feat(eval): log why a benchmark cell failed

The sweep printed error_kind=plan-evidence-invalid and nothing else, so
the reason a cell failed stayed in results.jsonl — an artifact uploaded
after the run, not something a watcher can read while it is still going.

Print the cell's error_detail next to its result line, redacted through
the same credential list as the artifact and bounded, since a
session-error detail carries stdout/stderr tails.

Co-authored-by: Cursor <cursoragent@cursor.com>

eval/tests/test_session_progress.py

@@ -184,3 +184,37 @@ def test_progress_sanitizes_a_hostile_tool_name() -> None:
 
     assert "FAKE-LOG-LINE" not in stream.getvalue()
     assert len(_drain_lines(stream)) == 1
+
+
+def test_cell_failure_detail_line_explains_why_a_cell_failed() -> None:
+    from workflow_bench.runner import cell_failure_detail_line
+
+    assert cell_failure_detail_line("t", "workflow", 0, {"error_kind": None}) is None
+    assert cell_failure_detail_line("t", "workflow", 0, {"error_kind": "x"}) is None
+
+    line = cell_failure_detail_line(
+        "trivial-version-alias",
+        "candidate_workflow",
+        1,
+        {
+            "error_kind": "plan-evidence-invalid",
+            "error_detail": "unauthorized workspace path\ntoken=sk-secret-value",
+        },
+        ("sk-secret-value",),
+    )
+    assert line is not None
+    assert line.startswith("[trivial-version-alias][candidate_workflow][run 1] detail: ")
+    assert "unauthorized workspace path" in line
+    assert "sk-secret-value" not in line
+    assert "\n" not in line
+
+
+def test_cell_failure_detail_line_bounds_a_huge_detail() -> None:
+    from workflow_bench.runner import MAX_CELL_DETAIL_CHARS, cell_failure_detail_line
+
+    line = cell_failure_detail_line(
+        "t", "workflow", 0, {"error_kind": "session-error", "error_detail": {"stdout_tail": "y" * 50_000}}
+    )
+    assert line is not None
+    assert "truncated" in line
+    assert len(line) < MAX_CELL_DETAIL_CHARS + 200

eval/workflow_bench/runner.py

@@ -1089,6 +1089,36 @@ def cell_progress_line(task_id: str, arm: str, run_idx: int, record: dict[str, A
     )
 
 
+# A failing cell's error_kind names the category; the detail names the cause.
+# Bounded because a session-error detail carries stdout/stderr tails.
+MAX_CELL_DETAIL_CHARS = 1200
+
+
+def cell_failure_detail_line(
+    task_id: str,
+    arm: str,
+    run_idx: int,
+    record: Mapping[str, Any],
+    secrets: Sequence[str] = (),
+) -> str | None:
+    """The redacted reason a cell failed, or None when it succeeded.
+
+    Without this the log says only ``error_kind=plan-evidence-invalid`` and the
+    reason stays locked in results.jsonl, which is an uploaded artifact rather
+    than something a watcher can read while the sweep is still running.
+    """
+    if not record.get("error_kind"):
+        return None
+    detail = record.get("error_detail")
+    if detail in (None, "", {}, []):
+        return None
+    rendered = detail if isinstance(detail, str) else json.dumps(detail, default=str, sort_keys=True)
+    rendered = redact_text(rendered, secrets).replace("\n", " ⏎ ")
+    if len(rendered) > MAX_CELL_DETAIL_CHARS:
+        rendered = f"{rendered[:MAX_CELL_DETAIL_CHARS]}…[truncated {len(rendered) - MAX_CELL_DETAIL_CHARS} chars]"
+    return f"[{task_id}][{arm}][run {run_idx}] detail: {rendered}"
+
+
 def aggregate(records: list[dict[str, Any]]) -> dict[str, Any]:
     """Median metrics + resolve rate across repeated runs of one task+arm.
 
@@ -1608,6 +1638,9 @@ def _run_sweep(
                     # sink was not).
                     fh.write(redact_text(json.dumps(record), credential_secrets(args)) + "\n")
                 print(cell_progress_line(task["id"], arm, run_idx, record))
+                failure = cell_failure_detail_line(task["id"], arm, run_idx, record, credential_secrets(args))
+                if failure:
+                    print(failure)
 
             outage_streak, outage_tripped = sweep_task_cells(
                 cells,