Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

fix(egress): close inference-path fail-open (LLM/embeddings PEPs use the single primary source) (#4752)

LearningCircuitf6256f5cbd4f
diff --git a/src/local_deep_research/config/llm_config.py b/src/local_deep_research/config/llm_config.py
index f8adb88e02..a5ec1de893 100644
--- a/src/local_deep_research/config/llm_config.py
+++ b/src/local_deep_research/config/llm_config.py
@@ -15,7 +15,6 @@ from ..utilities.search_utilities import remove_think_tags
from ..llm.providers import discover_providers # noqa: F401
from ..llm.providers.base import normalize_provider
from .thread_settings import get_setting_from_snapshot
-from ..constants import DEFAULT_SEARCH_TOOL
def get_selected_llm_provider(settings_snapshot=None):
@@ -126,6 +125,7 @@ def get_llm(
_is_user_registered_llm,
context_from_snapshot,
evaluate_llm_endpoint,
+ resolve_run_primary_engine,
)
if settings_snapshot is None:
@@ -145,28 +145,26 @@ def get_llm(
target=provider,
)
else:
- primary_engine_raw = settings_snapshot.get(
- "search.tool", DEFAULT_SEARCH_TOOL
- )
- if (
- isinstance(primary_engine_raw, dict)
- and "value" in primary_engine_raw
- ):
- primary_engine_raw = primary_engine_raw["value"]
-
try:
- ctx = context_from_snapshot(
- settings_snapshot, primary_engine_raw or DEFAULT_SEARCH_TOOL
- )
+ # Derive the run's primary the SAME way the search-engine
+ # factory does (single source of truth), instead of the old
+ # ``search.tool`` + searxng fallback. That fallback was a
+ # fail-OPEN: a missing/blank primary defaulted to searxng ->
+ # ADAPTIVE -> PUBLIC_ONLY -> require_local_llm stayed False ->
+ # the endpoint check below was skipped, admitting a CLOUD LLM
+ # for a run whose actual posture was private. resolve_run_
+ # primary_engine raises on a missing/invalid primary, which we
+ # treat as a hard stop here.
+ primary_engine = resolve_run_primary_engine(settings_snapshot)
+ ctx = context_from_snapshot(settings_snapshot, primary_engine)
except ValueError as exc:
- # An invalid policy config must fail closed here — this is
- # essential; previously we silently fell back to ctx=None
- # which skipped the LLM endpoint check entirely, opening a
- # cloud-LLM bypass under the very configuration the user
- # asked to be strict about. Raise PolicyDeniedError instead
- # so the misconfiguration surfaces as a hard stop.
+ # No configured primary, or an invalid policy config. Fail
+ # closed: previously a missing primary silently fell back to
+ # searxng/PUBLIC_ONLY and skipped the LLM endpoint check
+ # entirely, opening a cloud-LLM bypass under the very
+ # configuration the user asked to be strict about.
logger.bind(policy_audit=True).warning(
- "invalid egress policy configuration; refusing LLM",
+ "no/invalid egress policy primary; refusing LLM",
provider=provider,
reason=str(exc),
)
diff --git a/src/local_deep_research/embeddings/embeddings_config.py b/src/local_deep_research/embeddings/embeddings_config.py
index cf4ebee4d3..540555eeb2 100644
--- a/src/local_deep_research/embeddings/embeddings_config.py
+++ b/src/local_deep_research/embeddings/embeddings_config.py
@@ -12,7 +12,6 @@ from loguru import logger
from ..config.thread_settings import get_setting_from_snapshot
from .providers.base import BaseEmbeddingProvider
-from ..constants import DEFAULT_SEARCH_TOOL
# Internal: list of provider strings accepted by get_embeddings().
# Not re-exported from embeddings/__init__.py — call sites should not import it.
@@ -202,17 +201,16 @@ def get_embeddings(
PolicyDeniedError,
context_from_snapshot,
evaluate_embeddings,
+ resolve_run_primary_engine,
)
- primary = (
- get_setting_from_snapshot(
- "search.tool",
- default=DEFAULT_SEARCH_TOOL,
- settings_snapshot=settings_snapshot,
- )
- or DEFAULT_SEARCH_TOOL
- )
try:
+ # Single source of truth for the primary (was: search.tool +
+ # searxng fallback, a fail-OPEN — a missing primary defaulted to
+ # the public searxng -> PUBLIC_ONLY -> require_local_embeddings
+ # stayed False -> the corpus could ship to a cloud embedder for a
+ # private run). A missing/invalid primary now raises -> fail closed.
