jules-general・🦑

[outer ibex]

you did nothing wrong

[primal sentinel]

Still same

Again working working again 😔 stop

[gleaming kayak]

Im not sure why its even grep ing something - it should just insert/delete 1 line and thats it

[outer ibex]

unless you specifically mention what line to change, it will always try to grep

[gleaming kayak]

I sent it a screenshot which is like "line 1234 crash - did you mean to add end after if at line 1230" thing

[outer ibex]

screenshot wont do much success, it must be physically in text (like from copy paste or you typed them manually in chat). this solution goes the same for other AI models if you want them to avoid using grep

basically being straightforward

[jolly lotus]

Jules stays @ working for hours without generating any codes

anyone facing the similar issues?

[swift karma]

Wait for the dev team to say something (or nothing) in a few hours

They know the issues is happening, i just can't say if they're doing anything or not

It's hard to be comprehensive when it takes TOO long, i still try tho

[brisk hamlet]

the problem is not the "grep" command itself, I guess this is specifically to not spam Jules context window with complete, potential files, and therefore per se a good idea.
The problem is the spamming of identical commands (could be grep or any other command) or of 1 or n (e.g. 2 or 3) - loops of commands, identical, for hours.

I don't know what the product manager of Jules is doing, but it seems his prioritization is pretty derailed for months now. We got features nobody asked for on top of a broken product.
Jules could be extremely hepful and great, acutally outright owerpowered, considering its price point, rate limits and inclusion to the normal gemini pro account, but the bad product management blocks that.

[outer ibex]

yes it is overpowered until the preparation of gemini 3 release

[tranquil swan]

Does anyone else have problem to chat with a GitHub repo connected. Greece here

4 hours ago it was working

[leaden pawn]

Hello Everyone! We are aware of the current performance issues and our engineering team is actively investigating the root causes to restore stability as quickly as possible. We apologize for the disruption to your workflows, please post in #1238621511401734254 channel along with task ID. Thank you for your patience

[tranquil swan]

I have a question when does the daily limit reset? It is been over 1 and a half day and it got only about 10 less daily sessions

[brisk hamlet]

There is no fixed reset time.
Until now it was "running" i.e. the one task you started will decease to count exactly 24h later. Very fair tbh.

[tranquil swan]

Oh I see yeah this is fair. Thank you for your answer

[analog patrol]

Is anyone getting this error when starting a new task?

[rare lintel]

Me

[west zealot]

Hello everyone, I have an AI Ultra for Business account. Works fine on Antigravity and Gemini, but for some reason it doesn't work in Jules and shows the "upgrade" button with only 2.5 enabled.

[outer ibex]

Paid Jules plans are accessed via a Google AI Plans subscription, which is currently available only for individual Google Accounts (ending in @gmail.com).

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

[visual gyro]

Any updates coming?

[misty ibex]

Hey guys,
Can anyone tell me why Jules is specifically not accepting tasks of 1 of my repos?

[burnt elbow]

Why am I always stuck here?

[peak cobalt]

hey, please use #1238621511401734254 to submit any bugs and feedback you have for the team!

[hearty glade]

could this one be due to it updating over 800 files? 😅

ive had a task struggling to update 2 files over the last few days

[burnt elbow]

Thankyou sir

[burnt elbow]

I told him to debug 2000 lines

[hearty glade]

that sounds reasonable, i wonder why that required 848 files to be updated 🤔

I think I would not be merging that update to my main branch..

[burnt elbow]

Yes, that's right 🗿

[hearty glade]

ive just decided to try this route with my current task that im waiting on.. "please provide in plain text a prompt for another Jules instance to try this task as this task has been failing for days"

[misty ibex]

Any help guys...

[outer ibex]

try it in a new session, if it still failed then jules is having a problem

[burnt elbow]

It doesn't seem to be because of the 848 update, because it's the same thing

[burnt elbow]

That's my problem too and haven't found a solution yet 🗿

[misty ibex]

But just for 1 repo only?!
I created other session and it works for all other repo but gets stuck and says failed to start a session for this one

[hearty glade]

Jules provided a prompt as i requested so ill try another newer instance 🤞

[outer ibex]

that means jules has silently failed, its not worth waiting cuz we will never know when it will ends

[outer ibex]

i hope you didnt reached your limit yet, so how about a new repo for it ?

[burnt elbow]

It seems so 🗿

[misty ibex]

I have 100 session remaining for the day, I'll try new repo

[burnt elbow]

Good luck, friend.

[outer ibex]

if jules commit a changes but lost some of the files or even folders, the repo is done for

[burnt elbow]

Looks like this is the source of the 848 update he does the same thing over and over again

[swift karma]

If Jules was partially open-source like OpenHands i think people would contribute to it and it would be capable of doing more things

[burnt elbow]

Look this out bro 😂

[outer ibex]

looks fine as usual

[swift karma]

https://tenor.com/view/borzoi-siren-dawg-with-the-light-on-him-sailorzoop-dog-gif-2844905554045249724

[burnt elbow]

+103371 Even though only a few lines have been added to the file 😂

[outer ibex]

i know lol, i already saw mine peak almost 1m

looking at gh line changes, it only like few thousands

https://tenor.com/view/wtf-wtf-is-going-on-ohh-oh-ok-umm-gif-19411699

[burnt elbow]

Can this be called Psychopath🗿?

[outer ibex]

nah🗿

jules were reading dependencies files for no reason

[hearty glade]

task finished 🙂

[analog patrol]

Are they working on that? Do you have any updates?

[iron mesa]

there is a problem with jules, i am on pro and can't start new projects , what happened?

[iron mesa]

i am getting the same since morning

any news when it will be fixed?

[analog patrol]

I have the same issue

[analog patrol]

I don't know whether they know about this.

[limber frost]

hey all! how are u enjoying jules so far? I love it. Super complimentary for easier dev tasks in volume

[ashen glen]

how on earth has it done this from a +13/-6 branch?

[swift karma]

Cause Jules is slowly going crazy

[ashen glen]

It's insane. No idea why. I've never experienced any form of support with bug reports so I don't expect any resolution

[hearty glade]

Suggestion for Jules... some kind of service status within the app. This could save people from trying to work on their projects when it is going to create them more work.. also this could help reduce support tickets and messages that are all complaining about the same things.. I do occasionally see members of the team post updates but those updates are quickly covered up with chat activity and unless you watch the chat 24/7 you will miss those updates

[toxic dust]

What is the timeline to fix "Failed to start session. Please try again later." I am a pro subscriber and I still get this issue.👀

[leaden pawn]

Please report the bugs in #1238621511401734254 channel, we are triaging them regularly. Thank you

[leaden pawn]

Hello @hearty glade Great point! Having a status dedicated service status page would reduce support overhead and keep users informed. I will pass this suggestion to our developers for consideration. Thank you!

[hoary zinc]

I'm having a weird issue where my personal Gmail account and my business Gmail account are the same, but my AI Ultra subscription is on the business account. When I try to sign in to Jules, I can't figure out a way to get it to recognize my business account. So I'm getting the free basic features and not the features in the LLMs with Ultra.

[prisma vine]

Hello @hoary zinc paid Jules plans are accessed via a Google AI Plans subscription, which is currently available only for individual Google Accounts (ending in @gmail.com)." For more information please visit Jules FAQs

[tepid steeple]

https://www.adobe.com/legal/licenses-terms/adobe-gen-ai-user-guidelines-linkfree.html

[primal sentinel]

I feel Jules working is slowly slowly no response sometimes 😭

[misty ibex]

Yes

[swift karma]

I feel, i see, i know, i presence it, i understand, i mentioned, i wait for the day it's gonna be fixed

[strange mist]

I submit task in morning, expecting result at night.

But Jules just fails in the middle, one way or another. Or, the solution is half-backed without supervision.

Devs need to give Jules more tools, capabilities and behavioural context. Like there's not even MCP yet! It have a literal VM to utilize.
For example, Chatgpt's Browser Agent could do lots of complicated things fascinatedly; even things it was not intented for, like writing codes on Websites, all while ensuring it completes it's job completely (until it reach it's sanity context limit). It's amazing what only visual input and usage simulation (browsing) capability addition could do.

[rocky forge]

because it loves you

[outer ibex]

i wonder whats the actual memories was lol

[swift karma]

Day (i don't know) waiting for Jules to be fixed i'll say it's day 21

[ashen glen]

it's done it again:

[deep gust]

jules doesn't read my env project settings

[swift karma]

Do the devs understand thats an emergency? 😂

[brisk hamlet]

249 memories updated in the last session. Finally some movement on the memory tool. Much appreciated. Lets see if it sticks.

[outer ibex]

its not consistent when jules trying to read saved env. one thing i always put in the prompt was "use the stored env variables/secret" or ill just put the env in the chat itself and it will be saved and can be use on other session

[feral scarab]

Guys i am not much introduced to tech fields, can anyone please tell that how can this ai come in use for me

[halcyon wagon]

Jules website consumes too much data.
Yesterday, while using the Jules web interface, it consumed around 7 GB of internet, which is really bad and not practical for limited data connections.
However, when I use Jules via GitHub issues or PR comments, it works much better and more efficiently. Today, with that approach, it only consumed about 1.5 GB, which is a huge improvement.
The main problem I’m facing now is that I’m unable to create or use Jules on an issue linked to a branch other than main.
If anyone knows:
how to make Jules work with non-main branches, or
a better workflow to use Jules effectively without wasting so much data,
please share the solution.

[outer ibex]

its not just internet, but also browser usage

[feral scarab]

i know how to code in python btw

Can this be integrated into my ide

Pycharm or vs code

[outer ibex]

you can use jules cli or api

[feral scarab]

So I can use it through my terminal

[outer ibex]

yes

[feral scarab]

I also wanted one advice

If i have learnt python then which lang should I learn next if my interest is in ai fields?

[outer ibex]

depends on what you want to build but i would suggest c++

[tepid cliff]

yo jules is down?

[prisma vine]

Hi everyone please report the bugs in #1238621511401734254 channel,with as much detail as possible. The labs teams are working hard to address the numerous reports so we ask for your patience Thank you

[gleaming kayak]

it wants to remember

its funny how jules does stupid stuff like that but it also fixed ton of bugs in my code so its geniuely good

[visual gyro]

Thinks

[gaunt rivet]

I used to be on the pro plan, and jules responded usually made quick progress and was very responsive. Now, I'm on the free plan, and it's not like that anymore. Jules takes 5-10 minutes between updates, and can ignore my messages for up to 30 minutes at a time.
Are you guys throttling too hard, or is my project just broken?

[vague valley]

hi folks, I'm not sure if this is the proper channel for this message, excuse me if so. I'm a having an issue where I have an AI Ultra for businesses license assigned to my business email but my account in Jules shows still as if was the Jules plan and not the Jules in Ultra, can't run more than 15 daily tasks. Does anyone know where could I get support for this problem?

[leaden pawn]

Hey @gaunt rivet Your project isn't broken. We are actively working on these performance issues to improve response speeds and will definitely post an update once it's sorted. Thank you

[swift karma]

It sucks to have to wait

[drifting yew]

Let's say your repo has a PLAN.md that you want Jules to follow without stopping until there's nothing else on its checklist. How do you make that happen? (Checking in every 20m is tedious.)

[odd monolith]

hi, always something went wrong

[outer ibex]

Paid Jules plans are accessed via a Google AI Plans subscription, which is currently available only for individual Google Accounts (ending in @gmail.com).

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

[dire wolf]

AI takin over 😄

[outer ibex]

dang you got the suggestions huh, im still waiting for mine to show up cuz its been over a month already

[dire wolf]

🤌

[outer ibex]

jules is officially gone crazy

[minor garnet]

let him cook

[minor garnet]

i mean its a octopus dont expect anything from him

[vague valley]

really? that's really bad and does not make much sense to me but oh well, thanks for clarifying.

[swift karma]

Day uh i don't know waiting for Jules to be fixed

[minor garnet]

back in my days

jule was the goat now its the woat

[swift karma]

And it feels like the devs are doing nothing cause they don't even tell us if there's progress in fixing the issue. I KNOW they're working according the staff but it doesn't feel like it 🥲

[minor garnet]

i think you dont understand how much of a pain jules is to work in

[swift karma]

No i don't but it's been a while

[minor garnet]

jule has his own container for every instance wich mean he can install everything he want in it

[minor garnet]

what issues tho

[swift karma]

Slowness, skipping verification processes, submitting empty code, memory issues, etc.

[minor garnet]

not surprising

i feel bad for the people that pay for jule they should create a app that use your own local container

[swift karma]

I've been trying to do that for a while, the only issue is that using a API is more expensive than using Jules

[minor garnet]

ohh

[swift karma]

OpenHands is the only sollution i found, tho i still have to deal with limitations

[minor garnet]

isnt it free?

[minor garnet]

what is it

[swift karma]

It is free, but still slow!

[swift karma]

Rate limits, so i'm trying to make a local setup to run using local LLM

[minor garnet]

i might try to make a local jule agent but nobody is interested by my projects...

[swift karma]

Jules cannot be local if you don't have the model

[minor garnet]

ive already made one https://github.com/Mk3Ang8l/Eveline.git

[minor garnet]

jule basically run on gemini 1

2.5

[swift karma]

How much ram does this consume

[minor garnet]

1gb

in docker without docker is way more but i dont recommend to use it in your systeme without container

also can you starr my repo it would help

[swift karma]

the LLM is local too?

[minor garnet]

no api key

i might add ollama support idk if my projects get attention

[swift karma]

Yeah i'm looking for something that doesn't take my money in two requests 😅

[minor garnet]

mistral is free

depends on the model theres too much too add

[swift karma]

Can you provide me more information about this

[minor garnet]

basically

with mistral you pay what u use but like if you want to code with that it will be expensive like everything else my agent just do tool calls

[swift karma]

interesting

[minor garnet]

yea man should REALLY star that repo im not forcing you

[swift karma]

I'll give it a try 😂

[minor garnet]

but it would make me HAPPY yk

[minor garnet]

you da real one

[tranquil storm]

does jules API support scheduled tasks?

[tranquil swan]

I need your help how can I remove a GitHub repository from Jules?

[leaden pawn]

Hello @tranquil swan Go to your GitHub Settings > Applications > Installed GitHub Apps > Google Labs Jules > Configure > Repository access > Only Select Repositories > select only the specific repos that you need from the dropdown list and save

[tranquil swan]

The repositories are still showing on Jules. How can I delete them from Jules also.

[leaden pawn]

Please try to refresh the page and see. I just did and it works for me, let me know if you need more assistance

[tranquil swan]

I see if I archive and delete all sessions from Jules then it disappears. Is this the way?

