i had a feeling it might need to be that way, so i asked LoL~ Still, i think it's gret that it works without tailscale or any other dependencies. thaks for sharing such a greate project
Interesting approach. Making agentic AI accessible through visual UI instead of text commands is a real gap right now.
The dual interaction model — where both the user and the LLM can trigger the same functions — is a nice design choice. It avoids the "watch the AI work" problem where you're just a spectator.
Curious about the protocol design: how do you handle conflicts when the user and LLM try to act on the same element simultaneously? And is there a way for MUPs to communicate with each other, or is each one isolated?
The clarification protocol is a smart approach. In my experience running multiple Claude Code agents on the same codebase, the biggest gap isn't prompting — it's visibility. You don't know what each agent decided to do until you check the git log afterward.
Structured docs like ARCHITECTURE.md help agents make better decisions upfront, but I think there's also a need for runtime feedback — knowing which agent changed what, and whether it drifted from the original task while it's still running.
How does oh-my-agent handle multi-agent scenarios where two agents might touch overlapping files?
The attack chain you described highlights a gap that most teams overlook: AI-generated code passes functional tests but skips the "why this version?" review that experienced developers do instinctively.
I think the real issue is visibility. When AI generates a project, every dependency choice is implicit — there's no PR comment explaining why it pinned [email protected] instead of 14.2.1. In a human workflow, someone would have caught that during review.
Two things that have helped in my workflow:
1. Running `npm audit` as a post-generation step before even testing functionality
2. Treating AI-generated commits as "untrusted by default" — reviewing them with the same rigor as external contributor PRs