Tools for automation testing of desktop application, with a structural diff baked into every action

Because the alternative is writing one or two assertions per step (verify field X has value Y, verify button Z is now disabled), which collapses under volume. Once a regression suite covers a real desktop app, the assertions outnumber the actions and the suite becomes a maintenance pit. A structural diff inverts the model: every action records what changed in the accessibility tree, and the test only fails when the change does not match. The fact that Terminator computes the diff inside the agent (crates/terminator/src/ui_tree_diff.rs) instead of leaving it to user code is what lets you snapshot a UI flow with zero asserts.

Through aggressive preprocessing before the line diff runs. For JSON trees, remove_ids at ui_tree_diff.rs line 7 walks the tree and drops every key named 'id' or 'element_id'. For compact YAML trees, remove_ids_and_bounds_from_compact_yaml at line 40 applies two regexes: ' #[\w\-]+' to strip identifier suffixes and 'bounds: \[[^\]]+\],?\s*' to strip pixel rectangles. After both trees pass through that filter, only the structurally meaningful lines remain. A button that moved from x=412 to x=540 produces no diff. A button whose name changed from Submit to Save produces one '-' line and one '+' line. That distinction is the entire point.

On the result of every action. In the TypeScript SDK (docs/TERMINATOR_JS_API.md lines 470 to 500), ClickResult and ActionResult both expose an optional uiDiff field of shape { diff: string, treeBefore?: string, treeAfter?: string, hasChanges: boolean }. You opt in per call: element.click({ uiDiff: true }). The MCP agent surfaces the same shape on click_element, type_into_element, press_key, and a handful of other tools, so an AI coding assistant testing a desktop app sees the diff in its tool response and can decide whether the change matched intent. Without the flag the action runs without any tree capture and the field is undefined.

Unified-style line diff. Lines added in the after-tree are prefixed with '+ ', lines removed are prefixed with '- ', and equal lines are filtered out (ui_tree_diff.rs lines 86 to 96). The output is what Rust's similar crate produces from TextDiff::from_lines, which mirrors the behavior of Python's difflib.ndiff. That format matters because it is git-native: you can commit a baseline diff to your repo, diff again on the next run, and assert string equality. No image diffing library, no antialiasing tolerance, no font hinting workaround. The whole snapshot is a few hundred bytes of text per action and any reviewer can read it.

1500 milliseconds by default, configurable per call. The timing lives at lib.rs line 1814 inside execute_with_ui_diff: tokio::time::sleep(Duration::from_millis(settle_ms)) runs after the action and before the after-snapshot. UiDiffOptions exposes settle_delay_ms so a fast modal can cut it to 200ms and a slow data-bound view can stretch it to 4000ms. Tree capture itself uses TreeBuildConfig with timeout_per_operation_ms=100 and yield_every_n_elements=25 to avoid stalling the accessibility thread, which matters because UIAutomation on Windows is notoriously deadlock-prone if you query it while it is still mutating.

Then you fall back to the same primitives the diff is built on. The same accessibility-tree dump that feeds the diff is also exposed through get_window_tree, so a test that suspects a flaky snapshot can capture the raw tree before and after manually, run the same regex preprocessor, and inspect any line. Because Terminator goes through the OS accessibility APIs (UIAutomation on Windows, AX on macOS) and not through pixel scraping, the tree reflects what assistive tech sees, not what a screenshot library guesses. You still get false positives when the app emits transient roles during animation, which is what the settle delay is for.

It replaces them when your test runner is an AI coding assistant or a TypeScript script and the app is anything reachable through accessibility. Commercial desktop test tools were built for a recorder/playback workflow and a single-OS stack. Terminator is a developer framework, MIT licensed, with a Playwright-shaped API, an MCP server, and the diff-as-primitive design described above. If you have an existing TestComplete suite that drives a single Windows app and a team trained on its IDE, switching is a project. If you are starting a regression suite from scratch in 2026, or pointing Claude Code at a desktop app, Terminator is closer to the shape your test code actually wants.

Three ways. For TypeScript: npm install @mediar-ai/terminator, then construct a Desktop, locate elements with selectors that look like role:Button|name:Submit, and call .click({ uiDiff: true }) to read result.uiDiff. For Python: pip install terminator, identical shape. For an AI coding assistant: claude mcp add terminator 'npx -y terminator-mcp-agent@latest', and the assistant gets click_element / type_into_element / press_key tools that all carry the uiDiff flag. The Rust crate is published as terminator on crates.io if you want to embed the engine directly in a Rust test harness.