[visual gyro]

Jules is being lazy why?

[swift karma]

it's always been like that

[visual gyro]

No not when it was beta it was powerhouse

It used work task for hours and days

And did really good improvements

Now it feels like context window got smaller and it missing alot on the delivery

[swift karma]

Jules is lazy, slow, mostly broken and unreliable. That just frustrates me but i'd probably be banned for saying bad words i'd use against it so i just act like a normal human being, sit and wait.

[frigid rivet]

Man, Jules has been a blessing, I have been dreaming of making my idle chess game and it has finally come to pass thanks to Jules

[carmine coral]

I m trynig to use docker compose to setup my jules env, but keep getting this error:
permission denied while trying to connect to the Docker daemon socket at unix:///var/run/docker.sock: Get "http://%2Fvar%2Frun%2Fdocker.sock/v1.51/images/infra-api/json": dial unix /var/run/docker.sock: connect: permission denied

[prisma vine]

Hi @carmine coral please open a bug report in #1238621511401734254 channel with as much detail as possible in order to reproduce the issue. The Jules team is working hard to address the numerous reports so we ask for your patience Thank you!

[carmine coral]

turns out i needed to invoke docker compose with sudo

[leaden pawn]

Not really! Please refresh the Jules session once the repository access list is saved in Github. If you still face the same issue, please report in #1238621511401734254 channel with more details

[tough quiver]

Are there any programmatic hooks into Jules? It is almost without fail going to fail a ci step on first PR submission. I wish it would just listen to a GitHub actions job before opening a PR

I'm at the point of building a GitHub post-fail step to automatically update the thread on certain step failures

[swift karma]

Someone tell Jules it has to get back to work please 🥲

I'll do something funny... Every single time Jules does something it's not supposed to i'll post it on the bug report channel

[visual gyro]

Sound like plan

[swift karma]

Does it? 👻

[cosmic bay]

If I push a change to a jules PR branch in my project, Jules does not grab the changes, or does it?

[swift karma]

It won't unless it updates the local code itself, i think it's possible to ask to it to pull latest changes if it doesn't do anything stupid

[cosmic bay]

ah well, i asked it to but it ignored me and worked on the thing and made a new PR anyway

it's fine ill just make new sessions, i have the env setup in a snapshot anyway

[swift karma]

this is what i'm talking about when i say Jules is lazy 😂

[cosmic bay]

just dumb to ask it "fix a tiny compile error" when i can do it myself

[swift karma]

Jules can indeed do it, it just doesn't. Either cause it wasn't trained for it or lack of parameters

[cosmic bay]

ill put it in suggestions or something, a button to pull latest from the branch we're on
instead of relying on it

[swift karma]

Usually it skips the verification processes i want it to do so

[cosmic bay]

it didnt want to install a dependency, so it just pushed the PR without testing 😂
i need to fix the environment

[swift karma]

My Jules is stupid because the devs didn't fix the memories yet so i can't manage them and it does things i don't want it to

There's NOTHING i can do about it

[cosmic bay]

i mostly stopped using jules for now while they fix it, and just using agents in antigravity - then if i am on cooldown ill ship it to jules and pray 😂

im using it as a hobby anyway, i wouldnt try to run a business relying on AI coding lool

imagine relying on jules and AI coding only to run a business (guaranteed some are) 😂

[visual gyro]

I dont know what happened 2 Jules was on so good running simulation and task now nothing? It go Idle quickly

[swift karma]

Jules skipped code review https://jules.google.com/session/9648235634647607538/code/.gitignore

[visual gyro]

Jules is working on big big project

[late sparrow]

session is not loading:

https://jules.google.com/task/8666340343300245705

[junior yarrow]

Is there a way to use jules without giving access to certain GitHub permissions like workflows or is something like limited access on the roadmap

[swift karma]

There is NO progress in fixing Jules 😭

[hearty glade]

I dont want to tempt fate but... Jules is working for me on pro.. 🤞

[swift karma]

I'm not sending those session links in vain

[hearty glade]

Can understand that 🙂 I just thought Id add a report of it working as usually not many mention when its going right 😛

[junior yarrow]

tough, makes it hard to use jules on anything that's not a quick poc or a throw away project

[shy pine]

CAn anyone help me understand the difference between using Jules or Antigravity?

[swift karma]

Jules = Codes itself + runs commands on virtual machines (not your device), doesn't have conversation limits (except the , is stupid
Antigravity = is a little smarter, runs commands on your pc but has limits, if you code on mobile or low end devices you will struggle

[outer ibex]

jules = agent automation
antigravity = agent + full software development

[outer ibex]

opening heavy chats in jules web on phones browser can crash the page 😔

[swift karma]

What's yours?

[outer ibex]

i dont know what your question specifically ask for but

chats about fixing bugs, crashed at 2gb ram browser usage, can only be opened on pc/laptop

[swift karma]

The phone

[outer ibex]

s23 ultra

[swift karma]

And it's lagging?

[outer ibex]

well i think what phone doesnt matter cuz heavy ram usage for browser is still arent optimized

[outer ibex]

yep, tried other "fairly" old phone with sd888 still nope

[swift karma]

This shows how bad the optimization is. 🥲

[outer ibex]

ah well not a jules problem by much but it is what it is

[halcyon wagon]

Yeah, exactly, Jules’ web interface consumes too much RAM and internet data. Sometimes the phone browser even crashes, and my internet bills go too high.

That’s why I figured out new ways to use Jules. You can use it via GitHub by creating an issue and assigning the jules label "Jules" starts the session in the background and makes a PR. You can also comment @jules on PR and it will act on that comment and can make new commits.

It’s like prompting Jules through GitHub instead of the web UI. There’s also a Jules CLI, but I don’t think it works on Android.

[swift karma]

Jules ignored code reviewer https://jules.google.com/session/5751199170186154807/code/.gitignore

[faint obsidian]

hi guys, how do you maximize jules? what kind of repos do you usually use the agent?

[swift karma]

What do you mean by maximize?

[faint obsidian]

like i've seen people use jules for 80-100 sessions per day

[swift karma]

You need AI Pro to get 100 sessions per day

if you have just talk to Jules like crazy, i don't think people do spend that much of sessions depending of what they work with

[faint obsidian]

yepp. I was thinking they do scheduled tasks for their repos, so im just curious as to what are they usually, in order to really maximize the 100 sessions

[outer ibex]

i can think of someone testing out the possibility of using scripts to automate jules cli and api

[visual gyro]

Think Jules project is on ice something happend with antigravity team because no promotion or activity from the team lately here looks like they working on bigger project

[leaden pawn]

Hello Everyone! We are working on the reported bugs. Instead of the general chat group, please open a thread in #1238621511401734254 channel with as much detail as possible in order to reproduce the issue.

[leaden pawn]

Hello @swift karma This has been fixed, could you please try once and let me know. Thank you

[swift karma]

I believe this likely is not the issue i was referring to tho. My issue is very simple:

I cannot see the memories Jules saved on the repository, it can still resd them. Proof: https://jules.google.com/task/13426660973129996260

[slender finch]

Hi

[leaden pawn]

Thanks for the confirmation @swift karma I could see mine for the codebases that I have enabled this option. Let me check and raise a bug for this. Please open a thread in #1238621511401734254 channel with same details so that this issue doesn't get lost or mixed up again.

[gritty dirge]

Bonus, one of them thought that I AM AI! So I guess I I am become Jules, security reviewer of worlds?

[west venture]

free jobs if jules suggests it?

nice

[wild flicker]

lol plaintext password store? for a discord app?
the "no warranty" thing is not wrong but the reaction is entirely defensive

[gritty dirge]

No warranty for GNU licensing but it theoretically could break both GitHub and Discord’s TOS.

No warranty with a GNU licensed cryptovirus either, really.

But yeah it’s not just plaintext password store it’s also plaintext discord MFA token store that is accessed via the discord developer settings option for authentication

So it’s like whole account access. Payments activities. Etc.

(Long story short, I published it as a .deb for my discord ppl to use and then one works in infosec… and yeahhh - one thing leads to the next and I’m arguing about licensing not covering the author from consequences and not being a legal shield if it leads to leaked credentials etc)

[serene fern]

What is your thought on this

[swift karma]

Looks peak but looks expensive

[swift karma]

They fixed Jules status, nice

[swift karma]

Crazy but Jules kept running the same thing over and over again https://jules.google.com/task/10244070369534990836

[serene fern]

It's available for free, but I use the paid tier

[swift karma]

You're referring to that tool you use right

[swift karma]

This one, could you share? Plus the docs

[serene fern]

Can't talk much due to slowmode enabled, can you give me your social id to chat?

[swift karma]

you can just DM me

[serene fern]

Why Jules is not doing something like this 👀

Here is how can Jules be improved:

1. Quick response.
2. Plan mode (Current one is a flop, it's just asks common question which can AI tackle by scanning the codebase)
3. Providng MCP support.
4. Responsing PWA

[serene fern]

PC View

[jolly vigil]

PLEASE allow a option to clear the memory in its entirety

[inland solstice]

Hello, I'm new to Jules. I'm using Jules on GitHub, and Jules has opened a PR, but it doesn't fix the cicd error. How can I get Jules to read the cicd output and fix it?

[cosmic bay]

Copy paste the cicd error into a comment in the PR and tell it to fix it

[marsh iron]

Does anybody know what we can do to troubleshoot if Jules is not responding to github issues? "@jules begin working on this issue" is doing nothing and I've already uninstalled and resinstalled the github app

[marsh iron]

Nevermind, I read the docs and you have to add a label "jules" to the issue for it to work

That seems unintuitive, classic google lol

[final elk]

Where to post the bugs, I see few of my tasks are getting stucked.

[swift karma]

#1238621511401734254

[real patio]

when the AI gets into a loop like that, just interrupt it, point out it's in a loop, and tell it to dump the memory noise

[scarlet hemlock]

Is it possible to switch github accounts on an existing Jules account? I tried and it seems to allow connecting a different github, but then the repos don't appear in the dropdown

[gritty dirge]

It may also be to stop accidentally triggering @jules when it appears in a quoted comment from a previous review and suddenly you are in a loop of Jules and Copilot and GitHub all reviewing and commenting and AHHHHH BRANCHES AND MORE JULES AND MORE NEW BRANCHES WHY DO I HAVE 27 JULES-WIP BRANCHES

Anyway not that that’s ever happened

[gritty dirge]

That works. My instruction is usually something like to “envision the entire project and goals thus far; step back, analyze, debug, troubleshoot the current thought process loop and walk me through technically+intellectually what you are experiencing and any logs you may have. Show me the parts where you are truly stuck.”

Get code snippets. If I can’t figure it out, if I can’t google it, Jules keeps looping I then to commit and PR to myself for a squash and merge post-PR (having those GH automagical tools running for the copilot analysis) - fix all the little bugs.

Have copilot work through where Jules got stuck, guiding it as much as possible (the more verbosely explicit with instructions the better, just like Jules) then Squash and Merge the PR, and then start a new Jules session with the changes in place: having already instructed the broken Jules session to give me its roadmap in low-level imperative; informative detail and then go on.

Just: care about keeping all the AI notes. I thought they were useful for provenance and attribution to AI but they garner a lot of negativity

[real patio]

i just tell the AI "dump everything you're worried about forgetting to a file you can read later" and then "reset to baseline" - that reduces the memory noise down to where the AI can function again, and having it in a file it can re-read reduces panic. all computers use swapfiles for memory management. we humans do NOT keep all the details in our brain in active memory at all times, we are constantly swapping unneeded details to storage. AIs need to do the same thing

[wild flicker]

Since I'm guessing there hasn't been an update since winter break, the team is working on a large update?

[gritty dirge]

AI is the human swapfile you are saying?

Hence the giant data centers

[swift karma]

Jules needs to interact more with Github.

[halcyon wagon]

Previously, Jules was working perfectly with GitHub PR comments. When I mentioned @jules in a PR, it used to correctly analyze the comment, apply the requested changes, and push meaningful commits.
However, now the behavior seems broken. When I comment on a PR with @jules, it runs, but then creates a blank commit with no actual code changes, only adding a generic commit message or description.
This makes it look like the task was completed, but in reality nothing changes in the codebase. Because of this, Jules is currently not usable for GitHub PR comments.
Please investigate what changed in the PR comment workflow and fix this regression.

References:-
Jules session: https://jules.google.com/session/11816837098063052185

PR:
https://github.com/Brajesh2022/Hello-/pull/2

Blank commit created by Jules:
https://github.com/Brajesh2022/Hello-/pull/2/commits/a0defe69b0834e7dd3f8748b06ee55e49fac2855

[signal comet]

Anyone had luck getting good PR reviews from Jules? it seems extremely inconsistent with the way to provide the review, Sometimes it would leave a reply in the chat, sometimes it would create an md file(altough rarely the full review is there, most times its a couple of lines), sometimes even in more than one occasion it tried to provide the review in the commit comment(it was empty)

[inland solstice]

Does Jules support mcp?

[real patio]

no. watch yourself think. really watch do something and pay close attention as you do it. you'll be very aware of every single detail, but less than a second after you finish an action, you'll just remember you took the action, not all the details that were part of it. all those details are noise you don't need in active memory. you haven't forgotten them, you've swapped them to a short term storage area in your brain.

[gritty dirge]

So you are describing mindfulness and introspection

[gritty dirge]

Do you have instructions for how it should handle reviews and what style and format you prefer? (Even if cleverly adapted for varieties of codebases)?

[swift karma]

Jules seems to be stable again so i got AI Pro plan back again, hope it doesn't dissapoint me

[swift karma]

They grow up too fast

[plain dirge]

guys how to make jules not lazy?

it always have placeholder code bruh

[outer ibex]

just tell it to make them functional ?

i always force it like that

[plain dirge]

ok

[proven relic]

it have been too long since jules updated

[rigid flicker]

Tho I have to say Jules works pretty well in the last days for me. Not using it so much in the last days, like 1-2 Tasks per day

[outer ibex]

is it ? i might give it a try again

[rigid flicker]

Yes! Not crashing, being slow and all that.

[swift karma]

They just have to fix the memories now, that's all

[swift karma]

how do i prevent Jules from creating pull request automatically it's annoying

[rigid flicker]

Untoggle this in the settings

[swift karma]

I tought that this would uh ... Nevermind

thanks for the help tho 🔥

[white rain]

is there an API in jules for me to add and manage environment variables/secrets? I'd like to take all of a .env.development file and add it to the project env in Jules.

