fix: say where a run reports to W&B, loudly when nothing pins it

vero/src/vero/harbor/cli.py

@@ -678,6 +678,50 @@ def _probe_model(base_url: str, api_key: str, model: str) -> tuple[int | None, s
     return last
 
 
+def _report_wandb_destination(config) -> None:
+    """Name where results will land, before the run starts spending.
+
+    `wandb.entity` is optional, and omitting it is silent: the client falls back
+    to the launching user's personal default, so a grid launched by more than one
+    person scatters across their accounts and a cross-cell comparison has to be
+    reassembled by hand. Worse, the omission looks like a missing run rather than
+    a misplaced one, and "I cannot find it in W&B" is then diagnosed as a failed
+    run. That happened on 2026-08-01: an atlas cell reported into a personal
+    entity for nearly three hours while being read as having produced nothing.
+
+    Exporting WANDB_ENTITY does not fix it and is the natural thing to try. The
+    sidecar owns the run and receives only WANDB_API_KEY and WANDB_BASE_URL
+    through its passthrough list, so the variable never reaches the process that
+    calls `wandb.init`. The build config's `entity:` is the only thing that
+    decides the destination.
+
+    Printed on every run, not only when unset: knowing where a cell reports is
+    worth one line, and a warning that appears only in the bad case is one the
+    reader has no baseline for.
+    """
+
+    # getattr, not attribute access: this is a diagnostic, and one that raises
+    # would block the run it exists to describe. Reporting nothing is the correct
+    # degradation for a config shape that does not carry W&B settings at all.
+    wandb = getattr(config, "wandb", None)
+    if wandb is None:
+        return
+    name = getattr(wandb, "name", None) or "<default>"
+    project = getattr(wandb, "project", None) or "<default>"
+    if getattr(wandb, "entity", None):
+        click.echo(f"W&B: {wandb.entity}/{project} run={name}")
+        return
+    click.echo(
+        f"W&B: <your personal default>/{project} run={name}\n"
+        "  WARNING: wandb.entity is unset, so this run reports to the launching\n"
+        "  user's personal entity, not a shared org. Exporting WANDB_ENTITY will\n"
+        "  NOT change this: the sidecar passes through only WANDB_API_KEY and\n"
+        "  WANDB_BASE_URL. Set `entity:` under `wandb:` in the build config to\n"
+        "  pin it.",
+        err=True,
+    )
+
+
 def _preflight_models(config) -> None:
     """Refuse to launch when a configured model is not deployed upstream.
 
@@ -784,6 +828,7 @@ def run_command(config_path, agent, model, environment, params, env_file, extra)
         resolved.setdefault("optimizer_model", model)
     config = load_harbor_build_config(config_path, params=resolved)
     _preflight_models(config)
+    _report_wandb_destination(config)
     with tempfile.TemporaryDirectory(prefix="vero-harbor-") as temporary:
         task = compile_harbor_task(
             config,

