Just a thought, I'm not sure if AH plans to host Deepseek v4? They're good LLMs (I’m using Pro), and there's a tough confusing world out there with bugged providers, which makes users have a worse experience with OH. 🤔 It could be nice to offer some hosting which works as expected. (example issue)

Hi, I’m seeing an unexpected billing issue with the OpenHands LLM provider. I purchased $10 in credits and used openhands/kimi-k3, which is shown as a Free model in Agent Canvas. Agent Canvas reports a total cost of $0.0000, but my OpenHands Cloud balance decreased from $10.00 to about $7.47. I also tested this in a new conversation with a minimal prompt, and the balance continued to decrease while the conversation still showed $0.0000. Could you please check the billing records on my account and refund any credits that were charged incorrectly?

I recently received the honor of being recognized as an active contributor to the project. (tyty to everybody involved <3) Anybody who doesn't already know, I am among the large numbers of tech professionals between career posts. Due to that unemployment condition, I have an interest in getting my portfolio updated quickly--for potential employers, my hope is that participation here should be a strong indicator that I've kept current on my understanding of AI and coding harnesses. 😅 Anyway, that's why I was going through this stuff in the first place. That caused me to realize that openhands.dev is serving a stale community section. Upcoming Events still lists May 26, June 4, and June 18, and the contributor spotlight still references PRs #14437/#14444/#14261. There’s already a healthy daily sync in OpenHands/champions-list producing current contributor data for champions.hub.openhands.dev, so the contributor block may simply not be wired to that source. The event list would still need its own source.

Where can I submit feedback? I have issues with my account balance and the Critique Mode not enabling properly. I'd like to get in touch with the OpenHans team—who should I contact?

Cross-posting over here for anyone who may be open to give some team specific feedback https://openhands-ai.slack.com/archives/C06P5NCGSFP/p1787671947175249

Loading canvas with various automations etc can take 20-30 seconds which feels super slow

Also when automation fails it would be nice to have a notification capability so that I know (some run only once a week, and if it fails I don't realize perhaps until the next week)

Error logs are not super useful, the actual error trace is in the output window and the errors here are expected log statements

The open-source helm chart works great, but is there a particular reason a statefulset is used? A pvc (or two) mounted to the container is more flexible and easy to manage IMO.

My agent sees that Novita is the provider used by

openhands/deepseek-v4-pro

, and it doesn’t do prompt caching well. Occasionally I see a tiny cache read, but mostly I don’t. (!) Could we change it? cc: @millions-crayon-32880 @limited-student-10747

Grok Bot, Instinct… other personal agents are taking off out there. In the meantime, OpenAI appears to prepare their version, based on a “persistent model”. Apparently, it’s • “not really an app”, and • “accessible from almost anywhere”. And, of course, there be talk of agent identity and memory. All this is funny and a tad “oof” from me. It’s a bit like they heard me arguing where the future is going. Sorry! Need to say, though, me thinks that maybe it’s not in “learn a new web app so you can talk to an agent”. 😇

Hello, I've encountered quite a few errors while using DeepSeek V4 for work, and I'd like to report this issue. I'm not sure if this is an isolated case on my end or if others are experiencing it as well—could you let me know whether I'm the only one facing this problem? If it's specific to my setup, I'll try reinstalling it.

Would be great to have this button on the file viewer. Right now I have to copy all the contents manually, if needed.

Please add support for mermaid diagrams in rich editor view.

Hello, is the free DeepSeek-V4-Flash model no longer available for local use? I can only find this free model on the cloud.

