Rspeedy 0.15 brings Rspack/Rsbuild 2.0 to @LynxJS_org
Measured on real internal projects:
🚀 Builds are 17-39% faster
📦 Output is up to 11% smaller, with main-thread.js and template.js typically down 2-9%
@lightcap Typically for read apps and side projects. Hooks for input key redaction model can copy to clipboard and into keychain then write from keychain out without seeing the key
I love Hermes. But codex is the only piece of software that turns my entire machine into a simple chat. You have any idea how liberating it is to just whip out a phone, pay all my bills, follow up on emails, get api keys, setup whole servers, operate my discord, slack, other apps, deal with OTP - all while I’m taking a stroll in the park. There’s a lot I wish codex did better, but what it can do, it does the best.
Also like kanban awesome but not in desktop app. Dashboard ui is rough given the power of it.
Databases constantly corrupt and need recovery in kanban. Tickets made for kanban show up in main app chat. I’ll make thousands of tickets and it just floods any hope of finding actual top line threads
Hermes over telegram is incredibly good how I used it till desktop app. I added switch profile to the telegram app and server so one instance can switch to others. Problem is it’s difficult to see what going on. Great assistant, codex out the box tho can operate the whole machine. Has app capture, integrates with existing chrome sessions, I had to customize Hermes browser integration to add some missing tools needed to capture browser / be able to fully operate it. Thread organization is painful.
Don’t mean to be negative! I love Hermes and it’s my default operator. But it struggles to scale. Database concurrent write issues on kanban happen often. TUI rendering will get cut off / redraw cause other bugs like lost input control or only come characters received. Desktop app is huge improvement, I forked and added kanban ui into it. Would be nice if there was a way to make profiles behave a little more traditionally like codex. For instance I want to have a project with its own profile, kanban, and all threads organized under the project. I added a cursor provider but externally doesn’t integrate so had to do it in fork. Ideally more robust plugin hooks for the little things, hooks into desktop ui too would be nice. Hooks to register options that render in the config of dashboard. Profile rail doesn’t always switch. Would be nice if config crawled models has model explorer when picking in ui. In fork I added model routing and load balancing. So tickets get a profile and model route key. This let me split the load over multiple providers of the same model or like GLM and opus for frontend balances between the two.
Telegram and other app integration has areas where could be button vs text response. Subagent visibility has improved but loading the threads can be very slow in desktop app. No distinction between like ticket threads vs threads made in main app or tui. I have like 50k threads. While usually I can just start a new one and it’ll figure out what going on. It’s virtually impossible to navigate threads with large volumes. I’ll put like 1000 tickets into the kanban, rely on custom lcm to figure out what’s what. Which works, but it would be nice to have some organization vs everything’s just a thread in the sidebar and there’s tens of thousands. Hermes a fantastic assistant, look forward to a few iterations. The self improving design and plugin implementation is better than others, makes it very effective, but would be nice if it shipped with stronger browser impl, chrome extension into existing profile. Visibility is suboptimal in tui, gui, as async agent tho - it’s great. Maybe I’m using wrong
Holographic memory has by far been the best. This is a big reason I want to link them with lcm. Things can fade over time and distance themselves, naturally. So as you move from problem to problem, no longer helpful things eventually are forgotten, but can be focal for however long they need to be
@solomonneas@farzad_yz I also find LCM and holographic memory to be very useful. “Assign 1 agent per merged pr, learn from the maintainers on all merged PRs in the past year and make memories as needed” - was a game changer
Ultimately, I want to link the code graph to LCM to timeline aware git to memory storage. So that when memory is made or function is written, the code graph can pull up every session that involved that function. So model knows exactly why and how it was written like that. I then want to add something for pull requests that would show me the code changes along side the thread or excerpt from the threads so I can see what the model was thinking when it made these edits
@solomonneas@farzad_yz Thanks! Yeah I thought about adding embedding but it’s just so slow so I went the FTS route. Do think embedding would be useful during file edits to have hooks that warn when a newly written function looks similar to an existing one in the code base