ByteCode.NewsSubmit
Use quotes for an exact word.

Subscribe

RSSPodcast
Log in

Chat Logs

  1. bluestagas: if you like kilocode I'll check out if it has a server mode, because that's needed for repopack chat
  2. blueI only support opencode now because that's what I use. but I plan to add support for codex and claude code
  3. stagasit has "local model server" i find in the settings
  4. stagasis it that? or something else
  5. bluethat could be it
  6. blueit's like a daemon mode
  7. stagasthey might not have it
  8. stagashttps://ibb.co/XZ3m2MLQ
  9. nevetimage hosted at ImgBB
  10. stagasthat's all it has
  11. bluestagas: "Yes, Kilo Code uses a local server called kilo serve on your machine, as well as optional cloud services"
  12. blueapparently it's enough to run `kilo serve`, very cool
  13. stagasah nice
  14. bluebecause these servers are the prerequisite for hosted agents and so on
  15. stagasit's so much stuff it's hard to keep up with all that
  16. blueso basically the idea is, currently I have opencode running locally. and I use both its tui app and also it's a server. I have like four tmux sessions running three tabs (I have one tmux session per repopack project, one tab is implementer, one tab is reviewer, one tab is 'security reviewer')
  17. bluerepopack chat replaces the tui part. it connects to my local opencode server, so I can get rid of tmux and all those tabs and sessions that confuse me
  18. bluebut the server part still runs on my computer. so when I restart or whatever, everything is lost. also, my computer isn't the strongest
  19. blueso the idea of hosted agents is, I run the opencode server on some vps
  20. blueI can turn off this computer, go to sleep, and continue exactly where I stopped
  21. stagasyes i thought about running codex cli also remotely, didn't know it can be a server, so then the desktop/cli app connects to that?
  22. blueyes
  23. bluethat's the idea of repopack chat
  24. blueyou run codex cli as server, then repopack chat can connect to it
  25. blueyou can run it on your computer or any vps
  26. stagashope openai give us a reset to celebrate 6.2 :P
  27. stagasyou saw the launch article? it improves in all benchmarks over 6.1, finishes tasks with less tokens, and i see it is faster. so it's faster, cheaper and better
  28. blueI haven't seen. I will switch to it soon and also read the article, very exciting
  29. bluenice, opencode already has sol 6.2
  30. bluestagas: what is -astra? is this new?
  31. bluestronger than sol?
  32. stagasit's not new it was in 6.1 also
  33. stagasbut it's very expensive
  34. stagas$100 i think vs $20
  35. blueI also see in pricing that they no longer have terra and luna for 6.2
  36. bluemaybe they retired them and only have sol and astra now?
  37. stagasprolly nobody is using them since they don't deliver. luna 6.0 to change a variable is fine
  38. blueyup
  39. bluestagas: I got codex installed and connected. adding support for it to repopack chat
  40. stagasnice
  41. stagasbtw you saw this? https://runta.com/blog/introducing-frontierharness-eval-v1-1/
  42. stagasapparently codex cli (and probably the desktop app) uses the least tokens
  43. bluewell in that case stagas, it makes sense for repopakc chat to support codex too.. and I'll switch to it as well, since I'd be getting the same ui anyway via the web
  44. stagasi can't believe how good codex is at designing 3d meshes from photographs
  45. stagaslike one could just do that and sell them
  46. stagasit's very hard to find clean 3d meshes
  47. stagasfree
  48. blueinteresting
  49. stagasand i wouldn't be able to find these things anywhere, like wastebins, the trams, ubahn sbahn, police cars, ambulances, they're all from berlin photographs
  50. stagasauthentic
  51. stagasnevermind all the landmark buildings
  52. stagasthe skyscrapers the antenna
  53. phalethare those meshes triangular? in other words made of triangles, or is it more like a solid geometry made out of closed B-spline surface volumes?
