Upload 130 files

tests/test_v175_anonymous_eval_archive.py

@@ -22,25 +22,25 @@ def eval_settings() -> Settings:
     )
 
 
-def test_eval_bucket_env_is_passed_when_enabled(monkeypatch):
+def test_eval_record_env_is_passed_without_operator_bucket_mount(monkeypatch):
     monkeypatch.setattr(jobs, "settings", eval_settings())
     env = jobs._base_env(run_id="universal-20260606-000000-abcd1234", username="alice", bucket_source="alice/space-factory-runs", worker_script_path="/output/runs/x/_worker.py")
-    assert env["ASF_EVAL_ENABLED"] == "true"
-    assert env["ASF_EVAL_BUCKET_SOURCE"] == "huggingface/agentic-space-factory-evals"
-    assert env["ASF_EVAL_BUCKET_PATH"] == "evals"
-    assert env["ASF_EVAL_BUCKET_MOUNT"] == "/evals"
+    assert env["ASF_EVAL_RECORD_ENABLED"] == "true"
+    assert env["ASF_EVAL_PUBLISH_MODE"] == "backend"
     assert env["ASF_EVAL_SALT"] == "test-salt"
+    assert "ASF_EVAL_BUCKET_SOURCE" not in env
+    assert "ASF_EVAL_BUCKET_MOUNT" not in env
 
 
-def test_eval_bucket_is_mounted_as_separate_job_volume(monkeypatch):
+def test_eval_bucket_is_not_mounted_as_user_job_volume(monkeypatch):
     monkeypatch.setattr(jobs, "settings", eval_settings())
     volumes = jobs._job_volumes("alice/space-factory-runs")
     rendered = [vars(v) if hasattr(v, "__dict__") else v for v in volumes]
     text = repr(rendered)
     assert "alice/space-factory-runs" in text
-    assert "huggingface/agentic-space-factory-evals" in text
-    assert "evals" in text
-    assert "/evals" in text
+    assert "huggingface/agentic-space-factory-evals" not in text
+    assert "/evals" not in text
+    assert len(volumes) == 1
 
 
 def test_workers_publish_anonymized_eval_records_without_raw_code_or_prompts():

tests/test_v178_run_storage_eval_status_polish.py

@@ -0,0 +1,23 @@
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def test_run_storage_eval_disabled_copy_is_compact():
+    js = (ROOT / 'web/static/app.js').read_text()
+    assert "'Eval archive off · set env vars'" in js
+    assert 'Anonymous project eval archive disabled for this instance.' not in js
+
+
+def test_run_storage_eval_enabled_copy_is_compact_with_tooltip():
+    js = (ROOT / 'web/static/app.js').read_text()
+    assert 'Eval archive on ·' in js
+    assert 'ASF backend publishes anonymized eval records there after user Jobs write local run evals.' in js
+    assert 'Anonymous project eval archive enabled:' in js
+
+
+def test_eval_feedback_is_a_small_pill_not_long_wrapping_copy():
+    css = (ROOT / 'web/static/app.css').read_text()
+    assert '.eval-feedback{display:inline-flex' in css
+    assert 'border-radius:999px' in css
+    assert '.eval-feedback.off' in css

tests/test_v179_eval_activation_prescan_polish.py

@@ -0,0 +1,24 @@
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def test_eval_archive_off_message_is_actionable():
+    js = (ROOT / "web/static/app.js").read_text()
+    assert "Eval archive off · set env vars" in js
+    assert "ASF_EVAL_ENABLED=true" in js
+    assert "ASF_EVAL_BUCKET_SOURCE=<namespace>/<bucket>" in js
+
+
+def test_prescan_no_heavy_top_accent_bar():
+    css = (ROOT / "web/static/app.css").read_text()
+    assert ".model-prescan-card:before{display:none!important;content:none!important}" in css
+    assert "v179: calmer pre-scan surface" in css
+
+
+def test_readme_documents_eval_activation_and_keeps_metadata_first():
+    readme = (ROOT / "README.md").read_text()
+    assert readme.startswith("---\ntitle: Agentic Space Factory")
+    assert "## Anonymous eval archive activation" in readme
+    assert "ASF_EVAL_BUCKET_MOUNT=/evals" in readme
+    assert "Bucket path:   evals" in readme

