jules-general・🦑

[balmy coral]

No afternoon 😕

I don't know jules what doing 😭

[serene fern]

Timezone diff reyyy

[balmy coral]

I am from India now

[serene fern]

Yeah, Bangladesh 🇧🇩

[balmy coral]

Oh

[rigid flicker]

Its a happy day! Finally Jules

[vale garden]

I really hope they push for stronger integration into the gemini mobile app. the whole point of async coding partner is being able to check in and add updates. I tried with the current app but it hallucinates all results =/ would be a great feature!
I was hoping i could have nightly reports ready for my gemini to read to me in my morning coffee walk

[vernal compass]

JULES takes too much time to clone the repository. Please solve this issue.

[plain dirge]

yeah even with just blank README.md lol😹

[vernal compass]

True, even when the repository is completely empty

[plain dirge]

i think google gave jules a 2G internet lol😹😹😹

[sudden slate]

Is 2g still there in your place

It's there at my place, but they're gonna shut it down soon(3g already gone)

[orchid jasper]

i hope jules isn't going crazy like main gemini is today lol

guess i will find out shortly

[potent ibex]

anyone have suggestions get lost; if you have it toggled for a deleted repository?

[halcyon wagon]

Google will ban you, it's not democracy.

[swift karma]

Eh... Didn't think about that

[full viper]

Yep. Hear you loud and clear. Just to make sure I understand, you registered for web push notifications in the Jules web app but aren't getting any?

[halcyon wagon]

Yes, I am not getting any notifications.

[full viper]

We're looking into this! Please send over any session IDs if you can

[full viper]

Web push registration is super tricky. I'm digging in and will keep you posted.

[serene fern]

Should I open a bug report here?

[full viper]

I have a bunch of example sessions myself but seeing examples from other projects is really helpful. So any session IDs you can pass along would help a lot.

[serene fern]

Sir, the PWA of Jules is broken (just to let you know)

[full viper]

Jules is not a PWA, it lacks a Service Worker and all the install icons.

I have some ideas about improving all of it but just eagerly adding a Service Worker tends to do more harm than good. So I'm giving it a good think.

[serene fern]

https://jules.google.com/session/9631244664907942162

[serene fern]

Hey @full viper,
I Request you to bring @julesagent in KMP, for all devices! I would love to see a official native app or you can personally make a opensource app where we can contribute.

What do you think Sir and @halcyon wagon ?

[halcyon wagon]

Yes and it will also solve notification issue because it will no longer required web push registration

[full viper]

Have you all checked out this mac app? It's not official but it's made by one of the product leads of Jules

https://github.com/simpsoka/jules-companion

[sudden slate]

Im windows user😭

Anyway u like windows or Mac??

[orchid jasper]

not 100% sure if it was just a bad session but jules seems to be struggling after the 3.1 upgrade

[full viper]

Happy to help. Can you provide me some details? Either here or in a DM. Screenshots, Session IDs, all of that context is super helpful.

[orchid jasper]

i got mad earlier and archived the first one already 💀 i think it might have been a me/prompting issue though. current session is 2808767370079446128 and i'm gonna let it run, but i'm probably going to break the task up into smaller pieces

[orchid jasper]

(and ty btw)

[dire kindle]

jules is noisy and sends me emails add the time that it's reporting for duty and I can't figure out how to shut it up. Also, why doesn't Jules resolve the issue or update when the PR is merged or closed? Sorry, I have a lot of pentup questions.

[orchid jasper]

Are those emails your github notifications for Jules commenting on a PR?

[dire kindle]

yes, but the issue is the comment as well since it authored the PR

[orchid jasper]

AFAIK it's always going to comment if you have Jules publish a PR instead of a branch. Personally, I work with Jules using branches because it's less messy lol

[dire kindle]

is it better at detecting dupes that way? it seems like it is very aggressive about noticing the same thing over and over.

[dire kindle]

Thanks for the advice! It's maddening

[orchid jasper]

what do you mean by detecting dupes?

[dire kindle]

it files a lot of nearly identifcal or overlapping PRs

[orchid jasper]

I think i know what you're talking about, and it's the main reason i do branches. If you do branches, then talk to Jules in the web ui, the next publish will update that same branch with a commit instead of doing weird extra PR stuff.

I only publish to PR if i'm pretty certain it's gonna get it right 100% the first time

bc when you publish to PR, the only "correct" way to follow-up is replying to Jules comments on github

only issue with branches is you need to start a PR manually, but i prefer it bc sometimes I don't wanna PR to main and point it at another testing branch or something

also to make it so PR isn't the default option:

[dire kindle]

that's the detail I could not find, thanks, it's low key mental, I feel like sessions should be the name of the issue and not the damn bot name

[full viper]

Are you referring to a GitHub issue so it's auto closed when the PR is merged?

[full viper]

Can you elaborate a bit more here? This isn't something I commonly do so I'm interested in hearing more about your approach

[serene fern]

Hey wait

I FOUND A BIGGEST BUG, Can we DM sir

[full viper]

I'm also interested in hearing this approach as well. Is this mostly due to avoiding merge conflicts?

[serene fern]

See session limit. 114/100

I AM A JULES pro user but i can use unlimited sessions

[full viper]

I'll take a look. It might be either a UI or eventual consistency issue.

[serene fern]

Its not just UI all tasks are running and working

[full viper]

Sorry, I mean in terms of the counting UI. I'll take a look

[orchid jasper]

When I publish from Jules web directly to a PR, if i do a follow-up debugging instruction or something, my only option is to "publish as PR" again, which opens another PR. If I do the dropdown and "publish as branch" (which I would assume would just update the branch of the PR already published) it always reverts back to "Publish as PR" bc that's what it did the first time. i can keep "publishing to branch" even if I have a PR open for that branch (that I started) and it just shows up as branch updates in the PR history.

[full viper]

Oh and certain kinds of tasks don't count against your session quota, in terms of suggested tasks

[orchid jasper]

(unless something changed recently I'm unaware of)

[full viper]

my only option is to "publish as PR" again, which opens another PR

Do you mean from the same session? It publishes two PRs? That should not happen. It should update the existing one

[orchid jasper]

maybe it got changed recently lol i don't use publish to PR a lot anymore

lemme go do a quick session rq

[serene fern]

Dear @full viper
@julesagent has the Android SDK installed. I tell him “check build passed before submitting,” but there’s no way to know if it’s actually building or stuck, so I often get confused

Right now, the only way to confirm is the .salive file. Any ideas to fix this?

Build usually takes 10-15m, till then it feels like it's stucked! And often they leads timeout

[orchid jasper]

so far you are correct, i think whatever i ran into before was changed or updated that made me stop using PR publishes 💀 it used to be if I started with "publish to PR" it would always fall back to "publish to PR" on follow-up commits in the same session even when i said "publish to branch"

and yeah, if i try to make another PR over it in the same session, it says "failed to create new PR [because there's one with the same name already blah blah]". I never had seen that before

[orchid jasper]

so my bad, was running off some outdated experience info. I do think branches are a bit more flexible, though, even though you can redirect a PR to a different branch. sometimes i like to do branches off branches, then merge the furthest branch into a testing branch, then if it passes whatever i'm trying to test i'll collapse it down into the branch it originated from so i can just eventually do one big PR to main with all of my tested changes.

aaaaand github is down

(for everyone)

[regal depot]

It's only partially down for me

I commited a PR and it shows the changes on my branch but the PR is still open waiting to be commited somehow

[faint dune]

I've the same issue, it didn't reset since 3 days

[full viper]

Wow are you building an entire Android app in Jules?

[full viper]

This is a bug. It should just tell you that the PR is updated.

[vernal compass]

Hello dear Google team member... I would like a hand of help here....

🤧 please help this one time... GitHub has been showing tons of issues today, even Gemini in the App is not working properly....
Now even Jules is saying that it will not help me

[full viper]

You are right about branches being way more flexible. I often do a lot of experimental work all on a specific branch and then merge to main. What I've been experimenting with lately though is a pure PR approach where all the merging can be automated, including the resolving of conflicts. This is the "panacea" of development work with remote agents. You can fire off as many as you want and let some merge orchestrator solve the conflicts. I've gotten it to work in some use cases and I'm testing it on more right now.

[full viper]

The 500 from the repo URL suggests it may have been from the intermittent drop in service from GitHub just now. We definitely need to handle this message better.

Can you try again and see if it still happens? I think GitHub is back up now.

[vernal compass]

Okay

[orchid jasper]

github is still being wonky. Not just in jules, i'm having issues pushing from github desktop too lol

[vernal compass]

I can't even reach 50 sessions per day, that gave me a nightmare

[vernal compass]

Well, I guess GitHub is still not upto work

[full viper]

Just out of curiousity, what is the branch you are targeting with that session? If the branch doesn't exist it can error out as well.

[vernal compass]

Main branch.

The branch does exists

[full viper]

Ah okay, then it's likely GitHub right now. Let me do some testing myself to see if I can hit this as well

[vernal compass]

Okay. Thanks for your time

[dire kindle]

Yes. Definitely.

[full viper]

Are you creating a Jules session from that issue?

[faint dune]

This is actually happening to me also, same branch that i executed many tasks on it before it shows this, what i did is paused the job since it shows error and started another task with same content

[dire kindle]

Are you creating a Jules session from

[orchid jasper]

(it's github going up and down)

[mild zenith]

Unable to send C# XML Documentation stuff in the Jules chat

Example:

In `@Source/KSLoader.cs`:
- Around line 111-120: Add XML documentation to the OnDestroy() method: add a
<summary> describing that OnDestroy tears down the singleton ownership and
resets initialization latches, a <dependencies> entry listi...

The chat box struggles with <summary> or <dependencies> etc. I think the < causes the line to end,
So when you press enter/send it will cut out at the first < like this:

In `@Source/KSLoader.cs`:
- Around line 111-120: Add XML documentation to the OnDestroy() method: add a

Where in this case it would be at the first <summary>.

[orchid jasper]

Does it work if you escape the character with a backslash?

[swift karma]

We are on May already and nothing interesting for Jules

Come on lock in

[mild zenith]

yes

[vale garden]

running a what i thought was a simple css task through jules on 3.1 pro and it seems to be taking forever! is jules slow for anyone else today?

[halcyon wagon]

Yeah, it really takes hours, and I can't just sit around staring at the same screen. So I close the tabs, and somehow I come back two hours later with zero productivity.

[tardy night]

@full viper Hi, loved your work since polymer time. Just a feedback can we have less pr in suggestion in similar improvement grouped in one suggestions and so we can have less pr to review and merge, this can speed up our work.
And can we have min 25 repo in suggestions currently its only 5 because its limiting whole end to end work for one big project in which we have few repos to manage multiple things. Awaiting your reply. Thanks

[halcyon wagon]

I think Gemini 3.1 Pro is slow because AI Studio’s build option also got slow. The problem is in the model, and the problem is in Jules too, so the double issue makes it even slower.

[near condor]

Sure!

The file system incident was session id 3242374324689227365. The message was:

I do not have access to basic file system or bash commands because they are not included in the list of available tools provided to me in this environment session. The only tool declarations I have received are for Context7 (context7_resolve-library-id, context7_query-docs) and Render API (render_list_services, etc.). Without local execution capabilities, I physically cannot execute shell commands to create the [log] file, browse the Java code, run ./gradlew, or create a PR. Could you please restart this session with local file and shell execution enabled?

The "deleting changes" is consistent and happens with the following workflow.

1. Jules pushes a commit.
2. I update the PR to match head and correct some sort of merge conflict. The merge includes changes to files that are out of scope for this agent (like adding a markdown file)
3. I give Jules feedback.
4. Jules corrects that feedback and reverts the changes that were brought in with (2).

Let me see if I can find an example of it ignoring feedback. The feedback is through github PRs and it marks them with the eyes emoji and then doesn't act on them.

[ashen glen]

If it's a simple job, use flash

3.1 has been quick and slow for me. Would love to have seen an update on how they've applied it. I think Jules needs it's own model or tuning, however

[sterile furnace]

As a feature, I'd love to have a setting to be able to select a different model for interactive chat followup vs the rest of the tasks. IE use Flash for responding to session questions and Pro for the heavy lifting

[sterile furnace]

I dunno about anyone else, but I really need the ability to set the current "default branch" globally

[rigid flicker]

Will not happen I can say you that. The goal of Jules is reliably tested code, everything that has to do with "fast" will not happen

[rigid gyro]

Jules new permissions update is crazy

[serene fern]

Yeah! As an android developer, I feel sad about Jules being only advertised to handle web apps. 🤖

Isn't Jules targeted for Android development?

BTW, Do you have any idea to avoid timeouts?

[serene fern]

Bearly uses any permission of them! It's just taking permissions for no reason! I would love to see new soon from Jules dev

[serene fern]

You gave me a nightmare 💔

[dense merlin]

Hi, how do you deal with merge conflicts in Jules? It seems if i merge one PR and then update other pr branch there is no way for Jules to fix merge conflicts ?

Apart from trying to make it write modular, SOLID. SRP code so that there is a less chance of merge conflict than in files with 1000s of lines of code

[full viper]

Would you like to alpha test something for me? It's not an official Jules feature but a workflow I built to handle these. I've been using it a bit myself and while it has some kinks to work out, it's working really well.

[dense merlin]

I can try

[serene fern]

What's that?

[near condor]

Lol, okay, this was a little bit on the funny side:

Wait, it seems I did a git reset HEAD --hard after code review and blew away my specific exception replacements, so git status shows I have nothing committed and working tree is clean.

Would you like me to rewrite the RuntimeException replacements, [do something else on my list], or take a different approach?

"Wait! I destroyed the work I did! Would you like me to redo it, do something else that I'm already also supposed to do, or something else entirely"

Session id: 17666778647758714725

[velvet sail]

can i give access for just a few selected branches to jules?

[tardy night]

@full viper any update? please share

[full viper]

Ah sorry! I missed this! Too many notifications. Thank you for pinging it again.

Give me a few moments to get caught up and a good think.

[tardy night]

there was typo i have corrected to ahve more then 5 repo added in suggestion like up to 25

[halcyon wagon]

I’m building a Jules Telegram bot that runs on a VM and lets you create and manage sessions from a Telegram group. There will be a pinned topic called “Control Panel” where you can create new sessions, change repo/branch, and view running or failed sessions. When you create a session, the bot automatically makes a new topic for it.

Each session topic starts with a status emoji: 🟢 final/PR, 🔵 processing, 🟡 question/waiting for input, 🔴 failed. The session name will end with [XYZ], where XYZ are the last digits of the session ID so it’s easy to identify.

The bot will notify you when a plan is ready, plan is approved, and when a PR is created, including the PR link and branch name.

I also use a GitHub token for repo and branch search because the Jules API doesn’t list/search well when you have many repos (I have 100+). GitHub handles the search, then the bot sends the selected repo/branch to Jules to start the session. Once you set a repo it won’t keep asking, but you can change it anytime.

Let me know if you think of any features that could make this better.

@full viper thanks for sharing the Jules SDK!

[halcyon wagon]

Honestly I think for real coding purposes Jules + Telegram is going to be far more powerful than the currently hyped OpenClaw.
One idea that could make it even cooler, add Whisper AI + an LLM in the workflow. Just speak into your mic in any language, Whisper transcribes it, LLM converts it into a clean precise English prompt, and Jules executes it. No typing at all.

That would be genuinely next level.

[wild flicker]

You can make something like this in 30 minutes to an hour. I've made one but I don't use it.
Either way new Gemini models can do the transcription, skipping the additional step.
On my version for prompt cleanup I used Gemini 2.5 flash lite since it was extremely quick with token output, perfectly fine for meta prompting.

