Operating note4 min read
Taking Handover from a UI prototype to 348 PRs (and killing our ZIP64 engine)
Co-building handover.media across 348 PRs on TanStack Start, Convex, and Cloudflare Workers — why we built a full server-side ZIP64 pipeline, then ripped it out.
When people talk about building a product, they usually stop at week one: the stack diagram, the first working upload button, maybe a launch tweet.
Looking through the Git history of Handover — a client-facing delivery platform for sharing structured text, files, and media galleries with password protection and download tracking — week one was barely the warm-up. The real story sits inside the 348 pull requests and more than 1,800 commits that came after it.
Co-building a real product (humans and agents in the same commit log)
Handover isn’t a weekend toy. I’ve been co-building it alongside Edwin Towler since November 2025, starting from a front-end prototype, moving through an early Next.js + Supabase version, and eventually settling on three runtimes that each own one clear piece of the problem:
Browser
-> TanStack Start (UI, server functions, HTTP routes)
-> Convex (auth, deliveries, logical assets, quotas, permissions)
-> Storage adapter (Cloudflare R2 for source bytes, Stream for video playback)
Cloudflare Worker (same app worker)
-> TanStack Start SSR + server routes
-> Queue consumer: R2 object-create reconciliation -> Convex confirm
-> Scheduled jobs: bin cleanup, Stream deletion drain, storage recalculation
The rule that kept the whole codebase legible across 348 PRs is simple: Convex owns truth and authorisation; storage providers only own bytes. Deliveries and rich-text blocks never reference raw bucket keys directly — everything goes through logical asset IDs checked against Convex permissions at the boundary.
One funny sign of the times: if you run git shortlog -sn on the repo, right underneath my commits and Edwin’s commits sit Cursor Agent, Claude, and T3 Code with hundreds of commits between them. Having a written ARCHITECTURE.md and PRODUCT.md with hard invariants wasn’t bureaucratic documentation — it was what let two humans and several coding agents ship quickly without quietly undermining each other.
Killing our darlings: the ZIP64 reversal
My favourite engineering-meets-product moment in the entire repo is something that isn’t even user-facing anymore.
Early on, we assumed clients downloading a whole delivery or photo gallery needed a single server-generated ZIP file. On Cloudflare Workers — where you can’t just hold multi-gigabyte archives in RAM — doing that reliably is a serious systems problem.
So we engineered it properly:
- A durable Cloudflare Workflow (
prepared_exports) keyed by logical export ID and immutable source generation. - Bounded concurrent 4 MiB R2 range reads calculating CRC32 with 16 MiB durable checkpoints.
- Deterministic ZIP64 layout uploaded back to private R2 via concurrent 8 MiB multipart uploads.
- An independent ZIP64 verifier that parsed the central directory, local headers, ZIP64 extra fields, and recomputed CRC32 over the stored output bytes before Convex would ever mark an export
ready.
It was genuinely solid engineering. And then we stepped back and looked at the actual product experience: waiting for a server to pack gigabytes of already-uploaded originals into a duplicate temporary ZIP, paying to store and expire those copies, and dealing with giant single-file download drops wasn’t actually better for recipients.
So we made the call to kill server-side gallery and delivery ZIPs completely.
Instead, individual originals are downloadable the moment they finish uploading, and in Chrome or Edge recipients can click Download all to stream original files straight into a local folder via the File System Access API — zero server-side re-packing, zero duplicate storage, zero waiting for a background job.
Then we turned the remaining workflow machinery into an auditable, retryable admin cleanup tool that safely expires and purges any leftover prepared archives without ever touching a single customer original or messing up source-storage quotas. Throwing away clever code you’re proud of hurts for about ten minutes; living with a simpler product pays off every day after.
The sneaky bugs of going live (hndvr.io → handover.media)
When we moved production from hndvr.io to handover.media, the easy bits were spinning up a tiny standalone 308 redirect Worker (handover-hndvr-redirect) to preserve every old shared link’s path and query string, verifying send.handover.media in Resend, and temporarily allowing both origins in R2 CORS.
The sneaky catch? Passkeys. WebAuthn credentials are permanently bound to their relying-party ID (hndvr.io), so moving domains means existing passkeys literally cannot work on handover.media — users have to sign back in via magic link or Google and register a new passkey. Exactly the kind of detail you only appreciate once you migrate a live domain with real accounts.
The other 90%: everything outside the storage engine
Once the upload and download path is rock solid, the work stops looking like distributed systems and starts looking like running an actual business:
- Plain English everywhere. Internally a delivery status is
inactive,active, orexpired. Across every screen, filter, and menu, we locked the vocabulary to Draft, Live, and Expired — and recipients never see the word "draft" at all, just that a link isn’t available. - Real billing and abuse edges. Stripe pricing, trial-abuse protections, account suspension and delivery takedowns, 24-hour password unlock windows, and letting owners preview images inside their own drafts (PR #348) without weakening public read security.
- Respecting users while actually measuring the funnel. Cookieless PostHog until someone accepts the consent banner, a before-and-after hero replacing a clunky page tour, and three short plain-text lifecycle emails from Ed across a user’s first fortnight — off by default behind an env flag, gated by a rollout timestamp so they never spam older accounts, and carrying signed one-click unsubscribe links that work without logging in.
Launching a prototype is maybe 10% of building a product. The other 90% is being willing to delete your own clever features, sweating the passkey migration edge cases, and quietly polishing 348 pull requests until the thing feels effortless to the person opening a link.