Skip to the content
Krizaka Blog

35

The ergonomics of a cockpit: one screen per question

Nov 15, 202612 min read

What I faced, Teams of agents working at the same time, and no single place to look; What we built, a local cockpit where each screen answers one question, each figure with its age

Ink and watercolor: an empty chair at a small wooden desk with a brass lamp in a signal box, a wide window over the tracks of a station and a waiting train, a long panel of small green and red lamps above it, a kestrel on the desk looking out

I run Krizaka with teams of agents. They open tickets, write code, send pull requests through merge trains, and a watchdog keeps an eye on them. I look at one screen: the cockpit, an app that runs on my Mac and nowhere else.

In mid-November I rethought it around one simple idea: each screen answers one question, the one I ask at that moment of the day. If I cannot say a screen's question, the screen does not belong in the bar.

The screens are drawn by pen, from their code, not captured. Each number on a drawing points to a line of the legend. The figures come from the evening of 15 November 2026.

One question per screen

I put them here in the order I open them, not in the order of the menu.

The screen and its questionThe ergonomics choice
Today: what is waiting for me?one sentence at the top, and only what needs my hand
The age of figures: can I trust what I read?every block says how old its data is
The machine: is everything running?a verdict, then six rows to unfold
Alerts: what is broken, since when?one row per problem, with how many times it came back
What ships: what actually went out?one card per product, dated by its oldest source
Costs: what does it cost?a total that says “at least” when an amount is missing
Architecture: does the code still look like the diagram?the diagram drawn from the repositories, the gaps counted
The workshop: where is that tool?folded, because I do not go there every day

The cockpit reads GitHub, our local databases and our files. It never writes to GitHub: a ticket, a pull request or a required check is changed with GitHub's own tools, which enforce their rules on everyone, us included.

Morning: what is waiting for me?

I open the cockpit on “Today”. It is the screen I wanted before all the others.

  1. 1“Steer”: the bar's first group: Today, The machine, What ships, Costs. What I open every day.
  2. 2The top sentence: “38 things are waiting for you. Everything else moves without you.”
  3. 3“Sources to watch”: the strip names only the sources that are old or failed; it is gone when everything is up to date.
  4. 4Three figures: my actions, the decisions, the next date to keep.
  5. 5“Your actions”: my to-do note, then the tickets waiting for my hand; each row has one button, “Open”, and the card's foot says when the board was read.
  6. 6“Decisions to take”: what the CTO session asks me to settle, newest first.
  7. 7“Workshop”: the folded group: it opens on a click, or by itself when I am on one of its screens.
The cockpit's “Today” screen

On the evening of 15 November the sentence said 38. It counts what waits for a person, not what the agents are doing. When nothing waits, it says “Nothing is waiting for you. Everything moves without you.” The agents can write hundreds of lines a day; what limits the company is the queue of decisions they cannot take. So that queue is at the top, and the rest below.

“Your actions” shows only the urgent ones first. A “Show N more (not urgent)” button opens the rest, and the figure at the top says “urgent, out of N on the board”. I do not sort by hand every morning: the screen does it with the board's labels (founder, Blocked, “To validate – founder”).

The bar never changes shape. Four groups: Steer, Company, System, and the folded Workshop. Everything else sits behind a search (⌘K), which also finds the screens not filed yet. That is progressive disclosure: what serves every day is visible, what serves rarely is one click away.

Can I trust this figure?

A dashboard that shows a two-day-old figure without saying so lies by omission. In the cockpit every block carries the age of its data, and a “Refresh” button reads that block again from its source.

Before

  1. 1“2 d old”: the badge, with a word and an icon, not only a colour.
  2. 2“Refresh”: reads this block again from its source, and nothing else.

After

  1. 3“Up to date”: the same badge, with the date of the new gathering.
  2. 4The date of the data: it stays written next to it, even when all is well.
A block gone stale, then refreshed

The rules are decided in one place in the code, for every screen. Each source has its gathering cadence: 15 minutes for the board, 5 for the teams, one hour for the releases. A block turns “old” once it has gone past twice its cadence. If the last gathering failed, the badge says “Failed” with the date of the data still shown, whatever its age, and counts the failures in a row. Alerts and scheduled tasks are read at every visit.