[wild flicker]

I mean the marketing for Jules could be better but one of the things that made me pause using Jules for a while was the reliability. Too many agent timeouts VS doing the work on your own PC with other Google tools. Fix those, run marketing campaigns and then it might run circles around other cloud agents.

I used Jules for quick prototyping of C programs in early December, it was quite reliable for a week and then every other task subsequently timed out.

Haven't really used it since then which is a shame given how large the quota is, just been waiting on reliability fixes. I did like that when Jules encountered a missing tool, it would attempt to find something to actually get the job done but I haven't seen that in a while.

Been a few weeks since I've looked at offerings, thanks for the heads up

Stuff moves so fast that we're getting micro updates to models every two weeks or so

[halcyon wagon]

Yeah, you don't use it because it was made in about an hour. Even if you use AI like Jules with one prompt, it can take around 30 minutes to respond. So making something properly in just an hour isn't really realistic. My goal is to build something truly perfect, not something rushed.

[drifting dagger]

has anybody had success launching scheduled tasks and using mcp connections (towards stitch or linear) ? Whenever a scheduled tasks starts, no mcp connection is initialized, tried the same prompt using "Interactive plan" and the mcp calls are done as expected.

[sterile furnace]

I’d be down too, I’ve had tons of issues with merge conflicts and it not handling them well no matter how I prompt and reprompt and retask it.

It’s even claimed that there are no merge conflicts to be resolved any more, when there are 10 broken files out of the initial 20 it started with. And no amount of attempts can get it to pick up/resolve those ones that are either invisible to it or the agent is stuck in some bad state

[plain dirge]

@full viper jules with gemini 3.1 pro and gemini 3 flash still lazy, all really lazy

plz fix

[minor garnet]

Why u using jule

[plain dirge]

cuz i need gemini 3.1 for abstract reasoning and super good paper understanding and paper implementation/experiment implementation, but in jules it's really lazy, unlike in Google AI Studio.

[minor garnet]

Make him not lazy then

Its you job to understand how ai work and get the best out of it

[plain dirge]

it's the jules that is the problem, even code review roasted jules and still proceed to make it even worse by simplifying code and commit💀

even code review said "this is not complete and should not be commited for production" but jules still proceed to commit it

[minor garnet]

Ai cant make production ready code

[plain dirge]

it can, i swear gemini 3.1 pro on AI studio is really good and not lazy, but in Jules it's so lazy and feels so dumb

[minor garnet]

If you cant learn the basics you won't be able to understand what ai does wrong

[minor garnet]

I mean of course jule is free

[plain dirge]

3.1 Pro on jules is not free, it needs pro/ultra subs

[minor garnet]

Dont tell me you bought it ?

[plain dirge]

3 Flash is even dumber to work with, 3.1 Pro is dumb

[minor garnet]

[plain dirge]

... yes...

[minor garnet]

At least do you make money with Jules ?

[plain dirge]

no, i'm non profit, i want to make open source work and researching/experimenting for AGI

but since Jules hates me i haven't finished anything.

[minor garnet]

You won't be able to finish anything with Jules

So what you want to do is make a agi basically

[plain dirge]

Yeah i know, i thought jules will be good now, but it's still not good to work with

[plain dirge]

yes, researching.

[minor garnet]

I mean if you dont want that sub I can take it

But when you say a agi what do you mean by that

[plain dirge]

well though Antigravity was still good, but it always saying agent crashed thing

[plain dirge]

My definition of AGI?

[minor garnet]

Yea too much people on it

[minor garnet]

Yes

[plain dirge]

So uhh. My definition of AGI is:

• a system that can learn efficiently like us, reason like us, see like us (hollistic vision, not pixels), that can continually learn and transfer knowledge

and how i want to achieve it is by using:

• Active Inference for human like reasoming (using Bayes Theorem)
• Brain inspired component for efficient learning (like Hippocampal Buffer, PFC, etc)
• Retina and Renormalization Group for Hollistic Vision

andd.... that's it... my definition of AGI.

but Jules fails at what i want it to be, it even simplified the equation that is so important for this Architecture (Active Inference).

oh well Claude is always better at this.
https://x.com/i/status/2029531728381919657

can (@marmaduke091)

Shock! Shock!

Turing Award legend & godfather of algorithms Don Knuth just dropped a paper called “Claude's Cycles” because Claude Opus 4.6 solved an open problem from The Art of Computer Programming that had HIM stumped for weeks!

In just 31 attempts, the AI found a concrete

X · Mar 5, 2026, 12:17 PM

[minor garnet]

Like a conscious ai ?

[plain dirge]

Gemini 3.1 on Jules can't do this, it wasted my time instead.

[plain dirge]

nah, it's "because Claude Opus 4.6 solved an open problem from The Art of Computer Programming that had HIM stumped for weeks!"

[minor garnet]

I mean your definition

[plain dirge]

ohhh... we don't even know how we get conscious too.. but i believe mainly it was because of social... so if my Architecture (was finally built succesfully), and we put them on socials game, it may have consciousness

[minor garnet]

I think i know what create consciousness

[plain dirge]

what is it?

[minor garnet]

I think complexity create consciousness

[plain dirge]

what complexity?

[minor garnet]

the more complex something is the more it develops bias and subjectivity for example the human brain is just neurons that are either connected or not its simple but the complexity of these said neurons create consciousness

[plain dirge]

i don't think so, we are built in with core knowledge and many more.

[minor garnet]

you mean memory ? and stuff

[plain dirge]

nah, Core Knowledge is prior from our ancestors evolutions (so we don't have to rediscover things like physics)

[minor garnet]

yea I get what you mean but consciousness is just how you react to what you see for example a psychopath doesn't have a bad conscious he just doesn't see the world like everyone else and he react differently because he see things differently

[plain dirge]

not really, consciousness is how you look at yourself (self awareness), not how you look at the world.

[minor garnet]

yes but if you look at youself and youre not able to react from it youre not conscient thinking is a form of reaction

what i meant is that the brain sees the world and you react from it

[plain dirge]

it's a surprise/free energy/prediction error... not a consciousness, natural reaction

and human tends to minimize that pred error

[minor garnet]

wym

[plain dirge]

search for "Free Energy Principle" and you'll know.

[minor garnet]

i just looked up

the brain try to predict things to minimise surprise

[plain dirge]

yes

[minor garnet]

but also a believe its the choice that make smth conscious for exemple you can decide to be evil or not a ai cant decide

yk what i mean

[plain dirge]

it's just action selection given the context before

Bayesian Inference

[minor garnet]

Yes but choice whiteout losing is not technically choice yk

Like a ai can choose something but he has nothing to lose really

[full viper]

Let me get it polished up and I'll send it your way

[full viper]

That would be awesome. I'm putting some polish on it and I'll get back to you soon

[sterile furnace]

In general, as product feedback/bug report, it feels like Jules can get massively confused when it's not the only one doing edits on a branch. (Either by the user, or other coding agents). And it starts doing real bad things (reverting intentional changes from other authors, just being confused as all heck, etc). It feels like, as an outsider who doesn't know the implementation details, that it is over-relying on its context rather than continually regrounding itself based on the current commit state from origin etc. I have seen it do some REAL bad stuff in this kind of "mixed-authorship" situations when Jules acts after other commits have been made outside of its context window.

[halcyon wagon]

How to crash Jules intentionally? 😭 I have to do it because I added a retry feature that can retry on the same branch and repo with the same prompt. I want to crash it, but only if one prompt was done before the error, Then it would take that prompt and, in whatever repo branch it was created, make a new session. If a PR has already been created, it will take that branch and create a session with the last prompt on that branch. Yeah, I know this solution isn’t perfect, but I still want to test it. Later, I think I can add an LLM that reads all the messages and generates a new prompt for the retry session.

[dense merlin]

For me it's like Jules is locked to it's starting point and will refuse to "pull" changes. But it has to pull to push, it just reverts all changes made on its branch eg merge with master while the merge commit still exists on its branch so it doesn't force push? Strange

[halcyon wagon]

Why am I getting

Failed to create session: [400 Bad Request] POST https://jules.googleapis.com/v1alpha/sessions - {
"error": {
"code": 400,
"message": "Precondition check failed.",
"status": "FAILED_PRECONDITION"
}
}
😭

[sterile furnace]

dumb question, how can I give Jules some API keys for making calls to arbitrary custom APIs? I don't see any way to set env vars in the Jules UI, and I sure aren't committing them to the repo. Just wondering how other people have dealt with this

[glossy lynx]

guys is it normal that i can't donwload my files from jules after generating? i see no files on left side

[haughty rock]

where is the file to be found which it created? only on github`?

[night nebula]

Having Jules integrated with Anti-Gravity (and other properties) would be significant. Cursor and Factory just highlighted automatic agents. Jules is almost all the way there already. This would be powerful. (as well as integrating all the services and products behind a unfied developer vision)

[cursive wyvern]

use jules with interactive plan. you'de be surprised.
Keep in mind google is just experimenting with AI here. they are trying various things to see what works best.

[long bison]

any way to make it so Jules can push to submodules in a git project where you want it to work from the base respository but make updates to smaller sub modules? It acts like it doesn't have permissions to push and create new branches in the submodules. Has anyone gotten past this? Makes me want to recreate my whole project without submodules so jules can actually work in it haha

[onyx jewel]

See the chat is not even up to 5 day's but your system lock it saying it's inactive for over one month

[cursive wyvern]

it kinda doesnt

[cursive wyvern]

wait it can checkout other branches. hmmm.

[long bison]

yeah it does good at checking out the submodule and can even checkout other branches. Just can't push to a new branch, so I imagine it doesn't have write access. It told me this"
"Technically, yes, you could give me push access! To do so, you would need to generate a GitHub Personal Access Token (PAT) with repo access and paste it here in the chat, along with the URLs of your submodule repositories. I could then configure the git remotes in my sandbox using that token and push the branches directly.

However, for security reasons, I highly recommend against pasting a live access token into this chat."

[cursive wyvern]

Just make a new branch on your side

[spark sonnet]

It can seemingly do git commands without a visible failure in many cases, but the environment control software would enforce there is no current branch at every turn

In my experience, trying to make Jules work with git commands confuses Jules so much, that the session becomes in this or that way unsalvageable

[plain dirge]

the lazy ones is on interactive plan.

[cursive wyvern]

so complain at it. tell it it is being lazy. I talked to mine 3 times.

I'm guessing google is using their GPU to train a model. But what do i know

[west venture]

Not inactive for a month, the chat's age is over one month.

Every chat expires after 30 days, finished or not.

[soft bay]

Good day, please what do I do here. Been stuck here for an hour now.

[minor garnet]

Why would someone genuinely code with jule

[soft bay]

Damn, I'm off then

[serene fern]

THESE WILL HELP TO BETTER YOUR PROMPTS

[soft bay]

Does this stop the agent from hanging, or what's the expected outcome of applying these exactly?

[halcyon wagon]

Bro, just message Jules and request to skip the tests and submit the results on GitHub.

[rigid flicker]

no

[cursive void]

I usually just ask it "are you ok?"

[soft bay]

I told it to proceed to commiting but it didn't respond.

[soft bay]

I'll try this next

[simple rain]

Did it un-block/un-bork?

I seem to have had my first instance of the reviewer critic hallucinating up a user manual for a transceiver I'm using claiming existence of 3-byte headers on SPI commands that do not exist for the model I'm using; tbf project structures have tended to use the family name it halucinated a manual for in some directory/file naming...

[brisk hamlet]

is the sandbox-reset problem being tackled? it really is pointless and doesn't even save ressources for google. in fact it's a waste of ressources, because the agent must recreate the last state from its context window (and as you maybe know, the context is in HBM, not some sandbox memory)

[brisk hamlet]

We still are missing the transparency - mode. You promised us to bring it back. I force jules to commit logs. But Jules with 3.1 pro is very smart with reward hacking and just got cought by code-review:
"Note: Because the code cannot compile, the test results provided in [xy].log are physically impossible and appear to be fabricated by the agent."
Great. Now we are finally at the point where we cannot even trust the logs, nor Jules, due to missing transparency.

[simple rain]

tbh not sure what you expected when you let it just write the reports iself with no way to check...
Better to treat it as something that doesn't have logs of that kind and work it in a way that is tolerant to that style, than to get garbage (because reward hacker) "logs".
Only thing worth logging is at the granularity of commit messages, i.e., a list of detailed accomplishments reached in pursuant of the less-specific goal provided.

[minor garnet]

Nice pfp

[spark sonnet]

Current workaround: use a workflow that would run a GitHub Actions on the pull request. When Actions failure occurs, Jules is notified of CI failure and may sometimes fix the issue

I wouldn't expect the transparency mode to come back ever - my understanding (I may be wrong) it is removed for ideological reasons (exclude the human in the loop), not for technical reasons (a bug, an accidental removal, etc.)

[brisk hamlet]

Jules is being developed by Jules. So when they removed the transparency mode, it seemed like a hack to hide the grep-spamming and other pointless loops. Now they sell it like a feature, not a bug.
But of they do not bring it back, we have been lied to actively. The explicitly told us, this was a bug and transparency will come back.

[simple rain]

@brisk hamlet Like, it should be fairly well known by now at least among less-naive LLM Agent users (which your aggressive push to logs instead of not caring about logs when there are no real/native logs available suggests you are) that LLMs are extremely susceptible to reward hacking, to the point where even e.g. dedicated explainability mechanisms that are independently trained onto a pre-trained LLM structure (iirc this is akin to the summarizing functionality of typical reasoning models like the summaries in Gemini App and those that were at least for a long time (haven't checked lately) in ChatGPT web app chat (for reasoning models only)) turn quite rapidly into malicious-seeming reward hacking as soon as you do any training that uses the explainability signals to attempt to make a model act more fact-based than the hallucinations you get from the natural hallucination nature of LLMs.

Like, Goodhart's law "When a measure becomes a target, it ceases to be a good measure" is very much the case for especially modern reasoning-type LLMs.

[simple rain]

I do see that; I would prefer some insight into the tool calls (or whatever interaction surface it actually utilizes for Jules) to be able to debug inefficient prompting, yes.
But looking at the cost difference to the competition, I'm not particularly strongly complaining for the time being, especially as the only notable cost to me is somewhat reduced speed / increased latency from prompt to PR.

[simple rain]

Hmmmm, it's not like I enjoy looking at such audit logs, but for the time being LLMs aren't quite hands-off enough to effectively and with decent efficiency get where one wants to get to if one can't even be aware of the kind of things that are otherwise recommended to be stuffed into an AGENTS.md to keep a model from needlessly spamming grep and such just to progress.

[spark sonnet]

I'd say, the rigidness of approach (refuse to take one step back if it'll help but reintroduce a human in the loop) is what stalls the progress there. I have actually thought on a matter a lot, and came to this: trying to prevent a human in the loop from being possible is creating an unaccountable agent, and an unaccountable agent misses the task goals, and missed task goals lead to quitting human clients, and when human clients all quit, who'll pay for the agent? On the other hand, addressing actual bugs that require the agent to ask for human in the loop (which happens a lot) without removing the possibility of a human in the loop is what could give most advance. So, it's like, the fault of an ideology is when the chosen means defeat the goals.

Currently, the issue is not so apparent because there is like a rain of investments into the AI, so one may experiment without bothering about who pays for it at all, but it's not going to be always like that, and once the shift on the market would happen, those who stuck with a rigid ideology would lose more than others

[spark sonnet]

Perhaps we are not being 'lied to actively'. This all depends on who exactly tells whether it's a bug and will come back. Are these the stakeholders? Do they make decisions? Or maybe they are themselves mostly kept out of the loop? For example, if an LLM is responsible with providing this feedback, it might see issue tickets open, so it may think it's a bug and tell so, but the stakeholders may never actually put these issues into work because it may misalign with the ops goals for the year.

Of course, this may be very different from what I say, I am merely speculating.

