I noticed a Steer behavior in the Codex app that might be useful for staged workflows.
What I expected at first was that Steer might inject a new instruction directly into the current thinking / response.
What I’m seeing instead is different: it seems to behave more like a queued follow-up. The current thinking and response continue, and the steered message is only applied after that current turn finishes.
In other words, this feels more like queueing than inserting.
That turned out to be useful for me, because it lets me split a workflow across multiple follow-up calls without having to wait manually for each turn to finish.
A simple reproduction on my side was:
-
start a task
-
while Codex is still thinking / responding, send
go -
do that multiple times
In one test, I entered go five times. What happened was not “interrupt current reasoning and switch immediately,” but more like “finish the current turn, then consume the queued follow-ups one by one.”
That makes Steer surprisingly usable as a lightweight workflow sequencer.
I’m specifically talking about the Codex app here, not making a broader claim about every Codex surface.
Curious whether others are seeing the same thing, and whether this is intended behavior or just the current implementation.