Testing

cargo test

85 tests on Linux and 72 on Windows, with no test-only dependencies. 66 are shared. The rest are platform-specific terminal, path-lookup and key-decoding tests.

The scanner

Heaviest coverage: order preservation, object-form Deno tasks, escapes and surrogate pairs, structural bytes inside strings, comments, trailing commas, and syntax errors carrying an offset.

The picker loop

Driven by a scripted console. Tests feed it a fixed list of keypresses and read back the frames it drew. That covers navigation and clamping, the scrolling window, filtering, backspace, and redraw bookkeeping: every frame must rewind exactly the line count of the one before, or the picker would eat scrollback.

The most valuable test: the cursor indexes the filtered view, so committing must map back to the original index. Getting that wrong would run the wrong script, silently.

Key decoding

A pure function of what arrives from the OS, so both mappings are tested without a terminal. Cases include every spelling of Home and End, arrows, modified arrows, page keys, a bare Esc, truncated sequences and multi-byte UTF-8. An unrecognised sequence must be consumed through its final byte, or its tail would land in the filter.

Running real tools

run has integration tests that drive real npm and deno, covering PATH resolution, exit-code propagation and the working directory. They skip when the tool is missing, so the suite passes without a JavaScript toolchain.

Not tested

The Win32 layer itself (ReadConsoleInputW, SetConsoleMode, the screen-buffer query) needs a real console.