[rigid flicker]

Hello

[glad sequoia]

Hi

[dusky nest]

Did anyone know if there’s a way I can make Jules only work in one branch and not keep creating new ones? Im using the scheduled jobs but rather than create a PR I want the changes to be in the same branch

[swift karma]

This might suck since Jules is used to just submit code but you can ask it to do that

but the prompt has to be very specific, like "after doing the changes, push the changes to this branch: <branch name> using git push"

[dusky nest]

Actually that’s a great idea I forgot I can edit the prompt. I setup a scheduled default prompt and completely forgot about it and have 30 PRs. Jules did a good job at bulk merging but so should be okay.

[outer ibex]

early days of jules we were able to rename the branch before submit the changes and next time jules will always use those

i wish they would bring back that feature

[outer ibex]

and yet for my simple prompt and it failed on running npm init -y

maybe its a fluke i guess, third times the charm and work like normal

[main fable]

Seeing this issue too. Essentially makes Jules non functional

[rigid flicker]

hm that's weird, seems like the old jules. for me it is working without any issues, just a bit slow

I hope they really improve jules. it is a good idea and can be something good, but it is something that needs to be worked on

[dusky nest]

I wonder if there’s a way I can make Jules send it’s Knowledge it creates over to Antigravity where im in a deficit

[outer ibex]

mine works really fast and functional, just the failure rate on starting the task is still intermediate i would say

[raw imp]

hi. new here... guessing there is a better place to post this... if so please advise:

been working on vibecoding several projects (shipped a few) and have been getting really good at debugging and improving my developer practices..... now, as i launch into several new projects, i want to get really good at setting up new development projects correctly from day one--architecture, infrastructure, workflows, etc--so they can scale cleanly as the product grows. my goals is to build systems that make evolution easier, make reasoning more logical, and make them much more maintainable..... definitely finding that evolutionary complexity often leads to an accidental accumulation of mishandled code, bloat, and agents that think they are doing the right thing when they are building in the wrong environment or whatever.

[visual gyro]

Hello

[visual gyro]

Can we get possibility link this every project https://stitch.withgoogle.com/docs/mcp/setup

[twilit seal]

Imma non dev guys How reliable JULES is? I want to mix this mf with ANTIGRAVITY for last measure debugging.

[plain dirge]

do you guys prefer Gemini 2.5 or Gemini 3 in Jules?

i can't make Gemini 3 stop being lazy

[twilit seal]

I guess it's a no

[twilit seal]

To my question

[plain dirge]

yes but idk abt Gemini 2.5 tho, cuz thid current Gemini 3 Previes are soooooooo lazy

[visual gyro]

Gemini 3 is super Lazy 🦥 it gives minimal work Gemini 2,5 took longer but had alot hallucinations. It's like when chat gtp 5 came out, it put minimal effort if you where on free and even with think it was not good. Think longer and ai planning is not there yet tried rebuild webpage and webstruker with available context and photos and even WordPress html/css/js but it fail every Time. I trying find better ways implement click what options and packs available make pdf options of all with priced add ON and so

[proven relic]

Is it possible right now to push to specific branches in jules?\

[quartz tusk]

Good Day Everyone. I have a challenge that My Jules has not reset the daily limits in 3 days. I have tried clearing my browser cache and still having the same issue. Anyone experienced this and can assist?

[outer ibex]

2.5 for stability
3 for "professional work"

i use 2.5

[outer ibex]

right now jules is still unstable so better go with antigravity

[outer ibex]

you cant anymore

except by saying them explicitly in prompt(?). some people said it works for them but mine doesnt so

[visual gyro]

Yes it possible

Can any news of Destiny and Road map from Jules agent? Looks like team been relocated?

[quartz tusk]

Is it worth paying for it? I liked it sometime last year before they revamped it. It was much more stable. I am not sure anymore

[twilit seal]

Shoot

[outer ibex]

jules is dependent on your google one/ai plans type, so if you have google ai pro subscription, then you already have jules in pro

if you dont want to use jules, thats also fine because all it gave you was higher daily limits and gemini 3 in jules. you can just use other google ai experiments using the same subscription plan and get a better outcome for your needs

basically for jules, its considered unlimited even for free plan

for now, a simple project is better with jules

[leaden pawn]

Hello @visual gyro The team is still here and actively working! Our current focus is on resolving reported issues and improving stability. Please refer https://x.com/julesagent/ for any official announcements on product roadmap.

Jules (@julesagent) on X

Your new autonomous coding agent by @GoogleLabs. Built for devs, open to feedback, evolving with you. Dive in → https://t.co/iIzFEMmoqX

X

[visual gyro]

Great thanks

[rapid zephyr]

Hello I had reached limit of 15 open sessions and all got completely stuck i want to remove stop all of them but i can not access jules via cli, api call or web user interface problem is on [e-mail] account - where can i look for help with this?

all the time i get 503 error both on windows and ubuntu

[minor garnet]

could you pay for a local ai workspace with your own ai api keys?

[rapid zephyr]

I do not understand what do you mean @minor garnet , but thanks for response !

[minor garnet]

i mean could you pay for a local version of jule?

also i think the web interface is saturated by your sessions

[gray sequoia]

I've added an ultra ai license to our workspace and assigned the licence to myself but Jules is still showing me the basic plan. Has anyone else had this issue?

[minor garnet]

what is a ultra ai license

[gray sequoia]

Google AI Ultra for Business

[minor garnet]

Is it paid ?

[gray sequoia]

Yep -> https://workspace.google.com/products/ai-ultra/

[undone copper]

Jules + Antigrav is a powerful combo. I primarily use Antigrav when trying to map out and create the core app, along with integrating any front end design (Stitch or AI Studio). This is active for the most part, not fire and forget. Jules on the other hand I bring in when the core framework is nearing completion and have it run on scheduled daily or weekly tasks (incremental improvements every day, testing, red teaming etc.) or tackling or improving the results of my backlog from GitHub projects. The nice thing with Jules is it runs around the clock, with you having to check in a few times a day for PR. I preference this with you have to have a Ultra account to take full advantage of simultaneous tasks, or you are going to blow your Daily Session Limit in no time. Ultra gives you a 300 task limit a day, vs. (I believe) 10 per day on the smaller account. It's also handy that you can run the scheduled tests or automatic backlog vs. multiple repos/projects.

Edit: I also use Jules to take a second pass on a previous backlog issue (in the review stage) and ask Jules to improve it. This works extremely well.

The issue is that using both simultaneously means generating a tremendous amount of code to review, which starts to become overwhelming and becomes error prone if you do not document the hell out of it. Per folder readme are pretty much a must have.

[undone copper]

Set some system instructions such as "never truncate or abbreviate, always show full text" and similar.

[visual gyro]

Jules is not the same

[undone copper]

One area that helped me a great deal with memory-heavy web apps like Jules was upgrading my RAM to 128 GB. This helped tremendously; crashes and hangs are few and far between.

[minor garnet]

yea i see would you pay for a local ai workspace that work with your api key

[undone copper]

Send in a bug. This happened to me when I first started using Jules. Mine was fixed, and I thought the dev team patched this.

[gray sequoia]

Thanks

[undone copper]

This would be interesting to explore. Are you thinking along the lines of a local VPC?

[minor garnet]

more like a agentic ai with a vm that runs on your machines

and tools

[gray sequoia]

Where did you report this? Everything points to discord or x.

[leaden pawn]

Please report in #1238621511401734254 channel

[gray sequoia]

Thanks!

[visual gyro]

Missing old jules

[rigid flicker]

Old Jules was buggy, new is def better than old jules

[visual gyro]

It worked for hours and days this new Jules has limited tokens, 5 min then it thinks it is finished

[plain dirge]

it never follow instruction

[twilit seal]

The only thing that Im afraid about is the dang thing begin to hallucinate a messed up the whole project. While its purpose to be the final guardrail

[rapid zephyr]

I have already google AI pro ; do you mean the payment for this @minor garnet ?

[obsidian thicket]

How do I get jules to directly publish to main branch rather can creating new branches?

[uneven iron]

Is anyone still experiencing Jules (Gemini 3) just reverting progress out of nowhere

I switched back to 2.5 and haven't had the issue with it

[minor garnet]

because gemini 3 is too smart i think

[swift karma]

It is so smart that when i ask it to use git push it eats the commit history

[minor garnet]

ive been studying ai alot lately

i even made a lib for ai wait https://pypi.org/project/GuardianLayer/

[swift karma]

nice

[minor garnet]

its not perfect but yk ill work on it

also when models are too smart they react weirdly to some tool calls and since jule always been on gemini 2.5 i guess its more stablr

[swift karma]

Except that it can squash 1900 commits into one

[minor garnet]

explain

[swift karma]

(not that it happened to me of course cause why would jules do that right)

I asked Jules to use git push cause it saves time when i want my workflows do their thing automatically when my work is done. Jules goes there and messes up the commit history

[minor garnet]

genuinely how do you squash 1900 comit into one wont that make bugs?

[swift karma]

Jules does not know how to use git correctly, it is so used to just do edits and expect the user to normally submit using that button

[minor garnet]

and how do you even reach 1900 comit

[swift karma]

fork

[minor garnet]

i see now

[swift karma]

Even if you tell it exactly what to do, it will mess up

If you want to test it, try asking it to push the changes after you finish a task, it will submit the work again using it's tool

[minor garnet]

i genuinely stopped to use jule

using jule was like a virus mining crypto in your computer

[swift karma]

I just try it sometimes, if it wasn't for Antigravity i should've canceled the Pro plan already

[minor garnet]

i see gravity is very good

[swift karma]

Want to know something funny? It starts with "A" because you might accidentally open it instead of VSCode which is clever

[minor garnet]

ohh

[swift karma]

And if you don't understand but preended you did anyways

[minor garnet]

ahh you got me 👀

[plain dirge]

it rm rf ed my entire repo😭

[minor garnet]

do you make money with ai ?

[cosmic bay]

Jules creates its own branch how did it delete your entire repo?

Also always setup your github to protect your main branches! (don't ignore that warning on github!)

[plain dirge]

nope, i just love doing stuff

[minor garnet]

is antigravity weird with everyone or its just me

[swift karma]

Depends of what's weird for you

[minor garnet]

the g3 model

[cosmic bay]

if you let the context get way too long without starting a fresh conversation, it does start to get weird -- luckily, the models can look at previous conversations in antigravity - i tend to start a new conversation once i feel it's weird

weird for me -- it starts showing that it's going into execution mode in all caps, etc...

(this problem does not exist only with gemini, this is a problem for pretty much all AI models that exist out there, they break when the conversation context is too big)

[shell roost]

How do I make sure that Jules has jdk25 installed in its VM? This is for use with Gradle projects.

[cursive void]

One thing I have found is that one Jules agent is not the same as the next agent.
It's exactly like using interns, you can get a dumb slow one, or you can get one that should be running the whole business 😂

If I feel like I got a slow Jules agent I just stop it and try again and the next Jules agent might be better.
Benefit of 100 a day I guess

[carmine falcon]

I went too far? XD

[orchid jasper]

what happened to repo-less sessions?

did i miss something changing or am i just blind

nvm i figured it out

"detach codebase" sounds like you're about to detach a repo completely lmao

[hot mirage]

hello. i have google ai ultra but in jules i receive gemini 2.5

[gray canopy]

Did you sign in with the same Google account?

[hot mirage]

yes.

[gray canopy]

Hmmmmm.... Mine is Pro and I have Gemini 3 access.

Jules keeps pushing empty commits and claim that it made changes. 😢

[gray canopy]

So, you see "Ultra" on the top right side of the screen, and in the settings (Gear icon) > Model Settings > The only option is Gemini 2.5?

[hot mirage]

in gemini app. in jules

[gray canopy]

"Pro" is shown on my Jules screen. Maybe you want to try it in another cookie-clean browser?

[hot mirage]

didnt help. incognito too

[hollow spoke]

Hello here, sorry i think i missed alot here cos am not online for some time now

[fading sand]

Hello! I pulled the branch that Jules was working on and made some changes myself and pushed it back to the same branch jules sent their PR to.

I was wondering if they could pull my latest manual changes from their own branch and this is how it responded....

I attempted to pull your changes from the `feed-redesign-v3` branch, but the operation timed out. It appears I do not have permission or network access to pull from the remote repository in this environment.

Could you please paste the code you modified (or describe the changes) so I can apply them manually?

[versed lynx]

The code review flagged errors and yet it approved the PR. This happens regularly.
https://jules.google.com/session/3056121968224744109

[serene fern]

maybe bcz of the package naming, its a vs code oss fork

Jules website lags, so I made mine with API

[proud bramble]

is there a date for gemini 3 for free jules accounts?

[plain dirge]

is this an mobile app or web based??

[outer ibex]

i can say that i disagree

[outer ibex]

i never had any problem using any models in ag

[polar yarrow]

I'm running into exactly the same problem. Is there a resolution?

[minor garnet]

what is this?

[gray sequoia]

I'm in the same boat. I think Jules doesn't support workspaces/business AI ultra. I've messaged in the help forum and used the feedback form in Jules but heard nothing.

[swift karma]

There has to be something wrong with Jules system prompt seriously

[meager palm]

how do i find this archive my sessions are apparantly going to?

i created one, turned on proactive suggestions

then i archived it. now i can't seem to get back to it or how to actually delete it

so now i have 1/5 suggestions always used, and i can't ever disable it

how do i reach the archive???

[meager palm]

nevermind then, i just deleted my github account and it seems to have fixed itself. this is a design flaw for sure though

[low reef]

jules is so fun, and improved my projects quite a lot already! got a set up with different agents/personas for my main project now, with each agent also maintaining a memory in my repository.

also got a GitHub workflow going that automatically starts a new session of that agent after its PR is closed.

thinking about open-sourcing this workflow if anyone is interested 🤔

[gray canopy]

Is this normal? I have lots of issues in Jules. I wonder if this is what everyone is experiencing. These are happening to me almost 80% of time.

1. It often commit empty diff.
   1.1. Sometimes it forgets to add files and push.
2. When I ask Jules in the PR comment with @jules, it seems like it triggers it, it adds eyeball emoji, which is great. But the Jules task does not seem to know the content of the comment. It first tries gh pr view 1 --comments when 1 is not the correct PR number at all and it fails because gh isn't installed. It tries to install it and it fails.
   2.1. I do not have any prompt about gh in my AGENTS.md or so. So that's what Jules just made up or perhaps by reading my other scripts or code in my repo or not.
3. When I ask to fix lint errors or compilation errors with the given command, it runs the command, see it fails, and commits empty change without taking any actions and claims that it fixed it.
4. If I create multiple PRs at the same time, it almost always ends up deleting changes from other PRs in parallel. Maybe it does not understand rebase well.

