Read-only sessions a client can rely on (follow-up to #1063) #2827
SeanLF
started this conversation in
Protocol Suggestions
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
#1063 was closed to revisit with the orchestrator work. Is that tracked anywhere? Here is a client case for it.
I maintain via, which runs Codex (through codex-acp) and OpenCode as subagents of another coding agent. Its one hard promise is a read-only run: when the user asks for one, the agent can't change anything. ACP gives a client no way to ask for that or check it, so via works it out per agent:
read-only. Mode ids are free-form; I found this one in codex-acp's source.planmode still allows bash. via adds its own agent with every write tool denied, and turns off project config so a repo can't turn bash back on.model_configandspeedcategories, but the spec says categories must not be required for correctness.So each new agent means reading its source, and a wrong guess fails open.
What would fix it:
This doesn't ask ACP to sandbox anything; the security page rules that out. It asks for a defined name for a guarantee some agents already enforce.
Does this belong in the orchestrator work, or in config options? If config options, it would need a reserved meaning, since categories are UX only.
All reactions