  54. stagastriangles
  55. stagasi've made the first rewarding gameplay mechanic, you collect beer bottles you find under wastebins
  56. stagaswhich is so characteristic
  57. stagasyou'll probably use that score to exchange other goods
  58. stagaslike buy a doner kebap
  59. stagasor pay the entrance to the club
  60. phalethgermans are well rewarded for bottles at malls, would not leave them on streets
  61. phalethI guess if it's not made of too many triangles then those models should be efficient, but when it's a bottle with cylindrical shape or something like that not too sure
  62. phaleththe cool thing about those meshes is that they can be baked into wasm modules directly and then loading them in web browser is fast
  63. phalethno need for an obj file or whatever 3D mesh format
  64. stagasi have no idea how it's doing it, or if it's generating them at runtime, there are some precomputed like the hair and character rigs and walking/running animations that are json.gz
  65. stagasbut i'm spending a lot of time optimizing
  66. stagasso whatever the bottleneck (pun intended) we are solving it early
  67. stagasbecause my pc is crap and i want good fps as well
  68. phalethjson.gz, hmm, sounds like threejs is used behind the scenes
  69. phalethI remember digging into json.gz files alot when I worked with a three-js based viewer
  70. stagasnein threejs would have been a dead end. it's all custom shaders
  71. stagasi'm rendering a TON of stuff
  72. phalethah, ok, then it's just the format that is similar, I expect in those json.gz files you will find lots of arrays with numeric values that can be processed in parallel
  73. stagasjson.gz saves you from decoding the .gz because the browser does it, and saves you from having a mesh format decoder, which would be 5mb and would take time to decode
  74. stagasand the sizes are fine
  75. phalethyeah, size of an obj file is too much for the browser, that's for sure
  76. stagasthe walking/running/swimming and dance moves were all .fbx before
  77. phalethanyway, all those models are static, so you just need to separate out textures from there and the geometry itself can be baked into wasm directly as C code or however you can compile to wasm
  78. stagasand a few mb and with json they're down to kbs, idk why
  79. phalethcan't remember what is fbx used for, but you know all these formats are optimized for desktop applications, not for the web browser, on desktop the size does not matter too much
  80. stagasphaleth: have you been to Berlin?
  81. phalethnope
  82. stagashere is your chance https://berlin.stagas.me/ :D
  83. phalethheh
  84. stagascold start 17.9mb only
  85. stagas6s
  86. stagasand i haven't really spent time on optimizing that
  87. stagasi was focused on the game perf
  88. blueoh wow stagas that killed my browser
  89. stagaslol which browser
  90. bluebrave
  91. stagasodd
  92. bluebut I have a ton of agents running the background
  93. blueI guess they're eating a lot of memory
  94. bluebtw I switched to sol 6.2, it's so fast
  95. stagasyea
  96. blueI will soon try out codex because you said it's much better than opencode
  97. stagasit's a bit taking things too literally, need to explain more in the prompt
  98. stagaslike it does what i say, but it's not what i wanted :)
  99. stagasthey need to improve that
  100. blueheh
  101. stagasmaybe because i'm on Low most of the time
  102. stagasbut it's odd it has never crashed the browser here, and my pc is crap
  103. blueI'll try after a fresh start
  104. bluephaleth: repopack.com is now running primate 0.45.2, 922 should work now
  105. stagasi'm profiling the startup and will try to fix it and then i'll ask you to try again when you can
  106. stagasthere is a hang when shaders compile of 5s
  107. stagasthat might be it
  108. bluecould be. for now the entire browser froze, I needed to kill it from the terminal
  109. bluefor me*
  110. stagasah that has happened to me
  111. stagasi think i've found what it is
  112. blueawesome
  113. blue"On a local test project with 920 closed tasks, the Closed page’s HTML was 95% smaller, and median time to response headers fell from 1,062 ms to 59 ms. Browser, PostgreSQL, and docs tests passed."
  114. bluephaleth: ^
  115. blue(this is with pagination)
  116. stagasit's because in Berlin it's freezing so it's realistic the freeze
  117. stagas:P
  118. bluelol
  119. stagasshould be good now if you want to retry
  120. stagasomg this vps is so slow, i'm not getting >500kb/s
  121. phalethpagination, ok
  122. blueretrying stagas
  123. blueinteresting, it loads for a second or two, then freezes
  124. stagasuhm error in the console?
  125. bluethen goes back to loading after a while, then freezes again
  126. blueI'll try to open it, need to wait for unfreezing
  127. stagasyes some ops are blocking, you have to wait maybe
  128. blueoh wow it loaded!
