Error logs

.changeset/refusals-to-stderr.md

@@ -0,0 +1,9 @@
+---
+'@platforma-open/milaboratories.sort-seq-analysis.software': patch
+---
+
+Write refusals to stderr as well as stdout, so the platform's error dialog shows them
+
+The k8s runner fills a failed command's "Latest output" from the stderr file alone. A
+refusal printed only to stdout therefore reached the user as a blank "exited with code 1",
+with the actual reason sitting in the run's log panel.

software/src/main.py

@@ -14,8 +14,14 @@
 would multiply startup, leave the caller merging N manifests, and make the one-condition
 case a loop's degenerate iteration.
 
-**Diagnostics go to stdout**, including failures: the workflow layer does not capture
-stderr, so anything written there would be lost.
+**Diagnostics go to both streams, and that is deliberate.** They serve two different
+readers. Stdout is the run's audit trail: the block saves it as a log stream and the UI
+shows it, so the successful-run summary belongs there. Stderr is what the platform reads
+back when the command fails — the k8s runner fills the error dialog's "Latest output"
+from the stderr file alone. A refusal printed only to stdout therefore reaches the user
+as a blank "exited with code 1", with the reason sitting in a log panel nobody was told
+to open. So a refusal is written to both: stdout because it is part of the run's record,
+stderr because it is the failure the platform is about to report.
 """
 
 from __future__ import annotations
@@ -59,7 +65,12 @@ def main(argv: list[str] | None = None) -> int:
     except Refusal as refusal:
         # A data-value violation. Exit non-zero naming the offending values, having written
         # no file — nothing partial is produced.
-        print(f"REFUSED: {refusal}")
+        #
+        # Both streams, per the module docstring: stdout keeps the run's record complete,
+        # stderr is the only stream the platform quotes back in the error it shows.
+        message = f"REFUSED: {refusal}"
+        print(message)
+        print(message, file=sys.stderr)
         return 1
 
     _report(manifest)

software/tests/integration/test_cli.py

@@ -404,8 +404,12 @@ def test_over_summing_fractions_exits_non_zero_and_writes_nothing(tmp_path, caps
     assert code == 1
     assert manifest is None
     assert not out_dir.exists()
-    # Diagnostics go to stdout — the workflow layer does not capture stderr.
-    assert "REFUSED" in capsys.readouterr().out
+    # Both streams, and both are pinned. Stdout is the run's record; stderr is the only
+    # stream the platform quotes back in the error it shows the user, so a refusal that
+    # reaches stdout alone is reported as a blank non-zero exit.
+    captured = capsys.readouterr()
+    assert "REFUSED" in captured.out
+    assert "REFUSED" in captured.err
 
 
 def test_replicate_samples_exit_non_zero_and_write_nothing(tmp_path, capsys):
@@ -429,7 +433,9 @@ def test_replicate_samples_exit_non_zero_and_write_nothing(tmp_path, capsys):
     assert code == 1
     assert manifest is None
     assert not out_dir.exists()
-    assert "replicate" in capsys.readouterr().out
+    captured = capsys.readouterr()
+    assert "replicate" in captured.out
+    assert "replicate" in captured.err
 
 
 def test_sort_fraction_column_missing_from_the_reads_table_fails(tmp_path, capsys):
@@ -439,7 +445,9 @@ def test_sort_fraction_column_missing_from_the_reads_table_fails(tmp_path, capsy
 
     assert code == 1
     assert not out_dir.exists()
-    assert "absent" in capsys.readouterr().out
+    captured = capsys.readouterr()
+    assert "absent" in captured.out
+    assert "absent" in captured.err
 
 
 # ---------------------------------------------------------------------------