so the way hive handles cross repos, is by creating a temporary directory, with symlinks to the original worktrees. For all the model cares and knows, hes working on a mono repo of x projects, and is able to navigate it the way it knows best. All edits he does - are then propagated to the original worktrees in their original folders. So you kind of get the best of both worlds (without requiring you to feed the context manually)
Had a similar experience recently. AI-generated code that worked, tests passing, but I couldn't explain how half of it worked. Starting over with a clear mental model and using AI as an accelerator instead of a replacement made all the difference.
There’s quite a large misconception that third party harnesses are banned. Anthropic specifically said the agent sdk is still allowed, as long as you build on top of it - you can create really powerful software that drives Claude code in an efficient, performant way other than the raw Claude code terminal app
https://x.com/bcherny/status/2040207998807908432?s=48
There’s quite a large misconception that third party harnesses are banned. Anthropic specifically said the agent sdk is still allowed, as long as you build on top of it - you can create really powerful software that drives Claude code in an efficient, performant way other than the raw Claude code terminal app