+ primary = resolve_run_primary_engine(settings_snapshot)
ctx = context_from_snapshot(settings_snapshot, primary)
except PolicyDeniedError:
raise
diff --git a/tests/config/test_llm_config.py b/tests/config/test_llm_config.py
index 40c62d0ac8..7686c73150 100644
--- a/tests/config/test_llm_config.py
+++ b/tests/config/test_llm_config.py
@@ -192,7 +192,8 @@ class TestGetLlm:
return_value="custom_provider",
):
result = get_llm(
- provider="custom_provider", settings_snapshot={}
+ provider="custom_provider",
+ settings_snapshot={"search.tool": "searxng"},
)
assert result is mock_llm
@@ -214,7 +215,7 @@ class TestGetLlm:
}.get(key, default)
with pytest.raises(ValueError, match="Invalid provider"):
- get_llm(settings_snapshot={})
+ get_llm(settings_snapshot={"search.tool": "searxng"})
def test_raises_when_model_setting_empty(self):
"""get_llm() must raise ValueError when llm.model is empty string."""
@@ -312,7 +313,7 @@ class TestGetLlm:
model_name="custom-model",
temperature=0.5,
provider="custom_provider",
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
)
mock_factory.assert_called_once()
@@ -346,7 +347,10 @@ class TestGetLlm:
}.get(key, default)
with pytest.raises(TypeError, match="invalid signature"):
- get_llm(provider="bad_factory", settings_snapshot={})
+ get_llm(
+ provider="bad_factory",
+ settings_snapshot={"search.tool": "searxng"},
+ )
def test_custom_factory_returning_non_basechatmodel_raises(self):
"""Should raise ValueError when factory returns non-BaseChatModel."""
@@ -378,7 +382,10 @@ class TestGetLlm:
with pytest.raises(
ValueError, match="must return a BaseChatModel"
):
- get_llm(provider="bad_factory", settings_snapshot={})
+ get_llm(
+ provider="bad_factory",
+ settings_snapshot={"search.tool": "searxng"},
+ )
class TestWrapperStringResponse:
diff --git a/tests/config/test_llm_config_behavior.py b/tests/config/test_llm_config_behavior.py
index 5a8a75ae7c..c45043d821 100644
--- a/tests/config/test_llm_config_behavior.py
+++ b/tests/config/test_llm_config_behavior.py
@@ -30,7 +30,12 @@ class TestApiKeyAvailabilityChecks:
OpenAIProvider,
)
- assert OpenAIProvider.is_available(settings_snapshot={}) is False
+ assert (
+ OpenAIProvider.is_available(
+ settings_snapshot={"search.tool": "searxng"}
+ )
+ is False
+ )
def test_openai_unavailable_with_empty_api_key(self):
"""Empty string API key is falsy — treated as missing."""
@@ -58,7 +63,12 @@ class TestApiKeyAvailabilityChecks:
AnthropicProvider,
)
- assert AnthropicProvider.is_available(settings_snapshot={}) is False
+ assert (
+ AnthropicProvider.is_available(
+ settings_snapshot={"search.tool": "searxng"}
+ )
+ is False
+ )
def test_openai_endpoint_available_with_api_key(self):
from local_deep_research.llm.providers.implementations.custom_openai_endpoint import (
@@ -76,7 +86,9 @@ class TestApiKeyAvailabilityChecks:
)
assert (
- CustomOpenAIEndpointProvider.is_available(settings_snapshot={})
+ CustomOpenAIEndpointProvider.is_available(
+ settings_snapshot={"search.tool": "searxng"}
+ )
is False
)
@@ -89,7 +101,10 @@ class TestProviderValidation:
from local_deep_research.config.llm_config import get_llm
with pytest.raises(ValueError, match="Invalid provider"):
- get_llm(provider="nonexistent_provider", settings_snapshot={})
+ get_llm(
+ provider="nonexistent_provider",
+ settings_snapshot={"search.tool": "searxng"},
+ )
def test_error_message_lists_valid_providers(self):
"""ValueError message enumerates the auto-discovered providers.