tests/test_v181_backend_eval_archive_publish.py

@@ -0,0 +1,77 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from types import SimpleNamespace
+
+from src import eval_archive, eval_config, jobs
+
+
+def test_user_jobs_do_not_mount_private_operator_eval_bucket(monkeypatch):
+    monkeypatch.setattr(jobs, "settings", SimpleNamespace(
+        bucket_mount="/output",
+        bucket_runs_prefix="runs",
+        eval_enabled=True,
+        eval_bucket_source="fffiloni/agentic-space-factory-evals",
+        eval_bucket_path="evals",
+        eval_bucket_mount="/evals",
+        eval_salt="salt",
+        eval_include_redacted_tails=False,
+    ))
+    env = jobs._base_env(run_id="universal-20260606-000000-abcd1234", username="someone", bucket_source="someone/space-factory-runs", worker_script_path="/output/runs/x/_worker.py")
+    assert env["ASF_EVAL_RECORD_ENABLED"] == "true"
+    assert env["ASF_EVAL_PUBLISH_MODE"] == "backend"
+    assert "ASF_EVAL_BUCKET_SOURCE" not in env
+    assert "ASF_EVAL_BUCKET_MOUNT" not in env
+    volumes = jobs._job_volumes("someone/space-factory-runs")
+    assert len(volumes) == 1
+    assert "fffiloni/agentic-space-factory-evals" not in repr([vars(v) for v in volumes])
+
+
+def test_backend_publishes_user_bucket_eval_record_to_mounted_operator_archive(tmp_path, monkeypatch):
+    user_bucket_files: dict[str, str] = {}
+    root = "hf://buckets/alice/space-factory-runs/runs/universal-20260606-000000-abcd1234"
+    user_bucket_files[f"{root}/eval_record.json"] = json.dumps({
+        "schema_version": "1.0",
+        "phase": "final",
+        "anonymous_run_id": "run_abc123",
+        "finished_at": "2026-06-06T16:00:00+00:00",
+        "outcome": {"status": "success"},
+        "privacy": {"tokens_stored": False, "raw_prompts_stored": False, "generated_code_stored": False},
+        "timeline": [{"stage": "start", "status": "success"}],
+    })
+    user_bucket_files[f"{root}/events_compact.jsonl"] = '{"stage":"start","status":"success"}\n'
+
+    def fake_read_json(path, token=None):
+        text = user_bucket_files.get(path)
+        return json.loads(text) if text else None
+
+    def fake_read_text(path, token=None):
+        return user_bucket_files.get(path)
+
+    def fake_write_json(path, payload, token=None):
+        user_bucket_files[path] = json.dumps(payload)
+
+    monkeypatch.setattr(eval_archive, "read_json", fake_read_json)
+    monkeypatch.setattr(eval_archive, "read_text", fake_read_text)
+    monkeypatch.setattr(eval_archive, "write_json", fake_write_json)
+    monkeypatch.setattr(eval_archive, "effective_eval_config", lambda: {
+        "enabled": True,
+        "bucket_source": "fffiloni/agentic-space-factory-evals",
+        "bucket_path": "evals",
+        "job_mount_path": str(tmp_path),
+    })
+
+    result = eval_archive.maybe_publish_eval_record(
+        "universal-20260606-000000-abcd1234",
+        bucket_source="alice/space-factory-runs",
+        token="fake",
+        state={"status": "success"},
+    )
+    assert result["published"] is True
+    dest = tmp_path / "2026" / "06" / "06" / "run_abc123" / "eval_record.json"
+    assert dest.exists()
+    assert json.loads(dest.read_text())["anonymous_run_id"] == "run_abc123"
+    status_path = f"{root}/eval_publish_status.json"
+    assert status_path in user_bucket_files
+    assert json.loads(user_bucket_files[status_path])["publish_mode"] == "backend"