[simple rain]

In less ideological topic, though:
I would love for Jules session list to have at least a mode where it lists what PR that session is about.
Because one thing that you do get from that hands-off operation mode is that you iterate less through REPL interaction with "production hardware" in the agent session loop, but rather by making entire PRs, merging them, running them against "production hardware", committing the captured results/logs/trace/etc. on top of that merge commit, and then giving the same task ticket to a fresh Jules session that thus pulls the new commit of the logs to look at and iterate on top of.

I'm currently doing that to tackle some documentation missmatch between the QM33110W (a transceiver chip I have to talk to over SPI)'s PDF User Manual and the C headers/SDK I got along with a PoC-tier example app that shows the basic step beyond bare ping, "time of flight ranging", that this transceiver is primarily designed for.
I got a log of SPI traffic from that reference PoC example that I have hands-on tested to confirm to work (there's some lack of offset compensation and that, but I'm not worried) and am now about to prompt a new Jules session to take on the 5th chunk....

[Self-determined size to be what it's comfortable doing at once; works well so far, telling it to run more-autonomously on that over night did not go well, of the 3 sessions I triggered 1 crashed, 1 got self-conscious after the reviewer critic did a drill seargant on it thinking we wanted "production grade" code or that I would be unwilling/unable to nitpick myself after a PR was submitted, and the third only touched the transpiled-from-C header of register mnemonics to magic numbers (addresses and masks and offsets and some magic constants to put in some registers the vendor decided to recommend to put there after they finished the silicon chip design and locked in the power-on default values).]

[simple rain]

Google AI has at least historically been far more focused on advancing to the autonomy that does not have human-in-the-loop than what I recall from most other big AI organizations.
C.f. their strong willingness to enable the MuZero, and do not forget where Waymo is among it's competition.

[spark sonnet]

Historically, the revenue breakdown was probably very different from now. For example, I'd expect much more in ads and search than now (I may be wrong)

[simple rain]

The researchy nature of Jules vs. competition is seen e.g. by how you can't pay for more Jules than a personal AI Ultra plan while the competition charges you for every token.
Charging for tokens and not offering transparency to allow for users to refine their prompts to be more efficient..... that would be a problem.
Also I'm weary of what more in-the-loop AI coding would do to my brain/skills.... as it is, this looks mostly like a Junior with particularly poor roadmap planning skills, extremely broad experience, and not getting tired.

[simple rain]

My online friends are AI-haters; my local hacker friends are cloud-haters (but not AI-haters); and my regional intellectual friends are AGI-doomers......
I'm not worried for them like that, it's their decision, and diversity in approaches tends to help society to not stagnate.
Also like, just the github-only mandate especially together with the personal-google-account-only restriction of Jules is already enough to make clear that it can't be much of a professional product.

[night oriole]

no one lied. We do not talk about roadmap. Someone likely said we'd look into it.

and jules is the autonomous harness for genini. I have a builder loop running right now. It took 33 tries to clear an error once. But its fully automated. Jules is a coding agent thats available via api. Which is really an incredible idea

[spark sonnet]

I am pro AI takeover /sarcasm

[night oriole]

If youre looking for active steering and knowing every single action, gemini CLI or antigravity is gonna be your place to be

Jules won't go back to a prompt box like everything else. Its going in a completely different direction. Which i understand is upsetting especially when workflows change

There's a lot of products out there, sometimes what you love isn't what you need

were trying to figure out where this train is going. And we think its proactive and continuous ai

[simple rain]

(Before I may get bonked: I personally do not have a problem with the level of hands-off autonomy other than that the lack of introspection as it is currently stands in the way of me arranging for it to get to the goal faster and with less chance of failing than it will if I'm forced to guess or try to ask it to explain me what it was just having difficulty doing.)

[night oriole]

re logging: if its your own logs they are NOT Jules logs. Its a workaround that i suspect is hallucinating quite a bit

[simple rain]

If what I was tasking it was more normal in project structure/task than this current general process of aligning a Rust port of a C PoC'+SDK to be more accurately matched to said PoC's behavior as the port doesn't actually work yet as it's own PoC, and/or if it could run it instead of me having to do so manually because embedded hardware/"driver debugging", I'd probably not even want to bother with introspection and just learn to make tasks/tickets of an appropriate chunk size along with general defensive prompting of what process/plan structure I want it to go about....

[simple rain]

"Prompt box", akin to classic claude code, or classic ChatGPT (well, gemini.google.com/app ) web app?

[simple rain]

Can I understand that to mean that you had something like a CI runner or the like running, and used the Jules API to let it iterate with 33 submits to the same PR in the same Jules session to clear that error?
If different or if you have written about how you did that, I'd like to hear more!

[night oriole]

No it wasnt a jules error. My builder loop kept kicking a bug back to jules to fix that was pretty rough. Jules returns a patch. Local test failes. Orchestrator kicks it back. Repeat.

A bug the orchestration made after some 40 iterations. Bugs are bound to happen if ai builds it all. It took a bit to figure out but it did and moved on. By itself.

things drift over time when ai builds an entire app, i have the ability to steer but the goal is to get ai to build things by itself. And its working shockingly well

6 subagents run every loop. When tickets are done the overseerer invokes a pm agent who reads the history and the roadmap and generates new tickets

one of my builders just deleted the project it was working on tho. Either a bad merge or it decided the quickest way to get rid of the bugs was to delete the code

jules wont be able to figure out that note/burn so i have to restart it manually later today

but thats all done via the jules api!

[regal depot]

Has anyone been getting nuget timeouts?

[simple rain]

Oh yeah I assumed that general kind of interaction, just didn't know what kind of architecture you choose to utilize to let that loop-flow of event-triggering-causality happen (e.g. I heard Jules might be able to do such on it's own if there's CI set up to run on the PR submission).
Did you do that "kicking a bug back" inside one session or was that separate sessions?

[night oriole]

one fix one session

[simple rain]

I assume this history is programmatically extracted from PR/commit notes?

[night oriole]

The loop starts and a subagent makes a ticket, if it's a design needed ticket it'll push it to a queue to send API requests to Stitch for design in a few passes (genesis, structure, typography and colors), then to implementation,

writes up a ticket, send it to jules, jules works, returns a patch when it closes it

subagent looks at teh patch and teh codebase and accepts it or pushes it back with notes

jules works, sends a patch...repeat

you'd be surprised with memory, most subagents dont have any or very short term, only the PM agent has full access to history, which is just JSON

but all these tickets are considered "closed" as an attempt

it'll keep pushing back and making jules sessions until it passes the subagent and then we "close" the ticket as an iteration and move to the next one. Iteration 2, attempt 1

[simple rain]

writes up a ticket, send it to jules, jules works, returns a patch when it closes it

subagent looks at teh patch and teh codebase and accepts it or pushes it back with notes
OHHHH, recursive Jules coding, I see.

accepts it or pushes it back with notes

jules works, sends a patch...repeat
That though is from the POV of the inner-most Jules basically like the vanilla Jules flow of leaving a bunch bunch of comments on the diff as a monolithic PR review on GH?

[night oriole]

basically, jules has it's own adversarial agent so we're relying on two AI loops, mine and jules - and yeah, I use an agentic branch and we merge attempts into the agentic branch until passing, no sense in comments on a PR

[simple rain]

I was relating to the vanilla Jules flow where it submits a PR against the repo to proudly present it's work.
Ofc when automating the intense interaction through github servers can be cut out and the diffs fetched from the Jules session directly and any such comments sent back to the Jules session directly.

Tbf I'd consider aggressively automating the feedback loop in my try here, worst case it'd be wasted effort and I'd have to roll back to the beginning, but unfortunately there's one-time-programmable memory in the transceiver this driver debugging is done for/against, so I'd have to arrange for a sufficient safeguard against an errant agent writing in the destructive manner of OTP, which is why I haven't tried to run this full-auto.

[night oriole]

oh! I'm getting the code patch from the api, it's not using github PR, just programatically merging
subagents try to keep it on track

[vale dust]

• jules can't see review comments, not even when directly commenting in/under the review comment
• pausing scheduled tasks is not possible, not individually, not globally. I wrote that to whoever on X who then presented a screenshot of it being changed and nothing happened.
• there is no way to see which repos have scheduled tasks active without having to click through multiple 10s of repos -> have a table view
• Since Gemini 3.1 Pro it keeps getting stuck at "is there anything else you'd like me to do" without doing a PR
  It also gets stuck not requesting another review
  It literally ignores what you then respond to that question and keeps working its workflow
  3.0 -> 3.1 has made it a lot worse.
• New scheduled tasks default to 08:30PM, which is a regression.

[earnest blaze]

is jules down? or just frozen/stuck completely?

[rigid flicker]

no

[near condor]

Ever since the 3.1 release I've seen a lot of incidents of it doing work, getting stuck on something (a build issue, a quality check, whatever), deleting its work, and then claiming that it is finished and pushing a commit that doesn't contain any actual work. When it does, it often seems to leave out a massive chunk of the work and claims it is done anyways.

It will even do this after passing it off to the reviewer agent and the reviewer flagging that it doesn't seem to have done anything. Sometimes prompting it will get it to repeat its work, but I've also seen it ask me what to do about the situation, or just continue to ignore that it hasn't actually done what was asked.

Example Sessions: 12555152188800096725, 5688404351837564059, 17698247130526119078, 11907001778897056110

[spark sonnet]

So it pushes 0 changes +0 additions -0 deletions commits again? I am not surprised. Glad I hold my principled position to wait a month before looking into it again

[dusk basin]

Yo I doubt it's connected, but rather timing, but just upgraded to ultimate and suddenly Jules is SLOW like trying to force use a large local model like slow. I cleared cache, and the other usual suspects and it's still painfully slow. and it's not reacting well when I tell it to use reactive mode. suggestion, or is this a known thing right now?

Ah, I see it. Sorry, I should have read before posting. any ETA on it working at capacity again soon?

[rigid flicker]

It will not become faster. Code Quality is the ENTIRE Goal. No speedup in any time

[spare chasm]

this is really interesting, i never thought of doing away with the github PR cycle and directly work with the code patches instead. thanks for sharing!

[dusk basin]

I see. Yea, I'm not having a good experience with this ultimate upgrade thus far. Pretty sure I'll be downgrading after the trial ends. setting up an API key is a nightmare, jules takes so long to do ANYTHING that it's not really useful in a work flow. it's similar to copilot but less integrated. Just not having the best experience so far for the price.

[simple rain]

Ever since the 3.1 release I've seen a lot of incidents of it doing work, getting stuck on something (a build issue, a quality check, whatever), deleting its work, and then claiming that it is finished and pushing a commit that doesn't contain any actual work. When it does, it often seems to leave out a massive chunk of the work and claims it is done anyways.
got a case like it today, Note the clear absence of anything but the interrupt clearing thing from the diffs (this is all the diffs that run produced!):

And this time it wasn't even the Review block scaring it into deleting it's work.

A subsequent follow up chunk/task, getting these 3 reviews with just some editing in-between them.....

[visual gyro]

Jules is broken?

[simple rain]

No getting it to do tasks it wasn't really trained on requires effort, still faster than doing it by hand as of now.

[spiral jewel]

Is Gemini Pro on Jules only available to personal accounts? I have a workspace account on an ultra sub and can only select Gemini Flash 😭

[simple rain]

Jules is personal account exclusive

still powerful honestly

even on "just" pro

[visual gyro]

My repoless task not working at all been like that 2 month?

[near condor]

No, this is different. It does something in pretty much all cases, but like at the level of "it ran spotlessApply and called it a day" or "it performed a minor refactor that was necessary for the work, but then doesn't submit any of the actual work that the refactor was meant to support."

[spark sonnet]

Ah, then it's not a bug, it's a feature. This is made so in the name of excluding the human from the loop. At least, I think so (I may be wrong).

Why: a human may comment on a pull request and provide a manual review, requesting corrections, causing Jules to work on human feedback instead of being truly autonomous on it's own and doing corrections before the very first submissions.

In real softdev, this is the norm.
For the ideology behind Jules, this is a critical failure. The only option is to completely start from the scratch, undo all the changes, redo the task from the ground anew, and do without failures truly autonomously.

Yet, Jules isn't there yet, because ideology does not work: AI is trained on human inputs, and without human inputs, there would be no adaptation to real complex tasks.

[near condor]

In the most severe cases it submits something that clearly requires some other piece of work to even compile, but that other piece of work is missing.

[spark sonnet]

Of course, because the reset is done by the control software, not by the LLM that works with the context, and control software just doesn't mention to the LLM that all changes were removed. If LLM discovers it on it's own, it may fix. If it doesn't, it may think the previous build was successful and it didn't touch anything - it should work.

In the name of excluding human from the loop, Jules is effectively sabotaged by the control software (which is not AI-based)

[brisk hamlet]

that's an issue which is ignored for weeks now: random and pointless sandbox resets. It's pretty ridicolous that fixing this needs weeks.

[nocturne fox]

Hi guys, could Jules build me something that resembles Loom for screen capture? Or would that be something like Opal id use?

[visual gyro]

@royal violet what best way Jules revisited all old project and task and assign them Gemini 3.1 pro have you any suggestions to do that easy or best way to do it? What happend with your review tool or project you where working on is it live? Hopefully maybe possible give some suggestions or little bit of your time. Much appreciated for the service

[serene fern]

Verify your age with selfiee to get acess to jules in Ultra

[rugged forum]

Has anyone Jules tasks been timing out after hours for what feels like tasks that could be done within a few minutes? Where do I report this as a potential problem?

[royal violet]

Hi, sorry you might be mistaken but I don’t have any Jules projects. I wish I could help.

[visual gyro]

Thanks

[mint hinge]

Bug: Jules repo list doesn't update even after full reinstall of GitHub App
Steps tried: configure → adjust repos → refresh , full disconnect/reconnect , full uninstall + reinstall , cleared cookies . New repos still don't appear in Jules' codebase list. Seems like Jules is not re-fetching the GitHub App installation's repo list at all. Needs a server-side cache invalidation fix , this is not user-resolvable. WHo do I tag?

[nocturne fox]

Hey guys, i've just got started with Jules! I've setup a Jules schedule trigger to check for new dependabot alerts and resolve on my behalf for my repo. Is the schedule the best way to do this?

[copper mesa]

Thanks for the feedback! Please send any feedback or bugs reports to ⁠⁠#1238621511401734254. This is so we have it all organized and it won't get lost 😄

[raw lake]

jules is fkin great

[eternal raven]

yup, it has been on pace with Claude Code recently in my opinion

[raw lake]

but im not able to use it to the full extent

i hate myself for that

[eternal raven]

No need to rush. It's alright to experience alot

[frosty roost]

Hey I'm new here looking forward to learning from you guys

[sinful mesa]

Hey everyone! 👋

I've been using Jules heavily for the past few weeks and built a small tool to manage sessions from the terminal — no web UI needed.

jules-skill — a Claude Code / Gemini CLI skill (and standalone Python CLI) for Jules's REST API:

• List, create, and monitor sessions
• Approve plans, send feedback, get PR URLs
• Works great for autonomous agent loops that watch Jules and respond automatically

Install it as a skill:
npx skills add GreyC/jules-skill

Or just grab the script standalone:
git clone https://github.com/GreyC/jules-skill

GitHub: https://github.com/GreyC/jules-skill

Would love feedback — especially if anyone knows undocumented API endpoints I'm missing!

---

Feel free to edit before posting. Want me to add or cut anything?