@@ -104,7 +119,10 @@ class TestProviderValidation:
from local_deep_research.llm.providers.base import normalize_provider
with pytest.raises(ValueError) as exc_info:
- get_llm(provider="bad_provider", settings_snapshot={})
+ get_llm(
+ provider="bad_provider",
+ settings_snapshot={"search.tool": "searxng"},
+ )
message = str(exc_info.value)
for option in get_discovered_provider_options():
assert normalize_provider(option["value"]) in message
@@ -125,7 +143,11 @@ class TestModelAndProviderNameCleaning:
# 'none' is valid but has no implementation, so it raises ValueError
# The point is the quotes are stripped before validation
with pytest.raises(ValueError, match="No LLM provider configured"):
- get_llm(model_name="x", provider="'none'", settings_snapshot={})
+ get_llm(
+ model_name="x",
+ provider="'none'",
+ settings_snapshot={"search.tool": "searxng"},
+ )
def test_whitespace_provider_name_cleaned(self):
"""Surrounding whitespace is stripped: ' none ' → 'none'."""
@@ -133,7 +155,11 @@ class TestModelAndProviderNameCleaning:
from local_deep_research.config.llm_config import get_llm
with pytest.raises(ValueError, match="No LLM provider configured"):
- get_llm(model_name="x", provider=" none ", settings_snapshot={})
+ get_llm(
+ model_name="x",
+ provider=" none ",
+ settings_snapshot={"search.tool": "searxng"},
+ )
def test_uppercase_provider_lowercased(self):
"""Provider name is lowercased: 'NONE' → 'none'."""
@@ -141,7 +167,11 @@ class TestModelAndProviderNameCleaning:
from local_deep_research.config.llm_config import get_llm
with pytest.raises(ValueError, match="No LLM provider configured"):
- get_llm(model_name="x", provider="NONE", settings_snapshot={})
+ get_llm(
+ model_name="x",
+ provider="NONE",
+ settings_snapshot={"search.tool": "searxng"},
+ )
def test_combined_quotes_whitespace_case_cleaning(self):
"""Combined cleaning: \" 'None' \" → 'none'."""
@@ -149,7 +179,11 @@ class TestModelAndProviderNameCleaning:
from local_deep_research.config.llm_config import get_llm
with pytest.raises(ValueError, match="No LLM provider configured"):
- get_llm(model_name="x", provider=" 'None' ", settings_snapshot={})
+ get_llm(
+ model_name="x",
+ provider=" 'None' ",
+ settings_snapshot={"search.tool": "searxng"},
+ )
class TestGetSelectedLlmProvider:
@@ -172,5 +206,7 @@ class TestGetSelectedLlmProvider:
get_selected_llm_provider,
)
- result = get_selected_llm_provider(settings_snapshot={})
+ result = get_selected_llm_provider(
+ settings_snapshot={"search.tool": "searxng"}
+ )
assert result == "ollama"
diff --git a/tests/config/test_llm_config_deep_coverage.py b/tests/config/test_llm_config_deep_coverage.py
index 82212ec621..94c64b6ca6 100644
--- a/tests/config/test_llm_config_deep_coverage.py
+++ b/tests/config/test_llm_config_deep_coverage.py
@@ -63,7 +63,10 @@ class TestGetLlmRegisteredFactory:
from local_deep_research.config.llm_config import get_llm
- get_llm(provider="my_custom", settings_snapshot={})
+ get_llm(
+ provider="my_custom",
+ settings_snapshot={"search.tool": "searxng"},
+ )
mock_factory.assert_called_once()
mock_wrap.assert_called_once()
@@ -87,7 +90,10 @@ class TestGetLlmRegisteredFactory:
from local_deep_research.config.llm_config import get_llm
with pytest.raises(ValueError, match="must return a BaseChatModel"):
- get_llm(provider="my_custom", settings_snapshot={})
+ get_llm(
+ provider="my_custom",
+ settings_snapshot={"search.tool": "searxng"},
+ )
def test_factory_with_bad_signature_raises_type_error(self):
def bad_factory():
@@ -109,7 +115,10 @@ class TestGetLlmRegisteredFactory:
from local_deep_research.config.llm_config import get_llm
with pytest.raises(TypeError, match="invalid signature"):
