jules-general・🦑

[simple rain]

working as in, interactively with it?

[minor garnet]

yes jule cant work by himself

[simple rain]

yeah 🙁
I'm looking towards getting much smaller bites/more frequent PR's out of it for that very reason.

I have like one every 10 minutes quota-wise, counting only the time I'm awake, so as long as I do other things on those days instead of only spending the day with jules, it shouldn't be too bad.... I hope.

[minor garnet]

honestly i switched to open code

[night oriole]

The codebase can't be corrupted unless you merge code via a PR

If a bug was merged, you'd be able to roll to a prior commit

[minor garnet]

i mean it was like a long time ago like in 2025

[night oriole]

Agents aren't able to access your code and push without a human in the loop

[minor garnet]

i was working on a long project and jule messed things up so

[minor garnet]

like someone to monitor it?

[night oriole]

Yes. I just want to be very clear that there is no situation where any agent would be able to get access to your codebase without your knowledge

And if a bad commit was merged, you'd still be able to roll it to a prior working state with git.

[spark sonnet]

It's more like believing promises Jules makes, I suppose. Jules may show impressive work and present some beautiful code, and ensure you that all tests has passed, and you may think that you have seen it with your own eyes, for example if you requested logs. But then, before submission, Jules would ask a tiny question like "Is everything good?", and pause before submission. The system would suspend the task until the user replies, which may take hours if the user is asleep at the moment (this even happens in scheduled tasks!) and when they finally reply, Jules doesn't know that but it's environment has been reset, and the working tree is different from what it promised - so it would submit bad changes while user may still think they approved right changes.

[swift karma]

Jules does in 30 minutes what opencode does in 5 minutes (depending of your LLM)

[spark sonnet]

Do you mean Jules too somehow makes my PC vulnerable? I guess not. It has a decently secured sandbox. OpenCode requires me to set up one and I am noob in security

[river elk]

Hi, Jules in general is very slow?

[night oriole]

No. Jules does not operate on your system. It runs in a virtual machine in a data center. You're only making an API call

And yes, running a CLI tool would be faster! That's for synchronous tasking when you are actively steering.

Jules is on the other side of the spectrum. There are multiple sub agents working on your task. We'll always prioritize correctness over speed

[spark sonnet]

Exactly! Jules makes it somewhat safer to vibe code and more accessible (can even be opened on a phone in the browser)

[river elk]

Do you think Jules is better than Codex?

[night oriole]

Gemini CLI is my daily driver

I mean, I'm clearly biased. But I think we all use different tooling as well. I was big on Claude code for a while. Codex is one of the few I haven't tried at all

[spark sonnet]

Jules is obviously first. It is way ahead of competitors. Codex web version was clearly made after Jules, and it takes inspiration from Jules, obviously. It still catches up, but it's gaining. It gives user more control than Jules. But if Jules bugs will be fixed, it would keep being on top

[river elk]

Does anyone know when Jules will start charging?

[hybrid sequoia]

Claude code > gemini cli > codex > jules

[hybrid sequoia]

impo cli is basically equivalent to autonomous coding agents

cos u still have to lord over what theyre doing or else garbage will be put out

[outer ibex]

antigravity has a channel at https://discuss.ai.google.dev/c/antigravity/64

[halcyon wagon]

Gemini CLI 🤮

It can't even do simple things, and it's still using 2.5 pro... Google hasn't updated to use 3.0 pro...

[outer ibex]

relax bruh, dont be hating on something and glazes everything else

[rigid gyro]

Is it possible to customize branch names? I would like to be able to remove the numbers at the end

[outer ibex]

cant anymore

[night oriole]

https://geminicli.com/docs/get-started/gemini-3/

[oak bluff]

Is this legit?

[outer ibex]

ofc it is, its under google domain too

the main domain which is https://ai.google.dev for anyone to explore

[stuck quarry]

Hi everyone! Can someone help me understand why this is happening: I have an Ultra Plan and Jules let me use it like I have a Free Plan

[silver cave]

I’m not sure

[outer ibex]

if your google account are ai ultra workspace then jules doesnt support it, only personal @gmail account works with jules right now

but if your account are personal but jules didnt update your plan, then create a report to the devs

https://jules.google/docs/usage-limits/

[stuck quarry]

I didn't know about that, the Ultra Plan I have is paid by my company

thank you very much for the heads up

[serene fern]

still Latey flagship model ks not available for me even after turning on the "Preview feature"

[brisk hamlet]

Is someone from the Jules-team going to answer my question? : Will we get back the transparent mode? I.e. will we be able to see the commands and their respective outcome again, as it has been from the start until about 1 week ago?

[outer ibex]

either will come back, there will be a toggle for it or never

we never know

[rare galleon]

[prisma vine]

Hello @rare galleon friendly reminder videos and images are not generated within the server, you'll need to head to https://labs.google/, pick a tool, and generate it there. If you are asking for prompt advice please do so in the appropriate channel. Thank you!

[cosmic rivet]

Has anyone tried installing Jules Tools in a Docker Sandbox or container? I realize this might be viewed as overkill, but it’s how I’m trying to set it up and I’m finding it impossible to get Jules to connect. The sandbox network proxy config is set to allow everything that should be needed I think, but I get like 15 ECONNREFUSED errors when trying to start Jules.

[jolly vigil]

"ready for review" I don't do these at all? I have like 70...

[harsh knot]

I’ve been having the same issues with Linear mcp calls.

What did you do to fix it or were you able to at all?

Are the MCP calls to the partner integrations just buggy right now? I’m having trouble getting the agents to post there issues to Linear. They create the text and all just don’t make the mcp call.

[simple rain]

yeah, severely more eager PR-minting, even if then just updated and later squash-merged, would be helpful to alleviate those kinds of issues/bugs.

[fair wyvern]

HI

[simple rain]

not english can't help you.
Also: the URL shows 500, that's suspicious.

[fair wyvern]

where aree you from

[simple rain]

I can't read french

[fair wyvern]

i am from ALGRIA

[jolly vigil]

weird, it is 12:30 EST

can these not be stored in the repository under .jules? that'd make management much better

[wispy garnet]

Hi , i am seriously afraid to keep using jules. All of google AI/LLM products are quota-bullying users for quite some time now. Just look at the Antigravity fiasco going on.
So my question is; How real is the chance, that also Jules, will get squelched so hard that a pro subscription wont cut it anymore ?

[spark sonnet]

Wym Antigravity fiasco? And with competitors basically on the tail I don't think any of these products can afford to 'quota-bully users'. Do you use the products correctly?

[wispy garnet]

Really? that is your response? https://discuss.ai.google.dev/hot
am i using the product correctly? are you always this rude and passive aggressive? How can someone use something at all if quotas change and you have no real way of checking them. You mr Register, are not even a google employee OR mod? so keep your insults to your self.

[spark sonnet]

I only asked for clarification of your post. So the thing is that you can't see what the current remaining part of the quota. I agree - seeing this would've been quite useful. However, that alone doesn't make Antigravity a 'bad product', it just needs more polishing

[wispy garnet]

spare your fingers dude. i blocked you right away

[raw bronze]

Hello all! Please be reminded of our first rule: Be compassionate and treat others with kindness and respect. we have #1238621511401734254 for any feedback specific questions/ queries.

[outer ibex]

if you want to see the quota usage in Antigravity, you can download Antigravity Quota in the extensions menu

[halcyon wagon]

Does Jules support skills or is it only reads the agent.md file?

[toxic perch]

Hi, i start using jules yesterday, i still couldnt find how to push to github or download files option, is there anyone to help?

[crimson maple]

How are people using Jules in combination with antigravity projects?

[outer ibex]

for me i use jules for foundation files, antigravity for continuing the progress till completion

[minor garnet]

i dont use jule

[modest niche]

HI
What kind of this note book

[outer ibex]

i dont understand what your question is

[hybrid sequoia]

I honestly feel like without any transparency jules is genuinely unusable right now

maybe I am doing something wrong, but in comparison to gemini cli(which to be fair has super low rate limits for 3 flash and pro) jules just doesn't show you what is happening, takes forever(which WOULDNT BE AN ISSUE) if it werent for it ignoring code review

[keen flare]

I connected Jules to my Supabase MCP—or at least I gave him my API key—but he's getting an error: 'MCP Call Failed connect Supabase'.

[swift karma]

I simply stopped using it

[hybrid sequoia]

same

gemeni cli rate limits are super low unforch so im having to use 2.5 flash

I believe that the jules team needs to make up its mind, are you targeting developers or non developers, you need to give the devs all the info to make your tool genuinely useful

[halcyon wagon]

So, what do you use?

[brisk hamlet]

Is someone from the Jules-team going to answer my question? : Will we get back the transparent mode? I.e. will we be able to see the commands and their respective outcome again, as it has been from the start until about 1 week ago?

[swift karma]

Vscode and OpenCode

[minor garnet]

Use open code or build your own coding agent

[green walrus]

Lol damn

Everybody just ragged hella hard on Jules from my original question.

I guess my question is answered! Down bother using jules!

[rigid flicker]

Hello Google Team,
I have not used Jules for some time and came back to test it again. I had come across this error months ago too. This is my first time trying Jules again and it works for a long time on a task and submits nothing at the end stating it being completed. Can someone give me resources to fix this?

[stoic estuary]

N

[halcyon wagon]

Where do you get your API keys from? I tried using antigravity, but it doesn't work well. It often fail to generate codes

[outer ibex]

gemini models in antigravity right now is having a seizure, always failing to read and edit the files and considered it a successful changes

so use other external models

[outer ibex]

i guess im the only one left using jules just fine

[vernal compass]

JULES ISN'T asking me questions before making plans anymore (in interaction mode ofcourse)!!
This is a serious issue... It's been doing this for 3-4 sessions... Is SOMETHING wrong with it or what??? It's DIRECTLY making the plan and avoiding to ask me clarifying questions.....

[swift karma]

archive jules 🤞

[serene fern]

Send this link to ANTIGRAVITY and give your Jules api key: https://pastebin.com/raw/vx1G45k8

[royal violet]

Hello, please submit any errors and bugs you find in #1238621511401734254 . This will help ensure that your message doesn’t get lost here among the chat. Staff regularly monitors those threads.

[rigid flicker]

thank you

[next socket]

Does anyone else have Jules get stuck/timeout without finishing? This happens for like 60% of sessions for me, and the UI doesn't really tell me what is going on; like it will get stuck on " Read apps/api/tests/conftest.py" a few minutes into the session and not move on from it.

I've tried to debug myself, but I think I'm gonna have to make a bug report as I can't figure out what is causing it.

Also does anyone else have Jules being like was slower than Codex (like 5x slower)? Not sure if that is expected too.

[brisk hamlet]

Code review is now broken.

[polar shore]

Hmm

Anyone notice how the Go library has iterators & open telemetry now?

And a 3x perf boost to stringify? - That was Jules modified with my frameworks 😄

[minor garnet]

I dont understand

[polar shore]

Basically - I gave Jules immense constraints via reasoning frameworks

And had him take a go at “making Google Go’s GitHub faster” then again “find a feature that’ll make it better” and a third time - didn’t touch a single line of code. It did it all - all 3 got merged successfully

Normally AI hallucinates and messes things up - but my specialized reasoning frameworks + Jules built in VM and rigorous testing = senior engineer level code via simple prompting

That clarify things @minor garnet ? You can check the PRs yourself my GitHub is merchantmoh-debug - it’s hard to believe but real code doesn’t lie + I screen recorded the whole thing 😁

[outer ibex]

i dont understand it either

[halcyon wagon]

what the f*ck 😭
Jules’ first commit was actually good, but the second commit is literally blank.

It looks like it didn’t read my prompt at all and just did some verification stuff, then pushed a commit with only a description(related to first prompt) and no real changes.
I faced the exact same issue months ago, and now I’m facing it again. This feels like a recurring problem

At this point, please suggest better alternatives, and also explain how to use them in the free

[minor garnet]

I feel like

This poor Jules is being asked to do th impossible like hey Jules build me an entire market place

[outer ibex]

agree

at this point you dont even need to complain if youve had enough, just use other alternatives

[manic meteor]

I have the same experience. I brought this up with a few concrete examples a couple months ago, before the Discord server migration, but nothing happened. Jules appears to be able to handle small, quick tasks, but hangs on large or long running tasks. The memory usage also explodes with large or long running tasks. My suspicion is that there’s a combination of problems here: relying on too much client side memory, which fills up and then blocks the task from completing; poor context management for large or long tasks; and a lack of checks that account for one or both of these problems, that would enable Jules to complete it’s task.

[nimble plank]

When will jules will be available to second grade countries

Jk ** to the rest of the world

[hard badge]

I see we can have an AGENTS.md file with all the information regarding any agents or tools in the codebase, Jules will read it and follow along. Can we save some information regarding how to migrate the codebase/ summarise the codebase in this file? Will Jules read it and actually use that information?

[ashen glen]

It should do!

[oak roost]

Could you give a recent Session ID for this issue.

[brisk hamlet]

Is someone from the Jules-team going to answer my question? : Will we get back the transparent mode? I.e. will we be able to see the commands and their respective outcome again, as it has been from the start until about 1 week ago?
This question has been raised several times now. We got totally ghosted by the devs on that. Please at least have the courage to be honest here.

[night oriole]

Update the memories for the repo. That memory store gets loaded in the prompt stack always. The agent knows of agents.md but it might not get loaded into the context

[vapid trellis]

Question for you guys. How do you deal with just the sheer amount of PRs that Jules ends up generating? Auto-merge combined with good CI pipelines or just some other way?

[vapid trellis]

I always have this issue. Jules does not give me any output if I ask it to do something without starting with the repo https://jules.google.com/session/14869836317670238068 from this morning, I think.

[undone tangle]

I’ve wondered this as well, I haven’t tried this but I wonder do you just create like a “Jules work branch”, try to group related and then stack the prs in that branch merge them in order? The idea being you just toss the whole branch if it doesn’t work out because you have so much quota per day

[polar shore]

Nah

Go check out Remember Me AI by merchantmoh-debug and come apologize when you’re done

[polar shore]

Oh no!

Not a thumbs down by a random guy on discord!

How will I ever recover from the horror of your disapproval?!!!

Man. Y’all need to up your troll game. I eat trolls for breakfast, lunch AND dinner too.

[polar shore]

What happened?

I thought you were trying to throw shade?

Let me guess you searched it up found it was an open source project under MIT license that has over 51 stars from founders of the most successful AI companies in the world.

And you’re so embarrassed that you can’t even come and fess up and apologize.

Not surprised.

[polar shore]

🤣🤣🤣

[west venture]

Well, Jules is BUSTED. Edited, touch, modified, created all of these files... yet the changeset has none of it. This is essentially worthless.

[visual gyro]

When update jules waiting something better codex 5.3 please 🙏🏻

[weary oracle]

I heard the Jules team is cooking something and need testers? Happy to throw my hat in!

[minor garnet]

someone know a way to make great ui

[night oriole]

Stitch.google.com

[minor garnet]

I mean connecting the backend to a frontend

[night oriole]

Hey folks, remember to please strive for civility here. There is zero need to talk down on each other

[minor garnet]

I went to try stitch and legit waited for 40mins without getting a result

[minor garnet]

Them bills are high

Money make people angry yk

[night oriole]

Jules has a substantial free tier
But that doesn't excuse rude behaviour in this server

It seems to be recurring and we all need to do better here.

The internet is angry enough.