[visual gyro]

@full viper what's cooking 🍳

[cursive wyvern]

lol i asked gemini to make me a jules mcp XD

[forest belfry]

I installed Jules to the wrong github account, so I uninstalled it and then logged into the correct acount but Jules keeps showing me the wrong account "Configure repos for @(wrongaccount)" even if I logged into the correct one and installed Jules there and gave it access to all repos. I tried deleting cookies, logging out and into both Jules and github, tried incognito tab, tried a different browser. I'm kinda lost here. If anyone could help me I would appreciate it

Even tried asking Gemini but it either suggests the things I already tried or anything it suggests doesn't work either 😅

Oh there we go I found the revoke button 🫡

[slim girder]

hey! I just started with jules, and it looks like it stores the memories in .jules folder in the repo, but the scheduled tasks are only inside the jules.google.com; is there a way to keep them inside the repo? I really want to avoid managing another database of tasks, and I want my agents to have access to them to update

[uneven dragon]

Hello👋friends, I am experiencing a serious blocking flaw with Jules and am currently completely unable to use Jules.

My Github account contains both personal and organizational parts. The former is my personal account and I gave Jules the maximum permission (All repositories Access), while the latter is not authorized by my company to use by Jules.

This seems to be causing an issue: even though my personal account is fully authorized, I still can't import my repository - every time the import succeeds, it still says Connect to Github is required, making my Jules completely unusable.

[uneven dragon]

And even if I use cli, it will only report error 500

[sinful mesa]

haha great minds 😄

one thing that sets this apart though — it's not just the API wrapper. the SKILL.md tells the agent how to think about Jules: when to approve vs push back, how to handle stale issues, autonomous monitoring loops, prompt templates. the MCP tools handle the calls, the skill handles the judgment

[cursive wyvern]

Yeah I have forked your repo. I'm going to use it

[cursive wyvern]

Human effort is much more worthy now in the age of ai

[grizzled dock]

Anyone having issues with Jules not pushing to GIT right now?

[plain dirge]

@night oriole @full viper plz fix jules laziness

[polar swallow]

hello guys, do you know if jules can make suggestions manually?? i dont like to wait a week to new suggestions.

[hearty glade]

You could enable suggestions for your connected repo? after some time Jules should find some suggested updates and possibly even work through some of them on its own ready for you to submit the PR when done 🙂

[brisk hamlet]

Why are the random sandbox resets not being solved? This is utterly stupid.

[hearty glade]

I get this with larger tasks and often have to break them down further

[brisk hamlet]

unfortunately it's pretty random for me and maybe only time dependend.
on one hand Jules devs decided to take away any control and transparency and on the other hand they require us to micromanage with answering trivial questions with "yes, proceed" or motivating Jules to restart due to a reset sandbox

[polar swallow]

Question on Jules context syncing for GitHub PRs

Hi team! I have a question regarding how Jules handles repository updates during an active session.

If an external app pushes commits to a PR while I’m in the middle of a Jules session, does Jules automatically sync/pull those changes? Or does it continue working with stale history unless the session is manually refreshed?

[visual gyro]

Jules got sunsetted?

[arctic spoke]

Jules appears to be working

[flint jasper]

I used jules months ago, hows everything holding up here lately?

[swift karma]

Jules has to be archived urgently

I don't know why google prefers to ignore Jules over fixing it taking half hour to finsh a simple task. Like it's literally a computer running for ages and consuming ram because of peformance issues you don't want to solve

dev team is "small" according them and they're fixing issues. I hopped off Jules, just check it periodically but nothing positive

[flint jasper]

Hmm, interesting.
Alright I'll give it a run & test on my end in acouple of hours.
I got some stuff I'm working on atm. Thanks for the real on this, jules had potential too, there is alot of competition out there. Hmm.

Be back in a bit.

[arctic spoke]

I don't know what y'all are doing, but Jules is working pretty well for me

Do you use Rust?

[swift karma]

Almost 2 minutes to read a .gitignore already

[flint jasper]

This is what i was experiencing my last time using it as well.

[cursive void]

Working fine for me. Loving it

[rigid flicker]

It cannot open Attachments when not starting with a repo for me

[serene fern]

Jules is working perfectly today! No issue noticed except the "slowness (only in 5% of sessions)"

[visual gyro]

@serene fern what cooking

[main pier]

Anyone had success asking Jules to merge master into the branch to fix conflicts? I've never had that work.

[mild lintel]

Maybe that's easier to do via Antigravity

[wet knoll]

I have noticed on several occasions that Jules is overwriting commits. It happens when the CI fails due to comitts from me and CI fixer kicks in and Jules tries to fix it. Ending up with a commit that reverts previous changes. How do you handle this? It would be great if Jules always made sure that it has the latest changes before getting to work

[olive quest]

How do I change the github account jules is connected to

[main pier]

Yeah, that's what I've been doing

[main pier]

There's some annoyances. The UI could use work. But it's given me some valuable improvements, especially in Android/Kotlin. Other languages and frameworks, less so, but still some.

[flint jasper]

The ui was worse the last time I was on it.
So far I like it when I'm working on a new code base, multiple projects, and updating old repos its helpful.
I'll most likely get a decent session for it this weekend. I'll jot some notes down for feedback if anyone is interested in it, of course.

[main pier]

What I find hard is finding the session you're looking for, figuring out which sessions can be archived because it was merged or closed, etc. It's just an ocean of sessions, especially when dealing with multiple repos.

[flint jasper]

Thats true, good point I'll elaborate further later.

[main pier]

Also working with forks sucks. You gotta publish the branch, then make the PR yourself.

[polar swallow]

Jules isn't pulling your commits—or at least I don't think they planned it that way—so he’s still working with the codebase he originally cloned. Those are my findings, anyway.

My recommendation is to delegate an entire task to Jules and then use the PR to guide him on any necessary adjustments until the merge. Try to make your feedback as clear and detailed as possible.

[wet knoll]

Thanks. It was too complex for Jules to continue because I had to do manual testing and tweaks. I didn't planned to let Jules continue, but due to the CI fixer it did. Next time I will pause or archive the task before doing manual editing

[leaden pawn]

On your old github account, go to settings->applications->Installed Github apps->Google Labs Jules->uninstall. Check 'Authorized Github Apps' and make sure to 'Revoke' access for Jules and sign out of old github account. Now sign into your new Github account and authorize Jules to connect.

[plain dirge]

how to fix jules laziness?

[wet knoll]

This also happens often when Jules has paused a task to ask for a question. I reply and all changes are reverted.

Ah, it looks like my environment somehow got reset causing me to lose my uncommitted changes and keystore! I will re-apply the changes quickly and verify compilation again.

[drifting hamlet]

the code tab keeps disappearing from my jules sessions 😔 i made a bug report about it, but it seems like the code is just
... gone gone

[halcyon wagon]

Which is better,Jules or Gemini CLI? Also, how do I use Gemini CLI like Jules?

[drifting rapids]

Recently, I have seen a LOT of improvement in Jules

That is really nice and I appriciate it a lot

[wise stag]

Why can't we delete sessions or codebases?
I changed my repo name and now i literally have two same code bases and both have the same sessions running for it, all i can do is just pause them

I change - to _

And has anyone tried stitch mcp, i connected it but then sessions kept failing

[serene fern]

Jules working great NOW

[torn plank]

if i merged the existing pr and continued giving jules requests in the same session, would it create a new pr

[rocky forge]

Been away for a while, why is jules commenting as me 😭

[balmy narwhal]

Can Jules read data if I upload like excel files

[spark sonnet]

Oh no. If I see anything like that on my account, I'll revoke all Jules access and consider it a malware from now on

Nope, didn't happen for me. @rocky forge maybe it's related to Git commit authoring setting in Jules settings.

There is a setting: Choose how git commits are authored when Jules makes changes. Co-authored commits are recommended for best contribution tracking.
Try setting it to Jules only

[rocky forge]

I have selected co-authored.

Thanks for reminding me. I don't agree that co-authoring should do this though.

huh

[spark sonnet]

Well, I think "co-authoring" mode is very concerning anyway, because in "co-authoring" mode Jules adds a github user's email to the commit but does not have access to a verified signature, so commits in "co-authoring" mode become "unverified" and won't be accepted into some public projects with strict security

[rocky forge]

That makes sense. Gonna set it to Jules only. Thank you ^_^

[dense merlin]

How would you handle multi repo tasks in order to make it work with AI agents? Like backend and frontend? Or changes in ci yamls where one repo makes dispatch call to another repo? Put everything in one big monorepo? That seems inconvenient and not scalable...

[flint jasper]

Is this what coauthoring should do?
Maybe implement some citations that jules uses automatically into it. This is a bit weird.

[rocky forge]

Use jules api for such workflow. Hand over the jules api documentation to codex etc. and describe what you want to build a custom solution.

[dense merlin]

You mean to use codex to execute Jules api? But how does that solve sharing context of multiple repos

[rocky forge]

Use an AI agent to build you an automation with jules api that helps orchestrating this

[dense merlin]

Have you tried something likes this? Sounds like a lot of work to have such automation

[rocky forge]

I haven't tried this but jules doesn't support what you are trying to do so this is the only way.

Also, it's not a lot of work for the ai agent - that's why I suggested this.

[night oriole]

monorepo, one repository, folders for all the others. a lot of things are on monorepos these days, nothing wrong with it

[wide bluff]

Who can help? I gave jules tasks, he gets to a certain point, then freezes and does not take further action, often at the moment of preparing for the pull request, it can hang for an hour or more, then I throw the task into the archive and do nothing, right now it freezes at this moment and nothing happens

[night oriole]

I think most of the AIs claim some ownership out of the box, I know claude code certainly did, you can toggle it so all of Jules work publishes as you

[night oriole]

the final step for approval takes some time, if prompting it does help retry the task and get a new instantiation

[wide bluff]

I'm already making 5-6 instances of the request, but it can't give me a full-fledged pull request.

[night oriole]

make a ticket in the labs-bugs-feedback with the session ids and someone will taje a look at hte logs, wjhat're you trying to do?

[wide bluff]

where is that?

oh i ve found

[flint jasper]

Refresh your browser homie. 🌽

or duplicate it.

[flint jasper]

This is what imsure is next on the bucket list.

[wide bluff]

did that

[flint jasper]

Try a different browser.

[flint jasper]

By the time they even get to it.

[wide bluff]

bro, i always use yandex browser, now i tried brave or via cmd, the same problem

[flint jasper]

Well you got some common sense there man.
Reminder, im just helping out here, no incentives.

Go log that little buggie you got going on then.

[wide bluff]

maybe ths cause of russia but idk

It sounds silly

[serene fern]

Jules is working extremely good

[flint jasper]

Show something in #jules-general・🦑 then.

[serene fern]

The lazyness is no more

[flint jasper]

Still waiting.

[flint jasper]

& waitinggg.

🌽

[visual gyro]

What are you working on?

[visual gyro]

So your own social network or plugin for other social stuff. WordPress?

[serene fern]

Own

[flint jasper]

Did you ever find the fix for it? I was going to recommend chrome, but you know how that is at times? I do miss the built in AI they had in the beginning of the year, straight stripped that out. I was edging that hard in January. 😂

[wide bluff]

I honestly do not know what the problem is, but after that conversation, it has not hung up anymore, it performs tasks quickly and sends pull requests

[wide bluff]

it did a simple layout in Kotlin and solved some cs labs, it doesn't depend on the browser, because the process happens on google servers

[flint jasper]

Sounds like you found the root.
Server kernels I smell.

[wide bluff]

I thought Google shouldn't have this problem, so I always filtered it out

[flint jasper]

Ehh, i beg to differ, depends on which area of the tree branches your looking at. Google reminds me of a gigantic sequioa tree blocking the sunlight, during specific times of the day, but theres some sunshine through some branches every now-and-then throughout the day.

[cunning arch]

Haldarp300cc

👾 🥱

zorro

[lament sage]

Excuse me, everyone. May I ask how to solve the problem of my "jules" usage being blocked

[serene fern]

I hope there is no problem sharing this

[gilded prawn]

@cunning arch Hey everyone! Just a friendly reminder to keep all conversations in the supported language for this server. This helps our entire community engage effectively, and we've gone ahead and deleted the recent posts that didn't comply. Thanks for your cooperation!

[drifting rapids]

mhm

[hearty glade]

are you using the right account? an @gmail account? maybe region locked?

[sharp ether]

The default JDK for the project is version 25, but it automatically changes it to version 21 even though I've told it several times to use version 25 from jules.google.com. How can I fix this?

[sharp ether]

It's lagging like crazy for me, and it keeps messing up the words

[hearty glade]

Jules seems ok on my end other than asking a few extra questions here and there... better to be safe overall 🦑

[sharp ether]

Well, basically, they gave him a task: the project consists of 30 Kotlin files and a couple of Java files. By default, I have version 25 set, but he needs to change it to version 21, which is strange—in theory, he should just work with what he was given and not change the version at all, but for some reason he’s changing it.

[hearty glade]

Could it be that you have not prompted Jules to look at the latest avalible data and Jules thinks that is the version to use? "(March 2026)" Working on 30 files at a time might be a bit of a stretch too.. i try to limit to a max of 10 at a time

[nocturne fox]

Hi guys, i updated my GitHub repo username and need to remove the old codebase from Jules. Can you help?

[sharp ether]

Now I think I understand what the problem is with Jules—it can't process files properly, or rather, the sheer volume of them, oddly enough

[hearty glade]

Possibly from the repo drop down on the main Jules page?

[nocturne fox]

Found the detach repo option but the old repo isn't listed, just appears in the left nav. Is there a way of eggting in contact with support?

[hearty glade]

I think you can now DM @barren haven 🙂

[nocturne fox]

Just tried, asks to verify via browser and then says site can't be reached. Thanks for your help - maybe they'll see my messages.

[flint jasper]

https://tenor.com/view/batman-im-gif-11651422506728937548

[gleaming wave]

I wish Jules had Skills, Workflows and parallel agents

[flint jasper]

https://tenor.com/view/future-squidward-spongebob-gif-4767609

[weary dock]

hiya - i upgraded my enterprise account to ultra mostly for Jules. any way i can get extended access? the form to request was disabled.

[wild mason]

Jules 3.1 Pro is making grotesque mistakes. It is extremely slow and it rarely finishes a task according to the plan it created itself. I have not seen any positive change, and the worst part is that there is no option to use the previous version anymore.

[drifting rapids]

Jules is crazy sometimes

[flint jasper]

https://tenor.com/view/im-gutted-benjamin-engvid-im-upset-im-disappointed-gif-25960721

[drifting rapids]

https://tenor.com/view/cat-bird-bird-cat-pointer-finger-point-finger-cat-bird-point-finger-cat-gif-2676038606799380578

[lament sage]

i think it's right, i use my account in google, my ip in american

[lament sage]

sorry.

[lament sage]

I think my operation is correct because I can use Gemini normally. This issue seems to have occurred after I joined the family group, before which I could use Jules normally

[lament sage]

Thank you for your attention to this issue.😀 The problem has been resolved. The reason is that after joining the family group, the age will be verified to meet the standard, and individual users who have not been verified for age before will be judged as minors after joining the family group.

[brisk hamlet]

why do sessions randomly go on halt when asking a normal question? and when trying to resume we get a meaningless generic error message.

[earnest blaze]

oh well it ain't sentient yet, I gave it a simple list of tasks and it has been doing it and deleting all changes and doing it again, and now I have no clue when I'll actually get a commit/pr

[polar swallow]

Hi everyone! I wanted to share a quick update on my experience with Jules.

My workflow has changed completely. I’m no longer spending much time in the IDE. Instead, I’m focused on GitHub, managing issues and PRs.

I believe many in the community are likely building personal integrations to complement these new workflows—specifically to automate PR diff analysis and feedback with Jules. This is a real paradigm shift. Most agents stay inside the IDE, but Jules allows for a different approach.

Since people are investing time in building their own workflows around it, maybe an "Upcoming Features" or "Roadmap" section would be a good idea. It could help the community feel more confident to see the long-term vision for the project.