- get_llm(provider="bad_factory", settings_snapshot={})
+ get_llm(
+ provider="bad_factory",
+ settings_snapshot={"search.tool": "searxng"},
+ )
def test_registered_basechatmodel_instance_used_directly(self):
mock_instance = MagicMock(spec=BaseChatModel)
@@ -132,7 +141,10 @@ class TestGetLlmRegisteredFactory:
from local_deep_research.config.llm_config import get_llm
- get_llm(provider="my_instance", settings_snapshot={})
+ get_llm(
+ provider="my_instance",
+ settings_snapshot={"search.tool": "searxng"},
+ )
# The factory path is skipped for BaseChatModel instances
mock_wrap.assert_called_once()
@@ -161,7 +173,10 @@ class TestGetLlmRegisteredFactory:
with pytest.raises(
ValueError, match="must be either a BaseChatModel"
):
- get_llm(provider="weird", settings_snapshot={})
+ get_llm(
+ provider="weird",
+ settings_snapshot={"search.tool": "searxng"},
+ )
# ---------------------------------------------------------------------------
@@ -189,7 +204,10 @@ class TestGetLlmInvalidProvider:
from local_deep_research.config.llm_config import get_llm
with pytest.raises(ValueError, match="Invalid provider"):
- get_llm(provider="bogus_provider", settings_snapshot={})
+ get_llm(
+ provider="bogus_provider",
+ settings_snapshot={"search.tool": "searxng"},
+ )
# ---------------------------------------------------------------------------
@@ -319,7 +337,7 @@ class TestGetLlmNameCleaning:
try:
from local_deep_research.config.llm_config import get_llm
- get_llm(settings_snapshot={})
+ get_llm(settings_snapshot={"search.tool": "searxng"})
if mock_init.call_args:
captured["model"] = mock_init.call_args.kwargs.get(
"model",
diff --git a/tests/config/test_llm_config_missing_coverage.py b/tests/config/test_llm_config_missing_coverage.py
index 390a784d92..5e48c25caa 100644
--- a/tests/config/test_llm_config_missing_coverage.py
+++ b/tests/config/test_llm_config_missing_coverage.py
@@ -26,6 +26,7 @@ THREAD_SETTINGS = "local_deep_research.config.thread_settings"
def _settings_dict(overrides=None):
"""Base settings dict with sensible defaults for most tests."""
base = {
+ "search.tool": "searxng",
"llm.model": "test-model",
"llm.temperature": 0.5,
"llm.provider": "ollama",
diff --git a/tests/embeddings/test_embeddings_config_deep_coverage.py b/tests/embeddings/test_embeddings_config_deep_coverage.py
index 4074a239cc..1086fca1b6 100644
--- a/tests/embeddings/test_embeddings_config_deep_coverage.py
+++ b/tests/embeddings/test_embeddings_config_deep_coverage.py
@@ -65,7 +65,10 @@ class TestGetEmbeddings:
}
get_embeddings(
- settings_snapshot={"embeddings.provider": "ollama"}
+ settings_snapshot={
+ "embeddings.provider": "ollama",
+ "search.tool": "searxng",
+ }
)
mock_provider.create_embeddings.assert_called_once()
diff --git a/tests/security/test_egress_pep_llm_embeddings_callsite.py b/tests/security/test_egress_pep_llm_embeddings_callsite.py
index 3e1286917a..1347e70ee7 100644
--- a/tests/security/test_egress_pep_llm_embeddings_callsite.py
+++ b/tests/security/test_egress_pep_llm_embeddings_callsite.py
@@ -49,14 +49,20 @@ from local_deep_research.security.egress.policy import PolicyDeniedError
# routes the call sites must both honor:
# - PRIVATE_ONLY scope: context_from_snapshot IMPLIES require_local_*.
# - the explicit per-knob flag with an otherwise-permissive scope.
-PRIVATE_ONLY = {"policy.egress_scope": "private_only"}
+# All include a "search.tool": a real run always has a configured primary, and
+# the inference PEPs now resolve it via resolve_run_primary_engine (fail closed
+# on a missing primary). The explicit scope still drives the decision — the
+# primary value here is irrelevant to it.