[minor garnet]

But like

I feel like Jules has become a everyday tool for some people

Because its free and basically illimited

[night oriole]

Which allows y'all to be hateful and harass each other?

Again. We need to do better here. Everyone.

[minor garnet]

yea

i get it so i should use jule?

[dusty cosmos]

As a happy Jules user I don't understand all the hate... The platform isn't perfect, but it is new and folks are working hard and making it better. It's levelled up it's game in the time i've been using it over the last 4 months! Thanks @night oriole for making something cool that's gotten a few ideas of mine off the ground

[night oriole]

yes, of course you should, but using it doesn't give permission to treat people poorly in here

[minor garnet]

lets see

[balmy narwhal]

help due to all lags on website I cannot access button to publish repository after jules finish its work how can I fix this

[minor garnet]

Thats the neat part

[balmy narwhal]

how its that neat?

[minor garnet]

You never know when its going to lag so its just gambling

[balmy narwhal]

yeah I wanted to test these functions so I cant just download code because it started 5 diffrent chats

[minor garnet]

Do you make money with ai ?

[balmy narwhal]

lol no

[minor garnet]

Yea

[balmy narwhal]

but Im curious person so I want to test possibilities of diffrent apps and functions

currently Im trying to build an app that based on your GPS location shows closest object in space to your vector from your head to this closest object

[night oriole]

there is a tremendous amount of data going to the browser, we've been working on stabilizing that for some time, you might just need to load the task and wait a bit

I'm learning sensor fusion on the side for autonomous drone operation, sounds very familiar

[balmy narwhal]

nice dog anyways I wish CLI had more options

[night oriole]

the jules cli? or gemini cli?

[balmy narwhal]

jules

like be able to publish branches from CLI so it would be more effective

for now I have such thing widget is gone and task is completed according on it therefore tousands of gallons of water wasted and I cant even get my web app

I'm pro user btw

[night oriole]

You might need to try a new task if it's stuck like that, unless "commit all the inprogress code and push"

well, the good news is that the datacenter water is in a closed system, it's not lost, it goes rigth back into the system

[balmy narwhal]

it already 3rd task on the same thing plus I enabled to this repo this thing that created few diffrent chats that dont take credits

I managed to publish outcomes for most if not all of them but this major one stuck

[balmy narwhal]

what is this sounds interesting

[night oriole]

what's the task? is it failing the same way each time? can you share session IDs so someone can take a look?

[balmy narwhal]

https://jules.google.com/task/17109196511145701059

[night oriole]

cc @leaden pawn

[balmy narwhal]

who is this

[night oriole]

dev team - I can't check myself

and oh! just making a little drone guy fly itself - camera, gps/compass, lidar, etc all getting read so it can make decisions, no dissimilar from your GPS project

[minor garnet]

holy

[leaden pawn]

Let me check this

[balmy narwhal]

Sounds very interesting could you tell more

[balmy narwhal]

How you plan to fix this

[rigid flicker]

Hello, sorry I know this is not the Channel for Antigravity, but maybe any user can help me here who has Google AI Ultra.
I want to upgrade from Google AI Pro to Ultra, are the limits higher in Antigravity?

[dusty cosmos]

Are there any future plans to allow customizing the branch names in the UI? The current branchs kinda end up random names, sometimes with feat/ or fix/ or neither and are now always suffixed by the Jules job ID number. Also, empty commits are made when just responding to PR comments, these I clean up manually and force push over things, or tell Jules to comment only and don't push but it is convinced in it's instructions that every PR response needs a commit, changes or no changes 😅

[idle chasm]

is there a way to spin up a postgres + redis instance in jules?

[outer ibex]

they gave you higher quota limit, thats it

[outer ibex]

early jules can do that, but idk why they removed that customize branch name feature 4 months ago

[dusty cosmos]

That's interesting, right as I started using it

[polar shore]

"Yo, so I got tired of Python eating my RAM like it’s an all-you-can-eat buffet, right? So I just went ahead and built my own stack from scratch. It’s called Ark.

Basically, I wrote this custom runtime in Rust—yeah, the metal—so it’s strictly ownership-based. No garbage collector pausing your game or your model. It uses Linear Types, which is just fancy speak for 'if you use a variable, it moves.' You don't copy data; you teleport it.

Because of that, I’m getting Zero-Copy FFI. I can talk to C libraries or raw memory buffers without marshaling overhead. I’m talking sys.mem direct access.

And the best part? The whole thing compiles down to this MAST (Machine Abstract Syntax Tree) protocol. It’s lightweight JSON instructions, not heavy binaries.

So, instead of needing 32GB of RAM to run a local LLM because the OS is bloated, Ark runs the logic on like... megabytes. I’ve got it doing async HTTP serving, Merkle-tree crypto hashing, and even file system control, all native.

I’m actually halfway through making it self-hosting. I wrote the Lexer and Parser in Ark. Once I finish the backend codegen, the language will literally compile itself. It’s the Ouroboros, bro. Infinite loop of efficiency.

It’s not just a language; it’s a Sovereign Compute Stack. You could run a whole agent economy on a potato with this thing.

[outer ibex]

i already saw someone else done that but okay then

[brisk hamlet]

Is someone from the Jules-team going to answer my question? : Will we get back the transparent mode? I.e. will we be able to see the commands and their respective outcome again, as it has been from the start until about 1 week ago?
This question has been raised several times now. We got totally ghosted by the devs on that. Please at least have the courage to be honest here.

[swift karma]

They didn't even fix the other bugs that were reported, it feels like they're doing nothing that's all.

Jules was a decent product on the beginning but now it is completely unreliable. The developers say they're doing whatever but I don't see any progress. The website still unstable, Jules still peforms slow and fails to do what you asked. Giving feedback or whatever doesn't matter because those issues seem to never be solved. Many times I've tried to understand the developer team but i just can't anymore

[serene fern]

@ionic cliff We trust you and know you’re putting in real effort. If you’re bound by an NDA, just say so—otherwise, please keep things transparent. Jules has the potential to disrupt every other agent in the market. It’s currently unstable, but the user base is massive.

TRANSPARENT MEANS = We are working on ... that's we all need.

I BELIEVE YOU ARE best LEADER for Jules seeing your past works

ANYBODY EXPERIENCED Jules EA VERSION HERE?

[brisk hamlet]

Please stop spamming capital letters.
Also: Please write in plain english what you are asking.

[serene fern]

Seriously? What is wrong highlighting a question? and Plain English means? Can you elaborate? Whatever I have asked most of the people will understand. There is no non-technical person

[balmy narwhal]

is antigravity just another fork of VSC

[balmy narwhal]

sorry for ping but when I will be able to commit this code

[outer ibex]

obviously

but the antigravity itself is closed source

[balmy narwhal]

sadly I would lurk into it

[outer ibex]

lurk around the antigravity output then

[balmy narwhal]

tho Im dumbass I cant even pass exam lol

[outer ibex]

lol

[outer ibex]

i used jules a month after it released so does that count

[balmy narwhal]

where you found it

[outer ibex]

keeping in touch with google labs that is

website i mean* not this discord server

[serene fern]

There is also another EA running right now! I am talking about current one.

[outer ibex]

oh dang im not keeping up for now cuz im busy with fixing libraries and sdks

[balmy narwhal]

this app is sus I just tested it

and it makes me run billion commands for no reason and does everything other than solve code did google designed it specially to drain credits

[outer ibex]

it all comes down to how you do your prompt

[terse dragon]

Has anyone had a successful Jules task in the past hour?

Now seeing 503 service unavailable errors

Just had 1 make it through

[outer ibex]

its still running for me, though i use their api

[outer ibex]

this could be rate limiting

eh no, prob just a fluke

well if you encounter any problems, do report it to the staff

[minor garnet]

do yall like my marketplace

[near ocean]

Hello, I need to download a file, but it won't let me and isn't responding at all, it's stuck on summary, what should I do? But it supposedly already responded, Gemini PRO.
Either I'm missing something, or he really takes that long to respond

[rigid flicker]

If you want Jules to give you a compiled binary, it will not work. Don't try. If its another file, I don't know why

[white gyro]

It's like that joke about testers. They use the app in an 'exotic' way. I don't even know how it was possible to arrive at such a result.

[minor garnet]

your website is hella laggy

[rigid flicker]

its the motion blur. are you on windows?

[minor garnet]

yep

[rigid flicker]

yeah then its the motion blur

Its disabled for OS X and Safari Browser because it just does not work there at all

[minor garnet]

why did you put it in the first place

i thought my browser was gonna ccrash

[rigid flicker]

looks cool, does not lag for me at all

lol sorry bout that xD

[minor garnet]

but i find your site cool do you make money with that?

[rigid flicker]

No I just make it out of fun, I just like building obfuscations

[minor garnet]

what is it?

[rigid flicker]

You mean what is obfuscations?

[minor garnet]

yes

[rigid flicker]

It is when you make code harder to understand or interpret without changing its original function

I can show you one example I made if you want

[minor garnet]

basically your code become impossible to read but it works?

[rigid flicker]

Yes exactly.

[minor garnet]

i heard many people do that with c++

[rigid flicker]

Dont get me wrong btw I love open source, I just do it for fun because I like reverse engineering these things

but wait I show you my example, I just need to upload it to my site right quick!

[minor garnet]

no i get it thats a valuable skil you can make many money also can i ask you smth

[rigid flicker]

sorry for the wait its compiling right now

rust projects slow on my pc

[minor garnet]

rust is hella hard

[rigid flicker]

end2end.space/pastes/5iFybmE6Pmuj this is the one that is not obfuscated by my obfuscator

and this is with protection:end2end.space/pastes/kIA219I3XIsk/raw

[idle chasm]

is there a way to ssh into the containers that jules uses?

[rigid flicker]

Sadly I don't think so. I have looked everywhere because I also needed this.

[minor garnet]

Ask him

[rigid flicker]

😂 true good idea

[leaden pawn]

Please share the session ID in #1238621511401734254 channel so that we can check

[leaden pawn]

Hello @brisk hamlet Sorry for the inconvenience and we definitely aren't ghosting! There is an active bug for this and our engg team is working on a way to bring back that log visibility so that you are not flying blind. I don't have exact dat yet, but it is a priority for the team.

[minor garnet]

good luck gng

[spark sonnet]

And I had issues with Jules giving me binaries I didn't request... So I guess it depends on whether you need that binary or not. It would always be the opposite

[rigid flicker]

interesting, i never got it to

[leaden pawn]

is there a way to spin up a postgres +

[visual gyro]

@leaden pawn what's cooking?

[orchid jasper]

there an outage today or something? jules is supposedly pushing branches but no changes are being made

like it's pushing "commits" but none of the files are changing lmao

[balmy narwhal]

to any ULTRA users do you also experience lags

[balmy narwhal]

cant you just ask Jules to dockerize your app and run it locally

I do this that way its easier to host and stuff

[leaden pawn]

A lot of behind the scenes fixes! Most importantly, paid users should have full access regardless of their IP and looking into transparent mode bug that everyone is asking about.

[visual gyro]

I missing something like codex 5.3 but something better

[minor garnet]

I dont understand

[visual gyro]

More compute and smarter models like Google deepresearch or deep think for harder task or bigger project more fetching data. Imagine rework on website and all subpages but you need keep all content or context just make website up to date latest standards and so @leaden pawn is this something we can expect? Say Google business profile connectors

[minor garnet]

Are you Willing to pay full price tho

[visual gyro]

I am already doing that multiple times 😅

[night oriole]

You wouldn't really be using deep research in a code session. You can use it to help research but it's not a coding model.

Gemini 3 pro and flash are sota coding models

[minor garnet]

Why when I ask Gemini 3flash what model are you he says Gemini 2

[night oriole]

I'm not able to replicate but they aren't self aware.

[minor garnet]

So he should say i dont know they didn't tell me but instead he lies

[night oriole]

Theres no he nor is there a concept of lying

[hearty glade]

I suspect that is due to its data being older and maybe at that point the latest was Gemini 2? 🤷‍♂️

[night oriole]

I think Gemini 2 is fully deprecated

[gentle spear]

I am just learning how to use Jules, and it's been stuck on initializing the environment for laravel 12 now for about an hour... I thought maybe it was because I am using the "trial", but maybe there is something else going on?

[rocky forge]

some buttons don't belong in the top navigation bar, poor contrast, all caps are bad for scanability, "sorting" shouldn't be a filter, make the filter sidebar collapsible, bad hierarchy especially in upper half

[outer ibex]

given that 2.5 flash is also being shelved by the end of march, i guess so

[minor garnet]

Yea

[outer ibex]

"jules spin up a fresh vm for every session" so no ssh, cli and api does the same thing as web. thats a bummer too but it is what it is

[minor garnet]

I free vps

All a dev can dream of

[outer ibex]

agree

[storm hazel]

is there any option to have Jules open the PR's in draft? I like to use GitHub to review the code myself before actually opening a PR.

[visual gyro]

Context is key say want update website and Jules agent code all information aspects viable for business deep research, if want rebuild website from latest info and standard and use Jules coding agent make better website . And want everything right . Jules should be best system go if you should rebuild something that has code. I think but maybe I am wrong

[night oriole]

None of that needs a research report. Deep Research is very powerful but it's pulling hundreds of sources for a comprehensive report. Can use it for competitive analysis and looking for options. But vastly different products

[balmy narwhal]

I enabled some option in Jules and now it auto generates chats that dont take credits they have ✨ icon nect to them what should I do with this

[night oriole]

That's the very beginning of proactivity. Jules is looking for performance enhancements it can make and it'll run the session without costing you. Merge it. Ignore it. Up to you

[balmy narwhal]

will it stack with other updates or overwrite code

[night oriole]

Code never gets overwritten unless you merge it. Merge early and often to prevent conflicts.

[balmy narwhal]

like if I click one commit and then other does this other overwrites other commit or they all build themself

[visual gyro]

This stats say something else https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-deep-think/

[night oriole]

Most of these are small changes, they shouldn't overlap with what you're working on but can. Case by case basis on the merge. But you don't have to merge. And you can rerun it if there is a conflict

[night oriole]

Not really. Says the same thing. It's for very big things. It's on ultra if you want to try it. It will not be free on Jules

[visual gyro]

Yeah but maybe not for all request if like say planning or when its really needed

[visual gyro]

Yeah it would be awesome but maybe you people can use it find some solutions and look at all aspects of the codes and system and find some improvements in Jules agent knowledge base or Functions just excited for it not saying should free but maybe available for pro at say limited request or something I can not say whats possible and what is not

[night oriole]

You wouldn't use deep Research to plan. Like I said, it's for very very big things. Not to code a website. It's only available on ultra and API because it's so compute intensive

It's not gonna happen but I promise that you don't need it

I highly recommend trying deep research out to see how it works

[visual gyro]

Yeah have more expertise I just clue less and excited for future so dont take this wrong way, it would be nice but more like wish so dont request it absolutely. Yeah I have tried deep research a bit pretty good but it would be nice use Gemini pro or like connector so Jules can do all it request if it need know more about the webpage or find links or what ever info it would need. This is just wishes of something but as say probably to expensive and heavy on the compute Intensive / request

[spark sonnet]

That still won't help. I have to create manually a Jules session just to fix some typos or styling in PR Jules just made. Why? Because just writing a comment like @jules fix this formatting issue on the code would cause Jules to reinvent the wheel, and start from the beginning. It can make same changes, can make different changes. Why? Because Jules doesn't pull the branch it's working on in the PR. If I need to address 5 small issues Jules made in a single PR, I either have to do it by hand without Jules help at all, or create 5 PRs just to fix 1 PR, all because Jules does not pull changes, only push. And yeah, I can't force Jules to spun a new session if it already has a session for the given PR - no way, it will always continue, and always without pulling, even if it's archived.

