How it works¶
One process per download. Starting a row spawns the tool with its stdout piped, and a reader thread turns each line into a progress update, an output path, or a notice, which the event loop applies to the row by its stable id. Rows can be reordered or deleted while their process runs. Updates are looked up by id, so nothing is misrouted.
Progress is parsed, not guessed. Each tool has a parser for its own format:
aria2c's bracketed summary, yt-dlp's [download] lines, wget's dots. Torrent
rows also carry upload rate, session total, peers and seeders, and a row is
drawn as a torrent exactly when the tool reports a seeder count.
Both tools redraw with \r, so output is read by chunks and split on both
carriage returns and newlines. A line-based reader would sit there waiting for a
newline that never comes.
Processes are tracked by pid, not by a handle, because waiting on a child holds a lock that killing through the same lock would deadlock, which is how a quit hangs and orphans a download. The pid is written to the state file, so if the app dies without cleaning up, the next launch can kill what it left behind. It checks the process name first, so a recycled pid cannot take an unrelated process down with it.
The event loop polls for input at 200ms and draws every frame. Twice a second it samples aggregate speed for the sparkline and charges bandwidth quotas. Every fifteen seconds it runs the schedule pass: window and quota pauses, periodic re-syncs, and the actions a drained queue triggers.
Filling slots happens after anything that could free one: a finished download, a cancel, a pause, a resumed queue, a reorder. Each queue is filled independently up to its own limit, and a paused queue is skipped however many slots are free.
Exit codes are translated. aria2c exiting 13 means "the file already
exists", not "exit 13". wget's 8 means the server refused some of the crawl,
which is reported but does not fail the mirror.
Adding a backend¶
Implement Backend in src/models/, giving it a name, which urls it accepts,
the command to run, and a progress-line parser, then add it to backends() in
src/models/mod.rs. Spawning, output reading and progress reporting are shared.
Four optional hooks: reason turns an exit code into something a person can
read, tolerates marks an exit code that is not really a failure, notice
turns a line into a status message, and config_flag / credentials say how the
tool reads a login from a file. Add an OptSpec list in src/models/option.rs
and the settings panel gets a form for it.