ok i'm taking t3code for a spin. but like... responses are not streamed? compare with Hermes (also plugged in to Codex). what am I doing wrong?

proof that if u got distribution, product does not matter that much

ayo 💀

no shade to t3 code - just that i know no early stage founder (like me) would even consider launching without streaming built in

turns out u can

@thomasgauvin Have you ever considered the possibility that we did launch with this on and, after extensive testing and analysis, concluded that it was a better default to have it off? Would be crazy if we had a legacy section that also included this bad feature.

what in your testing and analysis of the user experience said that streaming in responses in worse? you've provided receipts for streaming requiring more bandwidth/network and for introducing engineering complexity, heard! but your explanations for worse user experience is that in your opinion it trains to wait for responses... but there are many valid cases for asking questions about the codebase or agent session that are not long running background tasks. so is t3code not intended for that? that is the 1% of time that streaming is actually useful for anyways, just the perspective of a potential new user looking for a new coding agent, I respect the work of you and your team and offer a different opinion on your product decision

you can enable it in settings -> legacy features

why legacy 😿

can someone share a video of the performance change between streaming on vs off? i can't perceive it from first impressions

The performance is mostly network-side. It's just significantly more data being sent back and forth and it changes the shape of events meaningfully in a way that requires much more work to keep it in sync. I also do genuinely believe it's a worse user experience. It trains you to wait for a thing that only happens 1% of the time, when the thing you should be doing is leaving and working on something else.

@maria_rcks so you never have one off questions about your code that can be answered with the existing context window?

@theo but then the animation is useless, distracts you from reading

@theo but maybe im not the intended user! i definitely don't think I'm the intended user for Hermes, but I appreciated that they streamed in the response

Your first mistake was installing this garbage

that's brutal yo i don't think it's garbage

I'm biased, not a fan of the prick with no morals it's associated with

@thomasgauvin Need to grow up buddy

@thomasgauvin you need to stop glazing buddy

In this case, there was nothing to stream as the model had not sent anything yet (it was still reasoning). You’d see the same in codex. Every paragraph and text block appears the moment it is ready. Separately, we don’t stream incomplete text because it is a really bad UX with even worse server side implications. We left it in as a legacy feature, but I ran the numbers and concluded you should never use it.

this is not true, it was actually not still thinking, you label it as 'thinking' while you buffer the response you could have streamed multiple paragraphs while achieving your objective of reducing websocket updates by not doing token by token streaming

you used a product built by Theo and you expected it to work? 💀

it came highly recommended by @roerohan

@_nahoc That's weird, I could swear streaming works for me but now I'm not sure anymore lol @shivamhwp @maria_rcks can comment too in case it's an issue

@_nahoc @shivamhwp @maria_rcks yeah she said it's a legacy feature

@_nahoc @shivamhwp @maria_rcks interesting i wonder why that is

@_nahoc @shivamhwp @maria_rcks the people said "perf"

Did they actually do this?😭 Doubt it tbh the repo has streaming so idk

bruh not streaming back responses is just laziness at this point

@NeuralAA You’re wrong. We have it. I measured the results. It is bad.

@thomasgauvin How can one be wrong about a subjective experience they’re used to? And most people are used to? Not having it is fairly opinionated I would say and not the standard UX

@theo @thomasgauvin It's not a subjective experience. Theo measured it and it's bad.

@NeuralAA @theo what did he measure? he measured that out of the 100% of the time the agent is working on a meaty task, only 1% of the time is spent streaming the response why does that justify not providing a good user experience for the 1% of the time? or for the simpler one off questions?

I prefer it this way since I don't care about reading the stupid prose from agents most of the time, I'd rather be pinged when it's done so I can scan and jump to reading the actual code or do an explainer skill etc

the composer looks so vibe coded 😭

you're not using Synara

Wow this thread is how I'm finding out people don't think streaming is valuable. Sometimes I catch it reasoning on things I didn't intend and want to steer it before it gets too late

the thing you're doing wrong is using t3code

i might have a solution, but people will get upset

you should check out if you want to really feel the future and obv the responses are streamed lol

why do responses need to be streamed?

> what am I doing wrong? Using t3code lol.

In settings, there is an option for legacy features. Clicking that shows all the legacy features, and from there you can turn streaming on or off.

thanks, i thought i had something wrong with my harness setup

That + some other things with the remote control is killing me I want to like the project but I just don't like the feel of the software

Hermes vs t3 code is not really apples-to-apples tbh. T3 code is not a harness.

I am pretty sure you can turn it on in General -> Legacy Features-> stream token by token.

You need to upgrade to Monocode.

Why are you watching your threads?

skill issue his agents couldn't make it good so he removed it

try @zuse_sh dude, it has streaming, t3 code is opinionated and they think streaming is not that good (not that its bad, just preference ig)

@trySynara and you will not be disappointed :)

I tried, not worth it, save your time

First thing I did after forking t3code

Yeah one of reasons I couldn’t use it, I perfer streamed responses.

streaming token responses generally heavy lift for any computer and once you start running multiple sessions, which T3 code is designed to help you do, you will start having system issues. generally, if T3 code is your tool of choice, the agent responses should be enough for you to use it effectively. @theo thats my interpretation anyway.

Theo said explicitly that he's tired of streaming

is t3code buffering the whole response, or is the codex adapter swallowing token events? the Hermes comparison makes this sound more like an integration bug than a model limitation.

Bruh are you actually watching the agent work?

Built for a different persona. Assumption is that you don't watch your agent working, just switch to another tab and multitask.

You can then it on in legacy settings (under general)

wow streaming responses is 'legacy' @NeuralAA

Idk why people get so opinionated and out of the norms with things they aren’t experts in Not wanting streaming is opinionated, and streaming is the UX for chats for a reason and it is better and more importantly expected For the subset of similarly opinionated people have the setting for them not the other way around no?

@realSalloxy i find it so weird the amount of people defending the no streaming

@NeuralAA Not just I don't defend it but I absolute hate the fact that's it's both off by default and doesn't work well when turned on. I was just explaining why he moved it to legacy settings but of course I would much rather have them fix it.

@NeuralAA yeah sorry wasn't targeted at any single person, thanks for sharing the context! I hope they get around to fixing it and getting it to a perf state they are happy with

@NeuralAA Np you're good. Just wanted to myself to the people frustrated with the lack of good streaming in T3 Code

This is one of those strong UI opinions of Theos along with not showing tool calls afaik Makes sense when you consider he literally leaves the thread the second the prompt is fired, doesn't quite feel right for me though But I'm a big fan of opiniated UI so I can't hate

i wish chatgpt would just open source the chatgpt/codex desktop app... let ppl mod it, bring in other models. it rlly is that good

it is the best