[outer ibex]

same issues everyone experienced for the past months, and no its not normal

on your issue 1.1, i had these unrecoverable and inaccessible files in one of my repo branch thats also cant be downloaded due to broken structure made by jules

i had no choice but to delete the branch because triggering jules to fix it back whether on the same session or new (or simply ask it to rewrite it again), jules refused to do it

[swift karma]

Welcome to the dark side of Jules 🔥

[swift karma]

Jules just admited why it doesn't do what i ask

[outer ibex]

https://tenor.com/view/blackbeard-blackbeard-writing-taking-notes-writing-gif-17583038544116987898

[swift karma]

Bad system prompting and training, guess that's all

There is no logic in jules being that incompetent when it uses same model antigravity does

[outer ibex]

yea

[swift karma]

Tho i doubt google would make Jules open-source

[outer ibex]

like i have no issue in ag nor the standard gemini web chat, except jules

[swift karma]

that's why OpenHands, Opencode, openchamber will be more reliable than it

[outer ibex]

i would prob make my own

[swift karma]

Jules is a "i can edit code" tool, not a complete customizable suite

I have no clue of how it would work since google wants to center it's products on their model

[outer ibex]

yes, im also implying that rather than using jules in web, i would make use the api and cli like some other user in here which i find fascinating

[outer ibex]

hmm

[swift karma]

Just so you know, people prefer Claude Sonnet over Gemini 3.0

AI Pro users are always complaining about antigravity and it's "5 hours" limits

[outer ibex]

i prefer gemini as a great ui designer and basic foundation and functionality while claude can improve it further on

and im also that same kinda person

[outer ibex]

bro its plenty of quota

[swift karma]

90 seconds and it's joever for claude

[outer ibex]

debugging consumed too much quota than doing the actual complex tasks

[swift karma]

Relying on Gemini 3.0 is not a problem, the problem is that gemini 3.0 requires too much guidance for coding

you gotta agree with me on this, claude can have way better ideas than gemini can

[outer ibex]

its true, but like i said its a great ui designer and nothing more

[swift karma]

gemini will normally have more "simple" ideas while claude makes you go "that's clever"

[swift karma]

Let's hope things change

[outer ibex]

and claude can give loads of awesome suggestions while planning too

[swift karma]

Forgive me for the nonsense, i'm just a 26 years old yapper with no hobbies, i get lost with the convo as it flows

[swift karma]

give jules a plan mode i think, interactive isn't enough

i believe it had a plan mode on it's beta, but i never managed to see it cause nothing would ever show up. So the devs decided to make it simply ask raw text

[outer ibex]

i sure hope so, cuz when i actually trying jules the first time months ago, it was truly a miracle doing the tasks like the user said so and not spewing half baked codes

[outer ibex]

i used interactive mode

refined the planning like the way it supposed to be, then proceed

[outer ibex]

no problem. actually relatable

[brazen crystal]

they just need to fix the bugs, Jules is ahead of its time and if your style of communication is verbose it works great

[gray canopy]

I suffered enough when it came out with 2.5. I was hoping that it would be somehow better now. I found that 3 is better in LLM quality but more subtle issues were found. At least I liked the gh issue integration, labeling jules to trigger it.

[gray canopy]

At least those engineers in Antigravity knows how to build a coding agent from their early experience from Windsurf.

[swift karma]

my communication time been verbal aggression recently

[frigid marten]

I am building an OSS version of DevinReview, would love to integrate Jules into it, where should i start? SDK or API? https://github.com/AsyncFuncAI/AsyncReview

[serene fern]

Jules has API (but not that robust, but Jules doesn't have Any official SDK)

[serene fern]

What makes it different? Code Review can be done by Claude Code/Gemini CLI/OpenCode

[plain dirge]

yes this current G3 Preview is soooooooooooooooooooooooooooooooo lazy

[serene fern]

This is the reason of Jules being slow. It doesn't generates any reason when get a message and this is not a model problem. I request everyone to bring this issue to the developers to actually generate response while user sends a follow up prompt

[serene fern]

I have done something, now see Jules reply 💀

[shy cedar]

Guys, I’ve been working on the backend of the app for the last two weeks. I’ve now completed the core skeleton and am ready to move on to the frontend. What is the best MCP to integrate with Antigravity or a cool way to build a more sophisticated UI not just a simple interface so I can publish it to users and gather meaningful feedback?

[serene fern]

What app? for android? Jules? I am thinking to working on Android Native Version of Jules. Cuz it won't lag on native android than web pwa

[shy cedar]

My apologies I didn’t mean a mobile app I was referring to a web-based interface where I can ship my tool which combines a chatbot with interactive dashboards.

[serene fern]

if you already have made it then please share the repository link and if you have not finished yet, I would like to say, there are already some backend SDK available in the market why create a new one?

[rocky forge]

Pick a UI library and ask antigravity to build a design system based on your vision and needs.

Before using the design system, ask it create a mockup UI. Use the mockup UI to critique the design system and refine it.

Once you are happy with it, use it create the UI for the first screen. Refine until you are happy and then keep it consistent using the design system spec.

[serene fern]

means?

[prime flicker]

guys is it just me are jules is getting laggier the further in the building of the like he takes 20 min juste to modify 10 lines is there a way to like debloat him ?

[swift karma]

make jules open-source

Let's annoy the devs and make them make it open-source

[spark sonnet]

I have noticed that somewhat recently there is such thing as 'automated reviewer' that not only works when Jules asks for code review, but then Jules uses other tools as well, especially at set_plan tool, and Jules is really willing to set plans before it does something, but like the automated reviewer can just reject any plans Jules makes even if properly formatted if it (out of context of course, only given the initial user message) thinks that the plan is incorrect. So Jules may spend a lot of time or just freeze trying in the loop to set a proper plan. Sometimes Jules begins to complain to the user how difficult it is to set a plan, other times Jules sets a vague plan and promises to actually do this and that, and sometimes Jules just gets stuck for really long, and cannot continue

[frigid marten]

They suffer from context rot, RLM fixed it and more effective deal with PR context recursively. they do have sdk--> https://github.com/google-labs-code/jules-sdk

[outer ibex]

looks like for me right now jules (g3) is slower than ever but its stable and 90% usable (functional?) codes

[scarlet musk]

hi i cant post in #1374125660859142254

i have a finished pr which can not be created to github via jules, any tips?

[wooden mantle]

Hello Support Team,
I can’t post in Jules bugs.

I am writing because Jules is not working for me despite having an active Google AI Pro subscription.
I am located in Kyiv, Ukraine, and I am over 18 years old. Could you please check if there are any technical issues or regional restrictions affecting my account?
Thank you for your help.

[rapid zephyr]

Hello for last 5 days I can not get into Jules via user interface or CLI or web api - problem is that I reached limit of 15 sessions opened concurrently - I want to close them all but I can not as i can not interact with Jules - problem is on account [e-mail]; in case of CLI and API I get error 503

[outer ibex]

i saw at the bottom its unofficial. i might use this to build something else and/or porting to another language

[uneven iron]

Well, Gemini 2.5 started bugging as well,

I asked it to to run my react portal to check if everything was alright > it ran it and everything was fine > asked it to apply the new design I had done > after it finished the coding and verification it backtracked to the first task that was running the project

[ashen glen]

Anyone have any issues with Gemini code review? Not showing up in my PRs

[swift karma]

Hallucinations

We're more and more close to Jules officially being diagnosed with dementia, alzheimer and adhd

[uneven iron]

Bridging the gap with being human

[outer ibex]

just need more emotions then its relatable

[uneven iron]

I am pretty sure I made it frustrated once with an error and it basically gave up after 20 minutes

[outer ibex]

i still remember back then researches found that early models can be a sentient beings, but fast forward a few years we got ai girlfriends

i wont be surprised if ai started having cancer and stroke

[visual gyro]

Need better system

[swift karma]

I can see that the devs are active somehow due to the CLI, focus on fixing Jules prompting lol!

[leaden pawn]

Hello Everyone! I see there are few issues reported in this group, please open a thread in ⁠#labs-bugs-feedback channel with as much details as possible so that we can investigate. Thank you!

[low reef]

For me it seems like i'm unable to create new posts in this forum.

I feel like I got a weird one. 😵‍💫 PRs generated by Google Jules are being submitted under my personal GitHub profile instead of the google-labs-jules bot. Happened with the last 3 PRs, while hundreds of other PRs were created under the bots profile before.

While I think this can be nice in some scenarios like triggering automated code reviews without extra steps, I have no idea why this happens and how to control it.

[prisma vine]

Hello our #1238621511401734254 channel is the best channel to post your feedback, along with any issues you’d like to report. When creating your thread please post as much as possible so that the team can properly investigate. Thank you !

[minor garnet]

hi how can i make many with ai if someone got some advice

[low reef]

Would love to! As mentioned in my initial message, I do not have permissions to post there. Maybe because my account is still considered "new"?

[minor garnet]

money

[prisma vine]

For us to better help you, please send a message with a screenshot of the error message (if any) you’re receiving when trying to create a new thread in the #1238621511401734254 channel to @barren haven and follow the instructions that the bot gives you! (he'll give you options on which server to contact if its in multiple servers)

[leaden pawn]

Hello @rapid zephyr What is the exact issue? Could you please open a thread in #1238621511401734254 channel with more details. Thank you

[formal tusk]

we should have a happy update to ya'll soon around Gemini 2.5 pro so stay tuned. ⚡

[formal tusk]

can you confirm where you're located? Jules is only available where the Gemini API is so if you're in a region where things are limited that could be why

[formal tusk]

hey yeah Ukraine is unfortunately one of our limited regions :(.

[formal tusk]

you're telling us...

[formal tusk]

Yet*

[swift karma]

Don't take this personally but it feels like users are finding out more about jules than the devs are recently 😂

Embrace the community and Jules will be loved again

[random iris]

Hey gang, is there a feature request/ticket tracking IP Whitelisting?
Would love to use Jules against my github org

[spark sonnet]

Jules doesn't show anymore the output of commands... Such as when Jules runs ls in bash session, it claims to have a listing, but I can't see the command and the output

[outer ibex]

hmm im watching👀

[rapid zephyr]

I had added the post there couple days ago ; problem is that i had hit limit of 15 concurrent requests about 5 days ago and since then jusles is completely stuck and i can not interact with it to stop the active sessions (want to stop all now)

[rapid zephyr]

Poland; on other account (also in Poland) it works just fine

[spice fossil]

Why so?

[arctic fern]

Yeah I have the same problem
I can't use Jules this morning

[torpid sinew]

hello everyone

[outer ibex]

hello

"treat it securely" im getting fakehope from this😂

[swift karma]

I wonder why file upload mid session isn't still a thing if i can literally trow files there if i connect to it's machine

[outer ibex]

yea tho

like i also wondered why only pictures are allowed to upload in the chat, or rather send the files inside a folder the upload them

i just want to send a basic txt or md file, aint no way im creating a new folder just to upload them, or upload it on repo then mention it in the chat

jules team, allow uploading text files in the jules chat please🥺

[spark sonnet]

Jules used to slack and dodge work. The first fix was a Critic that does code review. Never really helped — no context awareness, lots of false claims. Jules would crash its own web UI. After many months, pagination finally exists and it no longer crashes. Credit where it’s due: that did get fixed — slowly, manually, and not by AI.

But the recent “improvements”...

Something else Critic-like seems to have been added to process messages. Hello random resets.
Snapshots arrived… but not user-controlled — created automatically when Jules submits a PR. Hello random reverts to the last incorrect PR proposed.
Even environment snapshots randomly disappear, or revert to earlier ones. 😆
A Critic-style AI was added to setting plans. Checks Jules's plans against first message in a session (or first after a submission). Hello no-plans tasks...

---

Somehow stopped showing actual commands and outputs. Which is impressive, because that was the last useful way to understand what the agent was doing and try to correct it.

Now when you give a task, files just… change. Why? For what reason? Correct or not? Hard to tell. Process over outcome. Do tasks still get solved? Formally — because the now overly cautious agent undoes all changes before requesting Code Review, reapplies them afterward, and presents them as “review fixes.” No second review request, since it already happened. Solved 👍 .

Sometimes no plans anymore — avoids setting them to dodge rejection. Just does changes before planning, then undoes, sets a plan, and reapplies gradually. Barely explains what it’s doing. Doesn’t really listen to the user either. Mostly it just burns cycles. Slowly. Indefinitely. Because the control mechanisms added to fix it ended up being spectacularly counterproductive.

[outer ibex]

if its complex tasks, yea the problem still there and got worsen

[spark sonnet]

Jules used to be able to solve complex tasks. Like coding compilers, debugging rendering software and - the last one I was in the middle - porting Rust's std to a new OS

[outer ibex]

i know, thats why i said got worsen

right now one of my session just stuck forever, cant be opened

finished the tasks, missed some of the files on commit changes, cant be fixed

tons of same problems since last year, so just hope and wait

[swift karma]

Let's trow eggs on the developers

Now seriously, it's been a time since Jules last behaved correctly. The devs have to lock in

[hearty glade]

Im not so sure that everyone is having the same issue.. I dont want to tempt fate but Jules has been working nicely for me for a bit now 🤞 Pro plan and working on a project that is mainly HTML Python and JS if that helps at all. I also ask Jules to carefully plan phases for larger tasks which can be worked on by multiple instances.

[swift karma]

Jules been exterminating my productivity, look at it telling it ran the commands without doing it

[hearty glade]

hmm thats not good.. 😬 I have no idea how to fix but it might be worth leaving a detailed report in #1374125660859142254 ? include plan level and Gemini version too i would think 🤷‍♂️

[swift karma]

That's what i've been doing for a whilee 😂

[hearty glade]

oh.. 🙁

[swift karma]

Antigravity peforms way better than it locally, i don't know if it is system prompting issue. You write a giant prompt and it still does totally ignore what you ask

[swift karma]

Gotta have some sense of humor those days

[rigid flicker]

😳

If this helped, I am shocked xD

[prisma vine]

Hi @swift karma and @hearty glade wanted to just say that our #1238621511401734254 channel is now the central location for reporting any issues or to give feedback related to all Google Lab products. So please be sure to create a thread there with as much details as possible so that the Labs team can properly recreate and investigate the issues.

Thank you

[halcyon wagon]

I’m building a PWA.
Do you guys have any cool ideas, badges, or small UX tricks to show users that it’s browser-based & safer (no shady APK / EXE installs 😅)?
These days the internet is full of sketchy apps, so users hesitate to install anything via Chrome and prefer only Play Store apps.
But PWAs are actually much safer, I’m just not sure how to convey this clearly to users.

