Hi , I missed the community feedback meeting, so I wanted to follow up with a couple of questions here. I noticed that the project currently supports shell script (.sh) command hooks. Is there a roadmap for agent-level hooks in the future? Specifically, I'd like to know whether sub-agents in a multi-agent setup will be able to hook into each other's tool calls. My goal is to build a security agent that monitors other agents through hooks — I'm not sure if this is already supported or on the roadmap, and I may have missed it in the documentation. For a pre-configured MCP server, I couldn't find a way to restrict or filter the available toolset. Is that something that's currently supported or planned?

Hi there, to your point it looks like our hooks support may be focused on executing shell scripts: https://docs.openhands.dev/sdk/guides/hooks I'm curious if you can share more details about agent-level hooks? How does this differ and are there any good examples from other tools we can look at? Thanks for any info 🙏

@elegant-branch-54423 Hi Joe, The hooks fire the same way but instead of writing a script that explicitly handles the management, you hand the decision to an LLM with a plain instruction. Imagine you want to restrict grep from searching outside the project directory. With a command hook, you write a script that parses the payload, extracts the command, and checks it against a hardcoded list of blocked patterns like /etc/, ../, ~/. It works, but agent can behave a path variation your list didn't expect. With a prompt hook, you just describe the policy: "block any grep command that searches outside the current project directory." The LLM reads the actual command and applies judgment based on intent. I can abstract as hook generatıon via llm prompt. I need to mention that i havent tested out the agentic hooks yet.But I saw a documentation about it. prompt-based-hooks

Thanks for the feedback! I understand it now. I wonder if prompt-based can offer the same guarantees as shell-based/command-based hooks. Are there other use cases you have found for prompt-based? Also, this might be a great topic to bring to #C06R25BT5B2 or even create an issue in the SDK repo for some of our engineers to weigh in on! https://github.com/OpenHands/software-agent-sdk/

Hey Joe @elegant-branch-54423 , You're welcome! Took your suggestion and went digging through the SDK repo. Turns out there's already an issue covering part of what I had in mind — #2755, opened last week, flags the gap around prompt-based hooks (

HookType.PROMPT

is declared in the enum but has no implementation)However, agent-based hooks weren't mentioned there, so I just opened #2864 to cover that side. https://github.com/OpenHands/software-agent-sdk/issues/2755 https://github.com/OpenHands/software-agent-sdk/issues/2864

Thank you! Going to flag @modern-fireman-99755 on #2864 so he can also take a look and weigh in

I answered in the issue...