A desk that measured 162 context switches an hour

My next request had nowhere to go. The central hub screen was locked while it worked through another problem, and I was hunting for a place to type.

I had treated that blockage as a missing interface. I thought a fresh hub would get me moving, but it would have created another place for work to hide. The trouble was more literal: I could not see the working desk as a working desk. Sessions, windows, queues, and the text box that started the next job all existed, but they made me reconstruct the moment from different screens. I could drill into a department after I knew where to look. I could not reliably see what was happening across the monitors while it was happening.

The original requirement was more mundane than observability. I wanted a live record of what software and tabs were on each monitor, plus their relative locations and sizes. That scope mattered. It gave the desk a bounded contract. A map of four monitors did not need to infer intent before it could be useful. It could first reduce the scavenger hunt that starts when a session is visible somewhere but the route into it is not.

I started with the thing in front of me. The new desk drew four monitors to scale, then put a named tile for each session on the relevant space. Clicking a tile opened its details without making me hunt for that window. The window records gave the map a necessary distinction. rect described the operating system’s positioning space. visible_rect described the DWM frame on screen. Each record also carried onscreen, on_attention, and foreground status. A foreground window had keyboard focus and was marked as something the desk must never move.

That was supposed to be a visibility tool. It quickly became an attention recorder. The first measured hour showed 162 context switches, an average of 23 seconds on a window, and Discord taking 43% of the switches. I had expected the desk to make my working state easier to navigate. It instead made the fragmentation visible as a count that was hard to explain away. The map could show a healthy looking grid while the attention log showed that the grid was being abandoned again before a thought had time to settle.

Then Codex challenged the model underneath the picture. I had treated geometry as identity, which meant that where a tile sat could be used as evidence of which task and session it represented. Codex called that core design unsafe. The live sample produced confident but wrong task to session pairings, and the review found four things the feed was claiming without actually knowing them. A scaled rectangle can show placement. It cannot prove provenance. I had already been reading those pairings as fact for as long as the map existed, deciding which window to trust based on a tile position the tool was never actually in a position to verify. Once I saw that, a neat visual answer became an honest boundary: the desk could report what the display gave it and leave the unsupported connection unclaimed.

The review was not the only place the display exposed stale assumptions. The desk grid began sorting by session date and stopped skipping minimized windows. A display broker could place pages on the large monitor without taking focus or evicting what was pinned. Under that behavior were two state failures that a window tidy would have missed: a daily cron marked live sessions abandoned, and automation sessions never closed themselves. Fixing the picture without fixing those two sources would have made the desk more legible while preserving false status.

A preclose check found the same issue in the health path. The alarm called a healthy watcher dead. At that point, a dashboard that could draw every monitor, name each tile, and chart every switch was still only as useful as the state it declined to invent. I stopped treating the visualization as a control plane. It was an instrument with a precise job: show the displayed work, record observed attention, and make any inference earn its evidence.

The final scene was smaller than the four monitor map, and it’s the one I trust most precisely because it doesn’t try to explain itself: a desk lease stale for 65 minutes, three lanes queued behind it, no inferred story about why. Just the number, and the queue it was blocking.

dxdev LESSON · paste into your AI coding agent
LESSON: A Visibility Tool Can Turn Into an Attention Recorder You Didn't Ask For
SOURCE: dxdev.com/blog/2026-08-27_live-desk-attention-recording
WHAT HAPPENED: A desk map was built with a bounded contract (draw four monitors to scale, place a named tile per session, record onscreen/on_attention/foreground status) to fix a navigation problem, finding a window when it existed but the route into it didn't. Once running, it also measured 162 context switches in the first tracked hour, 23-second average attention per window. A separate review found the design had been treating window geometry as task identity, meaning the display could produce a confident but wrong task-to-session pairing, a rectangle proves placement, not provenance. The fix was narrowing the tool's claims to only what it could actually support: show displayed work, record observed attention, and leave any inferred task/session connection unclaimed rather than guessed.
THE RULE: A monitoring or visibility tool's on-screen position for an item (a window, a job, a request) is not the same fact as that item's identity or ownership; treating "where it's rendered" as "what it is" produces confident-looking output that can be wrong in ways nothing in the display will flag, so any inferred connection needs to be explicitly weaker than a directly-observed fact, not presented with the same confidence.
CHECK MY CODE, then report PASS or FAIL with file:line for each:
  1. Any dashboard or monitoring display that infers an entity's identity, task, or ownership from its screen position, sort order, or grouping, rather than from a directly verified source.
  2. Any status field (healthy/dead, active/stale) whose source is a heuristic (an age check, a cron-set flag) that has not been checked against ground truth recently, risking a healthy thing reported dead or an abandoned thing reported live.
  3. Any personal or operational metrics tool whose output revealed an uncomfortable real number (a fragmentation rate, an error rate, an abandonment rate) that hasn't yet been acted on.
THEN PRINT: a table (check, PASS/FAIL, evidence, fix) + a verdict (applies / partially / OUT_OF_SCOPE / no) + the single most important next action.