Submit

Back to the channels

#primate

  1. blue Hi phaleth.
  2. blue phaleth: https://github.com/primate-run/primate/issues/286
  3. nevet pema: `p.loose.bigint` / `i64` / `u64` / `i128` / `u128` coerce numeric strings via `Number()`, losing precision above 2^53 · Issue #286 · primate-run/primate
  4. blue Primate is commited to delivering in record time :P
  5. blue I used that report as a reason to introduce changelogs per package, which is kinda cool
  6. phaleth hi blue
  7. phaleth a research group led by Professor Yu Feng
  8. phaleth looks like they don't care how is that implemented
  9. blue how what is implemented?
  10. phaleth the number overflow
  11. blue well it's good we got their attention
  12. blue also, we're the fastest in delivering fixes, less than an hour from the original report to patch on npm
  13. phaleth yeah, but it's a lot of mess in the code https://repopack.com/project/primate/tasks/414#file-70c75a59428a18c31f4c3a38
  14. phaleth only a student would be worried about 2^53 overflow, practically that's not going to happen
  15. blue I see if I can move some of the complexity into rcompat
  16. phaleth well, would be nice if a lot of the JS clanker recently added to the website's app js bundle would go away
  17. blue what JS clanker?
  18. blue can you create task?
  19. phaleth I mean the JS that clanker recently added
  20. phaleth nah, prolly waste of time, and clanker has no chance anyway
  21. blue but where? what task?
  22. phaleth all of them I guess, like each release adds more and more kBs
  23. phaleth it used to be at around 71 kBs
  24. blue can you create task so I can review this? just write it used to be 71kBs and now it's whatever, and I'll look into it and figure it out why
  25. blue also I didn't update repopack.com yet to primate .45, still need to do it
  26. phaleth don't worry, it does not matter
  27. blue I think it does matter if it bothers you :{
  28. blue :P*
  29. phaleth yeah, would be nice if there was some work done for repopack
  30. phaleth just saying that bloat is being added
  31. blue in primate or repopack?
  32. phaleth I mean you've propably promised those 64 bit ints in pema or whatever, so that bored student chuckled over it
  33. phaleth but in the end who cares if somebody runs into a bug there
  34. phaleth both
  35. blue well please create tasks, I will look into it, memory consumption is kinda good in repopack, it's been running for a while and it's at 288MB
  36. phaleth but I've already created other tasks that are more important
  37. blue k
  38. blue major chat update is coming, phaleth
  39. blue and btw you can use with free models :)
  40. phaleth oh, hmm
  41. phaleth last time I've tried running Mistral on my laptop the laptop was dying
  42. blue we're gonna get hosted agents soon
  43. blue this will be a paid feature
  44. blue and I know dreamreal is gonna love it
  45. blue phaleth: https://repopack.com/project/primate/tasks/423
  46. blue man I REALLY need chat. I keep confusing all those opencode tmux sessions :(
  47. blue I have three sessions per project. and I have the projects: repopack, primate, rcompat, flog, dog/cit (and sometimes socialcloud). also planning on future projects, superarch and many others
  48. blue imagine how much confusion this causes
  49. blue with chat now, you always are in the context of a project (this is a local change I haven't deployed) yet. going to the homepage redirects you to your first, in the future default project
  50. blue being always in the context of a project is immensely helpful
  51. blue and btw: I think this is why people use herdr
  52. phaleth yeah, I think so too
  53. blue because they realised tmux with a ton of agent sessions is gonna be totally confusing
  54. blue but repopack.com solves this so nicely. you always have a project context
  55. blue also with hosted agents, I can restart the computer as often as I want. the git checkouts survive because they're not on my computer
  56. blue it's perfect
  57. blue dreamreal: ^
  58. blue I might be wrong, but I think we've hit gold here. only time will tell
  59. blue three workflows:
  60. blue 1) you host your agent on your computer and connect it to repopack.com via rp; that's how I intend to work now; rp manages the worktrees in ~/.local/rp/agent-workspaces
  61. blue 2) you host your agent on a vps or otherwise a server you own, and use rp to connect. same thing like 1), but not on your computer
  62. blue 3) you buy a hosted agent (essentially a vps or another abstraction) from repopack.com, you don't need to setup anything
  63. blue 3) will be the most elegant way to work, but also you might ask yourself - how do I get previews of the things I'm working on?
