Showing the working — `/details`

/details is a switch in the interface that decides how much of a turn you see. A turn is one request to the agent and what it does in reply. Turn it on to watch every tool call and every line of thinking while you debug; turn it off to read only what the agent said.

Try it

Type /details on in the interface to see everything the agent does.

/details            flips it
/details on         tool calls and thinking are shown
/details off        only what the agent says

You now see the agent's tool calls and thinking alongside what it says. Type /details off to hide them and read only what the agent says.

What each setting looks like

Off shows the transcript as the answer: what the agent said, turn after turn. This is the reading view. A nine-hour Ralph loop reads as a conversation.

On also shows every tool the agent reached for and every line it thought aloud. This is the debugging view. It is where you find out that the agent read the wrong file, or spent four minutes grepping.

Both settings draw from the same events: text, reasoning and tool. The switch only decides which of them are drawn. Nothing is discarded. Turning it back on does not recover what scrolled past, but the trace has all of it either way, always.

It is a screen setting, not an agent setting

/details changes nothing about the run. The agent is not told about the setting, and it does not think less because you are not looking. A flow that is running is not restarted. Nothing about the cycle or the trace is different.

From Python

/details does not exist outside the interface, because there is no screen. The event stream carries the same information:

python

for event in session.stream("write the tests"):
    if event.kind in ("text", "result"):
        print(event.text)              # /details off
    elif event.kind in ("reasoning", "tool"):
        ...                            # /details on adds these

Or over a whole agent, whichever session it came from:

python

def looking(agent, session, event):
    if event.kind == "tool":
        print(f"{agent.id} used {event.text}")

agent.watch(looking)

When to reach for it

On, the first time you run an unfamiliar flow, to find out what shape its turns are.
On, when a turn is taking far longer than it should.
Off, for a run you are reading rather than debugging, and for anything you will /export and show somebody.

Showing the working — /details ​

Try it ​

What each setting looks like ​

It is a screen setting, not an agent setting ​

From Python ​

When to reach for it ​

See also ​