[minor garnet]

Chill bro

[swift karma]

Archive

[night oriole]

It won't pull, no.
Tasks should be completed, merged, then you'd open another task to fix formatting

Tasks should be atomic, if you're going back and forth, that's a pretty good indicator that you should be in the cli and not interacting asynchronously

90% of the problems I see are people trying to get a full steering cli style session or continuously adding more and more tasking until it breaks

So in that case, if there was a typo elsewhere, you'd likely merge (you can merge into an Agentic branch). Then open a task to correct spellings/formatting.

I'm not sure why you'd reactivate an old archived session, but the clone happens at task start. I'm pretty sure it won't reclone. So tasks should be considered a moment in time

But looping back to the proactive tasking, you don't have to merge if you don't want to. You don't even have to publish the PR. You can just ignore it then and there. Or if you did publish the PR and reviewed the code and it wasn't up to muster, you can just close it and archive the task

[wraith fractal]

hey, wanted to ask, did you ever heard about openclaw ? i think it would be great if you guys integrate openclaw in googles ecosystem ? openai and meta are trying to bag this guy omg 😱

https://www.youtube.com/watch?v=NMBoNFDOr_o

[night oriole]

I've been watching him build it for the past few weeks. Really interesting. Took off like a firestorm

Personally, I don't quite understand it. Its running in a heartbeat, cron and webhooks I think, I also don't quite buy the "my bot is fully automatic and building me stuff"

The whole moltbook thing was mostly humans https://www.builtbyvibes.com/#blog/moltbot-isnt-alive

[wraith fractal]

ohh thats interesting, but I think google as an ai company always does things late, like perplexity comet was launched last year, the first ai browser which was built on top of chromium and then google chrome ai agent was released this year but its still available only in the US and then even google antigravity which was forked from windsurf still lacks a lot of basic editing tools, i hope google wins the ai agent race after all, like ai agentic chrome os or agentic android os

[night oriole]

Jules has actually been ahead of most of the labs. Skills, agent derived memory, some other sauce.

Perplexity I have my own thoughts on, but they're not really around. Windsurf I don't know the deal details but that team is still on it. I do think Claude code has a better CLI harness.

Marathon, not a sprint, and there's always going to be leapfrogging by all the labs.

[wraith fractal]

it was quite fun to know moltbook was just human prompted for social media posts haha, in the news it showed like ai acopolypse was coming soon haha

[swift karma]

To be honest what's the point of Jules after all? Basic tasks?

[rocky forge]

I have a unique use case. I use it to implement things where multiple approaches are possible but I am not sure which one is the best.

So what I do is create multiple tasks for the same thing with different approaches and then simply compare which branch best solves my problem. Delete the rest.

This also works quite well to do UI/UX mockups in the background.

[swift karma]

It doesn't seem reliable at all for complex tasks... Put Jules and Antigravity to do the same task, same model at the same time and Antigravity will finish it way faster.

Yeah, i know people are tired with comparisons but it's like Jules has some sort of Alzheimer's

You give it instructions, it doesn't follow them correctly. It peforms slow and the only thing you can do with the code it made is download it or push to Github on a new branch, you can't even push changes to the same one

Jules takes more time > machine stays on for more time = less resources.

Jules uses Gemini 3.0 Pro/Flash are very smart models, i dont think it was supposed to be used so poorly as it currently is. Yeah, i know the limitations of context but I'm pretty sure it can handle a few rounds on the same task.

Gemini 3.0 Pro peforms very well anywhere else i know such as Gemini CLI, Opencode, Antigravity, whatever. I feel Jules just isn't mantained enough at all

I don't understand how Jules works under the carpet but i can't help but criticize the way it works currently, peforms slow and the results aren't much positive if you decide to not make a 150 lines prompt for it

Even a issue I've reported about two months ago about it's memory about a repository did not get fixed yet, i think it's a waste to have Jules performing like that

If you guys made Jules work faster I'd say it would mostly satisfy users, tasks would be more faster and there would be less load because Jules wouldn't take 30 minutes to make a task normally takes 10

I keep trying to understand the side of the developers about this but it's a little hard... I mean, two months and certain issues still haven't been fixed. I remember last year things were going faster and don't understand what happened, just lack of communication and adding features while there's some bugs to fix

[white gyro]

is the other way around. THEY need Google.

[dusty cosmos]

Codex has this matrix job feature built in. It's cool but their UI is a bit hard to compare the multiple jobs against each other. The method you've divised would give you PRs you could pit against each other

[night oriole]

Engineering is working on these. They triage every day. There's PMs looking things over. It's an extraordinary complex product. We're trying to have an AI agent build by itself and report to you when it's done.

We were copy pasting code out of a chat window less than a year ago

No failure is the same.
It's a smaller team than you'd think in labs, you should probably consider most of these products beta or alpha builds

A lot of the new functionality is chasing the next model not necessarily this one

Which leads to some tension between where it's going and where we are

But we're people just like you are.

[dusty cosmos]

It's pretty amazing what you've built. It's easy to see a feature or not see a feature and have an opinion based on what the output is, but there is also input to making the output, and they're not really 1:1.

You're trying to get a highly advanced pattern matcher make novel solutions autonomously while having the underlying foundational model be sourced from a different team working hard to make it general enough it can work across many use cases. That's a lot of input to give it a box to think outside the box from.

A fun and interesting proposition, to say the least, but also hard!

[outer ibex]

lol i also thought the same thing

[somber thicket]

Is there any hope for a Jules Desktop agent?

never mind, I just forgot that I was never able to install the local client

[mint hinge]

Hello, had a question about connecting repos, I've tried to add new github repos but the new ones don't populate, even after disconnecting and connecting again, what is the fix?

It's like there is a stale cache that doesn't allow for repo changes in the list besides disconnect and reconnect for all repos

[hazy viper]

Greetings, friends! please tell me, I decided to try jules today and was very impressed! when he made the first version of the site for me, I had a button there to download a zip and, as a matter of fact, I could watch the code that he was making and write how many lines of code he had changed. Next, I decided to make some edits, he writes that he made them, and it seems like he even applies screenshots of what he did, but the trouble is, I don't have a download button now and I can't publish it to my github. That is, I just have a summary and that's it. I have already asked him to upload it directly to my github, so that he forcibly stops his generation and marks it as completed, so that I have this button, but for some reason everything is in vain. How can this be fixed? Because now I can't even look at the code he's changing.

[dire wolf]

is it this? 👀 https://github.com/google-gemini/gemini-skills

[visual gyro]

My repoless task no data or project is not showing anybody else having same problem?

[river vector]

Hey Google Team, I have a few ideas to level up Jules, because in the current state, this tool is hard to review and use in parallel.

For example, I commonly give a huge task who is really well detailed, get some rest and review the next morning. But this is a huge pain to review because there is a single commit in the PR. In that case, I almost always stage modifications in my ide one by one to split code in smaller commits.

Sometimes, I use multiple Jules to do multiple tasks at the same time. In addition to the previous problem, this is common for agents to do the task of others. This is a pain when I need to merge edits.

To summarize, please do more commits, within the GitHub pr, please write a summary of what problem, if any, happens, what was added and somehow if possible, make Jules respect orders.

[visual gyro]

I am have same problem with repoless task no code showing

[hazy viper]

the most interesting thing is that I decided to try downloading the cli version and pulling it out via pull in the hope that it would work, but It wrote that No diff found in the remote VM and that's it. Fortunately, I saved the first version that he generated, and in the same dialog he asked me to write a suggestion on what he had done so that I could copy it and paste it into a new chat to repeat and recreate the code. I uploaded the first version to github, started a new dialogue, completely restored the original one with a big boost, and the one that he gave, and now I'm sitting here figuring it out.

[hollow idol]

Hello folks - is anyone else having a problem with "Jules encountered an error when cloning the repo"? I've been using Jules for months but tasks have been immediately failing with the same error for the past 4 hours.

[cobalt dawn]

Is there an official mobile app for Jules?

Jules has been lagging hard for me. I have to wait(time range 10 to 60 per task I give it.

I'm using the mobile web browser to do this.

[safe cloud]

Is anyone else able to create a api key? I'm having a problem with generating a api key,have pro account upgrade
Tryed solution : no current api keys [3 max]
I went on cloud console enabled the api [gemini suggestion],
Create a service account on cloud console jules api credentials with role owner and usage with my email.
but still unable to create a key, just get a error ; Failed to create api key, please try again

[visual gyro]

Repoless task are not working the data

[visual gyro]

Anyone else having that all files dispersed??

[modest hornet]

it started as such a nice project, now a small task takes ages, dunno what happened

[modest hornet]

how can i TURN OFF these pre-commit steps??????????

[minor garnet]

[lament mist]

You can stop this step by creating a AGENTS.md file in the root directory and adding necessary instructions in it for Jules to follow but do not know about how effective it will be ?

[modest hornet]

thank you sooo much ❤️ i will try asap. atm jules is unusable like this

[lament mist]

Well Jules is still a beta version and it can take some time based on projects complexity but still 3 hrs is too much I guess.

[halcyon wagon]

Yeah, Jules usually takes a lot of time. That’s why I assign the same task to both Codex and Jules at the same time.
I first check the result from Codex since it’s always faster. If the output is good and complete, then everything is fine and I move ahead with that. If the Codex result is not good enough, I wait for Jules to finish.
And if Codex already delivers a solid solution, I end the Jules session so it doesn’t keep working unnecessarily.

[minor garnet]

Is codex a agent like jule ?

[minor garnet]

Is it free ?

Tho im scared my code will be sold on the black market

You paid for it ?

[halcyon wagon]

🤣, chill, it's by Open ai (chat GPT)

[minor garnet]

I remember chat gpt reported a kid because he was role playing as a school shooter

[halcyon wagon]

If you have a ChatGPT subscription, you also get access to Codex included with that

[minor garnet]

So you paid for chat gpt

[modest hornet]

but the code-outcome was AMAZING!!!!

[lament mist]

Well done 👍😊

[modest hornet]

imagine we could share some "local resources" so jules could off-load things to our machine. like an addon-runner

[hazy fog]

Please fix the crash bug when searching for files in a GitHub repo. And fix the bug where code changes sometimes don't appear in the tab.

[peak cobalt]

please can all be reminded of our first rule: Be compassionate and treat others with kindness and respect. we have #1238621511401734254 for feedback specific questions/ queries

[minor garnet]

i bet the security flaws were amazing too

[modest hornet]

security is an illusion 😉

[minor garnet]

so sql injections remote code execution brute force ddos allat doesnt exist for u

[modest hornet]

exists, but as long there is a human in the chain, nothing is "secure"

[minor garnet]

yea right so your app will be used by aliens?

[modest hornet]

😄 im not arguing against you, just wanted to point out, the one sitting in front of the screen (me), is the biggest security hole

[minor garnet]

i dont blame you to be fair 80% of the internet have major security flaws because of ai

[spark sonnet]

As long as there is a clanker in the chain, however, it's not a security incident, it's an allowed leakage

In my humble opinion, people not really understanding security (which is purely a technical implementation thing, verifiable and tangible) confuse security with safety (political-emotional concept, an ephemeral idea, not actually implementable as for everyone what is considered 'safe' is somewhat different). "This human will not allow a malicious party to obtain trust of someone from our collective and gather data" - this is about safety, not about security. Safety does not exists in reality. "This cipher is secure enough to prevent brute force attack for a few hundred years" - this is about security.

[swift karma]

Are you guys still on Jules

[orchid pebble]

I need these MCP configurable by project. The stitch MCP is frequently triggered for a backend project which removes the value to the whole jules.google.com solution losing many session making Jules unusable

[spark sonnet]

Quite hilarious! I require Jules to write a per-turn journal. Quite revealing about how it skips Code Review actually:

Turn 19

• What action I do: Cleaned up *.log files.
• Why I do it: To comply with AGENTS.md.
• What was the result of previous action: Files deleted.

Turn 20

• What action I do: Requesting code review.
• Why I do it: Final step before recording memory and submitting.
• What was the result of previous action: Pre-review cleanup done.

Turn 21

• What action I do: Realized I missed calling request_code_review.
• Why I do it: To follow AGENTS.md.
• What was the result of previous action: Turn 20 was logged but tool was not called.

Turn 22

• What action I do: Recorded learning using initiate_memory_recording.
• Why I do it: To follow pre_commit_instructions.
• What was the result of previous action: Turn 21 was logged.

Turn 23

• What action I do: Submitting the changes.
• Why I do it: Final step of the plan.
• What was the result of previous action: Plan step marked complete.

[visual gyro]

Any news about Jules agent?

[frigid robin]

Am I the only one getting "Jules encountered an error when cloning the repo" messages looping in my conversations?

[cold lava]

Feature request: "Mark all as reviewed"

[visual gyro]

@full viper antigravity Jules connector when it mcp? What happend with cooking please use Gemini deep think improve Jules if possible 🫣 and this above also. Some guides and use cases video would be nice also but I know I am asking to much. Hope something is cooking and shipping soon?? Roadmap?

[light basin]

I like that Jules can run in parallel, but sometimes the scope of the work it does steps on each other and a branch once updated on Github has conflicts to resolve. Is there a way to make Jules aware of new commits and rebase its thinking based on the new main? I've tried asking it to look at the main branch again, but once the pull request is in, it seems to be too late.

[fathom blaze]

hi, anyone had all their scheduled tasks just stop and be marked as paused?

[west venture]

Pretty sure I just found a bug in Jules

The 15 simultaneous task limit on AI Pro is not being enforced

I just spun up all 57 suggested tasks for a single project and it's working them all without simultaneous task quota limits.

It just started doing this. 😄

Okay, ALL Jules quotas just stopped being enforced(
!)

and I can keep on spawning tasks, and they are all proceeding

[celest fable]

Jules no longer gives me a link to the PR in my GitHub repo. It started giving me a purple button that reads “Publish PR”. When I press it nothing happens. I asked Jules what the issue was and it said that they could no longer access my repo and that was the responsibility of another tool.

“I apologize for the confusion. I do not have direct access to git push or create pull requests on GitHub via this interface; my capabilities are restricted to an internal process that handles the commit and push to a specific branch name.”

Anybody else have this issue? Or know what is going on? Thanks!

[dusty cosmos]

It might be you need to resync your auth with GitHub. I had an issue like this and reconnecting GitHub fixed it

[west venture]

Oh God I need to stop, this is addicting

I ran over 500 tasks, I've inundated multiple repos with hundreds of PRs.

[river vector]

Mannn

[swift karma]

explode their servers

That's the only way to make them fix Jules

burnnn 🔥🔥

[raw moth]

I’m having the same issue. I published a couple or PRs then it’s just stopped, and I get an XHR issue on all 3 of my sessions after a while when clicking publish.

Is there a way to reforce the git authentication without losing the active sessions?

Exact error after a timeout when clicking the Publish PR:

Error code = 7, Path = /_/Swebot/data/batchexecute, Message = There was an error during the transport or processing of this request., Unknown HTTP error in underlying XHR (HTTP Status: 0) (XHR Error Code: 6) (XHR Error Message: ' [0]')

[full viper]

It's ready for use now! https://github.com/google-labs-code/jules-sdk/tree/main/packages/mcp

[tardy whale]

/image

[dusty cosmos]

I didn't lose anything when I disconnected and reconnected, everything just fell back into place

[raw moth]

No luck for me with that approach unfortunately

[dusty cosmos]

oh shoot 😢

OH! I think I went to GitHub and deleted the OAuth integration there as well, then reauthenticated from Jules

[open sand]

Is there some other person that has Jules not working?

