What CmdTab has proved—and what it has not

This page separates reproducible automation from real-desktop acceptance work. The goal is not to turn test counts into marketing claims; it is to make the evidence and its limits inspectable.

Last reviewed: July 21, 2026

Product alignment commit: 0d042e9600e9

Automated evidence

Repeated checks that block regressions

Each item below is exercised by committed tests or verification scripts. A green check means the defined contract passed in that environment; it does not convert a simulator, model, or CI host into proof of every macOS desktop state.

Strict exact-window MRU regressions

Focused acceptance tests verify forward and reverse selection when the adjacent recent window belongs to the same application, ambiguous frontmost PID resolution, and preview-independent membership.

Environment: GitHub-hosted macOS 14 and macOS 15 runners

Complete Swift package suite

The complete SwiftPM test suite runs after the focused contract tests on both supported CI runner generations.

Environment: GitHub-hosted macOS 14 and macOS 15 runners

Reproducible pre-fix state-space model

The checked model covers 19,500 meaningful activation sequences and records 3,900 pre-fix PID-skip selection mismatches. It also covers 126 preview-success configurations and records 120 pre-fix completeness failures.

Environment: Dependency-free Python model; these figures describe the modeled state space, not a measured field failure rate.

Rendered production website verification

The production crawler and browser suites validate every maintained public route at desktop and mobile sizes, metadata, structured data, internal links, images, navigation, wide-table accessibility, private-route index protection, and the interactive demo.

Environment: Compiled Next.js production server and public cmdtab.net deployment

Downloadable artifacts

Inspect the model, matrix, and execution plan

These files are published from the same repository blobs used by the protected switcher review. The website build fails if the public model or matrix diverges from its canonical QA source.

Switcher model results (JSON)Machine-readable model counts and representative counterexamples for the pre-fix behavior.Open artifact →Comprehensive switcher test matrix (CSV)The 136-case matrix covering membership, MRU, session input, activation, Spaces, displays, permissions, performance, compatibility, and release checks.Open artifact →Implementation review and execution plan (Markdown)Detailed acceptance contract, findings, live test procedures, test architecture, and definition of done.Open artifact →

Manual boundary

Real macOS conditions that automation does not honestly replace

These scenarios require an interactive Mac, real permissions, real Spaces and displays, and focused-window evidence. Until each relevant row is executed, the public claim is limited to the automated contract above.

Accessibility and Screen Recording permission transitions on a real signed app
Selected identity versus the actual focused CGWindowID for keyboard and mouse commits
Cmd-`, Mission Control, Stage Manager, minimized windows, and native fullscreen behavior
Current Space, visible Spaces, all Spaces, and off-Space activation
Single-display, multi-display, mixed-scaling, mirror, disconnect, and reconnect behavior
Secure Input, event-tap recovery, sleep and wake, signing, notarization, and clean-account installation

How to interpret the numbers

State-space counts are not field failure rates

The model counts every selected configuration in a deliberately small synthetic universe. Its 3,900 selection mismatches and 120 completeness failures describe the old rules inside that universe. They do not mean that 20% or 95% of real CmdTab use failed. Their value is reproducibility: the repaired implementation must satisfy the explicit invariants for every covered state.