Hi everyone, I’ve been using OpenHands for about 2 months for my small business (with 10 repos mostly managing builds for SaaS repos). Performance was great initially, but since the free tier defaulted to DeepSeek, I've been stuck in endless loops fighting corrupted code outputs and file mangling. The issue is actively corrupting working files and injecting syntax errors directly into scripts. Because the tool's command echo shows the corruption before execution, it appears to be injected upstream by the model rather than the execution environment. Here are specific issues generated directly by the model: 1. Corrupted Shell Heredocs: Missing closing parentheses before execution causing

SyntaxError

(e.g.,

print('patched', name

missing

)

). 2. Variable Name Splicing: Single-letter prefixes injected into variables (e.g.,

SP

becomes

lSP

,

SEMI

becomes

undefined

), leading to

NameError

. 3. Mangled File Writes: Text following a colon or comma gets rewritten improperly during editor writes (e.g.,

deadline(an

instead of

deadline: an

). 4. Git Log Corruption: Injected hash marks and stripped spaces in standard commit messages (e.g.,

first:the, two-commas, two-semicolons.

). 5. Python Slice Syntax Errors: Standard slices like

data[-40:]

fail parsing with

SyntaxError: unmatched closing bracket

. 6. Upstream Injection: Tool command echoes display the mangled syntax prior to shell execution. Is this caused by the model/tokenizer proxy or an OpenHands environment parsing bug? Has anyone else encountered this on the free tier? I use OpenHands Cloud ~10 hours daily (and haven't been able to get the local setup running yet). I'd appreciate any insights or workarounds. Thanks, Ala

Hi team, I've been using OpenHands since launch and love it for tasks that Claude or GPT can't solve, OpenHands has always come through for me. I was away for a while, and when I came back and started using it again, I noticed the current model (DeepSeek) hallucinates a lot and often ends up not completing anything. This is a problem since I use it on my company's codebase, which I trust it with. I ended up switching to Codex to get the issue resolved. Could this be looked into? Also, BYOK feature it hasn't been working maybe I'm doing it wrong could someone help me sort that out too? Thanks!

Hi everyone! The contribution guide recommends posting larger UI proposals in

#agent-canvas

, but I can’t see that channel, so I’m sharing this here. We’re exploring self-hosted OpenHands for our platform at Zoey, and one feature we’d find very useful is an interactive human terminal directly inside the conversation UI. We simply want a human to be able to access and interact with the same machine and workspace directly. This would be useful when the agent is taking time, when we want to inspect the current state, run a quick command, review files or test something ourselves. At the moment, we can see the terminal commands executed by the agent, but we cannot type commands ourselves. I found the older request https://github.com/OpenHands/OpenHands/issues/2643, but it’s closed. Is this already supported, planned or tracked somewhere else?

Hello, I'm very sorry to bother you. Does OpenHands no longer offer free access to DeepSeek V4? Also, my Critique Mode doesn't seem to be working properly either.

I have been doing some experiments in the context of my research on condensation strategies with OpenHands harness along with gpt 5.6 luna on Swe Bench Verified. the idea is to bench the different strategies and compare it with the baseline( mostly task success rate, tokens usage, cost). that was the first initial results. in second experiments Task success rate was about 87% on easy tasks.

image.png

Hi. It's been a while 👋 I just wanted to try openhands cloud and login on app.all-hands.dev/login using my github. There is a warningin german(by browser lang) about bitbucket not working, but says github should be fine. But after I click on "Connect to Github" I get the attached error message

Hello, who can tell me what the current free LLMs are? Why is it completely empty here?

Google folks, casually: “Reinventing Kubernetes for agentic workloads with statefulness and fast resumption”. thinky spin Source

I'm using DeepSeek V4.1 Flash via OpenHands, and it is painfully slow — crawling at a mere 6 tokens/second. Is anyone else experiencing this?

