Minerva

An application under development for a single-user, browser-only spatial creative prototype with no sign-in. Minerva is a living atlas of ideas: visible inheritance, three views, contextual creative moves, Weave, Wander, Agent Drive, concurrent voice and usable outputs.

The root atlas provides the mall fixture, three perspectives, editable cards, relationships, layout controls, Wander, Weave, Expedition, typed and voice conversation, and downloads. Atlas data is saved per browser in IndexedDB. See the capability record for verified scope and the handoff for the current candidate.

The application will ship with editable demo data for What to do with a dead shopping mall, demonstrating every tool alongside fully functional capabilities and a blank-start path. The owner judges idea feasibility and creative usefulness while steering milestone chunks. See the demo contract.

Run locally

npx --yes [email protected] [email protected] npm ci
npx --yes [email protected] [email protected] npm run dev

Open localhost:3000. Dev/start bind to loopback; the prepared atlas needs no cloud account, provider key or sign-in. Model and voice calls require the Gateway configuration in .env.example.

The browser is the only client, not the only runtime. Keep the Next.js/TypeScript monolith, model route handlers, server-held credentials and Vercel AI Gateway for text and voice models. External REST/MCP APIs and accounts are excluded. Local loopback is the development loop; the release target is a public Vercel deployment for a bounded demonstration window used by a small judge panel, still with no sign-in. See setup.

Build and review

The default target is a hosted demonstration, defined in AGENTS.md and opened one task message at a time. Build prompts is a backlog catalog of 32 packages toward the complete product; a package applies only when the owner's task message names it.

Astra in Codex builds; the operator starts Fable 5.1 in Claude on a separate checkout of the exact candidate, using the review rubric. Both start at low effort. Use the existing handoff and capability record, not extra process machinery. See setup for copyable launch instructions.

Documentation

Concern	Source
Global working, authorization and review rules	AGENTS
Purpose, complete experience and falsifiers	INTENT
Required capabilities C01-C15 and acceptance scenarios	SPEC
State ownership, boundaries and open implementation choices	ARCHITECTURE
Living-atlas identity and interaction	DESIGN
Implementation evidence	CAPABILITIES
Backlog catalog of packages and milestone reviews	Build prompts
Reviewer system prompt and harness notes	Review rubric
Repository identity, sessions and local operation	Setup

Read the relevant source, not every document each session. Packages reference these contracts and specify outcomes, prerequisites, integration work and completion evidence. The application's short docs/HANDOFF.md names the current outcome and next step.

Check and contribute

npx --yes [email protected] [email protected] npm run check

check runs lint, typecheck, tests and build. Portable tests check document links, matching capability IDs and actual configuration behavior; they do not prove semantic consistency or product behavior. Package 1 removed the empty-seed restrictions. The retained checks allow application dependencies, files and capability progress. Review prose against its owning contract and add behavior coverage as capabilities arrive. ESLint 9 matches the installed Next.js plugin peer ranges; upgrade them together when compatible, without suppressing peer errors or removing lint rules.

Use a branch and PR; merge only after required CI. Repository setup and operating instructions are in setup. Report vulnerabilities privately under SECURITY.md.