  129. blueoh wow stagas this is incredible
  130. bluebtw after flying berghain I got stuck in a wall :P
  131. stagascome in
  132. stagasor you left
  133. blueI left, I'll rejoin soon
  134. stagas:D
  135. stagasthanks
  136. phalethblue: not sure if that's repopack or primate error https://repopack.com/project/repopack/tasks/928
  137. bluelooking phaleth
  138. phalethah, forgot to start db :D
  139. phalethwell, still got the error
  140. phalethok, there was one migration pending, now it seems to be running
  141. blueyeah, calnker said it's db reachability
  142. blueplease close 928 if it's not a problem anymore
  143. bluedreamreal: fyi; rp cli has switched to 30-day token sessions, extended indefinitely. the web ui session is 7-days, extended indefinitely as well
  144. bluethat should solve most of your timeout issues, I think
  145. phalethmaybe it is a primate problem, it could say that the db connection has been refused, but it threw the error in esbuild instead
  146. bluek, I'll consult with clanker
  147. phalethheh, the position of the pagination buttons isn't great, but now there is pagination :)
  148. blue929 fixed that
  149. blueredeploying
  150. phalethheh, ok
  151. bluehttps://repopack.com/project/primate/tasks/428
  152. bluephaleth: ^
  153. dreamrealooo thank you
  154. blueyw
  155. dreamrealdamn it why is the release channel not available
  156. dreamrealevery time you tell me something's updated, I go to release it, 503
  157. bluewhat release channel?
  158. dreamrealJust did rp update
  159. dreamrealgot a 503
  160. dreamrealthere it goes
  161. blue→ rp update
  162. blueupdated rp 2a19ad7a -> 9e9b6528 on channel unstable
  163. blue~ via  v26.13.0 took 3s
  164. blue→ rp update --channel stable
  165. blueupdated rp 9e9b6528 -> 0.2.0 on channel stable
  166. blue~ via  v26.13.0
  167. blue→ rp update --channel stable
  168. bluerp 0.2.0 is already current on channel stable
  169. blueoh. sorry. the 503 is me redeploying repopack.com
  170. blueI thought probably create a 0.2.1
  171. phalethdreamreal: heh, unlucky
  172. blueI should*
  173. bluethat happened to him TWICE today
  174. blueanyway, I'll cut rp 0.2.1, been a while. it's not needed for the session timeouts, but useful anyway
  175. phalethprimate takes a while to start
  176. dreamrealthen cut over to be more durable
  177. dreamrealit's happened with flog, it's happened with rp, I know it's like being hit with a quark but it's happened twice to me
  178. phalethmaybe in traefik there is some sort of configurable timeout, to tell it disconnect the prior container and switch to new one after some more time
  179. bluemaybe
  180. bluehonestly if we have one user encountering this timeouts twice in two days, we should make the cutover a bit more stable. that's on me I think, rps currently stops the runtime
  181. blueit does so so it can safely apply the migrations
  182. bluehowever, 1) not all updates have migrations, and that's detectable and more important 2) zero downtime is a possibility, it just needs planning
  183. dreamrealcould just be me being horribly unlucky with it, but it's not like you're going "hey update in the next five second" and bang, it doesn't work
  184. phaleth:D
  185. bluewell dreamreal I currently redeploy around 20 times a day, and phaleth doesn't really mess around with haproxy THAT often
  186. blueso yeah, the first time was you being unlucky
  187. bluethe second time should be avoidable with zero downtime cutover
  188. phalethI can tell repopack is being redeployed often lately, yeah
  189. phalethbut sometimes it's uptime goes up to two days, for example when clanker was idle
  190. bluebut also I have to be honest about: this is clear alpha/beta behaviour. we're moving fast, and while we're not breaking intentionally, we're also not holding back updates when we need them
  191. bluethis isn't the standard to expect of this service
  192. dreamrealI'm not complaining except when I am
  193. blueit's ok, you're the customer in this case, you have a right to complain
  194. bluewe should do better
  195. dreamrealeh, I'm not complaining as much as observing
  196. dreamrealjust in case you don't know
  197. phalethstackoverflow and other well established services still have such moments
  198. bluedreamreal: rp 0.2.1 is published
  199. dreamrealnobody noticed when SO goes down any more
  200. dreamrealbecause it has no users any more
  201. bluetrue
  202. phalethright
  203. bluethe thing about runtime is this, and it's like every other sort of negative feedback. you could be up 99.99% of the time (and repopack.com probably fulfills that metric, quantifiably), but it doesn't matter of users hit downtimes
  204. phalethstackoverflow prolly figures out how to scale itself down to zero when nobody needs it and then suddenly tries to become available returning 503
  205. bluelol
  206. bluephaleth: https://repopack.com/project/primate/tasks/428
  207. phalethcontabo alone is already bellow 99.99%
  208. blueya, true
  209. phalethhmm, why would clanker polute the quickstart?
  210. phalethotherwise looks good I guess
  211. phalethnow the pagination button are where one would think they might be
  212. blueyup
  213. phaleththe amount of JS didn't increase with primate 0.45.2, which is good
  214. blueyes
  215. stagasok the new domain is berlin.urt.one