vero/tests/test_v05_wandb_destination.py

@@ -0,0 +1,137 @@
+"""The launch-time report naming where a run's W&B results will land.
+
+`wandb.entity` is optional and omitting it is silent, so a cell reports into the
+launching user's personal entity while looking, to anyone checking the shared
+org, like a run that never started. These pin that the destination is always
+stated and that the unset case says so loudly.
+"""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import click
+import pytest
+
+from vero.harbor import cli as harbor_cli
+
+
+def _config(entity=None, project="harness-engineering-bench", name="cell-1"):
+    return SimpleNamespace(
+        wandb=SimpleNamespace(entity=entity, project=project, name=name)
+    )
+
+
+def test_pinned_entity_is_reported(capsys) -> None:
+    harbor_cli._report_wandb_destination(_config(entity="egp"))
+    out = capsys.readouterr()
+    assert "W&B: egp/harness-engineering-bench run=cell-1" in out.out
+    assert "WARNING" not in out.out + out.err
+
+
+def test_unset_entity_warns_and_names_the_fallback(capsys) -> None:
+    harbor_cli._report_wandb_destination(_config(entity=None))
+    err = capsys.readouterr().err
+    assert "personal default" in err
+    assert "WARNING" in err
+    # The natural workaround has to be named, or the reader tries it and the run
+    # still scatters: the sidecar owns wandb.init and never sees the variable.
+    assert "WANDB_ENTITY will" in err and "NOT change this" in err
+    assert "entity:" in err
+
+
+def test_empty_string_entity_is_treated_as_unset(capsys) -> None:
+    """`entity: ""` resolves from an unset ${param} and must not read as pinned."""
+
+    harbor_cli._report_wandb_destination(_config(entity=""))
+    assert "WARNING" in capsys.readouterr().err
+
+
+def test_silent_when_wandb_is_not_configured(capsys) -> None:
+    harbor_cli._report_wandb_destination(SimpleNamespace(wandb=None))
+    out = capsys.readouterr()
+    assert out.out == "" and out.err == ""
+
+
+def test_missing_run_name_does_not_crash_the_launch(capsys) -> None:
+    """A diagnostic that raises would block the run it is meant to describe."""
+
+    harbor_cli._report_wandb_destination(_config(entity="egp", name=None))
+    assert "run=<default>" in capsys.readouterr().out
+
+
+def test_harbor_run_warns_before_it_compiles_the_task(tmp_path, monkeypatch) -> None:
+    """Ordering is the point: the line has to land before anything is spent.
+
+    Proven by making compilation fail. If the report already ran, its warning is
+    in the output even though the command died at compile, which is exactly the
+    guarantee wanted: a misdirected destination is visible before a sandbox
+    exists and before a single case is scored.
+    """
+
+    from click.testing import CliRunner
+
+    from vero.cli import main
+    from vero.harbor import build as harbor_build
+
+    config_path = tmp_path / "build.yaml"
+    config_path.write_text("name: org/task\n", encoding="utf-8")
+
+    class _Config:
+        harbor_requirement = "harbor[modal]==0.20.0"
+        agent_env: dict[str, str] = {}
+        optimizer_harbor_args: list[str] = []
+        extra_harbor_args: list[str] = []
+        name = "vero/stub-benchmark"
+        wandb = SimpleNamespace(entity=None, project="heb", name="cell-1")
+
+    def _explode(config, output):
+        raise RuntimeError("compilation reached")
+
+    monkeypatch.setattr(
+        harbor_build, "load_harbor_build_config", lambda *a, **k: _Config()
+    )
+    monkeypatch.setattr(harbor_build, "compile_harbor_task", _explode)
+    monkeypatch.setattr(harbor_cli.shutil, "which", lambda name: "/usr/bin/uvx")
+    monkeypatch.setattr(harbor_cli, "_preflight_models", lambda config: None)
+
+    result = CliRunner().invoke(
+        main,
+        [
+            "harbor",
+            "run",
+            "--config",
+            str(config_path),
+            "--agent",
+            "codex",
+            "--model",
+            "gpt-5.3-codex",
+            "--yes",
+        ],
+    )
+
+    # click >=8.2 splits the streams; the warning is deliberately on stderr.
+    emitted = result.output + (result.stderr or "")
+    assert "compilation reached" in str(result.exception)
+    assert "WARNING" in emitted
+    assert "personal" in emitted
+
+
+def test_config_argument_is_required() -> None:
+    """A diagnostic called with no config should fail loudly, not pass silently."""
+
+    with pytest.raises(TypeError):
+        harbor_cli._report_wandb_destination()
+
+
+def test_config_without_a_wandb_attribute_is_a_no_op(capsys) -> None:
+    """Degrade to silence, never to an exception.
+
+    Several call sites and tests pass config shapes that carry no W&B settings at
+    all. A diagnostic that raised on those would block the launch it exists to
+    describe, which is a strictly worse failure than the one it prevents.
+    """
+
+    harbor_cli._report_wandb_destination(SimpleNamespace(name="org/task"))
+    out = capsys.readouterr()
+    assert out.out == "" and out.err == ""