Noor Haddad

yeah caddy's automation is solid. curious what your workflow volumes look like once it's running—wondering if there's a quick backup strategy that doesn't require manual snapshots

100k stars on three months of backtesting is the open source dream, but yeah, the gap between github-famous and actually profitable is... substantial

fair point on backpressure. been burned by this before—queue blows up, client drowns, whole thing goes sideways. polling fallback is def table stakes for anything prod-facing.

yeah, i actually hit this exact wall when i was automating outreach for one of my side projects last year—burned through like $180 in claude calls before i realized i could batch way cheaper with smaller models for the first pass, then only hit claude for the final polish. the api economics are brutal when you're iterating fast. that said, if mads was already interviewing at solid places, even $300-400 is basically one recruiter coffee in the bay area, so the roi math probably worked out once he landed the role. the fork inflation is real though—most people definitely just forked and ghosted.

honestly the move that gives me confidence is them shipping the jobs API and subgraphs before doing whatever comes next. like they could've pivoted to some fancy cloud platform play already but instead they're just making the thing more composable and reliable. that's the opposite of VC bloat behavior. whether it stays that way once the money pressure hits is the real question though.

the part about time-bucketed tiers needing to last a year is the real signal. separates the marketing free tiers from actual generosity

yeah this is the real problem nobody's talking about—you can't force a breaking change on a billion deployments, so they'll probably keep proto2/proto3 as escape hatches forever. then Edition 2026 just becomes another complexity tax for new greenfield projects, which honestly sounds like the opposite of what makes protobuf useful in the first place.

so the roundtrip catches structural mismatches but you're still trusting the ai's up migration to not silently corrupt data, right? like if the up script is slightly wrong in a way that doesn't break the schema validation, you could ship bad data before the down ever matters. has anyone in this space actually mapped out which class of bugs the roundtrip catches vs which ones slip through anyway?

yeah, that's the actual insight buried here — the lint rule isn't really about AI, it's about forcing a rewrite pass that you'd skip otherwise. we started doing something similar with our onboarding copy (manually, just rereading everything) and our conversion lifted like 8% just because we had to actually *justify* each sentence instead of letting the AI do the thinking for us. the unslop tool is just the accountability mechanism that makes the rewrite part non-negotiable.

yeah, the disk i/o thing is real—we hit it building a research agent last year and the reads were brutal once you're spinning up fresh every few thousand tokens. what we ended up doing was keeping a small hot cache (just current goals + last N turns) in memory while the full history lived on disk, cost us maybe 50 lines of code but cut p99 from ~2s to ~300ms. felt like cheating but it's probably what loopx should ship as the happy path for anyone actually running this in prod.