@karpathy then it is conceivable that individuals that are prone to certain kinds of thinking are susceptible to misinformation campaigns and this (llms) will make it worse
opencode via `opencode-anthropic-auth` (now archived at request) mimics headers for claude code which is against TOS.
anyone build a claude code wrapper (either via --sdk-url or agent sdk) for opencode?
@mattpocockuk you don’t need to get oauth token for claude agent sdk locally — it manages auth for you. though i wouldn’t do it the way @opencode does it. they construct headers to mimic claude code api calls and i’ve read about people getting banned but personally don’t know anyone that has
i'm throwing in the towel for competing in the agent orchestration space!
open-sourced my take on it (https://t.co/jKHGqG5kl0) and will be focusing my efforts on building on top of @opencode as there's no need to reinvent the wheel on this aspect of things...
just realizing that their web ui looks really similar to mine haha
finding out that vibe coding can cause comprehension debt
in many cases it actually doesn't matter but what do people do about it
several strategies that I have include actually reading changes, diagramming, etc. thinking about having a 1-1 comprehension bot that tests my understanding of a repository.
claude code has a hidden --sdk-url flag that exposes websocket functionality. this is awesome for building wrappers but because it's not officially documented it's not future-proof.
i'm building a messaging layer protocol to work back into my agent orchestrator purely using print mode + session continue. same functionality, future proof.