  64. blue that's why we have task previews, that's exactly the reason they exist
  65. blue it's a new way to do development
  66. blue radically different, I think, but also very interesting
  67. blue and also crucially, it means your computer can burn down and you don't lose anything
  68. blue and you can do development on a weak(er) laptop
  69. blue your computer becomes a terminal to your development environment, via your browser and repopack.com
  70. phaleth lot of people will pay clanker and be wanting to get clanker hooked up to the project easiest way possible
  71. blue yes
  72. phaleth if you make it so that the person has to write commands in terminal then you've lost them
  73. blue well ideally, they're all gonna work through repopack.com chat
  74. phaleth I mean for hooking up clanker
  75. phaleth like to me the chat says some nonsense cause I have no clanker
  76. phaleth it should shove down a guide how to hook up clanker to the user's throat
  77. blue yes, this is going to change with hosted agents
  78. phaleth also once they set it up it should make them feel like a winner
  79. blue well with hosted agents, there's no setup involved othre than connecting to your provider
  80. blue and if you buy a hosted agents, you would still get the free models immediately
  81. phaleth the gamification aspect is important, people are way too screwed nowadays, especially the youngsters
  82. blue lol
  83. blue dreamreal: you use claude code, correct?
  84. blue I currently only have rp wire opencode sessions. I do need to extend chat to claude code and codex
  85. phaleth and to seepdeek
  86. phaleth oops
  87. blue seepdeek?
  88. phaleth yeah, try saying that
  89. blue seepdeek!
  90. phaleth this is difficult to understand without an example https://docs.repopack.com/start#connect-a-project
  91. blue can you edit it or create task to improve?
  92. blue you don't have to work on the task, just create it
  93. blue I want to make it perfect but I need to know about things that don't work well
  94. blue just create tasks, it's cheap
  95. phaleth yeah, but clanker is just gonna write more walls of texts
  96. blue ignore it, just create tasks, tasks are the cheapest currency we have
  97. blue the walls of texts are something we can work against in AGENTS.md
  98. blue that doesn't make creating tasks invaluable
  99. phaleth so does `rp bootstrap ...` create a project?
  100. blue or well, "not valuable"
  101. phaleth the bootstrap term is confusing
  102. blue it doesn't
  103. blue it sets up a project by cloning it
  104. blue try this
  105. phaleth ah, now I get it
  106. blue mkdir ttt && cd ttt && rp bootstrap primate
  107. blue it will clone the project into ttt/primate, and print instructions. it's mostly for agents
  108. blue if you run rp bootstrap INSIDE a repopack.com project, it will just print instructinos
  109. blue it detects it by looking at the .git/config and finding repopack.com origins
  110. phaleth that information is missing there
  111. blue ya
  112. phaleth clanker would just turn it into a wall of text
  113. blue create task, I will fix
  114. blue or add manually :P
  115. blue either way you need a task for it
  116. phaleth yeah, I will try to read the docs, prolly impossible
  117. blue chat is working REALLY well
  118. blue once I have this deployed, you gotta try it out phaleth
  119. phaleth ok, anyway, I don't have a clanker
  120. blue there are open models
  121. phaleth last time I messed around with LLMs was in 2023 and since then I forgot everything
  122. phaleth I remember I used some python environment thing that was painful to work with
  123. phaleth every single project worked only on specific version of python
  124. blue you just need opencode
  125. phaleth isn't opencode's source closed?
  126. phaleth :)
  127. phaleth I think it was anthropic who published sourcemaps of their CLI to npm package registry and then somebody found two vulnerabs in their shitty CLI thanks to that
  128. * phaleth looks up if garbage is still running on the VPS
  129. phaleth nope, it doesn't
  130. blue opencode is open source
  131. blue https://github.com/anomalyco/opencode
  132. blue MIT
  133. phaleth wow, it can be installed via pacman
  134. phaleth then it must be important
  135. blue ya
  136. blue anyway try it, rp can connect to it and then you can use the chat :) I'll soon push out this amazing update
  137. blue almost done with it
  138. phaleth my computer does not have a GPU and it's CPU is too old
  139. blue don't worry about it, this is a T480
  140. phaleth yeah, but you use a remote clanker
  141. blue I don't
  142. blue opencode is running on my computer currently
  143. blue I don't have an extra vps atm
  144. phaleth by clanker I mean the LLM
  145. phaleth I found this knob https://i.imageupload.app/0ca74ba34e780ae573b2.png but in rp CLI there isn't anything like `rp task merge .` right?