That is Nielsen's first heuristic: the screen says what is going on. And the badge does not rely on colour alone, as WCAG 2.2 success criterion 1.4.1 asks: the word “old” reads in black and white too.

A full gathering takes about fifteen seconds. I chose a click and a visible wait over a screen that would refresh itself while I read it.

Is the machine healthy?

“The machine” has six tabs: Health, Delivery, Quality, Rules, Automations, Alerts. I almost only ever open the first.

  1. 1The tabs: the open tab is in the address, so a link brings me back to the same place.
  2. 2The verdict: “The machine is fine”, “needs a look” or “has a failure”, with the number of blocks to look at.
  3. 3One row per block: Teams, Trains, Crons (the scheduled tasks), Alerts, Locks, Cockpit refreshes, each with its verdict in words.
  4. 4The date of the source: on every row, not once at the top of the page.
  5. 5The fold: the row opens on its detail: the list of sessions, the table of tasks.
The Health tab of “The machine”

The question is closed: yes or no. The verdict answers first, in one sentence. The six rows come after, and their detail stays folded until I ask for it. On the evening of 15 November the page counted 13 team sessions at work, 4 red pull requests in the trains, 14 active scheduled tasks and 16 open alerts, 2 of them serious. The verdict said to take a look.

The “Locks” row shows the heavy-work slots in use and a session that holds one too long. The machine also limits itself: that evening a team launch waited because the Mac's five-minute load average was 86, above the brake of 60. I had nothing to decide.

What is broken, and since when?

On 15 November at 1:24 p.m., the krizaka-ui release stopped. On the cockpit it was one row of the Alerts tab: critical, raised by the watchdog, “main red: krizaka-ui release”, counted sixteen times between 2:15 and 6 p.m.

  1. 1The counts: open, acknowledged, resolved, at the head of the table.
  2. 2“Severity”: in a word and a colour; the table sorts on it.
  3. 3“Message”: the first line of what the watchdog or a session wrote.
  4. 4“Count”: how many times the alert came back: an alert counted sixteen times says the problem lasts.
  5. 5“State”: open, acknowledged by someone, or resolved, with its proof.
  6. 6The filters and the pages: to narrow the list, and 25 rows a page.
The Alerts tab of “The machine”

That evening, 196 pull requests had been merged in 24 hours. Nobody reads 196 notifications. An alert that comes back does not add a second row: it raises the count of the first. It is the idea of Google's book on reliability, in its chapter on monitoring: an alert must call for an action, and it is about the symptom (main is red); the cause is looked for afterwards.

The cause is exactly what the screen does not give. Mouloud, the CTO session that steers our teams, read the full log on GitHub: the runner had received a shutdown signal during the checks, out of memory. He wrote the ticket and launched a team; the fix was merged at 6:34 p.m. and the next release went through. My part was to see the row and let them work.

What actually went out?

“What ships” has three tabs: By product, Releases, Board. A project is judged by what reached people, not by what was planned.

  1. 1The tabs: by product, the releases of the last 30 days, then the board, from epic to task.
  2. 2One card per product: the tickets in progress, in review, next, and those blocked.
  3. 3The train: the open pull requests, and how many are red.
  4. 4The last release and the last deploy: what is really live.
  5. 5The card's date: that of its oldest source: a card does not look fresher than its worst data.
  6. 6The release link: it opens its page on GitHub; the cockpit does not write there.
The “What ships” screen, By product tab

A card mixes several sources: the board, the trains, the releases. If one of them is old, the whole card takes its date. When GitHub did not answer for a repository, the Releases tab names it instead of acting as if it had shipped nothing.

What does it cost?

“Costs” is a screen I show no figure of, here or anywhere. I describe it by what it does.