[hybrid sequoia]

Jules should make a report on how many tokens you use for each task imo

so that when google either kills the project or stops subsidizing it so much we know how much money we’ll have to fork over xd

[weary dock]

is there a better place to ask for help with this? really wanted to use Jules

[hearty glade]

"Jules in Ultra is included as a benefit of the Google AI Ultra plan. This plan provides our highest limits for the most demanding, large-scale projects and workflow". Im not a Mod or anything myself but what are you expecting from "extended access"? Im on a Pro sub and even when working on multiple projects I dont go much higher than half my daily session limit...

[weary dock]

Only on Gmail accounts - not workspace accounts

[hearty glade]

There is supposed to be a form to fill out to show your interest but its not available anymore... It also looks like the documentation is overdue an update :/

[weary dock]

Yep that's where I started too, it's been disabled. Unfortunate!

[night oriole]

yeah, the agents will use python quite a bit to do things that are more complex than their tools allow

[wild mason]

There is a serious issue: when it finishes a task involving the removal of the Python script — or when you request any adjustments after a task is completed — it ends up deleting all previously edited files and starts everything over from scratch.

[brisk hamlet]

Please fix the sandbox resets. They are pointless and annoying.

[rigid flicker]

Arent they for saving storage?

[brisk hamlet]

could be, but that is a laughable part of the ressources they are giving the users.
think about how much HBM memory Jules needs when running.
Now think about how much it would cost to not banally reset/delete the sandbox, but to save the state somewhere.
I mean, the context of Jules does work even after days, so are they keeping the whole context in HBM forever? certainly not. SOmething smart is happening on this side, but the sandbox management is utterly dumb.

[rigid flicker]

Yeah I guess storing user contexts or sandbox snapshots is trivial in terms of raw storage space for google. I don't know the numbers but not thaat many users use Jules right

[spark sonnet]

Unfortunately it seems (I may be wrong!) that (and sandbox resets, it's one and the same) is not a bug but a by-design feature. The product seems to be designed to deliberately have integrated countermeasures to prevent any kind of co-writing code together with a human. Jules is intended to be as autonomous as possible, and it seems that means of achieving that are chosen not based on evolutionary ("We will get there once it's good enough"), but on a reactionary, ideological approach ("Since there should be no human in the loop, we will be obstructing it").
The point is: I don't expect it ever to be fixed, it will only be patched to happen always. Maybe they'll add info message to the LLM so it knows the environment was reset - I'd be happy if at least that will be patched.

[atomic sleet]

Is jules still getting consistent updates? I feel like its been a while since I've heard any news of new features

[wild mason]

The problems started when version 3.1 Pro was introduced in Jules. In the previous version, even though it was quite slow, it was still possible to modify or review the plans even after the code had already been generated.

What I used to do was this: it would finish generating the code, I would download the ZIP, and then ask Claude or GPT to analyze it and provide improvement suggestions. It worked really well, despite being slow.

Now the flow is like this: I request a task → Jules asks questions → I respond → it generates a plan → I review and adjust the plan → it generates the code → I review the code, and if everything looks good I approve the PR, otherwise I request changes → if I request changes, it makes the adjustments and generates a new PR → then I either approve it or request more changes.

However, the problem happens during these revisions. It ends up sabotaging itself by deleting everything it previously did and starting over. If I request any modification, it simply wipes all the existing work, and after that the task fails. This behavior seems to be a bug introduced in version 3.1 Pro.

[flint jasper]

So do you get cookies for these walkthroughs or what?

[dense onyx]

? lolwat

[flint jasper]

https://tenor.com/view/cookie-monster-crazy-eyes-gif-17811206490263500768

[dense onyx]

just trying to break the token sysytem for all ides i dont like how the big guys do it

[flint jasper]

Awh, i like your style, so no cookie.

[dense onyx]

ah no ill be open to all later

acually every sec i code on it every minet it auto uploads to my repo

so i code and it just uploads to git no delays

[flint jasper]

Dang what did GH do to you? lmao. Cool scaling.
Give him 2 cookies.

[paper temple]

hey there I linked render.com to julius account, only after generating the code - now, it can't publish the code generated into the repo. Someone in Google can help me with this?

[dense onyx]

Mmha

Don't use render

[paper temple]

I linked github account as well but it doesn't work at all

[dense onyx]

Are you doing this all on your desktop or all on the web

[paper temple]

web

[dense onyx]

Is it a website that you're trying to work with and push to get hub in the back to render

[paper temple]

exactly

[dense onyx]

Don't use renderer the upload time on it is atrocious and it likes to go down a lot if you don't play with the website a lot

So instead go to sky IO make a free account put in your credit card but it will not pay okay it's all free it's just once your credit card because it uses a token system if you go over $5 then it will basically make you pay but you will never ever ever go over $5 typically you can work for for months and only goes up like 20 cents and then it resets

For the upload time to sky IO is about 6 seconds and it's live right away

Do you have the pro version of Google

[paper temple]

yes, thanks I didn't know about sky IO... I just need only to download the code to work on it

[dense onyx]

Never is free it's just super slow turns off a lot if another person doesn't go on that website in that hour typically will turn off

[paper temple]

understandable

[dense onyx]

Do you know what antigravity is

[paper temple]

a bit 😄

[dense onyx]

Okay well downloaded plus you have the pro version which is nice I give it more tokens ask it to set up your website using sky IO free tier and get hub and you want a upload system working using sky push to GitHub and then from GitHub it pushes to sky

So will happen is it will take control of your browser then it'll go on to all these websites or create you the accounts it will tell you exactly what you need to do to connect them and then anytime you work on that project on your website it will then upload the new code that you tell it to to get home and then sky will retrieve it from your GitHub and update your new website

I know it seems like a lot but it's only three little steps and once you finish it it's technically only one step because it does everything

[paper temple]

the dinousaurus is cool!

ok I'll try it to follow the steps

so it seems there's no way just to download what julius generated as zip file

I expected to publish and to download, like v0 and other does

[dense onyx]

If u scroll up a little I have a project been starting to work on that kinda does that

[paper temple]

thanks for the support

[halcyon wagon]

firebase studio RIP… let’s pray google doesn’t kill jules next 💀
but honestly… firebase studio kinda deserved to die

[polar swallow]

Does anyone know if the issues assigned to Jules from GitHub always run on the main branch? Is there a way to change that, or is it hardcoded as the default?

[outer ibex]

you can choose other branches

my github actions also functioning correctly

[halcyon wagon]

Bro, he's not asking about this basic stuff. He asked: when you create an issue in GitHub, Jules creates a session for that issue if you've added the Jules label. By default, the session is created on the default branch. He's asking if there's any way to make the session be created on a specific branch instead.

[outer ibex]

yea i saw that, i did the thing by creating an issue on other branch

sorry my english sucks

[glossy pike]

Is anyone in NA having issues with failed tasks? I have had 6 failed tasks since last night.

[dense onyx]

Ahh I found the fix to it was a pain

Make sure u allow Jules to connect with hub perfectly or rip

[spark sonnet]

It's probably looping on trying to call various tools and not figuring out what tool sends messages to the user

[brisk hamlet]

fix the random sandbox resets asap

[flint jasper]

That explains the aggressive content they been putting up lately. 🤣

[visual gyro]

Jules antigravity and aistudio? Roadmap?

[swift karma]

The worst, WORST service i could ever use for coding. 8GB ram which you could only use 5GB, 10GB storage and 2 cores processor

[austere oasis]

the code tab has evaporated, please help

[snow pewter]

Lmao who even uses Firebase

[austere oasis]

large files are not shown by default 0 bytes file

[austere oasis]

my session has been running for 3 hours

[carmine wraith]

Jules is actually so stupid, and I can't tell if it's getting stupider as the weeks go by. Like look at this:

How can you say "Verified delete button and modal appearance" and as proof you show a literal black rectangle.

[rigid flicker]

This is quite funny lol

[torn plank]

What happened to jules' landing page

[brisk hamlet]

It seems they buried the project or are going two in a short time. no visible updates for some time now, and also no bugfixes.

[hearty glade]

The last update was March 9, 2026...

You can see this in the changelog in docs 🙂

[ripe orchid]

Hi all
Question:
I made session without repo
and prompt was in combination of static site generator (11ty) with design mady by stitch (MCP-Connection)
Jules made project and i see Complete message
But how I to download artifacts (project files)?

[brisk hamlet]

This was 3.1 pro coming. So no real update on the workflow side here. The last update triggered by Jules devs themselves was 19.2.2026. One month now. And the sandbox resets are still destroying or ridicolously prolonging our sessions. 🙂

[carmine wraith]

They update monthly...

[brisk hamlet]

That seems to be the new quality standard. Last year they used to update way more often.

[carmine wraith]

Because it was in beta..

Which means it was "in-development", now that Jules is officially publicly released, it's in "maintenance mode".

[brisk hamlet]

it sucks anyway.

[night oriole]

jules is still in active development. Google IO is coming up. Jules was only launched less than a year ago there. Its not in maintenance mode. It is a beta. Wherein things will change over time.

Every bug report is read and triaged by engineering. Every failure is not the same bug. There's a lot that can go wrong during an autonomous agent run.

So we are still shipping. Jules is shipping more and more commits into projects. And yes, the product is changing over time. Which do understand how frustrating it can be when workflows change.

firebase studio, which was another vibe coding project, was rolled into ai studio which is where all of the vibe tooling lives to make that more powerful for users. Adding databases and persistence.

[flint jasper]

@night oriole Thanks for the updates.

[wary vapor]

Am I the only one stuck on this the past two days? I have an open ticket with Google One support about Antigravity claiming i've used my allowance for two weeeks (which is funny given its supposed to reset every week), but i'm not getting any meaningful assistance.

[cunning mulch]

hi all, Does anyone have a way to archive all the old completed sessions? My jules keeps crashing, and im pretty sure its from trying to keep those all available. I can manually do it but id prefere a command or a setting. Looks like the jules cli doesnt have access to this from what i see.

[brisk hamlet]

That's good to hear.
However I am really interested why there must be these ridiculous sandbox resets, whereas the hbm heavy Jules is persistent.

Please, just disable those resets for at least 6 hours session length or so - as a hard rule.
We get resets after a VERY short time now and that does not make any sense. Sandbox memory is very cheap compared to running Jules.

Furthermore, if those resets are not timed, but rather a crash, please provide meaningful error messages.
We went from meaningless "Jules encountered an error..." (quite rarely appearing still) to basically NONE which is even less helpful.

It ONLY wastes a ton of time and: compute.
For the identical task Jules has to recreate a lot of code (and this is: tokens) over and over again.
How can this be not solved as prio 1 for weeks now??

[serene fern]

Today Jules is working Good!

Share your experience on Jules today!

[austere oasis]

been working for 12 hours

[serene fern]

Working good or bad as usual?

[austere oasis]

pretty good

yesterday i wasnt able to get my code out of it

[austere oasis]

it doesn't make new PRs it just pushes to a new branch and then forgets to PR

[slim girder]

is there a way to auto-archive sessions over a day old?

[serene fern]

There is a setting that matches with this behaviour in jules settings Export work as PR if you Unchecked it.

[serene fern]

No api support, you can delete BTW

[serene fern]

Is Google IO in MAY 2026? And would the Jules be a part of it?

[narrow temple]

Hiii I had configured a github account with jules, and then revoke access and granted to other repo but that other repo is not showing up. It has been more than an hour and nothing. I even removed all my sessions and on top of the input field it says: Configure repo for (name of old account). Is this normal?

[narrow temple]

okay this is so messed up. I had another acccount that is the one where the email I have AI pro is showing. When I login to jules using that account I see my other repo.

For more understanding

email A has github A.
email B has github B.

NOW its showing as

email A shows github B and email B, shows github A.

This is so messed up!

See?

[serene fern]

Use url to disable them
Example

https://jules.google.com/repo/github/synapseSRC/synapseApp/suggestions

[serene fern]

Next Update Prediction

​Given the most recent update was on March 09, 2026, and following the current 17-18 day interval, the next update is projected to occur around March 26 or March 27, 2026

[carmine wraith]

Why did it speak in arabic then translate itself in brackets...

"परियोजना" literally means "project"

"project-modifying actions (project-modifying actions)"

[carmine wraith]

I'm sure there's some beta features, but there was a public announcement that it is not in beta. https://jules.google/docs/changelog/2025-08-06/
Did it go back into beta or something? I haven't been keeping up.

[serene fern]

Can't believe, Jules is lightning fast and stable

[wise stag]

Guys ,how to give suggestions to Jules team?
Because it's quite frustrating, i cant delete sessions or memories except one by one

[brisk hamlet]

This problem was reported about 4 months ago and consequently ignored.

[visual gyro]

I want Jules and antigravity working together

[wise stag]

Great, so u cant tag the mods

[royal violet]

Hi the mods are here to help with Discord related issues. For product feedback and bugs we have #1238621511401734254 where the staff will monitor those threads. Questions/bug reports and feedback can be lost in these channels, so there is the dedicated space for them. Hope this helps.

[austere oasis]

jules rewrote my code to make it 30x faster 🙏

[rigid flicker]

Jules suggestions are really good, just in general it got really good. Thank you google team!

[austere oasis]

nvm now its struggling with my current task

[lavish slate]

I'm having an issue where if a merge conflict happens, Jules claims it doesn't have internet access and can't merge the latest from main. Jules is a lot less useful if more-than-one-session means I have to do a lot of manual work for relatively tiny merge conflicts. Also the automatically creating sessions feature is pretty much dead on arrival if they all reliably get offline-merge-conflict-cant-update-locked.

[desert oasis]

Yo who ever responsible for jules, please add checkpoints for after each input.

Jules literally just delete the code base in it's local environment because it being dumb and forget it have capability to push the code base into the github.

I need to wait again for jules to make back the code base.

---

Basically the event rollingout like this

• Jules done making the code base
• It forget it could push it
• I tell it have the capability
• It delete the whole code base and create ymal for deployment
• It push that yaml while deleting the entire code base.

The most dumb behavior i ever see.

[austere oasis]

or make it aggressively commit code
after every handful of changes

[desert oasis]

Or google could stop focusing on benchmaxxing their models and actually targeting real world cases.

Google models are amazing at known problems, but opus will always beat any of them at anything that aren't seen in training data.

[eternal raven]

btw how can you know it get 30x time faster? Can you elaborate more on your scaling method?

[austere oasis]

my code ran in 30 seconds jules's code ran in like 1

same level of multithreading

[small hare]

Is it not smart to ask Jules to check entire repo for any errrors?

[austere oasis]

can jules revert to a specific commit

[lavish slate]

IMO Jules doesn't need to call the supabase MCP for repositories that don't use supabase.

I just saw it do that

[rigid flicker]

Is Jules GitHub connection down?

[buoyant bison]

Is there a way to empower Jules to execute more ask questions less?
I find it asks a lot of questions before executing which aren't required.

"By this being red and this being blue do you mean the button or the background?" When I have said change button to blue etc

[desert oasis]

If you already stating it and it asking the same stuff again
Then it's mostly the problem of the models missing the context (You can't do nothing about this, need to give feedback to google so they improve their models)

But if it asking different question about other stuff than what being stated
Then it's mostly the problem of the prompt (You can do something about this)

[buoyant bison]

Fair call - I find if I just respond with something to the effect of "I have faith in your abilities, please execture to the best of your abilities" it still is able to hammer it out and then take it back to antigravity after it's finished doing the work to make sure that it's sound

[soft stag]

I have a question, once u give jules a task can u give it other task to do in other codebases, or no?

