scripts-execution

Script execution — console and temp output

Applies when running or editing files under scripts/.

Visible console

• Stream train/decode/long jobs to a visible Cursor terminal; log-only redirect as the sole sink is not OK.
• Console + logs/<topic>.log (tee) is fine — terminal remains primary.
• Never launch a long job with > logs/... 2>&1, *> logs/..., or any equivalent that hides live output from the terminal.

From PowerShell, use:

& venv\Scripts\python.exe scripts/<script>.py <args> 2>&1 | Tee-Object -FilePath logs/<topic>_<run>.log
$code = $LASTEXITCODE
if ($code -ne 0) { exit $code }

PowerShell anti-pattern (hides the console): never combine a file redirect with Tee like 1> logs/out.json 2>&1 | Tee-Object ... or > logs/out.log 2>&1 | Tee-Object .... Stdout is bound to the file before the pipeline, so Tee (and the Cursor terminal) receive nothing — progress vanishes into the file only. That violates visible-console even if a log path looks correct.

For scripts that split JSON on stdout / progress on stderr (e.g. eval_ar_onset_offline.py):

& venv\Scripts\python.exe scripts/eval_ar_onset_offline.py <args> `
  2>&1 | Tee-Object -FilePath logs/<topic>_<run>.log

Keep one tee stream to the terminal + log; extract JSON from the log afterward if needed. Do not 1>-redirect stdout away from Tee to “save JSON cleanly” unless you also mirror progress to the console some other way.

If a tool backgrounds the command after its foreground timeout, the same tee pipeline must remain attached so the visible terminal continues receiving live output.

TensorBoard with training

When launching a training job that writes Keras TensorBoard callbacks (callback_root_dir / callbacks/.../logs):

1. Start TensorBoard in a separate terminal before or with the train command — do not wait for the user to ask.
2. Point --logdir at the stage log tree, not one run folder.
3. Tell the user the URL (http://localhost:6006/ by default) when training starts.