nothing gets lost
In short. Everything an agent put on your screen is in one list: the questions still waiting on top, answerable from the keyboard, then the day's history with what each one cost you. Nothing is clipped there, and an item that closed on its own timer while the desk was empty says so in its own row.
Read on if you leave agents running while you are in a meeting. Skip to When it stays quiet if what you want is fewer interruptions in the first place.
The inbox is an ordinary window with two sections. Pending first, then Recent: who asked, what for, when it arrived, and how it ended. The split is decided before the list reaches the screen, so a question can never sink below the answered noise.
Pending on top, answerable with the same keys the card takes. The outcome column is what the day cost you - including the one nobody got to.
One search box covers title, agent, project, kind and outcome. There is no level
filter and no per-agent dropdown, which is a smaller surface than it sounds: typing
release-bot beats finding a menu, and every row already carries its level as a
coloured edge. The footer counts the day, 7 interruptions today, beside a hint
line naming the keys that work right now.
j and k walk the pending run, and the hint under the
selected row names its keys in that item's own words instead of in general
terms: 1 eu-west · 2 us-east · enter dry run. Digits choose, y and
n confirm, Enter takes the default, c copies the
whole item ready to paste back at an agent. Answer one and the selection lands on
the next, so a queue built during a meeting clears without the mouse.
That vocabulary is decided in Go rather than in the window, which is what stops the list and a question on screen drifting apart as either one grows.
Two keys do differ, and knowing which saves a wasted press. Dismissing is d here and Esc on the card. Stopping a countdown is s here and Space there. Anything that needs you to type refuses to be answered in a list at all: Enter raises its real card instead.
On 2026-07-31 a countdown card came up while the desk was empty and expired on its own. Boris went looking for what it had said: "I've missed what you said - so I tried to look in the inbox and there doesn't seem to be a way to find it - this is a problem. I can't see the whole thing you wanted to tell me."
He was right, and the reason was a decision that had looked reasonable in writing. A row carried title, identity, age and outcome, and the body only as a hover tooltip on a resolved row, which truncated at 140 characters. The repo's own status file listed it under deliberate gaps: no per-item detail, promote it to a card to read the body. Promoting was not offered for an item that had already ended.
For a tool whose whole purpose is that a message is not lost, losing the message is the wrong failure.
Since 2026-08-05 every row opens a detail underneath it, through the same renderer the card used, so it reads the way it read on screen. Above it, both timestamps and how long the item stood. Below it, what was offered and which option went back, with the default and the taken one marked separately. Nothing is clamped, shortened or ellipsised.
The card that started it was a veto with a 946-character body. Its row says
proceeded and shows 140 characters. Opening it gives back the whole thing,
rendered, under arrived Aug 5 15:02, ended Aug 5 15:02, stood 8s.
stateDiagram-v2
[*] --> pending: an agent posts it
pending --> held: you are idle, presenting or recording
held --> pending: you come back
pending --> onscreen: its turn at the front of the queue
onscreen --> pending: Esc defers a question for five minutes
onscreen --> answered: you answer, and it ships three seconds later
onscreen --> dismissed: Esc on a notification
onscreen --> expired: the timeout ran out with nobody there
pending --> cancelled: the agent that asked went away
note right of pending
pending is the one state a daemon restart
reads back off disk and presents again
end note
A restart is the interesting branch. Pending items come back out of the store in arrival order, and a burst that was collapsed into one summary comes back collapsed, not as fifteen fresh cards.
What does not come back is the caller. A blocking question whose agent process is gone reads as waiting for a caller that reconnects, and an answer you give it is recorded in the history without reaching anybody. Better than pretending, and worth knowing before you spend a keystroke on it.
Resolved items are pruned once, at daemon start, and only below a level you set. The defaults keep 30 days of everything, and keep warnings, errors and anything urgent for good. Pending is never pruned at any age, which is the rule this page's title rests on.
| What | What happens to it |
|---|---|
| a pending item | kept until it is resolved, whatever its age |
| a resolved warning, error or urgent | kept, at the default keep_level
|
| a resolved info or success | evicted after 30 days |
| a credential | the history records that one was provided, never the value |
a line said with agentbox say
|
never written down at all |
Set keep_level = "info" and nothing is ever evicted. Setting
retention_days = 0 does not mean forever: it falls back to 30.
A line said on its own is not stored, because saying something out loud was never
an interruption to triage later. A line attached to a card is a different thing, and
that one is kept and read back under Said out loud.
Then there is the arithmetic of the day, which is the part that changes behaviour:
agentbox stats --since 7dTotals, how many of them blocked an agent, how many you answered, and the median time you took, broken down per agent and per day. It is usually one agent.
Next: how to make the noisy one stop, or the five levels and the sounds they make.