Well let me tell you
https://github.com/anomalyco/opencode/blob/dev/packages/open...
You literally send your first message “you are Claude code”
The fact that this ever worked was insane.
Headline is more like anthropic vibes a bug and finally catches it.
But actually the solution is checking out how the official client does it and then doing the same steps, though if people start doing this then Anthropic will probably start making it more difficult to monitor and reverse engineer.
It might not matter, as some people have a lot of expertice in this, but people might still get the message and move away to alternatives.
And if an open source tool would start to use those keys, their CI could just detect this automatically and change the keys and the obfuscation method. Probably quite doable with LLMs..
At some point it becomes easier to just reevaluate the business model. Or just make a superior product.
I believe the key issue here is that the product they're selling is all-you-can-eat API buffet for $200/month. The way they manage this is that they also sell the client for this, so they can more easily predict how much this is actually going to consume tokens (i.e. they can just put their new version of Claude Code to CI with some example scenarios and see it doesn't blow out their computing quota). If some third party client is also using the same subscription, it makes it much more difficult to make the deal affordable for them.
As I understand it, using the per-token API works just fine and I assume the reason people don't want to use it because it ends up costing more.
..right?