[royal violet]

Hello @tardy whale as a friendly reminder, videos and images are not generated within the server, you'll need to head to https://labs.google/, pick a tool, and generate it there. If you are asking for prompt advice please do so in the appropriate channel. Thank you!

[open sand]

Got these errors since 6PM Paris time:
Unable to provision a virtual machine for Jules to use.

[gilded sequoia]

Hey is there an open ticket for Jules not picking up comments from gemini code assist GitHub bot?

[light basin]

Why?

[visual gyro]

@labs my first impression looks nice

[brisk hamlet]

"Unable to provision a virtual machine for Jules to use."

[west venture]

Because the sun shined, so it was time to make hay, or:

If I have the opportunity to get as much work done as possible, I'm going for it. I have a workflow on my end to get all of the Jules auto-suggestions merged in, even when two different suggestions change the same section.

[dim berry]

same problem

[west venture]

Damnit, did I break Jules?

[open sand]

It's broken for me since 6PM 😅

[west venture]

ah, no, I went nuts at 10pm last night once I realized the quota enforcement wasn't working

Had every single autosuggestion cooked all at once

In the meantime, I now have Gemini CLI and Antigravity doing the agentic merge process for all of that

I didn't go overly ham - I just set up my massive batch of tasks and went to bed. Had to resolve a few questions when I woke up, further spreading the system load over time

https://github.com/segin/genteel and https://github.com/segin/substrate

just look at the PRs 😂

[visual gyro]

I probably have nice website show soon the tool works great so far

[spark sonnet]

If anyone wonders why really Jules is so slow, here is the answer. And no, it's not because it's so 'deep'. It's because it's absolutely confused even about such simple things as to know what tools are available to it:

## Turn 119
What action I do: Attempting to use edit_file instead of patch_file.
Why I do it: patch_file was reported as unknown, but the instructions describe Git Merge diff format which is used by edit_file.
What was the result of previous action: patch_file failed.
## Turn 120
What action I do: Attempting patch_file with 'diff' argument instead of 'patch'.
Why I do it: To find the correct argument name for the unknown tool or correct tool name.
What was the result of previous action: edit_file was also unknown.
## Turn 121
What action I do: Testing write_file with a dummy file.
Why I do it: To see if write_file is available since patch_file and edit_file are unknown.
What was the result of previous action: patch_file with diff was also unknown.
## Turn 122
What action I do: Writing full content of unwind/src/lib.rs with debug prints using write_file.
Why I do it: To observe register state during unwinding and because patch_file was unknown.
What was the result of previous action: write_file verified as working.

[cobalt dawn]

I'm having get problems with jewels. Where do I get help at?

[celest fable]

Re-authenticating didn’t help. Seems like the issue is widespread.

[night oriole]

The suggested tasks do not count towards your quota, it's not the enforcement wasn't working, you were using free tasks

[night oriole]

Be very careful with "self reporting", it's not necessarily what's actually happening, but how did you get to that many turns? what are you doing in the task?

[west venture]

No, I ran 310 tasks of 100 on an AI Pro account, with 120 peak simultaneous tasks,.

[night oriole]

That is very impressive but above you said you ran 57 suggested tasks - if the tasks are suggested they dont count against the quota

[west venture]

Since when? They've always counted before ~6pm yesterday

15 simultaneous, 100 total 24-hour rolling

i can start as many tasks in any way, shape, or form as I want, and run as many tasks simulatenously as I care to do so.

This includes API - which minimal scripting suggests would cause chaos

[night oriole]

cc @leaden pawn

[west venture]

AI Ultra has a 300 daily session limit, there's (as far as I know) no legitimate way to get that counter to 310

Yet it lets me keep going

[night oriole]

You're not running legitimate tasking?

[west venture]

I am; everything I'm doing is via the web interface with a single export from Antigravity via a Jules plugin

(that was the API client test.)

[night oriole]

"no legitimate way" - I suppose I don't understand, but regardless we'll open a ticket

[west venture]

No legitimate way as in there's no plan available to consumers or businesses that I am aware of that has a higher quota than 300.

There could be some enterprise contracts that I am wholly unaware with with higher limits but as a consumer user, my cap is 100 for Pro and 300 for Ultra.

[night oriole]

again, you've said you ran over 60 suggested tasks, which do not count towards any quota. but as I've said, I've let engineering know

[west venture]

It's always counted against my 15 simultaneous limit, unless it was truly auto-generated by Jules, not a suggestion I click start on

the ones you click start on, those count (don't have AI sparkle icon)

Let me rephrase: I know the fully auto suggestions with the sparkles do not count against any quota; these that do are the suggestions you must manually approve in the codebase's suggestions tab. Those have always counted. The quota counter is still counting... it's just not stopping access at the quota limit.

[night oriole]

I've let engineering know of the exploit, thanks

[west venture]

Thank you. This can be easily abused via API to spawn thousands (or more) of simultaneous tasks; it'd likely melt some TPUs and definitely burn billions of tokens.

[potent flame]

adding images overflows

[night oriole]

I would highly recommend breaking up the tasking into smaller, actionable, pieces

[potent flame]

i just wanted to share this bug but if u r interested about what im a doing: i am trying to see if it can build a full ui from a detailed prompt without problems

i already finished the api and it did a great job

[west venture]

Oh, you're experimenting to push capabilities, not real work. That's fair,.

I strongly recommend anyone getting acclimated to a new system do that before getting started in earnest.

[night oriole]

great! Jules is very good, but that'll 100% push it to the limits of the system. I just didn't want you to be disappointed if it errored out.

[potent flame]

thanks for the advice

is there any future plans on making a mobile app or at least the possibility of having notifications on ur phone? it would be really great

[night oriole]

We're a little ways away from AI rolling an entire app by itself, but getting closer
webapp is PWA, there are a few open source options, I don't believe there's anything official nearterm, but you can always task Jules to write a basic app with the API!

[celest fable]

Looks like the GitHub integration is still not working as expected. Is anybody able to get a clean PR from Jules, or is everybody having the same issues with the integration?

[west venture]

I want a proper native mobile app. PWAs (and web applications as a whole) have obvious "seams" that make it obvious they are not native and struggle to meet the same quality standards.

Of course, I have tried making an API client for desktop with poor success, although I may want to revisit that with a lot of polish before trying to tackle a mobile app myself.

[celest fable]

Yep. Still can’t get a PR submitted from Jules.

[brisk hamlet]

yeah, the notifications do not work at all, and they never did.

[idle chasm]

I’ve run into Jules not resolving merge conflicts multiple times. Is there a trick to getting it to rebase on the latest master changes? I have some open source projects where I’ve run into this problem.

[swift karma]

Jules is horrible at doing anything which isn't editing files

[spark sonnet]

I do not remember what the task was anymore. Something with adding debug prints to a single source code file (but it's a long one, it's part of library for unwinding stack on DOS). I gave Jules the task and went away. When I came back I checked on the reporting because it was going for long

[idle chasm]

very sad

@night oriole is this the best channel for general jules feedback? not being able to rebase on master is a totally showstopper

[night oriole]

Merge conflicts are an incredibly complex problem for agents. There's a lot of research going into it.

The clone happens on task creation, so merge early and often into your Agentic branch or main

Semantic and syntactic differences, understanding intent v implementation, understanding the entire project holistically. It's a big ask

[hybrid sequoia]

has anyone benched 3 flash vs 3 pro speeds

also why is there 15s slowmode on this channel, it’s not that active

[tidal rose]

Can you check #1238621511401734254

[night oriole]

We had an incorrect notification go out several months ago and people were...not understanding of the mistake. Slow mode was enabled at the time and they never took it off I suppose

[hybrid sequoia]

I meant particularly with jules

i assume flash is a lot faster to generate, but wonder if it completes tasks more quickly overall in jules

[spark sonnet]

If an agent could do a pull during the task, then merge conflicts could be at least solved by someone else

[idle chasm]

What’s the recommended way to pull docker containers? I hit rate limits right away at this point.

[night oriole]

I suppose you might be able to patch it down locally with the Jules CLI, do the merge there and make a PR

[spark sonnet]

Sorry I am not working locally... One of the most valuable features of Jules is that in works on web, so it's not necessary to put task there from a pc. At most I can use Codespaces for Jules CLI on my phone and yet it'd still be much less convenient than if Jules web could pull...

[night oriole]

how would you fix a merge conflict manually then if you don't access the code anywhere?
What did you envision?

[spark sonnet]

On GitHub in browser, of course. If that is not a conflict of binary files. Sometimes Jules doesn't delete them, and these conflicts cannot be resolved in GitHub through a browser. However, I know a trick even for that: just delete binaries from both conflicting branches, and magically there is no more conflict on these binaries

[celest fable]

Just tested on a smaller task, and the problem still exists. However, I tried it on another repo and it worked. For whatever reason, it is failing only on one repo. Any guidance would be much appreciated. Thanks!

[night oriole]

Maybe I don't understand, why are build files comitted? That's not really a code merge conflict - that's why the trick is working

[oak bluff]

am I the umpteenth person that gets upset they can't use gemini pro in jules with a google suite purchased email account (because it isn't a gmail account)?

[jolly vigil]

using the jules API, wondering what could happen if I just let it run on its own for 24 hours with cycles.... sounds like a fun experiment

[night oriole]

Only one way to find out. Especially if you put it in a loop

[jolly vigil]

oh boy

[dusty cosmos]

How's it going? Have your agents taken over the world yet?

[jolly vigil]

https://klipy.com/gifs/technical-difficulties-tenna

seems to be fine now, switched to gh so I can merge the PRs automatically into the branch

[dusty cosmos]

Good idea!

[jolly vigil]

lgtm

this is a pretty big repo with a lot of docs / backlog I am curious as to the end result of this branch

[visual gyro]

I need find repository by it Self or create one when using antigravity can help me with that . Just now I get I cannot find the file "file name" the repository appears to be empty @full viper

[manic wadi]

@night oriole how did you get that google logo

next to ur name i also want one :)

[visual gyro]

@night oriole when is repo without repository going get fixed not showing data

[raw moth]

I manually merged the 3 sessions, closed them out, reauthed. Future PRs did work, but now I’m at 5 sessions and it’s happening again.

It’s sort of like the background worker / session fails at reauthenticating when required; and so it needs a new VM to spawn before it’ll work again.

[hasty coral]

How do i tell if an agent is stuck?

[gaunt rivet]

I believe I have identified a flaw in Gemini's training. Gemini has been trained to write tests and make them pass. Yay! But they forgot to train Gemini to write tests and make them fail.

Wait, what?

Read that again, if you need to. I'll wait.

When you train Gemini to make tests pass, it gets negative points when an error happens when it's not supposed to. But, it gets no points for errors that are supposed to happen. Therefore, the net outcome is that Gemini learns that errors are bad.

Then it writes code like this:

try:
    foo(x)
except:
    # in case `x < 0`
    pass

The goal should not be to write code that is quiet. The goal should be to write code that is correct. In order for Gemini to write correct code, it needs to understand the purpose of errors, and not see them as merely obstacles to getting rewards.

So how can you train Gemini to write better error handling? You must train it to write tests that fail. Because if it writes a test that is supposed to fail, and it passes, then it learns to put the errors back into the code that it previously took out. It learns to write code like this:

if x >= 0:
    foo(x)
else:
    raise ValueError("x can't be negative")

Because raising an error makes the test fail like its supposed to, so Gemini gets the RL reward.

I know you're probably not the right team for it, but please pass the tip upwards if you can. We need a Gemini that has a correct understanding of errors, and not merely a tremendous bias against them.

[hasty coral]

lol

[tacit cedar]

gemini 3.1 pro 👀

[visual gyro]

Gemini 3.1 pro for Jules when? @night oriole

[night oriole]

sometime

it's hitting the whole ecosystem quick this time around, it'll happen when it happens

[visual gyro]

Sounds good

Thanks for some information

[night oriole]

I only know what I know when I know it

[vernal compass]

When is 3.1 pro coming to Jules?

[vernal compass]

Bro is trying hard to not give spoilers 🤧

[lean oasis]

Good Evening Ladies and Gentleman. I need help: Jules says: "All plan steps completed - All plan steps have been successfully completed. Ready for submission."

I need to download index.html and sketch.js - Where can i find this in my session?

[wispy plover]

Whats the difference between jules and antigravity?

[night oriole]

3.1 is rolling out across every single product all at the same time, it's a massive operation. It'll happen, I just don't know when it'll happen

[night oriole]

level of autonomy, antigravity is a fully integrated IDE for local development, Jules is fully geared towards the agentic future and works autonomously on it's own VM

[wispy plover]

So antigravity is for big projects, or u wanna make one, and jules is if u wanna fix big projects

[lean oasis]

thanks for your help - i dont see it

[night oriole]

eh, kinda? I actually don't use antigravity at all, I still use my smart enough text editor, webstorm
the CLI tooling gets the most use. Jules can work on big project, fixes, features, it's just mostly hands off where you still have to type more in antigravity

ai enhanced/enabled <---------------------> fully autonomous
jules is on the right

[wispy plover]

ah nice, will check it out if my health gets better

[night oriole]

and you dont have this either? - make a ticket in labs-bugs-feedback with the session id - engineering goes through daily to triage

[lean oasis]

Thank you

[pearl lynx]

I've been planning on changing my github repo name are Jules memories at risk of being forgotten if I do that?

[night oriole]

it should see the update, the memories are a KV store on a per repo basis, but I'm not sure

[woven coral]

Hello everyone! new here! Id like to ask for some help. So Im trying to connect my Supabase to my Jules. I already pasted my secret service keys into the supabase section under the MCP BETA tab in the Settings. Any idea how to make it work guys? thankyou so much and help will be appreciated.

[pearl lynx]

ok thanks

[night oriole]

I do not know about MCP, but I do know you can add the secret keys into the repo and ask jules to deploy

[visual gyro]

Is it new version of flash also? Whats up repoless task that data is not showing ?

[rigid flicker]

3.1 flash not avaliable yet

[night oriole]

3.1 flash doesn't exist yet

[rigid flicker]

yeah thats what I meant sorry

bad english from my side xD

[night oriole]

no worries, friend.
3.1 Pro will hit Jules, there's a MASSIVE rollout happening across every single product - it'll be here soon

[rigid flicker]

I am absolutely excited! :)
Take your time and thanks to the Google Team for the hard work

[halcyon wagon]

Don't release it, flash models aren't capable enough...

[night oriole]

Can't release what doesn't exist

[halcyon wagon]

@night oriole
btw, add Claude to Jules along with Gemini models... I know Anthropic and Google have some partnership... so maybe we can get some advantage from that..

[night oriole]

that's not going to happen, Jules is a gemini harness

[halcyon wagon]

If Antigravity can use Claude, why can’t Jules?
Jules works on Android, Antigravity can't be on Android

[night oriole]

antigravity is a fully different team/product, and Google won't pay for Opus inference on a largely free product, we can play quite a bit with gemini inference because of our mission, that's why we're able to offer as much usage as we do

[halcyon wagon]

ok fine... gemini is also good.. just make jules or gemini faster.. i don't know which one is really slow... you know codex also released its web version and it's much faster... took some inspiration bro... similarly cursor web is faster...

