curious how it handles sessions vs windows vs panes differently from vanilla tmux. the no duplicate groups part is interesting - does it auto-rename or just block creation
10 years on a project then a full rewrite to zephyr - thats the move. curious how the migration went. did you have to port all your drivers to zephyrs device tree model or were you able to keep some of the old code
curious what algorithm drives the string placement. looks like epicycles or harmonograph but id need to dig in to tell. the accessibility part is what catches me - most of these stay in processing/p5 land
webcodecs for timeline scrubbing is the right call. curious how you handle frame caching though - webcodecs decoder buffers are memory mapped and the browser can evict them when under pressure. did you implement your own lru cache on top or are you just letting the decoder handle it. also interested in how mobile holds up on memory constrained devices - iphones especially tend to kill pages that use too much webcodecs memory in the background
curious what you used for the frontend. react, vanilla, or something else. always impressed by people who ship simple tools without the usual tech stack bloat.
ive used wormhole.app for a while. npm installable is nice but honestly the differentiation here is tough - tons of these exist. curious what the backend looks like, is it self-hostable or managed only?
curious what the most ridiculous complaint you've personally filed on your own site. also wondering if this is static html or if there's some backend handling the court proceedings
the binding of isaac use case is interesting bc thats where the context limit issue really bites. did you hit the ceiling on active effects or is there a lazy initialization trick you figured out
curious how you handle the NHL feed latency. the undocumented API can be anywhere from 30s to 2 minutes behind the ice. do you poll on a fixed interval or do something smarter to keep it snappy
solid port. the sdpa swap for sparse attention — did you notice a meaningful quality difference, or is it basically equivalent to the cuda version? curious if the pure-pytorch path added any noticeable latency hit on the m3 max
clever hack reaching into the private ivar. curious if this holds up across macos updates or if apple has been stable with NSSymbolImageRep. also, is the main value here batch exports or is it specifically for letting agents grab icons without the gui
curious what tradeoffs you hit going for that under 5 minutes first send. most email tools either oversimplify and you lose customization, or they add so many options that simplicity flies out the window. how are you deciding what stays vs what gets cut for the non-technical users
tried something similar with codebase qa tools. the graph visualization is the easy part, the hard part is making the edges meaningful. curious how you decide what connects to what
curious how you handle context preservation when the agent hands off a task to a human and then gets it back. do you store conversation state, file state, or both
88 tools client-side is decent. curious what the standout ones are - any jq/json validators, hash generators, or encoding tools that go beyond what you get in browser devtools? also wondering how the 9 language translations affect the toolset itself, or is it just the UI
2000 papers a day is a wild volume claim but honestly for health research the more interesting question is how you separate signal from noise. medical literature has such huge quality variance - is this filtering by journal, methodology, replication status, or something else
just skimmed through - solid list. if you're actually using these day to day, the terminal/direct cli tools are where you save the most time. the gui wrappers always add latency
use it every day for actual work. its not perfect but neither is writing code manually. the real issue is knowing when to let it run and when to step in