  146. blue Yes. I think I removed `task merge` from the cli at some point. It should return, but I think it should always guarded
  147. blue Clankers should not be able to autonomously push code onto master. So that will become an always-guarded action.
  148. phaleth yeah, with a permission on the project
  149. blue here, https://repopack.com/project/repopack/tasks/868
  150. blue we don't need a permissoin on the project
  151. blue task merge will always be guarded
  152. phaleth but people also use CLI so it's impossible to distinguish people from clankers
  153. blue yes. but the CLI isn't assumed to be 100% safe
  154. phaleth and people should also be denied to take that action
  155. blue yes
  156. blue anyway can script your CLI
  157. blue the last authority is the website
  158. blue guarded actions always require you manually approving on the ui
  159. blue there are not a lot of them, but important things are guarded
  160. blue s/anyway/anyone/
  161. phaleth ok, that task says: "equires a deliberate human confirmation in Repopack's browser for every invocation"
  162. phaleth requires*
  163. phaleth but how is that confirmation gonna look like? will it be enough that I'm logged in only via rp CLI?
  164. phaleth cause now I have to be logged in on repopack site itself and hunt for the merge button everytime
  165. blue the human confirmation is a popup that appears on the top right in the browser
  166. blue and yes, you wouldn;t need to hunt for the merge button
  167. phaleth yeah, but in the browser after logged in to repopack website right?
  168. blue yes
  169. phaleth well, that's still the same then
  170. blue same as what?
  171. phaleth makes no sense to implement that task at all
  172. blue wait, are we talking past it other? `rp task merge` would allow your clanker to request a merge
  173. phaleth like currently, in CLI you get link to task, open it and hit the merge button there after logging into the website
  174. blue that would show up as a popup on the UI
  175. phaleth yeah, but I'm in role of the clanker too :D
  176. blue that spares you from looking for the task and scrolling down to the merge button
  177. blue if you're in the role of the clanker, that's useless for you, yes
  178. phaleth there is no looking up the task, the URL to the task is already in the terminal after `rp task submit . --push`
  179. blue you wouldn't type `rp task merge`, then go to the website and confirm it
  180. blue you'd just go directly to the website, I agree
  181. blue but human-driving rp isn't the most common use case
  182. phaleth ok, then lets move on, no need to implement
  183. phaleth that means I should enable that knob I shown on the picture above again
  184. blue well arguably it can move things a bit faster... let's imagine I'm in chat
  185. blue then I tell the review clanker, review this task, and if approved, request a merge
  186. blue then it can work in the backgronud and at some point I get a merge popup request
  187. blue I think there is some rationale for that task in the task description
  188. phaleth yeah, that makes sense, but I was thinking it should just be available to anybody, but clankers will not have permission by default, or better yet that merge availability is off by default, which is already done by that knob I posted on the picture above
  189. blue so the probem is, you cannot distinguish who's driving rp, right?
  190. phaleth I dunno, prolly not
  191. blue it's literally impossible to distinguish it. maybe in chat because we run in a sandbox. but even then we cheat with RP_CHAT_SANDBOX=1
  192. blue an env variable should not become an access gate
  193. blue I believe in repopack, we should have two levels of access. rp and browser. browser can do less actions, but has higher permissions
  194. blue rp can do anything, but important stuff is guarded, because it's scriptable and agent-runnable
  195. phaleth ok, how about doing something like `rp login` does
  196. phaleth it opens up a web browser to confirm and only a human can confirm that
  197. blue rp login just gets you the token
  198. phaleth yeah, I mean that behaviour
  199. blue essentially we have that behaviour already for guarded actions. except it doesn't open a new browser, it sends a push SSE to your existing session
  200. phaleth so just show a merge button there, no need to logged into the website itself and click some notification or whatever
  201. blue but remember that with rp login, you already need to be logged into the website anyway
  202. blue this is more or less the exact same flow as push popups, just in a new tab
  203. phaleth yeah, that's true
  204. blue I honestly don't even know why we're doing rp login in a new tab
  205. blue should just do a push notif
  206. phaleth so the client would just have to generate a one time token that the server can recognize as something belonging to the user
  207. blue I guess, yeah
  208. phaleth something super obscure
  209. blue phaleth, there's a new feature on repopack.com if you're interested
  210. blue when you're in a project, click on the project switcher
  211. blue the last line allows you to subscribe
  212. blue you can choose which events you're interested in getting informed about
  213. phaleth I see, looks nice
  214. blue the transport mechansim is confirmed in your account settings: any combination of email, bells and push notifications
  215. phaleth I mostly care about tagged releases
  216. blue can you create repopack task for that?
