fix(onboarding): pass the OpenRouter key to qsv describegpt via QSV_LLM_APIKEY

src/data_concierge/data_layer/onboard_index.py

@@ -31,7 +31,8 @@
 _PROJECT_ROOT = Path(__file__).resolve().parent.parent.parent.parent
 
 # qsv describegpt records its full invocation in the AI-generated descriptions,
-# and that command carries the OpenRouter API key on --api-key. Those
+# and onboarding runs before the key moved to QSV_LLM_APIKEY (see
+# qsv_profiling.run_qsv_describegpt) put the OpenRouter API key on --api-key. Those
 # descriptions are served over the Fair Store API, rendered in the data
 # dictionary UI, and fed into the agent's prompt — so the key would leak three
 # ways. Redact provider secrets and the key flag before any description is used.

src/data_concierge/data_layer/qsv_profiling.py

@@ -30,6 +30,7 @@
 import asyncio
 import csv
 import json
+import os
 from datetime import UTC, datetime
 from io import StringIO
 from pathlib import Path
@@ -65,20 +66,25 @@ async def run_qsv_describegpt(
     api_key: str,
     output_path: Path,
 ) -> dict | None:
-    """Run qsv describegpt and return parsed JSON output."""
+    """Run qsv describegpt and return parsed JSON output.
+
+    The API key goes to qsv in ``QSV_LLM_APIKEY``, not ``--api-key``: describegpt
+    copies its command line into the descriptions it generates, and argv is also
+    visible to other local users in ``ps``.
+    """
     cmd = [
         "qsv", "describegpt", str(csv_path),
         "--all",
         "--format", "json",
         "--base-url", "https://openrouter.ai/api/v1",
         "--model", "google/gemini-2.5-flash-lite",
-        "--api-key", api_key,
     ]
 
     proc = await asyncio.create_subprocess_exec(
         *cmd,
         stdout=asyncio.subprocess.PIPE,
         stderr=asyncio.subprocess.PIPE,
+        env={**os.environ, "QSV_LLM_APIKEY": api_key},
     )
     stdout, stderr = await proc.communicate()
 
@@ -399,8 +405,9 @@ def sync_to_storage(base_dir: Path, site_id: str, prefix: str = "ckan_onboard")
     backend when the admin panel starts a mirror on Cloud Run.
 
     JSON is scrubbed of provider keys on the way: qsv describegpt records its
-    own command line, ``--api-key`` included, in ``qsv_dict.json``,
-    ``meta.json`` and ``index.json``.
+    own command line in ``qsv_dict.json``, ``meta.json`` and ``index.json``,
+    and onboarding runs before the key moved to ``QSV_LLM_APIKEY`` passed it
+    on ``--api-key``.
     """
     synced = 0
     manifest = sync_manifest(base_dir, prefix)

tests/unit/test_qsv_profiling.py

@@ -0,0 +1,42 @@
+"""Tests for data_layer.qsv_profiling."""
+
+from __future__ import annotations
+
+import json
+import os
+import sys
+
+import pytest
+
+from data_concierge.data_layer.qsv_profiling import run_qsv_describegpt
+
+KEY = "sk-or-v1-0123456789abcdef"
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason="fake qsv is a POSIX shell script")
+async def test_describegpt_gets_the_api_key_from_the_environment_not_argv(tmp_path, monkeypatch):
+    # describegpt copies its argv into the descriptions it writes, so the key must not be in it.
+    record = tmp_path / "record.json"
+    fake = tmp_path / "bin" / "qsv"
+    fake.parent.mkdir()
+    fake.write_text(
+        f"#!{sys.executable}\n"
+        "import json, os, sys\n"
+        f"json.dump({{'argv': sys.argv[1:], 'key': os.environ.get('QSV_LLM_APIKEY')}},"
+        f" open({str(record)!r}, 'w'))\n"
+        "print(json.dumps({'description': 'ok'}))\n"
+    )
+    fake.chmod(0o755)
+    monkeypatch.setenv("PATH", f"{fake.parent}:{os.environ['PATH']}")
+    monkeypatch.delenv("QSV_LLM_APIKEY", raising=False)
+
+    out = tmp_path / "qsv_dict.json"
+    data = await run_qsv_describegpt(tmp_path / "data.csv", KEY, out)
+
+    assert data == {"description": "ok"}
+    assert json.loads(out.read_text()) == {"description": "ok"}
+    seen = json.loads(record.read_text())
+    assert seen["key"] == KEY
+    assert KEY not in " ".join(seen["argv"])
+    assert "--api-key" not in seen["argv"]
+    assert seen["argv"][:2] == ["describegpt", str(tmp_path / "data.csv")]