[bug] `!<cmd>` output invisible to Codex (defeats the entire Product UX purpose of the feature)

This one is pretty important to me too. One of my typical use cases is to kick off a lint via a command like !yarn lint. If I see that any issues were returned i could tell the LLM to work on those issues. It adds quite a bit of utility being able to directly get the output of the command into context then act on it.

This is WAI we currently don't inject output in the conversation history. We are having discussions around this. Closing this for now.

The workflow you've left us with:

# what openai built
!ls              → runs, output goes nowhere useful

# what we'd do anyway
Cmd+T            → new terminal tab
ls               → see output
Cmd+C, Cmd+V     → paste into chat if needed

So !cmd saves us... nothing. Arguably costs us, because nothing in the UI signals that the agent is blind to the output.

The feature only makes sense if the loop closes. Otherwise it's a shell escape that looks like integration but isn't. Worse than not having it, because it sets an expectation it doesn't fulfill.

"WAI" on this is essentially admitting the feature is decorative.

Proposal

Suggested behavior:

!cmd   → executes, output hidden from agent (current)
!!cmd  → executes, output injected into conversation

This preserves the current default while enabling the collaborative workflow that makes the feature useful. Minimal surface change, clear mental model.