Interesting direction. For a self-hosted agentic chat, I'd be curious how you handle long-running sessions and handoffs between users, since that's where these workflows usually get messy in real teams.
The peer-to-peer/outbound-only part is the most interesting bit to me. For coding agents, the tricky part is usually not starting a remote session, but making it clear when the agent is still running, blocked, or has changed files that need review before the next handoff.
Interesting benchmark idea. I'd be curious whether the scoring separates geometry correctness from code maintainability, since generated Three.js can look right while still being hard to edit or extend.
Nice idea, especially for onboarding to an unfamiliar repo. The hard part is usually keeping the explanation grounded in the actual code paths, so it would be useful if the map links every summary node back to the files/functions it was derived from.
This hits a real pain point in larger repos: the expensive part is often not one bad prompt, but stale logs and generated files silently getting pulled back into context. I’d be curious if the firewall mode supports per-task allowlists, because that seems useful for reducing false positives without making people maintain a huge global ignore file.
The maintenance-cost framing is the useful constraint. I’d rather see agents default to smaller diffs, test scaffolding, and explicit assumptions than maximize lines changed per prompt.