I’d like to make a note. And record my prediction! I saw a youtube interview with Robert, and he said OpenHands is scoped within SWE. He didn’t exactly say that personal agents are out of scope, but it felt like that to me. (I had to double check!). I want to expand on a different point of view on personal agents. So one way to look at the SWE scope is to say, ok, OpenHands is focused on SWE; in its entirety (from requirements gathering, triage to deploy, monitoring, feedback). But that could be said as both focus and limitation, as in, focus on SWE and limited to SWE (that is, no “booking flights”, you know, the favorite example since OpenAI’s old Atlas) I think the first, focus, is fine, sure; but the second, limiting, isn’t really true. Software has been eating the world because software is in almost every human endeavor. If an agent is good in making software, then it’s good in “booking flights” too. Sure, sometimes you may want some extra skills, but of course you can install them and it’s perfectly doable. In the measure personal agents mean “some agent that knows how to book me flights”, OpenHands can do it. Also, since it’s SWE, you can have it make the others! (I note that the other way around it’s not necessarily true: most specialized agents that I’ve seen - legal, customer service - have been, at least for a while, full of predefined workflows, which is to say, they’re not good in a different domain unless you strip them down first, and even then, it depends. Reduce them to bash, python, memory, and maybe browsing, if you can - if they even have them - then it can be done.) More importantly, I think that’s not what personal agents mean: I believe it’s primarily UX. It’s how they feel like, how you use them, not whether they can do stuff in a particular domain, SWE or non-SWE. It’s really just UX: your agent is defined not by what it can do, but by a few other characteristics which all come down to UX: • knows your priorities and preferences (which can be work or SWE) • accessible wherever you are (Slack, WhatsApp, Discord, GitHub, etc) • memory/learning. For example, my personal agent is literally an OpenHands agent, it works on agent-sdk REST interface. It can do everything OpenHands can do. Thanks to memory, access for context, and some bit of proactivity/personality, it _feels different_; which is to say, UX. It also feels like I can do more with it, within SWE, too. e.g. I never need to “select repo by clicking a UI element” (why? it knows them all) etc. Which is to say, smooth UX wise. I have, in the past, argued that OpenHands could return to its roots, and walk the path of personal AI engineer - it literally had in its system prompt that it’s “an AI software engineer” (since that’s what Cognition’s Devin campaign was claiming to be, back in the day). Doing so would imply to see it behave or feel like more like an entity, an “AI colleague”, than as “a web app”. I mean, it would imply, it seems to me, that the way interactivity works would be like a personal agent. (automations are the non-interactive side) I lost that argument. In a sense, we always walked a few steps on that path with the OpenHands Slack agent (since late 2024!), but we stopped at basics. You can mention it, that is, prompt it, and it can do its stuff as usual (yes, SWE stuff), and that’s about it. It doesn’t feel like a personal agent, and I don’t think - to my knowledge - that it feels like an “AI colleague”. I do want to say though: I notice, as we all do, that after the OpenClaw explosion (which went down! apparently), personal agents are actually showing up in the industry lately quite a bit. And TBH I’m not surprised: it’s my prediction that personal agents/AI colleagues, customizing them for both user and team and public, are only going to grow. People will try things, test limits, maybe get some unexpected UX right; while going to a web app to prompt them… is not part of the future. I could be wrong. Hey, it’s happened before! I want to record this prediction; as you may have seen, I’ve made the argument for a while; just want to be clear about it, and be burned at stake for it. Pinned. 😅

Hi again. I want just to give you some update on the experiments about context condensation strategies that I have run them on gpt 5.6 luna on easy tasks from the SWE verified bench.

image.png,image.png

EpochMask is an improvement of ObservationMaskingCondenser, and it looks very promising 😄

Hi again. always on the condensation(compact) strategies bench. I tried running gpt-6 Luna on 10 hard tasks from swe bench PRO v2. I compared the baseline (uncapped gpt luna 6) with two configs as input limit size : 64K and 32K. my goal is to track the cost and pass rate. The baseline solved 8/10 task for $0.99 ($0.12 per solve). Conversation window at 64K solves 9/10 tasks for $0.89 ($0.10 per solve). LLM summary solved too 9/10 for $1.03($0.11 per solve). I share also with you the full paper for more details( it's still unpublished). I would like to hear your feedback, questions or recommendations...