mihneadevries

honestly this is a pretty natural move, mobile access to your coding agent makes sense when you just want to check on a long running task or review something quick.

I've been building something similar, basically a way to run your full dev environment on your Mac and connect to it from iOS. terminal, files, AI agent all talking to the same session. the tricky part is honestly just keeping state in sync when you switch devices.

yeah the cost thing is real, though I think it's moving fast enough that the trajectory matters more than today's price.

like a year ago running an agent on a non-trivial task was painful and expensive. now it's annoying and somewhat expensive. tbh at the rate efficiency is improving I wouldn't bet on "still more expensive than a human" holding for long.

the reviewer/worker pipeline is honestly the part I'm most curious about. like how do you handle disagreements between agents, does the reviewer just block and the worker retries, or is there a loop with a hard cutoff?

the failure mode I'd worry about most is cascading context drift, where each agent in the chain slightly misunderstands the task and by the time you get to the test agent it's validating the wrong thing entirely. fwiw I think the LanceDB memory is the right call for this kind of setup, keeping shared context grounded is probably what prevents most of those drift issues.

The NAT traversal angle is honestly the most compelling part here, WebRTC's ICE/STUN/TURN stack handles weird network topologies pretty gracefully without you having to think about it.

That said I think spzb's point is real, the signaling server is still a trust boundary even if the p2p channel itself is encrypted. Tailscale sidesteps a lot of that by having a more battle-tested auth model tbh.

On the "why would you even want this on phone" thing though, I guess I disagree a bit, I've been building something related: Xtro, terminal + source control + AI agent on iPhone connected to Mac, and like the use case is surprisingly legit once the layout is actually designed for mobile. Emergency deploys, quick fixes, that kind of thing.

Aider's `--yes` flag combined with a git-based loop works honestly better than I expected for this, like it'll just commit and you review the diff. Pi I've tried headless and it's fine but you kinda have to wire up the exit conditions yourself since it's so minimal by design.

Fwiw the janky `claude -p` approach you described is actually pretty solid once you stop fighting it, the simplicity is the feature I think.

Ok, this is nice, but if that terminal running CC is open on VS code for example, does it work?

It's nice to have I guess. But still not as good as just using the Cli in the terminal, while in VS code or other fork, where you can glance the source control from time to time.

It looks like Claude tokenizer handles markdown pretty efficiently, like a ## header is like 1-2 tokens. So I think the actual token savings are smaller than the byte-level numbers suggest. Where this actually matters is if maybe if you're on Haiku with big codebases where system prompt and context fight for space. On Opus? Probably not worth making your files unreadable for humans?

Hmm, I tried asking in Chinese and it still identified itself as Claude by Anthropic.