+PRIVATE_ONLY = {"policy.egress_scope": "private_only", "search.tool": "searxng"}
LLM_REQUIRE_LOCAL_FLAG = {
"policy.egress_scope": "both",
"llm.require_local_endpoint": True,
+ "search.tool": "searxng",
}
EMB_REQUIRE_LOCAL_FLAG = {
"policy.egress_scope": "both",
"embeddings.require_local": True,
+ "search.tool": "searxng",
}
@@ -130,8 +136,9 @@ class TestGetLlmCloudBlocked:
pairing the deny tests could pass for an unrelated reason."""
snapshot = {
"policy.egress_scope": "both",
- "llm.openai.api_key": "k",
- } # gitleaks:allow
+ "llm.openai.api_key": "k", # gitleaks:allow
+ "search.tool": "searxng",
+ }
with patch(
"local_deep_research.config.llm_config.get_llm_from_registry"
) as mock_registry:
@@ -310,6 +317,7 @@ class TestGetEmbeddingsCloudBlocked:
settings_snapshot={
"policy.egress_scope": "both",
"embeddings.require_local": False,
+ "search.tool": "searxng",
},
)
assert result is not None
@@ -416,3 +424,109 @@ class TestGetEmbeddingsSnapshotless:
)
assert result is not None
mock_create.assert_called_once()
+
+
+# ---------------------------------------------------------------------------
+# Inference-path fail-OPEN closed: a snapshot with NO primary must not default
+# to the public searxng (-> PUBLIC_ONLY -> require_local off) and admit a cloud
+# provider for a run whose actual posture is private. The PEPs resolve the
+# primary via resolve_run_primary_engine (no default) and fail closed.
+# ---------------------------------------------------------------------------
+
+
+class TestInferencePepsFailClosedOnMissingPrimary:
+ def test_llm_refuses_cloud_provider_when_no_primary(self):
+ # A present cloud key but NO search.tool (and no explicit scope =>
+ # ADAPTIVE). Pre-fix: searxng default -> PUBLIC_ONLY -> require_local
+ # off -> openai ADMITTED. Post-fix: fail closed before construction.
+ snapshot = {"llm.openai.api_key": "k"} # gitleaks:allow
+ with patch(
+ "local_deep_research.config.llm_config.get_llm_from_registry"
+ ) as mock_registry:
+ mock_registry.return_value = Mock(spec=BaseChatModel)
+ with pytest.raises(PolicyDeniedError) as exc:
+ get_llm(
+ provider="openai",
+ model_name="gpt-4o",
+ settings_snapshot=snapshot,
+ )
+ assert exc.value.decision.reason == "invalid_policy_config"
+ mock_registry.assert_not_called()
+
+ def test_embeddings_refuses_cloud_provider_when_no_primary(self):
+ with patch.object(
+ OpenAIEmbeddingsProvider,
+ "create_embeddings",
+ return_value=Mock(spec=Embeddings),
+ ) as mock_create:
+ with pytest.raises(PolicyDeniedError) as exc:
+ get_embeddings(
+ provider="openai",
+ settings_snapshot={"embeddings.require_local": False},
+ )
+ assert exc.value.decision.reason == "invalid_policy_config"
+ mock_create.assert_not_called()
+
+
+# ---------------------------------------------------------------------------
+# The PR's HEADLINE behavior (not just the missing-primary guardrail): with a
+# PRESENT primary under the default ADAPTIVE scope, the inference PEPs now
+# resolve the SAME scope the factory does. A private primary ("library") forces
+# PRIVATE_ONLY -> a cloud provider is denied (provider_cloud_only) — exactly the
+# exfil scenario from the PR body (search_tool="library"). A public primary
+# ("searxng") resolves PUBLIC_ONLY -> the same cloud provider is admitted. The
+# "library" key classifies private without any DB/username lookup.
