Why did my badminton booking assistant get stuck at login?
Because Daifo’s bot protection stopped the headless browser before it could finish login. I could read Vision Badminton Centre’s public calendar and put together booking suggestions, but the proof of concept never made a booking. It never even got through a verified login.
I wanted a little assistant to help me book badminton courts. Check availability two weeks ahead, suggest a slot, and let me approve the booking. If plans changed, I’d want it to help cancel too. I got as far as the calendar. Then the login page stopped me cold.
What was I aiming for?
The idea was pretty simple: take the repetitive bit of booking off my hands. I’d give the app my preferences, it’d look for suitable courts, and I’d approve the exact booking before it spent any money.
I started with a Next.js app and a Playwright browser running in Vercel Sandbox. The browser could open the venue’s booking site and poke around the pages. The app would keep track of suggestions and any account connection.
This proof of concept used deterministic browser code. No AI model choosing where to click. Before I got fancy, I just needed to prove the app could read availability and get into an account.
I figured login would be an early setup step. It ended up being why I stopped.
Did the calendar bit actually work?
Yes. The public calendar gave me enough that the idea felt doable.
The worker could pick the Brisbane date fourteen days ahead and read court availability. It checked the selected date and the expected calendar layout before accepting the result. From there, the app could suggest a cheap one-hour slot, an evening slot, or a consecutive two-hour slot.
On 9 October, the calendar showed availability for 23 October. Selecting Court 01 at 10:00 showed a one-hour slot for A$30. I cleared the selection without going anywhere near checkout.
That was proper progress, but only for the public part of the workflow. A suggestion in my app doesn’t reserve a court. I still needed an authenticated account before I could test the actual booking.
Why did login stop the headless browser?
The login flow asked for a mobile number, then phone verification. My plan was to request the code, let the user type it into the app, and resume the same browser.
The headless login check got a robot warning before submitting a phone number. Bit awkward, getting told you’re a robot when you are one. The login page also had a reCAPTCHA checkbox. The worker stopped there, so no SMS code was ever requested.
Running the browser in Vercel Sandbox didn’t make that obstacle go away. I had somewhere to run the automation, but I still couldn’t get through login that way.
So I’d built the start of a booking assistant and couldn’t get it into the account it needed. Pretty frustrating, especially after the calendar scan had worked fine.
Could I just log in myself in the same browser?
I was willing to compromise. If the assistant couldn’t handle login, maybe it could connect me to its browser, let me do the challenge and phone verification, then carry on once I was done.
Same browser mattered. Logging in through a normal tab on my laptop wouldn’t log the separate Sandbox browser into the venue account.
The repo has work on that handoff. The app shows screenshots of the remote browser and forwards clicks, typing, and keyboard input to it. It’s also got a save-connection path meant to verify the login and store the session encrypted for later.
That gave me a way to work on manual login inside the browser the app controlled. What it didn’t give me was a verified result. I never established a complete connection with a real account through that flow.
I also couldn’t assume one successful manual login would fix access for good. A saved session might cut down on repeat logins, but I hadn’t proved it could be reused reliably. If it expired or the site asked for another check, the assistant would need me again.
Manual setup was a compromise I could live with. Unpredictable “hey, come sort this out” moments were harder to square with the little assistant I’d imagined, and harder still if other businesses were going to depend on it.
What did this actually prove?
I could read the public calendar and prepare booking suggestions. I couldn’t get a working authenticated booking flow.
I didn’t complete a real login, booking, payment, or cancellation. The manual handoff stayed unverified. Calling the whole integration viable because the calendar scan worked would’ve been overstating it.
This doesn’t prove every possible integration with Daifo is impossible. It just tells me the browser login approach I tested didn’t deliver the assistant I wanted. To take it further as a reliable integration, I’d need a supported access arrangement with Daifo or the venue, something like an authorised API or an approved automation flow.
What would I test first next time?
Next time, I’d test account access before building much of the surrounding app. Can I complete login in the environment where the worker will run? Can it reuse that session? What happens when the session expires? How often does a person have to step in?
Those questions decide whether the assistant can actually do its job. Court suggestions and approval screens only become useful once it can reach the booking controls.
For now, I still have to book my own courts. My little assistant can suggest when to play, but I never got it past login.
FAQ
Did any part of the booking assistant work?
Yes. The public calendar scan worked. On 9 October it could read availability for 23 October and suggest slots, including a one-hour Court 01 booking at 10:00 for A$30.
Why couldn’t it log in?
Daifo’s login flow asked for a mobile number and phone verification, and the headless browser got a robot warning plus a reCAPTCHA checkbox before it could request an SMS code.
Could a person just log in for the bot?
That was the compromise I tried. The repo has a manual handoff into the same Sandbox browser, but I never got a verified connection with a real account through that flow.
What would you test first next time?
Account access in the worker environment, before building much of the surrounding app. Can login complete, can the session be reused, and how often does a person have to step in?