When trying to parse images from the

openhands

CLI, the agent is trying to run this command from a directory that does not exist:

😭 ugh.. CLI doesn't have access to browser so it can't read images..

Issue created successfully! _Title_: [Bug]: CLI Unable to Parse Images Due to Browser Access Restriction _URL_: https://github.com/All-Hands-AI/OpenHands/issues/1

😅 seems like a bug

it happens!

@nutritious-balloon-65098 please create an issue that tracks this. We need a separate set of prompt for OpenHands CLI that removes the browser action space

Possibly related, getting this

DeprecationWarning

a lot:

/Users/seth/venvs/py3_12/lib/python3.12/site-packages/httpx/_models.py:408: DeprecationWarning: Use 'content=<...>' to upload raw bytes/text content.
  headers, stream = encode_request(

I'm on it! Xingyao Wang can track my progress here.

Are you on latest main? I tried to remove this one for a while and after adding it to suppress warning, it still didn't work 😞

No, I did just run

pip install openhands-ai

this morning though

hmm, you could also try

pip install git+<https://github.com/All-Hands-AI/OpenHands>

and see if it is better

Summary I have successfully created comprehensive documentation to track the issue of OpenHands CLI needing separate prompts that remove the browser action space. What I Accomplished Problem Analysis: I thoroughly explored the OpenHands codebase and identified the root cause: • CLI runtime automatically disables browsing (

enable_browsing = False

) • Browser tools are correctly excluded from the agent's tool list • However, system prompts remain unchanged regardless of runtime type • This causes the agent to still attempt browser-based solutions (like trying to parse images with browser actions) Documentation Created: 1.

CLI_PROMPT_ISSUE.md

- Complete GitHub issue content following the repository's feature request template, including: ◦ Detailed problem description with code references ◦ Root cause analysis ◦ Proposed technical solution ◦ Implementation plan with specific file paths ◦ Benefits and example scenarios 2.

CLI_PROMPT_IMPLEMENTATION_PLAN.md

- Comprehensive technical implementation guide with: ◦ Directory structure for CLI-specific prompts ◦ Code changes needed in

codeact_agent.py

and

PromptManager

◦ Detection logic for CLI mode ◦ Testing strategy and migration path 3.

create_cli_prompt_issue.py

- Python script that formats and displays the issue content for easy copying to GitHub 4.

ISSUE_SUMMARY.md

- Executive summary of the entire analysis and proposed solution Key Findings The issue occurs because while the CLI runtime correctly disables browser tools, the agent's system prompts don't change based on runtime type. This leads to the agent attempting browser-based image parsing that fails with "Browser functionality is not implemented in CLIRuntime" errors. Proposed Solution Create CLI-specific prompts in

openhands/agenthub/codeact_agent/prompts/cli/

that: • Remove all browser action references • Add explicit CLI limitations guidance • Suggest CLI-appropriate alternatives (command-line tools, Python libraries) • Focus on available tools (bash, file editing, etc.) Next Steps The issue is now fully documented and ready to be created in GitHub using the content from

CLI_PROMPT_ISSUE.md

. This will provide a clear tracking mechanism for implementing the separate CLI prompt system to resolve the browser action space conflicts in CLI mode.

I know you guys already have a long roadmap...but the actual fix would be letting the CLI handle images, no?

@nutritious-balloon-65098 please use GITHUB API to create a Github issue that describe the limitation of CLI cannot read image

I'm on it! Xingyao Wang can continue tracking my progress here.

@happy-helicopter-7895 yeah that's true, but i remember @full-greece-2387 recently add image reading support in the openhands-aci (https://github.com/All-Hands-AI/openhands-aci/blob/6bd03e6a68bf5fbf831116e438ae34ae796b88d2/openhands_aci/editor/md_converter.py#L849-L879) Maybe we can just get that working across the board and stop prompting agent to use browser to view images

That would be incredible! Context: I'm trying out OpenHands on building a complicated feature now that involves a wireframe, and Codex / Claude Code do currently read images. Earlier this morning OpenHands (with Claude) handled a task involving writing and running some data migrations very well (Codex was stuck).