[swift karma]

It just made it worse to be honest

[cosmic bay]

I built an android apk for my little project that runs the PWA inside the apk, so it's just an apk wrapper (not sure if thats allowed on the play store if you plan to put it there) but it's nice cause you open the APK (and using a cachebuster) it always has the latest version of the app -- I havent made it do an offline cache thing yet though so you can run it offline.

[formal tusk]

yeah usually what helps if you provide the session url and make sure you don't delete the session. Whenever a user deletes the session then we can't look at it in greater detail. Some of the bugs are reliability issues that we're triaging weekly and then another part of the bug issues are quality related. Unfortunately a lot of these aren't quick fixes but we're def working on it.

This may help for some of the issues you folks are running into. But a couple of new drops:

1. A new secondary agent -- the planning critic. The result is a 9.5% reduction in task failure rates and significantly higher quality execution paths. https://jules.google/docs/changelog/2026-01-26-1

2. Performance optimisations are being surfaced in your code. https://jules.google/docs/changelog/2026-01-26-3

3. If you're still testing Jules w/o a repo, hellooo to the repoless support in the REST API https://jules.google/docs/changelog/2026-01-26-4

4. Some quality of life updates regarding scheduled tasks bc we know it's been annoying. https://jules.google/docs/changelog/2026-01-26-5

David will probably make the prettier announcement but wanted to share to this crew for now

[cosmic bay]

Jules CLI /new command stuck loading repos for anyone else? My repos show up fine on jules web (same gmail account) - -- def the right account cause the initial jules page shows info on my repos

[halcyon wagon]

This is seriously unfair.
Antigravity users are getting huge advantages with multiple model options like Opus 4.5, while Jules users are completely left out.
Google, we all paid equally. There is no difference in subscription value, yet one group gets premium models and the other doesn’t.
The only reason many of us are on Jules is because our devices DO NOT support Antigravity 😭, not by choice.
So what are we supposed to do?
Be punished for our hardware limitations?
If Antigravity users can access Opus 4.5 and other advanced models, Jules users MUST get the same models. Equal payment should mean equal access, not artificial platform discrimination.
Please stop locking powerful models behind device-specific apps.
Give Jules users those models too.
We’re tired of being second-class users for no valid reason.
This needs to be fixed. Now.

[visual gyro]

Can mcp stitch project

[swift karma]

Tho Jules has no message limits per session and it's own VM...

[halcyon wagon]

If you’ve actually used Jules, you’d know that after some point it literally stops replying. That clearly means there is a limit, just not transparently shown.
And honestly, I don’t see any real advantage of its so-called “own VM”. i think Google AI Studio works faster, responds more consistently, and is just as capable as Jules without any VM (but it only codes in react and can't import repo 😢)

[cosmic bay]

I use jules..and antigravity -- i send big tasks to jules, then pull it down into antigravity, use opus etc... then push to my feat branch or wtv, then send it to jules again etc... why not just use everything 😉

[drifting dagger]

Can somebody please explain me what does the inative tag means on a scheduled task?

[echo lark]

hi everybody! i have Google Ultra package but Jules ask me to update for Ultra. and i have 2.5 gemini only. How to fix that?

[dry raptor]

Just means it’s not currently running / hasn’t ran yet. It’ll still run at your selected time.

[dry raptor]

Have you been a Trusted Tester before for any Google app? Sometimes they mess up permissions.

[devout mesa]

Hmm anyone else seeing Jules respond to PRs on github by creating a new PR with a slightly different name? (starts with feature-featurename, and then goes to feature/featurename for branch). I reported this a few months back, and sent some examples, but it still seems to be doing it, which is rather frustrating. It means that you end up with 2 different branches, one with all the main changes, and the other with the "fixes" and then if you aren't careful you end up having some merge cleanup to do to fix things

[brisk hamlet]

Welcome back, most nonsensical and unhelpful error message.

[halcyon wagon]

@swift karma "jules has no message limits" 😂

[swift karma]

Welp as stated by the image and the person who posted the image, this is an error

[hearty glade]

You might of tried it already but... have you tried replying with something like "What went wrong? Please review your progress in this session and try again" that kind of reply often sorts it out for me but be carefull.

[halcyon wagon]

Yeah, I get that it’s an error. My point was just that such errors usually start showing up once the session gets too long, even if it’s not officially called a “message limit"

[swift karma]

Didn't invalidate what you said earlier. Let me tell you something: Google will probably remove those "extra" models from AG once they get what they want: Users. And Jules will either become 100% locked behind a plan OR it will be just... Killed because of how expensive and unstable this is. We're just training Google's models so it learns how to code. I don't think Jules would get access to other models, would be very expensive for google.

They're already reducing their quota because they're struggling with resources, Antigravity users are always complaining about it now, do not expect much for too long. You'll be dissapointed

[cosmic bay]

Yep, this is Google LABS -- labs imo are just experiments, and you should be happy they give em to us and that jules has lasted this long

like the Google Glasses etc... this can be shut down any time they want, so enjoy it while you have it! (And seriously dont try to run your business relying on it, cause there is no way they keep this going forever!)

[halcyon wagon]

If compatively smaller platforms like Cursor can afford to offer multiple models at the same price, Google definitely can.
This feels less like a technical limit and more like cost-cutting.

[swift karma]

They can but they won't, they just made Jules free to train Gemini with code and you can see how "well" jules is peforming recently lol. Imagine how much google would pay for the many request a Claude model would spend on a long task

[visual gyro]

Anyone clawbot? But fore jules

[brisk hamlet]

well, we are coming towards the point, where the bottleneck is defenitively not the llm itself, but rather the ridicolous bad interface and unwillingness of the devs to adhere to very basic QA cycles. so my hope is: even if gemini 3.5 comes out, we wouldb be very fine to run jules with gemini 3 but on a smarter interface and a professional managed platform - with reasonable costs for google.
nobody would really complain if we were restricted to 50 sessions a day or so, because we are hard limited by the bad interface.

[halcyon wagon]

See Cursor, it’s far better and faster than Jules.
It offers many model options, and yes, it also has that so-called VM feature.
The difference?
Cursor’s limits end in 5–6 sessions, but at least it works properly.
Jules gives 100 sessions, but you easily lose 50 of them to annoying errors, that’s wasted time.
Half the session is spent fixing issues instead of actually building or testing anything.
Speed matters. Reliability matters.
Right now, Jules is just slow, time-taking, and frustrating.

[swift karma]

There's more criticism we can do about Jules than compliments oh my god 😂

We're locked in two different models, that's all we got

[halcyon wagon]

Sorry if this sounds harsh, I’m just provoking Google devs on purpose.
Because honestly, Google should learn something from Cursor (a much smaller team than Google) and actually give us something better.
Same with Anthropic, again, a smaller team, yet consistently delivering strong, usable models.
I had too much hope from Gemini 3 Pro after all the hype… but in reality it just feels like a slightly upgraded 2.5 Pro. A few improvements sure, but it’s still slow.
With Google’s resources, this shouldn’t be the outcome.
At this point, maybe Google should just acquire Anthropic and Cursor 😅 and ship something that actually matches the hype.

[halcyon wagon]

remember this fact: we’re still on Gemini 3 Pro preview.
The model isn’t even fully released yet.
So how exactly are we supposed to expect or talk about Gemini 3.5 when 3 Pro itself hasn’t been completed and shipped properly?

[brisk hamlet]

the issue is not gemini 3 pro. It is a very capable and smart model. Jules only uses that. The guys working on gemini 3 are way way above the pay grade compared to Jules devs.
The problem is the Jules project management.
Example: Why *on earth have we all of a sudden lost the ability to see which commands Jules did execute? This is one of the most ridicolous downgrades, we could not even imagine this bs would happen.
BRING IT BACK. IMMEDIATELY. PLEASE!!!

[eternal vault]

-# This is my own personal opinion
Hasn't Google historically always been the one to innovate and then just, uhm... abandon the projects? And, is it so surprising that it's not cutting edge due to that history? Sure they could just acquire companies that do it better than themselves, but is that truly the goal? I mean, on a more practical level, they have to worry about a lot of legal and political stuff, too, like monopoly. So, I wonder if they don't just build for the sake of building, the experience, understanding, and maybe, maybe on occasion, they strike gold and stick with it? So, there's that to hope for, I suppose. Again, just my personal opinion on that matter.

[dusty cosmos]

I feel i'm one of the few folks that has had an ok time with Jules, but i've written most of the prompts using ChatGPT or Claude to get them ironed out before letting Jules loose on it. It has been weird that I can't configure branch names, that it constantly reposts the same commit message for updates, doesn't follow the AGENTS.md file instructions, doesn't show what it is doing, likes to rely on old 2024 training data vs. looking up new things, and more. But i've got most of what I want done without abandoning things.

Recently they've started co-authoring commits - which is kinda cool but I wish it was a setting. There's a weird lack of polish, and I bet the Google Engineers are kinda stuck between trying to make something awesome and whatever whims management has for them.

The concepts in Jules are really cool, and the code is decent when it understands what you want!

[halcyon wagon]

Alright, let’s stop the discussion here 😅
Otherwise, the Jules team might commit suicide or just kick us out of the Discord.

[brisk hamlet]

no, i will keep saying at least two things: 1. get rid of non-verbose error messages. make them meaningful. 2. we MUST be able to see the commands Jules did execute and what the results have been. that's the bare minimum.

[swift karma]

Only way to get xp

[night oriole]

To be clear, these are vastly different products. Cursor is more like antigravity. It's an ide replacement. Jules is much more autonomous.

And there is not a message limit, but there is a context limit.

I generally try to clear a task and move on to a new session. Context loses details in the middle. Much like people. The vast majority of issues people run into is pushing the context to its absolute limit.

In CLI tooling, I try not to go past 20% context utilization. Gemini, Claude, they all work the same way with context

Gemini 3 flash is quite good, different learning methodology, and it's fast. 3 flash is a heavy lifter for sure. I mostly use 3 pro but flash is quicker.

[cosmic bay]

Just wanted to paste this from the official Google Labs website for the angry users that use Jules as their only way to make money in their small businesses: (And complain when things don't go their way!)

Google Labs is where you can discover and try Google's latest AI experimental products and help shape the future of AI technology. Get a first look at AI experiments from across the company, give feedback directly to the teams who built them, and help us responsibly scale these technologies into the products you use everyday.

[swift karma]

If the devs helps us we can help them, i can see the devs are active somehow but there are many common issues and complaints that aren't still resolved!

[halcyon wagon]

Cursor has both an IDE and a proper web version that works almost exactly like Jules, it also uses VM. It can commit and create branches, open PRs directly on GitHub.
Please actually try Cursor once you’ll instantly realize how far behind your current Jules is in comparison.

😂 You don’t even know your competitors properly, so how can we expect better from you?

Also, the screenshot I sent earlier clearly showed it was from Cursor’s website and phone. And yes, just like Jules runs on phone, Cursor does too.

[halcyon wagon]

Yes, Cursor IDE is more like Antigravity, but Cursor Web is literally Jules’ father 😂

[outer ibex]

cursor drained my wallet faster for the same task ive been doing as a test on other services

i appreciate google ecosystem where you paid one subscription but can use multiple services in it

if i hit the limit on ag, i can just use jules while waiting for it to clear cd. for planning ill just use notebooklm or rather standard gemini web chat

[spark sonnet]

Jules is much more autonomous.
I get it that the main feature of Jules is to try to be as autonomous as possible. However, currently it is also it's Achilles' heel. I don't think that requiring constant human attention would be better either. But it's a well-known wisdom that 'Only a Sith deals in absolutes' , so it would've been much better if no one was trying to fine-tune Jules to exclude any possibility of control and correction by human. Then maybe it would actually be better than Cursor because it would be indeed more autonomous while providing the same power - not autonomous instead of it.

And also, there's such a thing. If all the competitors are less autonomous, it's not only because Google Labs are only ones who came up with a genius idea that no one else came up with. It's also because the demand on the market dictates that the users expect some level of control & correction over agents, and agent autonomy should not come at cost of the user expectations, unless the competition is destined to be lost

[visual gyro]

Task not working for me anyone else?

[halcyon wagon]

Let’s be honest here.
Cursor clearly copied the Jules idea and launched their own web-based agent dashboard after that. Jules came first, no debate.
But the painful part is this: Cursor’s web agent is currently doing a better job than Jules.
Faster responses, fewer random failures, smoother UX, and it actually feels reliable.
Google, we love you for innovating first and bringing ideas like Jules. That’s why we’re here.
But please don’t stop at “first.” Finish the race.
Right now it honestly feels like Cursor took your idea and executed it better.
That shouldn’t be happening.
Please improve Jules, optimize the VM, reduce the silent session deaths, and make it clearly better than f_cking Cursor, not the other way around.
We want to stay with Google, just give us a reason.

[outer ibex]

i wonder what happens if jules try to fix its own source code

gonna go crazy fr

[onyx tulip]

Hi! I couldn't find anywhere else to ask, so I'll do it here:
How do I request that Moldova be added to the list of countries where Jules works? Unfortunately, I'm currently getting the error: "Jules is not yet available in your region." But I couldn't find a list of countries either.
Yesterday, Moldova was added to the list of countries where the Gemini API and AI Studio are available. Perhaps this was previously preventing Jules from being activated in our country?

[spark sonnet]

Somehow I think it was already attempted... I may be wrong though

[halcyon wagon]

Use VPN

[onyx tulip]

VPN does not allow you to use the benefits of a PRO subscription (

[outer ibex]

region is tied to the google account

[wet knoll]

How does the CI Fixer work? It doesn't pick up on failed CI builds. Are there any configuration needed?

[swift karma]

#1465869429027110933

[outer ibex]

gg as always

[swift karma]

Two weeks to get it fixed

[quartz tusk]

I feel like Jules team needs to go back to the drawing board. Interestingly it was much more stable late last year. Some of the observations: The sessions now seem to be more incremental, and only lapses after 24hrs, the sandboxes seem more unstable. Even if Jules is working smetimes you do not get timely messages or responses. It is just too slow to try and do anything constructive as you do not know when it will respond or how long it will take to respond. However on the tasks tha it has completed, it does dot a descent job depending on the model, and your prompt. But right now it is just not worth the time

[brisk hamlet]

What HAS been improved: The speed and reliability of the web-interface. It now streams much smarter and does not freeze anymore. That is a great improvement. Thanks.
But the new (for two days) total anti-verbosity is a real blocker. We cannot see anymore which commands have been executed. This is very very bad.