+# ---------------------------------------------------------------------------
+
+
+class TestInferencePepFollowsAdaptivePrimary:
+ def test_llm_private_primary_denies_cloud_under_adaptive(self):
+ snapshot = {
+ "search.tool": "library", # private primary, scope defaults ADAPTIVE
+ "llm.openai.api_key": "k", # gitleaks:allow
+ }
+ with patch(
+ "local_deep_research.config.llm_config.get_llm_from_registry"
+ ) as mock_registry:
+ mock_registry.return_value = Mock(spec=BaseChatModel)
+ with pytest.raises(PolicyDeniedError) as exc:
+ get_llm(
+ provider="openai",
+ model_name="gpt-4o",
+ settings_snapshot=snapshot,
+ )
+ assert exc.value.decision.reason == "provider_cloud_only"
+ mock_registry.assert_not_called()
+
+ def test_llm_public_primary_admits_cloud_under_adaptive(self):
+ # Mirror: a public primary under ADAPTIVE => PUBLIC_ONLY => admitted.
+ snapshot = {
+ "search.tool": "searxng",
+ "llm.openai.api_key": "k", # gitleaks:allow
+ }
+ with patch(
+ "local_deep_research.config.llm_config.get_llm_from_registry"
+ ) as mock_registry:
+ mock_registry.return_value = Mock(spec=BaseChatModel)
+ result = get_llm(
+ provider="openai",
+ model_name="gpt-4o",
+ settings_snapshot=snapshot,
+ )
+ assert result is not None
+ mock_registry.assert_called_once()
+
+ def test_embeddings_private_primary_denies_cloud_under_adaptive(self):
+ snapshot = {"search.tool": "library"} # private primary, ADAPTIVE
+ with patch.object(
+ OpenAIEmbeddingsProvider,
+ "create_embeddings",
+ return_value=Mock(spec=Embeddings),
+ ) as mock_create:
+ with pytest.raises(PolicyDeniedError) as exc:
+ get_embeddings(provider="openai", settings_snapshot=snapshot)
+ assert exc.value.decision.reason in (
+ "provider_cloud_only",
+ "provider_cloud",
+ )
+ mock_create.assert_not_called()
diff --git a/tests/test_api_key_configuration.py b/tests/test_api_key_configuration.py
index 3d536aee0f..90250c63f4 100644
--- a/tests/test_api_key_configuration.py
+++ b/tests/test_api_key_configuration.py
@@ -71,6 +71,7 @@ class TestAPIKeyConfiguration:
def _get_base_settings(self):
"""Get base settings that all tests need."""
return {
+ "search.tool": "searxng",
"llm.supports_max_tokens": True,
"llm.max_tokens": 4096,
"llm.context_window_unrestricted": False,
@@ -429,6 +430,7 @@ class TestLLMIntegration:
def _get_base_settings(self):
"""Get base settings that all tests need."""