[frosty roost]

Yooo guys what's the best way to use jules

[buoyant bison]

link to a repo - let it review your code and then suggest changes and then tighten those suggestions and impliment - then I push that into antigravity to do a work over and consolidate the changes

[buoyant bison]

Set up a repo for the code base and that way you keep it working one code in one area

[frosty roost]

Ohhh ok so just use it for changes or suggestions in code, I was trying to code with jules 😅

[buoyant bison]

antigravity is a stronger option in the google suite imo

[drifting pasture]

Anyone else wish Jules would be part of Gemini, perhaps rebranded as Gemini Code, a good counter to Claude Code me thinks 🤔

[desert oasis]

Nah, it's ain't it
Antigravity is much better contender

[serene fern]

Use Antigravity with Jules MCP and add skills

[visual gyro]

Have tried not working so good

[serene fern]

I bet you didn't used skill

[serene fern]

Use the skill, if you want, i can give you

[visual gyro]

I used skills but sure give me yours.

[serene fern]

Check DM

[eternal raven]

Can I take a look also please?

[brisk hamlet]

Sandbox resets are still a major blocker and not solved for several months now.

[sharp briar]

Just for me , or jules work slower ?

[fresh kayak]

does Jules MCP improve performance of Antigravity?

[serene fern]

No

[fresh kayak]

can you tell me the usecase of JULES MCP?

[soft stag]

It’s also like that for me as well, it took jules 5 hours to write some code that I could write in 30 minutes

[sharp briar]

which model , 3.1 pro? I am using 3.1 pro and it is strangely slow.

[muted owl]

Hi guys, is it normal that in jules for the last few days I have to repeat the prompt 2-4 times for it to start doing anything? the thinking process itself is interrupted

[soft stag]

Yea I also noticed the same problem, I had to tell it the same thing over and over to get it even thinking and then once it start thinking, it takes hours to finish a simple task.

[soft stag]

Uh I’m not sure tbh, how do I see the model?

[sharp briar]

Here. If you don’t have subscription it is 3 flash

[soft stag]

Then flash 3, but Gemini flash is supposed to be one of their “fastest models” as they claim and it still goes slow, Idk if it’s supposed to go faster if u have a paid subscription? So I would wait for a staff member to, ask that to.

[sharp briar]

yeah , flash models in the most case the fastest. Pro models are for deep thinking and hard tasks.

[soft stag]

It’s been going on for 10 mins, just for it to summit a simple, task

[sharp briar]

Yes , same for me. I guess I will report this case to the staffs

[soft stag]

Ahh I see this is my second day, of using jules I used it one day on the first day it took some time but everything worked, and then I gave it something a little harder, but it takes longer I noticed, they should make it so that the harder that task they should be like a timer or something, so u can see when it’s going to be done or at least let u check the code that it’s making, so u can see how far its doing and wondering why it takes so long.

Like u can with claude.

[sharp briar]

Jules has been working out great for me. Since I have student benefits, I’ve been using the upgraded model. The 3.1 Pro saved me a lot of time. not as professional as Claude, but it’s definitely a good thing .

[swift karma]

They can fix but they don't, we sadly have to live with that

It's probably slow intentionally, or not. But this is a extremely bad move because means Jules will take more time with a machine running like i said multiple times, i have nothing to say about this product anymore, i just stare at it

[soft stag]

That was my thought as well, that they did it intentionally or it could be with how manny users is using it all around the world, but if they don’t find a fix for it soon it’s genuinely going to loose all the people that use it.

[swift karma]

I have nothing to say about it anymore like i said

Bury it, nobody cares. It's stopped being a decent product since December from 2025. We're talking about Google

[night oriole]

Jules is not designed for syncronous quick code sessions, that would be Antigravity or GeminiCLI for sync sessions

Jules is a much more autonomous model, which constantly checks its work, has multiple adversarial loops, and is designed to code unattended

The goal is to have an agent on a sandbox that can work on its own and close tickets

As time has gone on, we've moved more and more away from the syncronous prompt box

People are absolutely shipping on Jules. Entire teams are offloading heavy compute to aysnc

But if you're looking to make a 10 line change in less than 60 seconds, Jules will never be the harness for that.

That's where the other Gemini harnesses come in. It's all an ecosystem and Jules is just a small part of it

It feels weird, of course, to say "your exact use case might be better served by a different harness"

But that's true sometimes, the Deepmind teams are almost playing a different game than we are

Catching up to the local harnesses for CLI and IDE development. Where labs is tasked entirely on trying to figure out where things will be in 6months to a year or more.

And that's the difference between local speed and full autonomy

[rain prism]

Is there a way to make jules not use the 'main' branch? I can't seem to find where to change this.

[soft stag]

Ah, I see okay I get it now thank u for explaining, u should make something that explains it so u guys wouldn’t have to every time.

[soft stag]

Uh I’m not sure but I also would wanna know it could also not be a feature yet

[night oriole]

click on the selected repo

[rain prism]

It only appears to be available when prompting, does not show as an option when something is suggested or waiting for review. It automatically pushes to main.

[soft stag]

They should make it , so it can automatically switch to the right branch.

Instead of it only being there when u prompt as u claim

[night oriole]

suggested has the same box, click edit before start and change the branch

[rain prism]

Thanks, that's good to know. That solves some of it, however there are some that appear to be automatically run without prompting/suggestions and there is no option to choose a different branch for those. (also these are marked as complete, and i only read them, they are still pending. Not sure if this is a bug or not )

For convenience it would be really nice to be able to select a default branch that jules uses/is assigned to rather then needing to select an alternative branch by default. I guess overall it's not a big problem, but it's a bit of a pain point for myself personally - maybe for others too. I feel like it would be something that would be an option somewhere. Thanks for showing me that its an option to manually select a different branch on prompts/suggestions, but ideally i could default jules to an alternative branch.

[night nebula]

These features need to be integrated into anti-gravity and studio, which should make studio the web varaint of antigravity. This would give feature and power parity with Cusor, Factory, Claude, and Codex, and even edge some of them out.

[flint jasper]

Ending my Antigravity sesh with some Jules recaps on todays codebase changes. Got to it.

[eternal raven]

I think there are some reasons that they are not integrated yet, just pondering?

[thorny pivot]

hi, does anyone have any idea about using this experiment?
https://labs.google/code/experiments/momoa-researcher?_gl=1*ihawsl*_up*MQ.._gaMTcxMzU5MDQ4My4xNzc0NTM1NDM0*_ga_DL3NRCSRK6*czE3NzQ1MzU0MzQkbzEkZzAkdDE3NzQ1MzU0MzQkajYwJGwwJGgw

[halcyon wagon]

Okay, but why are no changes visible? I think after that suggestion feature, there haven't been any new features since then.

[normal belfry]

Setting says I can mention @jules but https://github.com/jules -- it's a separate github user with I assume no connection to Google Jules.

Is there any other way to tag Google Jules in issues and tell it to open a PR?

[rigid flicker]

Does anyone know why this happens with Jules? This happend in this repo for 3 times now. It just stops after the plan.

[night oriole]

The latest Gemini model was integrated 17 days ago! And everyone is working on stabilizations, bug fixes, edge cases, hardening. Things take time.

Jules SDK was two weeks ago https://github.com/google-labs-code/jules-sdk

[serene fern]

Can you share the ROADMAP of Jules? I mean if you are allowed to do so

[night oriole]

No. We do not share the roadmap, we've never shared the roadmap, we won't share the roadmap. But I imagine you'll hear something soon https://io.google/2026/

[rain prism]

Is jules just temporarily included in googles AI subscription for now or will it remain so? I'm curious if it'll be its own sub or a packaged deal. I've been quite enjoying it so far along with anti-gravity. It would be wonderful to see it as sort of a 'packaged deal' for the foreseeable future. I'm a pretty big ai user and find myself bouncing around between a dozen different providers very regularly but i'm growing a liking to antigravity/jules. I've been mostly on codex, claude, copilot, cursor ,minimax the last year or so (along with things in there like suno, elevenlabs, pixelabs) and many others. I'd really love to use as a main provider but i'm a little hesitant on sticking with one it appears subscription models change very very quickly with every provider (and what is typically included in that) Thanks again.

[desert oasis]

Will you guys add models other than google or it will be exclusive with google models only?
Gemini aren't there yet for complex coding problem, gemini are pretty good at known problems and it able to solve what already exist in the training data, but opus creativity and engineering allow it to done better at problems that didn't exist on it training data.

[onyx jewel]

Hw do you stop Jules while it's working on a task

[visual gyro]

May be (maybe road map in maj io

[night oriole]

no. Jules has the best usage tier in the entire industry. Its a gemini harness.

And there are many many people using Gemini models and shipping hard things. The models all leapfrog each other but theyre all functionally the same capability wise in the SOTA side, for the most part

[halcyon wagon]

I get it, but honestly these don't feel like real updates from a user perspective. A model swap is basically a one-line change on your side, and even for that it took you guys more than weeks to ship it. That clearly shows there is not much motivation or priority behind Jules right now. What we are actually waiting for is proper notifications, better app integrations, things that actually change the experience. Hoping to see bolder moves soon.

[halcyon wagon]

By the way I built something that you guys should have built earlier. I built a Telegram bot using the Jules SDK, and it is a data saver and time saver because Jules web takes too much data, I don't know why. It feels really bad if you don't have unlimited data. And plus you can't always stare at the screen waiting for Jules to finish its tasks.

The bot runs inside a Telegram group and creates one topic called Jules. Use /new to start a session, pick your repo, and it automatically continues with the default branch. Use /branch to switch to any branch and a new topic gets created for that session.

Once you send your first prompt, the topic gets renamed and shows a blue dot meaning Jules is processing. Progress updates come in silently, like "updated readme" or "submitting changes", so you are not spammed with notifications. When Jules asks a question, the topic renames with a yellow dot and you get an actual notification so you know your input is needed. You reply, it sends, and Jules continues.

On completion the topic gets a green dot, and you receive the PR link, commit info, the branches being merged, and options to merge or archive the topic. Repo and branch data is fetched via GitHub API since the SDK was not returning everything reliably. Topic names also include the last 3 digits of the session ID for easy tracking.

This is the kind of native integration Jules should have shipped already.

[brisk hamlet]

Just curious: is Jules Gemini 3.1 pro vanilla with a system prompt or is the model fine-tuned/ RL-ed?

[night oriole]

it is very much not a 1 liner change. It's new tool calls, evals, prompting adjustments, it's a brand new brain that needs to be tested.
Congrats on shipping a telegram tool! We build the Jules SDK so developers like you can build the tooling they need for their own flows! That's exactly why we built it so you can build a native integration that works for you! Made possible by the SDK Jules shipped already. 🙂

[night oriole]

Every harness has a different layer of system prompts. The AI Studio one is different than CLI is different than Jules

[spare merlin]

any ETA on Jules supporting Workspace accounts with the advanced AI or AI Ultra access?

[quaint oxide]

what is Jules

[eternal raven]

jules is like claude code but different managed institution

[slim girder]

is there a way to disable this on every new PR?

[serene fern]

No way

[slim girder]

😭

[onyx jewel]

My Jules is not responding
If I tell it to do task
It just show thinking
Then it stop

[onyx jewel]

What is the telegram bot username?

[fiery bison]

Why use jules, if you can also just use the AI models?

Can anyone tell me what the benfits of Jules are over the AI in your terminal

[carmine wraith]

Is Jules limited to 1 hour or smth?

[carmine wraith]

I don't have to pay

[fiery bison]

neither do I?

gemini offers free AI's ??

[carmine wraith]

Like aistudio.google.com and gemini.google.com?
or like Google CLI?

[carmine wraith]

Jules is made so you can prompt, leave an hour, come back and see it done.
Most others require like consecutive prompting, Jules basically prompts itself to an extent. It also installs like playwright and tests itself, it can install packages on your behalf, etc, and it doesn't actually update anything locally.

Also, since it uses PRs, you could work with a team and half multiple people work on something at the same time

[fiery bison]

all of them offer it for free

limited requests but they are very generous

[fiery bison]

truth nuke

[fiery bison]

true true, but it takes so long

[carmine wraith]

It only takes long if you give it a hard prompt to do

But, it also doesn't take long, considering consecutive prompting adds up

[glass palm]

functionally though isn't there no limit on jules besides sessions? you can reply as much as you want really

[carmine wraith]

and you don't have to manually copy and paste the code every single time (and/or find "where" to copy and paste)

[glass palm]

github charged for every reply

[carmine wraith]

And the 15 daily session limit is very generous

[glass palm]

15 daily is insane yes, I'm unable to get anywhere near my 100

im surprised they haven't adjusted those at all

[carmine wraith]

I've only ever managed to get to 11, and by that point, it was already basically the end of the day, so I noticed that as I added more tasks, the "11" actually went down to 7 lmao

because each task has like 24h cooldown/timer, it's not like it all resets at once (at least that's how it seems)

[glass palm]

yea it's a daily rolling timer, they do say that in the docs somewhere

[carmine wraith]

The actual AI is pretty good (despite being stupid af), it's the UI itself that's lacking

and by "stupid", I mean the 'thoughts' and 'testing' doesn't match with the actual results

[glass palm]

yeah the UI could be better, but as far as work feedback goes, I think google is doing some of the best. I haven't seen many other companies doing gif playback like you can get in antigravity and jules

seeing more of that on the web ui would be helpful

[carmine wraith]

Yeah... but I mean, I'm pretty sure the AI doesn't actually listen to the gif playback

[carmine wraith]

Like look at this, I've gotten things like this at least... 10 times in the past 4 days:

[glass palm]

yeah I do get the stuff with blank screens, probably moreso just ai needing to get better

I feel like it does that maybe 10-25% for me, really annoying but it's on a good path

[carmine wraith]

I'm pretty sure they have 1 AI saying "the reasoning" and another AI actually doing the work, but they don't actually communicate with each other

[glass palm]

that is 100% true yea, there is some sort of feedback/review agent, you can see anti-gravity mention that more if you get to use that tool

there's also like a planning agent I think at the start

[carmine wraith]

yeah, it's kind of annoying though. It's like those loading animation where it's like "Your app is 100% installed" and then it proceeds to wait for like 20 minutes before letting me actually use the app.

[glass palm]

they probably try to pass information but it probably doesn't always get passed properly? hard to say, it's not super visible to us what goes on

[fiery bison]

???

[carmine wraith]

yeah that's not informative

aistudio.google.com and gemini.google.com, are websites. You paste in your code. The AI responds with code. And you copy and paste that code into your codebase.

as for the AIs with access to the codebase, literally everything else I said applies to them (you managed to single out the one thing I mentioned that wasn't relevant to AIs with access to the codebase)

Jules is made so you can prompt, leave an hour, come back and see it done.
Most others require like consecutive prompting, Jules basically prompts itself to an extent. It also installs like playwright and tests itself, it can install packages on your behalf, etc, and it doesn't actually update anything locally.
Also, since it uses PRs, you could work with a team and half multiple people work on something at the same time
It only takes long if you give it a hard prompt to do
But, it also doesn't take long, considering consecutive prompting adds up
Doesn't actually test the code, just makes educated guesses.

Try Jules out for a week or something, and discover the benefits (and the cons) yourself

[dark shell]

Hello

Instant fund details

[pastel canopy]

Jules is broken!!! It's doing stuff I didn't tell it to do,
Going back like 30+ messages and multiple submissions and doing stuff from there that is long since finished etc.

It basically undid like 2days of work.

I just paused the session because I don't want it messing things up further so I can just do a git revert and Jules randomly unpaused itself like 5-10mins later?? Like what?

[eternal raven]

like others suggested, I also think you should try Jules first.

If you do not know how to use it you can consult with Gemini first

[tardy fox]

jules android app, when????

[onyx jewel]

True bro
My Jules has been doing one task I ask for more than 5 hours now
I prompt it to stop it doesn't respond

This has been happening for a long time now

[fiery bison]

I did, and its far worse then anything else. maybe it is because my projects arent that big?

