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

I’ve experienced something similar, but I haven’t reached a firm conclusion yet. One thing I’ve noticed is that DeepSeek seems to become less reliable when overloaded with context. I also found that my

AGENTS.md

had become very bloated with historical and implementation context. I reduced it to mostly routing/instructions. I’m not sure yet if that was the cause, but DeepSeek seems to have improved after the reduction. I’ve also been experimenting with building a context map with progressive zoom-in/zoom-out, so the model only loads the context relevant to the current task instead of consuming the whole project context. Still experimenting, so I’d treat this as an observation rather than a conclusion.

its almost stopping after every paragraph and keep typing continue or finish the current task, is not helping. starting a new session and ask to catch up with the previous one also wont help if I have multiple tasks and already finished some of them. Today, i added 2 new global rules and it seems working so far but not the cleanest MD files yet and not preventing the sudden stops. 1. commit task once finished (helped with the multiple task prompt) 2. CRITICAL INSTRUCTION: Do not output raw parameter tags, DSML tags, or XML tool closing tags in text. Execute tool actions cleanly without echoing parser parameters. so far it seems giving better results but not the cleanest. At least I'm not wasting 3 hours on 15 minutes build

Hey all, I have found deepseek to be pretty reliable, but sometimes the providers cause issues. We’ll look into the providers that we’re using for the cloud and see if we need to change them up

Today it seems to be performing well. do you still experience the corruption?

Yesterday and today, the performance was better but not clean. I`m still having some mangling once in a while, but it can be managed. But the issue with the sudden stops, that I couldn`t manage till now.

For what is worth, I’m using openhands/Deepseek-v4-Pro (so the cloud provider) lately and it works very well for me

You didn`t face any of the issues that we posted? So what wrong that I`m doing to keep the cloud provider keep mangling and sabotaging my files? Today was absolute madness, to keep starting new chats and cant finish 1 complete task

I haven’t faced any of these issues, no. It might be that I’m using Pro, not Flash… (it’s very cheap, though not for free)

A useful workaround I found is a Fresh Authoring Recovery method. When the model starts repeatedly producing the same kind of malformed code, we stop trying to repair that corrupted output. Instead, we discard the bad generation, reload the last known-good source, and redo only the affected change from a clean starting point. This seems to help break the loop where each repair attempt keeps reinforcing the same syntax corruption. not really a fix since i think only you and me experience the same issue.

That is helping, yes. But the bigger issue when the corruption happens at the middle of the task or before it ends, especially if the prompt is multiple tasks. In that case a fresh sandbox will never get all the work done as carryover, or saving handoffs also will never all pending tasks completely to the next session. That is a big waste of time during full day building. But the new session is the only solution and some times the corruption is starting at the beginning of the new session as well. this is happening on daily bases, If I`m doing something wrong, please tell me.