding and additional Ludis settings can follow as explicit optional content rather than stowaways in the core product.

Try it

Built With

Share this project:

Updates