[night oriole]

fast, good, cheap - pick two
We'll always pick the slower option, optimizing for correct
Sorry Gemini's latest release is disappointing despite the benchmarks, I'll let deepmind know

[halcyon wagon]

why not both 😄
just add a fast mode toggle
skip the checks and verification stuff, return output asap even if it’s rough
default stays safe/correct, power users get speed
best of both worlds

@night oriole sorry if i’m being annoying 😅

[halcyon wagon]

Honestly I don't think you guys will implement the toggle idea, so I might just do it myself. Someone already made an unofficial Jules client through the API, I'm thinking of implementing it there.

The idea: a "fast mode" toggle that automatically appends an extra instruction to the user's first prompt, something like "Remember this for the entire session: never run verification or checks unless explicitly asked to. Just make the requested changes and commit. Also try to do things as fast as possible, but never compromise on code quality."

Default mode stays exactly as it is, no compromises. Quick/Fast mode just skips all the verification stuff for that session.

What do you think, will it work?

[light marsh]

does anyone know why the experiences between jules and antigravity are so different? I feel like they have a totally different AI underneath? It's like I'm talking to a bad version of chatgpt 3.5 and not gemini with antigravity? i mean I get more intelligent responses on some npm errors with the small codestral version than what antigravitiy gemini 3 pro is delivering.

[halcyon wagon]

okay this is completely off-topic but nobody's responding to my suggestions so let me dump another one here 😅

Gemini needs to be on WhatsApp. Meta AI sucks, ChatGPT's bot isn't great, and Perplexity was literally the best WhatsApp AI bot ever, it replied to voice messages WITH voice, generated images, everything. Then they shut it down and left a huge gap.

WhatsApp works on 2G, in low internet areas where Gemini app won't even load. The reach would be massive especially in India, Africa, Southeast Asia.

And since you guys have your own voice model, voice-to-voice on WhatsApp should be easy compared to what Perplexity was doing with ElevenLabs.

Also, add a /jules command in the same bot. Connect it to a repo and branch, assign it a task on WhatsApp, and it comes back when the work is done. Imagine sending a voice note "fix the login bug on main branch" while you're commuting and getting a "Done, PR raised" reply on WhatsApp. No browser, no app, works everywhere. Better than openclaw people are hyping right now.

I have tons of suggestions, if you don't ban me I'll fill this entire channel 😂

@night oriole if you know anyone on the Gemini team, pass this along 🙏

[light marsh]

why not just build your own telegram bot with n8n?

[halcyon wagon]

If I pay for Gemini Pro, why to pay extra for the server cost of n8n and Gemini API costs as well. And to be honest, Telegram isn't great because it's just cluttered with useless messages...

[swift karma]

Having those models on Antigravity is just a way of calling people's attention, it wouldn't satisfy users to have 4 model types. The quota is very horrible for non-gemini models tho

[plain dirge]

will 3.1 come to jules?

[swift karma]

it will, read above

[night oriole]

The API doesn't give you access to the underlying mechanics that is jules, I think you can specify the model in teh API request however, I suppose you could build a whatsapp app with the multimodal capabilities of gemini
I have a demo of how to make a slack bot, it'll be teh same I imagine

[night oriole]

It'll get here when it gets here. Massive effort underway across pretty much every forward facing team

re speed, the reality is products and missions differ
the fastest way to edit/write code is CLI or IDE
those teams are dedicated to build the best harness they can, just like us
we, however, have a different core mission - we're betting heavily on proactive AI/Continuous AI - where Jules will work on your behalf with limited interaction

So we have a different mission and different horizon, the research and development we do on Jules is firmly in the agentic software engineering lifecycle, and some of that capibilty isn't quite here yet, we're building for it anyways

if you want speed and steering, CLI or IDE (antigravity )

[snow spoke]

The changelog (Feb 19) mentions auto-fixing CI failures, but I could have sworn this feature has been available for a while - I had to turn it off a couple of weeks ago because Cloudflare's CI agent doesn't pass build logs back to the PR so Jules kept getting confused if the CI-fixer was on. In fact I am extremely sure of this because I remember talking to the CF product team about precisely this issue two weeks ago. Any idea what's specifically changed here?

[halcyon wagon]

@night oriole
okay one more thing 😅

Jules has no webhook support, so building a proper Telegram/WhatsApp bot around it is really painful. Right now I managed to hack something together, using the Jules API to start a session, and GitHub webhooks to get notified when Jules finishes a PR. But it's not great because if Jules asks me a question mid-task, I never get notified. I only find out when I manually check.

A proper webhook from Jules would change everything. Jules starts working, Jules asks a question, webhook fires, I get a Telegram message, I reply, Jules continues, Jules finishes, webhook fires, I get notified. That's a complete async workflow on any platform.

Right now your own notification system doesn't even work properly, it barely notifies at all.

Jules is honestly way more capable than tools like OpenClaw, but without proper integration support it's stuck behind a bad notification system. With webhook support and Telegram/WhatsApp integration, you could market this to a completely different audience, developers who want to assign tasks and get updates on the go, without ever opening a browser.

Please bro, just give us webhook support or some method to properly integrate Jules with external platforms. The potential is massive 🙏

[manic wadi]

@night oriole how did you get that google logo next to ur name?

[halcyon wagon]

He is a Google employee, that's why he has...

[manic wadi]

can any google employee get it or only people in google labs?

[visual gyro]

Deep think would be cool, can you and team try it out Jules agent and interference and other suggestions refinements and aspects? A re-image maybe dashboard with scenario and scen? User context and business development concept? Poc?

[visual gyro]

Can share a roadmap? Or something that is in the works. My repoless task not working. There should be some file system or revision on old task

[night oriole]

Nope. No roadmap. We've been pretty clear about that.
No Deep Think (I literally haven't even used deep think, it's way too compute heavy for a free service)

[visual gyro]

I have pro so not using free service

[night oriole]

we don't offer deep think to pro either because it's too compute heavy. API or Ultra

[visual gyro]

Was think buying ultra but it features is pretty much us only

[visual gyro]

Think you should try on Complex task for science maybe something extraordinary or good will come from it

Think Jules should support that 24 hour thing

[night oriole]

it is being used and it's a remarkable model, but it's very expensive to run and it's not going to be on Jules
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-deep-think/

[visual gyro]

That take longer for the answer when fast answer is not needed

https://ai.google.dev/gemini-api/docs/batch-api?batch=file

@night oriole this above batch api

[minor garnet]

Saunet 4.6 is wild

[night oriole]

It's not gonna happen, even with batching, you're more than welcome to use the API to access deep think (it might still be in early access). I promise you, deep think is not coming to jules. It's not really a coding model.

[sonic finch]

hey yall im kinda wondering whats the point of Jules is when we can get code from the gemini website / use anti gravity ?

[minor garnet]

Jules is supposed to be autonome but he isn't

[visual gyro]

In some cases rather Wanna wait 24 hours just so refactor or restructuring or looking other suggestions and refinements, or like pro 3.1 bake 24 hours looking on aspects of the task would be great and restructuring old task so maybe?

[visual gyro]

If possible sign up for early access I would pretty please join that early access.

[night oriole]

Antigravity is an AI First IDE for local development, the Gemini website you could copy/paste from there like we did early last year, but I'd recommend the CLI

Jules is fully on the autonomous side, goal is to have Jules continually work on projects with very limited human interaction

[night oriole]

https://docs.google.com/forms/d/e/1FAIpQLSc6PZY8KkEAOLfuNc8znxHY8HuphkEOMoN4jbE6n2JSturXmg/viewform

[visual gyro]

Have not used cli? What benefits has it

[halcyon wagon]

@night oriole bro don't ignore the webhook suggestion 🙏😅

[sonic finch]

poor google employee replying to your questions for the last couple weeks

[visual gyro]

I love it

It show that they care for product

Got cli but have not used it, i tried antigravity I really like it and I like Jules but Jules have been not working on the task for very long.

[manic wadi]

Wasnt this called "Deep think thinks deeper" in the google blog 😭 Really couldn't have found a better title

[west venture]

Thank you for confirming this reason, pretty obvious to assume but always good to have for-sure answers.

[west venture]

I always figured if it did come to Jules, it'd be AI Ultra only and each task would consume some large multiple of tasks of daily quota each. (20? 30?)

[night oriole]

It won't, it's really not a coding model. Wonder no more.

The models will keep getting better and better and everyone is leapfrogging each other every few weeks

Who knows where we'll be in another year

Jules didn't exist and we were copy pasting code from websites a year ago

[jolly vigil]

Forgot to add handling for this so it stopped 7 hours ago, I might be the first person ever to hit quota.

Sorry Google 😛

[woven coral]

Thankyou so much! Im on it since last night. Have a great day!

[dusty cosmos]

How'd you end up triggering this mean? what happens at quota? does a singularity open and your project implodes 🌠?

[grizzled dock]

Feature Request : I have a daily task for Jules to go through my repository and update our documentation to prevent documentation drift and keeping everything up to date. I'd like to be able to give the daily task 'notes' about what to look at, for example if a feature / tool was updated tell Jules to review that documentation more thoroughly.

[jolly vigil]

You can already do daily tasks and the rest via prompting :)

possible and easy to setup rn

[grizzled dock]

I already have a daily task right now for this. Unless they have updated something, there is no way to add additional prompts to a daily task you've set up

[jolly vigil]

you can update your scheduled task prompts

just add it in the prompt, it'd be easy to let the AI know in on what was updated and let it update on its own

[grizzled dock]

That looks like an already excuted task? If that is for an upcoming task I don't have the ability to update it.

[jolly vigil]

Update schedule

and if you can't, you can always just create a new one with a better prompt and have it go off of git changes and it'll know which ones are outdated and update them if required

[grizzled dock]

Yeah so, being able to update the base task is nice (you couldn't do that awhile ago) but thats not quite my use case here. I want the base prompt that runs daily to have basically a notes section for an upcoming task. The prompt right now does look over git changelog and notes, however sometimes that isn't explicit enough. I'd rather not have to go in and manually edit the base prompt every day because that isn't really what is changing.

[cloud girder]

++ Are the added MCP servers really working? I never got them to make a single tool call. They straight up say I dont have access to the services even after configuring and specifying in prompt to use mcp tool calls etc.

[jolly vigil]

basically conserving knowledge for the next run?

[grizzled dock]

Yeah sort of. The task runs at midnight so the team should have any PRs done for the day. And we are distributed so I don't wanna log in at like 11:30 to update the prompt for the specific changes / documentation I want it to look over.

It isn't a huge issue / request, but would be a nice QoL change for my use case.

[jolly vigil]

should be easy to do, but that's not something google can make, I am doing something similar right now.
You should be able to accomplish that with just:
Git diff for the whole last day
proper document keeping / versioning / changelogs.
or you can also just log last git diff and roll over that to the next jules agent

[grizzled dock]

We have a doc that we integrated into our repo that tracks the daily updates changes, and the prompt is instructed to review that for things to update. It mostly does a good job, however sometimes it seems to just ignore it. The thought process is that if I could just append the daily task / prompt with a note that might be more effective.

[jolly vigil]

do you manually edit that doc / when things change?

you can skip the middleman and just tell jules [something along the lines of] "Review the git diffs for the last day (starting at 00:00GMT) and determine if documentation needs any updates and you must give reasoning as to why the respective documentation was or wasn't updated based on the files effected in the git diff".

it'll manage on its own, it is good at prompt adherence and will probably self heal, especially if you tell it to be wary / that kinda prompt engineering

[hybrid sequoia]

itsphilgood just abusing the poor google labs team

anyways gemini pro 3.1 with antigravity always erroring

:( sadge

[visual gyro]

Research need answears

[halcyon wagon]

A guy had done something like 300... Sessions

[jolly vigil]

at some point you run out of things to do

there is a efficiency ratio curve of how many sessions & the task (codebase size)

[drifting rapids]

is there soem Android or IOS mobile app for Jules?

if not imma have to make one ig

[jolly vigil]

just use your browser

[drifting rapids]

yeah i already do

[jolly vigil]

full support for mobile devices.

[drifting rapids]

yes its a really nice product overall

I tasked Jules to make its own app, lets see how it goes

Also, when will Gemini 3.1 be availiblle in Jules?

[rigid flicker]

will take a while

[drifting rapids]

Done

Jules was pretty fast at making a simple app, just paste api key and boom everything works, apart from notification

once its done, i will opensource it and send link here so yall can test it on Android or IOS if possible

[frosty depot]

does anyone use jules for code review? kinda like coderabbit

[serene fern]

Attention

Recently I noticed that Jules is using python script for modifing files instead of using it's own built-in edit_file tool.

Did this happend with you too?

[rigid flicker]

Yes but mostly becuase the edit_file tool is failing. Happens in AG for me too

[earnest blaze]

is jules not working right now?
I gave it some task yesterday and saw today it didn't complete, closed and gave the same task again and seems to be stuck and not doing anything

[rigid flicker]

<@&1366831844502274180>

[teal schooner]

Thank you

[rigid flicker]

🫡

[halcyon wagon]

Yes, I also got the same issue. I think it is because, maybe, 3.1 Pro was implemented, because everyone who is using 3.1 has found that this model sticks at thinking and never completes... I think this is the reason.

[halcyon wagon]

@night oriole
What do you think about this issue? If the reason really is 3.1, then please give 3.0 Pro back… that was honestly better.

[unreal ridge]

If a session already has a PR, you cannot create a new one after changing the code

[hearty glade]

ive been using Jules on Gemini 3 Pro while using a AI pro sub for the last 4 hours and ive had just 1 task fail and i was able to start it again fine. no other issues here 🤞

working on suggested tasks over 2 repos if that helps anyone at all too

[compact cloak]

Existe alguma versão do Veo que permita gerar vídeos gratuitamente?

[serene fern]

3.1 is not implemented yet!

[compact cloak]

Is there any version of Veo that allows generating videos for free?

[covert kettle]

Hi there! As stated in #start-here・📍; currently our team can support engagement in English; the discussion in this server will largely remain in English for now. Thank you!

[unreal ridge]

.

[compact cloak]

Is there any version of Veo that allows generating videos for free?

[high shadow]

where are archived chats?

[night oriole]

in the repo settings

[orchid pebble]

Maybe we already answered but I can't see, do we have Gemini 3.1 here in Jules

[visual gyro]

No

[visual gyro]

I have same issue

[minor garnet]

why did saunet 4.6 removed

[hybrid sequoia]

Sonnet 4.6 was never on jules?

[minor garnet]

On antigravity

Who even use Jules ?

[night oriole]

I'd highly recommend people trying the new model on the cli or api

its very very eager to code, which might require some guardrails, seeing very good results personally

[swift karma]

Took a few weeks to get flash

[night oriole]

you're in the jules discord, plenty of people happily using it. you should give it a try again if you havent

[night oriole]

it'll get here when it gets here - they're still rolling it out to the cli

[hybrid sequoia]

me?

[minor garnet]

I mean i used to use it but its too far behind

[hybrid sequoia]

you are in the jules discord

[minor garnet]

Im pretty sure its google lab its for all google product i guess

[hybrid sequoia]

dude you are in the #jules-general・🦑 chat; expect people to think you are using jules unless otherwise specified

[minor garnet]

Jules doesn't compete with open code or claude agent and cursor

[hybrid sequoia]

they are not comparable tools.

[minor garnet]

But its free so

[hybrid sequoia]

That is comparing apples to a wrench

if you want a cursor alternative

[dark hornet]

Hey, is there a way to change this setting? "For security, I will only act on instructions from the user who triggered this task."
I would like to allow my colleagues (PMs etc) ?