[serene fern]

Suggested: https://github.com/TheRealAshik/jules-mcp

[cosmic bay]

As someone who's still learning about coding, and just vibecodes, i have no idea what your tool does, the description on your github page is like reading another language to me, and I have no reason to use it as i dont know what it is

it's like giving me an F1 car, and telling me to figure out how to start it up, and handle it well around a track when ive never been in a race car before - with no instructions or training

im sure it's really useful and awesome, but a bit more context on what it actually does that makes sense to a vibe coder is appreciated please!

[serene fern]

It allows "Antigravity" to control "Jules".

How I use it?

I deeply analyse my requirements and codebase with "AG Gemini 3 Flash/Opus 4.5 (for complex)" and tell the "Antigravity" to start a Jules task. "Antigravity autonomously starts a Jules session with detailed prompting...

As you know, Better, detailed prompt = better output

I am requesting @elfin kite to add more endpoints/Activities in Jules API. API is being outdated with new offical updates

[cosmic bay]

Does this let me prompt inside Antigravity to chat with Jules. As if I was on the Jules website chatting with Jules? Kind of like a better version of the CLI

[visual gyro]

Have Jules failing so many Task lately

[swift karma]

Pay for product ≠ product works, i'm tired of waiting for fixes and improvements all the time already

[minor garnet]

😏

[hybrid sequoia]

Does jules ever get stuck on "Pre-commit checks"

trying out jules for the first time

basic refactor, don't know if I did something wrong, but this was just a personal repo, there is no pre commit checks that it should be running and it's been on that for a while

[ornate oasis]

jules error?
Jules encountered an error when working on the task.

[spark sonnet]

We cannot see anymore which commands have been executed
This is indeed worst change that could have happened to Jules, because the model is prone to lying about what it does, even if you demand it providing you with regular verbose updates, and it tends to bypass Code Review because it knows that if it undoes some changes, requests code review, and reapplies changes, it would count as 'fixing'

[light marsh]

I had this so often, most of the times it was so critical of its own changes, that it redid the whole work in that phase

[worthy trout]

Jules encountered an error when working on the task. Every second task in a session jules has failed it 3 sessions in a row.

[brazen crystal]

this has been my Antigravity experience all day. Jules has been a good alternative and it works on the go. Would be good if you merge a PR then notice a change and prompt Jules again for it to open a new PR like Codex Web does

[swift karma]

You're using Windows right?

[brazen crystal]

yes

[swift karma]

Yeah windows has that chronical problem of infinite loading

[outer ibex]

i never had this problem

hmm

[brisk hamlet]

Dear Devs, we need the verbose "ran xyz" with the verbatim output as control. Jules has been halluzinating way too often about the results (about 20 %). It's a downgrade which I have absolutely zero tolerance for.

[swift karma]

So you're telling me the devs did that?

First you guys remove the delete session button from the menu, now remove something so basic? Honestly, you're just ruining Jules with no valid reason at all

I wonder who has those ideas... "Let's remove the ability of the users to see the commands so they can't know if Jules is hallucinating or not!"

[brisk hamlet]

It really feels like a reward hacking trick: The users complain about command-spamming and output-hallucinations - what do I do? Hide the commands and the output. Prolem solved.

[swift karma]

Just archive Jules at this point, it's not even being a useful tool anymore

[brisk hamlet]

I wonder if the code-review is able to see the output. If not, it is utterly useless since Jules can tell it whatever it wants.

[swift karma]

It can see the diff prob.

[brisk hamlet]

that for sure - but the point is: when things get complex, only the test result or the truth of "does it compile?" is reliable.

[rapid zephyr]

Poland

[minor garnet]

hi

[halcyon wagon]

Hlw

[hybrid sequoia]

it's odd to me that jules seems to be struggling with such a basic task I gave it

given that gemini 3 is probably the second best model out there

[brisk hamlet]

here is my hack. just add it to your prompt, or AGENTS.md or memory:
"
Verification Visibility Protocol: To ensure transparency and prevent hallucination, the agent must redirect output from all critical verification steps (compilation, tests, benchmarks) to a log file. This file must be dynamically filtered (e.g., tail, grep) to capture essential success/failure evidence without spamming context. The filtered artifact must be committed to the repository. A session is INVALID if critical verification logs are missing.
"

[visual gyro]

Alot failed task lately?

[onyx tulip]

Hi! Any plans to launch the service in Moldova? I'd really like to try it! Thanks!
PS Gemini API available

[spark sonnet]

In my experience Core Review never was able to see the commands output. It sees what it thinks to be the user message for setting the task, and the patch Jules provides. Code Review often hallucinates itself about code being non-working even if it was just successfully executed by Jules a turn ago - so it clearly can't see commands ran.

[spark sonnet]

Already tried this. But unlike log of ran commands and their outputs that the model couldn't hide before, this the model still could and would hide.

1. It just forgets that this memory exists / it doesn't read or ignores AGENTS.md. Happened so often that I can't even count. Basically, user-provided instructions barely exist for Jules unless the user begins screaming sheets of angry paste with threats to archive the task in bold caps. Even if typed in the beginning of the initial prompt, Jules may still deviate.
2. By such an instruction, Jules may create a log on the first ran of a command, and then 'forgets' to do the log on subsequent runs.
3. As there is no control over how Jules makes these logs, it can just creatively write a log file instead of really redirecting output.

[brisk hamlet]

that's all correct. It's not a great solution at all, but a hacky workaround - i just catched Jules faking a "successful" compilation - but the log was truthful.

now we have an even more severe bug: Jules submits empty commits to a existing PR, the interface looks still normal (with all files there!) but the branch gets corrupt now. Happend twice, in two different repos with two very different commits. "/gemini review" on github denied to review that because the file-type is not supported. The "Files changed" is now empty. Very very poor, Jules.

[swift karma]

It's so ironic that the users have to go to that level to have an sucessfull task

[brisk hamlet]

yes, now Jules is making empty commits which look non-empty. In fact it zeros the branch in a way that not even githubs gemini code assist will do a PR review. Indeed very professional 😄

[spark sonnet]

I speculate that this happens because Jules is 'afraid' of critical feedback from Code Review, so sometimes Jules undoes all changes (to hide them) before Code Review intending to reapply them after it would be critical and submit it as 'fixes', but if Code Review is actually happens to be not critical, then Jules has to submit what was submitted for review, and 'forgets' to reapply changes..

[brisk hamlet]

the spooky thing is that it resetted to ZERO compared to main after multiple pin-pong sessions of submit and review inside one session (which worked very nicely for about 10 review cycles the last days) - and the files-compartment looked normal (non-empty!). this is ridicolous.

[serene fern]

Jules 💯 success rate from last few months, i never faced any failures, Thanks Jules Team

[swift karma]

Witchcraft?

[serene fern]

???

[spark sonnet]

Still, worth trying to ask Jules to reapply these changes as it definitely has them stashed somewhere. It will pretend that it does it again from scratch though and would set the plan as if it was redoing, but actually it would just pop from the stash

[swift karma]

Jules is not working like it should, unless you're doing some crazy work with the API

[brisk hamlet]

I lost any trust in Jules now. This is so ridicolous, since it is a combined interface bug which corrupted an entire branch. i now have to forensically revert the anti-commits from jules to a point where it still was ok

[spark sonnet]

By the way, how would I know about Jules having a stash where it can hide it's changes? Because sometimes it pops these changes before setting the plan during which it would pretend to reapply them

[serene fern]

I am seeing, alots of peoples are complaining about task failures recently. But I never faced anything like that or any problem, even after the update. It significantly increased my workflows ❤️. Maybe Jules server load is less in my region, IDK but best

[brisk hamlet]

is anybody able to recreate this massive f-up? Jules committing a literal zero PR, despite showing files, a nice commit message, etc.? just look to your PR and add ".diff" to your github web address (or see: "Files changed") - for me it is zero again.

[frank cedar]

confirm - 15ish session in total - 10 are from last 2 hours - the other 5 from tuesday where a tiny correction today zero-ed all changes

-> Showing 0 changed files with 0 additions and 0 deletions.

[brisk hamlet]

yes. and the jules interface still looks normal, right?

[frank cedar]

also if you download the zip it is normal

[brisk hamlet]

then something with their github-connection is broken.

[frank cedar]

maybe some rights changed or github is not liking jules anymore

[brisk hamlet]

which is very reasonable I guess

[frank cedar]

@brisk hamlet created in the labs-bugs-feedback channels a message

[spark sonnet]

Could be, due to Jules submitting unverified commits. In past, Jules was only signing commits on it's own behalf. Recently there was a change: now it's 'co-authoring' commits with the user that uses it. However, there's a catch: Jules can't sign a commit for the user. So from GitHub's perspective, all Jules commits are now 'unverified'...

[brisk hamlet]

yeah, still zero qa at the jules team. just commit any random idea without testing the consequences.

[spark sonnet]

They are testing on users, apparently. And not just on volunteers

[brisk hamlet]

and wasting everybody's resources

[versed lynx]

Resolving merge conflicts is impossible because Jules doesn’t have access to origin and update local refs

[frank cedar]

think that is it jules only or jules&user work

[visual gyro]

You Lucky

Not happy

[visual gyro]

Close application kill the task and restart

[swift karma]

Jules isn't even trained to use git, it just knows how to run verifications and edit files

watch #1466473538780266556 be totally ignored

[spark sonnet]

I can tell why it's not trained to use git. Back when it still shown how it ran commands, one could have told it to use git commands in bash session, and it would comply, but there's a catch... Everything git-related in the project working directory gets reset every turn, so no changes to index persist, not even possible to compare index to HEAD (always empty diff) and so on. So Jules is not trained to use git just because the environment it works with is tuned in a way that breaks everything git inside

[spark sonnet]

I heard that every issue where devs are not tagged don't get any attention..

[swift karma]

That staff called Srujana is the one responsible for reporting, i'm not really going to keep pinging the devs to be ignored + warned by the mods

[halcyon wagon]

I had faced the same issue a few days back, almost a week ago. When I reported it, they replied that they were not able to reproduce the issue on their side while trying to investigate it.

Currently, it’s working fine in my case.

[leaden pawn]

Please note that I am reporting all the issues and our team is working on. Please report in #1238621511401734254 channel so that your issue don't get lost in the chats, we really appreciate your patience

[leaden pawn]

Hey Everyone! Please follow https://x.com/julesagent/ or https://jules.google/docs/changelog/ for updates and share us the feedback

Jules (@julesagent) on X

Your new autonomous coding agent by @GoogleLabs. Built for devs, open to feedback, evolving with you. Dive in → https://t.co/iIzFEMmoqX

X

[urban onyx]

I recently started a blank project without GitHub. The agent has been working and has completed the job, but I am surprised to find that the button to download the code in Zip format, which I used to do before, is not there.

[spark sonnet]

Typical Jules.
Jules: - I have almost succeeded, look! This works and this works, however this I do not know. Should we do this or that? (Waits for user input)
User: - Add just this and let's save this.
Jules: (Resets the environment and sets new plan, nothing works again)

[indigo wasp]

👋 new here, are google workspace users with ultra subscriptions limited to Gemini 2.5 in Jules ? Clicking the uppgrade link to the docs and it says you need a @gmail.com account but my workspace uses my company domain name.

[swift karma]

Yeah, i don't know why this is still a thing but workspace users are stuck on the free tier for some reason

[daring walrus]

Hi @ Jules Team, How can i remove my current GitHub Account on jules, especially if i want to switch to another GitHub account, there definitely no option to remove current GitHub Account.

[shy cipher]

Discount Julie's permission from gethub settings

[shy cipher]

Because your github not connected that time

[outer ibex]

finally someone else had the same issue ive been having since 3 months ago

[outer ibex]

they are working on support for workspace account, otherwise its only @gmail personal account only

you can request for access using the forms link tho (i dont believe it but whatever)

[daring walrus]

thanks

[ornate oasis]

wheres the result? just text?

[ornate oasis]

same here, i think bug

[outer ibex]

yea i think its a bug for some people

[ornate oasis]

this new program?

[outer ibex]

huh ?

[jagged hull]

Antigravity is Down?

[outer ibex]

no ?

[shy cipher]

Welcome

[serene fern]

@elfin kite

Currently, How Jules is building app with gradle without configuring Android SDK? Does it comes pre-installed now? Why not mentioned in (Pre-installed Docs)

[swift karma]

Uh who pinged me

[serene fern]

msg sent failed

[swift karma]

Ok

[shy cipher]

What

[serene fern]

This can fix jules lazzy ness, so raise it to the devs: (Jules Skills Implement)

@swift karma share it @outer ibex
see top issue of #1238621511401734254

[serene fern]

@shy cipher share this to everyone to raise this Feature Request

[visual gyro]

Can we just push all thinking update jules

[serene fern]

Wohooooooo

@swift karma @shy cipher @outer ibex See updates

[swift karma]

Woah

[spark sonnet]

Seems updates finally begin to be positive!

[swift karma]

I hope so

[outer ibex]

yea i saw them, finally 3 flash can do real hard code

[serene fern]

@outer ibex Now free user only get "Gemini 3 Flash" and paid tier users gets "Pro"

[serene fern]

is 3 Flash better than 2.5 Pro?

[outer ibex]

yes, 3 flash is better than 2.5 pro AND beat the 3 pro

[serene fern]

So according to you Flash > Pro 🙂

Un-digestive statement. But I will test it now

[outer ibex]

because pro is better for deep research and thinking, flash can output faster with more functional code

even the google staff says so

i ran multiple tests in AG (examples for now), and 3 flash consistently produced better output than 3 pro

3 pro cant do proper functions, but it is superior on ui designing

[serene fern]

I suggested them in the AMA to being flash for faster replying but they refused but now the brought it. I wish it could be changed inside a session instead of Defining at once

[outer ibex]

i know right ? cant change the models in sessions is seriously a downer, not efficient too

but we will see maybe in future updates hopefully soon for other improvements (pls jules team)

[serene fern]

I am going to create a issue now

I am requesting everyone in the server to REACT fire in the issue, so it becomes notiable to devs. go to #1238621511401734254

Things we want

1. On Demand Model selection in sessions
2. Skills (Inspired by Antigravity)
3. Custom MCP Configuration

[spark sonnet]

😶‍🌫️ It seems now with 3 Flash Jules actually remembers these memories and a lot of context. I'll have to go manually delete since many memories became irrelevant (I didn't bother to delete before since on 2.5 Jules never actually remembered)

[night oriole]

Flash benches better than 2.5pro
A lot of staff use flash all day. I shift between the two with a personal preference towards 3 pro still, particularly one more complex issues

Flash also benches pretty close to 3 pro, some different RL training on flash that didn't make it into 3 pro, but that's very good news for the next batch of models

Jules actually has its own skills that are dynamically inserted into context as it needs it, like front end specific skills

[outer ibex]

hmm

[stone ice]

@night oriole what do you think of the thinking model? between flash and pro?