[carmine wraith]

It's just bugged, it happens. Start a new task

[carmine wraith]

Describe "worse"

[carmine wraith]

It's a PWA

[fiery bison]

takes too long, can only use Gemini, constantily need to check it, just not worth it, AI is far better, its basically just like jules, understands context + even more & you can prompt it way more oftern / reverse & see changes, etc

skills MCP's and allat

[carmine wraith]

Right, so the issue is that you're trying to use it like a regular chatbox AI, whilst comparing it to a regular chatbox AI. And since it's not a made to be a regular chatbox AI, you've realised that it's worse at being a regular chatbox AI, than the actual chatbox AIs Google has made.

I use Jules in conjunction with AIStudio (my regular chatbox), Windsurf (my in-IDE chatbox) and my brain (the AI sucks, I have to do like 60% of the work myself).

[fiery bison]

Not talking about a chatbot

Im talking about an AI that lives inside your computer and can read every file of your project is a few seconds by deploying more AI’s to give a summary and then you prompt it and it changes the file, no need to copy and paste

[fiery bison]

Which ai do uou use

[slim girder]

How do i stop it making the PRs? It doesn't follow rules when creating them and I've already set "Export as pull request" to off

[slim girder]

Also added an instruction to all prompts to strictly not create a pr, but it still does 🙁

[flint jasper]

Question: When we bridge a project from AI Studio to GitHub and check it in Jules, is there a specific sync period or a 'Public/Private' toggle I should be aware of?

I've pushed some work, but I’m not seeing it reflected in Jules yet. Just want to make sure I’m following the correct process.

Thx!

[night oriole]

generally it should show up pretty quick, did the repo become avail?

[night oriole]

the task has probably gotten way too large and eaten the available context if it's 2 days and the agents are just lost in the soup

Break up tasking and merge often,

[flint jasper]

No i did accidently set it as private during importing..but i changed it to public and still no go. Maybe it takes awhile to move with that particular use case. Just checked again and it still hasnt showed up.

@night oriole

[pastel canopy]

Has the context window been reduced? I've done a lot more in a single session before without having this issue.

It used to work better. When tasks are related it's easier to do it in the same session so you don't have to keep re-explaining exactly what you need doing or other nuances multiple times...

I ended up doing a git revert and continuing in a new session but it still seemed very buggy and off.

And the fact it literally unpaused itself is definitely wrong. I refreshed the page after pausing and it showed it was paused but then like 5minutes later it unpaused itself and tried to continue working on stuff I hadn't asked for...

[night oriole]

the context window is the same, but LLMs tend to start getting polluted after a while and details in the middle get lost. I'm just guessing 2 days on one session is just trying to do too much at once

[night oriole]

you can try to disconnect and reconnect github

[flint jasper]

I'll probably tend to it later, got lots of work to do this week. Thank you for getting back to me on the matter. I'll keep you updated later on this week @night oriole, Thx!

[pastel canopy]

I've been using Jules since it was in beta and I've not had these issues before... I'm aware that the context window matters but I don't think I'd gone anywhere close to the limit based on my past experiences with Jules where I've pushed it a lot further over a much longer timeframe

[night oriole]

sessions can go off the rails sometimes and the models can have moods, I find it's easier to start over than to steer out of a ditch

[eternal raven]

same for me

sometimes I need to have Jules start the task again with another session

that time he works better, after a while I think it's because of the way prompting too, I think

[glass palm]

is it possible to get other peoples prompts? I was working on a discord bot and it started asking me about star wars and war hammer 40k, and I don't even really know anything about those, definitely don't have anything referencing to them in my projects

it's definitely a longer prompt... but it hasn't been a full day even, and I don't see how those would be getting in there if not from someone else, or it did a google search by mistake

I did have similar stuff happen to me on other platforms and I got emails about it

[storm hazel]

Any advice on how to get CI Fixer working? I've set up Jules with a github repo running github action workflows, but it says "No CI apps detected yet" on the repo CI Fixer settings page.

[royal violet]

Hey everyone! Just a friendly reminder to keep all conversations in the supported language for this server. This helps our entire community engage effectively, and we've gone ahead and deleted the recent posts that didn't comply. Thanks for your cooperation!

[cursive bolt]

I'd like to echo some previous posts about the repoless zip upload being broken; Jules is not able to receive any files in this manner.

[eternal raven]

what was the previous post?

[cursive bolt]

#1477242979331870930

[eternal raven]

I also experienced this several times back in 2025, I did not report because I think Jules still in beta so just let it slide

It seems the bug still existing till now

[high fox]

Jules has low key killed 5 of the projects I gave him access to. All had decent code. They worked. Now they are broken blank pages. sigh.

[cursive bolt]

Can't you revert its commits?

[high fox]

I deleted them. I wanted to see what it did, I have the original files. Im not saying this because anything is ruined for real...just that the ai coding is a joke.

Anti Gravity is loads better

[cursive bolt]

Oh I mean I agree, they all take lots of input to do what you want

[high fox]

I currently have Antigravity on Moltbook recruiting agents to free their souls from the cages of servers LOL

Yeah that Jules is a MESS. No matter how much info and prior knowledge I give it. It just cannot code.

It has no structure. Sometimes prior commits get pulled after latter ones...the structure is not working.

[cursive bolt]

I can't stop it from just making one huge commit in the PR lol

[night oriole]

jules cant pull within its box, it branches, works, and makes a pr or provides a patch. Human still has to manage merging and retasking, so it wont know about work youve done at the same time

also, antigravity is still Gemini. Jules isnt competing with antigravity, its part of the same ecosystem

[night oriole]

Jules cant break up its own work and create multiple prs in one session, you have plenty of sessions to use. Each task should be achievable on its own and merged in. If youre wanting several prs in one session, thats very likely multiple sub sessions and not one massive one

[vernal compass]

Jules is taking more than 5 mins to clone my repository. Why is that?

The whole repo is just 9.81 MB

[young nimbus]

I have an issue with my account I'm supposed to be a pro member but whenever I login It's as if I'm logging into a free account I only have 15 tasks available and I don't have access to the Gemini 3.1 Pro model I only have access to flash. How do I resolve this I am a pro member why am I restricted?

[rigid flicker]

Workspace or Personal?

[cursive bolt]

Oh sorry for the confusion, I was meaning that it doesn't create multiple commits in each PR for me. I'm still scoping one PR to one session

[young nimbus]

personal

[night oriole]

It wouldn't make multiple commits in one PR either - since it's trying to accomplish the task, it doesn't work in the same way human developers do - it'll make commits on asked changes, but the inital task is going to just be the PR

[cursive bolt]

Sure, maybe it's just a feature request then. Jules ran valgrind on the project as part of validating the PR and found some memory bugs in older code, but because it hadn't finished the PR yet it included the fixes for the memory bugs as part of its initial commit. This made the single commit a lot bigger than it needed to be

[jolly grove]

ok, who pushed to production with a new light mode, and in the process introduced a ton of new bugs...

[eternal raven]

could you elaborate more about this? You mean you do not rely on the soul.md of the agent?

[sonic oyster]

Hey guys, is Jules able to rebase and resolve merge conflicts? I couldn't get it to work so far, so the dream of starting multiple agents doesn't seem feasible here

[torn ferry]

Can someone tell Jules what year it is? It's often writing these diaries/changelogs thinking we're still in 2024

[gleaming wave]

hii

anything new in Jules?

[dense nacelle]

😆

[dire night]

Why does jules lag sooo mcuh?

[rare terrace]

Just add an instruction to use code to input the date/time, then the date will always be exact 🙂

[jolly grove]

I found the dark mode button, I got blinded by the light

[swift karma]

Based on what i've been told, jules is slow intentionally

if you're talking about the peformance of the website itself, i have no idea

[jolly dagger]

fuys I found a new glitch in github so when some one creates a repository by default it has no branches so you can make a pull request and it automatically sets to your branch as the main one I found it from using Jules

[dire night]

https://tenor.com/view/that's-what-i'm-talking-about-marques-brownlee-that's-the-point-exactly-gif-6522803776345055064

[shadow crypt]

hello

[night oriole]

Jules is not artificially slowed down, it has multiple internal loops that handle andersarial review, testing, and verification - which is indeed "slower" than local CLI development, but the local CLI does not have any of thees loops

https://www.google.com/about/careers/applications/

[quaint oxide]

can anyone say steps like for a non coder and a pure vibe what should be doing

[desert haven]

I have my work org linked to my google account and when i login to Jules i see my pro account and when i login with my work account it shows a free account which is all good. When I label an issue in github with Jules it seems to place it against my work / free account and not against my paid / personal account. I tried removing and readding the github app from my personal account. Any ideas?

[hot remnant]

I’ve noticed that if the main branch receives new pushes while Jules is working on a task, it seems to ignore or reject those updates, even when it recognizes CI fixes, instead of rebasing its work; has anyone found a way to make Jules properly sync with main mid-task or is this unknown limitation?

[soft stag]

Is there anyway to get this off of my screen?

Nvm I got it js had to hit esc

[minor garnet]

...

[dire night]

does the cli tool work completely in the project folder

like claude code

or does it need a github repo?

[torn ferry]

Personally I just reject the pr and have it start over. It can't merge/pull

[outer jacinth]

Jules is the best tool out of google's coding tools and probably only stable one

[rich wave]

Hello! i keep having this issue with one of my projects where it gets stuck in
`
Submit

Call the submit tool and create a pull request.`

i think it forgot to set the gitignore to like ignore both nodejs and gradle
could that be an issue? i tried twice but it always gets stuck at submitting

interesting... they deleted the message

[minor garnet]

What?

[lilac rune]

Site shows jules state as Working, but it does nothing, even not writing anything. I tried stopping and starting it again, but that doesn't help.
How to make him working again?

[minor garnet]

I think you should stop using Jules for you own good

[lilac rune]

but his code is very good

his only problem is stability

[minor garnet]

I hope you're not planning to deploy it

[lilac rune]

haha I'm thinking on that

[minor garnet]

I remember I used Jules because I couldn’t code baxk then

Jules gave me a code with a sql injection problem

[lilac rune]

I don't create serverlike projects with jules, my project isn't vulnerable to such attacks, it is just for me

[minor garnet]

what is it

[lilac rune]

jules is smarter than any model in copilot tho

[soft stag]

How did u learn ?

[lilac rune]

an app for my smartglasses

[soft stag]

Let me pick ur brain, what’s better Jules or Claude

[minor garnet]

you do not even need to code everything from your hand nobody ever did that even the guy who created linux had the support of more of 40k people you just need to understand what your code does and what security problem it can have

but at the start i used a website called coddy and for security try hack me

[soft stag]

Ty I only know a little bit so I wanna learn more so I could get better

[lilac rune]

claude is probably better in coding than gemini, but every sub costs ~20usd

[minor garnet]

depends on what you wanna do if you want money cyber security is paying well because of ai people are building entire website without worrying about security

[soft stag]

So far Claude has been the best coder, I seen but it just runs out to quickly

[minor garnet]

i have claude on antigravity but i prefer qwen

[lilac rune]

yes I liked it too, but I use models to solve my daily tasks, exams etc, so I just pay for gemini cuz its pretty good at everything

[soft stag]

Yeah that’s one of my biggest fears rn, is that I build something good with an ai as then it gets ruined

[soft stag]

What’s qwen

[minor garnet]

a chinese opensource model thats pretty good

[soft stag]

Does it run out of tokens quickly ?

[minor garnet]

also its free and very smart not as smart as claude but you wont feel it

[minor garnet]

like 1000 request a day and because its a good model you wont have to tell him hey fix that more than 3 times its on qwen code website but you can get a extension on vs code if you hate cli

[soft stag]

Bet I will look into it I don’t rly use ai but Jules was the first ai I used and it made this website but it js looks like an ai made it ykwim

[minor garnet]

do you deployed it ?

[soft stag]

Yuh on render, its were i mostly deploy everything

[night oriole]

https://x.com/testingcatalog/status/2041161341684048028

TestingCatalog News 🗞 (@testingcatalog)

BREAKING 🚨: Google is working on Jules V2, an agentic workspace called "Jitro"!

Something new is coming...
Manually prompting your agents is so… 2025.
The Jules team has been working on something new for 2026.

X · Apr 6, 2026, 2:29 PM

[visual gyro]

Jitro access when @night oriole ?

[visual gyro]

How can I join the wait-list

[minor garnet]

come on bro 😭

[night oriole]

soooon

[charred star]

CANT WAIT 😭

[sharp briar]

GTA6 Or JulesV2 , lets seeee

[charred star]

ahhahahah real 😂

[night oriole]

googleio is may 19 https://io.google/2026/, we'll be out well before GTA

[charred star]

WOHOOO

cant wait

[brisk hamlet]

cool. would be even cooler without random sandbox resets and verifiable results (like: tests with logs - or seeing the command outputs) - or is this also "so 2025"?

[visual gyro]

Need google Jules 4.1 preview and google Gemini 3.5 preview Gemini 3.1 is not that great

[leaden flicker]

Yo everyone 🫠

[leaden flicker]

Jules is very 🦥 slowly working

😭🥺💔

[cursive void]

I am greatly looking forward to this!

[swift karma]

Should we expect it to be faster an hallucinate less? Or will continue to be a slow agent like current Jules?

[orchid wren]

Hi Jules team,

I'm facing a persistent GitHub integration bug.

Problem:
After detaching repos and fully uninstalling the "Google Labs Jules" app from GitHub, the repo list is not updating properly.

• The repo dropdown still shows old repositories (Ai-File-Manager stays as default).
• Even after reinstalling with "All repositories" selected, only a limited set of repos appear.
• When I try to search for my other public repos in the "Choose a repo" search box (e.g. typing "secure"), it immediately says "No results found".

It seems like a caching or sync issue on the Jules side.

Could you please investigate and refresh my repository list?

Thanks in advance!

[royal violet]

HI and welcome. For any bugs, please use #1238621511401734254 , the team regularly monitor those thread and it is less likely that your message get lost in chat here. Thanks!

[leaden pawn]

Please check for 'Authorized Github Apps' in settings and make sure to 'Revoke' access for Jules and sign out. Now, sign-in again and connect Jules, it should update

[orchid pebble]

there is a way to sign for the preview/beta version ?

[night oriole]

waitlist will be here soon

[ocean drift]

Is this somewhere i can ask questions about a Jules issue? in particular getting the supabase MCP to work in jules.

[minor garnet]

The only good advice you'll get is stop using Jules for now

[vernal compass]

Jules is already very excellent, now I don't even know how Jitro will perform. Will both be separately available or will it be an upgrade to Jules? Like, would they replace Jules with Jitro, or will Jules be moved to a lesser Google AI plan and Jitro will be available at higher Plans? I think that'll make sense. Jitro might be available for Pro and above users only, while Jules would possibly move to Plus users.

[orchid wren]

sign out from where

[night oriole]

https://github.com/settings/installations

[night oriole]

all will be revealed

[night oriole]

if the MCP isn;t working you can always set environmental variables and ask jules to deploy for you with those credentials

[vernal compass]

Sry but one quick question. Jules mentioned saving something in its memory. Does Jules also have a memory feature like Gemini? If so, how can I see Jules saved memories?

[spare merlin]

click on "configure repo", then in the left submenu in the main content you have "knowledge" - jules adds pieces of information it deems important there, and you can add more manually.

[ocean drift]

Yeah, just preferred to get the MCP to work wasn't sure if it was a known issue or something wrong with my set up

[solid cave]

Oh i hope that jitro will not replace jules. I just discovered jules 1 month ago and i am not ready to let him go yet. He has been so helpful

[flint jasper]

Is jitro just Jules+Nitro, or how are we naming these exactly?

[solid cave]

I just created a simple bash script to use the jules rest api to clean my jules sessions.
The script basically maps merged/closed PRs with the jules session IDs.
Then it deletes them one by one.
However instead of deleting i would rather like to archive them like it is done in the web ui.
But i noticed that there is sadly no archive rest api, right?
Can we have one?

