Hono + Postgres + Electric + Better Auth, on Bun. Anonymous-first identities that upgrade in place.

Full as-built design memo: docs/index.html (published to GitHub Pages once enabled — Settings → Pages → Deploy from branch main, folder /docs). docs/design.md is a plaintext quick-reference.

src/
 index.ts # Hono entrypoint (Bun + node:http dispatcher)
 auth.ts # Better Auth (anonymous + accountLinking)
 env.ts # env validation
 db/
 client.ts # drizzle on bun:sql
 migrate.ts # drizzle migrator (bun-sql)
 schema/ # auth + agent tables
 routes/
 inference.ts # OpenAI-compatible streaming gateway
 sync.ts # Electric shape proxy
 health.ts
 adapters/
 openai.ts # passthrough
 anthropic.ts # OpenAI <-> Anthropic Messages translator
 quota/ # atomic conditional debit + reconcile
 middleware/
 auth.ts # session middleware
 admin/
 index.ts # AdminJS resources
 auth.ts # Express middleware gating /admin behind Better Auth
public/
 index.html # minimal demo page

cp .env.example .env
docker compose up -d # postgres + electric
bun install
bun run db:generate
bun run db:migrate
bun run dev # bun --hot src/index.ts
open http://localhost:3000

bun runs TypeScript directly; there is no build step. .env is auto-loaded.

• POST /api/auth/sign-in/anonymous — seeds user, virtual_keys, quota_buckets.
• POST /api/auth/sign-in/social — OAuth + onLinkAccount re-targets FKs in one transaction.
• POST /api/inference — estimate → atomic debit → SSE stream → reconcile + insert message/tool calls in one tx.
• GET /api/sync/:table — proxies to Electric with account_id = <user.id> pinned server-side.
• GET /api/datasets + GET /api/datasets/:id/download — server-side proxy over a curated catalog of public CSV/Parquet sample sets; the browser writes the bytes into the Pyodide MEMFS at /data/<file>.
• GET /api/me, /healthz, /readyz.
• GET /admin — AdminJS over @adminjs/sql, gated by Better Auth + ADMIN_EMAILS allowlist.

The Datasets menu lists a curated catalog (seaborn-data + vega-datasets, plus a sample Parquet). Selecting one downloads it through the server proxy (Kaggle isn't usable directly — its files need an authenticated API token; the catalog is structured so a Kaggle source can be added later behind KAGGLE_USERNAME/KAGGLE_KEY) and writes it into the Pyodide worker's filesystem at /data/<filename>. Mounted paths are injected into the agent's system prompt so it can pd.read_csv("/data/iris.csv").

/data is an Emscripten IDBFS mount, so files (mounted datasets and anything the agent writes from Python) survive a page reload: syncfs(true) restores from IndexedDB on worker init, syncfs(false) persists after every write. Mounted-dataset metadata is mirrored to localStorage so the chips and the agent's file awareness restore instantly on reload without forcing the ~10 MB Pyodide download; the first time the worker actually comes up it reconciles that optimistic list against the real FS (pruning anything missing, surfacing anything the agent wrote). A Clear button in the Datasets panel wipes /data and the metadata.

• DB driver is bun:sql (built-in Postgres client) via drizzle-orm/bun-sql.
• @hono/node-server is retained for the node:http top-level dispatcher (Bun's node compat handles it); rewriting the dispatch as a fetch-to-Express bridge wasn't worth the bridge code.
• package.json pins every @tiptap/* to 2.1.13 via overrides — AdminJS's design-system declares ^2.1.13 which bun otherwise resolves to a newer API-incompatible version.

• Browser agent (pi-agent-core + Pyodide).
• ECS Fargate IaC.
• Initial drizzle migration files (bun run db:generate creates them).
• Anthropic prompt-token estimation uses chars/4; swap to /v1/messages/count_tokens for precision.