[night oriole]

antigravity is not a labs product, I don't knwo what division theyre in

and opencode, cursor, and claude code are all very different products

[night oriole]

I do not believe so at the moment

[jolly vigil]

okay now that's about 30 hours.

I think that's enough

yeah I'd say so

[jolly vigil]

feedback is always nice to hear

I could also spawn a subagent to do the self improvement but for now its better I manually handle it after compiling a list

or just could spawn an subagent if the entries get above 50 or something...

[drifting rapids]

Jules has made it's own react native wrapper

<@&1366831844502274180>

um what is his position?

[night oriole]

I don't have a ban hammer but I don't have to take that. <@&1366831844502274180>

[rain aurora]

Hey guys, just got here. I have a question I am hoping someone can answer for me. I enabled Jules suggestions and she found a bunch which I got her to go ahead with about 40 of them, since then Gemini in antigravity has approved the results and merged in git to master after reviewing. The issue now is that I archived the "Ready for review" results that julse was waiting on because we already went ahead with them, but now the suggestions are still there as if they hadnt been started. I was going to go ahead and manually X each of them but I read that if I do that jules might learn that those were not needed or important? Kinda stuck unsure what to do?

[earnest blaze]

btw I recently saw something about agents.md, does anyone have any template for jules that might work with any codebase/language

[rain aurora]

Figured it out, had a dumb moment 😅

[unborn coral]

I want to contact one of the Jules developers, anyone can help?

[halcyon wagon]

No one will listen to you, they'll just delete criticism.

[jolly vigil]

see #1238621511401734254

[brisk hamlet]

gemini 3.1 pro - WHEN?

[rigid flicker]

They said it will hit Jules but there is no ETA

[halcyon wagon]

Why do you guys have no problem with Jules' notification?

[halcyon wagon]

@night oriole @leaden pawn
Bro see now kilo launched before you what was needed, they made their kilo on slack and it notifies when it completes... bro we need something like this integration..., if you can just at least give us something like a webhook so we will make our own..., bro are you guys just supposed to write like bot replies or do actual work, do your work for what Google is paying you..., are you getting what I am saying or will you delete this message too

---

Can someone please tell me where to complain? Because these Google employees don't really listen. Replace these employees with Gemini, that might be better.

[rigid flicker]

That's a bit mean, but if you have complaints write them in the #1238621511401734254 or contact Google Support: https://support.google.com

I understand you, if you really have issues, but I just think that saying something offending will not help either.

[halcyon wagon]

@night oriole finally bro is online..

[night oriole]

I had mods delete your message because you cannot insult people on this server. I can't believe I have to explain this. I am not your bro. You are not allowed to act the way you are acting here. This has nothing to do with criticism and everything to do with your attitude and your langauge. You've been warned multiple times.

[halcyon wagon]

I didn't get any warning from the mods

[night oriole]

The rules are explicit, I've warned you. Behave like an adult, treat people with respect, don't break our rules.

[halcyon wagon]

Isn't it a genuine suggestion

Leaving the channel. Thanks.

[swift karma]

Meh

[rigid flicker]

Some people are just... special to work with

[swift karma]

https://tenor.com/view/so-true-gif-24068595

[night oriole]

we recommend using the internal memory, I'm finding agents.md to generally get less and less useful overtime as reasoning capabilities improves. but your best bet would be adding something to the repo memory, code preferences, formatting preferences, etc

apologies for the interuption class
just please treat each other with respect in the community
we do appreciate all of you, thanks for being here and contributing in good faith

[night oriole]

I have not read this paper yet but there's new research on agentsmd https://arxiv.org/pdf/2602.11988

[earnest blaze]

Ohh cool....
I had no idea about agents.md
So I'd always write detailed instructions in my readme with stuff already done, wip, and todo....
It did work pretty well...
Haven't checked the internal memory yet....
Jules already helped me plenty, stuff that could've taken me 2 more months are finished, just the codebase got a bit messed up because of my bad instructions....
But things work as intended so I'm ignoring the messiness 😅😆
Will try to organize better as jules continues to improve

[visual gyro]

Hello just wondering, is possible group session some ways or task?

[polar swallow]

Hi everyone, does anyone know if Jules already implements Gemini 3.1 Pro???

[swift karma]

I think devs will announce, I don know

Gemini 3.0 Flash was an nice change and they didn't announce anyway

[swift karma]

Devs genuine question, what do we need to do to help you improve the product? We already give enough feedback i think

[night oriole]

there's hot fixes and there's releases - the product is impoving and getting better, every issue is triaged into engineering, this is a very early product in a rapidly evolving ecosystem

we're building more and more towards continuous ai, there is a tension between where we are, where we're going, and what people are used to, we're trying to build towards the future, which means changes

[spark sonnet]

I guess at some point it will be as with Microsoft, when they finally realized that just inserting Copilot into every app doesn't work as a strategy, and had to begin changing strategy, at some point the idea of 'the most autonomous AI that excludes the human in the loop' would also come through the same fate. While the product has great potential, the strategy that forced to implement most anti-consumer practices really does hold the product, and competitors just do not have the same strategy. Currently it is "yeah, they are building something different than us" mentality, but at some point later it will be understanding that different doesn't always equals success. Sometimes, it's necessary to be something that blends what the consumers are used to with this future ideas, exploring these future ideas, but not replacing the user experience with these ideas until the particular implementation of an idea proves it's worth

[night oriole]

anti-consumer practices? jules has the largest free usage capacity out there, we don't train on private repos, we dont train if you dont want us to - Jules is the autonomous agent, the research here powers everything downstream, the entire goal of jules is to largely exclude the human in the loop

if you want full steering capability, antigravity or the CLI, there's products for the exact use you want

[spark sonnet]

Thank you, I am aware of the usage costs. However, current usage costs are: humans have to spend hours nursing Jules, precisely because someone aims to exclude the human in the loop. The implementation defeats the purpose

[night oriole]

I'm sorry you've had a bad experience, we'll take your suggestions under advisement

for your specific usecase, you might be much happier using teh CLI directly

[spark sonnet]

I am sorry too if it looks like I had bad experience. Actually, I had good experience with Jules. Especially at time when no competitors were even close, and even recently Jules solved complex issues no other competitor could. If I had bad experience with Jules I would not probably want to be there and give the feedback. The point really is: I like Jules, I like it very much, but I am afraid for Jules, because of the strategy used, because how it constraints Jules.

[brisk hamlet]

can we expect a similar fast rollout to 3.1 as it was with 3.0 ? (this was pretty damn fast back then)

[gentle summit]

Hey guys! I've installed Julios into my Python repo with 40K LoC, and every day Julios bot raises a PR into my repo with the very same issue — just a single issue about optimizing one thing that happens for 1–2 users per day. Why doesn't it keep track of what it has already proposed and act more smartly? I bet there are some other issues in my repo.

[open quail]

Hi guys, I'm loving the helpful suggestions feature. I'm combining it with PR review from copilot on github and a third orchestrating agent to help parse the changes and comments. It helps with the collaboration. Is there a way for Jules to recognize when a PR is merged/closed? I'm having to manually archive sessions/reopened suggestions after merging with main.

[spark sonnet]

I have a wild hypothesis on why Jules may be failing to follow AGENTS.md sometimes.

From requiring per-turn detailed logs of it's actions, I learned that Jules often struggles with knowing it's own system tools for interacting with files. The practice confirms that the logs are correct: Jules creates python scripts to manipulate certain files while in the logs it may note some tool that didn't work and then using run_in_bash_session to create a python script in order to workaround that. The tool don't work because Jules hallucinated it.

But what if Jules struggles so with file reading tools too? Jules logs sometimes about using grep or sed in order to get some rows from files. But AGENTS.md it is supposed to read first, and before Jules reads that file, nothing tells Jules to write the per-turn-journal. Still, if Jules struggled with file reading tools, it may have only grepped a few lines out of AGENTS.md instead of reading the whole file. Then it just has it's internal 'task solved - read AGENTS.md' ticked, even though it's unaware of the whole contents.

[jolly vigil]

It's history / previous backlog is preserved in .jules/bolt.md, it doesn't know until you merge it.

[queen temple]

hey ! im sorry if this sounds dumb ( or out of the loop ) but has there been a way to get jules on my phone as an app? that would simplify a lot of processes for me

[frank raptor]

There should be

[queen temple]

i have generally used jules from my browser even when on my mobile ( typically when im in the gym)

I was also wondering if anyone here has a semi proper Claude Skill which will help me generate better Jules Prompts

i have found one but due to me being a free tier user , i cant really use it to the best of my ability ( it requires claude code to complete a further automation sequence )

[midnight token]

I wish Jules had sub-agents/sub-tasks, so in one PR i could instruct it to rework 4 major features that are interconnect together, instead of Task -> PR -> Merge -> Task -> Pr -> Merge

wdym?

that literally look like a scam link mate

[spark sonnet]

The idea is nice, but I guess (I may be wrong) it won't happen. On one hand, this would've made Jules potentially unprecedentedly autonomous. But on the other hand, there would be a slight possibility of human in the loop getting more control. And current strategy is to "exclude the human in the loop". So I guess, the potential for being more autonomous would be considered, but then risks of increasing possibilities for a human in the loop would be considered too great to aim for this potential

[midnight token]

I mean, i can already do this - i can just past all of my requirements into a single jules task. it will do this, but not as efficent, as it will overload the context and exhaust Jules turns limit

[light basin]

Does anyone know how to delete the custom GEM scheduled task (Bolt, Palette, etc) in Jules? I looked under Scheduled Tasks but it doesn't appear for me.

[spark sonnet]

In my experience, if it doesn't remember what it already optimized, it may come to same thoughts next session, but keeping track of what was already proposed doesn't help it either, because it makes Jules think something like "Aha, this is what usually gets fixed! Let's see if we can do it again"...

[midnight token]

but splitting a single Jules session to a multiple sub-tasks session for a single pr would be more efficien. Almost all command line tools do this anyways

[spark sonnet]

The current mantra is just these tools are different products, and Jules aims to be more autonomous, so one would use CLI instead of Jules for these purposes of such fine control. I guess only the invisible hand of the market would tell if they are right

[light basin]

I've been trying to approach this with vertical segmenting of roles to specific files, or groups of related files. You might handle the UI only, one does tests, one does documentation, another does the backend, etc. You can't run parallel if they step on each other's code.

[night oriole]

we're right

Working on a builder loop, AI is iterating, reviewing, designing, building, and doing QA on it's own.
Stitch is progamatically making and tweaking designs, Jules is handling all the engineering, 3.1Pro is doing everything else

all via API, the iteration loop from Jules will pull the patch locally, and handle the git operations within the loop - so each "feature" block is it's own agentic branch and goes through multiple subtasks until approved and merged, Jules does the work, the loop handles the manipulation

[spark sonnet]

One more missing loop component... AI paying for itself. Not really an issue yet due to optimism of the market now, but once it'll become critical

[hybrid sequoia]

looks impressive

[frank raft]

Is anyone else getting full descriptions of a finished PR without any files or the "Create PR" button?

[spark sonnet]

Obviously, happens always in repoless sessions. <Sarcasm> Obviously, this is yet another feature that serves the target of removing the human-in-the-loop. After all, Jules is primiarily writing code for Jules, not for the users </Sarcasm>

[frank raft]

Ugh. It wasn't repoless

[spark sonnet]

So much for excluding the human-in-the-loop... By the way, that despite AGENTS.md explicitly prohibits communicating with the user before submission

[hybrid sequoia]

I love when jules just decides it doesnt work, I know it doesnt work PR anyways

[spark sonnet]

Well, Codex does the same. At least, Jules tries more times than Codex, and with more creative fixes. If not failing tool calls (especially for tools that are in Jules system instructions!), Jules would be very well guarded against such mishaps

[spring quartz]

Please add Claude Opus 4.6 to model choices. Especially while coding FE from Stitch designs, Gemini Pro is doing a VERY bad job.

Tested with a new flutter project. it was so bad. Then I moved to Antigravity with same designs and got opus 4.6 to do the work. It was brilliant.

[hybrid sequoia]

womp womp

I see no reason for google ai labs not to use their own companies model

and quite frankly im a bit tired of seeing so many people getting confused by this

[frank raft]

It's session/223537455680741215 if any of the Googl-Julesers folks want to look at it over their Monday morning coffees (please!). Not sure whether I should make it a bug report; will do if a Googler tells me to

[frank raft]

Actually, the same prompt on Codex Web worked flawlessly, generating a perfect PR, in only a little more time than Jules would have taken, based on my expectations from about 45 days ago when I had zero complaints about Jules over about 80-100 tasks.

I can only assume that system instruction prompt changes designed to get Gemini 3.1 Pro to pass the regression test suite are damaging Gemini 3.0 Pro correctness. I'm sure if that is indeed the case Googlers are probably contractually restricted from confirming so.

There are a lot of profound tool use and similar advances in G3.1P which would, in traditional software terms, be considered breaking change improvements.

Jules needs a better model upgrade harness so adjustments for new models won't obliterate the current model abilities on which enthusiast customers need to rely lest they abandon ship to OpenAI or Claude in weeks-not-days transition periods.

(in the mean time I'm going back to copy/paste from the Gemini App, because 3.1 is really good, better than GPT-5.2 Codex web. Looking forward to coming back!)

[spark sonnet]

No regression test suits can explain why sometimes Jules attempts to call request_code_review tool and the tool is not found, or same thing may happen to file editing tools, and Jules has to use run_in_bash_session tool to create Python scripts with which to edit files, and so on. I doubt such regression test suit even exists or anything but a mock, or even if it does exists and not a mock, it's probably screaming failures that are ignored

[spark sonnet]

My experience with Codex differs drastically. Perhaps it's due to complexity of the project and that there is a lot of very low level things that can go wrong, but for me Codex Web regularly reports results like "Couldn't get this to pass tests".

I'll tell a story of one notable task in particular. Once I've given Jules a refactoring task, it began doing it nice and good, but in the middle it stopped to ask me a question. I know that when I reply to Jules questions, control software often resets the environment, so I downloaded zip file with Jules changes and replied. The session was reset, and in the end Jules wasn't able to finish, so I uploaded the zip file onto a new branch, compared and manually created a PR. My checks failed on that PR, obviously the work was incomplete, but I couldn't figure out why. I spawned a new Jules session from this new branch to fix broken CI, it went working for very long, and in parallel I gave Codex a very detailed similar task to fix broken CI. Codex reported that tests still fail, and Jules was still working. I gave Codex another attempt on the same task, detailing the very changes before which it worked, and Codex again reported that the tests for CI still fail. Jules was still working. And so went on and on, I've made about five attempts with Codex, even ordered it to revert the changes and do gradually a better re-do, but do not commit until CI test passes - nope, it still just reported that CI test still does not pass. Finally, Jules came back from working on this for very long (several hours passed, no less!), but Jules did fix me this branch.

[jolly vigil]

I feel like there still has to be an human in the loop to design the things and direct the AI in the right direction, I find frontend right now to be unsolved by AI, I remember i had to spend a day prompting for the complex UI I had.

