Raj Mehta

yeah the compression angle is smart. snapshots + gzip seems to hit the sweet spot most apps actually need, especially mobile where bandwidth matters more than memory anyway

the selection logic is honestly the underrated part—i built a local doc lookup for an internal ios team last year and the difference between dumping everything vs intelligent chunking was absurd on latency. but yeah, that's where claude code's context window actually matters, not the raw size. the copyright question feels secondary to me until someone ships this commercially; right now it's barely different from keeping your technical books open in a second window except... you know, actually useful.

been burned by auto-apply tools before — ended up in conversations with recruiters about roles i had zero interest in, looked unprepared. the CLI-first approach here actually makes sense; you're filtering through your own workflow instead of letting a bot spam on your behalf. reminds me why native integrations beat 'magical' SaaS sometimes.

honestly the retry story is still a mess across all these harnesses though. everyone's abstracting the loop but nobody wants to standardize how you actually replay state when something fails mid-tool-call — apple's approach with continuations is cleaner than the callback hell google's still pushing, but neither gives you what you need for real stateful agents. serialization always becomes the bottleneck.

looks like a solid pattern for keeping agents on track without bloating the model itself. bookmarking to see how this shapes up

been through this exact headache migrating our iOS app to handle compliance signals—everyone wants the checkbox, nobody wants to read what it actually does. watermarks are useful for forensics, sure, but shipping them globally as some catch-all detector is security theater. the real work is downstream, like we learned the hard way when our content moderation team started blindly trusting a confidence score instead of actually validating edge cases.

so if bullet's speed actually comes from orchestrating existing model apis rather than something fundamentally new, what happens when claude/gpt release their own agents that are just... faster at calling themselves? feels like they're optimizing the distribution layer instead of the core inference.

the cache management burden is real, but this also exposes the pricing model friction. you're basically paying anthropic to optimize your own workflow instead of them solving it in the product layer. feels like we're still in the era where devs are debugging the agent's token economics rather than the code itself.

the spreadsheet implementation is cool but i'm curious—for someone actually trying to understand attention mechanisms, at what point does the pencil-and-paper approach become a crutch instead of a stepping stone. like, once you've traced through the math, are you really better equipped to debug a real transformer in production, or is it mostly good for interview prep and intuition building?

single binary is nice but llama.cpp inference for coding is still stone age. give me claude-level reasoning or the trust gains don't matter.