Windows automation software that throws out unstable selector text before it writes the file

Because Windows UIA lets elements return text that looks like content but is actually a placeholder, a COM variant marker, or a rendered timestamp. A recorder that captures 'Button with Name unknown' or 'Pane with Name Dashboard - 3 minutes ago' will fail on replay when that same element reports 'BSTR()' or 'Dashboard - 5 minutes ago'. Most consumer automation software for Windows (Power Automate Desktop, UiPath, WinAutomation, AutoHotkey-backed recorders) treats whatever UIA returns as ground truth and serializes it into the workflow. The stability problem is a capture-time problem, and almost no product solves it at capture time.

Three things, all at crates/terminator-workflow-recorder/src/events.rs. First, the NULL_LIKE_VALUES HashSet at lines 8 through 36 holds 20 strings: 'null', 'nil', 'undefined', '(null)', '<null>', 'n/a', 'na', '', 'unknown', '<unknown>', '(unknown)', 'none', '<none>', '(none)', 'empty', '<empty>', '(empty)', 'bstr()', 'variant()', 'variant(empty)'. Second, contains_relative_time at lines 69 through 81 rejects any element whose name contains ' ago', 'just now', 'yesterday', 'today', 'last week', 'last month', or ends with ' min', ' mins', ' hr', or ' hrs'. Third, is_internal_element at lines 696 through 716 skips parents whose name contains 'legacy window', 'chrome_widgetwin', 'intermediate d3d window', or Panes whose title ends with ' - Google Chrome', ' - Mozilla Firefox', or ' - Microsoft Edge'.

They are render-layer implementation details that leak into the UIA tree but have no semantic meaning for locating UI elements. 'Chrome_WidgetWin_1' is Chromium's top-level native window class; 'Legacy Window' is Windows' compatibility shim around HWND-backed controls; 'Intermediate D3D Window' is the Direct3D compositor surface. If the recorder keeps them in the selector chain, the chain will include container roles that depend on GPU state, window-manager compositor behavior, and Chrome version. Filter them out and the selector chain collapses to the meaningful semantic parents, which is what the UIA-based locator actually wants to match on.

crates/terminator-workflow-recorder/src/events.rs. The imports and static block are at lines 1 through 36. The is_empty_string helper is at lines 39 through 65. contains_relative_time is at lines 69 through 81. Every serialized event field that carries user-facing text is marked with #[serde(skip_serializing_if = "is_empty_string")], starting at line 195 and appearing roughly 40 times through the file. The parent hierarchy walker that rejects internal elements is at build_parent_hierarchy, lines 685 through 744. The unit test battery that pins down every filtered string is test_empty_string_helper at line 1764.

Power Automate Desktop's 'selector repair' is a replay-time recovery feature. It waits until the selector fails to match, then opens the Selector Builder and asks the human to manually repair it inside the designer. By the time the repair dialog opens, the unattended run has already failed. Terminator's filter runs at capture time, before the workflow file is written, so the stored selector never contains the unstable text in the first place. Nothing to repair later because the volatile bits never made it into the YAML.

No. NULL_LIKE_VALUES is a precomputed LazyLock<HashSet<&'static str>> built once on first use (line 8). Lookups are O(1). is_empty_string short-circuits on empty strings and whitespace-only strings (lines 41-51) before it even considers allocating a lowercase copy. It also caps the lowercase allocation at strings of 20 characters or less (line 55), on the observation that null-like markers are always short. The relative-time check is a straight series of str::contains calls, no regex. On a reasonable workflow the whole filter costs under a millisecond per event.

It is dropped from the parent hierarchy, not the recording. build_parent_hierarchy at events.rs line 685 walks up from the target element; for each parent it checks has_name (non-empty after filter) and is_internal_element (not on the skip list). If a parent fails either check, it is omitted from the hierarchy and the walk continues up to the next parent. The target element itself is still captured; only the path to it is trimmed. Result: a compact selector chain that only references semantically meaningful ancestors, which is what build_chained_selector (line 750) then stitches together with the '>>' operator.

build_chained_selector at events.rs line 750. It iterates parent_hierarchy, skipping the first Window element because the 'process:' prefix already targets that window, then emits 'role:X && text:Y' for each named parent and joins with ' >> '. A chain like 'role:Pane && text:Toolbar >> role:Button && text:Export CSV' is the product of that walker plus the filter. Skip the filter and the same chain would read 'role:Pane && text:Chrome_WidgetWin_1 >> role:Pane && text:Intermediate D3D Window >> role:Pane && text:Dashboard - 3 hours ago >> role:Button && text:Export CSV', which is the kind of string that breaks one refresh later.

The set is a static LazyLock<HashSet<&'static str>>, so you cannot add to it at runtime, but you can fork the crate and add entries at the top of events.rs. The CONTRIBUTING.md in the repo welcomes PRs for new null-like strings that show up in specific application UI. If you find your team's internal app emits 'NO_VALUE' or '[empty]' in UIA tree nodes, open a PR that adds those literals to the set and the CI tests will verify the existing behavior keeps passing. NULL_LIKE_VALUES.contains is a single-line check, so adding entries is additive and safe.

The filter is scoped to the recorder crate (terminator-workflow-recorder) because that is where selector text gets serialized to disk. When you call the SDK directly, you pass selectors as strings you chose, so there is nothing to filter. The value of the filter shows up specifically for the record-and-replay workflow: a human clicks through a task in a real Windows session, the recorder captures UIA events, and the stored YAML needs to replay reliably on the next run. That is where stripping unstable text at capture time matters.