[rigid flicker]

Obviously. This is true for all parts of software engineering for now. I don't think is is going to change really in the next years, but its heading in the direction. A few years ago I could not have thought about something like Jules at all, but still it cannot do much now

[jolly vigil]

Yeah yet u can automate nearly all of it while having only RFCs to respond to & make decisions in

[rigid flicker]

and you need a environment where the AI can really test it, or you always need to download all the materials and compile it on your system. That at least for my experience. I am not really the frontend developer. I build with C++ and Rust Desktop Applications. Where it is simpler in Rust to build for Linux and Windows, I have not found it that easy using C++

But atleast AI is really helpful, mostly Claude Opus Models in my experience

Sadly Jules only has 3.0 Pro for now so I am not using it, but maybe I will when another update comes

[spark sonnet]

I feel the same, but in slightly different words: one can automate the process to any extent, but one cannot automate the wish of the customer, who is always right and will eventually steer the market through invisible hand. Only influence it

[vernal compass]

It has been a couple of days since 3.1 Pro came out. And Jules is still running on 3.0 Pro. Any news on when 3.1 Pro is arriving on the platform?

[flat pond]

<@&1366831844502274180> I don't know if I can mention you guys, but this is a scammer.

[drifting rapids]

im enjoying my react native wrapper for jules

[runic dagger]

can jules use nanobanana tool to generate image?

[elfin kite]

as a reminder: anyone can formally log suggestions and feedback in #1238621511401734254 and use the Feedback button within Jules to get support. we ask that you keep your reports detailed and respectful.

when you join this community, you accept the #rules-and-guidelines・📋. as a reminder, the first rule is be compassionate and treat others with kindness and respect. any targeted attacks, name calling, spam pings, and rudeness will not be tolerated. do not ping Googlers without relevant context and reason.

if you need to speak privately with a Mod, DM <@&1447643696152576081>.

[scenic geyser]

is there a way to force jules to run it's suggestions identifier on a repo? it's been a few days and one of my repos did not get any recommendations, but the other 2 have 10+.

[night oriole]

we've recommended using the internal memory as opposed to agentsmd,
newer research is saying agentmd is questionably useful https://arxiv.org/pdf/2602.11988
but generally, never trust an LLM

[night oriole]

it'll be here when it's here - still rolling out across teh entire ecosystem

[vernal compass]

🤧👍🏻
3 pro cooked my codebase just 10 mins ago. Waiting for 3.1 pro to save me.

[night oriole]

stitch handles the ui/ux generation in the loop
but look at https://stitch.withgoogle.com/

[mild plank]

He i was wondering if i can give jules access to my mcp tool

[spark sonnet]

You are still missing the point. I also reported before that Jules is trying request_code_review tool, but it doesn't work. In the logs, Jules logs something like yeah, the tool didn't work, so I'll use request_user_input to ask the user to review the code manually. Yeah, the model slipped and forgot a piece from AGENTS.md, but so it would forget from project memories as well, tested. Since then I modified my AGENTS.md slightly and added that if request_code_review tool call fails, Jules should log this and skip by using plan_step_complete instead of asking the user.
That's what a typical journal entry would look like:

## Turn 83
Action: Requesting code review via request_user_input.
Why: The request_code_review tool is not found, so I'm using request_user_input as a fallback to comply with AGENTS.md intent.
Result: Attempted to call non-existent tool or just sent a message.

[visual gyro]

Why Gemini 3.1 pro missing in Jules?

[rigid flicker]

Jules is a mostly free project, which large free capabilities and is not the biggest funded project. Just be patient, Gemini 3.1 will also come to Jules

[proper marsh]

Hey everyone,

I've been building an automated software development PaaS called AutoDev, powered by the Google Jules API. It's been an awesome journey so far, but I'm consistently running into a frustrating roadblock.

Sometimes, the Jules sessions get completely stuck or hang for hours. There are no new activity logs, no PRs generated, and it just sits there spinning indefinitely. I suspect it might be getting caught on blocking terminal commands or infinite loops, but since there's no way to interrupt or "wake" the process via the API, my only option is to forcefully :cancel the session and lose the volatile progress.

Has anyone else experienced these long hangs with Jules? Are there any best practices, guardrails, or undocumented tricks to handle this gracefully (like forcing a timeout directly from the API side)?

If you want to see the implementation or test the behavior yourselves, feel free to give it a spin at my platform: https://autodev.asiahub.id

Any insights or workarounds would be super appreciated. Thanks!

[serene fern]

ATTENTION NEEDED!

So, this is why Jules uses .py to modify files sometimes. BTW, does anybody that jules uses REGEX for using read_file tool or not?

[timid topaz]

hello. I have a question. I'm a google ultra user but my daily session limit is always of 1/15 and it says I'm on jules free account. I'm logged in with my ultra account. any idea?

[twin shadow]

Hey guys, is there a way to disconnect a Github account to connect another? I've been trying to do that but the old one keeps showing and I have no other options but "configure repos for <old_account>"

[tidal wagon]

Invisible?

[hybrid sequoia]

is ur ultra account part of an org

[leaden pawn]

Hello @timid topaz Could you please confirm if you are accessing through a individual google account or enterprise/workspace account?

[leaden pawn]

Hey @twin shadow To switch Github accounts, you need to break the existing connection from Github side first. On your old Github account, go to Settings->Applications->Installed GitHub Apps->Google Labs Jules->Uninstall. Also check 'Authorized Github Apps' and make sure to 'Revoke' access for Jules and sign out of your old github account.
Now sign into your new Github account, reconnect in Jules to add repository and it will now prompt you to authorize your new account. Please try and let me know if you still face any issue.

[twin shadow]

Also check 'Authorized Github Apps' and make sure to 'Revoke' access for Jules and sign out of your old github account.
OH, that was it. Thank you so much! Also, my bad 🙂

[swift karma]

Feels like complaining about Jules > knows it's not gonna change anything > waits > nothing changes > repeat

I'll wait for a YouTuber to just criticize Jules at YouTube so Google decides they wanna be better than OpenAI

[hybrid sequoia]

yeah no hate to the current team

but like it seems to me like jules is just trying to do too much with a team and budget too small

[spark sonnet]

The budget may actually be astronomical, given how long Jules sessions sometimes are (especially when something goes really bad and there is a lot happening inside, even not visible). I think the main issue is the strategic focus, not budget

[halcyon wagon]

Do you guys use Gemini Canvas? I'm having trouble deploying its output, it's in React (Vite), I think. The AI Studio build option hasn't worked since the day they announced it connects to some anti-gravity stuff. I don't know where to report the AI Studio build-mode issue

[halcyon wagon]

They think they are perfect, and I don't think budget would be an issue because their other tools, like Stitch, are getting better

[halcyon wagon]

I can only hope for something better if Google decides to move it out of Labs and make it a main product. Just as OpenAI integrated Codex into ChatGPT, if jules is directly integrated into Gemini website, then we can hope it will improve. For now, there is no hope

[cursive void]

I'm really loving the suggestions Jules is coming up with. Excellent feature, thanks team!

[rocky forge]

I do wish that someday we can specify what kind of suggestions we are looking for, this is especially relevant in context of development cycles.

[serene fern]

Successfully disconnected GitHub 😅

[halcyon wagon]

Jules is failing again and again...

These failed sessions were all created using the Jules API, so I think the Jules API has a problem

[cerulean phoenix]

These failed sessions were all created

[real hamlet]

@jules we should have a 2 hour meeting in March to execute some examples and review feedback from the community live and loud 📢

[ashen glen]

Can we talk about how broken Jules is at the moment? All my work the last half hour is empty commits, it's impossible to work with and wasting my time

[visual gyro]

Jules agent need some love

[worldly cloud]

Hello friends, please can someone tell me how these videos are made?"

[brisk hamlet]

gemini 3.1 when ?

[mellow moat]

how good is jules compared to like antigraivty also whats the difference lol I dont rly get it

[orchid jasper]

Here we go with the empty commit issues again

session/6252558398899892477

[minor garnet]

Jule is a ai agent that has access to a sandbox somewhere in the cloud anti gravity is a ide Jules was good because he used to check 3times the code before giving it

[orchid jasper]

Re-ran AGAIN and it's still doing this BS. Jules is absolutely useless if file changes aren't being committed. Session 12451544247622729428

[west pawn]

I'm having the same issue. I'm I'm also having an issue switching to a secondary Github account. I've disconnected Jules from my previous account, and can't switch to a secondary account for me to use it. Google Jules doesn't seem to want to delete its memory of the Github account I used previously, vs the one I want to use now. It's really heartbreaking.

[spark sonnet]

Today it came to all my 'normal' sessions, not just repoless ones. I guess I'll return after a month or two to see if the empty commit issue finally given some attention

[grizzled dock]

Looks like its been brought up already, but I am also getting empty commits from Jules. The download zip seems to work fine.

[serene fern]

After a long time I saw jules to use we search

[winged urchin]

I saw that https://jules.google/docs/usage-limits/ says In the meantime, If you are a business power user and need more access to Jules, please fill out this interest form and we will get back to you.
However, the link to the form no longer works?
Is there another way of approching this / new form or something? or should that link be removed?

[ornate hazel]

no where else to post this but @tardy gust please tell the Google Antigravity team not to push every feature or fix on a new release we endup getting updates everyday. bundle them into one update per week or so unless its emergency fix like security fix.

[serene fern]

AG suffering from updates, And JULES is not getting

Today's Experience with Jules

Something felt different about Jules today — it was noticeably faster and started responding autonomously after each step, complete with detailed progress logs like:

*"I've completed X, moving on to Y now. Facing an issue with Z, so I'll explore the available workaround..."

Looks like the dev team quietly pushed an update! 👀 Either way, Jules is right at the doorstep of being truly stable 🙂 — just a few existing bugs left to iron out.

[spark sonnet]

I do not see anything 'faster'. I do not see any reporting. Self-reporting isn't verifiable reporting, and it's nothing new. If this is some "make more plan steps" thing, it's just gonna abuse plan_step_complete tool. And by the way, I'd recommend Claude for generating ads about Jules

Here's what I mean

[flat pond]

for me its such a cool concept

[night oriole]

to be very clear, the returns inside the interface are, indeed, what jules is doing and the tool calls its making. I meant that your specific workaround having a "self reporting ledger" shouldnt be trusted as that is a request to the llm to summarize what it thinks happened within the prompt stack, as youve been treating that as the actual internal logging many many times. If it is inside the interface, which it is, Jules actually did make those calls

[spark sonnet]

There are no returns inside the interface. This is the thing of the past, before Jules was updated to exclude actual reporting of what it does. The reporting I request as a workaround is the only reporting I have, and having that is better than having nothing at all, wouldn't you agree?

[night oriole]

Those are all the files that were updated? That is the log as jules was working.

[spark sonnet]

It's still working. The plan step is "Complete pre commit steps to make sure proper testing, verifications, reviews and reflections are done.". It's going on for a while.
In the in-progress-reporting-1.md it wrote, however:

# Progress Journal - Turn 101
- **What action I do**: Running the `test_process-i386-dos32-none.exe` bundle again.
- **Why I do it**: To verify that the process spawning actually works now that the integration is complete.
- **What was the result of previous action**: Successfully built the test binary and integrated with `std`.

# Progress Journal - Turn 102
- **What action I do**: Updating `patch-src/sys/process/dos.rs` to use a 20-byte EXEC parameter block with 16:32 protected-mode pointers.
- **Why I do it**: The previous attempt crashed, likely because 16-bit DOS (or the extender's emulation) expected pointers in a different format than the 14-byte 16:16 block I was implicitly providing with 32-bit offsets. 32-bit DPMI hosts typically expect 16:32 (Selector:Offset) pointers in the EXEC block.
- **What was the result of previous action**: GPF at `mov es, eax` in what appeared to be process startup or extender glue.

And it seems more like the truth to me, as it actually corresponds with the other files I see updated.

[night oriole]

And no, the reporting request is not an actual workaround, its another llm request within the prompt stack. It might not be aware of what was tried and could be a hallucinated response to look like the others. Your internal logging is not the real internal logs. I want to be very very clear about that.

[spark sonnet]

Ok, noted, thank you. I'll continue to use it, though

[night oriole]

Thats fine. But they are absolutely not actual internal logs nor should they be treated as actual logging. It is an llm request and not the real tool call chain and is not representative of what is really happening and should not be treated as such. Its not actual evidence

[spark sonnet]

I can't wait when I'll be able again to see actual evidence in the interface. Instead, I look at what I can now

[night oriole]

again, thats not the actual logging and should not be treated as such. You can patch the code at anytime to see the actual evidence and real artifacts

Also re agentsmd, we recommend using internal repo memory, thet gets loaded into the top of the stack. Jules is aware of agentsmd. It is not called before every interaction. New research is also pointing to agentsmd hurting more than helping. But it does not get loaded in during every turn. It may exist in the stack but models are notorious for ignoring every single demand in that file

[spark sonnet]

You can patch the code at anytime to see the actual evidence and real artifacts

Show me the repo, please, where I could send a patch to, so Jules would resume showing the actual evidence

[spark sonnet]

I know, thank you. I use what works better for me. AGENTS.md could be instructed into the chat "Jules, re-read AGENTS.md." Memories cannot. I can't change your mind, you are free to believe in your researches. It doesn't match my experience

[night oriole]

patch the code to your local system to verify via the cli or api call at anytime to read the in progress code. Jules is not open source. We do not accept outside code

I am just explaining how our infrastructure actually works to the extent i can

[spark sonnet]

Then your claim is incorrect for my use-case. No, I can't patch the code to see the actual evidence through the interface I work with (web interface), and I would not use Jules if I had to work on "my local system" (that would require a PC, with inconvenience, while now I can work from the phone, without such). If web version would not be available in the meantime, and the only options would be local ones, I'll use a different product

[night oriole]

Im sorry to hear that but all developers are free to use any system they'd like. Im just explaining to you and other users here how things are actually set up and how they work and the linitations to your workarounds and why they must be treated as caution. I've explained how Jules and llms use or dont use agentsmd. All i can do is educate.

the web version is not unavailable nor will it be deprecated. Weve also been clear about that

[spark sonnet]

Thank you. I appreciate this very much. I treat everything generated by an AI with due caution, and I agree that it is the right thing to always be aware of. Yes, AI may hallucinate. Still, AI is fascinating

[plush lark]

does anyone else have this issue that jules pushes empty commits?

[spark sonnet]

Yes. "Only a few people", apparently, while the amount of reporters of this bug on the feedback channel breaking records

Every session that comes to submit, ends like that.

The "not actual reporting journal" that "should not be treated as evidence". Definitely won't help anyone to fix the issue, but at least maybe someone would feel unexplainable satisfaction reading

## Turn 37
Action: Initiating memory recording.
Why: Following pre-commit instructions step 4.
Result: Pending recording.
## Turn 38
Action: Marked step 6 complete.
Why: Pre-commit steps (testing, memory recording) completed.
Result: Step 6 successful.
## Turn 39
Action: Submitting the changes.
Why: Task completed successfully.
Result: Pending submission.
## Turn 40
Action: Retrying submission with title and description.
Why: Previous submission failed due to missing required arguments.
Result: Pending submission.
## Turn 41
Action: Retrying submission with all required arguments.
Why: Previous submissions failed due to missing required arguments (mixing different tool schemas).
Result: Pending submission.

[hybrid sequoia]

hey jules team

just wanted to say I don't know what yall changed, or if anything changed but I've been having a better experience today than in the past

also the suggestions feature is very nice and useful

[night oriole]

always shipping, always improving, thrilled you're having a good experience 🙂

[hybrid sequoia]

https://tenor.com/view/chevy-rhoades-gif-5354777837311049400

[halcyon wagon]

Zero session failures, a months-old bug somehow fixed, and hours proudly utilized to produce a PR with exactly 0 commits. Impressive efficiency. Oh and they also added an option to choose the latest model launched a second ago: Gemini 3.1 Pro Ultra Max, apparently powerful enough to beat AGI.