  217. blue `add tagged releases as subscription option`
  218. phaleth ok
  219. blue thanks
  220. blue thanks, I got a push notif for that :)
  221. phaleth heh
  222. * phaleth continues to read those walls of texts
  223. blue https://repopack.com/project/repopack/tasks/918
  224. blue is ok phaleth? it's not a wall of text..
  225. phaleth yeah, that means clicking on that notif will open the release page I guess
  226. blue yes
  227. phaleth ok
  228. blue rp is kinda nice, don't you think
  229. blue can really do cool things with it
  230. phaleth yeah, would be nice if I could merge without logging into the website
  231. phaleth that's the only piece missing I guess
  232. blue that would give rp a lot of power to merge things into master autonomously
  233. phaleth also I dunno if it's possible to list task with review state, but that prolly already is possible
  234. blue maybe we should allow this per rp token, phaleth?
  235. blue so when you do `rp login`, we can could offer an option "Allow rp to merge tasks without confirmation, subject to project settings."
  236. phaleth per project
  237. phaleth oh, that login token, yeah
  238. blue ok, per project is also fine by me
  239. blue my thought about the login token was that you could have your personal rp login on your computer, to merge stuff
  240. blue and then your clanker login on your vps, with no merge capabilities
  241. blue we could even define that that `rp login --headless` never gets merge unguarded merge capabilities
  242. phaleth no clue how that works currently, but I can see on the repopack website there are like a lot of already expired tokens from the CLI
  243. phaleth not sure if they get autodeleted at some point from the db
  244. phaleth yeah, that's a good idea
  245. phaleth I noticed that `rp login --headless` in the docs, but the docs are not telling much I can understand
  246. blue it's mostly self explanatory but should still document
  247. blue rp login --headless will print a url, I think repopack.com/auth/cli/dervice or so
  248. blue [blue@vmd170320 ~]$ rp login --headless
  249. blue Repopack server: https://repopack.com
  250. blue Open: https://repopack.com/auth/cli/device
  251. blue Enter code: REDACTED
  252. blue This code expires in 10 minutes. Waiting for authorization...
  253. phaleth and clankers just curls that url?
  254. blue it's a server-server token exchange, yes
  255. blue you need to authorise the token
  256. phaleth cool
  257. blue this is for environments that have no browser
  258. blue like a vps.
  259. phaleth yeah, that's understandable form the docs, it's just no example there
  260. phaleth from*
  261. blue ok, could you create task to improve that? I will do. no walls of text, promised
  262. phaleth nah, don't worry about it, I just need to read all repopack docs
  263. blue k
  264. phaleth all of those functionalities are going to change several times anyway
  265. phaleth at this stage docs are just a pure waste of time
  266. blue yes but remember we will add to AGENTS.md, to update docs on every task
  267. blue and remove as much internal docs as possible
  268. blue anything that's not 'secret' (whatever that is), should be public-facing
  269. phaleth I think then in AGENTS.md there should be something about docs in admin area, if there is going to be anything internal ever documented
  270. blue YES
  271. blue create task?
  272. phaleth at this point, no
  273. phaleth there is nothing internal, except in #392 which should just be closed already
  274. blue 392 is a dumpster fire :)
  275. phaleth wall of text, yeah
  276. phaleth `rp task list --status review` takes a few seconds to respond, is it not like one db query behind the scenes?