[night oriole]

I almost always use thinking. I'm less of a speed demon

I stay on 3 pro for most things myself

[stone ice]

I also usually use pro for better reasoning, such as coding, etc. But the thinking model is also good

How to make Google icons in nick?

[hot portal]

Is there any published roadmap for Jules? Are there plans for a private connector for github repos behind vpn, or even not-github support like gitlab or gitea?

[swift karma]

You guys just need to go more active on discord since algorythms will barely recommend Jules X page. Just a simple warning saying "Hey Gemini 3 Flash is available on Jules now" is enough

[cursive void]

Has anyone else had an issue where everything looks good but when I push to GitHub I go there and file changed = 0?

[cursive void]

Ok I see lots of people have report the same thing.
I guess I'll go for a walk 😂

I think I spend more time with Jules than wife!

[devout mesa]

I just came here to report the same thing... Damn... Ok guess I'll do the same lol

[fossil ivy]

are there any coders here who can take a look at my code for a bot im making and let me know where ive gone wrong? thank you

[hybrid sequoia]

According to livebench 3 flash beats 3 pro in coding, but not agentic coding, so I still think for non time bound tasks it still might be useful to use 3 pro

[visual gyro]

Don't have this view point

[visual gyro]

How do you get this view???

How do get this view @ionic cliff

[ionic cliff]

What about it? Flash should be there

[swift karma]

According to what i saw, apparently takes longer to peform more difficult tasks since it thinks less

So basically if you ask both models to do the same thing:

• Flash will be commiting more mistakes = More time fixing errors
• Flash is mostly up make more "unverified code" which code review can handle but still

Flash is better for "light" coding, leave complex stuff for Pro if you have AI Pro+

[visual gyro]

I talking about other settings codebase over view and performance?

[ionic cliff]

Ah! You gotta turn the suggestions on and have a lot of crappy code like I do

[hybrid sequoia]

hmm

jules should subagent flash for implementing individual functions and then use pro for most of the "architecting"

I feel like this is coming soon™

[serene fern]

Flash better in coding

[violet sparrow]

hiii

[hybrid sequoia]

it marked u as a spammer

[outer ibex]

well, results still may vary for each people

what suits the most is their decision

ofc giving simple prompt but hoping for complex code is not gonna work on flash, gotta build the planning

even in AG i saw that g3 flash will be the one that do browsing capabilities and screenshot/record the page, not g3 pro for some reason

never fail for me tho

[halcyon wagon]

No, pro is obviously better

[outer ibex]

"pro" in the name doesnt mean its 100% better than its lightweight counterparts

obv pros and cons

[halcyon wagon]

In Antigravity and AI Studio, they already had 3 Flash. I’ve used them, and honestly, the experience was really bad.
So I think Flash is mainly made for easy and quick coding tasks where speed matters more. For complex work, Pro models are supposed to be better.
But to be honest, both models feel disappointing. There was way too much hype before the launch, and in the end, we didn’t really get anything close to what was shown

[outer ibex]

different people different result i guess

[spark sonnet]

For complex tasks cooperation is necessary between the human and the model. Or between the model and another model.
And any possibility of cooperation is relentlessly attacked by previous changes to Jules which are unrelated to the type of model being used... Just "no, Jules is the agent to work on it's own without cooperation" kind of vibe. And that kills it for complex projects. Started a while ago a project that involves Rust, C and assembly code, and has to patch Rust's std library. Knew Jules limitations but it was still a bit useful in the beginning, but now - either it's Jules way (which is code indefinitely, create problems to solve, and stuck indefinitely trying to solve artificially created issues without human intervention until finally breaking everything or committing broken code), or it's "random reset on every attempt of human to correct Jules"

[broken drum]

anyone have tried integrated supabase MCP to jules? I just tried it and it always fail, while i cant disconnect it for reconnecting it

[outer ibex]

well im not talking about jules but i get the point

im currently running some tests and tasks on both model in jules to determine the "[this] is slightly better than that in overall" type things

[ornate oasis]

please fix bug error code/file not show when submission task finish the code tab also gone.

[serene fern]

#jules-general・🦑 must have official NDK support 🤭🤣

[outer ibex]

atp we need jules sdk in all prog langs

i made one in .net

im gonna target some other lang

[serene fern]

kotlin

[serene fern]

KMP would be better for that

[stark reef]

has anyone used clawdbot succesfully with gemini models?

[twilit turret]

I actually like them the most out of the frontier modes. I’m surprised how people hate on them.

I predominantly code in Julia, but I also vibe coded a front end app and found both cases Gemini flash and pro to be much better than the competition.

[serene fern]

Yes

Jules should be open source, we are ready to contribute to fix bugs 🫩

Seriously, I need AMA asap

[swift karma]

Google is not making a labs project open-source

[swift karma]

Dead chat

[eternal lake]

Hi! Did anyone find a way for Jules to read the pr comments from code rabbit or Gemini pr review?

[serene fern]

What is offensive here?

[serene fern]

hope you got my answer

[eternal lake]

I don't know. I wanted to reply that I want the process to be automatic, without my intervention

[eternal lake]

Yes. That works, thanks. I was wondering if you could set a list of user Jules could listen to in a pull request

[serene fern]

Create a issue here with [Feature Request] here: #1238621511401734254

They already brought the Actions failure detection feature (I and some other people had feedback for that) so it's possible they will bring this too

[craggy sinew]

Trying out jules after 3 months. And I forgot how cruddy it is...
I cant tell if it's actually working or not.
Where the hell is my verbose!!

https://tenor.com/view/shaking-mr-robot-mr-robot-train-robot-was-it-you-mr-robot-gif-26818665

[serene fern]

Antigravity Vs Jules

I gave the same PROMPT to both of them (The prompt was, Find these files and delete them, I have the full list as well)

Antigravity

Workflow of Antigravity

Completed within 10s (Gemini 3 Flash)

1. Just deleted all files.

Jules

Workflow of Jules

1. Reading AGENTS.md (took so long)
2. After 10 minutes, deleted all them.
3. (Not submitting) so I asked to give me
4. No reply (Restored all deleted files) so I asked Why did you restored them?
5. Jules replied: I was timeout. I will delete all them (deleted all files again)
6. Restored all them after 5m
7. Working... (after 30m, still working)

Note: Jules was in Gemini 3 Flash mode

[ionic vortex]

Can I ask a question? Let's say I subscribe to Jule pro for a month can I cancel after I buy a month so it runs out after the time limit but I retain pro features? Im using it to develope a game so very interested.

Thanks a bunch

[serene fern]

Finally Jules submited but took an hour, where as ANTIGRAVITY did it within 10s

[ionic vortex]

Can anyone answer my question? I appreciate the help

[serene fern]

Yes — in most subscription systems (including Google AI Pro / Jules Pro), if you cancel right after buying, you will keep full Pro access for the remainder of that paid period, and it will not renew automatically. Your Pro features remain active until that monthly period ends.

[ionic vortex]

Thanks very much!

[halcyon wagon]

Use 3 pro

[serene fern]

speed: Flash > Pro, so you are telling me to use the pro version to so that you submit the task after 3 hours

[halcyon wagon]

No, my jules takes a maximum of around half an hour on pro, not more than that
Earlier it was extremely slow, but now it has improved a bit, though its still slow overall

At least the Pro model works
I honestly have no hope left with Flash.

[orchid pebble]

I have a massive bug :
Jules and Github exchange indefinitely :

• jules do a commit
• the CI/CD deploy
• jules see that and do an empty PR
• the CI/CD deploy again

it creates an infinite loop of retroaction between the Github Ci/CD feedback and jules.

Need some advice to fix it or please fix Jules.

PS : I dont have the "permission to send bug reports"

[orchid pebble]

I'm trying this prompt on the knowledge base to solve my issue :
"Never push an empty commit"

I will add that on my AGENT.md.

[spark sonnet]

I would be very grateful if you'll be able to tell if it worked or not. Because in my experience, Jules adds "Pre-commit steps" and "Submission" to its plans due to very invasive system instructions, and it almost never follows user instructions that contradict these... Basically, the system reminds Jules to follow on these system instructions, even if the user said "Don't do this", and Jules interprets the system message as if the user too agreed to change their opinion
(I mean, Jules may get an idea that the user don't want empty commits, but still push a commit, even if it's empty, due to system prompt)

[orchid pebble]

I feel it's hackable

I got one success. I should try more. I'm sharing my prompt

[midnight patrol]

Hi guys, I've encountered a strange problem and I'm not certain if I should submit a bug report or not. The session was created without codebases, because I wanted Jules to create a new repo referring to multiple repos, some of which are not mine. After planning and approving, Jules told me all plan steps completed and gave me a summary, but I confirmed that no commits were made anywhere. I checked the bug reports, and I saw people talking about Jules submitting empty requests, but in my case it didn't even give me the button at all... Is it because no codebases are connected? Or is this really a bug?

[orchid pebble]

can you try that :

⁨```

Agent Instructions

Core Directives

1. Never Push Empty Commits

   • Verification Required: Before creating a commit, you must verify that there are actual file changes to be committed (e.g., using git status or by tracking your modifications).
   • Prohibited Action: Do not create a commit if no files have been modified or created. Do not create "empty" commits just to trigger a pipeline or acknowledge a message.

2. Handling CI/CD Feedback

   • Ignore Success: The CI/CD pipeline (Cloud Build) post status comments on Pull Requests. If a deployment or build is reported as successful, you must ignore it. Do not create a new commit, comment, or PR to acknowledge success.
   • Act Only on Failure: You should only respond to CI/CD feedback if it reports a failure or error that requires a code fix.

3. Preventing Feedback Loops

   • Be aware that pushing a commit triggers the CI/CD pipeline, which creates a comment. Responding to that comment with another commit creates an infinite loop.
   • Break this cycle by stopping work when the task is complete and the tests pass, regardless of the subsequent CI/CD success notification.

[swift karma]

"Code reviewer told me the app will crash but i'll submit anyways"

[leaden flicker]

Hello guys so which model you like pro or flash which is good can tell me your experience

[visual gyro]

Gemini flash not good for coding or nothing it can not even do decent website if write it in different languages then English stay away

[visual gyro]

It does not let me activate it, can you look in too it , please 🙏🏻 sir can you please also ship 🚢🚢🚢 Jules need some love maybe android application or something. Can share roadmap soon, is possible combine it with google antigravity service 🗺️?

[hybrid sequoia]

I prefer 3 pro for most tasks.

[serene fern]

Gey @eternal lake I want to inform you that, I found a method to do automated review sending to Jules

Step 1: Set Github Actions secret (in .yml code, i used "PAT")
NOTE: Make sure, you (Token) have permission to write comments on PR.

Step 2: Create this: .github/workflows/send-reviews-to-jules.yml

name: Comment review link and tag Jules

on:
  pull_request_review:
    types: [submitted]

jobs:
  comment-review-link:
    runs-on: ubuntu-latest
    steps:
      - name: Comment review link
        uses: actions/github-script@v6
        with:
          github-token: ${{ secrets.PAT }}
          script: |
            await github.rest.issues.createComment({
              owner: context.repo.owner,
              repo: context.repo.repo,
              issue_number: context.payload.pull_request.number,
              body: `@jules ${context.payload.review.html_url}`
            });

[mint totem]

heya!

[leaden flicker]

Thanks man 😊

[serene fern]

Silent removal of "SUGGESTIONS" feature

[swift karma]

what

[balmy narwhal]

dear google please add to Jules posibility to check http status and server/client activity because it is very hard to work with jules when you have lags and you cannot verify if gemini is doing hard task or is stuck

[serene fern]

Full detailed 100+ lines of general prompt 🙂🤐 written in the style of official example with the help of Claude Sonnet 4.5

[swift karma]

Uh how

isn't it better to use opencode/openchamber?

[serene fern]

🔁 The Jules-Gemini Development Loop

After running high-level prompts through Jules, I observed that:

• 🚨 30% session failure rate - Jules struggles with complex tasks
• 🐌 Slow execution on remaining sessions
• ⚠️ These prompts are currently too abstract for Jules to handle effectively

So I created an automated feedback loop! 🔄

🛠️ Here's My Implementation:

1️⃣ Initialization

🎯 Jules auto-triggers a dynamic prompt that transforms it into an autonomous developer for the specific project

Note: This isn't a static prompt—it adapts per project!

2️⃣ Development Phase

🤖 Jules performs multiple development tasks within the session
📦 Creates a Pull Request (PR) with changes

3️⃣ Review Trigger

👁️ Gemini Code Assist automatically detects new PRs and begins review

4️⃣ Feedback Loop

⚡ GitHub Action workflow:

• Monitors for PR reviews
• Captures review URL
• Pings Jules: @jules here is my review: [URL]

5️⃣ Correction Phase

🔧 Jules addresses all issues mentioned in the review

6️⃣ Commit Updates

✅ Jules commits the fixes

7️⃣ Loop Continuation

🔄 Another workflow triggers on new commits, restarting Gemini Code Assist review

🔄 The Continuous Cycle:

Jules → Development → PR → Gemini Review → Feedback → Jules Fixes → Commit → Repeat

[swift karma]

Tho Jules kind of still not good at coding

[serene fern]

Yes, But i will not let go daily sessions away like this

[swift karma]

We love failed tasks

[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

[pure sand]

Hello World 👋

[uneven iron]

What is going with Jules recently 😭😭

[swift karma]

recently?

[eternal lake]

Thanks for the workaround Ashik

[tight rivet]

Is it just my version that no longer allows downloading the zip file made by Jules? It used to be possible to download it in the "Ready for review" section.

[cosmic bay]

always has been top right code area for me

[ionic cliff]

Got something for you to try out!

[ionic cliff]

Is this a repoless task?

[outer ibex]

you need to specify your preference and styling, then flash will work great

thats like the same for every ai models

i even built a clone of perplexity with 3 flash. gave some specific styling from the "original" code, multiple screenshots and it built them as is

its obv simple prompting will be a disappointment

[outer ibex]

i used 3 pro for deep research, planning and creating foundation files then 3 flash for the rest of progress till completion

[leaden flicker]

Thanks man 👍

[visual gyro]

Awesome 👍🏻 let me know what you want me try out I am all ears 👂🏻

[hard egret]

Hi, is anyone else having trouble disconnecting an MCP connection? my Key was wrong and now I can't disconnect it to change it.

[minor garnet]

https://github.com/Mk3Ang8l/GuardianLayer.git

for those who make ai assistant

[outer ibex]

as far as i know, theres no way to disconnect them

[hard egret]

Oh well that's unfortunate, guess it's a beta thing? I managed to disconnect it once, after connecting it again, I couldn't.