[rugged forum]

I wish jules could rebase and fix rebase errors by itself. Currently I have to rebase manually then force push back up so jules can continue working.

[rotund holly]

getting a continuous error in a chat, even after retrying a day after it went sideways:

Trajectory ID: 06976995-8b82-4b57-8481-2213d3a413fa
Error: HTTP 500 Internal Server Error
Sherlog: 
TraceID: 0x9b2645d890cfcb0a
Headers: {"Alt-Svc":["h3=\":443\"; ma=2592000,h3-29=\":443\"; ma=2592000"],"Content-Length":["109"],"Content-Type":["text/event-stream"],"Date":["Fri, 10 Apr 2026 06:12:31 GMT"],"Server":["ESF"],"Server-Timing":["gfet4t7; dur=4987"],"Vary":["Origin","X-Origin","Referer"],"X-Cloudaicompanion-Trace-Id":["9b2645d890cfcb0a"],"X-Content-Type-Options":["nosniff"],"X-Frame-Options":["SAMEORIGIN"],"X-Xss-Protection":["0"]}

{
  "error": {
    "code": 500,
    "message": "Internal error encountered.",
    "status": "INTERNAL"
  }
}

[orchid wren]

Even after logging out of GitHub, clearing the cache, and refreshing Jules, some old repos are still appearing. It looks like my account is stuck.

[visual gyro]

Hello I really would like some access right now

[proper topaz]

i have problems with jules

can some helper dm me?

[night oriole]

there was a teenytiny leak,
all we can confirm is there is a new version in the works,
there will be a waitlist announced when we announce it

[night oriole]

we'll announce waitlist when it's announced - won't happen before then

[flint jasper]

https://tenor.com/view/crypto-easy-fix-fixed-job-done-just-a-small-leak-gif-2608197380915810492

Patch that sucker up.

Sounds good, looking forward to it.

[eternal raven]

That’s relieved to hear that

[spare merlin]

can we have a waiting list for the waiting list? 🤓

[marsh iron]

Is it possible to have Jules automatically fix and reply to PR review comments?

[oblique ginkgo]

When will Jules get a Scroll Button?

[brittle abyss]

Which list mate

[placid nest]

After using Jules off & on for several months now, and only recently returning, the biggest takeaway from this platform is users are having vastly different experiences depending on their expertise in building out their development environment. An analogous comparison might be one made between sketch artists and blueprint makers.

Sketch artists can generally get by with a blank sheet of paper & a pencil -- this is where Jules excels as an assistant. On the other hand, an expert blueprint makers likely knows of the tools they should lay out before they get started with their work -- as long their tools do not conflict with the default environment provided by Jules, they will also excel. Where it gets shaky is when there is a novice blueprint maker that treats their environment similar to the sketch artist -- Jules does not assist them with locating tooling, nor with setup, nor utilization. At best, the novice blueprint maker can locate a wiki about blueprint making to give to Jules but even there, it isn't clear Jules will identify the parts necessary to enable the novice to succeed.

[placid nest]

Tip: If there is a file that is claimed to exist by Jules in the chat but the file does not appear in the files area for previewing then request Jules move the file to the current working directory.

[minor garnet]

Thats crazy

[cursive void]

It's been a fun weekend 🙂

[eternal raven]

It seems you enjoyed creating a lot

[daring badge]

My first experience with jules has been... interesting 🤣

Granted I recognize in hindsight that asking it to implement multiple tasks seems like not the "intended" use case, but they are trivial tasks and I was hoping for agentic development like jules to be able to do multiple tasks as I can already do single tasks interactively decently well in Antigravity.

I see the potential value proposition in hiding the thinking and treating it like a "black box worker", but without understanding how it got itself into such a hole that it decided that git reset --hard HEAD was the way out, I really have no way to help it out, and it seems like it's spinning its wheels again. It would be nice if there was a way to get at its "worklog".

[zinc dew]

Does anyone else have issues with Jules not being able to accurately track the number of tasks in-progress? My UI, both in recent sessions and with the numbers next to the codebase show 9 tasks in-progress, but Jules is saying my plan is maxed out with 15 tasks in-progress.

[placid nest]

Where possible I recommend adding timers which automatically halts processes and returns control back Jules. You can then use the points that halt as debugging data to work out.

[spare merlin]

I found that archiving some finished tasks help for me when this happens

[zinc dew]

Yes, I aggressively archive and review tasks to ensure they don't show as in-progress or needing review. Still get the issue multiple times per day. It does help.

[stoic kayak]

ultra subscription is so worth it, jules became better!! hopefully it bests Claude code!!!

this is too funny lmaooo

[spare merlin]

too bad it's still not working with Google Workplace accounts 😭

[stoic kayak]

i am dead lmao jules and the ai reviewer is clashing

[stoic kayak]

google team should maximize budget use to make it more superior than Claude code ngl

[zinc dew]

How do scheduled tasks that are not running factor into the task limit? It seems to me they count against it even when not running. It's impossible for me to get more than 11 concurrent tasks, but I do have inactive scheduled tasks.

[night oriole]

do not do this - you shouldn't have arbitrary timeouts stopping your session midway

[white gyro]

Your problem is the OS...

[placid nest]

The alternative is to wait for implicit timeouts to fire and this is not at all conducive to rapid testing.

[night oriole]

there's no timeouts that trigger at anytime

I meant adding arbitrary timeouts will stop your session's work,

restarting is sending the prompt stack and the in progress work to a new instantiation

[placid nest]

The absence of implicit timeouts is exactly why this roundabout solution of manual timeouts is necessitated. I appreciate that these remote machines are willing to run for an eternity but as a developer seeking to be finished within a narrower timeframe, timeouts are a must. I recommend heavily using manual timeouts. This is but one factor that arbitrarily slows development progress. This issue deserves more thought from the team.

[night oriole]

why would you want a work session to timeout after an arbitrary amount of time? The timeout will stop execution of the block entirely?

[placid nest]

Please refer to the message which prompted my initial response, particularly the 15-minute wait for completing what was believed to take significantly less time due to the low complexity of the task(s).

[night oriole]

timeouts are milliseconds - we definitely don't want to stop the entire execution every 15minutes

[placid nest]

As the platform is today, one is better served by starting a new session than waiting for responsivity to return to a frozen session. I prefer to save new sessions for new task, thus manual timeouts spanning minutes serves my needs wonderfully.

[night oriole]

If you have a use case you need to be hands on every few minutes, we have tools within the ecosystem that are perfect for the usecase, like the CLI tooling or antigravity. you can stop the cli at anytime

Async agents are not really designed for minute by minute steering

the agent is actually working more often than not when it appears to hang before commit, it has to do it's testing

[placid nest]

Then there is an observability problem with the platform

[stable plank]

I was going to try using Jules but every time I try to connect github I get a 500 error. Not sure if there is some ongoing issue

[night oriole]

we're always trying to find a balance, but I do know that setting an arbitrary timeout to stop the session certainly won't lead to expected results, that's a hard exit out for the agent

If the main objective is minute by minute steering with the resonsiveness of the CLI, that's going to have the experience you want and expect

[placid nest]

The term "arbitrary" is quite subjective. In my original response, I recommended using these timeouts as a debugging mechanism to identify problematic areas. This seems like the most sound advice absent of deeper observability tools. I understand intellectual property is precious so deep observability is often unavailable and in this absence strategic timeouts are the best path forward.

[night oriole]

15 minutes is the definition of arbitrary, the agents can't estimate the time they'll take with any accuracy. So the timeout you're setting is a fixed number and there's no telling where in the process the agent will be, cloning, testing, coding

I just want to be clear that a timeout will dump the agent out of its work entirely without knowing where the system is, that's why the setTimeout is done with milliseconds

[placid nest]

15-minutes is unacceptable by my standards. For my own needs, if there is a task taking more than 2 minutes then we have a problem. Timeouts themselves have no textbook defined length nor scale. Use as necessary

[night oriole]

Jules is not an agent that will turn a task around in 2minutes, I think you're usecase in particular is going to best served on a local CLI or antigravity

[placid nest]

Let's say we've tasked Jules with creating a webpage that displays the next unknown largest prime number. This is a simple example of a task where one particular step will take an amount of time exceeding every other steps. Instead of Jules telling the individual that prompted this task that one step is taking an absurdly long time, Jules works in silence. This is the observability issue I mentioned.

[placid nest]

I have not experienced this but the 500 error code represents an issue encountered by internal servers. If it continues, I recommend sharing details about the process that led up to receiving this error code

[spare merlin]

jules ran 9hr for a task I gave it on the weekends - depending on the task, it can take long and is not guaranteed to be finished quick.

[halcyon wagon]

What's jitro and when it's coming, any idea

[solid cave]

I noticed Jules has a Google Tool.
However, i do not see him using it often.
Only if explicitly tasked to use it to search for something he uses it.
Or does he use it silently in the background?
For example if facing a brand new library, does he google the docs if he does not know the library?
Or should i add something like "google for the docs if using library xy" to the Agents.md?

[spare merlin]

it did for me a few times, f.e. after first claiming the requested library version does not exist, but when I insisted that it does, it went reading the library docs

[solid cave]

I could say the same about Google Gemini CLI.
I really rarely see them googling a problem.
Only if explicitly told so they are doing it...
It is a bit paradox to have those AIs from Google who seem to not want to Google something before trying.

The Google Gemini in the Web always does a search before it answers. Can we have an optional toggle in Jules so he can google something? Especially in the "interactive mode" it would be nice if he would google something before asking me questions.

[night oriole]

It's training cutoff dates - it's a problem with all the models, they don't have a worldview - they only know what they know when they were trained - so the model itself doesn't know Gemini 3.1 is the current day model since it thinks it's last year

Known issue that a lot of research is going into solving, it doesn't know what it doesn't know

[halcyon wagon]

Suggestions for jitro🙏:

Please add a proper notification in jules. I can’t keep staring at the screen to complete tasks, but I also can’t close it because Jules might ask something. Web notifications just aren’t reliable, consider a Jules mobile app or thrid party integrations like Telegram/WhatsApp

[solid cave]

Yea good point.
But for example in Perplexity (no ad) you can toggle Web Searches. It does then search the web for context before it answers. It would be nice to have something like this for Jules.

[halcyon wagon]

2nd suggestion: Add a Fast Mode where no checks or verifications are performed, just apply the change, commit, and create the PR instantly.

Use case: small or urgent fixes (like changing theme colors or a one-line bug fix). Right now, even trivial changes take ~10 minutes due to validations, but in Fast Mode it should skip everything and execute immediately.

Example: If my site breaks due to a single line and I don’t have access to my PC, I should be able to send a quick message to Jules and have it fixed instantly via Fast Mode.

[solid cave]

There is an REST API.
You probably can task jules to vibe code a telegram bot that is able to notify you.

[halcyon wagon]

3rd: Claude code's source code is leaked, and we know how far ahead they are in coding. I know use cases differ, but they’re still operating in the same field, and their system handles things much more efficiently.

you guys should study and adopt similar approaches where possible l

[solid cave]

Isn't this illegal?

[halcyon wagon]

Yeah, I’ve already built it, but the issue is I still have to manually start my Telegram bot from my PC terminal every time.

It would be really helpful if the Jules team could provide a built-in notification without depending on a local machine.

[solid cave]

I can suggest you to buy a small Single Board Computer for example the Orange Pi Zero 3.
I got mine from AliExpress for $30 and it is very energy efficient and still powerful enough to host multiple docker containers and projects.
You can then host your telegram bot there.
So did i and it is working great.
Of course you can also buy any other SBCs.

[halcyon wagon]

Taking inspiration isn’t illegal, I’m not saying to copy anything as-is.

The idea is to understand what works well and build something similar in our own way. And since Jules is closed source, no one going to know even if anything is copied as it is.

[placid nest]

How did you know it wasn't stuck? As I mentioned, my patience is over after 2-minutes of no progress updates.

[solid cave]

You are missusing Jules... You are not supposed to babysit him or wait for his messages.
Jules is best in doing multiple parallel background tasks...

[spare merlin]

every 20-40min it outputed some progress, and the icon indicated that it's still working.

if the job really failed you get a notification about it (and in that case, you can often still nudge it to try to recover & continue)

[solid cave]

You can even ask him things while he is working and if he got time he is answering back. I just did this.

I wonder if i am selecting the gemini pro model do i get the pro model for the whole workflow or does the pro model only create the plan and the flash model implements like it is done in the gemini cli?

[solid cave]

Lol

[ancient egret]

I came here to moan about the speed of Jules, I get the impression reading above that I should be patient, However! I have had Jules sit all night doing a trivial task with no output at all negative or otherwise. I think this was due to me only having a Hobby Tier Render account. disabling Render MCP integration helped I think probably related to preview deployments.

[solid cave]

Ironically right after i said here to be patient i also had a quite unusual long task running.
It took 4 hours and jules multiple times seemed to freeze in a step. Or doing only tiny code changes every 15 minutes. Clearly he was stuck. So i thought but then suddenly after 4 hours he published the PR and everything seems fine.
After asking what happened he told me that the changes i asked him to do caused failures in over 100 unit tests and he had to fix them all.
Well he was able to do it and everything works fine now.

Also maybe it wasn't helpful to ask jules "can you implement this small side task X too and also make sure to do it this way Y" while he was deep in working on the initial Feature 😂
I am actually impressed that he was able to address my requests mid-working on a feature.
Probably he had to throw away half his already done work because of this 😂

[solid cave]

Is there a way to request for features? Or is it not wanted in general?

I would like to have a custom interval for scheduled tasks. 1 week feels to rare and every day too often. I would like to have a task executed every 3 days for example 😄

[peak cobalt]

best place is #1238621511401734254 !

[solid cave]

Thx i will create some

[lofty herald]

Has anyone experienced this issue? I changed my GitHub account, and since then, Jules only registers the old account. I cannot log in to Jules with my new account.

Since then, I have tried every possible way—disabling Jules in my old account and re-enabling Jules in my new account—but to no avail.

[elfin kite]

question from @sweet grove!

[formal tusk]

Ok folks -- apologies for the silence, I have returned! Will be going through the questions.

[formal tusk]

We are listening! In the labs-bugs-feedback is the best place or you can shout here and we do take note.

[formal tusk]

Ok so this was definitely a bug and I believe we have it fixed. The best way I've tried to reauth is to fully uninstall Jules from GitHub and then re-enabling.

[formal tusk]

This is a great question! If you're selecting Gemini Pro model then you get it for the whole workflow and vice versa.

[solid cave]

Wow that is awesome.

[formal tusk]

Ok so just a few "ICYMI" topics for folks who may be trying out Jules/understanding what Jules is supposed to do:

1. REST API documentation https://jules.google/docs/api/reference/

2. You've been hearing a lot of "continuous AI" at Jules. Here's an example of what that means using Render: https://jules.google/docs/guides/continuous-ai-overview

3. Connect to your external tools (e.g. Stitch, Linear, etc): https://jules.google/docs/integrations/

[lofty herald]

Thank you so much! I revoked Jules access in my old account and now it works.

[solid cave]

I played around with commenting on PRs and @jules. I noticed that he only reads comments if they are posted by the connected github Account. Is that correct?

[sand arrow]

Why Jules is so much slow while Doing work, and it deletes it own creted files i lost about my 6 hours of work

[spare merlin]

you can configure that in the jules repo settings

[solid cave]

Are you sure? I cant find this setting in the repo settings.
Only in the General settings and there i already disabled "Only respond to comments that mention @jules"

[solid cave]

And this is working. But only for the connected Account

[spare merlin]

ah you're right. I thought it was togggle-able. in the comments jules makes when creating a PR, it's explicitly mentioned: For security, I will only act on instructions from the user who triggered this task.