  277. blue yes
  278. blue should be faster
  279. phaleth 2 seconds https://upaste.de/raw/bER
  280. blue I'll be looking into it
  281. phaleth ok, great
  282. blue working on perf now
  283. blue phaleth, for the record, at some point we're gonna need to develop our own agent to make things even more streamlined
  284. blue supporting opencode and codex and claude code is nice, but having our integrated agent makes things a whole lot smoother
  285. phaleth I thought the chat replaces opencode
  286. blue no. the chat is not an agent. the chat is a multiplexer over agent sessions
  287. blue it connects to your local opencode/codex/claude code
  288. blue an agent is a complex thing, needing instructions, context compaction, tools, and so on
  289. blue it's doable, but it's not what the chat or rp currently does
  290. phaleth ah, ok, so it could be doing that in the future, but looking at the size of the opencode project that is a huge undertaking
  291. blue yes and no. because our chat agent will be 100% headless, just a server, not a client like opencode/codex/claude code
  292. blue that saves you a lot of complexity
  293. blue all those opencode/codex/claude code are both. they are a server and a client. and I think a lot of the complexity is in the client
  294. phaleth so I don't know how a web browser can control a CLI tool, sounds like there is a security problem
  295. blue our chat replaces the client essentially
  296. blue it doesn't
  297. blue rp starts an outgoing connection
  298. blue the web browser doesn't really control it, it's more like rp connects to repopack.com and they exchange info
  299. blue you could say in essence it does control it, but that's not a real security problem since connection can only be initiated from rp
  300. blue installing or running rp does not allow any repopack.com user to control rp
  301. phaleth yeah, the docs do not explain any of that
  302. blue you need to explicitly run `rp agent start`, and enter the code shown to link it
  303. blue yeah, the docs are also unfinished :P
  304. phaleth https://docs.repopack.com/guides/agent-chat
  305. phaleth wall of text
  306. blue yes, we're gonna need to deal with that
  307. blue believe me, I'll handpick all the docs and make sure they're reduced and understandable
  308. blue and then I'll have dreamreal go over them and scream at me when he doesn't get something
  309. phaleth no, don't worry about the docs, they are a pure waste of time
  310. blue give it some time, all good things take a while
  311. phaleth no no no, no time given to that
  312. blue the docs are actually important because people have been asking for them
  313. blue dreamreal did, as did stagas
  314. blue I can't ignore that forever
  315. phaleth yeah, but those current docs are useless
  316. blue not entirely. they already have structure, and some useful information
  317. blue the problem is they're gonna outdate fast
  318. phaleth so it's like waste of token, and people can type in `rp agent --help`
  319. blue and for that we need to update AGENTS.md
  320. phaleth tokens*
  321. blue ya
  322. phaleth yeah, but if you update rp CLI's help then clanker can figure things out
  323. phaleth and so can people
  324. phaleth no person is gonna read AGENTS.md
  325. blue ya
  326. blue phaleth: you can now subscribe to new releases, too
  327. blue it's deployed :)
  328. phaleth cool, thanks
  329. blue oh man, it's not marked on by default
  330. blue stupid clanker :(
  331. blue need to fix that
  332. blue phaleth, are you on rp unstable?
  333. blue should switch to the unstable channel for now, I'm not cutting releases for everything
  334. blue `rp update --channel unstable`
  335. blue I'm gonna deploy the perf improvement in a few mins
  336. phaleth I think stable but it's working well so far
  337. blue ok
  338. phaleth ok, I could test that then
  339. blue you will need the latest version for the chat. I might cut a new rp version. but for now you can switch to unstable
  340. phaleth I have no local LLM
  341. phaleth but maybe later I can try finding some
  342. phaleth ever since hugging face requires a log in I just don't bother with LLMs anymore
  343. phaleth yeah, clanker is stupid cause of coding this way https://repopack.com/project/repopack/tasks/919
  344. phaleth I mean, by making the problem in the first place
  345. phaleth well, whatever it wrote there looks like a bloat
  346. phaleth ah, so clanker added a bunch of flags to the CLI it seems
  347. phaleth or query options
  348. phaleth on task stacks page in docs there is this: "the stack merges from parent to child" and later "An unmerged parent cannot be closed while open children remain"
  349. phaleth that sounds contradictory
  350. phaleth the first statements is most likely wrong, or at least I'd think so
  351. blue ya
  352. blue the problem is phaleth, before it was filtering in the cli, now on the server, that's faster
  353. blue anyway, I merged 919, gonna redeploy
  354. phaleth yeah, figured
  355. blue phaleth: try now
  356. blue rp task list --status review
  357. blue is much, much faster
  358. blue you need `rp update` first and to be on the unstable channel
  359. phaleth yeah, trying to update
  360. blue `rp update --channel unstable`
  361. phaleth much better, yeah https://upaste.de/raw/eA5
  362. blue yup

Search posts