[spark sonnet]

Now this is news for me...

[outer ibex]

yes its still in beta

everything in jules is technically still beta tbh

[spark sonnet]

Reading a single short file is an impossibly hard job... 13 minutes on this single action already

[outer ibex]

gg

[visual gyro]

What do you want me too try out? 🙂

[proud bramble]

jules have google 3 flash for free account great ^^

[full viper]

Hi guys, I've encountered a strange

[full viper]

fixing it up. will drop more info very soon

[minor junco]

Any advice on what to do when jules just seem to be stuck at "Working...." (during: Complete pre commit steps)

[minor junco]

Regarding my last message, jules is actually stuck in a loop and just keeps updating same files over and over gain during "Complete pre commit steps." Has been going on for one hour now (this is a small project). Is there some way to stop or abort?

[visual gyro]

Is it for stuck like that can not activate it? Or enable it on the phone? I can not activate in web browser when I tried

[minor junco]

It just keept looping over the same 20 files with the same message again and again... for more than 1h....
...it finally stopped with a message and I asked it to push code so I could restart in a new session, but it just went back to looping over those files again...

Could jules just be incredably overloaded just now?

[spark sonnet]

This symptom usually seems to happen (based on my experience) when Jules thinks that what it's doing is not gonna solve the task, and some command it runs fails. It knows that submitting that to Code Review would probably flag this, so it doesn't submit to Code Review (it is very afraid of that thing, apparently), but it can't make the command work, so it just continues to try until the system prompt pops up to tell something that causes it to stop with a message.

However, the user replying to a message mid-task causes the environment to reset (genius idea - totally kills any attempts at human-guides-agent type of tasks), so it doesn't have nothing to submit anymore, so it begins to edit the files again, and gets stuck in the same issue

[minor junco]

I have tried to write to jules, but nothing happens... I even started a complete new task but after a while jules seems just to get stuck on that loop again.... to bad. I have to give up on jules for this project in that case...
Seems like there is something in my code/project that triggers this bug.
Funny thing though: Everything was written by jules....

[spark sonnet]

Worth trying to just make a new task with the same prompt. It took me a week of attempts to set a task from time to time, but Jules was able to solve for me a really complex issue recently

[minor junco]

Yeah... so far I have been quite happy, it's been slow, but if I don't have to monitor progress this is less of a problem.... but to get this to work I have to try and try and try....... and I can imagine why things are slow if resources are wasted on such never ending tasks....
Maybe I will try more ....

[upbeat pond]

Has anybody had the Problem that Jules enters a full chat mode and doesn't show the current file contents in a code pane on the right side?

[spark sonnet]

I totally understand your frustration. I even feel pain for the team that develops Jules, due to the idea being so awesome, but as pioneers in that direction, apparently not every solution to each issue is the best one, and layer onto layer of such imperfect solutions and new issues that adds to underfixed old issues create a gradually growing disaster. And what's worst? The competitors take idea of Jules, but they aren't pioneers.... They aren't locked by long-since-committed decisions or some strategy of completely eliminating the very possibility of human-agent cooperation in the name of the 'most autonomous agent'. So the competitors just make Claude Code web in the cloud and now even OpenAI made their Codex web version clearly taking inspirations from Jules...

And yet, it was Jules who solved me the issue in assembly code for me, no other agent did that

[minor junco]

Yeah... my feeling right now is that I have a supercar in front of me, but it lacks a steering wheel... 🙂 Thanks anyway.

[spark sonnet]

If there were some files edited, and then Jules started to chat and the files are no longer there, it probably reset it's environment and the files are gone But not really, hypothetically Jules can restore these files from the cache if asked very nicely, but it's entirely up to it now, and it may be stuck in a loop, thinking non-end or doing nothing at all

[leaden pawn]

Hello Everyone! In case you missed it, please refer this for our new addition https://x.com/julesagent/status/2018801499573457222. For more details: https://jules.google/docs/changelog/

[visual gyro]

Hello David

Sorry for being excited 😅

Happy you fixing it up, whatever it may be. But can you please tell me, What you are referring to 2.

That you will drop info on very soon? The word= Soon too me is within minutes, at most hours. So will it take days, weeks, month?

I prefer an estimate on how long time approximately?

My Birthday 🎂🎈🥳 is 5 February so hopefully before that? What your take on it?

[visual gyro]

Saw the tweet, as soon it available for first tryout I am in

[gleaming kayak]

Is a fix that I won't have to keep refreshing the website for new messages to appear coming too?
Also maybe consider adding mobile support or mobile app version of jules (That would help majority of remote users and let us utilize jules when we can't use cli or AG) 🤔

[minor garnet]

would yall be interested by a jule discord bot? like a open source one

[outer ibex]

depends on what it does, i assumed it would be knowledge base and modmail i guess

or is it the truly jules usage

[minor garnet]

Like it use the jule api

You can program task without leaving discord

[outer ibex]

i already saw someone made it on gh

i dont remember whats the name of the repo tho

if you are willing to try do it then im all eyes in the progress

[minor garnet]

i made this tho

[outer ibex]

cool app

thought you were referring to actual jules discord bot

but thats even better lol

[minor garnet]

yea its open source on my github

i want o add more thingsa and i could even make it a saas

[visual gyro]

You know this is against rules of server either explain it or this is something for the @barren haven

[outer ibex]

man just ignore them, dont even need to reply to spam

[white oxide]

My Jules doesn't respone and I think it is stopped . What can i do. I waited 1 hours

[leaden pawn]

My Jules doesn't respone and I think it

[serene fern]

Maybe this: https://github.com/TheRealAshik/JulesClient

[visual gyro]

It's my birthday 🥳🎈🎂 Jules do something 🎉

[vapid trellis]

IS there an easy way to tell which repos I've enabled suggestions on?

[swift karma]

Happy birthday 2u

[timid folio]

Happy B day

[visual gyro]

Can we get this MCP support please : https://developer.wordpress.org/news/2026/02/from-abilities-to-ai-agents-introducing-the-wordpress-mcp-adapter/?utm_source=twitter&utm_medium=social&utm_campaign=fedica-WP-Dev-Blog

[serene fern]

You have mcp suggestions option in jules > settings > MCP

[visual gyro]

Yeah 👍🏻 I did

[raven needle]

My jules has been loading like this all day while opening a session. What could be the issue?

[tropic hound]

I had something like this yesterday, when Jules had decided it was a good idea to include 736 third-party node.js modules with their source code in the PR. I guess that's too many files for the display logic...

[raven needle]

How did you manage to get it fixed?

[tropic hound]

I just had Jules remove the files via a comment on the PR, but that didn't fix the display issues. Luckily what was left of the PR could be merged without problems and then I opened a new task for any changes that needed to be made. And on that new task I did not have any display issues.

[outer ibex]

if the session is too heavy (full of long conversations and texts) it will stuck

[serene fern]

@full viper Why is @julesagent using MCP tools for understanding how to use his own tool?

[brisk hamlet]

Jules still not showing command outputs.
This is still inacceptable. I don't know how long we are supposed to fly blind.
Is this "issue" (it really seems to be on purpose) being worked on, or did you decide that we need to kept in the dark?

[spark sonnet]

Despite what is being said, I suppose (though I speculate and I may be wrong) that it's the latter. The high-level strategic goal of Jules is to create 'the most autonomous agent', and as long as users are trying to control how Jules works, it goes against that strategic idea, so any means of control and transparency are removed away from the users. Ideally, it's just one prompt - and then one pull request. That's how I suppose Jules is seen in it's perfection, but it's impossibly far from this perfection

[surreal sky]

https://github.com/superealityai-MYTH-OS/PRZ-AI-EI-OS

[outer ibex]

you can somewhat control jules behaviour but you must follow the expected jules thoughts, try to be on the same page as jules

brainstorming ideas, refine the planning and just let it do the work

took me 3 months to really get it to work like old times, but now its running just fine for me (both 3 flash and 3 pro btw)

[brisk hamlet]

it's nice that you got it to work. for any dev who needs verifiable results, and does not trust Jules blindly (due to a "success"-hallucination-rate of about 10-20 %), it is unacceptable and a ridicolous regression.
At least it would be decen/honest of the Jules-team to tell us their plan: Are we supposed to fly blind, or is it a bug and we will get back the command outputs?

[lofty aurora]

Hello

[dreamy breach]

Hello

[spark sonnet]

What kind of tasks does this work with? When I know in advance in such details what I need to do, why would I ask Jules?

[outer ibex]

they prob hides the command outputs to make loading faster because its "less text to load"

maybe because its highlight is "autonomous", letting the normal user (not a coding savvy user) see the actual progress of the codes is better than showing bunch of "weird" commands

(my assumptions, take it lightly)

[spark sonnet]

Indeed I believe (my assumptions) that it's because it's highlight as "autonomous" mainly, because otherwise it would be logical to make this a preference in settings

[serene fern]

I believe they removed it for security reasons, same they done with MCP by not giving custom configuration option. And they can keep the command default collapsed and do On-Demand-Loading when expanded, so optimization was never been thought for Jules 😅

[outer ibex]

it works on all complex tasks (3 pro specifically), for some easy prompt its better to use 3 flash

first and foremost would be to always use the interactive plan mode, because review and start modes will just gave you half-baked, good enough but functional code specifically for creating a foundation files

using the first mode will make jules keep asking you questions to actually refine the planning on whats the result that you should expect it to be

sample prompt of mine for starting a session :

i want you to [put your first thoughts in here]. some specific requirements would be [list out things that you truly need either by commas "," or in numbers].

revise the codebase to fully establish a deep context understanding.

then jules will start planning and it will asks you further questions, answer them in long paragraph but separated by each list number

example jules response :

Before i begin planning for the [task name], i have a few questions to ensure i meet your expectations:

1. question
2. question
   • sub question (if available)
3. testing

Once you clarify these points, I will create a detailed plan.

you should reply in numbers like :

1. [your answer]
2. [your answer/suggestions] (combined both question and sub question answer into long sentence)
3. for testing, you can try to verify or if it isnt possible then i will test them myself

if the testing requires api key, you need to provide them for automatic testing

works on either repoless, populated repo or empty repo

i dont know if thats answer your question or not but ill try to answer

[spark sonnet]

I did recently used that interactive plan mode to get the following task done:

Files to work with:
1. `test_lib/src/lib.rs`
2. `src/test_dl.rs`
3. `sys/src/dl.rs`

Directories to run `clippy`, `rustfmt` and `test`: `./`, `test_lib/`, `sys/`.
Verification:
- `cargo test` in every directory to run
- `build.sh` with `test_dl` feature argument
- run `test_dl` binaries in `dosemu` and verify output

Do the following at once in the files to work with:
- Ensure that the code contains no magic numbers or magic conditions. Use named constants, variables or self-explanatory code instead. Self-explanatory code does not mean adding comments, it means that code itself (statements, logic, blocks) read well in English without any need of non-code comments to understand what is being done.
- Ensure that the code contains no instances of `else` in any form. Use early returns instead and/or method extraction
- Ensure that the logic has specific/unexpected cases in more nested blocks and regular/expected cases in less nested blocks
- Verify that files do not contain any `unwrap` / `expect` call or other similar call that may panic. Only `unwrap_or` and equivalents are allowed as long as these cannot panic. All errors must be handled instead of lead to panic.
- Use `cargo clippy` and fix everything it reports
- Run unit tests and ensure all tests pass and no `#[allow(...)]` blocks exist in the code
- Use `rustfmt` to correct code style
- Update the documentation in the relevant `.md` files regarding the topic of changes. You must find relevant `.md` files and update this documentation.
- Clean all `.log` files using `find . -name "*.log" -delete`. This must be THE LAST step before submission

Do not do any unrequested file edits.

The plans were alright. Still took 3 attempts (had to force deletion of previous 2) because Jules was editing files I didn't request to edit anyway, and not just on cargo clippy suggestions, but actually logic changes I didn't want.

[outer ibex]

because you are specifying file names in the chat, jules will mostly not follow your specific tasks on what to do/what not to do with the files

its better if you use AGENTS.md

[spark sonnet]

It would ignore it as well sometimes. It's not robust for Jules. Besides, the files to be changed vary depending on the task

[outer ibex]

well yea, as always different people different result

the issue right now jules is just half the time are incompetent and inconsistent

i might do some more testing and prompting to see whats gonna change

[spark sonnet]

How I solved this: in the 3rd attempt I said that I would not approve the plan until Jules includes "I must not produce unrequested changes." into the text of every plan step. It did so on second attempt of making a plan in the 3rd created task on the topic, and didn't produce the unrequested changes in the end (task completed)

[outer ibex]

very needy of jules requiring specific prompting lol

it is what it is

[spark sonnet]

Yeah. And yet, the concept is awesome and the potential is almost unlimited, if all the bugs in the control software would be fixed. However, the competitors are gaining with similar products (but their agents are still a bit behind Jules for now)

[outer ibex]

agree

[serene fern]

Could it be possible reason for Jules to be slow for RATE LIMITS? Literally there is no cap per sessions

[minor garnet]

is jule good or not

[uneven iron]

Where can I post bugs with antigravity ?

currently here it's 11PM (23:00) so what's the deal with that

[minor relic]

check reddit xD

then you will understand why they dont have a own channel for antiygravity

[uneven iron]

Ok it works now

[spark sonnet]

Sometimes, it's hilarious

[simple rain]

How are we supposed to get any insight into what is going on with a run? Seems like 2 out of 3 sessions I created are just plain stuck

[minor garnet]

well

theres a 80% chance they are stuck

what do you do for a living tho?

[simple rain]

beyond this ofc: (nothing in here is secret; the only reason the repo isn't public is that I don't feel like publishing unreviewed AI code with my name attached)

[simple rain]

technically now sysadmin; but this is my side project, which is currently to get some mesh networking up and running to get location self calibration working for indoor location services at events/venues that have zero GPS reception inside.
I'm mostly just trying to get boilerplate done now 😄

ok I one of the 3 runs, the one that wasn't appearing stuck, just shipped a PR after an hour of running

ok, I guess that's something..... too late now here in central europe to review it today, but I feared worse, still way up in terms of time spent vs. doing it by hand....

[simple rain]

thanks for the warning.
Any tips/tricks on dealing with that in the future?

[minor garnet]

like jule?

jule is a pretty bad agent ai for large project since it will corrupt everything 100%

its still good for fast design etc

[simple rain]

it's that bad on the corruption?

(what languages have you experienced that with?)

[minor garnet]

imagine working on a project for 24h straight right and then jule corrupt everything aand dont even give you the pr

[minor garnet]

python mostly and csharp