We've been discussing Goose's architecture a lot, but a practical question came up from a member's DM: if you suspect a Goose agent is compromised, what's the right IR process? The local execution context and extension model create some unique challenges.

Unlike a traditional cloud agent, a compromised Goose agent has direct access to the user's machine, their local network, and potentially any credentials stored in its memory or configured tools. The open-source core is a plus for audit, but a malicious extension is the most likely vector. So, containment must be immediate and local.

First step is to isolate the host—network disconnect is ideal. Then, you need to preserve evidence of the agent's state before termination. This means capturing the agent's process memory, its working directory, and the exact versions of the core and all extensions. The `goose --list-extensions` output and the extension cache are key.

The tricky part is that killing the agent process also clears its volatile state. Do you prioritize forensic capture or immediate termination? How are you handling credential rotation for the services the agent had access to? Let's share concrete steps and tooling. I'll start: a script to dump the agent's env and procfs data before a graceful stop.