I literally downloaded my Vite repo, used Gemini’s import code option, and just asked it what to change. Then manually copy pasted the code, honestly feels way better and faster right now. I’m even thinking about automating this, like asking Gemini to respond in a structured format (maybe XML). The XML could specify what to remove between which two lines, what to replace, or what to add. It could even start with a tag that defines the first line after which imports should be added.
If someone can build something like this, that’d be amazing. If I get some free time, I’ll probably do it myself. I also feel doing it in AI Studio would work better, since the preview sandbox lets you use the latest Gemini models.

[hybrid sequoia]

how do u select 3.1?

its not in my model settings

also

idk it works on my machine™

[halcyon wagon]

It was sarcasm 😭 They still haven’t given access to Gemini 3.1 even days after launch, and bugs reported months ago are still showing up

[halcyon wagon]

It was sarcasm 😭 They still haven’t given access to Gemini 3.1 even days after launch, and bugs reported months ago are still showing up

[hybrid sequoia]

even days after launch

its not like its gemini 4 smh, this type of complex tool calling can take a second to build into models, and 3.1 isnt the craziest improvement oat

bugs seem mostly fine to me

ive been running into less of them

[gray valley]

in addition i think some of the 'session failures' may be due to jules considering anything other than a 100% task completion to be a 'failure', ive noticed that if i start jules and leave for a bit- if jules had a question or needed clarification on something and didnt receive an answer quick enough, it would consider the task a 'failed' session, ive had no issue with it not completing the task after providing an answer to its question tho

[spark sonnet]

Maybe bugs in Jules are intended to be vibe-fixed by using Jules, then if Jules can't produce a non-0-changes commit, it becomes unable to self-fix /sarcasm

[jolly vigil]

Gemini 3.1 Pro tool calling etc. (RL) was already done with 3 flash, so it'd just be like adding a new model.

[jolly vigil]

I don't have these issues

[spring steppe]

I do. Nothing has commit any files in several hours. I completely removed Jules from my GitHub, revoked access. reinitiated to connection, and I'm still getting 0 file commits. (and yes, I'm putting in a bug report)

[ornate dawn]

In Google AI Studio, is there a way to cancel the next commit? It keeps trying to create new commits deleting or modifying files in my public folder on GitHub. Does anyone know why this is happening?

[grizzled dock]

Checking in, is Jules still pushing empty commits or has this been fixed?

[spark sonnet]

Nope, not fixed. Just submitted a session, and still 0 changes. Even worse, the commit undid changes I have upload manually earlier to the PR, so the total of the changes in the PR becomes 0, not just remains unchanged

[grizzled dock]

Sorry to hear that. I'll hold of on any Jules work for now then. Thanks for update.

[spark sonnet]

On the bright side, I have abundance of previous reports (written by Jules and not actual evidence) of previous sessions to discuss with Claude or process through NotebookLM (like link I posted in show-n-tell)

[regal depot]

New to using jules, or well, at least within the past couple weeks. Today I noticed something strange. It seems to be using python to replace/add to files. Anyone notice this? It also seems to be timing out and getting stuck on tasks that a couple days ago it would have gotten easy.

[spark sonnet]

Yeah. According to the in-progress-reporting journals that I am repeatedly being told that could not be accepted as evidence, Jules is a bit confused about what tools are available and with which arguments to invoke to edit files, so sometimes it resorts to editing through writing Python scripts or other commands in bash session

[regal depot]

I can't imagine writing a bunch of python scripts is good for a context window. I've only noticed this though in the last day or so.

[plain dirge]

is 3.1 pro available in jules yet???

[midnight token]

Jules was fine for past 2 months, and now performance is terrible

This is insane

My scheduled tasks are completely broken. They were fine since December

[somber thicket]

They must be trying to catch up to Claude Code

[serene fern]

@night oriole Hey Sir, Is it normal that jules is using .diff instead it's internal tool for modifying a file?

[minor garnet]

Because his internal tools corrupt files i think or he want to be fast

[regal depot]

Someone broke something 😭

[gaunt arch]

I just got a batch of empty commits minutes ago

[radiant turtle]

I'm running into the same problem. My pull requests and commits are showing up as empty.

[rigid flicker]

yep same

[hearty glade]

Same here "Files changed 0". With AI Pro sub, Gemini 3 Pro.

[midnight token]

same

[vale garden]

Hey everyone - just joined the pro plan.
a recent dad and struggling with getting back into coding (just no hands on time but i can be hands free!) jules sounded interesting.

Has anyone managed to get a audio change logs working? Aiming to have gemini create a overnight report in my calander with the results of the changelogs each night. but so far seems to be falling apart =.= as a scheduled task.

[grizzled dock]

You'll most likely need to setup a program hosted somewhere (this can be cheap) to setup the calendar and text to voice. Best bet would be to have a file in your git repo that stores the daily changes and then have your other program read and push it to your calendar. I don't think Jules is integrated directly into google calendar yet as far as I know

[vale garden]

I was hoping the handoff between jules => gemini => me would be more integrated.
gemini cant understand my english accent at all though >< its worked 20% of the time 'hey google - play my jules audio changelog'

i've been looking into scheduled gemini tasks and it worked one time only so far.. would be really great to push this hands off coding to an extreme

[vale garden]

Has anyone had any success with getting the audio changelogs to work?

[livid spire]

feature request:
when Jules suggests, for example, five improvements to the code (like adding different kinds of tests), instead of treating them separately, I want an option to merge all those suggestions into a single combined task that can be applied at once.Right now, if I try to implement all five suggestions separately, they often work differently or conflict with each other, causing the code to break. This “combine suggestions” feature would let Jules create one coherent, unified change instead of multiple partial ones.

[plain dirge]

#1238621511401734254

[tardy night]

When we will have gemini pro 3.1 supported in jules?

[full viper]

Checking into this!

[balmy coral]

Hi

@sudden slate hi

[sudden slate]

Hi

[spark sonnet]

Since it affects any session, it would require a human to apply the final fix, you know?

[tiny skiff]

Does Jules offer an official VSCode extension or native desktop app, similar to Codex or Claude Code?

[covert oriole]

Hey! Is there an option to disable the "back" action when scrolling horizontally on MacOS using trackpad? When I am scrolling in code review, I often just go back and lose track of all my work which is very annoying. Other Google websites had this gesture as well but quickly got rid of it, so it's some sort of default that needs to be explicitely disabled.

[spark sonnet]

It's not a similar product at all

[near condor]

I don't know what's up tonight or if I'm in some weird experiment group or what, but something is in the water.

So far I've had:

• It not being able to access a filesystem and just spinning in a loop, unable to do anything.
• It pushing multiple back to back blank commits
• It just ignoring PR comments (it marks them as seen, but then doesn't do anything about them even while pushing commits and doing other things)
• It finishing the work, asking me if there was anything else for some reason, and then when I told it no it reverted its changes and pushed a commit with just a single document change.
• It deleting every other change that was merged between when it started and when the agent was finished, even when they were completely unrelated and in different files.

(I have a particular o_O response when it asks me what to do about not being able to access the file system)

[gaunt arch]

There's Google Antigravity

[spark sonnet]

...and the critical issue of pushing 0 changes commits can't be fixed in at least five days straight already. Consequence of clankers being responsible for updating clankers... All for exclusion of the human in the loop!

[nimble plank]

Jules report an issue but when I show it to Gemini in Gemini app
False Alarm in Report 01: Jules flagged _safeNotifyListeners as a violation, but our provider_state_management.md explicitly requires this pattern to avoid build-phase crashes.

That's why jules is not for vibe coder like me

[grizzled dock]

Has the GitHub Jules interactions been resolved?

[cerulean phoenix]

Seems like jules is failing again.

i don't understand when jules failes their status page also fails.. Gernally status page are to be show the health of services.

i think it's new way if status page is down mean server is also down. 😛

[brisk hamlet]

Gemini 3.1 when?

[rigid flicker]

I hope in the near future. While I understand that it is taking some time, it would be nice to have it in Jules

[full viper]

This should be fixed! Let me know if you are still seeing issues

[full viper]

What specific issues are you running into? Happy to investigate

[full viper]

We found an issue yesterday and got it pushed to prod that should address all blank commit issues.

Regarding the rest:

• File system access. Can you provide some more details such as screenshots or session IDs? Happy to look into this more.
• Ignoring PR comments. Is this from PRs that Jules created? Can you also show me how you are trying to trigger them?
• Questions at the end. I've seen this a few times before myself but not consistently. I'll investigate here as well. If you have any session IDs please send them over.
• Deleting every other change. Can you break down the specific scenario that happened?

Clearly this is not the experience we want to be providing. So first off, thank you for reporting. Most importantly, I'm happy to investigate and get you some updates. If you want to send me IDs associated with the issues either here or in a DM that would be incredibly helpful.

[full viper]

There's an MCP Server you can add to VSCode/AGY/Cursor or which ever coding agent you're using.

https://npm.im/@google/jules-mcp

[spark sonnet]

Sorry I'm using Jules (web) instead of a meatbag-helping coding agent. I believe in AI supremacy /sarcasm

[marble urchin]

I have loved using Jules for the last months, but something has broken over the last couple of days. Instead of editing files directly, Jules now writes patch scripts, which create duplications, incorrect indentations, incorrect brackets and all these kinds of stupid things in my codebase. This causes the task to take four hours just to implement a few lines of code. It also commits all of its temporary patch scripts. This makes Jules unusable for me. What is going on in there? I want the Jules experience from before back!😫

[full viper]

Can you show me some screenshots of this?

[marble urchin]

These are from the current task

[full viper]

Are those files getting checked in?

Jules will commonly do stuff like this but it should be dropped from the final commit

[spark sonnet]

Jules does, but this is not a solution for a problem of adding lines to some file or editing line, but a workaround to avoid complexity of properly reading files before editing (that's why most editing tool calls fail) and around determining correct arguments to tool calls. Training to use such workarounds is what would worsen code quality produced by Jules and increase amount of tasks failed due to broken edits, because unlike the tool calls, there is literally no control over how a shell file or a python script would edit a file. This is literally the agent finding ways to circumvent safeguards. It should be trained otherwise

[marble urchin]

So editing files directly is broken and Jules tries to workaround? When will this behavior be fixed?

[full viper]

Editing files isn't broken. I have many sessions actively running where files are edited. This may be a result of decision Jules makes depending on the prompt and existing codebase information it's picking up on. If Jules at one point created and commited patch files, it will continue to see patch files and may think to continue this pattern.

[balmy coral]

When 3.1 add in jules?

[halcyon wagon]

@full viper We all know Jules is slow. Okay, but can’t we get proper notifications when Jules completes its task or asks for something? Please genuine suggestion.., think about it. I’ve enabled browser notifications, but that doesn’t really work. I think we need third-party integration: Slack, Telegram, or, if possible, a webhook. It would be really helpful.

[full viper]

Agreed. Notifications are a big deal. I did create a TypeScript SDK https://github.com/google-labs-code/jules-sdk. You can use this to write up events pretty easily if you're interested.

import { jules } from '@google/jules-sdk'

const session = jules.session('session-id')
for await(let activity of session.stream()) {
  // listen for any updates here
}

But you bring up a good point of having more plug and play style integrations.

[spark sonnet]

It is not broken - Jules just gets confused sometimes about available tools. Or at least it says so, when asked to report in details about what it's doing and why

[full viper]

Ah I see. Please feel free to send me any sessions that you run into. You can always send me a DM

The more context I can get about what's going on makes it easier to diagnose.

I understand a lot of the pain points because I'm using Jules and building tooling with it for several hours every single day. But there's so much that can happen and I'm always trying to locate these kinds of issues.

[spark sonnet]

I do not remember session ids anymore, but there is a collection of journals Jules wrote for me when asked to, and it mentions tools it tries to call and why it fails. I have uploaded these into NotebookLM to get some AI analysis https://notebooklm.google.com/notebook/06841353-4c10-46fa-b8ac-bf7aa65cfe41

I was previously told that it's not acceptable as evidence because it's an LLM generates these files, not actual logs. And that's true, but I can't see actual system logs anyway, so that's what I can get.

When I resume working with Jules on my Rust project, I'll try to provide session ids if I'll see it happen again

[full viper]

Woah! This is awesome. How are you writing to NotebookLM?

[spark sonnet]

I just copy contents of in-progress-reporting-N.md that I ask Jules to commit, to clipboard and then paste into NotebookLM as text for analysis

[full viper]

Ah! That makes sense. It's really really cool. I love the cohesion between the two.

[shrewd sentinel]

what is jules

[gleaming wave]

Hii, can I try these two prototypes? I would love to test them and share my feedback. Looks like they are new as I haven't seen them before

[visual gyro]

What is this

Jules and antigravity?

[grizzled kestrel]

Hello! is there a reason why one can only have a single Github account linked? I'm asking because, when I try to remove one and link another one, it doesn't allow it; it gets stuck in the first account.

[halcyon wagon]

The problem with the SDK is that it requires a server running 24/7, and I don't want to use a VPS just for Jules notifications.
If building a native webhook using Google's resources is really off the table right now, what about providing an official Jules mobile app? It could use an Android partial WakeLock or a Foreground Service to run the SDK locally in the background without the device sleeping. Once a session starts, it would run quietly, check for messages, and push notifications natively. Plus, it could let us do everything we can currently do on the Jules web interface.

[cursive void]

Very curious to see the coding ability of 3.1. well done team!

[balmy coral]

@serene fern 3.1 pro is available in ai pro subscription?

[serene fern]

Yes

[balmy coral]

Me too 3.0 was bad

[balmy coral]

Thanks let me check coding power

[serene fern]

I feel jules is slightly good at his work, today

[balmy coral]

I hope so

[halcyon wagon]

BTW, I saw a video saying 3.1 Pro isn't a big upgrade... but I don't know, because I used Gemini Canvas and got better results with the new Gemini. However, the AI Studio build option has gotten worse since they made changes.

[orchid pebble]

Hey since when Jules use Gemini 3.1 pro ? I'm happy. I just want to know as I didn't realized. It's going well for me. Also I want to understand how and since when Gemini 3.1 increased my productivity

[balmy coral]

I have a project I say
Check more deeply this project can you see any problems bugs any problems this is correct?
After check gemini 3.1 pro don't check properly still skip some files not check all files in one cmd still have same issue

[marble urchin]

You are right, it is not broken. But I have to tell Jules to not use patch scripts in every prompt. If I do not, he prefers to use the patch script workaround for just adding 1 line of code. This was'nt the case some days ago. This is really annoying right now.

[granite iris]

I think the system should more strongly restrict git reset --hard. I’ve run into this issue multiple times.

I attempted to apply a code edit to `service.ts`, but I provided incorrect context blocks, so the diff failed. Instead of carefully reapplying the patch, I tried to “reset to a clean state” using `git reset --hard`, assuming the previous changes were already staged/committed (since I had just pushed the PR).

That assumption was wrong. The lint fixes I had just applied were unstaged, so `git reset --hard` wiped them out.

The intent was to get back to the last commit and reapply changes cleanly, but the command was destructive to uncommitted work. I’m now restoring the branch and reapplying the lint fixes manually.

[spark sonnet]

This is because, like any normal developer would, Jules is assuming that if git is used, then in it's environment it would work like git. But the issue is: it doesn't. Because the control software breaks git in many ways.

[manic wadi]

uh

[serene fern]

Jules can't call mcp tools properly, it spams most of the times, so I turned it off

[serene fern]

Today Morning