Skip to content

TESTING.md

Latest commit

 

History

History
100 lines (85 loc) · 4.9 KB

TESTING.md

File metadata and controls

100 lines (85 loc) · 4.9 KB
Raw

Testing SysMon

Five layers, cheapest first. The live layers deliberately use the machine as fixture (a monitor passing only synthetic data proves nothing).

layer command needs
parser units + fixtures cargo test -p sysmon-core --lib nothing (pure)
settings migration cargo test -p sysmon-app --lib nothing
accuracy cross-checks cargo test -p sysmon-core --test accuracy a live Linux box
live probes (smoke, icons, net attribution) cargo test -p sysmon-core --test live_smoke --test live_icons --test live_net_attribution live box; net test wants WAN + skips offline
UI interactions cargo test -p sysmon-app --test ui_kittest GPU or lavapipe (wgpu)
full e2e scripts/e2e.sh out/ target/release/sysmon Xvfb, openbox, xdotool, imagemagick, jq

cargo test runs everything but the e2e script. All 41 green is the resting state.

The accuracy suite (crates/sysmon-core/tests/accuracy.rs)

Each check compares a collector against an independent authority (free -b, df -B1, ps, /proc//sys re-reads) under a justified tolerance. The law: a red check means fix the collector. Never widen a tolerance to pass. Identities and the tolerance table: docs/dev/ACCURACY.md. The CPU check pins a spinner thread — brief 100%-of-one-core during the run is the test working.

The kittest suite (crates/sysmon-app/tests/ui_kittest.rs)

Drives the real SysMonApp (real sampler, live /proc) through AccessKit — clicks and typing land the way assistive tech would. Patterns to keep:

  • Construct with Harness::builder().build_eframe(…), scratch XDG_CONFIG_HOME set first (the app saves settings on changes).
  • Query by label (get_by_label("Processes")), by Role::TextInput for the filter, get_by_label_contains for count labels. Custom-painted widgets are findable only because glyph_button registers WidgetInfo::labeled. Do likewise for every new hand-painted control.
  • Assert on harness.state() (the app struct) and on the UI where possible (e.g. panel_fill color after a theme switch).
  • One #[test] fn, sequential — harnesses mutate process-global env; parallel harnesses race.
  • After clicking anything animated (glyph buttons, menu rows — they ease via animate_bool for ~0.1 s), use run_steps(8), never run(): run() demands quiescence within 4 steps and the easing legitimately repaints for ~6.
  • click_secondary() opens context menus; click_modifiers does Ctrl+click multi-select. Off-screen columns (horizontal scroll) can't be clicked — set_size the harness wider first.

The e2e script (scripts/e2e.sh)

One command, full live receipts: boots Xvfb + openbox, launches the GUI with scratch config/socket, screenshots all 10 themes × 2 pages through ctl shot, opens a pop-out and docks it by drag-drop, checks single-instance raise, probe-via-socket, tap streaming, exit codes, clean socket removal. Screenshots land in the output dir and regenerate README images (fresh-from-this-build is the law for release screenshots).

⚠ GUI-testing gotchas (each one cost a debugging loop)

  • Bare Xvfb has no window manager: no keyboard-focus semantics (typed keys vanish — xdotool windowfocus --sync or run openbox), and windows cannot move at all.
  • Synthetic XTEST titlebar drags do not drive openbox's interactive move. To emulate a user drag for dock testing: xdotool mousedown 1 + step the window with xdotool windowmove
    • mouseup 1 — the same signal triad (motion + button held + release position) the app's dock logic reads.
  • Never blind-click a row of a live-sorted process table to test kill flows — rows re-sort between coordinate lookup and click. Spawn a sacrificial process (sleep 3000 &) and filter by its exact PID first; the confirm modal must name the sacrifice before you click confirm.
  • Scratch env or you drive the real thing: XDG_CONFIG_HOME (settings writes!) and XDG_RUNTIME_DIR (socket ownership!) both. A ctl theme test once silently rewrote the user's saved theme.
  • pkill -f <pattern> kills your own test shell when the pattern appears in its command text (exit 144, script dies mid-run). Match readlink /proc/<pid>/exe instead.
  • Xvfb renders via llvmpipe (software): CPU figures there are worst-case, not real-GPU truth. Label them; the resource table in RECEIPTS.md does.
  • ctl shot captures the main viewport only; compose pop-outs with import -window root on the Xvfb display.

Resource measurement recipe

# own CPU% over a window, no tools needed:
clk=$(getconf CLK_TCK)
t0=$(awk '{print $14+$15}' /proc/<pid>/stat); sleep 60
t1=$(awk '{print $14+$15}' /proc/<pid>/stat)
echo "$(( (t1-t0) ))" # ticks; % of one core = ticks/clk/60*100
awk '/VmRSS/{print $2}' /proc/<pid>/status   # RSS KiB

Let the GUI settle ~8 s first (icon index + first paints). Baseline numbers to beat are in docs/dev/RECEIPTS.md.