  1. 1“Refresh”: top right, as on every screen that keeps a copy.
  2. 2The month's total: hatched here; on screen it says “at least” when an amount is missing.
  3. 3“N amounts missing”: the badge that says why the total is a minimum.
  4. 4The fixed part: what is written by hand; what is still to confirm is listed, never guessed.
  5. 5The stale block: its source is a report run by hand: the screen says so instead of showing it as fresh.
  6. 6“Estimates”: what is estimated is kept apart from what is known, and written as such.
The “Costs” screen, without its amounts

Everywhere, an incomplete figure says so. A total that leaves out a line without warning, I would believe.

Does the code still look like the diagram?

“Architecture” answers a question I did not ask often enough: is the diagram I have in my head still true?

  1. 1The facts at the top: how many repositories are in step with main, how many gaps, and the date they were read.
  2. 2One tab per system: Sovkora, Verikora, Folioviv, the libraries, the sites, the dependencies, the code graph.
  3. 3The drawing: the systems (C1), then their containers (C2), drawn from what each repository declares.
  4. 4A flagged box: for instance a library our repositories use at a version behind the latest.
  5. 5The text version: the same drawing in sentences, under the diagram, for whoever does not see the picture.
The “Architecture” screen

The diagram is not drawn by hand: it comes from each repository's architecture folder, and the screen counts where main no longer looks like it. Without that comparison, the diagram would have aged without my seeing it. The architecture of our agents is told in another chapter.

The company, less often

The “Company” group serves less often, and I draw none of it.

“Company” keeps the company's sheet, our roles and the brand. “Business & quotes” is where I prepare an offer for a client; what is saved there goes to our local files, never to GitHub. “Community” shows whether the publishing calendar is kept, planned against published, with the date each network was last read. “Documents” lists our notes and documents; a private document shows only its name there.

The workshop, folded

The Workshop is folded in the bar because I go there when I need a tool, not to know where the company stands. It has five screens: Workspace, Commands, Secrets & users, Media library and Coordination.

  1. 1“Local only”: the header's badge: nothing on this screen leaves the Mac.
  2. 2The two tabs: the test accounts, by app, and the cloud secrets.
  3. 3One name per row: the screen holds names only; the value is never read into the page.
  4. 4Where the value would be: hatched on the drawing; on screen it does not exist, there is nothing underneath.
  5. 5“Copy”: the Mac itself puts the value on the clipboard; the page only learns “Copied”, and the clipboard is cleared after 30 seconds.
The “Secrets & users” screen, in the Workshop

Masking a value that is displayed is not enough: a blur can be undone. Here the value never reaches the page, so there is nothing to recover.

“Commands” is a closed list of 11 local commands (file the inbox, re-index the code, refresh the indicators…). Each one has a single button, “Run”, which opens a confirmation; those that cannot be undone say so before the click, and the output scrolls with secrets masked. “Workspace” shows our folders, the free disk space and the bin. “Coordination” shows the waves of work and the tasks the agents share. “Media library” lists the pictures and films we made.

What the screen does not do

The cockpit exists only on this Mac. It listens only on the machine, sends nothing to anyone, and I do not look at the company from my phone. That is on purpose: nothing from our folders goes to a third party.

It shows; it does not explain. The release alert said main was red, since when and how many times; the why was in GitHub's log. And figures age: that evening two badges said “2 d old”. Refreshing them takes a click.

How these drawings are made

The screens are drawn from the cockpit's code, read on 16 November 2026: the order of the blocks, their names and their buttons are the code's. They are not to scale and hold no data. The figures (38, 13, 86, 196, sixteen times) were read on the cockpit on the evening of 15 November. Nothing was timed, and nobody but me uses this screen: this is the ergonomics of a tool for one person, not a study.

Chapter 35

Results

things waiting for me on “Today”, on the evening of 15 November

38

Morning: what is waiting for me?

team sessions at work on the health page

13

Is the machine healthy?

the Mac's load average that held a team launch back, above the brake of 60

86

Is the machine healthy?

pull requests merged in 24 hours, that evening

196

What is broken, and since when?