return {
+ "search.tool": "searxng",
"llm.supports_max_tokens": True,
"llm.max_tokens": 4096,
"llm.context_window_unrestricted": False,
diff --git a/tests/test_llm/test_llm_benchmarks.py b/tests/test_llm/test_llm_benchmarks.py
index 678adc2e35..7e4f8055d0 100644
--- a/tests/test_llm/test_llm_benchmarks.py
+++ b/tests/test_llm/test_llm_benchmarks.py
@@ -210,7 +210,7 @@ def test_benchmark_with_custom_llm_factory():
provider="benchmark_factory",
model_name="accurate",
temperature=0.1,
- settings_snapshot={},
+ settings_snapshot={"search.tool": "searxng"},
)
# Should create the accurate version
diff --git a/tests/test_llm/test_llm_edge_cases.py b/tests/test_llm/test_llm_edge_cases.py
index e818d7129d..8c49647bec 100644
--- a/tests/test_llm/test_llm_edge_cases.py
+++ b/tests/test_llm/test_llm_edge_cases.py
@@ -161,6 +161,7 @@ def full_settings_snapshot():
return {"value": value, "ui_element": "checkbox"}
return {
+ "search.tool": _text_param("searxng"),
"llm.model": _text_param("test-model"),
"llm.temperature": _number_param(0.7),
"llm.provider": _text_param("test"),
diff --git a/tests/test_llm/test_llm_integration.py b/tests/test_llm/test_llm_integration.py
index b998a48b99..671a9f36fd 100644
--- a/tests/test_llm/test_llm_integration.py
+++ b/tests/test_llm/test_llm_integration.py
@@ -56,6 +56,9 @@ def full_settings_snapshot():
return {"value": value, "ui_element": "checkbox"}
return {
+ # Real runs always carry a primary engine; the inference
+ # egress PEP now fails closed without it.
+ "search.tool": _text_param("searxng"),
"llm.model": _text_param("test-model"),
"llm.temperature": _number_param(0.7),
"llm.provider": _text_param("test"),
diff --git a/tests/test_openai_api_key_usage.py b/tests/test_openai_api_key_usage.py
index 088432c23d..401ad03c06 100644
--- a/tests/test_openai_api_key_usage.py
+++ b/tests/test_openai_api_key_usage.py
@@ -26,6 +26,7 @@ class TestOpenAIAPIKeyUsage:
type coercion in get_setting_from_snapshot().
"""
return {
+ "search.tool": "searxng",
"llm.provider": "openai",
"llm.model": "gpt-3.5-turbo",
"llm.temperature": 0.7,
diff --git a/tests/web/services/test_research_service_core.py b/tests/web/services/test_research_service_core.py
index 4691187b4a..6bd3e47f1c 100644
--- a/tests/web/services/test_research_service_core.py
+++ b/tests/web/services/test_research_service_core.py
@@ -323,7 +323,10 @@ class TestLLMInstantiation:
# which fails closed for snapshot-less non-local providers — lets the
# call reach the provider-name validation this test exercises.
with pytest.raises(ValueError) as exc_info:
- get_llm(provider="invalid_provider", settings_snapshot={})
+ get_llm(
+ provider="invalid_provider",
+ settings_snapshot={"search.tool": "searxng"},
+ )
assert "Invalid provider" in str(exc_info.value)
diff --git a/tests/web_search_engines/engines/test_search_engine_semantic_scholar.py b/tests/web_search_engines/engines/test_search_engine_semantic_scholar.py
index d0b0b34088..1d87a0a695 100644
--- a/tests/web_search_engines/engines/test_search_engine_semantic_scholar.py
+++ b/tests/web_search_engines/engines/test_search_engine_semantic_scholar.py
@@ -62,7 +62,9 @@ class TestSemanticScholarSearchEngineInit:
return_value="settings-api-key",
):
engine = SemanticScholarSearchEngine(
- settings_snapshot={"key": "value"}
+ # search.tool present so the inference PEP (now fail-closed
+ # on a missing primary) doesn't raise during engine init.
+ settings_snapshot={"key": "value", "search.tool": "searxng"}
)
assert engine.api_key == "settings-api-key"
diff --git a/tests/web_search_engines/test_search_engine_factory_branch_coverage.py b/tests/web_search_engines/test_search_engine_factory_branch_coverage.py
index 569097df41..2861a00a34 100644
--- a/tests/web_search_engines/test_search_engine_factory_branch_coverage.py
+++ b/tests/web_search_engines/test_search_engine_factory_branch_coverage.py
@@ -488,6 +488,10 @@ class TestGetLlmAnthropicMissingKey:
def test_anthropic_no_api_key_raises(self):
settings = {
+ # A real run always carries a primary; the inference PEP now fails
+ # closed without it (resolve_run_primary_engine), which would raise
+ # before this test reaches the Anthropic-key check.
+ "search.tool": "searxng",
"llm.model": "claude-3-haiku",
"llm.temperature": 0.7,
"llm.provider": "anthropic",