Build your own brief-to-app generator (a personal Lovable)

Build a Node 22 CLI called vibe. Requirements:
- package.json with "type": "module" and only two dependencies: commander for argument parsing and dotenv for loading config.
- src/index.js registers three subcommands: new <brief>, undo, preview, each stubbed with a console.log so node src/index.js new "test" runs and exits 0 today.
- src/lib/config.js reads MODEL_API_KEY and VIBE_MODEL_PROVIDER from a root .env file and exits with a clear message naming the missing variable.
- src/lib/slug.js converts a brief string to a lowercase, hyphen-only, filesystem-safe slug capped at 40 characters.
- An npm test script that invokes the CLI once and asserts exit code 0.
- Honest pain warning: resist adding chalk, ora, or any UI library now, every extra dependency is something later steps must debug.
- Out of scope: model calls, web UI, telemetry, progress spinners.

Add the brief-to-files engine to the vibe CLI from step 2. Requirements:
- src/lib/generate.js calls the provider named in .env using the official SDK, @anthropic-ai/sdk or openai, with the key read from MODEL_API_KEY and never logged.
- The system prompt demands exactly one JSON object back: {"files":[{"path":"src/App.tsx","contents":"..."}]}, no prose outside the JSON.
- Strip any markdown fences, then JSON.parse; on failure write the raw response to .vibe/last-response.txt and exit 1.
- Write every file under ./apps/<slug>/ using src/lib/slug.js, rejecting any path containing .. or leading slashes.
- Generated apps are Vite + React + TypeScript with package.json, vite.config.ts, index.html, src/main.tsx, and src/App.tsx at minimum.
- Generated apps persist data with better-sqlite3 into a single app.db file, declared in the generated package.json.
- Honest pain warning: models wrap JSON in ```json fences far more often than you expect, handle that case before anything else works.
- Out of scope: streaming responses, multi-provider fallback, conversation memory.

Add a verify-and-repair loop to src/lib/generate.js. Requirements:
- After files are written, run npm install, then npm run build, then npx tsc --noEmit inside ./apps/<slug> using child_process.execSync with output captured.
- On any failure, send the combined stdout and stderr back to the model asking only for changed files in the same files-array JSON format.
- Retry the full generate-and-check cycle at most 3 times, then print the last compiler error and stop.
- Print one status line per attempt: attempt number, failing command, and error line count.
- Record the attempt count in .vibe/session.json, creating the file if missing, so reruns start clean instead of inheriting stale counts.
- Honest pain warning: tsc output can overflow the model's context window, truncate each error blob to its first 4000 characters before sending.
- Out of scope: unit tests for generated apps, linting, formatting tools.

Add git safety to every vibe round. Requirements:
- When vibe new scaffolds ./apps/<slug>, run git init and make an initial empty commit on main.
- After a round passes the verify loop from step 4, print git diff --stat for the pending changes, then git add -A and git commit with message round-<N>: <first 50 chars of the brief>.
- vibe undo runs git revert HEAD inside the most recently used app directory and refuses to run when git status reports a dirty worktree.
- On a dirty tree, print instructions to stash or discard instead of attempting the revert.
- Honest pain warning: revert on a dirty tree fails with a message nobody enjoys reading, catching it early is the whole feature.
- Verify by generating twice and confirming git log shows the initial commit plus two round commits.
- Out of scope: pushing to remotes, pull requests, merge conflict resolution.

Add conversation memory and cost accounting to vibe. Requirements:
- Store the transcript, current file tree, and last diff stat in .vibe/session.json inside each app directory, rewriting it after every round.
- On follow-up rounds to an existing app, resend only files the user names by path plus the session summary, never the whole tree.
- Read token usage from each API response, normalize the Anthropic and OpenAI usage field shapes in one tested function, and store totals in session.json.
- Compute dollars per round from src/lib/prices.json, which maps model name to input and output price per million tokens, and print the figure with 4 decimal places.
- Print an inline warning whenever a single round exceeds 0.25 USD so bill surprises surface immediately.
- Add vibe new --continue <slug> to resume the stored session instead of starting cold.
- Honest pain warning: usage fields differ between providers and drift across versions, pin SDK versions in package.json.
- Out of scope: embeddings, vector stores, semantic memory search.

Finish vibe with a preview command and a README. Requirements:
- vibe preview spawns npx vite --port 5173 inside the named app directory with stdio inherited, and exits when the server process exits.
- Refuse to preview an app whose last verify round failed, printing which command failed instead.
- README.md at the repo root documents the required .env variables, a typical total cost per generated app computed from src/lib/prices.json, and a plain paragraph stating that output quality equals the model behind the key.
- Include a 60-second quickstart: install, set the key, generate one app, run preview, open localhost:5173.
- Verify with npm test plus one manual preview run against an app generated in earlier rounds.
- Honest pain warning: port 5173 is Vite's default and collides if another instance runs, detect and report the busy port.
- Out of scope: hosted URLs, HTTPS certificates, deployment targets, Docker.